Copying IBM's comments from BZ 662782.  I will have them reverse mirror this one so they can comment directly on this one.



IBM Bug Proxy 2011-02-25 13:00:45 EST

------- Comment From pbadari.com 2011-02-25 12:52 EDT-------
I would like to suggest that instead of considering all NFS changes that are
targetted for RHEL6.1,
lets focus on specific fixes we need for RHEL6.0-z stream.

I will provide list of fixes we need for z-stream:

1) fix similar to the one descrined in bug 67632 (for RHEL5 series)

@@ -709,7 +709,7 @@ static ssize_t nfs_direct_write(struct k
return -ENOMEM;
nfs_alloc_commit_data(dreq);

-       if (dreq->commit_data == NULL || count<  wsize)
+       if (dreq->commit_data == NULL || count<= wsize)
sync = FLUSH_STABLE;

dreq->inode = inode;

2) we may have another one to improve O_SYNC performance. let me look at that
also.

ltcbz 67632 is rhbz 643441/677172 ltc bz 67632  -  RIT00404862- pure nfs client
performance using odirect )

------- Comment From pbadari.com 2011-02-25 16:42 EDT-------
We have 2 use-case scenario fixes here:

1) O_DIRECT: above simple fix where STABLE -> UNSTABLE cut off happens.

2) O_SYNC: There are lot of changes on how O_SYNC is handled between RHEL5 -> RHEL6.
One major change is, O_SYNC no longer generates STABLE writes. It always generates
UNSTABLE followed by COMMIT. This is causing significant performance degrade with
our backend SoNAS/GPFS (since it has larger blocksize).

Here is the proposed fix (currently under validation) to generate STABLE writes like
before (RHEL5) till ->wsize.

Created attachment 481091 [details]
generate stable writes with O_SYNC


------- Comment (attachment only) From pbadari.com 2011-02-25 16:42 EDT-------

------- Comment From tpnoonan.com 2011-02-26 11:25 EDT-------
can these 2 fixes go into rhel6.0.z with kernel freeze 3/2/11? thanks

(In reply to comment #8)
> ------- Comment From tpnoonan.com 2011-02-26 11:25 EDT-------
> can these 2 fixes go into rhel6.0.z with kernel freeze 3/2/11? thanks

Question, has requested patch been posted to upstream?

Also has there been any performance testing to make sure
this does not cause a regression with non-O_SYNC writes?

------- Comment From pbadari.com 2011-03-01 16:07 EDT-------
1) O_DIRECT patch is well tested and is upstream.

2) O_SYNC problem was recently found. Neil Brown proposed this patch and still being worked
upstream. https://patchwork.kernel.org/patch/565831/

------- Comment From ffilz.com 2011-03-01 16:13 EDT-------
Another patch that should be considered is:

Bug 68522  -  RIT1613663- NFS client has troubles with fileid with bit 31 (or bit 63) set

------- Comment From ffilz.com 2011-03-02 10:59 EDT-------
The impact o the bit 31/63 bug is that 32 bit clients will get unxpected failures if the NFS server has inode numbers with bit 31 or bit 63 set. The compatibility code tries to fix up these inode numbers to fit into 32 bits, but then causes the resulting 32 bit quantity to be sign extended to 64 bits, which of course then doesn't fit in 32 bits.

If I remember correctly, this issue will arise even with 32 bit inode numbers when bit 31 is set.