[Intel-gfx] [PATCH v3] drm/i915: Ensure associated VMAs are inactive when contexts are destroyed
Tvrtko Ursulin
tvrtko.ursulin at linux.intel.com
Thu Nov 19 04:28:29 PST 2015
Hi,
On 19/11/15 12:13, Daniel Vetter wrote:
> On Thu, Nov 19, 2015 at 09:42:17AM +0000, Tvrtko Ursulin wrote:
>>
>> On 19/11/15 09:17, Daniel Vetter wrote:
>>> On Wed, Nov 18, 2015 at 05:18:30PM +0000, Tvrtko Ursulin wrote:
>>>>
>>>> On 17/11/15 17:56, Daniel Vetter wrote:
>>>>> On Tue, Nov 17, 2015 at 05:24:01PM +0000, Tvrtko Ursulin wrote:
>>>>>>
>>>>>> On 17/11/15 17:08, Daniel Vetter wrote:
>>>>>>> On Tue, Nov 17, 2015 at 04:54:50PM +0000, Tvrtko Ursulin wrote:
>>>>>>>>
>>>>>>>> On 17/11/15 16:39, Daniel Vetter wrote:
>>>>>>>>> On Tue, Nov 17, 2015 at 04:27:12PM +0000, Tvrtko Ursulin wrote:
>>>>>>>>>> From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>>>>>>>>>
>>>>>>>>>> In the following commit:
>>>>>>>>>>
>>>>>>>>>> commit e9f24d5fb7cf3628b195b18ff3ac4e37937ceeae
>>>>>>>>>> Author: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>>>>>>>>> Date: Mon Oct 5 13:26:36 2015 +0100
>>>>>>>>>>
>>>>>>>>>> drm/i915: Clean up associated VMAs on context destruction
>>>>>>>>>>
>>>>>>>>>> I added a WARN_ON assertion that VM's active list must be empty
>>>>>>>>>> at the time of owning context is getting freed, but that turned
>>>>>>>>>> out to be a wrong assumption.
>>>>>>>>>>
>>>>>>>>>> Due ordering of operations in i915_gem_object_retire__read, where
>>>>>>>>>> contexts are unreferenced before VMAs are moved to the inactive
>>>>>>>>>> list, the described situation can in fact happen.
>>>>>>>>>>
>>>>>>>>>> It feels wrong to do things in such order so this fix makes sure
>>>>>>>>>> a reference to context is held until the move to inactive list
>>>>>>>>>> is completed.
>>>>>>>>>>
>>>>>>>>>> v2: Rather than hold a temporary context reference move the
>>>>>>>>>> request unreference to be the last operation. (Daniel Vetter)
>>>>>>>>>>
>>>>>>>>>> v3: Fix use after free. (Chris Wilson)
>>>>>>>>>>
>>>>>>>>>> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>>>>>>>>> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=92638
>>>>>>>>>> Cc: Michel Thierry <michel.thierry at intel.com>
>>>>>>>>>> Cc: Chris Wilson <chris at chris-wilson.co.uk>
>>>>>>>>>> Cc: Daniel Vetter <daniel.vetter at ffwll.ch>
>>>>>>>>>> ---
>>>>>>>>>> drivers/gpu/drm/i915/i915_gem.c | 33 ++++++++++++++++++---------------
>>>>>>>>>> 1 file changed, 18 insertions(+), 15 deletions(-)
>>>>>>>>>>
>>>>>>>>>> diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
>>>>>>>>>> index 98c83286ab68..094ac17a712d 100644
>>>>>>>>>> --- a/drivers/gpu/drm/i915/i915_gem.c
>>>>>>>>>> +++ b/drivers/gpu/drm/i915/i915_gem.c
>>>>>>>>>> @@ -2404,29 +2404,32 @@ i915_gem_object_retire__read(struct drm_i915_gem_object *obj, int ring)
>>>>>>>>>> RQ_BUG_ON(!(obj->active & (1 << ring)));
>>>>>>>>>>
>>>>>>>>>> list_del_init(&obj->ring_list[ring]);
>>>>>>>>>> - i915_gem_request_assign(&obj->last_read_req[ring], NULL);
>>>>>>>>>>
>>>>>>>>>> if (obj->last_write_req && obj->last_write_req->ring->id == ring)
>>>>>>>>>> i915_gem_object_retire__write(obj);
>>>>>>>>>>
>>>>>>>>>> obj->active &= ~(1 << ring);
>>>>>>>>>> - if (obj->active)
>>>>>>>>>> - return;
>>>>>>>>>
>>>>>>>>> if (obj->active) {
>>>>>>>>> i915_gem_request_assign(&obj->last_read_req[ring], NULL);
>>>>>>>>> return;
>>>>>>>>> }
>>>>>>>>>
>>>>>>>>> Would result in less churn in the code and drop the unecessary indent
>>>>>>>>> level. Also comment is missing as to why we need to do things in a
>>>>>>>>> specific order.
>>>>>>>>
>>>>>>>> Actually I think I changed my mind and that v1 is the way to go.
>>>>>>>>
>>>>>>>> Just re-ordering the code here still makes it possible for the context
>>>>>>>> destructor to run with VMAs on the active list I think.
>>>>>>>>
>>>>>>>> If we hold the context then it is 100% clear it is not possible.
>>>>>>>
>>>>>>> request_assign _is_ the function which adjust the refcounts for us, which
>>>>>>> means if we drop that reference too early then grabbing a temp reference
>>>>>>> is just papering over the real bug.
>>>>>>>
>>>>>>> Written out your patch looks something like
>>>>>>>
>>>>>>> a_reference(a);
>>>>>>> a_unreference(a);
>>>>>>>
>>>>>>> /* more cleanup code that should get run before a_unreference but isn't */
>>>>>>>
>>>>>>> a_unrefernce(a); /* for real this time */
>>>>>>>
>>>>>>> Unfortunately foo_assign is a new pattern and not well-established, so
>>>>>>> that connection isn't clear. Maybe we should rename it to
>>>>>>> foo_reference_assign to make it more obvious. Or just drop the pretense
>>>>>>> and open-code it since we unconditionally assign NULL as the new pointer
>>>>>>> value, and we know the current value of the pointer is non-NULL. So
>>>>>>> there's really no benefit to the helper here, it only obfuscates. And
>>>>>>> since that obfuscation tripped you up it's time to remove it ;-)
>>>>>>
>>>>>> Then foo_reference_unreference_assign. :)
>>>>>>
>>>>>> But seriously, I think it is more complicated that..
>>>>>>
>>>>>> The thing it trips over is that moving VMAs to inactive does not correspond
>>>>>> in time to request retirement. But in fact VMAs are moved to inactive only
>>>>>> when all requests associated with an object are done.
>>>>>>
>>>>>> This is the unintuitive thing I was working around. To make sure when
>>>>>> context destructor runs there are not active VMAs for that VM.
>>>>>>
>>>>>> I don't know how to guarantee that with what you propose. Perhaps I am
>>>>>> missing something?
>>>>>
>>>>> Ok, my example was slightly off, since we have 2 objects:
>>>>>
>>>>> b_reference(a->b);
>>>>> a_unreference(a); /* might unref a->b if it's the last reference */
>>>>>
>>>>> /* more cleanup code that should get run before a_unreference but isn't */
>>>>>
>>>>> b_unrefernce(a->b); /* for real this time */
>>>>>
>>>>> Holding the ref to a makes sure that b doesn't disappear. We rely on that
>>>>> in a fundamental way (a request really needs the ctx to stick around), and
>>>>> the bug really is that we drop the ref to a too early. That it's the
>>>>> releasing of a->b which is eventually blowing things up doesn't really
>>>>> matter.
>>>>>
>>>>> Btw would it be possible to have an igt for this? I should be possible to
>>>>> hit this with some varian of gem_unref_active_buffers.
>>>>
>>>> I was trying to do that today and it is proving to be a bit tricky.
>>>>
>>>> I need a blitter workload which will run for long enough for the retire
>>>> worker to run. So I'll try and build a bit bb tomorrow which will do that.
>>>>
>>>> Alternatively I did this:
>>>>
>>>> diff --git a/drivers/gpu/drm/i915/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/i915_gem_execbuffer.c
>>>> index 6ed7d63a0688..db51e4b42a20 100644
>>>> --- a/drivers/gpu/drm/i915/i915_gem_execbuffer.c
>>>> +++ b/drivers/gpu/drm/i915/i915_gem_execbuffer.c
>>>> @@ -699,6 +699,8 @@ i915_gem_execbuffer_reserve(struct intel_engine_cs *ring,
>>>> int retry;
>>>>
>>>> i915_gem_retire_requests_ring(ring);
>>>> + if (i915.enable_execlists)
>>>> + intel_execlists_retire_requests(ring);
>>>>
>>>> vm = list_first_entry(vmas, struct i915_vma, exec_list)->vm;
>>>>
>>>> And that enables me to trigger the WARN from my igt even with the current
>>>> shorter blitter workload (copy 64Mb).
>>>>
>>>> Going back to the original problem, how about something like this hunk for a fix?
>>>>
>>>> @@ -2413,19 +2416,36 @@ static void
>>>> i915_gem_object_retire__read(struct drm_i915_gem_object *obj, int ring)
>>>> {
>>>> struct i915_vma *vma;
>>>> + struct i915_hw_ppgtt *ppgtt;
>>>>
>>>> RQ_BUG_ON(obj->last_read_req[ring] == NULL);
>>>> RQ_BUG_ON(!(obj->active & (1 << ring)));
>>>>
>>>> list_del_init(&obj->ring_list[ring]);
>>>> +
>>>> + ppgtt = obj->last_read_req[ring]->ctx->ppgtt;
>>>> + if (ppgtt) {
>>>> + list_for_each_entry(vma, &obj->vma_list, vma_link) {
>>>> + if (vma->vm == &ppgtt->base &&
>>>> + !list_empty(&vma->mm_list)) {
>>>> + list_move_tail(&vma->mm_list,
>>>> + &vma->vm->inactive_list);
>>>> + }
>>>> + }
>>>> + }
>>>> +
>>>> i915_gem_request_assign(&obj->last_read_req[ring], NULL);
>>>>
>>>> This moves VMAs immediately to inactive as requests are retired and avoids
>>>> the problem with them staying on active for undefined amount of time.
>>>
>>> You can't put active objects onto the inactive list, i.e. the obj->active
>>
>> Hm it is inactive in this VM so who would care?
>
> The shrinker will fall over because of some long standing design mistake.
> Well, that's only the obj->active vs. vma->active problem.
>
> The real one is that you can have piles of concurrent read requests on a
> specific vma even, and retiring the first one of those doesn't make the
> vma inactive. So even with the active tracking design mistake fixed your
> suggestion wouldn't work.
But it is *last_read_req* and it is getting assigned to NULL meaning
there are no other ones at that point.
>> Perhaps then the fix is simply to remove the
>> WARN_ON(!list_empty(&ppgtt->base.active_list)) from the context destructor.
>>
>> If there are active VMAs at that point, they'll get cleaned up when they are
>> retired and there is no leak.
>>
>>> check is non-optional. And the if (ppgtt) case is abstraction violation.
>>> I really don't get why we can't just move the unref to the right place ...
>>
>> I don't see where.
>
> Please explain why the below change (which is the one I've been proposing,
> and which Chris suggested too) doesn't work:
>
>
> diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
> index 9552647a925d..d16b5ca042fa 100644
> --- a/drivers/gpu/drm/i915/i915_gem.c
> +++ b/drivers/gpu/drm/i915/i915_gem.c
> @@ -2375,14 +2375,15 @@ i915_gem_object_retire__read(struct drm_i915_gem_object *obj, int ring)
> RQ_BUG_ON(!(obj->active & (1 << ring)));
>
> list_del_init(&obj->ring_list[ring]);
> - i915_gem_request_assign(&obj->last_read_req[ring], NULL);
>
> if (obj->last_write_req && obj->last_write_req->ring->id == ring)
> i915_gem_object_retire__write(obj);
>
> obj->active &= ~(1 << ring);
> - if (obj->active)
> + if (obj->active) {
> + i915_gem_request_assign(&obj->last_read_req[ring], NULL);
> return;
Effectively my test code a bit annotated:
static void
test_retire_vma_not_inactive(int fd)
{
uint32_t ctx_id;
uint32_t src, dst, dst2;
uint32_t blit_bb, store_bb;
igt_require(HAS_BLT_RING(intel_get_drm_devid(fd)));
ctx_id = gem_context_create(fd);
igt_assert(ctx_id);
src = gem_create(fd, BO_SIZE);
igt_assert(src);
dst = gem_create(fd, BO_SIZE);
igt_assert(dst);
dst2 = gem_create(fd, 4096);
igt_assert(dst2);
blit_bb = blit(fd, dst, src, 0);
// The above blit takes long time to complete, src bo is last_read_req.
store_bb = store(fd, dst2, src, ctx_id);
// This is a very quick one puts the same src bo as last_read_req on
render ring.
gem_sync(fd, store_bb);
gem_sync(fd, dst2);
gem_close(fd, store_bb);
store_bb = store(fd, dst2, src, ctx_id);
gem_sync(fd, store_bb);
gem_sync(fd, dst2);
gem_close(fd, store_bb);
gem_close(fd, dst2);
// I was doing two in hope to trigger retirement but turns out the key
is in execlist retirement so never mind. Key is req is retired but VMA
not put on the inactive list.
gem_context_destroy(fd, ctx_id);
// Now context destruction runs with VMA on the active list -> WARN.
gem_sync(fd, blit_bb);
gem_close(fd, blit_bb);
gem_close(fd, src);
gem_close(fd, dst);
}
> + }
>
> /* Bump our place on the bound list to keep it roughly in LRU order
> * so that we don't steal from recently used but inactive objects
> @@ -2396,6 +2397,7 @@ i915_gem_object_retire__read(struct drm_i915_gem_object *obj, int ring)
> list_move_tail(&vma->mm_list, &vma->vm->inactive_list);
> }
>
> + /* Only unref once we're on the inactive list. */
> + i915_gem_request_assign(&obj->last_read_req[ring], NULL);
> i915_gem_request_assign(&obj->last_fenced_req, NULL);
> drm_gem_object_unreference(&obj->base);
> }
Regards,
Tvrtko
More information about the Intel-gfx
mailing list