[3544] in linux-scsi channel archive
Re: SMP 2.1.90-pre3 SCSI kernel panic
daemon@ATHENA.MIT.EDU (Doug Ledford)
Tue Mar 17 04:54:28 1998
Date: Tue, 17 Mar 1998 03:51:53 -0600
From: Doug Ledford <dledford@dialnet.net>
To: sistema@readysoft.es
CC: linux-scsi@vger.rutgers.edu
sistema@readysoft.es wrote:
> Iīve been punnishing the system with heavy swapping, both with
> 2.0.33+aic-5.0.[78] and 2.1.90-pre3+aic-5.0.7.
>
> Iīve made the tests with swap on and off, with both kernels. Remapping
> bits were on.
>
> Swap off
> --------
> No problems with any of them. It looks like the swap partition has got
> physical errors, as I donīt get failures without swap.
>
> Swap on
> -------
> 2.0.33 recovers from errors, but other partitions in the same drive
> start failing when swap starts.
>
> 2.1.90-pre doesnīt recover from all the errors and finally the system
> locks up; iīts not a hrd lock as I can use the magic keys, but I canīt
> even shutdown the machine.
I would say both of these are merely different symptoms of the same
problem. Namely, when a swap sector is bad and data can't be swapped back
in, the system is pretty well screwed. We've written something we expected
to be able to get back (and indeed have to be able to get back in order to
continue) and now we find out we can't. It's no better than trying to
compile kernels on a machine with bogus RAM that causes gcc to give SIG 11
errors all of the time.
Secondly, there are a lot of drives out there with firmware that starts to
flake out if the error is bad enough. Some of the drives I've seen in the
recent past would completely wedge the entire SCSI bus so hard that the only
solution was to power down the system before it would work again when it hit
particularly bad portions of the disk.
> My conclusions
> --------------
> Thereīs something wrong in the SCSI error recovery in 2.1.89 and
> greater kernels.
When you get a bad sector in a swap partition that you've already written
info to, there is *no* good recovery method. The best case scenario is you
kill a program or two and go on, the worst case scenario is all hell breaks
loose because programs are passing all kinds of bogus data, and the kernel
itself could have data there (specifically, things like the vmalloc call in
net/unix come to mind as things that have swappable kernel mem, but I could
be wrong if vmalloc doesn't actually give you swappable mem, I haven't
looked into it but I was under the impression that it does).
> Iīll turn on debugging this midday, with 2.1.90-pre3 and see what I get.
>
> Thanks for your interest.
> Pau
You can try debugging, and we can try to make the system more resilient, but
the bottom line is, if you have bad sectors in a swap partition that aren't
going away with the AWRE and ARRE bits set on the drive, then you either
have to quit using that partition or low level format the drive and hope
they don't come back. There are certain things that simply can not be
worked around in a sane manner without building hardware that has every
single circuit and bit duplicated in one place or another (and does that
even qualify as sane?)
--
Doug Ledford <dledford@dialnet.net>
Opinions expressed are my own, but
they should be everybody's.
-
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to majordomo@vger.rutgers.edu
