project-base/server-skynet-source-3rd-jemalloc
2
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

refactor profiling. only use a bytes till next sample variable.

Browse Source
This commit is contained in:
Ben Maurer 2014-04-15 13:47:13 -07:00
parent a7619b7fa5
commit 6c39f9e059
3 changed files with 134 additions and 154 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
include/jemalloc/internal/private_symbols.txt
View File

@ -299,6 +299,7 @@ prof_idump
prof_interval
prof_lookup
prof_malloc
prof_malloc_record_object
prof_mdump
prof_postfork_child
prof_postfork_parent

182
include/jemalloc/internal/prof.h
View File

@ -177,8 +177,7 @@ struct prof_tdata_s {
/* Sampling state. */
uint64_t prng_state;
uint64_t threshold;
uint64_t accum;
uint64_t bytes_until_sample;
/* State used to avoid dumping while operating on prof internals. */
bool enq;
@ -239,6 +238,7 @@ bool prof_boot2(void);
void prof_prefork(void);
void prof_postfork_parent(void);
void prof_postfork_child(void);
void prof_sample_threshold_update(prof_tdata_t *prof_tdata);
#endif /* JEMALLOC_H_EXTERNS */
/******************************************************************************/
@ -250,49 +250,13 @@ void prof_postfork_child(void);
\
assert(size == s2u(size)); \
\
prof_tdata = prof_tdata_get(true); \
if ((uintptr_t)prof_tdata <= (uintptr_t)PROF_TDATA_STATE_MAX) { \
if (prof_tdata != NULL) \
if (!opt_prof_active || \
prof_sample_accum_update(size, false, &prof_tdata)) { \
ret = (prof_thr_cnt_t *)(uintptr_t)1U; \
else \
ret = NULL; \
break; \
} \
\
if (opt_prof_active == false) { \
/* Sampling is currently inactive, so avoid sampling. */\
ret = (prof_thr_cnt_t *)(uintptr_t)1U; \
} else if (opt_lg_prof_sample == 0) { \
/* Don't bother with sampling logic, since sampling */\
/* interval is 1. */\
bt_init(&bt, prof_tdata->vec); \
prof_backtrace(&bt, nignore); \
ret = prof_lookup(&bt); \
} else { \
if (prof_tdata->threshold == 0) { \
/* Initialize. Seed the prng differently for */\
/* each thread. */\
prof_tdata->prng_state = \
(uint64_t)(uintptr_t)&size; \
prof_sample_threshold_update(prof_tdata); \
} \
\
/* Determine whether to capture a backtrace based on */\
/* whether size is enough for prof_accum to reach */\
/* prof_tdata->threshold. However, delay updating */\
/* these variables until prof_{m,re}alloc(), because */\
/* we don't know for sure that the allocation will */\
/* succeed. */\
/* */\
/* Use subtraction rather than addition to avoid */\
/* potential integer overflow. */\
if (size >= prof_tdata->threshold - \
prof_tdata->accum) { \
bt_init(&bt, prof_tdata->vec); \
prof_backtrace(&bt, nignore); \
ret = prof_lookup(&bt); \
} else \
ret = (prof_thr_cnt_t *)(uintptr_t)1U; \
} \
} while (0)
@ -300,10 +264,13 @@ void prof_postfork_child(void);
malloc_tsd_protos(JEMALLOC_ATTR(unused), prof_tdata, prof_tdata_t *)
prof_tdata_t *prof_tdata_get(bool create);
void prof_sample_threshold_update(prof_tdata_t *prof_tdata);
void prof_sample_accum_update(size_t size, bool commit,
prof_tdata_t **prof_tdata_out);
prof_ctx_t *prof_ctx_get(const void *ptr);
void prof_ctx_set(const void *ptr, prof_ctx_t *ctx);
bool prof_sample_accum_update(size_t size);
void prof_malloc_record_object(const void *ptr, size_t usize,
prof_thr_cnt_t *cnt)
void prof_malloc(const void *ptr, size_t usize, prof_thr_cnt_t *cnt);
void prof_realloc(const void *ptr, size_t usize, prof_thr_cnt_t *cnt,
size_t old_usize, prof_ctx_t *old_ctx);
@ -330,55 +297,6 @@ prof_tdata_get(bool create)
return (prof_tdata);
}
JEMALLOC_INLINE void
prof_sample_threshold_update(prof_tdata_t *prof_tdata)
{
/*
* The body of this function is compiled out unless heap profiling is
* enabled, so that it is possible to compile jemalloc with floating
* point support completely disabled. Avoiding floating point code is
* important on memory-constrained systems, but it also enables a
* workaround for versions of glibc that don't properly save/restore
* floating point registers during dynamic lazy symbol loading (which
* internally calls into whatever malloc implementation happens to be
* integrated into the application). Note that some compilers (e.g.
* gcc 4.8) may use floating point registers for fast memory moves, so
* jemalloc must be compiled with such optimizations disabled (e.g.
* -mno-sse) in order for the workaround to be complete.
*/
#ifdef JEMALLOC_PROF
uint64_t r;
double u;
cassert(config_prof);
/*
* Compute sample threshold as a geometrically distributed random
* variable with mean (2^opt_lg_prof_sample).
*
* __ __
* | log(u) | 1
* prof_tdata->threshold = | -------- |, where p = -------------------
* | log(1-p) | opt_lg_prof_sample
* 2
*
* For more information on the math, see:
*
* Non-Uniform Random Variate Generation
* Luc Devroye
* Springer-Verlag, New York, 1986
* pp 500
* (http://luc.devroye.org/rnbookindex.html)
*/
prng64(r, 53, prof_tdata->prng_state,
UINT64_C(6364136223846793005), UINT64_C(1442695040888963407));
u = (double)r * (1.0/9007199254740992.0L);
prof_tdata->threshold = (uint64_t)(log(u) /
log(1.0 - (1.0 / (double)((uint64_t)1U << opt_lg_prof_sample))))
+ (uint64_t)1U;
#endif
}
JEMALLOC_INLINE prof_ctx_t *
prof_ctx_get(const void *ptr)
{
@ -415,56 +333,37 @@ prof_ctx_set(const void *ptr, prof_ctx_t *ctx)
}
JEMALLOC_INLINE bool
prof_sample_accum_update(size_t size)
prof_sample_accum_update(size_t size, bool commit,
prof_tdata_t **prof_tdata_out)
{
prof_tdata_t *prof_tdata;
cassert(config_prof);
/* Sampling logic is unnecessary if the interval is 1. */
assert(opt_lg_prof_sample != 0);
prof_tdata = prof_tdata_get(false);
prof_tdata = prof_tdata_get(true);
if ((uintptr_t)prof_tdata <= (uintptr_t)PROF_TDATA_STATE_MAX)
prof_tdata = NULL;
if (prof_tdata_out != NULL)
*prof_tdata_out = prof_tdata;
if (prof_tdata == NULL)
return (true);
/* Take care to avoid integer overflow. */
if (size >= prof_tdata->threshold - prof_tdata->accum) {
prof_tdata->accum -= (prof_tdata->threshold - size);
/* Compute new sample threshold. */
prof_sample_threshold_update(prof_tdata);
while (prof_tdata->accum >= prof_tdata->threshold) {
prof_tdata->accum -= prof_tdata->threshold;
prof_sample_threshold_update(prof_tdata);
}
return (false);
} else {
prof_tdata->accum += size;
if (prof_tdata->bytes_until_sample >= size) {
if (commit)
prof_tdata->bytes_until_sample -= size;
return (true);
} else {
/* Compute new sample threshold. */
if (commit)
prof_sample_threshold_update(prof_tdata);
return (false);
}
}
JEMALLOC_INLINE void
prof_malloc(const void *ptr, size_t usize, prof_thr_cnt_t *cnt)
{
cassert(config_prof);
assert(ptr != NULL);
assert(usize == isalloc(ptr, true));
if (opt_lg_prof_sample != 0) {
if (prof_sample_accum_update(usize)) {
/*
* Don't sample. For malloc()-like allocation, it is
* always possible to tell in advance how large an
* object's usable size will be, so there should never
* be a difference between the usize passed to
* PROF_ALLOC_PREP() and prof_malloc().
*/
assert((uintptr_t)cnt == (uintptr_t)1U);
}
}
if ((uintptr_t)cnt > (uintptr_t)1U) {
prof_malloc_record_object(const void *ptr, size_t usize, prof_thr_cnt_t *cnt) {
prof_ctx_set(ptr, cnt->ctx);
cnt->epoch++;
@ -484,7 +383,30 @@ prof_malloc(const void *ptr, size_t usize, prof_thr_cnt_t *cnt)
/*********/
mb_write();
/*********/
} else
}
JEMALLOC_INLINE void
prof_malloc(const void *ptr, size_t usize, prof_thr_cnt_t *cnt)
{
cassert(config_prof);
assert(ptr != NULL);
assert(usize == isalloc(ptr, true));
if (prof_sample_accum_update(usize, true, NULL)) {
/*
* Don't sample. For malloc()-like allocation, it is
* always possible to tell in advance how large an
* object's usable size will be, so there should never
* be a difference between the usize passed to
* PROF_ALLOC_PREP() and prof_malloc().
*/
assert((uintptr_t)cnt == (uintptr_t)1U);
}
if ((uintptr_t)cnt > (uintptr_t)1U)
prof_malloc_record_object(ptr, usize, cnt);
else
prof_ctx_set(ptr, (prof_ctx_t *)(uintptr_t)1U);
}
@ -499,8 +421,7 @@ prof_realloc(const void *ptr, size_t usize, prof_thr_cnt_t *cnt,
if (ptr != NULL) {
assert(usize == isalloc(ptr, true));
if (opt_lg_prof_sample != 0) {
if (prof_sample_accum_update(usize)) {
if (prof_sample_accum_update(usize, true, NULL)) {
/*
* Don't sample. The usize passed to
* PROF_ALLOC_PREP() was larger than what
@ -512,7 +433,6 @@ prof_realloc(const void *ptr, size_t usize, prof_thr_cnt_t *cnt,
cnt = (prof_thr_cnt_t *)(uintptr_t)1U;
}
}
}
if ((uintptr_t)old_ctx > (uintptr_t)1U) {
told_cnt = prof_lookup(old_ctx->bt);

65
src/prof.c
View File

@ -645,6 +645,66 @@ prof_lookup(prof_bt_t *bt)
return (ret.p);
}
void
prof_sample_threshold_update(prof_tdata_t *prof_tdata)
{
/*
* The body of this function is compiled out unless heap profiling is
* enabled, so that it is possible to compile jemalloc with floating
* point support completely disabled. Avoiding floating point code is
* important on memory-constrained systems, but it also enables a
* workaround for versions of glibc that don't properly save/restore
* floating point registers during dynamic lazy symbol loading (which
* internally calls into whatever malloc implementation happens to be
* integrated into the application). Note that some compilers (e.g.
* gcc 4.8) may use floating point registers for fast memory moves, so
* jemalloc must be compiled with such optimizations disabled (e.g.
* -mno-sse) in order for the workaround to be complete.
*/
#ifdef JEMALLOC_PROF
uint64_t r;
double u;
if (!config_prof)
return;
if (prof_tdata == NULL)
prof_tdata = prof_tdata_get(false);
if (opt_lg_prof_sample == 0) {
prof_tdata->bytes_until_sample = 0;
return;
}
/*
* Compute sample threshold as a geometrically distributed random
* variable with mean (2^opt_lg_prof_sample).
*
* __ __
* | log(u) | 1
* prof_tdata->threshold = | -------- |, where p = -------------------
* | log(1-p) | opt_lg_prof_sample
* 2
*
* For more information on the math, see:
*
* Non-Uniform Random Variate Generation
* Luc Devroye
* Springer-Verlag, New York, 1986
* pp 500
* (http://luc.devroye.org/rnbookindex.html)
*/
prng64(r, 53, prof_tdata->prng_state,
UINT64_C(6364136223846793005), UINT64_C(1442695040888963407));
u = (double)r * (1.0/9007199254740992.0L);
prof_tdata->bytes_until_sample = (uint64_t)(log(u) /
log(1.0 - (1.0 / (double)((uint64_t)1U << opt_lg_prof_sample))))
+ (uint64_t)1U;
#endif
}
#ifdef JEMALLOC_JET
size_t
prof_bt_count(void)
@ -1224,9 +1284,8 @@ prof_tdata_init(void)
return (NULL);
}
prof_tdata->prng_state = 0;
prof_tdata->threshold = 0;
prof_tdata->accum = 0;
prof_tdata->prng_state = (uint64_t)(uintptr_t)prof_tdata;
prof_sample_threshold_update(prof_tdata);
prof_tdata->enq = false;
prof_tdata->enq_idump = false;