On Fri, Apr 26, 2019 at 05:11:14PM +0200, Christoph Hellwig wrote: > On Thu, Apr 25, 2019 at 09:00:31AM +0800, Ming Lei wrote: > > The issue is driver(NVMe) specific, the race window is just between > > between blk_cleanup_queue() and removing the ns from the controller namspace > > list in nvme_ns_remove() > > And I wouldn't be surprised if others have the same issue. SCSI hasn't such issue, and loop/null/virtio-blk doesn't too. > > > > > blk_mq_init_queue() does hold one refcount, and its counter-part is > > blk_cleanup_queue(). > > > > It is simply ugly to ask blk_mq_init_queue() to grab a refcnt for driver, > > then who is the counter-part for releasing the extra refcount? > > Well, the problem is exactly that blk_cleanup_queue drops the reference. > If move the blk_put_queue() call from the end of it to the callers the > callers can keep the reference as long as they need them, and we wouldn't > need an extra reference. It depends on driver. Still no difference between your suggestion and the way in this patch, given driver specific change is a must. Even it is more clean to hold the queue refocunt by drivers explicitly because we usually do the get/put pair in one place, instead that getting refcnt in one subsystem, and doing put in another place. I am going to drop this patch in V8 since my original plan is to fix block layer's race in this patchset. Thanks, Ming
