ea51e97bb8
This can be used to take pressure off a more centralized, worse-sharded allocator without requiring a full break of the arena abstraction.
449 lines
14 KiB
C
449 lines
14 KiB
C
#ifndef JEMALLOC_INTERNAL_CACHE_BIN_H
|
|
#define JEMALLOC_INTERNAL_CACHE_BIN_H
|
|
|
|
#include "jemalloc/internal/ql.h"
|
|
#include "jemalloc/internal/sz.h"
|
|
|
|
/*
|
|
* The cache_bins are the mechanism that the tcache and the arena use to
|
|
* communicate. The tcache fills from and flushes to the arena by passing a
|
|
* cache_bin_t to fill/flush. When the arena needs to pull stats from the
|
|
* tcaches associated with it, it does so by iterating over its
|
|
* cache_bin_array_descriptor_t objects and reading out per-bin stats it
|
|
* contains. This makes it so that the arena need not know about the existence
|
|
* of the tcache at all.
|
|
*/
|
|
|
|
/*
|
|
* The size in bytes of each cache bin stack. We also use this to indicate
|
|
* *counts* of individual objects.
|
|
*/
|
|
typedef uint16_t cache_bin_sz_t;
|
|
|
|
/*
|
|
* That implies the following value, for the maximum number of items in any
|
|
* individual bin. The cache bins track their bounds looking just at the low
|
|
* bits of a pointer, compared against a cache_bin_sz_t. So that's
|
|
* 1 << (sizeof(cache_bin_sz_t) * 8)
|
|
* bytes spread across pointer sized objects to get the maximum.
|
|
*/
|
|
#define CACHE_BIN_NCACHED_MAX (((size_t)1 << sizeof(cache_bin_sz_t) * 8) \
|
|
/ sizeof(void *) - 1)
|
|
|
|
/*
|
|
* This lives inside the cache_bin (for locality reasons), and is initialized
|
|
* alongside it, but is otherwise not modified by any cache bin operations.
|
|
* It's logically public and maintained by its callers.
|
|
*/
|
|
typedef struct cache_bin_stats_s cache_bin_stats_t;
|
|
struct cache_bin_stats_s {
|
|
/*
|
|
* Number of allocation requests that corresponded to the size of this
|
|
* bin.
|
|
*/
|
|
uint64_t nrequests;
|
|
};
|
|
|
|
/*
|
|
* Read-only information associated with each element of tcache_t's tbins array
|
|
* is stored separately, mainly to reduce memory usage.
|
|
*/
|
|
typedef struct cache_bin_info_s cache_bin_info_t;
|
|
struct cache_bin_info_s {
|
|
cache_bin_sz_t ncached_max;
|
|
};
|
|
|
|
/*
|
|
* Responsible for caching allocations associated with a single size.
|
|
*/
|
|
typedef struct cache_bin_s cache_bin_t;
|
|
struct cache_bin_s {
|
|
/*
|
|
* The stack grows down. Whenever the bin is nonempty, the head points
|
|
* to an array entry containing a valid allocation. When it is empty,
|
|
* the head points to one element past the owned array.
|
|
*/
|
|
void **stack_head;
|
|
/*
|
|
* cur_ptr and stats are both modified frequently. Let's keep them
|
|
* close so that they have a higher chance of being on the same
|
|
* cacheline, thus less write-backs.
|
|
*/
|
|
cache_bin_stats_t tstats;
|
|
|
|
/*
|
|
* The low bits of the address of the first item in the stack that
|
|
* hasn't been used since the last GC, to track the low water mark (min
|
|
* # of cached items).
|
|
*
|
|
* Since the stack grows down, this is a higher address than
|
|
* low_bits_full.
|
|
*/
|
|
uint16_t low_bits_low_water;
|
|
|
|
/*
|
|
* The low bits of the value that stack_head will take on when the array
|
|
* is full. (But remember that stack_head always points to a valid item
|
|
* when the array is nonempty -- this is in the array).
|
|
*
|
|
* Recall that since the stack grows down, this is the lowest address in
|
|
* the array.
|
|
*/
|
|
uint16_t low_bits_full;
|
|
|
|
/*
|
|
* The low bits of the value that stack_head will take on when the array
|
|
* is empty.
|
|
*
|
|
* The stack grows down -- this is one past the highest address in the
|
|
* array.
|
|
*/
|
|
uint16_t low_bits_empty;
|
|
};
|
|
|
|
/*
|
|
* The cache_bins live inside the tcache, but the arena (by design) isn't
|
|
* supposed to know much about tcache internals. To let the arena iterate over
|
|
* associated bins, we keep (with the tcache) a linked list of
|
|
* cache_bin_array_descriptor_ts that tell the arena how to find the bins.
|
|
*/
|
|
typedef struct cache_bin_array_descriptor_s cache_bin_array_descriptor_t;
|
|
struct cache_bin_array_descriptor_s {
|
|
/*
|
|
* The arena keeps a list of the cache bins associated with it, for
|
|
* stats collection.
|
|
*/
|
|
ql_elm(cache_bin_array_descriptor_t) link;
|
|
/* Pointers to the tcache bins. */
|
|
cache_bin_t *bins;
|
|
};
|
|
|
|
static inline void
|
|
cache_bin_array_descriptor_init(cache_bin_array_descriptor_t *descriptor,
|
|
cache_bin_t *bins) {
|
|
ql_elm_new(descriptor, link);
|
|
descriptor->bins = bins;
|
|
}
|
|
|
|
/* Returns ncached_max: Upper limit on ncached. */
|
|
static inline cache_bin_sz_t
|
|
cache_bin_info_ncached_max(cache_bin_info_t *info) {
|
|
return info->ncached_max;
|
|
}
|
|
|
|
/*
|
|
* Internal.
|
|
*
|
|
* Asserts that the pointer associated with earlier is <= the one associated
|
|
* with later.
|
|
*/
|
|
static inline void
|
|
cache_bin_assert_earlier(cache_bin_t *bin, uint16_t earlier, uint16_t later) {
|
|
if (earlier > later) {
|
|
assert(bin->low_bits_full > bin->low_bits_empty);
|
|
}
|
|
}
|
|
|
|
/*
|
|
* Internal.
|
|
*
|
|
* Does difference calculations that handle wraparound correctly. Earlier must
|
|
* be associated with the position earlier in memory.
|
|
*/
|
|
static inline uint16_t
|
|
cache_bin_diff(cache_bin_t *bin, uint16_t earlier, uint16_t later) {
|
|
cache_bin_assert_earlier(bin, earlier, later);
|
|
return later - earlier;
|
|
}
|
|
|
|
/* Number of items currently cached in the bin. */
|
|
static inline cache_bin_sz_t
|
|
cache_bin_ncached_get(cache_bin_t *bin, cache_bin_info_t *info) {
|
|
cache_bin_sz_t diff = cache_bin_diff(bin,
|
|
(uint16_t)(uintptr_t)bin->stack_head, bin->low_bits_empty);
|
|
cache_bin_sz_t n = diff / sizeof(void *);
|
|
|
|
assert(n <= cache_bin_info_ncached_max(info));
|
|
assert(n == 0 || *(bin->stack_head) != NULL);
|
|
|
|
return n;
|
|
}
|
|
|
|
/*
|
|
* Internal.
|
|
*
|
|
* A pointer to the position one past the end of the backing array.
|
|
*/
|
|
static inline void **
|
|
cache_bin_empty_position_get(cache_bin_t *bin, cache_bin_info_t *info) {
|
|
cache_bin_sz_t diff = cache_bin_diff(bin,
|
|
(uint16_t)(uintptr_t)bin->stack_head, bin->low_bits_empty);
|
|
uintptr_t empty_bits = (uintptr_t)bin->stack_head + diff;
|
|
void **ret = (void **)empty_bits;
|
|
|
|
assert(ret >= bin->stack_head);
|
|
|
|
return ret;
|
|
}
|
|
|
|
/*
|
|
* As the name implies. This is important since it's not correct to try to
|
|
* batch fill a nonempty cache bin.
|
|
*/
|
|
static inline void
|
|
cache_bin_assert_empty(cache_bin_t *bin, cache_bin_info_t *info) {
|
|
assert(cache_bin_ncached_get(bin, info) == 0);
|
|
assert(cache_bin_empty_position_get(bin, info) == bin->stack_head);
|
|
}
|
|
|
|
/*
|
|
* Get low water, but without any of the correctness checking we do for the
|
|
* caller-usable version, if we are temporarily breaking invariants (like
|
|
* ncached >= low_water during flush).
|
|
*/
|
|
static inline cache_bin_sz_t
|
|
cache_bin_low_water_get_internal(cache_bin_t *bin, cache_bin_info_t *info) {
|
|
return cache_bin_diff(bin, bin->low_bits_low_water,
|
|
bin->low_bits_empty) / sizeof(void *);
|
|
}
|
|
|
|
/* Returns the numeric value of low water in [0, ncached]. */
|
|
static inline cache_bin_sz_t
|
|
cache_bin_low_water_get(cache_bin_t *bin, cache_bin_info_t *info) {
|
|
cache_bin_sz_t low_water = cache_bin_low_water_get_internal(bin, info);
|
|
assert(low_water <= cache_bin_info_ncached_max(info));
|
|
assert(low_water <= cache_bin_ncached_get(bin, info));
|
|
|
|
cache_bin_assert_earlier(bin, (uint16_t)(uintptr_t)bin->stack_head,
|
|
bin->low_bits_low_water);
|
|
|
|
return low_water;
|
|
}
|
|
|
|
/*
|
|
* Indicates that the current cache bin position should be the low water mark
|
|
* going forward.
|
|
*/
|
|
static inline void
|
|
cache_bin_low_water_set(cache_bin_t *bin) {
|
|
bin->low_bits_low_water = (uint16_t)(uintptr_t)bin->stack_head;
|
|
}
|
|
|
|
JEMALLOC_ALWAYS_INLINE void *
|
|
cache_bin_alloc_impl(cache_bin_t *bin, bool *success, bool adjust_low_water) {
|
|
/*
|
|
* success (instead of ret) should be checked upon the return of this
|
|
* function. We avoid checking (ret == NULL) because there is never a
|
|
* null stored on the avail stack (which is unknown to the compiler),
|
|
* and eagerly checking ret would cause pipeline stall (waiting for the
|
|
* cacheline).
|
|
*/
|
|
|
|
/*
|
|
* This may read from the empty position; however the loaded value won't
|
|
* be used. It's safe because the stack has one more slot reserved.
|
|
*/
|
|
void *ret = *bin->stack_head;
|
|
uint16_t low_bits = (uint16_t)(uintptr_t)bin->stack_head;
|
|
void **new_head = bin->stack_head + 1;
|
|
|
|
/*
|
|
* Note that the low water mark is at most empty; if we pass this check,
|
|
* we know we're non-empty.
|
|
*/
|
|
if (likely(low_bits != bin->low_bits_low_water)) {
|
|
bin->stack_head = new_head;
|
|
*success = true;
|
|
return ret;
|
|
}
|
|
if (!adjust_low_water) {
|
|
*success = false;
|
|
return NULL;
|
|
}
|
|
/*
|
|
* In the fast-path case where we call alloc_easy and then alloc, the
|
|
* previous checking and computation is optimized away -- we didn't
|
|
* actually commit any of our operations.
|
|
*/
|
|
if (likely(low_bits != bin->low_bits_empty)) {
|
|
bin->stack_head = new_head;
|
|
bin->low_bits_low_water = (uint16_t)(uintptr_t)new_head;
|
|
*success = true;
|
|
return ret;
|
|
}
|
|
*success = false;
|
|
return NULL;
|
|
}
|
|
|
|
/*
|
|
* Allocate an item out of the bin, failing if we're at the low-water mark.
|
|
*/
|
|
JEMALLOC_ALWAYS_INLINE void *
|
|
cache_bin_alloc_easy(cache_bin_t *bin, bool *success) {
|
|
/* We don't look at info if we're not adjusting low-water. */
|
|
return cache_bin_alloc_impl(bin, success, false);
|
|
}
|
|
|
|
/*
|
|
* Allocate an item out of the bin, even if we're currently at the low-water
|
|
* mark (and failing only if the bin is empty).
|
|
*/
|
|
JEMALLOC_ALWAYS_INLINE void *
|
|
cache_bin_alloc(cache_bin_t *bin, bool *success) {
|
|
return cache_bin_alloc_impl(bin, success, true);
|
|
}
|
|
|
|
/*
|
|
* Free an object into the given bin. Fails only if the bin is full.
|
|
*/
|
|
JEMALLOC_ALWAYS_INLINE bool
|
|
cache_bin_dalloc_easy(cache_bin_t *bin, void *ptr) {
|
|
uint16_t low_bits = (uint16_t)(uintptr_t)bin->stack_head;
|
|
if (unlikely(low_bits == bin->low_bits_full)) {
|
|
return false;
|
|
}
|
|
|
|
bin->stack_head--;
|
|
*bin->stack_head = ptr;
|
|
cache_bin_assert_earlier(bin, bin->low_bits_full,
|
|
(uint16_t)(uintptr_t)bin->stack_head);
|
|
|
|
return true;
|
|
}
|
|
|
|
/**
|
|
* Filling and flushing are done in batch, on arrays of void *s. For filling,
|
|
* the arrays go forward, and can be accessed with ordinary array arithmetic.
|
|
* For flushing, we work from the end backwards, and so need to use special
|
|
* accessors that invert the usual ordering.
|
|
*
|
|
* This is important for maintaining first-fit; the arena code fills with
|
|
* earliest objects first, and so those are the ones we should return first for
|
|
* cache_bin_alloc calls. When flushing, we should flush the objects that we
|
|
* wish to return later; those at the end of the array. This is better for the
|
|
* first-fit heuristic as well as for cache locality; the most recently freed
|
|
* objects are the ones most likely to still be in cache.
|
|
*
|
|
* This all sounds very hand-wavey and theoretical, but reverting the ordering
|
|
* on one or the other pathway leads to measurable slowdowns.
|
|
*/
|
|
|
|
typedef struct cache_bin_ptr_array_s cache_bin_ptr_array_t;
|
|
struct cache_bin_ptr_array_s {
|
|
cache_bin_sz_t n;
|
|
void **ptr;
|
|
};
|
|
|
|
/*
|
|
* Declare a cache_bin_ptr_array_t sufficient for nval items.
|
|
*
|
|
* In the current implementation, this could be just part of a
|
|
* cache_bin_ptr_array_init_... call, since we reuse the cache bin stack memory.
|
|
* Indirecting behind a macro, though, means experimenting with linked-list
|
|
* representations is easy (since they'll require an alloca in the calling
|
|
* frame).
|
|
*/
|
|
#define CACHE_BIN_PTR_ARRAY_DECLARE(name, nval) \
|
|
cache_bin_ptr_array_t name; \
|
|
name.n = (nval)
|
|
|
|
/*
|
|
* Start a fill. The bin must be empty, and This must be followed by a
|
|
* finish_fill call before doing any alloc/dalloc operations on the bin.
|
|
*/
|
|
static inline void
|
|
cache_bin_init_ptr_array_for_fill(cache_bin_t *bin, cache_bin_info_t *info,
|
|
cache_bin_ptr_array_t *arr, cache_bin_sz_t nfill) {
|
|
assert(cache_bin_ncached_get(bin, info) == 0);
|
|
arr->ptr = cache_bin_empty_position_get(bin, info) - nfill;
|
|
}
|
|
|
|
/*
|
|
* While nfill in cache_bin_init_ptr_array_for_fill is the number we *intend* to
|
|
* fill, nfilled here is the number we actually filled (which may be less, in
|
|
* case of OOM.
|
|
*/
|
|
static inline void
|
|
cache_bin_finish_fill(cache_bin_t *bin, cache_bin_info_t *info,
|
|
cache_bin_ptr_array_t *arr, cache_bin_sz_t nfilled) {
|
|
assert(cache_bin_ncached_get(bin, info) == 0);
|
|
void **empty_position = cache_bin_empty_position_get(bin, info);
|
|
if (nfilled < arr->n) {
|
|
memmove(empty_position - nfilled, empty_position - arr->n,
|
|
nfilled * sizeof(void *));
|
|
}
|
|
bin->stack_head = empty_position - nfilled;
|
|
}
|
|
|
|
/* Same deal, but with flush. */
|
|
static inline void
|
|
cache_bin_init_ptr_array_for_flush(cache_bin_t *bin, cache_bin_info_t *info,
|
|
cache_bin_ptr_array_t *arr, cache_bin_sz_t nflush) {
|
|
arr->ptr = cache_bin_empty_position_get(bin, info) - 1;
|
|
assert(cache_bin_ncached_get(bin, info) == 0
|
|
|| *arr->ptr != NULL);
|
|
}
|
|
|
|
/*
|
|
* These accessors are used by the flush pathways -- they reverse ordinary array
|
|
* ordering. See the note above.
|
|
*/
|
|
JEMALLOC_ALWAYS_INLINE void *
|
|
cache_bin_ptr_array_get(cache_bin_ptr_array_t *arr, cache_bin_sz_t n) {
|
|
return *(arr->ptr - n);
|
|
}
|
|
|
|
JEMALLOC_ALWAYS_INLINE void
|
|
cache_bin_ptr_array_set(cache_bin_ptr_array_t *arr, cache_bin_sz_t n, void *p) {
|
|
*(arr->ptr - n) = p;
|
|
}
|
|
|
|
static inline void
|
|
cache_bin_finish_flush(cache_bin_t *bin, cache_bin_info_t *info,
|
|
cache_bin_ptr_array_t *arr, cache_bin_sz_t nflushed) {
|
|
unsigned rem = cache_bin_ncached_get(bin, info) - nflushed;
|
|
memmove(bin->stack_head + nflushed, bin->stack_head,
|
|
rem * sizeof(void *));
|
|
bin->stack_head = bin->stack_head + nflushed;
|
|
if (cache_bin_ncached_get(bin, info)
|
|
< cache_bin_low_water_get_internal(bin, info)) {
|
|
bin->low_bits_low_water = (uint16_t)(uintptr_t)bin->stack_head;
|
|
}
|
|
}
|
|
|
|
/*
|
|
* Initialize a cache_bin_info to represent up to the given number of items in
|
|
* the cache_bins it is associated with.
|
|
*/
|
|
void cache_bin_info_init(cache_bin_info_t *bin_info,
|
|
cache_bin_sz_t ncached_max);
|
|
/*
|
|
* Given an array of initialized cache_bin_info_ts, determine how big an
|
|
* allocation is required to initialize a full set of cache_bin_ts.
|
|
*/
|
|
void cache_bin_info_compute_alloc(cache_bin_info_t *infos, szind_t ninfos,
|
|
size_t *size, size_t *alignment);
|
|
|
|
/*
|
|
* Actually initialize some cache bins. Callers should allocate the backing
|
|
* memory indicated by a call to cache_bin_compute_alloc. They should then
|
|
* preincrement, call init once for each bin and info, and then call
|
|
* cache_bin_postincrement. *alloc_cur will then point immediately past the end
|
|
* of the allocation.
|
|
*/
|
|
void cache_bin_preincrement(cache_bin_info_t *infos, szind_t ninfos,
|
|
void *alloc, size_t *cur_offset);
|
|
void cache_bin_postincrement(cache_bin_info_t *infos, szind_t ninfos,
|
|
void *alloc, size_t *cur_offset);
|
|
void cache_bin_init(cache_bin_t *bin, cache_bin_info_t *info, void *alloc,
|
|
size_t *cur_offset);
|
|
|
|
/*
|
|
* If a cache bin was zero initialized (either because it lives in static or
|
|
* thread-local storage, or was memset to 0), this function indicates whether or
|
|
* not cache_bin_init was called on it.
|
|
*/
|
|
bool cache_bin_still_zero_initialized(cache_bin_t *bin);
|
|
|
|
#endif /* JEMALLOC_INTERNAL_CACHE_BIN_H */
|