Re: [PATCH] scsi_lib.c: continue after MEDIUM_ERROR

[Ric Wheeler <ric@emc.com>]

James Bottomley wrote:
> On Fri, 2007-02-02 at 14:42 +0000, Alan wrote:
>   
> > > The interesting point of this question is about the typically pattern of 
> > > IO errors. On a read, it is safe to assume that you will have issues 
> > > with some bounded numbers of adjacent sectors.
> > >       
> > Which in theory you can get by asking the drive for the real sector size
> > from the ATA7 info. (We ought to dig this out more as its relevant for
> > partition layout too).
> >     

Actually, my point is that damage typically impacts a cluster of disk 
sectors that are adjacent. Think of a drive that has junk on the platter 
or a some such thing - the contamination is likely to be localized.
> >     
> > > I really like the idea of being able to set this kind of policy on a per 
> > > drive instance since what you want here will change depending on what 
> > > your system requirements are, what the system is trying to do (i.e., 
> > > when trying to recover a failing but not dead yet disk, IO errors should 
> > > be as quick as possible and we should choose an IO scheduler that does 
> > > not combine IO's).
> > >       
> > That seems to be arguing for a bounded "live" time including retry run
> > time for a command. That's also more intuitive for real time work and for
> > end user setup. "Either work or fail within n seconds"
> >     
>
> Actually, then I think perhaps we use the allowed retries for this ...
>   
I really am not a big retry fan for most modern drives - the drive will 
try really, really hard to complete an IO for us and multiple retries 
can just slow down the higher level application from recovering.
> So you would fail a single sector and count it against the retries.
> When you've done this allowed retries times, you fail the rest of the
> request.
>
> James
>
>   
I think that we need to play with some of these possible solutions on 
some  real-world bad drives and see how they react. 

We should definitely talk more about this at the workshop ;-)

ric

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/