The futex init function is called init(). This is a pain in the neck
when debugging when you code dies in ... init :-)
This renames it to futex_init().
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Not all architectures implement futex_atomic_cmpxchg_inatomic(). The default
implementation returns -ENOSYS, which is currently not handled inside of the
futex guts.
Futex PI calls and robust list exits with a held futex result in an endless
loop in the futex code on architectures which have no support.
Fixing up every place where futex_atomic_cmpxchg_inatomic() is called would
add a fair amount of extra if/else constructs to the already complex code. It
is also not possible to disable the robust feature before user space tries to
register robust lists.
Compile time disabling is not a good idea either, as there are already
architectures with runtime detection of futex_atomic_cmpxchg_inatomic support.
Detect the functionality at runtime instead by calling
cmpxchg_futex_value_locked() with a NULL pointer from the futex initialization
code. This is guaranteed to fail, but the call of
futex_atomic_cmpxchg_inatomic() happens with pagefaults disabled.
On architectures, which use the asm-generic implementation or have a runtime
CPU feature detection, a -ENOSYS return value disables the PI/robust features.
On architectures with a working implementation the call returns -EFAULT and
the PI/robust features are enabled.
The relevant syscalls return -ENOSYS and the robust list exit code is blocked,
when the detection fails.
Fixes http://lkml.org/lkml/2008/2/11/149
Originally reported by: Lennart Buytenhek
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Ingo Molnar <mingo@elte.hu>
Cc: Lennert Buytenhek <buytenh@wantstofly.org>
Cc: Riku Voipio <riku.voipio@movial.fi>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
When the futex init code fails to initialize the futex pseudo file system it
returns early without initializing the hash queues. Should the boot succeed
then a futex syscall which tries to enqueue a waiter on the hashqueue will
crash due to the unitilialized plist heads.
Initialize the hash queues before the filesystem.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Ingo Molnar <mingo@elte.hu>
Cc: Lennert Buytenhek <buytenh@wantstofly.org>
Cc: Riku Voipio <riku.voipio@movial.fi>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Various user space callers ask for relative timeouts. While we fixed
that overflow issue in hrtimer_start(), the sites which convert
relative user space values to absolute timeouts themself were uncovered.
Instead of putting overflow checks into each place add a function
which does the sanity checking and convert all affected callers to use
it.
Thanks to Frans Pop, who reported the problem and tested the fixes.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Ingo Molnar <mingo@elte.hu>
Tested-by: Frans Pop <elendil@planet.nl>
To allow the implementation of optimized rw-locks in user space, glibc
needs a possibility to select waiters for wakeup depending on a bitset
mask.
This requires two new futex OPs: FUTEX_WAIT_BITS and FUTEX_WAKE_BITS
These OPs are basically the same as FUTEX_WAIT and FUTEX_WAKE plus an
additional argument - a bitset. Further the FUTEX_WAIT_BITS OP is
expecting an absolute timeout value instead of the relative one, which
is used for the FUTEX_WAIT OP.
FUTEX_WAIT_BITS calls into the kernel with a bitset. The bitset is
stored in the futex_q structure, which is used to enqueue the waiter
into the hashed futex waitqueue.
FUTEX_WAKE_BITS also calls into the kernel with a bitset. The wakeup
function logically ANDs the bitset with the bitset stored in each
waiters futex_q structure. If the result is zero (i.e. none of the set
bits in the bitsets is matching), then the waiter is not woken up. If
the result is not zero (i.e. one of the set bits in the bitsets is
matching), then the waiter is woken.
The bitset provided by the caller must be non zero. In case the
provided bitset is zero the kernel returns EINVAL.
Internaly the new OPs are only extensions to the existing FUTEX_WAIT
and FUTEX_WAKE functions. The existing OPs hand a bitset with all bits
set into the futex_wait() and futex_wake() functions.
Signed-off-by: Thomas Gleixner <tgxl@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The WARN_ON() in the fixup return path of futex_lock_pi() can
trigger with false positives.
The following scenario happens:
t1 holds the futex and t2 and t3 are blocked on the kernel side rt_mutex.
t1 releases the futex (and the rt_mutex) and assigned t2 to be the next
owner of the futex.
t2 is interrupted and returns w/o acquiring the rt_mutex, before t1 can
release the rtmutex.
t1 releases the rtmutex and t3 becomes the pending owner of the rtmutex.
t2 notices that it is the designated owner (user space variable) and
fails to acquire the rt_mutex via trylock, because it is not allowed to
steal the rt_mutex from t3. Now it looks at the rt_mutex pending owner (t3)
and assigns the futex and the pi_state to it.
During the fixup t4 steals the rtmutex from t3.
t2 returns from the fixup and the owner of the rt_mutex has changed from
t3 to t4.
There is no need to do another round of fixups from t2. The important
part (t2 is not returning as the user space visible owner) is
done. The further fixups are done, before either t3 or t4 return to
user space.
For the user space it is not relevant which task (t3 or t4) is the real
owner, as long as those are both in the kernel, which is guaranteed by
the serialization of the hash bucket lock. Both tasks (which ever returns
first to userspace - t4 because it locked the rt_mutex or t3 due to a signal)
are going through the lock_futex_pi() return path where the ownership is
fixed before the return to user space.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
this patch:
commit 37bb6cb409
Author: Peter Zijlstra <a.p.zijlstra@chello.nl>
Date: Fri Jan 25 21:08:32 2008 +0100
hrtimer: unlock hrtimer_wakeup
Broke hrtimer_init_sleeper() users. It forgot to fix up the futex
caller of this function to detect the failed queueing and messed up
the do_nanosleep() caller in that it could leak a TASK_INTERRUPTIBLE
state.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Roland Westrelin did a great analysis of a long standing thinko in the
return path of futex_lock_pi.
While we fixed the lock steal case long ago, which was easy to trigger,
we never had a test case which exposed this problem and stupidly never
thought about the reverse lock stealing scenario and the return to user
space with a stale state.
When a blocked tasks returns from rt_mutex_timed_locked without holding
the rt_mutex (due to a signal or timeout) and at the same time the task
holding the futex is releasing the futex and assigning the ownership of
the futex to the returning task, then it might happen that a third task
acquires the rt_mutex before the final rt_mutex_trylock() of the
returning task happens under the futex hash bucket lock. The returning
task returns to user space with ETIMEOUT or EINTR, but the user space
futex value is assigned to this task. The task which acquired the
rt_mutex fixes the user space futex value right after the hash bucket
lock has been released by the returning task, but for a short period of
time the user space value is wrong.
Detailed description is available at:
https://bugzilla.redhat.com/show_bug.cgi?id=400541
The fix for this is the same as we do when the rt_mutex was acquired by
a higher priority task via lock stealing from the designated new owner.
In that case we already fix the user space value and the internal
pi_state up before we return. This mechanism can be used to fixup the
above corner case as well. When the returning task, which failed to
acquire the rt_mutex, notices that it is the designated owner of the
futex, then it fixes up the stale user space value and the pi_state,
before returning to user space. This happens with the futex hash bucket
lock held, so the task which acquired the rt_mutex is guaranteed to be
blocked on the hash bucket lock. We can access the rt_mutex owner, which
gives us the pid of the new owner, safely here as the owner is not able
to modify (release) it while waiting on the hash bucket lock.
Rename the "curr" argument of fixup_pi_state_owner() to "newowner" to
avoid confusion with current and add the check for the stale state into
the failure path of rt_mutex_trylock() in the return path of
unlock_futex_pi(). If the situation is detected use
fixup_pi_state_owner() to assign everything to the owner of the
rt_mutex.
Pointed-out-and-tested-by: Roland Westrelin <roland.westrelin@sun.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
David Holmes found a bug in the -rt tree with respect to
pthread_cond_timedwait. After trying his test program on the latest git
from mainline, I found the bug was there too. The bug he was seeing
that his test program showed, was that if one were to do a "Ctrl-Z" on a
process that was in the pthread_cond_timedwait, and then did a "bg" on
that process, it would return with a "-ETIMEDOUT" but early. That is,
the timer would go off early.
Looking into this, I found the source of the problem. And it is a rather
nasty bug at that.
Here's the relevant code from kernel/futex.c: (not in order in the file)
[...]
smlinkage long sys_futex(u32 __user *uaddr, int op, u32 val,
struct timespec __user *utime, u32 __user *uaddr2,
u32 val3)
{
struct timespec ts;
ktime_t t, *tp = NULL;
u32 val2 = 0;
int cmd = op & FUTEX_CMD_MASK;
if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI)) {
if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
return -EFAULT;
if (!timespec_valid(&ts))
return -EINVAL;
t = timespec_to_ktime(ts);
if (cmd == FUTEX_WAIT)
t = ktime_add(ktime_get(), t);
tp = &t;
}
[...]
return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);
}
[...]
long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
u32 __user *uaddr2, u32 val2, u32 val3)
{
int ret;
int cmd = op & FUTEX_CMD_MASK;
struct rw_semaphore *fshared = NULL;
if (!(op & FUTEX_PRIVATE_FLAG))
fshared = ¤t->mm->mmap_sem;
switch (cmd) {
case FUTEX_WAIT:
ret = futex_wait(uaddr, fshared, val, timeout);
[...]
static int futex_wait(u32 __user *uaddr, struct rw_semaphore *fshared,
u32 val, ktime_t *abs_time)
{
[...]
struct restart_block *restart;
restart = ¤t_thread_info()->restart_block;
restart->fn = futex_wait_restart;
restart->arg0 = (unsigned long)uaddr;
restart->arg1 = (unsigned long)val;
restart->arg2 = (unsigned long)abs_time;
restart->arg3 = 0;
if (fshared)
restart->arg3 |= ARG3_SHARED;
return -ERESTART_RESTARTBLOCK;
[...]
static long futex_wait_restart(struct restart_block *restart)
{
u32 __user *uaddr = (u32 __user *)restart->arg0;
u32 val = (u32)restart->arg1;
ktime_t *abs_time = (ktime_t *)restart->arg2;
struct rw_semaphore *fshared = NULL;
restart->fn = do_no_restart_syscall;
if (restart->arg3 & ARG3_SHARED)
fshared = ¤t->mm->mmap_sem;
return (long)futex_wait(uaddr, fshared, val, abs_time);
}
So when the futex_wait is interrupt by a signal we break out of the
hrtimer code and set up or return from signal. This code does not return
back to userspace, so we set up a RESTARTBLOCK. The bug here is that we
save the "abs_time" which is a pointer to the stack variable "ktime_t t"
from sys_futex.
This returns and unwinds the stack before we get to call our signal. On
return from the signal we go to futex_wait_restart, where we update all
the parameters for futex_wait and call it. But here we have a problem
where abs_time is no longer valid.
I verified this with print statements, and sure enough, what abs_time
was set to ends up being garbage when we get to futex_wait_restart.
The solution I did to solve this (with input from Linus Torvalds)
was to add unions to the restart_block to allow system calls to
use the restart with specific parameters. This way the futex code now
saves the time in a 64bit value in the restart block instead of storing
it on the stack.
Note: I'm a bit nervious to add "linux/types.h" and use u32 and u64
in thread_info.h, when there's a #ifdef __KERNEL__ just below that.
Not sure what that is there for. If this turns out to be a problem, I've
tested this with using "unsigned int" for u32 and "unsigned long long" for
u64 and it worked just the same. I'm using u32 and u64 just to be
consistent with what the futex code uses.
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
The following functions can now become static again:
- get_futex_key()
- get_futex_key_refs()
- drop_futex_key_refs()
Signed-off-by: Adrian Bunk <bunk@kernel.org>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
The find_task_by_something is a set of macros are used to find task by pid
depending on what kind of pid is proposed - global or virtual one. All of
them are wrappers above the most generic one - find_task_by_pid_type_ns() -
and just substitute some args for it.
It turned out, that dereferencing the current->nsproxy->pid_ns construction
and pushing one more argument on the stack inline cause kernel text size to
grow.
This patch moves all this stuff out-of-line into kernel/pid.c. Together
with the next patch it saves a bit less than 400 bytes from the .text
section.
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Cc: Sukadev Bhattiprolu <sukadev@us.ibm.com>
Cc: Oleg Nesterov <oleg@tv-sign.ru>
Cc: Paul Menage <menage@google.com>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This is the largest patch in the set. Make all (I hope) the places where
the pid is shown to or get from user operate on the virtual pids.
The idea is:
- all in-kernel data structures must store either struct pid itself
or the pid's global nr, obtained with pid_nr() call;
- when seeking the task from kernel code with the stored id one
should use find_task_by_pid() call that works with global pids;
- when showing pid's numerical value to the user the virtual one
should be used, but however when one shows task's pid outside this
task's namespace the global one is to be used;
- when getting the pid from userspace one need to consider this as
the virtual one and use appropriate task/pid-searching functions.
[akpm@linux-foundation.org: build fix]
[akpm@linux-foundation.org: nuther build fix]
[akpm@linux-foundation.org: yet nuther build fix]
[akpm@linux-foundation.org: remove unneeded casts]
Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Signed-off-by: Alexey Dobriyan <adobriyan@openvz.org>
Cc: Sukadev Bhattiprolu <sukadev@us.ibm.com>
Cc: Oleg Nesterov <oleg@tv-sign.ru>
Cc: Paul Menage <menage@google.com>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Get rid of sparse related warnings from places that use integer as NULL
pointer.
[akpm@linux-foundation.org: coding-style fixes]
Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org>
Cc: Andi Kleen <ak@suse.de>
Cc: Jeff Garzik <jeff@garzik.org>
Cc: Matt Mackall <mpm@selenic.com>
Cc: Ian Kent <raven@themaw.net>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Davide Libenzi <davidel@xmailserver.org>
Cc: Stephen Smalley <sds@tycho.nsa.gov>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Right now futexfs and inotifyfs have one magic 0xBAD1DEA, that looks a
little bit confusing. Use 0xBAD1DEA as magic for futexfs and 0x2BAD1DEA as
magic for inotifyfs.
Signed-off-by: Andrey Mirkin <major@openvz.org>
Acked-by: Randy Dunlap <randy.dunlap@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Calling handle_futex_death in exit_robust_list for the different robust
mutexes of a thread basically frees the mutex. Another thread might grab
the lock immediately which updates the next pointer of the mutex.
fetch_robust_entry over the next pointer might therefore branch into the
robust mutex list of a different thread. This can cause two problems: 1)
some mutexes held by the dead thread are not getting freed and 2) some
mutexs held by a different thread are freed.
The next point need to be read before calling handle_futex_death.
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
Acked-by: Ingo Molnar <mingo@elte.hu>
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The fourth argument of sys_futex is ignored when op == FUTEX_WAKE_OP,
but futex_wake_op expects it as its nr_wake2 parameter.
The only user of this operation in glibc is always passing 1, so this
bug had no consequences so far.
Signed-off-by: Andreas Schwab <schwab@suse.de>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Ulrich Drepper <drepper@redhat.com>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This patch completes Linus's wish that the fault return codes be made into
bit flags, which I agree makes everything nicer. This requires requires
all handle_mm_fault callers to be modified (possibly the modifications
should go further and do things like fault accounting in handle_mm_fault --
however that would be for another patch).
[akpm@linux-foundation.org: fix alpha build]
[akpm@linux-foundation.org: fix s390 build]
[akpm@linux-foundation.org: fix sparc build]
[akpm@linux-foundation.org: fix sparc64 build]
[akpm@linux-foundation.org: fix ia64 build]
Signed-off-by: Nick Piggin <npiggin@suse.de>
Cc: Richard Henderson <rth@twiddle.net>
Cc: Ivan Kokshaysky <ink@jurassic.park.msu.ru>
Cc: Russell King <rmk@arm.linux.org.uk>
Cc: Ian Molton <spyro@f2s.com>
Cc: Bryan Wu <bryan.wu@analog.com>
Cc: Mikael Starvik <starvik@axis.com>
Cc: David Howells <dhowells@redhat.com>
Cc: Yoshinori Sato <ysato@users.sourceforge.jp>
Cc: "Luck, Tony" <tony.luck@intel.com>
Cc: Hirokazu Takata <takata@linux-m32r.org>
Cc: Geert Uytterhoeven <geert@linux-m68k.org>
Cc: Roman Zippel <zippel@linux-m68k.org>
Cc: Greg Ungerer <gerg@uclinux.org>
Cc: Matthew Wilcox <willy@debian.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Paul Mundt <lethal@linux-sh.org>
Cc: Kazumoto Kojima <kkojima@rr.iij4u.or.jp>
Cc: Richard Curnow <rc@rc0.org.uk>
Cc: William Lee Irwin III <wli@holomorphy.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Jeff Dike <jdike@addtoit.com>
Cc: Paolo 'Blaisorblade' Giarrusso <blaisorblade@yahoo.it>
Cc: Miles Bader <uclinux-v850@lsi.nec.co.jp>
Cc: Chris Zankel <chris@zankel.net>
Acked-by: Kyle McMartin <kyle@mcmartin.ca>
Acked-by: Haavard Skinnemoen <hskinnemoen@atmel.com>
Acked-by: Ralf Baechle <ralf@linux-mips.org>
Acked-by: Andi Kleen <ak@muc.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
[ Still apparently needs some ARM and PPC loving - Linus ]
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The return value of futex_find_get_task() needs to be -ESRCH in case
that the search fails. This was part of the original futex fixes and
got accidentally dropped, when the futex-tidy-up patch was split out.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Stable Team <stable@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This reverts commit d0aa7a70bf.
It not only introduced user space visible changes to the futex syscall,
it is also non-functional and there is no way to fix it proper before
the 2.6.22 release.
The breakage report ( http://lkml.org/lkml/2007/5/12/17 ) went
unanswered, and unfortunately it turned out that the concept is not
feasible at all. It violates the rtmutex semantics badly by introducing
a virtual owner, which hacks around the coupling of the user-space
pi_futex and the kernel internal rt_mutex representation.
At the moment the only safe option is to remove it fully as it contains
user-space visible changes to broken kernel code, which we do not want
to expose in the 2.6.22 release.
The patch reverts the original patch mostly 1:1, but contains a couple
of trivial manual cleanups which were necessary due to patches, which
touched the same area of code later.
Verified against the glibc tests and my own PI futex tests.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Ingo Molnar <mingo@elte.hu>
Acked-by: Ulrich Drepper <drepper@redhat.com>
Cc: Pierre Peiffer <pierre.peiffer@bull.net>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
1. New entries can be added to tsk->pi_state_list after task completed
exit_pi_state_list(). The result is memory leakage and deadlocks.
2. handle_mm_fault() is called under spinlock. The result is obvious.
3. results in self-inflicted deadlock inside glibc.
Sometimes futex_lock_pi returns -ESRCH, when it is not expected
and glibc enters to for(;;) sleep() to simulate deadlock. This problem
is quite obvious and I think the patch is right. Though it looks like
each "if" in futex_lock_pi() got some stupid special case "else if". :-)
4. sometimes futex_lock_pi() returns -EDEADLK,
when nobody has the lock. The reason is also obvious (see comment
in the patch), but correct fix is far beyond my comprehension.
I guess someone already saw this, the chunk:
if (rt_mutex_trylock(&q.pi_state->pi_mutex))
ret = 0;
is obviously from the same opera. But it does not work, because the
rtmutex is really taken at this point: wake_futex_pi() of previous
owner reassigned it to us. My fix works. But it looks very stupid.
I would think about removal of shift of ownership in wake_futex_pi()
and making all the work in context of process taking lock.
From: Thomas Gleixner <tglx@linutronix.de>
Fix 1) Avoid the tasklist lock variant of the exit race fix by adding
an additional state transition to the exit code.
This fixes also the issue, when a task with recursive segfaults
is not able to release the futexes.
Fix 2) Cleanup the lookup_pi_state() failure path and solve the -ESRCH
problem finally.
Fix 3) Solve the fixup_pi_state_owner() problem which needs to do the fixup
in the lock protected section by using the in_atomic userspace access
functions.
This removes also the ugly lock drop / unqueue inside of fixup_pi_state()
Fix 4) Fix a stale lock in the error path of futex_wake_pi()
Added some error checks for verification.
The -EDEADLK problem is solved by the rtmutex fixups.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Ingo Molnar <mingo@elte.hu>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Ulrich Drepper <drepper@redhat.com>
Cc: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Analysis of current linux futex code :
--------------------------------------
A central hash table futex_queues[] holds all contexts (futex_q) of waiting
threads.
Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to
perform lookups or insert/deletion of a futex_q.
When a futex_wait() is done, calling thread has to :
1) - Obtain a read lock on mmap_sem to be able to validate the user pointer
(calling find_vma()). This validation tells us if the futex uses
an inode based store (mapped file), or mm based store (anonymous mem)
2) - compute a hash key
3) - Atomic increment of reference counter on an inode or a mm_struct
4) - lock part of futex_queues[] hash table
5) - perform the test on value of futex.
(rollback is value != expected_value, returns EWOULDBLOCK)
(various loops if test triggers mm faults)
6) queue the context into hash table, release the lock got in 4)
7) - release the read_lock on mmap_sem
<block>
8) Eventually unqueue the context (but rarely, as this part may be done
by the futex_wake())
Futexes were designed to improve scalability but current implementation has
various problems :
- Central hashtable :
This means scalability problems if many processes/threads want to use
futexes at the same time.
This means NUMA unbalance because this hashtable is located on one node.
- Using mmap_sem on every futex() syscall :
Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic
ops on mmap_sem, dirtying cache line :
- lot of cache line ping pongs on SMP configurations.
mmap_sem is also extensively used by mm code (page faults, mmap()/munmap())
Highly threaded processes might suffer from mmap_sem contention.
mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded
programs because of contention on the mmap_sem cache line.
- Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter:
It's also a cache line ping pong on SMP. It also increases mmap_sem hold time
because of cache misses.
Most of these scalability problems come from the fact that futexes are in
one global namespace. As we use a central hash table, we must make sure
they are all using the same reference (given by the mm subsystem). We
chose to force all futexes be 'shared'. This has a cost.
But fact is POSIX defined PRIVATE and SHARED, allowing clear separation,
and optimal performance if carefuly implemented. Time has come for linux
to have better threading performance.
The goal is to permit new futex commands to avoid :
- Taking the mmap_sem semaphore, conflicting with other subsystems.
- Modifying a ref_count on mm or an inode, still conflicting with mm or fs.
This is possible because, for one process using PTHREAD_PROCESS_PRIVATE
futexes, we only need to distinguish futexes by their virtual address, no
matter the underlying mm storage is.
If glibc wants to exploit this new infrastructure, it should use new
_PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes. And be
prepared to fallback on old subcommands for old kernels. Using one global
variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK.
PTHREAD_PROCESS_SHARED futexes should still use the old subcommands.
Compatibility with old applications is preserved, they still hit the
scalability problems, but new applications can fly :)
Note : the same SHARED futex (mapped on a file) can be used by old binaries
*and* new binaries, because both binaries will use the old subcommands.
Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic,
as this is the default semantic. Almost all applications should benefit
of this changes (new kernel and updated libc)
Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine)
/* calling futex_wait(addr, value) with value != *addr */
433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes)
424 cycles per futex(FUTEX_WAIT) call (using one futex)
334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes)
334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex)
For reference :
187 cycles per getppid() call
188 cycles per umask() call
181 cycles per ni_syscall() call
Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Pierre Peiffer <pierre.peiffer@bull.net>
Cc: "Ulrich Drepper" <drepper@gmail.com>
Cc: "Nick Piggin" <nickpiggin@yahoo.com.au>
Cc: "Ingo Molnar" <mingo@elte.hu>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
