With the dentry LRUs being per-sb structures, there is no real need for
a global dentry_lru_lock. The locking can be made more fine-grained by
moving to a per-sb LRU lock, isolating the LRU operations of different
filesytsems completely from each other. The need for this is independent
of any performance consideration that may arise: in the interest of
abstracting the lru operations away, it is mandatory that each lru works
around its own lock instead of a global lock for all of them.
[glommer@openvz.org: updated changelog ]
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Glauber Costa <glommer@openvz.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
This series reworks our current object cache shrinking infrastructure in
two main ways:
* Noticing that a lot of users copy and paste their own version of LRU
lists for objects, we put some effort in providing a generic version.
It is modeled after the filesystem users: dentries, inodes, and xfs
(for various tasks), but we expect that other users could benefit in
the near future with little or no modification. Let us know if you
have any issues.
* The underlying list_lru being proposed automatically and
transparently keeps the elements in per-node lists, and is able to
manipulate the node lists individually. Given this infrastructure, we
are able to modify the up-to-now hammer called shrink_slab to proceed
with node-reclaim instead of always searching memory from all over like
it has been doing.
Per-node lru lists are also expected to lead to less contention in the lru
locks on multi-node scans, since we are now no longer fighting for a
global lock. The locks usually disappear from the profilers with this
change.
Although we have no official benchmarks for this version - be our guest to
independently evaluate this - earlier versions of this series were
performance tested (details at
http://permalink.gmane.org/gmane.linux.kernel.mm/100537) yielding no
visible performance regressions while yielding a better qualitative
behavior in NUMA machines.
With this infrastructure in place, we can use the list_lru entry point to
provide memcg isolation and per-memcg targeted reclaim. Historically,
those two pieces of work have been posted together. This version presents
only the infrastructure work, deferring the memcg work for a later time,
so we can focus on getting this part tested. You can see more about the
history of such work at http://lwn.net/Articles/552769/
Dave Chinner (18):
dcache: convert dentry_stat.nr_unused to per-cpu counters
dentry: move to per-sb LRU locks
dcache: remove dentries from LRU before putting on dispose list
mm: new shrinker API
shrinker: convert superblock shrinkers to new API
list: add a new LRU list type
inode: convert inode lru list to generic lru list code.
dcache: convert to use new lru list infrastructure
list_lru: per-node list infrastructure
shrinker: add node awareness
fs: convert inode and dentry shrinking to be node aware
xfs: convert buftarg LRU to generic code
xfs: rework buffer dispose list tracking
xfs: convert dquot cache lru to list_lru
fs: convert fs shrinkers to new scan/count API
drivers: convert shrinkers to new count/scan API
shrinker: convert remaining shrinkers to count/scan API
shrinker: Kill old ->shrink API.
Glauber Costa (7):
fs: bump inode and dentry counters to long
super: fix calculation of shrinkable objects for small numbers
list_lru: per-node API
vmscan: per-node deferred work
i915: bail out earlier when shrinker cannot acquire mutex
hugepage: convert huge zero page shrinker to new shrinker API
list_lru: dynamically adjust node arrays
This patch:
There are situations in very large machines in which we can have a large
quantity of dirty inodes, unused dentries, etc. This is particularly true
when umounting a filesystem, where eventually since every live object will
eventually be discarded.
Dave Chinner reported a problem with this while experimenting with the
shrinker revamp patchset. So we believe it is time for a change. This
patch just moves int to longs. Machines where it matters should have a
big long anyway.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Pull vfs pile 3 (of many) from Al Viro:
"Waiman's conversion of d_path() and bits related to it,
kern_path_mountpoint(), several cleanups and fixes (exportfs
one is -stable fodder, IMO).
There definitely will be more... ;-/"
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs:
split read_seqretry_or_unlock(), convert d_walk() to resulting primitives
dcache: Translating dentry into pathname without taking rename_lock
autofs4 - fix device ioctl mount lookup
introduce kern_path_mountpoint()
rename user_path_umountat() to user_path_mountpoint_at()
take unlazy_walk() into umount_lookup_last()
Kill indirect include of file.h from eventfd.h, use fdget() in cgroup.c
prune_super(): sb->s_op is never NULL
exportfs: don't assume that ->iterate() won't feed us too long entries
afs: get rid of redundant ->d_name.len checks
Separate "check if we need to retry" from "unlock if we are done and
had seq_writelock"; that allows to use these guys in d_walk(), where
we need to recheck every time we ascend back to parent, but do *not*
want to unlock until the very end. Lift rcu_read_lock/rcu_read_unlock
out into callers.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
When running the AIM7's short workload, Linus' lockref patch eliminated
most of the spinlock contention. However, there were still some left:
8.46% reaim [kernel.kallsyms] [k] _raw_spin_lock
|--42.21%-- d_path
| proc_pid_readlink
| SyS_readlinkat
| SyS_readlink
| system_call
| __GI___readlink
|
|--40.97%-- sys_getcwd
| system_call
| __getcwd
The big one here is the rename_lock (seqlock) contention in d_path()
and the getcwd system call. This patch will eliminate the need to take
the rename_lock while translating dentries into the full pathnames.
The need to take the rename_lock is to make sure that no rename
operation can be ongoing while the translation is in progress. However,
only one thread can take the rename_lock thus blocking all the other
threads that need it even though the translation process won't make
any change to the dentries.
This patch will replace the writer's write_seqlock/write_sequnlock
sequence of the rename_lock of the callers of the prepend_path() and
__dentry_path() functions with the reader's read_seqbegin/read_seqretry
sequence within these 2 functions. As a result, the code will have to
retry if one or more rename operations had been performed. In addition,
RCU read lock will be taken during the translation process to make sure
that no dentries will go away. To prevent live-lock from happening,
the code will switch back to take the rename_lock if read_seqretry()
fails for three times.
To further reduce spinlock contention, this patch does not take the
dentry's d_lock when copying the filename from the dentries. Instead,
it treats the name pointer and length as unreliable and just copy
the string byte-by-byte over until it hits a null byte or the end of
string as specified by the length. This should avoid stepping into
invalid memory address. The error cases are left to be handled by
the sequence number check.
The following code re-factoring are also made:
1. Move prepend('/') into prepend_name() to remove one conditional
check.
2. Move the global root check in prepend_path() back to the top of
the while loop.
With this patch, the _raw_spin_lock will now account for only 1.2%
of the total CPU cycles for the short workload. This patch also has
the effect of reducing the effect of running perf on its profile
since the perf command itself can be a heavy user of the d_path()
function depending on the complexity of the workload.
When taking the perf profile of the high-systime workload, the amount
of spinlock contention contributed by running perf without this patch
was about 16%. With this patch, the spinlock contention caused by
the running of perf will go away and we will have a more accurate
perf profile.
Signed-off-by: Waiman Long <Waiman.Long@hp.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
This simplifies the RCU to refcounting code in particular.
I was originally intending to leave this for later, but walking through
all the dput() logic (see previous commit), I realized that the dput()
"might_sleep()" check was misleadingly weak. And I removed it as
misleading, both for performance profiling and for debugging.
However, the might_sleep() debugging case is actually true: the final
dput() can indeed sleep, if the inode of the dentry that you are
releasing ends up sleeping at iput time (see dentry_iput()). So the
problem with the might_sleep() in dput() wasn't that it wasn't true, it
was that it wasn't actually testing and triggering on the interesting
case.
In particular, just about *any* dput() can indeed sleep, if you happen
to race with another thread deleting the file in question, and you then
lose the race to the be the last dput() for that file. But because it's
a very rare race, the debugging code would never trigger it in practice.
Why is this problematic? The new d_rcu_to_refcount() (see commit
15570086b5: "vfs: reimplement d_rcu_to_refcount() using
lockref_get_or_lock()") does a dput() for the failure case, and it does
it under the RCU lock. So potentially sleeping really is a bug.
But there's no way I'm going to fix this with the previous complicated
"lockref_get_or_lock()" interface. And rather than revert to the old
and crufty nested dentry locking code (which did get this right by
delaying the reference count updates until they were verified to be
safe), let's make forward progress.
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This is me being a bit OCD after all the dentry optimization work this
merge window: profiles end up showing 'dput()' as a rather expensive
operation, and there were two unrelated bad reasons for that.
The first reason was reading d_lockref.count for debugging purposes,
which touches the lockref cacheline (for reads) before really need to.
More importantly, the debugging test in question is _wrong_, and has
hidden bugs. It's true that we can only sleep when the count goes down
to zero, but the test as-is hides the much more subtle bug that happens
if we race with somebody else deleting the file.
Anyway we _will_ touch that cacheline, but let's do it for a write and
in the right routine (ie in "lockref_put_or_lock()") which annotates the
costs better. So remove the misleading debug code.
The other was an unnecessary access to the cacheline that contains the
d_lru list, just to check whether we already were on the LRU list or
not. This is exactly what we have d_flags for, so that we can avoid
touching extra cache lines for the common case. So just add another bit
for "is this dentry on the LRU".
Finally, mark the tests properly likely/unlikely, so that the common
fast-paths are dense in the instruction stream.
This makes the profiles look much saner.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
We check submounts before doing d_drop() on a non-empty directory dentry in
NFS (have_submounts()), but we do not exclude a racing mount. Nor do we
prevent mounts to be added to the disconnected subtree using relative paths
after the d_drop().
This patch fixes these issues by checking for unlinked (unhashed, non-root)
ancestors before proceeding with the mount. This is done with rename
seqlock taken for write and with ->d_lock grabbed on each ancestor in turn,
including our dentry itself. This ensures that the only one of
check_submounts_and_drop() or has_unlinked_ancestor() can succeed.
Signed-off-by: Miklos Szeredi <miklos@szeredi.hu>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
We check submounts before doing d_drop() on a non-empty directory dentry in
NFS (have_submounts()), but we do not exclude a racing mount.
Process A: have_submounts() -> returns false
Process B: mount() -> success
Process A: d_drop()
This patch prepares the ground for the fix by doing the following
operations all under the same rename lock:
have_submounts()
shrink_dcache_parent()
d_drop()
This is actually an optimization since have_submounts() and
shrink_dcache_parent() both traverse the same dentry tree separately.
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
CC: David Howells <dhowells@redhat.com>
CC: Steven Whitehouse <swhiteho@redhat.com>
CC: Trond Myklebust <Trond.Myklebust@netapp.com>
CC: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
This one replaces three instances open coded tree walking (have_submounts,
select_parent, d_genocide) with a common helper.
In addition to slightly reducing the kernel size, this simplifies the
callers and makes them less bug prone.
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
It shouldn't matter when we decrement the refcount during the walk as long
as we do it exactly once.
Restructure d_genocide() to do the killing on entering the dentry instead
of when leaving it. This helps creating a common helper for tree walking.
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Pull vfs pile 1 from Al Viro:
"Unfortunately, this merge window it'll have a be a lot of small piles -
my fault, actually, for not keeping #for-next in anything that would
resemble a sane shape ;-/
This pile: assorted fixes (the first 3 are -stable fodder, IMO) and
cleanups + %pd/%pD formats (dentry/file pathname, up to 4 last
components) + several long-standing patches from various folks.
There definitely will be a lot more (starting with Miklos'
check_submount_and_drop() series)"
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs: (26 commits)
direct-io: Handle O_(D)SYNC AIO
direct-io: Implement generic deferred AIO completions
add formats for dentry/file pathnames
kvm eventfd: switch to fdget
powerpc kvm: use fdget
switch fchmod() to fdget
switch epoll_ctl() to fdget
switch copy_module_from_fd() to fdget
git simplify nilfs check for busy subtree
ibmasmfs: don't bother passing superblock when not needed
don't pass superblock to hypfs_{mkdir,create*}
don't pass superblock to hypfs_diag_create_files
don't pass superblock to hypfs_vm_create_files()
oprofile: get rid of pointless forward declarations of struct super_block
oprofilefs_create_...() do not need superblock argument
oprofilefs_mkdir() doesn't need superblock argument
don't bother with passing superblock to oprofile_create_stats_files()
oprofile: don't bother with passing superblock to ->create_files()
don't bother passing sb to oprofile_create_files()
coh901318: don't open-code simple_read_from_buffer()
...
The d_prune dentry operation is used to notify filesystem when VFS
about to prune a hashed dentry from the dcache. There are three
code paths that prune dentries: shrink_dcache_for_umount_subtree(),
prune_dcache_sb() and d_prune_aliases(). For the d_prune_aliases()
case, VFS unhashes the dentry first, then call the d_prune dentry
operation. This confuses ceph_d_prune() (ceph uses the d_prune
dentry operation to maintain a flag indicating whether the complete
contents of a directory are in the dcache, pruning unhashed dentry
does not affect dir's completeness)
This patch fixes the issue by calling the d_prune dentry operation
in d_prune_aliases(), before unhashing the dentry. Also make VFS
only call the d_prune dentry operation for hashed dentry, to avoid
calling the d_prune dentry operation twice when dentry is pruned
by d_prune_aliases().
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
This moves __d_rcu_to_refcount() from <linux/dcache.h> into fs/namei.c
and re-implements it using the lockref infrastructure instead. It also
adds a lot of comments about what is actually going on, because turning
a dentry that was looked up using RCU into a long-lived reference
counted entry is one of the more subtle parts of the rcu walk.
We also used to be _particularly_ subtle in unlazy_walk() where we
re-validate both the dentry and its parent using the same sequence
count. We used to do it by nesting the locks and then verifying the
sequence count just once.
That was silly, because nested locking is expensive, but the sequence
count check is not. So this just re-validates the dentry and the parent
separately, avoiding the nested locking, and making the lockref lookup
possible.
Acked-by: Waiman Long <waiman.long@hp.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
A valid parent pointer is always going to have a non-zero reference
count, but if we look up the parent optimistically without locking, we
have to protect against the (very unlikely) race against renaming
changing the parent from under us.
We do that by using lockref_get_not_zero(), and then re-checking the
parent pointer after getting a valid reference.
[ This is a re-implementation of a chunk from the original patch by
Waiman Long: "dcache: Enable lockless update of dentry's refcount".
I've completely rewritten the patch-series and split it up, but I'm
attributing this part to Waiman as it's close enough to his earlier
patch - Linus ]
Signed-off-by: Waiman Long <Waiman.Long@hp.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This just replaces the dentry count/lock combination with the lockref
structure that contains both a count and a spinlock, and does the
mechanical conversion to use the lockref infrastructure.
There are no semantic changes here, it's purely syntactic. The
reference lockref implementation uses the spinlock exactly the same way
that the old dcache code did, and the bulk of this patch is just
expanding the internal "d_count" use in the dcache code to use
"d_lockref.count" instead.
This is purely preparation for the real change to make the reference
count updates be lockless during the 3.12 merge window.
[ As with the previous commit, this is a rewritten version of a concept
originally from Waiman, so credit goes to him, blame for any errors
goes to me.
Waiman's patch had some semantic differences for taking advantage of
the lockless update in dget_parent(), while this patch is
intentionally a pure search-and-replace change with no semantic
changes. - Linus ]
Signed-off-by: Waiman Long <Waiman.Long@hp.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
dynamic_dname() is both too much and too little for those - the
output may be well in excess of 64 bytes dynamic_dname() assumes
to be enough (thanks to ashmem feeding really long names to
shmem_file_setup()) and vsnprintf() is an overkill for those
guys.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Pull second set of VFS changes from Al Viro:
"Assorted f_pos race fixes, making do_splice_direct() safe to call with
i_mutex on parent, O_TMPFILE support, Jeff's locks.c series,
->d_hash/->d_compare calling conventions changes from Linus, misc
stuff all over the place."
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs: (63 commits)
Document ->tmpfile()
ext4: ->tmpfile() support
vfs: export lseek_execute() to modules
lseek_execute() doesn't need an inode passed to it
block_dev: switch to fixed_size_llseek()
cpqphp_sysfs: switch to fixed_size_llseek()
tile-srom: switch to fixed_size_llseek()
proc_powerpc: switch to fixed_size_llseek()
ubi/cdev: switch to fixed_size_llseek()
pci/proc: switch to fixed_size_llseek()
isapnp: switch to fixed_size_llseek()
lpfc: switch to fixed_size_llseek()
locks: give the blocked_hash its own spinlock
locks: add a new "lm_owner_key" lock operation
locks: turn the blocked_list into a hashtable
locks: convert fl_link to a hlist_node
locks: avoid taking global lock if possible when waking up blocked waiters
locks: protect most of the file_lock handling with i_lock
locks: encapsulate the fl_link list handling
locks: make "added" in __posix_lock_file a bool
...
Instances either don't look at it at all (the majority of cases) or
only want it to find the superblock (which can be had as dentry->d_sb).
A few cases that want more are actually safe with dentry->d_inode -
the only precaution needed is the check that it hadn't been replaced with
NULL by rmdir() or by overwriting rename(), which case should be simply
treated as cache miss.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
the only remaining caller (in ncpfs) is guaranteed to return 0 -
we only hit it if we'd just checked that there's no dentry with
such name.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
I've restricted atomic_open to only operate on regular files, although
I still don't understand why atomic_open should not be possible also for
directories on GFS2. That can always be added in later though, if it
makes sense.
The ->atomic_open function can be passed negative dentries, which
in most cases means either ENOENT (->lookup) or a call to d_instantiate
(->create). In the GFS2 case though, we need to actually perform the
look up, since we do not know whether there has been a new inode created
on another node. The look up calls d_splice_alias which then tries to
rehash the dentry - so the solution here is to simply check for that
in d_splice_alias. The same issue is likely to affect any other cluster
filesystem implementing ->atomic_open
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: "J. Bruce Fields" <bfields fieldses org>
Cc: Jeff Layton <jlayton@redhat.com>
When pruning a dentry, its ancestor dentry can also be pruned. But
the ancestor dentry does not go through dput(), so it does not get
put on the dentry LRU. Hence associating d_prune with removing the
dentry from the LRU is the wrong.
The fix is remove dentry_lru_prune(). Call file system's d_prune()
callback directly when pruning dentries.
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Dave Chinner
Glauber Costa
Linus Torvalds
Al Viro
Waiman Long
Miklos Szeredi
Miklos Szeredi
Yan, Zheng
Steven Whitehouse
Wei Yongjun