FEX-Emu/linux
1
0
Fork 0
mirror of https://github.com/FEX-Emu/linux.git synced 2024-12-23 01:40:30 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity
dfd4e3ec24
linux/include/asm-i386/futex.h

136 lines
3.1 KiB
C
Raw Normal View History

[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup 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>
2005-09-06 22:16:25 +00:00
#ifndef _ASM_FUTEX_H
#define _ASM_FUTEX_H
#ifdef __KERNEL__
#include <linux/futex.h>
#include <asm/errno.h>
#include <asm/system.h>
#include <asm/processor.h>
#include <asm/uaccess.h>
#define __futex_atomic_op1(insn, ret, oldval, uaddr, oparg) \
__asm__ __volatile ( \
"1: " insn "\n" \
"2: .section .fixup,\"ax\"\n\
3: mov %3, %1\n\
jmp 2b\n\
.previous\n\
.section __ex_table,\"a\"\n\
.align 8\n\
.long 1b,3b\n\
.previous" \
: "=r" (oldval), "=r" (ret), "=m" (*uaddr) \
: "i" (-EFAULT), "m" (*uaddr), "0" (oparg), "1" (0))
#define __futex_atomic_op2(insn, ret, oldval, uaddr, oparg) \
__asm__ __volatile ( \
"1: movl %2, %0\n\
movl %0, %3\n" \
insn "\n" \
"2: " LOCK_PREFIX "cmpxchgl %3, %2\n\
jnz 1b\n\
3: .section .fixup,\"ax\"\n\
4: mov %5, %1\n\
jmp 3b\n\
.previous\n\
.section __ex_table,\"a\"\n\
.align 8\n\
.long 1b,4b,2b,4b\n\
.previous" \
: "=&a" (oldval), "=&r" (ret), "=m" (*uaddr), \
"=&r" (tem) \
: "r" (oparg), "i" (-EFAULT), "m" (*uaddr), "1" (0))
static inline int
futex_atomic_op_inuser (int encoded_op, int __user *uaddr)
{
int op = (encoded_op >> 28) & 7;
int cmp = (encoded_op >> 24) & 15;
int oparg = (encoded_op << 8) >> 20;
int cmparg = (encoded_op << 20) >> 20;
int oldval = 0, ret, tem;
if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28))
oparg = 1 << oparg;
if (! access_ok (VERIFY_WRITE, uaddr, sizeof(int)))
return -EFAULT;
inc_preempt_count();
if (op == FUTEX_OP_SET)
__futex_atomic_op1("xchgl %0, %2", ret, oldval, uaddr, oparg);
else {
[PATCH] uml: remove leftover from patch revertal I added this line to share this file with UML, but now it's no longer shared so remove this useless leftover. Signed-off-by: Paolo 'Blaisorblade' Giarrusso <blaisorblade@yahoo.it> Acked-by: Jeff Dike <jdike@addtoit.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-19 01:42:52 +00:00
#ifndef CONFIG_X86_BSWAP
[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup 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>
2005-09-06 22:16:25 +00:00
if (boot_cpu_data.x86 == 3)
ret = -ENOSYS;
else
#endif
switch (op) {
case FUTEX_OP_ADD:
__futex_atomic_op1(LOCK_PREFIX "xaddl %0, %2", ret,
oldval, uaddr, oparg);
break;
case FUTEX_OP_OR:
__futex_atomic_op2("orl %4, %3", ret, oldval, uaddr,
oparg);
break;
case FUTEX_OP_ANDN:
__futex_atomic_op2("andl %4, %3", ret, oldval, uaddr,
~oparg);
break;
case FUTEX_OP_XOR:
__futex_atomic_op2("xorl %4, %3", ret, oldval, uaddr,
oparg);
break;
default:
ret = -ENOSYS;
}
}
dec_preempt_count();
if (!ret) {
switch (cmp) {
case FUTEX_OP_CMP_EQ: ret = (oldval == cmparg); break;
case FUTEX_OP_CMP_NE: ret = (oldval != cmparg); break;
case FUTEX_OP_CMP_LT: ret = (oldval < cmparg); break;
case FUTEX_OP_CMP_GE: ret = (oldval >= cmparg); break;
case FUTEX_OP_CMP_LE: ret = (oldval <= cmparg); break;
case FUTEX_OP_CMP_GT: ret = (oldval > cmparg); break;
default: ret = -ENOSYS;
}
}
return ret;
}
[PATCH] lightweight robust futexes: arch defaults 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>
2006-03-27 09:16:21 +00:00
static inline int
futex_atomic_cmpxchg_inuser(int __user *uaddr, int oldval, int newval)
{
[PATCH] lightweight robust futexes: i386 i386: add the futex_atomic_cmpxchg_inuser() assembly implementation, and wire up the new syscalls. 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>
2006-03-27 09:16:25 +00:00
if (!access_ok(VERIFY_WRITE, uaddr, sizeof(int)))
return -EFAULT;
__asm__ __volatile__(
"1: " LOCK_PREFIX "cmpxchgl %3, %1 \n"
"2: .section .fixup, \"ax\" \n"
"3: mov %2, %0 \n"
" jmp 2b \n"
" .previous \n"
" .section __ex_table, \"a\" \n"
" .align 8 \n"
" .long 1b,3b \n"
" .previous \n"
: "=a" (oldval), "=m" (*uaddr)
: "i" (-EFAULT), "r" (newval), "0" (oldval)
: "memory"
);
return oldval;
[PATCH] lightweight robust futexes: arch defaults 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>
2006-03-27 09:16:21 +00:00
}
[PATCH] FUTEX_WAKE_OP: pthread_cond_signal() speedup 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>
2005-09-06 22:16:25 +00:00
#endif
#endif
Reference in New Issue Copy Permalink