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

SEC: Implement thread affinity.

Browse Source 
For now, just have every thread pick a shard once and stick with it.
This commit is contained in:
David Goldblatt 2020-10-22 14:13:09 -07:00 committed by David Goldblatt
parent d16849c91d
commit ea32060f9c
2 changed files with 16 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
include/jemalloc/internal/tsd.h
View File

@ -77,6 +77,7 @@ typedef ql_elm(tsd_t) tsd_link_t;
O(iarena, arena_t *, arena_t *) \ O(iarena, arena_t *, arena_t *) \
O(arena, arena_t *, arena_t *) \ O(arena, arena_t *, arena_t *) \
O(arenas_tdata, arena_tdata_t *, arena_tdata_t *)\ O(arenas_tdata, arena_tdata_t *, arena_tdata_t *)\
O(sec_shard, uint8_t, uint8_t) \
O(binshards, tsd_binshards_t, tsd_binshards_t)\ O(binshards, tsd_binshards_t, tsd_binshards_t)\
O(tsd_link, tsd_link_t, tsd_link_t) \ O(tsd_link, tsd_link_t, tsd_link_t) \
O(in_hook, bool, bool) \ O(in_hook, bool, bool) \
@ -106,6 +107,7 @@ typedef ql_elm(tsd_t) tsd_link_t;
/* iarena */ NULL, \ /* iarena */ NULL, \
/* arena */ NULL, \ /* arena */ NULL, \
/* arenas_tdata */ NULL, \ /* arenas_tdata */ NULL, \
/* sec_shard */ (uint8_t)-1, \
/* binshards */ TSD_BINSHARDS_ZERO_INITIALIZER, \ /* binshards */ TSD_BINSHARDS_ZERO_INITIALIZER, \
/* tsd_link */ {NULL}, \ /* tsd_link */ {NULL}, \
/* in_hook */ false, \ /* in_hook */ false, \

22
src/sec.c
View File

@ -61,14 +61,20 @@ sec_shard_pick(tsdn_t *tsdn, sec_t *sec) {
return &sec->shards[0]; return &sec->shards[0];
} }
tsd_t *tsd = tsdn_tsd(tsdn); tsd_t *tsd = tsdn_tsd(tsdn);
/* uint8_t *idxp = tsd_sec_shardp_get(tsd);
* Use the trick from Daniel Lemire's "A fast alternative to the modulo if (*idxp == (uint8_t)-1) {
* reduction. Use a 64 bit number to store 32 bits, since we'll /*
* deliberately overflow when we multiply by the number of shards. * First use; initialize using the trick from Daniel Lemire's
*/ * "A fast alternative to the modulo reduction. Use a 64 bit
uint64_t rand32 = prng_lg_range_u64(tsd_prng_statep_get(tsd), 32); * number to store 32 bits, since we'll deliberately overflow
uint32_t idx = (uint32_t)((rand32 * (uint64_t)sec->nshards) >> 32); * when we multiply by the number of shards.
return &sec->shards[idx]; */
uint64_t rand32 = prng_lg_range_u64(tsd_prng_statep_get(tsd), 32);
uint32_t idx = (uint32_t)((rand32 * (uint64_t)sec->nshards) >> 32);
assert(idx < (uint32_t)sec->nshards);
*idxp = (uint8_t)idx;
}
return &sec->shards[*idxp];
} }
static edata_t * static edata_t *