This adds a d_prune dentry operation that is called by the VFS prior to
pruning (i.e. unhashing and killing) a hashed dentry from the dcache.
Wrap dentry_lru_del() and use the new _prune() helper in the cases where we
are about to unhash and kill the dentry.

This will be used by Ceph to maintain a flag indicating whether the
complete contents of a directory are contained in the dcache, allowing it
to satisfy lookups and readdir without addition server communication.

Renumber a few DCACHE_* #defines to group DCACHE_OP_PRUNE with the other
DCACHE_OP_ bits.
Signed-off-by: NSage Weil <sage@newdream.net>
Signed-off-by: NChristoph Hellwig <hch@lst.de>

Gcc tends to generate better code with small integers, including the
DCACHE_xyz flag tests - so move the common ones to be first in the list.
Also just remove the unused DCACHE_INOTIFY_PARENT_WATCHED and
DCACHE_AUTOFS_PENDING values, their users no longer exists in the source
tree.

And add a "unlikely()" to the DCACHE_OP_COMPARE test, since we want the
common case to be a nice straight-line fall-through.
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Fix new kernel-doc warning in fs/dcache.c:

  Warning(fs/dcache.c:797): No description found for parameter 'sb'
Signed-off-by: NRandy Dunlap <rdunlap@xenotime.net>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Reorganise shrink_dcache_for_umount_subtree() in light of the demise of
dcache_lock.  Without that dcache_lock, there is no need for the batching of
removal of dentries from the system under it (we wanted to make intensive use
of the locked data whilst we held it, but didn't want to hold it for long at a
time).

This works, provided the preceding patch is correct in its removal of locking
on dentry->d_lock on the basis that no one should be locking these dentries any
more as the whole superblock is defunct.

With this patch, the calls to dentry_lru_del() and __d_shrink() are placed at
the point where each dentry is detached handled.

It is possible that, as an alternative, the batching should still be done -
but only for dentry_lru_del() of all a dentry's children in one go.  In such a
case, the batching would be done under dcache_lru_lock.
Signed-off-by: NDavid Howells <dhowells@redhat.com>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

Locks of the dcache_lock were replaced by locks of dentry->d_lock in commits
such as:

	23044507
	2fd6b7f5

as part of the RCU-based pathwalk changes, despite the fact that the caller
(shrink_dcache_for_umount()) notes in the banner comment the reasons that
d_lock is not necessary in these functions:

/*
 * destroy the dentries attached to a superblock on unmounting
 * - we don't need to use dentry->d_lock because:
 *   - the superblock is detached from all mountings and open files, so the
 *     dentry trees will not be rearranged by the VFS
 *   - s_umount is write-locked, so the memory pressure shrinker will ignore
 *     any dentries belonging to this superblock that it comes across
 *   - the filesystem itself is no longer permitted to rearrange the dentries
 *     in this superblock
 */

So remove these locks.  If the locks are actually necessary, then this banner
comment should be altered instead.

The hash table chains are protected by 1-bit locks in the hash table heads, so
those shouldn't be a problem.

Note that to make this work, __d_drop() has to be split so that the RCUwalk
barrier can be avoided.  This causes problems otherwise as it has an assertion
that dentry->d_lock is locked - but there is no need for that as no one else
can be trying to access this dentry, except to step over it (and that should
be handled by d_free(), I think).
Signed-off-by: NDavid Howells <dhowells@redhat.com>
Cc: Nick Piggin <npiggin@kernel.dk>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

Remove the detached-dentry counter from shrink_dcache_for_umount_subtree() as
the value it computes is no longer used as of commit
312d3ca8 which made the nr_dentry counters
summed per-CPU rather than global atomic.
Signed-off-by: NDavid Howells <dhowells@redhat.com>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

Adding a comment to d_materialise_unique per Al's request...

d_move and __d_move have some pretty substantial locking requirements,
but they are not clearly documented. Add some comments spelling them
out. Also, document the requirement for the i_mutex of the parent in
d_materialise_unique.

Cc: Al Viro <viro@ZenIV.linux.org.uk>
Signed-off-by: NJeff Layton <jlayton@redhat.com>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

It seems to hurt performance in real life.  Yes, the inode will be used
later, but the conditional doesn't seem to predict all that well
(negative dentries are not uncommon) and it looks like the cost of
prefetching is simply higher than depending on the cache doing the right
thing.

As usual.
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

->d_parent is never NULL...
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

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

With context based shrinkers, we can implement a per-superblock
shrinker that shrinks the caches attached to the superblock. We
currently have global shrinkers for the inode and dentry caches that
split up into per-superblock operations via a coarse proportioning
method that does not batch very well.  The global shrinkers also
have a dependency - dentries pin inodes - so we have to be very
careful about how we register the global shrinkers so that the
implicit call order is always correct.

With a per-sb shrinker callout, we can encode this dependency
directly into the per-sb shrinker, hence avoiding the need for
strictly ordering shrinker registrations. We also have no need for
any proportioning code for the shrinker subsystem already provides
this functionality across all shrinkers. Allowing the shrinker to
operate on a single superblock at a time means that we do less
superblock list traversals and locking and reclaim should batch more
effectively. This should result in less CPU overhead for reclaim and
potentially faster reclaim of items from each filesystem.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

... and simplify the living hell out of callers
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

New helper (non-exported, fs/internal.h-only): __d_alloc(sb, name).
Allocates dentry, sets its ->d_sb to given superblock and sets
->d_op accordingly.  Old d_alloc(NULL, name) callers are converted
to that (all of them know what superblock they want).  d_alloc()
itself is left only for parent != NULl case; uses __d_alloc(),
inserts result into the list of parent's children.

Note that now ->d_sb is assign-once and never NULL *and*
->d_parent is never NULL either.
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

Btrfs (and I'd venture most other fs's) stores its indexes in nice disk order
for readdir, but unfortunately in the case of anything that stats the files in
order that readdir spits back (like oh say ls) that means we still have to do
the normal lookup of the file, which means looking up our other index and then
looking up the inode.  What I want is a way to create dummy dentries when we
find them in readdir so that when ls or anything else subsequently does a
stat(), we already have the location information in the dentry and can go
straight to the inode itself.  The lookup stuff just assumes that if it finds a
dentry it is done, it doesn't perform a lookup.  So add a DCACHE_NEED_LOOKUP
flag so that the lookup code knows it still needs to run i_op->lookup() on the
parent to get the inode for the dentry.  I have tested this with btrfs and I
went from something that looks like this

http://people.redhat.com/jwhiter/ls-noreada.png

To this

http://people.redhat.com/jwhiter/ls-good.png

Thats a savings of 1300 seconds, or 22 minutes.  That is a significant savings.
Thanks,
Signed-off-by: NJosef Bacik <josef@redhat.com>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

Both __d_unalias() and __d_materialise_dentry() need loop prevention.
Grab rename_lock in caller, check for loops there...

As a side benefit, we have dentry_lock_for_move() called only under
rename_lock, which seriously reduces deadlock potential of the
execrable "locking order" used for ->d_lock.
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

Change each shrinker's API by consolidating the existing parameters into
shrink_control struct.  This will simplify any further features added w/o
touching each file of shrinker.

[akpm@linux-foundation.org: fix build]
[akpm@linux-foundation.org: fix warning]
[kosaki.motohiro@jp.fujitsu.com: fix up new shrinker API]
[akpm@linux-foundation.org: fix xfs warning]
[akpm@linux-foundation.org: update gfs2]
Signed-off-by: NYing Han <yinghan@google.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Acked-by: NPavel Emelyanov <xemul@openvz.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Acked-by: NRik van Riel <riel@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Commit e66eed65 ("list: remove prefetching from regular list
iterators") removed the include of prefetch.h from list.h, which
uncovered several cases that had apparently relied on that rather
obscure header file dependency.

So this fixes things up a bit, using

   grep -L linux/prefetch.h $(git grep -l '[^a-z_]prefetchw*(' -- '*.[ch]')
   grep -L 'prefetchw*(' $(git grep -l 'linux/prefetch.h' -- '*.[ch]')

to guide us in finding files that either need <linux/prefetch.h>
inclusion, or have it despite not needing it.

There are more of them around (mostly network drivers), but this gets
many core ones.
Reported-by: NStephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Now that the whole dcache_hash_bucket crap is gone, go all the way and
also remove the weird locking layering violations for locking the hash
buckets.  Add hlist_bl_lock/unlock helpers to move the locking into the
list abstraction instead of requiring each caller to open code it.
After all allowing for the bit locks is the whole point of these helpers
over the plain hlist variant.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

The dentry hashing rules have been really quite complicated for a long
while, in odd ways.  That made functions like __d_drop() very fragile
and non-obvious.

In particular, whether a dentry was hashed or not was indicated with an
explicit DCACHE_UNHASHED bit.  That's despite the fact that the hash
abstraction that the dentries use actually have a 'is this entry hashed
or not' model (which is a simple test of the 'pprev' pointer).

The reason that was done is because we used the normal 'is this entry
unhashed' model to mark whether the dentry had _ever_ been hashed in the
dentry hash tables, and that logic goes back many years (commit
b3423415: "dcache: avoid RCU for never-hashed dentries").

That, in turn, meant that __d_drop had totally different unhashing logic
for the dentry hash table case and for the anonymous dcache case,
because in order to use the "is this dentry hashed" logic as a flag for
whether it had ever been on the RCU hash table, we had to unhash such a
dentry differently so that we'd never think that it wasn't 'unhashed'
and wouldn't be free'd correctly.

That's just insane.  It made the logic really hard to follow, when there
were two different kinds of "unhashed" states, and one of them (the one
that used "list_bl_unhashed()") really had nothing at all to do with
being unhashed per se, but with a very subtle lifetime rule instead.

So turn all of it around, and make it logical.

Instead of having a DENTRY_UNHASHED bit in d_flags to indicate whether
the dentry is on the hash chains or not, use the hash chain unhashed
logic for that.  Suddenly "d_unhashed()" just uses "list_bl_unhashed()",
and everything makes sense.

And for the lifetime rule, just use an explicit DENTRY_RCUACCEES bit.
If we ever insert the dentry into the dentry hash table so that it is
visible to RCU lookup, we mark it DENTRY_RCUACCESS to show that it now
needs the RCU lifetime rules.  Now suddently that test at dentry free
time makes sense too.

And because unhashing now is sane and doesn't depend on where the dentry
got unhashed from (because the dentry hash chain details doesn't have
some subtle side effects), we can re-unify the __d_drop() logic and use
common code for the unhashing.

Also fix one more open-coded hash chain bit_spin_lock() that I missed in
the previous chain locking cleanup commit.
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

It's a useless abstraction for 'hlist_bl_head', and it doesn't actually
help anything - quite the reverse.  All the users end up having to know
about the hlist_bl_head details anyway, using 'struct hlist_bl_node *'
etc. So it just makes the code look confusing.

And the cost of it is extra '&b->head' syntactic noise, but more
importantly it spuriously makes the hash table dentry list look
different from the per-superblock DCACHE_DISCONNECTED dentry list.

As a result, the code ended up using ad-hoc locking for one case and
special helper functions for what is really another totally identical
case in the very same function.

Make it all look and work the same.
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

The case we should be verifying when updating the dentry name is that
the _parent_ inode (the directory) semaphore is held, not the semaphore
for the dentry itself.  It's the directory locking that rename and
readdir() etc all care about.

The comment just above even says so - but then the BUG_ON() still
checked the dentry inode itself.

Very few people noticed, because this helper function really isn't used
for very much, so you had to be using ncpfs to ever hit it.

I think I should just remove the BUG_ON (the function really has just
one user), but let's run with it fixed for a while before getting rid of
it entirely.
Reported-and-tested-by: NBongani Hlope <bonganih@bankservafrica.com>
Reported-and-tested-by: NBernd Feige <bernd.feige@uniklinik-freiburg.de>
Cc: Petr Vandrovec <petr@vandrovec.name>,
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Nick Piggin <npiggin@kernel.dk>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

While trying to track down some NFS problems with BTRFS, I kept noticing I was
getting -EACCESS for no apparent reason.  Eric Paris and printk() helped me
figure out that it was SELinux that was giving me grief, with the following
denial

type=AVC msg=audit(1290013638.413:95): avc:  denied  { 0x800000 } for  pid=1772
comm="nfsd" name="" dev=sda1 ino=256 scontext=system_u:system_r:kernel_t:s0
tcontext=system_u:object_r:unlabeled_t:s0 tclass=file

Turns out this is because in d_obtain_alias if we can't find an alias we create
one and do all the normal instantiation stuff, but we don't do the
security_d_instantiate.

Usually we are protected from getting a hashed dentry that hasn't yet run
security_d_instantiate() by the parent's i_mutex, but obviously this isn't an
option there, so in order to deal with the case that a second thread comes in
and finds our new dentry before we get to run security_d_instantiate(), we go
ahead and call it if we find a dentry already.  Eric assures me that this is ok
as the code checks to see if the dentry has been initialized already so calling
security_d_instantiate() against the same dentry multiple times is ok.  With
this patch I'm no longer getting errant -EACCESS values.
Signed-off-by: NJosef Bacik <josef@redhat.com>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

The new vfs locking scheme introduced in 2.6.38 breaks NFS sillyrename
because the latter relies on being able to determine the parent
directory of the dentry in the ->iput() callback in order to send the
appropriate unlink rpc call.

Looking at the code that cares about races with dput(), there doesn't
seem to be anything that specifically uses d_parent as a test for
whether or not there is a race:
  - __d_lookup_rcu(), __d_lookup() all test for d_hashed() after d_parent
  - shrink_dcache_for_umount() is safe since nothing else can rearrange
    the dentries in that super block.
  - have_submount(), select_parent() and d_genocide() can test for a
    deletion if we set the DCACHE_DISCONNECTED flag when the dentry
    is removed from the parent's d_subdirs list.
Signed-off-by: NTrond Myklebust <Trond.Myklebust@netapp.com>
Cc: stable@kernel.org (2.6.38, needs commit c826cb7d "dcache.c:
	create helper function for duplicated functionality" )
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

This creates a helper function for he "try to ascend into the parent
directory" case, which was written out in triplicate before.  With all
the locking and subtle sequence number stuff, we really don't want to
duplicate that kind of code.
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Without this patch, inodes are not promptly freed on last close of an
unlinked file by an nfs client:

	client$ mount -tnfs4 server:/export/ /mnt/
	client$ tail -f /mnt/FOO
	...
	server$ df -i /export
	server$ rm /export/FOO
	(^C the tail -f)
	server$ df -i /export
	server$ echo 2 >/proc/sys/vm/drop_caches
	server$ df -i /export

the df's will show that the inode is not freed on the filesystem until
the last step, when it could have been freed after killing the client's
tail -f. On-disk data won't be deallocated either, leading to possible
spurious ENOSPC.

This occurs because when the client does the close, it arrives in a
compound with a putfh and a close, processed like:

	- putfh: look up the filehandle.  The only alias found for the
	  inode will be DCACHE_UNHASHED alias referenced by the filp
	  this, so it creates a new DCACHE_DISCONECTED dentry and
	  returns that instead.
	- close: closes the existing filp, which is destroyed
	  immediately by dput() since it's DCACHE_UNHASHED.
	- end of the compound: release the reference
	  to the current filehandle, and dput() the new
	  DCACHE_DISCONECTED dentry, which gets put on the
	  unused list instead of being destroyed immediately.

Nick Piggin suggested fixing this by allowing d_obtain_alias to return
the unhashed dentry that is referenced by the filp, instead of making it
create a new dentry.

Leave __d_find_alias() alone to avoid changing behavior of other
callers.

Also nfsd doesn't need all the checks of __d_find_alias(); any dentry,
hashed or unhashed, disconnected or not, should work.
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

dcache-locking.txt is not exist any more, and the path was not
correct anyway. Fix it.
Signed-off-by: NNamhyung Kim <namhyung@gmail.com>
Signed-off-by: NJiri Kosina <jkosina@suse.cz>

Fix new fs/dcache.c kernel-doc warnings:

  Warning(fs/dcache.c:184): No description found for parameter 'dentry'
  Warning(fs/dcache.c:296): No description found for parameter 'parent'
  Warning(fs/dcache.c:1985): No description found for parameter 'dparent'
  Warning(fs/dcache.c:1985): Excess function parameter 'parent' description in 'd_validate'
Signed-off-by: NRandy Dunlap <randy.dunlap@oracle.com>
Cc:	Alexander Viro <viro@zeniv.linux.org.uk>
Cc:	Nick Piggin <npiggin@kernel.dk>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Add a dentry op (d_automount) to handle automounting directories rather than
abusing the follow_link() inode operation.  The operation is keyed off a new
dentry flag (DCACHE_NEED_AUTOMOUNT).

This also makes it easier to add an AT_ flag to suppress terminal segment
automount during pathwalk and removes the need for the kludge code in the
pathwalk algorithm to handle directories with follow_link() semantics.

The ->d_automount() dentry operation:

	struct vfsmount *(*d_automount)(struct path *mountpoint);

takes a pointer to the directory to be mounted upon, which is expected to
provide sufficient data to determine what should be mounted.  If successful, it
should return the vfsmount struct it creates (which it should also have added
to the namespace using do_add_mount() or similar).  If there's a collision with
another automount attempt, NULL should be returned.  If the directory specified
by the parameter should be used directly rather than being mounted upon,
-EISDIR should be returned.  In any other case, an error code should be
returned.

The ->d_automount() operation is called with no locks held and may sleep.  At
this point the pathwalk algorithm will be in ref-walk mode.

Within fs/namei.c itself, a new pathwalk subroutine (follow_automount()) is
added to handle mountpoints.  It will return -EREMOTE if the automount flag was
set, but no d_automount() op was supplied, -ELOOP if we've encountered too many
symlinks or mountpoints, -EISDIR if the walk point should be used without
mounting and 0 if successful.  The path will be updated to point to the mounted
filesystem if a successful automount took place.

__follow_mount() is replaced by follow_managed() which is more generic
(especially with the patch that adds ->d_manage()).  This handles transits from
directories during pathwalk, including automounting and skipping over
mountpoints (and holding processes with the next patch).

__follow_mount_rcu() will jump out of RCU-walk mode if it encounters an
automount point with nothing mounted on it.

follow_dotdot*() does not handle automounts as you don't want to trigger them
whilst following "..".

I've also extracted the mount/don't-mount logic from autofs4 and included it
here.  It makes the mount go ahead anyway if someone calls open() or creat(),
tries to traverse the directory, tries to chdir/chroot/etc. into the directory,
or sticks a '/' on the end of the pathname.  If they do a stat(), however,
they'll only trigger the automount if they didn't also say O_NOFOLLOW.

I've also added an inode flag (S_AUTOMOUNT) so that filesystems can mark their
inodes as automount points.  This flag is automatically propagated to the
dentry as DCACHE_NEED_AUTOMOUNT by __d_instantiate().  This saves NFS and could
save AFS a private flag bit apiece, but is not strictly necessary.  It would be
preferable to do the propagation in d_set_d_op(), but that doesn't normally
have access to the inode.

[AV: fixed breakage in case if __follow_mount_rcu() fails and nameidata_drop_rcu()
succeeds in RCU case of do_lookup(); we need to fall through to non-RCU case after
that, rather than just returning with ungrabbed *path]
Signed-off-by: NDavid Howells <dhowells@redhat.com>
Was-Acked-by: NIan Kent <raven@themaw.net>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

It's indicative of a real problem, and it actually triggers with
autofs4, but the BUG_ON() is excessive.  The autofs4 case is being fixed
(to only set d_op in the ->lookup method) but not merged yet.  In the
meantime this gets the code limping along.
Reported-by: NAlex Elder <aelder@sgi.com>
Cc: Ian Kent <raven@themaw.net>
Cc: Nick Piggin <npiggin@kernel.dk>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Fix function kernel-doc warning for prepend_path():

Warning(fs/dcache.c:1924): missing initial short description on line:
Signed-off-by: NRandy Dunlap <randy.dunlap@oracle.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

Fix function parameter kernel-doc for d_validate():

Warning(fs/dcache.c:1495): No description found for parameter 'parent'
Warning(fs/dcache.c:1495): Excess function parameter 'dparent' description in 'd_validate'
Signed-off-by: NRandy Dunlap <randy.dunlap@oracle.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

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

The standard memcmp function on a Westmere system shows up hot in
profiles in the `git diff` workload (both parallel and single threaded),
and it is likely due to the costs associated with trapping into
microcode, and little opportunity to improve memory access (dentry
name is not likely to take up more than a cacheline).

So replace it with an open-coded byte comparison. This increases code
size by 8 bytes in the critical __d_lookup_rcu function, but the
speedup is huge, averaging 10 runs of each:

git diff st   user   sys   elapsed  CPU
before        1.15   2.57  3.82      97.1
after         1.14   2.35  3.61      96.8

git diff mt   user   sys   elapsed  CPU
before        1.27   3.85  1.46     349
after         1.26   3.54  1.43     333

Elapsed time for single threaded git diff at 95.0% confidence:
        -0.21  +/- 0.01
        -5.45% +/- 0.24%

It's -0.66% +/- 0.06% elapsed time on my Opteron, so rep cmp costs on the
fam10h seem to be relatively smaller, but there is still a win.
Signed-off-by: NNick Piggin <npiggin@kernel.dk>

This makes single threaded git diff -1.25% +/- 0.05% elapsed time on my
2s12c24t Westmere system, and -0.86% +/- 0.05% on my 2s8c Barcelona, by
prefetching the important first cacheline of the inode in while we do the
actual name compare and other operations on the dentry.

There was no measurable slowdown in the single file stat case, or the creat
case (where negative dentries would be common).
Signed-off-by: NNick Piggin <npiggin@kernel.dk>

Regardless of how much we possibly try to scale dcache, there is likely
always going to be some fundamental contention when adding or removing children
under the same parent. Pseudo filesystems do not seem need to have connected
dentries because by definition they are disconnected.
Signed-off-by: NNick Piggin <npiggin@kernel.dk>

dcache_inode_lock can be replaced with per-inode locking. Use existing
inode->i_lock for this. This is slightly non-trivial because we sometimes
need to find the inode from the dentry, which requires d_inode to be
stabilised (either with refcount or d_lock).
Signed-off-by: NNick Piggin <npiggin@kernel.dk>

We can turn the dcache hash locking from a global dcache_hash_lock into
per-bucket locking.
Signed-off-by: NNick Piggin <npiggin@kernel.dk>

Put dentry and inode fields into top of data structure.  This allows RCU path
traversal to perform an RCU dentry lookup in a path walk by touching only the
first 56 bytes of the dentry.

We also fit in 8 bytes of inline name in the first 64 bytes, so for short
names, only 64 bytes needs to be touched to perform the lookup. We should
get rid of the hash->prev pointer from the first 64 bytes, and fit 16 bytes
of name in there, which will take care of 81% rather than 32% of the kernel
tree.

inode is also rearranged so that RCU lookup will only touch a single cacheline
in the inode, plus one in the i_ops structure.

This is important for directory component lookups in RCU path walking. In the
kernel source, directory names average is around 6 chars, so this works.

When we reach the last element of the lookup, we need to lock it and take its
refcount which requires another cacheline access.

Align dentry and inode operations structs, so members will be at predictable
offsets and we can group common operations into head of structure.
Signed-off-by: NNick Piggin <npiggin@kernel.dk>

Reduce some branches and memory accesses in dcache lookup by adding dentry
flags to indicate common d_ops are set, rather than having to check them.
This saves a pointer memory access (dentry->d_op) in common path lookup
situations, and saves another pointer load and branch in cases where we
have d_op but not the particular operation.

Patched with:

git grep -E '[.>]([[:space:]])*d_op([[:space:]])*=' | xargs sed -e 's/\([^\t ]*\)->d_op = \(.*\);/d_set_d_op(\1, \2);/' -e 's/\([^\t ]*\)\.d_op = \(.*\);/d_set_d_op(\&\1, \2);/' -i
Signed-off-by: NNick Piggin <npiggin@kernel.dk>

Rather than keep a d_mounted count in the dentry, set a dentry flag instead.
The flag can be cleared by checking the hash table to see if there are any
mounts left, which is not time critical because it is performed at detach time.

The mounted state of a dentry is only used to speculatively take a look in the
mount hash table if it is set -- before following the mount, vfsmount lock is
taken and mount re-checked without races.

This saves 4 bytes on 32-bit, nothing on 64-bit but it does provide a hole I
might use later (and some configs have larger than 32-bit spinlocks which might
make use of the hole).

Autofs4 conversion and changelog by Ian Kent <raven@themaw.net>:
In autofs4, when expring direct (or offset) mounts we need to ensure that we
block user path walks into the autofs mount, which is covered by another mount.
To do this we clear the mounted status so that follows stop before walking into
the mount and are essentially blocked until the expire is completed. The
automount daemon still finds the correct dentry for the umount due to the
follow mount logic in fs/autofs4/root.c:autofs4_follow_link(), which is set as
an inode operation for direct and offset mounts only and is called following
the lookup that stopped at the covered mount.

At the end of the expire the covering mount probably has gone away so the
mounted status need not be restored. But we need to check this and only restore
the mounted status if the expire failed.

XXX: autofs may not work right if we have other mounts go over the top of it?
Signed-off-by: NNick Piggin <npiggin@kernel.dk>