Re: xfs and page allocation failures

[Nathan Scott]

Hi there,

On Tue, Apr 06, 2004 at 11:33:35AM +0200, Oliver Kiddle wrote:
> I posted back in January about problems I was having with 2.6.1. Thanks
> for the help back then. I'm still having problems with the same
> machine, though.
> 
> 2.6.3 is usable (just three page allocation failures printed when I run
> xfsdump). xfsdump crashed 2.6.4 and I never got around to trying to
> capture the console output. I tried 2.6.5 yesterday and have had two
> issues. The first, relatively harmless problem was two page allocation
> failures printed when running xfsdump (output is below).
> 
> Then, this morning, mountd was no longer working. NFS was still happily
> working where clients had mounted the filesystems. rpcinfo -u was
> getting a timeout for mountd. I tried first restarting rpc.mountd which
> had no effect and then tried `/etc/init.d/nfs-kernel-server restart',
> also to no effect.
> 
> I've attached the relevant part of the dmesg output below.
> 
> Thanks
> 
> Oliver
> 
> st0: Block limits 1 - 16777215 bytes.
> xfsdump: page allocation failure. order:9, mode:0xd0
> Call Trace:
>  [<c012d79f>] __alloc_pages+0x2e0/0x325
>  [<c02ad92d>] enlarge_buffer+0xcf/0x182

This is xfsdump making requests of the SCSI tape driver with buffers
which are too large for it to pin down, the warning is harmless and
ST chugs on.  I believe the warnings will be fixed in future kernels,
and xfsdump/xfsrestore will need some tweaks to use more appropriately
sized buffers.

> Unable to handle kernel NULL pointer dereference at virtual address 00000004
> EIP is at free_block+0x48/0xc8
> Process kswapd0 (pid: 7, threadinfo=c1b74000 task=c1b791e0)
> Call Trace:
>  [<c0130296>] cache_flusharray+0x3f/0xbc
>  [<c013045f>] kmem_cache_free+0x48/0x4c
>  [<c02119fa>] linvfs_destroy_inode+0x1b/0x1f

This looks like a use-after-free - freeing an inode for a second
time, I think - I haven't come across this one before.  I expect
the second oops and pagebuf warnings will be a follow-on effect
from this initial failure.  The initial failure will need some 
more investigation - if you can reliably hit it pls let me know
what/how you're doing that.

thanks.

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