Doing TLB invalidation cause performance regressions, like: [424.370996] i915 0000:00:02.0: [drm] *ERROR* rcs0 TLB invalidation did not complete in 4ms!
As reported at: https://gitlab.freedesktop.org/drm/intel/-/issues/6424
as this is an expensive operation. So, reduce the need of it by: - checking if the engine is awake; - checking if the engine is not wedged; - batching operations.
Additionally, add a workaround for a known hardware issue on some GPUs.
In order to double-check that this series won't be introducing any regressions, I used this new IGT test:
https://patchwork.freedesktop.org/patch/495684/?series=106757&rev=1
Checking the results for 3 different patchsets, on Broadwell:
1) On the top of drm-tip (2022y-07m-14d-08h-35m-36) - e. g. with TLB invalidation and serialization patches:
$ sudo build/tests/gem_exec_tlb|grep Subtest Subtest close-clear: SUCCESS (10.490s) Subtest madv-clear: SUCCESS (10.484s) Subtest u-unmap-clear: SUCCESS (10.527s) Subtest u-shrink-clear: SUCCESS (10.506s) Subtest close-dumb: SUCCESS (10.165s) Subtest madv-dumb: SUCCESS (10.177s) Subtest u-unmap-dumb: SUCCESS (10.172s) Subtest u-shrink-dumb: SUCCESS (10.172s)
2) With the new version of the batch TLB invalidation patches from this series:
$ sudo build/tests/gem_exec_tlb|grep Subtest Subtest close-clear: SUCCESS (10.483s) Subtest madv-clear: SUCCESS (10.495s) Subtest u-unmap-clear: SUCCESS (10.545s) Subtest u-shrink-clear: SUCCESS (10.508s) Subtest close-dumb: SUCCESS (10.172s) Subtest madv-dumb: SUCCESS (10.169s) Subtest u-unmap-dumb: SUCCESS (10.174s) Subtest u-shrink-dumb: SUCCESS (10.176s)
3) Changing the TLB invalidation routine to do nothing[1]:
$ sudo ~/freedesktop-igt/build/tests/gem_exec_tlb|grep Subtest (gem_exec_tlb:1958) CRITICAL: Test assertion failure function check_bo, file ../tests/i915/gem_exec_tlb.c:384: (gem_exec_tlb:1958) CRITICAL: Failed assertion: !sq (gem_exec_tlb:1958) CRITICAL: Found deadbeef in a new (clear) buffer after 3 tries! (gem_exec_tlb:1956) CRITICAL: Test assertion failure function check_bo, file ../tests/i915/gem_exec_tlb.c:384: (gem_exec_tlb:1956) CRITICAL: Failed assertion: !sq (gem_exec_tlb:1956) CRITICAL: Found deadbeef in a new (clear) buffer after 89 tries! (gem_exec_tlb:1957) CRITICAL: Test assertion failure function check_bo, file ../tests/i915/gem_exec_tlb.c:384: (gem_exec_tlb:1957) CRITICAL: Failed assertion: !sq (gem_exec_tlb:1957) CRITICAL: Found deadbeef in a new (clear) buffer after 256 tries! (gem_exec_tlb:1960) CRITICAL: Test assertion failure function check_bo, file ../tests/i915/gem_exec_tlb.c:384: (gem_exec_tlb:1960) CRITICAL: Failed assertion: !sq (gem_exec_tlb:1960) CRITICAL: Found deadbeef in a new (clear) buffer after 845 tries! (gem_exec_tlb:1961) CRITICAL: Test assertion failure function check_bo, file ../tests/i915/gem_exec_tlb.c:384: (gem_exec_tlb:1961) CRITICAL: Failed assertion: !sq (gem_exec_tlb:1961) CRITICAL: Found deadbeef in a new (clear) buffer after 1138 tries! (gem_exec_tlb:1954) CRITICAL: Test assertion failure function check_bo, file ../tests/i915/gem_exec_tlb.c:384: (gem_exec_tlb:1954) CRITICAL: Failed assertion: !sq (gem_exec_tlb:1954) CRITICAL: Found deadbeef in a new (clear) buffer after 1359 tries! (gem_exec_tlb:1955) CRITICAL: Test assertion failure function check_bo, file ../tests/i915/gem_exec_tlb.c:384: (gem_exec_tlb:1955) CRITICAL: Failed assertion: !sq (gem_exec_tlb:1955) CRITICAL: Found deadbeef in a new (clear) buffer after 1794 tries! (gem_exec_tlb:1959) CRITICAL: Test assertion failure function check_bo, file ../tests/i915/gem_exec_tlb.c:384: (gem_exec_tlb:1959) CRITICAL: Failed assertion: !sq (gem_exec_tlb:1959) CRITICAL: Found deadbeef in a new (clear) buffer after 2139 tries! Dynamic subtest smem0 failed. **** DEBUG **** (gem_exec_tlb:1944) DEBUG: 2M hole:200000 contains poison:6b6b6b6b (gem_exec_tlb:1944) DEBUG: Running writer for 200000 at 300000 on bcs0 (gem_exec_tlb:1944) DEBUG: Closing hole:200000 on rcs0, sample:deadbeef (gem_exec_tlb:1944) DEBUG: Rechecking hole:200000, sample:6b6b6b6b **** END **** Subtest close-clear: FAIL (10.434s) Subtest madv-clear: SUCCESS (10.479s) Subtest u-unmap-clear: SUCCESS (10.512s)
In summary, the test does properly detect fail when TLB cache invalidation doesn't happen, as shown at result (3). It also shows that both current drm-tip and drm-tip with this series applied don't have TLB invalidation cache issues.
[1] I applied this patch on the top of drm-tip:
diff --git a/drivers/gpu/drm/i915/gt/intel_gt.c b/drivers/gpu/drm/i915/gt/intel_gt.c index 68c2b0d8f187..0aefcd7be5e9 100644 --- a/drivers/gpu/drm/i915/gt/intel_gt.c +++ b/drivers/gpu/drm/i915/gt/intel_gt.c @@ -930,0 +931,3 @@ void intel_gt_invalidate_tlbs(struct intel_gt *gt) + // HACK: don't do TLB invalidations!!! + return; +
Regards, Mauro
Chris Wilson (4): drm/i915/gt: Ignore TLB invalidations on idle engines drm/i915/gt: Invalidate TLB of the OA unit at TLB invalidations drm/i915/gt: Skip TLB invalidations once wedged drm/i915/gt: Batch TLB invalidations
Mauro Carvalho Chehab (2): drm/i915/gt: document with_intel_gt_pm_if_awake() drm/i915/gt: describe the new tlb parameter at i915_vma_resource
.../gpu/drm/i915/gem/i915_gem_object_types.h | 3 +- drivers/gpu/drm/i915/gem/i915_gem_pages.c | 25 +++--- drivers/gpu/drm/i915/gt/intel_gt.c | 77 +++++++++++++++---- drivers/gpu/drm/i915/gt/intel_gt.h | 12 ++- drivers/gpu/drm/i915/gt/intel_gt_pm.h | 11 +++ drivers/gpu/drm/i915/gt/intel_gt_types.h | 18 ++++- drivers/gpu/drm/i915/gt/intel_ppgtt.c | 8 +- drivers/gpu/drm/i915/i915_vma.c | 33 ++++++-- drivers/gpu/drm/i915/i915_vma.h | 1 + drivers/gpu/drm/i915/i915_vma_resource.c | 9 ++- drivers/gpu/drm/i915/i915_vma_resource.h | 6 +- 11 files changed, 163 insertions(+), 40 deletions(-)