On Wed, Mar 22, 2017 at 11:58:17AM -0400, Keith Busch wrote:
> On Tue, Mar 21, 2017 at 11:03:59PM -0400, Jens Axboe wrote:
> > On 03/21/2017 10:14 PM, Ming Lei wrote:
> > > When iterating busy requests in timeout handler,
> > > if the STARTED flag of one request isn't set, that means
> > > the request is being processed in block layer or driver, and
> > > isn't submitted to hardware yet.
> > >
> > > In current implementation of blk_mq_check_expired(),
> > > if the request queue becomes dying, un-started requests are
> > > handled as being completed/freed immediately. This way is
> > > wrong, and can cause rq corruption or double allocation[1][2],
> > > when doing I/O and removing&resetting NVMe device at the sametime.
> >
> > I agree, completing it looks bogus. If the request is in a scheduler or
> > on a software queue, this won't end well at all. Looks like it was
> > introduced by this patch:
> >
> > commit eb130dbfc40eabcd4e10797310bda6b9f6dd7e76
> > Author: Keith Busch <keith.busch@xxxxxxxxx>
> > Date: Thu Jan 8 08:59:53 2015 -0700
> >
> > blk-mq: End unstarted requests on a dying queue
> >
> > Before that, we just ignored it. Keith?
>
> The above was intended for a stopped hctx on a dying queue such that
> there's nothing in flight to the driver. Nvme had been relying on this
> to end unstarted requests so we may progress when a controller dies.
So the brokenness started just from the begining.
>
> We've since obviated the need: we restart the hw queues to flush entered
> requests to failure, so we don't need that brokenness.
Looks the following commit need to be backported too if we port this patch.
commit 69d9a99c258eb1d6478fd9608a2070890797eed7
Author: Keith Busch <keith.busch@xxxxxxxxx>
Date: Wed Feb 24 09:15:56 2016 -0700
NVMe: Move error handling to failed reset handler
Thanks,
Ming
