[Intel-gfx] [PATCH] drm/i915: Wait one vblank before sending hotplug event to userspace

[Chris Wilson]

Quoting Mika Kahola (2018-12-04 09:46:39)
> Occasionally, we get the following error in our CI runs
> 
> [853.132830] Workqueue: events i915_hotplug_work_func [i915]
> [853.132844] RIP: 0010:drm_wait_one_vblank+0x19b/0x1b0
> [853.132852] Code: fe ff ff e8 b7 4e a6 ff 48 89 e6 4c 89 ff e8 6c 5f ab ff 45 85 ed 0f 85
> 15 ff ff ff 89 ee 48 c7 c7 e8 03 10 82 e8 b5 4b a6 ff <0f> 0b e9 00 ff ff ff 0f 1f 40 00 66
> 2e 0f 1f 84 00 00 00 00 00 8b
> [853.132859] RSP: 0018:ffffc9000146bca0 EFLAGS: 00010286
> [853.132866] RAX: 0000000000000000 RBX: ffff88849ef00000 RCX: 0000000000000000
> [853.132873] RDX: 0000000000000007 RSI: ffffffff820c6f58 RDI: 00000000ffffffff
> [853.132879] RBP: 0000000000000000 R08: 000000007ffc637a R09: 0000000000000000
> [853.132884] R10: 0000000000000000 R11: 0000000000000000 R12: 0000000000000000
> [853.132890] R13: 0000000000000000 R14: 000000000000d0c2 R15: ffff8884a491e680
> [853.132897] FS:  0000000000000000(0000) GS:ffff8884afe80000(0000) knlGS:0000000000000000
> [853.132904] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> [853.132910] CR2: 00007f63bf0df000 CR3: 0000000005210006 CR4: 0000000000760ee0
> [853.132916] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
> [853.132922] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
> [853.132927] PKRU: 55555554
> [853.132932] Call Trace:
> [853.132949]  ? wait_woken+0xa0/0xa0
> [853.133068]  intel_dp_retrain_link+0x130/0x190 [i915]
> [853.133176]  intel_ddi_hotplug+0x54/0x2e0 [i915]
> [853.133298]  i915_hotplug_work_func+0x1a9/0x240 [i915]
> [853.133324]  process_one_work+0x262/0x630
> [853.133349]  worker_thread+0x37/0x380
> [853.133365]  ? process_one_work+0x630/0x630
> [853.133373]  kthread+0x119/0x130
> [853.133383]  ? kthread_park+0x80/0x80
> [853.133400]  ret_from_fork+0x3a/0x50
> [853.133433] irq event stamp: 1426928
> 
> I suspect that this is caused by a racy condition when retraining the
> DisplayPort link. My proposal is to wait for one additional vblank
> event before we send out a hotplug event to userspace for reprobing.

Problem is just by waiting for the next vblank doesn't rule out hitting
the same race with another/delayed retraining. If you want
serialisation, please do add some -- and it may be sensible for the
hotplug to wait for the vblank after any ongoing work has finished.
-Chris