about summary refs log tree commit diff
diff options
context:
space:
mode:
authorDJ Delorie <dj@delorie.com>2017-07-06 13:37:30 -0400
committerDJ Delorie <dj@delorie.com>2017-07-06 13:37:30 -0400
commitd5c3fafc4307c9b7a4c7d5cb381fcdbfad340bcc (patch)
tree380cfbc329860434d6b29825bd02ba5f0c7d4b30
parent3cefdd7310a5d1fad45648d9346e47df9c185fdc (diff)
downloadglibc-d5c3fafc4307c9b7a4c7d5cb381fcdbfad340bcc.tar.gz
glibc-d5c3fafc4307c9b7a4c7d5cb381fcdbfad340bcc.tar.xz
glibc-d5c3fafc4307c9b7a4c7d5cb381fcdbfad340bcc.zip
Add per-thread cache to malloc
* config.make.in: Enable experimental malloc option.
* configure.ac: Likewise.
* configure: Regenerate.
* manual/install.texi: Document it.
* INSTALL: Regenerate.
* malloc/Makefile: Likewise.
* malloc/malloc.c: Add per-thread cache (tcache).
(tcache_put): New.
(tcache_get): New.
(tcache_thread_freeres): New.
(tcache_init): New.
(__libc_malloc): Use cached chunks if available.
(__libc_free): Initialize tcache if needed.
(__libc_realloc): Likewise.
(__libc_calloc): Likewise.
(_int_malloc): Prefill tcache when appropriate.
(_int_free): Likewise.
(do_set_tcache_max): New.
(do_set_tcache_count): New.
(do_set_tcache_unsorted_limit): New.
* manual/probes.texi: Document new probes.
* malloc/arena.c: Add new tcache tunables.
* elf/dl-tunables.list: Likewise.
* manual/tunables.texi: Document them.
* NEWS: Mention the per-thread cache.
-rw-r--r--ChangeLog28
-rw-r--r--INSTALL6
-rw-r--r--NEWS8
-rw-r--r--config.make.in2
-rwxr-xr-xconfigure13
-rw-r--r--configure.ac7
-rw-r--r--elf/dl-tunables.list12
-rw-r--r--malloc/Makefile5
-rw-r--r--malloc/arena.c11
-rw-r--r--malloc/malloc.c350
-rw-r--r--manual/install.texi6
-rw-r--r--manual/probes.texi19
-rw-r--r--manual/tunables.texi32
13 files changed, 490 insertions, 9 deletions
diff --git a/ChangeLog b/ChangeLog
index ac598984e3..1c51b3be29 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,31 @@
+2017-07-06  DJ Delorie  <dj@delorie.com>
+
+	* config.make.in: Enable experimental malloc option.
+	* configure.ac: Likewise.
+	* configure: Regenerate.
+	* manual/install.texi: Document it.
+	* INSTALL: Regenerate.
+	* malloc/Makefile: Likewise.
+	* malloc/malloc.c: Add per-thread cache (tcache).
+	(tcache_put): New.
+	(tcache_get): New.
+	(tcache_thread_freeres): New.
+	(tcache_init): New.
+	(__libc_malloc): Use cached chunks if available.
+	(__libc_free): Initialize tcache if needed.
+	(__libc_realloc): Likewise.
+	(__libc_calloc): Likewise.
+	(_int_malloc): Prefill tcache when appropriate.
+	(_int_free): Likewise.
+	(do_set_tcache_max): New.
+	(do_set_tcache_count): New.
+	(do_set_tcache_unsorted_limit): New.
+	* manual/probes.texi: Document new probes.
+	* malloc/arena.c: Add new tcache tunables.
+	* elf/dl-tunables.list: Likewise.
+	* manual/tunables.texi: Document them.
+	* NEWS: Mention the per-thread cache.
+
 2017-07-06  Joseph Myers  <joseph@codesourcery.com>
 
 	* iconvdata/tst-loading.c (TIMEOUT): Define to 30.
diff --git a/INSTALL b/INSTALL
index 5c745244a6..0ff87ed603 100644
--- a/INSTALL
+++ b/INSTALL
@@ -200,6 +200,12 @@ will be used, and CFLAGS sets optimization options for the compiler.
      libnss_nisplus are not built at all.  Use this option to enable
      libnsl with all depending NSS modules and header files.
 
+'--disable-experimental-malloc'
+     By default, a per-thread cache is enabled in 'malloc'.  While this
+     cache can be disabled on a per-application basis using tunables
+     (set glibc.malloc.tcache_count to zero), this option can be used to
+     remove it from the build completely.
+
 '--build=BUILD-SYSTEM'
 '--host=HOST-SYSTEM'
      These options are for cross-compiling.  If you specify both options
diff --git a/NEWS b/NEWS
index 616b3244e7..1266276fcd 100644
--- a/NEWS
+++ b/NEWS
@@ -9,6 +9,14 @@ Version 2.26
 
 Major new features:
 
+* A per-thread cache has been added to malloc. Access to the cache requires
+  no locks and therefore significantly accelerates the fast path to allocate
+  and free small amounts of memory. Refilling an empty cache requires locking
+  the underlying arena. Performance measurements show significant gains in a
+  wide variety of user workloads. Workloads were captured using a special
+  instrumented malloc and analyzed with a malloc simulator. Contributed by
+  DJ Delorie with the help of Florian Weimer, and Carlos O'Donell.
+
 * Unicode 10.0.0 Support: Character encoding, character type info, and
   transliteration tables are all updated to Unicode 10.0.0, using
   generator scripts contributed by Mike FABIAN (Red Hat).
diff --git a/config.make.in b/config.make.in
index dadabf9b6a..5a4a054612 100644
--- a/config.make.in
+++ b/config.make.in
@@ -78,6 +78,8 @@ multi-arch = @multi_arch@
 
 mach-interface-list = @mach_interface_list@
 
+experimental-malloc = @experimental_malloc@
+
 nss-crypt = @libc_cv_nss_crypt@
 static-nss-crypt = @libc_cv_static_nss_crypt@
 
diff --git a/configure b/configure
index 8e5c5bdde0..d8e1c50e11 100755
--- a/configure
+++ b/configure
@@ -674,6 +674,7 @@ build_obsolete_nsl
 link_obsolete_rpc
 libc_cv_static_nss_crypt
 libc_cv_nss_crypt
+experimental_malloc
 enable_werror
 all_warnings
 force_install
@@ -779,6 +780,7 @@ enable_kernel
 enable_all_warnings
 enable_werror
 enable_multi_arch
+enable_experimental_malloc
 enable_nss_crypt
 enable_obsolete_rpc
 enable_obsolete_nsl
@@ -1450,6 +1452,8 @@ Optional Features:
   --disable-werror        do not build with -Werror
   --enable-multi-arch     enable single DSO with optimizations for multiple
                           architectures
+  --disable-experimental-malloc
+                          disable experimental malloc features
   --enable-nss-crypt      enable libcrypt to use nss
   --enable-obsolete-rpc   build and install the obsolete RPC code for
                           link-time usage
@@ -3522,6 +3526,15 @@ else
 fi
 
 
+# Check whether --enable-experimental-malloc was given.
+if test "${enable_experimental_malloc+set}" = set; then :
+  enableval=$enable_experimental_malloc; experimental_malloc=$enableval
+else
+  experimental_malloc=yes
+fi
+
+
+
 # Check whether --enable-nss-crypt was given.
 if test "${enable_nss_crypt+set}" = set; then :
   enableval=$enable_nss_crypt; nss_crypt=$enableval
diff --git a/configure.ac b/configure.ac
index 7e03019c4f..77456aa8d9 100644
--- a/configure.ac
+++ b/configure.ac
@@ -313,6 +313,13 @@ AC_ARG_ENABLE([multi-arch],
 	      [multi_arch=$enableval],
 	      [multi_arch=default])
 
+AC_ARG_ENABLE([experimental-malloc],
+	      AC_HELP_STRING([--disable-experimental-malloc],
+			     [disable experimental malloc features]),
+	      [experimental_malloc=$enableval],
+	      [experimental_malloc=yes])
+AC_SUBST(experimental_malloc)
+
 AC_ARG_ENABLE([nss-crypt],
 	      AC_HELP_STRING([--enable-nss-crypt],
 			     [enable libcrypt to use nss]),
diff --git a/elf/dl-tunables.list b/elf/dl-tunables.list
index df4f9622b4..c188c6ad52 100644
--- a/elf/dl-tunables.list
+++ b/elf/dl-tunables.list
@@ -76,6 +76,18 @@ glibc {
       minval: 1
       security_level: SXID_IGNORE
     }
+    tcache_max {
+      type: SIZE_T
+      security_level: SXID_ERASE
+    }
+    tcache_count {
+      type: SIZE_T
+      security_level: SXID_ERASE
+    }
+    tcache_unsorted_limit {
+      type: SIZE_T
+      security_level: SXID_ERASE
+    }
   }
   tune {
     hwcap_mask {
diff --git a/malloc/Makefile b/malloc/Makefile
index b50de7cd6c..3fa395b949 100644
--- a/malloc/Makefile
+++ b/malloc/Makefile
@@ -189,6 +189,11 @@ tst-malloc-usable-static-ENV = $(tst-malloc-usable-ENV)
 tst-malloc-usable-tunables-ENV = GLIBC_TUNABLES=glibc.malloc.check=3
 tst-malloc-usable-static-tunables-ENV = $(tst-malloc-usable-tunables-ENV)
 
+ifeq ($(experimental-malloc),yes)
+CPPFLAGS-malloc.c += -DUSE_TCACHE=1
+else
+CPPFLAGS-malloc.c += -DUSE_TCACHE=0
+endif
 # Uncomment this for test releases.  For public releases it is too expensive.
 #CPPFLAGS-malloc.o += -DMALLOC_DEBUG=1
 
diff --git a/malloc/arena.c b/malloc/arena.c
index 660d638c93..dc14fae152 100644
--- a/malloc/arena.c
+++ b/malloc/arena.c
@@ -236,6 +236,11 @@ TUNABLE_CALLBACK_FNDECL (set_perturb_byte, int32_t)
 TUNABLE_CALLBACK_FNDECL (set_trim_threshold, size_t)
 TUNABLE_CALLBACK_FNDECL (set_arena_max, size_t)
 TUNABLE_CALLBACK_FNDECL (set_arena_test, size_t)
+#if USE_TCACHE
+TUNABLE_CALLBACK_FNDECL (set_tcache_max, size_t)
+TUNABLE_CALLBACK_FNDECL (set_tcache_count, size_t)
+TUNABLE_CALLBACK_FNDECL (set_tcache_unsorted_limit, size_t)
+#endif
 #else
 /* Initialization routine. */
 #include <string.h>
@@ -322,6 +327,12 @@ ptmalloc_init (void)
   TUNABLE_GET (mmap_max, int32_t, TUNABLE_CALLBACK (set_mmaps_max));
   TUNABLE_GET (arena_max, size_t, TUNABLE_CALLBACK (set_arena_max));
   TUNABLE_GET (arena_test, size_t, TUNABLE_CALLBACK (set_arena_test));
+#if USE_TCACHE
+  TUNABLE_GET (tcache_max, size_t, TUNABLE_CALLBACK (set_tcache_max));
+  TUNABLE_GET (tcache_count, size_t, TUNABLE_CALLBACK (set_tcache_count));
+  TUNABLE_GET (tcache_unsorted_limit, size_t,
+	       TUNABLE_CALLBACK (set_tcache_unsorted_limit));
+#endif
   __libc_lock_unlock (main_arena.mutex);
 #else
   const char *s = NULL;
diff --git a/malloc/malloc.c b/malloc/malloc.c
index aa45626093..2527e25047 100644
--- a/malloc/malloc.c
+++ b/malloc/malloc.c
@@ -238,6 +238,9 @@
 /* For ALIGN_UP et. al.  */
 #include <libc-pointer-arith.h>
 
+/* For DIAG_PUSH/POP_NEEDS_COMMENT et al.  */
+#include <libc-diag.h>
+
 #include <malloc/malloc-internal.h>
 
 /*
@@ -296,6 +299,30 @@ __malloc_assert (const char *assertion, const char *file, unsigned int line,
 }
 #endif
 
+#if USE_TCACHE
+/* We want 64 entries.  This is an arbitrary limit, which tunables can reduce.  */
+# define TCACHE_MAX_BINS		64
+# define MAX_TCACHE_SIZE	tidx2usize (TCACHE_MAX_BINS-1)
+
+/* Only used to pre-fill the tunables.  */
+# define tidx2usize(idx)	(((size_t) idx) * MALLOC_ALIGNMENT + MINSIZE - SIZE_SZ)
+
+/* When "x" is from chunksize().  */
+# define csize2tidx(x) (((x) - MINSIZE + MALLOC_ALIGNMENT - 1) / MALLOC_ALIGNMENT)
+/* When "x" is a user-provided size.  */
+# define usize2tidx(x) csize2tidx (request2size (x))
+
+/* With rounding and alignment, the bins are...
+   idx 0   bytes 0..24 (64-bit) or 0..12 (32-bit)
+   idx 1   bytes 25..40 or 13..20
+   idx 2   bytes 41..56 or 21..28
+   etc.  */
+
+/* This is another arbitrary limit, which tunables can change.  Each
+   tcache bin will hold at most this number of chunks.  */
+# define TCACHE_FILL_COUNT 7
+#endif
+
 
 /*
   REALLOC_ZERO_BYTES_FREES should be set if a call to
@@ -1712,6 +1739,17 @@ struct malloc_par
 
   /* First address handed out by MORECORE/sbrk.  */
   char *sbrk_base;
+
+#if USE_TCACHE
+  /* Maximum number of buckets to use.  */
+  size_t tcache_bins;
+  size_t tcache_max_bytes;
+  /* Maximum number of chunks in each bucket.  */
+  size_t tcache_count;
+  /* Maximum number of chunks to remove from the unsorted list, which
+     aren't used to prefill the cache.  */
+  size_t tcache_unsorted_limit;
+#endif
 };
 
 /* There are several instances of this struct ("arenas") in this
@@ -1750,6 +1788,13 @@ static struct malloc_par mp_ =
   .trim_threshold = DEFAULT_TRIM_THRESHOLD,
 #define NARENAS_FROM_NCORES(n) ((n) * (sizeof (long) == 4 ? 2 : 8))
   .arena_test = NARENAS_FROM_NCORES (1)
+#if USE_TCACHE
+  ,
+  .tcache_count = TCACHE_FILL_COUNT,
+  .tcache_bins = TCACHE_MAX_BINS,
+  .tcache_max_bytes = tidx2usize (TCACHE_MAX_BINS-1),
+  .tcache_unsorted_limit = 0 /* No limit.  */
+#endif
 };
 
 /* Maximum size of memory handled in fastbins.  */
@@ -2875,6 +2920,124 @@ mremap_chunk (mchunkptr p, size_t new_size)
 
 /*------------------------ Public wrappers. --------------------------------*/
 
+#if USE_TCACHE
+
+/* We overlay this structure on the user-data portion of a chunk when
+   the chunk is stored in the per-thread cache.  */
+typedef struct tcache_entry
+{
+  struct tcache_entry *next;
+} tcache_entry;
+
+/* There is one of these for each thread, which contains the
+   per-thread cache (hence "tcache_perthread_struct").  Keeping
+   overall size low is mildly important.  Note that COUNTS and ENTRIES
+   are redundant (we could have just counted the linked list each
+   time), this is for performance reasons.  */
+typedef struct tcache_perthread_struct
+{
+  char counts[TCACHE_MAX_BINS];
+  tcache_entry *entries[TCACHE_MAX_BINS];
+} tcache_perthread_struct;
+
+static __thread char tcache_shutting_down = 0;
+static __thread tcache_perthread_struct *tcache = NULL;
+
+/* Caller must ensure that we know tc_idx is valid and there's room
+   for more chunks.  */
+static void
+tcache_put (mchunkptr chunk, size_t tc_idx)
+{
+  tcache_entry *e = (tcache_entry *) chunk2mem (chunk);
+  assert (tc_idx < TCACHE_MAX_BINS);
+  e->next = tcache->entries[tc_idx];
+  tcache->entries[tc_idx] = e;
+  ++(tcache->counts[tc_idx]);
+}
+
+/* Caller must ensure that we know tc_idx is valid and there's
+   available chunks to remove.  */
+static void *
+tcache_get (size_t tc_idx)
+{
+  tcache_entry *e = tcache->entries[tc_idx];
+  assert (tc_idx < TCACHE_MAX_BINS);
+  assert (tcache->entries[tc_idx] > 0);
+  tcache->entries[tc_idx] = e->next;
+  --(tcache->counts[tc_idx]);
+  return (void *) e;
+}
+
+static void __attribute__ ((section ("__libc_thread_freeres_fn")))
+tcache_thread_freeres (void)
+{
+  int i;
+  tcache_perthread_struct *tcache_tmp = tcache;
+
+  if (!tcache)
+    return;
+
+  tcache = NULL;
+
+  for (i = 0; i < TCACHE_MAX_BINS; ++i)
+    {
+      while (tcache_tmp->entries[i])
+	{
+	  tcache_entry *e = tcache_tmp->entries[i];
+	  tcache_tmp->entries[i] = e->next;
+	  __libc_free (e);
+	}
+    }
+
+  __libc_free (tcache_tmp);
+
+  tcache_shutting_down = 1;
+}
+text_set_element (__libc_thread_subfreeres, tcache_thread_freeres);
+
+static void
+tcache_init(void)
+{
+  mstate ar_ptr;
+  void *victim = 0;
+  const size_t bytes = sizeof (tcache_perthread_struct);
+
+  if (tcache_shutting_down)
+    return;
+
+  arena_get (ar_ptr, bytes);
+  victim = _int_malloc (ar_ptr, bytes);
+  if (!victim && ar_ptr != NULL)
+    {
+      ar_ptr = arena_get_retry (ar_ptr, bytes);
+      victim = _int_malloc (ar_ptr, bytes);
+    }
+
+
+  if (ar_ptr != NULL)
+    __libc_lock_unlock (ar_ptr->mutex);
+
+  /* In a low memory situation, we may not be able to allocate memory
+     - in which case, we just keep trying later.  However, we
+     typically do this very early, so either there is sufficient
+     memory, or there isn't enough memory to do non-trivial
+     allocations anyway.  */
+  if (victim)
+    {
+      tcache = (tcache_perthread_struct *) victim;
+      memset (tcache, 0, sizeof (tcache_perthread_struct));
+    }
+
+}
+
+#define MAYBE_INIT_TCACHE() \
+  if (__glibc_unlikely (tcache == NULL)) \
+    tcache_init();
+
+#else
+#define MAYBE_INIT_TCACHE()
+#endif
+
 void *
 __libc_malloc (size_t bytes)
 {
@@ -2885,6 +3048,23 @@ __libc_malloc (size_t bytes)
     = atomic_forced_read (__malloc_hook);
   if (__builtin_expect (hook != NULL, 0))
     return (*hook)(bytes, RETURN_ADDRESS (0));
+#if USE_TCACHE
+  /* int_free also calls request2size, be careful to not pad twice.  */
+  size_t tbytes = request2size (bytes);
+  size_t tc_idx = csize2tidx (tbytes);
+
+  MAYBE_INIT_TCACHE ();
+
+  DIAG_PUSH_NEEDS_COMMENT;
+  if (tc_idx < mp_.tcache_bins
+      /*&& tc_idx < TCACHE_MAX_BINS*/ /* to appease gcc */
+      && tcache
+      && tcache->entries[tc_idx] != NULL)
+    {
+      return tcache_get (tc_idx);
+    }
+  DIAG_POP_NEEDS_COMMENT;
+#endif
 
   arena_get (ar_ptr, bytes);
 
@@ -2944,6 +3124,8 @@ __libc_free (void *mem)
       return;
     }
 
+  MAYBE_INIT_TCACHE ();
+
   ar_ptr = arena_for_chunk (p);
   _int_free (ar_ptr, p, 0);
 }
@@ -2981,7 +3163,10 @@ __libc_realloc (void *oldmem, size_t bytes)
   if (chunk_is_mmapped (oldp))
     ar_ptr = NULL;
   else
-    ar_ptr = arena_for_chunk (oldp);
+    {
+      MAYBE_INIT_TCACHE ();
+      ar_ptr = arena_for_chunk (oldp);
+    }
 
   /* Little security check which won't hurt performance: the allocator
      never wrapps around at the end of the address space.  Therefore
@@ -3206,6 +3391,8 @@ __libc_calloc (size_t n, size_t elem_size)
 
   sz = bytes;
 
+  MAYBE_INIT_TCACHE ();
+
   arena_get (av, sz);
   if (av)
     {
@@ -3336,6 +3523,10 @@ _int_malloc (mstate av, size_t bytes)
   mchunkptr fwd;                    /* misc temp for linking */
   mchunkptr bck;                    /* misc temp for linking */
 
+#if USE_TCACHE
+  size_t tcache_unsorted_count;	    /* count of unsorted chunks processed */
+#endif
+
   const char *errstr = NULL;
 
   /*
@@ -3365,19 +3556,22 @@ _int_malloc (mstate av, size_t bytes)
      can try it without checking, which saves some time on this fast path.
    */
 
+#define REMOVE_FB(fb, victim, pp)			\
+  do							\
+    {							\
+      victim = pp;					\
+      if (victim == NULL)				\
+	break;						\
+    }							\
+  while ((pp = catomic_compare_and_exchange_val_acq (fb, victim->fd, victim)) \
+	 != victim);					\
+
   if ((unsigned long) (nb) <= (unsigned long) (get_max_fast ()))
     {
       idx = fastbin_index (nb);
       mfastbinptr *fb = &fastbin (av, idx);
       mchunkptr pp = *fb;
-      do
-        {
-          victim = pp;
-          if (victim == NULL)
-            break;
-        }
-      while ((pp = catomic_compare_and_exchange_val_acq (fb, victim->fd, victim))
-             != victim);
+      REMOVE_FB (fb, victim, pp);
       if (victim != 0)
         {
           if (__builtin_expect (fastbin_index (chunksize (victim)) != idx, 0))
@@ -3388,6 +3582,26 @@ _int_malloc (mstate av, size_t bytes)
               return NULL;
             }
           check_remalloced_chunk (av, victim, nb);
+#if USE_TCACHE
+	  /* While we're here, if we see other chunks of the same size,
+	     stash them in the tcache.  */
+	  size_t tc_idx = csize2tidx (nb);
+	  if (tcache && tc_idx < mp_.tcache_bins)
+	    {
+	      mchunkptr tc_victim;
+
+	      /* While bin not empty and tcache not full, copy chunks over.  */
+	      while (tcache->counts[tc_idx] < mp_.tcache_count
+		     && (pp = *fb) != NULL)
+		{
+		  REMOVE_FB (fb, tc_victim, pp);
+		  if (tc_victim != 0)
+		    {
+		      tcache_put (tc_victim, tc_idx);
+	            }
+		}
+	    }
+#endif
           void *p = chunk2mem (victim);
           alloc_perturb (p, bytes);
           return p;
@@ -3426,6 +3640,32 @@ _int_malloc (mstate av, size_t bytes)
               if (av != &main_arena)
 		set_non_main_arena (victim);
               check_malloced_chunk (av, victim, nb);
+#if USE_TCACHE
+	  /* While we're here, if we see other chunks of the same size,
+	     stash them in the tcache.  */
+	  size_t tc_idx = csize2tidx (nb);
+	  if (tcache && tc_idx < mp_.tcache_bins)
+	    {
+	      mchunkptr tc_victim;
+
+	      /* While bin not empty and tcache not full, copy chunks over.  */
+	      while (tcache->counts[tc_idx] < mp_.tcache_count
+		     && (tc_victim = last (bin)) != bin)
+		{
+		  if (tc_victim != 0)
+		    {
+		      bck = tc_victim->bk;
+		      set_inuse_bit_at_offset (tc_victim, nb);
+		      if (av != &main_arena)
+			set_non_main_arena (tc_victim);
+		      bin->bk = bck;
+		      bck->fd = bin;
+
+		      tcache_put (tc_victim, tc_idx);
+	            }
+		}
+	    }
+#endif
               void *p = chunk2mem (victim);
               alloc_perturb (p, bytes);
               return p;
@@ -3464,6 +3704,16 @@ _int_malloc (mstate av, size_t bytes)
      otherwise need to expand memory to service a "small" request.
    */
 
+#if USE_TCACHE
+  INTERNAL_SIZE_T tcache_nb = 0;
+  size_t tc_idx = csize2tidx (nb);
+  if (tcache && tc_idx < mp_.tcache_bins)
+    tcache_nb = nb;
+  int return_cached = 0;
+
+  tcache_unsorted_count = 0;
+#endif
+
   for (;; )
     {
       int iters = 0;
@@ -3524,10 +3774,26 @@ _int_malloc (mstate av, size_t bytes)
               set_inuse_bit_at_offset (victim, size);
               if (av != &main_arena)
 		set_non_main_arena (victim);
+#if USE_TCACHE
+	      /* Fill cache first, return to user only if cache fills.
+		 We may return one of these chunks later.  */
+	      if (tcache_nb
+		  && tcache->counts[tc_idx] < mp_.tcache_count)
+		{
+		  tcache_put (victim, tc_idx);
+		  return_cached = 1;
+		  continue;
+		}
+	      else
+		{
+#endif
               check_malloced_chunk (av, victim, nb);
               void *p = chunk2mem (victim);
               alloc_perturb (p, bytes);
               return p;
+#if USE_TCACHE
+		}
+#endif
             }
 
           /* place chunk in bin */
@@ -3594,11 +3860,31 @@ _int_malloc (mstate av, size_t bytes)
           fwd->bk = victim;
           bck->fd = victim;
 
+#if USE_TCACHE
+      /* If we've processed as many chunks as we're allowed while
+	 filling the cache, return one of the cached ones.  */
+      ++tcache_unsorted_count;
+      if (return_cached
+	  && mp_.tcache_unsorted_limit > 0
+	  && tcache_unsorted_count > mp_.tcache_unsorted_limit)
+	{
+	  return tcache_get (tc_idx);
+	}
+#endif
+
 #define MAX_ITERS       10000
           if (++iters >= MAX_ITERS)
             break;
         }
 
+#if USE_TCACHE
+      /* If all the small chunks we found ended up cached, return one now.  */
+      if (return_cached)
+	{
+	  return tcache_get (tc_idx);
+	}
+#endif
+
       /*
          If a large request, scan through the chunks of current bin in
          sorted order to find smallest that fits.  Use the skip list for this.
@@ -3884,6 +4170,20 @@ _int_free (mstate av, mchunkptr p, int have_lock)
 
   check_inuse_chunk(av, p);
 
+#if USE_TCACHE
+  {
+    size_t tc_idx = csize2tidx (size);
+
+    if (tcache
+	&& tc_idx < mp_.tcache_bins
+	&& tcache->counts[tc_idx] < mp_.tcache_count)
+      {
+	tcache_put (p, tc_idx);
+	return;
+      }
+  }
+#endif
+
   /*
     If eligible, place chunk on a fastbin so it can be found
     and used quickly in malloc.
@@ -4845,6 +5145,38 @@ do_set_arena_max (size_t value)
   return 1;
 }
 
+#if USE_TCACHE
+static inline int
+__always_inline
+do_set_tcache_max (size_t value)
+{
+  if (value >= 0 && value <= MAX_TCACHE_SIZE)
+    {
+      LIBC_PROBE (memory_tunable_tcache_max_bytes, 2, value, mp_.tcache_max_bytes);
+      mp_.tcache_max_bytes = value;
+      mp_.tcache_bins = csize2tidx (request2size(value)) + 1;
+    }
+  return 1;
+}
+
+static inline int
+__always_inline
+do_set_tcache_count (size_t value)
+{
+  LIBC_PROBE (memory_tunable_tcache_count, 2, value, mp_.tcache_count);
+  mp_.tcache_count = value;
+  return 1;
+}
+
+static inline int
+__always_inline
+do_set_tcache_unsorted_limit (size_t value)
+{
+  LIBC_PROBE (memory_tunable_tcache_unsorted_limit, 2, value, mp_.tcache_unsorted_limit);
+  mp_.tcache_unsorted_limit = value;
+  return 1;
+}
+#endif
 
 int
 __libc_mallopt (int param_number, int value)
diff --git a/manual/install.texi b/manual/install.texi
index 03eb2dd93b..b8deb9ceba 100644
--- a/manual/install.texi
+++ b/manual/install.texi
@@ -232,6 +232,12 @@ libnss_nisplus are not built at all.
 Use this option to enable libnsl with all depending NSS modules and
 header files.
 
+@item --disable-experimental-malloc
+By default, a per-thread cache is enabled in @code{malloc}.  While
+this cache can be disabled on a per-application basis using tunables
+(set glibc.malloc.tcache_count to zero), this option can be used to
+remove it from the build completely.
+
 @item --build=@var{build-system}
 @itemx --host=@var{host-system}
 These options are for cross-compiling.  If you specify both options and
diff --git a/manual/probes.texi b/manual/probes.texi
index eb91c62703..96acaed206 100644
--- a/manual/probes.texi
+++ b/manual/probes.texi
@@ -231,6 +231,25 @@ dynamic brk/mmap thresholds.  Argument @var{$arg1} and @var{$arg2} are
 the adjusted mmap and trim thresholds, respectively.
 @end deftp
 
+@deftp Probe memory_tunable_tcache_max_bytes (int @var{$arg1}, int @var{$arg2})
+This probe is triggered when the @code{glibc.malloc.tcache_max}
+tunable is set.  Argument @var{$arg1} is the requested value, and
+@var{$arg2} is the previous value of this tunable.
+@end deftp
+
+@deftp Probe memory_tunable_tcache_count (int @var{$arg1}, int @var{$arg2})
+This probe is triggered when the @code{glibc.malloc.tcache_count}
+tunable is set.  Argument @var{$arg1} is the requested value, and
+@var{$arg2} is the previous value of this tunable.
+@end deftp
+
+@deftp Probe memory_tunable_tcache_unsorted_limit (int @var{$arg1}, int @var{$arg2})
+This probe is triggered when the
+@code{glibc.malloc.tcache_unsorted_limit} tunable is set.  Argument
+@var{$arg1} is the requested value, and @var{$arg2} is the previous
+value of this tunable.
+@end deftp
+
 @node Mathematical Function Probes
 @section Mathematical Function Probes
 
diff --git a/manual/tunables.texi b/manual/tunables.texi
index 9331b03702..b16d591b90 100644
--- a/manual/tunables.texi
+++ b/manual/tunables.texi
@@ -193,6 +193,38 @@ systems the limit is twice the number of cores online and on 64-bit systems, it
 is 8 times the number of cores online.
 @end deftp
 
+@deftp Tunable glibc.malloc.tcache_max
+The maximum size of a request (in bytes) which may be met via the
+per-thread cache.  The default (and maximum) value is 1032 bytes on
+64-bit systems and 516 bytes on 32-bit systems.
+@end deftp
+
+@deftp Tunable glibc.malloc.tcache_count
+The maximum number of chunks of each size to cache. The default is 7.
+There is no upper limit, other than available system memory.  If set
+to zero, the per-thread cache is effectively disabled.
+
+The approximate maximum overhead of the per-thread cache is thus equal
+to the number of bins times the chunk count in each bin times the size
+of each chunk.  With defaults, the approximate maximum overhead of the
+per-thread cache is approximately 236 KB on 64-bit systems and 118 KB
+on 32-bit systems.
+@end deftp
+
+@deftp Tunable glibc.malloc.tcache_unsorted_limit
+When the user requests memory and the request cannot be met via the
+per-thread cache, the arenas are used to meet the request.  At this
+time, additional chunks will be moved from existing arena lists to
+pre-fill the corresponding cache.  While copies from the fastbins,
+smallbins, and regular bins are bounded and predictable due to the bin
+sizes, copies from the unsorted bin are not bounded, and incur
+additional time penalties as they need to be sorted as they're
+scanned.  To make scanning the unsorted list more predictable and
+bounded, the user may set this tunable to limit the number of chunks
+that are scanned from the unsorted list while searching for chunks to
+pre-fill the per-thread cache with.  The default, or when set to zero,
+is no limit.
+
 @node Hardware Capability Tunables
 @section Hardware Capability Tunables
 @cindex hardware capability tunables