Edit report at http://bugs.php.net/bug.php?id=53310&edit=1
ID: 53310
User updated by: stefan at whocares dot de
Reported by: stefan at whocares dot de
Summary: fpm_atomic.h uses SPARC v9 only code, doesn't work
on v8
Status: Analyzed
Type: Feature/Change Request
Package: FPM related
Operating System: Linux (Debian for Sparc)
PHP Version: 5.3.3
Block user comment: N
New Comment:
Well, I blatantly copied from PostgreSQL's s_lock.h and came up with
this:
diff -Nau fpm_atomic.h.org fpm_atomic.h
--- fpm_atomic.h.org 2009-12-14 09:18:53.000000000 +0000
+++ fpm_atomic.h 2010-11-15 01:50:31.000000000 +0000
@@ -82,7 +82,7 @@
#endif /* defined (__GNUC__) &&... */
#elif ( __sparc__ || __sparc ) /* Marcin Ochab */
-
+#if (__sparc_v9__)
#if (__arch64__ || __arch64)
typedef uint64_t atomic_uint_t;
typedef volatile atomic_uint_t atomic_t;
@@ -118,7 +118,23 @@
}
/* }}} */
#endif
+#else /* sparcv9 */
+typedef uint32_t atomic_uint_t;
+typedef volatile atomic_uint_t atomic_t;
+static inline int atomic_cas_32(atomic_t *lock) /* {{{ */
+{
+ register atomic_uint_t _res;
+ __asm__ __volatile__("ldstub [%2], %0" : "=r"(_res),
"+m"(*lock) : "r"(lock) : "memory");
+ return (int) _res;
+}
+/* }}} */
+
+static inline atomic_uint_t atomic_cmp_set(atomic_t *lock,
atomic_uint_t old,
atomic_uint_t set) /* {{{ */
+{
+ return (atomic_cas_32(lock)==0);
+}
+/* }}} */
#else
#error unsupported processor. please write a patch and send it to me
Rationale:
If I'm reading the original code correctly, there's no actual locking
done but
instead the code only tests whether it could acquire a lock. 'ldstub'
works such
that it returns the current value of the memory region specified and
sets it to
all '1' afterwards. Thus, if the return value is '-1' the lock was
already set
by another process whereas if it's '0' we acquired the lock. Well, at
least in
my certainly flawed logic ;) Since ldstub is atomic I didn't see a need
to
explicitly "lock;" the code.
The patch should leave the 'cas' code intact when being compiled on v9
type
SPARC systems. Tested (for successful compilation only!) on Debian
(etch) using
gcc 3.3.5. Thus I believe further testing is necessary to verify this is
actually working.
Well, please test and incorporate if you feel the code is doing what
it's
supposed to do.
Previous Comments:
------------------------------------------------------------------------
[2010-11-15 00:53:49] fat@php.net
As the sparc documentation says
(http://developers.sun.com/solaris/articles/atomic_sparc/#CAS):
The SPARC v9 manual introduced the newest atomic instruction: compare
and swap
(cas)
I don't know how to fix this right now. If you know someone who can,
he's
welcome. I've already asked for help.
wait and see
------------------------------------------------------------------------
[2010-11-15 00:21:50] stefan at whocares dot de
Description:
------------
Compiling with PHP-FPM enabled on an older SPARC system will result in
/tmp/cc6w5Fh0.s: Assembler messages:
/tmp/cc6w5Fh0.s:39: Error: Architecture mismatch on "cas".
/tmp/cc6w5Fh0.s:39: (Requires v9|v9a|v9b; requested architecture is
sparclite.)
Unfortunately my knowledge of SPARC assembly language isn't nearly good
enough to fix that. I know that the v9 "cas" opcode does an atomic
"compare and swap" operation but I wouldn't know how to translate that
into v8 code.
Test script:
---------------
Copy /sapi/fpm/fpm/fpm_atomic.h to fpm_atomic.c and add bogus main()
function:
int main () {
int result;
atomic_t mylock;
result = fpm_spinlock(&mylock, 1);
}
Compile using "gcc -mcpu=v8 fpm_atomic.c" will result in error message
given.
Expected result:
----------------
Should compile without error.
Actual result:
--------------
sparky:~# gcc -mcpu=v8 fpm_atomic.c
/tmp/cciAbMrC.s: Assembler messages:
/tmp/cciAbMrC.s:121: Error: Architecture mismatch on "cas".
/tmp/cciAbMrC.s:121: (Requires v9|v9a|v9b; requested architecture is
sparclite.)
sparky:~#
------------------------------------------------------------------------
--
Edit this bug report at http://bugs.php.net/bug.php?id=53310&edit=1