CI has been sporadically reporting the following issue triggered by
igt@i915_selftest@live@hangcheck on ADL-P and similar machines:
<6> [414.049203] i915: Running intel_hangcheck_live_selftests/igt_reset_evict_fence
...
<6> [414.068804] i915 0000:00:02.0: [drm] GT0: GUC: submission enabled
<6> [414.068812] i915 0000:00:02.0: [drm] GT0: GUC: SLPC enabled
<3> [414.070354] Unable to pin Y-tiled fence; err:-4
<3> [414.071282] i915_vma_revoke_fence:301 GEM_BUG_ON(!i915_active_is_idle(&fence->active))
...
<4>[ 609.603992] ------------[ cut here ]------------
<2>[ 609.603995] kernel BUG at drivers/gpu/drm/i915/gt/intel_ggtt_fencing.c:301!
<4>[ 609.604003] invalid opcode: 0000 [#1] PREEMPT SMP NOPTI
<4>[ 609.604006] CPU: 0 PID: 268 Comm: kworker/u64:3 Tainted: G U W 6.9.0-CI_DRM_14785-g1ba62f8cea9c+ #1
<4>[ 609.604008] Hardware name: Intel Corporation Alder Lake Client Platform/AlderLake-P DDR4 RVP, BIOS RPLPFWI1.R00.4035.A00.2301200723 01/20/2023
<4>[ 609.604010] Workqueue: i915 __i915_gem_free_work [i915]
<4>[ 609.604149] RIP: 0010:i915_vma_revoke_fence+0x187/0x1f0 [i915]
...
<4>[ 609.604271] Call Trace:
<4>[ 609.604273] <TASK>
...
<4>[ 609.604716] __i915_vma_evict+0x2e9/0x550 [i915]
<4>[ 609.604852] __i915_vma_unbind+0x7c/0x160 [i915]
<4>[ 609.604977] force_unbind+0x24/0xa0 [i915]
<4>[ 609.605098] i915_vma_destroy+0x2f/0xa0 [i915]
<4>[ 609.605210] __i915_gem_object_pages_fini+0x51/0x2f0 [i915]
<4>[ 609.605330] __i915_gem_free_objects.isra.0+0x6a/0xc0 [i915]
<4>[ 609.605440] process_scheduled_works+0x351/0x690
...
In the past, there were similar failures reported by CI from other IGT
tests, observed on other platforms.
Before commit 63baf4f3d587 ("drm/i915/gt: Only wait for GPU activity
before unbinding a GGTT fence"), i915_vma_revoke_fence() was waiting for
idleness of vma->active via fence_update(). That commit introduced
vma->fence->active in order for the fence_update() to be able to wait
selectively on that one instead of vma->active since only idleness of
fence registers was needed. But then, another commit 0d86ee35097a
("drm/i915/gt: Make fence revocation unequivocal") replaced the call to
fence_update() in i915_vma_revoke_fence() with only fence_write(), and
also added that GEM_BUG_ON(!i915_active_is_idle(&fence->active)) in front.
No justification was provided on why we might then expect idleness of
vma->fence->active without first waiting on it.
The issue can be potentially caused by a race among revocation of fence
registers on one side and sequential execution of signal callbacks invoked
on completion of a request that was using them on the other, still
processed in parallel to revocation of those fence registers. Fix it by
waiting for idleness of vma->fence->active in i915_vma_revoke_fence().
Fixes: 0d86ee35097a ("drm/i915/gt: Make fence revocation unequivocal")
Closes: https://gitlab.freedesktop.org/drm/intel/issues/10021
Signed-off-by: Janusz Krzysztofik <janusz.krzysztofik@xxxxxxxxxxxxxxx>
Cc: stable@xxxxxxxxxxxxxxx # v5.8+
---
drivers/gpu/drm/i915/gt/intel_ggtt_fencing.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/drivers/gpu/drm/i915/gt/intel_ggtt_fencing.c b/drivers/gpu/drm/i915/gt/intel_ggtt_fencing.c
index 40371b8a9bbbd..93bc1cc1ee7e6 100644
--- a/drivers/gpu/drm/i915/gt/intel_ggtt_fencing.c
+++ b/drivers/gpu/drm/i915/gt/intel_ggtt_fencing.c
@@ -298,6 +298,7 @@ void i915_vma_revoke_fence(struct i915_vma *vma)
return;
GEM_BUG_ON(fence->vma != vma);
+ i915_active_wait(&fence->active);
GEM_BUG_ON(!i915_active_is_idle(&fence->active));
GEM_BUG_ON(atomic_read(&fence->pin_count));
--
2.45.1
