Re: [PATCH] aoe: Do not free a device while its work is queued or running

[Jens Axboe <axboe@kernel.dk>]

On 10/29/26 9:45 AM, Bart Van Assche wrote:
> Serialize flush() with a mutex because dropping devlist_lock during pass
> one otherwise allows a concurrent flush() call to enter pass two and free
> a device that is still being torn down.
> 
> Set DEVFL_TKILL under d->lock before unlocking and calling
> aoedev_downdev() because commit 430380b4637a ("block: aoe: Fix kernel
> crash due to atomic sleep when exiting") dropped d->lock and devlist_lock
> before setting DEVFL_TKILL, allowing d->work to be re-queued while
> aoedev_downdev() sleeps.
> 
> Guard queue_work(aoe_wq, &d->work) with a DEVFL_TKILL check under d->lock
> in rexmit_timer(), ataid_complete(), and aoeblk_gdalloc() so that timers,
> ATA ID completions, and gendisk allocation failures cannot re-queue
> d->work once teardown has started.
> 
> Cancel d->work synchronously in flush() before aoedev_downdev() and in
> freedev() after timer_delete_sync() so that pending work cannot run
> concurrently with device teardown or outlive the freed struct aoedev.
> 
> Check DEVFL_TKILL under d->lock in aoecmd_sleepwork() and aoeblk_gdalloc()
> instead of triggering WARN_ON() or setting DEVFL_UP on a device that is
> being killed.

Thanks Bart, I've replaced the previous patch with this one.

-- 
Jens Axboe