Re: [PATCH 5/6] drm/i915/gt: Serialize GRDOM access between multiple engine resets

From: Tvrtko Ursulin
Date: Thu Jun 30 2022 - 04:16:43 EST


On 30/06/2022 08:32, Mauro Carvalho Chehab wrote:
Em Wed, 29 Jun 2022 17:02:59 +0100
Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxxxxxxxx> escreveu:

On 29/06/2022 16:30, Mauro Carvalho Chehab wrote:
On Tue, 28 Jun 2022 16:49:23 +0100
Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxxxxxxxx> wrote:
.. which for me means a different patch 1, followed by patch 6 (moved
to be patch 2) would be ideal stable material.

Then we have the current patch 2 which is open/unknown (to me at least).

And the rest seem like optimisations which shouldn't be tagged as fixes.

Apart from patch 5 which should be cc: stable, but no fixes as agreed.

Could you please double check if what I am suggesting here is feasible
to implement and if it is just send those minimal patches out alone?

Tested and porting just those 3 patches are enough to fix the Broadwell
bug.

So, I submitted a v2 of this series with just those. They all need to
be backported to stable.

I would really like to give even a smaller fix a try. Something like, although not even compile tested:

commit 4d5e94aef164772f4d85b3b4c1a46eac9a2bd680
Author: Chris Wilson <chris.p.wilson@xxxxxxxxx>
Date: Wed Jun 29 16:25:24 2022 +0100

drm/i915/gt: Serialize TLB invalidates with GT resets
Avoid trying to invalidate the TLB in the middle of performing an
engine reset, as this may result in the reset timing out. Currently,
the TLB invalidate is only serialised by its own mutex, forgoing the
uncore lock, but we can take the uncore->lock as well to serialise
the mmio access, thereby serialising with the GDRST.
Tested on a NUC5i7RYB, BIOS RYBDWi35.86A.0380.2019.0517.1530 with
i915 selftest/hangcheck.
Cc: stable@xxxxxxxxxxxxxxx
Fixes: 7938d61591d3 ("drm/i915: Flush TLBs before releasing backing store")
Reported-by: Mauro Carvalho Chehab <mchehab@xxxxxxxxxx>
Tested-by: Mauro Carvalho Chehab <mchehab@xxxxxxxxxx>
Reviewed-by: Mauro Carvalho Chehab <mchehab@xxxxxxxxxx>
Signed-off-by: Chris Wilson <chris.p.wilson@xxxxxxxxx>
Cc: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxxxxxxxx>
Acked-by: Thomas Hellström <thomas.hellstrom@xxxxxxxxxxxxxxx>
Reviewed-by: Andi Shyti <andi.shyti@xxxxxxxxx>
Signed-off-by: Mauro Carvalho Chehab <mchehab@xxxxxxxxxx>
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx>

diff --git a/drivers/gpu/drm/i915/gt/intel_gt.c b/drivers/gpu/drm/i915/gt/intel_gt.c
index 8da3314bb6bf..aaadd0b02043 100644
--- a/drivers/gpu/drm/i915/gt/intel_gt.c
+++ b/drivers/gpu/drm/i915/gt/intel_gt.c
@@ -952,7 +952,23 @@ void intel_gt_invalidate_tlbs(struct intel_gt *gt)
mutex_lock(&gt->tlb_invalidate_lock);
intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL);
+ spin_lock_irq(&uncore->lock); /* serialise invalidate with GT reset */
+
+ for_each_engine(engine, gt, id) {
+ struct reg_and_bit rb;
+
+ rb = get_reg_and_bit(engine, regs == gen8_regs, regs, num);
+ if (!i915_mmio_reg_offset(rb.reg))
+ continue;
+
+ intel_uncore_write_fw(uncore, rb.reg, rb.bit);
+ }
+
+ spin_unlock_irq(&uncore->lock);
+
for_each_engine(engine, gt, id) {
+ struct reg_and_bit rb;
+
/*
* HW architecture suggest typical invalidation time at 40us,
* with pessimistic cases up to 100us and a recommendation to
@@ -960,13 +976,11 @@ void intel_gt_invalidate_tlbs(struct intel_gt *gt)
*/
const unsigned int timeout_us = 100;
const unsigned int timeout_ms = 4;
- struct reg_and_bit rb;
rb = get_reg_and_bit(engine, regs == gen8_regs, regs, num);
if (!i915_mmio_reg_offset(rb.reg))
continue;
- intel_uncore_write_fw(uncore, rb.reg, rb.bit);
if (__intel_wait_for_register_fw(uncore,
rb.reg, rb.bit, 0,
timeout_us, timeout_ms,


This won't work, as it is not serializing TLB cache invalidation with
i915 resets. Besides that, this is more or less merging patches 1 and 3,

Could you explain why you think it is not doing exactly that? In both versions end result is TLB flush requests are under the uncore lock and waits are outside it.

placing patches with different rationales altogether. Upstream rule is
to have one logical change per patch.

I don't think it applies in this case. It is simply splitting into two loops so lock can be held across all mmio writes. I think of it this way - what is the rationale for sending only the first patch to stable? What does it _fix_ on it's own?

If this works it would be least painful to backport. The other improvements can then be devoid of the fixes tag.

From backport PoV, it wouldn't make any difference applying one patch
or two. See, intel_gt_invalidate_tlbs() function doesn't exist before
changeset 7938d61591d3 ("drm/i915: Flush TLBs before releasing backing store"),
so, it shouldn't have merge conflicts while backporting it, maybe except
if some functions it calls (or parameters) have changed. On such case,
the backport fix should be trivial, and the end result of backporting
one folded patch or two would be the same.

Yes a lot of things changed. Not least engine and GT pm code. Note that TLB flushing was backported all the way to 4.4 so any hunk you don't strictly need can and will bite you. I have attached a tarball of patches for you to explore. :)
Regards,

Tvrtko

If any conflict happens, I can help doing the backports.

I still think that other TLB patches are needed/desired upstream, but
I'll submit them on a separate series. Let's fix the regression first ;-)

Yep, that's exactly right.

Regards,

Tvrtko

Attachment: tlbflush-220114-patches.tar.gz
Description: application/gzip