Dump heap profile backtraces in a stable order.

Also iterate over per thread stats in a stable order, which prepares the
way for stable ordering of per thread heap profile dumps.
This commit is contained in:
Jason Evans 2014-08-16 12:58:55 -07:00
parent ab532e9799
commit 3a81cbd2d4
2 changed files with 119 additions and 62 deletions

View File

@ -1,6 +1,7 @@
/******************************************************************************/ /******************************************************************************/
#ifdef JEMALLOC_H_TYPES #ifdef JEMALLOC_H_TYPES
typedef uint64_t prof_thr_uid_t;
typedef struct prof_bt_s prof_bt_t; typedef struct prof_bt_s prof_bt_t;
typedef struct prof_cnt_s prof_cnt_t; typedef struct prof_cnt_s prof_cnt_t;
typedef struct prof_thr_cnt_s prof_thr_cnt_t; typedef struct prof_thr_cnt_s prof_thr_cnt_t;
@ -81,15 +82,17 @@ struct prof_cnt_s {
}; };
struct prof_thr_cnt_s { struct prof_thr_cnt_s {
/* Linkage into prof_ctx_t's cnts_ql. */ prof_thr_uid_t thr_uid;
ql_elm(prof_thr_cnt_t) cnts_link;
/* Linkage into prof_ctx_t's thr_cnts. */
rb_node(prof_thr_cnt_t) thr_cnt_link;
/* /*
* Associated context. If a thread frees an object that it did not * Associated context. If a thread frees an object that it did not
* allocate, it is possible that the context is not cached in the * allocate, it is possible that the context is not present in the
* thread's hash table, in which case it must be able to look up the * thread's hash table, in which case it must be able to look up the
* context, insert a new prof_thr_cnt_t into the thread's hash table, * context, insert a new prof_thr_cnt_t into the thread's hash table,
* and link it into the prof_ctx_t's cnts_ql. * and link it into the prof_ctx_t's thr_cnts.
*/ */
prof_ctx_t *ctx; prof_ctx_t *ctx;
@ -113,9 +116,10 @@ struct prof_thr_cnt_s {
/* Profiling counters. */ /* Profiling counters. */
prof_cnt_t cnts; prof_cnt_t cnts;
}; };
typedef rb_tree(prof_thr_cnt_t) prof_thr_cnt_tree_t;
struct prof_ctx_s { struct prof_ctx_s {
/* Protects nlimbo, cnt_merged, and cnts_ql. */ /* Protects nlimbo, cnt_merged, and thr_cnts. */
malloc_mutex_t *lock; malloc_mutex_t *lock;
/* /*
@ -136,13 +140,13 @@ struct prof_ctx_s {
prof_cnt_t cnt_merged; prof_cnt_t cnt_merged;
/* /*
* List of profile counters, one for each thread that has allocated in * Tree of profile counters, one for each thread that has allocated in
* this context. * this context.
*/ */
ql_head(prof_thr_cnt_t) cnts_ql; prof_thr_cnt_tree_t thr_cnts;
/* Linkage for list of contexts to be dumped. */ /* Linkage for tree of contexts to be dumped. */
ql_elm(prof_ctx_t) dump_link; rb_node(prof_ctx_t) dump_link;
/* Associated backtrace. */ /* Associated backtrace. */
prof_bt_t bt; prof_bt_t bt;
@ -150,7 +154,7 @@ struct prof_ctx_s {
/* Backtrace vector, variable size, referred to by bt. */ /* Backtrace vector, variable size, referred to by bt. */
void *vec[1]; void *vec[1];
}; };
typedef ql_head(prof_ctx_t) prof_ctx_list_t; typedef rb_tree(prof_ctx_t) prof_ctx_tree_t;
struct prof_tdata_s { struct prof_tdata_s {
/* /*

View File

@ -77,6 +77,33 @@ static bool prof_booted = false;
/******************************************************************************/ /******************************************************************************/
JEMALLOC_INLINE_C int
prof_thr_cnt_comp(const prof_thr_cnt_t *a, const prof_thr_cnt_t *b)
{
prof_thr_uid_t a_uid = a->thr_uid;
prof_thr_uid_t b_uid = b->thr_uid;
return ((a_uid > b_uid) - (a_uid < b_uid));
}
rb_gen(static UNUSED, thr_cnt_tree_, prof_thr_cnt_tree_t, prof_thr_cnt_t,
thr_cnt_link, prof_thr_cnt_comp)
JEMALLOC_INLINE_C int
prof_ctx_comp(const prof_ctx_t *a, const prof_ctx_t *b)
{
unsigned a_len = a->bt.len;
unsigned b_len = b->bt.len;
unsigned comp_len = (a_len < b_len) ? a_len : b_len;
int ret = memcmp(a->bt.vec, b->bt.vec, comp_len * sizeof(void *));
if (ret == 0)
ret = (a_len > b_len) - (a_len < b_len);
return (ret);
}
rb_gen(static UNUSED, ctx_tree_, prof_ctx_tree_t, prof_ctx_t, dump_link,
prof_ctx_comp)
void void
bt_init(prof_bt_t *bt, void **vec) bt_init(prof_bt_t *bt, void **vec)
{ {
@ -369,9 +396,8 @@ prof_ctx_create(prof_bt_t *bt)
* prof_ctx_merge()/prof_ctx_destroy(). * prof_ctx_merge()/prof_ctx_destroy().
*/ */
ctx->nlimbo = 1; ctx->nlimbo = 1;
ql_elm_new(ctx, dump_link);
memset(&ctx->cnt_merged, 0, sizeof(prof_cnt_t)); memset(&ctx->cnt_merged, 0, sizeof(prof_cnt_t));
ql_new(&ctx->cnts_ql); thr_cnt_tree_new(&ctx->thr_cnts);
/* Duplicate bt. */ /* Duplicate bt. */
memcpy(ctx->vec, bt->vec, bt->len * sizeof(void *)); memcpy(ctx->vec, bt->vec, bt->len * sizeof(void *));
ctx->bt.vec = ctx->vec; ctx->bt.vec = ctx->vec;
@ -397,8 +423,8 @@ prof_ctx_destroy(prof_ctx_t *ctx)
assert((uintptr_t)prof_tdata > (uintptr_t)PROF_TDATA_STATE_MAX); assert((uintptr_t)prof_tdata > (uintptr_t)PROF_TDATA_STATE_MAX);
prof_enter(prof_tdata); prof_enter(prof_tdata);
malloc_mutex_lock(ctx->lock); malloc_mutex_lock(ctx->lock);
if (ql_first(&ctx->cnts_ql) == NULL && ctx->cnt_merged.curobjs == 0 && if (thr_cnt_tree_first(&ctx->thr_cnts) == NULL &&
ctx->nlimbo == 1) { ctx->cnt_merged.curobjs == 0 && ctx->nlimbo == 1) {
assert(ctx->cnt_merged.curbytes == 0); assert(ctx->cnt_merged.curbytes == 0);
assert(ctx->cnt_merged.accumobjs == 0); assert(ctx->cnt_merged.accumobjs == 0);
assert(ctx->cnt_merged.accumbytes == 0); assert(ctx->cnt_merged.accumbytes == 0);
@ -433,9 +459,9 @@ prof_ctx_merge(prof_ctx_t *ctx, prof_thr_cnt_t *cnt)
ctx->cnt_merged.curbytes += cnt->cnts.curbytes; ctx->cnt_merged.curbytes += cnt->cnts.curbytes;
ctx->cnt_merged.accumobjs += cnt->cnts.accumobjs; ctx->cnt_merged.accumobjs += cnt->cnts.accumobjs;
ctx->cnt_merged.accumbytes += cnt->cnts.accumbytes; ctx->cnt_merged.accumbytes += cnt->cnts.accumbytes;
ql_remove(&ctx->cnts_ql, cnt, cnts_link); thr_cnt_tree_remove(&ctx->thr_cnts, cnt);
if (opt_prof_accum == false && ql_first(&ctx->cnts_ql) == NULL && if (opt_prof_accum == false && thr_cnt_tree_first(&ctx->thr_cnts) ==
ctx->cnt_merged.curobjs == 0 && ctx->nlimbo == 0) { NULL && ctx->cnt_merged.curobjs == 0 && ctx->nlimbo == 0) {
/* /*
* Increment ctx->nlimbo in order to keep another thread from * Increment ctx->nlimbo in order to keep another thread from
* winning the race to destroy ctx while this one has ctx->lock * winning the race to destroy ctx while this one has ctx->lock
@ -540,7 +566,6 @@ prof_lookup(prof_bt_t *bt)
prof_ctx_destroy(ctx); prof_ctx_destroy(ctx);
return (NULL); return (NULL);
} }
ql_elm_new(ret.p, cnts_link);
ret.p->ctx = ctx; ret.p->ctx = ctx;
ret.p->epoch = 0; ret.p->epoch = 0;
memset(&ret.p->cnts, 0, sizeof(prof_cnt_t)); memset(&ret.p->cnts, 0, sizeof(prof_cnt_t));
@ -551,7 +576,7 @@ prof_lookup(prof_bt_t *bt)
return (NULL); return (NULL);
} }
malloc_mutex_lock(ctx->lock); malloc_mutex_lock(ctx->lock);
ql_tail_insert(&ctx->cnts_ql, ret.p, cnts_link); thr_cnt_tree_insert(&ctx->thr_cnts, ret.p);
ctx->nlimbo--; ctx->nlimbo--;
malloc_mutex_unlock(ctx->lock); malloc_mutex_unlock(ctx->lock);
} }
@ -745,28 +770,12 @@ prof_dump_printf(bool propagate_err, const char *format, ...)
return (ret); return (ret);
} }
static void static prof_thr_cnt_t *
prof_dump_ctx_prep(prof_ctx_t *ctx, prof_cnt_t *cnt_all, size_t *leak_nctx, ctx_sum_iter(prof_thr_cnt_tree_t *thr_cnts, prof_thr_cnt_t *thr_cnt, void *arg)
prof_ctx_list_t *ctx_ql)
{ {
prof_thr_cnt_t *thr_cnt; prof_ctx_t *ctx = (prof_ctx_t *)arg;
prof_cnt_t tcnt;
cassert(config_prof);
malloc_mutex_lock(ctx->lock);
/*
* Increment nlimbo so that ctx won't go away before dump.
* Additionally, link ctx into the dump list so that it is included in
* prof_dump()'s second pass.
*/
ctx->nlimbo++;
ql_tail_insert(ctx_ql, ctx, dump_link);
memcpy(&ctx->cnt_summed, &ctx->cnt_merged, sizeof(prof_cnt_t));
ql_foreach(thr_cnt, &ctx->cnts_ql, cnts_link) {
volatile unsigned *epoch = &thr_cnt->epoch; volatile unsigned *epoch = &thr_cnt->epoch;
prof_cnt_t tcnt;
while (true) { while (true) {
unsigned epoch0 = *epoch; unsigned epoch0 = *epoch;
@ -788,8 +797,30 @@ prof_dump_ctx_prep(prof_ctx_t *ctx, prof_cnt_t *cnt_all, size_t *leak_nctx,
ctx->cnt_summed.accumobjs += tcnt.accumobjs; ctx->cnt_summed.accumobjs += tcnt.accumobjs;
ctx->cnt_summed.accumbytes += tcnt.accumbytes; ctx->cnt_summed.accumbytes += tcnt.accumbytes;
} }
return (NULL);
} }
static void
prof_dump_ctx_prep(prof_ctx_t *ctx, prof_cnt_t *cnt_all, size_t *leak_nctx,
prof_ctx_tree_t *ctxs)
{
cassert(config_prof);
malloc_mutex_lock(ctx->lock);
/*
* Increment nlimbo so that ctx won't go away before dump.
* Additionally, link ctx into the dump list so that it is included in
* prof_dump()'s second pass.
*/
ctx->nlimbo++;
ctx_tree_insert(ctxs, ctx);
memcpy(&ctx->cnt_summed, &ctx->cnt_merged, sizeof(prof_cnt_t));
thr_cnt_tree_iter(&ctx->thr_cnts, NULL, ctx_sum_iter, (void *)ctx);
if (ctx->cnt_summed.curobjs != 0) if (ctx->cnt_summed.curobjs != 0)
(*leak_nctx)++; (*leak_nctx)++;
@ -829,25 +860,24 @@ prof_dump_header(bool propagate_err, const prof_cnt_t *cnt_all)
} }
static void static void
prof_dump_ctx_cleanup_locked(prof_ctx_t *ctx, prof_ctx_list_t *ctx_ql) prof_dump_ctx_cleanup_locked(prof_ctx_t *ctx, prof_ctx_tree_t *ctxs)
{ {
ctx->nlimbo--; ctx->nlimbo--;
ql_remove(ctx_ql, ctx, dump_link);
} }
static void static void
prof_dump_ctx_cleanup(prof_ctx_t *ctx, prof_ctx_list_t *ctx_ql) prof_dump_ctx_cleanup(prof_ctx_t *ctx, prof_ctx_tree_t *ctxs)
{ {
malloc_mutex_lock(ctx->lock); malloc_mutex_lock(ctx->lock);
prof_dump_ctx_cleanup_locked(ctx, ctx_ql); prof_dump_ctx_cleanup_locked(ctx, ctxs);
malloc_mutex_unlock(ctx->lock); malloc_mutex_unlock(ctx->lock);
} }
static bool static bool
prof_dump_ctx(bool propagate_err, prof_ctx_t *ctx, const prof_bt_t *bt, prof_dump_ctx(bool propagate_err, prof_ctx_t *ctx, const prof_bt_t *bt,
prof_ctx_list_t *ctx_ql) prof_ctx_tree_t *ctxs)
{ {
bool ret; bool ret;
unsigned i; unsigned i;
@ -895,7 +925,7 @@ prof_dump_ctx(bool propagate_err, prof_ctx_t *ctx, const prof_bt_t *bt,
ret = false; ret = false;
label_return: label_return:
prof_dump_ctx_cleanup_locked(ctx, ctx_ql); prof_dump_ctx_cleanup_locked(ctx, ctxs);
malloc_mutex_unlock(ctx->lock); malloc_mutex_unlock(ctx->lock);
return (ret); return (ret);
} }
@ -966,6 +996,26 @@ prof_leakcheck(const prof_cnt_t *cnt_all, size_t leak_nctx,
} }
} }
static prof_ctx_t *
prof_ctx_dump_iter(prof_ctx_tree_t *ctxs, prof_ctx_t *ctx, void *arg)
{
bool propagate_err = *(bool *)arg;
if (prof_dump_ctx(propagate_err, ctx, &ctx->bt, ctxs))
return (ctx_tree_next(ctxs, ctx));
return (NULL);
}
static prof_ctx_t *
prof_ctx_cleanup_iter(prof_ctx_tree_t *ctxs, prof_ctx_t *ctx, void *arg)
{
prof_dump_ctx_cleanup(ctx, ctxs);
return (NULL);
}
static bool static bool
prof_dump(bool propagate_err, const char *filename, bool leakcheck) prof_dump(bool propagate_err, const char *filename, bool leakcheck)
{ {
@ -977,7 +1027,8 @@ prof_dump(bool propagate_err, const char *filename, bool leakcheck)
void *v; void *v;
} ctx; } ctx;
size_t leak_nctx; size_t leak_nctx;
prof_ctx_list_t ctx_ql; prof_ctx_tree_t ctxs;
prof_ctx_t *cleanup_start = NULL;
cassert(config_prof); cassert(config_prof);
@ -990,10 +1041,10 @@ prof_dump(bool propagate_err, const char *filename, bool leakcheck)
/* Merge per thread profile stats, and sum them in cnt_all. */ /* Merge per thread profile stats, and sum them in cnt_all. */
memset(&cnt_all, 0, sizeof(prof_cnt_t)); memset(&cnt_all, 0, sizeof(prof_cnt_t));
leak_nctx = 0; leak_nctx = 0;
ql_new(&ctx_ql); ctx_tree_new(&ctxs);
prof_enter(prof_tdata); prof_enter(prof_tdata);
for (tabind = 0; ckh_iter(&bt2ctx, &tabind, NULL, &ctx.v) == false;) for (tabind = 0; ckh_iter(&bt2ctx, &tabind, NULL, &ctx.v) == false;)
prof_dump_ctx_prep(ctx.p, &cnt_all, &leak_nctx, &ctx_ql); prof_dump_ctx_prep(ctx.p, &cnt_all, &leak_nctx, &ctxs);
prof_leave(prof_tdata); prof_leave(prof_tdata);
/* Create dump file. */ /* Create dump file. */
@ -1005,10 +1056,10 @@ prof_dump(bool propagate_err, const char *filename, bool leakcheck)
goto label_write_error; goto label_write_error;
/* Dump per ctx profile stats. */ /* Dump per ctx profile stats. */
while ((ctx.p = ql_first(&ctx_ql)) != NULL) { cleanup_start = ctx_tree_iter(&ctxs, NULL, prof_ctx_dump_iter,
if (prof_dump_ctx(propagate_err, ctx.p, &ctx.p->bt, &ctx_ql)) (void *)&propagate_err);
if (cleanup_start != NULL)
goto label_write_error; goto label_write_error;
}
/* Dump /proc/<pid>/maps if possible. */ /* Dump /proc/<pid>/maps if possible. */
if (prof_dump_maps(propagate_err)) if (prof_dump_maps(propagate_err))
@ -1026,8 +1077,10 @@ prof_dump(bool propagate_err, const char *filename, bool leakcheck)
label_write_error: label_write_error:
prof_dump_close(propagate_err); prof_dump_close(propagate_err);
label_open_close_error: label_open_close_error:
while ((ctx.p = ql_first(&ctx_ql)) != NULL) if (cleanup_start != NULL) {
prof_dump_ctx_cleanup(ctx.p, &ctx_ql); ctx_tree_iter(&ctxs, cleanup_start, prof_ctx_cleanup_iter,
NULL);
}
malloc_mutex_unlock(&prof_dump_mtx); malloc_mutex_unlock(&prof_dump_mtx);
return (true); return (true);
} }