I discovered on oprofile hunting on a SMP platform that dentry lookups were
slowed down because d_hash_mask, d_hash_shift and dentry_hashtable were in
a cache line that contained inodes_stat. So each time inodes_stats is
changed by a cpu, other cpus have to refill their cache line.
This patch moves some variables to the __read_mostly section, in order to
avoid false sharing. RCU dentry lookups can go full speed.
Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Currently lockd directly access the file_lock_list from fs/locks.c.
It does so to mark locks granted or reclaimable. This is very
suboptimal, because a) lockd needs to poke into locks.c internals, and
b) it needs to iterate over all locks in the system for marking locks
granted or reclaimable.
This patch adds lists for granted and reclaimable locks to the nlm_host
structure instead, and adds locks to those.
nlmclnt_lock:
now adds the lock to h_granted instead of setting the
NFS_LCK_GRANTED, still O(1)
nlmclnt_mark_reclaim:
goes away completely, replaced by a list_splice_init.
Complexity reduced from O(locks in the system) to O(1)
reclaimer:
iterates over h_reclaim now, complexity reduced from
O(locks in the system) to O(locks per nlm_host)
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
The caller of posix_test_lock() should never need to look at the lock
private data, so do not copy that information. This also means that there
is no need to call the fl_release_private methods.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
posix_test_lock() returns a pointer to a struct file_lock which is unprotected
and can be removed while in use by the caller. Move the conflicting lock from
the return to a parameter, and copy the conflicting lock.
In most cases the caller ends up putting the copy of the conflicting lock on
the stack. On i386, sizeof(struct file_lock) appears to be about 100 bytes.
We're assuming that's reasonable.
Signed-off-by: Andy Adamson <andros@citi.umich.edu>
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
posix_lock_file() is used to add a blocked lock to Lockd's block, so
posix_block_lock() is no longer needed.
Signed-off-by: Andy Adamson <andros@citi.umich.edu>
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
If the server receives an NLM cancel call and finds no waiting lock to
cancel, then chances are the lock has already been applied, and the client
just hadn't yet processed the NLM granted callback before it sent the
cancel.
The Open Group text, for example, perimts a server to return either success
(LCK_GRANTED) or failure (LCK_DENIED) in this case. But returning an error
seems more helpful; the client may be able to use it to recognize that a
race has occurred and to recover from the race.
So, modify the relevant functions to return an error in this case.
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Currently when lockd gets an NLM_CANCEL request, it also does an unlock for
the same range. This is incorrect.
The Open Group documentation says that "This procedure cancels an
*outstanding* blocked lock request." (Emphasis mine.)
Also, consider a client that holds a lock on the first byte of a file, and
requests a lock on the entire file. If the client cancels that request
(perhaps because the requesting process is signalled), the server shouldn't
apply perform an unlock on the entire file, since that will also remove the
previous lock that the client was already granted.
Or consider a lock request that actually *downgraded* an exclusive lock to
a shared lock.
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
The patch
http://linux.bkbits.net:8080/linux-2.6/diffs/fs/locks.c@1.70??nav=index.html
introduced a pretty nasty memory leak in the lease code. When freeing
the lease, the code in locks_delete_lock() will correctly clean up
the fasync queue, but when we return to fcntl_setlease(), the freed
fasync entry will be reinstated.
This patch ensures that we skip the call to fasync_helper() when we're
freeing up the lease.
Signed-off-by: J. Bruce Fields <bfields@fieldses.org>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
We currently fail Connectathon test 6.10 in the case of 32-bit locks due
to incorrect error checking.
Also add support for l->l_len < 0 to 64-bit locks.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
[PATCH] Fix miscompare in __posix_lock_file
If an application requests the same lock twice, the
kernel should just leave the existing lock in place.
Currently, it will install a second lock of the same type.
Signed-off-by: Olaf Kirch <okir@suse.de>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
With the new fdtable locking rules, you have to protect fdtable with either
->file_lock or rcu_read_lock/unlock(). There are some places where we
aren't doing either. This patch fixes those places.
Signed-off-by: Dipankar Sarma <dipankar@in.ibm.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
In order for the RCU to work, the file table array, sets and their sizes must
be updated atomically. Instead of ensuring this through too many memory
barriers, we put the arrays and their sizes in a separate structure. This
patch takes the first step of putting the file table elements in a separate
structure fdtable that is embedded withing files_struct. It also changes all
the users to refer to the file table using files_fdtable() macro. Subsequent
applciation of RCU becomes easier after this.
Signed-off-by: Dipankar Sarma <dipankar@in.ibm.com>
Signed-Off-By: David Howells <dhowells@redhat.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
I believe that there is a problem with the handling of POSIX locks, which
the attached patch should address.
The problem appears to be a race between fcntl(2) and close(2). A
multithreaded application could close a file descriptor at the same time as
it is trying to acquire a lock using the same file descriptor. I would
suggest that that multithreaded application is not providing the proper
synchronization for itself, but the OS should still behave correctly.
SUS3 (Single UNIX Specification Version 3, read: POSIX) indicates that when
a file descriptor is closed, that all POSIX locks on the file, owned by the
process which closed the file descriptor, should be released.
The trick here is when those locks are released. The current code releases
all locks which exist when close is processing, but any locks in progress
are handled when the last reference to the open file is released.
There are three cases to consider.
One is the simple case, a multithreaded (mt) process has a file open and
races to close it and acquire a lock on it. In this case, the close will
release one reference to the open file and when the fcntl is done, it will
release the other reference. For this situation, no locks should exist on
the file when both the close and fcntl operations are done. The current
system will handle this case because the last reference to the open file is
being released.
The second case is when the mt process has dup(2)'d the file descriptor.
The close will release one reference to the file and the fcntl, when done,
will release another, but there will still be at least one more reference
to the open file. One could argue that the existence of a lock on the file
after the close has completed is okay, because it was acquired after the
close operation and there is still a way for the application to release the
lock on the file, using an existing file descriptor.
The third case is when the mt process has forked, after opening the file
and either before or after becoming an mt process. In this case, each
process would hold a reference to the open file. For each process, this
degenerates to first case above. However, the lock continues to exist
until both processes have released their references to the open file. This
lock could block other lock requests.
The changes to release the lock when the last reference to the open file
aren't quite right because they would allow the lock to exist as long as
there was a reference to the open file. This is too long.
The new proposed solution is to add support in the fcntl code path to
detect a race with close and then to release the lock which was just
acquired when such as race is detected. This causes locks to be released
in a timely fashion and for the system to conform to the POSIX semantic
specification.
This was tested by instrumenting a kernel to detect the handling locks and
then running a program which generates case #3 above. A dangling lock
could be reliably generated. When the changes to detect the close/fcntl
race were added, a dangling lock could no longer be generated.
Cc: Matthew Wilcox <willy@debian.org>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
We're dereferencing `flp' and then we're testing it for NULLness.
Either the compiler accidentally saved us or the existing null-pointer checdk
is redundant.
This defect was found automatically by Coverity Prevent, a static analysis tool.
Signed-off-by: Zaur Kambarov <zkambarov@coverity.com>
Cc: Matthew Wilcox <willy@debian.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This patch makes some needlessly global identifiers static.
Signed-off-by: Adrian Bunk <bunk@stusta.de>
Acked-by: Arjan van de Ven <arjanv@infradead.org>
Acked-by: Trond Myklebust <trond.myklebust@fys.uio.no>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Initial git repository build. I'm not bothering with the full history,
even though we have it. We can create a separate "historical" git
archive of that later if we want to, and in the meantime it's about
3.2GB when imported into git - space that would just make the early
git days unnecessarily complicated, when we don't have a lot of good
infrastructure for it.
Let it rip!
Eric Dumazet