|
We are pleased to announce "lightweight userspace priority inheritance" (PI)
support for futexes. The following patchset and glibc patch implements it,
ontop of the robust-futexes patchset which is included in 2.6.16-mm1.
We are calling it lightweight for 3 reasons:
- in the user-space fastpath a PI-enabled futex involves no kernel work
(or any other PI complexity) at all. No registration, no extra kernel
calls - just pure fast atomic ops in userspace.
- in the slowpath (in the lock-contention case), the system call and
scheduling pattern is in fact better than that of normal futexes, due to
the 'integrated' nature of FUTEX_LOCK_PI. [more about that further down]
- the in-kernel PI implementation is streamlined around the mutex
abstraction, with strict rules that keep the implementation relatively
simple: only a single owner may own a lock (i.e. no read-write lock
support), only the owner may unlock a lock, no recursive locking, etc.
Priority Inheritance - why, oh why???
-------------------------------------
Many of you heard the horror stories about the evil PI code circling Linux for
years, which makes no real sense at all and is only used by buggy applications
and which has horrible overhead. Some of you have dreaded this very moment,
when someone actually submits working PI code ;-)
So why would we like to see PI support for futexes?
We'd like to see it done purely for technological reasons. We dont think it's
a buggy concept, we think it's useful functionality to offer to applications,
which functionality cannot be achieved in other ways. We also think it's the
right thing to do, and we think we've got the right arguments and the right
numbers to prove that. We also believe that we can address all the
counter-arguments as well. For these reasons (and the reasons outlined below)
we are submitting this patch-set for upstream kernel inclusion.
What are the benefits of PI?
The short reply:
----------------
User-space PI helps achieving/improving determinism for user-space
applications. In the best-case, it can help achieve determinism and
well-bound latencies. Even in the worst-case, PI will improve the statistical
distribution of locking related application delays.
The longer reply:
-----------------
Firstly, sharing locks between multiple tasks is a common programming
technique that often cannot be replaced with lockless algorithms. As we can
see it in the kernel [which is a quite complex program in itself], lockless
structures are rather the exception than the norm - the current ratio of
lockless vs. locky code for shared data structures is somewhere between 1:10
and 1:100. Lockless is hard, and the complexity of lockless algorithms often
endangers to ability to do robust reviews of said code. I.e. critical RT
apps often choose lock structures to protect critical data structures, instead
of lockless algorithms. Furthermore, there are cases (like shared hardware,
or other resource limits) where lockless access is mathematically impossible.
Media players (such as Jack) are an example of reasonable application design
with multiple tasks (with multiple priority levels) sharing short-held locks:
for example, a highprio audio playback thread is combined with medium-prio
construct-audio-data threads and low-prio display-colory-stuff threads. Add
video and decoding to the mix and we've got even more priority levels.
So once we accept that synchronization objects (locks) are an unavoidable fact
of life, and once we accept that multi-task userspace apps have a very fair
expectation of being able to use locks, we've got to think about how to offer
the option of a deterministic locking implementation to user-space.
Most of the technical counter-arguments against doing priority inheritance
only apply to kernel-space locks. But user-space locks are different, there
we cannot disable interrupts or make the task non-preemptible in a critical
section, so the 'use spinlocks' argument does not apply (user-space spinlocks
have the same priority inversion problems as other user-space locking
constructs). Fact is, pretty much the only technique that currently enables
good determinism for userspace locks (such as futex-based pthread mutexes) is
priority inheritance:
Currently (without PI), if a high-prio and a low-prio task shares a lock [this
is a quite common scenario for most non-trivial RT applications], even if all
critical sections are coded carefully to be deterministic (i.e. all critical
sections are short in duration and only execute a limited number of
instructions), the kernel cannot guarantee any deterministic execution of the
high-prio task: any medium-priority task could preempt the low-prio task while
it holds the shared lock and executes the critical section, and could delay it
indefinitely.
Implementation:
---------------
As mentioned before, the userspace fastpath of PI-enabled pthread mutexes
involves no kernel work at all - they behave quite similarly to normal
futex-based locks: a 0 value means unlocked, and a value==TID means locked.
(This is the same method as used by list-based robust futexes.) Userspace uses
atomic ops to lock/unlock these mutexes without entering the kernel.
To handle the slowpath, we have added two new futex ops:
FUTEX_LOCK_PI
FUTEX_UNLOCK_PI
If the lock-acquire fastpath fails, [i.e. an atomic transition from 0 to TID
fails], then FUTEX_LOCK_PI is called. The kernel does all the remaining work:
if there is no futex-queue attached to the futex address yet then the code
looks up the task that owns the futex [it has put its own TID into the futex
value], and attaches a 'PI state' structure to the futex-queue. The pi_state
includes an rt-mutex, which is a PI-aware, kernel-based synchronization
object. The 'other' task is made the owner of the rt-mutex, and the
FUTEX_WAITERS bit is atomically set in the futex value. Then this task tries
to lock the rt-mutex, on which it blocks. Once it returns, it has the mutex
acquired, and it sets the futex value to its own TID and returns. Userspace
has no other work to perform - it now owns the lock, and futex value contains
FUTEX_WAITERS|TID.
If the unlock side fastpath succeeds, [i.e. userspace manages to do a TID ->
0 atomic transition of the futex value], then no kernel work is triggered.
If the unlock fastpath fails (because the FUTEX_WAITERS bit is set), then
FUTEX_UNLOCK_PI is called, and the kernel unlocks the futex on the behalf of
userspace - and it also unlocks the attached pi_state->rt_mutex and thus wakes
up any potential waiters.
Note that under this approach, contrary to other PI-futex approaches, there is
no prior 'registration' of a PI-futex. [which is not quite possible anyway,
due to existing ABI properties of pthread mutexes.]
Also, under this scheme, 'robustness' and 'PI' are two orthogonal properties
of futexes, and all four combinations are possible: futex, robust-futex,
PI-futex, robust+PI-futex.
glibc support:
--------------
Ulrich Drepper and Jakub Jelinek have written glibc support for PI-futexes
(and robust futexes), enabling robust and PI (PTHREAD_PRIO_INHERIT) POSIX
mutexes. (PTHREAD_PRIO_PROTECT support will be added later on too, no
additional kernel changes are needed for that). [NOTE: The glibc patch is
obviously inofficial and unsupported without matching upstream kernel
functionality.]
the patch-queue and the glibc patch can also be downloaded from:
http://redhat.com/~mingo/PI-futex-patches/
Many thanks go to the people who helped us create this kernel feature: Steven
Rostedt, Esben Nielsen, Benedikt Spranger, Daniel Walker, John Cooper, Arjan
van de Ven, Oleg Nesterov and others. Credits for related prior projects goes
to Dirk Grambow, Inaky Perez-Gonzalez, Bill Huey and many others.
Clean up the futex code, before adding more features to it:
- use u32 as the futex field type - that's the ABI
- use __user and pointers to u32 instead of unsigned long
- code style / comment style cleanups
- rename hash-bucket name from 'bh' to 'hb'.
I checked the pre and post futex.o object files to make sure this
patch has no code effects.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>
Cc: Ulrich Drepper <drepper@redhat.com>
Cc: Jakub Jelinek <jakub@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>
|
