[Intel-gfx] [PATCH 32/34] drm/i915: Use HW semaphores for inter-engine synchronisation on gen8+
Chris Wilson
chris at chris-wilson.co.uk
Mon Jan 21 22:21:15 UTC 2019
Having introduced per-context seqno, we know have a means to identity
progress across the system without feel of rollback as befell the
global_seqno. That is we can program a MI_SEMAPHORE_WAIT operation in
advance of submission safe in the knowledge that our target seqno and
address is stable.
However, since we are telling the GPU to busy-spin on the target address
until it matches the signaling seqno, we only want to do so when we are
sure that busy-spin will be completed quickly. To achieve this we only
submit the request to HW once the signaler is itself executing (modulo
preemption causing us to wait longer), and we only do so for default and
above priority requests (so that idle priority tasks never themselves
hog the GPU waiting for others).
But what AB-BA deadlocks? If you remove B, there can be no deadlock...
The issue is that with a deep ELSP queue, we can queue up a pair of
AB-BA on different engines, thus forming a classic mutual exclusion
deadlock. We side-step that issue by restricting the queue depth to
avoid having multiple semaphores in flight and so we only ever take one
set of locks at a time.
Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
---
drivers/gpu/drm/i915/i915_request.c | 139 +++++++++++++++++++++-
drivers/gpu/drm/i915/i915_request.h | 1 +
drivers/gpu/drm/i915/i915_scheduler.c | 1 +
drivers/gpu/drm/i915/i915_scheduler.h | 1 +
drivers/gpu/drm/i915/i915_sw_fence.c | 4 +-
drivers/gpu/drm/i915/i915_sw_fence.h | 3 +
drivers/gpu/drm/i915/intel_gpu_commands.h | 5 +
drivers/gpu/drm/i915/intel_lrc.c | 13 +-
8 files changed, 163 insertions(+), 4 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_request.c b/drivers/gpu/drm/i915/i915_request.c
index 099c6f994b99..b7554a399c39 100644
--- a/drivers/gpu/drm/i915/i915_request.c
+++ b/drivers/gpu/drm/i915/i915_request.c
@@ -22,8 +22,9 @@
*
*/
-#include <linux/prefetch.h>
#include <linux/dma-fence-array.h>
+#include <linux/irq_work.h>
+#include <linux/prefetch.h>
#include <linux/sched.h>
#include <linux/sched/clock.h>
#include <linux/sched/signal.h>
@@ -331,6 +332,66 @@ void i915_request_retire_upto(struct i915_request *rq)
} while (tmp != rq);
}
+struct execute_cb {
+ struct list_head link;
+ struct irq_work work;
+ struct i915_sw_fence *fence;
+};
+
+static void irq_execute_cb(struct irq_work *wrk)
+{
+ struct execute_cb *cb = container_of(wrk, typeof(*cb), work);
+
+ i915_sw_fence_complete(cb->fence);
+ kfree(cb);
+}
+
+static void __notify_execute_cb(struct i915_request *rq)
+{
+ struct execute_cb *cb;
+
+ lockdep_assert_held(&rq->lock);
+
+ if (list_empty(&rq->execute_cb))
+ return;
+
+ list_for_each_entry(cb, &rq->execute_cb, link)
+ irq_work_queue(&cb->work);
+
+ INIT_LIST_HEAD(&rq->execute_cb);
+}
+
+static int
+i915_request_await_execution(struct i915_request *rq,
+ struct i915_request *signal,
+ gfp_t gfp)
+{
+ struct execute_cb *cb;
+ unsigned long flags;
+
+ if (test_bit(I915_FENCE_FLAG_ACTIVE, &signal->fence.flags))
+ return 0;
+
+ cb = kmalloc(sizeof(*cb), gfp);
+ if (!cb)
+ return -ENOMEM;
+
+ cb->fence = &rq->submit;
+ i915_sw_fence_await(cb->fence);
+ init_irq_work(&cb->work, irq_execute_cb);
+
+ spin_lock_irqsave(&signal->lock, flags);
+ if (test_bit(I915_FENCE_FLAG_ACTIVE, &signal->fence.flags)) {
+ i915_sw_fence_complete(cb->fence);
+ kfree(cb);
+ } else {
+ list_add_tail(&cb->link, &signal->execute_cb);
+ }
+ spin_unlock_irqrestore(&signal->lock, flags);
+
+ return 0;
+}
+
static void move_to_timeline(struct i915_request *request,
struct i915_timeline *timeline)
{
@@ -377,6 +438,7 @@ void __i915_request_submit(struct i915_request *request)
if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &request->fence.flags) &&
!intel_engine_enable_signaling(request))
intel_engine_queue_breadcrumbs(engine);
+ __notify_execute_cb(request);
spin_unlock(&request->lock);
engine->emit_fini_breadcrumb(request,
@@ -621,6 +683,7 @@ i915_request_alloc(struct intel_engine_cs *engine, struct i915_gem_context *ctx)
tl->fence_context, seqno);
INIT_LIST_HEAD(&rq->active_list);
+ INIT_LIST_HEAD(&rq->execute_cb);
rq->i915 = i915;
rq->engine = engine;
rq->gem_context = ctx;
@@ -693,6 +756,77 @@ i915_request_alloc(struct intel_engine_cs *engine, struct i915_gem_context *ctx)
return ERR_PTR(ret);
}
+static int
+emit_semaphore_wait(struct i915_request *to,
+ struct i915_request *from,
+ gfp_t gfp)
+{
+ u32 *cs;
+
+ GEM_BUG_ON(i915_timeline_is_global(from->timeline));
+ GEM_BUG_ON(!from->timeline->has_initial_breadcrumb);
+
+ /*
+ * If we know our signaling request has started, we know that it
+ * must, at least, have passed its initial breadcrumb and that its
+ * seqno can only increase, therefore any change in its breadcrumb
+ * must indicate completion. By using a "not equal to start" compare
+ * we avoid the murky issue of how to handle seqno wraparound in an
+ * async environment (short answer, we must stop the world whenever
+ * any context wraps!) as the likelihood of missing one request then
+ * seeing the same start value for a new request is 1 in 2^31, and
+ * even then we know that the new request has started and is in
+ * progress, so we are sure it will complete soon enough (not to
+ * worry about).
+ */
+ if (i915_request_started(from)) {
+ cs = intel_ring_begin(to, 4);
+ if (IS_ERR(cs))
+ return PTR_ERR(cs);
+
+ *cs++ = MI_SEMAPHORE_WAIT |
+ MI_SEMAPHORE_GLOBAL_GTT |
+ MI_SEMAPHORE_POLL |
+ MI_SEMAPHORE_SAD_NEQ_SDD;
+ *cs++ = from->fence.seqno - 1;
+ *cs++ = i915_timeline_seqno_address(from->timeline);
+ *cs++ = 0;
+
+ intel_ring_advance(to, cs);
+ } else {
+ int err;
+
+ err = i915_request_await_execution(to, from, gfp);
+ if (err)
+ return err;
+
+ cs = intel_ring_begin(to, 4);
+ if (IS_ERR(cs))
+ return PTR_ERR(cs);
+
+ /*
+ * Using greater-than-or-equal here means we have to worry
+ * about seqno wraparound. To side step that issue, we swap
+ * the timeline HWSP upon wrapping, so that everyone listening
+ * for the old (pre-wrap) values do not see the much smaller
+ * (post-wrap) values than they were expecting (and so wait
+ * forever).
+ */
+ *cs++ = MI_SEMAPHORE_WAIT |
+ MI_SEMAPHORE_GLOBAL_GTT |
+ MI_SEMAPHORE_POLL |
+ MI_SEMAPHORE_SAD_GTE_SDD;
+ *cs++ = from->fence.seqno;
+ *cs++ = i915_timeline_seqno_address(from->timeline);
+ *cs++ = 0;
+
+ intel_ring_advance(to, cs);
+ }
+
+ to->sched.semaphore = true;
+ return 0;
+}
+
static int
i915_request_await_request(struct i915_request *to, struct i915_request *from)
{
@@ -716,6 +850,9 @@ i915_request_await_request(struct i915_request *to, struct i915_request *from)
ret = i915_sw_fence_await_sw_fence_gfp(&to->submit,
&from->submit,
I915_FENCE_GFP);
+ } else if (HAS_EXECLISTS(to->i915) &&
+ to->gem_context->sched.priority >= I915_PRIORITY_NORMAL) {
+ ret = emit_semaphore_wait(to, from, I915_FENCE_GFP);
} else {
ret = i915_sw_fence_await_dma_fence(&to->submit,
&from->fence, 0,
diff --git a/drivers/gpu/drm/i915/i915_request.h b/drivers/gpu/drm/i915/i915_request.h
index 8f78ac97b8d6..d76be15ba0dd 100644
--- a/drivers/gpu/drm/i915/i915_request.h
+++ b/drivers/gpu/drm/i915/i915_request.h
@@ -109,6 +109,7 @@ struct i915_request {
*/
struct i915_sw_fence submit;
wait_queue_entry_t submitq;
+ struct list_head execute_cb;
/*
* A list of everyone we wait upon, and everyone who waits upon us.
diff --git a/drivers/gpu/drm/i915/i915_scheduler.c b/drivers/gpu/drm/i915/i915_scheduler.c
index fb5d953430e5..c41a36e48d12 100644
--- a/drivers/gpu/drm/i915/i915_scheduler.c
+++ b/drivers/gpu/drm/i915/i915_scheduler.c
@@ -29,6 +29,7 @@ void i915_sched_node_init(struct i915_sched_node *node)
INIT_LIST_HEAD(&node->waiters_list);
INIT_LIST_HEAD(&node->link);
node->attr.priority = I915_PRIORITY_INVALID;
+ node->semaphore = false;
}
static struct i915_dependency *
diff --git a/drivers/gpu/drm/i915/i915_scheduler.h b/drivers/gpu/drm/i915/i915_scheduler.h
index dbe9cb7ecd82..d764cf10536f 100644
--- a/drivers/gpu/drm/i915/i915_scheduler.h
+++ b/drivers/gpu/drm/i915/i915_scheduler.h
@@ -72,6 +72,7 @@ struct i915_sched_node {
struct list_head waiters_list; /* those after us, they depend upon us */
struct list_head link;
struct i915_sched_attr attr;
+ bool semaphore;
};
struct i915_dependency {
diff --git a/drivers/gpu/drm/i915/i915_sw_fence.c b/drivers/gpu/drm/i915/i915_sw_fence.c
index 7c58b049ecb5..8d1400d378d7 100644
--- a/drivers/gpu/drm/i915/i915_sw_fence.c
+++ b/drivers/gpu/drm/i915/i915_sw_fence.c
@@ -192,7 +192,7 @@ static void __i915_sw_fence_complete(struct i915_sw_fence *fence,
__i915_sw_fence_notify(fence, FENCE_FREE);
}
-static void i915_sw_fence_complete(struct i915_sw_fence *fence)
+void i915_sw_fence_complete(struct i915_sw_fence *fence)
{
debug_fence_assert(fence);
@@ -202,7 +202,7 @@ static void i915_sw_fence_complete(struct i915_sw_fence *fence)
__i915_sw_fence_complete(fence, NULL);
}
-static void i915_sw_fence_await(struct i915_sw_fence *fence)
+void i915_sw_fence_await(struct i915_sw_fence *fence)
{
debug_fence_assert(fence);
WARN_ON(atomic_inc_return(&fence->pending) <= 1);
diff --git a/drivers/gpu/drm/i915/i915_sw_fence.h b/drivers/gpu/drm/i915/i915_sw_fence.h
index 0e055ea0179f..6dec9e1d1102 100644
--- a/drivers/gpu/drm/i915/i915_sw_fence.h
+++ b/drivers/gpu/drm/i915/i915_sw_fence.h
@@ -79,6 +79,9 @@ int i915_sw_fence_await_reservation(struct i915_sw_fence *fence,
unsigned long timeout,
gfp_t gfp);
+void i915_sw_fence_await(struct i915_sw_fence *fence);
+void i915_sw_fence_complete(struct i915_sw_fence *fence);
+
static inline bool i915_sw_fence_signaled(const struct i915_sw_fence *fence)
{
return atomic_read(&fence->pending) <= 0;
diff --git a/drivers/gpu/drm/i915/intel_gpu_commands.h b/drivers/gpu/drm/i915/intel_gpu_commands.h
index 105e2a9e874a..bafad94f751b 100644
--- a/drivers/gpu/drm/i915/intel_gpu_commands.h
+++ b/drivers/gpu/drm/i915/intel_gpu_commands.h
@@ -106,7 +106,12 @@
#define MI_SEMAPHORE_TARGET(engine) ((engine)<<15)
#define MI_SEMAPHORE_WAIT MI_INSTR(0x1c, 2) /* GEN8+ */
#define MI_SEMAPHORE_POLL (1<<15)
+#define MI_SEMAPHORE_SAD_GT_SDD (0<<12)
#define MI_SEMAPHORE_SAD_GTE_SDD (1<<12)
+#define MI_SEMAPHORE_SAD_LT_SDD (2<<12)
+#define MI_SEMAPHORE_SAD_LTE_SDD (3<<12)
+#define MI_SEMAPHORE_SAD_EQ_SDD (4<<12)
+#define MI_SEMAPHORE_SAD_NEQ_SDD (5<<12)
#define MI_STORE_DWORD_IMM MI_INSTR(0x20, 1)
#define MI_STORE_DWORD_IMM_GEN4 MI_INSTR(0x20, 2)
#define MI_MEM_VIRTUAL (1 << 22) /* 945,g33,965 */
diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c
index 3d8fffa1b6dc..b59cfec1d5d4 100644
--- a/drivers/gpu/drm/i915/intel_lrc.c
+++ b/drivers/gpu/drm/i915/intel_lrc.c
@@ -332,7 +332,7 @@ __unwind_incomplete_requests(struct intel_engine_cs *engine)
* stream, so give it the equivalent small priority bump to prevent
* it being gazumped a second time by another peer.
*/
- if (!(prio & I915_PRIORITY_NEWCLIENT)) {
+ if (!(prio & I915_PRIORITY_NEWCLIENT) && i915_request_started(active)) {
prio |= I915_PRIORITY_NEWCLIENT;
active->sched.attr.priority = prio;
list_move_tail(&active->sched.link,
@@ -516,6 +516,17 @@ static bool can_merge_rq(const struct i915_request *prev,
{
GEM_BUG_ON(need_preempt(prev->engine, prev, rq_prio(next)));
+ /*
+ * To avoid AB-BA deadlocks, we simply restrict ourselves to only
+ * submitting one semaphore (think HW spinlock) to HW at a time. This
+ * prevents the execution callback on a later sempahore from being
+ * queued on another engine, so no cycle can be formed. Preemption
+ * rules should mean that if this semaphore is preempted, its
+ * dependency chain is preserved and suitably promoted via PI.
+ */
+ if (prev->sched.semaphore && !i915_request_started(prev))
+ return false;
+
if (!can_merge_ctx(prev->hw_context, next->hw_context))
return false;
--
2.20.1
More information about the Intel-gfx
mailing list