Re: [PATCH v2 1/2] nvme/host/pci: Fix a race in controller removal

["Singh, Balbir" <sblbir@amazon.com>]

On Mon, 2019-09-16 at 09:49 +0200, Christoph Hellwig wrote:
> On Fri, Sep 13, 2019 at 11:36:30PM +0000, Balbir Singh wrote:
> > This race is hard to hit in general, now that we
> > have the shutdown_lock in both nvme_reset_work() and
> > nvme_dev_disable()
> > 
> > The real issue is that after doing all the setup work
> > in nvme_reset_work(), when get another timeout (nvme_timeout()),
> > then we proceed to disable the controller. This causes
> > the reset work to only partially progress and then fail.
> > 
> > Depending on the progress made, we call into
> > nvme_remove_dead_ctrl(), which does another
> > nvme_dev_disable() freezing the block-mq queues.
> > 
> > I've noticed a race with udevd with udevd trying to re-read
> > the partition table, it ends up with the bd_mutex held and
> > it ends up waiting in blk_queue_enter(), since we froze
> > the queues in nvme_dev_disable(). nvme_kill_queues() calls
> > revalidate_disk() and ends up waiting on the bd_mutex
> > resulting in a deadlock.
> > 
> > Allow the hung tasks a chance by unfreezing the queues after
> > setting dying bits on the queue, then call revalidate_disk()
> > to update the disk size.
> > 
> > NOTE: I've seen this race when the controller does not
> > respond to IOs or abort requests, but responds to other
> > commands and even signals it's ready after its reset,
> > but still drops IO. I've tested this by emulating the
> > behaviour in the driver.
> > 
> > Signed-off-by: Balbir Singh <sblbir@amzn.com>
> > ---
> > 
> > Changelog:
> > - Rely on blk_set_queue_dying to do the wake_all()
> > 
> >  drivers/nvme/host/core.c | 8 +++++++-
> >  1 file changed, 7 insertions(+), 1 deletion(-)
> > 
> > diff --git a/drivers/nvme/host/core.c b/drivers/nvme/host/core.c
> > index b45f82d58be8..f6ddb58a7013 100644
> > --- a/drivers/nvme/host/core.c
> > +++ b/drivers/nvme/host/core.c
> > @@ -103,10 +103,16 @@ static void nvme_set_queue_dying(struct
> > nvme_ns *ns)
> >  	 */
> >  	if (!ns->disk || test_and_set_bit(NVME_NS_DEAD, &ns->flags))
> >  		return;
> > -	revalidate_disk(ns->disk);
> >  	blk_set_queue_dying(ns->queue);
> >  	/* Forcibly unquiesce queues to avoid blocking dispatch */
> >  	blk_mq_unquiesce_queue(ns->queue);
> > +	/*
> > +	 * revalidate_disk, after all pending IO is cleaned up
> > +	 * by blk_set_queue_dying, largely any races with blk parittion
> > +	 * reads that might come in after freezing the queues,
> > otherwise
> > +	 * we'll end up waiting up on bd_mutex, creating a deadlock.
> > +	 */
> > +	revalidate_disk(ns->disk);
> 
> The patch looks fine to me, but the comments looks a little strange.
> How do we trigger the partition scan?  Is someone opening the device
> again after we froze it?

The race itself is between the timeout for an aborted IO (iod->aborted
= 1) causing a reset and the subsequent disable controller, eventually
leading to IO errors on the device.

To reproduce the issue I run fio on the block device, when due to
timeout fio gets errors, it closes the device (/dev/nvme1), which
triggers udevd to reread the partitions.

But independent of this scenario, the race can be triggered by anything
scanning or reading the partition at the time of reset/disable
controller.

Balbir Singh.
_______________________________________________
Linux-nvme mailing list
Linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme