1 From 0d1622d7f526311d87d7da2ee7dd14b73e45d3fc Mon Sep 17 00:00:00 2001
2 From: Avi Kivity <avi@redhat.com>
3 Date: Sat, 13 Feb 2010 10:33:12 +0200
4 Subject: x86-64, rwsem: Avoid store forwarding hazard in __downgrade_write
6 From: Avi Kivity <avi@redhat.com>
8 commit 0d1622d7f526311d87d7da2ee7dd14b73e45d3fc upstream.
10 The Intel Architecture Optimization Reference Manual states that a short
11 load that follows a long store to the same object will suffer a store
12 forwading penalty, particularly if the two accesses use different addresses.
13 Trivially, a long load that follows a short store will also suffer a penalty.
15 __downgrade_write() in rwsem incurs both penalties: the increment operation
16 will not be able to reuse a recently-loaded rwsem value, and its result will
17 not be reused by any recently-following rwsem operation.
19 A comment in the code states that this is because 64-bit immediates are
20 special and expensive; but while they are slightly special (only a single
21 instruction allows them), they aren't expensive: a test shows that two loops,
22 one loading a 32-bit immediate and one loading a 64-bit immediate, both take
23 1.5 cycles per iteration.
25 Fix this by changing __downgrade_write to use the same add instruction on
26 i386 and on x86_64, so that it uses the same operand size as all the other
29 Signed-off-by: Avi Kivity <avi@redhat.com>
30 LKML-Reference: <1266049992-17419-1-git-send-email-avi@redhat.com>
31 Signed-off-by: H. Peter Anvin <hpa@zytor.com>
32 Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
35 arch/x86/include/asm/rwsem.h | 25 +++++--------------------
36 1 file changed, 5 insertions(+), 20 deletions(-)
38 --- a/arch/x86/include/asm/rwsem.h
39 +++ b/arch/x86/include/asm/rwsem.h
40 @@ -232,34 +232,19 @@ static inline void __up_write(struct rw_
42 static inline void __downgrade_write(struct rw_semaphore *sem)
45 -# if RWSEM_WAITING_BIAS != -0x100000000
46 -# error "This code assumes RWSEM_WAITING_BIAS == -2^32"
49 - /* 64-bit immediates are special and expensive, and not needed here */
50 - asm volatile("# beginning __downgrade_write\n\t"
51 - LOCK_PREFIX "incl 4(%1)\n\t"
52 - /* transitions 0xZZZZZZZZ00000001 -> 0xYYYYYYYY00000001 */
54 - " call call_rwsem_downgrade_wake\n"
56 - "# ending __downgrade_write\n"
61 asm volatile("# beginning __downgrade_write\n\t"
62 LOCK_PREFIX _ASM_ADD "%2,(%1)\n\t"
63 - /* transitions 0xZZZZ0001 -> 0xYYYY0001 */
65 + * transitions 0xZZZZ0001 -> 0xYYYY0001 (i386)
66 + * 0xZZZZZZZZ00000001 -> 0xYYYYYYYY00000001 (x86_64)
69 " call call_rwsem_downgrade_wake\n"
71 "# ending __downgrade_write\n"
73 - : "a" (sem), "i" (-RWSEM_WAITING_BIAS)
74 + : "a" (sem), "er" (-RWSEM_WAITING_BIAS)