Linus had mentioned that he wished there were a read_trylock that could be
used when taking locks out of order to avoid the usually unnecessary full
treatment with its interrupt-blocked overhead when you unlock and relock.
This came up when adding a lock hierarchy dance to part of do_sigaction,
and the use of it is demonstrated by the patch below (forgive the fuzz).
The following header patches add the read_trylock macro, implementing it
fully for i386 and leaving other architectures with an always-fail fallback.
Other architectures include/asm-foo/spinlock.h files should define
_raw_read_trylock appropriately and define HAVE_ARCH_RAW_READ_TRYLOCK. But
until they do, machine-independent code can use read_trylock for optimization
purposes (i.e. where it falls back to using read_lock) and architectures that
haven't implemented it will just not be getting the optimization.
The machine-dependent function is simple enough to add to each architecture.
Even if I didn't know the x86 as well as I do, writing the new function by
looking at the existing assembly code in _raw_read_lock is easy enough
without really grokking the architecture (just tedious, so I'm not doing it
myself).
Enjoy,
Roland
Index: linux-2.5/kernel/signal.c
===================================================================
RCS file: /home/cvs/linux-2.5/kernel/signal.c,v
retrieving revision 1.98
diff -p -u -r1.98 signal.c
--- linux-2.5/kernel/signal.c 7 Aug 2003 20:06:12 -0000 1.98
+++ linux-2.5/kernel/signal.c 12 Aug 2003 21:38:15 -0000
@@ -2255,9 +2267,11 @@ do_sigaction(int sig, const struct k_sig
* dance to maintain the lock hierarchy.
*/
struct task_struct *t = current;
- spin_unlock_irq(&t->sighand->siglock);
- read_lock(&tasklist_lock);
- spin_lock_irq(&t->sighand->siglock);
+ if (!read_trylock(&tasklist_lock)) {
+ spin_unlock_irq(&t->sighand->siglock);
+ read_lock(&tasklist_lock);
+ spin_lock_irq(&t->sighand->siglock);
+ }
*k = *act;
sigdelsetmask(&k->sa.sa_mask,
sigmask(SIGKILL) | sigmask(SIGSTOP));
Index: linux-2.5/include/linux/spinlock.h
===================================================================
RCS file: /home/cvs/linux-2.5/include/linux/spinlock.h,v
retrieving revision 1.27
diff -b -p -u -r1.27 spinlock.h
--- linux-2.5/include/linux/spinlock.h 26 Jun 2003 01:19:21 -0000 1.27
+++ linux-2.5/include/linux/spinlock.h 12 Aug 2003 21:58:24 -0000
@@ -38,6 +38,10 @@
#ifdef CONFIG_SMP
#include <asm/spinlock.h>
+# ifndef HAVE_ARCH_RAW_READ_TRYLOCK
+# define _raw_read_trylock(lock) ({ (void)(lock); (0); })
+# endif
+
#else
#if !defined(CONFIG_PREEMPT) && !defined(CONFIG_DEBUG_SPINLOCK)
@@ -178,6 +182,7 @@ typedef struct {
#define rwlock_init(lock) do { (void)(lock); } while(0)
#define _raw_read_lock(lock) do { (void)(lock); } while(0)
#define _raw_read_unlock(lock) do { (void)(lock); } while(0)
+#define _raw_read_trylock(lock) ({ (void)(lock); (1); })
#define _raw_write_lock(lock) do { (void)(lock); } while(0)
#define _raw_write_unlock(lock) do { (void)(lock); } while(0)
#define _raw_write_trylock(lock) ({ (void)(lock); (1); })
@@ -194,8 +199,8 @@ typedef struct {
#define write_trylock(lock) ({preempt_disable();_raw_write_trylock(lock) ? \
1 : ({preempt_enable(); 0;});})
-
-/* Where's read_trylock? */
+#define read_trylock(lock) ({preempt_disable();_raw_read_trylock(lock) ? \
+ 1 : ({preempt_enable(); 0;});})
#if defined(CONFIG_SMP) && defined(CONFIG_PREEMPT)
void __preempt_spin_lock(spinlock_t *lock);
Index: linux-2.5/include/asm-i386/spinlock.h
===================================================================
RCS file: /home/cvs/linux-2.5/include/asm-i386/spinlock.h,v
retrieving revision 1.11
diff -b -p -u -r1.11 spinlock.h
--- linux-2.5/include/asm-i386/spinlock.h 19 May 2003 17:30:30 -0000 1.11
+++ linux-2.5/include/asm-i386/spinlock.h 12 Aug 2003 21:58:24 -0000
@@ -195,5 +195,22 @@ static inline int _raw_write_trylock(rwl
atomic_add(RW_LOCK_BIAS, count);
return 0;
}
+
+static inline int _raw_read_trylock(rwlock_t *lock)
+{
+ unsigned char locked;
+ asm volatile(LOCK "subl %2,%0\n\t"
+ "setns %1\n\t"
+ "js 2f\n\t"
+ "1:\n"
+ LOCK_SECTION_START("")
+ "2:\t" LOCK "incl %0\n\t"
+ "jmp 1b\n"
+ LOCK_SECTION_END
+ : "=m" (lock->lock), "=qm" (locked)
+ : "ir" (1), "0" (lock->lock) : "memory");
+ return locked;
+}
+#define HAVE_ARCH_RAW_READ_TRYLOCK 1
#endif /* __ASM_SPINLOCK_H */
On Tue, Aug 12, 2003 at 11:14:17PM -0700, Roland McGrath wrote:
> The following header patches add the read_trylock macro, implementing it
> fully for i386 and leaving other architectures with an always-fail fallback.
> Other architectures include/asm-foo/spinlock.h files should define
> _raw_read_trylock appropriately and define HAVE_ARCH_RAW_READ_TRYLOCK. But
> until they do, machine-independent code can use read_trylock for optimization
> purposes (i.e. where it falls back to using read_lock) and architectures that
> haven't implemented it will just not be getting the optimization.
I strongy dislike the always-fail case. Just let the other arches fail
until the arch maintainers update them.
Roland McGrath <[email protected]> wrote:
>
> The following header patches add the read_trylock macro, implementing it
> fully for i386 and leaving other architectures with an always-fail fallback.
Looks reasonable, and we do need read_trylock() for symmetry at least. And
I did need it for the CONFIG_PREEMT low-latency read_lock() implementation.
Please note that the lockmeter patch in -mm kernels has implementations of
_raw_read_trylock() for sparc64, alpha and ia64. ppc64 already implements
it.
So it would be better if someone could sweep all those together, implement
the necessary stubs for uniprocessor builds on all architectures and
apologetically break the build on the remaining SMP architectures.
That would appear to be mips, parisc, s390, sparc and x86_64.
> So it would be better if someone could sweep all those together, implement
> the necessary stubs for uniprocessor builds on all architectures and
> apologetically break the build on the remaining SMP architectures.
The uniprocessor stub is in linux/spinlock.h and already included in my patch.
All that's required is supplying _raw_read_trylock in asm/spinlock.h.
On Wed, 13 Aug 2003, Andrew Morton wrote:
> So it would be better if someone could sweep all those together, implement
> the necessary stubs for uniprocessor builds on all architectures and
> apologetically break the build on the remaining SMP architectures.
>
> That would appear to be mips, parisc, s390, sparc and x86_64.
Well, the code for mips will be less complicated even. I can supply it
if the i386 patch goes in.
--
+ Maciej W. Rozycki, Technical University of Gdansk, Poland +
+--------------------------------------------------------------+
+ e-mail: [email protected], PGP key available +