Re: [Intel-gfx] [PATCH v5 01/63] drm/i915: Do not share hwsp across contexts any more, v6

[Maarten Lankhorst <maarten.lankhorst@linux.intel.com>]

Op 25-11-2020 om 11:44 schreef Chris Wilson:
> Quoting Maarten Lankhorst (2020-11-25 10:39:09)
>> -               } while (!__igt_timeout(end_time, NULL));
>> -               WRITE_ONCE(*(u32 *)tl->hwsp_seqno, 0xdeadbeef);
>> +               } while (!__igt_timeout(end_time, NULL) &&
>> +                        count < (PAGE_SIZE / TIMELINE_SEQNO_BYTES - 1) / 2);
> Which part of this is NOT PAPERING OVER THE EXACT ISSUE I nacked this
> patch over?
>
> This series is still a joke. Go back to the design until you do not
> break ABI.
> -Chris

Thanks for your feedback, I've taken a look at this code and I will remove this hunk.
There is no need to limit to only 1 iteration.

Setting TIMELINE_SEQNO_BYTES to 128 (so we wrap the hwsp every 32 times, or every 16
test loops because we start at a clear cacheline) shows the code works as intended.

[  100.455041] i915: Running intel_timeline_live_selftests/live_hwsp_read
[  100.962336] rcs0: simulated 114 wraps
[  101.466285] bcs0: simulated 114 wraps
[  101.968282] vcs0: simulated 127 wraps
[  102.470052] vecs0: simulated 134 wraps
[  102.524057] i915: Running intel_timeline_live_selftests/live_hwsp_rollover_kernel

_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/intel-gfx