fs: scale files_lock

Improve scalability of files_lock by adding per-cpu, per-sb files lists,
protected with an lglock. The lglock provides fast access to the per-cpu lists
to add and remove files. It also provides a snapshot of all the per-cpu lists
(although this is very slow).

One difficulty with this approach is that a file can be removed from the list
by another CPU. We must track which per-cpu list the file is on with a new
variale in the file struct (packed into a hole on 64-bit archs). Scalability
could suffer if files are frequently removed from different cpu's list.

However loads with frequent removal of files imply short interval between
adding and removing the files, and the scheduler attempts to avoid moving
processes too far away. Also, even in the case of cross-CPU removal, the
hardware has much more opportunity to parallelise cacheline transfers with N
cachelines than with 1.

A worst-case test of 1 CPU allocating files subsequently being freed by N CPUs
degenerates to contending on a single lock, which is no worse than before. When
more than one CPU are allocating files, even if they are always freed by
different CPUs, there will be more parallelism than the single-lock case.

Testing results:

On a 2 socket, 8 core opteron, I measure the number of times the lock is taken
to remove the file, the number of times it is removed by the same CPU that
added it, and the number of times it is removed by the same node that added it.

Booting:    locks=  25049 cpu-hits=  23174 (92.5%) node-hits=  23945 (95.6%)
kbuild -j16 locks=2281913 cpu-hits=2208126 (96.8%) node-hits=2252674 (98.7%)
dbench 64   locks=4306582 cpu-hits=4287247 (99.6%) node-hits=4299527 (99.8%)

So a file is removed from the same CPU it was added by over 90% of the time.
It remains within the same node 95% of the time.

Tim Chen ran some numbers for a 64 thread Nehalem system performing a compile.

                throughput
2.6.34-rc2      24.5
+patch          24.9

                us      sys     idle    IO wait (in %)
2.6.34-rc2      51.25   28.25   17.25   3.25
+patch          53.75   18.5    19      8.75

So significantly less CPU time spent in kernel code, higher idle time and
slightly higher throughput.

Single threaded performance difference was within the noise of microbenchmarks.
That is not to say penalty does not exist, the code is larger and more memory
accesses required so it will be slightly slower.

Cc: linux-kernel@vger.kernel.org
Cc: Tim Chen <tim.c.chen@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>

just delay __put_super() a bit

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>

If sget() finds a matching superblock being set up, it'll
grab an active reference to it and grab s_umount.  That's
fine - we'll wait for completion of foofs_get_sb() that way.
However, if said foofs_get_sb() fails we'll end up holding
the halfway-created superblock.  deactivate_locked_super()
called by foofs_get_sb() will just unlock the sucker since
we are holding another active reference to it.

What we need is a way to tell if superblock has been successfully
set up.  Unfortunately, neither ->s_root nor the check for
MS_ACTIVE quite fit.  Cheap and easy way, suitable for backport:
new flag set by the (only) caller of ->get_sb().  If that flag
isn't present by the time sget() grabbed s_umount on preexisting
superblock it has found, it's seeing a stillborn and should
just bury it with deactivate_locked_super() (and repeat the search).

Longer term we want to set that flag in ->get_sb() instances (and
check for it to distinguish between "sget() found us a live sb"
and "sget() has allocated an sb, we need to set it up" in there,
instead of checking ->s_root as we do now).

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Cc: stable@kernel.org

Fix an obscure AB-BA deadlock in get_sb_bdev().

When a superblock is mounted more than once get_sb_bdev() calls
close_bdev_exclusive() to drop the extra bdev reference while holding
s_umount.  However, sb->s_umount nests inside bd_mutex during
__invalidate_device() and close_bdev_exclusive() acquires bd_mutex during
blkdev_put(); thus creating an AB-BA deadlock.

This condition doesn't trigger frequently.  For this condition to be
visible to lockdep, the filesystem must occupy the whole device (as
__invalidate_device() only grabs bd_mutex for the whole device), the FS
must be mounted more than once and partition rescan should be issued while
the FS is still mounted.

Fix it by dropping s_umount over close_bdev_exclusive().

Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: Ciprian Docan <docan@eden.rutgers.edu>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Acked-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>

list_for_each_entry_safe is not suitable to protect against concurrent
modification of the list. 6754af6 introduced a race in sb walking.

list_for_each_entry can use the trick of pinning the current entry in
the list before we drop and retake the lock because it subsequently
follows cur->next. However list_for_each_entry_safe saves n=cur->next
for following before entering the loop body, so when the lock is
dropped, n may be deleted.

Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: John Stultz <johnstul@us.ibm.com>
Cc: Frank Mayhar <fmayhar@google.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

* 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jack/linux-fs-2.6:
  quota: Convert quota statistics to generic percpu_counter
  ext3 uses rb_node = NULL; to zero rb_root.
  quota: Fixup dquot_transfer
  reiserfs: Fix resuming of quotas on remount read-write
  pohmelfs: Remove dead quota code
  ufs: Remove dead quota code
  udf: Remove dead quota code
  quota: rename default quotactl methods to dquot_
  quota: explicitly set ->dq_op and ->s_qcop
  quota: drop remount argument to ->quota_on and ->quota_off
  quota: move unmount handling into the filesystem
  quota: kill the vfs_dq_off and vfs_dq_quota_on_remount wrappers
  quota: move remount handling into the filesystem
  ocfs2: Fix use after free on remount read-only

Fix up conflicts in fs/ext4/super.c and fs/ufs/file.c

Fix fs/super.c kernel-doc warning and function notation:
Warning(fs/super.c:957): No description found for parameter 'sb'

Signed-off-by: Randy Dunlap <randy.dunlap@oracle.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>

Only set the quota operation vectors if the filesystem actually supports
quota instead of doing it for all filesystems in alloc_super().

[Jan Kara: Export dquot_operations and vfs_quotactl_ops]

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>

Currently the VFS calls into the quotactl interface for unmounting
filesystems.  This means filesystems with their own quota handling
can't easily distinguish between user-space originating quotaoff
and an unount.  Instead move the responsibily of the unmount handling
into the filesystem to be consistent with all other dquot handling.

Note that we do call dquot_disable a lot later now, e.g. after
a sync_filesystem.  But this is fine as the quota code does all its
writes via blockdev's mapping and that is synced even later.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>

Currently do_remount_sb calls into the dquot code to tell it about going
from rw to ro and ro to rw.  Move this code into the filesystem to
not depend on the dquot code in the VFS - note ocfs2 already ignores
these calls and handles remount by itself.  This gets rid of overloading
the quotactl calls and allows to unify the VFS and XFS codepaths in
that area later.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>