The i915_request fence wait behaves differently for timeout = 0 compared to expected dma-fence behavior.
i915 behavior: - Unsignaled: -ETIME - Signaled: 0 (= timeout)
Expected: - Unsignaled: 0 - Signaled: 1
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com --- drivers/gpu/drm/i915/i915_request.c | 57 ++++++++++++++++++++++++----- drivers/gpu/drm/i915/i915_request.h | 5 +++ 2 files changed, 52 insertions(+), 10 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_request.c b/drivers/gpu/drm/i915/i915_request.c index 820a1f38b271..42cd17357771 100644 --- a/drivers/gpu/drm/i915/i915_request.c +++ b/drivers/gpu/drm/i915/i915_request.c @@ -96,9 +96,9 @@ static signed long i915_fence_wait(struct dma_fence *fence, bool interruptible, signed long timeout) { - return i915_request_wait(to_request(fence), - interruptible | I915_WAIT_PRIORITY, - timeout); + return i915_request_wait_timeout(to_request(fence), + interruptible | I915_WAIT_PRIORITY, + timeout); }
struct kmem_cache *i915_request_slab_cache(void) @@ -1857,23 +1857,27 @@ static void request_wait_wake(struct dma_fence *fence, struct dma_fence_cb *cb) }
/** - * i915_request_wait - wait until execution of request has finished + * i915_request_wait_timeout - wait until execution of request has finished * @rq: the request to wait upon * @flags: how to wait * @timeout: how long to wait in jiffies * - * i915_request_wait() waits for the request to be completed, for a + * i915_request_wait_timeout() waits for the request to be completed, for a * maximum of @timeout jiffies (with MAX_SCHEDULE_TIMEOUT implying an * unbounded wait). * * Returns the remaining time (in jiffies) if the request completed, which may - * be zero or -ETIME if the request is unfinished after the timeout expires. + * be zero if the request is unfinished after the timeout expires. + * If the timeout is 0, it will return 1 if the fence is signaled. + * * May return -EINTR is called with I915_WAIT_INTERRUPTIBLE and a signal is * pending before the request completes. + * + * NOTE: This function has the same wait semantics as dma-fence. */ -long i915_request_wait(struct i915_request *rq, - unsigned int flags, - long timeout) +long i915_request_wait_timeout(struct i915_request *rq, + unsigned int flags, + long timeout) { const int state = flags & I915_WAIT_INTERRUPTIBLE ? TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE; @@ -1883,7 +1887,7 @@ long i915_request_wait(struct i915_request *rq, GEM_BUG_ON(timeout < 0);
if (dma_fence_is_signaled(&rq->fence)) - return timeout; + return timeout ?: 1;
if (!timeout) return -ETIME; @@ -1992,6 +1996,39 @@ long i915_request_wait(struct i915_request *rq, return timeout; }
+/** + * i915_request_wait - wait until execution of request has finished + * @rq: the request to wait upon + * @flags: how to wait + * @timeout: how long to wait in jiffies + * + * i915_request_wait() waits for the request to be completed, for a + * maximum of @timeout jiffies (with MAX_SCHEDULE_TIMEOUT implying an + * unbounded wait). + * + * Returns the remaining time (in jiffies) if the request completed, which may + * be zero or -ETIME if the request is unfinished after the timeout expires. + * May return -EINTR is called with I915_WAIT_INTERRUPTIBLE and a signal is + * pending before the request completes. + * + * NOTE: This function behaves differently from dma-fence wait semantics for + * timeout = 0. It returns 0 on success, and -ETIME if not signaled. + */ +long i915_request_wait(struct i915_request *rq, + unsigned int flags, + long timeout) +{ + long ret = i915_request_wait_timeout(rq, flags, timeout); + + if (!ret) + return -ETIME; + + if (ret > 0 && !timeout) + return 0; + + return ret; +} + static int print_sched_attr(const struct i915_sched_attr *attr, char *buf, int x, int len) { diff --git a/drivers/gpu/drm/i915/i915_request.h b/drivers/gpu/drm/i915/i915_request.h index dc359242d1ae..3c6e8acd1457 100644 --- a/drivers/gpu/drm/i915/i915_request.h +++ b/drivers/gpu/drm/i915/i915_request.h @@ -414,6 +414,11 @@ void i915_request_unsubmit(struct i915_request *request);
void i915_request_cancel(struct i915_request *rq, int error);
+long i915_request_wait_timeout(struct i915_request *rq, + unsigned int flags, + long timeout) + __attribute__((nonnull(1))); + long i915_request_wait(struct i915_request *rq, unsigned int flags, long timeout)
From: Christian König christian.koenig@amd.com
Simplifying the code a bit.
Signed-off-by: Christian König christian.koenig@amd.com [mlankhorst: Handle timeout = 0 correctly, use new i915_request_wait_timeout.] Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com --- drivers/gpu/drm/i915/gem/i915_gem_wait.c | 65 ++++++++---------------- 1 file changed, 20 insertions(+), 45 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_wait.c b/drivers/gpu/drm/i915/gem/i915_gem_wait.c index f909aaa09d9c..840c13706999 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_wait.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_wait.c @@ -25,7 +25,7 @@ i915_gem_object_wait_fence(struct dma_fence *fence, return timeout;
if (dma_fence_is_i915(fence)) - return i915_request_wait(to_request(fence), flags, timeout); + return i915_request_wait_timeout(to_request(fence), flags, timeout);
return dma_fence_wait_timeout(fence, flags & I915_WAIT_INTERRUPTIBLE, @@ -37,58 +37,29 @@ i915_gem_object_wait_reservation(struct dma_resv *resv, unsigned int flags, long timeout) { - struct dma_fence *excl; - bool prune_fences = false; - - if (flags & I915_WAIT_ALL) { - struct dma_fence **shared; - unsigned int count, i; - int ret; - - ret = dma_resv_get_fences(resv, &excl, &count, &shared); - if (ret) - return ret; - - for (i = 0; i < count; i++) { - timeout = i915_gem_object_wait_fence(shared[i], - flags, timeout); - if (timeout < 0) - break; - - dma_fence_put(shared[i]); - } - - for (; i < count; i++) - dma_fence_put(shared[i]); - kfree(shared); + struct dma_resv_iter cursor; + struct dma_fence *fence; + long ret = timeout ?: 1; + + dma_resv_iter_begin(&cursor, resv, flags & I915_WAIT_ALL); + dma_resv_for_each_fence_unlocked(&cursor, fence) { + ret = i915_gem_object_wait_fence(fence, flags, timeout); + if (ret <= 0) + break;
- /* - * If both shared fences and an exclusive fence exist, - * then by construction the shared fences must be later - * than the exclusive fence. If we successfully wait for - * all the shared fences, we know that the exclusive fence - * must all be signaled. If all the shared fences are - * signaled, we can prune the array and recover the - * floating references on the fences/requests. - */ - prune_fences = count && timeout >= 0; - } else { - excl = dma_resv_get_excl_unlocked(resv); + if (timeout) + timeout = ret; } - - if (excl && timeout >= 0) - timeout = i915_gem_object_wait_fence(excl, flags, timeout); - - dma_fence_put(excl); + dma_resv_iter_end(&cursor);
/* * Opportunistically prune the fences iff we know they have *all* been * signaled. */ - if (prune_fences) + if (timeout > 0) dma_resv_prune(resv);
- return timeout; + return ret; }
static void fence_set_priority(struct dma_fence *fence, @@ -196,7 +167,11 @@ i915_gem_object_wait(struct drm_i915_gem_object *obj,
timeout = i915_gem_object_wait_reservation(obj->base.resv, flags, timeout); - return timeout < 0 ? timeout : 0; + + if (timeout < 0) + return timeout; + + return !timeout ? -ETIME : 0; }
static inline unsigned long nsecs_to_jiffies_timeout(const u64 n)
Am 21.10.21 um 12:35 schrieb Maarten Lankhorst:
From: Christian König christian.koenig@amd.com
Simplifying the code a bit.
Signed-off-by: Christian König christian.koenig@amd.com [mlankhorst: Handle timeout = 0 correctly, use new i915_request_wait_timeout.] Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
LGTM, do you want to push it or should I pick it up into drm-misc-next?
Christian.
drivers/gpu/drm/i915/gem/i915_gem_wait.c | 65 ++++++++---------------- 1 file changed, 20 insertions(+), 45 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_wait.c b/drivers/gpu/drm/i915/gem/i915_gem_wait.c index f909aaa09d9c..840c13706999 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_wait.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_wait.c @@ -25,7 +25,7 @@ i915_gem_object_wait_fence(struct dma_fence *fence, return timeout;
if (dma_fence_is_i915(fence))
return i915_request_wait(to_request(fence), flags, timeout);
return i915_request_wait_timeout(to_request(fence), flags, timeout);return dma_fence_wait_timeout(fence, flags & I915_WAIT_INTERRUPTIBLE,
@@ -37,58 +37,29 @@ i915_gem_object_wait_reservation(struct dma_resv *resv, unsigned int flags, long timeout) {
- struct dma_fence *excl;
- bool prune_fences = false;
- if (flags & I915_WAIT_ALL) {
struct dma_fence **shared;unsigned int count, i;int ret;ret = dma_resv_get_fences(resv, &excl, &count, &shared);if (ret)return ret;for (i = 0; i < count; i++) {timeout = i915_gem_object_wait_fence(shared[i],flags, timeout);if (timeout < 0)break;dma_fence_put(shared[i]);}for (; i < count; i++)dma_fence_put(shared[i]);kfree(shared);
- struct dma_resv_iter cursor;
- struct dma_fence *fence;
- long ret = timeout ?: 1;
- dma_resv_iter_begin(&cursor, resv, flags & I915_WAIT_ALL);
- dma_resv_for_each_fence_unlocked(&cursor, fence) {
ret = i915_gem_object_wait_fence(fence, flags, timeout);if (ret <= 0)break;
/** If both shared fences and an exclusive fence exist,* then by construction the shared fences must be later* than the exclusive fence. If we successfully wait for* all the shared fences, we know that the exclusive fence* must all be signaled. If all the shared fences are* signaled, we can prune the array and recover the* floating references on the fences/requests.*/prune_fences = count && timeout >= 0;- } else {
excl = dma_resv_get_excl_unlocked(resv);
if (timeout)timeout = ret;
- if (excl && timeout >= 0)
timeout = i915_gem_object_wait_fence(excl, flags, timeout);- dma_fence_put(excl);
dma_resv_iter_end(&cursor);
/*
- Opportunistically prune the fences iff we know they have *all* been
- signaled.
*/
- if (prune_fences)
- if (timeout > 0) dma_resv_prune(resv);
- return timeout;
return ret; }
static void fence_set_priority(struct dma_fence *fence,
@@ -196,7 +167,11 @@ i915_gem_object_wait(struct drm_i915_gem_object *obj,
timeout = i915_gem_object_wait_reservation(obj->base.resv, flags, timeout);
- return timeout < 0 ? timeout : 0;
if (timeout < 0)
return timeout;return !timeout ? -ETIME : 0; }
static inline unsigned long nsecs_to_jiffies_timeout(const u64 n)
Op 21-10-2021 om 12:38 schreef Christian König:
Am 21.10.21 um 12:35 schrieb Maarten Lankhorst:
From: Christian König christian.koenig@amd.com
Simplifying the code a bit.
Signed-off-by: Christian König christian.koenig@amd.com [mlankhorst: Handle timeout = 0 correctly, use new i915_request_wait_timeout.] Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
LGTM, do you want to push it or should I pick it up into drm-misc-next?
I think it can be applied to drm-intel-gt-next, after a backmerge. It needs patch 1 too, which fixes
i915_request_wait semantics when used in dma-fence. It exports a dma-fence compatible i915_request_wait_timeout function, used in this patch.
On 21/10/2021 12:06, Maarten Lankhorst wrote:
Op 21-10-2021 om 12:38 schreef Christian König:
Am 21.10.21 um 12:35 schrieb Maarten Lankhorst:
From: Christian König christian.koenig@amd.com
Simplifying the code a bit.
Signed-off-by: Christian König christian.koenig@amd.com [mlankhorst: Handle timeout = 0 correctly, use new i915_request_wait_timeout.] Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
LGTM, do you want to push it or should I pick it up into drm-misc-next?
I think it can be applied to drm-intel-gt-next, after a backmerge. It needs patch 1 too, which fixes
i915_request_wait semantics when used in dma-fence. It exports a dma-fence compatible i915_request_wait_timeout function, used in this patch.
I don't think my open has been resolved, at least I haven't seen a reply from Daniel on the topic of potential for infinite waits with untrusted clients after this change. +Daniel
Regards,
Tvrtko
Am 21.10.21 um 13:13 schrieb Tvrtko Ursulin:
On 21/10/2021 12:06, Maarten Lankhorst wrote:
Op 21-10-2021 om 12:38 schreef Christian König:
Am 21.10.21 um 12:35 schrieb Maarten Lankhorst:
From: Christian König christian.koenig@amd.com
Simplifying the code a bit.
Signed-off-by: Christian König christian.koenig@amd.com [mlankhorst: Handle timeout = 0 correctly, use new i915_request_wait_timeout.] Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
LGTM, do you want to push it or should I pick it up into drm-misc-next?
I think it can be applied to drm-intel-gt-next, after a backmerge. It needs patch 1 too, which fixes
i915_request_wait semantics when used in dma-fence. It exports a dma-fence compatible i915_request_wait_timeout function, used in this patch.
What about the other i915 patches? I guess you then want to merge them through drm-intel-gt-next as well.
I don't think my open has been resolved, at least I haven't seen a reply from Daniel on the topic of potential for infinite waits with untrusted clients after this change. +Daniel
Please resolve that internally and let me know the result. I'm fine to use any of the possible approaches, I just need to know which one.
Regards, Christian.
Regards,
Tvrtko
On Thu, Oct 28, 2021 at 10:41:38AM +0200, Christian König wrote:
Am 21.10.21 um 13:13 schrieb Tvrtko Ursulin:
On 21/10/2021 12:06, Maarten Lankhorst wrote:
Op 21-10-2021 om 12:38 schreef Christian König:
Am 21.10.21 um 12:35 schrieb Maarten Lankhorst:
From: Christian König christian.koenig@amd.com
Simplifying the code a bit.
Signed-off-by: Christian König christian.koenig@amd.com [mlankhorst: Handle timeout = 0 correctly, use new i915_request_wait_timeout.] Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
LGTM, do you want to push it or should I pick it up into drm-misc-next?
I think it can be applied to drm-intel-gt-next, after a backmerge. It needs patch 1 too, which fixes
i915_request_wait semantics when used in dma-fence. It exports a dma-fence compatible i915_request_wait_timeout function, used in this patch.
What about the other i915 patches? I guess you then want to merge them through drm-intel-gt-next as well.
I don't think my open has been resolved, at least I haven't seen a reply from Daniel on the topic of potential for infinite waits with untrusted clients after this change. +Daniel
Please resolve that internally and let me know the result. I'm fine to use any of the possible approaches, I just need to know which one.
I thought I explained this in the patch set from Maarten. This isn't an issue, since the exact same thing can happen if you get interrupts and stuff.
The only proper fix for bounding the waits is a) compositor grabs a stable set of dma_fence from the dma-buf through the proposed fence export ioctl b) compositor waits on that fence (or drm_syncobj).
Everything else is cargo-culted nonsense, and very much includes that igt patch that's floating around internally.
I can also whack this into drm-next if this is stuck in this silly bikeshed. -Daniel
On 28/10/2021 16:30, Daniel Vetter wrote:
On Thu, Oct 28, 2021 at 10:41:38AM +0200, Christian König wrote:
Am 21.10.21 um 13:13 schrieb Tvrtko Ursulin:
On 21/10/2021 12:06, Maarten Lankhorst wrote:
Op 21-10-2021 om 12:38 schreef Christian König:
Am 21.10.21 um 12:35 schrieb Maarten Lankhorst:
From: Christian König christian.koenig@amd.com
Simplifying the code a bit.
Signed-off-by: Christian König christian.koenig@amd.com [mlankhorst: Handle timeout = 0 correctly, use new i915_request_wait_timeout.] Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
LGTM, do you want to push it or should I pick it up into drm-misc-next?
I think it can be applied to drm-intel-gt-next, after a backmerge. It needs patch 1 too, which fixes
i915_request_wait semantics when used in dma-fence. It exports a dma-fence compatible i915_request_wait_timeout function, used in this patch.
What about the other i915 patches? I guess you then want to merge them through drm-intel-gt-next as well.
I don't think my open has been resolved, at least I haven't seen a reply from Daniel on the topic of potential for infinite waits with untrusted clients after this change. +Daniel
Please resolve that internally and let me know the result. I'm fine to use any of the possible approaches, I just need to know which one.
I thought I explained this in the patch set from Maarten. This isn't an issue, since the exact same thing can happen if you get interrupts and stuff.
Ah were you trying to point out all this time the infinite wait just got moved from inside the "old" dma_resv_get_fences to the new iterator caller?
Regards,
Tvrtko
The only proper fix for bounding the waits is a) compositor grabs a stable set of dma_fence from the dma-buf through the proposed fence export ioctl b) compositor waits on that fence (or drm_syncobj).
Everything else is cargo-culted nonsense, and very much includes that igt patch that's floating around internally.
I can also whack this into drm-next if this is stuck in this silly bikeshed. -Daniel
Am 01.11.21 um 10:41 schrieb Tvrtko Ursulin:
On 28/10/2021 16:30, Daniel Vetter wrote:
On Thu, Oct 28, 2021 at 10:41:38AM +0200, Christian König wrote:
Am 21.10.21 um 13:13 schrieb Tvrtko Ursulin:
On 21/10/2021 12:06, Maarten Lankhorst wrote:
Op 21-10-2021 om 12:38 schreef Christian König:
Am 21.10.21 um 12:35 schrieb Maarten Lankhorst: > From: Christian König christian.koenig@amd.com > > Simplifying the code a bit. > > Signed-off-by: Christian König christian.koenig@amd.com > [mlankhorst: Handle timeout = 0 correctly, use new > i915_request_wait_timeout.] > Signed-off-by: Maarten Lankhorst > maarten.lankhorst@linux.intel.com
LGTM, do you want to push it or should I pick it up into drm-misc-next?
I think it can be applied to drm-intel-gt-next, after a backmerge. It needs patch 1 too, which fixes
i915_request_wait semantics when used in dma-fence. It exports a dma-fence compatible i915_request_wait_timeout function, used in this patch.
What about the other i915 patches? I guess you then want to merge them through drm-intel-gt-next as well.
I don't think my open has been resolved, at least I haven't seen a reply from Daniel on the topic of potential for infinite waits with untrusted clients after this change. +Daniel
Please resolve that internally and let me know the result. I'm fine to use any of the possible approaches, I just need to know which one.
I thought I explained this in the patch set from Maarten. This isn't an issue, since the exact same thing can happen if you get interrupts and stuff.
Ah were you trying to point out all this time the infinite wait just got moved from inside the "old" dma_resv_get_fences to the new iterator caller?
At least I think so, yes. But Daniel needs to answer that.
Regards,
Tvrtko
The only proper fix for bounding the waits is a) compositor grabs a stable set of dma_fence from the dma-buf through the proposed fence export ioctl b) compositor waits on that fence (or drm_syncobj).
Everything else is cargo-culted nonsense, and very much includes that igt patch that's floating around internally.
I can also whack this into drm-next if this is stuck in this silly bikeshed.
Can we move forward with those patches? I still don't see them in drm-misc-next.
I you want I can also pick Maartens modified version here up and send it out with the reset of the i915 patches once more.
Everything else is already pushed.
Thanks, Christian.
-Daniel
On Thu, Nov 11, 2021 at 12:36:47PM +0100, Christian König wrote:
Am 01.11.21 um 10:41 schrieb Tvrtko Ursulin:
On 28/10/2021 16:30, Daniel Vetter wrote:
On Thu, Oct 28, 2021 at 10:41:38AM +0200, Christian König wrote:
Am 21.10.21 um 13:13 schrieb Tvrtko Ursulin:
On 21/10/2021 12:06, Maarten Lankhorst wrote:
Op 21-10-2021 om 12:38 schreef Christian König: > Am 21.10.21 um 12:35 schrieb Maarten Lankhorst: > > From: Christian König christian.koenig@amd.com > > > > Simplifying the code a bit. > > > > Signed-off-by: Christian König christian.koenig@amd.com > > [mlankhorst: Handle timeout = 0 correctly, use new > > i915_request_wait_timeout.] > > Signed-off-by: Maarten Lankhorst > > maarten.lankhorst@linux.intel.com > > LGTM, do you want to push it or should I pick it up > into drm-misc-next?
I think it can be applied to drm-intel-gt-next, after a backmerge. It needs patch 1 too, which fixes
i915_request_wait semantics when used in dma-fence. It exports a dma-fence compatible i915_request_wait_timeout function, used in this patch.
What about the other i915 patches? I guess you then want to merge them through drm-intel-gt-next as well.
I don't think my open has been resolved, at least I haven't seen a reply from Daniel on the topic of potential for infinite waits with untrusted clients after this change. +Daniel
Please resolve that internally and let me know the result. I'm fine to use any of the possible approaches, I just need to know which one.
I thought I explained this in the patch set from Maarten. This isn't an issue, since the exact same thing can happen if you get interrupts and stuff.
Ah were you trying to point out all this time the infinite wait just got moved from inside the "old" dma_resv_get_fences to the new iterator caller?
At least I think so, yes. But Daniel needs to answer that.
Well maybe there's also an infinite wait inside the old dma_resv_get_fences, what I mean is that when you have some signals interrupting you, then the infinite wait is already there due to a retry loop outside of the syscall even.
Anyway _any_ userspace which wants to use this wait on a shared bo and waits to be safe against the other end adding more rendering has to use something else (like the sync_file export ioctl on the dma-buf that Jason typed). Trying to make this ioctl here against fence addition is just bs.
Regards,
Tvrtko
The only proper fix for bounding the waits is a) compositor grabs a stable set of dma_fence from the dma-buf through the proposed fence export ioctl b) compositor waits on that fence (or drm_syncobj).
Everything else is cargo-culted nonsense, and very much includes that igt patch that's floating around internally.
I can also whack this into drm-next if this is stuck in this silly bikeshed.
Can we move forward with those patches? I still don't see them in drm-misc-next.
I you want I can also pick Maartens modified version here up and send it out with the reset of the i915 patches once more.
Everything else is already pushed.
Please push to drm-misc-next or wherever (assuming CI is happy) and feel free to add my ack. -Daniel
The signaled bit is already used for quick testing if a fence is signaled.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com --- drivers/gpu/drm/i915/Makefile | 1 - drivers/gpu/drm/i915/dma_resv_utils.c | 17 ----------------- drivers/gpu/drm/i915/dma_resv_utils.h | 13 ------------- drivers/gpu/drm/i915/gem/i915_gem_shrinker.c | 3 --- drivers/gpu/drm/i915/gem/i915_gem_wait.c | 8 -------- 5 files changed, 42 deletions(-) delete mode 100644 drivers/gpu/drm/i915/dma_resv_utils.c delete mode 100644 drivers/gpu/drm/i915/dma_resv_utils.h
diff --git a/drivers/gpu/drm/i915/Makefile b/drivers/gpu/drm/i915/Makefile index 467872cca027..b87e3ed10d86 100644 --- a/drivers/gpu/drm/i915/Makefile +++ b/drivers/gpu/drm/i915/Makefile @@ -60,7 +60,6 @@ i915-y += i915_drv.o \
# core library code i915-y += \ - dma_resv_utils.o \ i915_memcpy.o \ i915_mm.o \ i915_sw_fence.o \ diff --git a/drivers/gpu/drm/i915/dma_resv_utils.c b/drivers/gpu/drm/i915/dma_resv_utils.c deleted file mode 100644 index 7df91b7e4ca8..000000000000 --- a/drivers/gpu/drm/i915/dma_resv_utils.c +++ /dev/null @@ -1,17 +0,0 @@ -// SPDX-License-Identifier: MIT -/* - * Copyright © 2020 Intel Corporation - */ - -#include <linux/dma-resv.h> - -#include "dma_resv_utils.h" - -void dma_resv_prune(struct dma_resv *resv) -{ - if (dma_resv_trylock(resv)) { - if (dma_resv_test_signaled(resv, true)) - dma_resv_add_excl_fence(resv, NULL); - dma_resv_unlock(resv); - } -} diff --git a/drivers/gpu/drm/i915/dma_resv_utils.h b/drivers/gpu/drm/i915/dma_resv_utils.h deleted file mode 100644 index b9d8fb5f8367..000000000000 --- a/drivers/gpu/drm/i915/dma_resv_utils.h +++ /dev/null @@ -1,13 +0,0 @@ -/* SPDX-License-Identifier: MIT */ -/* - * Copyright © 2020 Intel Corporation - */ - -#ifndef DMA_RESV_UTILS_H -#define DMA_RESV_UTILS_H - -struct dma_resv; - -void dma_resv_prune(struct dma_resv *resv); - -#endif /* DMA_RESV_UTILS_H */ diff --git a/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c b/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c index 5ab136ffdeb2..af3eb7fd951d 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c @@ -15,7 +15,6 @@
#include "gt/intel_gt_requests.h"
-#include "dma_resv_utils.h" #include "i915_trace.h"
static bool swap_available(void) @@ -229,8 +228,6 @@ i915_gem_shrink(struct i915_gem_ww_ctx *ww, i915_gem_object_unlock(obj); }
- dma_resv_prune(obj->base.resv); - scanned += obj->base.size >> PAGE_SHIFT; skip: i915_gem_object_put(obj); diff --git a/drivers/gpu/drm/i915/gem/i915_gem_wait.c b/drivers/gpu/drm/i915/gem/i915_gem_wait.c index 840c13706999..1592d95c3ead 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_wait.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_wait.c @@ -10,7 +10,6 @@
#include "gt/intel_engine.h"
-#include "dma_resv_utils.h" #include "i915_gem_ioctls.h" #include "i915_gem_object.h"
@@ -52,13 +51,6 @@ i915_gem_object_wait_reservation(struct dma_resv *resv, } dma_resv_iter_end(&cursor);
- /* - * Opportunistically prune the fences iff we know they have *all* been - * signaled. - */ - if (timeout > 0) - dma_resv_prune(resv); - return ret; }
On Thu, 21 Oct 2021 at 11:36, Maarten Lankhorst maarten.lankhorst@linux.intel.com wrote:
The signaled bit is already used for quick testing if a fence is signaled.
Why do we need this change? Can you add some more details to the commit please?
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
drivers/gpu/drm/i915/Makefile | 1 - drivers/gpu/drm/i915/dma_resv_utils.c | 17 ----------------- drivers/gpu/drm/i915/dma_resv_utils.h | 13 ------------- drivers/gpu/drm/i915/gem/i915_gem_shrinker.c | 3 --- drivers/gpu/drm/i915/gem/i915_gem_wait.c | 8 -------- 5 files changed, 42 deletions(-) delete mode 100644 drivers/gpu/drm/i915/dma_resv_utils.c delete mode 100644 drivers/gpu/drm/i915/dma_resv_utils.h
diff --git a/drivers/gpu/drm/i915/Makefile b/drivers/gpu/drm/i915/Makefile index 467872cca027..b87e3ed10d86 100644 --- a/drivers/gpu/drm/i915/Makefile +++ b/drivers/gpu/drm/i915/Makefile @@ -60,7 +60,6 @@ i915-y += i915_drv.o \
# core library code i915-y += \
dma_resv_utils.o \ i915_memcpy.o \ i915_mm.o \ i915_sw_fence.o \diff --git a/drivers/gpu/drm/i915/dma_resv_utils.c b/drivers/gpu/drm/i915/dma_resv_utils.c deleted file mode 100644 index 7df91b7e4ca8..000000000000 --- a/drivers/gpu/drm/i915/dma_resv_utils.c +++ /dev/null @@ -1,17 +0,0 @@ -// SPDX-License-Identifier: MIT -/*
- Copyright © 2020 Intel Corporation
- */
-#include <linux/dma-resv.h>
-#include "dma_resv_utils.h"
-void dma_resv_prune(struct dma_resv *resv) -{
if (dma_resv_trylock(resv)) {if (dma_resv_test_signaled(resv, true))dma_resv_add_excl_fence(resv, NULL);dma_resv_unlock(resv);}-} diff --git a/drivers/gpu/drm/i915/dma_resv_utils.h b/drivers/gpu/drm/i915/dma_resv_utils.h deleted file mode 100644 index b9d8fb5f8367..000000000000 --- a/drivers/gpu/drm/i915/dma_resv_utils.h +++ /dev/null @@ -1,13 +0,0 @@ -/* SPDX-License-Identifier: MIT */ -/*
- Copyright © 2020 Intel Corporation
- */
-#ifndef DMA_RESV_UTILS_H -#define DMA_RESV_UTILS_H
-struct dma_resv;
-void dma_resv_prune(struct dma_resv *resv);
-#endif /* DMA_RESV_UTILS_H */ diff --git a/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c b/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c index 5ab136ffdeb2..af3eb7fd951d 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c @@ -15,7 +15,6 @@
#include "gt/intel_gt_requests.h"
-#include "dma_resv_utils.h" #include "i915_trace.h"
static bool swap_available(void) @@ -229,8 +228,6 @@ i915_gem_shrink(struct i915_gem_ww_ctx *ww, i915_gem_object_unlock(obj); }
dma_resv_prune(obj->base.resv);
Like here, why do we want to drop this? Later in the series it looks like it gets added back, just in a slightly different form.
scanned += obj->base.size >> PAGE_SHIFT;skip: i915_gem_object_put(obj); diff --git a/drivers/gpu/drm/i915/gem/i915_gem_wait.c b/drivers/gpu/drm/i915/gem/i915_gem_wait.c index 840c13706999..1592d95c3ead 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_wait.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_wait.c @@ -10,7 +10,6 @@
#include "gt/intel_engine.h"
-#include "dma_resv_utils.h" #include "i915_gem_ioctls.h" #include "i915_gem_object.h"
@@ -52,13 +51,6 @@ i915_gem_object_wait_reservation(struct dma_resv *resv, } dma_resv_iter_end(&cursor);
/** Opportunistically prune the fences iff we know they have *all* been* signaled.*/if (timeout > 0)dma_resv_prune(resv);return ret;}
-- 2.33.0
Op 21-10-2021 om 16:43 schreef Matthew Auld:
On Thu, 21 Oct 2021 at 11:36, Maarten Lankhorst maarten.lankhorst@linux.intel.com wrote:
The signaled bit is already used for quick testing if a fence is signaled.
Why do we need this change? Can you add some more details to the commit please?
It's a terrible abuse of dma-fence api, and in the common case where the object is already locked by the caller, the trylock will fail.
If it were useful, the core dma-api would have exposed the same functionality.
On top of that, the fact that i915 has a dma_resv_utils.c file should be a warning that the functionality either belongs in core, or is not very useful at all. In this case the latter.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
drivers/gpu/drm/i915/Makefile | 1 - drivers/gpu/drm/i915/dma_resv_utils.c | 17 ----------------- drivers/gpu/drm/i915/dma_resv_utils.h | 13 ------------- drivers/gpu/drm/i915/gem/i915_gem_shrinker.c | 3 --- drivers/gpu/drm/i915/gem/i915_gem_wait.c | 8 -------- 5 files changed, 42 deletions(-) delete mode 100644 drivers/gpu/drm/i915/dma_resv_utils.c delete mode 100644 drivers/gpu/drm/i915/dma_resv_utils.h
diff --git a/drivers/gpu/drm/i915/Makefile b/drivers/gpu/drm/i915/Makefile index 467872cca027..b87e3ed10d86 100644 --- a/drivers/gpu/drm/i915/Makefile +++ b/drivers/gpu/drm/i915/Makefile @@ -60,7 +60,6 @@ i915-y += i915_drv.o \
# core library code i915-y += \
dma_resv_utils.o \ i915_memcpy.o \ i915_mm.o \ i915_sw_fence.o \diff --git a/drivers/gpu/drm/i915/dma_resv_utils.c b/drivers/gpu/drm/i915/dma_resv_utils.c deleted file mode 100644 index 7df91b7e4ca8..000000000000 --- a/drivers/gpu/drm/i915/dma_resv_utils.c +++ /dev/null @@ -1,17 +0,0 @@ -// SPDX-License-Identifier: MIT -/*
- Copyright © 2020 Intel Corporation
- */
-#include <linux/dma-resv.h>
-#include "dma_resv_utils.h"
-void dma_resv_prune(struct dma_resv *resv) -{
if (dma_resv_trylock(resv)) {if (dma_resv_test_signaled(resv, true))dma_resv_add_excl_fence(resv, NULL);dma_resv_unlock(resv);}-} diff --git a/drivers/gpu/drm/i915/dma_resv_utils.h b/drivers/gpu/drm/i915/dma_resv_utils.h deleted file mode 100644 index b9d8fb5f8367..000000000000 --- a/drivers/gpu/drm/i915/dma_resv_utils.h +++ /dev/null @@ -1,13 +0,0 @@ -/* SPDX-License-Identifier: MIT */ -/*
- Copyright © 2020 Intel Corporation
- */
-#ifndef DMA_RESV_UTILS_H -#define DMA_RESV_UTILS_H
-struct dma_resv;
-void dma_resv_prune(struct dma_resv *resv);
-#endif /* DMA_RESV_UTILS_H */ diff --git a/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c b/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c index 5ab136ffdeb2..af3eb7fd951d 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c @@ -15,7 +15,6 @@
#include "gt/intel_gt_requests.h"
-#include "dma_resv_utils.h" #include "i915_trace.h"
static bool swap_available(void) @@ -229,8 +228,6 @@ i915_gem_shrink(struct i915_gem_ww_ctx *ww, i915_gem_object_unlock(obj); }
dma_resv_prune(obj->base.resv);Like here, why do we want to drop this? Later in the series it looks like it gets added back, just in a slightly different form.
scanned += obj->base.size >> PAGE_SHIFT;skip: i915_gem_object_put(obj); diff --git a/drivers/gpu/drm/i915/gem/i915_gem_wait.c b/drivers/gpu/drm/i915/gem/i915_gem_wait.c index 840c13706999..1592d95c3ead 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_wait.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_wait.c @@ -10,7 +10,6 @@
#include "gt/intel_engine.h"
-#include "dma_resv_utils.h" #include "i915_gem_ioctls.h" #include "i915_gem_object.h"
@@ -52,13 +51,6 @@ i915_gem_object_wait_reservation(struct dma_resv *resv, } dma_resv_iter_end(&cursor);
/** Opportunistically prune the fences iff we know they have *all* been* signaled.*/if (timeout > 0)dma_resv_prune(resv);return ret;}
-- 2.33.0
When reworking the code to move the eviction fence to the object, the best code is removed code.
Remove some functions that are unused, and change the function definition if it's only used in 1 place.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com Reviewed-by: Niranjana Vishwanathapura niranjana.vishwanathapura@intel.com --- drivers/gpu/drm/i915/i915_active.c | 28 +++------------------------- drivers/gpu/drm/i915/i915_active.h | 17 +---------------- drivers/gpu/drm/i915/i915_vma.c | 2 +- drivers/gpu/drm/i915/i915_vma.h | 2 -- 4 files changed, 5 insertions(+), 44 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_active.c b/drivers/gpu/drm/i915/i915_active.c index 3103c1e1fd14..ee2b3a375362 100644 --- a/drivers/gpu/drm/i915/i915_active.c +++ b/drivers/gpu/drm/i915/i915_active.c @@ -426,8 +426,9 @@ replace_barrier(struct i915_active *ref, struct i915_active_fence *active) return true; }
-int i915_active_ref(struct i915_active *ref, u64 idx, struct dma_fence *fence) +int i915_active_add_request(struct i915_active *ref, struct i915_request *rq) { + struct dma_fence *fence = &rq->fence; struct i915_active_fence *active; int err;
@@ -436,7 +437,7 @@ int i915_active_ref(struct i915_active *ref, u64 idx, struct dma_fence *fence) if (err) return err;
- active = active_instance(ref, idx); + active = active_instance(ref, i915_request_timeline(rq)->fence_context); if (!active) { err = -ENOMEM; goto out; @@ -477,29 +478,6 @@ __i915_active_set_fence(struct i915_active *ref, return prev; }
-static struct i915_active_fence * -__active_fence(struct i915_active *ref, u64 idx) -{ - struct active_node *it; - - it = __active_lookup(ref, idx); - if (unlikely(!it)) { /* Contention with parallel tree builders! */ - spin_lock_irq(&ref->tree_lock); - it = __active_lookup(ref, idx); - spin_unlock_irq(&ref->tree_lock); - } - GEM_BUG_ON(!it); /* slot must be preallocated */ - - return &it->base; -} - -struct dma_fence * -__i915_active_ref(struct i915_active *ref, u64 idx, struct dma_fence *fence) -{ - /* Only valid while active, see i915_active_acquire_for_context() */ - return __i915_active_set_fence(ref, __active_fence(ref, idx), fence); -} - struct dma_fence * i915_active_set_exclusive(struct i915_active *ref, struct dma_fence *f) { diff --git a/drivers/gpu/drm/i915/i915_active.h b/drivers/gpu/drm/i915/i915_active.h index 5fcdb0e2bc9e..7eb44132183a 100644 --- a/drivers/gpu/drm/i915/i915_active.h +++ b/drivers/gpu/drm/i915/i915_active.h @@ -164,26 +164,11 @@ void __i915_active_init(struct i915_active *ref, __i915_active_init(ref, active, retire, flags, &__mkey, &__wkey); \ } while (0)
-struct dma_fence * -__i915_active_ref(struct i915_active *ref, u64 idx, struct dma_fence *fence); -int i915_active_ref(struct i915_active *ref, u64 idx, struct dma_fence *fence); - -static inline int -i915_active_add_request(struct i915_active *ref, struct i915_request *rq) -{ - return i915_active_ref(ref, - i915_request_timeline(rq)->fence_context, - &rq->fence); -} +int i915_active_add_request(struct i915_active *ref, struct i915_request *rq);
struct dma_fence * i915_active_set_exclusive(struct i915_active *ref, struct dma_fence *f);
-static inline bool i915_active_has_exclusive(struct i915_active *ref) -{ - return rcu_access_pointer(ref->excl.fence); -} - int __i915_active_wait(struct i915_active *ref, int state); static inline int i915_active_wait(struct i915_active *ref) { diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c index 90546fa58fc1..1187f1956c20 100644 --- a/drivers/gpu/drm/i915/i915_vma.c +++ b/drivers/gpu/drm/i915/i915_vma.c @@ -1220,7 +1220,7 @@ __i915_request_await_bind(struct i915_request *rq, struct i915_vma *vma) return __i915_request_await_exclusive(rq, &vma->active); }
-int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request *rq) +static int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request *rq) { int err;
diff --git a/drivers/gpu/drm/i915/i915_vma.h b/drivers/gpu/drm/i915/i915_vma.h index 648dbe744c96..b882fd7b5f99 100644 --- a/drivers/gpu/drm/i915/i915_vma.h +++ b/drivers/gpu/drm/i915/i915_vma.h @@ -55,8 +55,6 @@ static inline bool i915_vma_is_active(const struct i915_vma *vma) /* do not reserve memory to prevent deadlocks */ #define __EXEC_OBJECT_NO_RESERVE BIT(31)
-int __must_check __i915_vma_move_to_active(struct i915_vma *vma, - struct i915_request *rq); int __must_check _i915_vma_move_to_active(struct i915_vma *vma, struct i915_request *rq, struct dma_fence *fence,
Use a single null-terminated array for simplicity instead of a linked list. This might slightly speed up execbuf when many vma's may be marked as capture, but definitely removes an allocation from a signaling path.
We are not allowed to allocate memory in eb_move_to_gpu, but we can't enforce it yet through annotations.
Changes since v1: - Rebase on top of multi-batchbuffer changes.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com Reviewed-by: Niranjana Vishwanathapura niranjana.vishwanathapura@intel.com #v1 --- .../gpu/drm/i915/gem/i915_gem_execbuffer.c | 29 ++++++++++++------- drivers/gpu/drm/i915/i915_gpu_error.c | 9 +++--- drivers/gpu/drm/i915/i915_request.c | 9 ++---- drivers/gpu/drm/i915/i915_request.h | 7 +---- 4 files changed, 26 insertions(+), 28 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c index 9c323666bd7c..eaacadd2d2e5 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c @@ -265,6 +265,9 @@ struct i915_execbuffer { /* number of batches in execbuf IOCTL */ unsigned int num_batches;
+ /* Number of objects with EXEC_OBJECT_CAPTURE set */ + unsigned int capture_count; + /** list of vma not yet bound during reservation phase */ struct list_head unbound;
@@ -909,6 +912,9 @@ static int eb_lookup_vmas(struct i915_execbuffer *eb) goto err; }
+ if (eb->exec[i].flags & EXEC_OBJECT_CAPTURE) + eb->capture_count++; + err = eb_validate_vma(eb, &eb->exec[i], vma); if (unlikely(err)) { i915_vma_put(vma); @@ -1906,19 +1912,11 @@ static int eb_move_to_gpu(struct i915_execbuffer *eb) assert_vma_held(vma);
if (flags & EXEC_OBJECT_CAPTURE) { - struct i915_capture_list *capture; + eb->capture_count--;
for_each_batch_create_order(eb, j) { - if (!eb->requests[j]) - break; - - capture = kmalloc(sizeof(*capture), GFP_KERNEL); - if (capture) { - capture->next = - eb->requests[j]->capture_list; - capture->vma = vma; - eb->requests[j]->capture_list = capture; - } + if (eb->requests[j]->capture_list) + eb->requests[j]->capture_list[eb->capture_count] = vma; } }
@@ -3130,6 +3128,14 @@ eb_requests_create(struct i915_execbuffer *eb, struct dma_fence *in_fence, return out_fence; }
+ if (eb->capture_count) { + eb->requests[i]->capture_list = + kvcalloc(eb->capture_count + 1, + sizeof(*eb->requests[i]->capture_list), + GFP_KERNEL | __GFP_NOWARN); + } + + /* * Only the first request added (committed to backend) has to * take the in fences into account as all subsequent requests @@ -3197,6 +3203,7 @@ i915_gem_do_execbuffer(struct drm_device *dev,
eb.fences = NULL; eb.num_fences = 0; + eb.capture_count = 0;
memset(eb.requests, 0, sizeof(struct i915_request *) * ARRAY_SIZE(eb.requests)); diff --git a/drivers/gpu/drm/i915/i915_gpu_error.c b/drivers/gpu/drm/i915/i915_gpu_error.c index 2a2d7643b551..45104bb12a98 100644 --- a/drivers/gpu/drm/i915/i915_gpu_error.c +++ b/drivers/gpu/drm/i915/i915_gpu_error.c @@ -1356,10 +1356,10 @@ capture_user(struct intel_engine_capture_vma *capture, const struct i915_request *rq, gfp_t gfp) { - struct i915_capture_list *c; + int i;
- for (c = rq->capture_list; c; c = c->next) - capture = capture_vma(capture, c->vma, "user", gfp); + for (i = 0; rq->capture_list[i]; i++) + capture = capture_vma(capture, rq->capture_list[i], "user", gfp);
return capture; } @@ -1407,7 +1407,8 @@ intel_engine_coredump_add_request(struct intel_engine_coredump *ee, * by userspace. */ vma = capture_vma(vma, rq->batch, "batch", gfp); - vma = capture_user(vma, rq, gfp); + if (rq->capture_list) + vma = capture_user(vma, rq, gfp); vma = capture_vma(vma, rq->ring->vma, "ring", gfp); vma = capture_vma(vma, rq->context->state, "HW context", gfp);
diff --git a/drivers/gpu/drm/i915/i915_request.c b/drivers/gpu/drm/i915/i915_request.c index 42cd17357771..9b28e6f97c4b 100644 --- a/drivers/gpu/drm/i915/i915_request.c +++ b/drivers/gpu/drm/i915/i915_request.c @@ -188,15 +188,10 @@ void i915_request_notify_execute_cb_imm(struct i915_request *rq)
static void free_capture_list(struct i915_request *request) { - struct i915_capture_list *capture; + struct i915_vma **capture;
capture = fetch_and_zero(&request->capture_list); - while (capture) { - struct i915_capture_list *next = capture->next; - - kfree(capture); - capture = next; - } + kvfree(capture); }
static void __i915_request_fill(struct i915_request *rq, u8 val) diff --git a/drivers/gpu/drm/i915/i915_request.h b/drivers/gpu/drm/i915/i915_request.h index 3c6e8acd1457..2d04b25dca36 100644 --- a/drivers/gpu/drm/i915/i915_request.h +++ b/drivers/gpu/drm/i915/i915_request.h @@ -48,11 +48,6 @@ struct drm_i915_gem_object; struct drm_printer; struct i915_request;
-struct i915_capture_list { - struct i915_capture_list *next; - struct i915_vma *vma; -}; - #define RQ_TRACE(rq, fmt, ...) do { \ const struct i915_request *rq__ = (rq); \ ENGINE_TRACE(rq__->engine, "fence %llx:%lld, current %d " fmt, \ @@ -299,7 +294,7 @@ struct i915_request { * active reference - all objects on this list must also be * on the active_list (of their final request). */ - struct i915_capture_list *capture_list; + struct i915_vma **capture_list;
/** Time at which this request was emitted, in jiffies. */ unsigned long emitted_jiffies;
gen6_ppgtt_unpin_all is unused, kill it.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com Reviewed-by: Matthew Auld matthew.auld@intel.com --- drivers/gpu/drm/i915/gt/gen6_ppgtt.c | 11 ----------- drivers/gpu/drm/i915/gt/gen6_ppgtt.h | 1 - 2 files changed, 12 deletions(-)
diff --git a/drivers/gpu/drm/i915/gt/gen6_ppgtt.c b/drivers/gpu/drm/i915/gt/gen6_ppgtt.c index 890191f286e3..9fdbd9d3372b 100644 --- a/drivers/gpu/drm/i915/gt/gen6_ppgtt.c +++ b/drivers/gpu/drm/i915/gt/gen6_ppgtt.c @@ -405,17 +405,6 @@ void gen6_ppgtt_unpin(struct i915_ppgtt *base) i915_vma_unpin(ppgtt->vma); }
-void gen6_ppgtt_unpin_all(struct i915_ppgtt *base) -{ - struct gen6_ppgtt *ppgtt = to_gen6_ppgtt(base); - - if (!atomic_read(&ppgtt->pin_count)) - return; - - i915_vma_unpin(ppgtt->vma); - atomic_set(&ppgtt->pin_count, 0); -} - struct i915_ppgtt *gen6_ppgtt_create(struct intel_gt *gt) { struct i915_ggtt * const ggtt = gt->ggtt; diff --git a/drivers/gpu/drm/i915/gt/gen6_ppgtt.h b/drivers/gpu/drm/i915/gt/gen6_ppgtt.h index 6a61a5c3a85a..ab0eecb086dd 100644 --- a/drivers/gpu/drm/i915/gt/gen6_ppgtt.h +++ b/drivers/gpu/drm/i915/gt/gen6_ppgtt.h @@ -71,7 +71,6 @@ static inline struct gen6_ppgtt *to_gen6_ppgtt(struct i915_ppgtt *base)
int gen6_ppgtt_pin(struct i915_ppgtt *base, struct i915_gem_ww_ctx *ww); void gen6_ppgtt_unpin(struct i915_ppgtt *base); -void gen6_ppgtt_unpin_all(struct i915_ppgtt *base); void gen6_ppgtt_enable(struct intel_gt *gt); void gen7_ppgtt_enable(struct intel_gt *gt); struct i915_ppgtt *gen6_ppgtt_create(struct intel_gt *gt);
We currently have to special case vma->obj being NULL because of gen6 ppgtt and mock_engine. Fix gen6 ppgtt, so we may soon be able to remove a few checks. As the object only exists as a fake object pointing to ggtt, we have no backing storage, so no real object is created. It just has to look real enough.
Also kill pin_mutex, it's not compatible with ww locking, and we can use the vm lock instead.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com --- drivers/gpu/drm/i915/gem/i915_gem_internal.c | 44 ++++--- drivers/gpu/drm/i915/gt/gen6_ppgtt.c | 122 +++++++++++-------- drivers/gpu/drm/i915/gt/gen6_ppgtt.h | 1 - drivers/gpu/drm/i915/i915_drv.h | 4 + 4 files changed, 99 insertions(+), 72 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_internal.c b/drivers/gpu/drm/i915/gem/i915_gem_internal.c index a57a6b7013c2..c5150a1ee3d2 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_internal.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_internal.c @@ -145,24 +145,10 @@ static const struct drm_i915_gem_object_ops i915_gem_object_internal_ops = { .put_pages = i915_gem_object_put_pages_internal, };
-/** - * i915_gem_object_create_internal: create an object with volatile pages - * @i915: the i915 device - * @size: the size in bytes of backing storage to allocate for the object - * - * Creates a new object that wraps some internal memory for private use. - * This object is not backed by swappable storage, and as such its contents - * are volatile and only valid whilst pinned. If the object is reaped by the - * shrinker, its pages and data will be discarded. Equally, it is not a full - * GEM object and so not valid for access from userspace. This makes it useful - * for hardware interfaces like ringbuffers (which are pinned from the time - * the request is written to the time the hardware stops accessing it), but - * not for contexts (which need to be preserved when not active for later - * reuse). Note that it is not cleared upon allocation. - */ struct drm_i915_gem_object * -i915_gem_object_create_internal(struct drm_i915_private *i915, - phys_addr_t size) +__i915_gem_object_create_internal(struct drm_i915_private *i915, + const struct drm_i915_gem_object_ops *ops, + phys_addr_t size) { static struct lock_class_key lock_class; struct drm_i915_gem_object *obj; @@ -179,7 +165,7 @@ i915_gem_object_create_internal(struct drm_i915_private *i915, return ERR_PTR(-ENOMEM);
drm_gem_private_object_init(&i915->drm, &obj->base, size); - i915_gem_object_init(obj, &i915_gem_object_internal_ops, &lock_class, 0); + i915_gem_object_init(obj, ops, &lock_class, 0); obj->mem_flags |= I915_BO_FLAG_STRUCT_PAGE;
/* @@ -199,3 +185,25 @@ i915_gem_object_create_internal(struct drm_i915_private *i915,
return obj; } + +/** + * i915_gem_object_create_internal: create an object with volatile pages + * @i915: the i915 device + * @size: the size in bytes of backing storage to allocate for the object + * + * Creates a new object that wraps some internal memory for private use. + * This object is not backed by swappable storage, and as such its contents + * are volatile and only valid whilst pinned. If the object is reaped by the + * shrinker, its pages and data will be discarded. Equally, it is not a full + * GEM object and so not valid for access from userspace. This makes it useful + * for hardware interfaces like ringbuffers (which are pinned from the time + * the request is written to the time the hardware stops accessing it), but + * not for contexts (which need to be preserved when not active for later + * reuse). Note that it is not cleared upon allocation. + */ +struct drm_i915_gem_object * +i915_gem_object_create_internal(struct drm_i915_private *i915, + phys_addr_t size) +{ + return __i915_gem_object_create_internal(i915, &i915_gem_object_internal_ops, size); +} diff --git a/drivers/gpu/drm/i915/gt/gen6_ppgtt.c b/drivers/gpu/drm/i915/gt/gen6_ppgtt.c index 9fdbd9d3372b..5caa1703716e 100644 --- a/drivers/gpu/drm/i915/gt/gen6_ppgtt.c +++ b/drivers/gpu/drm/i915/gt/gen6_ppgtt.c @@ -262,13 +262,10 @@ static void gen6_ppgtt_cleanup(struct i915_address_space *vm) { struct gen6_ppgtt *ppgtt = to_gen6_ppgtt(i915_vm_to_ppgtt(vm));
- __i915_vma_put(ppgtt->vma); - gen6_ppgtt_free_pd(ppgtt); free_scratch(vm);
mutex_destroy(&ppgtt->flush); - mutex_destroy(&ppgtt->pin_mutex);
free_pd(&ppgtt->base.vm, ppgtt->base.pd); } @@ -331,37 +328,6 @@ static const struct i915_vma_ops pd_vma_ops = { .unbind_vma = pd_vma_unbind, };
-static struct i915_vma *pd_vma_create(struct gen6_ppgtt *ppgtt, int size) -{ - struct i915_ggtt *ggtt = ppgtt->base.vm.gt->ggtt; - struct i915_vma *vma; - - GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE)); - GEM_BUG_ON(size > ggtt->vm.total); - - vma = i915_vma_alloc(); - if (!vma) - return ERR_PTR(-ENOMEM); - - i915_active_init(&vma->active, NULL, NULL, 0); - - kref_init(&vma->ref); - mutex_init(&vma->pages_mutex); - vma->vm = i915_vm_get(&ggtt->vm); - vma->ops = &pd_vma_ops; - vma->private = ppgtt; - - vma->size = size; - vma->fence_size = size; - atomic_set(&vma->flags, I915_VMA_GGTT); - vma->ggtt_view.type = I915_GGTT_VIEW_ROTATED; /* prevent fencing */ - - INIT_LIST_HEAD(&vma->obj_link); - INIT_LIST_HEAD(&vma->closed_link); - - return vma; -} - int gen6_ppgtt_pin(struct i915_ppgtt *base, struct i915_gem_ww_ctx *ww) { struct gen6_ppgtt *ppgtt = to_gen6_ppgtt(base); @@ -378,24 +344,84 @@ int gen6_ppgtt_pin(struct i915_ppgtt *base, struct i915_gem_ww_ctx *ww) if (atomic_add_unless(&ppgtt->pin_count, 1, 0)) return 0;
- if (mutex_lock_interruptible(&ppgtt->pin_mutex)) - return -EINTR; + /* grab the ppgtt resv to pin the object */ + err = i915_vm_lock_objects(&ppgtt->base.vm, ww); + if (err) + return err;
/* * PPGTT PDEs reside in the GGTT and consists of 512 entries. The * allocator works in address space sizes, so it's multiplied by page * size. We allocate at the top of the GTT to avoid fragmentation. */ - err = 0; - if (!atomic_read(&ppgtt->pin_count)) + if (!atomic_read(&ppgtt->pin_count)) { err = i915_ggtt_pin(ppgtt->vma, ww, GEN6_PD_ALIGN, PIN_HIGH); + + GEM_BUG_ON(ppgtt->vma->fence); + clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(ppgtt->vma)); + } if (!err) atomic_inc(&ppgtt->pin_count); - mutex_unlock(&ppgtt->pin_mutex);
return err; }
+static int pd_dummy_obj_get_pages(struct drm_i915_gem_object *obj) +{ + obj->mm.pages = ZERO_SIZE_PTR; + return 0; +} + +static void pd_dummy_obj_put_pages(struct drm_i915_gem_object *obj, + struct sg_table *pages) +{ +} + +static const struct drm_i915_gem_object_ops pd_dummy_obj_ops = { + .name = "pd_dummy_obj", + .flags = I915_GEM_OBJECT_IS_SHRINKABLE, + .get_pages = pd_dummy_obj_get_pages, + .put_pages = pd_dummy_obj_put_pages, +}; + +static struct i915_page_directory * +gen6_alloc_top_pd(struct gen6_ppgtt *ppgtt) +{ + struct i915_ggtt * const ggtt = ppgtt->base.vm.gt->ggtt; + struct i915_page_directory *pd; + int err; + + pd = __alloc_pd(I915_PDES); + if (unlikely(!pd)) + return ERR_PTR(-ENOMEM); + + pd->pt.base = __i915_gem_object_create_internal(ppgtt->base.vm.gt->i915, &pd_dummy_obj_ops, I915_PDES * SZ_4K); + if (IS_ERR(pd->pt.base)) { + err = PTR_ERR(pd->pt.base); + pd->pt.base = NULL; + goto err_pd; + } + + pd->pt.base->base.resv = i915_vm_resv_get(&ppgtt->base.vm); + pd->pt.base->shares_resv_from = &ppgtt->base.vm; + + ppgtt->vma = i915_vma_instance(pd->pt.base, &ggtt->vm, NULL); + if (IS_ERR(ppgtt->vma)) { + err = PTR_ERR(ppgtt->vma); + ppgtt->vma = NULL; + goto err_pd; + } + + /* The dummy object we create is special, override ops.. */ + ppgtt->vma->ops = &pd_vma_ops; + ppgtt->vma->private = ppgtt; + return pd; + +err_pd: + free_pd(&ppgtt->base.vm, pd); + return ERR_PTR(err); +} + void gen6_ppgtt_unpin(struct i915_ppgtt *base) { struct gen6_ppgtt *ppgtt = to_gen6_ppgtt(base); @@ -416,7 +442,6 @@ struct i915_ppgtt *gen6_ppgtt_create(struct intel_gt *gt) return ERR_PTR(-ENOMEM);
mutex_init(&ppgtt->flush); - mutex_init(&ppgtt->pin_mutex);
ppgtt_init(&ppgtt->base, gt, 0); ppgtt->base.vm.pd_shift = ilog2(SZ_4K * SZ_4K / sizeof(gen6_pte_t)); @@ -431,19 +456,13 @@ struct i915_ppgtt *gen6_ppgtt_create(struct intel_gt *gt) ppgtt->base.vm.alloc_pt_dma = alloc_pt_dma; ppgtt->base.vm.pte_encode = ggtt->vm.pte_encode;
- ppgtt->base.pd = __alloc_pd(I915_PDES); - if (!ppgtt->base.pd) { - err = -ENOMEM; - goto err_free; - } - err = gen6_ppgtt_init_scratch(ppgtt); if (err) - goto err_pd; + goto err_free;
- ppgtt->vma = pd_vma_create(ppgtt, GEN6_PD_SIZE); - if (IS_ERR(ppgtt->vma)) { - err = PTR_ERR(ppgtt->vma); + ppgtt->base.pd = gen6_alloc_top_pd(ppgtt); + if (IS_ERR(ppgtt->base.pd)) { + err = PTR_ERR(ppgtt->base.pd); goto err_scratch; }
@@ -451,10 +470,7 @@ struct i915_ppgtt *gen6_ppgtt_create(struct intel_gt *gt)
err_scratch: free_scratch(&ppgtt->base.vm); -err_pd: - free_pd(&ppgtt->base.vm, ppgtt->base.pd); err_free: - mutex_destroy(&ppgtt->pin_mutex); kfree(ppgtt); return ERR_PTR(err); } diff --git a/drivers/gpu/drm/i915/gt/gen6_ppgtt.h b/drivers/gpu/drm/i915/gt/gen6_ppgtt.h index ab0eecb086dd..5e5cf2ec3309 100644 --- a/drivers/gpu/drm/i915/gt/gen6_ppgtt.h +++ b/drivers/gpu/drm/i915/gt/gen6_ppgtt.h @@ -19,7 +19,6 @@ struct gen6_ppgtt { u32 pp_dir;
atomic_t pin_count; - struct mutex pin_mutex;
bool scan_for_unused_pt; }; diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h index 12256218634f..a8d733a7e1d9 100644 --- a/drivers/gpu/drm/i915/i915_drv.h +++ b/drivers/gpu/drm/i915/i915_drv.h @@ -1933,6 +1933,10 @@ int i915_gem_evict_vm(struct i915_address_space *vm); struct drm_i915_gem_object * i915_gem_object_create_internal(struct drm_i915_private *dev_priv, phys_addr_t size); +struct drm_i915_gem_object * +__i915_gem_object_create_internal(struct drm_i915_private *dev_priv, + const struct drm_i915_gem_object_ops *ops, + phys_addr_t size);
/* i915_gem_tiling.c */ static inline bool i915_gem_object_needs_bit17_swizzle(struct drm_i915_gem_object *obj)
On Thu, 21 Oct 2021 at 11:36, Maarten Lankhorst maarten.lankhorst@linux.intel.com wrote:
We currently have to special case vma->obj being NULL because of gen6 ppgtt and mock_engine. Fix gen6 ppgtt, so we may soon be able to remove a few checks. As the object only exists as a fake object pointing to ggtt, we have no backing storage, so no real object is created. It just has to look real enough.
Also kill pin_mutex, it's not compatible with ww locking, and we can use the vm lock instead.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
drivers/gpu/drm/i915/gem/i915_gem_internal.c | 44 ++++--- drivers/gpu/drm/i915/gt/gen6_ppgtt.c | 122 +++++++++++-------- drivers/gpu/drm/i915/gt/gen6_ppgtt.h | 1 - drivers/gpu/drm/i915/i915_drv.h | 4 + 4 files changed, 99 insertions(+), 72 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_internal.c b/drivers/gpu/drm/i915/gem/i915_gem_internal.c index a57a6b7013c2..c5150a1ee3d2 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_internal.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_internal.c @@ -145,24 +145,10 @@ static const struct drm_i915_gem_object_ops i915_gem_object_internal_ops = { .put_pages = i915_gem_object_put_pages_internal, };
-/**
- i915_gem_object_create_internal: create an object with volatile pages
- @i915: the i915 device
- @size: the size in bytes of backing storage to allocate for the object
- Creates a new object that wraps some internal memory for private use.
- This object is not backed by swappable storage, and as such its contents
- are volatile and only valid whilst pinned. If the object is reaped by the
- shrinker, its pages and data will be discarded. Equally, it is not a full
- GEM object and so not valid for access from userspace. This makes it useful
- for hardware interfaces like ringbuffers (which are pinned from the time
- the request is written to the time the hardware stops accessing it), but
- not for contexts (which need to be preserved when not active for later
- reuse). Note that it is not cleared upon allocation.
- */
struct drm_i915_gem_object * -i915_gem_object_create_internal(struct drm_i915_private *i915,
phys_addr_t size)+__i915_gem_object_create_internal(struct drm_i915_private *i915,
const struct drm_i915_gem_object_ops *ops,phys_addr_t size){ static struct lock_class_key lock_class; struct drm_i915_gem_object *obj; @@ -179,7 +165,7 @@ i915_gem_object_create_internal(struct drm_i915_private *i915, return ERR_PTR(-ENOMEM);
drm_gem_private_object_init(&i915->drm, &obj->base, size);
i915_gem_object_init(obj, &i915_gem_object_internal_ops, &lock_class, 0);
i915_gem_object_init(obj, ops, &lock_class, 0); obj->mem_flags |= I915_BO_FLAG_STRUCT_PAGE; /*@@ -199,3 +185,25 @@ i915_gem_object_create_internal(struct drm_i915_private *i915,
return obj;}
+/**
- i915_gem_object_create_internal: create an object with volatile pages
- @i915: the i915 device
- @size: the size in bytes of backing storage to allocate for the object
- Creates a new object that wraps some internal memory for private use.
- This object is not backed by swappable storage, and as such its contents
- are volatile and only valid whilst pinned. If the object is reaped by the
- shrinker, its pages and data will be discarded. Equally, it is not a full
- GEM object and so not valid for access from userspace. This makes it useful
- for hardware interfaces like ringbuffers (which are pinned from the time
- the request is written to the time the hardware stops accessing it), but
- not for contexts (which need to be preserved when not active for later
- reuse). Note that it is not cleared upon allocation.
- */
+struct drm_i915_gem_object * +i915_gem_object_create_internal(struct drm_i915_private *i915,
phys_addr_t size)+{
return __i915_gem_object_create_internal(i915, &i915_gem_object_internal_ops, size);+} diff --git a/drivers/gpu/drm/i915/gt/gen6_ppgtt.c b/drivers/gpu/drm/i915/gt/gen6_ppgtt.c index 9fdbd9d3372b..5caa1703716e 100644 --- a/drivers/gpu/drm/i915/gt/gen6_ppgtt.c +++ b/drivers/gpu/drm/i915/gt/gen6_ppgtt.c @@ -262,13 +262,10 @@ static void gen6_ppgtt_cleanup(struct i915_address_space *vm) { struct gen6_ppgtt *ppgtt = to_gen6_ppgtt(i915_vm_to_ppgtt(vm));
__i915_vma_put(ppgtt->vma);gen6_ppgtt_free_pd(ppgtt); free_scratch(vm); mutex_destroy(&ppgtt->flush);mutex_destroy(&ppgtt->pin_mutex); free_pd(&ppgtt->base.vm, ppgtt->base.pd);} @@ -331,37 +328,6 @@ static const struct i915_vma_ops pd_vma_ops = { .unbind_vma = pd_vma_unbind, };
-static struct i915_vma *pd_vma_create(struct gen6_ppgtt *ppgtt, int size) -{
struct i915_ggtt *ggtt = ppgtt->base.vm.gt->ggtt;struct i915_vma *vma;GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));GEM_BUG_ON(size > ggtt->vm.total);vma = i915_vma_alloc();if (!vma)return ERR_PTR(-ENOMEM);i915_active_init(&vma->active, NULL, NULL, 0);kref_init(&vma->ref);mutex_init(&vma->pages_mutex);vma->vm = i915_vm_get(&ggtt->vm);vma->ops = &pd_vma_ops;vma->private = ppgtt;vma->size = size;vma->fence_size = size;atomic_set(&vma->flags, I915_VMA_GGTT);vma->ggtt_view.type = I915_GGTT_VIEW_ROTATED; /* prevent fencing */INIT_LIST_HEAD(&vma->obj_link);INIT_LIST_HEAD(&vma->closed_link);return vma;-}
int gen6_ppgtt_pin(struct i915_ppgtt *base, struct i915_gem_ww_ctx *ww) { struct gen6_ppgtt *ppgtt = to_gen6_ppgtt(base); @@ -378,24 +344,84 @@ int gen6_ppgtt_pin(struct i915_ppgtt *base, struct i915_gem_ww_ctx *ww) if (atomic_add_unless(&ppgtt->pin_count, 1, 0)) return 0;
if (mutex_lock_interruptible(&ppgtt->pin_mutex))return -EINTR;
/* grab the ppgtt resv to pin the object */err = i915_vm_lock_objects(&ppgtt->base.vm, ww);if (err)return err; /* * PPGTT PDEs reside in the GGTT and consists of 512 entries. The * allocator works in address space sizes, so it's multiplied by page * size. We allocate at the top of the GTT to avoid fragmentation. */
err = 0;if (!atomic_read(&ppgtt->pin_count))
if (!atomic_read(&ppgtt->pin_count)) { err = i915_ggtt_pin(ppgtt->vma, ww, GEN6_PD_ALIGN, PIN_HIGH);GEM_BUG_ON(ppgtt->vma->fence);clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(ppgtt->vma));} if (!err) atomic_inc(&ppgtt->pin_count);
mutex_unlock(&ppgtt->pin_mutex); return err;}
+static int pd_dummy_obj_get_pages(struct drm_i915_gem_object *obj) +{
obj->mm.pages = ZERO_SIZE_PTR;return 0;+}
+static void pd_dummy_obj_put_pages(struct drm_i915_gem_object *obj,
struct sg_table *pages)+{ +}
+static const struct drm_i915_gem_object_ops pd_dummy_obj_ops = {
.name = "pd_dummy_obj",.flags = I915_GEM_OBJECT_IS_SHRINKABLE,
I would assume we don't want this dummy object to be considered for shrinking?
.get_pages = pd_dummy_obj_get_pages,.put_pages = pd_dummy_obj_put_pages,+};
+static struct i915_page_directory * +gen6_alloc_top_pd(struct gen6_ppgtt *ppgtt) +{
struct i915_ggtt * const ggtt = ppgtt->base.vm.gt->ggtt;struct i915_page_directory *pd;int err;pd = __alloc_pd(I915_PDES);if (unlikely(!pd))return ERR_PTR(-ENOMEM);pd->pt.base = __i915_gem_object_create_internal(ppgtt->base.vm.gt->i915, &pd_dummy_obj_ops, I915_PDES * SZ_4K);
Overly long line?
Otherwise, Reviewed-by: Matthew Auld matthew.auld@intel.com
if (IS_ERR(pd->pt.base)) {err = PTR_ERR(pd->pt.base);pd->pt.base = NULL;goto err_pd;}pd->pt.base->base.resv = i915_vm_resv_get(&ppgtt->base.vm);pd->pt.base->shares_resv_from = &ppgtt->base.vm;ppgtt->vma = i915_vma_instance(pd->pt.base, &ggtt->vm, NULL);if (IS_ERR(ppgtt->vma)) {err = PTR_ERR(ppgtt->vma);ppgtt->vma = NULL;goto err_pd;}/* The dummy object we create is special, override ops.. */ppgtt->vma->ops = &pd_vma_ops;ppgtt->vma->private = ppgtt;return pd;+err_pd:
free_pd(&ppgtt->base.vm, pd);return ERR_PTR(err);+}
void gen6_ppgtt_unpin(struct i915_ppgtt *base) { struct gen6_ppgtt *ppgtt = to_gen6_ppgtt(base); @@ -416,7 +442,6 @@ struct i915_ppgtt *gen6_ppgtt_create(struct intel_gt *gt) return ERR_PTR(-ENOMEM);
mutex_init(&ppgtt->flush);
mutex_init(&ppgtt->pin_mutex); ppgtt_init(&ppgtt->base, gt, 0); ppgtt->base.vm.pd_shift = ilog2(SZ_4K * SZ_4K / sizeof(gen6_pte_t));@@ -431,19 +456,13 @@ struct i915_ppgtt *gen6_ppgtt_create(struct intel_gt *gt) ppgtt->base.vm.alloc_pt_dma = alloc_pt_dma; ppgtt->base.vm.pte_encode = ggtt->vm.pte_encode;
ppgtt->base.pd = __alloc_pd(I915_PDES);if (!ppgtt->base.pd) {err = -ENOMEM;goto err_free;}err = gen6_ppgtt_init_scratch(ppgtt); if (err)goto err_pd;
goto err_free;
ppgtt->vma = pd_vma_create(ppgtt, GEN6_PD_SIZE);if (IS_ERR(ppgtt->vma)) {err = PTR_ERR(ppgtt->vma);
ppgtt->base.pd = gen6_alloc_top_pd(ppgtt);if (IS_ERR(ppgtt->base.pd)) {err = PTR_ERR(ppgtt->base.pd); goto err_scratch; }@@ -451,10 +470,7 @@ struct i915_ppgtt *gen6_ppgtt_create(struct intel_gt *gt)
err_scratch: free_scratch(&ppgtt->base.vm); -err_pd:
free_pd(&ppgtt->base.vm, ppgtt->base.pd);err_free:
mutex_destroy(&ppgtt->pin_mutex); kfree(ppgtt); return ERR_PTR(err);} diff --git a/drivers/gpu/drm/i915/gt/gen6_ppgtt.h b/drivers/gpu/drm/i915/gt/gen6_ppgtt.h index ab0eecb086dd..5e5cf2ec3309 100644 --- a/drivers/gpu/drm/i915/gt/gen6_ppgtt.h +++ b/drivers/gpu/drm/i915/gt/gen6_ppgtt.h @@ -19,7 +19,6 @@ struct gen6_ppgtt { u32 pp_dir;
atomic_t pin_count;
struct mutex pin_mutex; bool scan_for_unused_pt;}; diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h index 12256218634f..a8d733a7e1d9 100644 --- a/drivers/gpu/drm/i915/i915_drv.h +++ b/drivers/gpu/drm/i915/i915_drv.h @@ -1933,6 +1933,10 @@ int i915_gem_evict_vm(struct i915_address_space *vm); struct drm_i915_gem_object * i915_gem_object_create_internal(struct drm_i915_private *dev_priv, phys_addr_t size); +struct drm_i915_gem_object * +__i915_gem_object_create_internal(struct drm_i915_private *dev_priv,
const struct drm_i915_gem_object_ops *ops,phys_addr_t size);/* i915_gem_tiling.c */ static inline bool i915_gem_object_needs_bit17_swizzle(struct drm_i915_gem_object *obj) -- 2.33.0
This allows us to finally get rid of all the assumptions that vma->obj is NULL.
Changes since v1: - Ensure the mock_ring vma is pinned to prevent a fault. - Pin it high to avoid failure in evict_for_vma selftest.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com --- drivers/gpu/drm/i915/gt/mock_engine.c | 38 ++++++++++++++++++++------- 1 file changed, 28 insertions(+), 10 deletions(-)
diff --git a/drivers/gpu/drm/i915/gt/mock_engine.c b/drivers/gpu/drm/i915/gt/mock_engine.c index 8b89215afe46..bb99fc03f503 100644 --- a/drivers/gpu/drm/i915/gt/mock_engine.c +++ b/drivers/gpu/drm/i915/gt/mock_engine.c @@ -35,9 +35,31 @@ static void mock_timeline_unpin(struct intel_timeline *tl) atomic_dec(&tl->pin_count); }
+static struct i915_vma *create_ring_vma(struct i915_ggtt *ggtt, int size) +{ + struct i915_address_space *vm = &ggtt->vm; + struct drm_i915_private *i915 = vm->i915; + struct drm_i915_gem_object *obj; + struct i915_vma *vma; + + obj = i915_gem_object_create_internal(i915, size); + if (IS_ERR(obj)) + return ERR_CAST(obj); + + vma = i915_vma_instance(obj, vm, NULL); + if (IS_ERR(vma)) + goto err; + + return vma; + +err: + i915_gem_object_put(obj); + return vma; +} + static struct intel_ring *mock_ring(struct intel_engine_cs *engine) { - const unsigned long sz = PAGE_SIZE / 2; + const unsigned long sz = PAGE_SIZE; struct intel_ring *ring;
ring = kzalloc(sizeof(*ring) + sz, GFP_KERNEL); @@ -50,15 +72,11 @@ static struct intel_ring *mock_ring(struct intel_engine_cs *engine) ring->vaddr = (void *)(ring + 1); atomic_set(&ring->pin_count, 1);
- ring->vma = i915_vma_alloc(); - if (!ring->vma) { + ring->vma = create_ring_vma(engine->gt->ggtt, PAGE_SIZE); + if (IS_ERR(ring->vma)) { kfree(ring); return NULL; } - i915_active_init(&ring->vma->active, NULL, NULL, 0); - __set_bit(I915_VMA_GGTT_BIT, __i915_vma_flags(ring->vma)); - __set_bit(DRM_MM_NODE_ALLOCATED_BIT, &ring->vma->node.flags); - ring->vma->node.size = sz;
intel_ring_update_space(ring);
@@ -67,8 +85,7 @@ static struct intel_ring *mock_ring(struct intel_engine_cs *engine)
static void mock_ring_free(struct intel_ring *ring) { - i915_active_fini(&ring->vma->active); - i915_vma_free(ring->vma); + i915_vma_put(ring->vma);
kfree(ring); } @@ -125,6 +142,7 @@ static void mock_context_unpin(struct intel_context *ce)
static void mock_context_post_unpin(struct intel_context *ce) { + i915_vma_unpin(ce->ring->vma); }
static void mock_context_destroy(struct kref *ref) @@ -169,7 +187,7 @@ static int mock_context_alloc(struct intel_context *ce) static int mock_context_pre_pin(struct intel_context *ce, struct i915_gem_ww_ctx *ww, void **unused) { - return 0; + return i915_vma_pin_ww(ce->ring->vma, ww, 0, 0, PIN_GLOBAL | PIN_HIGH); }
static int mock_context_pin(struct intel_context *ce, void *unused)
On Thu, 21 Oct 2021 at 11:36, Maarten Lankhorst maarten.lankhorst@linux.intel.com wrote:
This allows us to finally get rid of all the assumptions that vma->obj is NULL.
Changes since v1:
- Ensure the mock_ring vma is pinned to prevent a fault.
- Pin it high to avoid failure in evict_for_vma selftest.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
drivers/gpu/drm/i915/gt/mock_engine.c | 38 ++++++++++++++++++++------- 1 file changed, 28 insertions(+), 10 deletions(-)
diff --git a/drivers/gpu/drm/i915/gt/mock_engine.c b/drivers/gpu/drm/i915/gt/mock_engine.c index 8b89215afe46..bb99fc03f503 100644 --- a/drivers/gpu/drm/i915/gt/mock_engine.c +++ b/drivers/gpu/drm/i915/gt/mock_engine.c @@ -35,9 +35,31 @@ static void mock_timeline_unpin(struct intel_timeline *tl) atomic_dec(&tl->pin_count); }
+static struct i915_vma *create_ring_vma(struct i915_ggtt *ggtt, int size) +{
struct i915_address_space *vm = &ggtt->vm;struct drm_i915_private *i915 = vm->i915;struct drm_i915_gem_object *obj;struct i915_vma *vma;obj = i915_gem_object_create_internal(i915, size);if (IS_ERR(obj))return ERR_CAST(obj);
We didn't want to use the dummy object here also? I guess meh?
vma = i915_vma_instance(obj, vm, NULL);if (IS_ERR(vma))goto err;return vma;+err:
i915_gem_object_put(obj);return vma;+}
static struct intel_ring *mock_ring(struct intel_engine_cs *engine) {
const unsigned long sz = PAGE_SIZE / 2;
const unsigned long sz = PAGE_SIZE;
Is that significant?
Reviewed-by: Matthew Auld matthew.auld@intel.com
struct intel_ring *ring; ring = kzalloc(sizeof(*ring) + sz, GFP_KERNEL);@@ -50,15 +72,11 @@ static struct intel_ring *mock_ring(struct intel_engine_cs *engine) ring->vaddr = (void *)(ring + 1); atomic_set(&ring->pin_count, 1);
ring->vma = i915_vma_alloc();if (!ring->vma) {
ring->vma = create_ring_vma(engine->gt->ggtt, PAGE_SIZE);if (IS_ERR(ring->vma)) { kfree(ring); return NULL; }
i915_active_init(&ring->vma->active, NULL, NULL, 0);__set_bit(I915_VMA_GGTT_BIT, __i915_vma_flags(ring->vma));__set_bit(DRM_MM_NODE_ALLOCATED_BIT, &ring->vma->node.flags);ring->vma->node.size = sz; intel_ring_update_space(ring);@@ -67,8 +85,7 @@ static struct intel_ring *mock_ring(struct intel_engine_cs *engine)
static void mock_ring_free(struct intel_ring *ring) {
i915_active_fini(&ring->vma->active);i915_vma_free(ring->vma);
i915_vma_put(ring->vma); kfree(ring);} @@ -125,6 +142,7 @@ static void mock_context_unpin(struct intel_context *ce)
static void mock_context_post_unpin(struct intel_context *ce) {
i915_vma_unpin(ce->ring->vma);}
static void mock_context_destroy(struct kref *ref) @@ -169,7 +187,7 @@ static int mock_context_alloc(struct intel_context *ce) static int mock_context_pre_pin(struct intel_context *ce, struct i915_gem_ww_ctx *ww, void **unused) {
return 0;
return i915_vma_pin_ww(ce->ring->vma, ww, 0, 0, PIN_GLOBAL | PIN_HIGH);}
static int mock_context_pin(struct intel_context *ce, void *unused)
2.33.0
Op 21-10-2021 om 17:57 schreef Matthew Auld:
On Thu, 21 Oct 2021 at 11:36, Maarten Lankhorst maarten.lankhorst@linux.intel.com wrote:
This allows us to finally get rid of all the assumptions that vma->obj is NULL.
Changes since v1:
- Ensure the mock_ring vma is pinned to prevent a fault.
- Pin it high to avoid failure in evict_for_vma selftest.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
drivers/gpu/drm/i915/gt/mock_engine.c | 38 ++++++++++++++++++++------- 1 file changed, 28 insertions(+), 10 deletions(-)
diff --git a/drivers/gpu/drm/i915/gt/mock_engine.c b/drivers/gpu/drm/i915/gt/mock_engine.c index 8b89215afe46..bb99fc03f503 100644 --- a/drivers/gpu/drm/i915/gt/mock_engine.c +++ b/drivers/gpu/drm/i915/gt/mock_engine.c @@ -35,9 +35,31 @@ static void mock_timeline_unpin(struct intel_timeline *tl) atomic_dec(&tl->pin_count); }
+static struct i915_vma *create_ring_vma(struct i915_ggtt *ggtt, int size) +{
struct i915_address_space *vm = &ggtt->vm;struct drm_i915_private *i915 = vm->i915;struct drm_i915_gem_object *obj;struct i915_vma *vma;obj = i915_gem_object_create_internal(i915, size);if (IS_ERR(obj))return ERR_CAST(obj);We didn't want to use the dummy object here also? I guess meh?
vma = i915_vma_instance(obj, vm, NULL);if (IS_ERR(vma))goto err;return vma;+err:
i915_gem_object_put(obj);return vma;+}
static struct intel_ring *mock_ring(struct intel_engine_cs *engine) {
const unsigned long sz = PAGE_SIZE / 2;
const unsigned long sz = PAGE_SIZE;Is that significant?
Reviewed-by: Matthew Auld matthew.auld@intel.com
vma->node.size has to be page aligned, since we create an actual vma, it was required to bump the size.
~Maarten
struct intel_ring *ring; ring = kzalloc(sizeof(*ring) + sz, GFP_KERNEL);@@ -50,15 +72,11 @@ static struct intel_ring *mock_ring(struct intel_engine_cs *engine) ring->vaddr = (void *)(ring + 1); atomic_set(&ring->pin_count, 1);
ring->vma = i915_vma_alloc();if (!ring->vma) {
ring->vma = create_ring_vma(engine->gt->ggtt, PAGE_SIZE);if (IS_ERR(ring->vma)) { kfree(ring); return NULL; }
i915_active_init(&ring->vma->active, NULL, NULL, 0);__set_bit(I915_VMA_GGTT_BIT, __i915_vma_flags(ring->vma));__set_bit(DRM_MM_NODE_ALLOCATED_BIT, &ring->vma->node.flags);ring->vma->node.size = sz; intel_ring_update_space(ring);@@ -67,8 +85,7 @@ static struct intel_ring *mock_ring(struct intel_engine_cs *engine)
static void mock_ring_free(struct intel_ring *ring) {
i915_active_fini(&ring->vma->active);i915_vma_free(ring->vma);
i915_vma_put(ring->vma); kfree(ring);} @@ -125,6 +142,7 @@ static void mock_context_unpin(struct intel_context *ce)
static void mock_context_post_unpin(struct intel_context *ce) {
i915_vma_unpin(ce->ring->vma);}
static void mock_context_destroy(struct kref *ref) @@ -169,7 +187,7 @@ static int mock_context_alloc(struct intel_context *ce) static int mock_context_pre_pin(struct intel_context *ce, struct i915_gem_ww_ctx *ww, void **unused) {
return 0;
return i915_vma_pin_ww(ce->ring->vma, ww, 0, 0, PIN_GLOBAL | PIN_HIGH);}
static int mock_context_pin(struct intel_context *ce, void *unused)
2.33.0
vma->obj and vma->resv are now never NULL, and some checks can be removed.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com --- drivers/gpu/drm/i915/gt/intel_context.c | 2 +- .../gpu/drm/i915/gt/intel_ring_submission.c | 2 +- drivers/gpu/drm/i915/i915_vma.c | 48 ++++++++----------- drivers/gpu/drm/i915/i915_vma.h | 3 -- 4 files changed, 22 insertions(+), 33 deletions(-)
diff --git a/drivers/gpu/drm/i915/gt/intel_context.c b/drivers/gpu/drm/i915/gt/intel_context.c index 5634d14052bc..e0220ac0e9b6 100644 --- a/drivers/gpu/drm/i915/gt/intel_context.c +++ b/drivers/gpu/drm/i915/gt/intel_context.c @@ -219,7 +219,7 @@ int __intel_context_do_pin_ww(struct intel_context *ce, */
err = i915_gem_object_lock(ce->timeline->hwsp_ggtt->obj, ww); - if (!err && ce->ring->vma->obj) + if (!err) err = i915_gem_object_lock(ce->ring->vma->obj, ww); if (!err && ce->state) err = i915_gem_object_lock(ce->state->obj, ww); diff --git a/drivers/gpu/drm/i915/gt/intel_ring_submission.c b/drivers/gpu/drm/i915/gt/intel_ring_submission.c index 586dca1731ce..3e6fac0340ef 100644 --- a/drivers/gpu/drm/i915/gt/intel_ring_submission.c +++ b/drivers/gpu/drm/i915/gt/intel_ring_submission.c @@ -1357,7 +1357,7 @@ int intel_ring_submission_setup(struct intel_engine_cs *engine) err = i915_gem_object_lock(timeline->hwsp_ggtt->obj, &ww); if (!err && gen7_wa_vma) err = i915_gem_object_lock(gen7_wa_vma->obj, &ww); - if (!err && engine->legacy.ring->vma->obj) + if (!err) err = i915_gem_object_lock(engine->legacy.ring->vma->obj, &ww); if (!err) err = intel_timeline_pin(timeline, &ww); diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c index 1187f1956c20..aebfc232b58b 100644 --- a/drivers/gpu/drm/i915/i915_vma.c +++ b/drivers/gpu/drm/i915/i915_vma.c @@ -40,12 +40,12 @@
static struct kmem_cache *slab_vmas;
-struct i915_vma *i915_vma_alloc(void) +static struct i915_vma *i915_vma_alloc(void) { return kmem_cache_zalloc(slab_vmas, GFP_KERNEL); }
-void i915_vma_free(struct i915_vma *vma) +static void i915_vma_free(struct i915_vma *vma) { return kmem_cache_free(slab_vmas, vma); } @@ -426,10 +426,8 @@ int i915_vma_bind(struct i915_vma *vma,
work->base.dma.error = 0; /* enable the queue_work() */
- if (vma->obj) { - __i915_gem_object_pin_pages(vma->obj); - work->pinned = i915_gem_object_get(vma->obj); - } + __i915_gem_object_pin_pages(vma->obj); + work->pinned = i915_gem_object_get(vma->obj); } else { vma->ops->bind_vma(vma->vm, NULL, vma, cache_level, bind_flags); } @@ -670,7 +668,7 @@ i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags) }
color = 0; - if (vma->obj && i915_vm_has_cache_coloring(vma->vm)) + if (i915_vm_has_cache_coloring(vma->vm)) color = vma->obj->cache_level;
if (flags & PIN_OFFSET_FIXED) { @@ -795,17 +793,14 @@ static bool try_qad_pin(struct i915_vma *vma, unsigned int flags) static int vma_get_pages(struct i915_vma *vma) { int err = 0; - bool pinned_pages = false; + bool pinned_pages = true;
if (atomic_add_unless(&vma->pages_count, 1, 0)) return 0;
- if (vma->obj) { - err = i915_gem_object_pin_pages(vma->obj); - if (err) - return err; - pinned_pages = true; - } + err = i915_gem_object_pin_pages(vma->obj); + if (err) + return err;
/* Allocations ahoy! */ if (mutex_lock_interruptible(&vma->pages_mutex)) { @@ -838,8 +833,8 @@ static void __vma_put_pages(struct i915_vma *vma, unsigned int count) if (atomic_sub_return(count, &vma->pages_count) == 0) { vma->ops->clear_pages(vma); GEM_BUG_ON(vma->pages); - if (vma->obj) - i915_gem_object_unpin_pages(vma->obj); + + i915_gem_object_unpin_pages(vma->obj); } mutex_unlock(&vma->pages_mutex); } @@ -875,7 +870,7 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, int err;
#ifdef CONFIG_PROVE_LOCKING - if (debug_locks && !WARN_ON(!ww) && vma->resv) + if (debug_locks && !WARN_ON(!ww)) assert_vma_held(vma); #endif
@@ -983,7 +978,7 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
GEM_BUG_ON(!vma->pages); err = i915_vma_bind(vma, - vma->obj ? vma->obj->cache_level : 0, + vma->obj->cache_level, flags, work); if (err) goto err_remove; @@ -1037,7 +1032,7 @@ int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, GEM_BUG_ON(!i915_vma_is_ggtt(vma));
#ifdef CONFIG_LOCKDEP - WARN_ON(!ww && vma->resv && dma_resv_held(vma->resv)); + WARN_ON(!ww && dma_resv_held(vma->resv)); #endif
do { @@ -1116,6 +1111,7 @@ void i915_vma_reopen(struct i915_vma *vma) void i915_vma_release(struct kref *ref) { struct i915_vma *vma = container_of(ref, typeof(*vma), ref); + struct drm_i915_gem_object *obj = vma->obj;
if (drm_mm_node_allocated(&vma->node)) { mutex_lock(&vma->vm->mutex); @@ -1126,15 +1122,11 @@ void i915_vma_release(struct kref *ref) } GEM_BUG_ON(i915_vma_is_active(vma));
- if (vma->obj) { - struct drm_i915_gem_object *obj = vma->obj; - - spin_lock(&obj->vma.lock); - list_del(&vma->obj_link); - if (!RB_EMPTY_NODE(&vma->obj_node)) - rb_erase(&vma->obj_node, &obj->vma.tree); - spin_unlock(&obj->vma.lock); - } + spin_lock(&obj->vma.lock); + list_del(&vma->obj_link); + if (!RB_EMPTY_NODE(&vma->obj_node)) + rb_erase(&vma->obj_node, &obj->vma.tree); + spin_unlock(&obj->vma.lock);
__i915_vma_remove_closed(vma); i915_vm_put(vma->vm); diff --git a/drivers/gpu/drm/i915/i915_vma.h b/drivers/gpu/drm/i915/i915_vma.h index b882fd7b5f99..423e0df81c87 100644 --- a/drivers/gpu/drm/i915/i915_vma.h +++ b/drivers/gpu/drm/i915/i915_vma.h @@ -416,9 +416,6 @@ static inline void i915_vma_clear_scanout(struct i915_vma *vma) list_for_each_entry(V, &(OBJ)->vma.list, obj_link) \ for_each_until(!i915_vma_is_ggtt(V))
-struct i915_vma *i915_vma_alloc(void); -void i915_vma_free(struct i915_vma *vma); - struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma); void i915_vma_make_shrinkable(struct i915_vma *vma); void i915_vma_make_purgeable(struct i915_vma *vma);
On Thu, 21 Oct 2021 at 11:37, Maarten Lankhorst maarten.lankhorst@linux.intel.com wrote:
vma->obj and vma->resv are now never NULL, and some checks can be removed.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
drivers/gpu/drm/i915/gt/intel_context.c | 2 +- .../gpu/drm/i915/gt/intel_ring_submission.c | 2 +- drivers/gpu/drm/i915/i915_vma.c | 48 ++++++++----------- drivers/gpu/drm/i915/i915_vma.h | 3 -- 4 files changed, 22 insertions(+), 33 deletions(-)
diff --git a/drivers/gpu/drm/i915/gt/intel_context.c b/drivers/gpu/drm/i915/gt/intel_context.c index 5634d14052bc..e0220ac0e9b6 100644 --- a/drivers/gpu/drm/i915/gt/intel_context.c +++ b/drivers/gpu/drm/i915/gt/intel_context.c @@ -219,7 +219,7 @@ int __intel_context_do_pin_ww(struct intel_context *ce, */
err = i915_gem_object_lock(ce->timeline->hwsp_ggtt->obj, ww);
if (!err && ce->ring->vma->obj)
if (!err) err = i915_gem_object_lock(ce->ring->vma->obj, ww); if (!err && ce->state) err = i915_gem_object_lock(ce->state->obj, ww);diff --git a/drivers/gpu/drm/i915/gt/intel_ring_submission.c b/drivers/gpu/drm/i915/gt/intel_ring_submission.c index 586dca1731ce..3e6fac0340ef 100644 --- a/drivers/gpu/drm/i915/gt/intel_ring_submission.c +++ b/drivers/gpu/drm/i915/gt/intel_ring_submission.c @@ -1357,7 +1357,7 @@ int intel_ring_submission_setup(struct intel_engine_cs *engine) err = i915_gem_object_lock(timeline->hwsp_ggtt->obj, &ww); if (!err && gen7_wa_vma) err = i915_gem_object_lock(gen7_wa_vma->obj, &ww);
if (!err && engine->legacy.ring->vma->obj)
if (!err) err = i915_gem_object_lock(engine->legacy.ring->vma->obj, &ww); if (!err) err = intel_timeline_pin(timeline, &ww);diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c index 1187f1956c20..aebfc232b58b 100644 --- a/drivers/gpu/drm/i915/i915_vma.c +++ b/drivers/gpu/drm/i915/i915_vma.c @@ -40,12 +40,12 @@
static struct kmem_cache *slab_vmas;
-struct i915_vma *i915_vma_alloc(void) +static struct i915_vma *i915_vma_alloc(void) { return kmem_cache_zalloc(slab_vmas, GFP_KERNEL); }
-void i915_vma_free(struct i915_vma *vma) +static void i915_vma_free(struct i915_vma *vma) {
I assume this belongs in the previous patch?
Otherwise, Reviewed-by: Matthew Auld matthew.auld@intel.com
return kmem_cache_free(slab_vmas, vma);} @@ -426,10 +426,8 @@ int i915_vma_bind(struct i915_vma *vma,
work->base.dma.error = 0; /* enable the queue_work() */
if (vma->obj) {__i915_gem_object_pin_pages(vma->obj);work->pinned = i915_gem_object_get(vma->obj);}
__i915_gem_object_pin_pages(vma->obj);work->pinned = i915_gem_object_get(vma->obj); } else { vma->ops->bind_vma(vma->vm, NULL, vma, cache_level, bind_flags); }@@ -670,7 +668,7 @@ i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags) }
color = 0;
if (vma->obj && i915_vm_has_cache_coloring(vma->vm))
if (i915_vm_has_cache_coloring(vma->vm)) color = vma->obj->cache_level; if (flags & PIN_OFFSET_FIXED) {@@ -795,17 +793,14 @@ static bool try_qad_pin(struct i915_vma *vma, unsigned int flags) static int vma_get_pages(struct i915_vma *vma) { int err = 0;
bool pinned_pages = false;
bool pinned_pages = true; if (atomic_add_unless(&vma->pages_count, 1, 0)) return 0;
if (vma->obj) {err = i915_gem_object_pin_pages(vma->obj);if (err)return err;pinned_pages = true;}
err = i915_gem_object_pin_pages(vma->obj);if (err)return err; /* Allocations ahoy! */ if (mutex_lock_interruptible(&vma->pages_mutex)) {@@ -838,8 +833,8 @@ static void __vma_put_pages(struct i915_vma *vma, unsigned int count) if (atomic_sub_return(count, &vma->pages_count) == 0) { vma->ops->clear_pages(vma); GEM_BUG_ON(vma->pages);
if (vma->obj)i915_gem_object_unpin_pages(vma->obj);
i915_gem_object_unpin_pages(vma->obj); } mutex_unlock(&vma->pages_mutex);} @@ -875,7 +870,7 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, int err;
#ifdef CONFIG_PROVE_LOCKING
if (debug_locks && !WARN_ON(!ww) && vma->resv)
if (debug_locks && !WARN_ON(!ww)) assert_vma_held(vma);#endif
@@ -983,7 +978,7 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
GEM_BUG_ON(!vma->pages); err = i915_vma_bind(vma,
vma->obj ? vma->obj->cache_level : 0,
vma->obj->cache_level, flags, work); if (err) goto err_remove;@@ -1037,7 +1032,7 @@ int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, GEM_BUG_ON(!i915_vma_is_ggtt(vma));
#ifdef CONFIG_LOCKDEP
WARN_ON(!ww && vma->resv && dma_resv_held(vma->resv));
WARN_ON(!ww && dma_resv_held(vma->resv));#endif
do {@@ -1116,6 +1111,7 @@ void i915_vma_reopen(struct i915_vma *vma) void i915_vma_release(struct kref *ref) { struct i915_vma *vma = container_of(ref, typeof(*vma), ref);
struct drm_i915_gem_object *obj = vma->obj; if (drm_mm_node_allocated(&vma->node)) { mutex_lock(&vma->vm->mutex);@@ -1126,15 +1122,11 @@ void i915_vma_release(struct kref *ref) } GEM_BUG_ON(i915_vma_is_active(vma));
if (vma->obj) {struct drm_i915_gem_object *obj = vma->obj;spin_lock(&obj->vma.lock);list_del(&vma->obj_link);if (!RB_EMPTY_NODE(&vma->obj_node))rb_erase(&vma->obj_node, &obj->vma.tree);spin_unlock(&obj->vma.lock);}
spin_lock(&obj->vma.lock);list_del(&vma->obj_link);if (!RB_EMPTY_NODE(&vma->obj_node))rb_erase(&vma->obj_node, &obj->vma.tree);spin_unlock(&obj->vma.lock); __i915_vma_remove_closed(vma); i915_vm_put(vma->vm);diff --git a/drivers/gpu/drm/i915/i915_vma.h b/drivers/gpu/drm/i915/i915_vma.h index b882fd7b5f99..423e0df81c87 100644 --- a/drivers/gpu/drm/i915/i915_vma.h +++ b/drivers/gpu/drm/i915/i915_vma.h @@ -416,9 +416,6 @@ static inline void i915_vma_clear_scanout(struct i915_vma *vma) list_for_each_entry(V, &(OBJ)->vma.list, obj_link) \ for_each_until(!i915_vma_is_ggtt(V))
-struct i915_vma *i915_vma_alloc(void); -void i915_vma_free(struct i915_vma *vma);
struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma); void i915_vma_make_shrinkable(struct i915_vma *vma); void i915_vma_make_purgeable(struct i915_vma *vma); -- 2.33.0
Call drop_pages with the gem object lock held, instead of the other way around. This will allow us to drop the vma bindings with the gem object lock held.
We plan to require the object lock for unpinning in the future, and this is an easy target.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com Reviewed-by: Niranjana Vishwanathapura niranjana.vishwanathapura@intel.com --- drivers/gpu/drm/i915/gem/i915_gem_shrinker.c | 42 ++++++++++---------- 1 file changed, 22 insertions(+), 20 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c b/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c index af3eb7fd951d..d3f29a66cb36 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c @@ -36,8 +36,8 @@ static bool can_release_pages(struct drm_i915_gem_object *obj) return swap_available() || obj->mm.madv == I915_MADV_DONTNEED; }
-static bool unsafe_drop_pages(struct drm_i915_gem_object *obj, - unsigned long shrink, bool trylock_vm) +static int drop_pages(struct drm_i915_gem_object *obj, + unsigned long shrink, bool trylock_vm) { unsigned long flags;
@@ -208,26 +208,28 @@ i915_gem_shrink(struct i915_gem_ww_ctx *ww,
spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
- err = 0; - if (unsafe_drop_pages(obj, shrink, trylock_vm)) { - /* May arrive from get_pages on another bo */ - if (!ww) { - if (!i915_gem_object_trylock(obj)) - goto skip; - } else { - err = i915_gem_object_lock(obj, ww); - if (err) - goto skip; - } - - if (!__i915_gem_object_put_pages(obj)) { - try_to_writeback(obj, shrink); - count += obj->base.size >> PAGE_SHIFT; - } - if (!ww) - i915_gem_object_unlock(obj); + /* May arrive from get_pages on another bo */ + if (!ww) { + if (!i915_gem_object_trylock(obj)) + goto skip; + } else { + err = i915_gem_object_lock(obj, ww); + if (err) + goto skip; }
+ if (drop_pages(obj, shrink, trylock_vm) && + !__i915_gem_object_put_pages(obj)) { + try_to_writeback(obj, shrink); + count += obj->base.size >> PAGE_SHIFT; + } + + if (dma_resv_test_signaled(obj->base.resv, true)) + dma_resv_add_excl_fence(obj->base.resv, NULL); + + if (!ww) + i915_gem_object_unlock(obj); + scanned += obj->base.size >> PAGE_SHIFT; skip: i915_gem_object_put(obj);
On Thu, 21 Oct 2021 at 11:37, Maarten Lankhorst maarten.lankhorst@linux.intel.com wrote:
Call drop_pages with the gem object lock held, instead of the other way around. This will allow us to drop the vma bindings with the gem object lock held.
We plan to require the object lock for unpinning in the future, and this is an easy target.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com Reviewed-by: Niranjana Vishwanathapura niranjana.vishwanathapura@intel.com
drivers/gpu/drm/i915/gem/i915_gem_shrinker.c | 42 ++++++++++---------- 1 file changed, 22 insertions(+), 20 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c b/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c index af3eb7fd951d..d3f29a66cb36 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c @@ -36,8 +36,8 @@ static bool can_release_pages(struct drm_i915_gem_object *obj) return swap_available() || obj->mm.madv == I915_MADV_DONTNEED; }
-static bool unsafe_drop_pages(struct drm_i915_gem_object *obj,
unsigned long shrink, bool trylock_vm)+static int drop_pages(struct drm_i915_gem_object *obj,
unsigned long shrink, bool trylock_vm){ unsigned long flags;
@@ -208,26 +208,28 @@ i915_gem_shrink(struct i915_gem_ww_ctx *ww,
spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
err = 0;if (unsafe_drop_pages(obj, shrink, trylock_vm)) {/* May arrive from get_pages on another bo */if (!ww) {if (!i915_gem_object_trylock(obj))goto skip;} else {err = i915_gem_object_lock(obj, ww);if (err)goto skip;}if (!__i915_gem_object_put_pages(obj)) {try_to_writeback(obj, shrink);count += obj->base.size >> PAGE_SHIFT;}if (!ww)i915_gem_object_unlock(obj);
/* May arrive from get_pages on another bo */if (!ww) {if (!i915_gem_object_trylock(obj))goto skip;} else {err = i915_gem_object_lock(obj, ww);if (err)goto skip; }if (drop_pages(obj, shrink, trylock_vm) &&!__i915_gem_object_put_pages(obj)) {try_to_writeback(obj, shrink);count += obj->base.size >> PAGE_SHIFT;}if (dma_resv_test_signaled(obj->base.resv, true))dma_resv_add_excl_fence(obj->base.resv, NULL);
I assume we want to rip out resv_prune here in the series, or something? Instead of randomly adding this back here.
if (!ww)i915_gem_object_unlock(obj);scanned += obj->base.size >> PAGE_SHIFT;skip: i915_gem_object_put(obj); -- 2.33.0
Op 21-10-2021 om 18:12 schreef Matthew Auld:
On Thu, 21 Oct 2021 at 11:37, Maarten Lankhorst maarten.lankhorst@linux.intel.com wrote:
Call drop_pages with the gem object lock held, instead of the other way around. This will allow us to drop the vma bindings with the gem object lock held.
We plan to require the object lock for unpinning in the future, and this is an easy target.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com Reviewed-by: Niranjana Vishwanathapura niranjana.vishwanathapura@intel.com
drivers/gpu/drm/i915/gem/i915_gem_shrinker.c | 42 ++++++++++---------- 1 file changed, 22 insertions(+), 20 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c b/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c index af3eb7fd951d..d3f29a66cb36 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c @@ -36,8 +36,8 @@ static bool can_release_pages(struct drm_i915_gem_object *obj) return swap_available() || obj->mm.madv == I915_MADV_DONTNEED; }
-static bool unsafe_drop_pages(struct drm_i915_gem_object *obj,
unsigned long shrink, bool trylock_vm)+static int drop_pages(struct drm_i915_gem_object *obj,
unsigned long shrink, bool trylock_vm){ unsigned long flags;
@@ -208,26 +208,28 @@ i915_gem_shrink(struct i915_gem_ww_ctx *ww,
spin_unlock_irqrestore(&i915->mm.obj_lock, flags);
err = 0;if (unsafe_drop_pages(obj, shrink, trylock_vm)) {/* May arrive from get_pages on another bo */if (!ww) {if (!i915_gem_object_trylock(obj))goto skip;} else {err = i915_gem_object_lock(obj, ww);if (err)goto skip;}if (!__i915_gem_object_put_pages(obj)) {try_to_writeback(obj, shrink);count += obj->base.size >> PAGE_SHIFT;}if (!ww)i915_gem_object_unlock(obj);
/* May arrive from get_pages on another bo */if (!ww) {if (!i915_gem_object_trylock(obj))goto skip;} else {err = i915_gem_object_lock(obj, ww);if (err)goto skip; }if (drop_pages(obj, shrink, trylock_vm) &&!__i915_gem_object_put_pages(obj)) {try_to_writeback(obj, shrink);count += obj->base.size >> PAGE_SHIFT;}if (dma_resv_test_signaled(obj->base.resv, true))dma_resv_add_excl_fence(obj->base.resv, NULL);I assume we want to rip out resv_prune here in the series, or something? Instead of randomly adding this back here.
Oh yeah, this hunk can be removed safely. It's stale and shouldn't be here. :)
if (!ww)i915_gem_object_unlock(obj);scanned += obj->base.size >> PAGE_SHIFT;skip: i915_gem_object_put(obj); -- 2.33.0
Resetting will clear the CONTEXT_VALID_BIT, so wait until after that to test.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com Reviewed-by: Niranjana Vishwanathapura niranjana.vishwanathapura@intel.com --- drivers/gpu/drm/i915/gt/intel_engine_pm.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)
diff --git a/drivers/gpu/drm/i915/gt/intel_engine_pm.c b/drivers/gpu/drm/i915/gt/intel_engine_pm.c index a1334b48dde7..849fbb229bd3 100644 --- a/drivers/gpu/drm/i915/gt/intel_engine_pm.c +++ b/drivers/gpu/drm/i915/gt/intel_engine_pm.c @@ -52,8 +52,6 @@ static int __engine_unpark(struct intel_wakeref *wf) /* Discard stale context state from across idling */ ce = engine->kernel_context; if (ce) { - GEM_BUG_ON(test_bit(CONTEXT_VALID_BIT, &ce->flags)); - /* Flush all pending HW writes before we touch the context */ while (unlikely(intel_context_inflight(ce))) intel_engine_flush_submission(engine); @@ -68,6 +66,9 @@ static int __engine_unpark(struct intel_wakeref *wf) ce->timeline->seqno, READ_ONCE(*ce->timeline->hwsp_seqno), ce->ring->emit); + + GEM_BUG_ON(test_bit(CONTEXT_VALID_BIT, &ce->flags)); + GEM_BUG_ON(ce->timeline->seqno != READ_ONCE(*ce->timeline->hwsp_seqno)); }
On Thu, 21 Oct 2021 at 11:37, Maarten Lankhorst maarten.lankhorst@linux.intel.com wrote:
Resetting will clear the CONTEXT_VALID_BIT, so wait until after that to test.
AFAIK this seems to be fixing something earlier in the series(maybe patch 7?) i.e without this patch we seem to trigger the BUG_ON. If so, this needs to be much earlier in the series?
Also this probably needs some more commentary in the commit message for whether moving the BUG_ON matters, or if this is potentially papering over something significant?
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com Reviewed-by: Niranjana Vishwanathapura niranjana.vishwanathapura@intel.com
drivers/gpu/drm/i915/gt/intel_engine_pm.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)
diff --git a/drivers/gpu/drm/i915/gt/intel_engine_pm.c b/drivers/gpu/drm/i915/gt/intel_engine_pm.c index a1334b48dde7..849fbb229bd3 100644 --- a/drivers/gpu/drm/i915/gt/intel_engine_pm.c +++ b/drivers/gpu/drm/i915/gt/intel_engine_pm.c @@ -52,8 +52,6 @@ static int __engine_unpark(struct intel_wakeref *wf) /* Discard stale context state from across idling */ ce = engine->kernel_context; if (ce) {
GEM_BUG_ON(test_bit(CONTEXT_VALID_BIT, &ce->flags));/* Flush all pending HW writes before we touch the context */ while (unlikely(intel_context_inflight(ce))) intel_engine_flush_submission(engine);@@ -68,6 +66,9 @@ static int __engine_unpark(struct intel_wakeref *wf) ce->timeline->seqno, READ_ONCE(*ce->timeline->hwsp_seqno), ce->ring->emit);
GEM_BUG_ON(test_bit(CONTEXT_VALID_BIT, &ce->flags));GEM_BUG_ON(ce->timeline->seqno != READ_ONCE(*ce->timeline->hwsp_seqno)); }-- 2.33.0
It's just an alias to vma->obj->base.resv, no need to duplicate it.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com Reviewed-by: Niranjana Vishwanathapura niranjana.vishwanathapura@intel.com --- drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c | 4 ++-- drivers/gpu/drm/i915/i915_vma.c | 9 ++++----- drivers/gpu/drm/i915/i915_vma.h | 6 +++--- drivers/gpu/drm/i915/i915_vma_types.h | 1 - 4 files changed, 9 insertions(+), 11 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c index eaacadd2d2e5..e614591ca510 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c @@ -1007,7 +1007,7 @@ static int eb_validate_vmas(struct i915_execbuffer *eb) }
if (!(ev->flags & EXEC_OBJECT_WRITE)) { - err = dma_resv_reserve_shared(vma->resv, 1); + err = dma_resv_reserve_shared(vma->obj->base.resv, 1); if (err) return err; } @@ -2173,7 +2173,7 @@ static int eb_parse(struct i915_execbuffer *eb) goto err_trampoline; }
- err = dma_resv_reserve_shared(shadow->resv, 1); + err = dma_resv_reserve_shared(shadow->obj->base.resv, 1); if (err) goto err_trampoline;
diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c index aebfc232b58b..ac09b685678a 100644 --- a/drivers/gpu/drm/i915/i915_vma.c +++ b/drivers/gpu/drm/i915/i915_vma.c @@ -116,7 +116,6 @@ vma_create(struct drm_i915_gem_object *obj, vma->vm = i915_vm_get(vm); vma->ops = &vm->vma_ops; vma->obj = obj; - vma->resv = obj->base.resv; vma->size = obj->base.size; vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
@@ -1032,7 +1031,7 @@ int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, GEM_BUG_ON(!i915_vma_is_ggtt(vma));
#ifdef CONFIG_LOCKDEP - WARN_ON(!ww && dma_resv_held(vma->resv)); + WARN_ON(!ww && dma_resv_held(vma->obj->base.resv)); #endif
do { @@ -1251,19 +1250,19 @@ int _i915_vma_move_to_active(struct i915_vma *vma, }
if (fence) { - dma_resv_add_excl_fence(vma->resv, fence); + dma_resv_add_excl_fence(vma->obj->base.resv, fence); obj->write_domain = I915_GEM_DOMAIN_RENDER; obj->read_domains = 0; } } else { if (!(flags & __EXEC_OBJECT_NO_RESERVE)) { - err = dma_resv_reserve_shared(vma->resv, 1); + err = dma_resv_reserve_shared(vma->obj->base.resv, 1); if (unlikely(err)) return err; }
if (fence) { - dma_resv_add_shared_fence(vma->resv, fence); + dma_resv_add_shared_fence(vma->obj->base.resv, fence); obj->write_domain = 0; } } diff --git a/drivers/gpu/drm/i915/i915_vma.h b/drivers/gpu/drm/i915/i915_vma.h index 423e0df81c87..9a931ecb09e5 100644 --- a/drivers/gpu/drm/i915/i915_vma.h +++ b/drivers/gpu/drm/i915/i915_vma.h @@ -232,16 +232,16 @@ static inline void __i915_vma_put(struct i915_vma *vma) kref_put(&vma->ref, i915_vma_release); }
-#define assert_vma_held(vma) dma_resv_assert_held((vma)->resv) +#define assert_vma_held(vma) dma_resv_assert_held((vma)->obj->base.resv)
static inline void i915_vma_lock(struct i915_vma *vma) { - dma_resv_lock(vma->resv, NULL); + dma_resv_lock(vma->obj->base.resv, NULL); }
static inline void i915_vma_unlock(struct i915_vma *vma) { - dma_resv_unlock(vma->resv); + dma_resv_unlock(vma->obj->base.resv); }
int __must_check diff --git a/drivers/gpu/drm/i915/i915_vma_types.h b/drivers/gpu/drm/i915/i915_vma_types.h index 80e93bf00f2e..8a0decb19bcc 100644 --- a/drivers/gpu/drm/i915/i915_vma_types.h +++ b/drivers/gpu/drm/i915/i915_vma_types.h @@ -178,7 +178,6 @@ struct i915_vma { const struct i915_vma_ops *ops;
struct drm_i915_gem_object *obj; - struct dma_resv *resv; /** Alias of obj->resv */
struct sg_table *pages; void __iomem *iomap;
Big delta, but boils down to moving set_pages to i915_vma.c, and removing the special handling, all callers use the defaults anyway. We only remap in ggtt, so default case will fall through.
Because we still don't require locking in i915_vma_unpin(), handle this by using xchg in get_pages(), as it's locked with obj->mutex, and cmpxchg in unpin, which only fails if we race a against a new pin.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com --- drivers/gpu/drm/i915/display/intel_dpt.c | 2 - drivers/gpu/drm/i915/gt/gen6_ppgtt.c | 15 - drivers/gpu/drm/i915/gt/intel_ggtt.c | 345 ---------------- drivers/gpu/drm/i915/gt/intel_gtt.c | 13 - drivers/gpu/drm/i915/gt/intel_gtt.h | 7 - drivers/gpu/drm/i915/gt/intel_ppgtt.c | 12 - drivers/gpu/drm/i915/i915_vma.c | 388 ++++++++++++++++-- drivers/gpu/drm/i915/i915_vma.h | 3 + drivers/gpu/drm/i915/i915_vma_types.h | 1 - drivers/gpu/drm/i915/selftests/i915_gem_gtt.c | 12 +- drivers/gpu/drm/i915/selftests/mock_gtt.c | 4 - 11 files changed, 368 insertions(+), 434 deletions(-)
diff --git a/drivers/gpu/drm/i915/display/intel_dpt.c b/drivers/gpu/drm/i915/display/intel_dpt.c index 8f7b1f7534a4..ef428f3fc538 100644 --- a/drivers/gpu/drm/i915/display/intel_dpt.c +++ b/drivers/gpu/drm/i915/display/intel_dpt.c @@ -221,8 +221,6 @@ intel_dpt_create(struct intel_framebuffer *fb)
vm->vma_ops.bind_vma = dpt_bind_vma; vm->vma_ops.unbind_vma = dpt_unbind_vma; - vm->vma_ops.set_pages = ggtt_set_pages; - vm->vma_ops.clear_pages = clear_pages;
vm->pte_encode = gen8_ggtt_pte_encode;
diff --git a/drivers/gpu/drm/i915/gt/gen6_ppgtt.c b/drivers/gpu/drm/i915/gt/gen6_ppgtt.c index 5caa1703716e..5c048b4ccd4d 100644 --- a/drivers/gpu/drm/i915/gt/gen6_ppgtt.c +++ b/drivers/gpu/drm/i915/gt/gen6_ppgtt.c @@ -270,19 +270,6 @@ static void gen6_ppgtt_cleanup(struct i915_address_space *vm) free_pd(&ppgtt->base.vm, ppgtt->base.pd); }
-static int pd_vma_set_pages(struct i915_vma *vma) -{ - vma->pages = ERR_PTR(-ENODEV); - return 0; -} - -static void pd_vma_clear_pages(struct i915_vma *vma) -{ - GEM_BUG_ON(!vma->pages); - - vma->pages = NULL; -} - static void pd_vma_bind(struct i915_address_space *vm, struct i915_vm_pt_stash *stash, struct i915_vma *vma, @@ -322,8 +309,6 @@ static void pd_vma_unbind(struct i915_address_space *vm, struct i915_vma *vma) }
static const struct i915_vma_ops pd_vma_ops = { - .set_pages = pd_vma_set_pages, - .clear_pages = pd_vma_clear_pages, .bind_vma = pd_vma_bind, .unbind_vma = pd_vma_unbind, }; diff --git a/drivers/gpu/drm/i915/gt/intel_ggtt.c b/drivers/gpu/drm/i915/gt/intel_ggtt.c index f17383e76eb7..6da57199bb33 100644 --- a/drivers/gpu/drm/i915/gt/intel_ggtt.c +++ b/drivers/gpu/drm/i915/gt/intel_ggtt.c @@ -20,9 +20,6 @@ #include "intel_gtt.h" #include "gen8_ppgtt.h"
-static int -i915_get_ggtt_vma_pages(struct i915_vma *vma); - static void i915_ggtt_color_adjust(const struct drm_mm_node *node, unsigned long color, u64 *start, @@ -875,21 +872,6 @@ static int ggtt_probe_common(struct i915_ggtt *ggtt, u64 size) return 0; }
-int ggtt_set_pages(struct i915_vma *vma) -{ - int ret; - - GEM_BUG_ON(vma->pages); - - ret = i915_get_ggtt_vma_pages(vma); - if (ret) - return ret; - - vma->page_sizes = vma->obj->mm.page_sizes; - - return 0; -} - static void gen6_gmch_remove(struct i915_address_space *vm) { struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm); @@ -950,8 +932,6 @@ static int gen8_gmch_probe(struct i915_ggtt *ggtt)
ggtt->vm.vma_ops.bind_vma = ggtt_bind_vma; ggtt->vm.vma_ops.unbind_vma = ggtt_unbind_vma; - ggtt->vm.vma_ops.set_pages = ggtt_set_pages; - ggtt->vm.vma_ops.clear_pages = clear_pages;
ggtt->vm.pte_encode = gen8_ggtt_pte_encode;
@@ -1100,8 +1080,6 @@ static int gen6_gmch_probe(struct i915_ggtt *ggtt)
ggtt->vm.vma_ops.bind_vma = ggtt_bind_vma; ggtt->vm.vma_ops.unbind_vma = ggtt_unbind_vma; - ggtt->vm.vma_ops.set_pages = ggtt_set_pages; - ggtt->vm.vma_ops.clear_pages = clear_pages;
return ggtt_probe_common(ggtt, size); } @@ -1145,8 +1123,6 @@ static int i915_gmch_probe(struct i915_ggtt *ggtt)
ggtt->vm.vma_ops.bind_vma = ggtt_bind_vma; ggtt->vm.vma_ops.unbind_vma = ggtt_unbind_vma; - ggtt->vm.vma_ops.set_pages = ggtt_set_pages; - ggtt->vm.vma_ops.clear_pages = clear_pages;
if (unlikely(ggtt->do_idle_maps)) drm_notice(&i915->drm, @@ -1294,324 +1270,3 @@ void i915_ggtt_resume(struct i915_ggtt *ggtt)
intel_ggtt_restore_fences(ggtt); } - -static struct scatterlist * -rotate_pages(struct drm_i915_gem_object *obj, unsigned int offset, - unsigned int width, unsigned int height, - unsigned int src_stride, unsigned int dst_stride, - struct sg_table *st, struct scatterlist *sg) -{ - unsigned int column, row; - unsigned int src_idx; - - for (column = 0; column < width; column++) { - unsigned int left; - - src_idx = src_stride * (height - 1) + column + offset; - for (row = 0; row < height; row++) { - st->nents++; - /* - * We don't need the pages, but need to initialize - * the entries so the sg list can be happily traversed. - * The only thing we need are DMA addresses. - */ - sg_set_page(sg, NULL, I915_GTT_PAGE_SIZE, 0); - sg_dma_address(sg) = - i915_gem_object_get_dma_address(obj, src_idx); - sg_dma_len(sg) = I915_GTT_PAGE_SIZE; - sg = sg_next(sg); - src_idx -= src_stride; - } - - left = (dst_stride - height) * I915_GTT_PAGE_SIZE; - - if (!left) - continue; - - st->nents++; - - /* - * The DE ignores the PTEs for the padding tiles, the sg entry - * here is just a conenience to indicate how many padding PTEs - * to insert at this spot. - */ - sg_set_page(sg, NULL, left, 0); - sg_dma_address(sg) = 0; - sg_dma_len(sg) = left; - sg = sg_next(sg); - } - - return sg; -} - -static noinline struct sg_table * -intel_rotate_pages(struct intel_rotation_info *rot_info, - struct drm_i915_gem_object *obj) -{ - unsigned int size = intel_rotation_info_size(rot_info); - struct drm_i915_private *i915 = to_i915(obj->base.dev); - struct sg_table *st; - struct scatterlist *sg; - int ret = -ENOMEM; - int i; - - /* Allocate target SG list. */ - st = kmalloc(sizeof(*st), GFP_KERNEL); - if (!st) - goto err_st_alloc; - - ret = sg_alloc_table(st, size, GFP_KERNEL); - if (ret) - goto err_sg_alloc; - - st->nents = 0; - sg = st->sgl; - - for (i = 0 ; i < ARRAY_SIZE(rot_info->plane); i++) - sg = rotate_pages(obj, rot_info->plane[i].offset, - rot_info->plane[i].width, rot_info->plane[i].height, - rot_info->plane[i].src_stride, - rot_info->plane[i].dst_stride, - st, sg); - - return st; - -err_sg_alloc: - kfree(st); -err_st_alloc: - - drm_dbg(&i915->drm, "Failed to create rotated mapping for object size %zu! (%ux%u tiles, %u pages)\n", - obj->base.size, rot_info->plane[0].width, - rot_info->plane[0].height, size); - - return ERR_PTR(ret); -} - -static struct scatterlist * -remap_pages(struct drm_i915_gem_object *obj, - unsigned int offset, unsigned int alignment_pad, - unsigned int width, unsigned int height, - unsigned int src_stride, unsigned int dst_stride, - struct sg_table *st, struct scatterlist *sg) -{ - unsigned int row; - - if (alignment_pad) { - st->nents++; - - /* - * The DE ignores the PTEs for the padding tiles, the sg entry - * here is just a convenience to indicate how many padding PTEs - * to insert at this spot. - */ - sg_set_page(sg, NULL, alignment_pad * 4096, 0); - sg_dma_address(sg) = 0; - sg_dma_len(sg) = alignment_pad * 4096; - sg = sg_next(sg); - } - - for (row = 0; row < height; row++) { - unsigned int left = width * I915_GTT_PAGE_SIZE; - - while (left) { - dma_addr_t addr; - unsigned int length; - - /* - * We don't need the pages, but need to initialize - * the entries so the sg list can be happily traversed. - * The only thing we need are DMA addresses. - */ - - addr = i915_gem_object_get_dma_address_len(obj, offset, &length); - - length = min(left, length); - - st->nents++; - - sg_set_page(sg, NULL, length, 0); - sg_dma_address(sg) = addr; - sg_dma_len(sg) = length; - sg = sg_next(sg); - - offset += length / I915_GTT_PAGE_SIZE; - left -= length; - } - - offset += src_stride - width; - - left = (dst_stride - width) * I915_GTT_PAGE_SIZE; - - if (!left) - continue; - - st->nents++; - - /* - * The DE ignores the PTEs for the padding tiles, the sg entry - * here is just a conenience to indicate how many padding PTEs - * to insert at this spot. - */ - sg_set_page(sg, NULL, left, 0); - sg_dma_address(sg) = 0; - sg_dma_len(sg) = left; - sg = sg_next(sg); - } - - return sg; -} - -static noinline struct sg_table * -intel_remap_pages(struct intel_remapped_info *rem_info, - struct drm_i915_gem_object *obj) -{ - unsigned int size = intel_remapped_info_size(rem_info); - struct drm_i915_private *i915 = to_i915(obj->base.dev); - struct sg_table *st; - struct scatterlist *sg; - unsigned int gtt_offset = 0; - int ret = -ENOMEM; - int i; - - /* Allocate target SG list. */ - st = kmalloc(sizeof(*st), GFP_KERNEL); - if (!st) - goto err_st_alloc; - - ret = sg_alloc_table(st, size, GFP_KERNEL); - if (ret) - goto err_sg_alloc; - - st->nents = 0; - sg = st->sgl; - - for (i = 0 ; i < ARRAY_SIZE(rem_info->plane); i++) { - unsigned int alignment_pad = 0; - - if (rem_info->plane_alignment) - alignment_pad = ALIGN(gtt_offset, rem_info->plane_alignment) - gtt_offset; - - sg = remap_pages(obj, - rem_info->plane[i].offset, alignment_pad, - rem_info->plane[i].width, rem_info->plane[i].height, - rem_info->plane[i].src_stride, rem_info->plane[i].dst_stride, - st, sg); - - gtt_offset += alignment_pad + - rem_info->plane[i].dst_stride * rem_info->plane[i].height; - } - - i915_sg_trim(st); - - return st; - -err_sg_alloc: - kfree(st); -err_st_alloc: - - drm_dbg(&i915->drm, "Failed to create remapped mapping for object size %zu! (%ux%u tiles, %u pages)\n", - obj->base.size, rem_info->plane[0].width, - rem_info->plane[0].height, size); - - return ERR_PTR(ret); -} - -static noinline struct sg_table * -intel_partial_pages(const struct i915_ggtt_view *view, - struct drm_i915_gem_object *obj) -{ - struct sg_table *st; - struct scatterlist *sg, *iter; - unsigned int count = view->partial.size; - unsigned int offset; - int ret = -ENOMEM; - - st = kmalloc(sizeof(*st), GFP_KERNEL); - if (!st) - goto err_st_alloc; - - ret = sg_alloc_table(st, count, GFP_KERNEL); - if (ret) - goto err_sg_alloc; - - iter = i915_gem_object_get_sg_dma(obj, view->partial.offset, &offset); - GEM_BUG_ON(!iter); - - sg = st->sgl; - st->nents = 0; - do { - unsigned int len; - - len = min(sg_dma_len(iter) - (offset << PAGE_SHIFT), - count << PAGE_SHIFT); - sg_set_page(sg, NULL, len, 0); - sg_dma_address(sg) = - sg_dma_address(iter) + (offset << PAGE_SHIFT); - sg_dma_len(sg) = len; - - st->nents++; - count -= len >> PAGE_SHIFT; - if (count == 0) { - sg_mark_end(sg); - i915_sg_trim(st); /* Drop any unused tail entries. */ - - return st; - } - - sg = __sg_next(sg); - iter = __sg_next(iter); - offset = 0; - } while (1); - -err_sg_alloc: - kfree(st); -err_st_alloc: - return ERR_PTR(ret); -} - -static int -i915_get_ggtt_vma_pages(struct i915_vma *vma) -{ - int ret; - - /* - * The vma->pages are only valid within the lifespan of the borrowed - * obj->mm.pages. When the obj->mm.pages sg_table is regenerated, so - * must be the vma->pages. A simple rule is that vma->pages must only - * be accessed when the obj->mm.pages are pinned. - */ - GEM_BUG_ON(!i915_gem_object_has_pinned_pages(vma->obj)); - - switch (vma->ggtt_view.type) { - default: - GEM_BUG_ON(vma->ggtt_view.type); - fallthrough; - case I915_GGTT_VIEW_NORMAL: - vma->pages = vma->obj->mm.pages; - return 0; - - case I915_GGTT_VIEW_ROTATED: - vma->pages = - intel_rotate_pages(&vma->ggtt_view.rotated, vma->obj); - break; - - case I915_GGTT_VIEW_REMAPPED: - vma->pages = - intel_remap_pages(&vma->ggtt_view.remapped, vma->obj); - break; - - case I915_GGTT_VIEW_PARTIAL: - vma->pages = intel_partial_pages(&vma->ggtt_view, vma->obj); - break; - } - - ret = 0; - if (IS_ERR(vma->pages)) { - ret = PTR_ERR(vma->pages); - vma->pages = NULL; - drm_err(&vma->vm->i915->drm, - "Failed to get pages for VMA view type %u (%d)!\n", - vma->ggtt_view.type, ret); - } - return ret; -} diff --git a/drivers/gpu/drm/i915/gt/intel_gtt.c b/drivers/gpu/drm/i915/gt/intel_gtt.c index 67d14afa6623..12eed5fcb17a 100644 --- a/drivers/gpu/drm/i915/gt/intel_gtt.c +++ b/drivers/gpu/drm/i915/gt/intel_gtt.c @@ -221,19 +221,6 @@ void i915_address_space_init(struct i915_address_space *vm, int subclass) INIT_LIST_HEAD(&vm->bound_list); }
-void clear_pages(struct i915_vma *vma) -{ - GEM_BUG_ON(!vma->pages); - - if (vma->pages != vma->obj->mm.pages) { - sg_free_table(vma->pages); - kfree(vma->pages); - } - vma->pages = NULL; - - memset(&vma->page_sizes, 0, sizeof(vma->page_sizes)); -} - void *__px_vaddr(struct drm_i915_gem_object *p) { enum i915_map_type type; diff --git a/drivers/gpu/drm/i915/gt/intel_gtt.h b/drivers/gpu/drm/i915/gt/intel_gtt.h index bc6750263359..306e7645fdc5 100644 --- a/drivers/gpu/drm/i915/gt/intel_gtt.h +++ b/drivers/gpu/drm/i915/gt/intel_gtt.h @@ -206,9 +206,6 @@ struct i915_vma_ops { */ void (*unbind_vma)(struct i915_address_space *vm, struct i915_vma *vma); - - int (*set_pages)(struct i915_vma *vma); - void (*clear_pages)(struct i915_vma *vma); };
struct i915_address_space { @@ -594,10 +591,6 @@ release_pd_entry(struct i915_page_directory * const pd, const struct drm_i915_gem_object * const scratch); void gen6_ggtt_invalidate(struct i915_ggtt *ggtt);
-int ggtt_set_pages(struct i915_vma *vma); -int ppgtt_set_pages(struct i915_vma *vma); -void clear_pages(struct i915_vma *vma); - void ppgtt_bind_vma(struct i915_address_space *vm, struct i915_vm_pt_stash *stash, struct i915_vma *vma, diff --git a/drivers/gpu/drm/i915/gt/intel_ppgtt.c b/drivers/gpu/drm/i915/gt/intel_ppgtt.c index 4396bfd630d8..083b3090c69c 100644 --- a/drivers/gpu/drm/i915/gt/intel_ppgtt.c +++ b/drivers/gpu/drm/i915/gt/intel_ppgtt.c @@ -289,16 +289,6 @@ void i915_vm_free_pt_stash(struct i915_address_space *vm, } }
-int ppgtt_set_pages(struct i915_vma *vma) -{ - GEM_BUG_ON(vma->pages); - - vma->pages = vma->obj->mm.pages; - vma->page_sizes = vma->obj->mm.page_sizes; - - return 0; -} - void ppgtt_init(struct i915_ppgtt *ppgtt, struct intel_gt *gt, unsigned long lmem_pt_obj_flags) { @@ -315,6 +305,4 @@ void ppgtt_init(struct i915_ppgtt *ppgtt, struct intel_gt *gt,
ppgtt->vm.vma_ops.bind_vma = ppgtt_bind_vma; ppgtt->vm.vma_ops.unbind_vma = ppgtt_unbind_vma; - ppgtt->vm.vma_ops.set_pages = ppgtt_set_pages; - ppgtt->vm.vma_ops.clear_pages = clear_pages; } diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c index ac09b685678a..bacc8d68e495 100644 --- a/drivers/gpu/drm/i915/i915_vma.c +++ b/drivers/gpu/drm/i915/i915_vma.c @@ -112,7 +112,6 @@ vma_create(struct drm_i915_gem_object *obj, return ERR_PTR(-ENOMEM);
kref_init(&vma->ref); - mutex_init(&vma->pages_mutex); vma->vm = i915_vm_get(vm); vma->ops = &vm->vma_ops; vma->obj = obj; @@ -789,10 +788,343 @@ static bool try_qad_pin(struct i915_vma *vma, unsigned int flags) return pinned; }
-static int vma_get_pages(struct i915_vma *vma) + + +static struct scatterlist * +rotate_pages(struct drm_i915_gem_object *obj, unsigned int offset, + unsigned int width, unsigned int height, + unsigned int src_stride, unsigned int dst_stride, + struct sg_table *st, struct scatterlist *sg) { - int err = 0; - bool pinned_pages = true; + unsigned int column, row; + unsigned int src_idx; + + for (column = 0; column < width; column++) { + unsigned int left; + + src_idx = src_stride * (height - 1) + column + offset; + for (row = 0; row < height; row++) { + st->nents++; + /* + * We don't need the pages, but need to initialize + * the entries so the sg list can be happily traversed. + * The only thing we need are DMA addresses. + */ + sg_set_page(sg, NULL, I915_GTT_PAGE_SIZE, 0); + sg_dma_address(sg) = + i915_gem_object_get_dma_address(obj, src_idx); + sg_dma_len(sg) = I915_GTT_PAGE_SIZE; + sg = sg_next(sg); + src_idx -= src_stride; + } + + left = (dst_stride - height) * I915_GTT_PAGE_SIZE; + + if (!left) + continue; + + st->nents++; + + /* + * The DE ignores the PTEs for the padding tiles, the sg entry + * here is just a conenience to indicate how many padding PTEs + * to insert at this spot. + */ + sg_set_page(sg, NULL, left, 0); + sg_dma_address(sg) = 0; + sg_dma_len(sg) = left; + sg = sg_next(sg); + } + + return sg; +} + +static noinline struct sg_table * +intel_rotate_pages(struct intel_rotation_info *rot_info, + struct drm_i915_gem_object *obj) +{ + unsigned int size = intel_rotation_info_size(rot_info); + struct drm_i915_private *i915 = to_i915(obj->base.dev); + struct sg_table *st; + struct scatterlist *sg; + int ret = -ENOMEM; + int i; + + /* Allocate target SG list. */ + st = kmalloc(sizeof(*st), GFP_KERNEL); + if (!st) + goto err_st_alloc; + + ret = sg_alloc_table(st, size, GFP_KERNEL); + if (ret) + goto err_sg_alloc; + + st->nents = 0; + sg = st->sgl; + + for (i = 0 ; i < ARRAY_SIZE(rot_info->plane); i++) + sg = rotate_pages(obj, rot_info->plane[i].offset, + rot_info->plane[i].width, rot_info->plane[i].height, + rot_info->plane[i].src_stride, + rot_info->plane[i].dst_stride, + st, sg); + + return st; + +err_sg_alloc: + kfree(st); +err_st_alloc: + + drm_dbg(&i915->drm, "Failed to create rotated mapping for object size %zu! (%ux%u tiles, %u pages)\n", + obj->base.size, rot_info->plane[0].width, + rot_info->plane[0].height, size); + + return ERR_PTR(ret); +} + +static struct scatterlist * +remap_pages(struct drm_i915_gem_object *obj, + unsigned int offset, unsigned int alignment_pad, + unsigned int width, unsigned int height, + unsigned int src_stride, unsigned int dst_stride, + struct sg_table *st, struct scatterlist *sg) +{ + unsigned int row; + + if (alignment_pad) { + st->nents++; + + /* + * The DE ignores the PTEs for the padding tiles, the sg entry + * here is just a convenience to indicate how many padding PTEs + * to insert at this spot. + */ + sg_set_page(sg, NULL, alignment_pad * 4096, 0); + sg_dma_address(sg) = 0; + sg_dma_len(sg) = alignment_pad * 4096; + sg = sg_next(sg); + } + + for (row = 0; row < height; row++) { + unsigned int left = width * I915_GTT_PAGE_SIZE; + + while (left) { + dma_addr_t addr; + unsigned int length; + + /* + * We don't need the pages, but need to initialize + * the entries so the sg list can be happily traversed. + * The only thing we need are DMA addresses. + */ + + addr = i915_gem_object_get_dma_address_len(obj, offset, &length); + + length = min(left, length); + + st->nents++; + + sg_set_page(sg, NULL, length, 0); + sg_dma_address(sg) = addr; + sg_dma_len(sg) = length; + sg = sg_next(sg); + + offset += length / I915_GTT_PAGE_SIZE; + left -= length; + } + + offset += src_stride - width; + + left = (dst_stride - width) * I915_GTT_PAGE_SIZE; + + if (!left) + continue; + + st->nents++; + + /* + * The DE ignores the PTEs for the padding tiles, the sg entry + * here is just a conenience to indicate how many padding PTEs + * to insert at this spot. + */ + sg_set_page(sg, NULL, left, 0); + sg_dma_address(sg) = 0; + sg_dma_len(sg) = left; + sg = sg_next(sg); + } + + return sg; +} + +static noinline struct sg_table * +intel_remap_pages(struct intel_remapped_info *rem_info, + struct drm_i915_gem_object *obj) +{ + unsigned int size = intel_remapped_info_size(rem_info); + struct drm_i915_private *i915 = to_i915(obj->base.dev); + struct sg_table *st; + struct scatterlist *sg; + unsigned int gtt_offset = 0; + int ret = -ENOMEM; + int i; + + /* Allocate target SG list. */ + st = kmalloc(sizeof(*st), GFP_KERNEL); + if (!st) + goto err_st_alloc; + + ret = sg_alloc_table(st, size, GFP_KERNEL); + if (ret) + goto err_sg_alloc; + + st->nents = 0; + sg = st->sgl; + + for (i = 0 ; i < ARRAY_SIZE(rem_info->plane); i++) { + unsigned int alignment_pad = 0; + + if (rem_info->plane_alignment) + alignment_pad = ALIGN(gtt_offset, rem_info->plane_alignment) - gtt_offset; + + sg = remap_pages(obj, + rem_info->plane[i].offset, alignment_pad, + rem_info->plane[i].width, rem_info->plane[i].height, + rem_info->plane[i].src_stride, rem_info->plane[i].dst_stride, + st, sg); + + gtt_offset += alignment_pad + + rem_info->plane[i].dst_stride * rem_info->plane[i].height; + } + + i915_sg_trim(st); + + return st; + +err_sg_alloc: + kfree(st); +err_st_alloc: + + drm_dbg(&i915->drm, "Failed to create remapped mapping for object size %zu! (%ux%u tiles, %u pages)\n", + obj->base.size, rem_info->plane[0].width, + rem_info->plane[0].height, size); + + return ERR_PTR(ret); +} + +static noinline struct sg_table * +intel_partial_pages(const struct i915_ggtt_view *view, + struct drm_i915_gem_object *obj) +{ + struct sg_table *st; + struct scatterlist *sg, *iter; + unsigned int count = view->partial.size; + unsigned int offset; + int ret = -ENOMEM; + + st = kmalloc(sizeof(*st), GFP_KERNEL); + if (!st) + goto err_st_alloc; + + ret = sg_alloc_table(st, count, GFP_KERNEL); + if (ret) + goto err_sg_alloc; + + iter = i915_gem_object_get_sg_dma(obj, view->partial.offset, &offset); + GEM_BUG_ON(!iter); + + sg = st->sgl; + st->nents = 0; + do { + unsigned int len; + + len = min(sg_dma_len(iter) - (offset << PAGE_SHIFT), + count << PAGE_SHIFT); + sg_set_page(sg, NULL, len, 0); + sg_dma_address(sg) = + sg_dma_address(iter) + (offset << PAGE_SHIFT); + sg_dma_len(sg) = len; + + st->nents++; + count -= len >> PAGE_SHIFT; + if (count == 0) { + sg_mark_end(sg); + i915_sg_trim(st); /* Drop any unused tail entries. */ + + return st; + } + + sg = __sg_next(sg); + iter = __sg_next(iter); + offset = 0; + } while (1); + +err_sg_alloc: + kfree(st); +err_st_alloc: + return ERR_PTR(ret); +} + +static int +__i915_vma_get_pages(struct i915_vma *vma) +{ + struct sg_table *pages; + int ret; + + /* + * The vma->pages are only valid within the lifespan of the borrowed + * obj->mm.pages. When the obj->mm.pages sg_table is regenerated, so + * must be the vma->pages. A simple rule is that vma->pages must only + * be accessed when the obj->mm.pages are pinned. + */ + GEM_BUG_ON(!i915_gem_object_has_pinned_pages(vma->obj)); + + switch (vma->ggtt_view.type) { + default: + GEM_BUG_ON(vma->ggtt_view.type); + fallthrough; + case I915_GGTT_VIEW_NORMAL: + pages = vma->obj->mm.pages; + break; + + case I915_GGTT_VIEW_ROTATED: + pages = + intel_rotate_pages(&vma->ggtt_view.rotated, vma->obj); + break; + + case I915_GGTT_VIEW_REMAPPED: + pages = + intel_remap_pages(&vma->ggtt_view.remapped, vma->obj); + break; + + case I915_GGTT_VIEW_PARTIAL: + pages = intel_partial_pages(&vma->ggtt_view, vma->obj); + break; + } + + ret = 0; + /* gen6 ppgtt doesn't have backing pages, special-case it */ + if (IS_ERR(pages) && pages != ERR_PTR(-ENODEV)) { + ret = PTR_ERR(pages); + pages = NULL; + drm_err(&vma->vm->i915->drm, + "Failed to get pages for VMA view type %u (%d)!\n", + vma->ggtt_view.type, ret); + } + + pages = xchg(&vma->pages, pages); + + /* did we race against a put_pages? */ + if (pages && pages != vma->obj->mm.pages) { + sg_free_table(vma->pages); + kfree(vma->pages); + } + + return ret; +} + +I915_SELFTEST_EXPORT int i915_vma_get_pages(struct i915_vma *vma) +{ + int err;
if (atomic_add_unless(&vma->pages_count, 1, 0)) return 0; @@ -801,25 +1133,17 @@ static int vma_get_pages(struct i915_vma *vma) if (err) return err;
- /* Allocations ahoy! */ - if (mutex_lock_interruptible(&vma->pages_mutex)) { - err = -EINTR; - goto unpin; - } + err = __i915_vma_get_pages(vma); + if (err) + goto err_unpin;
- if (!atomic_read(&vma->pages_count)) { - err = vma->ops->set_pages(vma); - if (err) - goto unlock; - pinned_pages = false; - } + vma->page_sizes = vma->obj->mm.page_sizes; atomic_inc(&vma->pages_count);
-unlock: - mutex_unlock(&vma->pages_mutex); -unpin: - if (pinned_pages) - __i915_gem_object_unpin_pages(vma->obj); + return 0; + +err_unpin: + __i915_gem_object_unpin_pages(vma->obj);
return err; } @@ -827,18 +1151,22 @@ static int vma_get_pages(struct i915_vma *vma) static void __vma_put_pages(struct i915_vma *vma, unsigned int count) { /* We allocate under vma_get_pages, so beware the shrinker */ - mutex_lock_nested(&vma->pages_mutex, SINGLE_DEPTH_NESTING); + struct sg_table *pages = READ_ONCE(vma->pages); + GEM_BUG_ON(atomic_read(&vma->pages_count) < count); + if (atomic_sub_return(count, &vma->pages_count) == 0) { - vma->ops->clear_pages(vma); - GEM_BUG_ON(vma->pages); + if (pages == cmpxchg(&vma->pages, pages, NULL) && + pages != vma->obj->mm.pages) { + sg_free_table(pages); + kfree(pages); + }
i915_gem_object_unpin_pages(vma->obj); } - mutex_unlock(&vma->pages_mutex); }
-static void vma_put_pages(struct i915_vma *vma) +I915_SELFTEST_EXPORT void i915_vma_put_pages(struct i915_vma *vma) { if (atomic_add_unless(&vma->pages_count, -1, 1)) return; @@ -868,10 +1196,8 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, unsigned int bound; int err;
-#ifdef CONFIG_PROVE_LOCKING - if (debug_locks && !WARN_ON(!ww)) - assert_vma_held(vma); -#endif + assert_vma_held(vma); + GEM_BUG_ON(!ww);
BUILD_BUG_ON(PIN_GLOBAL != I915_VMA_GLOBAL_BIND); BUILD_BUG_ON(PIN_USER != I915_VMA_LOCAL_BIND); @@ -882,7 +1208,7 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, if (try_qad_pin(vma, flags & I915_VMA_BIND_MASK)) return 0;
- err = vma_get_pages(vma); + err = i915_vma_get_pages(vma); if (err) return err;
@@ -1007,7 +1333,7 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, err_rpm: if (wakeref) intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref); - vma_put_pages(vma); + i915_vma_put_pages(vma); return err; }
diff --git a/drivers/gpu/drm/i915/i915_vma.h b/drivers/gpu/drm/i915/i915_vma.h index 9a931ecb09e5..32719431b3df 100644 --- a/drivers/gpu/drm/i915/i915_vma.h +++ b/drivers/gpu/drm/i915/i915_vma.h @@ -431,4 +431,7 @@ static inline int i915_vma_sync(struct i915_vma *vma) void i915_vma_module_exit(void); int i915_vma_module_init(void);
+I915_SELFTEST_DECLARE(int i915_vma_get_pages(struct i915_vma *vma)); +I915_SELFTEST_DECLARE(void i915_vma_put_pages(struct i915_vma *vma)); + #endif diff --git a/drivers/gpu/drm/i915/i915_vma_types.h b/drivers/gpu/drm/i915/i915_vma_types.h index 8a0decb19bcc..412d3e78a30d 100644 --- a/drivers/gpu/drm/i915/i915_vma_types.h +++ b/drivers/gpu/drm/i915/i915_vma_types.h @@ -261,7 +261,6 @@ struct i915_vma { #define I915_VMA_PAGES_BIAS 24 #define I915_VMA_PAGES_ACTIVE (BIT(24) | 1) atomic_t pages_count; /* number of active binds to the pages */ - struct mutex pages_mutex; /* protect acquire/release of backing pages */
/** * Support different GGTT views into the same object. diff --git a/drivers/gpu/drm/i915/selftests/i915_gem_gtt.c b/drivers/gpu/drm/i915/selftests/i915_gem_gtt.c index 46f4236039a9..4574fb51b656 100644 --- a/drivers/gpu/drm/i915/selftests/i915_gem_gtt.c +++ b/drivers/gpu/drm/i915/selftests/i915_gem_gtt.c @@ -1953,7 +1953,9 @@ static int igt_cs_tlb(void *arg) goto end; }
- err = vma->ops->set_pages(vma); + i915_gem_object_lock(bbe, NULL); + err = i915_vma_get_pages(vma); + i915_gem_object_unlock(bbe); if (err) goto end;
@@ -1994,7 +1996,7 @@ static int igt_cs_tlb(void *arg) i915_request_put(rq); }
- vma->ops->clear_pages(vma); + i915_vma_put_pages(vma);
err = context_sync(ce); if (err) { @@ -2009,7 +2011,9 @@ static int igt_cs_tlb(void *arg) goto end; }
- err = vma->ops->set_pages(vma); + i915_gem_object_lock(act, NULL); + err = i915_vma_get_pages(vma); + i915_gem_object_unlock(act); if (err) goto end;
@@ -2047,7 +2051,7 @@ static int igt_cs_tlb(void *arg) } end_spin(batch, count - 1);
- vma->ops->clear_pages(vma); + i915_vma_put_pages(vma);
err = context_sync(ce); if (err) { diff --git a/drivers/gpu/drm/i915/selftests/mock_gtt.c b/drivers/gpu/drm/i915/selftests/mock_gtt.c index cc047ec594f9..096679014d99 100644 --- a/drivers/gpu/drm/i915/selftests/mock_gtt.c +++ b/drivers/gpu/drm/i915/selftests/mock_gtt.c @@ -86,8 +86,6 @@ struct i915_ppgtt *mock_ppgtt(struct drm_i915_private *i915, const char *name)
ppgtt->vm.vma_ops.bind_vma = mock_bind_ppgtt; ppgtt->vm.vma_ops.unbind_vma = mock_unbind_ppgtt; - ppgtt->vm.vma_ops.set_pages = ppgtt_set_pages; - ppgtt->vm.vma_ops.clear_pages = clear_pages;
return ppgtt; } @@ -126,8 +124,6 @@ void mock_init_ggtt(struct drm_i915_private *i915, struct i915_ggtt *ggtt)
ggtt->vm.vma_ops.bind_vma = mock_bind_ggtt; ggtt->vm.vma_ops.unbind_vma = mock_unbind_ggtt; - ggtt->vm.vma_ops.set_pages = ggtt_set_pages; - ggtt->vm.vma_ops.clear_pages = clear_pages;
i915_address_space_init(&ggtt->vm, VM_CLASS_GGTT); i915->gt.ggtt = ggtt;
On Thu, 21 Oct 2021 at 11:36, Maarten Lankhorst maarten.lankhorst@linux.intel.com wrote:
Big delta, but boils down to moving set_pages to i915_vma.c, and removing the special handling, all callers use the defaults anyway. We only remap in ggtt, so default case will fall through.
Because we still don't require locking in i915_vma_unpin(), handle this by using xchg in get_pages(), as it's locked with obj->mutex, and cmpxchg in unpin, which only fails if we race a against a new pin.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
drivers/gpu/drm/i915/display/intel_dpt.c | 2 - drivers/gpu/drm/i915/gt/gen6_ppgtt.c | 15 - drivers/gpu/drm/i915/gt/intel_ggtt.c | 345 ---------------- drivers/gpu/drm/i915/gt/intel_gtt.c | 13 - drivers/gpu/drm/i915/gt/intel_gtt.h | 7 - drivers/gpu/drm/i915/gt/intel_ppgtt.c | 12 - drivers/gpu/drm/i915/i915_vma.c | 388 ++++++++++++++++-- drivers/gpu/drm/i915/i915_vma.h | 3 + drivers/gpu/drm/i915/i915_vma_types.h | 1 - drivers/gpu/drm/i915/selftests/i915_gem_gtt.c | 12 +- drivers/gpu/drm/i915/selftests/mock_gtt.c | 4 - 11 files changed, 368 insertions(+), 434 deletions(-)
diff --git a/drivers/gpu/drm/i915/display/intel_dpt.c b/drivers/gpu/drm/i915/display/intel_dpt.c index 8f7b1f7534a4..ef428f3fc538 100644 --- a/drivers/gpu/drm/i915/display/intel_dpt.c +++ b/drivers/gpu/drm/i915/display/intel_dpt.c @@ -221,8 +221,6 @@ intel_dpt_create(struct intel_framebuffer *fb)
vm->vma_ops.bind_vma = dpt_bind_vma; vm->vma_ops.unbind_vma = dpt_unbind_vma;
vm->vma_ops.set_pages = ggtt_set_pages;vm->vma_ops.clear_pages = clear_pages; vm->pte_encode = gen8_ggtt_pte_encode;diff --git a/drivers/gpu/drm/i915/gt/gen6_ppgtt.c b/drivers/gpu/drm/i915/gt/gen6_ppgtt.c index 5caa1703716e..5c048b4ccd4d 100644 --- a/drivers/gpu/drm/i915/gt/gen6_ppgtt.c +++ b/drivers/gpu/drm/i915/gt/gen6_ppgtt.c @@ -270,19 +270,6 @@ static void gen6_ppgtt_cleanup(struct i915_address_space *vm) free_pd(&ppgtt->base.vm, ppgtt->base.pd); }
-static int pd_vma_set_pages(struct i915_vma *vma) -{
vma->pages = ERR_PTR(-ENODEV);return 0;-}
-static void pd_vma_clear_pages(struct i915_vma *vma) -{
GEM_BUG_ON(!vma->pages);vma->pages = NULL;-}
static void pd_vma_bind(struct i915_address_space *vm, struct i915_vm_pt_stash *stash, struct i915_vma *vma, @@ -322,8 +309,6 @@ static void pd_vma_unbind(struct i915_address_space *vm, struct i915_vma *vma) }
static const struct i915_vma_ops pd_vma_ops = {
.set_pages = pd_vma_set_pages,.clear_pages = pd_vma_clear_pages, .bind_vma = pd_vma_bind, .unbind_vma = pd_vma_unbind,}; diff --git a/drivers/gpu/drm/i915/gt/intel_ggtt.c b/drivers/gpu/drm/i915/gt/intel_ggtt.c index f17383e76eb7..6da57199bb33 100644 --- a/drivers/gpu/drm/i915/gt/intel_ggtt.c +++ b/drivers/gpu/drm/i915/gt/intel_ggtt.c @@ -20,9 +20,6 @@ #include "intel_gtt.h" #include "gen8_ppgtt.h"
-static int -i915_get_ggtt_vma_pages(struct i915_vma *vma);
static void i915_ggtt_color_adjust(const struct drm_mm_node *node, unsigned long color, u64 *start, @@ -875,21 +872,6 @@ static int ggtt_probe_common(struct i915_ggtt *ggtt, u64 size) return 0; }
-int ggtt_set_pages(struct i915_vma *vma) -{
int ret;GEM_BUG_ON(vma->pages);ret = i915_get_ggtt_vma_pages(vma);if (ret)return ret;vma->page_sizes = vma->obj->mm.page_sizes;return 0;-}
static void gen6_gmch_remove(struct i915_address_space *vm) { struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm); @@ -950,8 +932,6 @@ static int gen8_gmch_probe(struct i915_ggtt *ggtt)
ggtt->vm.vma_ops.bind_vma = ggtt_bind_vma; ggtt->vm.vma_ops.unbind_vma = ggtt_unbind_vma;
ggtt->vm.vma_ops.set_pages = ggtt_set_pages;ggtt->vm.vma_ops.clear_pages = clear_pages; ggtt->vm.pte_encode = gen8_ggtt_pte_encode;@@ -1100,8 +1080,6 @@ static int gen6_gmch_probe(struct i915_ggtt *ggtt)
ggtt->vm.vma_ops.bind_vma = ggtt_bind_vma; ggtt->vm.vma_ops.unbind_vma = ggtt_unbind_vma;
ggtt->vm.vma_ops.set_pages = ggtt_set_pages;ggtt->vm.vma_ops.clear_pages = clear_pages; return ggtt_probe_common(ggtt, size);} @@ -1145,8 +1123,6 @@ static int i915_gmch_probe(struct i915_ggtt *ggtt)
ggtt->vm.vma_ops.bind_vma = ggtt_bind_vma; ggtt->vm.vma_ops.unbind_vma = ggtt_unbind_vma;
ggtt->vm.vma_ops.set_pages = ggtt_set_pages;ggtt->vm.vma_ops.clear_pages = clear_pages; if (unlikely(ggtt->do_idle_maps)) drm_notice(&i915->drm,@@ -1294,324 +1270,3 @@ void i915_ggtt_resume(struct i915_ggtt *ggtt)
intel_ggtt_restore_fences(ggtt);}
-static struct scatterlist * -rotate_pages(struct drm_i915_gem_object *obj, unsigned int offset,
unsigned int width, unsigned int height,unsigned int src_stride, unsigned int dst_stride,struct sg_table *st, struct scatterlist *sg)-{
unsigned int column, row;unsigned int src_idx;for (column = 0; column < width; column++) {unsigned int left;src_idx = src_stride * (height - 1) + column + offset;for (row = 0; row < height; row++) {st->nents++;/** We don't need the pages, but need to initialize* the entries so the sg list can be happily traversed.* The only thing we need are DMA addresses.*/sg_set_page(sg, NULL, I915_GTT_PAGE_SIZE, 0);sg_dma_address(sg) =i915_gem_object_get_dma_address(obj, src_idx);sg_dma_len(sg) = I915_GTT_PAGE_SIZE;sg = sg_next(sg);src_idx -= src_stride;}left = (dst_stride - height) * I915_GTT_PAGE_SIZE;if (!left)continue;st->nents++;/** The DE ignores the PTEs for the padding tiles, the sg entry* here is just a conenience to indicate how many padding PTEs* to insert at this spot.*/sg_set_page(sg, NULL, left, 0);sg_dma_address(sg) = 0;sg_dma_len(sg) = left;sg = sg_next(sg);}return sg;-}
-static noinline struct sg_table * -intel_rotate_pages(struct intel_rotation_info *rot_info,
struct drm_i915_gem_object *obj)-{
unsigned int size = intel_rotation_info_size(rot_info);struct drm_i915_private *i915 = to_i915(obj->base.dev);struct sg_table *st;struct scatterlist *sg;int ret = -ENOMEM;int i;/* Allocate target SG list. */st = kmalloc(sizeof(*st), GFP_KERNEL);if (!st)goto err_st_alloc;ret = sg_alloc_table(st, size, GFP_KERNEL);if (ret)goto err_sg_alloc;st->nents = 0;sg = st->sgl;for (i = 0 ; i < ARRAY_SIZE(rot_info->plane); i++)sg = rotate_pages(obj, rot_info->plane[i].offset,rot_info->plane[i].width, rot_info->plane[i].height,rot_info->plane[i].src_stride,rot_info->plane[i].dst_stride,st, sg);return st;-err_sg_alloc:
kfree(st);-err_st_alloc:
drm_dbg(&i915->drm, "Failed to create rotated mapping for object size %zu! (%ux%u tiles, %u pages)\n",obj->base.size, rot_info->plane[0].width,rot_info->plane[0].height, size);return ERR_PTR(ret);-}
-static struct scatterlist * -remap_pages(struct drm_i915_gem_object *obj,
unsigned int offset, unsigned int alignment_pad,unsigned int width, unsigned int height,unsigned int src_stride, unsigned int dst_stride,struct sg_table *st, struct scatterlist *sg)-{
unsigned int row;if (alignment_pad) {st->nents++;/** The DE ignores the PTEs for the padding tiles, the sg entry* here is just a convenience to indicate how many padding PTEs* to insert at this spot.*/sg_set_page(sg, NULL, alignment_pad * 4096, 0);sg_dma_address(sg) = 0;sg_dma_len(sg) = alignment_pad * 4096;sg = sg_next(sg);}for (row = 0; row < height; row++) {unsigned int left = width * I915_GTT_PAGE_SIZE;while (left) {dma_addr_t addr;unsigned int length;/** We don't need the pages, but need to initialize* the entries so the sg list can be happily traversed.* The only thing we need are DMA addresses.*/addr = i915_gem_object_get_dma_address_len(obj, offset, &length);length = min(left, length);st->nents++;sg_set_page(sg, NULL, length, 0);sg_dma_address(sg) = addr;sg_dma_len(sg) = length;sg = sg_next(sg);offset += length / I915_GTT_PAGE_SIZE;left -= length;}offset += src_stride - width;left = (dst_stride - width) * I915_GTT_PAGE_SIZE;if (!left)continue;st->nents++;/** The DE ignores the PTEs for the padding tiles, the sg entry* here is just a conenience to indicate how many padding PTEs* to insert at this spot.*/sg_set_page(sg, NULL, left, 0);sg_dma_address(sg) = 0;sg_dma_len(sg) = left;sg = sg_next(sg);}return sg;-}
-static noinline struct sg_table * -intel_remap_pages(struct intel_remapped_info *rem_info,
struct drm_i915_gem_object *obj)-{
unsigned int size = intel_remapped_info_size(rem_info);struct drm_i915_private *i915 = to_i915(obj->base.dev);struct sg_table *st;struct scatterlist *sg;unsigned int gtt_offset = 0;int ret = -ENOMEM;int i;/* Allocate target SG list. */st = kmalloc(sizeof(*st), GFP_KERNEL);if (!st)goto err_st_alloc;ret = sg_alloc_table(st, size, GFP_KERNEL);if (ret)goto err_sg_alloc;st->nents = 0;sg = st->sgl;for (i = 0 ; i < ARRAY_SIZE(rem_info->plane); i++) {unsigned int alignment_pad = 0;if (rem_info->plane_alignment)alignment_pad = ALIGN(gtt_offset, rem_info->plane_alignment) - gtt_offset;sg = remap_pages(obj,rem_info->plane[i].offset, alignment_pad,rem_info->plane[i].width, rem_info->plane[i].height,rem_info->plane[i].src_stride, rem_info->plane[i].dst_stride,st, sg);gtt_offset += alignment_pad +rem_info->plane[i].dst_stride * rem_info->plane[i].height;}i915_sg_trim(st);return st;-err_sg_alloc:
kfree(st);-err_st_alloc:
drm_dbg(&i915->drm, "Failed to create remapped mapping for object size %zu! (%ux%u tiles, %u pages)\n",obj->base.size, rem_info->plane[0].width,rem_info->plane[0].height, size);return ERR_PTR(ret);-}
-static noinline struct sg_table * -intel_partial_pages(const struct i915_ggtt_view *view,
struct drm_i915_gem_object *obj)-{
struct sg_table *st;struct scatterlist *sg, *iter;unsigned int count = view->partial.size;unsigned int offset;int ret = -ENOMEM;st = kmalloc(sizeof(*st), GFP_KERNEL);if (!st)goto err_st_alloc;ret = sg_alloc_table(st, count, GFP_KERNEL);if (ret)goto err_sg_alloc;iter = i915_gem_object_get_sg_dma(obj, view->partial.offset, &offset);GEM_BUG_ON(!iter);sg = st->sgl;st->nents = 0;do {unsigned int len;len = min(sg_dma_len(iter) - (offset << PAGE_SHIFT),count << PAGE_SHIFT);sg_set_page(sg, NULL, len, 0);sg_dma_address(sg) =sg_dma_address(iter) + (offset << PAGE_SHIFT);sg_dma_len(sg) = len;st->nents++;count -= len >> PAGE_SHIFT;if (count == 0) {sg_mark_end(sg);i915_sg_trim(st); /* Drop any unused tail entries. */return st;}sg = __sg_next(sg);iter = __sg_next(iter);offset = 0;} while (1);-err_sg_alloc:
kfree(st);-err_st_alloc:
return ERR_PTR(ret);-}
-static int -i915_get_ggtt_vma_pages(struct i915_vma *vma) -{
int ret;/** The vma->pages are only valid within the lifespan of the borrowed* obj->mm.pages. When the obj->mm.pages sg_table is regenerated, so* must be the vma->pages. A simple rule is that vma->pages must only* be accessed when the obj->mm.pages are pinned.*/GEM_BUG_ON(!i915_gem_object_has_pinned_pages(vma->obj));switch (vma->ggtt_view.type) {default:GEM_BUG_ON(vma->ggtt_view.type);fallthrough;case I915_GGTT_VIEW_NORMAL:vma->pages = vma->obj->mm.pages;return 0;case I915_GGTT_VIEW_ROTATED:vma->pages =intel_rotate_pages(&vma->ggtt_view.rotated, vma->obj);break;case I915_GGTT_VIEW_REMAPPED:vma->pages =intel_remap_pages(&vma->ggtt_view.remapped, vma->obj);break;case I915_GGTT_VIEW_PARTIAL:vma->pages = intel_partial_pages(&vma->ggtt_view, vma->obj);break;}ret = 0;if (IS_ERR(vma->pages)) {ret = PTR_ERR(vma->pages);vma->pages = NULL;drm_err(&vma->vm->i915->drm,"Failed to get pages for VMA view type %u (%d)!\n",vma->ggtt_view.type, ret);}return ret;-} diff --git a/drivers/gpu/drm/i915/gt/intel_gtt.c b/drivers/gpu/drm/i915/gt/intel_gtt.c index 67d14afa6623..12eed5fcb17a 100644 --- a/drivers/gpu/drm/i915/gt/intel_gtt.c +++ b/drivers/gpu/drm/i915/gt/intel_gtt.c @@ -221,19 +221,6 @@ void i915_address_space_init(struct i915_address_space *vm, int subclass) INIT_LIST_HEAD(&vm->bound_list); }
-void clear_pages(struct i915_vma *vma) -{
GEM_BUG_ON(!vma->pages);if (vma->pages != vma->obj->mm.pages) {sg_free_table(vma->pages);kfree(vma->pages);}vma->pages = NULL;memset(&vma->page_sizes, 0, sizeof(vma->page_sizes));-}
void *__px_vaddr(struct drm_i915_gem_object *p) { enum i915_map_type type; diff --git a/drivers/gpu/drm/i915/gt/intel_gtt.h b/drivers/gpu/drm/i915/gt/intel_gtt.h index bc6750263359..306e7645fdc5 100644 --- a/drivers/gpu/drm/i915/gt/intel_gtt.h +++ b/drivers/gpu/drm/i915/gt/intel_gtt.h @@ -206,9 +206,6 @@ struct i915_vma_ops { */ void (*unbind_vma)(struct i915_address_space *vm, struct i915_vma *vma);
int (*set_pages)(struct i915_vma *vma);void (*clear_pages)(struct i915_vma *vma);};
struct i915_address_space { @@ -594,10 +591,6 @@ release_pd_entry(struct i915_page_directory * const pd, const struct drm_i915_gem_object * const scratch); void gen6_ggtt_invalidate(struct i915_ggtt *ggtt);
-int ggtt_set_pages(struct i915_vma *vma); -int ppgtt_set_pages(struct i915_vma *vma); -void clear_pages(struct i915_vma *vma);
void ppgtt_bind_vma(struct i915_address_space *vm, struct i915_vm_pt_stash *stash, struct i915_vma *vma, diff --git a/drivers/gpu/drm/i915/gt/intel_ppgtt.c b/drivers/gpu/drm/i915/gt/intel_ppgtt.c index 4396bfd630d8..083b3090c69c 100644 --- a/drivers/gpu/drm/i915/gt/intel_ppgtt.c +++ b/drivers/gpu/drm/i915/gt/intel_ppgtt.c @@ -289,16 +289,6 @@ void i915_vm_free_pt_stash(struct i915_address_space *vm, } }
-int ppgtt_set_pages(struct i915_vma *vma) -{
GEM_BUG_ON(vma->pages);vma->pages = vma->obj->mm.pages;vma->page_sizes = vma->obj->mm.page_sizes;return 0;-}
void ppgtt_init(struct i915_ppgtt *ppgtt, struct intel_gt *gt, unsigned long lmem_pt_obj_flags) { @@ -315,6 +305,4 @@ void ppgtt_init(struct i915_ppgtt *ppgtt, struct intel_gt *gt,
ppgtt->vm.vma_ops.bind_vma = ppgtt_bind_vma; ppgtt->vm.vma_ops.unbind_vma = ppgtt_unbind_vma;
ppgtt->vm.vma_ops.set_pages = ppgtt_set_pages;ppgtt->vm.vma_ops.clear_pages = clear_pages;} diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c index ac09b685678a..bacc8d68e495 100644 --- a/drivers/gpu/drm/i915/i915_vma.c +++ b/drivers/gpu/drm/i915/i915_vma.c @@ -112,7 +112,6 @@ vma_create(struct drm_i915_gem_object *obj, return ERR_PTR(-ENOMEM);
kref_init(&vma->ref);
mutex_init(&vma->pages_mutex); vma->vm = i915_vm_get(vm); vma->ops = &vm->vma_ops; vma->obj = obj;@@ -789,10 +788,343 @@ static bool try_qad_pin(struct i915_vma *vma, unsigned int flags) return pinned; }
-static int vma_get_pages(struct i915_vma *vma)
+static struct scatterlist * +rotate_pages(struct drm_i915_gem_object *obj, unsigned int offset,
unsigned int width, unsigned int height,unsigned int src_stride, unsigned int dst_stride,struct sg_table *st, struct scatterlist *sg){
int err = 0;bool pinned_pages = true;
unsigned int column, row;unsigned int src_idx;for (column = 0; column < width; column++) {unsigned int left;src_idx = src_stride * (height - 1) + column + offset;for (row = 0; row < height; row++) {st->nents++;/** We don't need the pages, but need to initialize* the entries so the sg list can be happily traversed.* The only thing we need are DMA addresses.*/sg_set_page(sg, NULL, I915_GTT_PAGE_SIZE, 0);sg_dma_address(sg) =i915_gem_object_get_dma_address(obj, src_idx);sg_dma_len(sg) = I915_GTT_PAGE_SIZE;sg = sg_next(sg);src_idx -= src_stride;}left = (dst_stride - height) * I915_GTT_PAGE_SIZE;if (!left)continue;st->nents++;/** The DE ignores the PTEs for the padding tiles, the sg entry* here is just a conenience to indicate how many padding PTEs* to insert at this spot.*/sg_set_page(sg, NULL, left, 0);sg_dma_address(sg) = 0;sg_dma_len(sg) = left;sg = sg_next(sg);}return sg;+}
+static noinline struct sg_table * +intel_rotate_pages(struct intel_rotation_info *rot_info,
struct drm_i915_gem_object *obj)+{
unsigned int size = intel_rotation_info_size(rot_info);struct drm_i915_private *i915 = to_i915(obj->base.dev);struct sg_table *st;struct scatterlist *sg;int ret = -ENOMEM;int i;/* Allocate target SG list. */st = kmalloc(sizeof(*st), GFP_KERNEL);if (!st)goto err_st_alloc;ret = sg_alloc_table(st, size, GFP_KERNEL);if (ret)goto err_sg_alloc;st->nents = 0;sg = st->sgl;for (i = 0 ; i < ARRAY_SIZE(rot_info->plane); i++)sg = rotate_pages(obj, rot_info->plane[i].offset,rot_info->plane[i].width, rot_info->plane[i].height,rot_info->plane[i].src_stride,rot_info->plane[i].dst_stride,st, sg);return st;+err_sg_alloc:
kfree(st);+err_st_alloc:
drm_dbg(&i915->drm, "Failed to create rotated mapping for object size %zu! (%ux%u tiles, %u pages)\n",obj->base.size, rot_info->plane[0].width,rot_info->plane[0].height, size);return ERR_PTR(ret);+}
+static struct scatterlist * +remap_pages(struct drm_i915_gem_object *obj,
unsigned int offset, unsigned int alignment_pad,unsigned int width, unsigned int height,unsigned int src_stride, unsigned int dst_stride,struct sg_table *st, struct scatterlist *sg)+{
unsigned int row;if (alignment_pad) {st->nents++;/** The DE ignores the PTEs for the padding tiles, the sg entry* here is just a convenience to indicate how many padding PTEs* to insert at this spot.*/sg_set_page(sg, NULL, alignment_pad * 4096, 0);sg_dma_address(sg) = 0;sg_dma_len(sg) = alignment_pad * 4096;sg = sg_next(sg);}for (row = 0; row < height; row++) {unsigned int left = width * I915_GTT_PAGE_SIZE;while (left) {dma_addr_t addr;unsigned int length;/** We don't need the pages, but need to initialize* the entries so the sg list can be happily traversed.* The only thing we need are DMA addresses.*/addr = i915_gem_object_get_dma_address_len(obj, offset, &length);length = min(left, length);st->nents++;sg_set_page(sg, NULL, length, 0);sg_dma_address(sg) = addr;sg_dma_len(sg) = length;sg = sg_next(sg);offset += length / I915_GTT_PAGE_SIZE;left -= length;}offset += src_stride - width;left = (dst_stride - width) * I915_GTT_PAGE_SIZE;if (!left)continue;st->nents++;/** The DE ignores the PTEs for the padding tiles, the sg entry* here is just a conenience to indicate how many padding PTEs* to insert at this spot.*/sg_set_page(sg, NULL, left, 0);sg_dma_address(sg) = 0;sg_dma_len(sg) = left;sg = sg_next(sg);}return sg;+}
+static noinline struct sg_table * +intel_remap_pages(struct intel_remapped_info *rem_info,
struct drm_i915_gem_object *obj)+{
unsigned int size = intel_remapped_info_size(rem_info);struct drm_i915_private *i915 = to_i915(obj->base.dev);struct sg_table *st;struct scatterlist *sg;unsigned int gtt_offset = 0;int ret = -ENOMEM;int i;/* Allocate target SG list. */st = kmalloc(sizeof(*st), GFP_KERNEL);if (!st)goto err_st_alloc;ret = sg_alloc_table(st, size, GFP_KERNEL);if (ret)goto err_sg_alloc;st->nents = 0;sg = st->sgl;for (i = 0 ; i < ARRAY_SIZE(rem_info->plane); i++) {unsigned int alignment_pad = 0;if (rem_info->plane_alignment)alignment_pad = ALIGN(gtt_offset, rem_info->plane_alignment) - gtt_offset;sg = remap_pages(obj,rem_info->plane[i].offset, alignment_pad,rem_info->plane[i].width, rem_info->plane[i].height,rem_info->plane[i].src_stride, rem_info->plane[i].dst_stride,st, sg);gtt_offset += alignment_pad +rem_info->plane[i].dst_stride * rem_info->plane[i].height;}i915_sg_trim(st);return st;+err_sg_alloc:
kfree(st);+err_st_alloc:
drm_dbg(&i915->drm, "Failed to create remapped mapping for object size %zu! (%ux%u tiles, %u pages)\n",obj->base.size, rem_info->plane[0].width,rem_info->plane[0].height, size);return ERR_PTR(ret);+}
+static noinline struct sg_table * +intel_partial_pages(const struct i915_ggtt_view *view,
struct drm_i915_gem_object *obj)+{
struct sg_table *st;struct scatterlist *sg, *iter;unsigned int count = view->partial.size;unsigned int offset;int ret = -ENOMEM;st = kmalloc(sizeof(*st), GFP_KERNEL);if (!st)goto err_st_alloc;ret = sg_alloc_table(st, count, GFP_KERNEL);if (ret)goto err_sg_alloc;iter = i915_gem_object_get_sg_dma(obj, view->partial.offset, &offset);GEM_BUG_ON(!iter);sg = st->sgl;st->nents = 0;do {unsigned int len;len = min(sg_dma_len(iter) - (offset << PAGE_SHIFT),count << PAGE_SHIFT);sg_set_page(sg, NULL, len, 0);sg_dma_address(sg) =sg_dma_address(iter) + (offset << PAGE_SHIFT);sg_dma_len(sg) = len;st->nents++;count -= len >> PAGE_SHIFT;if (count == 0) {sg_mark_end(sg);i915_sg_trim(st); /* Drop any unused tail entries. */return st;}sg = __sg_next(sg);iter = __sg_next(iter);offset = 0;} while (1);+err_sg_alloc:
kfree(st);+err_st_alloc:
return ERR_PTR(ret);+}
+static int +__i915_vma_get_pages(struct i915_vma *vma) +{
struct sg_table *pages;int ret;/** The vma->pages are only valid within the lifespan of the borrowed* obj->mm.pages. When the obj->mm.pages sg_table is regenerated, so* must be the vma->pages. A simple rule is that vma->pages must only* be accessed when the obj->mm.pages are pinned.*/GEM_BUG_ON(!i915_gem_object_has_pinned_pages(vma->obj));switch (vma->ggtt_view.type) {default:GEM_BUG_ON(vma->ggtt_view.type);fallthrough;case I915_GGTT_VIEW_NORMAL:pages = vma->obj->mm.pages;break;case I915_GGTT_VIEW_ROTATED:pages =intel_rotate_pages(&vma->ggtt_view.rotated, vma->obj);break;case I915_GGTT_VIEW_REMAPPED:pages =intel_remap_pages(&vma->ggtt_view.remapped, vma->obj);break;case I915_GGTT_VIEW_PARTIAL:pages = intel_partial_pages(&vma->ggtt_view, vma->obj);break;}ret = 0;/* gen6 ppgtt doesn't have backing pages, special-case it */if (IS_ERR(pages) && pages != ERR_PTR(-ENODEV)) {
Where does this -ENODEV come from? AFAIK for the gen6 thing mm.pages = ZERO_SIZE_PTR. Also, just checking, I assume we always hit the VIEW_NORMAL for that?
ret = PTR_ERR(pages);pages = NULL;drm_err(&vma->vm->i915->drm,"Failed to get pages for VMA view type %u (%d)!\n",vma->ggtt_view.type, ret);
Can we just return here?
}pages = xchg(&vma->pages, pages);/* did we race against a put_pages? */if (pages && pages != vma->obj->mm.pages) {sg_free_table(vma->pages);kfree(vma->pages);}
What is this race exactly, can we add some more details here please? So we can more easily evaluate the lockless trickery here.
On Thu, 21 Oct 2021 at 18:30, Matthew Auld matthew.william.auld@gmail.com wrote:
On Thu, 21 Oct 2021 at 11:36, Maarten Lankhorst maarten.lankhorst@linux.intel.com wrote:
Big delta, but boils down to moving set_pages to i915_vma.c, and removing the special handling, all callers use the defaults anyway. We only remap in ggtt, so default case will fall through.
Because we still don't require locking in i915_vma_unpin(), handle this by using xchg in get_pages(), as it's locked with obj->mutex, and cmpxchg in unpin, which only fails if we race a against a new pin.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
drivers/gpu/drm/i915/display/intel_dpt.c | 2 - drivers/gpu/drm/i915/gt/gen6_ppgtt.c | 15 - drivers/gpu/drm/i915/gt/intel_ggtt.c | 345 ---------------- drivers/gpu/drm/i915/gt/intel_gtt.c | 13 - drivers/gpu/drm/i915/gt/intel_gtt.h | 7 - drivers/gpu/drm/i915/gt/intel_ppgtt.c | 12 - drivers/gpu/drm/i915/i915_vma.c | 388 ++++++++++++++++-- drivers/gpu/drm/i915/i915_vma.h | 3 + drivers/gpu/drm/i915/i915_vma_types.h | 1 - drivers/gpu/drm/i915/selftests/i915_gem_gtt.c | 12 +- drivers/gpu/drm/i915/selftests/mock_gtt.c | 4 - 11 files changed, 368 insertions(+), 434 deletions(-)
diff --git a/drivers/gpu/drm/i915/display/intel_dpt.c b/drivers/gpu/drm/i915/display/intel_dpt.c index 8f7b1f7534a4..ef428f3fc538 100644 --- a/drivers/gpu/drm/i915/display/intel_dpt.c +++ b/drivers/gpu/drm/i915/display/intel_dpt.c @@ -221,8 +221,6 @@ intel_dpt_create(struct intel_framebuffer *fb)
vm->vma_ops.bind_vma = dpt_bind_vma; vm->vma_ops.unbind_vma = dpt_unbind_vma;
vm->vma_ops.set_pages = ggtt_set_pages;vm->vma_ops.clear_pages = clear_pages; vm->pte_encode = gen8_ggtt_pte_encode;diff --git a/drivers/gpu/drm/i915/gt/gen6_ppgtt.c b/drivers/gpu/drm/i915/gt/gen6_ppgtt.c index 5caa1703716e..5c048b4ccd4d 100644 --- a/drivers/gpu/drm/i915/gt/gen6_ppgtt.c +++ b/drivers/gpu/drm/i915/gt/gen6_ppgtt.c @@ -270,19 +270,6 @@ static void gen6_ppgtt_cleanup(struct i915_address_space *vm) free_pd(&ppgtt->base.vm, ppgtt->base.pd); }
-static int pd_vma_set_pages(struct i915_vma *vma) -{
vma->pages = ERR_PTR(-ENODEV);return 0;-}
-static void pd_vma_clear_pages(struct i915_vma *vma) -{
GEM_BUG_ON(!vma->pages);vma->pages = NULL;-}
static void pd_vma_bind(struct i915_address_space *vm, struct i915_vm_pt_stash *stash, struct i915_vma *vma, @@ -322,8 +309,6 @@ static void pd_vma_unbind(struct i915_address_space *vm, struct i915_vma *vma) }
static const struct i915_vma_ops pd_vma_ops = {
.set_pages = pd_vma_set_pages,.clear_pages = pd_vma_clear_pages, .bind_vma = pd_vma_bind, .unbind_vma = pd_vma_unbind,}; diff --git a/drivers/gpu/drm/i915/gt/intel_ggtt.c b/drivers/gpu/drm/i915/gt/intel_ggtt.c index f17383e76eb7..6da57199bb33 100644 --- a/drivers/gpu/drm/i915/gt/intel_ggtt.c +++ b/drivers/gpu/drm/i915/gt/intel_ggtt.c @@ -20,9 +20,6 @@ #include "intel_gtt.h" #include "gen8_ppgtt.h"
-static int -i915_get_ggtt_vma_pages(struct i915_vma *vma);
static void i915_ggtt_color_adjust(const struct drm_mm_node *node, unsigned long color, u64 *start, @@ -875,21 +872,6 @@ static int ggtt_probe_common(struct i915_ggtt *ggtt, u64 size) return 0; }
-int ggtt_set_pages(struct i915_vma *vma) -{
int ret;GEM_BUG_ON(vma->pages);ret = i915_get_ggtt_vma_pages(vma);if (ret)return ret;vma->page_sizes = vma->obj->mm.page_sizes;return 0;-}
static void gen6_gmch_remove(struct i915_address_space *vm) { struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm); @@ -950,8 +932,6 @@ static int gen8_gmch_probe(struct i915_ggtt *ggtt)
ggtt->vm.vma_ops.bind_vma = ggtt_bind_vma; ggtt->vm.vma_ops.unbind_vma = ggtt_unbind_vma;
ggtt->vm.vma_ops.set_pages = ggtt_set_pages;ggtt->vm.vma_ops.clear_pages = clear_pages; ggtt->vm.pte_encode = gen8_ggtt_pte_encode;@@ -1100,8 +1080,6 @@ static int gen6_gmch_probe(struct i915_ggtt *ggtt)
ggtt->vm.vma_ops.bind_vma = ggtt_bind_vma; ggtt->vm.vma_ops.unbind_vma = ggtt_unbind_vma;
ggtt->vm.vma_ops.set_pages = ggtt_set_pages;ggtt->vm.vma_ops.clear_pages = clear_pages; return ggtt_probe_common(ggtt, size);} @@ -1145,8 +1123,6 @@ static int i915_gmch_probe(struct i915_ggtt *ggtt)
ggtt->vm.vma_ops.bind_vma = ggtt_bind_vma; ggtt->vm.vma_ops.unbind_vma = ggtt_unbind_vma;
ggtt->vm.vma_ops.set_pages = ggtt_set_pages;ggtt->vm.vma_ops.clear_pages = clear_pages; if (unlikely(ggtt->do_idle_maps)) drm_notice(&i915->drm,@@ -1294,324 +1270,3 @@ void i915_ggtt_resume(struct i915_ggtt *ggtt)
intel_ggtt_restore_fences(ggtt);}
-static struct scatterlist * -rotate_pages(struct drm_i915_gem_object *obj, unsigned int offset,
unsigned int width, unsigned int height,unsigned int src_stride, unsigned int dst_stride,struct sg_table *st, struct scatterlist *sg)-{
unsigned int column, row;unsigned int src_idx;for (column = 0; column < width; column++) {unsigned int left;src_idx = src_stride * (height - 1) + column + offset;for (row = 0; row < height; row++) {st->nents++;/** We don't need the pages, but need to initialize* the entries so the sg list can be happily traversed.* The only thing we need are DMA addresses.*/sg_set_page(sg, NULL, I915_GTT_PAGE_SIZE, 0);sg_dma_address(sg) =i915_gem_object_get_dma_address(obj, src_idx);sg_dma_len(sg) = I915_GTT_PAGE_SIZE;sg = sg_next(sg);src_idx -= src_stride;}left = (dst_stride - height) * I915_GTT_PAGE_SIZE;if (!left)continue;st->nents++;/** The DE ignores the PTEs for the padding tiles, the sg entry* here is just a conenience to indicate how many padding PTEs* to insert at this spot.*/sg_set_page(sg, NULL, left, 0);sg_dma_address(sg) = 0;sg_dma_len(sg) = left;sg = sg_next(sg);}return sg;-}
-static noinline struct sg_table * -intel_rotate_pages(struct intel_rotation_info *rot_info,
struct drm_i915_gem_object *obj)-{
unsigned int size = intel_rotation_info_size(rot_info);struct drm_i915_private *i915 = to_i915(obj->base.dev);struct sg_table *st;struct scatterlist *sg;int ret = -ENOMEM;int i;/* Allocate target SG list. */st = kmalloc(sizeof(*st), GFP_KERNEL);if (!st)goto err_st_alloc;ret = sg_alloc_table(st, size, GFP_KERNEL);if (ret)goto err_sg_alloc;st->nents = 0;sg = st->sgl;for (i = 0 ; i < ARRAY_SIZE(rot_info->plane); i++)sg = rotate_pages(obj, rot_info->plane[i].offset,rot_info->plane[i].width, rot_info->plane[i].height,rot_info->plane[i].src_stride,rot_info->plane[i].dst_stride,st, sg);return st;-err_sg_alloc:
kfree(st);-err_st_alloc:
drm_dbg(&i915->drm, "Failed to create rotated mapping for object size %zu! (%ux%u tiles, %u pages)\n",obj->base.size, rot_info->plane[0].width,rot_info->plane[0].height, size);return ERR_PTR(ret);-}
-static struct scatterlist * -remap_pages(struct drm_i915_gem_object *obj,
unsigned int offset, unsigned int alignment_pad,unsigned int width, unsigned int height,unsigned int src_stride, unsigned int dst_stride,struct sg_table *st, struct scatterlist *sg)-{
unsigned int row;if (alignment_pad) {st->nents++;/** The DE ignores the PTEs for the padding tiles, the sg entry* here is just a convenience to indicate how many padding PTEs* to insert at this spot.*/sg_set_page(sg, NULL, alignment_pad * 4096, 0);sg_dma_address(sg) = 0;sg_dma_len(sg) = alignment_pad * 4096;sg = sg_next(sg);}for (row = 0; row < height; row++) {unsigned int left = width * I915_GTT_PAGE_SIZE;while (left) {dma_addr_t addr;unsigned int length;/** We don't need the pages, but need to initialize* the entries so the sg list can be happily traversed.* The only thing we need are DMA addresses.*/addr = i915_gem_object_get_dma_address_len(obj, offset, &length);length = min(left, length);st->nents++;sg_set_page(sg, NULL, length, 0);sg_dma_address(sg) = addr;sg_dma_len(sg) = length;sg = sg_next(sg);offset += length / I915_GTT_PAGE_SIZE;left -= length;}offset += src_stride - width;left = (dst_stride - width) * I915_GTT_PAGE_SIZE;if (!left)continue;st->nents++;/** The DE ignores the PTEs for the padding tiles, the sg entry* here is just a conenience to indicate how many padding PTEs* to insert at this spot.*/sg_set_page(sg, NULL, left, 0);sg_dma_address(sg) = 0;sg_dma_len(sg) = left;sg = sg_next(sg);}return sg;-}
-static noinline struct sg_table * -intel_remap_pages(struct intel_remapped_info *rem_info,
struct drm_i915_gem_object *obj)-{
unsigned int size = intel_remapped_info_size(rem_info);struct drm_i915_private *i915 = to_i915(obj->base.dev);struct sg_table *st;struct scatterlist *sg;unsigned int gtt_offset = 0;int ret = -ENOMEM;int i;/* Allocate target SG list. */st = kmalloc(sizeof(*st), GFP_KERNEL);if (!st)goto err_st_alloc;ret = sg_alloc_table(st, size, GFP_KERNEL);if (ret)goto err_sg_alloc;st->nents = 0;sg = st->sgl;for (i = 0 ; i < ARRAY_SIZE(rem_info->plane); i++) {unsigned int alignment_pad = 0;if (rem_info->plane_alignment)alignment_pad = ALIGN(gtt_offset, rem_info->plane_alignment) - gtt_offset;sg = remap_pages(obj,rem_info->plane[i].offset, alignment_pad,rem_info->plane[i].width, rem_info->plane[i].height,rem_info->plane[i].src_stride, rem_info->plane[i].dst_stride,st, sg);gtt_offset += alignment_pad +rem_info->plane[i].dst_stride * rem_info->plane[i].height;}i915_sg_trim(st);return st;-err_sg_alloc:
kfree(st);-err_st_alloc:
drm_dbg(&i915->drm, "Failed to create remapped mapping for object size %zu! (%ux%u tiles, %u pages)\n",obj->base.size, rem_info->plane[0].width,rem_info->plane[0].height, size);return ERR_PTR(ret);-}
-static noinline struct sg_table * -intel_partial_pages(const struct i915_ggtt_view *view,
struct drm_i915_gem_object *obj)-{
struct sg_table *st;struct scatterlist *sg, *iter;unsigned int count = view->partial.size;unsigned int offset;int ret = -ENOMEM;st = kmalloc(sizeof(*st), GFP_KERNEL);if (!st)goto err_st_alloc;ret = sg_alloc_table(st, count, GFP_KERNEL);if (ret)goto err_sg_alloc;iter = i915_gem_object_get_sg_dma(obj, view->partial.offset, &offset);GEM_BUG_ON(!iter);sg = st->sgl;st->nents = 0;do {unsigned int len;len = min(sg_dma_len(iter) - (offset << PAGE_SHIFT),count << PAGE_SHIFT);sg_set_page(sg, NULL, len, 0);sg_dma_address(sg) =sg_dma_address(iter) + (offset << PAGE_SHIFT);sg_dma_len(sg) = len;st->nents++;count -= len >> PAGE_SHIFT;if (count == 0) {sg_mark_end(sg);i915_sg_trim(st); /* Drop any unused tail entries. */return st;}sg = __sg_next(sg);iter = __sg_next(iter);offset = 0;} while (1);-err_sg_alloc:
kfree(st);-err_st_alloc:
return ERR_PTR(ret);-}
-static int -i915_get_ggtt_vma_pages(struct i915_vma *vma) -{
int ret;/** The vma->pages are only valid within the lifespan of the borrowed* obj->mm.pages. When the obj->mm.pages sg_table is regenerated, so* must be the vma->pages. A simple rule is that vma->pages must only* be accessed when the obj->mm.pages are pinned.*/GEM_BUG_ON(!i915_gem_object_has_pinned_pages(vma->obj));switch (vma->ggtt_view.type) {default:GEM_BUG_ON(vma->ggtt_view.type);fallthrough;case I915_GGTT_VIEW_NORMAL:vma->pages = vma->obj->mm.pages;return 0;case I915_GGTT_VIEW_ROTATED:vma->pages =intel_rotate_pages(&vma->ggtt_view.rotated, vma->obj);break;case I915_GGTT_VIEW_REMAPPED:vma->pages =intel_remap_pages(&vma->ggtt_view.remapped, vma->obj);break;case I915_GGTT_VIEW_PARTIAL:vma->pages = intel_partial_pages(&vma->ggtt_view, vma->obj);break;}ret = 0;if (IS_ERR(vma->pages)) {ret = PTR_ERR(vma->pages);vma->pages = NULL;drm_err(&vma->vm->i915->drm,"Failed to get pages for VMA view type %u (%d)!\n",vma->ggtt_view.type, ret);}return ret;-} diff --git a/drivers/gpu/drm/i915/gt/intel_gtt.c b/drivers/gpu/drm/i915/gt/intel_gtt.c index 67d14afa6623..12eed5fcb17a 100644 --- a/drivers/gpu/drm/i915/gt/intel_gtt.c +++ b/drivers/gpu/drm/i915/gt/intel_gtt.c @@ -221,19 +221,6 @@ void i915_address_space_init(struct i915_address_space *vm, int subclass) INIT_LIST_HEAD(&vm->bound_list); }
-void clear_pages(struct i915_vma *vma) -{
GEM_BUG_ON(!vma->pages);if (vma->pages != vma->obj->mm.pages) {sg_free_table(vma->pages);kfree(vma->pages);}vma->pages = NULL;memset(&vma->page_sizes, 0, sizeof(vma->page_sizes));-}
void *__px_vaddr(struct drm_i915_gem_object *p) { enum i915_map_type type; diff --git a/drivers/gpu/drm/i915/gt/intel_gtt.h b/drivers/gpu/drm/i915/gt/intel_gtt.h index bc6750263359..306e7645fdc5 100644 --- a/drivers/gpu/drm/i915/gt/intel_gtt.h +++ b/drivers/gpu/drm/i915/gt/intel_gtt.h @@ -206,9 +206,6 @@ struct i915_vma_ops { */ void (*unbind_vma)(struct i915_address_space *vm, struct i915_vma *vma);
int (*set_pages)(struct i915_vma *vma);void (*clear_pages)(struct i915_vma *vma);};
struct i915_address_space { @@ -594,10 +591,6 @@ release_pd_entry(struct i915_page_directory * const pd, const struct drm_i915_gem_object * const scratch); void gen6_ggtt_invalidate(struct i915_ggtt *ggtt);
-int ggtt_set_pages(struct i915_vma *vma); -int ppgtt_set_pages(struct i915_vma *vma); -void clear_pages(struct i915_vma *vma);
void ppgtt_bind_vma(struct i915_address_space *vm, struct i915_vm_pt_stash *stash, struct i915_vma *vma, diff --git a/drivers/gpu/drm/i915/gt/intel_ppgtt.c b/drivers/gpu/drm/i915/gt/intel_ppgtt.c index 4396bfd630d8..083b3090c69c 100644 --- a/drivers/gpu/drm/i915/gt/intel_ppgtt.c +++ b/drivers/gpu/drm/i915/gt/intel_ppgtt.c @@ -289,16 +289,6 @@ void i915_vm_free_pt_stash(struct i915_address_space *vm, } }
-int ppgtt_set_pages(struct i915_vma *vma) -{
GEM_BUG_ON(vma->pages);vma->pages = vma->obj->mm.pages;vma->page_sizes = vma->obj->mm.page_sizes;return 0;-}
void ppgtt_init(struct i915_ppgtt *ppgtt, struct intel_gt *gt, unsigned long lmem_pt_obj_flags) { @@ -315,6 +305,4 @@ void ppgtt_init(struct i915_ppgtt *ppgtt, struct intel_gt *gt,
ppgtt->vm.vma_ops.bind_vma = ppgtt_bind_vma; ppgtt->vm.vma_ops.unbind_vma = ppgtt_unbind_vma;
ppgtt->vm.vma_ops.set_pages = ppgtt_set_pages;ppgtt->vm.vma_ops.clear_pages = clear_pages;} diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c index ac09b685678a..bacc8d68e495 100644 --- a/drivers/gpu/drm/i915/i915_vma.c +++ b/drivers/gpu/drm/i915/i915_vma.c @@ -112,7 +112,6 @@ vma_create(struct drm_i915_gem_object *obj, return ERR_PTR(-ENOMEM);
kref_init(&vma->ref);
mutex_init(&vma->pages_mutex); vma->vm = i915_vm_get(vm); vma->ops = &vm->vma_ops; vma->obj = obj;@@ -789,10 +788,343 @@ static bool try_qad_pin(struct i915_vma *vma, unsigned int flags) return pinned; }
-static int vma_get_pages(struct i915_vma *vma)
+static struct scatterlist * +rotate_pages(struct drm_i915_gem_object *obj, unsigned int offset,
unsigned int width, unsigned int height,unsigned int src_stride, unsigned int dst_stride,struct sg_table *st, struct scatterlist *sg){
int err = 0;bool pinned_pages = true;
unsigned int column, row;unsigned int src_idx;for (column = 0; column < width; column++) {unsigned int left;src_idx = src_stride * (height - 1) + column + offset;for (row = 0; row < height; row++) {st->nents++;/** We don't need the pages, but need to initialize* the entries so the sg list can be happily traversed.* The only thing we need are DMA addresses.*/sg_set_page(sg, NULL, I915_GTT_PAGE_SIZE, 0);sg_dma_address(sg) =i915_gem_object_get_dma_address(obj, src_idx);sg_dma_len(sg) = I915_GTT_PAGE_SIZE;sg = sg_next(sg);src_idx -= src_stride;}left = (dst_stride - height) * I915_GTT_PAGE_SIZE;if (!left)continue;st->nents++;/** The DE ignores the PTEs for the padding tiles, the sg entry* here is just a conenience to indicate how many padding PTEs* to insert at this spot.*/sg_set_page(sg, NULL, left, 0);sg_dma_address(sg) = 0;sg_dma_len(sg) = left;sg = sg_next(sg);}return sg;+}
+static noinline struct sg_table * +intel_rotate_pages(struct intel_rotation_info *rot_info,
struct drm_i915_gem_object *obj)+{
unsigned int size = intel_rotation_info_size(rot_info);struct drm_i915_private *i915 = to_i915(obj->base.dev);struct sg_table *st;struct scatterlist *sg;int ret = -ENOMEM;int i;/* Allocate target SG list. */st = kmalloc(sizeof(*st), GFP_KERNEL);if (!st)goto err_st_alloc;ret = sg_alloc_table(st, size, GFP_KERNEL);if (ret)goto err_sg_alloc;st->nents = 0;sg = st->sgl;for (i = 0 ; i < ARRAY_SIZE(rot_info->plane); i++)sg = rotate_pages(obj, rot_info->plane[i].offset,rot_info->plane[i].width, rot_info->plane[i].height,rot_info->plane[i].src_stride,rot_info->plane[i].dst_stride,st, sg);return st;+err_sg_alloc:
kfree(st);+err_st_alloc:
drm_dbg(&i915->drm, "Failed to create rotated mapping for object size %zu! (%ux%u tiles, %u pages)\n",obj->base.size, rot_info->plane[0].width,rot_info->plane[0].height, size);return ERR_PTR(ret);+}
+static struct scatterlist * +remap_pages(struct drm_i915_gem_object *obj,
unsigned int offset, unsigned int alignment_pad,unsigned int width, unsigned int height,unsigned int src_stride, unsigned int dst_stride,struct sg_table *st, struct scatterlist *sg)+{
unsigned int row;if (alignment_pad) {st->nents++;/** The DE ignores the PTEs for the padding tiles, the sg entry* here is just a convenience to indicate how many padding PTEs* to insert at this spot.*/sg_set_page(sg, NULL, alignment_pad * 4096, 0);sg_dma_address(sg) = 0;sg_dma_len(sg) = alignment_pad * 4096;sg = sg_next(sg);}for (row = 0; row < height; row++) {unsigned int left = width * I915_GTT_PAGE_SIZE;while (left) {dma_addr_t addr;unsigned int length;/** We don't need the pages, but need to initialize* the entries so the sg list can be happily traversed.* The only thing we need are DMA addresses.*/addr = i915_gem_object_get_dma_address_len(obj, offset, &length);length = min(left, length);st->nents++;sg_set_page(sg, NULL, length, 0);sg_dma_address(sg) = addr;sg_dma_len(sg) = length;sg = sg_next(sg);offset += length / I915_GTT_PAGE_SIZE;left -= length;}offset += src_stride - width;left = (dst_stride - width) * I915_GTT_PAGE_SIZE;if (!left)continue;st->nents++;/** The DE ignores the PTEs for the padding tiles, the sg entry* here is just a conenience to indicate how many padding PTEs* to insert at this spot.*/sg_set_page(sg, NULL, left, 0);sg_dma_address(sg) = 0;sg_dma_len(sg) = left;sg = sg_next(sg);}return sg;+}
+static noinline struct sg_table * +intel_remap_pages(struct intel_remapped_info *rem_info,
struct drm_i915_gem_object *obj)+{
unsigned int size = intel_remapped_info_size(rem_info);struct drm_i915_private *i915 = to_i915(obj->base.dev);struct sg_table *st;struct scatterlist *sg;unsigned int gtt_offset = 0;int ret = -ENOMEM;int i;/* Allocate target SG list. */st = kmalloc(sizeof(*st), GFP_KERNEL);if (!st)goto err_st_alloc;ret = sg_alloc_table(st, size, GFP_KERNEL);if (ret)goto err_sg_alloc;st->nents = 0;sg = st->sgl;for (i = 0 ; i < ARRAY_SIZE(rem_info->plane); i++) {unsigned int alignment_pad = 0;if (rem_info->plane_alignment)alignment_pad = ALIGN(gtt_offset, rem_info->plane_alignment) - gtt_offset;sg = remap_pages(obj,rem_info->plane[i].offset, alignment_pad,rem_info->plane[i].width, rem_info->plane[i].height,rem_info->plane[i].src_stride, rem_info->plane[i].dst_stride,st, sg);gtt_offset += alignment_pad +rem_info->plane[i].dst_stride * rem_info->plane[i].height;}i915_sg_trim(st);return st;+err_sg_alloc:
kfree(st);+err_st_alloc:
drm_dbg(&i915->drm, "Failed to create remapped mapping for object size %zu! (%ux%u tiles, %u pages)\n",obj->base.size, rem_info->plane[0].width,rem_info->plane[0].height, size);return ERR_PTR(ret);+}
+static noinline struct sg_table * +intel_partial_pages(const struct i915_ggtt_view *view,
struct drm_i915_gem_object *obj)+{
struct sg_table *st;struct scatterlist *sg, *iter;unsigned int count = view->partial.size;unsigned int offset;int ret = -ENOMEM;st = kmalloc(sizeof(*st), GFP_KERNEL);if (!st)goto err_st_alloc;ret = sg_alloc_table(st, count, GFP_KERNEL);if (ret)goto err_sg_alloc;iter = i915_gem_object_get_sg_dma(obj, view->partial.offset, &offset);GEM_BUG_ON(!iter);sg = st->sgl;st->nents = 0;do {unsigned int len;len = min(sg_dma_len(iter) - (offset << PAGE_SHIFT),count << PAGE_SHIFT);sg_set_page(sg, NULL, len, 0);sg_dma_address(sg) =sg_dma_address(iter) + (offset << PAGE_SHIFT);sg_dma_len(sg) = len;st->nents++;count -= len >> PAGE_SHIFT;if (count == 0) {sg_mark_end(sg);i915_sg_trim(st); /* Drop any unused tail entries. */return st;}sg = __sg_next(sg);iter = __sg_next(iter);offset = 0;} while (1);+err_sg_alloc:
kfree(st);+err_st_alloc:
return ERR_PTR(ret);+}
+static int +__i915_vma_get_pages(struct i915_vma *vma) +{
struct sg_table *pages;int ret;/** The vma->pages are only valid within the lifespan of the borrowed* obj->mm.pages. When the obj->mm.pages sg_table is regenerated, so* must be the vma->pages. A simple rule is that vma->pages must only* be accessed when the obj->mm.pages are pinned.*/GEM_BUG_ON(!i915_gem_object_has_pinned_pages(vma->obj));switch (vma->ggtt_view.type) {default:GEM_BUG_ON(vma->ggtt_view.type);fallthrough;case I915_GGTT_VIEW_NORMAL:pages = vma->obj->mm.pages;break;case I915_GGTT_VIEW_ROTATED:pages =intel_rotate_pages(&vma->ggtt_view.rotated, vma->obj);break;case I915_GGTT_VIEW_REMAPPED:pages =intel_remap_pages(&vma->ggtt_view.remapped, vma->obj);break;case I915_GGTT_VIEW_PARTIAL:pages = intel_partial_pages(&vma->ggtt_view, vma->obj);break;}ret = 0;/* gen6 ppgtt doesn't have backing pages, special-case it */if (IS_ERR(pages) && pages != ERR_PTR(-ENODEV)) {Where does this -ENODEV come from? AFAIK for the gen6 thing mm.pages = ZERO_SIZE_PTR. Also, just checking, I assume we always hit the VIEW_NORMAL for that?
ret = PTR_ERR(pages);pages = NULL;drm_err(&vma->vm->i915->drm,"Failed to get pages for VMA view type %u (%d)!\n",vma->ggtt_view.type, ret);Can we just return here?
}pages = xchg(&vma->pages, pages);/* did we race against a put_pages? */if (pages && pages != vma->obj->mm.pages) {sg_free_table(vma->pages);kfree(vma->pages);}What is this race exactly, can we add some more details here please? So we can more easily evaluate the lockless trickery here.
Ok, the lockless stuff just gets removed by the end of the series.
On 22-10-2021 12:59, Matthew Auld wrote:
On Thu, 21 Oct 2021 at 18:30, Matthew Auld matthew.william.auld@gmail.com wrote:
On Thu, 21 Oct 2021 at 11:36, Maarten Lankhorst maarten.lankhorst@linux.intel.com wrote:
Big delta, but boils down to moving set_pages to i915_vma.c, and removing the special handling, all callers use the defaults anyway. We only remap in ggtt, so default case will fall through.
Because we still don't require locking in i915_vma_unpin(), handle this by using xchg in get_pages(), as it's locked with obj->mutex, and cmpxchg in unpin, which only fails if we race a against a new pin.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
drivers/gpu/drm/i915/display/intel_dpt.c | 2 - drivers/gpu/drm/i915/gt/gen6_ppgtt.c | 15 - drivers/gpu/drm/i915/gt/intel_ggtt.c | 345 ---------------- drivers/gpu/drm/i915/gt/intel_gtt.c | 13 - drivers/gpu/drm/i915/gt/intel_gtt.h | 7 - drivers/gpu/drm/i915/gt/intel_ppgtt.c | 12 - drivers/gpu/drm/i915/i915_vma.c | 388 ++++++++++++++++-- drivers/gpu/drm/i915/i915_vma.h | 3 + drivers/gpu/drm/i915/i915_vma_types.h | 1 - drivers/gpu/drm/i915/selftests/i915_gem_gtt.c | 12 +- drivers/gpu/drm/i915/selftests/mock_gtt.c | 4 - 11 files changed, 368 insertions(+), 434 deletions(-)
diff --git a/drivers/gpu/drm/i915/display/intel_dpt.c b/drivers/gpu/drm/i915/display/intel_dpt.c index 8f7b1f7534a4..ef428f3fc538 100644 --- a/drivers/gpu/drm/i915/display/intel_dpt.c +++ b/drivers/gpu/drm/i915/display/intel_dpt.c @@ -221,8 +221,6 @@ intel_dpt_create(struct intel_framebuffer *fb)
vm->vma_ops.bind_vma = dpt_bind_vma; vm->vma_ops.unbind_vma = dpt_unbind_vma;
vm->vma_ops.set_pages = ggtt_set_pages;vm->vma_ops.clear_pages = clear_pages; vm->pte_encode = gen8_ggtt_pte_encode;diff --git a/drivers/gpu/drm/i915/gt/gen6_ppgtt.c b/drivers/gpu/drm/i915/gt/gen6_ppgtt.c index 5caa1703716e..5c048b4ccd4d 100644 --- a/drivers/gpu/drm/i915/gt/gen6_ppgtt.c +++ b/drivers/gpu/drm/i915/gt/gen6_ppgtt.c @@ -270,19 +270,6 @@ static void gen6_ppgtt_cleanup(struct i915_address_space *vm) free_pd(&ppgtt->base.vm, ppgtt->base.pd); }
-static int pd_vma_set_pages(struct i915_vma *vma) -{
vma->pages = ERR_PTR(-ENODEV);return 0;-}
-static void pd_vma_clear_pages(struct i915_vma *vma) -{
GEM_BUG_ON(!vma->pages);vma->pages = NULL;-}
static void pd_vma_bind(struct i915_address_space *vm, struct i915_vm_pt_stash *stash, struct i915_vma *vma, @@ -322,8 +309,6 @@ static void pd_vma_unbind(struct i915_address_space *vm, struct i915_vma *vma) }
static const struct i915_vma_ops pd_vma_ops = {
.set_pages = pd_vma_set_pages,.clear_pages = pd_vma_clear_pages, .bind_vma = pd_vma_bind, .unbind_vma = pd_vma_unbind,}; diff --git a/drivers/gpu/drm/i915/gt/intel_ggtt.c b/drivers/gpu/drm/i915/gt/intel_ggtt.c index f17383e76eb7..6da57199bb33 100644 --- a/drivers/gpu/drm/i915/gt/intel_ggtt.c +++ b/drivers/gpu/drm/i915/gt/intel_ggtt.c @@ -20,9 +20,6 @@ #include "intel_gtt.h" #include "gen8_ppgtt.h"
-static int -i915_get_ggtt_vma_pages(struct i915_vma *vma);
static void i915_ggtt_color_adjust(const struct drm_mm_node *node, unsigned long color, u64 *start, @@ -875,21 +872,6 @@ static int ggtt_probe_common(struct i915_ggtt *ggtt, u64 size) return 0; }
-int ggtt_set_pages(struct i915_vma *vma) -{
int ret;GEM_BUG_ON(vma->pages);ret = i915_get_ggtt_vma_pages(vma);if (ret)return ret;vma->page_sizes = vma->obj->mm.page_sizes;return 0;-}
static void gen6_gmch_remove(struct i915_address_space *vm) { struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm); @@ -950,8 +932,6 @@ static int gen8_gmch_probe(struct i915_ggtt *ggtt)
ggtt->vm.vma_ops.bind_vma = ggtt_bind_vma; ggtt->vm.vma_ops.unbind_vma = ggtt_unbind_vma;
ggtt->vm.vma_ops.set_pages = ggtt_set_pages;ggtt->vm.vma_ops.clear_pages = clear_pages; ggtt->vm.pte_encode = gen8_ggtt_pte_encode;@@ -1100,8 +1080,6 @@ static int gen6_gmch_probe(struct i915_ggtt *ggtt)
ggtt->vm.vma_ops.bind_vma = ggtt_bind_vma; ggtt->vm.vma_ops.unbind_vma = ggtt_unbind_vma;
ggtt->vm.vma_ops.set_pages = ggtt_set_pages;ggtt->vm.vma_ops.clear_pages = clear_pages; return ggtt_probe_common(ggtt, size);} @@ -1145,8 +1123,6 @@ static int i915_gmch_probe(struct i915_ggtt *ggtt)
ggtt->vm.vma_ops.bind_vma = ggtt_bind_vma; ggtt->vm.vma_ops.unbind_vma = ggtt_unbind_vma;
ggtt->vm.vma_ops.set_pages = ggtt_set_pages;ggtt->vm.vma_ops.clear_pages = clear_pages; if (unlikely(ggtt->do_idle_maps)) drm_notice(&i915->drm,@@ -1294,324 +1270,3 @@ void i915_ggtt_resume(struct i915_ggtt *ggtt)
intel_ggtt_restore_fences(ggtt);}
-static struct scatterlist * -rotate_pages(struct drm_i915_gem_object *obj, unsigned int offset,
unsigned int width, unsigned int height,unsigned int src_stride, unsigned int dst_stride,struct sg_table *st, struct scatterlist *sg)-{
unsigned int column, row;unsigned int src_idx;for (column = 0; column < width; column++) {unsigned int left;src_idx = src_stride * (height - 1) + column + offset;for (row = 0; row < height; row++) {st->nents++;/** We don't need the pages, but need to initialize* the entries so the sg list can be happily traversed.* The only thing we need are DMA addresses.*/sg_set_page(sg, NULL, I915_GTT_PAGE_SIZE, 0);sg_dma_address(sg) =i915_gem_object_get_dma_address(obj, src_idx);sg_dma_len(sg) = I915_GTT_PAGE_SIZE;sg = sg_next(sg);src_idx -= src_stride;}left = (dst_stride - height) * I915_GTT_PAGE_SIZE;if (!left)continue;st->nents++;/** The DE ignores the PTEs for the padding tiles, the sg entry* here is just a conenience to indicate how many padding PTEs* to insert at this spot.*/sg_set_page(sg, NULL, left, 0);sg_dma_address(sg) = 0;sg_dma_len(sg) = left;sg = sg_next(sg);}return sg;-}
-static noinline struct sg_table * -intel_rotate_pages(struct intel_rotation_info *rot_info,
struct drm_i915_gem_object *obj)-{
unsigned int size = intel_rotation_info_size(rot_info);struct drm_i915_private *i915 = to_i915(obj->base.dev);struct sg_table *st;struct scatterlist *sg;int ret = -ENOMEM;int i;/* Allocate target SG list. */st = kmalloc(sizeof(*st), GFP_KERNEL);if (!st)goto err_st_alloc;ret = sg_alloc_table(st, size, GFP_KERNEL);if (ret)goto err_sg_alloc;st->nents = 0;sg = st->sgl;for (i = 0 ; i < ARRAY_SIZE(rot_info->plane); i++)sg = rotate_pages(obj, rot_info->plane[i].offset,rot_info->plane[i].width, rot_info->plane[i].height,rot_info->plane[i].src_stride,rot_info->plane[i].dst_stride,st, sg);return st;-err_sg_alloc:
kfree(st);-err_st_alloc:
drm_dbg(&i915->drm, "Failed to create rotated mapping for object size %zu! (%ux%u tiles, %u pages)\n",obj->base.size, rot_info->plane[0].width,rot_info->plane[0].height, size);return ERR_PTR(ret);-}
-static struct scatterlist * -remap_pages(struct drm_i915_gem_object *obj,
unsigned int offset, unsigned int alignment_pad,unsigned int width, unsigned int height,unsigned int src_stride, unsigned int dst_stride,struct sg_table *st, struct scatterlist *sg)-{
unsigned int row;if (alignment_pad) {st->nents++;/** The DE ignores the PTEs for the padding tiles, the sg entry* here is just a convenience to indicate how many padding PTEs* to insert at this spot.*/sg_set_page(sg, NULL, alignment_pad * 4096, 0);sg_dma_address(sg) = 0;sg_dma_len(sg) = alignment_pad * 4096;sg = sg_next(sg);}for (row = 0; row < height; row++) {unsigned int left = width * I915_GTT_PAGE_SIZE;while (left) {dma_addr_t addr;unsigned int length;/** We don't need the pages, but need to initialize* the entries so the sg list can be happily traversed.* The only thing we need are DMA addresses.*/addr = i915_gem_object_get_dma_address_len(obj, offset, &length);length = min(left, length);st->nents++;sg_set_page(sg, NULL, length, 0);sg_dma_address(sg) = addr;sg_dma_len(sg) = length;sg = sg_next(sg);offset += length / I915_GTT_PAGE_SIZE;left -= length;}offset += src_stride - width;left = (dst_stride - width) * I915_GTT_PAGE_SIZE;if (!left)continue;st->nents++;/** The DE ignores the PTEs for the padding tiles, the sg entry* here is just a conenience to indicate how many padding PTEs* to insert at this spot.*/sg_set_page(sg, NULL, left, 0);sg_dma_address(sg) = 0;sg_dma_len(sg) = left;sg = sg_next(sg);}return sg;-}
-static noinline struct sg_table * -intel_remap_pages(struct intel_remapped_info *rem_info,
struct drm_i915_gem_object *obj)-{
unsigned int size = intel_remapped_info_size(rem_info);struct drm_i915_private *i915 = to_i915(obj->base.dev);struct sg_table *st;struct scatterlist *sg;unsigned int gtt_offset = 0;int ret = -ENOMEM;int i;/* Allocate target SG list. */st = kmalloc(sizeof(*st), GFP_KERNEL);if (!st)goto err_st_alloc;ret = sg_alloc_table(st, size, GFP_KERNEL);if (ret)goto err_sg_alloc;st->nents = 0;sg = st->sgl;for (i = 0 ; i < ARRAY_SIZE(rem_info->plane); i++) {unsigned int alignment_pad = 0;if (rem_info->plane_alignment)alignment_pad = ALIGN(gtt_offset, rem_info->plane_alignment) - gtt_offset;sg = remap_pages(obj,rem_info->plane[i].offset, alignment_pad,rem_info->plane[i].width, rem_info->plane[i].height,rem_info->plane[i].src_stride, rem_info->plane[i].dst_stride,st, sg);gtt_offset += alignment_pad +rem_info->plane[i].dst_stride * rem_info->plane[i].height;}i915_sg_trim(st);return st;-err_sg_alloc:
kfree(st);-err_st_alloc:
drm_dbg(&i915->drm, "Failed to create remapped mapping for object size %zu! (%ux%u tiles, %u pages)\n",obj->base.size, rem_info->plane[0].width,rem_info->plane[0].height, size);return ERR_PTR(ret);-}
-static noinline struct sg_table * -intel_partial_pages(const struct i915_ggtt_view *view,
struct drm_i915_gem_object *obj)-{
struct sg_table *st;struct scatterlist *sg, *iter;unsigned int count = view->partial.size;unsigned int offset;int ret = -ENOMEM;st = kmalloc(sizeof(*st), GFP_KERNEL);if (!st)goto err_st_alloc;ret = sg_alloc_table(st, count, GFP_KERNEL);if (ret)goto err_sg_alloc;iter = i915_gem_object_get_sg_dma(obj, view->partial.offset, &offset);GEM_BUG_ON(!iter);sg = st->sgl;st->nents = 0;do {unsigned int len;len = min(sg_dma_len(iter) - (offset << PAGE_SHIFT),count << PAGE_SHIFT);sg_set_page(sg, NULL, len, 0);sg_dma_address(sg) =sg_dma_address(iter) + (offset << PAGE_SHIFT);sg_dma_len(sg) = len;st->nents++;count -= len >> PAGE_SHIFT;if (count == 0) {sg_mark_end(sg);i915_sg_trim(st); /* Drop any unused tail entries. */return st;}sg = __sg_next(sg);iter = __sg_next(iter);offset = 0;} while (1);-err_sg_alloc:
kfree(st);-err_st_alloc:
return ERR_PTR(ret);-}
-static int -i915_get_ggtt_vma_pages(struct i915_vma *vma) -{
int ret;/** The vma->pages are only valid within the lifespan of the borrowed* obj->mm.pages. When the obj->mm.pages sg_table is regenerated, so* must be the vma->pages. A simple rule is that vma->pages must only* be accessed when the obj->mm.pages are pinned.*/GEM_BUG_ON(!i915_gem_object_has_pinned_pages(vma->obj));switch (vma->ggtt_view.type) {default:GEM_BUG_ON(vma->ggtt_view.type);fallthrough;case I915_GGTT_VIEW_NORMAL:vma->pages = vma->obj->mm.pages;return 0;case I915_GGTT_VIEW_ROTATED:vma->pages =intel_rotate_pages(&vma->ggtt_view.rotated, vma->obj);break;case I915_GGTT_VIEW_REMAPPED:vma->pages =intel_remap_pages(&vma->ggtt_view.remapped, vma->obj);break;case I915_GGTT_VIEW_PARTIAL:vma->pages = intel_partial_pages(&vma->ggtt_view, vma->obj);break;}ret = 0;if (IS_ERR(vma->pages)) {ret = PTR_ERR(vma->pages);vma->pages = NULL;drm_err(&vma->vm->i915->drm,"Failed to get pages for VMA view type %u (%d)!\n",vma->ggtt_view.type, ret);}return ret;-} diff --git a/drivers/gpu/drm/i915/gt/intel_gtt.c b/drivers/gpu/drm/i915/gt/intel_gtt.c index 67d14afa6623..12eed5fcb17a 100644 --- a/drivers/gpu/drm/i915/gt/intel_gtt.c +++ b/drivers/gpu/drm/i915/gt/intel_gtt.c @@ -221,19 +221,6 @@ void i915_address_space_init(struct i915_address_space *vm, int subclass) INIT_LIST_HEAD(&vm->bound_list); }
-void clear_pages(struct i915_vma *vma) -{
GEM_BUG_ON(!vma->pages);if (vma->pages != vma->obj->mm.pages) {sg_free_table(vma->pages);kfree(vma->pages);}vma->pages = NULL;memset(&vma->page_sizes, 0, sizeof(vma->page_sizes));-}
void *__px_vaddr(struct drm_i915_gem_object *p) { enum i915_map_type type; diff --git a/drivers/gpu/drm/i915/gt/intel_gtt.h b/drivers/gpu/drm/i915/gt/intel_gtt.h index bc6750263359..306e7645fdc5 100644 --- a/drivers/gpu/drm/i915/gt/intel_gtt.h +++ b/drivers/gpu/drm/i915/gt/intel_gtt.h @@ -206,9 +206,6 @@ struct i915_vma_ops { */ void (*unbind_vma)(struct i915_address_space *vm, struct i915_vma *vma);
int (*set_pages)(struct i915_vma *vma);void (*clear_pages)(struct i915_vma *vma);};
struct i915_address_space { @@ -594,10 +591,6 @@ release_pd_entry(struct i915_page_directory * const pd, const struct drm_i915_gem_object * const scratch); void gen6_ggtt_invalidate(struct i915_ggtt *ggtt);
-int ggtt_set_pages(struct i915_vma *vma); -int ppgtt_set_pages(struct i915_vma *vma); -void clear_pages(struct i915_vma *vma);
void ppgtt_bind_vma(struct i915_address_space *vm, struct i915_vm_pt_stash *stash, struct i915_vma *vma, diff --git a/drivers/gpu/drm/i915/gt/intel_ppgtt.c b/drivers/gpu/drm/i915/gt/intel_ppgtt.c index 4396bfd630d8..083b3090c69c 100644 --- a/drivers/gpu/drm/i915/gt/intel_ppgtt.c +++ b/drivers/gpu/drm/i915/gt/intel_ppgtt.c @@ -289,16 +289,6 @@ void i915_vm_free_pt_stash(struct i915_address_space *vm, } }
-int ppgtt_set_pages(struct i915_vma *vma) -{
GEM_BUG_ON(vma->pages);vma->pages = vma->obj->mm.pages;vma->page_sizes = vma->obj->mm.page_sizes;return 0;-}
void ppgtt_init(struct i915_ppgtt *ppgtt, struct intel_gt *gt, unsigned long lmem_pt_obj_flags) { @@ -315,6 +305,4 @@ void ppgtt_init(struct i915_ppgtt *ppgtt, struct intel_gt *gt,
ppgtt->vm.vma_ops.bind_vma = ppgtt_bind_vma; ppgtt->vm.vma_ops.unbind_vma = ppgtt_unbind_vma;
ppgtt->vm.vma_ops.set_pages = ppgtt_set_pages;ppgtt->vm.vma_ops.clear_pages = clear_pages;} diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c index ac09b685678a..bacc8d68e495 100644 --- a/drivers/gpu/drm/i915/i915_vma.c +++ b/drivers/gpu/drm/i915/i915_vma.c @@ -112,7 +112,6 @@ vma_create(struct drm_i915_gem_object *obj, return ERR_PTR(-ENOMEM);
kref_init(&vma->ref);
mutex_init(&vma->pages_mutex); vma->vm = i915_vm_get(vm); vma->ops = &vm->vma_ops; vma->obj = obj;@@ -789,10 +788,343 @@ static bool try_qad_pin(struct i915_vma *vma, unsigned int flags) return pinned; }
-static int vma_get_pages(struct i915_vma *vma)
+static struct scatterlist * +rotate_pages(struct drm_i915_gem_object *obj, unsigned int offset,
unsigned int width, unsigned int height,unsigned int src_stride, unsigned int dst_stride,struct sg_table *st, struct scatterlist *sg){
int err = 0;bool pinned_pages = true;
unsigned int column, row;unsigned int src_idx;for (column = 0; column < width; column++) {unsigned int left;src_idx = src_stride * (height - 1) + column + offset;for (row = 0; row < height; row++) {st->nents++;/** We don't need the pages, but need to initialize* the entries so the sg list can be happily traversed.* The only thing we need are DMA addresses.*/sg_set_page(sg, NULL, I915_GTT_PAGE_SIZE, 0);sg_dma_address(sg) =i915_gem_object_get_dma_address(obj, src_idx);sg_dma_len(sg) = I915_GTT_PAGE_SIZE;sg = sg_next(sg);src_idx -= src_stride;}left = (dst_stride - height) * I915_GTT_PAGE_SIZE;if (!left)continue;st->nents++;/** The DE ignores the PTEs for the padding tiles, the sg entry* here is just a conenience to indicate how many padding PTEs* to insert at this spot.*/sg_set_page(sg, NULL, left, 0);sg_dma_address(sg) = 0;sg_dma_len(sg) = left;sg = sg_next(sg);}return sg;+}
+static noinline struct sg_table * +intel_rotate_pages(struct intel_rotation_info *rot_info,
struct drm_i915_gem_object *obj)+{
unsigned int size = intel_rotation_info_size(rot_info);struct drm_i915_private *i915 = to_i915(obj->base.dev);struct sg_table *st;struct scatterlist *sg;int ret = -ENOMEM;int i;/* Allocate target SG list. */st = kmalloc(sizeof(*st), GFP_KERNEL);if (!st)goto err_st_alloc;ret = sg_alloc_table(st, size, GFP_KERNEL);if (ret)goto err_sg_alloc;st->nents = 0;sg = st->sgl;for (i = 0 ; i < ARRAY_SIZE(rot_info->plane); i++)sg = rotate_pages(obj, rot_info->plane[i].offset,rot_info->plane[i].width, rot_info->plane[i].height,rot_info->plane[i].src_stride,rot_info->plane[i].dst_stride,st, sg);return st;+err_sg_alloc:
kfree(st);+err_st_alloc:
drm_dbg(&i915->drm, "Failed to create rotated mapping for object size %zu! (%ux%u tiles, %u pages)\n",obj->base.size, rot_info->plane[0].width,rot_info->plane[0].height, size);return ERR_PTR(ret);+}
+static struct scatterlist * +remap_pages(struct drm_i915_gem_object *obj,
unsigned int offset, unsigned int alignment_pad,unsigned int width, unsigned int height,unsigned int src_stride, unsigned int dst_stride,struct sg_table *st, struct scatterlist *sg)+{
unsigned int row;if (alignment_pad) {st->nents++;/** The DE ignores the PTEs for the padding tiles, the sg entry* here is just a convenience to indicate how many padding PTEs* to insert at this spot.*/sg_set_page(sg, NULL, alignment_pad * 4096, 0);sg_dma_address(sg) = 0;sg_dma_len(sg) = alignment_pad * 4096;sg = sg_next(sg);}for (row = 0; row < height; row++) {unsigned int left = width * I915_GTT_PAGE_SIZE;while (left) {dma_addr_t addr;unsigned int length;/** We don't need the pages, but need to initialize* the entries so the sg list can be happily traversed.* The only thing we need are DMA addresses.*/addr = i915_gem_object_get_dma_address_len(obj, offset, &length);length = min(left, length);st->nents++;sg_set_page(sg, NULL, length, 0);sg_dma_address(sg) = addr;sg_dma_len(sg) = length;sg = sg_next(sg);offset += length / I915_GTT_PAGE_SIZE;left -= length;}offset += src_stride - width;left = (dst_stride - width) * I915_GTT_PAGE_SIZE;if (!left)continue;st->nents++;/** The DE ignores the PTEs for the padding tiles, the sg entry* here is just a conenience to indicate how many padding PTEs* to insert at this spot.*/sg_set_page(sg, NULL, left, 0);sg_dma_address(sg) = 0;sg_dma_len(sg) = left;sg = sg_next(sg);}return sg;+}
+static noinline struct sg_table * +intel_remap_pages(struct intel_remapped_info *rem_info,
struct drm_i915_gem_object *obj)+{
unsigned int size = intel_remapped_info_size(rem_info);struct drm_i915_private *i915 = to_i915(obj->base.dev);struct sg_table *st;struct scatterlist *sg;unsigned int gtt_offset = 0;int ret = -ENOMEM;int i;/* Allocate target SG list. */st = kmalloc(sizeof(*st), GFP_KERNEL);if (!st)goto err_st_alloc;ret = sg_alloc_table(st, size, GFP_KERNEL);if (ret)goto err_sg_alloc;st->nents = 0;sg = st->sgl;for (i = 0 ; i < ARRAY_SIZE(rem_info->plane); i++) {unsigned int alignment_pad = 0;if (rem_info->plane_alignment)alignment_pad = ALIGN(gtt_offset, rem_info->plane_alignment) - gtt_offset;sg = remap_pages(obj,rem_info->plane[i].offset, alignment_pad,rem_info->plane[i].width, rem_info->plane[i].height,rem_info->plane[i].src_stride, rem_info->plane[i].dst_stride,st, sg);gtt_offset += alignment_pad +rem_info->plane[i].dst_stride * rem_info->plane[i].height;}i915_sg_trim(st);return st;+err_sg_alloc:
kfree(st);+err_st_alloc:
drm_dbg(&i915->drm, "Failed to create remapped mapping for object size %zu! (%ux%u tiles, %u pages)\n",obj->base.size, rem_info->plane[0].width,rem_info->plane[0].height, size);return ERR_PTR(ret);+}
+static noinline struct sg_table * +intel_partial_pages(const struct i915_ggtt_view *view,
struct drm_i915_gem_object *obj)+{
struct sg_table *st;struct scatterlist *sg, *iter;unsigned int count = view->partial.size;unsigned int offset;int ret = -ENOMEM;st = kmalloc(sizeof(*st), GFP_KERNEL);if (!st)goto err_st_alloc;ret = sg_alloc_table(st, count, GFP_KERNEL);if (ret)goto err_sg_alloc;iter = i915_gem_object_get_sg_dma(obj, view->partial.offset, &offset);GEM_BUG_ON(!iter);sg = st->sgl;st->nents = 0;do {unsigned int len;len = min(sg_dma_len(iter) - (offset << PAGE_SHIFT),count << PAGE_SHIFT);sg_set_page(sg, NULL, len, 0);sg_dma_address(sg) =sg_dma_address(iter) + (offset << PAGE_SHIFT);sg_dma_len(sg) = len;st->nents++;count -= len >> PAGE_SHIFT;if (count == 0) {sg_mark_end(sg);i915_sg_trim(st); /* Drop any unused tail entries. */return st;}sg = __sg_next(sg);iter = __sg_next(iter);offset = 0;} while (1);+err_sg_alloc:
kfree(st);+err_st_alloc:
return ERR_PTR(ret);+}
+static int +__i915_vma_get_pages(struct i915_vma *vma) +{
struct sg_table *pages;int ret;/** The vma->pages are only valid within the lifespan of the borrowed* obj->mm.pages. When the obj->mm.pages sg_table is regenerated, so* must be the vma->pages. A simple rule is that vma->pages must only* be accessed when the obj->mm.pages are pinned.*/GEM_BUG_ON(!i915_gem_object_has_pinned_pages(vma->obj));switch (vma->ggtt_view.type) {default:GEM_BUG_ON(vma->ggtt_view.type);fallthrough;case I915_GGTT_VIEW_NORMAL:pages = vma->obj->mm.pages;break;case I915_GGTT_VIEW_ROTATED:pages =intel_rotate_pages(&vma->ggtt_view.rotated, vma->obj);break;case I915_GGTT_VIEW_REMAPPED:pages =intel_remap_pages(&vma->ggtt_view.remapped, vma->obj);break;case I915_GGTT_VIEW_PARTIAL:pages = intel_partial_pages(&vma->ggtt_view, vma->obj);break;}ret = 0;/* gen6 ppgtt doesn't have backing pages, special-case it */if (IS_ERR(pages) && pages != ERR_PTR(-ENODEV)) {Where does this -ENODEV come from? AFAIK for the gen6 thing mm.pages = ZERO_SIZE_PTR. Also, just checking, I assume we always hit the VIEW_NORMAL for that?
It was from when I was still converting the code. Looks like it can be removed now because setting pages to -ENODEV caused too many changes, and be unconditional IS_ERR.
ret = PTR_ERR(pages);pages = NULL;drm_err(&vma->vm->i915->drm,"Failed to get pages for VMA view type %u (%d)!\n",vma->ggtt_view.type, ret);Can we just return here?
Yeah, should be harmless. Out of paranoia I set pages to 0 below.
}pages = xchg(&vma->pages, pages);/* did we race against a put_pages? */if (pages && pages != vma->obj->mm.pages) {sg_free_table(vma->pages);kfree(vma->pages);}What is this race exactly, can we add some more details here please? So we can more easily evaluate the lockless trickery here.
Ok, the lockless stuff just gets removed by the end of the series.
Correct!
Will fix the ENODEV thing and send a new version.
~Maarten
i915_vma_wait_for_bind needs the vma lock held, fix the caller.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com --- drivers/gpu/drm/i915/i915_vma.c | 40 +++++++++++++++++++++++---------- 1 file changed, 28 insertions(+), 12 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c index bacc8d68e495..2877dcd62acb 100644 --- a/drivers/gpu/drm/i915/i915_vma.c +++ b/drivers/gpu/drm/i915/i915_vma.c @@ -1348,23 +1348,15 @@ static void flush_idle_contexts(struct intel_gt *gt) intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT); }
-int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, - u32 align, unsigned int flags) +static int __i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, + u32 align, unsigned int flags) { struct i915_address_space *vm = vma->vm; int err;
- GEM_BUG_ON(!i915_vma_is_ggtt(vma)); - -#ifdef CONFIG_LOCKDEP - WARN_ON(!ww && dma_resv_held(vma->obj->base.resv)); -#endif - do { - if (ww) - err = i915_vma_pin_ww(vma, ww, 0, align, flags | PIN_GLOBAL); - else - err = i915_vma_pin(vma, 0, align, flags | PIN_GLOBAL); + err = i915_vma_pin_ww(vma, ww, 0, align, flags | PIN_GLOBAL); + if (err != -ENOSPC) { if (!err) { err = i915_vma_wait_for_bind(vma); @@ -1383,6 +1375,30 @@ int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, } while (1); }
+int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, + u32 align, unsigned int flags) +{ + struct i915_gem_ww_ctx _ww; + int err; + + GEM_BUG_ON(!i915_vma_is_ggtt(vma)); + + if (ww) + return __i915_ggtt_pin(vma, ww, align, flags); + +#ifdef CONFIG_LOCKDEP + WARN_ON(dma_resv_held(vma->obj->base.resv)); +#endif + + for_i915_gem_ww(&_ww, err, true) { + err = i915_gem_object_lock(vma->obj, &_ww); + if (!err) + err = __i915_ggtt_pin(vma, &_ww, align, flags); + } + + return err; +} + static void __vma_close(struct i915_vma *vma, struct intel_gt *gt) { /*
On Thu, 21 Oct 2021 at 11:37, Maarten Lankhorst maarten.lankhorst@linux.intel.com wrote:
i915_vma_wait_for_bind needs the vma lock held, fix the caller.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
drivers/gpu/drm/i915/i915_vma.c | 40 +++++++++++++++++++++++---------- 1 file changed, 28 insertions(+), 12 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c index bacc8d68e495..2877dcd62acb 100644 --- a/drivers/gpu/drm/i915/i915_vma.c +++ b/drivers/gpu/drm/i915/i915_vma.c @@ -1348,23 +1348,15 @@ static void flush_idle_contexts(struct intel_gt *gt) intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT); }
-int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
u32 align, unsigned int flags)+static int __i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
u32 align, unsigned int flags){ struct i915_address_space *vm = vma->vm; int err;
GEM_BUG_ON(!i915_vma_is_ggtt(vma));-#ifdef CONFIG_LOCKDEP
WARN_ON(!ww && dma_resv_held(vma->obj->base.resv));-#endif
do {if (ww)err = i915_vma_pin_ww(vma, ww, 0, align, flags | PIN_GLOBAL);elseerr = i915_vma_pin(vma, 0, align, flags | PIN_GLOBAL);
err = i915_vma_pin_ww(vma, ww, 0, align, flags | PIN_GLOBAL);if (err != -ENOSPC) { if (!err) { err = i915_vma_wait_for_bind(vma);@@ -1383,6 +1375,30 @@ int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, } while (1); }
+int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
u32 align, unsigned int flags)+{
struct i915_gem_ww_ctx _ww;int err;GEM_BUG_ON(!i915_vma_is_ggtt(vma));if (ww)return __i915_ggtt_pin(vma, ww, align, flags);+#ifdef CONFIG_LOCKDEP
WARN_ON(dma_resv_held(vma->obj->base.resv));
Hmm, so this can't trigger, say if shrinker grabs the lock, or some other concurrent user?
+#endif
for_i915_gem_ww(&_ww, err, true) {err = i915_gem_object_lock(vma->obj, &_ww);if (!err)err = __i915_ggtt_pin(vma, &_ww, align, flags);}return err;+}
static void __vma_close(struct i915_vma *vma, struct intel_gt *gt) { /* -- 2.33.0
On 21-10-2021 19:39, Matthew Auld wrote:
On Thu, 21 Oct 2021 at 11:37, Maarten Lankhorst maarten.lankhorst@linux.intel.com wrote:
i915_vma_wait_for_bind needs the vma lock held, fix the caller.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
drivers/gpu/drm/i915/i915_vma.c | 40 +++++++++++++++++++++++---------- 1 file changed, 28 insertions(+), 12 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c index bacc8d68e495..2877dcd62acb 100644 --- a/drivers/gpu/drm/i915/i915_vma.c +++ b/drivers/gpu/drm/i915/i915_vma.c @@ -1348,23 +1348,15 @@ static void flush_idle_contexts(struct intel_gt *gt) intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT); }
-int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
u32 align, unsigned int flags)+static int __i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
u32 align, unsigned int flags){ struct i915_address_space *vm = vma->vm; int err;
GEM_BUG_ON(!i915_vma_is_ggtt(vma));-#ifdef CONFIG_LOCKDEP
WARN_ON(!ww && dma_resv_held(vma->obj->base.resv));-#endif
do {if (ww)err = i915_vma_pin_ww(vma, ww, 0, align, flags | PIN_GLOBAL);elseerr = i915_vma_pin(vma, 0, align, flags | PIN_GLOBAL);
err = i915_vma_pin_ww(vma, ww, 0, align, flags | PIN_GLOBAL);if (err != -ENOSPC) { if (!err) { err = i915_vma_wait_for_bind(vma);@@ -1383,6 +1375,30 @@ int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, } while (1); }
+int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
u32 align, unsigned int flags)+{
struct i915_gem_ww_ctx _ww;int err;GEM_BUG_ON(!i915_vma_is_ggtt(vma));if (ww)return __i915_ggtt_pin(vma, ww, align, flags);+#ifdef CONFIG_LOCKDEP
WARN_ON(dma_resv_held(vma->obj->base.resv));Hmm, so this can't trigger, say if shrinker grabs the lock, or some other concurrent user?
No, it checks internally in lockdep that the current task doesn't hold the lock. Others can lock just fine.
dma_resv_is_locked() would check if anyone locked it. :)
~Maarten
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com --- drivers/gpu/drm/i915/i915_gem.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c index 981e383d1a5d..6aa9e465b48e 100644 --- a/drivers/gpu/drm/i915/i915_gem.c +++ b/drivers/gpu/drm/i915/i915_gem.c @@ -931,7 +931,14 @@ i915_gem_object_ggtt_pin_ww(struct drm_i915_gem_object *obj, goto new_vma; }
- ret = i915_vma_unbind(vma); + ret = 0; + if (!ww) + ret = i915_gem_object_lock_interruptible(obj, NULL); + if (!ret) { + ret = i915_vma_unbind(vma); + if (!ww) + i915_gem_object_unlock(obj); + } if (ret) return ERR_PTR(ret); }
On Thu, 21 Oct 2021 at 11:37, Maarten Lankhorst maarten.lankhorst@linux.intel.com wrote:
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
Needs a proper commit message.
drivers/gpu/drm/i915/i915_gem.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c index 981e383d1a5d..6aa9e465b48e 100644 --- a/drivers/gpu/drm/i915/i915_gem.c +++ b/drivers/gpu/drm/i915/i915_gem.c @@ -931,7 +931,14 @@ i915_gem_object_ggtt_pin_ww(struct drm_i915_gem_object *obj, goto new_vma; }
ret = i915_vma_unbind(vma);
ret = 0;if (!ww)ret = i915_gem_object_lock_interruptible(obj, NULL);if (!ret) {ret = i915_vma_unbind(vma);if (!ww)i915_gem_object_unlock(obj);}
There is also a wait_for_bind below. Do we need the lock for that also?
if (ret) return ERR_PTR(ret); }-- 2.33.0
On 21-10-2021 19:48, Matthew Auld wrote:
On Thu, 21 Oct 2021 at 11:37, Maarten Lankhorst maarten.lankhorst@linux.intel.com wrote:
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
Needs a proper commit message.
What about this?
drivers/gpu/drm/i915/i915_gem.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c index 981e383d1a5d..6aa9e465b48e 100644 --- a/drivers/gpu/drm/i915/i915_gem.c +++ b/drivers/gpu/drm/i915/i915_gem.c @@ -931,7 +931,14 @@ i915_gem_object_ggtt_pin_ww(struct drm_i915_gem_object *obj, goto new_vma; }
ret = i915_vma_unbind(vma);
ret = 0;if (!ww)ret = i915_gem_object_lock_interruptible(obj, NULL);if (!ret) {ret = i915_vma_unbind(vma);if (!ww)i915_gem_object_unlock(obj);}There is also a wait_for_bind below. Do we need the lock for that also?
Hmm good find.
Not sure if required in this patch series. I have some patches on top that require the lock because of async binding / unbinding, that would need it for sure.
I will fix this function to use WARN_ON(!ww), and add ww handling to i915_gem_object_ggtt_pin(). That should fix all issues without special casing !ww.
~Maarten
In the next commit, we don't evict when refcount = 0, so we need to call drain freed objects, because we want to pin new bo's in the same place, causing a test failure.
Furthermore, since each subtest is separated, it's a lot better to use i915_live_selftests, so each subtest starts with a clean slate, and a clean address space.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com --- .../gpu/drm/i915/gem/selftests/huge_pages.c | 127 +++++++++++------- 1 file changed, 79 insertions(+), 48 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/selftests/huge_pages.c b/drivers/gpu/drm/i915/gem/selftests/huge_pages.c index b2003133deaf..493509f90b35 100644 --- a/drivers/gpu/drm/i915/gem/selftests/huge_pages.c +++ b/drivers/gpu/drm/i915/gem/selftests/huge_pages.c @@ -22,6 +22,21 @@ #include "selftests/mock_region.h" #include "selftests/i915_random.h"
+struct i915_gem_context *hugepage_ctx(struct drm_i915_private *i915, struct file *file) +{ + struct i915_gem_context *ctx = live_context(i915, file); + struct i915_address_space *vm; + + if (IS_ERR(ctx)) + return ctx; + + vm = ctx->vm; + if (vm) + WRITE_ONCE(vm->scrub_64K, true); + + return ctx; +} + static const unsigned int page_sizes[] = { I915_GTT_PAGE_SIZE_2M, I915_GTT_PAGE_SIZE_64K, @@ -959,6 +974,8 @@ static int igt_mock_ppgtt_64K(void *arg) __i915_gem_object_put_pages(obj); i915_gem_object_unlock(obj); i915_gem_object_put(obj); + + i915_gem_drain_freed_objects(i915); } }
@@ -1080,10 +1097,6 @@ static int __igt_write_huge(struct intel_context *ce, if (IS_ERR(vma)) return PTR_ERR(vma);
- err = i915_vma_unbind(vma); - if (err) - return err; - err = i915_vma_pin(vma, size, 0, flags | offset); if (err) { /* @@ -1117,7 +1130,7 @@ static int __igt_write_huge(struct intel_context *ce, return err; }
-static int igt_write_huge(struct i915_gem_context *ctx, +static int igt_write_huge(struct drm_i915_private *i915, struct drm_i915_gem_object *obj) { struct i915_gem_engines *engines; @@ -1127,6 +1140,8 @@ static int igt_write_huge(struct i915_gem_context *ctx, IGT_TIMEOUT(end_time); unsigned int max_page_size; unsigned int count; + struct i915_gem_context *ctx; + struct file *file; u64 max; u64 num; u64 size; @@ -1134,6 +1149,16 @@ static int igt_write_huge(struct i915_gem_context *ctx, int i, n; int err = 0;
+ file = mock_file(i915); + if (IS_ERR(file)) + return PTR_ERR(file); + + ctx = hugepage_ctx(i915, file); + if (IS_ERR(ctx)) { + err = PTR_ERR(ctx); + goto out; + } + GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
size = obj->base.size; @@ -1153,7 +1178,7 @@ static int igt_write_huge(struct i915_gem_context *ctx, } i915_gem_context_unlock_engines(ctx); if (!n) - return 0; + goto out;
/* * To keep things interesting when alternating between engines in our @@ -1215,6 +1240,8 @@ static int igt_write_huge(struct i915_gem_context *ctx,
kfree(order);
+out: + fput(file); return err; }
@@ -1277,8 +1304,7 @@ static u32 igt_random_size(struct rnd_state *prng,
static int igt_ppgtt_smoke_huge(void *arg) { - struct i915_gem_context *ctx = arg; - struct drm_i915_private *i915 = ctx->i915; + struct drm_i915_private *i915 = arg; struct drm_i915_gem_object *obj; I915_RND_STATE(prng); struct { @@ -1302,6 +1328,7 @@ static int igt_ppgtt_smoke_huge(void *arg) u32 min = backends[i].min; u32 max = backends[i].max; u32 size = max; + try_again: size = igt_random_size(&prng, min, rounddown_pow_of_two(size));
@@ -1336,7 +1363,7 @@ static int igt_ppgtt_smoke_huge(void *arg) goto out_unpin; }
- err = igt_write_huge(ctx, obj); + err = igt_write_huge(i915, obj); if (err) { pr_err("%s write-huge failed with size=%u, i=%d\n", __func__, size, i); @@ -1363,8 +1390,7 @@ static int igt_ppgtt_smoke_huge(void *arg)
static int igt_ppgtt_sanity_check(void *arg) { - struct i915_gem_context *ctx = arg; - struct drm_i915_private *i915 = ctx->i915; + struct drm_i915_private *i915 = arg; unsigned int supported = INTEL_INFO(i915)->page_sizes; struct { igt_create_fn fn; @@ -1431,7 +1457,7 @@ static int igt_ppgtt_sanity_check(void *arg) if (pages) obj->mm.page_sizes.sg = pages;
- err = igt_write_huge(ctx, obj); + err = igt_write_huge(i915, obj);
i915_gem_object_lock(obj, NULL); i915_gem_object_unpin_pages(obj); @@ -1458,15 +1484,27 @@ static int igt_ppgtt_sanity_check(void *arg)
static int igt_tmpfs_fallback(void *arg) { - struct i915_gem_context *ctx = arg; - struct drm_i915_private *i915 = ctx->i915; + struct drm_i915_private *i915 = arg; + struct i915_address_space *vm; + struct i915_gem_context *ctx; struct vfsmount *gemfs = i915->mm.gemfs; - struct i915_address_space *vm = i915_gem_context_get_eb_vm(ctx); struct drm_i915_gem_object *obj; struct i915_vma *vma; + struct file *file; u32 *vaddr; int err = 0;
+ file = mock_file(i915); + if (IS_ERR(file)) + return PTR_ERR(file); + + ctx = hugepage_ctx(i915, file); + if (IS_ERR(ctx)) { + err = PTR_ERR(ctx); + goto out; + } + vm = i915_gem_context_get_eb_vm(ctx); + /* * Make sure that we don't burst into a ball of flames upon falling back * to tmpfs, which we rely on if on the off-chance we encouter a failure @@ -1510,33 +1548,47 @@ static int igt_tmpfs_fallback(void *arg) i915->mm.gemfs = gemfs;
i915_vm_put(vm); +out: + fput(file); return err; }
static int igt_shrink_thp(void *arg) { - struct i915_gem_context *ctx = arg; - struct drm_i915_private *i915 = ctx->i915; - struct i915_address_space *vm = i915_gem_context_get_eb_vm(ctx); + struct drm_i915_private *i915 = arg; + struct i915_address_space *vm; + struct i915_gem_context *ctx; struct drm_i915_gem_object *obj; struct i915_gem_engines_iter it; struct intel_context *ce; struct i915_vma *vma; + struct file *file; unsigned int flags = PIN_USER; unsigned int n; bool should_swap; - int err = 0; + int err; + + if (!igt_can_allocate_thp(i915)) { + pr_info("missing THP support, skipping\n"); + return 0; + } + + file = mock_file(i915); + if (IS_ERR(file)) + return PTR_ERR(file); + + ctx = hugepage_ctx(i915, file); + if (IS_ERR(ctx)) { + err = PTR_ERR(ctx); + goto out; + } + vm = i915_gem_context_get_eb_vm(ctx);
/* * Sanity check shrinking huge-paged object -- make sure nothing blows * up. */
- if (!igt_can_allocate_thp(i915)) { - pr_info("missing THP support, skipping\n"); - goto out_vm; - } - obj = i915_gem_object_create_shmem(i915, SZ_2M); if (IS_ERR(obj)) { err = PTR_ERR(obj); @@ -1626,7 +1678,8 @@ static int igt_shrink_thp(void *arg) i915_gem_object_put(obj); out_vm: i915_vm_put(vm); - +out: + fput(file); return err; }
@@ -1687,10 +1740,6 @@ int i915_gem_huge_page_live_selftests(struct drm_i915_private *i915) SUBTEST(igt_ppgtt_smoke_huge), SUBTEST(igt_ppgtt_sanity_check), }; - struct i915_gem_context *ctx; - struct i915_address_space *vm; - struct file *file; - int err;
if (!HAS_PPGTT(i915)) { pr_info("PPGTT not supported, skipping live-selftests\n"); @@ -1700,23 +1749,5 @@ int i915_gem_huge_page_live_selftests(struct drm_i915_private *i915) if (intel_gt_is_wedged(&i915->gt)) return 0;
- file = mock_file(i915); - if (IS_ERR(file)) - return PTR_ERR(file); - - ctx = live_context(i915, file); - if (IS_ERR(ctx)) { - err = PTR_ERR(ctx); - goto out_file; - } - - vm = ctx->vm; - if (vm) - WRITE_ONCE(vm->scrub_64K, true); - - err = i915_subtests(tests, ctx); - -out_file: - fput(file); - return err; + return i915_live_subtests(tests, i915); }
On Thu, 21 Oct 2021 at 11:37, Maarten Lankhorst maarten.lankhorst@linux.intel.com wrote:
In the next commit, we don't evict when refcount = 0, so we need to call drain freed objects, because we want to pin new bo's in the same place, causing a test failure.
Furthermore, since each subtest is separated, it's a lot better to use i915_live_selftests, so each subtest starts with a clean slate, and a clean address space.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
Reviewed-by: Matthew Auld matthew.auld@intel.com
Hi Maarten,
I love your patch! Perhaps something to improve:
[auto build test WARNING on drm-tip/drm-tip] [cannot apply to drm-intel/for-linux-next drm-exynos/exynos-drm-next tegra-drm/drm/tegra/for-next airlied/drm-next v5.15-rc6 next-20211021] [If your patch is applied to the wrong git tree, kindly drop us a note. And when submitting patch, we suggest to use '--base' as documented in https://git-scm.com/docs/git-format-patch]
url: https://github.com/0day-ci/linux/commits/Maarten-Lankhorst/drm-i915-Fix-i915... base: git://anongit.freedesktop.org/drm/drm-tip drm-tip config: x86_64-randconfig-a016-20211021 (attached as .config) compiler: gcc-9 (Debian 9.3.0-22) 9.3.0 reproduce (this is a W=1 build): # https://github.com/0day-ci/linux/commit/536df9470ecdddde67095f4f40f37752d944... git remote add linux-review https://github.com/0day-ci/linux git fetch --no-tags linux-review Maarten-Lankhorst/drm-i915-Fix-i915_request-fence-wait-semantics/20211021-183839 git checkout 536df9470ecdddde67095f4f40f37752d944c343 # save the attached .config to linux build tree make W=1 ARCH=x86_64
If you fix the issue, kindly add following tag as appropriate Reported-by: kernel test robot lkp@intel.com
All warnings (new ones prefixed by >>):
In file included from drivers/gpu/drm/i915/gem/i915_gem_object.c:737:
drivers/gpu/drm/i915/gem/selftests/huge_pages.c:25:26: warning: no previous prototype for 'hugepage_ctx' [-Wmissing-prototypes]
25 | struct i915_gem_context *hugepage_ctx(struct drm_i915_private *i915, struct file *file) | ^~~~~~~~~~~~
vim +/hugepage_ctx +25 drivers/gpu/drm/i915/gem/selftests/huge_pages.c
24
25 struct i915_gem_context *hugepage_ctx(struct drm_i915_private *i915, struct file *file)
26 { 27 struct i915_gem_context *ctx = live_context(i915, file); 28 struct i915_address_space *vm; 29 30 if (IS_ERR(ctx)) 31 return ctx; 32 33 vm = ctx->vm; 34 if (vm) 35 WRITE_ONCE(vm->scrub_64K, true); 36 37 return ctx; 38 } 39
--- 0-DAY CI Kernel Test Service, Intel Corporation https://lists.01.org/hyperkitty/list/kbuild-all@lists.01.org
Hi Maarten,
I love your patch! Yet something to improve:
[auto build test ERROR on drm-tip/drm-tip] [cannot apply to drm-intel/for-linux-next drm-exynos/exynos-drm-next tegra-drm/drm/tegra/for-next airlied/drm-next v5.15-rc6 next-20211021] [If your patch is applied to the wrong git tree, kindly drop us a note. And when submitting patch, we suggest to use '--base' as documented in https://git-scm.com/docs/git-format-patch]
url: https://github.com/0day-ci/linux/commits/Maarten-Lankhorst/drm-i915-Fix-i915... base: git://anongit.freedesktop.org/drm/drm-tip drm-tip config: x86_64-randconfig-a011-20211021 (attached as .config) compiler: gcc-9 (Debian 9.3.0-22) 9.3.0 reproduce (this is a W=1 build): # https://github.com/0day-ci/linux/commit/536df9470ecdddde67095f4f40f37752d944... git remote add linux-review https://github.com/0day-ci/linux git fetch --no-tags linux-review Maarten-Lankhorst/drm-i915-Fix-i915_request-fence-wait-semantics/20211021-183839 git checkout 536df9470ecdddde67095f4f40f37752d944c343 # save the attached .config to linux build tree make W=1 ARCH=x86_64
If you fix the issue, kindly add following tag as appropriate Reported-by: kernel test robot lkp@intel.com
All errors (new ones prefixed by >>):
In file included from drivers/gpu/drm/i915/gem/i915_gem_object.c:737:
drivers/gpu/drm/i915/gem/selftests/huge_pages.c:25:26: error: no previous prototype for 'hugepage_ctx' [-Werror=missing-prototypes]
25 | struct i915_gem_context *hugepage_ctx(struct drm_i915_private *i915, struct file *file) | ^~~~~~~~~~~~ cc1: all warnings being treated as errors
vim +/hugepage_ctx +25 drivers/gpu/drm/i915/gem/selftests/huge_pages.c
24
25 struct i915_gem_context *hugepage_ctx(struct drm_i915_private *i915, struct file *file)
26 { 27 struct i915_gem_context *ctx = live_context(i915, file); 28 struct i915_address_space *vm; 29 30 if (IS_ERR(ctx)) 31 return ctx; 32 33 vm = ctx->vm; 34 if (vm) 35 WRITE_ONCE(vm->scrub_64K, true); 36 37 return ctx; 38 } 39
--- 0-DAY CI Kernel Test Service, Intel Corporation https://lists.01.org/hyperkitty/list/kbuild-all@lists.01.org
In the next commit, we don't evict when refcount = 0.
igt_vm_isolation() continuously tries to pin/unpin at same address, but also calls put() on the object, which means the object may not be unpinned in time.
Instead of this, re-use the same object over and over, so they can be unbound as required.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com --- .../drm/i915/gem/selftests/i915_gem_context.c | 54 +++++++++++-------- 1 file changed, 32 insertions(+), 22 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/selftests/i915_gem_context.c b/drivers/gpu/drm/i915/gem/selftests/i915_gem_context.c index b32f7fed2d9c..3fc595b57cf4 100644 --- a/drivers/gpu/drm/i915/gem/selftests/i915_gem_context.c +++ b/drivers/gpu/drm/i915/gem/selftests/i915_gem_context.c @@ -1481,10 +1481,10 @@ static int check_scratch(struct i915_address_space *vm, u64 offset)
static int write_to_scratch(struct i915_gem_context *ctx, struct intel_engine_cs *engine, + struct drm_i915_gem_object *obj, u64 offset, u32 value) { struct drm_i915_private *i915 = ctx->i915; - struct drm_i915_gem_object *obj; struct i915_address_space *vm; struct i915_request *rq; struct i915_vma *vma; @@ -1497,15 +1497,9 @@ static int write_to_scratch(struct i915_gem_context *ctx, if (err) return err;
- obj = i915_gem_object_create_internal(i915, PAGE_SIZE); - if (IS_ERR(obj)) - return PTR_ERR(obj); - cmd = i915_gem_object_pin_map_unlocked(obj, I915_MAP_WB); - if (IS_ERR(cmd)) { - err = PTR_ERR(cmd); - goto out; - } + if (IS_ERR(cmd)) + return PTR_ERR(cmd);
*cmd++ = MI_STORE_DWORD_IMM_GEN4; if (GRAPHICS_VER(i915) >= 8) { @@ -1569,17 +1563,19 @@ static int write_to_scratch(struct i915_gem_context *ctx, i915_vma_unpin(vma); out_vm: i915_vm_put(vm); -out: - i915_gem_object_put(obj); + + if (!err) + err = i915_gem_object_wait(obj, 0, MAX_SCHEDULE_TIMEOUT); + return err; }
static int read_from_scratch(struct i915_gem_context *ctx, struct intel_engine_cs *engine, + struct drm_i915_gem_object *obj, u64 offset, u32 *value) { struct drm_i915_private *i915 = ctx->i915; - struct drm_i915_gem_object *obj; struct i915_address_space *vm; const u32 result = 0x100; struct i915_request *rq; @@ -1594,10 +1590,6 @@ static int read_from_scratch(struct i915_gem_context *ctx, if (err) return err;
- obj = i915_gem_object_create_internal(i915, PAGE_SIZE); - if (IS_ERR(obj)) - return PTR_ERR(obj); - if (GRAPHICS_VER(i915) >= 8) { const u32 GPR0 = engine->mmio_base + 0x600;
@@ -1615,7 +1607,7 @@ static int read_from_scratch(struct i915_gem_context *ctx, cmd = i915_gem_object_pin_map_unlocked(obj, I915_MAP_WB); if (IS_ERR(cmd)) { err = PTR_ERR(cmd); - goto out; + goto err_unpin; }
memset(cmd, POISON_INUSE, PAGE_SIZE); @@ -1651,7 +1643,7 @@ static int read_from_scratch(struct i915_gem_context *ctx, cmd = i915_gem_object_pin_map_unlocked(obj, I915_MAP_WB); if (IS_ERR(cmd)) { err = PTR_ERR(cmd); - goto out; + goto err_unpin; }
memset(cmd, POISON_INUSE, PAGE_SIZE); @@ -1722,8 +1714,10 @@ static int read_from_scratch(struct i915_gem_context *ctx, i915_vma_unpin(vma); out_vm: i915_vm_put(vm); -out: - i915_gem_object_put(obj); + + if (!err) + err = i915_gem_object_wait(obj, 0, MAX_SCHEDULE_TIMEOUT); + return err; }
@@ -1765,6 +1759,7 @@ static int igt_vm_isolation(void *arg) u64 vm_total; u32 expected; int err; + struct drm_i915_gem_object *obj_a, *obj_b;
if (GRAPHICS_VER(i915) < 7) return 0; @@ -1810,6 +1805,18 @@ static int igt_vm_isolation(void *arg) vm_total = ctx_a->vm->total; GEM_BUG_ON(ctx_b->vm->total != vm_total);
+ obj_a = i915_gem_object_create_internal(i915, PAGE_SIZE); + if (IS_ERR(obj_a)) { + err = PTR_ERR(obj_a); + goto out_file; + } + + obj_b = i915_gem_object_create_internal(i915, PAGE_SIZE); + if (IS_ERR(obj_b)) { + err = PTR_ERR(obj_b); + goto put_a; + } + count = 0; num_engines = 0; for_each_uabi_engine(engine, i915) { @@ -1832,10 +1839,10 @@ static int igt_vm_isolation(void *arg) I915_GTT_PAGE_SIZE, vm_total, sizeof(u32), alignof_dword);
- err = write_to_scratch(ctx_a, engine, + err = write_to_scratch(ctx_a, engine, obj_a, offset, 0xdeadbeef); if (err == 0) - err = read_from_scratch(ctx_b, engine, + err = read_from_scratch(ctx_b, engine, obj_b, offset, &value); if (err) goto out_file; @@ -1858,6 +1865,9 @@ static int igt_vm_isolation(void *arg) pr_info("Checked %lu scratch offsets across %lu engines\n", count, num_engines);
+ i915_gem_object_put(obj_b); +put_a: + i915_gem_object_put(obj_a); out_file: if (igt_live_test_end(&t)) err = -EIO;
Now that freeing objects takes the object lock when destroying the backing pages, we can confidently take the object lock even for dead objects.
Use this fact to take the object lock in the shrinker, without requiring a reference to the object, so all calls to unbind take the object lock.
This is the last step to requiring the object lock for vma_unbind.
Changes since v1: - No longer require the refcount, as every freed object now holds the lock when unbinding VMA's.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com --- drivers/gpu/drm/i915/gem/i915_gem_shrinker.c | 6 ++++ drivers/gpu/drm/i915/i915_gem_evict.c | 34 +++++++++++++++++--- 2 files changed, 35 insertions(+), 5 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c b/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c index d3f29a66cb36..34c12e5983eb 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c @@ -403,12 +403,18 @@ i915_gem_shrinker_vmap(struct notifier_block *nb, unsigned long event, void *ptr list_for_each_entry_safe(vma, next, &i915->ggtt.vm.bound_list, vm_link) { unsigned long count = vma->node.size >> PAGE_SHIFT; + struct drm_i915_gem_object *obj = vma->obj;
if (!vma->iomap || i915_vma_is_active(vma)) continue;
+ if (!i915_gem_object_trylock(obj)) + continue; + if (__i915_vma_unbind(vma) == 0) freed_pages += count; + + i915_gem_object_unlock(obj); } mutex_unlock(&i915->ggtt.vm.mutex);
diff --git a/drivers/gpu/drm/i915/i915_gem_evict.c b/drivers/gpu/drm/i915/i915_gem_evict.c index 2b73ddb11c66..286efa462eca 100644 --- a/drivers/gpu/drm/i915/i915_gem_evict.c +++ b/drivers/gpu/drm/i915/i915_gem_evict.c @@ -58,6 +58,9 @@ mark_free(struct drm_mm_scan *scan, if (i915_vma_is_pinned(vma)) return false;
+ if (!i915_gem_object_trylock(vma->obj)) + return false; + list_add(&vma->evict_link, unwind); return drm_mm_scan_add_block(scan, &vma->node); } @@ -178,6 +181,7 @@ i915_gem_evict_something(struct i915_address_space *vm, list_for_each_entry_safe(vma, next, &eviction_list, evict_link) { ret = drm_mm_scan_remove_block(&scan, &vma->node); BUG_ON(ret); + i915_gem_object_unlock(vma->obj); }
/* @@ -222,10 +226,12 @@ i915_gem_evict_something(struct i915_address_space *vm, * of any of our objects, thus corrupting the list). */ list_for_each_entry_safe(vma, next, &eviction_list, evict_link) { - if (drm_mm_scan_remove_block(&scan, &vma->node)) + if (drm_mm_scan_remove_block(&scan, &vma->node)) { __i915_vma_pin(vma); - else + } else { list_del(&vma->evict_link); + i915_gem_object_unlock(vma->obj); + } }
/* Unbinding will emit any required flushes */ @@ -234,16 +240,22 @@ i915_gem_evict_something(struct i915_address_space *vm, __i915_vma_unpin(vma); if (ret == 0) ret = __i915_vma_unbind(vma); + + i915_gem_object_unlock(vma->obj); }
while (ret == 0 && (node = drm_mm_scan_color_evict(&scan))) { vma = container_of(node, struct i915_vma, node);
+ /* If we find any non-objects (!vma), we cannot evict them */ - if (vma->node.color != I915_COLOR_UNEVICTABLE) + if (vma->node.color != I915_COLOR_UNEVICTABLE && + i915_gem_object_trylock(vma->obj)) { ret = __i915_vma_unbind(vma); - else - ret = -ENOSPC; /* XXX search failed, try again? */ + i915_gem_object_unlock(vma->obj); + } else { + ret = -ENOSPC; + } }
return ret; @@ -333,6 +345,11 @@ int i915_gem_evict_for_node(struct i915_address_space *vm, break; }
+ if (!i915_gem_object_trylock(vma->obj)) { + ret = -ENOSPC; + break; + } + /* * Never show fear in the face of dragons! * @@ -350,6 +367,8 @@ int i915_gem_evict_for_node(struct i915_address_space *vm, __i915_vma_unpin(vma); if (ret == 0) ret = __i915_vma_unbind(vma); + + i915_gem_object_unlock(vma->obj); }
return ret; @@ -393,6 +412,9 @@ int i915_gem_evict_vm(struct i915_address_space *vm) if (i915_vma_is_pinned(vma)) continue;
+ if (!i915_gem_object_trylock(vma->obj)) + continue; + __i915_vma_pin(vma); list_add(&vma->evict_link, &eviction_list); } @@ -406,6 +428,8 @@ int i915_gem_evict_vm(struct i915_address_space *vm) ret = __i915_vma_unbind(vma); if (ret != -EINTR) /* "Get me out of here!" */ ret = 0; + + i915_gem_object_unlock(vma->obj); } } while (ret == 0);
On Thu, 21 Oct 2021 at 11:37, Maarten Lankhorst maarten.lankhorst@linux.intel.com wrote:
Now that freeing objects takes the object lock when destroying the backing pages, we can confidently take the object lock even for dead objects.
Use this fact to take the object lock in the shrinker, without requiring a reference to the object, so all calls to unbind take the object lock.
This is the last step to requiring the object lock for vma_unbind.
For the eviction what is the reason for only trylock here, assuming we are given a ww context? Maybe the back off is annoying? And the full lock version comes later?
Changes since v1:
- No longer require the refcount, as every freed object now holds the lock when unbinding VMA's.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
drivers/gpu/drm/i915/gem/i915_gem_shrinker.c | 6 ++++ drivers/gpu/drm/i915/i915_gem_evict.c | 34 +++++++++++++++++--- 2 files changed, 35 insertions(+), 5 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c b/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c index d3f29a66cb36..34c12e5983eb 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c @@ -403,12 +403,18 @@ i915_gem_shrinker_vmap(struct notifier_block *nb, unsigned long event, void *ptr list_for_each_entry_safe(vma, next, &i915->ggtt.vm.bound_list, vm_link) { unsigned long count = vma->node.size >> PAGE_SHIFT;
struct drm_i915_gem_object *obj = vma->obj; if (!vma->iomap || i915_vma_is_active(vma)) continue;if (!i915_gem_object_trylock(obj))continue;if (__i915_vma_unbind(vma) == 0) freed_pages += count;i915_gem_object_unlock(obj); } mutex_unlock(&i915->ggtt.vm.mutex);diff --git a/drivers/gpu/drm/i915/i915_gem_evict.c b/drivers/gpu/drm/i915/i915_gem_evict.c index 2b73ddb11c66..286efa462eca 100644 --- a/drivers/gpu/drm/i915/i915_gem_evict.c +++ b/drivers/gpu/drm/i915/i915_gem_evict.c @@ -58,6 +58,9 @@ mark_free(struct drm_mm_scan *scan, if (i915_vma_is_pinned(vma)) return false;
if (!i915_gem_object_trylock(vma->obj))return false;list_add(&vma->evict_link, unwind); return drm_mm_scan_add_block(scan, &vma->node);} @@ -178,6 +181,7 @@ i915_gem_evict_something(struct i915_address_space *vm, list_for_each_entry_safe(vma, next, &eviction_list, evict_link) { ret = drm_mm_scan_remove_block(&scan, &vma->node); BUG_ON(ret);
i915_gem_object_unlock(vma->obj); } /*@@ -222,10 +226,12 @@ i915_gem_evict_something(struct i915_address_space *vm, * of any of our objects, thus corrupting the list). */ list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
if (drm_mm_scan_remove_block(&scan, &vma->node))
if (drm_mm_scan_remove_block(&scan, &vma->node)) { __i915_vma_pin(vma);
else
} else { list_del(&vma->evict_link);i915_gem_object_unlock(vma->obj);} } /* Unbinding will emit any required flushes */@@ -234,16 +240,22 @@ i915_gem_evict_something(struct i915_address_space *vm, __i915_vma_unpin(vma); if (ret == 0) ret = __i915_vma_unbind(vma);
i915_gem_object_unlock(vma->obj); } while (ret == 0 && (node = drm_mm_scan_color_evict(&scan))) { vma = container_of(node, struct i915_vma, node);/* If we find any non-objects (!vma), we cannot evict them */
if (vma->node.color != I915_COLOR_UNEVICTABLE)
if (vma->node.color != I915_COLOR_UNEVICTABLE &&i915_gem_object_trylock(vma->obj)) { ret = __i915_vma_unbind(vma);
elseret = -ENOSPC; /* XXX search failed, try again? */
i915_gem_object_unlock(vma->obj);} else {ret = -ENOSPC;} } return ret;@@ -333,6 +345,11 @@ int i915_gem_evict_for_node(struct i915_address_space *vm, break; }
if (!i915_gem_object_trylock(vma->obj)) {ret = -ENOSPC;break;}/* * Never show fear in the face of dragons! *@@ -350,6 +367,8 @@ int i915_gem_evict_for_node(struct i915_address_space *vm, __i915_vma_unpin(vma); if (ret == 0) ret = __i915_vma_unbind(vma);
i915_gem_object_unlock(vma->obj); } return ret;@@ -393,6 +412,9 @@ int i915_gem_evict_vm(struct i915_address_space *vm) if (i915_vma_is_pinned(vma)) continue;
if (!i915_gem_object_trylock(vma->obj))continue;__i915_vma_pin(vma); list_add(&vma->evict_link, &eviction_list); }@@ -406,6 +428,8 @@ int i915_gem_evict_vm(struct i915_address_space *vm) ret = __i915_vma_unbind(vma); if (ret != -EINTR) /* "Get me out of here!" */ ret = 0;
i915_gem_object_unlock(vma->obj); } } while (ret == 0);-- 2.33.0
Op 21-10-2021 om 19:59 schreef Matthew Auld:
On Thu, 21 Oct 2021 at 11:37, Maarten Lankhorst maarten.lankhorst@linux.intel.com wrote:
Now that freeing objects takes the object lock when destroying the backing pages, we can confidently take the object lock even for dead objects.
Use this fact to take the object lock in the shrinker, without requiring a reference to the object, so all calls to unbind take the object lock.
This is the last step to requiring the object lock for vma_unbind.
For the eviction what is the reason for only trylock here, assuming we are given a ww context? Maybe the back off is annoying? And the full lock version comes later?
2 reasons:
1. We can't take the full lock, because we already hold vm->mutex, which may be held inside dma_resv_lock. This inverts the locking, and is also why we could not keep obj->mm.lock. Until locking for vm is reworked, you cannot do this anyway.
Lockdep will complain about the following lock cycle: dma_resv_lock -> vm->mutex -> dma_resv_lock, and will eventually deadlock.
2. Until locking or delayed destroy is reworked, we cannot call a blocking dma_resv_lock for objects in the list when the refcount may be 0.
"[PATCH 25/28] drm/i915: Require object lock when freeing pages during destruction".
When destroying the object, we will take dma_resv_lock in blocking mode one last time, then unbind all its vma's. The fact we're holding vm->mutex prevents the object from disappearing, because its vma is not yet unbound. This is how we can get away with unbinding dead objects currently, before and after the changes. This also means we can only trylock, because we can only trylock inside vm->mutex.
If we start reworking vm locking, we may need to handle waiting on dead objects better. It's worth noting that TTM has to handle the exact same race, which can be seen inside ttm_bo_cleanup_refs().
Changes since v1:
- No longer require the refcount, as every freed object now holds the lock when unbinding VMA's.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
drivers/gpu/drm/i915/gem/i915_gem_shrinker.c | 6 ++++ drivers/gpu/drm/i915/i915_gem_evict.c | 34 +++++++++++++++++--- 2 files changed, 35 insertions(+), 5 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c b/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c index d3f29a66cb36..34c12e5983eb 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c @@ -403,12 +403,18 @@ i915_gem_shrinker_vmap(struct notifier_block *nb, unsigned long event, void *ptr list_for_each_entry_safe(vma, next, &i915->ggtt.vm.bound_list, vm_link) { unsigned long count = vma->node.size >> PAGE_SHIFT;
struct drm_i915_gem_object *obj = vma->obj; if (!vma->iomap || i915_vma_is_active(vma)) continue;if (!i915_gem_object_trylock(obj))continue;if (__i915_vma_unbind(vma) == 0) freed_pages += count;i915_gem_object_unlock(obj); } mutex_unlock(&i915->ggtt.vm.mutex);diff --git a/drivers/gpu/drm/i915/i915_gem_evict.c b/drivers/gpu/drm/i915/i915_gem_evict.c index 2b73ddb11c66..286efa462eca 100644 --- a/drivers/gpu/drm/i915/i915_gem_evict.c +++ b/drivers/gpu/drm/i915/i915_gem_evict.c @@ -58,6 +58,9 @@ mark_free(struct drm_mm_scan *scan, if (i915_vma_is_pinned(vma)) return false;
if (!i915_gem_object_trylock(vma->obj))return false;list_add(&vma->evict_link, unwind); return drm_mm_scan_add_block(scan, &vma->node);} @@ -178,6 +181,7 @@ i915_gem_evict_something(struct i915_address_space *vm, list_for_each_entry_safe(vma, next, &eviction_list, evict_link) { ret = drm_mm_scan_remove_block(&scan, &vma->node); BUG_ON(ret);
i915_gem_object_unlock(vma->obj); } /*@@ -222,10 +226,12 @@ i915_gem_evict_something(struct i915_address_space *vm, * of any of our objects, thus corrupting the list). */ list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
if (drm_mm_scan_remove_block(&scan, &vma->node))
if (drm_mm_scan_remove_block(&scan, &vma->node)) { __i915_vma_pin(vma);
else
} else { list_del(&vma->evict_link);i915_gem_object_unlock(vma->obj);} } /* Unbinding will emit any required flushes */@@ -234,16 +240,22 @@ i915_gem_evict_something(struct i915_address_space *vm, __i915_vma_unpin(vma); if (ret == 0) ret = __i915_vma_unbind(vma);
i915_gem_object_unlock(vma->obj); } while (ret == 0 && (node = drm_mm_scan_color_evict(&scan))) { vma = container_of(node, struct i915_vma, node);/* If we find any non-objects (!vma), we cannot evict them */
if (vma->node.color != I915_COLOR_UNEVICTABLE)
if (vma->node.color != I915_COLOR_UNEVICTABLE &&i915_gem_object_trylock(vma->obj)) { ret = __i915_vma_unbind(vma);
elseret = -ENOSPC; /* XXX search failed, try again? */
i915_gem_object_unlock(vma->obj);} else {ret = -ENOSPC;} } return ret;@@ -333,6 +345,11 @@ int i915_gem_evict_for_node(struct i915_address_space *vm, break; }
if (!i915_gem_object_trylock(vma->obj)) {ret = -ENOSPC;break;}/* * Never show fear in the face of dragons! *@@ -350,6 +367,8 @@ int i915_gem_evict_for_node(struct i915_address_space *vm, __i915_vma_unpin(vma); if (ret == 0) ret = __i915_vma_unbind(vma);
i915_gem_object_unlock(vma->obj); } return ret;@@ -393,6 +412,9 @@ int i915_gem_evict_vm(struct i915_address_space *vm) if (i915_vma_is_pinned(vma)) continue;
if (!i915_gem_object_trylock(vma->obj))continue;__i915_vma_pin(vma); list_add(&vma->evict_link, &eviction_list); }@@ -406,6 +428,8 @@ int i915_gem_evict_vm(struct i915_address_space *vm) ret = __i915_vma_unbind(vma); if (ret != -EINTR) /* "Get me out of here!" */ ret = 0;
i915_gem_object_unlock(vma->obj); } } while (ret == 0);-- 2.33.0
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com --- .../gpu/drm/i915/gem/i915_gem_execbuffer.c | 2 +- drivers/gpu/drm/i915/gem/i915_gem_object.h | 8 ++++-- drivers/gpu/drm/i915/gem/i915_gem_shrinker.c | 4 +-- drivers/gpu/drm/i915/gem/i915_gem_stolen.c | 2 +- drivers/gpu/drm/i915/gt/intel_engine_pm.c | 2 +- drivers/gpu/drm/i915/gt/intel_ggtt.c | 2 +- drivers/gpu/drm/i915/gt/mock_engine.c | 2 +- drivers/gpu/drm/i915/gt/selftest_hangcheck.c | 2 +- drivers/gpu/drm/i915/gt/selftest_migrate.c | 2 +- drivers/gpu/drm/i915/gvt/aperture_gm.c | 2 +- drivers/gpu/drm/i915/i915_drv.h | 5 +++- drivers/gpu/drm/i915/i915_gem_evict.c | 15 ++++++----- drivers/gpu/drm/i915/i915_gem_gtt.c | 8 +++--- drivers/gpu/drm/i915/i915_gem_gtt.h | 3 +++ drivers/gpu/drm/i915/i915_vgpu.c | 2 +- drivers/gpu/drm/i915/i915_vma.c | 13 +++++---- .../gpu/drm/i915/selftests/i915_gem_evict.c | 27 ++++++++++++------- drivers/gpu/drm/i915/selftests/i915_gem_gtt.c | 14 +++++----- 18 files changed, 70 insertions(+), 45 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c index e614591ca510..bbf2a10738f7 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c @@ -764,7 +764,7 @@ static int eb_reserve(struct i915_execbuffer *eb) case 1: /* Too fragmented, unbind everything and retry */ mutex_lock(&eb->context->vm->mutex); - err = i915_gem_evict_vm(eb->context->vm); + err = i915_gem_evict_vm(eb->context->vm, &eb->ww); mutex_unlock(&eb->context->vm->mutex); if (err) return err; diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.h b/drivers/gpu/drm/i915/gem/i915_gem_object.h index 59201801cec5..accd8fad0ed7 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_object.h +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.h @@ -211,9 +211,13 @@ static inline int i915_gem_object_lock_interruptible(struct drm_i915_gem_object return __i915_gem_object_lock(obj, ww, true); }
-static inline bool i915_gem_object_trylock(struct drm_i915_gem_object *obj) +static inline bool i915_gem_object_trylock(struct drm_i915_gem_object *obj, + struct i915_gem_ww_ctx *ww) { - return dma_resv_trylock(obj->base.resv); + if (!ww) + return dma_resv_trylock(obj->base.resv); + else + return ww_mutex_trylock(&obj->base.resv->lock, &ww->ctx); }
static inline void i915_gem_object_unlock(struct drm_i915_gem_object *obj) diff --git a/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c b/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c index 34c12e5983eb..5375f3f9f016 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c @@ -210,7 +210,7 @@ i915_gem_shrink(struct i915_gem_ww_ctx *ww,
/* May arrive from get_pages on another bo */ if (!ww) { - if (!i915_gem_object_trylock(obj)) + if (!i915_gem_object_trylock(obj, NULL)) goto skip; } else { err = i915_gem_object_lock(obj, ww); @@ -408,7 +408,7 @@ i915_gem_shrinker_vmap(struct notifier_block *nb, unsigned long event, void *ptr if (!vma->iomap || i915_vma_is_active(vma)) continue;
- if (!i915_gem_object_trylock(obj)) + if (!i915_gem_object_trylock(obj, NULL)) continue;
if (__i915_vma_unbind(vma) == 0) diff --git a/drivers/gpu/drm/i915/gem/i915_gem_stolen.c b/drivers/gpu/drm/i915/gem/i915_gem_stolen.c index ddd37ccb1362..d629ef5abf9a 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_stolen.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_stolen.c @@ -653,7 +653,7 @@ static int __i915_gem_object_create_stolen(struct intel_memory_region *mem, cache_level = HAS_LLC(mem->i915) ? I915_CACHE_LLC : I915_CACHE_NONE; i915_gem_object_set_cache_coherency(obj, cache_level);
- if (WARN_ON(!i915_gem_object_trylock(obj))) + if (WARN_ON(!i915_gem_object_trylock(obj, NULL))) return -EBUSY;
i915_gem_object_init_memory_region(obj, mem); diff --git a/drivers/gpu/drm/i915/gt/intel_engine_pm.c b/drivers/gpu/drm/i915/gt/intel_engine_pm.c index 849fbb229bd3..d66c8bd50ae3 100644 --- a/drivers/gpu/drm/i915/gt/intel_engine_pm.c +++ b/drivers/gpu/drm/i915/gt/intel_engine_pm.c @@ -26,7 +26,7 @@ static void dbg_poison_ce(struct intel_context *ce) int type = i915_coherent_map_type(ce->engine->i915, obj, true); void *map;
- if (!i915_gem_object_trylock(obj)) + if (!i915_gem_object_trylock(obj, NULL)) return;
map = i915_gem_object_pin_map(obj, type); diff --git a/drivers/gpu/drm/i915/gt/intel_ggtt.c b/drivers/gpu/drm/i915/gt/intel_ggtt.c index 6da57199bb33..df29fd992b45 100644 --- a/drivers/gpu/drm/i915/gt/intel_ggtt.c +++ b/drivers/gpu/drm/i915/gt/intel_ggtt.c @@ -487,7 +487,7 @@ static int ggtt_reserve_guc_top(struct i915_ggtt *ggtt) GEM_BUG_ON(ggtt->vm.total <= GUC_GGTT_TOP); size = ggtt->vm.total - GUC_GGTT_TOP;
- ret = i915_gem_gtt_reserve(&ggtt->vm, &ggtt->uc_fw, size, + ret = i915_gem_gtt_reserve(&ggtt->vm, NULL, &ggtt->uc_fw, size, GUC_GGTT_TOP, I915_COLOR_UNEVICTABLE, PIN_NOEVICT); if (ret) diff --git a/drivers/gpu/drm/i915/gt/mock_engine.c b/drivers/gpu/drm/i915/gt/mock_engine.c index bb99fc03f503..240265354274 100644 --- a/drivers/gpu/drm/i915/gt/mock_engine.c +++ b/drivers/gpu/drm/i915/gt/mock_engine.c @@ -17,7 +17,7 @@ static int mock_timeline_pin(struct intel_timeline *tl) { int err;
- if (WARN_ON(!i915_gem_object_trylock(tl->hwsp_ggtt->obj))) + if (WARN_ON(!i915_gem_object_trylock(tl->hwsp_ggtt->obj, NULL))) return -EBUSY;
err = intel_timeline_pin_map(tl); diff --git a/drivers/gpu/drm/i915/gt/selftest_hangcheck.c b/drivers/gpu/drm/i915/gt/selftest_hangcheck.c index 7e2d99dd012d..572f942c3c5d 100644 --- a/drivers/gpu/drm/i915/gt/selftest_hangcheck.c +++ b/drivers/gpu/drm/i915/gt/selftest_hangcheck.c @@ -1378,7 +1378,7 @@ static int evict_vma(void *data) complete(&arg->completion);
mutex_lock(&vm->mutex); - err = i915_gem_evict_for_node(vm, &evict, 0); + err = i915_gem_evict_for_node(vm, NULL, &evict, 0); mutex_unlock(&vm->mutex);
return err; diff --git a/drivers/gpu/drm/i915/gt/selftest_migrate.c b/drivers/gpu/drm/i915/gt/selftest_migrate.c index 12ef2837c89b..78993c760f12 100644 --- a/drivers/gpu/drm/i915/gt/selftest_migrate.c +++ b/drivers/gpu/drm/i915/gt/selftest_migrate.c @@ -464,7 +464,7 @@ create_init_lmem_internal(struct intel_gt *gt, size_t sz, bool try_lmem) return obj; }
- i915_gem_object_trylock(obj); + i915_gem_object_trylock(obj, NULL); err = i915_gem_object_pin_pages(obj); if (err) { i915_gem_object_unlock(obj); diff --git a/drivers/gpu/drm/i915/gvt/aperture_gm.c b/drivers/gpu/drm/i915/gvt/aperture_gm.c index 0d6d59871308..c08098a167e9 100644 --- a/drivers/gpu/drm/i915/gvt/aperture_gm.c +++ b/drivers/gpu/drm/i915/gvt/aperture_gm.c @@ -63,7 +63,7 @@ static int alloc_gm(struct intel_vgpu *vgpu, bool high_gm)
mutex_lock(>->ggtt->vm.mutex); mmio_hw_access_pre(gt); - ret = i915_gem_gtt_insert(>->ggtt->vm, node, + ret = i915_gem_gtt_insert(>->ggtt->vm, NULL, node, size, I915_GTT_PAGE_SIZE, I915_COLOR_UNEVICTABLE, start, end, flags); diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h index a8d733a7e1d9..22c891720c6d 100644 --- a/drivers/gpu/drm/i915/i915_drv.h +++ b/drivers/gpu/drm/i915/i915_drv.h @@ -1920,14 +1920,17 @@ i915_gem_vm_lookup(struct drm_i915_file_private *file_priv, u32 id)
/* i915_gem_evict.c */ int __must_check i915_gem_evict_something(struct i915_address_space *vm, + struct i915_gem_ww_ctx *ww, u64 min_size, u64 alignment, unsigned long color, u64 start, u64 end, unsigned flags); int __must_check i915_gem_evict_for_node(struct i915_address_space *vm, + struct i915_gem_ww_ctx *ww, struct drm_mm_node *node, unsigned int flags); -int i915_gem_evict_vm(struct i915_address_space *vm); +int i915_gem_evict_vm(struct i915_address_space *vm, + struct i915_gem_ww_ctx *ww);
/* i915_gem_internal.c */ struct drm_i915_gem_object * diff --git a/drivers/gpu/drm/i915/i915_gem_evict.c b/drivers/gpu/drm/i915/i915_gem_evict.c index 286efa462eca..24f5e3345e43 100644 --- a/drivers/gpu/drm/i915/i915_gem_evict.c +++ b/drivers/gpu/drm/i915/i915_gem_evict.c @@ -51,6 +51,7 @@ static int ggtt_flush(struct intel_gt *gt)
static bool mark_free(struct drm_mm_scan *scan, + struct i915_gem_ww_ctx *ww, struct i915_vma *vma, unsigned int flags, struct list_head *unwind) @@ -58,7 +59,7 @@ mark_free(struct drm_mm_scan *scan, if (i915_vma_is_pinned(vma)) return false;
- if (!i915_gem_object_trylock(vma->obj)) + if (!i915_gem_object_trylock(vma->obj, ww)) return false;
list_add(&vma->evict_link, unwind); @@ -101,6 +102,7 @@ static bool defer_evict(struct i915_vma *vma) */ int i915_gem_evict_something(struct i915_address_space *vm, + struct i915_gem_ww_ctx *ww, u64 min_size, u64 alignment, unsigned long color, u64 start, u64 end, @@ -173,7 +175,7 @@ i915_gem_evict_something(struct i915_address_space *vm, continue; }
- if (mark_free(&scan, vma, flags, &eviction_list)) + if (mark_free(&scan, ww, vma, flags, &eviction_list)) goto found; }
@@ -250,7 +252,7 @@ i915_gem_evict_something(struct i915_address_space *vm,
/* If we find any non-objects (!vma), we cannot evict them */ if (vma->node.color != I915_COLOR_UNEVICTABLE && - i915_gem_object_trylock(vma->obj)) { + i915_gem_object_trylock(vma->obj, ww)) { ret = __i915_vma_unbind(vma); i915_gem_object_unlock(vma->obj); } else { @@ -273,6 +275,7 @@ i915_gem_evict_something(struct i915_address_space *vm, * memory in e.g. the shrinker. */ int i915_gem_evict_for_node(struct i915_address_space *vm, + struct i915_gem_ww_ctx *ww, struct drm_mm_node *target, unsigned int flags) { @@ -345,7 +348,7 @@ int i915_gem_evict_for_node(struct i915_address_space *vm, break; }
- if (!i915_gem_object_trylock(vma->obj)) { + if (!i915_gem_object_trylock(vma->obj, ww)) { ret = -ENOSPC; break; } @@ -386,7 +389,7 @@ int i915_gem_evict_for_node(struct i915_address_space *vm, * To clarify: This is for freeing up virtual address space, not for freeing * memory in e.g. the shrinker. */ -int i915_gem_evict_vm(struct i915_address_space *vm) +int i915_gem_evict_vm(struct i915_address_space *vm, struct i915_gem_ww_ctx *ww) { int ret = 0;
@@ -412,7 +415,7 @@ int i915_gem_evict_vm(struct i915_address_space *vm) if (i915_vma_is_pinned(vma)) continue;
- if (!i915_gem_object_trylock(vma->obj)) + if (!i915_gem_object_trylock(vma->obj, ww)) continue;
__i915_vma_pin(vma); diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.c b/drivers/gpu/drm/i915/i915_gem_gtt.c index cd5f2348a187..c810da476c2b 100644 --- a/drivers/gpu/drm/i915/i915_gem_gtt.c +++ b/drivers/gpu/drm/i915/i915_gem_gtt.c @@ -93,6 +93,7 @@ void i915_gem_gtt_finish_pages(struct drm_i915_gem_object *obj, * asked to wait for eviction and interrupted. */ int i915_gem_gtt_reserve(struct i915_address_space *vm, + struct i915_gem_ww_ctx *ww, struct drm_mm_node *node, u64 size, u64 offset, unsigned long color, unsigned int flags) @@ -117,7 +118,7 @@ int i915_gem_gtt_reserve(struct i915_address_space *vm, if (flags & PIN_NOEVICT) return -ENOSPC;
- err = i915_gem_evict_for_node(vm, node, flags); + err = i915_gem_evict_for_node(vm, ww, node, flags); if (err == 0) err = drm_mm_reserve_node(&vm->mm, node);
@@ -184,6 +185,7 @@ static u64 random_offset(u64 start, u64 end, u64 len, u64 align) * asked to wait for eviction and interrupted. */ int i915_gem_gtt_insert(struct i915_address_space *vm, + struct i915_gem_ww_ctx *ww, struct drm_mm_node *node, u64 size, u64 alignment, unsigned long color, u64 start, u64 end, unsigned int flags) @@ -269,7 +271,7 @@ int i915_gem_gtt_insert(struct i915_address_space *vm, */ offset = random_offset(start, end, size, alignment ?: I915_GTT_MIN_ALIGNMENT); - err = i915_gem_gtt_reserve(vm, node, size, offset, color, flags); + err = i915_gem_gtt_reserve(vm, ww, node, size, offset, color, flags); if (err != -ENOSPC) return err;
@@ -277,7 +279,7 @@ int i915_gem_gtt_insert(struct i915_address_space *vm, return -ENOSPC;
/* Randomly selected placement is pinned, do a search */ - err = i915_gem_evict_something(vm, size, alignment, color, + err = i915_gem_evict_something(vm, ww, size, alignment, color, start, end, flags); if (err) return err; diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.h b/drivers/gpu/drm/i915/i915_gem_gtt.h index c9b0ee5e1d23..e4938aba3fe9 100644 --- a/drivers/gpu/drm/i915/i915_gem_gtt.h +++ b/drivers/gpu/drm/i915/i915_gem_gtt.h @@ -16,6 +16,7 @@
struct drm_i915_gem_object; struct i915_address_space; +struct i915_gem_ww_ctx;
int __must_check i915_gem_gtt_prepare_pages(struct drm_i915_gem_object *obj, struct sg_table *pages); @@ -23,11 +24,13 @@ void i915_gem_gtt_finish_pages(struct drm_i915_gem_object *obj, struct sg_table *pages);
int i915_gem_gtt_reserve(struct i915_address_space *vm, + struct i915_gem_ww_ctx *ww, struct drm_mm_node *node, u64 size, u64 offset, unsigned long color, unsigned int flags);
int i915_gem_gtt_insert(struct i915_address_space *vm, + struct i915_gem_ww_ctx *ww, struct drm_mm_node *node, u64 size, u64 alignment, unsigned long color, u64 start, u64 end, unsigned int flags); diff --git a/drivers/gpu/drm/i915/i915_vgpu.c b/drivers/gpu/drm/i915/i915_vgpu.c index 31a105bc1792..c97323973f9b 100644 --- a/drivers/gpu/drm/i915/i915_vgpu.c +++ b/drivers/gpu/drm/i915/i915_vgpu.c @@ -197,7 +197,7 @@ static int vgt_balloon_space(struct i915_ggtt *ggtt, drm_info(&dev_priv->drm, "balloon space: range [ 0x%lx - 0x%lx ] %lu KiB.\n", start, end, size / 1024); - ret = i915_gem_gtt_reserve(&ggtt->vm, node, + ret = i915_gem_gtt_reserve(&ggtt->vm, NULL, node, size, start, I915_COLOR_UNEVICTABLE, 0); if (!ret) diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c index 2877dcd62acb..3275b2364785 100644 --- a/drivers/gpu/drm/i915/i915_vma.c +++ b/drivers/gpu/drm/i915/i915_vma.c @@ -623,7 +623,8 @@ bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long color) * 0 on success, negative error code otherwise. */ static int -i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags) +i915_vma_insert(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, + u64 size, u64 alignment, u64 flags) { unsigned long color; u64 start, end; @@ -675,7 +676,7 @@ i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags) range_overflows(offset, size, end)) return -EINVAL;
- ret = i915_gem_gtt_reserve(vma->vm, &vma->node, + ret = i915_gem_gtt_reserve(vma->vm, ww, &vma->node, size, offset, color, flags); if (ret) @@ -714,7 +715,7 @@ i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags) size = round_up(size, I915_GTT_PAGE_SIZE_2M); }
- ret = i915_gem_gtt_insert(vma->vm, &vma->node, + ret = i915_gem_gtt_insert(vma->vm, ww, &vma->node, size, alignment, color, start, end, flags); if (ret) @@ -1293,7 +1294,7 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, goto err_unlock;
if (!(bound & I915_VMA_BIND_MASK)) { - err = i915_vma_insert(vma, size, alignment, flags); + err = i915_vma_insert(vma, ww, size, alignment, flags); if (err) goto err_active;
@@ -1369,7 +1370,9 @@ static int __i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, /* Unlike i915_vma_pin, we don't take no for an answer! */ flush_idle_contexts(vm->gt); if (mutex_lock_interruptible(&vm->mutex) == 0) { - i915_gem_evict_vm(vm); + + /* We pass NULL ww here, as we don't want to unbind locked objects */ + i915_gem_evict_vm(vm, NULL); mutex_unlock(&vm->mutex); } } while (1); diff --git a/drivers/gpu/drm/i915/selftests/i915_gem_evict.c b/drivers/gpu/drm/i915/selftests/i915_gem_evict.c index f99bb0113726..4591be1a8803 100644 --- a/drivers/gpu/drm/i915/selftests/i915_gem_evict.c +++ b/drivers/gpu/drm/i915/selftests/i915_gem_evict.c @@ -117,7 +117,7 @@ static int igt_evict_something(void *arg)
/* Everything is pinned, nothing should happen */ mutex_lock(&ggtt->vm.mutex); - err = i915_gem_evict_something(&ggtt->vm, + err = i915_gem_evict_something(&ggtt->vm, NULL, I915_GTT_PAGE_SIZE, 0, 0, 0, U64_MAX, 0); @@ -132,11 +132,12 @@ static int igt_evict_something(void *arg)
/* Everything is unpinned, we should be able to evict something */ mutex_lock(&ggtt->vm.mutex); - err = i915_gem_evict_something(&ggtt->vm, + err = i915_gem_evict_something(&ggtt->vm, NULL, I915_GTT_PAGE_SIZE, 0, 0, 0, U64_MAX, 0); mutex_unlock(&ggtt->vm.mutex); + if (err) { pr_err("i915_gem_evict_something failed on a full GGTT with err=%d\n", err); @@ -204,7 +205,7 @@ static int igt_evict_for_vma(void *arg)
/* Everything is pinned, nothing should happen */ mutex_lock(&ggtt->vm.mutex); - err = i915_gem_evict_for_node(&ggtt->vm, &target, 0); + err = i915_gem_evict_for_node(&ggtt->vm, NULL, &target, 0); mutex_unlock(&ggtt->vm.mutex); if (err != -ENOSPC) { pr_err("i915_gem_evict_for_node on a full GGTT returned err=%d\n", @@ -216,7 +217,7 @@ static int igt_evict_for_vma(void *arg)
/* Everything is unpinned, we should be able to evict the node */ mutex_lock(&ggtt->vm.mutex); - err = i915_gem_evict_for_node(&ggtt->vm, &target, 0); + err = i915_gem_evict_for_node(&ggtt->vm, NULL, &target, 0); mutex_unlock(&ggtt->vm.mutex); if (err) { pr_err("i915_gem_evict_for_node returned err=%d\n", @@ -297,7 +298,7 @@ static int igt_evict_for_cache_color(void *arg)
/* Remove just the second vma */ mutex_lock(&ggtt->vm.mutex); - err = i915_gem_evict_for_node(&ggtt->vm, &target, 0); + err = i915_gem_evict_for_node(&ggtt->vm, NULL, &target, 0); mutex_unlock(&ggtt->vm.mutex); if (err) { pr_err("[0]i915_gem_evict_for_node returned err=%d\n", err); @@ -310,7 +311,7 @@ static int igt_evict_for_cache_color(void *arg) target.color = I915_CACHE_L3_LLC;
mutex_lock(&ggtt->vm.mutex); - err = i915_gem_evict_for_node(&ggtt->vm, &target, 0); + err = i915_gem_evict_for_node(&ggtt->vm, NULL, &target, 0); mutex_unlock(&ggtt->vm.mutex); if (!err) { pr_err("[1]i915_gem_evict_for_node returned err=%d\n", err); @@ -331,6 +332,7 @@ static int igt_evict_vm(void *arg) { struct intel_gt *gt = arg; struct i915_ggtt *ggtt = gt->ggtt; + struct i915_gem_ww_ctx ww; LIST_HEAD(objects); int err;
@@ -342,7 +344,7 @@ static int igt_evict_vm(void *arg)
/* Everything is pinned, nothing should happen */ mutex_lock(&ggtt->vm.mutex); - err = i915_gem_evict_vm(&ggtt->vm); + err = i915_gem_evict_vm(&ggtt->vm, NULL); mutex_unlock(&ggtt->vm.mutex); if (err) { pr_err("i915_gem_evict_vm on a full GGTT returned err=%d]\n", @@ -352,9 +354,14 @@ static int igt_evict_vm(void *arg)
unpin_ggtt(ggtt);
+ i915_gem_ww_ctx_init(&ww, false); mutex_lock(&ggtt->vm.mutex); - err = i915_gem_evict_vm(&ggtt->vm); + err = i915_gem_evict_vm(&ggtt->vm, &ww); mutex_unlock(&ggtt->vm.mutex); + + /* no -EDEADLK handling; can't happen with vm.mutex in place */ + i915_gem_ww_ctx_fini(&ww); + if (err) { pr_err("i915_gem_evict_vm on a full GGTT returned err=%d]\n", err); @@ -402,7 +409,7 @@ static int igt_evict_contexts(void *arg) /* Reserve a block so that we know we have enough to fit a few rq */ memset(&hole, 0, sizeof(hole)); mutex_lock(&ggtt->vm.mutex); - err = i915_gem_gtt_insert(&ggtt->vm, &hole, + err = i915_gem_gtt_insert(&ggtt->vm, NULL, &hole, PRETEND_GGTT_SIZE, 0, I915_COLOR_UNEVICTABLE, 0, ggtt->vm.total, PIN_NOEVICT); @@ -422,7 +429,7 @@ static int igt_evict_contexts(void *arg) goto out_locked; }
- if (i915_gem_gtt_insert(&ggtt->vm, &r->node, + if (i915_gem_gtt_insert(&ggtt->vm, NULL, &r->node, 1ul << 20, 0, I915_COLOR_UNEVICTABLE, 0, ggtt->vm.total, PIN_NOEVICT)) { diff --git a/drivers/gpu/drm/i915/selftests/i915_gem_gtt.c b/drivers/gpu/drm/i915/selftests/i915_gem_gtt.c index 4574fb51b656..6b1db83119b3 100644 --- a/drivers/gpu/drm/i915/selftests/i915_gem_gtt.c +++ b/drivers/gpu/drm/i915/selftests/i915_gem_gtt.c @@ -1378,7 +1378,7 @@ static int igt_gtt_reserve(void *arg) }
mutex_lock(&ggtt->vm.mutex); - err = i915_gem_gtt_reserve(&ggtt->vm, &vma->node, + err = i915_gem_gtt_reserve(&ggtt->vm, NULL, &vma->node, obj->base.size, total, obj->cache_level, @@ -1430,7 +1430,7 @@ static int igt_gtt_reserve(void *arg) }
mutex_lock(&ggtt->vm.mutex); - err = i915_gem_gtt_reserve(&ggtt->vm, &vma->node, + err = i915_gem_gtt_reserve(&ggtt->vm, NULL, &vma->node, obj->base.size, total, obj->cache_level, @@ -1477,7 +1477,7 @@ static int igt_gtt_reserve(void *arg) I915_GTT_MIN_ALIGNMENT);
mutex_lock(&ggtt->vm.mutex); - err = i915_gem_gtt_reserve(&ggtt->vm, &vma->node, + err = i915_gem_gtt_reserve(&ggtt->vm, NULL, &vma->node, obj->base.size, offset, obj->cache_level, @@ -1552,7 +1552,7 @@ static int igt_gtt_insert(void *arg) /* Check a couple of obviously invalid requests */ for (ii = invalid_insert; ii->size; ii++) { mutex_lock(&ggtt->vm.mutex); - err = i915_gem_gtt_insert(&ggtt->vm, &tmp, + err = i915_gem_gtt_insert(&ggtt->vm, NULL, &tmp, ii->size, ii->alignment, I915_COLOR_UNEVICTABLE, ii->start, ii->end, @@ -1594,7 +1594,7 @@ static int igt_gtt_insert(void *arg) }
mutex_lock(&ggtt->vm.mutex); - err = i915_gem_gtt_insert(&ggtt->vm, &vma->node, + err = i915_gem_gtt_insert(&ggtt->vm, NULL, &vma->node, obj->base.size, 0, obj->cache_level, 0, ggtt->vm.total, 0); @@ -1654,7 +1654,7 @@ static int igt_gtt_insert(void *arg) }
mutex_lock(&ggtt->vm.mutex); - err = i915_gem_gtt_insert(&ggtt->vm, &vma->node, + err = i915_gem_gtt_insert(&ggtt->vm, NULL, &vma->node, obj->base.size, 0, obj->cache_level, 0, ggtt->vm.total, 0); @@ -1703,7 +1703,7 @@ static int igt_gtt_insert(void *arg) }
mutex_lock(&ggtt->vm.mutex); - err = i915_gem_gtt_insert(&ggtt->vm, &vma->node, + err = i915_gem_gtt_insert(&ggtt->vm, NULL, &vma->node, obj->base.size, 0, obj->cache_level, 0, ggtt->vm.total, 0);
On Thu, 21 Oct 2021 at 11:37, Maarten Lankhorst maarten.lankhorst@linux.intel.com wrote:
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
Needs a proper commit message.
Op 21-10-2021 om 20:03 schreef Matthew Auld:
On Thu, 21 Oct 2021 at 11:37, Maarten Lankhorst maarten.lankhorst@linux.intel.com wrote:
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
Needs a proper commit message.
What about this?
drm/i915: Pass trylock context to callers
We are moving away from short term pinning, and need a way to evict objects locked by the current context. Pass the ww context to all eviction functions, so that they can evict objects that are already locked by the current ww context.
On top of that, this slightly improves ww handling because the locked objects are marked as locked by the correct ww.
Now that we require locking to evict, multiple vmas from the same object might not be evicted. This is expected and required, because execbuf will move to short-term pinning by using the lock only. This will cause these tests to fail, because they create a ton of vma's for the same object.
Unbind manually to prevent spurious -ENOSPC in those mock tests.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com --- drivers/gpu/drm/i915/selftests/i915_vma.c | 17 ++++++++++++++++- 1 file changed, 16 insertions(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/i915/selftests/i915_vma.c b/drivers/gpu/drm/i915/selftests/i915_vma.c index 1f10fe36619b..5c5809dfe9b2 100644 --- a/drivers/gpu/drm/i915/selftests/i915_vma.c +++ b/drivers/gpu/drm/i915/selftests/i915_vma.c @@ -691,7 +691,11 @@ static int igt_vma_rotate_remap(void *arg) }
i915_vma_unpin(vma); - + err = i915_vma_unbind(vma); + if (err) { + pr_err("Unbinding returned %i\n", err); + goto out_object; + } cond_resched(); } } @@ -848,6 +852,11 @@ static int igt_vma_partial(void *arg)
i915_vma_unpin(vma); nvma++; + err = i915_vma_unbind(vma); + if (err) { + pr_err("Unbinding returned %i\n", err); + goto out_object; + }
cond_resched(); } @@ -882,6 +891,12 @@ static int igt_vma_partial(void *arg)
i915_vma_unpin(vma);
+ err = i915_vma_unbind(vma); + if (err) { + pr_err("Unbinding returned %i\n", err); + goto out_object; + } + count = 0; list_for_each_entry(vma, &obj->vma.list, obj_link) count++;
Be thorough..
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com --- drivers/gpu/drm/i915/i915_drv.h | 1 + 1 file changed, 1 insertion(+)
diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h index 22c891720c6d..7c5ed5957fe2 100644 --- a/drivers/gpu/drm/i915/i915_drv.h +++ b/drivers/gpu/drm/i915/i915_drv.h @@ -1819,6 +1819,7 @@ static inline void i915_gem_drain_freed_objects(struct drm_i915_private *i915) */ while (atomic_read(&i915->mm.free_count)) { flush_work(&i915->mm.free_work); + flush_delayed_work(&i915->bdev.wq); rcu_barrier(); } }
On Thu, 21 Oct 2021 at 11:37, Maarten Lankhorst maarten.lankhorst@linux.intel.com wrote:
Be thorough..
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
Is this strictly needed for something? Needs a proper commit message anyway.
drivers/gpu/drm/i915/i915_drv.h | 1 + 1 file changed, 1 insertion(+)
diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h index 22c891720c6d..7c5ed5957fe2 100644 --- a/drivers/gpu/drm/i915/i915_drv.h +++ b/drivers/gpu/drm/i915/i915_drv.h @@ -1819,6 +1819,7 @@ static inline void i915_gem_drain_freed_objects(struct drm_i915_private *i915) */ while (atomic_read(&i915->mm.free_count)) { flush_work(&i915->mm.free_work);
flush_delayed_work(&i915->bdev.wq); rcu_barrier(); }}
2.33.0
i915_gem_execbuf will call i915_gem_evict_vm() after failing to pin all objects in the first round. We are about to remove those short-term pins, but even without those the objects are still locked. Add a special case to allow i915_gem_evict_vm to evict locked objects as well.
This might also allow multiple objects sharing the same resv to be evicted.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com --- drivers/gpu/drm/i915/i915_gem_evict.c | 23 ++++++++++++++++++++++- 1 file changed, 22 insertions(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/i915/i915_gem_evict.c b/drivers/gpu/drm/i915/i915_gem_evict.c index 24f5e3345e43..f502a617b35c 100644 --- a/drivers/gpu/drm/i915/i915_gem_evict.c +++ b/drivers/gpu/drm/i915/i915_gem_evict.c @@ -410,21 +410,42 @@ int i915_gem_evict_vm(struct i915_address_space *vm, struct i915_gem_ww_ctx *ww) do { struct i915_vma *vma, *vn; LIST_HEAD(eviction_list); + LIST_HEAD(locked_eviction_list);
list_for_each_entry(vma, &vm->bound_list, vm_link) { if (i915_vma_is_pinned(vma)) continue;
+ /* + * If we already own the lock, trylock fails. In case the resv + * is shared among multiple objects, we still need the object ref. + */ + if (ww && (dma_resv_locking_ctx(vma->obj->base.resv) == &ww->ctx)) { + __i915_vma_pin(vma); + list_add(&vma->evict_link, &locked_eviction_list); + continue; + } + if (!i915_gem_object_trylock(vma->obj, ww)) continue;
__i915_vma_pin(vma); list_add(&vma->evict_link, &eviction_list); } - if (list_empty(&eviction_list)) + if (list_empty(&eviction_list) && list_empty(&locked_eviction_list)) break;
ret = 0; + /* Unbind locked objects first, before unlocking the eviction_list */ + list_for_each_entry_safe(vma, vn, &locked_eviction_list, evict_link) { + __i915_vma_unpin(vma); + + if (ret == 0) + ret = __i915_vma_unbind(vma); + if (ret != -EINTR) /* "Get me out of here!" */ + ret = 0; + } + list_for_each_entry_safe(vma, vn, &eviction_list, evict_link) { __i915_vma_unpin(vma); if (ret == 0)
Now that we cannot unbind kill the currently locked object directly because we're removing short term pinning, we may have to unbind the object from gtt manually, using a i915_gem_evict_vm() call.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com --- drivers/gpu/drm/i915/gem/i915_gem_mman.c | 18 ++++++++++++++++-- 1 file changed, 16 insertions(+), 2 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_mman.c b/drivers/gpu/drm/i915/gem/i915_gem_mman.c index 65fc6ff5f59d..6d557bb9926f 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_mman.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_mman.c @@ -357,8 +357,22 @@ static vm_fault_t vm_fault_gtt(struct vm_fault *vmf) vma = i915_gem_object_ggtt_pin_ww(obj, &ww, &view, 0, 0, flags); }
- /* The entire mappable GGTT is pinned? Unexpected! */ - GEM_BUG_ON(vma == ERR_PTR(-ENOSPC)); + /* + * The entire mappable GGTT is pinned? Unexpected! + * Try to evict the object we locked too, as normally we skip it + * due to lack of short term pinning inside execbuf. + */ + if (vma == ERR_PTR(-ENOSPC)) { + ret = mutex_lock_interruptible(&ggtt->vm.mutex); + if (!ret) { + ret = i915_gem_evict_vm(&ggtt->vm, &ww); + mutex_unlock(&ggtt->vm.mutex); + } + if (ret) + goto err_reset; + vma = i915_gem_object_ggtt_pin_ww(obj, &ww, &view, 0, 0, flags); + } + GEM_WARN_ON(vma == ERR_PTR(-ENOSPC)); } if (IS_ERR(vma)) { ret = PTR_ERR(vma);
We want to remove more members of i915_vma, which requires the locking to be held more often.
Start requiring gem object lock for i915_vma_unbind, as it's one of the callers that may unpin pages.
Some special care is needed when evicting, because the last reference to the object may be held by the VMA, so after __i915_vma_unbind, vma may be garbage, and we need to cache vma->obj before unlocking.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com --- drivers/gpu/drm/i915/display/intel_fb_pin.c | 2 +- .../gpu/drm/i915/gem/selftests/huge_pages.c | 2 +- .../i915/gem/selftests/i915_gem_client_blt.c | 2 +- .../drm/i915/gem/selftests/i915_gem_mman.c | 6 +++ drivers/gpu/drm/i915/gt/intel_ggtt.c | 45 ++++++++++++++++--- drivers/gpu/drm/i915/i915_gem.c | 2 + drivers/gpu/drm/i915/i915_vma.c | 27 ++++++++++- drivers/gpu/drm/i915/i915_vma.h | 1 + drivers/gpu/drm/i915/selftests/i915_gem_gtt.c | 22 ++++----- drivers/gpu/drm/i915/selftests/i915_vma.c | 8 ++-- 10 files changed, 91 insertions(+), 26 deletions(-)
diff --git a/drivers/gpu/drm/i915/display/intel_fb_pin.c b/drivers/gpu/drm/i915/display/intel_fb_pin.c index 3f77f3013584..1c0ff2ef3937 100644 --- a/drivers/gpu/drm/i915/display/intel_fb_pin.c +++ b/drivers/gpu/drm/i915/display/intel_fb_pin.c @@ -47,7 +47,7 @@ intel_pin_fb_obj_dpt(struct drm_framebuffer *fb, goto err;
if (i915_vma_misplaced(vma, 0, alignment, 0)) { - ret = i915_vma_unbind(vma); + ret = i915_vma_unbind_unlocked(vma); if (ret) { vma = ERR_PTR(ret); goto err; diff --git a/drivers/gpu/drm/i915/gem/selftests/huge_pages.c b/drivers/gpu/drm/i915/gem/selftests/huge_pages.c index 493509f90b35..97473ac7c7f7 100644 --- a/drivers/gpu/drm/i915/gem/selftests/huge_pages.c +++ b/drivers/gpu/drm/i915/gem/selftests/huge_pages.c @@ -646,7 +646,7 @@ static int igt_mock_ppgtt_misaligned_dma(void *arg) * pages. */ for (offset = 4096; offset < page_size; offset += 4096) { - err = i915_vma_unbind(vma); + err = i915_vma_unbind_unlocked(vma); if (err) goto out_unpin;
diff --git a/drivers/gpu/drm/i915/gem/selftests/i915_gem_client_blt.c b/drivers/gpu/drm/i915/gem/selftests/i915_gem_client_blt.c index 8402ed925a69..8fb5be799b3c 100644 --- a/drivers/gpu/drm/i915/gem/selftests/i915_gem_client_blt.c +++ b/drivers/gpu/drm/i915/gem/selftests/i915_gem_client_blt.c @@ -318,7 +318,7 @@ static int pin_buffer(struct i915_vma *vma, u64 addr) int err;
if (drm_mm_node_allocated(&vma->node) && vma->node.start != addr) { - err = i915_vma_unbind(vma); + err = i915_vma_unbind_unlocked(vma); if (err) return err; } diff --git a/drivers/gpu/drm/i915/gem/selftests/i915_gem_mman.c b/drivers/gpu/drm/i915/gem/selftests/i915_gem_mman.c index 6d30cdfa80f3..e69e8861352d 100644 --- a/drivers/gpu/drm/i915/gem/selftests/i915_gem_mman.c +++ b/drivers/gpu/drm/i915/gem/selftests/i915_gem_mman.c @@ -165,7 +165,9 @@ static int check_partial_mapping(struct drm_i915_gem_object *obj, kunmap(p);
out: + i915_gem_object_lock(obj, NULL); __i915_vma_put(vma); + i915_gem_object_unlock(obj); return err; }
@@ -259,7 +261,9 @@ static int check_partial_mappings(struct drm_i915_gem_object *obj, if (err) return err;
+ i915_gem_object_lock(obj, NULL); __i915_vma_put(vma); + i915_gem_object_unlock(obj);
if (igt_timeout(end_time, "%s: timed out after tiling=%d stride=%d\n", @@ -1349,7 +1353,9 @@ static int __igt_mmap_revoke(struct drm_i915_private *i915, * for other objects. Ergo we have to revoke the previous mmap PTE * access as it no longer points to the same object. */ + i915_gem_object_lock(obj, NULL); err = i915_gem_object_unbind(obj, I915_GEM_OBJECT_UNBIND_ACTIVE); + i915_gem_object_unlock(obj); if (err) { pr_err("Failed to unbind object!\n"); goto out_unmap; diff --git a/drivers/gpu/drm/i915/gt/intel_ggtt.c b/drivers/gpu/drm/i915/gt/intel_ggtt.c index df29fd992b45..44bc30458cac 100644 --- a/drivers/gpu/drm/i915/gt/intel_ggtt.c +++ b/drivers/gpu/drm/i915/gt/intel_ggtt.c @@ -118,22 +118,45 @@ void i915_ggtt_suspend(struct i915_ggtt *ggtt) struct i915_vma *vma, *vn; int open;
+retry: + i915_gem_drain_freed_objects(ggtt->vm.i915); + mutex_lock(&ggtt->vm.mutex);
/* Skip rewriting PTE on VMA unbind. */ open = atomic_xchg(&ggtt->vm.open, 0);
list_for_each_entry_safe(vma, vn, &ggtt->vm.bound_list, vm_link) { + struct drm_i915_gem_object *obj = vma->obj; + GEM_BUG_ON(!drm_mm_node_allocated(&vma->node)); - i915_vma_wait_for_bind(vma);
- if (i915_vma_is_pinned(vma)) + if (i915_vma_is_pinned(vma) || !i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND)) continue;
- if (!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND)) { - __i915_vma_evict(vma); - drm_mm_remove_node(&vma->node); + /* unlikely to race when GPU is idle, so no worry about slowpath.. */ + if (!i915_gem_object_trylock(obj, NULL)) { + atomic_set(&ggtt->vm.open, open); + + i915_gem_object_get(obj); + mutex_unlock(&ggtt->vm.mutex); + + i915_gem_object_lock(obj, NULL); + open = i915_vma_unbind(vma); + i915_gem_object_unlock(obj); + + GEM_WARN_ON(open); + + i915_gem_object_put(obj); + goto retry; } + + i915_vma_wait_for_bind(vma); + + __i915_vma_evict(vma); + drm_mm_remove_node(&vma->node); + + i915_gem_object_unlock(obj); }
ggtt->vm.clear_range(&ggtt->vm, 0, ggtt->vm.total); @@ -725,11 +748,21 @@ static void ggtt_cleanup_hw(struct i915_ggtt *ggtt) atomic_set(&ggtt->vm.open, 0);
flush_workqueue(ggtt->vm.i915->wq); + i915_gem_drain_freed_objects(ggtt->vm.i915);
mutex_lock(&ggtt->vm.mutex);
- list_for_each_entry_safe(vma, vn, &ggtt->vm.bound_list, vm_link) + list_for_each_entry_safe(vma, vn, &ggtt->vm.bound_list, vm_link) { + struct drm_i915_gem_object *obj = vma->obj; + bool trylock; + + trylock = i915_gem_object_trylock(obj, NULL); + WARN_ON(!trylock); + WARN_ON(__i915_vma_unbind(vma)); + if (trylock) + i915_gem_object_unlock(obj); + }
if (drm_mm_node_allocated(&ggtt->error_capture)) drm_mm_remove_node(&ggtt->error_capture); diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c index 6aa9e465b48e..d45abf05987d 100644 --- a/drivers/gpu/drm/i915/i915_gem.c +++ b/drivers/gpu/drm/i915/i915_gem.c @@ -118,6 +118,8 @@ int i915_gem_object_unbind(struct drm_i915_gem_object *obj, struct i915_vma *vma; int ret;
+ assert_object_held(obj); + if (list_empty(&obj->vma.list)) return 0;
diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c index 3275b2364785..982bb8b739d3 100644 --- a/drivers/gpu/drm/i915/i915_vma.c +++ b/drivers/gpu/drm/i915/i915_vma.c @@ -1508,8 +1508,16 @@ void i915_vma_parked(struct intel_gt *gt) struct drm_i915_gem_object *obj = vma->obj; struct i915_address_space *vm = vma->vm;
- INIT_LIST_HEAD(&vma->closed_link); - __i915_vma_put(vma); + if (i915_gem_object_trylock(obj, NULL)) { + INIT_LIST_HEAD(&vma->closed_link); + __i915_vma_put(vma); + i915_gem_object_unlock(obj); + } else { + /* back you go.. */ + spin_lock_irq(>->closed_lock); + list_add(&vma->closed_link, >->closed_vma); + spin_unlock_irq(>->closed_lock); + }
i915_gem_object_put(obj); i915_vm_close(vm); @@ -1625,6 +1633,7 @@ int _i915_vma_move_to_active(struct i915_vma *vma, void __i915_vma_evict(struct i915_vma *vma) { GEM_BUG_ON(i915_vma_is_pinned(vma)); + assert_object_held_shared(vma->obj);
if (i915_vma_is_map_and_fenceable(vma)) { /* Force a pagefault for domain tracking on next user access */ @@ -1670,6 +1679,7 @@ int __i915_vma_unbind(struct i915_vma *vma) int ret;
lockdep_assert_held(&vma->vm->mutex); + assert_object_held_shared(vma->obj);
if (!drm_mm_node_allocated(&vma->node)) return 0; @@ -1701,6 +1711,8 @@ int i915_vma_unbind(struct i915_vma *vma) intel_wakeref_t wakeref = 0; int err;
+ assert_object_held_shared(vma->obj); + /* Optimistic wait before taking the mutex */ err = i915_vma_sync(vma); if (err) @@ -1731,6 +1743,17 @@ int i915_vma_unbind(struct i915_vma *vma) return err; }
+int i915_vma_unbind_unlocked(struct i915_vma *vma) +{ + int err; + + i915_gem_object_lock(vma->obj, NULL); + err = i915_vma_unbind(vma); + i915_gem_object_unlock(vma->obj); + + return err; +} + struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma) { i915_gem_object_make_unshrinkable(vma->obj); diff --git a/drivers/gpu/drm/i915/i915_vma.h b/drivers/gpu/drm/i915/i915_vma.h index 32719431b3df..da69ecb1b860 100644 --- a/drivers/gpu/drm/i915/i915_vma.h +++ b/drivers/gpu/drm/i915/i915_vma.h @@ -214,6 +214,7 @@ void i915_vma_revoke_mmap(struct i915_vma *vma); void __i915_vma_evict(struct i915_vma *vma); int __i915_vma_unbind(struct i915_vma *vma); int __must_check i915_vma_unbind(struct i915_vma *vma); +int __must_check i915_vma_unbind_unlocked(struct i915_vma *vma); void i915_vma_unlink_ctx(struct i915_vma *vma); void i915_vma_close(struct i915_vma *vma); void i915_vma_reopen(struct i915_vma *vma); diff --git a/drivers/gpu/drm/i915/selftests/i915_gem_gtt.c b/drivers/gpu/drm/i915/selftests/i915_gem_gtt.c index 6b1db83119b3..de16897d3d9a 100644 --- a/drivers/gpu/drm/i915/selftests/i915_gem_gtt.c +++ b/drivers/gpu/drm/i915/selftests/i915_gem_gtt.c @@ -385,7 +385,7 @@ static void close_object_list(struct list_head *objects,
vma = i915_vma_instance(obj, vm, NULL); if (!IS_ERR(vma)) - ignored = i915_vma_unbind(vma); + ignored = i915_vma_unbind_unlocked(vma);
list_del(&obj->st_link); i915_gem_object_put(obj); @@ -496,7 +496,7 @@ static int fill_hole(struct i915_address_space *vm, goto err; }
- err = i915_vma_unbind(vma); + err = i915_vma_unbind_unlocked(vma); if (err) { pr_err("%s(%s) (forward) unbind of vma.node=%llx + %llx failed with err=%d\n", __func__, p->name, vma->node.start, vma->node.size, @@ -569,7 +569,7 @@ static int fill_hole(struct i915_address_space *vm, goto err; }
- err = i915_vma_unbind(vma); + err = i915_vma_unbind_unlocked(vma); if (err) { pr_err("%s(%s) (backward) unbind of vma.node=%llx + %llx failed with err=%d\n", __func__, p->name, vma->node.start, vma->node.size, @@ -655,7 +655,7 @@ static int walk_hole(struct i915_address_space *vm, goto err_put; }
- err = i915_vma_unbind(vma); + err = i915_vma_unbind_unlocked(vma); if (err) { pr_err("%s unbind failed at %llx + %llx with err=%d\n", __func__, addr, vma->size, err); @@ -732,13 +732,13 @@ static int pot_hole(struct i915_address_space *vm, pr_err("%s incorrect at %llx + %llx\n", __func__, addr, vma->size); i915_vma_unpin(vma); - err = i915_vma_unbind(vma); + err = i915_vma_unbind_unlocked(vma); err = -EINVAL; goto err_obj; }
i915_vma_unpin(vma); - err = i915_vma_unbind(vma); + err = i915_vma_unbind_unlocked(vma); GEM_BUG_ON(err); }
@@ -832,13 +832,13 @@ static int drunk_hole(struct i915_address_space *vm, pr_err("%s incorrect at %llx + %llx\n", __func__, addr, BIT_ULL(size)); i915_vma_unpin(vma); - err = i915_vma_unbind(vma); + err = i915_vma_unbind_unlocked(vma); err = -EINVAL; goto err_obj; }
i915_vma_unpin(vma); - err = i915_vma_unbind(vma); + err = i915_vma_unbind_unlocked(vma); GEM_BUG_ON(err);
if (igt_timeout(end_time, @@ -906,7 +906,7 @@ static int __shrink_hole(struct i915_address_space *vm, pr_err("%s incorrect at %llx + %llx\n", __func__, addr, size); i915_vma_unpin(vma); - err = i915_vma_unbind(vma); + err = i915_vma_unbind_unlocked(vma); err = -EINVAL; break; } @@ -1465,7 +1465,7 @@ static int igt_gtt_reserve(void *arg) goto out; }
- err = i915_vma_unbind(vma); + err = i915_vma_unbind_unlocked(vma); if (err) { pr_err("i915_vma_unbind failed with err=%d!\n", err); goto out; @@ -1647,7 +1647,7 @@ static int igt_gtt_insert(void *arg) GEM_BUG_ON(!drm_mm_node_allocated(&vma->node)); offset = vma->node.start;
- err = i915_vma_unbind(vma); + err = i915_vma_unbind_unlocked(vma); if (err) { pr_err("i915_vma_unbind failed with err=%d!\n", err); goto out; diff --git a/drivers/gpu/drm/i915/selftests/i915_vma.c b/drivers/gpu/drm/i915/selftests/i915_vma.c index 5c5809dfe9b2..2c73f7448df7 100644 --- a/drivers/gpu/drm/i915/selftests/i915_vma.c +++ b/drivers/gpu/drm/i915/selftests/i915_vma.c @@ -340,7 +340,7 @@ static int igt_vma_pin1(void *arg)
if (!err) { i915_vma_unpin(vma); - err = i915_vma_unbind(vma); + err = i915_vma_unbind_unlocked(vma); if (err) { pr_err("Failed to unbind single page from GGTT, err=%d\n", err); goto out; @@ -691,7 +691,7 @@ static int igt_vma_rotate_remap(void *arg) }
i915_vma_unpin(vma); - err = i915_vma_unbind(vma); + err = i915_vma_unbind_unlocked(vma); if (err) { pr_err("Unbinding returned %i\n", err); goto out_object; @@ -852,7 +852,7 @@ static int igt_vma_partial(void *arg)
i915_vma_unpin(vma); nvma++; - err = i915_vma_unbind(vma); + err = i915_vma_unbind_unlocked(vma); if (err) { pr_err("Unbinding returned %i\n", err); goto out_object; @@ -891,7 +891,7 @@ static int igt_vma_partial(void *arg)
i915_vma_unpin(vma);
- err = i915_vma_unbind(vma); + err = i915_vma_unbind_unlocked(vma); if (err) { pr_err("Unbinding returned %i\n", err); goto out_object;
TTM already requires this, and we require it for delayed destroy.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com --- drivers/gpu/drm/i915/gem/i915_gem_object.c | 5 +++++ 1 file changed, 5 insertions(+)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.c b/drivers/gpu/drm/i915/gem/i915_gem_object.c index 1e426a42a36c..d549f34829d0 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_object.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.c @@ -257,6 +257,8 @@ static void __i915_gem_object_free_mmaps(struct drm_i915_gem_object *obj) */ void __i915_gem_object_pages_fini(struct drm_i915_gem_object *obj) { + assert_object_held(obj); + if (!list_empty(&obj->vma.list)) { struct i915_vma *vma;
@@ -323,7 +325,10 @@ static void __i915_gem_free_objects(struct drm_i915_private *i915, obj->ops->delayed_free(obj); continue; } + + i915_gem_object_lock(obj, NULL); __i915_gem_object_pages_fini(obj); + i915_gem_object_unlock(obj); __i915_gem_free_object(obj);
/* But keep the pointer alive for RCU-protected lookups */
On Thu, 21 Oct 2021 at 11:37, Maarten Lankhorst maarten.lankhorst@linux.intel.com wrote:
TTM already requires this, and we require it for delayed destroy.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
Reviewed-by: Matthew Auld matthew.auld@intel.com
This duck tape workaround is no longer required, unbind and destroy are fixed to take the obj->resv mutex before destroying and obj->mm.lock has been removed, always requiring obj->resv as well.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com --- drivers/gpu/drm/i915/gem/i915_gem_object.c | 4 ++-- drivers/gpu/drm/i915/gem/i915_gem_object.h | 14 -------------- drivers/gpu/drm/i915/gem/i915_gem_pages.c | 12 ++++++------ drivers/gpu/drm/i915/gem/i915_gem_userptr.c | 2 +- drivers/gpu/drm/i915/i915_vma.c | 6 +++--- 5 files changed, 12 insertions(+), 26 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.c b/drivers/gpu/drm/i915/gem/i915_gem_object.c index d549f34829d0..a4fb76b7841d 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_object.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.c @@ -548,7 +548,7 @@ bool i915_gem_object_has_struct_page(const struct drm_i915_gem_object *obj) #ifdef CONFIG_LOCKDEP if (IS_DGFX(to_i915(obj->base.dev)) && i915_gem_object_evictable((void __force *)obj)) - assert_object_held_shared(obj); + assert_object_held(obj); #endif return obj->mem_flags & I915_BO_FLAG_STRUCT_PAGE; } @@ -567,7 +567,7 @@ bool i915_gem_object_has_iomem(const struct drm_i915_gem_object *obj) #ifdef CONFIG_LOCKDEP if (IS_DGFX(to_i915(obj->base.dev)) && i915_gem_object_evictable((void __force *)obj)) - assert_object_held_shared(obj); + assert_object_held(obj); #endif return obj->mem_flags & I915_BO_FLAG_IOMEM; } diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.h b/drivers/gpu/drm/i915/gem/i915_gem_object.h index accd8fad0ed7..ef4f2568d46a 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_object.h +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.h @@ -158,20 +158,6 @@ i915_gem_object_put(struct drm_i915_gem_object *obj)
#define assert_object_held(obj) dma_resv_assert_held((obj)->base.resv)
-/* - * If more than one potential simultaneous locker, assert held. - */ -static inline void assert_object_held_shared(const struct drm_i915_gem_object *obj) -{ - /* - * Note mm list lookup is protected by - * kref_get_unless_zero(). - */ - if (IS_ENABLED(CONFIG_LOCKDEP) && - kref_read(&obj->base.refcount) > 0) - assert_object_held(obj); -} - static inline int __i915_gem_object_lock(struct drm_i915_gem_object *obj, struct i915_gem_ww_ctx *ww, bool intr) diff --git a/drivers/gpu/drm/i915/gem/i915_gem_pages.c b/drivers/gpu/drm/i915/gem/i915_gem_pages.c index 8eb1c3a6fc9c..dbd226b0ea49 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_pages.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_pages.c @@ -19,7 +19,7 @@ void __i915_gem_object_set_pages(struct drm_i915_gem_object *obj, bool shrinkable; int i;
- assert_object_held_shared(obj); + assert_object_held(obj);
if (i915_gem_object_is_volatile(obj)) obj->mm.madv = I915_MADV_DONTNEED; @@ -94,7 +94,7 @@ int ____i915_gem_object_get_pages(struct drm_i915_gem_object *obj) struct drm_i915_private *i915 = to_i915(obj->base.dev); int err;
- assert_object_held_shared(obj); + assert_object_held(obj);
if (unlikely(obj->mm.madv != I915_MADV_WILLNEED)) { drm_dbg(&i915->drm, @@ -121,7 +121,7 @@ int __i915_gem_object_get_pages(struct drm_i915_gem_object *obj)
assert_object_held(obj);
- assert_object_held_shared(obj); + assert_object_held(obj);
if (unlikely(!i915_gem_object_has_pages(obj))) { GEM_BUG_ON(i915_gem_object_has_pinned_pages(obj)); @@ -168,7 +168,7 @@ void i915_gem_object_truncate(struct drm_i915_gem_object *obj) /* Try to discard unwanted pages */ void i915_gem_object_writeback(struct drm_i915_gem_object *obj) { - assert_object_held_shared(obj); + assert_object_held(obj); GEM_BUG_ON(i915_gem_object_has_pages(obj));
if (obj->ops->writeback) @@ -199,7 +199,7 @@ __i915_gem_object_unset_pages(struct drm_i915_gem_object *obj) { struct sg_table *pages;
- assert_object_held_shared(obj); + assert_object_held(obj);
pages = fetch_and_zero(&obj->mm.pages); if (IS_ERR_OR_NULL(pages)) @@ -229,7 +229,7 @@ int __i915_gem_object_put_pages(struct drm_i915_gem_object *obj) return -EBUSY;
/* May be called by shrinker from within get_pages() (on another bo) */ - assert_object_held_shared(obj); + assert_object_held(obj);
i915_gem_object_release_mmap_offset(obj);
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_userptr.c b/drivers/gpu/drm/i915/gem/i915_gem_userptr.c index 3173c9f9a040..a315c010f635 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_userptr.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_userptr.c @@ -109,7 +109,7 @@ static void i915_gem_object_userptr_drop_ref(struct drm_i915_gem_object *obj) { struct page **pvec = NULL;
- assert_object_held_shared(obj); + assert_object_held(obj);
if (!--obj->userptr.page_ref) { pvec = obj->userptr.pvec; diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c index 982bb8b739d3..8131dbf89048 100644 --- a/drivers/gpu/drm/i915/i915_vma.c +++ b/drivers/gpu/drm/i915/i915_vma.c @@ -1633,7 +1633,7 @@ int _i915_vma_move_to_active(struct i915_vma *vma, void __i915_vma_evict(struct i915_vma *vma) { GEM_BUG_ON(i915_vma_is_pinned(vma)); - assert_object_held_shared(vma->obj); + assert_object_held(vma->obj);
if (i915_vma_is_map_and_fenceable(vma)) { /* Force a pagefault for domain tracking on next user access */ @@ -1679,7 +1679,7 @@ int __i915_vma_unbind(struct i915_vma *vma) int ret;
lockdep_assert_held(&vma->vm->mutex); - assert_object_held_shared(vma->obj); + assert_object_held(vma->obj);
if (!drm_mm_node_allocated(&vma->node)) return 0; @@ -1711,7 +1711,7 @@ int i915_vma_unbind(struct i915_vma *vma) intel_wakeref_t wakeref = 0; int err;
- assert_object_held_shared(vma->obj); + assert_object_held(vma->obj);
/* Optimistic wait before taking the mutex */ err = i915_vma_sync(vma);
Now that we require the object lock for all ops, some code handling race conditions can be removed.
This is required to not take short-term pins inside execbuf.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com Acked-by: Niranjana Vishwanathapura niranjana.vishwanathapura@intel.com --- drivers/gpu/drm/i915/i915_vma.c | 40 +++++---------------------------- 1 file changed, 5 insertions(+), 35 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c index 8131dbf89048..65168db534f0 100644 --- a/drivers/gpu/drm/i915/i915_vma.c +++ b/drivers/gpu/drm/i915/i915_vma.c @@ -749,7 +749,6 @@ i915_vma_detach(struct i915_vma *vma) static bool try_qad_pin(struct i915_vma *vma, unsigned int flags) { unsigned int bound; - bool pinned = true;
bound = atomic_read(&vma->flags); do { @@ -759,34 +758,10 @@ static bool try_qad_pin(struct i915_vma *vma, unsigned int flags) if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR))) return false;
- if (!(bound & I915_VMA_PIN_MASK)) - goto unpinned; - GEM_BUG_ON(((bound + 1) & I915_VMA_PIN_MASK) == 0); } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
return true; - -unpinned: - /* - * If pin_count==0, but we are bound, check under the lock to avoid - * racing with a concurrent i915_vma_unbind(). - */ - mutex_lock(&vma->vm->mutex); - do { - if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR))) { - pinned = false; - break; - } - - if (unlikely(flags & ~bound)) { - pinned = false; - break; - } - } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1)); - mutex_unlock(&vma->vm->mutex); - - return pinned; }
@@ -1112,13 +1087,7 @@ __i915_vma_get_pages(struct i915_vma *vma) vma->ggtt_view.type, ret); }
- pages = xchg(&vma->pages, pages); - - /* did we race against a put_pages? */ - if (pages && pages != vma->obj->mm.pages) { - sg_free_table(vma->pages); - kfree(vma->pages); - } + vma->pages = pages;
return ret; } @@ -1152,13 +1121,14 @@ I915_SELFTEST_EXPORT int i915_vma_get_pages(struct i915_vma *vma) static void __vma_put_pages(struct i915_vma *vma, unsigned int count) { /* We allocate under vma_get_pages, so beware the shrinker */ - struct sg_table *pages = READ_ONCE(vma->pages); + struct sg_table *pages = vma->pages;
GEM_BUG_ON(atomic_read(&vma->pages_count) < count);
if (atomic_sub_return(count, &vma->pages_count) == 0) { - if (pages == cmpxchg(&vma->pages, pages, NULL) && - pages != vma->obj->mm.pages) { + vma->pages = NULL; + + if (pages != vma->obj->mm.pages) { sg_free_table(pages); kfree(pages); }
Add a flag PIN_VALIDATE, to indicate we don't need to pin and only protected by the object lock.
This removes the need to unpin, which is done by just releasing the lock.
eb_reserve is slightly reworked for readability, but the same steps are still done: - First pass pins with NONBLOCK. - Second pass unbinds all objects first, then pins. - Third pass is only called when not all objects are softpinned, and unbinds all objects, then calls i915_gem_evict_vm(), then pins.
When evicting the entire vm in eb_reserve() we do temporarily pin objects that are marked with EXEC_OBJECT_PINNED. This is because they are already at their destination, and i915_gem_evict_vm() would otherwise unbind them.
However, we reduce the visibility of those pins by limiting the pin to our call to i915_gem_evict_vm() only, and pin with vm->mutex held, instead of the entire duration of the execbuf.
Not sure the latter matters, one can hope.. In theory we could kill the pinning by adding an extra flag to the vma to temporarily prevent unbinding for gtt for i915_gem_evict_vm only, but I think that might be overkill. We're still holding the object lock, and we don't have blocking eviction yet. It's likely sufficient to simply enforce EXEC_OBJECT_PINNED for all objects on >= gen12.
Changes since v1: - Split out eb_reserve() into separate functions for readability. Changes since v2: - Make batch buffer mappable on platforms where only GGTT is available, to prevent moving the batch buffer during relocations. Changes since v3: - Preserve current behavior for batch buffer, instead be cautious when calling i915_gem_object_ggtt_pin_ww, and re-use the current batch vma if it's inside ggtt and map-and-fenceable.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com --- .../gpu/drm/i915/gem/i915_gem_execbuffer.c | 252 ++++++++++-------- drivers/gpu/drm/i915/i915_gem_gtt.h | 1 + drivers/gpu/drm/i915/i915_vma.c | 24 +- 3 files changed, 161 insertions(+), 116 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c index bbf2a10738f7..19f91143cfcf 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c @@ -439,7 +439,7 @@ eb_pin_vma(struct i915_execbuffer *eb, else pin_flags = entry->offset & PIN_OFFSET_MASK;
- pin_flags |= PIN_USER | PIN_NOEVICT | PIN_OFFSET_FIXED; + pin_flags |= PIN_USER | PIN_NOEVICT | PIN_OFFSET_FIXED | PIN_VALIDATE; if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_GTT)) pin_flags |= PIN_GLOBAL;
@@ -457,17 +457,15 @@ eb_pin_vma(struct i915_execbuffer *eb, entry->pad_to_size, entry->alignment, eb_pin_flags(entry, ev->flags) | - PIN_USER | PIN_NOEVICT); + PIN_USER | PIN_NOEVICT | PIN_VALIDATE); if (unlikely(err)) return err; }
if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_FENCE)) { err = i915_vma_pin_fence(vma); - if (unlikely(err)) { - i915_vma_unpin(vma); + if (unlikely(err)) return err; - }
if (vma->fence) ev->flags |= __EXEC_OBJECT_HAS_FENCE; @@ -483,13 +481,9 @@ eb_pin_vma(struct i915_execbuffer *eb, static inline void eb_unreserve_vma(struct eb_vma *ev) { - if (!(ev->flags & __EXEC_OBJECT_HAS_PIN)) - return; - if (unlikely(ev->flags & __EXEC_OBJECT_HAS_FENCE)) __i915_vma_unpin_fence(ev->vma);
- __i915_vma_unpin(ev->vma); ev->flags &= ~__EXEC_OBJECT_RESERVED; }
@@ -682,10 +676,8 @@ static int eb_reserve_vma(struct i915_execbuffer *eb,
if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_FENCE)) { err = i915_vma_pin_fence(vma); - if (unlikely(err)) { - i915_vma_unpin(vma); + if (unlikely(err)) return err; - }
if (vma->fence) ev->flags |= __EXEC_OBJECT_HAS_FENCE; @@ -697,85 +689,129 @@ static int eb_reserve_vma(struct i915_execbuffer *eb, return 0; }
-static int eb_reserve(struct i915_execbuffer *eb) +static int eb_evict_vm(struct i915_execbuffer *eb) +{ + const unsigned int count = eb->buffer_count; + unsigned int i; + int err; + + err = mutex_lock_interruptible(&eb->context->vm->mutex); + if (err) + return err; + + /* pin to protect against i915_gem_evict_vm evicting below */ + for (i = 0; i < count; i++) { + struct eb_vma *ev = &eb->vma[i]; + + if (ev->flags & __EXEC_OBJECT_HAS_PIN) + __i915_vma_pin(ev->vma); + } + + /* Too fragmented, unbind everything and retry */ + err = i915_gem_evict_vm(eb->context->vm, &eb->ww); + + /* unpin objects.. */ + for (i = 0; i < count; i++) { + struct eb_vma *ev = &eb->vma[i]; + + if (ev->flags & __EXEC_OBJECT_HAS_PIN) + i915_vma_unpin(ev->vma); + } + + mutex_unlock(&eb->context->vm->mutex); + + return err; +} + +static bool eb_unbind(struct i915_execbuffer *eb) { const unsigned int count = eb->buffer_count; - unsigned int pin_flags = PIN_USER | PIN_NONBLOCK; + unsigned int i; struct list_head last; + bool unpinned = false; + + /* Resort *all* the objects into priority order */ + INIT_LIST_HEAD(&eb->unbound); + INIT_LIST_HEAD(&last); + + for (i = 0; i < count; i++) { + struct eb_vma *ev = &eb->vma[i]; + unsigned int flags = ev->flags; + + if (flags & EXEC_OBJECT_PINNED && + flags & __EXEC_OBJECT_HAS_PIN) + continue; + + unpinned = true; + eb_unreserve_vma(ev); + + if (flags & EXEC_OBJECT_PINNED) + /* Pinned must have their slot */ + list_add(&ev->bind_link, &eb->unbound); + else if (flags & __EXEC_OBJECT_NEEDS_MAP) + /* Map require the lowest 256MiB (aperture) */ + list_add_tail(&ev->bind_link, &eb->unbound); + else if (!(flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS)) + /* Prioritise 4GiB region for restricted bo */ + list_add(&ev->bind_link, &last); + else + list_add_tail(&ev->bind_link, &last); + } + + list_splice_tail(&last, &eb->unbound); + return unpinned; +} + +static int eb_reserve(struct i915_execbuffer *eb) +{ struct eb_vma *ev; - unsigned int i, pass; + unsigned int pass; int err = 0; + bool unpinned;
/* * Attempt to pin all of the buffers into the GTT. - * This is done in 3 phases: + * This is done in 2 phases: * - * 1a. Unbind all objects that do not match the GTT constraints for - * the execbuffer (fenceable, mappable, alignment etc). - * 1b. Increment pin count for already bound objects. - * 2. Bind new objects. - * 3. Decrement pin count. + * 1. Unbind all objects that do not match the GTT constraints for + * the execbuffer (fenceable, mappable, alignment etc). + * 2. Bind new objects. * * This avoid unnecessary unbinding of later objects in order to make * room for the earlier objects *unless* we need to defragment. + * + * Defragmenting is skipped if all objects are pinned at a fixed location. */ - pass = 0; - do { - list_for_each_entry(ev, &eb->unbound, bind_link) { - err = eb_reserve_vma(eb, ev, pin_flags); - if (err) - break; - } - if (err != -ENOSPC) - return err; + for (pass = 0; pass <= 2; pass++) { + int pin_flags = PIN_USER | PIN_VALIDATE;
- /* Resort *all* the objects into priority order */ - INIT_LIST_HEAD(&eb->unbound); - INIT_LIST_HEAD(&last); - for (i = 0; i < count; i++) { - unsigned int flags; + if (pass == 0) + pin_flags |= PIN_NONBLOCK;
- ev = &eb->vma[i]; - flags = ev->flags; - if (flags & EXEC_OBJECT_PINNED && - flags & __EXEC_OBJECT_HAS_PIN) - continue; + if (pass >= 1) + unpinned = eb_unbind(eb);
- eb_unreserve_vma(ev); - - if (flags & EXEC_OBJECT_PINNED) - /* Pinned must have their slot */ - list_add(&ev->bind_link, &eb->unbound); - else if (flags & __EXEC_OBJECT_NEEDS_MAP) - /* Map require the lowest 256MiB (aperture) */ - list_add_tail(&ev->bind_link, &eb->unbound); - else if (!(flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS)) - /* Prioritise 4GiB region for restricted bo */ - list_add(&ev->bind_link, &last); - else - list_add_tail(&ev->bind_link, &last); - } - list_splice_tail(&last, &eb->unbound); - - switch (pass++) { - case 0: - break; + if (pass == 2) { + /* No point in defragmenting gtt if all is pinned */ + if (!unpinned) + return -ENOSPC;
- case 1: - /* Too fragmented, unbind everything and retry */ - mutex_lock(&eb->context->vm->mutex); - err = i915_gem_evict_vm(eb->context->vm, &eb->ww); - mutex_unlock(&eb->context->vm->mutex); + err = eb_evict_vm(eb); if (err) return err; - break; + }
- default: - return -ENOSPC; + list_for_each_entry(ev, &eb->unbound, bind_link) { + err = eb_reserve_vma(eb, ev, pin_flags); + if (err) + break; }
- pin_flags = PIN_USER; - } while (1); + if (err != -ENOSPC) + break; + } + + return err; }
static int eb_select_context(struct i915_execbuffer *eb) @@ -1184,10 +1220,11 @@ static void *reloc_kmap(struct drm_i915_gem_object *obj, return vaddr; }
-static void *reloc_iomap(struct drm_i915_gem_object *obj, +static void *reloc_iomap(struct i915_vma *batch, struct i915_execbuffer *eb, unsigned long page) { + struct drm_i915_gem_object *obj = batch->obj; struct reloc_cache *cache = &eb->reloc_cache; struct i915_ggtt *ggtt = cache_to_ggtt(cache); unsigned long offset; @@ -1197,7 +1234,7 @@ static void *reloc_iomap(struct drm_i915_gem_object *obj, intel_gt_flush_ggtt_writes(ggtt->vm.gt); io_mapping_unmap_atomic((void __force __iomem *) unmask_page(cache->vaddr)); } else { - struct i915_vma *vma; + struct i915_vma *vma = ERR_PTR(-ENODEV); int err;
if (i915_gem_object_is_tiled(obj)) @@ -1210,10 +1247,23 @@ static void *reloc_iomap(struct drm_i915_gem_object *obj, if (err) return ERR_PTR(err);
- vma = i915_gem_object_ggtt_pin_ww(obj, &eb->ww, NULL, 0, 0, - PIN_MAPPABLE | - PIN_NONBLOCK /* NOWARN */ | - PIN_NOEVICT); + /* + * i915_gem_object_ggtt_pin_ww may attempt to remove the batch + * VMA from the object list because we no longer pin. + * + * Only attempt to pin the batch buffer to ggtt if the current batch + * is not inside ggtt, or the batch buffer is not misplaced. + */ + if (!i915_is_ggtt(batch->vm)) { + vma = i915_gem_object_ggtt_pin_ww(obj, &eb->ww, NULL, 0, 0, + PIN_MAPPABLE | + PIN_NONBLOCK /* NOWARN */ | + PIN_NOEVICT); + } else if (i915_vma_is_map_and_fenceable(batch)) { + __i915_vma_pin(batch); + vma = batch; + } + if (vma == ERR_PTR(-EDEADLK)) return vma;
@@ -1251,7 +1301,7 @@ static void *reloc_iomap(struct drm_i915_gem_object *obj, return vaddr; }
-static void *reloc_vaddr(struct drm_i915_gem_object *obj, +static void *reloc_vaddr(struct i915_vma *vma, struct i915_execbuffer *eb, unsigned long page) { @@ -1263,9 +1313,9 @@ static void *reloc_vaddr(struct drm_i915_gem_object *obj, } else { vaddr = NULL; if ((cache->vaddr & KMAP) == 0) - vaddr = reloc_iomap(obj, eb, page); + vaddr = reloc_iomap(vma, eb, page); if (!vaddr) - vaddr = reloc_kmap(obj, cache, page); + vaddr = reloc_kmap(vma->obj, cache, page); }
return vaddr; @@ -1306,7 +1356,7 @@ relocate_entry(struct i915_vma *vma, void *vaddr;
repeat: - vaddr = reloc_vaddr(vma->obj, eb, + vaddr = reloc_vaddr(vma, eb, offset >> PAGE_SHIFT); if (IS_ERR(vaddr)) return PTR_ERR(vaddr); @@ -2074,7 +2124,7 @@ shadow_batch_pin(struct i915_execbuffer *eb, if (IS_ERR(vma)) return vma;
- err = i915_vma_pin_ww(vma, &eb->ww, 0, 0, flags); + err = i915_vma_pin_ww(vma, &eb->ww, 0, 0, flags | PIN_VALIDATE); if (err) return ERR_PTR(err);
@@ -2088,7 +2138,7 @@ static struct i915_vma *eb_dispatch_secure(struct i915_execbuffer *eb, struct i9 * batch" bit. Hence we need to pin secure batches into the global gtt. * hsw should have this fixed, but bdw mucks it up again. */ if (eb->batch_flags & I915_DISPATCH_SECURE) - return i915_gem_object_ggtt_pin_ww(vma->obj, &eb->ww, NULL, 0, 0, 0); + return i915_gem_object_ggtt_pin_ww(vma->obj, &eb->ww, NULL, 0, 0, PIN_VALIDATE);
return NULL; } @@ -2139,13 +2189,12 @@ static int eb_parse(struct i915_execbuffer *eb)
err = i915_gem_object_lock(pool->obj, &eb->ww); if (err) - goto err; + return err;
shadow = shadow_batch_pin(eb, pool->obj, eb->context->vm, PIN_USER); - if (IS_ERR(shadow)) { - err = PTR_ERR(shadow); - goto err; - } + if (IS_ERR(shadow)) + return PTR_ERR(shadow); + intel_gt_buffer_pool_mark_used(pool); i915_gem_object_set_readonly(shadow->obj); shadow->private = pool; @@ -2157,25 +2206,21 @@ static int eb_parse(struct i915_execbuffer *eb) shadow = shadow_batch_pin(eb, pool->obj, &eb->gt->ggtt->vm, PIN_GLOBAL); - if (IS_ERR(shadow)) { - err = PTR_ERR(shadow); - shadow = trampoline; - goto err_shadow; - } + if (IS_ERR(shadow)) + return PTR_ERR(shadow); + shadow->private = pool;
eb->batch_flags |= I915_DISPATCH_SECURE; }
batch = eb_dispatch_secure(eb, shadow); - if (IS_ERR(batch)) { - err = PTR_ERR(batch); - goto err_trampoline; - } + if (IS_ERR(batch)) + return PTR_ERR(batch);
err = dma_resv_reserve_shared(shadow->obj->base.resv, 1); if (err) - goto err_trampoline; + return err;
err = intel_engine_cmd_parser(eb->context->engine, eb->batches[0]->vma, @@ -2183,7 +2228,7 @@ static int eb_parse(struct i915_execbuffer *eb) eb->batch_len[0], shadow, trampoline); if (err) - goto err_unpin_batch; + return err;
eb->batches[0] = &eb->vma[eb->buffer_count++]; eb->batches[0]->vma = i915_vma_get(shadow); @@ -2202,17 +2247,6 @@ static int eb_parse(struct i915_execbuffer *eb) eb->batches[0]->vma = i915_vma_get(batch); } return 0; - -err_unpin_batch: - if (batch) - i915_vma_unpin(batch); -err_trampoline: - if (trampoline) - i915_vma_unpin(trampoline); -err_shadow: - i915_vma_unpin(shadow); -err: - return err; }
static int eb_request_submit(struct i915_execbuffer *eb, @@ -3337,8 +3371,6 @@ i915_gem_do_execbuffer(struct drm_device *dev,
err_vma: eb_release_vmas(&eb, true); - if (eb.trampoline) - i915_vma_unpin(eb.trampoline); WARN_ON(err == -EDEADLK); i915_gem_ww_ctx_fini(&eb.ww);
diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.h b/drivers/gpu/drm/i915/i915_gem_gtt.h index e4938aba3fe9..8c2f57eb5dda 100644 --- a/drivers/gpu/drm/i915/i915_gem_gtt.h +++ b/drivers/gpu/drm/i915/i915_gem_gtt.h @@ -44,6 +44,7 @@ int i915_gem_gtt_insert(struct i915_address_space *vm, #define PIN_HIGH BIT_ULL(5) #define PIN_OFFSET_BIAS BIT_ULL(6) #define PIN_OFFSET_FIXED BIT_ULL(7) +#define PIN_VALIDATE BIT_ULL(8) /* validate placement only, no need to call unpin() */
#define PIN_GLOBAL BIT_ULL(10) /* I915_VMA_GLOBAL_BIND */ #define PIN_USER BIT_ULL(11) /* I915_VMA_LOCAL_BIND */ diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c index 65168db534f0..0706731b211d 100644 --- a/drivers/gpu/drm/i915/i915_vma.c +++ b/drivers/gpu/drm/i915/i915_vma.c @@ -751,6 +751,15 @@ static bool try_qad_pin(struct i915_vma *vma, unsigned int flags) unsigned int bound;
bound = atomic_read(&vma->flags); + + if (flags & PIN_VALIDATE) { + flags &= I915_VMA_BIND_MASK; + + return (flags & bound) == flags; + } + + /* with the lock mandatory for unbind, we don't race here */ + flags &= I915_VMA_BIND_MASK; do { if (unlikely(flags & ~bound)) return false; @@ -1176,7 +1185,7 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, GEM_BUG_ON(!(flags & (PIN_USER | PIN_GLOBAL)));
/* First try and grab the pin without rebinding the vma */ - if (try_qad_pin(vma, flags & I915_VMA_BIND_MASK)) + if (try_qad_pin(vma, flags)) return 0;
err = i915_vma_get_pages(vma); @@ -1255,7 +1264,8 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, }
if (unlikely(!(flags & ~bound & I915_VMA_BIND_MASK))) { - __i915_vma_pin(vma); + if (!(flags & PIN_VALIDATE)) + __i915_vma_pin(vma); goto err_unlock; }
@@ -1284,8 +1294,10 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, atomic_add(I915_VMA_PAGES_ACTIVE, &vma->pages_count); list_move_tail(&vma->vm_link, &vma->vm->bound_list);
- __i915_vma_pin(vma); - GEM_BUG_ON(!i915_vma_is_pinned(vma)); + if (!(flags & PIN_VALIDATE)) { + __i915_vma_pin(vma); + GEM_BUG_ON(!i915_vma_is_pinned(vma)); + } GEM_BUG_ON(!i915_vma_is_bound(vma, flags)); GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
@@ -1538,8 +1550,6 @@ static int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request * { int err;
- GEM_BUG_ON(!i915_vma_is_pinned(vma)); - /* Wait for the vma to be bound before we start! */ err = __i915_request_await_bind(rq, vma); if (err) @@ -1558,6 +1568,8 @@ int _i915_vma_move_to_active(struct i915_vma *vma,
assert_object_held(obj);
+ GEM_BUG_ON(!vma->pages); + err = __i915_vma_move_to_active(vma, rq); if (unlikely(err)) return err;
On Thu, 21 Oct 2021 at 11:37, Maarten Lankhorst maarten.lankhorst@linux.intel.com wrote:
Add a flag PIN_VALIDATE, to indicate we don't need to pin and only protected by the object lock.
This removes the need to unpin, which is done by just releasing the lock.
eb_reserve is slightly reworked for readability, but the same steps are still done:
- First pass pins with NONBLOCK.
- Second pass unbinds all objects first, then pins.
- Third pass is only called when not all objects are softpinned, and unbinds all objects, then calls i915_gem_evict_vm(), then pins.
When evicting the entire vm in eb_reserve() we do temporarily pin objects that are marked with EXEC_OBJECT_PINNED. This is because they are already at their destination, and i915_gem_evict_vm() would otherwise unbind them.
However, we reduce the visibility of those pins by limiting the pin to our call to i915_gem_evict_vm() only, and pin with vm->mutex held, instead of the entire duration of the execbuf.
Not sure the latter matters, one can hope.. In theory we could kill the pinning by adding an extra flag to the vma to temporarily prevent unbinding for gtt for i915_gem_evict_vm only, but I think that might be overkill. We're still holding the object lock, and we don't have blocking eviction yet. It's likely sufficient to simply enforce EXEC_OBJECT_PINNED for all objects on >= gen12.
Changes since v1:
- Split out eb_reserve() into separate functions for readability.
Changes since v2:
- Make batch buffer mappable on platforms where only GGTT is available, to prevent moving the batch buffer during relocations.
Changes since v3:
- Preserve current behavior for batch buffer, instead be cautious when calling i915_gem_object_ggtt_pin_ww, and re-use the current batch vma if it's inside ggtt and map-and-fenceable.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
.../gpu/drm/i915/gem/i915_gem_execbuffer.c | 252 ++++++++++-------- drivers/gpu/drm/i915/i915_gem_gtt.h | 1 + drivers/gpu/drm/i915/i915_vma.c | 24 +- 3 files changed, 161 insertions(+), 116 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c index bbf2a10738f7..19f91143cfcf 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c @@ -439,7 +439,7 @@ eb_pin_vma(struct i915_execbuffer *eb, else pin_flags = entry->offset & PIN_OFFSET_MASK;
pin_flags |= PIN_USER | PIN_NOEVICT | PIN_OFFSET_FIXED;
pin_flags |= PIN_USER | PIN_NOEVICT | PIN_OFFSET_FIXED | PIN_VALIDATE; if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_GTT)) pin_flags |= PIN_GLOBAL;@@ -457,17 +457,15 @@ eb_pin_vma(struct i915_execbuffer *eb, entry->pad_to_size, entry->alignment, eb_pin_flags(entry, ev->flags) |
PIN_USER | PIN_NOEVICT);
PIN_USER | PIN_NOEVICT | PIN_VALIDATE); if (unlikely(err)) return err; } if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_FENCE)) { err = i915_vma_pin_fence(vma);
if (unlikely(err)) {i915_vma_unpin(vma);
if (unlikely(err)) return err;
} if (vma->fence) ev->flags |= __EXEC_OBJECT_HAS_FENCE;@@ -483,13 +481,9 @@ eb_pin_vma(struct i915_execbuffer *eb, static inline void eb_unreserve_vma(struct eb_vma *ev) {
if (!(ev->flags & __EXEC_OBJECT_HAS_PIN))return;if (unlikely(ev->flags & __EXEC_OBJECT_HAS_FENCE)) __i915_vma_unpin_fence(ev->vma);__i915_vma_unpin(ev->vma); ev->flags &= ~__EXEC_OBJECT_RESERVED;}
@@ -682,10 +676,8 @@ static int eb_reserve_vma(struct i915_execbuffer *eb,
if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_FENCE)) { err = i915_vma_pin_fence(vma);
if (unlikely(err)) {i915_vma_unpin(vma);
if (unlikely(err)) return err;
} if (vma->fence) ev->flags |= __EXEC_OBJECT_HAS_FENCE;@@ -697,85 +689,129 @@ static int eb_reserve_vma(struct i915_execbuffer *eb, return 0; }
-static int eb_reserve(struct i915_execbuffer *eb) +static int eb_evict_vm(struct i915_execbuffer *eb) +{
const unsigned int count = eb->buffer_count;unsigned int i;int err;err = mutex_lock_interruptible(&eb->context->vm->mutex);if (err)return err;/* pin to protect against i915_gem_evict_vm evicting below */for (i = 0; i < count; i++) {struct eb_vma *ev = &eb->vma[i];if (ev->flags & __EXEC_OBJECT_HAS_PIN)__i915_vma_pin(ev->vma);}/* Too fragmented, unbind everything and retry */err = i915_gem_evict_vm(eb->context->vm, &eb->ww);/* unpin objects.. */for (i = 0; i < count; i++) {struct eb_vma *ev = &eb->vma[i];if (ev->flags & __EXEC_OBJECT_HAS_PIN)i915_vma_unpin(ev->vma);}mutex_unlock(&eb->context->vm->mutex);return err;+}
+static bool eb_unbind(struct i915_execbuffer *eb) { const unsigned int count = eb->buffer_count;
unsigned int pin_flags = PIN_USER | PIN_NONBLOCK;
unsigned int i; struct list_head last;bool unpinned = false;/* Resort *all* the objects into priority order */INIT_LIST_HEAD(&eb->unbound);INIT_LIST_HEAD(&last);for (i = 0; i < count; i++) {struct eb_vma *ev = &eb->vma[i];unsigned int flags = ev->flags;if (flags & EXEC_OBJECT_PINNED &&flags & __EXEC_OBJECT_HAS_PIN)continue;unpinned = true;eb_unreserve_vma(ev);if (flags & EXEC_OBJECT_PINNED)/* Pinned must have their slot */list_add(&ev->bind_link, &eb->unbound);else if (flags & __EXEC_OBJECT_NEEDS_MAP)/* Map require the lowest 256MiB (aperture) */list_add_tail(&ev->bind_link, &eb->unbound);else if (!(flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS))/* Prioritise 4GiB region for restricted bo */list_add(&ev->bind_link, &last);elselist_add_tail(&ev->bind_link, &last);}list_splice_tail(&last, &eb->unbound);return unpinned;+}
+static int eb_reserve(struct i915_execbuffer *eb) +{ struct eb_vma *ev;
unsigned int i, pass;
unsigned int pass; int err = 0;bool unpinned; /* * Attempt to pin all of the buffers into the GTT.
* This is done in 3 phases:
* This is done in 2 phases: *
* 1a. Unbind all objects that do not match the GTT constraints for* the execbuffer (fenceable, mappable, alignment etc).* 1b. Increment pin count for already bound objects.* 2. Bind new objects.* 3. Decrement pin count.
* 1. Unbind all objects that do not match the GTT constraints for* the execbuffer (fenceable, mappable, alignment etc).* 2. Bind new objects. * * This avoid unnecessary unbinding of later objects in order to make * room for the earlier objects *unless* we need to defragment.** Defragmenting is skipped if all objects are pinned at a fixed location. */
pass = 0;do {list_for_each_entry(ev, &eb->unbound, bind_link) {err = eb_reserve_vma(eb, ev, pin_flags);if (err)break;}if (err != -ENOSPC)return err;
for (pass = 0; pass <= 2; pass++) {int pin_flags = PIN_USER | PIN_VALIDATE;
/* Resort *all* the objects into priority order */INIT_LIST_HEAD(&eb->unbound);INIT_LIST_HEAD(&last);for (i = 0; i < count; i++) {unsigned int flags;
if (pass == 0)pin_flags |= PIN_NONBLOCK;
ev = &eb->vma[i];flags = ev->flags;if (flags & EXEC_OBJECT_PINNED &&flags & __EXEC_OBJECT_HAS_PIN)continue;
if (pass >= 1)unpinned = eb_unbind(eb);
eb_unreserve_vma(ev);if (flags & EXEC_OBJECT_PINNED)/* Pinned must have their slot */list_add(&ev->bind_link, &eb->unbound);else if (flags & __EXEC_OBJECT_NEEDS_MAP)/* Map require the lowest 256MiB (aperture) */list_add_tail(&ev->bind_link, &eb->unbound);else if (!(flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS))/* Prioritise 4GiB region for restricted bo */list_add(&ev->bind_link, &last);elselist_add_tail(&ev->bind_link, &last);}list_splice_tail(&last, &eb->unbound);switch (pass++) {case 0:break;
if (pass == 2) {/* No point in defragmenting gtt if all is pinned */if (!unpinned)return -ENOSPC;
Can this ever happen? If everything is already pinned where it's meant to be, then how did we get here?
case 1:/* Too fragmented, unbind everything and retry */mutex_lock(&eb->context->vm->mutex);err = i915_gem_evict_vm(eb->context->vm, &eb->ww);mutex_unlock(&eb->context->vm->mutex);
err = eb_evict_vm(eb); if (err) return err;
break;
}
default:return -ENOSPC;
list_for_each_entry(ev, &eb->unbound, bind_link) {err = eb_reserve_vma(eb, ev, pin_flags);if (err)break; }
pin_flags = PIN_USER;} while (1);
if (err != -ENOSPC)break;}return err;}
static int eb_select_context(struct i915_execbuffer *eb) @@ -1184,10 +1220,11 @@ static void *reloc_kmap(struct drm_i915_gem_object *obj, return vaddr; }
-static void *reloc_iomap(struct drm_i915_gem_object *obj, +static void *reloc_iomap(struct i915_vma *batch, struct i915_execbuffer *eb, unsigned long page) {
struct drm_i915_gem_object *obj = batch->obj; struct reloc_cache *cache = &eb->reloc_cache; struct i915_ggtt *ggtt = cache_to_ggtt(cache); unsigned long offset;@@ -1197,7 +1234,7 @@ static void *reloc_iomap(struct drm_i915_gem_object *obj, intel_gt_flush_ggtt_writes(ggtt->vm.gt); io_mapping_unmap_atomic((void __force __iomem *) unmask_page(cache->vaddr)); } else {
struct i915_vma *vma;
struct i915_vma *vma = ERR_PTR(-ENODEV); int err; if (i915_gem_object_is_tiled(obj))@@ -1210,10 +1247,23 @@ static void *reloc_iomap(struct drm_i915_gem_object *obj, if (err) return ERR_PTR(err);
vma = i915_gem_object_ggtt_pin_ww(obj, &eb->ww, NULL, 0, 0,PIN_MAPPABLE |PIN_NONBLOCK /* NOWARN */ |PIN_NOEVICT);
/** i915_gem_object_ggtt_pin_ww may attempt to remove the batch* VMA from the object list because we no longer pin.** Only attempt to pin the batch buffer to ggtt if the current batch* is not inside ggtt, or the batch buffer is not misplaced.*/if (!i915_is_ggtt(batch->vm)) {vma = i915_gem_object_ggtt_pin_ww(obj, &eb->ww, NULL, 0, 0,PIN_MAPPABLE |PIN_NONBLOCK /* NOWARN */ |PIN_NOEVICT);} else if (i915_vma_is_map_and_fenceable(batch)) {__i915_vma_pin(batch);vma = batch;}if (vma == ERR_PTR(-EDEADLK)) return vma;@@ -1251,7 +1301,7 @@ static void *reloc_iomap(struct drm_i915_gem_object *obj, return vaddr; }
-static void *reloc_vaddr(struct drm_i915_gem_object *obj, +static void *reloc_vaddr(struct i915_vma *vma, struct i915_execbuffer *eb, unsigned long page) { @@ -1263,9 +1313,9 @@ static void *reloc_vaddr(struct drm_i915_gem_object *obj, } else { vaddr = NULL; if ((cache->vaddr & KMAP) == 0)
vaddr = reloc_iomap(obj, eb, page);
vaddr = reloc_iomap(vma, eb, page); if (!vaddr)
vaddr = reloc_kmap(obj, cache, page);
vaddr = reloc_kmap(vma->obj, cache, page); } return vaddr;@@ -1306,7 +1356,7 @@ relocate_entry(struct i915_vma *vma, void *vaddr;
repeat:
vaddr = reloc_vaddr(vma->obj, eb,
vaddr = reloc_vaddr(vma, eb, offset >> PAGE_SHIFT); if (IS_ERR(vaddr)) return PTR_ERR(vaddr);@@ -2074,7 +2124,7 @@ shadow_batch_pin(struct i915_execbuffer *eb, if (IS_ERR(vma)) return vma;
err = i915_vma_pin_ww(vma, &eb->ww, 0, 0, flags);
err = i915_vma_pin_ww(vma, &eb->ww, 0, 0, flags | PIN_VALIDATE); if (err) return ERR_PTR(err);@@ -2088,7 +2138,7 @@ static struct i915_vma *eb_dispatch_secure(struct i915_execbuffer *eb, struct i9 * batch" bit. Hence we need to pin secure batches into the global gtt. * hsw should have this fixed, but bdw mucks it up again. */ if (eb->batch_flags & I915_DISPATCH_SECURE)
return i915_gem_object_ggtt_pin_ww(vma->obj, &eb->ww, NULL, 0, 0, 0);
return i915_gem_object_ggtt_pin_ww(vma->obj, &eb->ww, NULL, 0, 0, PIN_VALIDATE); return NULL;} @@ -2139,13 +2189,12 @@ static int eb_parse(struct i915_execbuffer *eb)
err = i915_gem_object_lock(pool->obj, &eb->ww); if (err)
goto err;
return err; shadow = shadow_batch_pin(eb, pool->obj, eb->context->vm, PIN_USER);
if (IS_ERR(shadow)) {err = PTR_ERR(shadow);goto err;}
if (IS_ERR(shadow))return PTR_ERR(shadow);intel_gt_buffer_pool_mark_used(pool); i915_gem_object_set_readonly(shadow->obj); shadow->private = pool;@@ -2157,25 +2206,21 @@ static int eb_parse(struct i915_execbuffer *eb) shadow = shadow_batch_pin(eb, pool->obj, &eb->gt->ggtt->vm, PIN_GLOBAL);
if (IS_ERR(shadow)) {err = PTR_ERR(shadow);shadow = trampoline;goto err_shadow;}
if (IS_ERR(shadow))return PTR_ERR(shadow);shadow->private = pool; eb->batch_flags |= I915_DISPATCH_SECURE; } batch = eb_dispatch_secure(eb, shadow);
if (IS_ERR(batch)) {err = PTR_ERR(batch);goto err_trampoline;}
if (IS_ERR(batch))return PTR_ERR(batch); err = dma_resv_reserve_shared(shadow->obj->base.resv, 1); if (err)
goto err_trampoline;
return err; err = intel_engine_cmd_parser(eb->context->engine, eb->batches[0]->vma,@@ -2183,7 +2228,7 @@ static int eb_parse(struct i915_execbuffer *eb) eb->batch_len[0], shadow, trampoline); if (err)
goto err_unpin_batch;
return err; eb->batches[0] = &eb->vma[eb->buffer_count++]; eb->batches[0]->vma = i915_vma_get(shadow);@@ -2202,17 +2247,6 @@ static int eb_parse(struct i915_execbuffer *eb) eb->batches[0]->vma = i915_vma_get(batch); } return 0;
-err_unpin_batch:
if (batch)i915_vma_unpin(batch);-err_trampoline:
if (trampoline)i915_vma_unpin(trampoline);-err_shadow:
i915_vma_unpin(shadow);-err:
return err;}
static int eb_request_submit(struct i915_execbuffer *eb, @@ -3337,8 +3371,6 @@ i915_gem_do_execbuffer(struct drm_device *dev,
err_vma: eb_release_vmas(&eb, true);
if (eb.trampoline)i915_vma_unpin(eb.trampoline); WARN_ON(err == -EDEADLK); i915_gem_ww_ctx_fini(&eb.ww);diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.h b/drivers/gpu/drm/i915/i915_gem_gtt.h index e4938aba3fe9..8c2f57eb5dda 100644 --- a/drivers/gpu/drm/i915/i915_gem_gtt.h +++ b/drivers/gpu/drm/i915/i915_gem_gtt.h @@ -44,6 +44,7 @@ int i915_gem_gtt_insert(struct i915_address_space *vm, #define PIN_HIGH BIT_ULL(5) #define PIN_OFFSET_BIAS BIT_ULL(6) #define PIN_OFFSET_FIXED BIT_ULL(7) +#define PIN_VALIDATE BIT_ULL(8) /* validate placement only, no need to call unpin() */
#define PIN_GLOBAL BIT_ULL(10) /* I915_VMA_GLOBAL_BIND */ #define PIN_USER BIT_ULL(11) /* I915_VMA_LOCAL_BIND */ diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c index 65168db534f0..0706731b211d 100644 --- a/drivers/gpu/drm/i915/i915_vma.c +++ b/drivers/gpu/drm/i915/i915_vma.c @@ -751,6 +751,15 @@ static bool try_qad_pin(struct i915_vma *vma, unsigned int flags) unsigned int bound;
bound = atomic_read(&vma->flags);
if (flags & PIN_VALIDATE) {flags &= I915_VMA_BIND_MASK;return (flags & bound) == flags;}/* with the lock mandatory for unbind, we don't race here */flags &= I915_VMA_BIND_MASK; do { if (unlikely(flags & ~bound)) return false;@@ -1176,7 +1185,7 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, GEM_BUG_ON(!(flags & (PIN_USER | PIN_GLOBAL)));
/* First try and grab the pin without rebinding the vma */
if (try_qad_pin(vma, flags & I915_VMA_BIND_MASK))
if (try_qad_pin(vma, flags)) return 0; err = i915_vma_get_pages(vma);@@ -1255,7 +1264,8 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, }
if (unlikely(!(flags & ~bound & I915_VMA_BIND_MASK))) {
__i915_vma_pin(vma);
if (!(flags & PIN_VALIDATE))__i915_vma_pin(vma); goto err_unlock; }@@ -1284,8 +1294,10 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, atomic_add(I915_VMA_PAGES_ACTIVE, &vma->pages_count); list_move_tail(&vma->vm_link, &vma->vm->bound_list);
__i915_vma_pin(vma);GEM_BUG_ON(!i915_vma_is_pinned(vma));
if (!(flags & PIN_VALIDATE)) {__i915_vma_pin(vma);GEM_BUG_ON(!i915_vma_is_pinned(vma));} GEM_BUG_ON(!i915_vma_is_bound(vma, flags)); GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));@@ -1538,8 +1550,6 @@ static int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request * { int err;
GEM_BUG_ON(!i915_vma_is_pinned(vma));/* Wait for the vma to be bound before we start! */ err = __i915_request_await_bind(rq, vma); if (err)@@ -1558,6 +1568,8 @@ int _i915_vma_move_to_active(struct i915_vma *vma,
assert_object_held(obj);
GEM_BUG_ON(!vma->pages);err = __i915_vma_move_to_active(vma, rq); if (unlikely(err)) return err;-- 2.33.0
On 25-10-2021 17:02, Matthew Auld wrote:
On Thu, 21 Oct 2021 at 11:37, Maarten Lankhorst maarten.lankhorst@linux.intel.com wrote:
Add a flag PIN_VALIDATE, to indicate we don't need to pin and only protected by the object lock.
This removes the need to unpin, which is done by just releasing the lock.
eb_reserve is slightly reworked for readability, but the same steps are still done:
- First pass pins with NONBLOCK.
- Second pass unbinds all objects first, then pins.
- Third pass is only called when not all objects are softpinned, and unbinds all objects, then calls i915_gem_evict_vm(), then pins.
When evicting the entire vm in eb_reserve() we do temporarily pin objects that are marked with EXEC_OBJECT_PINNED. This is because they are already at their destination, and i915_gem_evict_vm() would otherwise unbind them.
However, we reduce the visibility of those pins by limiting the pin to our call to i915_gem_evict_vm() only, and pin with vm->mutex held, instead of the entire duration of the execbuf.
Not sure the latter matters, one can hope.. In theory we could kill the pinning by adding an extra flag to the vma to temporarily prevent unbinding for gtt for i915_gem_evict_vm only, but I think that might be overkill. We're still holding the object lock, and we don't have blocking eviction yet. It's likely sufficient to simply enforce EXEC_OBJECT_PINNED for all objects on >= gen12.
Changes since v1:
- Split out eb_reserve() into separate functions for readability.
Changes since v2:
- Make batch buffer mappable on platforms where only GGTT is available, to prevent moving the batch buffer during relocations.
Changes since v3:
- Preserve current behavior for batch buffer, instead be cautious when calling i915_gem_object_ggtt_pin_ww, and re-use the current batch vma if it's inside ggtt and map-and-fenceable.
Signed-off-by: Maarten Lankhorst maarten.lankhorst@linux.intel.com
.../gpu/drm/i915/gem/i915_gem_execbuffer.c | 252 ++++++++++-------- drivers/gpu/drm/i915/i915_gem_gtt.h | 1 + drivers/gpu/drm/i915/i915_vma.c | 24 +- 3 files changed, 161 insertions(+), 116 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c index bbf2a10738f7..19f91143cfcf 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c @@ -439,7 +439,7 @@ eb_pin_vma(struct i915_execbuffer *eb, else pin_flags = entry->offset & PIN_OFFSET_MASK;
pin_flags |= PIN_USER | PIN_NOEVICT | PIN_OFFSET_FIXED;
pin_flags |= PIN_USER | PIN_NOEVICT | PIN_OFFSET_FIXED | PIN_VALIDATE; if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_GTT)) pin_flags |= PIN_GLOBAL;@@ -457,17 +457,15 @@ eb_pin_vma(struct i915_execbuffer *eb, entry->pad_to_size, entry->alignment, eb_pin_flags(entry, ev->flags) |
PIN_USER | PIN_NOEVICT);
PIN_USER | PIN_NOEVICT | PIN_VALIDATE); if (unlikely(err)) return err; } if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_FENCE)) { err = i915_vma_pin_fence(vma);
if (unlikely(err)) {i915_vma_unpin(vma);
if (unlikely(err)) return err;
} if (vma->fence) ev->flags |= __EXEC_OBJECT_HAS_FENCE;@@ -483,13 +481,9 @@ eb_pin_vma(struct i915_execbuffer *eb, static inline void eb_unreserve_vma(struct eb_vma *ev) {
if (!(ev->flags & __EXEC_OBJECT_HAS_PIN))return;if (unlikely(ev->flags & __EXEC_OBJECT_HAS_FENCE)) __i915_vma_unpin_fence(ev->vma);__i915_vma_unpin(ev->vma); ev->flags &= ~__EXEC_OBJECT_RESERVED;}
@@ -682,10 +676,8 @@ static int eb_reserve_vma(struct i915_execbuffer *eb,
if (unlikely(ev->flags & EXEC_OBJECT_NEEDS_FENCE)) { err = i915_vma_pin_fence(vma);
if (unlikely(err)) {i915_vma_unpin(vma);
if (unlikely(err)) return err;
} if (vma->fence) ev->flags |= __EXEC_OBJECT_HAS_FENCE;@@ -697,85 +689,129 @@ static int eb_reserve_vma(struct i915_execbuffer *eb, return 0; }
-static int eb_reserve(struct i915_execbuffer *eb) +static int eb_evict_vm(struct i915_execbuffer *eb) +{
const unsigned int count = eb->buffer_count;unsigned int i;int err;err = mutex_lock_interruptible(&eb->context->vm->mutex);if (err)return err;/* pin to protect against i915_gem_evict_vm evicting below */for (i = 0; i < count; i++) {struct eb_vma *ev = &eb->vma[i];if (ev->flags & __EXEC_OBJECT_HAS_PIN)__i915_vma_pin(ev->vma);}/* Too fragmented, unbind everything and retry */err = i915_gem_evict_vm(eb->context->vm, &eb->ww);/* unpin objects.. */for (i = 0; i < count; i++) {struct eb_vma *ev = &eb->vma[i];if (ev->flags & __EXEC_OBJECT_HAS_PIN)i915_vma_unpin(ev->vma);}mutex_unlock(&eb->context->vm->mutex);return err;+}
+static bool eb_unbind(struct i915_execbuffer *eb) { const unsigned int count = eb->buffer_count;
unsigned int pin_flags = PIN_USER | PIN_NONBLOCK;
unsigned int i; struct list_head last;bool unpinned = false;/* Resort *all* the objects into priority order */INIT_LIST_HEAD(&eb->unbound);INIT_LIST_HEAD(&last);for (i = 0; i < count; i++) {struct eb_vma *ev = &eb->vma[i];unsigned int flags = ev->flags;if (flags & EXEC_OBJECT_PINNED &&flags & __EXEC_OBJECT_HAS_PIN)continue;unpinned = true;eb_unreserve_vma(ev);if (flags & EXEC_OBJECT_PINNED)/* Pinned must have their slot */list_add(&ev->bind_link, &eb->unbound);else if (flags & __EXEC_OBJECT_NEEDS_MAP)/* Map require the lowest 256MiB (aperture) */list_add_tail(&ev->bind_link, &eb->unbound);else if (!(flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS))/* Prioritise 4GiB region for restricted bo */list_add(&ev->bind_link, &last);elselist_add_tail(&ev->bind_link, &last);}list_splice_tail(&last, &eb->unbound);return unpinned;+}
+static int eb_reserve(struct i915_execbuffer *eb) +{ struct eb_vma *ev;
unsigned int i, pass;
unsigned int pass; int err = 0;bool unpinned; /* * Attempt to pin all of the buffers into the GTT.
* This is done in 3 phases:
* This is done in 2 phases: *
* 1a. Unbind all objects that do not match the GTT constraints for* the execbuffer (fenceable, mappable, alignment etc).* 1b. Increment pin count for already bound objects.* 2. Bind new objects.* 3. Decrement pin count.
* 1. Unbind all objects that do not match the GTT constraints for* the execbuffer (fenceable, mappable, alignment etc).* 2. Bind new objects. * * This avoid unnecessary unbinding of later objects in order to make * room for the earlier objects *unless* we need to defragment.** Defragmenting is skipped if all objects are pinned at a fixed location. */
pass = 0;do {list_for_each_entry(ev, &eb->unbound, bind_link) {err = eb_reserve_vma(eb, ev, pin_flags);if (err)break;}if (err != -ENOSPC)return err;
for (pass = 0; pass <= 2; pass++) {int pin_flags = PIN_USER | PIN_VALIDATE;
/* Resort *all* the objects into priority order */INIT_LIST_HEAD(&eb->unbound);INIT_LIST_HEAD(&last);for (i = 0; i < count; i++) {unsigned int flags;
if (pass == 0)pin_flags |= PIN_NONBLOCK;
ev = &eb->vma[i];flags = ev->flags;if (flags & EXEC_OBJECT_PINNED &&flags & __EXEC_OBJECT_HAS_PIN)continue;
if (pass >= 1)unpinned = eb_unbind(eb);
eb_unreserve_vma(ev);if (flags & EXEC_OBJECT_PINNED)/* Pinned must have their slot */list_add(&ev->bind_link, &eb->unbound);else if (flags & __EXEC_OBJECT_NEEDS_MAP)/* Map require the lowest 256MiB (aperture) */list_add_tail(&ev->bind_link, &eb->unbound);else if (!(flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS))/* Prioritise 4GiB region for restricted bo */list_add(&ev->bind_link, &last);elselist_add_tail(&ev->bind_link, &last);}list_splice_tail(&last, &eb->unbound);switch (pass++) {case 0:break;
if (pass == 2) {/* No point in defragmenting gtt if all is pinned */if (!unpinned)return -ENOSPC;Can this ever happen? If everything is already pinned where it's meant to be, then how did we get here?
Hmm no, in previous versions it could, but it looks like I removed that in a refactor. Seems like dead code.
case 1:/* Too fragmented, unbind everything and retry */mutex_lock(&eb->context->vm->mutex);err = i915_gem_evict_vm(eb->context->vm, &eb->ww);mutex_unlock(&eb->context->vm->mutex);
err = eb_evict_vm(eb); if (err) return err;
break;
}
default:return -ENOSPC;
list_for_each_entry(ev, &eb->unbound, bind_link) {err = eb_reserve_vma(eb, ev, pin_flags);if (err)break; }
pin_flags = PIN_USER;} while (1);
if (err != -ENOSPC)break;}return err;}
static int eb_select_context(struct i915_execbuffer *eb) @@ -1184,10 +1220,11 @@ static void *reloc_kmap(struct drm_i915_gem_object *obj, return vaddr; }
-static void *reloc_iomap(struct drm_i915_gem_object *obj, +static void *reloc_iomap(struct i915_vma *batch, struct i915_execbuffer *eb, unsigned long page) {
struct drm_i915_gem_object *obj = batch->obj; struct reloc_cache *cache = &eb->reloc_cache; struct i915_ggtt *ggtt = cache_to_ggtt(cache); unsigned long offset;@@ -1197,7 +1234,7 @@ static void *reloc_iomap(struct drm_i915_gem_object *obj, intel_gt_flush_ggtt_writes(ggtt->vm.gt); io_mapping_unmap_atomic((void __force __iomem *) unmask_page(cache->vaddr)); } else {
struct i915_vma *vma;
struct i915_vma *vma = ERR_PTR(-ENODEV); int err; if (i915_gem_object_is_tiled(obj))@@ -1210,10 +1247,23 @@ static void *reloc_iomap(struct drm_i915_gem_object *obj, if (err) return ERR_PTR(err);
vma = i915_gem_object_ggtt_pin_ww(obj, &eb->ww, NULL, 0, 0,PIN_MAPPABLE |PIN_NONBLOCK /* NOWARN */ |PIN_NOEVICT);
/** i915_gem_object_ggtt_pin_ww may attempt to remove the batch* VMA from the object list because we no longer pin.** Only attempt to pin the batch buffer to ggtt if the current batch* is not inside ggtt, or the batch buffer is not misplaced.*/if (!i915_is_ggtt(batch->vm)) {vma = i915_gem_object_ggtt_pin_ww(obj, &eb->ww, NULL, 0, 0,PIN_MAPPABLE |PIN_NONBLOCK /* NOWARN */ |PIN_NOEVICT);} else if (i915_vma_is_map_and_fenceable(batch)) {__i915_vma_pin(batch);vma = batch;}if (vma == ERR_PTR(-EDEADLK)) return vma;@@ -1251,7 +1301,7 @@ static void *reloc_iomap(struct drm_i915_gem_object *obj, return vaddr; }
-static void *reloc_vaddr(struct drm_i915_gem_object *obj, +static void *reloc_vaddr(struct i915_vma *vma, struct i915_execbuffer *eb, unsigned long page) { @@ -1263,9 +1313,9 @@ static void *reloc_vaddr(struct drm_i915_gem_object *obj, } else { vaddr = NULL; if ((cache->vaddr & KMAP) == 0)
vaddr = reloc_iomap(obj, eb, page);
vaddr = reloc_iomap(vma, eb, page); if (!vaddr)
vaddr = reloc_kmap(obj, cache, page);
vaddr = reloc_kmap(vma->obj, cache, page); } return vaddr;@@ -1306,7 +1356,7 @@ relocate_entry(struct i915_vma *vma, void *vaddr;
repeat:
vaddr = reloc_vaddr(vma->obj, eb,
vaddr = reloc_vaddr(vma, eb, offset >> PAGE_SHIFT); if (IS_ERR(vaddr)) return PTR_ERR(vaddr);@@ -2074,7 +2124,7 @@ shadow_batch_pin(struct i915_execbuffer *eb, if (IS_ERR(vma)) return vma;
err = i915_vma_pin_ww(vma, &eb->ww, 0, 0, flags);
err = i915_vma_pin_ww(vma, &eb->ww, 0, 0, flags | PIN_VALIDATE); if (err) return ERR_PTR(err);@@ -2088,7 +2138,7 @@ static struct i915_vma *eb_dispatch_secure(struct i915_execbuffer *eb, struct i9 * batch" bit. Hence we need to pin secure batches into the global gtt. * hsw should have this fixed, but bdw mucks it up again. */ if (eb->batch_flags & I915_DISPATCH_SECURE)
return i915_gem_object_ggtt_pin_ww(vma->obj, &eb->ww, NULL, 0, 0, 0);
return i915_gem_object_ggtt_pin_ww(vma->obj, &eb->ww, NULL, 0, 0, PIN_VALIDATE); return NULL;} @@ -2139,13 +2189,12 @@ static int eb_parse(struct i915_execbuffer *eb)
err = i915_gem_object_lock(pool->obj, &eb->ww); if (err)
goto err;
return err; shadow = shadow_batch_pin(eb, pool->obj, eb->context->vm, PIN_USER);
if (IS_ERR(shadow)) {err = PTR_ERR(shadow);goto err;}
if (IS_ERR(shadow))return PTR_ERR(shadow);intel_gt_buffer_pool_mark_used(pool); i915_gem_object_set_readonly(shadow->obj); shadow->private = pool;@@ -2157,25 +2206,21 @@ static int eb_parse(struct i915_execbuffer *eb) shadow = shadow_batch_pin(eb, pool->obj, &eb->gt->ggtt->vm, PIN_GLOBAL);
if (IS_ERR(shadow)) {err = PTR_ERR(shadow);shadow = trampoline;goto err_shadow;}
if (IS_ERR(shadow))return PTR_ERR(shadow);shadow->private = pool; eb->batch_flags |= I915_DISPATCH_SECURE; } batch = eb_dispatch_secure(eb, shadow);
if (IS_ERR(batch)) {err = PTR_ERR(batch);goto err_trampoline;}
if (IS_ERR(batch))return PTR_ERR(batch); err = dma_resv_reserve_shared(shadow->obj->base.resv, 1); if (err)
goto err_trampoline;
return err; err = intel_engine_cmd_parser(eb->context->engine, eb->batches[0]->vma,@@ -2183,7 +2228,7 @@ static int eb_parse(struct i915_execbuffer *eb) eb->batch_len[0], shadow, trampoline); if (err)
goto err_unpin_batch;
return err; eb->batches[0] = &eb->vma[eb->buffer_count++]; eb->batches[0]->vma = i915_vma_get(shadow);@@ -2202,17 +2247,6 @@ static int eb_parse(struct i915_execbuffer *eb) eb->batches[0]->vma = i915_vma_get(batch); } return 0;
-err_unpin_batch:
if (batch)i915_vma_unpin(batch);-err_trampoline:
if (trampoline)i915_vma_unpin(trampoline);-err_shadow:
i915_vma_unpin(shadow);-err:
return err;}
static int eb_request_submit(struct i915_execbuffer *eb, @@ -3337,8 +3371,6 @@ i915_gem_do_execbuffer(struct drm_device *dev,
err_vma: eb_release_vmas(&eb, true);
if (eb.trampoline)i915_vma_unpin(eb.trampoline); WARN_ON(err == -EDEADLK); i915_gem_ww_ctx_fini(&eb.ww);diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.h b/drivers/gpu/drm/i915/i915_gem_gtt.h index e4938aba3fe9..8c2f57eb5dda 100644 --- a/drivers/gpu/drm/i915/i915_gem_gtt.h +++ b/drivers/gpu/drm/i915/i915_gem_gtt.h @@ -44,6 +44,7 @@ int i915_gem_gtt_insert(struct i915_address_space *vm, #define PIN_HIGH BIT_ULL(5) #define PIN_OFFSET_BIAS BIT_ULL(6) #define PIN_OFFSET_FIXED BIT_ULL(7) +#define PIN_VALIDATE BIT_ULL(8) /* validate placement only, no need to call unpin() */
#define PIN_GLOBAL BIT_ULL(10) /* I915_VMA_GLOBAL_BIND */ #define PIN_USER BIT_ULL(11) /* I915_VMA_LOCAL_BIND */ diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c index 65168db534f0..0706731b211d 100644 --- a/drivers/gpu/drm/i915/i915_vma.c +++ b/drivers/gpu/drm/i915/i915_vma.c @@ -751,6 +751,15 @@ static bool try_qad_pin(struct i915_vma *vma, unsigned int flags) unsigned int bound;
bound = atomic_read(&vma->flags);
if (flags & PIN_VALIDATE) {flags &= I915_VMA_BIND_MASK;return (flags & bound) == flags;}/* with the lock mandatory for unbind, we don't race here */flags &= I915_VMA_BIND_MASK; do { if (unlikely(flags & ~bound)) return false;@@ -1176,7 +1185,7 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, GEM_BUG_ON(!(flags & (PIN_USER | PIN_GLOBAL)));
/* First try and grab the pin without rebinding the vma */
if (try_qad_pin(vma, flags & I915_VMA_BIND_MASK))
if (try_qad_pin(vma, flags)) return 0; err = i915_vma_get_pages(vma);@@ -1255,7 +1264,8 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, }
if (unlikely(!(flags & ~bound & I915_VMA_BIND_MASK))) {
__i915_vma_pin(vma);
if (!(flags & PIN_VALIDATE))__i915_vma_pin(vma); goto err_unlock; }@@ -1284,8 +1294,10 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww, atomic_add(I915_VMA_PAGES_ACTIVE, &vma->pages_count); list_move_tail(&vma->vm_link, &vma->vm->bound_list);
__i915_vma_pin(vma);GEM_BUG_ON(!i915_vma_is_pinned(vma));
if (!(flags & PIN_VALIDATE)) {__i915_vma_pin(vma);GEM_BUG_ON(!i915_vma_is_pinned(vma));} GEM_BUG_ON(!i915_vma_is_bound(vma, flags)); GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));@@ -1538,8 +1550,6 @@ static int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request * { int err;
GEM_BUG_ON(!i915_vma_is_pinned(vma));/* Wait for the vma to be bound before we start! */ err = __i915_request_await_bind(rq, vma); if (err)@@ -1558,6 +1568,8 @@ int _i915_vma_move_to_active(struct i915_vma *vma,
assert_object_held(obj);
GEM_BUG_ON(!vma->pages);err = __i915_vma_move_to_active(vma, rq); if (unlikely(err)) return err;-- 2.33.0
dri-devel@lists.freedesktop.org
-
Christian König -
Daniel Vetter -
kernel test robot -
Maarten Lankhorst -
Matthew Auld -
Tvrtko Ursulin