|
This patchset provides a new (written from scratch) implementation of robust
futexes, called "lightweight robust futexes". We believe this new
implementation is faster and simpler than the vma-based robust futex solutions
presented before, and we'd like this patchset to be adopted in the upstream
kernel. This is version 1 of the patchset.
Background
----------
What are robust futexes? To answer that, we first need to understand what
futexes are: normal futexes are special types of locks that in the
noncontended case can be acquired/released from userspace without having to
enter the kernel.
A futex is in essence a user-space address, e.g. a 32-bit lock variable
field. If userspace notices contention (the lock is already owned and someone
else wants to grab it too) then the lock is marked with a value that says
"there's a waiter pending", and the sys_futex(FUTEX_WAIT) syscall is used to
wait for the other guy to release it. The kernel creates a 'futex queue'
internally, so that it can later on match up the waiter with the waker -
without them having to know about each other. When the owner thread releases
the futex, it notices (via the variable value) that there were waiter(s)
pending, and does the sys_futex(FUTEX_WAKE) syscall to wake them up. Once all
waiters have taken and released the lock, the futex is again back to
'uncontended' state, and there's no in-kernel state associated with it. The
kernel completely forgets that there ever was a futex at that address. This
method makes futexes very lightweight and scalable.
"Robustness" is about dealing with crashes while holding a lock: if a process
exits prematurely while holding a pthread_mutex_t lock that is also shared
with some other process (e.g. yum segfaults while holding a pthread_mutex_t,
or yum is kill -9-ed), then waiters for that lock need to be notified that the
last owner of the lock exited in some irregular way.
To solve such types of problems, "robust mutex" userspace APIs were created:
pthread_mutex_lock() returns an error value if the owner exits prematurely -
and the new owner can decide whether the data protected by the lock can be
recovered safely.
There is a big conceptual problem with futex based mutexes though: it is the
kernel that destroys the owner task (e.g. due to a SEGFAULT), but the kernel
cannot help with the cleanup: if there is no 'futex queue' (and in most cases
there is none, futexes being fast lightweight locks) then the kernel has no
information to clean up after the held lock! Userspace has no chance to clean
up after the lock either - userspace is the one that crashes, so it has no
opportunity to clean up. Catch-22.
In practice, when e.g. yum is kill -9-ed (or segfaults), a system reboot is
needed to release that futex based lock. This is one of the leading
bugreports against yum.
To solve this problem, 'Robust Futex' patches were created and presented on
lkml: the one written by Todd Kneisel and David Singleton is the most advanced
at the moment. These patches all tried to extend the futex abstraction by
registering futex-based locks in the kernel - and thus give the kernel a
chance to clean up.
E.g. in David Singleton's robust-futex-6.patch, there are 3 new syscall
variants to sys_futex(): FUTEX_REGISTER, FUTEX_DEREGISTER and FUTEX_RECOVER.
The kernel attaches such robust futexes to vmas (via
vma->vm_file->f_mapping->robust_head), and at do_exit() time, all vmas are
searched to see whether they have a robust_head set.
Lots of work went into the vma-based robust-futex patch, and recently it has
improved significantly, but unfortunately it still has two fundamental
problems left:
- they have quite complex locking and race scenarios. The vma-based
patches had been pending for years, but they are still not completely
reliable.
- they have to scan _every_ vma at sys_exit() time, per thread!
The second disadvantage is a real killer: pthread_exit() takes around 1
microsecond on Linux, but with thousands (or tens of thousands) of vmas every
pthread_exit() takes a millisecond or more, also totally destroying the CPU's
L1 and L2 caches!
This is very much noticeable even for normal process sys_exit_group() calls:
the kernel has to do the vma scanning unconditionally! (this is because the
kernel has no knowledge about how many robust futexes there are to be cleaned
up, because a robust futex might have been registered in another task, and the
futex variable might have been simply mmap()-ed into this process's address
space).
This huge overhead forced the creation of CONFIG_FUTEX_ROBUST, but worse than
that: the overhead makes robust futexes impractical for any type of generic
Linux distribution.
So it became clear to us, something had to be done. Last week, when Thomas
Gleixner tried to fix up the vma-based robust futex patch in the -rt tree, he
found a handful of new races and we were talking about it and were analyzing
the situation. At that point a fundamentally different solution occured to
me. This patchset (written in the past couple of days) implements that new
solution. Be warned though - the patchset does things we normally dont do in
Linux, so some might find the approach disturbing. Parental advice
recommended ;-)
New approach to robust futexes
------------------------------
At the heart of this new approach there is a per-thread private list of robust
locks that userspace is holding (maintained by glibc) - which userspace list
is registered with the kernel via a new syscall [this registration happens at
most once per thread lifetime]. At do_exit() time, the kernel checks this
user-space list: are there any robust futex locks to be cleaned up?
In the common case, at do_exit() time, there is no list registered, so the
cost of robust futexes is just a simple current->robust_list != NULL
comparison. If the thread has registered a list, then normally the list is
empty. If the thread/process crashed or terminated in some incorrect way then
the list might be non-empty: in this case the kernel carefully walks the list
[not trusting it], and marks all locks that are owned by this thread with the
FUTEX_OWNER_DEAD bit, and wakes up one waiter (if any).
The list is guaranteed to be private and per-thread, so it's lockless. There
is one race possible though: since adding to and removing from the list is
done after the futex is acquired by glibc, there is a few instructions window
for the thread (or process) to die there, leaving the futex hung. To protect
against this possibility, userspace (glibc) also maintains a simple per-thread
'list_op_pending' field, to allow the kernel to clean up if the thread dies
after acquiring the lock, but just before it could have added itself to the
list. Glibc sets this list_op_pending field before it tries to acquire the
futex, and clears it after the list-add (or list-remove) has finished.
That's all that is needed - all the rest of robust-futex cleanup is done in
userspace [just like with the previous patches].
Ulrich Drepper has implemented the necessary glibc support for this new
mechanism, which fully enables robust mutexes. (Ulrich plans to commit these
changes to glibc-HEAD later today.)
Key differences of this userspace-list based approach, compared to the vma
based method:
- it's much, much faster: at thread exit time, there's no need to loop
over every vma (!), which the VM-based method has to do. Only a very
simple 'is the list empty' op is done.
- no VM changes are needed - 'struct address_space' is left alone.
- no registration of individual locks is needed: robust mutexes dont need
any extra per-lock syscalls. Robust mutexes thus become a very lightweight
primitive - so they dont force the application designer to do a hard choice
between performance and robustness - robust mutexes are just as fast.
- no per-lock kernel allocation happens.
- no resource limits are needed.
- no kernel-space recovery call (FUTEX_RECOVER) is needed.
- the implementation and the locking is "obvious", and there are no
interactions with the VM.
Performance
-----------
I have benchmarked the time needed for the kernel to process a list of 1
million (!) held locks, using the new method [on a 2GHz CPU]:
- with FUTEX_WAIT set [contended mutex]: 130 msecs
- without FUTEX_WAIT set [uncontended mutex]: 30 msecs
I have also measured an approach where glibc does the lock notification [which
it currently does for !pshared robust mutexes], and that took 256 msecs -
clearly slower, due to the 1 million FUTEX_WAKE syscalls userspace had to do.
(1 million held locks are unheard of - we expect at most a handful of locks to
be held at a time. Nevertheless it's nice to know that this approach scales
nicely.)
Implementation details
----------------------
The patch adds two new syscalls: one to register the userspace list, and one
to query the registered list pointer:
asmlinkage long
sys_set_robust_list(struct robust_list_head __user *head,
size_t len);
asmlinkage long
sys_get_robust_list(int pid, struct robust_list_head __user **head_ptr,
size_t __user *len_ptr);
List registration is very fast: the pointer is simply stored in
current->robust_list. [Note that in the future, if robust futexes become
widespread, we could extend sys_clone() to register a robust-list head for new
threads, without the need of another syscall.]
So there is virtually zero overhead for tasks not using robust futexes, and
even for robust futex users, there is only one extra syscall per thread
lifetime, and the cleanup operation, if it happens, is fast and
straightforward. The kernel doesnt have any internal distinction between
robust and normal futexes.
If a futex is found to be held at exit time, the kernel sets the highest bit
of the futex word:
#define FUTEX_OWNER_DIED 0x40000000
and wakes up the next futex waiter (if any). User-space does the rest of
the cleanup.
Otherwise, robust futexes are acquired by glibc by putting the TID into the
futex field atomically. Waiters set the FUTEX_WAITERS bit:
#define FUTEX_WAITERS 0x80000000
and the remaining bits are for the TID.
Testing, architecture support
-----------------------------
I've tested the new syscalls on x86 and x86_64, and have made sure the parsing
of the userspace list is robust [ ;-) ] even if the list is deliberately
corrupted.
i386 and x86_64 syscalls are wired up at the moment, and Ulrich has tested the
new glibc code (on x86_64 and i386), and it works for his robust-mutex
testcases.
All other architectures should build just fine too - but they wont have the
new syscalls yet.
Architectures need to implement the new futex_atomic_cmpxchg_inuser() inline
function before writing up the syscalls (that function returns -ENOSYS right
now).
This patch:
Add placeholder futex_atomic_cmpxchg_inuser() implementations to every
architecture that supports futexes. It returns -ENOSYS.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Arjan van de Ven <arjan@infradead.org>
Acked-by: Ulrich Drepper <drepper@redhat.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
|
|
ATM pthread_cond_signal is unnecessarily slow, because it wakes one waiter
(which at least on UP usually means an immediate context switch to one of
the waiter threads). This waiter wakes up and after a few instructions it
attempts to acquire the cv internal lock, but that lock is still held by
the thread calling pthread_cond_signal. So it goes to sleep and eventually
the signalling thread is scheduled in, unlocks the internal lock and wakes
the waiter again.
Now, before 2003-09-21 NPTL was using FUTEX_REQUEUE in pthread_cond_signal
to avoid this performance issue, but it was removed when locks were
redesigned to the 3 state scheme (unlocked, locked uncontended, locked
contended).
Following scenario shows why simply using FUTEX_REQUEUE in
pthread_cond_signal together with using lll_mutex_unlock_force in place of
lll_mutex_unlock is not enough and probably why it has been disabled at
that time:
The number is value in cv->__data.__lock.
thr1 thr2 thr3
0 pthread_cond_wait
1 lll_mutex_lock (cv->__data.__lock)
0 lll_mutex_unlock (cv->__data.__lock)
0 lll_futex_wait (&cv->__data.__futex, futexval)
0 pthread_cond_signal
1 lll_mutex_lock (cv->__data.__lock)
1 pthread_cond_signal
2 lll_mutex_lock (cv->__data.__lock)
2 lll_futex_wait (&cv->__data.__lock, 2)
2 lll_futex_requeue (&cv->__data.__futex, 0, 1, &cv->__data.__lock)
# FUTEX_REQUEUE, not FUTEX_CMP_REQUEUE
2 lll_mutex_unlock_force (cv->__data.__lock)
0 cv->__data.__lock = 0
0 lll_futex_wake (&cv->__data.__lock, 1)
1 lll_mutex_lock (cv->__data.__lock)
0 lll_mutex_unlock (cv->__data.__lock)
# Here, lll_mutex_unlock doesn't know there are threads waiting
# on the internal cv's lock
Now, I believe it is possible to use FUTEX_REQUEUE in pthread_cond_signal,
but it will cost us not one, but 2 extra syscalls and, what's worse, one of
these extra syscalls will be done for every single waiting loop in
pthread_cond_*wait.
We would need to use lll_mutex_unlock_force in pthread_cond_signal after
requeue and lll_mutex_cond_lock in pthread_cond_*wait after lll_futex_wait.
Another alternative is to do the unlocking pthread_cond_signal needs to do
(the lock can't be unlocked before lll_futex_wake, as that is racy) in the
kernel.
I have implemented both variants, futex-requeue-glibc.patch is the first
one and futex-wake_op{,-glibc}.patch is the unlocking inside of the kernel.
The kernel interface allows userland to specify how exactly an unlocking
operation should look like (some atomic arithmetic operation with optional
constant argument and comparison of the previous futex value with another
constant).
It has been implemented just for ppc*, x86_64 and i?86, for other
architectures I'm including just a stub header which can be used as a
starting point by maintainers to write support for their arches and ATM
will just return -ENOSYS for FUTEX_WAKE_OP. The requeue patch has been
(lightly) tested just on x86_64, the wake_op patch on ppc64 kernel running
32-bit and 64-bit NPTL and x86_64 kernel running 32-bit and 64-bit NPTL.
With the following benchmark on UP x86-64 I get:
for i in nptl-orig nptl-requeue nptl-wake_op; do echo time elf/ld.so --library-path .:$i /tmp/bench; \
for j in 1 2; do echo ( time elf/ld.so --library-path .:$i /tmp/bench ) 2>&1; done; done
time elf/ld.so --library-path .:nptl-orig /tmp/bench
real 0m0.655s user 0m0.253s sys 0m0.403s
real 0m0.657s user 0m0.269s sys 0m0.388s
time elf/ld.so --library-path .:nptl-requeue /tmp/bench
real 0m0.496s user 0m0.225s sys 0m0.271s
real 0m0.531s user 0m0.242s sys 0m0.288s
time elf/ld.so --library-path .:nptl-wake_op /tmp/bench
real 0m0.380s user 0m0.176s sys 0m0.204s
real 0m0.382s user 0m0.175s sys 0m0.207s
The benchmark is at:
http://sourceware.org/ml/libc-alpha/2005-03/txt00001.txt
Older futex-requeue-glibc.patch version is at:
http://sourceware.org/ml/libc-alpha/2005-03/txt00002.txt
Older futex-wake_op-glibc.patch version is at:
http://sourceware.org/ml/libc-alpha/2005-03/txt00003.txt
Will post a new version (just x86-64 fixes so that the patch
applies against pthread_cond_signal.S) to libc-hacker ml soon.
Attached is the kernel FUTEX_WAKE_OP patch as well as a simple-minded
testcase that will not test the atomicity of the operation, but at least
check if the threads that should have been woken up are woken up and
whether the arithmetic operation in the kernel gave the expected results.
Acked-by: Ingo Molnar <mingo@redhat.com>
Cc: Ulrich Drepper <drepper@redhat.com>
Cc: Jamie Lokier <jamie@shareable.org>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Yoichi Yuasa <yuasa@hh.iij4u.or.jp>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
|
