Re: Infinite retries reading the partition table

[Luben Tuikov <ltuikov@yahoo.com>]

--- Andrew Morton <akpm@osdl.org> wrote:
> On Thu, 30 Nov 2006 22:34:57 -0800 (PST)
> Luben Tuikov <ltuikov@yahoo.com> wrote:
> 
> > --- Andrew Morton <akpm@osdl.org> wrote:
> > > On Wed, 29 Nov 2006 17:22:48 -0800 (PST)
> > > Luben Tuikov <ltuikov@yahoo.com> wrote:
> > > 
> > > > Suppose reading sector 0 always reports an error,
> > > > sense key HARDWARE ERROR.
> > > > 
> > > > What I'm observing is that the request to read sector 0,
> > > > reading partition information, is retried forever, ad infinitum.
> > > > 
> > > > Does anyone have a patch to resolve this? (2.6.19-rc6)
> > > > 
> > > 
> > > Please send a backtrace so we can see where the offending loop occurs.
> > 
> > I posted a patch to linux-scsi
> 
> hm.  Does sending patches to linux-scsi get them applied?  It might, I
> don't know.

Good question -- I don't know either.

> > which resolves this issue:
> > http://marc.theaimsgroup.com/?l=linux-scsi&m=116485834119885&w=2
> 
> That looks like it prevents the IO error.  But why was an IO error causing
> an infinite loop?   What piece of code was initiating the retries?

Here is what happens: sector 0 is broken -- the device cannot read
the media at that location.  The device properly returns a certain
type of uncorrectable MEDIUM ERROR (ASC: UNRECOVERABLE READ ERR).

SCSI Core loops around its retries (which this patch fixes) and
eventually gives up and sends it for "completion".  This is what
happens when scsi_check_sense() returns NEEDS_RETRY to
scsi_decide_disposition() to scsi_softirq().  The first chunk
of the patch fixes this.

We end up in scsi_io_completion(), where good_bytes = 0, and
result = 0x08000002 (DRIVER SENSE and CHECK CONDITION).

This statement in scsi_io_completion() causes the infinite retry loop:
   if (scsi_end_request(cmd, 1, good_bytes, !!result) == NULL)
         return;
substitute to get: scsi_end_request(cmd, uptodate=1, uptodate bytes=0, retry=1)
Yeah, but it doesn't make sense to call scsi_end_request() with uptodate=1 and
uptodate bytes = 0.  This causes the infinite retry, since the code
tries to re-read the whole xfer size (0 bytes were uptodate and retry=1),
from the bad media.

That is, we want to set uptodate=1 iff there was at least 1 byte up to date.
Else if nothing was read, uptodate bytes = 0, then we should pass
uptodate = 0, uptodate_bytes = total xfer, to mean the whole xfer is
not uptodate; and retry iff there was no error. (This is the very bottom
of the function.)

... I know, I know, but that's what we've got.

See this commit 03aba2f79594ca94d159c8bab454de9bcc385b76 as well.

      Luben