romenskiy2012/jemalloc
1
0
Fork 0
mirror of https://github.com/jemalloc/jemalloc.git synced 2026-04-22 19:02:13 +03:00
Code Issues Projects Releases Packages Wiki Activity

[thread_event] Remove macros from thread_event and replace with dynamic event objects

Browse source
This commit is contained in:
Slobodan Predolac 2025-03-14 06:34:05 -07:00 • committed by guangli-dai
parent 1972241cd2
commit 153fab2b00
17 changed files with 455 additions and 318 deletions
Download patch file Download diff file Expand all files Collapse all files

17
include/jemalloc/internal/peak_event.h
View file

@ -4,6 +4,14 @@
#include "jemalloc/internal/jemalloc_preamble.h"
#include "jemalloc/internal/tsd_types.h"
/*
* Update every 64K by default. We're not exposing this as a configuration
* option for now; we don't want to bind ourselves too tightly to any particular
* performance requirements for small values, or guarantee that we'll even be
* able to provide fine-grained accuracy.
*/
#define PEAK_EVENT_WAIT (64 * 1024)
/*
* While peak.h contains the simple helper struct that tracks state, this
* contains the allocator tie-ins (and knows about tsd, the event module, etc.).
@ -15,13 +23,6 @@ void peak_event_update(tsd_t *tsd);
void peak_event_zero(tsd_t *tsd);
uint64_t peak_event_max(tsd_t *tsd);
/* Manual hooks. */
/* The activity-triggered hooks. */
uint64_t peak_alloc_new_event_wait(tsd_t *tsd);
uint64_t peak_alloc_postponed_event_wait(tsd_t *tsd);
void peak_alloc_event_handler(tsd_t *tsd, uint64_t elapsed);
uint64_t peak_dalloc_new_event_wait(tsd_t *tsd);
uint64_t peak_dalloc_postponed_event_wait(tsd_t *tsd);
void peak_dalloc_event_handler(tsd_t *tsd, uint64_t elapsed);
extern te_base_cb_t peak_te_handler;
#endif /* JEMALLOC_INTERNAL_PEAK_EVENT_H */

63
include/jemalloc/internal/prof_externs.h
View file

@ -5,6 +5,7 @@
#include "jemalloc/internal/base.h"
#include "jemalloc/internal/mutex.h"
#include "jemalloc/internal/prof_hook.h"
#include "jemalloc/internal/thread_event_registry.h"
extern bool opt_prof;
extern bool opt_prof_active;
@ -104,9 +105,65 @@ void prof_prefork1(tsdn_t *tsdn);
void prof_postfork_parent(tsdn_t *tsdn);
void prof_postfork_child(tsdn_t *tsdn);
/* Only accessed by thread event. */
uint64_t prof_sample_new_event_wait(tsd_t *tsd);
uint64_t prof_sample_postponed_event_wait(tsd_t *tsd);
void prof_sample_event_handler(tsd_t *tsd, uint64_t elapsed);
uint64_t tsd_prof_sample_event_wait_get(tsd_t *tsd);
/*
* The lookahead functionality facilitates events to be able to lookahead, i.e.
* without touching the event counters, to determine whether an event would be
* triggered. The event counters are not advanced until the end of the
* allocation / deallocation calls, so the lookahead can be useful if some
* preparation work for some event must be done early in the allocation /
* deallocation calls.
*
* Currently only the profiling sampling event needs the lookahead
* functionality, so we don't yet define general purpose lookahead functions.
*
* Surplus is a terminology referring to the amount of bytes beyond what's
* needed for triggering an event, which can be a useful quantity to have in
* general when lookahead is being called.
*
* This function returns true if allocation of usize would go above the next
* trigger for prof event, and false otherwise.
* If function returns true surplus will contain number of bytes beyond that
* trigger.
*/
JEMALLOC_ALWAYS_INLINE bool
te_prof_sample_event_lookahead_surplus(tsd_t *tsd, size_t usize,
size_t *surplus) {
if (surplus != NULL) {
/*
* This is a dead store: the surplus will be overwritten before
* any read. The initialization suppresses compiler warnings.
* Meanwhile, using SIZE_MAX to initialize is good for
* debugging purpose, because a valid surplus value is strictly
* less than usize, which is at most SIZE_MAX.
*/
*surplus = SIZE_MAX;
}
if (unlikely(!tsd_nominal(tsd) || tsd_reentrancy_level_get(tsd) > 0)) {
return false;
}
/* The subtraction is intentionally susceptible to underflow. */
uint64_t accumbytes = tsd_thread_allocated_get(tsd) + usize -
tsd_thread_allocated_last_event_get(tsd);
uint64_t sample_wait = tsd_prof_sample_event_wait_get(tsd);
if (accumbytes < sample_wait) {
return false;
}
assert(accumbytes - sample_wait < (uint64_t)usize);
if (surplus != NULL) {
*surplus = (size_t)(accumbytes - sample_wait);
}
return true;
}
JEMALLOC_ALWAYS_INLINE bool
te_prof_sample_event_lookahead(tsd_t *tsd, size_t usize) {
return te_prof_sample_event_lookahead_surplus(tsd, usize, NULL);
}
extern te_base_cb_t prof_sample_te_handler;
#endif /* JEMALLOC_INTERNAL_PROF_EXTERNS_H */

5
include/jemalloc/internal/prof_threshold.h
View file

@ -3,9 +3,6 @@
#include "jemalloc/internal/tsd_types.h"
/* The activity-triggered hooks. */
uint64_t prof_threshold_new_event_wait(tsd_t *tsd);
uint64_t prof_threshold_postponed_event_wait(tsd_t *tsd);
void prof_threshold_event_handler(tsd_t *tsd, uint64_t elapsed);
extern te_base_cb_t prof_threshold_te_handler;
#endif /* JEMALLOC_INTERNAL_THRESHOLD_EVENT_H */

5
include/jemalloc/internal/stats.h
View file

@ -3,6 +3,7 @@
#include "jemalloc/internal/jemalloc_preamble.h"
#include "jemalloc/internal/jemalloc_internal_types.h"
#include "jemalloc/internal/thread_event_registry.h"
#include "jemalloc/internal/tsd_types.h"
/* OPTION(opt, var_name, default, set_value_to) */
@ -43,9 +44,7 @@ extern char opt_stats_interval_opts[stats_print_tot_num_options+1];
#define STATS_INTERVAL_ACCUM_BATCH_MAX (4 << 20)
/* Only accessed by thread event. */
uint64_t stats_interval_new_event_wait(tsd_t *tsd);
uint64_t stats_interval_postponed_event_wait(tsd_t *tsd);
void stats_interval_event_handler(tsd_t *tsd, uint64_t elapsed);
extern te_base_cb_t stats_interval_te_handler;
/* Implements je_malloc_stats_print. */
void stats_print(write_cb_t *write_cb, void *cbopaque, const char *opts);

3
include/jemalloc/internal/tcache_externs.h
View file

@ -6,6 +6,7 @@
#include "jemalloc/internal/cache_bin.h"
#include "jemalloc/internal/sz.h"
#include "jemalloc/internal/tcache_types.h"
#include "jemalloc/internal/thread_event_registry.h"
extern bool opt_tcache;
extern size_t opt_tcache_max;
@ -89,4 +90,6 @@ uint64_t tcache_gc_dalloc_new_event_wait(tsd_t *tsd);
uint64_t tcache_gc_dalloc_postponed_event_wait(tsd_t *tsd);
void tcache_gc_dalloc_event_handler(tsd_t *tsd, uint64_t elapsed);
extern te_base_cb_t tcache_gc_te_handler;
#endif /* JEMALLOC_INTERNAL_TCACHE_EXTERNS_H */

84
include/jemalloc/internal/thread_event.h
View file

@ -49,29 +49,12 @@ void te_event_trigger(tsd_t *tsd, te_ctx_t *ctx);
void te_recompute_fast_threshold(tsd_t *tsd);
void tsd_te_init(tsd_t *tsd);
/*
* List of all events, in the following format:
* E(event, (condition), is_alloc_event)
*/
#define ITERATE_OVER_ALL_EVENTS \
E(tcache_gc, (opt_tcache_gc_incr_bytes > 0), true) \
E(prof_sample, (config_prof && opt_prof), true) \
E(prof_threshold, config_stats, true) \
E(stats_interval, (opt_stats_interval >= 0), true) \
E(tcache_gc_dalloc, (opt_tcache_gc_incr_bytes > 0), false) \
E(peak_alloc, config_stats, true) \
E(peak_dalloc, config_stats, false)
#define E(event, condition_unused, is_alloc_event_unused) \
C(event##_event_wait)
/* List of all thread event counters. */
#define ITERATE_OVER_ALL_COUNTERS \
C(thread_allocated) \
C(thread_allocated_last_event) \
ITERATE_OVER_ALL_EVENTS \
C(prof_sample_last_event) \
C(stats_interval_last_event)
#define ITERATE_OVER_ALL_COUNTERS \
C(thread_allocated) \
C(thread_allocated_last_event) \
C(prof_sample_last_event) \
C(stats_interval_last_event)
/* Getters directly wrap TSD getters. */
#define C(counter) \
@ -99,12 +82,6 @@ counter##_set(tsd_t *tsd, uint64_t v) { \
ITERATE_OVER_ALL_COUNTERS
#undef C
/*
* For generating _event_wait getter / setter functions for each individual
* event.
*/
#undef E
/*
* The malloc and free fastpath getters -- use the unsafe getters since tsd may
* be non-nominal, in which case the fast_threshold will be set to 0. This
@ -221,57 +198,6 @@ te_ctx_get(tsd_t *tsd, te_ctx_t *ctx, bool is_alloc) {
}
}
/*
* The lookahead functionality facilitates events to be able to lookahead, i.e.
* without touching the event counters, to determine whether an event would be
* triggered. The event counters are not advanced until the end of the
* allocation / deallocation calls, so the lookahead can be useful if some
* preparation work for some event must be done early in the allocation /
* deallocation calls.
*
* Currently only the profiling sampling event needs the lookahead
* functionality, so we don't yet define general purpose lookahead functions.
*
* Surplus is a terminology referring to the amount of bytes beyond what's
* needed for triggering an event, which can be a useful quantity to have in
* general when lookahead is being called.
*/
JEMALLOC_ALWAYS_INLINE bool
te_prof_sample_event_lookahead_surplus(tsd_t *tsd, size_t usize,
size_t *surplus) {
if (surplus != NULL) {
/*
* This is a dead store: the surplus will be overwritten before
* any read. The initialization suppresses compiler warnings.
* Meanwhile, using SIZE_MAX to initialize is good for
* debugging purpose, because a valid surplus value is strictly
* less than usize, which is at most SIZE_MAX.
*/
*surplus = SIZE_MAX;
}
if (unlikely(!tsd_nominal(tsd) || tsd_reentrancy_level_get(tsd) > 0)) {
return false;
}
/* The subtraction is intentionally susceptible to underflow. */
uint64_t accumbytes = tsd_thread_allocated_get(tsd) + usize -
tsd_thread_allocated_last_event_get(tsd);
uint64_t sample_wait = tsd_prof_sample_event_wait_get(tsd);
if (accumbytes < sample_wait) {
return false;
}
assert(accumbytes - sample_wait < (uint64_t)usize);
if (surplus != NULL) {
*surplus = (size_t)(accumbytes - sample_wait);
}
return true;
}
JEMALLOC_ALWAYS_INLINE bool
te_prof_sample_event_lookahead(tsd_t *tsd, size_t usize) {
return te_prof_sample_event_lookahead_surplus(tsd, usize, NULL);
}
JEMALLOC_ALWAYS_INLINE void
te_event_advance(tsd_t *tsd, size_t usize, bool is_alloc) {
te_assert_invariants(tsd);

58
include/jemalloc/internal/thread_event_registry.h Normal file
View file

@ -0,0 +1,58 @@
#ifndef JEMALLOC_INTERNAL_THREAD_EVENT_REGISTRY_H
#define JEMALLOC_INTERNAL_THREAD_EVENT_REGISTRY_H
#include "jemalloc/internal/jemalloc_preamble.h"
#include "jemalloc/internal/tsd.h"
/* "te" is short for "thread_event" */
enum te_alloc_e {
#ifdef JEMALLOC_PROF
te_alloc_prof_sample,
#endif
te_alloc_stats_interval,
#ifdef JEMALLOC_STATS
te_alloc_prof_threshold,
#endif
te_alloc_tcache_gc,
#ifdef JEMALLOC_STATS
te_alloc_peak,
te_alloc_last = te_alloc_peak,
#else
te_alloc_last = te_alloc_tcache_gc,
#endif
te_alloc_count = te_alloc_last + 1
};
typedef enum te_alloc_e te_alloc_t;
enum te_dalloc_e {
te_dalloc_tcache_gc,
#ifdef JEMALLOC_STATS
te_dalloc_peak,
te_dalloc_last = te_dalloc_peak,
#else
te_dalloc_last = te_dalloc_tcache_gc,
#endif
te_dalloc_count = te_dalloc_last + 1
};
typedef enum te_dalloc_e te_dalloc_t;
/* These will live in tsd */
typedef struct te_data_s te_data_t;
struct te_data_s {
uint64_t alloc_wait[te_alloc_count];
uint64_t dalloc_wait[te_dalloc_count];
};
#define TE_DATA_INITIALIZER { {0}, {0} }
typedef struct te_base_cb_s te_base_cb_t;
struct te_base_cb_s {
bool (*enabled)(void);
uint64_t (*new_event_wait)(tsd_t *tsd);
uint64_t (*postponed_event_wait)(tsd_t *tsd);
void (*event_handler)(tsd_t *tsd);
};
extern te_base_cb_t *te_alloc_handlers[te_alloc_count];
extern te_base_cb_t *te_dalloc_handlers[te_dalloc_count];
#endif /* JEMALLOC_INTERNAL_THREAD_EVENT_REGISTRY_H */

19
include/jemalloc/internal/tsd_internals.h
View file

@ -15,6 +15,7 @@
#include "jemalloc/internal/rtree_tsd.h"
#include "jemalloc/internal/tcache_structs.h"
#include "jemalloc/internal/tcache_types.h"
#include "jemalloc/internal/thread_event_registry.h"
#include "jemalloc/internal/tsd_types.h"
#include "jemalloc/internal/util.h"
#include "jemalloc/internal/witness.h"
@ -68,15 +69,9 @@ typedef ql_elm(tsd_t) tsd_link_t;
O(thread_allocated_next_event, uint64_t, uint64_t) \
O(thread_deallocated_last_event, uint64_t, uint64_t) \
O(thread_deallocated_next_event, uint64_t, uint64_t) \
O(tcache_gc_event_wait, uint64_t, uint64_t) \
O(tcache_gc_dalloc_event_wait, uint64_t, uint64_t) \
O(prof_sample_event_wait, uint64_t, uint64_t) \
O(te_data, te_data_t, te_data_t) \
O(prof_sample_last_event, uint64_t, uint64_t) \
O(prof_threshold_event_wait, uint64_t, uint64_t) \
O(stats_interval_event_wait, uint64_t, uint64_t) \
O(stats_interval_last_event, uint64_t, uint64_t) \
O(peak_alloc_event_wait, uint64_t, uint64_t) \
O(peak_dalloc_event_wait, uint64_t, uint64_t) \
O(stats_interval_last_event, uint64_t, uint64_t) \
O(prof_tdata, prof_tdata_t *, prof_tdata_t *) \
O(prng_state, uint64_t, uint64_t) \
O(san_extents_until_guard_small, uint64_t, uint64_t) \
@ -102,15 +97,9 @@ typedef ql_elm(tsd_t) tsd_link_t;
/* thread_allocated_next_event */ 0, \
/* thread_deallocated_last_event */ 0, \
/* thread_deallocated_next_event */ 0, \
/* tcache_gc_event_wait */ 0, \
/* tcache_gc_dalloc_event_wait */ 0, \
/* prof_sample_event_wait */ 0, \
/* te_data */ TE_DATA_INITIALIZER, \
/* prof_sample_last_event */ 0, \
/* prof_threshold_event_wait */ 0, \
/* stats_interval_event_wait */ 0, \
/* stats_interval_last_event */ 0, \
/* peak_alloc_event_wait */ 0, \
/* peak_dalloc_event_wait */ 0, \
/* prof_tdata */ NULL, \
/* prng_state */ 0, \
/* san_extents_until_guard_small */ 0, \