summaryrefslogtreecommitdiff
path: root/arch/sparc/include/asm
diff options
context:
space:
mode:
authorDavid S. Miller <davem@davemloft.net>2012-10-29 00:04:47 +0400
committerDavid S. Miller <davem@davemloft.net>2012-10-29 00:04:47 +0400
commit187818cd6a5ab6343eac47e52da2f3e40c544b98 (patch)
tree09b55dad78359ae7e047f3eb533844adaa4094c6 /arch/sparc/include/asm
parente9b9eb59ffcdee09ec96b040f85c919618f4043e (diff)
downloadlinux-187818cd6a5ab6343eac47e52da2f3e40c544b98.tar.xz
sparc64: Improvde documentation and readability of atomic backoff code.
Document what's going on in asm/backoff.h with a large and descriptive comment. Refer to it above the cpu_relax() definition in asm/processor_64.h Rename the pause patching section to have "3insn" in it's name like the other patching sections do. Based upon feedback from Sam Ravnborg. Signed-off-by: David S. Miller <davem@davemloft.net>
Diffstat (limited to 'arch/sparc/include/asm')
-rw-r--r--arch/sparc/include/asm/backoff.h42
-rw-r--r--arch/sparc/include/asm/processor_64.h7
2 files changed, 47 insertions, 2 deletions
diff --git a/arch/sparc/include/asm/backoff.h b/arch/sparc/include/asm/backoff.h
index 20f01df0871b..4e02086b839c 100644
--- a/arch/sparc/include/asm/backoff.h
+++ b/arch/sparc/include/asm/backoff.h
@@ -1,6 +1,46 @@
#ifndef _SPARC64_BACKOFF_H
#define _SPARC64_BACKOFF_H
+/* The macros in this file implement an exponential backoff facility
+ * for atomic operations.
+ *
+ * When multiple threads compete on an atomic operation, it is
+ * possible for one thread to be continually denied a successful
+ * completion of the compare-and-swap instruction. Heavily
+ * threaded cpu implementations like Niagara can compound this
+ * problem even further.
+ *
+ * When an atomic operation fails and needs to be retried, we spin a
+ * certain number of times. At each subsequent failure of the same
+ * operation we double the spin count, realizing an exponential
+ * backoff.
+ *
+ * When we spin, we try to use an operation that will cause the
+ * current cpu strand to block, and therefore make the core fully
+ * available to any other other runnable strands. There are two
+ * options, based upon cpu capabilities.
+ *
+ * On all cpus prior to SPARC-T4 we do three dummy reads of the
+ * condition code register. Each read blocks the strand for something
+ * between 40 and 50 cpu cycles.
+ *
+ * For SPARC-T4 and later we have a special "pause" instruction
+ * available. This is implemented using writes to register %asr27.
+ * The cpu will block the number of cycles written into the register,
+ * unless a disrupting trap happens first. SPARC-T4 specifically
+ * implements pause with a granularity of 8 cycles. Each strand has
+ * an internal pause counter which decrements every 8 cycles. So the
+ * chip shifts the %asr27 value down by 3 bits, and writes the result
+ * into the pause counter. If a value smaller than 8 is written, the
+ * chip blocks for 1 cycle.
+ *
+ * To achieve the same amount of backoff as the three %ccr reads give
+ * on earlier chips, we shift the backoff value up by 7 bits. (Three
+ * %ccr reads block for about 128 cycles, 1 << 7 == 128) We write the
+ * whole amount we want to block into the pause register, rather than
+ * loop writing 128 each time.
+ */
+
#define BACKOFF_LIMIT (4 * 1024)
#ifdef CONFIG_SMP
@@ -16,7 +56,7 @@
88: rd %ccr, %g0; \
rd %ccr, %g0; \
rd %ccr, %g0; \
- .section .pause_patch,"ax"; \
+ .section .pause_3insn_patch,"ax";\
.word 88b; \
sllx tmp, 7, tmp; \
wr tmp, 0, %asr27; \
diff --git a/arch/sparc/include/asm/processor_64.h b/arch/sparc/include/asm/processor_64.h
index 9cdf52eec48a..721e25f0e2ea 100644
--- a/arch/sparc/include/asm/processor_64.h
+++ b/arch/sparc/include/asm/processor_64.h
@@ -196,11 +196,16 @@ extern unsigned long get_wchan(struct task_struct *task);
#define KSTK_EIP(tsk) (task_pt_regs(tsk)->tpc)
#define KSTK_ESP(tsk) (task_pt_regs(tsk)->u_regs[UREG_FP])
+/* Please see the commentary in asm/backoff.h for a description of
+ * what these instructions are doing and how they have been choosen.
+ * To make a long story short, we are trying to yield the current cpu
+ * strand during busy loops.
+ */
#define cpu_relax() asm volatile("\n99:\n\t" \
"rd %%ccr, %%g0\n\t" \
"rd %%ccr, %%g0\n\t" \
"rd %%ccr, %%g0\n\t" \
- ".section .pause_patch,\"ax\"\n\t"\
+ ".section .pause_3insn_patch,\"ax\"\n\t"\
".word 99b\n\t" \
"wr %%g0, 128, %%asr27\n\t" \
"nop\n\t" \