drm/i915: Make sure engines are idle during GPU idling in LR mode
Subsystems (2)
PATCH 1/2 drm/i915: Make sure engines are idle during GPU idling in LR mode
We assume that the GPU is idle once receiving the seqno via the last request's user interrupt. In execlist mode the corresponding context completed interrupt can be delayed though and until this latter interrupt arrives we consider the request to be pending on the ELSP submit port. This can cause a problem during system suspend where this last request will be seen by the resume code as still pending. Such pending requests are normally replayed after a GPU reset, but during resume we reset both SW and HW tracking of the ring head/tail pointers, so replaying the pending request with its stale tale pointer will leave the ring in an inconsistent state. A subsequent request submission can lead then to the GPU executing from uninitialized area in the ring behind the above stale tail pointer.
Fix this by making sure any pending request on the ELSP port is completed before suspending. I used a polling wait since the completion time I measured was <1ms and since normally we only need to wait during system suspend. GPU idling during runtime suspend is scheduled with a delay (currently 50-100ms) after the retirement of the last request at which point the context completed interrupt must have arrived already.
The chance of this bug was increased by
commit 1c777c5d1dcdf8fa0223fcff35fb387b5bb9517a
Author: Imre Deak <[email protected]> Date: Wed Oct 12 17:46:37 2016 +0300
drm/i915/hsw: Fix GPU hang during resume from S3-devices state
but it could happen even without the explicit GPU reset, since we disable interrupts afterwards during the suspend sequence.
Cc: Chris Wilson <[email protected]>
Cc: Mika Kuoppala <[email protected]>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98470
--- drivers/gpu/drm/i915/i915_gem.c | 3 +++ drivers/gpu/drm/i915/intel_lrc.c | 12 ++++++++++++ drivers/gpu/drm/i915/intel_lrc.h | 1 + 3 files changed, 16 insertions(+)
drivers/gpu/drm/i915/i915_gem.c +3 -0
@@ -2766,6 +2766,9 @@ i915_gem_idle_work_handler(struct work_struct *work)if (dev_priv->gt.active_requests) goto out_unlock; +if (i915.enable_execlists) +intel_lr_wait_engines_idle (dev_priv); +for_each_engine (engine, dev_priv, id)i915_gem_batch_pool_fini (&engine->batch_pool);
drivers/gpu/drm/i915/intel_lrc.c +12 -0
@@ -522,6 +522,18 @@ static bool execlists_elsp_idle(struct intel_engine_cs *engine)return !engine->execlist_port[0 ].request; } +void intel_lr_wait_engines_idle (struct drm_i915_private *dev_priv) +{ +struct intel_engine_cs *engine; +enum intel_engine_id id; + +for_each_engine (engine, dev_priv, id) { +if (wait_for (execlists_elsp_idle (engine),10 )) +DRM_ERROR ("Timeout waiting for engine %s to idle\n" , + engine->name); + } +} +static bool execlists_elsp_ready (struct intel_engine_cs *engine) {int port;
drivers/gpu/drm/i915/intel_lrc.h +1 -0
@@ -87,6 +87,7 @@ void intel_lr_context_unpin(struct i915_gem_context *ctx,struct drm_i915_private ; +void intel_lr_wait_engines_idle (struct drm_i915_private *dev_priv);void intel_lr_context_resume (struct drm_i915_private *dev_priv);uint64_t intel_lr_context_descriptor (struct i915_gem_context *ctx,struct intel_engine_cs *engine);
Signature · 6 lines
2.5.0 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH 2/2 drm/i915: Add assert for no pending GPU requests during resume in LR mode
During resume we will reset the SW/HW tracking for each ring head/tail pointers and so are not prepared to replay any pending requests (as opposed to GPU reset time). Add an assert for this.
--- drivers/gpu/drm/i915/intel_lrc.c | 2 ++ 1 file changed, 2 insertions(+)
drivers/gpu/drm/i915/intel_lrc.c +2 -0
@@ -2158,6 +2158,8 @@ void intel_lr_context_resume(struct drm_i915_private *dev_priv)if (WARN_ON (IS_ERR (reg)))continue ; +WARN_ON (!execlists_elsp_idle (engine)); + reg +=LRC_STATE_PN *PAGE_SIZE /sizeof (*reg); reg[CTX_RING_HEAD +1 ] =0 ; reg[CTX_RING_TAIL +1 ] =0 ;
Signature · 6 lines
2.5.0 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [PATCH 2/2] drm/i915: Add assert for no pending GPU requests during resume in LR mode
Quoted 3 lines · Imre Deak
During resume we will reset the SW/HW tracking for each ring head/tail pointers and so are not prepared to replay any pending requests (as opposed to GPU reset time). Add an assert for this.
A bit belated. You could just put WARN_ON(dev_priv->gt.awake) in i915_gem_resume() to mirror the assertions that we are idle on suspend.
Quoted 13 lines
Cc Chris Wilson <[email protected]>Cc Mika Kuoppala <[email protected]>Signed-off-by Imre Deak <[email protected]>--- drivers/gpu/drm/i915/intel_lrc.c | 2 ++ 1 file changed, 2 insertions(+)drivers/gpu/drm/i915/intel_lrc.c +0
-0
drivers/gpu/drm/i915/intel_lrc.c void intel_lr_context_resume(struct drm_i915_private *dev_priv) · line 2161
Quoted 4 lines
drivers/gpu/drm/i915/intel_lrc.c +1
-0@@ -2158,6 +2158,8 @@ void intel_lr_context_resume(struct drm_i915_private *dev_priv)if (WARN_ON (IS_ERR (reg)))continue ; +WARN_ON (!execlists_elsp_idle (engine));
I don't think this check here documents the assumption of touching the context. What you want is this assertion in suspend.
-Chris
Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Patchwork
CI✗ Fi.CI.BAT: warning for series starting with [1/2] drm/i915: Make sure engines are idle during GPU idling in LR mode
In reply to Imre Deak== Series Details ==
== Series Details ==
Series: series starting with [1/2] drm/i915: Make sure engines are idle during GPU idling in LR mode
URL : https://patchwork.freedesktop.org/series/14789/
State : warning
== Summary ==
Series 14789v1 Series without cover letter
https://patchwork.freedesktop.org/api/1.0/series/14789/revisions/1/mbox/
Test drv_module_reload_basic:
pass -> DMESG-WARN (fi-skl-6770hq)
fi-bdw-5557u total:241 pass:226 dwarn:0 dfail:0 fail:0 skip:15 fi-bsw-n3050 total:241 pass:201 dwarn:0 dfail:0 fail:0 skip:40 fi-bxt-t5700 total:241 pass:213 dwarn:0 dfail:0 fail:0 skip:28 fi-byt-j1900 total:241 pass:213 dwarn:0 dfail:0 fail:0 skip:28 fi-byt-n2820 total:241 pass:209 dwarn:0 dfail:0 fail:0 skip:32 fi-hsw-4770 total:241 pass:221 dwarn:0 dfail:0 fail:0 skip:20 fi-hsw-4770r total:241 pass:221 dwarn:0 dfail:0 fail:0 skip:20 fi-ilk-650 total:241 pass:188 dwarn:0 dfail:0 fail:0 skip:53 fi-ivb-3520m total:241 pass:219 dwarn:0 dfail:0 fail:0 skip:22 fi-ivb-3770 total:241 pass:219 dwarn:0 dfail:0 fail:0 skip:22 fi-kbl-7200u total:241 pass:219 dwarn:0 dfail:0 fail:0 skip:22 fi-skl-6260u total:241 pass:227 dwarn:0 dfail:0 fail:0 skip:14 fi-skl-6700hq total:241 pass:220 dwarn:0 dfail:0 fail:0 skip:21 fi-skl-6700k total:241 pass:219 dwarn:1 dfail:0 fail:0 skip:21 fi-skl-6770hq total:241 pass:226 dwarn:1 dfail:0 fail:0 skip:14 fi-snb-2520m total:241 pass:209 dwarn:0 dfail:0 fail:0 skip:32 fi-snb-2600 total:241 pass:208 dwarn:0 dfail:0 fail:0 skip:33
8a1c897ee1d0d93b8e70991becb6f9f8488dba1b drm-intel-nightly: 2016y-11m-03d-13h-53m-22s UTC integration manifest 48b2b9f drm/i915: Add assert for no pending GPU requests during resume in LR mode
f38c913 drm/i915: Make sure engines are idle during GPU idling in LR mode
== Logs ==
For more details see: https://intel-gfx-ci.01.org/CI/Patchwork_2899/
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Quoted 62 lines · Imre Deak
We assume that the GPU is idle once receiving the seqno via the last request's user interrupt. In execlist mode the corresponding context completed interrupt can be delayed though and until this latter interrupt arrives we consider the request to be pending on the ELSP submit port. This can cause a problem during system suspend where this last request will be seen by the resume code as still pending. Such pending requests are normally replayed after a GPU reset, but during resume we reset both SW and HW tracking of the ring head/tail pointers, so replaying the pending request with its stale tale pointer will leave the ring in an inconsistent state. A subsequent request submission can lead then to the GPU executing from uninitialized area in the ring behind the above stale tail pointer.
Fix this by making sure any pending request on the ELSP port is completed before suspending. I used a polling wait since the completion time I measured was <1ms and since normally we only need to wait during system suspend. GPU idling during runtime suspend is scheduled with a delay (currently 50-100ms) after the retirement of the last request at which point the context completed interrupt must have arrived already.
The chance of this bug was increased by
commit
1c777c5d1dcdf8fa0223fcff35fb387b5bb9517aAuthor: Imre Deak <[email protected]> Date: Wed Oct 12 17:46:37 2016 +0300drm/i915/hsw: Fix GPU hang during resume from S3-devices state
but it could happen even without the explicit GPU reset, since we disable interrupts afterwards during the suspend sequence.
Cc: Chris Wilson <[email protected]>
Cc: Mika Kuoppala <[email protected]>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98470Signed-off-by Imre Deak <[email protected]>--- drivers/gpu/drm/i915/i915_gem.c | 3 +++ drivers/gpu/drm/i915/intel_lrc.c | 12 ++++++++++++ drivers/gpu/drm/i915/intel_lrc.h | 1 + 3 files changed, 16 insertions(+)drivers/gpu/drm/i915/i915_gem.c +3
-0@@ -2766,6 +2766,9 @@ i915_gem_idle_work_handler(struct work_struct *work)if (dev_priv->gt.active_requests) goto out_unlock; +if (i915.enable_execlists) +intel_lr_wait_engines_idle (dev_priv); +for_each_engine (engine, dev_priv, id)i915_gem_batch_pool_fini (&engine->batch_pool);drivers/gpu/drm/i915/intel_lrc.c +0
-0@@ -522,6 +522,18 @@ static bool execlists_elsp_idle(struct intel_engine_cs *engine)return !engine->execlist_port[0 ].request; }
drivers/gpu/drm/i915/intel_lrc.c static bool execlists_elsp_idle(struct intel_engine_cs *engine) · line 533
Quoted 10 lines
drivers/gpu/drm/i915/intel_lrc.c +9
-0+void intel_lr_wait_engines_idle (struct drm_i915_private *dev_priv) +{ +struct intel_engine_cs *engine; +enum intel_engine_id id; + +for_each_engine (engine, dev_priv, id) { +if (wait_for (execlists_elsp_idle (engine),10 )) +DRM_ERROR ("Timeout waiting for engine %s to idle\n" , + engine->name);
Just noticed engine names are currently like "render ring",etc, so you can drop the 'engine' from the message.
Quoted 12 lines
drivers/gpu/drm/i915/intel_lrc.c +3
-0+ } +} +static bool execlists_elsp_ready (struct intel_engine_cs *engine) {int port;drivers/gpu/drm/i915/intel_lrc.h +0
-0
drivers/gpu/drm/i915/intel_lrc.h void intel_lr_context_unpin(struct i915_gem_context *ctx, · line 93
Quoted 7 lines
drivers/gpu/drm/i915/intel_lrc.h +1
-0@@ -87,6 +87,7 @@ void intel_lr_context_unpin(struct i915_gem_context *ctx,struct drm_i915_private ; +void intel_lr_wait_engines_idle (struct drm_i915_private *dev_priv);void intel_lr_context_resume (struct drm_i915_private *dev_priv);uint64_t intel_lr_context_descriptor (struct i915_gem_context *ctx,struct intel_engine_cs *engine);
Regards,
Tvrtko
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Quoted 46 lines · Imre Deak
We assume that the GPU is idle once receiving the seqno via the last request's user interrupt. In execlist mode the corresponding context completed interrupt can be delayed though and until this latter interrupt arrives we consider the request to be pending on the ELSP submit port. This can cause a problem during system suspend where this last request will be seen by the resume code as still pending. Such pending requests are normally replayed after a GPU reset, but during resume we reset both SW and HW tracking of the ring head/tail pointers, so replaying the pending request with its stale tale pointer will leave the ring in an inconsistent state. A subsequent request submission can lead then to the GPU executing from uninitialized area in the ring behind the above stale tail pointer.
Fix this by making sure any pending request on the ELSP port is completed before suspending. I used a polling wait since the completion time I measured was <1ms and since normally we only need to wait during system suspend. GPU idling during runtime suspend is scheduled with a delay (currently 50-100ms) after the retirement of the last request at which point the context completed interrupt must have arrived already.
The chance of this bug was increased by
commit
1c777c5d1dcdf8fa0223fcff35fb387b5bb9517aAuthor: Imre Deak <[email protected]> Date: Wed Oct 12 17:46:37 2016 +0300drm/i915/hsw: Fix GPU hang during resume from S3-devices state
but it could happen even without the explicit GPU reset, since we disable interrupts afterwards during the suspend sequence.
Cc: Chris Wilson <[email protected]>
Cc: Mika Kuoppala <[email protected]>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98470Signed-off-by Imre Deak <[email protected]>--- drivers/gpu/drm/i915/i915_gem.c | 3 +++ drivers/gpu/drm/i915/intel_lrc.c | 12 ++++++++++++ drivers/gpu/drm/i915/intel_lrc.h | 1 + 3 files changed, 16 insertions(+)drivers/gpu/drm/i915/i915_gem.c +0
-0
drivers/gpu/drm/i915/i915_gem.c i915_gem_idle_work_handler(struct work_struct *work) · line 2770
Quoted 5 lines
drivers/gpu/drm/i915/i915_gem.c +2
-0@@ -2766,6 +2766,9 @@ i915_gem_idle_work_handler(struct work_struct *work)if (dev_priv->gt.active_requests) goto out_unlock; +if (i915.enable_execlists) +intel_lr_wait_engines_idle (dev_priv);
Idle work handler... So runtime suspend. Anyway this is not an ideal place for a stall under struct_mutex (even if 16x10us, it's the principle!).
Move this to before the first READ_ONCE(dev_priv->gt.active_requests); so we stall before taking the lock, and skip if any new requests arrive whilst waiting.
(Also i915.enable_execlists is forbidden. But meh)
static struct drm_i915_gem_request * execlists_active_port(struct intel_engine_cs *engine)
{
struct drm_i915_gem_request *request;
request = READ_ONCE(engine->execlist_port[1]);
if (request)
return request;
return READ_ONCE(engine->execlist_port[0]);
}/* Wait for execlists to settle, but bail if any new requests come in */
for_each_engine(engine, dev_priv, id) {
struct drm_i915_gem_request *request;
request = execlists_active_port(engine);
if (!request)
continue;
if (wait_for(execlists_active_port(engine) != request, 10))
DRM_ERROR("Timeout waiting for %s to idle\n", engine->name);
}Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Quoted 42 lines · Chris Wilson
Quoted 41 lines · Imre Deak
We assume that the GPU is idle once receiving the seqno via the last request's user interrupt. In execlist mode the corresponding context completed interrupt can be delayed though and until this latter interrupt arrives we consider the request to be pending on the ELSP submit port. This can cause a problem during system suspend where this last request will be seen by the resume code as still pending. Such pending requests are normally replayed after a GPU reset, but during resume we reset both SW and HW tracking of the ring head/tail pointers, so replaying the pending request with its stale tale pointer will leave the ring in an inconsistent state. A subsequent request submission can lead then to the GPU executing from uninitialized area in the ring behind the above stale tail pointer.
Fix this by making sure any pending request on the ELSP port is completed before suspending. I used a polling wait since the completion time I measured was <1ms and since normally we only need to wait during system suspend. GPU idling during runtime suspend is scheduled with a delay (currently 50-100ms) after the retirement of the last request at which point the context completed interrupt must have arrived already.
The chance of this bug was increased by
commit
1c777c5d1dcdf8fa0223fcff35fb387b5bb9517a
Author: Imre Deak <[email protected]>
Date: Wed Oct 12 17:46:37 2016 +0300drm/i915/hsw: Fix GPU hang during resume from S3-devices state
but it could happen even without the explicit GPU reset, since we disable interrupts afterwards during the suspend sequence.
Cc: Chris Wilson <[email protected]>
Cc: Mika Kuoppala <[email protected]>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98470Signed-off-by Imre Deak <[email protected]>---
drivers/gpu/drm/i915/i915_gem.c | 3 +++drivers/gpu/drm/i915/intel_lrc.c| 12 ++++++++++++drivers/gpu/drm/i915/intel_lrc.h| 1 +3 files changed, 16 insertions(+)
Quoted 14 lines
drivers/gpu/drm/i915/i915_gem.c i915_gem_idle_work_handler(struct work_struct *work) · line 2770
Quoted 10 lines
drivers/gpu/drm/i915/i915_gem.c +0
-0@@ -2766,6 +2766,9 @@ i915_gem_idle_work_handler(struct work_struct *work) if (dev_priv->gt.active_requests) goto out_unlock;drivers/gpu/drm/i915/i915_gem.c +2
-0+if (i915.enable_execlists) +intel_lr_wait_engines_idle (dev_priv);Idle work handler... So runtime suspend.
Anyway this is not an ideal place for a stall understruct_mutex(even if 16x10us, it's the principle!).
During runtime suspend this won't add any overhead since the context done interrupt happened already (unless there is a bug somewhere else).
Quoted 19 lines
Move this to before the first
READ_ONCE(dev_priv->gt.active_requests); so we stall before taking the lock, and skip if any new requests arrive whilst waiting.(Also i915.enable_execlists is forbidden. But meh)
static struct
drm_i915_gem_request*execlists_active_port(struct intel_engine_cs *engine){ struct drm_i915_gem_request *request; request = READ_ONCE(engine->execlist_port[1]); if (request) return request; return READ_ONCE(engine->execlist_port[0]); }/* Wait for execlists to settle, but bail if any new requests come in */
Quoted 10 lines
for_each_engine(engine, dev_priv, id) {
struct drm_i915_gem_request *request;
request = execlists_active_port(engine);
if (!request)
continue;
if (wait_for(execlists_active_port(engine) != request, 10))
DRM_ERROR("Timeout waiting for %s to idle\n", engine->name);
}Hm, but we still need to re-check and bail out if not idle with struct_mutex held, since gt.active_requests could go 0->1->0 before taking struct_mutex? I can rewrite things with that check added, using the above.
--Imre
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Quoted 58 lines · Imre Deak
Quoted 42 lines · Chris Wilson
Quoted 41 lines · Imre Deak
We assume that the GPU is idle once receiving the seqno via the last request's user interrupt. In execlist mode the corresponding context completed interrupt can be delayed though and until this latter interrupt arrives we consider the request to be pending on the ELSP submit port. This can cause a problem during system suspend where this last request will be seen by the resume code as still pending. Such pending requests are normally replayed after a GPU reset, but during resume we reset both SW and HW tracking of the ring head/tail pointers, so replaying the pending request with its stale tale pointer will leave the ring in an inconsistent state. A subsequent request submission can lead then to the GPU executing from uninitialized area in the ring behind the above stale tail pointer.
Fix this by making sure any pending request on the ELSP port is completed before suspending. I used a polling wait since the completion time I measured was <1ms and since normally we only need to wait during system suspend. GPU idling during runtime suspend is scheduled with a delay (currently 50-100ms) after the retirement of the last request at which point the context completed interrupt must have arrived already.
The chance of this bug was increased by
commit
1c777c5d1dcdf8fa0223fcff35fb387b5bb9517a
Author: Imre Deak <[email protected]>
Date: Wed Oct 12 17:46:37 2016 +0300drm/i915/hsw: Fix GPU hang during resume from S3-devices state
but it could happen even without the explicit GPU reset, since we disable interrupts afterwards during the suspend sequence.
Cc: Chris Wilson <[email protected]>
Cc: Mika Kuoppala <[email protected]>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98470Signed-off-by Imre Deak <[email protected]>---
drivers/gpu/drm/i915/i915_gem.c | 3 +++drivers/gpu/drm/i915/intel_lrc.c| 12 ++++++++++++drivers/gpu/drm/i915/intel_lrc.h| 1 +3 files changed, 16 insertions(+)Quoted 14 lines
drivers/gpu/drm/i915/i915_gem.c i915_gem_idle_work_handler(struct work_struct *work) · line 2770
Quoted 10 lines
drivers/gpu/drm/i915/i915_gem.c +0
-0@@ -2766,6 +2766,9 @@ i915_gem_idle_work_handler(struct work_struct *work) if (dev_priv->gt.active_requests) goto out_unlock;drivers/gpu/drm/i915/i915_gem.c +2
-0+if (i915.enable_execlists) +intel_lr_wait_engines_idle (dev_priv);Idle work handler... So runtime suspend.
Anyway this is not an ideal place for a stall understruct_mutex(even if 16x10us, it's the principle!).
Quoted 2 lines
During runtime suspend this won't add any overhead since the context done interrupt happened already (unless there is a bug somewhere else).
Where is that guaranteed? I thought we only serialised with the pm interrupts. Remember this happens before rpm suspend, since gem_idle_work_handler is responsible for dropping the GPU wakelock.
Quoted 26 lines
Quoted 26 lines
Move this to before the first
READ_ONCE(dev_priv->gt.active_requests); so we stall before taking the lock, and skip if any new requests arrive whilst waiting.(Also i915.enable_execlists is forbidden. But meh)
static struct
drm_i915_gem_request*execlists_active_port(struct intel_engine_cs *engine){ struct drm_i915_gem_request *request; request = READ_ONCE(engine->execlist_port[1]); if (request) return request; return READ_ONCE(engine->execlist_port[0]); }/* Wait for execlists to settle, but bail if any new requests come in */
for_each_engine(engine, dev_priv, id) { struct drm_i915_gem_request *request; request = execlists_active_port(engine); if (!request) continue;
Quoted 8 lines
Quoted 3 lines
if (wait_for(execlists_active_port(engine) != request, 10)) DRM_ERROR("Timeout waiting for %s to idle\n", engine->name); }Hm, but we still need to re-check and bail out if not idle with
struct_mutexheld, since gt.active_requests could go 0->1->0 before takingstruct_mutex? I can rewrite things with that check added, using the above.
Hmm, apparently we don't care ;) If the context-done interrupt is serialised with runtime suspend, then we don't need a wait here at all. On the system path there are no new requests and we are just flushing the idle worker.
But yes, for the sake of correctness do both an unlocked wait followed by a locked wait.
-Chris
Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Quoted 59 lines · Chris Wilson
Quoted 58 lines · Imre Deak
Quoted 56 lines · Chris Wilson
Quoted 49 lines · Imre Deak
We assume that the GPU is idle once receiving the seqno via the last request's user interrupt. In execlist mode the corresponding context completed interrupt can be delayed though and until this latter interrupt arrives we consider the request to be pending on the ELSP submit port. This can cause a problem during system suspend where this last request will be seen by the resume code as still pending. Such pending requests are normally replayed after a GPU reset, but during resume we reset both SW and HW tracking of the ring head/tail pointers, so replaying the pending request with its stale tale pointer will leave the ring in an inconsistent state. A subsequent request submission can lead then to the GPU executing from uninitialized area in the ring behind the above stale tail pointer.
Fix this by making sure any pending request on the ELSP port is completed before suspending. I used a polling wait since the completion time I measured was <1ms and since normally we only need to wait during system suspend. GPU idling during runtime suspend is scheduled with a delay (currently 50-100ms) after the retirement of the last request at which point the context completed interrupt must have arrived already.
The chance of this bug was increased by
commit
1c777c5d1dcdf8fa0223fcff35fb387b5bb9517a
Author: Imre Deak <[email protected]>
Date: Wed Oct 12 17:46:37 2016 +0300drm/i915/hsw: Fix GPU hang during resume from S3-devices state
but it could happen even without the explicit GPU reset, since we disable interrupts afterwards during the suspend sequence.
Cc: Chris Wilson <[email protected]>
Cc: Mika Kuoppala <[email protected]>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98470Signed-off-by Imre Deak <[email protected]>---
drivers/gpu/drm/i915/i915_gem.c | 3 +++drivers/gpu/drm/i915/intel_lrc.c| 12 ++++++++++++drivers/gpu/drm/i915/intel_lrc.h| 1 +3 files changed, 16 insertions(+)drivers/gpu/drm/i915/i915_gem.c +0
-0@@ -2766,6 +2766,9 @@ i915_gem_idle_work_handler(struct work_struct *work) if (dev_priv->gt.active_requests) goto out_unlock;drivers/gpu/drm/i915/i915_gem.c i915_gem_idle_work_handler(struct work_struct *work) · line 2770
Quoted 2 lines
drivers/gpu/drm/i915/i915_gem.c +2
-0+if (i915.enable_execlists) +intel_lr_wait_engines_idle (dev_priv);Idle work handler... So runtime suspend.
Anyway this is not an ideal place for a stall understruct_mutex(even if 16x10us, it's the principle!).
Quoted 6 lines
Quoted 2 lines
During runtime suspend this won't add any overhead since the context done interrupt happened already (unless there is a bug somewhere else).
Where is that guaranteed? I thought we only serialised with the pm interrupts. Remember this happens before rpm suspend, since
gem_idle_work_handleris responsible for dropping the GPU wakelock.
I meant that the 100msec after the last request signals completion and this handler is scheduled is normally enough for the context complete interrupt to get delivered. But yea, it's not a guarantee.
Quoted 35 lines
Quoted 26 lines
Quoted 26 lines
Move this to before the first
READ_ONCE(dev_priv->gt.active_requests); so we stall before taking the lock, and skip if any new requests arrive whilst waiting.(Also i915.enable_execlists is forbidden. But meh)
static struct
drm_i915_gem_request*execlists_active_port(struct intel_engine_cs *engine){ struct drm_i915_gem_request *request; request = READ_ONCE(engine->execlist_port[1]); if (request) return request; return READ_ONCE(engine->execlist_port[0]); }/* Wait for execlists to settle, but bail if any new requests come in */
for_each_engine(engine, dev_priv, id) { struct drm_i915_gem_request *request; request = execlists_active_port(engine); if (!request) continue;Quoted 8 lines
Quoted 3 lines
if (wait_for(execlists_active_port(engine) != request, 10)) DRM_ERROR("Timeout waiting for %s to idle\n", engine->name); }Hm, but we still need to re-check and bail out if not idle with
struct_mutexheld, since gt.active_requests could go 0->1->0 before takingstruct_mutex? I can rewrite things with that check added, using the above.
Quoted 7 lines
Hmm, apparently we don't care ;) If the context-done interrupt is serialised with runtime suspend, then we don't need a wait here at all. On the system path there are no new requests and we are just flushing the idle worker.
But yes, for the sake of correctness do both an unlocked wait followed by a locked wait.
Ok, will resend with this change. I used gt.active_requests to detect a new request during the unlocked wait, by that we can avoid polling in some cases.
--Imre
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Quoted 8 lines · Imre Deak
Quoted 3 lines · Chris Wilson
Where is that guaranteed? I thought we only serialised with the pm interrupts. Remember this happens before rpm suspend, since
gem_idle_work_handleris responsible for dropping the GPU wakelock.I meant that the 100msec after the last request signals completion and this handler is scheduled is normally enough for the context complete interrupt to get delivered. But yea, it's not a guarantee.
If only it was that deterministic! The idle_worker was scheduled 100ms after some retire_worker, just not necessarily the most recent. So it could be running exactly as active_requests -> 0 and so before the context-interrupt.
Anyway, it was a good find!
-Chris
Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Quoted 14 lines · Chris Wilson
Quoted 7 lines · Imre Deak
Quoted 5 lines · Chris Wilson
Where is that guaranteed? I thought we only serialised with the pm
interrupts. Remember this happens before rpm suspend, sincegem_idle_work_handleris responsible for dropping the GPU wakelock.Quoted 5 lines
I meant that the 100msec after the last request signals completion and
this handler is scheduled is normally enough for the context complete
interrupt to get delivered. But yea, it's not a guarantee.
Quoted 5 lines
If only it was that deterministic! The
idle_workerwas scheduled 100ms
after someretire_worker, just not necessarily the most recent. So it could be running exactly asactive_requests-> 0 and so before the context-interrupt.
Right, but we don't poll in that case, so there is no overhead.
Quoted 3 lines
Anyway, it was a good find!
-Chris
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Quoted 16 lines · Imre Deak
Quoted 15 lines · Chris Wilson
Quoted 13 lines · Imre Deak
Quoted 5 lines · Chris Wilson
Where is that guaranteed? I thought we only serialised with the pm
interrupts. Remember this happens before rpm suspend, sincegem_idle_work_handleris responsible for dropping the GPU wakelock.I meant that the 100msec after the last request signals completion
and
this handler is scheduled is normally enough for the context complete
interrupt to get delivered. But yea, it's not a guarantee.
Quoted 8 lines
Quoted 6 lines
If only it was that deterministic! The
idle_workerwas scheduled 100ms
after someretire_worker, just not necessarily the most recent. So it
could be running exactly asactive_requests-> 0 and so before the context-interrupt.Right, but we don't poll in that case, so there is no overhead.
Ok, there is a small window in the idle_worker after the unlocked poll and before taking the lock where a new request could be submitted and retired. In that case active_requests could be 0 after taking the lock and we'd have the poll overhead there.
We could detect this by the fact that there is a new idle_worker pending and bail out in that case. We shouldn't idle the GPU in that case anyway.
Quoted 7 lines
Quoted 3 lines
Anyway, it was a good find!
-Chris_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx