* Fw: [PATCH]: Fix sparc atomic operations
@ 2006-03-25 20:46 David S. Miller
2006-03-25 21:03 ` Ulrich Drepper
0 siblings, 1 reply; 2+ messages in thread
From: David S. Miller @ 2006-03-25 20:46 UTC (permalink / raw)
To: libc-hacker
[-- Attachment #1: Type: Text/Plain, Size: 7 bytes --]
Ping?
[-- Attachment #2: Type: Message/Rfc822, Size: 5558 bytes --]
From: "David S. Miller" <davem@davemloft.net>
To: libc-hacker@sources.redhat.com
Cc: jakub@redhat.com
Subject: [PATCH]: Fix sparc atomic operations
Date: Fri, 24 Mar 2006 02:06:30 -0800 (PST)
Message-ID: <20060324.020630.43672826.davem@davemloft.net>
The sparc atomics were missing "memory" clobbers in their
asm statements. This caused miscompilation of the various
pthread atomic primitives.
In one case, pthread_mutex_lock() was miscompiled such that
the ->owner field of the mutex was sampled before we acquired
the lock, thus triggering the "owner == 0" assertion if things
were just right.
Surprisingly nothing in the nptl tests caught this, instead I
discovered it when trying to build MONO, which seemed to be
able to trigger this case quite readily. :-)
Please apply, thanks.
2006-03-24 David S. Miller <davem@sunset.davemloft.net>
* sysdeps/sparc/sparc32/bits/atomic.h
(__v9_compare_and_exchange_val_32_acq): Add "memory" clobber.
* sysdeps/sparc/sparc32/sparcv9/bits/atomic.h
(__arch_compare_and_exchange_val_32_acq, atomic_exchange_acq):
Likewise.
* sysdeps/sparc/sparc64/bits/atomic.h
(__arch_compare_and_exchange_val_32_acq,
__arch_compare_and_exchange_val_64_acq, atomic_exchange_acq):
Likewise.
--- ./sysdeps/sparc/sparc32/bits/atomic.h.~1~ 2006-01-03 16:05:07.000000000 -0800
+++ ./sysdeps/sparc/sparc32/bits/atomic.h 2006-03-24 00:00:06.000000000 -0800
@@ -122,7 +122,7 @@ volatile unsigned char __sparc32_atomic_
__asm __volatile (".word 0xcde05005" \
: "+r" (__acev_tmp), "=m" (*__acev_mem) \
: "r" (__acev_oldval), "m" (*__acev_mem), \
- "r" (__acev_mem)); \
+ "r" (__acev_mem) : "memory"); \
__acev_tmp; })
#endif
--- ./sysdeps/sparc/sparc32/sparcv9/bits/atomic.h.~1~ 2006-01-03 16:05:07.000000000 -0800
+++ ./sysdeps/sparc/sparc32/sparcv9/bits/atomic.h 2006-03-24 00:00:32.000000000 -0800
@@ -59,7 +59,7 @@ typedef uintmax_t uatomic_max_t;
__asm __volatile ("cas [%4], %2, %0" \
: "=r" (__acev_tmp), "=m" (*__acev_mem) \
: "r" (oldval), "m" (*__acev_mem), "r" (__acev_mem), \
- "0" (newval)); \
+ "0" (newval) : "memory"); \
__acev_tmp; })
/* This can be implemented if needed. */
@@ -74,7 +74,7 @@ typedef uintmax_t uatomic_max_t;
if (sizeof (*(mem)) == 4) \
__asm ("swap %0, %1" \
: "=m" (*__memp), "=r" (__oldval) \
- : "m" (*__memp), "1" (__value)); \
+ : "m" (*__memp), "1" (__value) : "memory"); \
else \
abort (); \
__oldval; })
--- ./sysdeps/sparc/sparc64/bits/atomic.h.~1~ 2006-01-03 16:05:07.000000000 -0800
+++ ./sysdeps/sparc/sparc64/bits/atomic.h 2006-03-24 00:01:04.000000000 -0800
@@ -59,7 +59,7 @@ typedef uintmax_t uatomic_max_t;
__asm __volatile ("cas [%4], %2, %0" \
: "=r" (__acev_tmp), "=m" (*__acev_mem) \
: "r" (oldval), "m" (*__acev_mem), "r" (__acev_mem), \
- "0" (newval)); \
+ "0" (newval) : "memory"); \
__acev_tmp; })
#define __arch_compare_and_exchange_val_64_acq(mem, newval, oldval) \
@@ -69,7 +69,7 @@ typedef uintmax_t uatomic_max_t;
__asm __volatile ("casx [%4], %2, %0" \
: "=r" (__acev_tmp), "=m" (*__acev_mem) \
: "r" ((long) (oldval)), "m" (*__acev_mem), \
- "r" (__acev_mem), "0" ((long) (newval))); \
+ "r" (__acev_mem), "0" ((long) (newval)) : "memory"); \
__acev_tmp; })
#define atomic_exchange_acq(mem, newvalue) \
@@ -80,7 +80,7 @@ typedef uintmax_t uatomic_max_t;
if (sizeof (*(mem)) == 4) \
__asm ("swap %0, %1" \
: "=m" (*__memp), "=r" (__oldval) \
- : "m" (*__memp), "1" (__value)); \
+ : "m" (*__memp), "1" (__value) : "memory"); \
else \
{ \
__val = *__memp; \
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2006-03-25 21:03 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2006-03-25 20:46 Fw: [PATCH]: Fix sparc atomic operations David S. Miller
2006-03-25 21:03 ` Ulrich Drepper
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).