staging/lustre: Limit reply buffer size
authorBrian Behlendorf <behlendorf1@llnl.gov>
Sun, 27 Apr 2014 17:06:47 +0000 (13:06 -0400)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Sun, 27 Apr 2014 17:24:54 +0000 (10:24 -0700)
commit44779340c6dcb8c858955c02b3aeef7d6b28a684
treebfbd8f6ec331e58db7df23d342f46ed8f7b044fb
parente69cd00cb3b3ce49954947ee3f78ba887d7ac8d0
staging/lustre: Limit reply buffer size

When allocating a reply buffer for the striping information don't
assume the unlikely worst case.  Instead, assume the common case
and size the buffer based on the observed default ea/cookie size.

The default size is initialized to a single stripe and allowed to
grow up to an entire page if needed.  This means that for smallish
filesystems (less than ~21 OSTs) where the worst case striping
information can fit in a single page there is effectively no
change.  Only for larger filesystem will the default be less than
the maximum.  This has a number of advantages.

* By limiting the default reply buffer size we avoid always
  vmalloc()'ing the buffer because it exceeds four pages in size
  and instead kmalloc() it.  This prevents the client from
  thrashing on the global vmalloc() spin lock.

* A reply buffer of exactly the right size (no larger) is allocated
  in the overflow case.  These larger reply buffers are still
  unlikely to exceed the 16k limit where a vmalloc() will occur.

* Saves memory in the common case.  Wide striped files exceeded
  the default are expected to be the exception.

The reason this patch works is because the ptlrpc layer is smart
enough to reallocate the reply buffer when an overflow occurs.
Therefore the client doesn't have to drop the incoming reply and
send a new request with a larger reply buffer.

It's also worth mentioning that the reply buffer always contains
a significant amount of extra padding because they are rounded up
to the nearest power of two.  This means that even files striped
wider than the default have a good chance of fitting in the
allocated reply buffer.

Also remove client eadatasize check in mdt xattr packing because
as said above client can handle -EOVERFLOW.

Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Lai Siyao <lai.siyao@intel.com>
Reviewed-on: http://review.whamcloud.com/6339
Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-3338
Reviewed-by: James Simmons <uja.ornl@gmail.com>
Reviewed-by: Andreas Dilger <andreas.dilger@intel.com>
Reviewed-by: Bob Glossman <bob.glossman@intel.com>
Signed-off-by: Oleg Drokin <oleg.drokin@intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
14 files changed:
drivers/staging/lustre/lustre/include/lustre_mdc.h
drivers/staging/lustre/lustre/include/obd.h
drivers/staging/lustre/lustre/include/obd_class.h
drivers/staging/lustre/lustre/lclient/lcommon_misc.c
drivers/staging/lustre/lustre/llite/dir.c
drivers/staging/lustre/lustre/llite/file.c
drivers/staging/lustre/lustre/llite/llite_internal.h
drivers/staging/lustre/lustre/llite/llite_lib.c
drivers/staging/lustre/lustre/llite/llite_nfs.c
drivers/staging/lustre/lustre/llite/lproc_llite.c
drivers/staging/lustre/lustre/lmv/lmv_obd.c
drivers/staging/lustre/lustre/mdc/mdc_locks.c
drivers/staging/lustre/lustre/mdc/mdc_reint.c
drivers/staging/lustre/lustre/mdc/mdc_request.c