Re: [PATCH] drm/nouveau: idle the channel before destroying an engine object via abi16

[Alex <xnderv@yandex.com>]

On the Sashiko AI review of this patch:

> Can calling nouveau_channel_idle(obj->chan->chan) here race with jobs
> running on obj->chan->sched when VM_BIND is enabled?

Thanks, the race is real.

Legacy PUSHBUF submissions take cli->mutex (nouveau_abi16_get()), so they
are serialized with nouveau_abi16_ioctl_del(), which is why the Chromium
case was fixed by the plain nouveau_channel_idle().  With VM_BIND the
EXEC jobs are pushed from the scheduler workqueue without cli->mutex, so
emitting a fence from the ioctl could indeed race with
nouveau_exec_job_run(), and jobs still queued in the entity would not be
covered by the fence at all.

v2 (sent as a reply to this thread) waits for the channel's scheduler to
drain first -- the same job-list wait nouveau_sched_fini() already does,
factored out as nouveau_sched_wait_idle() -- and then idles the channel.
New EXEC submissions cannot sneak in meanwhile because the EXEC ioctl
also holds cli->mutex.

Alex