Re: [PATCH 1/2] pktcdvd: Fix pkt_setup_dev() error path

[Luis Chamberlain <mcgrof@kernel.org>]

On Sat, Apr 25, 2020 at 05:17:00PM +0800, Ming Lei wrote:
> On Fri, Apr 24, 2020 at 07:39:47PM -0600, Luis Chamberlain wrote:
> > So I hopped on a time machine to revise some old collateral due to
> > 523e1d399ce ("block: make gendisk hold a reference to its queue")
> > merged on v3.2 which added the conditional check for the disk->queue
> > before calling blk_put_queue() on release_disk(). I started wondering
> > *why* the conditional was added, but I checked the original patch and
> > I could not find discussion around it.
> > 
> > Tejun, do you call why you added that conditional on
> > 
> > if (disk->queue)
> >   blk_put_queue(disk->queue);
> > 
> > This patch however struck me as one I should highlight, since I'm
> > reviewing all this now and dealing with adding error paths on
> > add_disk(). Below some notes.
> 
> disk->queue is assigned by drivers, I guess that is why the check
> is needed, given the disk may be released in error path before driver
> assigns queue to it.
> 
> Also some driver may only allocate disk and not add disk, then not
> necessary to assign disk->queue, such as drivers/scsi/sg.c

Jeesh. Ugh. Yes I see, thanks this helps.

> > As we have it now drivers *do* call blk_cleanup_queue() on error paths
> > prior to add_disk(). An example today is on drivers/block/loop.c where
> > in loop_add(), if alloc_disk() fails we call  blk_cleanup_queue()
> > *but* this error path *never* called put_disk() as
> > drivers/block/pktcdvd.c did on error, and that is because it doesn't
> > need to as the last error-path-induced call was alloc_disk(). So it
> > doesn't need to free the disk as its not created on the error path of
> > loop_add().
> > 
> > This will of course change once we make add_disk() return int, and
> > capture errors, and it brings the question if we want to follow
> > similar strategy for other drivers, however note that blk_put_queue()
> > doesn't do everything blk_cleanup_queue() does, and in fact
> > blk_cleanup_queue() states it sets up "the appropriate flags" *and*
> > then calls blk_put_queue().
> > 
> > We'll have a a bit more collateral evolutions if we embrace the
> > strategy in this commit, for those drivers that wish to start taking
> > advantage of the error checks, but other then considering this, I
> > thought it would be good to think about the fact that *today* we call
> > blk_cleanup_queue() on error paths *without* the disk being yet
> > associated either. This, in spite of the fact that the way we designed
> 
> Some drivers may have only request queue, and not have disk, such as
> NVMe's admin queue, so I think blk_cleanup_queue() has to cover this
> case.

Alright, also useful, thanks.

> > the queue, it sits on top of the disk from a kobject perspective once
> > registered. Since we call blk_cleanup_queue() on error paths today --
> > without a disk parent being possible -- it means nothing on
> > blk_cleanup_queue() should not rely on it having a functional disk. Do
> > we want to keep it that way? If we keep the practice of drivers using
> 
> Yes, see the reason above.

Alright, the patch I replied to was a case where blk_queue_cleanup() was
removed due to a crash even though this driver both add_disk() and
assigned the queue before. Although this patch didn't come with a full
kernel splat and only:

Kernel BUG at 00000000e98fd882 [verbose debug info unavailable]

I can only guess that this was likely a double put of the queue, once
at blk_cleanup_queue() and another with the last put on disk_release().

I'll consider these things when extending the error paths, thanks for
the feedback.

  Luis