server-skynet-source-3rd-je.../src/eset.c
David Goldblatt dcb7b83fac Eset: Cache summary information for heap edatas.
This lets us do a single array scan to find first fits, instead of taking a
cache miss per examined size class.
2021-08-02 15:02:49 -07:00

283 lines
8.6 KiB
C

#include "jemalloc/internal/jemalloc_preamble.h"
#include "jemalloc/internal/jemalloc_internal_includes.h"
#include "jemalloc/internal/eset.h"
#define ESET_NPSIZES (SC_NPSIZES + 1)
static void
eset_bin_init(eset_bin_t *bin) {
edata_heap_new(&bin->heap);
/*
* heap_min doesn't need initialization; it gets filled in when the bin
* goes from non-empty to empty.
*/
}
static void
eset_bin_stats_init(eset_bin_stats_t *bin_stats) {
atomic_store_zu(&bin_stats->nextents, 0, ATOMIC_RELAXED);
atomic_store_zu(&bin_stats->nbytes, 0, ATOMIC_RELAXED);
}
void
eset_init(eset_t *eset, extent_state_t state) {
for (unsigned i = 0; i < ESET_NPSIZES; i++) {
eset_bin_init(&eset->bins[i]);
eset_bin_stats_init(&eset->bin_stats[i]);
}
fb_init(eset->bitmap, ESET_NPSIZES);
edata_list_inactive_init(&eset->lru);
eset->state = state;
}
size_t
eset_npages_get(eset_t *eset) {
return atomic_load_zu(&eset->npages, ATOMIC_RELAXED);
}
size_t
eset_nextents_get(eset_t *eset, pszind_t pind) {
return atomic_load_zu(&eset->bin_stats[pind].nextents, ATOMIC_RELAXED);
}
size_t
eset_nbytes_get(eset_t *eset, pszind_t pind) {
return atomic_load_zu(&eset->bin_stats[pind].nbytes, ATOMIC_RELAXED);
}
static void
eset_stats_add(eset_t *eset, pszind_t pind, size_t sz) {
size_t cur = atomic_load_zu(&eset->bin_stats[pind].nextents,
ATOMIC_RELAXED);
atomic_store_zu(&eset->bin_stats[pind].nextents, cur + 1,
ATOMIC_RELAXED);
cur = atomic_load_zu(&eset->bin_stats[pind].nbytes, ATOMIC_RELAXED);
atomic_store_zu(&eset->bin_stats[pind].nbytes, cur + sz,
ATOMIC_RELAXED);
}
static void
eset_stats_sub(eset_t *eset, pszind_t pind, size_t sz) {
size_t cur = atomic_load_zu(&eset->bin_stats[pind].nextents,
ATOMIC_RELAXED);
atomic_store_zu(&eset->bin_stats[pind].nextents, cur - 1,
ATOMIC_RELAXED);
cur = atomic_load_zu(&eset->bin_stats[pind].nbytes, ATOMIC_RELAXED);
atomic_store_zu(&eset->bin_stats[pind].nbytes, cur - sz,
ATOMIC_RELAXED);
}
void
eset_insert(eset_t *eset, edata_t *edata) {
assert(edata_state_get(edata) == eset->state);
size_t size = edata_size_get(edata);
size_t psz = sz_psz_quantize_floor(size);
pszind_t pind = sz_psz2ind(psz);
edata_cmp_summary_t edata_cmp_summary = edata_cmp_summary_get(edata);
if (edata_heap_empty(&eset->bins[pind].heap)) {
fb_set(eset->bitmap, ESET_NPSIZES, (size_t)pind);
/* Only element is automatically the min element. */
eset->bins[pind].heap_min = edata_cmp_summary;
} else {
/*
* There's already a min element; update the summary if we're
* about to insert a lower one.
*/
if (edata_cmp_summary_comp(edata_cmp_summary,
eset->bins[pind].heap_min) < 0) {
eset->bins[pind].heap_min = edata_cmp_summary;
}
}
edata_heap_insert(&eset->bins[pind].heap, edata);
if (config_stats) {
eset_stats_add(eset, pind, size);
}
edata_list_inactive_append(&eset->lru, edata);
size_t npages = size >> LG_PAGE;
/*
* All modifications to npages hold the mutex (as asserted above), so we
* don't need an atomic fetch-add; we can get by with a load followed by
* a store.
*/
size_t cur_eset_npages =
atomic_load_zu(&eset->npages, ATOMIC_RELAXED);
atomic_store_zu(&eset->npages, cur_eset_npages + npages,
ATOMIC_RELAXED);
}
void
eset_remove(eset_t *eset, edata_t *edata) {
assert(edata_state_get(edata) == eset->state ||
edata_state_in_transition(edata_state_get(edata)));
size_t size = edata_size_get(edata);
size_t psz = sz_psz_quantize_floor(size);
pszind_t pind = sz_psz2ind(psz);
if (config_stats) {
eset_stats_sub(eset, pind, size);
}
edata_cmp_summary_t edata_cmp_summary = edata_cmp_summary_get(edata);
edata_heap_remove(&eset->bins[pind].heap, edata);
if (edata_heap_empty(&eset->bins[pind].heap)) {
fb_unset(eset->bitmap, ESET_NPSIZES, (size_t)pind);
} else {
/*
* This is a little weird; we compare if the summaries are
* equal, rather than if the edata we removed was the heap
* minimum. The reason why is that getting the heap minimum
* can cause a pairing heap merge operation. We can avoid this
* if we only update the min if it's changed, in which case the
* summaries of the removed element and the min element should
* compare equal.
*/
if (edata_cmp_summary_comp(edata_cmp_summary,
eset->bins[pind].heap_min) == 0) {
eset->bins[pind].heap_min = edata_cmp_summary_get(
edata_heap_first(&eset->bins[pind].heap));
}
}
edata_list_inactive_remove(&eset->lru, edata);
size_t npages = size >> LG_PAGE;
/*
* As in eset_insert, we hold eset->mtx and so don't need atomic
* operations for updating eset->npages.
*/
size_t cur_extents_npages =
atomic_load_zu(&eset->npages, ATOMIC_RELAXED);
assert(cur_extents_npages >= npages);
atomic_store_zu(&eset->npages,
cur_extents_npages - (size >> LG_PAGE), ATOMIC_RELAXED);
}
/*
* Find an extent with size [min_size, max_size) to satisfy the alignment
* requirement. For each size, try only the first extent in the heap.
*/
static edata_t *
eset_fit_alignment(eset_t *eset, size_t min_size, size_t max_size,
size_t alignment) {
pszind_t pind = sz_psz2ind(sz_psz_quantize_ceil(min_size));
pszind_t pind_max = sz_psz2ind(sz_psz_quantize_ceil(max_size));
for (pszind_t i =
(pszind_t)fb_ffs(eset->bitmap, ESET_NPSIZES, (size_t)pind);
i < pind_max;
i = (pszind_t)fb_ffs(eset->bitmap, ESET_NPSIZES, (size_t)i + 1)) {
assert(i < SC_NPSIZES);
assert(!edata_heap_empty(&eset->bins[i].heap));
edata_t *edata = edata_heap_first(&eset->bins[i].heap);
uintptr_t base = (uintptr_t)edata_base_get(edata);
size_t candidate_size = edata_size_get(edata);
assert(candidate_size >= min_size);
uintptr_t next_align = ALIGNMENT_CEILING((uintptr_t)base,
PAGE_CEILING(alignment));
if (base > next_align || base + candidate_size <= next_align) {
/* Overflow or not crossing the next alignment. */
continue;
}
size_t leadsize = next_align - base;
if (candidate_size - leadsize >= min_size) {
return edata;
}
}
return NULL;
}
/*
* Do first-fit extent selection, i.e. select the oldest/lowest extent that is
* large enough.
*
* lg_max_fit is the (log of the) maximum ratio between the requested size and
* the returned size that we'll allow. This can reduce fragmentation by
* avoiding reusing and splitting large extents for smaller sizes. In practice,
* it's set to opt_lg_extent_max_active_fit for the dirty eset and SC_PTR_BITS
* for others.
*/
static edata_t *
eset_first_fit(eset_t *eset, size_t size, bool exact_only,
unsigned lg_max_fit) {
edata_t *ret = NULL;
edata_cmp_summary_t ret_summ JEMALLOC_CC_SILENCE_INIT({0});
pszind_t pind = sz_psz2ind(sz_psz_quantize_ceil(size));
if (exact_only) {
return edata_heap_empty(&eset->bins[pind].heap) ? NULL :
edata_heap_first(&eset->bins[pind].heap);
}
for (pszind_t i =
(pszind_t)fb_ffs(eset->bitmap, ESET_NPSIZES, (size_t)pind);
i < ESET_NPSIZES;
i = (pszind_t)fb_ffs(eset->bitmap, ESET_NPSIZES, (size_t)i + 1)) {
assert(!edata_heap_empty(&eset->bins[i].heap));
if (lg_max_fit == SC_PTR_BITS) {
/*
* We'll shift by this below, and shifting out all the
* bits is undefined. Decreasing is safe, since the
* page size is larger than 1 byte.
*/
lg_max_fit = SC_PTR_BITS - 1;
}
if ((sz_pind2sz(i) >> lg_max_fit) > size) {
break;
}
if (ret == NULL || edata_cmp_summary_comp(
eset->bins[i].heap_min, ret_summ) < 0) {
/*
* We grab the edata as early as possible, even though
* we might change it later. Practically, a large
* portion of eset_fit calls succeed at the first valid
* index, so this doesn't cost much, and we get the
* effect of prefetching the edata as early as possible.
*/
edata_t *edata = edata_heap_first(&eset->bins[i].heap);
assert(edata_size_get(edata) >= size);
assert(ret == NULL || edata_snad_comp(edata, ret) < 0);
assert(ret == NULL || edata_cmp_summary_comp(
eset->bins[i].heap_min,
edata_cmp_summary_get(edata)) == 0);
ret = edata;
ret_summ = eset->bins[i].heap_min;
}
if (i == SC_NPSIZES) {
break;
}
assert(i < SC_NPSIZES);
}
return ret;
}
edata_t *
eset_fit(eset_t *eset, size_t esize, size_t alignment, bool exact_only,
unsigned lg_max_fit) {
size_t max_size = esize + PAGE_CEILING(alignment) - PAGE;
/* Beware size_t wrap-around. */
if (max_size < esize) {
return NULL;
}
edata_t *edata = eset_first_fit(eset, max_size, exact_only, lg_max_fit);
if (alignment > PAGE && edata == NULL) {
/*
* max_size guarantees the alignment requirement but is rather
* pessimistic. Next we try to satisfy the aligned allocation
* with sizes in [esize, max_size).
*/
edata = eset_fit_alignment(eset, esize, max_size, alignment);
}
return edata;
}