5389146191
Remove opt.lg_prof_bt_max, and hard code it to 7. The original intention of this option was to enable faster backtracing by limiting backtrace depth. However, this makes graphical pprof output very difficult to interpret. In practice, decreasing sampling frequency is a better mechanism for limiting profiling overhead.
542 lines
15 KiB
C
542 lines
15 KiB
C
/******************************************************************************/
|
|
#ifdef JEMALLOC_H_TYPES
|
|
|
|
typedef struct prof_bt_s prof_bt_t;
|
|
typedef struct prof_cnt_s prof_cnt_t;
|
|
typedef struct prof_thr_cnt_s prof_thr_cnt_t;
|
|
typedef struct prof_ctx_s prof_ctx_t;
|
|
typedef struct prof_tdata_s prof_tdata_t;
|
|
|
|
/* Option defaults. */
|
|
#define PROF_PREFIX_DEFAULT "jeprof"
|
|
#define LG_PROF_SAMPLE_DEFAULT 0
|
|
#define LG_PROF_INTERVAL_DEFAULT -1
|
|
|
|
/*
|
|
* Hard limit on stack backtrace depth. The version of prof_backtrace() that
|
|
* is based on __builtin_return_address() necessarily has a hard-coded number
|
|
* of backtrace frame handlers, and should be kept in sync with this setting.
|
|
*/
|
|
#define PROF_BT_MAX 128
|
|
|
|
/* Maximum number of backtraces to store in each per thread LRU cache. */
|
|
#define PROF_TCMAX 1024
|
|
|
|
/* Initial hash table size. */
|
|
#define PROF_CKH_MINITEMS 64
|
|
|
|
/* Size of memory buffer to use when writing dump files. */
|
|
#define PROF_DUMP_BUF_SIZE 65536
|
|
|
|
#endif /* JEMALLOC_H_TYPES */
|
|
/******************************************************************************/
|
|
#ifdef JEMALLOC_H_STRUCTS
|
|
|
|
struct prof_bt_s {
|
|
/* Backtrace, stored as len program counters. */
|
|
void **vec;
|
|
unsigned len;
|
|
};
|
|
|
|
#ifdef JEMALLOC_PROF_LIBGCC
|
|
/* Data structure passed to libgcc _Unwind_Backtrace() callback functions. */
|
|
typedef struct {
|
|
prof_bt_t *bt;
|
|
unsigned nignore;
|
|
unsigned max;
|
|
} prof_unwind_data_t;
|
|
#endif
|
|
|
|
struct prof_cnt_s {
|
|
/*
|
|
* Profiling counters. An allocation/deallocation pair can operate on
|
|
* different prof_thr_cnt_t objects that are linked into the same
|
|
* prof_ctx_t cnts_ql, so it is possible for the cur* counters to go
|
|
* negative. In principle it is possible for the *bytes counters to
|
|
* overflow/underflow, but a general solution would require something
|
|
* like 128-bit counters; this implementation doesn't bother to solve
|
|
* that problem.
|
|
*/
|
|
int64_t curobjs;
|
|
int64_t curbytes;
|
|
uint64_t accumobjs;
|
|
uint64_t accumbytes;
|
|
};
|
|
|
|
struct prof_thr_cnt_s {
|
|
/* Linkage into prof_ctx_t's cnts_ql. */
|
|
ql_elm(prof_thr_cnt_t) cnts_link;
|
|
|
|
/* Linkage into thread's LRU. */
|
|
ql_elm(prof_thr_cnt_t) lru_link;
|
|
|
|
/*
|
|
* Associated context. If a thread frees an object that it did not
|
|
* allocate, it is possible that the context is not cached in the
|
|
* thread's hash table, in which case it must be able to look up the
|
|
* context, insert a new prof_thr_cnt_t into the thread's hash table,
|
|
* and link it into the prof_ctx_t's cnts_ql.
|
|
*/
|
|
prof_ctx_t *ctx;
|
|
|
|
/*
|
|
* Threads use memory barriers to update the counters. Since there is
|
|
* only ever one writer, the only challenge is for the reader to get a
|
|
* consistent read of the counters.
|
|
*
|
|
* The writer uses this series of operations:
|
|
*
|
|
* 1) Increment epoch to an odd number.
|
|
* 2) Update counters.
|
|
* 3) Increment epoch to an even number.
|
|
*
|
|
* The reader must assure 1) that the epoch is even while it reads the
|
|
* counters, and 2) that the epoch doesn't change between the time it
|
|
* starts and finishes reading the counters.
|
|
*/
|
|
unsigned epoch;
|
|
|
|
/* Profiling counters. */
|
|
prof_cnt_t cnts;
|
|
};
|
|
|
|
struct prof_ctx_s {
|
|
/* Associated backtrace. */
|
|
prof_bt_t *bt;
|
|
|
|
/* Protects cnt_merged and cnts_ql. */
|
|
malloc_mutex_t lock;
|
|
|
|
/* Temporary storage for summation during dump. */
|
|
prof_cnt_t cnt_summed;
|
|
|
|
/* When threads exit, they merge their stats into cnt_merged. */
|
|
prof_cnt_t cnt_merged;
|
|
|
|
/*
|
|
* List of profile counters, one for each thread that has allocated in
|
|
* this context.
|
|
*/
|
|
ql_head(prof_thr_cnt_t) cnts_ql;
|
|
};
|
|
|
|
struct prof_tdata_s {
|
|
/*
|
|
* Hash of (prof_bt_t *)-->(prof_thr_cnt_t *). Each thread keeps a
|
|
* cache of backtraces, with associated thread-specific prof_thr_cnt_t
|
|
* objects. Other threads may read the prof_thr_cnt_t contents, but no
|
|
* others will ever write them.
|
|
*
|
|
* Upon thread exit, the thread must merge all the prof_thr_cnt_t
|
|
* counter data into the associated prof_ctx_t objects, and unlink/free
|
|
* the prof_thr_cnt_t objects.
|
|
*/
|
|
ckh_t bt2cnt;
|
|
|
|
/* LRU for contents of bt2cnt. */
|
|
ql_head(prof_thr_cnt_t) lru_ql;
|
|
|
|
/* Backtrace vector, used for calls to prof_backtrace(). */
|
|
void **vec;
|
|
|
|
/* Sampling state. */
|
|
uint64_t prn_state;
|
|
uint64_t threshold;
|
|
uint64_t accum;
|
|
};
|
|
|
|
#endif /* JEMALLOC_H_STRUCTS */
|
|
/******************************************************************************/
|
|
#ifdef JEMALLOC_H_EXTERNS
|
|
|
|
extern bool opt_prof;
|
|
/*
|
|
* Even if opt_prof is true, sampling can be temporarily disabled by setting
|
|
* opt_prof_active to false. No locking is used when updating opt_prof_active,
|
|
* so there are no guarantees regarding how long it will take for all threads
|
|
* to notice state changes.
|
|
*/
|
|
extern bool opt_prof_active;
|
|
extern size_t opt_lg_prof_sample; /* Mean bytes between samples. */
|
|
extern ssize_t opt_lg_prof_interval; /* lg(prof_interval). */
|
|
extern bool opt_prof_gdump; /* High-water memory dumping. */
|
|
extern bool opt_prof_leak; /* Dump leak summary at exit. */
|
|
extern bool opt_prof_accum; /* Report cumulative bytes. */
|
|
extern char opt_prof_prefix[PATH_MAX + 1];
|
|
|
|
/*
|
|
* Profile dump interval, measured in bytes allocated. Each arena triggers a
|
|
* profile dump when it reaches this threshold. The effect is that the
|
|
* interval between profile dumps averages prof_interval, though the actual
|
|
* interval between dumps will tend to be sporadic, and the interval will be a
|
|
* maximum of approximately (prof_interval * narenas).
|
|
*/
|
|
extern uint64_t prof_interval;
|
|
|
|
/*
|
|
* If true, promote small sampled objects to large objects, since small run
|
|
* headers do not have embedded profile context pointers.
|
|
*/
|
|
extern bool prof_promote;
|
|
|
|
/* Thread-specific backtrace cache, used to reduce bt2ctx contention. */
|
|
#ifndef NO_TLS
|
|
extern __thread prof_tdata_t *prof_tdata_tls
|
|
JEMALLOC_ATTR(tls_model("initial-exec"));
|
|
# define PROF_TCACHE_GET() prof_tdata_tls
|
|
# define PROF_TCACHE_SET(v) do { \
|
|
prof_tdata_tls = (v); \
|
|
pthread_setspecific(prof_tdata_tsd, (void *)(v)); \
|
|
} while (0)
|
|
#else
|
|
# define PROF_TCACHE_GET() \
|
|
((prof_tdata_t *)pthread_getspecific(prof_tdata_tsd))
|
|
# define PROF_TCACHE_SET(v) do { \
|
|
pthread_setspecific(prof_tdata_tsd, (void *)(v)); \
|
|
} while (0)
|
|
#endif
|
|
/*
|
|
* Same contents as b2cnt_tls, but initialized such that the TSD destructor is
|
|
* called when a thread exits, so that prof_tdata_tls contents can be merged,
|
|
* unlinked, and deallocated.
|
|
*/
|
|
extern pthread_key_t prof_tdata_tsd;
|
|
|
|
void bt_init(prof_bt_t *bt, void **vec);
|
|
void prof_backtrace(prof_bt_t *bt, unsigned nignore);
|
|
prof_thr_cnt_t *prof_lookup(prof_bt_t *bt);
|
|
void prof_idump(void);
|
|
bool prof_mdump(const char *filename);
|
|
void prof_gdump(void);
|
|
prof_tdata_t *prof_tdata_init(void);
|
|
void prof_boot0(void);
|
|
void prof_boot1(void);
|
|
bool prof_boot2(void);
|
|
|
|
#endif /* JEMALLOC_H_EXTERNS */
|
|
/******************************************************************************/
|
|
#ifdef JEMALLOC_H_INLINES
|
|
|
|
#define PROF_ALLOC_PREP(nignore, size, ret) do { \
|
|
prof_tdata_t *prof_tdata; \
|
|
prof_bt_t bt; \
|
|
\
|
|
assert(size == s2u(size)); \
|
|
\
|
|
prof_tdata = PROF_TCACHE_GET(); \
|
|
if (prof_tdata == NULL) { \
|
|
prof_tdata = prof_tdata_init(); \
|
|
if (prof_tdata == NULL) { \
|
|
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->prn_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)
|
|
|
|
#ifndef JEMALLOC_ENABLE_INLINE
|
|
void prof_sample_threshold_update(prof_tdata_t *prof_tdata);
|
|
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(const void *ptr, size_t size, prof_thr_cnt_t *cnt);
|
|
void prof_realloc(const void *ptr, size_t size, prof_thr_cnt_t *cnt,
|
|
size_t old_size, prof_ctx_t *old_ctx);
|
|
void prof_free(const void *ptr, size_t size);
|
|
#endif
|
|
|
|
#if (defined(JEMALLOC_ENABLE_INLINE) || defined(JEMALLOC_PROF_C_))
|
|
JEMALLOC_INLINE void
|
|
prof_sample_threshold_update(prof_tdata_t *prof_tdata)
|
|
{
|
|
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://cg.scs.carleton.ca/~luc/rnbookindex.html)
|
|
*/
|
|
prn64(r, 53, prof_tdata->prn_state,
|
|
(uint64_t)6364136223846793005LLU, (uint64_t)1442695040888963407LLU);
|
|
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;
|
|
}
|
|
|
|
JEMALLOC_INLINE prof_ctx_t *
|
|
prof_ctx_get(const void *ptr)
|
|
{
|
|
prof_ctx_t *ret;
|
|
arena_chunk_t *chunk;
|
|
|
|
cassert(config_prof);
|
|
assert(ptr != NULL);
|
|
|
|
chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr);
|
|
if (chunk != ptr) {
|
|
/* Region. */
|
|
ret = arena_prof_ctx_get(ptr);
|
|
} else
|
|
ret = huge_prof_ctx_get(ptr);
|
|
|
|
return (ret);
|
|
}
|
|
|
|
JEMALLOC_INLINE void
|
|
prof_ctx_set(const void *ptr, prof_ctx_t *ctx)
|
|
{
|
|
arena_chunk_t *chunk;
|
|
|
|
cassert(config_prof);
|
|
assert(ptr != NULL);
|
|
|
|
chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr);
|
|
if (chunk != ptr) {
|
|
/* Region. */
|
|
arena_prof_ctx_set(ptr, ctx);
|
|
} else
|
|
huge_prof_ctx_set(ptr, ctx);
|
|
}
|
|
|
|
JEMALLOC_INLINE bool
|
|
prof_sample_accum_update(size_t size)
|
|
{
|
|
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_TCACHE_GET();
|
|
assert(prof_tdata != NULL);
|
|
|
|
/* 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;
|
|
return (true);
|
|
}
|
|
}
|
|
|
|
JEMALLOC_INLINE void
|
|
prof_malloc(const void *ptr, size_t size, prof_thr_cnt_t *cnt)
|
|
{
|
|
|
|
cassert(config_prof);
|
|
assert(ptr != NULL);
|
|
assert(size == isalloc(ptr));
|
|
|
|
if (opt_lg_prof_sample != 0) {
|
|
if (prof_sample_accum_update(size)) {
|
|
/*
|
|
* 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 size passed to
|
|
* PROF_ALLOC_PREP() and prof_malloc().
|
|
*/
|
|
assert((uintptr_t)cnt == (uintptr_t)1U);
|
|
}
|
|
}
|
|
|
|
if ((uintptr_t)cnt > (uintptr_t)1U) {
|
|
prof_ctx_set(ptr, cnt->ctx);
|
|
|
|
cnt->epoch++;
|
|
/*********/
|
|
mb_write();
|
|
/*********/
|
|
cnt->cnts.curobjs++;
|
|
cnt->cnts.curbytes += size;
|
|
if (opt_prof_accum) {
|
|
cnt->cnts.accumobjs++;
|
|
cnt->cnts.accumbytes += size;
|
|
}
|
|
/*********/
|
|
mb_write();
|
|
/*********/
|
|
cnt->epoch++;
|
|
/*********/
|
|
mb_write();
|
|
/*********/
|
|
} else
|
|
prof_ctx_set(ptr, (prof_ctx_t *)(uintptr_t)1U);
|
|
}
|
|
|
|
JEMALLOC_INLINE void
|
|
prof_realloc(const void *ptr, size_t size, prof_thr_cnt_t *cnt,
|
|
size_t old_size, prof_ctx_t *old_ctx)
|
|
{
|
|
prof_thr_cnt_t *told_cnt;
|
|
|
|
cassert(config_prof);
|
|
assert(ptr != NULL || (uintptr_t)cnt <= (uintptr_t)1U);
|
|
|
|
if (ptr != NULL) {
|
|
assert(size == isalloc(ptr));
|
|
if (opt_lg_prof_sample != 0) {
|
|
if (prof_sample_accum_update(size)) {
|
|
/*
|
|
* Don't sample. The size passed to
|
|
* PROF_ALLOC_PREP() was larger than what
|
|
* actually got allocated, so a backtrace was
|
|
* captured for this allocation, even though
|
|
* its actual size was insufficient to cross
|
|
* the sample threshold.
|
|
*/
|
|
cnt = (prof_thr_cnt_t *)(uintptr_t)1U;
|
|
}
|
|
}
|
|
}
|
|
|
|
if ((uintptr_t)old_ctx > (uintptr_t)1U) {
|
|
told_cnt = prof_lookup(old_ctx->bt);
|
|
if (told_cnt == NULL) {
|
|
/*
|
|
* It's too late to propagate OOM for this realloc(),
|
|
* so operate directly on old_cnt->ctx->cnt_merged.
|
|
*/
|
|
malloc_mutex_lock(&old_ctx->lock);
|
|
old_ctx->cnt_merged.curobjs--;
|
|
old_ctx->cnt_merged.curbytes -= old_size;
|
|
malloc_mutex_unlock(&old_ctx->lock);
|
|
told_cnt = (prof_thr_cnt_t *)(uintptr_t)1U;
|
|
}
|
|
} else
|
|
told_cnt = (prof_thr_cnt_t *)(uintptr_t)1U;
|
|
|
|
if ((uintptr_t)told_cnt > (uintptr_t)1U)
|
|
told_cnt->epoch++;
|
|
if ((uintptr_t)cnt > (uintptr_t)1U) {
|
|
prof_ctx_set(ptr, cnt->ctx);
|
|
cnt->epoch++;
|
|
} else
|
|
prof_ctx_set(ptr, (prof_ctx_t *)(uintptr_t)1U);
|
|
/*********/
|
|
mb_write();
|
|
/*********/
|
|
if ((uintptr_t)told_cnt > (uintptr_t)1U) {
|
|
told_cnt->cnts.curobjs--;
|
|
told_cnt->cnts.curbytes -= old_size;
|
|
}
|
|
if ((uintptr_t)cnt > (uintptr_t)1U) {
|
|
cnt->cnts.curobjs++;
|
|
cnt->cnts.curbytes += size;
|
|
if (opt_prof_accum) {
|
|
cnt->cnts.accumobjs++;
|
|
cnt->cnts.accumbytes += size;
|
|
}
|
|
}
|
|
/*********/
|
|
mb_write();
|
|
/*********/
|
|
if ((uintptr_t)told_cnt > (uintptr_t)1U)
|
|
told_cnt->epoch++;
|
|
if ((uintptr_t)cnt > (uintptr_t)1U)
|
|
cnt->epoch++;
|
|
/*********/
|
|
mb_write(); /* Not strictly necessary. */
|
|
}
|
|
|
|
JEMALLOC_INLINE void
|
|
prof_free(const void *ptr, size_t size)
|
|
{
|
|
prof_ctx_t *ctx = prof_ctx_get(ptr);
|
|
|
|
cassert(config_prof);
|
|
|
|
if ((uintptr_t)ctx > (uintptr_t)1) {
|
|
assert(size == isalloc(ptr));
|
|
prof_thr_cnt_t *tcnt = prof_lookup(ctx->bt);
|
|
|
|
if (tcnt != NULL) {
|
|
tcnt->epoch++;
|
|
/*********/
|
|
mb_write();
|
|
/*********/
|
|
tcnt->cnts.curobjs--;
|
|
tcnt->cnts.curbytes -= size;
|
|
/*********/
|
|
mb_write();
|
|
/*********/
|
|
tcnt->epoch++;
|
|
/*********/
|
|
mb_write();
|
|
/*********/
|
|
} else {
|
|
/*
|
|
* OOM during free() cannot be propagated, so operate
|
|
* directly on cnt->ctx->cnt_merged.
|
|
*/
|
|
malloc_mutex_lock(&ctx->lock);
|
|
ctx->cnt_merged.curobjs--;
|
|
ctx->cnt_merged.curbytes -= size;
|
|
malloc_mutex_unlock(&ctx->lock);
|
|
}
|
|
}
|
|
}
|
|
#endif
|
|
|
|
#endif /* JEMALLOC_H_INLINES */
|
|
/******************************************************************************/
|