]> err.no Git - linux-2.6/commitdiff
[PATCH] mutex subsystem, add include/asm-arm/mutex.h
authorNicolas Pitre <nico@cam.org>
Mon, 9 Jan 2006 23:59:18 +0000 (15:59 -0800)
committerIngo Molnar <mingo@hera.kernel.org>
Mon, 9 Jan 2006 23:59:18 +0000 (15:59 -0800)
add the ARM version of mutex.h, which is optimized in assembly for
ARMv6, and uses the xchg implementation on pre-ARMv6.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
include/asm-arm/mutex.h [new file with mode: 0644]

diff --git a/include/asm-arm/mutex.h b/include/asm-arm/mutex.h
new file mode 100644 (file)
index 0000000..6caa59f
--- /dev/null
@@ -0,0 +1,128 @@
+/*
+ * include/asm-arm/mutex.h
+ *
+ * ARM optimized mutex locking primitives
+ *
+ * Please look into asm-generic/mutex-xchg.h for a formal definition.
+ */
+#ifndef _ASM_MUTEX_H
+#define _ASM_MUTEX_H
+
+#if __LINUX_ARM_ARCH__ < 6
+/* On pre-ARMv6 hardware the swp based implementation is the most efficient. */
+# include <asm-generic/mutex-xchg.h>
+#else
+
+/*
+ * Attempting to lock a mutex on ARMv6+ can be done with a bastardized
+ * atomic decrement (it is not a reliable atomic decrement but it satisfies
+ * the defined semantics for our purpose, while being smaller and faster
+ * than a real atomic decrement or atomic swap.  The idea is to attempt
+ * decrementing the lock value only once.  If once decremented it isn't zero,
+ * or if its store-back fails due to a dispute on the exclusive store, we
+ * simply bail out immediately through the slow path where the lock will be
+ * reattempted until it succeeds.
+ */
+#define __mutex_fastpath_lock(count, fail_fn)                          \
+do {                                                                   \
+       int __ex_flag, __res;                                           \
+                                                                       \
+       typecheck(atomic_t *, count);                                   \
+       typecheck_fn(fastcall void (*)(atomic_t *), fail_fn);           \
+                                                                       \
+       __asm__ (                                                       \
+               "ldrex  %0, [%2]        \n"                             \
+               "sub    %0, %0, #1      \n"                             \
+               "strex  %1, %0, [%2]    \n"                             \
+                                                                       \
+               : "=&r" (__res), "=&r" (__ex_flag)                      \
+               : "r" (&(count)->counter)                               \
+               : "cc","memory" );                                      \
+                                                                       \
+       if (unlikely(__res || __ex_flag))                               \
+               fail_fn(count);                                         \
+} while (0)
+
+#define __mutex_fastpath_lock_retval(count, fail_fn)                   \
+({                                                                     \
+       int __ex_flag, __res;                                           \
+                                                                       \
+       typecheck(atomic_t *, count);                                   \
+       typecheck_fn(fastcall int (*)(atomic_t *), fail_fn);            \
+                                                                       \
+       __asm__ (                                                       \
+               "ldrex  %0, [%2]        \n"                             \
+               "sub    %0, %0, #1      \n"                             \
+               "strex  %1, %0, [%2]    \n"                             \
+                                                                       \
+               : "=&r" (__res), "=&r" (__ex_flag)                      \
+               : "r" (&(count)->counter)                               \
+               : "cc","memory" );                                      \
+                                                                       \
+       __res |= __ex_flag;                                             \
+       if (unlikely(__res != 0))                                       \
+               __res = fail_fn(count);                                 \
+       __res;                                                          \
+})
+
+/*
+ * Same trick is used for the unlock fast path. However the original value,
+ * rather than the result, is used to test for success in order to have
+ * better generated assembly.
+ */
+#define __mutex_fastpath_unlock(count, fail_fn)                                \
+do {                                                                   \
+       int __ex_flag, __res, __orig;                                   \
+                                                                       \
+       typecheck(atomic_t *, count);                                   \
+       typecheck_fn(fastcall void (*)(atomic_t *), fail_fn);           \
+                                                                       \
+       __asm__ (                                                       \
+               "ldrex  %0, [%3]        \n"                             \
+               "add    %1, %0, #1      \n"                             \
+               "strex  %2, %1, [%3]    \n"                             \
+                                                                       \
+               : "=&r" (__orig), "=&r" (__res), "=&r" (__ex_flag)      \
+               : "r" (&(count)->counter)                               \
+               : "cc","memory" );                                      \
+                                                                       \
+       if (unlikely(__orig || __ex_flag))                              \
+               fail_fn(count);                                         \
+} while (0)
+
+/*
+ * If the unlock was done on a contended lock, or if the unlock simply fails
+ * then the mutex remains locked.
+ */
+#define __mutex_slowpath_needs_to_unlock()     1
+
+/*
+ * For __mutex_fastpath_trylock we use another construct which could be
+ * described as a "single value cmpxchg".
+ *
+ * This provides the needed trylock semantics like cmpxchg would, but it is
+ * lighter and less generic than a true cmpxchg implementation.
+ */
+static inline int
+__mutex_fastpath_trylock(atomic_t *count, int (*fail_fn)(atomic_t *))
+{
+       int __ex_flag, __res, __orig;
+
+       __asm__ (
+
+               "1: ldrex       %0, [%3]        \n"
+               "subs           %1, %0, #1      \n"
+               "strexeq        %2, %1, [%3]    \n"
+               "movlt          %0, #0          \n"
+               "cmpeq          %2, #0          \n"
+               "bgt            1b              \n"
+
+               : "=&r" (__orig), "=&r" (__res), "=&r" (__ex_flag)
+               : "r" (&count->counter)
+               : "cc", "memory" );
+
+       return __orig;
+}
+
+#endif
+#endif