Impose ordering on accesses of d_inode and d_flags to avoid the need to do
this:

	if (!dentry->d_inode || d_is_negative(dentry)) {

when this:

	if (d_is_negative(dentry)) {

should suffice.

This check is especially problematic if a dentry can have its type field set
to something other than DENTRY_MISS_TYPE when d_inode is NULL (as in
unionmount).

What we really need to do is stick a write barrier between setting d_inode and
setting d_flags and a read barrier between reading d_flags and reading
d_inode.
Signed-off-by: NDavid Howells <dhowells@redhat.com>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

On a distributed filesystem it's possible for lookup to discover that a
directory it just found is already cached elsewhere in the directory
heirarchy.  The dcache won't let us keep the directory in both places,
so we have to move the dentry to the new location from the place we
previously had it cached.

If the parent has changed, then this requires all the same locks as we'd
need to do a cross-directory rename.  But we're already in lookup
holding one parent's i_mutex, so it's too late to acquire those locks in
the right order.

The (unreliable) solution in __d_unalias is to trylock() the required
locks and return -EBUSY if it fails.

I see no particular reason for returning -EBUSY, and -ESTALE is already
the result of some other lookup races on NFS.  I think -ESTALE is the
more helpful error return.  It also allows us to take advantage of the
logic Jeff Layton added in c6a94284 "vfs: fix renameat to retry on
ESTALE errors" and ancestors, which hopefully resolves some of these
errors before they're returned to userspace.

I can reproduce these cases using NFS with:

	ssh root@$client '
		mount -olookupcache=pos '$server':'$export' /mnt/
		mkdir /mnt/TO
		mkdir /mnt/DIR
		touch /mnt/DIR/test.txt
		while true; do
			strace -e open cat /mnt/DIR/test.txt 2>&1 | grep EBUSY
		done
	'
	ssh root@$server '
		while true; do
			mv $export/DIR $export/TO/DIR
			mv $export/TO/DIR $export/DIR
		done
	'

It also helps to add some other concurrent use of the directory on the
client (e.g., "ls /mnt/TO").  And you can replace the server-side mv's
by client-side mv's that are repeatedly killed.  (If the client is
interrupted while waiting for the RENAME response then it's left with a
dentry that has to go under one parent or the other, but it doesn't yet
know which.)
Acked-by: NJeff Layton <jlayton@primarydata.com>
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

Split DCACHE_FILE_TYPE into DCACHE_REGULAR_TYPE (dentries representing regular
files) and DCACHE_SPECIAL_TYPE (representing blockdev, chardev, FIFO and
socket files).

d_is_reg() and d_is_special() are added to detect these subtypes and
d_is_file() is left as the union of the two.

This allows a number of places that use S_ISREG(dentry->d_inode->i_mode) to
use d_is_reg(dentry) instead.
Signed-off-by: NDavid Howells <dhowells@redhat.com>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

Add a DCACHE_FALLTHRU flag to indicate that, in a layered filesystem, this is
a virtual dentry that covers another one in a lower layer that should be used
instead.  This may be recorded on medium if directory integration is stored
there.

The flag can be set with d_set_fallthru() and tested with d_is_fallthru().

Original-author: Valerie Aurora <vaurora@redhat.com>
Signed-off-by: NDavid Howells <dhowells@redhat.com>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

We need to manually unpoison rounded up allocation size for dname to avoid
kasan's reports in dentry_string_cmp().  When CONFIG_DCACHE_WORD_ACCESS=y
dentry_string_cmp may access few bytes beyound requested in kmalloc()
size.

dentry_string_cmp() relates on that fact that dentry allocated using
kmalloc and kmalloc internally round up allocation size.  So this is not a
bug, but this makes kasan to complain about such accesses.  To avoid such
reports we mark rounded up allocation size in shadow as accessible.
Signed-off-by: NAndrey Ryabinin <a.ryabinin@samsung.com>
Reported-by: NDmitry Vyukov <dvyukov@google.com>
Cc: Konstantin Serebryany <kcc@google.com>
Cc: Dmitry Chernenkov <dmitryc@google.com>
Signed-off-by: NAndrey Konovalov <adech.fo@gmail.com>
Cc: Yuri Gribov <tetra2005@gmail.com>
Cc: Konstantin Khlebnikov <koct9i@gmail.com>
Cc: Sasha Levin <sasha.levin@oracle.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Currently, the isolate callback passed to the list_lru_walk family of
functions is supposed to just delete an item from the list upon returning
LRU_REMOVED or LRU_REMOVED_RETRY, while nr_items counter is fixed by
__list_lru_walk_one after the callback returns.  Since the callback is
allowed to drop the lock after removing an item (it has to return
LRU_REMOVED_RETRY then), the nr_items can be less than the actual number
of elements on the list even if we check them under the lock.  This makes
it difficult to move items from one list_lru_one to another, which is
required for per-memcg list_lru reparenting - we can't just splice the
lists, we have to move entries one by one.

This patch therefore introduces helpers that must be used by callback
functions to isolate items instead of raw list_del/list_move.  These are
list_lru_isolate and list_lru_isolate_move.  They not only remove the
entry from the list, but also fix the nr_items counter, making sure
nr_items always reflects the actual number of elements on the list if
checked under the appropriate lock.
Signed-off-by: NVladimir Davydov <vdavydov@parallels.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Tejun Heo <tj@kernel.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Dave Chinner <david@fromorbit.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Kmem accounting of memcg is unusable now, because it lacks slab shrinker
support.  That means when we hit the limit we will get ENOMEM w/o any
chance to recover.  What we should do then is to call shrink_slab, which
would reclaim old inode/dentry caches from this cgroup.  This is what
this patch set is intended to do.

Basically, it does two things.  First, it introduces the notion of
per-memcg slab shrinker.  A shrinker that wants to reclaim objects per
cgroup should mark itself as SHRINKER_MEMCG_AWARE.  Then it will be
passed the memory cgroup to scan from in shrink_control->memcg.  For
such shrinkers shrink_slab iterates over the whole cgroup subtree under
the target cgroup and calls the shrinker for each kmem-active memory
cgroup.

Secondly, this patch set makes the list_lru structure per-memcg.  It's
done transparently to list_lru users - everything they have to do is to
tell list_lru_init that they want memcg-aware list_lru.  Then the
list_lru will automatically distribute objects among per-memcg lists
basing on which cgroup the object is accounted to.  This way to make FS
shrinkers (icache, dcache) memcg-aware we only need to make them use
memcg-aware list_lru, and this is what this patch set does.

As before, this patch set only enables per-memcg kmem reclaim when the
pressure goes from memory.limit, not from memory.kmem.limit.  Handling
memory.kmem.limit is going to be tricky due to GFP_NOFS allocations, and
it is still unclear whether we will have this knob in the unified
hierarchy.

This patch (of 9):

NUMA aware slab shrinkers use the list_lru structure to distribute
objects coming from different NUMA nodes to different lists.  Whenever
such a shrinker needs to count or scan objects from a particular node,
it issues commands like this:

        count = list_lru_count_node(lru, sc->nid);
        freed = list_lru_walk_node(lru, sc->nid, isolate_func,
                                   isolate_arg, &sc->nr_to_scan);

where sc is an instance of the shrink_control structure passed to it
from vmscan.

To simplify this, let's add special list_lru functions to be used by
shrinkers, list_lru_shrink_count() and list_lru_shrink_walk(), which
consolidate the nid and nr_to_scan arguments in the shrink_control
structure.

This will also allow us to avoid patching shrinkers that use list_lru
when we make shrink_slab() per-memcg - all we will have to do is extend
the shrink_control structure to include the target memcg and make
list_lru_shrink_{count,walk} handle this appropriately.
Signed-off-by: NVladimir Davydov <vdavydov@parallels.com>
Suggested-by: NDave Chinner <david@fromorbit.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Greg Thelen <gthelen@google.com>
Cc: Glauber Costa <glommer@gmail.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

We can be more aggressive about this, if we are clever and careful. This is subtle.
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

no users left
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

In "d_prune_alias(): just lock the parent and call __dentry_kill()" the old
dget + d_drop + dput has been replaced with lock_parent + __dentry_kill;
unfortunately, dput() does more than just killing dentry - it also drops the
reference to parent.  New variant leaks that reference and needs dput(parent)
after killing the child off.
Signed-off-by: NYan, Zheng <zyan@redhat.com>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

This patch fixes kmemcheck warning in switch_names. The function
switch_names swaps inline names of two dentries. It swaps full arrays
d_iname, no matter how many bytes are really used by the strings. Reading
data beyond string ends results in kmemcheck warning.

We fix the bug by marking both arrays as fully initialized.
Signed-off-by: NMikulas Patocka <mpatocka@redhat.com>
Cc: stable@vger.kernel.org # v3.15
Signed-off-by: NAl 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>

... by not hitting rename_retry for reasons other than rename having
happened.  In other words, do _not_ restart when finding that
between unlocking the child and locking the parent the former got
into __dentry_kill().  Skip the killed siblings instead...
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

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

d_splice_alias() callers expect it to either stash the inode reference
into a new alias, or drop the inode reference.  That makes it possible
to just return d_splice_alias() result from ->lookup() instance, without
any extra housekeeping required.

Unfortunately, that should include the failure exits.  If d_splice_alias()
returns an error, it leaves the dentry it has been given negative and
thus it *must* drop the inode reference.  Easily fixed, but it goes way
back and will need backporting.

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

never used outside and it's too low-level for legitimate uses outside
of fs/dcache.c anyway
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

Fixed coding style in dcache.c
Signed-off-by: NDaeseok Youn <daeseok.youn@gmail.com>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

the only in-tree instance checks d_unhashed() anyway,
out-of-tree code can preserve the current behaviour by
adding such check if they want it and we get an ability
to use it in cases where we *want* to be notified of
killing being inevitable before ->d_lock is dropped,
whether it's unhashed or not.  In particular, autofs
would benefit from that.
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

The only reason for games with ->d_prune() was __d_drop(), which
was needed only to force dput() into killing the sucker off.

Note that lock_parent() can be called under ->i_lock and won't
drop it, so dentry is safe from somebody managing to kill it
under us - it won't happen while we are holding ->i_lock.

__dentry_kill() is called only with ->d_lockref.count being 0
(here and when picked from shrink list) or 1 (dput() and dropping
the ancestors in shrink_dentry_list()), so it will never be called
twice - the first thing it's doing is making ->d_lockref.count
negative and once that happens, nothing will increment it.
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

Now that d_invalidate can no longer fail, stop returning a useless
return code.  For the few callers that checked the return code update
remove the handling of d_invalidate failure.
Reviewed-by: NMiklos Szeredi <miklos@szeredi.hu>
Signed-off-by: N"Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

Now that d_invalidate is the only caller of check_submounts_and_drop,
expand check_submounts_and_drop inline in d_invalidate.
Reviewed-by: NMiklos Szeredi <miklos@szeredi.hu>
Signed-off-by: N"Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

With the introduction of mount namespaces and bind mounts it became
possible to access files and directories that on some paths are mount
points but are not mount points on other paths.  It is very confusing
when rm -rf somedir returns -EBUSY simply because somedir is mounted
somewhere else.  With the addition of user namespaces allowing
unprivileged mounts this condition has gone from annoying to allowing
a DOS attack on other users in the system.

The possibility for mischief is removed by updating the vfs to support
rename, unlink and rmdir on a dentry that is a mountpoint and by
lazily unmounting mountpoints on deleted dentries.

In particular this change allows rename, unlink and rmdir system calls
on a dentry without a mountpoint in the current mount namespace to
succeed, and it allows rename, unlink, and rmdir performed on a
distributed filesystem to update the vfs cache even if when there is a
mount in some namespace on the original dentry.

There are two common patterns of maintaining mounts: Mounts on trusted
paths with the parent directory of the mount point and all ancestory
directories up to / owned by root and modifiable only by root
(i.e. /media/xxx, /dev, /dev/pts, /proc, /sys, /sys/fs/cgroup/{cpu,
cpuacct, ...}, /usr, /usr/local).  Mounts on unprivileged directories
maintained by fusermount.

In the case of mounts in trusted directories owned by root and
modifiable only by root the current parent directory permissions are
sufficient to ensure a mount point on a trusted path is not removed
or renamed by anyone other than root, even if there is a context
where the there are no mount points to prevent this.

In the case of mounts in directories owned by less privileged users
races with users modifying the path of a mount point are already a
danger.  fusermount already uses a combination of chdir,
/proc/<pid>/fd/NNN, and UMOUNT_NOFOLLOW to prevent these races.  The
removable of global rename, unlink, and rmdir protection really adds
nothing new to consider only a widening of the attack window, and
fusermount is already safe against unprivileged users modifying the
directory simultaneously.

In principle for perfect userspace programs returning -EBUSY for
unlink, rmdir, and rename of dentires that have mounts in the local
namespace is actually unnecessary.  Unfortunately not all userspace
programs are perfect so retaining -EBUSY for unlink, rmdir and rename
of dentries that have mounts in the current mount namespace plays an
important role of maintaining consistency with historical behavior and
making imperfect userspace applications hard to exploit.

v2: Remove spurious old_dentry.
v3: Optimized shrink_submounts_and_drop
    Removed unsued afs label
v4: Simplified the changes to check_submounts_and_drop
    Do not rename check_submounts_and_drop shrink_submounts_and_drop
    Document what why we need atomicity in check_submounts_and_drop
    Rely on the parent inode mutex to make d_revalidate and d_invalidate
    an atomic unit.
v5: Refcount the mountpoint to detach in case of simultaneous
    renames.
Reviewed-by: NMiklos Szeredi <miklos@szeredi.hu>
Signed-off-by: N"Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

The current comments in d_invalidate about what and why it is doing
what it is doing are wildly off-base.  Which is not surprising as
the comments date back to last minute bug fix of the 2.2 kernel.

The big fat lie of a comment said: If it's a directory, we can't drop
it for fear of somebody re-populating it with children (even though
dropping it would make it unreachable from that root, we still might
repopulate it if it was a working directory or similar).

[AV] What we really need to avoid is multiple dentry aliases of the
same directory inode; on all filesystems that have ->d_revalidate()
we either declare all positive dentries always valid (and thus never
fed to d_invalidate()) or use d_materialise_unique() and/or d_splice_alias(),
which take care of alias prevention.

The current rules are:
- To prevent mount point leaks dentries that are mount points or that
  have childrent that are mount points may not be be unhashed.
- All dentries may be unhashed.
- Directories may be rehashed with d_materialise_unique

check_submounts_and_drop implements this already for well maintained
remote filesystems so implement the current rules in d_invalidate
by just calling check_submounts_and_drop.

The one difference between d_invalidate and check_submounts_and_drop
is that d_invalidate must respect it when a d_revalidate method has
earlier called d_drop so preserve the d_unhashed check in
d_invalidate.
Reviewed-by: NMiklos Szeredi <miklos@szeredi.hu>
Signed-off-by: N"Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

d_drop or check_submounts_and_drop called from d_revalidate can result
in renamed directories with child dentries being unhashed.  These
renamed and drop directory dentries can be rehashed after
d_materialise_unique uses d_find_alias to find them.
Reviewed-by: NMiklos Szeredi <miklos@szeredi.hu>
Signed-off-by: N"Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

* external dentry names get a small structure prepended to them
(struct external_name).
* it contains an atomic refcount, matching the number of struct dentry
instances that have ->d_name.name pointing to that external name.  The
first thing free_dentry() does is decrementing refcount of external name,
so the instances that are between the call of free_dentry() and
RCU-delayed actual freeing do not contribute.
* __d_move(x, y, false) makes the name of x equal to the name of y,
external or not.  If y has an external name, extra reference is grabbed
and put into x->d_name.name.  If x used to have an external name, the
reference to the old name is dropped and, should it reach zero, freeing
is scheduled via kfree_rcu().
* free_dentry() in dentry with external name decrements the refcount of
that name and, should it reach zero, does RCU-delayed call that will
free both the dentry and external name.  Otherwise it does what it
used to do, except that __d_free() doesn't even look at ->d_name.name;
it simply frees the dentry.

All non-RCU accesses to dentry external name are safe wrt freeing since they
all should happen before free_dentry() is called.  RCU accesses might run
into a dentry seen by free_dentry() or into an old name that got already
dropped by __d_move(); however, in both cases dentry must have been
alive and refer to that name at some point after we'd done rcu_read_lock(),
which means that any freeing must be still pending.
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

AFAICS, prepend_name() is broken on SMP alpha.  Disclaimer: I don't have
SMP alpha boxen to reproduce it on.  However, it really looks like the race
is real.

CPU1: d_path() on /mnt/ramfs/<255-character>/foo
CPU2: mv /mnt/ramfs/<255-character> /mnt/ramfs/<63-character>

CPU2 does d_alloc(), which allocates an external name, stores the name there
including terminating NUL, does smp_wmb() and stores its address in
dentry->d_name.name.  It proceeds to d_add(dentry, NULL) and d_move()
old dentry over to that.  ->d_name.name value ends up in that dentry.

In the meanwhile, CPU1 gets to prepend_name() for that dentry.  It fetches
->d_name.name and ->d_name.len; the former ends up pointing to new name
(64-byte kmalloc'ed array), the latter - 255 (length of the old name).
Nothing to force the ordering there, and normally that would be OK, since we'd
run into the terminating NUL and stop.  Except that it's alpha, and we'd need
a data dependency barrier to guarantee that we see that store of NUL
__d_alloc() has done.  In a similar situation dentry_cmp() would survive; it
does explicit smp_read_barrier_depends() after fetching ->d_name.name.
prepend_name() doesn't and it risks walking past the end of kmalloc'ed object
and possibly oops due to taking a page fault in kernel mode.

Cc: stable@vger.kernel.org # 3.12+
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

Only exchange source and destination filenames
if flags contain RENAME_EXCHANGE.
In case if executable file was running and replaced by
other file /proc/PID/exe should still show correct file name,
not the old name of the file by which it was replaced.

The scenario when this bug manifests itself was like this:
* ALT Linux uses rpm and start-stop-daemon;
* during a package upgrade rpm creates a temporary file
  for an executable to rename it upon successful unpacking;
* start-stop-daemon is run subsequently and it obtains
  the (nonexistant) temporary filename via /proc/PID/exe
  thus failing to identify the running process.

Note that "long" filenames (> DNAiME_INLINE_LEN) are still
exchanged without RENAME_EXCHANGE and this behaviour exists
long enough (should be fixed too apparently).
So this patch is just an interim workaround that restores
behavior for "short" names as it was before changes
introduced by commit da1ce067 ("vfs: add cross-rename").

See https://lkml.org/lkml/2014/9/7/6 for details.

AV: the comments about being more careful with ->d_name.hash
than with ->d_name.name are from back in 2.3.40s; they
became obsolete by 2.3.60s, when we started to unhash the
target instead of swapping hash chain positions followed
by d_delete() as we used to do when dcache was first
introduced.
Acked-by: NMiklos Szeredi <mszeredi@suse.cz>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: linux-fsdevel@vger.kernel.org
Cc: stable@vger.kernel.org
Fixes: da1ce067 "vfs: add cross-rename"
Signed-off-by: NMikhail Efremov <sem@altlinux.org>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

and do it along with ->d_name.len there
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

... renaming it into dentry_unlock_for_move() and making it more
symmetric with dentry_lock_for_move().
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

it folds into __d_move() now
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

... thus making it much closer to (now unreachable, BTW) IS_ROOT(dentry)
case in __d_move().  A bit more and it'll fold in.
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

list_del() + list_add() is a slightly pessimised list_move()
list_del() + INIT_LIST_HEAD() is a slightly pessimised list_del_init()

Interleaving those makes the resulting code even worse.  And harder to follow...
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

... and get rid of duplicate BUG_ON() there
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

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

and lock the right list there

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

Josef Bacik found a performance regression between 3.2 and 3.10 and
narrowed it down to commit bfcfaa77 ("vfs: use 'unsigned long'
accesses for dcache name comparison and hashing"). He reports:

 "The test case is essentially

      for (i = 0; i < 1000000; i++)
              mkdir("a$i");

  On xfs on a fio card this goes at about 20k dir/sec with 3.2, and 12k
  dir/sec with 3.10.  This is because we spend waaaaay more time in
  __d_lookup on 3.10 than in 3.2.

  The new hashing function for strings is suboptimal for <
  sizeof(unsigned long) string names (and hell even > sizeof(unsigned
  long) string names that I've tested).  I broke out the old hashing
  function and the new one into a userspace helper to get real numbers
  and this is what I'm getting:

      Old hash table had 1000000 entries, 0 dupes, 0 max dupes
      New hash table had 12628 entries, 987372 dupes, 900 max dupes
      We had 11400 buckets with a p50 of 30 dupes, p90 of 240 dupes, p99 of 567 dupes for the new hash

  My test does the hash, and then does the d_hash into a integer pointer
  array the same size as the dentry hash table on my system, and then
  just increments the value at the address we got to see how many
  entries we overlap with.

  As you can see the old hash function ended up with all 1 million
  entries in their own bucket, whereas the new one they are only
  distributed among ~12.5k buckets, which is why we're using so much
  more CPU in __d_lookup".

The reason for this hash regression is two-fold:

 - On 64-bit architectures the down-mixing of the original 64-bit
   word-at-a-time hash into the final 32-bit hash value is very
   simplistic and suboptimal, and just adds the two 32-bit parts
   together.

   In particular, because there is no bit shuffling and the mixing
   boundary is also a byte boundary, similar character patterns in the
   low and high word easily end up just canceling each other out.

 - the old byte-at-a-time hash mixed each byte into the final hash as it
   hashed the path component name, resulting in the low bits of the hash
   generally being a good source of hash data.  That is not true for the
   word-at-a-time case, and the hash data is distributed among all the
   bits.

The fix is the same in both cases: do a better job of mixing the bits up
and using as much of the hash data as possible.  We already have the
"hash_32|64()" functions to do that.
Reported-by: NJosef Bacik <jbacik@fb.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Chris Mason <clm@fb.com>
Cc: linux-fsdevel@vger.kernel.org
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Signed-off-by: NFengguang Wu <fengguang.wu@intel.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

I believe this can only happen in the case of a corrupted filesystem.
So -EIO looks like the appropriate error.
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

If we get to this point and discover the dentry is not a root dentry, or
not DCACHE_DISCONNECTED--great, we always prefer that anyway.
Signed-off-by: NJ. Bruce Fields <bfields@redhat.com>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>