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

Optimize large deallocation.

Browse Source 
We eagerly coalesce large buffers when deallocating, however the previous logic
around this introduced extra lock overhead -- when coalescing we always lock the
neighbors even if they are active, while for active extents nothing can be done.

This commit checks if the neighbor extents are potentially active before
locking, and avoids locking if possible.  This speeds up large_dalloc by ~20%.
It also fixes some undesired behavior: we could stop coalescing because a small
buffer was merged, while a large neighbor was ignored on the other side.
This commit is contained in:
Qi Wang 2018-10-23 13:50:42 -07:00 committed by Qi Wang
parent 8dabf81df1
commit d66f976628
1 changed files with 41 additions and 17 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

58
src/extent.c
View File

@ -134,13 +134,16 @@ typedef enum {
static lock_result_t
extent_rtree_leaf_elm_try_lock(tsdn_t *tsdn, rtree_leaf_elm_t *elm,
extent_t **result) {
extent_t **result, bool inactive_only) {
extent_t *extent1 = rtree_leaf_elm_extent_read(tsdn, &extents_rtree,
elm, true);
if (extent1 == NULL) {
/* Slab implies active extents and should be skipped. */
if (extent1 == NULL || (inactive_only && rtree_leaf_elm_slab_read(tsdn,
&extents_rtree, elm, true))) {
return lock_result_no_extent;
}
/*
* It's possible that the extent changed out from under us, and with it
* the leaf->extent mapping. We have to recheck while holding the lock.
@ -163,7 +166,8 @@ extent_rtree_leaf_elm_try_lock(tsdn_t *tsdn, rtree_leaf_elm_t *elm,
* address, and NULL otherwise.
*/
static extent_t *
extent_lock_from_addr(tsdn_t *tsdn, rtree_ctx_t *rtree_ctx, void *addr) {
extent_lock_from_addr(tsdn_t *tsdn, rtree_ctx_t *rtree_ctx, void *addr,
bool inactive_only) {
extent_t *ret = NULL;
rtree_leaf_elm_t *elm = rtree_leaf_elm_lookup(tsdn, &extents_rtree,
rtree_ctx, (uintptr_t)addr, false, false);
@ -172,7 +176,8 @@ extent_lock_from_addr(tsdn_t *tsdn, rtree_ctx_t *rtree_ctx, void *addr) {
}
lock_result_t lock_result;
do {
lock_result = extent_rtree_leaf_elm_try_lock(tsdn, elm, &ret);
lock_result = extent_rtree_leaf_elm_try_lock(tsdn, elm, &ret,
inactive_only);
} while (lock_result == lock_result_failure);
return ret;
}
@ -917,7 +922,8 @@ extent_recycle_extract(tsdn_t *tsdn, arena_t *arena,
extent_hooks_assure_initialized(arena, r_extent_hooks);
extent_t *extent;
if (new_addr != NULL) {
extent = extent_lock_from_addr(tsdn, rtree_ctx, new_addr);
extent = extent_lock_from_addr(tsdn, rtree_ctx, new_addr,
false);
if (extent != NULL) {
/*
* We might null-out extent to report an error, but we
@ -1088,8 +1094,8 @@ extent_recycle_split(tsdn_t *tsdn, arena_t *arena,
extent_deregister_no_gdump_sub(tsdn, to_leak);
extents_leak(tsdn, arena, r_extent_hooks, extents,
to_leak, growing_retained);
assert(extent_lock_from_addr(tsdn, rtree_ctx, leak)
== NULL);
assert(extent_lock_from_addr(tsdn, rtree_ctx, leak,
false) == NULL);
}
return NULL;
}
@ -1567,9 +1573,15 @@ extent_coalesce(tsdn_t *tsdn, arena_t *arena, extent_hooks_t **r_extent_hooks,
}
static extent_t *
extent_try_coalesce(tsdn_t *tsdn, arena_t *arena,
extent_try_coalesce_impl(tsdn_t *tsdn, arena_t *arena,
extent_hooks_t **r_extent_hooks, rtree_ctx_t *rtree_ctx, extents_t *extents,
extent_t *extent, bool *coalesced, bool growing_retained) {
extent_t *extent, bool *coalesced, bool growing_retained,
bool inactive_only) {
/*
* We avoid checking / locking inactive neighbors for large size
* classes, since they are eagerly coalesced on deallocation which can
* cause lock contention.
*/
/*
* Continue attempting to coalesce until failure, to protect against
* races with other threads that are thwarted by this one.
@ -1580,7 +1592,7 @@ extent_try_coalesce(tsdn_t *tsdn, arena_t *arena,
/* Try to coalesce forward. */
extent_t *next = extent_lock_from_addr(tsdn, rtree_ctx,
extent_past_get(extent));
extent_past_get(extent), inactive_only);
if (next != NULL) {
/*
* extents->mtx only protects against races for
@ -1606,7 +1618,7 @@ extent_try_coalesce(tsdn_t *tsdn, arena_t *arena,
/* Try to coalesce backward. */
extent_t *prev = extent_lock_from_addr(tsdn, rtree_ctx,
extent_before_get(extent));
extent_before_get(extent), inactive_only);
if (prev != NULL) {
bool can_coalesce = extent_can_coalesce(arena, extents,
extent, prev);
@ -1632,6 +1644,22 @@ extent_try_coalesce(tsdn_t *tsdn, arena_t *arena,
return extent;
}
static extent_t *
extent_try_coalesce(tsdn_t *tsdn, arena_t *arena,
extent_hooks_t **r_extent_hooks, rtree_ctx_t *rtree_ctx, extents_t *extents,
extent_t *extent, bool *coalesced, bool growing_retained) {
return extent_try_coalesce_impl(tsdn, arena, r_extent_hooks, rtree_ctx,
extents, extent, coalesced, growing_retained, false);
}
static extent_t *
extent_try_coalesce_large(tsdn_t *tsdn, arena_t *arena,
extent_hooks_t **r_extent_hooks, rtree_ctx_t *rtree_ctx, extents_t *extents,
extent_t *extent, bool *coalesced, bool growing_retained) {
return extent_try_coalesce_impl(tsdn, arena, r_extent_hooks, rtree_ctx,
extents, extent, coalesced, growing_retained, true);
}
/*
* Does the metadata management portions of putting an unused extent into the
* given extents_t (coalesces, deregisters slab interiors, the heap operations).
@ -1664,16 +1692,12 @@ extent_record(tsdn_t *tsdn, arena_t *arena, extent_hooks_t **r_extent_hooks,
} else if (extent_size_get(extent) >= SC_LARGE_MINCLASS) {
/* Always coalesce large extents eagerly. */
bool coalesced;
size_t prev_size;
do {
prev_size = extent_size_get(extent);
assert(extent_state_get(extent) == extent_state_active);
extent = extent_try_coalesce(tsdn, arena,
extent = extent_try_coalesce_large(tsdn, arena,
r_extent_hooks, rtree_ctx, extents, extent,
&coalesced, growing_retained);
} while (coalesced &&
extent_size_get(extent)
>= prev_size + SC_LARGE_MINCLASS);
} while (coalesced);
}
extent_deactivate_locked(tsdn, arena, extents, extent);