[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>
2005-09-07 02:16:25 +04:00
|
|
|
#ifndef _ASM_FUTEX_H
|
|
|
|
|
#define _ASM_FUTEX_H
|
|
|
|
|
|
2006-09-27 01:00:56 +04:00
|
|
|
#include <linux/futex.h>
|
2008-04-30 11:54:49 +04:00
|
|
|
#include <linux/uaccess.h>
|
2006-09-27 01:00:56 +04:00
|
|
|
#include <asm/errno.h>
|
[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>
2005-09-07 02:16:25 +04:00
|
|
|
|
2006-09-27 01:00:56 +04:00
|
|
|
#define __futex_atomic_op1(insn, ret, oldval, uaddr, oparg) \
|
|
|
|
|
do { \
|
|
|
|
|
register unsigned long r8 __asm ("r8") = 0; \
|
|
|
|
|
__asm__ __volatile__( \
|
|
|
|
|
" mf;; \n" \
|
|
|
|
|
"[1:] " insn ";; \n" \
|
|
|
|
|
" .xdata4 \"__ex_table\", 1b-., 2f-. \n" \
|
|
|
|
|
"[2:]" \
|
|
|
|
|
: "+r" (r8), "=r" (oldval) \
|
|
|
|
|
: "r" (uaddr), "r" (oparg) \
|
|
|
|
|
: "memory"); \
|
|
|
|
|
ret = r8; \
|
|
|
|
|
} while (0)
|
|
|
|
|
|
|
|
|
|
#define __futex_atomic_op2(insn, ret, oldval, uaddr, oparg) \
|
|
|
|
|
do { \
|
|
|
|
|
register unsigned long r8 __asm ("r8") = 0; \
|
|
|
|
|
int val, newval; \
|
|
|
|
|
do { \
|
|
|
|
|
__asm__ __volatile__( \
|
|
|
|
|
" mf;; \n" \
|
|
|
|
|
"[1:] ld4 %3=[%4];; \n" \
|
|
|
|
|
" mov %2=%3 \n" \
|
|
|
|
|
insn ";; \n" \
|
|
|
|
|
" mov ar.ccv=%2;; \n" \
|
|
|
|
|
"[2:] cmpxchg4.acq %1=[%4],%3,ar.ccv;; \n" \
|
|
|
|
|
" .xdata4 \"__ex_table\", 1b-., 3f-.\n" \
|
|
|
|
|
" .xdata4 \"__ex_table\", 2b-., 3f-.\n" \
|
|
|
|
|
"[3:]" \
|
|
|
|
|
: "+r" (r8), "=r" (val), "=&r" (oldval), \
|
|
|
|
|
"=&r" (newval) \
|
|
|
|
|
: "r" (uaddr), "r" (oparg) \
|
|
|
|
|
: "memory"); \
|
|
|
|
|
if (unlikely (r8)) \
|
|
|
|
|
break; \
|
|
|
|
|
} while (unlikely (val != oldval)); \
|
|
|
|
|
ret = r8; \
|
|
|
|
|
} while (0)
|
|
|
|
|
|
|
|
|
|
static inline int
|
2011-03-11 05:50:58 +03:00
|
|
|
futex_atomic_op_inuser (int encoded_op, u32 __user *uaddr)
|
2006-09-27 01:00:56 +04:00
|
|
|
{
|
|
|
|
|
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;
|
|
|
|
|
if (encoded_op & (FUTEX_OP_OPARG_SHIFT << 28))
|
|
|
|
|
oparg = 1 << oparg;
|
|
|
|
|
|
2011-03-11 05:50:58 +03:00
|
|
|
if (! access_ok (VERIFY_WRITE, uaddr, sizeof(u32)))
|
2006-09-27 01:00:56 +04:00
|
|
|
return -EFAULT;
|
|
|
|
|
|
2006-12-07 07:32:20 +03:00
|
|
|
pagefault_disable();
|
2006-09-27 01:00:56 +04:00
|
|
|
|
|
|
|
|
switch (op) {
|
|
|
|
|
case FUTEX_OP_SET:
|
|
|
|
|
__futex_atomic_op1("xchg4 %1=[%2],%3", ret, oldval, uaddr,
|
|
|
|
|
oparg);
|
|
|
|
|
break;
|
|
|
|
|
case FUTEX_OP_ADD:
|
|
|
|
|
__futex_atomic_op2("add %3=%3,%5", ret, oldval, uaddr, oparg);
|
|
|
|
|
break;
|
|
|
|
|
case FUTEX_OP_OR:
|
|
|
|
|
__futex_atomic_op2("or %3=%3,%5", ret, oldval, uaddr, oparg);
|
|
|
|
|
break;
|
|
|
|
|
case FUTEX_OP_ANDN:
|
|
|
|
|
__futex_atomic_op2("and %3=%3,%5", ret, oldval, uaddr,
|
|
|
|
|
~oparg);
|
|
|
|
|
break;
|
|
|
|
|
case FUTEX_OP_XOR:
|
|
|
|
|
__futex_atomic_op2("xor %3=%3,%5", ret, oldval, uaddr, oparg);
|
|
|
|
|
break;
|
|
|
|
|
default:
|
|
|
|
|
ret = -ENOSYS;
|
|
|
|
|
}
|
|
|
|
|
|
2006-12-07 07:32:20 +03:00
|
|
|
pagefault_enable();
|
2006-09-27 01:00:56 +04:00
|
|
|
|
|
|
|
|
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;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static inline int
|
2011-03-11 05:50:58 +03:00
|
|
|
futex_atomic_cmpxchg_inatomic(u32 *uval, u32 __user *uaddr,
|
|
|
|
|
u32 oldval, u32 newval)
|
2006-09-27 01:00:56 +04:00
|
|
|
{
|
2011-03-11 05:50:58 +03:00
|
|
|
if (!access_ok(VERIFY_WRITE, uaddr, sizeof(u32)))
|
2006-09-27 01:00:56 +04:00
|
|
|
return -EFAULT;
|
|
|
|
|
|
|
|
|
|
{
|
Wrong asm register contraints in the futex implementation
The Linux Kernel contains some inline assembly source code which has
wrong asm register constraints in arch/ia64/include/asm/futex.h.
I observed this on Kernel 3.2.23 but it is also true on the most
recent Kernel 3.9-rc1.
File arch/ia64/include/asm/futex.h:
static inline int
futex_atomic_cmpxchg_inatomic(u32 *uval, u32 __user *uaddr,
u32 oldval, u32 newval)
{
if (!access_ok(VERIFY_WRITE, uaddr, sizeof(u32)))
return -EFAULT;
{
register unsigned long r8 __asm ("r8");
unsigned long prev;
__asm__ __volatile__(
" mf;; \n"
" mov %0=r0 \n"
" mov ar.ccv=%4;; \n"
"[1:] cmpxchg4.acq %1=[%2],%3,ar.ccv \n"
" .xdata4 \"__ex_table\", 1b-., 2f-. \n"
"[2:]"
: "=r" (r8), "=r" (prev)
: "r" (uaddr), "r" (newval),
"rO" ((long) (unsigned) oldval)
: "memory");
*uval = prev;
return r8;
}
}
The list of output registers is
: "=r" (r8), "=r" (prev)
The constraint "=r" means that the GCC has to maintain that these vars
are in registers and contain valid info when the program flow leaves
the assembly block (output registers).
But "=r" also means that GCC can put them in registers that are used
as input registers. Input registers are uaddr, newval, oldval on the
example.
The second assembly instruction
" mov %0=r0 \n"
is the first one which writes to a register; it sets %0 to 0. %0 means
the first register operand; it is r8 here. (The r0 is read-only and
always 0 on the Itanium; it can be used if an immediate zero value is
needed.)
This instruction might overwrite one of the other registers which are
still needed.
Whether it really happens depends on how GCC decides what registers it
uses and how it optimizes the code.
The objdump utility can give us disassembly.
The futex_atomic_cmpxchg_inatomic() function is inline, so we have to
look for a module that uses the funtion. This is the
cmpxchg_futex_value_locked() function in
kernel/futex.c:
static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
u32 uval, u32 newval)
{
int ret;
pagefault_disable();
ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
pagefault_enable();
return ret;
}
Now the disassembly. At first from the Kernel package 3.2.23 which has
been compiled with GCC 4.4, remeber this Kernel seemed to work:
objdump -d linux-3.2.23/debian/build/build_ia64_none_mckinley/kernel/futex.o
0000000000000230 <cmpxchg_futex_value_locked>:
230: 0b 18 80 1b 18 21 [MMI] adds r3=3168,r13;;
236: 80 40 0d 00 42 00 adds r8=40,r3
23c: 00 00 04 00 nop.i 0x0;;
240: 0b 50 00 10 10 10 [MMI] ld4 r10=[r8];;
246: 90 08 28 00 42 00 adds r9=1,r10
24c: 00 00 04 00 nop.i 0x0;;
250: 09 00 00 00 01 00 [MMI] nop.m 0x0
256: 00 48 20 20 23 00 st4 [r8]=r9
25c: 00 00 04 00 nop.i 0x0;;
260: 08 10 80 06 00 21 [MMI] adds r2=32,r3
266: 00 00 00 02 00 00 nop.m 0x0
26c: 02 08 f1 52 extr.u r16=r33,0,61
270: 05 40 88 00 08 e0 [MLX] addp4 r8=r34,r0
276: ff ff 0f 00 00 e0 movl r15=0xfffffffbfff;;
27c: f1 f7 ff 65
280: 09 70 00 04 18 10 [MMI] ld8 r14=[r2]
286: 00 00 00 02 00 c0 nop.m 0x0
28c: f0 80 1c d0 cmp.ltu p6,p7=r15,r16;;
290: 08 40 fc 1d 09 3b [MMI] cmp.eq p8,p9=-1,r14
296: 00 00 00 02 00 40 nop.m 0x0
29c: e1 08 2d d0 cmp.ltu p10,p11=r14,r33
2a0: 56 01 10 00 40 10 [BBB] (p10) br.cond.spnt.few 2e0
<cmpxchg_futex_value_locked+0xb0>
2a6: 02 08 00 80 21 03 (p08) br.cond.dpnt.few 2b0
<cmpxchg_futex_value_locked+0x80>
2ac: 40 00 00 41 (p06) br.cond.spnt.few 2e0
<cmpxchg_futex_value_locked+0xb0>
2b0: 0a 00 00 00 22 00 [MMI] mf;;
2b6: 80 00 00 00 42 00 mov r8=r0
2bc: 00 00 04 00 nop.i 0x0
2c0: 0b 00 20 40 2a 04 [MMI] mov.m ar.ccv=r8;;
2c6: 10 1a 85 22 20 00 cmpxchg4.acq r33=[r33],r35,ar.ccv
2cc: 00 00 04 00 nop.i 0x0;;
2d0: 10 00 84 40 90 11 [MIB] st4 [r32]=r33
2d6: 00 00 00 02 00 00 nop.i 0x0
2dc: 20 00 00 40 br.few 2f0
<cmpxchg_futex_value_locked+0xc0>
2e0: 09 40 c8 f9 ff 27 [MMI] mov r8=-14
2e6: 00 00 00 02 00 00 nop.m 0x0
2ec: 00 00 04 00 nop.i 0x0;;
2f0: 0b 58 20 1a 19 21 [MMI] adds r11=3208,r13;;
2f6: 20 01 2c 20 20 00 ld4 r18=[r11]
2fc: 00 00 04 00 nop.i 0x0;;
300: 0b 88 fc 25 3f 23 [MMI] adds r17=-1,r18;;
306: 00 88 2c 20 23 00 st4 [r11]=r17
30c: 00 00 04 00 nop.i 0x0;;
310: 11 00 00 00 01 00 [MIB] nop.m 0x0
316: 00 00 00 02 00 80 nop.i 0x0
31c: 08 00 84 00 br.ret.sptk.many b0;;
The lines
2b0: 0a 00 00 00 22 00 [MMI] mf;;
2b6: 80 00 00 00 42 00 mov r8=r0
2bc: 00 00 04 00 nop.i 0x0
2c0: 0b 00 20 40 2a 04 [MMI] mov.m ar.ccv=r8;;
2c6: 10 1a 85 22 20 00 cmpxchg4.acq r33=[r33],r35,ar.ccv
2cc: 00 00 04 00 nop.i 0x0;;
are the instructions of the assembly block.
The line
2b6: 80 00 00 00 42 00 mov r8=r0
sets the r8 register to 0 and after that
2c0: 0b 00 20 40 2a 04 [MMI] mov.m ar.ccv=r8;;
prepares the 'oldvalue' for the cmpxchg but it takes it from r8. This
is wrong.
What happened here is what I explained above: An input register is
overwritten which is still needed.
The register operand constraints in futex.h are wrong.
(The problem doesn't occur when the Kernel is compiled with GCC 4.6.)
The attached patch fixes the register operand constraints in futex.h.
The code after patching of it:
static inline int
futex_atomic_cmpxchg_inatomic(u32 *uval, u32 __user *uaddr,
u32 oldval, u32 newval)
{
if (!access_ok(VERIFY_WRITE, uaddr, sizeof(u32)))
return -EFAULT;
{
register unsigned long r8 __asm ("r8") = 0;
unsigned long prev;
__asm__ __volatile__(
" mf;; \n"
" mov ar.ccv=%4;; \n"
"[1:] cmpxchg4.acq %1=[%2],%3,ar.ccv \n"
" .xdata4 \"__ex_table\", 1b-., 2f-. \n"
"[2:]"
: "+r" (r8), "=&r" (prev)
: "r" (uaddr), "r" (newval),
"rO" ((long) (unsigned) oldval)
: "memory");
*uval = prev;
return r8;
}
}
I also initialized the 'r8' var with the C programming language.
The _asm qualifier on the definition of the 'r8' var forces GCC to use
the r8 processor register for it.
I don't believe that we should use inline assembly for zeroing out a
local variable.
The constraint is
"+r" (r8)
what means that it is both an input register and an output register.
Note that the page fault handler will modify the r8 register which
will be the return value of the function.
The real fix is
"=&r" (prev)
The & means that GCC must not use any of the input registers to place
this output register in.
Patched the Kernel 3.2.23 and compiled it with GCC4.4:
0000000000000230 <cmpxchg_futex_value_locked>:
230: 0b 18 80 1b 18 21 [MMI] adds r3=3168,r13;;
236: 80 40 0d 00 42 00 adds r8=40,r3
23c: 00 00 04 00 nop.i 0x0;;
240: 0b 50 00 10 10 10 [MMI] ld4 r10=[r8];;
246: 90 08 28 00 42 00 adds r9=1,r10
24c: 00 00 04 00 nop.i 0x0;;
250: 09 00 00 00 01 00 [MMI] nop.m 0x0
256: 00 48 20 20 23 00 st4 [r8]=r9
25c: 00 00 04 00 nop.i 0x0;;
260: 08 10 80 06 00 21 [MMI] adds r2=32,r3
266: 20 12 01 10 40 00 addp4 r34=r34,r0
26c: 02 08 f1 52 extr.u r16=r33,0,61
270: 05 40 00 00 00 e1 [MLX] mov r8=r0
276: ff ff 0f 00 00 e0 movl r15=0xfffffffbfff;;
27c: f1 f7 ff 65
280: 09 70 00 04 18 10 [MMI] ld8 r14=[r2]
286: 00 00 00 02 00 c0 nop.m 0x0
28c: f0 80 1c d0 cmp.ltu p6,p7=r15,r16;;
290: 08 40 fc 1d 09 3b [MMI] cmp.eq p8,p9=-1,r14
296: 00 00 00 02 00 40 nop.m 0x0
29c: e1 08 2d d0 cmp.ltu p10,p11=r14,r33
2a0: 56 01 10 00 40 10 [BBB] (p10) br.cond.spnt.few 2e0
<cmpxchg_futex_value_locked+0xb0>
2a6: 02 08 00 80 21 03 (p08) br.cond.dpnt.few 2b0
<cmpxchg_futex_value_locked+0x80>
2ac: 40 00 00 41 (p06) br.cond.spnt.few 2e0
<cmpxchg_futex_value_locked+0xb0>
2b0: 0b 00 00 00 22 00 [MMI] mf;;
2b6: 00 10 81 54 08 00 mov.m ar.ccv=r34
2bc: 00 00 04 00 nop.i 0x0;;
2c0: 09 58 8c 42 11 10 [MMI] cmpxchg4.acq r11=[r33],r35,ar.ccv
2c6: 00 00 00 02 00 00 nop.m 0x0
2cc: 00 00 04 00 nop.i 0x0;;
2d0: 10 00 2c 40 90 11 [MIB] st4 [r32]=r11
2d6: 00 00 00 02 00 00 nop.i 0x0
2dc: 20 00 00 40 br.few 2f0
<cmpxchg_futex_value_locked+0xc0>
2e0: 09 40 c8 f9 ff 27 [MMI] mov r8=-14
2e6: 00 00 00 02 00 00 nop.m 0x0
2ec: 00 00 04 00 nop.i 0x0;;
2f0: 0b 88 20 1a 19 21 [MMI] adds r17=3208,r13;;
2f6: 30 01 44 20 20 00 ld4 r19=[r17]
2fc: 00 00 04 00 nop.i 0x0;;
300: 0b 90 fc 27 3f 23 [MMI] adds r18=-1,r19;;
306: 00 90 44 20 23 00 st4 [r17]=r18
30c: 00 00 04 00 nop.i 0x0;;
310: 11 00 00 00 01 00 [MIB] nop.m 0x0
316: 00 00 00 02 00 80 nop.i 0x0
31c: 08 00 84 00 br.ret.sptk.many b0;;
Much better.
There is a
270: 05 40 00 00 00 e1 [MLX] mov r8=r0
which was generated by C code r8 = 0. Below
2b6: 00 10 81 54 08 00 mov.m ar.ccv=r34
what means that oldval is no longer overwritten.
This is Debian bug#702641
(http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=702641).
The patch is applicable on Kernel 3.9-rc1, 3.2.23 and many other versions.
Signed-off-by: Stephan Schreiber <info@fs-driver.org>
Cc: stable@vger.kernel.org
Signed-off-by: Tony Luck <tony.luck@intel.com>
2013-03-20 02:22:27 +04:00
|
|
|
register unsigned long r8 __asm ("r8") = 0;
|
2011-03-11 05:48:51 +03:00
|
|
|
unsigned long prev;
|
2006-09-27 01:00:56 +04:00
|
|
|
__asm__ __volatile__(
|
|
|
|
|
" mf;; \n"
|
2012-04-17 03:28:01 +04:00
|
|
|
" mov ar.ccv=%4;; \n"
|
|
|
|
|
"[1:] cmpxchg4.acq %1=[%2],%3,ar.ccv \n"
|
2006-09-27 01:00:56 +04:00
|
|
|
" .xdata4 \"__ex_table\", 1b-., 2f-. \n"
|
|
|
|
|
"[2:]"
|
Wrong asm register contraints in the futex implementation
The Linux Kernel contains some inline assembly source code which has
wrong asm register constraints in arch/ia64/include/asm/futex.h.
I observed this on Kernel 3.2.23 but it is also true on the most
recent Kernel 3.9-rc1.
File arch/ia64/include/asm/futex.h:
static inline int
futex_atomic_cmpxchg_inatomic(u32 *uval, u32 __user *uaddr,
u32 oldval, u32 newval)
{
if (!access_ok(VERIFY_WRITE, uaddr, sizeof(u32)))
return -EFAULT;
{
register unsigned long r8 __asm ("r8");
unsigned long prev;
__asm__ __volatile__(
" mf;; \n"
" mov %0=r0 \n"
" mov ar.ccv=%4;; \n"
"[1:] cmpxchg4.acq %1=[%2],%3,ar.ccv \n"
" .xdata4 \"__ex_table\", 1b-., 2f-. \n"
"[2:]"
: "=r" (r8), "=r" (prev)
: "r" (uaddr), "r" (newval),
"rO" ((long) (unsigned) oldval)
: "memory");
*uval = prev;
return r8;
}
}
The list of output registers is
: "=r" (r8), "=r" (prev)
The constraint "=r" means that the GCC has to maintain that these vars
are in registers and contain valid info when the program flow leaves
the assembly block (output registers).
But "=r" also means that GCC can put them in registers that are used
as input registers. Input registers are uaddr, newval, oldval on the
example.
The second assembly instruction
" mov %0=r0 \n"
is the first one which writes to a register; it sets %0 to 0. %0 means
the first register operand; it is r8 here. (The r0 is read-only and
always 0 on the Itanium; it can be used if an immediate zero value is
needed.)
This instruction might overwrite one of the other registers which are
still needed.
Whether it really happens depends on how GCC decides what registers it
uses and how it optimizes the code.
The objdump utility can give us disassembly.
The futex_atomic_cmpxchg_inatomic() function is inline, so we have to
look for a module that uses the funtion. This is the
cmpxchg_futex_value_locked() function in
kernel/futex.c:
static int cmpxchg_futex_value_locked(u32 *curval, u32 __user *uaddr,
u32 uval, u32 newval)
{
int ret;
pagefault_disable();
ret = futex_atomic_cmpxchg_inatomic(curval, uaddr, uval, newval);
pagefault_enable();
return ret;
}
Now the disassembly. At first from the Kernel package 3.2.23 which has
been compiled with GCC 4.4, remeber this Kernel seemed to work:
objdump -d linux-3.2.23/debian/build/build_ia64_none_mckinley/kernel/futex.o
0000000000000230 <cmpxchg_futex_value_locked>:
230: 0b 18 80 1b 18 21 [MMI] adds r3=3168,r13;;
236: 80 40 0d 00 42 00 adds r8=40,r3
23c: 00 00 04 00 nop.i 0x0;;
240: 0b 50 00 10 10 10 [MMI] ld4 r10=[r8];;
246: 90 08 28 00 42 00 adds r9=1,r10
24c: 00 00 04 00 nop.i 0x0;;
250: 09 00 00 00 01 00 [MMI] nop.m 0x0
256: 00 48 20 20 23 00 st4 [r8]=r9
25c: 00 00 04 00 nop.i 0x0;;
260: 08 10 80 06 00 21 [MMI] adds r2=32,r3
266: 00 00 00 02 00 00 nop.m 0x0
26c: 02 08 f1 52 extr.u r16=r33,0,61
270: 05 40 88 00 08 e0 [MLX] addp4 r8=r34,r0
276: ff ff 0f 00 00 e0 movl r15=0xfffffffbfff;;
27c: f1 f7 ff 65
280: 09 70 00 04 18 10 [MMI] ld8 r14=[r2]
286: 00 00 00 02 00 c0 nop.m 0x0
28c: f0 80 1c d0 cmp.ltu p6,p7=r15,r16;;
290: 08 40 fc 1d 09 3b [MMI] cmp.eq p8,p9=-1,r14
296: 00 00 00 02 00 40 nop.m 0x0
29c: e1 08 2d d0 cmp.ltu p10,p11=r14,r33
2a0: 56 01 10 00 40 10 [BBB] (p10) br.cond.spnt.few 2e0
<cmpxchg_futex_value_locked+0xb0>
2a6: 02 08 00 80 21 03 (p08) br.cond.dpnt.few 2b0
<cmpxchg_futex_value_locked+0x80>
2ac: 40 00 00 41 (p06) br.cond.spnt.few 2e0
<cmpxchg_futex_value_locked+0xb0>
2b0: 0a 00 00 00 22 00 [MMI] mf;;
2b6: 80 00 00 00 42 00 mov r8=r0
2bc: 00 00 04 00 nop.i 0x0
2c0: 0b 00 20 40 2a 04 [MMI] mov.m ar.ccv=r8;;
2c6: 10 1a 85 22 20 00 cmpxchg4.acq r33=[r33],r35,ar.ccv
2cc: 00 00 04 00 nop.i 0x0;;
2d0: 10 00 84 40 90 11 [MIB] st4 [r32]=r33
2d6: 00 00 00 02 00 00 nop.i 0x0
2dc: 20 00 00 40 br.few 2f0
<cmpxchg_futex_value_locked+0xc0>
2e0: 09 40 c8 f9 ff 27 [MMI] mov r8=-14
2e6: 00 00 00 02 00 00 nop.m 0x0
2ec: 00 00 04 00 nop.i 0x0;;
2f0: 0b 58 20 1a 19 21 [MMI] adds r11=3208,r13;;
2f6: 20 01 2c 20 20 00 ld4 r18=[r11]
2fc: 00 00 04 00 nop.i 0x0;;
300: 0b 88 fc 25 3f 23 [MMI] adds r17=-1,r18;;
306: 00 88 2c 20 23 00 st4 [r11]=r17
30c: 00 00 04 00 nop.i 0x0;;
310: 11 00 00 00 01 00 [MIB] nop.m 0x0
316: 00 00 00 02 00 80 nop.i 0x0
31c: 08 00 84 00 br.ret.sptk.many b0;;
The lines
2b0: 0a 00 00 00 22 00 [MMI] mf;;
2b6: 80 00 00 00 42 00 mov r8=r0
2bc: 00 00 04 00 nop.i 0x0
2c0: 0b 00 20 40 2a 04 [MMI] mov.m ar.ccv=r8;;
2c6: 10 1a 85 22 20 00 cmpxchg4.acq r33=[r33],r35,ar.ccv
2cc: 00 00 04 00 nop.i 0x0;;
are the instructions of the assembly block.
The line
2b6: 80 00 00 00 42 00 mov r8=r0
sets the r8 register to 0 and after that
2c0: 0b 00 20 40 2a 04 [MMI] mov.m ar.ccv=r8;;
prepares the 'oldvalue' for the cmpxchg but it takes it from r8. This
is wrong.
What happened here is what I explained above: An input register is
overwritten which is still needed.
The register operand constraints in futex.h are wrong.
(The problem doesn't occur when the Kernel is compiled with GCC 4.6.)
The attached patch fixes the register operand constraints in futex.h.
The code after patching of it:
static inline int
futex_atomic_cmpxchg_inatomic(u32 *uval, u32 __user *uaddr,
u32 oldval, u32 newval)
{
if (!access_ok(VERIFY_WRITE, uaddr, sizeof(u32)))
return -EFAULT;
{
register unsigned long r8 __asm ("r8") = 0;
unsigned long prev;
__asm__ __volatile__(
" mf;; \n"
" mov ar.ccv=%4;; \n"
"[1:] cmpxchg4.acq %1=[%2],%3,ar.ccv \n"
" .xdata4 \"__ex_table\", 1b-., 2f-. \n"
"[2:]"
: "+r" (r8), "=&r" (prev)
: "r" (uaddr), "r" (newval),
"rO" ((long) (unsigned) oldval)
: "memory");
*uval = prev;
return r8;
}
}
I also initialized the 'r8' var with the C programming language.
The _asm qualifier on the definition of the 'r8' var forces GCC to use
the r8 processor register for it.
I don't believe that we should use inline assembly for zeroing out a
local variable.
The constraint is
"+r" (r8)
what means that it is both an input register and an output register.
Note that the page fault handler will modify the r8 register which
will be the return value of the function.
The real fix is
"=&r" (prev)
The & means that GCC must not use any of the input registers to place
this output register in.
Patched the Kernel 3.2.23 and compiled it with GCC4.4:
0000000000000230 <cmpxchg_futex_value_locked>:
230: 0b 18 80 1b 18 21 [MMI] adds r3=3168,r13;;
236: 80 40 0d 00 42 00 adds r8=40,r3
23c: 00 00 04 00 nop.i 0x0;;
240: 0b 50 00 10 10 10 [MMI] ld4 r10=[r8];;
246: 90 08 28 00 42 00 adds r9=1,r10
24c: 00 00 04 00 nop.i 0x0;;
250: 09 00 00 00 01 00 [MMI] nop.m 0x0
256: 00 48 20 20 23 00 st4 [r8]=r9
25c: 00 00 04 00 nop.i 0x0;;
260: 08 10 80 06 00 21 [MMI] adds r2=32,r3
266: 20 12 01 10 40 00 addp4 r34=r34,r0
26c: 02 08 f1 52 extr.u r16=r33,0,61
270: 05 40 00 00 00 e1 [MLX] mov r8=r0
276: ff ff 0f 00 00 e0 movl r15=0xfffffffbfff;;
27c: f1 f7 ff 65
280: 09 70 00 04 18 10 [MMI] ld8 r14=[r2]
286: 00 00 00 02 00 c0 nop.m 0x0
28c: f0 80 1c d0 cmp.ltu p6,p7=r15,r16;;
290: 08 40 fc 1d 09 3b [MMI] cmp.eq p8,p9=-1,r14
296: 00 00 00 02 00 40 nop.m 0x0
29c: e1 08 2d d0 cmp.ltu p10,p11=r14,r33
2a0: 56 01 10 00 40 10 [BBB] (p10) br.cond.spnt.few 2e0
<cmpxchg_futex_value_locked+0xb0>
2a6: 02 08 00 80 21 03 (p08) br.cond.dpnt.few 2b0
<cmpxchg_futex_value_locked+0x80>
2ac: 40 00 00 41 (p06) br.cond.spnt.few 2e0
<cmpxchg_futex_value_locked+0xb0>
2b0: 0b 00 00 00 22 00 [MMI] mf;;
2b6: 00 10 81 54 08 00 mov.m ar.ccv=r34
2bc: 00 00 04 00 nop.i 0x0;;
2c0: 09 58 8c 42 11 10 [MMI] cmpxchg4.acq r11=[r33],r35,ar.ccv
2c6: 00 00 00 02 00 00 nop.m 0x0
2cc: 00 00 04 00 nop.i 0x0;;
2d0: 10 00 2c 40 90 11 [MIB] st4 [r32]=r11
2d6: 00 00 00 02 00 00 nop.i 0x0
2dc: 20 00 00 40 br.few 2f0
<cmpxchg_futex_value_locked+0xc0>
2e0: 09 40 c8 f9 ff 27 [MMI] mov r8=-14
2e6: 00 00 00 02 00 00 nop.m 0x0
2ec: 00 00 04 00 nop.i 0x0;;
2f0: 0b 88 20 1a 19 21 [MMI] adds r17=3208,r13;;
2f6: 30 01 44 20 20 00 ld4 r19=[r17]
2fc: 00 00 04 00 nop.i 0x0;;
300: 0b 90 fc 27 3f 23 [MMI] adds r18=-1,r19;;
306: 00 90 44 20 23 00 st4 [r17]=r18
30c: 00 00 04 00 nop.i 0x0;;
310: 11 00 00 00 01 00 [MIB] nop.m 0x0
316: 00 00 00 02 00 80 nop.i 0x0
31c: 08 00 84 00 br.ret.sptk.many b0;;
Much better.
There is a
270: 05 40 00 00 00 e1 [MLX] mov r8=r0
which was generated by C code r8 = 0. Below
2b6: 00 10 81 54 08 00 mov.m ar.ccv=r34
what means that oldval is no longer overwritten.
This is Debian bug#702641
(http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=702641).
The patch is applicable on Kernel 3.9-rc1, 3.2.23 and many other versions.
Signed-off-by: Stephan Schreiber <info@fs-driver.org>
Cc: stable@vger.kernel.org
Signed-off-by: Tony Luck <tony.luck@intel.com>
2013-03-20 02:22:27 +04:00
|
|
|
: "+r" (r8), "=&r" (prev)
|
2006-09-27 01:00:56 +04:00
|
|
|
: "r" (uaddr), "r" (newval),
|
|
|
|
|
"rO" ((long) (unsigned) oldval)
|
|
|
|
|
: "memory");
|
2011-03-11 05:48:51 +03:00
|
|
|
*uval = prev;
|
2006-09-27 01:00:56 +04:00
|
|
|
return r8;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
#endif /* _ASM_FUTEX_H */
|
