diff options
Diffstat (limited to 'arch/sparc/include/asm/backoff.h')
-rw-r--r-- | arch/sparc/include/asm/backoff.h | 69 |
1 files changed, 59 insertions, 10 deletions
diff --git a/arch/sparc/include/asm/backoff.h b/arch/sparc/include/asm/backoff.h index db3af0d30fb1..4e02086b839c 100644 --- a/arch/sparc/include/asm/backoff.h +++ b/arch/sparc/include/asm/backoff.h | |||
@@ -1,6 +1,46 @@ | |||
1 | #ifndef _SPARC64_BACKOFF_H | 1 | #ifndef _SPARC64_BACKOFF_H |
2 | #define _SPARC64_BACKOFF_H | 2 | #define _SPARC64_BACKOFF_H |
3 | 3 | ||
4 | /* The macros in this file implement an exponential backoff facility | ||
5 | * for atomic operations. | ||
6 | * | ||
7 | * When multiple threads compete on an atomic operation, it is | ||
8 | * possible for one thread to be continually denied a successful | ||
9 | * completion of the compare-and-swap instruction. Heavily | ||
10 | * threaded cpu implementations like Niagara can compound this | ||
11 | * problem even further. | ||
12 | * | ||
13 | * When an atomic operation fails and needs to be retried, we spin a | ||
14 | * certain number of times. At each subsequent failure of the same | ||
15 | * operation we double the spin count, realizing an exponential | ||
16 | * backoff. | ||
17 | * | ||
18 | * When we spin, we try to use an operation that will cause the | ||
19 | * current cpu strand to block, and therefore make the core fully | ||
20 | * available to any other other runnable strands. There are two | ||
21 | * options, based upon cpu capabilities. | ||
22 | * | ||
23 | * On all cpus prior to SPARC-T4 we do three dummy reads of the | ||
24 | * condition code register. Each read blocks the strand for something | ||
25 | * between 40 and 50 cpu cycles. | ||
26 | * | ||
27 | * For SPARC-T4 and later we have a special "pause" instruction | ||
28 | * available. This is implemented using writes to register %asr27. | ||
29 | * The cpu will block the number of cycles written into the register, | ||
30 | * unless a disrupting trap happens first. SPARC-T4 specifically | ||
31 | * implements pause with a granularity of 8 cycles. Each strand has | ||
32 | * an internal pause counter which decrements every 8 cycles. So the | ||
33 | * chip shifts the %asr27 value down by 3 bits, and writes the result | ||
34 | * into the pause counter. If a value smaller than 8 is written, the | ||
35 | * chip blocks for 1 cycle. | ||
36 | * | ||
37 | * To achieve the same amount of backoff as the three %ccr reads give | ||
38 | * on earlier chips, we shift the backoff value up by 7 bits. (Three | ||
39 | * %ccr reads block for about 128 cycles, 1 << 7 == 128) We write the | ||
40 | * whole amount we want to block into the pause register, rather than | ||
41 | * loop writing 128 each time. | ||
42 | */ | ||
43 | |||
4 | #define BACKOFF_LIMIT (4 * 1024) | 44 | #define BACKOFF_LIMIT (4 * 1024) |
5 | 45 | ||
6 | #ifdef CONFIG_SMP | 46 | #ifdef CONFIG_SMP |
@@ -11,16 +51,25 @@ | |||
11 | #define BACKOFF_LABEL(spin_label, continue_label) \ | 51 | #define BACKOFF_LABEL(spin_label, continue_label) \ |
12 | spin_label | 52 | spin_label |
13 | 53 | ||
14 | #define BACKOFF_SPIN(reg, tmp, label) \ | 54 | #define BACKOFF_SPIN(reg, tmp, label) \ |
15 | mov reg, tmp; \ | 55 | mov reg, tmp; \ |
16 | 88: brnz,pt tmp, 88b; \ | 56 | 88: rd %ccr, %g0; \ |
17 | sub tmp, 1, tmp; \ | 57 | rd %ccr, %g0; \ |
18 | set BACKOFF_LIMIT, tmp; \ | 58 | rd %ccr, %g0; \ |
19 | cmp reg, tmp; \ | 59 | .section .pause_3insn_patch,"ax";\ |
20 | bg,pn %xcc, label; \ | 60 | .word 88b; \ |
21 | nop; \ | 61 | sllx tmp, 7, tmp; \ |
22 | ba,pt %xcc, label; \ | 62 | wr tmp, 0, %asr27; \ |
23 | sllx reg, 1, reg; | 63 | clr tmp; \ |
64 | .previous; \ | ||
65 | brnz,pt tmp, 88b; \ | ||
66 | sub tmp, 1, tmp; \ | ||
67 | set BACKOFF_LIMIT, tmp; \ | ||
68 | cmp reg, tmp; \ | ||
69 | bg,pn %xcc, label; \ | ||
70 | nop; \ | ||
71 | ba,pt %xcc, label; \ | ||
72 | sllx reg, 1, reg; | ||
24 | 73 | ||
25 | #else | 74 | #else |
26 | 75 | ||