diff options
author | Peter Zijlstra <peterz@infradead.org> | 2018-04-05 11:05:35 +0200 |
---|---|---|
committer | Ingo Molnar <mingo@kernel.org> | 2018-05-04 10:02:39 +0200 |
commit | c427f69564e2a844c5fcf2804042609342513da0 (patch) | |
tree | 2db714fd8600eb96734dd88c2c6f9bb4beff7b3e /include/asm-generic | |
parent | de5b55c1d4e30740009864eb35ce4ed856aac01d (diff) | |
download | linux-exynos-c427f69564e2a844c5fcf2804042609342513da0.tar.gz linux-exynos-c427f69564e2a844c5fcf2804042609342513da0.tar.bz2 linux-exynos-c427f69564e2a844c5fcf2804042609342513da0.zip |
locking/mutex: Optimize __mutex_trylock_fast()
Use try_cmpxchg to avoid the pointless TEST instruction..
And add the (missing) atomic_long_try_cmpxchg*() wrappery.
On x86_64 this gives:
0000000000000710 <mutex_lock>: 0000000000000710 <mutex_lock>:
710: 65 48 8b 14 25 00 00 mov %gs:0x0,%rdx 710: 65 48 8b 14 25 00 00 mov %gs:0x0,%rdx
717: 00 00 717: 00 00
715: R_X86_64_32S current_task 715: R_X86_64_32S current_task
719: 31 c0 xor %eax,%eax 719: 31 c0 xor %eax,%eax
71b: f0 48 0f b1 17 lock cmpxchg %rdx,(%rdi) 71b: f0 48 0f b1 17 lock cmpxchg %rdx,(%rdi)
720: 48 85 c0 test %rax,%rax 720: 75 02 jne 724 <mutex_lock+0x14>
723: 75 02 jne 727 <mutex_lock+0x17> 722: f3 c3 repz retq
725: f3 c3 repz retq 724: eb da jmp 700 <__mutex_lock_slowpath>
727: eb d7 jmp 700 <__mutex_lock_slowpath> 726: 66 2e 0f 1f 84 00 00 nopw %cs:0x0(%rax,%rax,1)
729: 0f 1f 80 00 00 00 00 nopl 0x0(%rax) 72d: 00 00 00
On ARM64 this gives:
000000000000638 <mutex_lock>: 0000000000000638 <mutex_lock>:
638: d5384101 mrs x1, sp_el0 638: d5384101 mrs x1, sp_el0
63c: d2800002 mov x2, #0x0 63c: d2800002 mov x2, #0x0
640: f9800011 prfm pstl1strm, [x0] 640: f9800011 prfm pstl1strm, [x0]
644: c85ffc03 ldaxr x3, [x0] 644: c85ffc03 ldaxr x3, [x0]
648: ca020064 eor x4, x3, x2 648: ca020064 eor x4, x3, x2
64c: b5000064 cbnz x4, 658 <mutex_lock+0x20> 64c: b5000064 cbnz x4, 658 <mutex_lock+0x20>
650: c8047c01 stxr w4, x1, [x0] 650: c8047c01 stxr w4, x1, [x0]
654: 35ffff84 cbnz w4, 644 <mutex_lock+0xc> 654: 35ffff84 cbnz w4, 644 <mutex_lock+0xc>
658: b40000c3 cbz x3, 670 <mutex_lock+0x38> 658: b5000043 cbnz x3, 660 <mutex_lock+0x28>
65c: a9bf7bfd stp x29, x30, [sp,#-16]! 65c: d65f03c0 ret
660: 910003fd mov x29, sp 660: a9bf7bfd stp x29, x30, [sp,#-16]!
664: 97ffffef bl 620 <__mutex_lock_slowpath> 664: 910003fd mov x29, sp
668: a8c17bfd ldp x29, x30, [sp],#16 668: 97ffffee bl 620 <__mutex_lock_slowpath>
66c: d65f03c0 ret 66c: a8c17bfd ldp x29, x30, [sp],#16
670: d65f03c0 ret 670: d65f03c0 ret
Reported-by: Matthew Wilcox <mawilcox@microsoft.com>
Acked-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Diffstat (limited to 'include/asm-generic')
-rw-r--r-- | include/asm-generic/atomic-long.h | 17 |
1 files changed, 17 insertions, 0 deletions
diff --git a/include/asm-generic/atomic-long.h b/include/asm-generic/atomic-long.h index 5b2b0b5ea06d..87d14476edc2 100644 --- a/include/asm-generic/atomic-long.h +++ b/include/asm-generic/atomic-long.h @@ -25,6 +25,7 @@ typedef atomic64_t atomic_long_t; #define ATOMIC_LONG_INIT(i) ATOMIC64_INIT(i) #define ATOMIC_LONG_PFX(x) atomic64 ## x +#define ATOMIC_LONG_TYPE s64 #else @@ -32,6 +33,7 @@ typedef atomic_t atomic_long_t; #define ATOMIC_LONG_INIT(i) ATOMIC_INIT(i) #define ATOMIC_LONG_PFX(x) atomic ## x +#define ATOMIC_LONG_TYPE int #endif @@ -90,6 +92,21 @@ ATOMIC_LONG_ADD_SUB_OP(sub, _release) #define atomic_long_cmpxchg(l, old, new) \ (ATOMIC_LONG_PFX(_cmpxchg)((ATOMIC_LONG_PFX(_t) *)(l), (old), (new))) + +#define atomic_long_try_cmpxchg_relaxed(l, old, new) \ + (ATOMIC_LONG_PFX(_try_cmpxchg_relaxed)((ATOMIC_LONG_PFX(_t) *)(l), \ + (ATOMIC_LONG_TYPE *)(old), (ATOMIC_LONG_TYPE)(new))) +#define atomic_long_try_cmpxchg_acquire(l, old, new) \ + (ATOMIC_LONG_PFX(_try_cmpxchg_acquire)((ATOMIC_LONG_PFX(_t) *)(l), \ + (ATOMIC_LONG_TYPE *)(old), (ATOMIC_LONG_TYPE)(new))) +#define atomic_long_try_cmpxchg_release(l, old, new) \ + (ATOMIC_LONG_PFX(_try_cmpxchg_release)((ATOMIC_LONG_PFX(_t) *)(l), \ + (ATOMIC_LONG_TYPE *)(old), (ATOMIC_LONG_TYPE)(new))) +#define atomic_long_try_cmpxchg(l, old, new) \ + (ATOMIC_LONG_PFX(_try_cmpxchg)((ATOMIC_LONG_PFX(_t) *)(l), \ + (ATOMIC_LONG_TYPE *)(old), (ATOMIC_LONG_TYPE)(new))) + + #define atomic_long_xchg_relaxed(v, new) \ (ATOMIC_LONG_PFX(_xchg_relaxed)((ATOMIC_LONG_PFX(_t) *)(v), (new))) #define atomic_long_xchg_acquire(v, new) \ |