Merge branch 'dev'5.0.0

1 files changed, 313 insertions, 223 deletions

diff --git a/src/tcache.c b/src/tcache.c
index e3b04be..6355805 100644
--- a/src/tcache.c
+++ b/src/tcache.c
@@ -1,5 +1,10 @@
-#define	JEMALLOC_TCACHE_C_
-#include "jemalloc/internal/jemalloc_internal.h"
+#define JEMALLOC_TCACHE_C_
+#include "jemalloc/internal/jemalloc_preamble.h"
+#include "jemalloc/internal/jemalloc_internal_includes.h"
+
+#include "jemalloc/internal/assert.h"
+#include "jemalloc/internal/mutex.h"
+#include "jemalloc/internal/size_classes.h"
 
 /******************************************************************************/
 /* Data. */
@@ -27,19 +32,20 @@ static malloc_mutex_t	tcaches_mtx;
 /******************************************************************************/
 
 size_t
-tcache_salloc(tsdn_t *tsdn, const void *ptr)
-{
-
-	return (arena_salloc(tsdn, ptr, false));
+tcache_salloc(tsdn_t *tsdn, const void *ptr) {
+	return arena_salloc(tsdn, ptr);
 }
 
 void
-tcache_event_hard(tsd_t *tsd, tcache_t *tcache)
-{
+tcache_event_hard(tsd_t *tsd, tcache_t *tcache) {
 	szind_t binind = tcache->next_gc_bin;
-	tcache_bin_t *tbin = &tcache->tbins[binind];
-	tcache_bin_info_t *tbin_info = &tcache_bin_info[binind];
 
+	tcache_bin_t *tbin;
+	if (binind < NBINS) {
+		tbin = tcache_small_bin_get(tcache, binind);
+	} else {
+		tbin = tcache_large_bin_get(tcache, binind);
+	}
 	if (tbin->low_water > 0) {
 		/*
 		 * Flush (ceiling) 3/4 of the objects below the low water mark.
@@ -48,71 +54,80 @@ tcache_event_hard(tsd_t *tsd, tcache_t *tcache)
 			tcache_bin_flush_small(tsd, tcache, tbin, binind,
 			    tbin->ncached - tbin->low_water + (tbin->low_water
 			    >> 2));
+			/*
+			 * Reduce fill count by 2X.  Limit lg_fill_div such that
+			 * the fill count is always at least 1.
+			 */
+			tcache_bin_info_t *tbin_info = &tcache_bin_info[binind];
+			if ((tbin_info->ncached_max >>
+			     (tcache->lg_fill_div[binind] + 1)) >= 1) {
+				tcache->lg_fill_div[binind]++;
+			}
 		} else {
 			tcache_bin_flush_large(tsd, tbin, binind, tbin->ncached
 			    - tbin->low_water + (tbin->low_water >> 2), tcache);
 		}
-		/*
-		 * Reduce fill count by 2X.  Limit lg_fill_div such that the
-		 * fill count is always at least 1.
-		 */
-		if ((tbin_info->ncached_max >> (tbin->lg_fill_div+1)) >= 1)
-			tbin->lg_fill_div++;
 	} else if (tbin->low_water < 0) {
 		/*
-		 * Increase fill count by 2X.  Make sure lg_fill_div stays
-		 * greater than 0.
+		 * Increase fill count by 2X for small bins.  Make sure
+		 * lg_fill_div stays greater than 0.
 		 */
-		if (tbin->lg_fill_div > 1)
-			tbin->lg_fill_div--;
+		if (binind < NBINS && tcache->lg_fill_div[binind] > 1) {
+			tcache->lg_fill_div[binind]--;
+		}
 	}
 	tbin->low_water = tbin->ncached;
 
 	tcache->next_gc_bin++;
-	if (tcache->next_gc_bin == nhbins)
+	if (tcache->next_gc_bin == nhbins) {
 		tcache->next_gc_bin = 0;
+	}
 }
 
 void *
 tcache_alloc_small_hard(tsdn_t *tsdn, arena_t *arena, tcache_t *tcache,
-    tcache_bin_t *tbin, szind_t binind, bool *tcache_success)
-{
+    tcache_bin_t *tbin, szind_t binind, bool *tcache_success) {
 	void *ret;
 
-	arena_tcache_fill_small(tsdn, arena, tbin, binind, config_prof ?
-	    tcache->prof_accumbytes : 0);
-	if (config_prof)
+	assert(tcache->arena != NULL);
+	arena_tcache_fill_small(tsdn, arena, tcache, tbin, binind,
+	    config_prof ? tcache->prof_accumbytes : 0);
+	if (config_prof) {
 		tcache->prof_accumbytes = 0;
+	}
 	ret = tcache_alloc_easy(tbin, tcache_success);
 
-	return (ret);
+	return ret;
 }
 
 void
 tcache_bin_flush_small(tsd_t *tsd, tcache_t *tcache, tcache_bin_t *tbin,
-    szind_t binind, unsigned rem)
-{
-	arena_t *arena;
-	void *ptr;
-	unsigned i, nflush, ndeferred;
+    szind_t binind, unsigned rem) {
 	bool merged_stats = false;
 
 	assert(binind < NBINS);
 	assert(rem <= tbin->ncached);
 
-	arena = arena_choose(tsd, NULL);
+	arena_t *arena = tcache->arena;
 	assert(arena != NULL);
-	for (nflush = tbin->ncached - rem; nflush > 0; nflush = ndeferred) {
+	unsigned nflush = tbin->ncached - rem;
+	VARIABLE_ARRAY(extent_t *, item_extent, nflush);
+	/* Look up extent once per item. */
+	for (unsigned i = 0 ; i < nflush; i++) {
+		item_extent[i] = iealloc(tsd_tsdn(tsd), *(tbin->avail - 1 - i));
+	}
+
+	while (nflush > 0) {
 		/* Lock the arena bin associated with the first object. */
-		arena_chunk_t *chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(
-		    *(tbin->avail - 1));
-		arena_t *bin_arena = extent_node_arena_get(&chunk->node);
+		extent_t *extent = item_extent[0];
+		arena_t *bin_arena = extent_arena_get(extent);
 		arena_bin_t *bin = &bin_arena->bins[binind];
 
 		if (config_prof && bin_arena == arena) {
 			if (arena_prof_accum(tsd_tsdn(tsd), arena,
-			    tcache->prof_accumbytes))
+			    tcache->prof_accumbytes)) {
 				prof_idump(tsd_tsdn(tsd));
+			}
 			tcache->prof_accumbytes = 0;
 		}
 
@@ -124,18 +139,15 @@ tcache_bin_flush_small(tsd_t *tsd, tcache_t *tcache, tcache_bin_t *tbin,
 			bin->stats.nrequests += tbin->tstats.nrequests;
 			tbin->tstats.nrequests = 0;
 		}
-		ndeferred = 0;
-		for (i = 0; i < nflush; i++) {
-			ptr = *(tbin->avail - 1 - i);
-			assert(ptr != NULL);
-			chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr);
-			if (extent_node_arena_get(&chunk->node) == bin_arena) {
-				size_t pageind = ((uintptr_t)ptr -
-				    (uintptr_t)chunk) >> LG_PAGE;
-				arena_chunk_map_bits_t *bitselm =
-				    arena_bitselm_get_mutable(chunk, pageind);
+		unsigned ndeferred = 0;
+		for (unsigned i = 0; i < nflush; i++) {
+			void *ptr = *(tbin->avail - 1 - i);
+			extent = item_extent[i];
+			assert(ptr != NULL && extent != NULL);
+
+			if (extent_arena_get(extent) == bin_arena) {
 				arena_dalloc_bin_junked_locked(tsd_tsdn(tsd),
-				    bin_arena, chunk, ptr, bitselm);
+				    bin_arena, extent, ptr);
 			} else {
 				/*
 				 * This object was allocated via a different
@@ -144,11 +156,13 @@ tcache_bin_flush_small(tsd_t *tsd, tcache_t *tcache, tcache_bin_t *tbin,
 				 * handled in a future pass.
 				 */
 				*(tbin->avail - 1 - ndeferred) = ptr;
+				item_extent[ndeferred] = extent;
 				ndeferred++;
 			}
 		}
 		malloc_mutex_unlock(tsd_tsdn(tsd), &bin->lock);
 		arena_decay_ticks(tsd_tsdn(tsd), bin_arena, nflush - ndeferred);
+		nflush = ndeferred;
 	}
 	if (config_stats && !merged_stats) {
 		/*
@@ -166,58 +180,72 @@ tcache_bin_flush_small(tsd_t *tsd, tcache_t *tcache, tcache_bin_t *tbin,
 	memmove(tbin->avail - rem, tbin->avail - tbin->ncached, rem *
 	    sizeof(void *));
 	tbin->ncached = rem;
-	if ((int)tbin->ncached < tbin->low_water)
+	if ((low_water_t)tbin->ncached < tbin->low_water) {
 		tbin->low_water = tbin->ncached;
+	}
 }
 
 void
 tcache_bin_flush_large(tsd_t *tsd, tcache_bin_t *tbin, szind_t binind,
-    unsigned rem, tcache_t *tcache)
-{
-	arena_t *arena;
-	void *ptr;
-	unsigned i, nflush, ndeferred;
+    unsigned rem, tcache_t *tcache) {
 	bool merged_stats = false;
 
 	assert(binind < nhbins);
 	assert(rem <= tbin->ncached);
 
-	arena = arena_choose(tsd, NULL);
+	arena_t *arena = tcache->arena;
 	assert(arena != NULL);
-	for (nflush = tbin->ncached - rem; nflush > 0; nflush = ndeferred) {
+	unsigned nflush = tbin->ncached - rem;
+	VARIABLE_ARRAY(extent_t *, item_extent, nflush);
+	/* Look up extent once per item. */
+	for (unsigned i = 0 ; i < nflush; i++) {
+		item_extent[i] = iealloc(tsd_tsdn(tsd), *(tbin->avail - 1 - i));
+	}
+
+	while (nflush > 0) {
 		/* Lock the arena associated with the first object. */
-		arena_chunk_t *chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(
-		    *(tbin->avail - 1));
-		arena_t *locked_arena = extent_node_arena_get(&chunk->node);
+		extent_t *extent = item_extent[0];
+		arena_t *locked_arena = extent_arena_get(extent);
 		UNUSED bool idump;
 
-		if (config_prof)
+		if (config_prof) {
 			idump = false;
-		malloc_mutex_lock(tsd_tsdn(tsd), &locked_arena->lock);
+		}
+
+		malloc_mutex_lock(tsd_tsdn(tsd), &locked_arena->large_mtx);
+		for (unsigned i = 0; i < nflush; i++) {
+			void *ptr = *(tbin->avail - 1 - i);
+			assert(ptr != NULL);
+			extent = item_extent[i];
+			if (extent_arena_get(extent) == locked_arena) {
+				large_dalloc_prep_junked_locked(tsd_tsdn(tsd),
+				    extent);
+			}
+		}
 		if ((config_prof || config_stats) && locked_arena == arena) {
 			if (config_prof) {
-				idump = arena_prof_accum_locked(arena,
+				idump = arena_prof_accum(tsd_tsdn(tsd), arena,
 				    tcache->prof_accumbytes);
 				tcache->prof_accumbytes = 0;
 			}
 			if (config_stats) {
 				merged_stats = true;
-				arena->stats.nrequests_large +=
-				    tbin->tstats.nrequests;
-				arena->stats.lstats[binind - NBINS].nrequests +=
-				    tbin->tstats.nrequests;
+				arena_stats_large_nrequests_add(tsd_tsdn(tsd),
+				    &arena->stats, binind,
+				    tbin->tstats.nrequests);
 				tbin->tstats.nrequests = 0;
 			}
 		}
-		ndeferred = 0;
-		for (i = 0; i < nflush; i++) {
-			ptr = *(tbin->avail - 1 - i);
-			assert(ptr != NULL);
-			chunk = (arena_chunk_t *)CHUNK_ADDR2BASE(ptr);
-			if (extent_node_arena_get(&chunk->node) ==
-			    locked_arena) {
-				arena_dalloc_large_junked_locked(tsd_tsdn(tsd),
-				    locked_arena, chunk, ptr);
+		malloc_mutex_unlock(tsd_tsdn(tsd), &locked_arena->large_mtx);
+
+		unsigned ndeferred = 0;
+		for (unsigned i = 0; i < nflush; i++) {
+			void *ptr = *(tbin->avail - 1 - i);
+			extent = item_extent[i];
+			assert(ptr != NULL && extent != NULL);
+
+			if (extent_arena_get(extent) == locked_arena) {
+				large_dalloc_finish(tsd_tsdn(tsd), extent);
 			} else {
 				/*
 				 * This object was allocated via a different
@@ -226,55 +254,56 @@ tcache_bin_flush_large(tsd_t *tsd, tcache_bin_t *tbin, szind_t binind,
 				 * in a future pass.
 				 */
 				*(tbin->avail - 1 - ndeferred) = ptr;
+				item_extent[ndeferred] = extent;
 				ndeferred++;
 			}
 		}
-		malloc_mutex_unlock(tsd_tsdn(tsd), &locked_arena->lock);
-		if (config_prof && idump)
+		if (config_prof && idump) {
 			prof_idump(tsd_tsdn(tsd));
+		}
 		arena_decay_ticks(tsd_tsdn(tsd), locked_arena, nflush -
 		    ndeferred);
+		nflush = ndeferred;
 	}
 	if (config_stats && !merged_stats) {
 		/*
 		 * The flush loop didn't happen to flush to this thread's
 		 * arena, so the stats didn't get merged.  Manually do so now.
 		 */
-		malloc_mutex_lock(tsd_tsdn(tsd), &arena->lock);
-		arena->stats.nrequests_large += tbin->tstats.nrequests;
-		arena->stats.lstats[binind - NBINS].nrequests +=
-		    tbin->tstats.nrequests;
+		arena_stats_large_nrequests_add(tsd_tsdn(tsd), &arena->stats,
+		    binind, tbin->tstats.nrequests);
 		tbin->tstats.nrequests = 0;
-		malloc_mutex_unlock(tsd_tsdn(tsd), &arena->lock);
 	}
 
 	memmove(tbin->avail - rem, tbin->avail - tbin->ncached, rem *
 	    sizeof(void *));
 	tbin->ncached = rem;
-	if ((int)tbin->ncached < tbin->low_water)
+	if ((low_water_t)tbin->ncached < tbin->low_water) {
 		tbin->low_water = tbin->ncached;
+	}
 }
 
-static void
-tcache_arena_associate(tsdn_t *tsdn, tcache_t *tcache, arena_t *arena)
-{
+void
+tcache_arena_associate(tsdn_t *tsdn, tcache_t *tcache, arena_t *arena) {
+	assert(tcache->arena == NULL);
+	tcache->arena = arena;
 
 	if (config_stats) {
 		/* Link into list of extant tcaches. */
-		malloc_mutex_lock(tsdn, &arena->lock);
+		malloc_mutex_lock(tsdn, &arena->tcache_ql_mtx);
 		ql_elm_new(tcache, link);
 		ql_tail_insert(&arena->tcache_ql, tcache, link);
-		malloc_mutex_unlock(tsdn, &arena->lock);
+		malloc_mutex_unlock(tsdn, &arena->tcache_ql_mtx);
 	}
 }
 
 static void
-tcache_arena_dissociate(tsdn_t *tsdn, tcache_t *tcache, arena_t *arena)
-{
-
+tcache_arena_dissociate(tsdn_t *tsdn, tcache_t *tcache) {
+	arena_t *arena = tcache->arena;
+	assert(arena != NULL);
 	if (config_stats) {
 		/* Unlink from list of extant tcaches. */
-		malloc_mutex_lock(tsdn, &arena->lock);
+		malloc_mutex_lock(tsdn, &arena->tcache_ql_mtx);
 		if (config_debug) {
 			bool in_ql = false;
 			tcache_t *iter;
@@ -288,150 +317,216 @@ tcache_arena_dissociate(tsdn_t *tsdn, tcache_t *tcache, arena_t *arena)
 		}
 		ql_remove(&arena->tcache_ql, tcache, link);
 		tcache_stats_merge(tsdn, tcache, arena);
-		malloc_mutex_unlock(tsdn, &arena->lock);
+		malloc_mutex_unlock(tsdn, &arena->tcache_ql_mtx);
 	}
+	tcache->arena = NULL;
 }
 
 void
-tcache_arena_reassociate(tsdn_t *tsdn, tcache_t *tcache, arena_t *oldarena,
-    arena_t *newarena)
-{
+tcache_arena_reassociate(tsdn_t *tsdn, tcache_t *tcache, arena_t *arena) {
+	tcache_arena_dissociate(tsdn, tcache);
+	tcache_arena_associate(tsdn, tcache, arena);
+}
+
+bool
+tsd_tcache_enabled_data_init(tsd_t *tsd) {
+	/* Called upon tsd initialization. */
+	tsd_tcache_enabled_set(tsd, opt_tcache);
+	tsd_slow_update(tsd);
+
+	if (opt_tcache) {
+		/* Trigger tcache init. */
+		tsd_tcache_data_init(tsd);
+	}
 
-	tcache_arena_dissociate(tsdn, tcache, oldarena);
-	tcache_arena_associate(tsdn, tcache, newarena);
+	return false;
 }
 
-tcache_t *
-tcache_get_hard(tsd_t *tsd)
-{
-	arena_t *arena;
+/* Initialize auto tcache (embedded in TSD). */
+static void
+tcache_init(tsd_t *tsd, tcache_t *tcache, void *avail_stack) {
+	memset(&tcache->link, 0, sizeof(ql_elm(tcache_t)));
+	tcache->prof_accumbytes = 0;
+	tcache->next_gc_bin = 0;
+	tcache->arena = NULL;
+
+	ticker_init(&tcache->gc_ticker, TCACHE_GC_INCR);
 
-	if (!tcache_enabled_get()) {
-		if (tsd_nominal(tsd))
-			tcache_enabled_set(false); /* Memoize. */
-		return (NULL);
+	size_t stack_offset = 0;
+	assert((TCACHE_NSLOTS_SMALL_MAX & 1U) == 0);
+	memset(tcache->tbins_small, 0, sizeof(tcache_bin_t) * NBINS);
+	memset(tcache->tbins_large, 0, sizeof(tcache_bin_t) * (nhbins - NBINS));
+	unsigned i = 0;
+	for (; i < NBINS; i++) {
+		tcache->lg_fill_div[i] = 1;
+		stack_offset += tcache_bin_info[i].ncached_max * sizeof(void *);
+		/*
+		 * avail points past the available space.  Allocations will
+		 * access the slots toward higher addresses (for the benefit of
+		 * prefetch).
+		 */
+		tcache_small_bin_get(tcache, i)->avail =
+		    (void **)((uintptr_t)avail_stack + (uintptr_t)stack_offset);
+	}
+	for (; i < nhbins; i++) {
+		stack_offset += tcache_bin_info[i].ncached_max * sizeof(void *);
+		tcache_large_bin_get(tcache, i)->avail =
+		    (void **)((uintptr_t)avail_stack + (uintptr_t)stack_offset);
 	}
-	arena = arena_choose(tsd, NULL);
-	if (unlikely(arena == NULL))
-		return (NULL);
-	return (tcache_create(tsd_tsdn(tsd), arena));
+	assert(stack_offset == stack_nelms * sizeof(void *));
 }
 
+/* Initialize auto tcache (embedded in TSD). */
+bool
+tsd_tcache_data_init(tsd_t *tsd) {
+	tcache_t *tcache = tsd_tcachep_get_unsafe(tsd);
+	assert(tcache_small_bin_get(tcache, 0)->avail == NULL);
+	size_t size = stack_nelms * sizeof(void *);
+	/* Avoid false cacheline sharing. */
+	size = sz_sa2u(size, CACHELINE);
+
+	void *avail_array = ipallocztm(tsd_tsdn(tsd), size, CACHELINE, true,
+	    NULL, true, arena_get(TSDN_NULL, 0, true));
+	if (avail_array == NULL) {
+		return true;
+	}
+
+	tcache_init(tsd, tcache, avail_array);
+	/*
+	 * Initialization is a bit tricky here.  After malloc init is done, all
+	 * threads can rely on arena_choose and associate tcache accordingly.
+	 * However, the thread that does actual malloc bootstrapping relies on
+	 * functional tsd, and it can only rely on a0.  In that case, we
+	 * associate its tcache to a0 temporarily, and later on
+	 * arena_choose_hard() will re-associate properly.
+	 */
+	tcache->arena = NULL;
+	arena_t *arena;
+	if (!malloc_initialized()) {
+		/* If in initialization, assign to a0. */
+		arena = arena_get(tsd_tsdn(tsd), 0, false);
+		tcache_arena_associate(tsd_tsdn(tsd), tcache, arena);
+	} else {
+		arena = arena_choose(tsd, NULL);
+		/* This may happen if thread.tcache.enabled is used. */
+		if (tcache->arena == NULL) {
+			tcache_arena_associate(tsd_tsdn(tsd), tcache, arena);
+		}
+	}
+	assert(arena == tcache->arena);
+
+	return false;
+}
+
+/* Created manual tcache for tcache.create mallctl. */
 tcache_t *
-tcache_create(tsdn_t *tsdn, arena_t *arena)
-{
+tcache_create_explicit(tsd_t *tsd) {
 	tcache_t *tcache;
 	size_t size, stack_offset;
-	unsigned i;
 
-	size = offsetof(tcache_t, tbins) + (sizeof(tcache_bin_t) * nhbins);
+	size = sizeof(tcache_t);
 	/* Naturally align the pointer stacks. */
 	size = PTR_CEILING(size);
 	stack_offset = size;
 	size += stack_nelms * sizeof(void *);
 	/* Avoid false cacheline sharing. */
-	size = sa2u(size, CACHELINE);
+	size = sz_sa2u(size, CACHELINE);
 
-	tcache = ipallocztm(tsdn, size, CACHELINE, true, NULL, true,
+	tcache = ipallocztm(tsd_tsdn(tsd), size, CACHELINE, true, NULL, true,
 	    arena_get(TSDN_NULL, 0, true));
-	if (tcache == NULL)
-		return (NULL);
-
-	tcache_arena_associate(tsdn, tcache, arena);
-
-	ticker_init(&tcache->gc_ticker, TCACHE_GC_INCR);
-
-	assert((TCACHE_NSLOTS_SMALL_MAX & 1U) == 0);
-	for (i = 0; i < nhbins; i++) {
-		tcache->tbins[i].lg_fill_div = 1;
-		stack_offset += tcache_bin_info[i].ncached_max * sizeof(void *);
-		/*
-		 * avail points past the available space.  Allocations will
-		 * access the slots toward higher addresses (for the benefit of
-		 * prefetch).
-		 */
-		tcache->tbins[i].avail = (void **)((uintptr_t)tcache +
-		    (uintptr_t)stack_offset);
+	if (tcache == NULL) {
+		return NULL;
 	}
 
-	return (tcache);
+	tcache_init(tsd, tcache,
+	    (void *)((uintptr_t)tcache + (uintptr_t)stack_offset));
+	tcache_arena_associate(tsd_tsdn(tsd), tcache, arena_ichoose(tsd, NULL));
+
+	return tcache;
 }
 
 static void
-tcache_destroy(tsd_t *tsd, tcache_t *tcache)
-{
-	arena_t *arena;
-	unsigned i;
-
-	arena = arena_choose(tsd, NULL);
-	tcache_arena_dissociate(tsd_tsdn(tsd), tcache, arena);
+tcache_flush_cache(tsd_t *tsd, tcache_t *tcache) {
+	assert(tcache->arena != NULL);
 
-	for (i = 0; i < NBINS; i++) {
-		tcache_bin_t *tbin = &tcache->tbins[i];
+	for (unsigned i = 0; i < NBINS; i++) {
+		tcache_bin_t *tbin = tcache_small_bin_get(tcache, i);
 		tcache_bin_flush_small(tsd, tcache, tbin, i, 0);
 
-		if (config_stats && tbin->tstats.nrequests != 0) {
-			arena_bin_t *bin = &arena->bins[i];
-			malloc_mutex_lock(tsd_tsdn(tsd), &bin->lock);
-			bin->stats.nrequests += tbin->tstats.nrequests;
-			malloc_mutex_unlock(tsd_tsdn(tsd), &bin->lock);
+		if (config_stats) {
+			assert(tbin->tstats.nrequests == 0);
 		}
 	}
-
-	for (; i < nhbins; i++) {
-		tcache_bin_t *tbin = &tcache->tbins[i];
+	for (unsigned i = NBINS; i < nhbins; i++) {
+		tcache_bin_t *tbin = tcache_large_bin_get(tcache, i);
 		tcache_bin_flush_large(tsd, tbin, i, 0, tcache);
 
-		if (config_stats && tbin->tstats.nrequests != 0) {
-			malloc_mutex_lock(tsd_tsdn(tsd), &arena->lock);
-			arena->stats.nrequests_large += tbin->tstats.nrequests;
-			arena->stats.lstats[i - NBINS].nrequests +=
-			    tbin->tstats.nrequests;
-			malloc_mutex_unlock(tsd_tsdn(tsd), &arena->lock);
+		if (config_stats) {
+			assert(tbin->tstats.nrequests == 0);
 		}
 	}
 
 	if (config_prof && tcache->prof_accumbytes > 0 &&
-	    arena_prof_accum(tsd_tsdn(tsd), arena, tcache->prof_accumbytes))
+	    arena_prof_accum(tsd_tsdn(tsd), tcache->arena,
+	    tcache->prof_accumbytes)) {
 		prof_idump(tsd_tsdn(tsd));
-
-	idalloctm(tsd_tsdn(tsd), tcache, NULL, true, true);
+	}
 }
 
 void
-tcache_cleanup(tsd_t *tsd)
-{
-	tcache_t *tcache;
-
-	if (!config_tcache)
-		return;
+tcache_flush(void) {
+	tsd_t *tsd = tsd_fetch();
+	assert(tcache_available(tsd));
+	tcache_flush_cache(tsd, tsd_tcachep_get(tsd));
+}
 
-	if ((tcache = tsd_tcache_get(tsd)) != NULL) {
-		tcache_destroy(tsd, tcache);
-		tsd_tcache_set(tsd, NULL);
+static void
+tcache_destroy(tsd_t *tsd, tcache_t *tcache, bool tsd_tcache) {
+	tcache_flush_cache(tsd, tcache);
+	tcache_arena_dissociate(tsd_tsdn(tsd), tcache);
+
+	if (tsd_tcache) {
+		/* Release the avail array for the TSD embedded auto tcache. */
+		void *avail_array =
+		    (void *)((uintptr_t)tcache_small_bin_get(tcache, 0)->avail -
+		    (uintptr_t)tcache_bin_info[0].ncached_max * sizeof(void *));
+		idalloctm(tsd_tsdn(tsd), avail_array, NULL, NULL, true, true);
+	} else {
+		/* Release both the tcache struct and avail array. */
+		idalloctm(tsd_tsdn(tsd), tcache, NULL, NULL, true, true);
 	}
 }
 
+/* For auto tcache (embedded in TSD) only. */
 void
-tcache_enabled_cleanup(tsd_t *tsd)
-{
+tcache_cleanup(tsd_t *tsd) {
+	tcache_t *tcache = tsd_tcachep_get(tsd);
+	if (!tcache_available(tsd)) {
+		assert(tsd_tcache_enabled_get(tsd) == false);
+		if (config_debug) {
+			assert(tcache_small_bin_get(tcache, 0)->avail == NULL);
+		}
+		return;
+	}
+	assert(tsd_tcache_enabled_get(tsd));
+	assert(tcache_small_bin_get(tcache, 0)->avail != NULL);
 
-	/* Do nothing. */
+	tcache_destroy(tsd, tcache, true);
+	if (config_debug) {
+		tcache_small_bin_get(tcache, 0)->avail = NULL;
+	}
 }
 
 void
-tcache_stats_merge(tsdn_t *tsdn, tcache_t *tcache, arena_t *arena)
-{
+tcache_stats_merge(tsdn_t *tsdn, tcache_t *tcache, arena_t *arena) {
 	unsigned i;
 
 	cassert(config_stats);
 
-	malloc_mutex_assert_owner(tsdn, &arena->lock);
-
 	/* Merge and reset tcache stats. */
 	for (i = 0; i < NBINS; i++) {
 		arena_bin_t *bin = &arena->bins[i];
-		tcache_bin_t *tbin = &tcache->tbins[i];
+		tcache_bin_t *tbin = tcache_small_bin_get(tcache, i);
 		malloc_mutex_lock(tsdn, &bin->lock);
 		bin->stats.nrequests += tbin->tstats.nrequests;
 		malloc_mutex_unlock(tsdn, &bin->lock);
@@ -439,10 +534,9 @@ tcache_stats_merge(tsdn_t *tsdn, tcache_t *tcache, arena_t *arena)
 	}
 
 	for (; i < nhbins; i++) {
-		malloc_large_stats_t *lstats = &arena->stats.lstats[i - NBINS];
-		tcache_bin_t *tbin = &tcache->tbins[i];
-		arena->stats.nrequests_large += tbin->tstats.nrequests;
-		lstats->nrequests += tbin->tstats.nrequests;
+		tcache_bin_t *tbin = tcache_large_bin_get(tcache, i);
+		arena_stats_large_nrequests_add(tsdn, &arena->stats, i,
+		    tbin->tstats.nrequests);
 		tbin->tstats.nrequests = 0;
 	}
 }
@@ -454,8 +548,8 @@ tcaches_create_prep(tsd_t *tsd) {
 	malloc_mutex_lock(tsd_tsdn(tsd), &tcaches_mtx);
 
 	if (tcaches == NULL) {
-		tcaches = base_alloc(tsd_tsdn(tsd), sizeof(tcache_t *) *
-		    (MALLOCX_TCACHE_MAX+1));
+		tcaches = base_alloc(tsd_tsdn(tsd), b0get(), sizeof(tcache_t *)
+		    * (MALLOCX_TCACHE_MAX+1), CACHELINE);
 		if (tcaches == NULL) {
 			err = true;
 			goto label_return;
@@ -475,27 +569,22 @@ label_return:
 
 bool
 tcaches_create(tsd_t *tsd, unsigned *r_ind) {
+	witness_assert_depth(tsdn_witness_tsdp_get(tsd_tsdn(tsd)), 0);
+
 	bool err;
-	arena_t *arena;
-	tcache_t *tcache;
-	tcaches_t *elm;
 
 	if (tcaches_create_prep(tsd)) {
 		err = true;
 		goto label_return;
 	}
 
-	arena = arena_ichoose(tsd, NULL);
-	if (unlikely(arena == NULL)) {
-		err = true;
-		goto label_return;
-	}
-	tcache = tcache_create(tsd_tsdn(tsd), arena);
+	tcache_t *tcache = tcache_create_explicit(tsd);
 	if (tcache == NULL) {
 		err = true;
 		goto label_return;
 	}
 
+	tcaches_t *elm;
 	malloc_mutex_lock(tsd_tsdn(tsd), &tcaches_mtx);
 	if (tcaches_avail != NULL) {
 		elm = tcaches_avail;
@@ -512,69 +601,70 @@ tcaches_create(tsd_t *tsd, unsigned *r_ind) {
 
 	err = false;
 label_return:
-	malloc_mutex_assert_not_owner(tsd_tsdn(tsd), &tcaches_mtx);
+	witness_assert_depth(tsdn_witness_tsdp_get(tsd_tsdn(tsd)), 0);
 	return err;
 }
 
-static void
-tcaches_elm_flush(tsd_t *tsd, tcaches_t *elm) {
+static tcache_t *
+tcaches_elm_remove(tsd_t *tsd, tcaches_t *elm) {
 	malloc_mutex_assert_owner(tsd_tsdn(tsd), &tcaches_mtx);
 
 	if (elm->tcache == NULL) {
-		return;
+		return NULL;
 	}
-	tcache_destroy(tsd, elm->tcache);
+	tcache_t *tcache = elm->tcache;
 	elm->tcache = NULL;
+	return tcache;
 }
 
 void
 tcaches_flush(tsd_t *tsd, unsigned ind) {
 	malloc_mutex_lock(tsd_tsdn(tsd), &tcaches_mtx);
-	tcaches_elm_flush(tsd, &tcaches[ind]);
+	tcache_t *tcache = tcaches_elm_remove(tsd, &tcaches[ind]);
 	malloc_mutex_unlock(tsd_tsdn(tsd), &tcaches_mtx);
+	if (tcache != NULL) {
+		tcache_destroy(tsd, tcache, false);
+	}
 }
 
 void
 tcaches_destroy(tsd_t *tsd, unsigned ind) {
-	tcaches_t *elm;
-
 	malloc_mutex_lock(tsd_tsdn(tsd), &tcaches_mtx);
-	elm = &tcaches[ind];
-	tcaches_elm_flush(tsd, elm);
+	tcaches_t *elm = &tcaches[ind];
+	tcache_t *tcache = tcaches_elm_remove(tsd, elm);
 	elm->next = tcaches_avail;
 	tcaches_avail = elm;
 	malloc_mutex_unlock(tsd_tsdn(tsd), &tcaches_mtx);
+	if (tcache != NULL) {
+		tcache_destroy(tsd, tcache, false);
+	}
 }
 
 bool
 tcache_boot(tsdn_t *tsdn) {
-	unsigned i;
-
-	cassert(config_tcache);
-
-	/*
-	 * If necessary, clamp opt_lg_tcache_max, now that large_maxclass is
-	 * known.
-	 */
-	if (opt_lg_tcache_max < 0 || (ZU(1) << opt_lg_tcache_max) < SMALL_MAXCLASS)
+	/* If necessary, clamp opt_lg_tcache_max. */
+	if (opt_lg_tcache_max < 0 || (ZU(1) << opt_lg_tcache_max) <
+	    SMALL_MAXCLASS) {
 		tcache_maxclass = SMALL_MAXCLASS;
-	else if ((ZU(1) << opt_lg_tcache_max) > large_maxclass)
-		tcache_maxclass = large_maxclass;
-	else
+	} else {
 		tcache_maxclass = (ZU(1) << opt_lg_tcache_max);
+	}
 
-	if (malloc_mutex_init(&tcaches_mtx, "tcaches", WITNESS_RANK_TCACHES)) {
+	if (malloc_mutex_init(&tcaches_mtx, "tcaches", WITNESS_RANK_TCACHES,
+	    malloc_mutex_rank_exclusive)) {
 		return true;
 	}
 
-	nhbins = size2index(tcache_maxclass) + 1;
+	nhbins = sz_size2index(tcache_maxclass) + 1;
 
 	/* Initialize tcache_bin_info. */
-	tcache_bin_info = (tcache_bin_info_t *)base_alloc(tsdn, nhbins *
-	    sizeof(tcache_bin_info_t));
-	if (tcache_bin_info == NULL)
-		return (true);
+	tcache_bin_info = (tcache_bin_info_t *)base_alloc(tsdn, b0get(), nhbins
+	    * sizeof(tcache_bin_info_t), CACHELINE);
+	if (tcache_bin_info == NULL) {
+		return true;
+	}
 	stack_nelms = 0;
+	unsigned i;
 	for (i = 0; i < NBINS; i++) {
 		if ((arena_bin_info[i].nregs << 1) <= TCACHE_NSLOTS_SMALL_MIN) {
 			tcache_bin_info[i].ncached_max =
@@ -594,7 +684,7 @@ tcache_boot(tsdn_t *tsdn) {
 		stack_nelms += tcache_bin_info[i].ncached_max;
 	}
 
-	return (false);
+	return false;
 }
 
 void