[dm-devel] dm: track io errors per mapped device

[kj at orbekk.com]

On Fri, May 8, 2020, at 17:09, Mike Snitzer wrote:
> On Fri, May 08 2020 at  3:22pm -0400,
> kj at orbekk.com <kj at orbekk.com> wrote:
> 
> > On Thu, May 7, 2020, at 21:12, Chaitanya Kulkarni wrote:
> > > On 05/07/2020 04:06 PM, Kjetil Orbekk wrote:
> > > > +		if (tio->error)
> > > > +			atomic_inc(&md->ioerr_cnt);
> > > 
> > > Given that there are so many errors how would user know what
> > > kind of error is generated and how many times?
> > 
> > The intended use case is to provide an easy way to check if errors
> > have occurred at all, and then the user needs to investigate using
> > other means. I replied with more detail to Alasdair's email.
> 
> But most operations initiated by the user that fail will be felt by the
> upper layer that the user is interfacing with.
> 
> Only exception that springs to mind is dm-writecache's writeback that
> occurs after writes have already been acknowledged back to the upper
> layers -- in that case dm-writecache provides an error flag that is
> exposed via writecache_status.
> 
> Anyway, just not seeing why you need a upper-layer use-case agnostic
> flag to know an error occurred in DM.  That layers that interface with
> the DM device will have been notified of all errors.

It's used as a signal by a probing process which is not in the IO path itself.

> And why just for DM devices?  Why not all block devices (in which case
> DM would get this feature "for free")?

This sounds like a good idea to me. Looks like I could add this to disk_stats and expose it through the block/<device>/stats file. I'll try to see if this is feasible.