Eliminate false sharing in malloc due to statistic collection

Currently stats are collected in a MAXCPU-sized array which is not
aligned and suffers enormous false-sharing. Fix the problem by
utilizing per-cpu allocation.

The counter(9) API is not used here as it is too incomplete and does
not provide a win over per-cpu zone sized for malloc stats struct. In
particular stats are being reported for each cpu separately by just
copying what is supposed to be an array element for given cpu.

This eliminates significant false-sharing during malloc-heavy tests
e.g. on Skylake. See the review for details.

Reviewed by:	markj
Approved by:	re (kib)
Differential Revision:	https://reviews.freebsd.org/D17289
This commit is contained in:
Mateusz Guzik 2018-09-23 19:00:06 +00:00
parent 078a49a077
commit 9afff6b1c0
Notes: svn2git 2020-12-20 02:59:44 +00:00
svn path=/head/; revision=338899
2 changed files with 28 additions and 9 deletions

View File

@ -61,6 +61,7 @@ __FBSDID("$FreeBSD$");
#include <sys/vmmeter.h>
#include <sys/proc.h>
#include <sys/sbuf.h>
#include <sys/smp.h>
#include <sys/sysctl.h>
#include <sys/time.h>
#include <sys/vmem.h>
@ -173,6 +174,7 @@ struct {
* declare malloc types.
*/
static uma_zone_t mt_zone;
static uma_zone_t mt_stats_zone;
u_long vm_kmem_size;
SYSCTL_ULONG(_vm, OID_AUTO, kmem_size, CTLFLAG_RDTUN, &vm_kmem_size, 0,
@ -368,7 +370,7 @@ malloc_type_zone_allocated(struct malloc_type *mtp, unsigned long size,
critical_enter();
mtip = mtp->ks_handle;
mtsp = &mtip->mti_stats[curcpu];
mtsp = zpcpu_get(mtip->mti_stats);
if (size > 0) {
mtsp->mts_memalloced += size;
mtsp->mts_numallocs++;
@ -411,7 +413,7 @@ malloc_type_freed(struct malloc_type *mtp, unsigned long size)
critical_enter();
mtip = mtp->ks_handle;
mtsp = &mtip->mti_stats[curcpu];
mtsp = zpcpu_get(mtip->mti_stats);
mtsp->mts_memfreed += size;
mtsp->mts_numfrees++;
@ -953,6 +955,9 @@ mallocinit(void *dummy)
if (kmem_zmax < PAGE_SIZE || kmem_zmax > KMEM_ZMAX)
kmem_zmax = KMEM_ZMAX;
mt_stats_zone = uma_zcreate("mt_stats_zone",
sizeof(struct malloc_type_stats), NULL, NULL, NULL, NULL,
UMA_ALIGN_PTR, UMA_ZONE_PCPU);
mt_zone = uma_zcreate("mt_zone", sizeof(struct malloc_type_internal),
#ifdef INVARIANTS
mtrash_ctor, mtrash_dtor, mtrash_init, mtrash_fini,
@ -995,6 +1000,7 @@ malloc_init(void *data)
panic("malloc_init: bad malloc type magic");
mtip = uma_zalloc(mt_zone, M_WAITOK | M_ZERO);
mtip->mti_stats = uma_zalloc_pcpu(mt_stats_zone, M_WAITOK | M_ZERO);
mtp->ks_handle = mtip;
mtp_set_subzone(mtp);
@ -1042,8 +1048,8 @@ malloc_uninit(void *data)
* Look for memory leaks.
*/
temp_allocs = temp_bytes = 0;
for (i = 0; i < MAXCPU; i++) {
mtsp = &mtip->mti_stats[i];
for (i = 0; i <= mp_maxid; i++) {
mtsp = zpcpu_get_cpu(mtip->mti_stats, i);
temp_allocs += mtsp->mts_numallocs;
temp_allocs -= mtsp->mts_numfrees;
temp_bytes += mtsp->mts_memalloced;
@ -1056,6 +1062,7 @@ malloc_uninit(void *data)
}
slab = vtoslab((vm_offset_t) mtip & (~UMA_SLAB_MASK));
uma_zfree_pcpu(mt_stats_zone, mtip->mti_stats);
uma_zfree_arg(mt_zone, mtip, slab);
}
@ -1077,6 +1084,7 @@ sysctl_kern_malloc_stats(SYSCTL_HANDLER_ARGS)
{
struct malloc_type_stream_header mtsh;
struct malloc_type_internal *mtip;
struct malloc_type_stats *mtsp, zeromts;
struct malloc_type_header mth;
struct malloc_type *mtp;
int error, i;
@ -1089,6 +1097,8 @@ sysctl_kern_malloc_stats(SYSCTL_HANDLER_ARGS)
sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL);
mtx_lock(&malloc_mtx);
bzero(&zeromts, sizeof(zeromts));
/*
* Insert stream header.
*/
@ -1114,10 +1124,17 @@ sysctl_kern_malloc_stats(SYSCTL_HANDLER_ARGS)
/*
* Insert type statistics for each CPU.
*/
for (i = 0; i < MAXCPU; i++) {
(void)sbuf_bcat(&sbuf, &mtip->mti_stats[i],
sizeof(mtip->mti_stats[i]));
for (i = 0; i <= mp_maxid; i++) {
mtsp = zpcpu_get_cpu(mtip->mti_stats, i);
(void)sbuf_bcat(&sbuf, mtsp, sizeof(*mtsp));
}
/*
* Fill in the missing CPUs.
*/
for (; i < MAXCPU; i++) {
(void)sbuf_bcat(&sbuf, &zeromts, sizeof(zeromts));
}
}
mtx_unlock(&malloc_mtx);
error = sbuf_finish(&sbuf);
@ -1170,6 +1187,7 @@ malloc_type_list(malloc_type_list_func_t *func, void *arg)
DB_SHOW_COMMAND(malloc, db_show_malloc)
{
struct malloc_type_internal *mtip;
struct malloc_type_internal *mtsp;
struct malloc_type *mtp;
uint64_t allocs, frees;
uint64_t alloced, freed;
@ -1183,7 +1201,8 @@ DB_SHOW_COMMAND(malloc, db_show_malloc)
frees = 0;
alloced = 0;
freed = 0;
for (i = 0; i < MAXCPU; i++) {
for (i = 0; i <= mp_maxid; i++) {
mtsp = zpcpu_get_cpu(mtip->mti_stats, i);
allocs += mtip->mti_stats[i].mts_numallocs;
frees += mtip->mti_stats[i].mts_numfrees;
alloced += mtip->mti_stats[i].mts_memalloced;

View File

@ -101,7 +101,7 @@ struct malloc_type_internal {
uint32_t mti_probes[DTMALLOC_PROBE_MAX];
/* DTrace probe ID array. */
u_char mti_zone;
struct malloc_type_stats mti_stats[MAXCPU];
struct malloc_type_stats *mti_stats;
};
/*