[Cbe-oss-dev] NFS corruption with 2.6.16 and 2.6.17-rc1/rc2

[Dan Burcaw]

A follow up from an associate:

"The posting incorrectly listed the checksum for /h/abono2/words as
634329572 when the actual checksum was 2429808809 you will have to update
the list posting. 

In addition I think that I'm very close to tracking down the cause of the
problem. I believe that there is a constant that is used by NFS to limit the
max request size and in the pre 2.6.16 releases the actual request size
would never exceeded 32768 even when the read size was set to 65536.
Therefore when the response for each nfs request was received the REPLY
would fulfill the whole request. In post 2.6.16 releases this is not the
case. NFS now tries to request 65536 bytes per request when the read size
exceeds 65535. The network appliance responds to these requests with the 1st
32768 bytes of the block. LINUX appears to recognize the missing data and
proceeds to issue a new read request for the rest of the block with the
correct offset and size. However the processing for of the reply to the 2nd
half doesn't get handled correctly. Even though the correct data for the
second half is received it doesn't get added to the inode cache. I first
thought that this was because the 1st half of the lookahead read request
preceeded the 2nd half of the 1st request but at the present moment I don¹t
think that this is the case but I will do more testing that should confirm
this and hopefully help me track down which constant is causing the
problem."



On 4/26/06 12:23 PM, "Dan Burcaw" <dburcaw at terrasoftsolutions.com> wrote:

> Folks,
> 
> Using 2.6.16, 2.6.17-rc1, and linux-2.6.17-rc2 we are seeing what appears to
> be NFS corruption.
> 
> Please note, this is on Cell processor hardware.
> 
> Here is the output:
> 
> [root at cell_03 ~]# uname -a
> Linux cell_03 2.6.17-0.yhpc.0.1.3cell #1 SMP Wed Apr 26 06:05:55
> EDT 2006 ppc64 ppc64 ppc64 GNU/Linux
> 
> [root at cell_03 ~]# mount | grep abono
> 10.0.0.1:/home/abono on /h/abono type nfs
> (rw,nosuid,nodev,rsize=65535,wsize=65535,addr=10.0.0.1)
> 10.0.0.1:/home/abono on /h/abono2 type nfs
> (rw,nosuid,nodev,rsize=65536,wsize=65536,addr=10.0.0.1)
> 
> [root at cell_03 ~]# cksum /h/abono/words /h/abono/words2
> 634329572 206662 /h/abono/words
> 2429808809 206662 /h/abono/words2
> 
> [root at cell_03 ~]# cksum /h/abono2/words /h/abono2/words2
> 634329572 206662 /h/abono2/words
> 2429808809 206662 /h/abono2/words2
> 
> 
> Ideas?
> 
> Thanks,
> Dan Burcaw
> 
> 
> _______________________________________________
> tss-mercury mailing list
> tss-mercury at lists.terrasoftsolutions.com
> http://lists.terrasoftsolutions.com/mailman/listinfo/tss-mercury
> 
> !DSPAM:444fc6ee83828987581379!