libglusterfs: make memory pools more thread-friendly

Early multiplexing tests revealed *massive* contention on certain
pools' global locks - especially for dictionaries and secondarily for
call stubs.  For the thread counts that multiplexing can create, a
more lock-free solution is clearly needed.  Also, the current mem-pool
implementation does a poor job releasing memory back to the system,
artificially inflating memory usage to match whatever the worst case
was since the process started.  This is bad in general, but especially
so for multiplexing where there are more pools and a major point of
the whole exercise is to reduce memory consumption.

The basic ideas for the new design are these

  There is one pool, globally, for each power-of-two size range.
  Every attempt to create a new pool within this range will instead
  add a reference to the existing pool.

  Instead of adding pools for each translator within each multiplexed
  brick (potentially infinite and quite possibly thousands), we
  allocate one set of size-based pools per *thread* (hundreds at
  worst).

  Each per-thread pool is divided into hot and cold lists.  Every
  allocation first attempts to use the hot list, then the cold list.
  When objects are freed, they always go on the hot list.

  There is one global "pool sweeper" thread, which periodically
  reclaims everything in each pool's cold list and then "demotes" the
  current hot list to be the new cold list.

  For normal allocation activity, only a per-thread lock need be
  taken, and even that only to guard against very rare contention from
  the pool sweeper.  When threads start and stop, a global lock must
  be taken to add them to the pool sweeper's list.  Lock contention is
  therefore extremely low, and the hot/cold lists also provide good
  locality.

A more complete explanation (of a similar earlier design) can be found
here:

 http://www.gluster.org/pipermail/gluster-devel/2016-October/051160.html

Backport of:
> Change-Id: I5bc8a1ba57cfb553998f979a498886e0d006e665
> BUG: 1385758
> Reviewed-on: https://review.gluster.org/15645

BUG: 1418091
Change-Id: Id09bbea41f65fcd245822607bc204f3a34904dc2
Signed-off-by: Jeff Darcy <jdarcy@redhat.com>
Reviewed-on: https://review.gluster.org/16531
Smoke: Gluster Build System <jenkins@build.gluster.org>
NetBSD-regression: NetBSD Build System <jenkins@build.gluster.org>
CentOS-regression: Gluster Build System <jenkins@build.gluster.org>
Reviewed-by: Shyamsundar Ranganathan <srangana@redhat.com>

5 files changed, 375 insertions, 238 deletions

diff --git a/glusterfsd/src/glusterfsd.c b/glusterfsd/src/glusterfsd.c
index 1f7b63e7594..f402246e78e 100644
--- a/glusterfsd/src/glusterfsd.c
+++ b/glusterfsd/src/glusterfsd.c
@@ -2472,6 +2472,13 @@ main (int argc, char *argv[])
         if (ret)
                 goto out;
 
+        /*
+         * If we do this before daemonize, the pool-sweeper thread dies with
+         * the parent, but we want to do it as soon as possible after that in
+         * case something else depends on pool allocations.
+         */
+        mem_pools_init ();
+
 #ifdef GF_LINUX_HOST_OS
         ret = set_oom_score_adj (ctx);
         if (ret)
diff --git a/libglusterfs/src/mem-pool.c b/libglusterfs/src/mem-pool.c
index b11e4c72e12..deaf286c298 100644
--- a/libglusterfs/src/mem-pool.c
+++ b/libglusterfs/src/mem-pool.c
@@ -338,8 +338,20 @@ free:
 }
 
 
-/* Based on the mem-type that is used for the allocation, GF_FREE can be
+/*
+ * Based on the mem-type that is used for the allocation, GF_FREE can be
  * called, or something more intelligent for the structure can be done.
+ *
+ * NOTE: this will not work for allocations from a memory pool.  It never did,
+ * because those allocations never set the type in the first place.  Any caller
+ * that relies on knowing whether a particular type was allocated via a pool or
+ * not is *BROKEN*, or will be any time either this module or the module
+ * "owning" the type changes.  The proper way to handle this, assuming the
+ * caller is not smart enough to call a type-specific free function themselves,
+ * would be to create a callback interface where destructors for specific types
+ * can be registered so that code *here* (GF_FREE, mem_put, etc.) can do the
+ * right thing.  That allows type-specific behavior without creating the kind
+ * of fragile coupling that we have now.
  */
 int
 gf_get_mem_type (void *ptr)
@@ -358,80 +370,201 @@ gf_get_mem_type (void *ptr)
 }
 
 
+#define POOL_SMALLEST   7       /* i.e. 128 */
+#define POOL_LARGEST    20      /* i.e. 1048576 */
+#define NPOOLS          (POOL_LARGEST - POOL_SMALLEST + 1)
+#define N_COLD_LISTS    1024
+#define POOL_SWEEP_SECS 30
 
-struct mem_pool *
-mem_pool_new_fn (unsigned long sizeof_type,
-                 unsigned long count, char *name)
+static pthread_key_t            pool_key;
+static pthread_mutex_t          pool_lock       = PTHREAD_MUTEX_INITIALIZER;
+static struct list_head         pool_threads;
+static pthread_mutex_t          pool_free_lock  = PTHREAD_MUTEX_INITIALIZER;
+static struct list_head         pool_free_threads;
+static struct mem_pool          pools[NPOOLS];
+static size_t                   pool_list_size;
+static unsigned long            sweep_times;
+static unsigned long            sweep_usecs;
+static unsigned long            frees_to_system;
+
+typedef struct {
+        struct list_head        death_row;
+        pooled_obj_hdr_t        *cold_lists[N_COLD_LISTS];
+        unsigned int            n_cold_lists;
+} sweep_state_t;
+
+
+void
+collect_garbage (sweep_state_t *state, per_thread_pool_list_t *pool_list)
 {
-        struct mem_pool  *mem_pool = NULL;
-        unsigned long     padded_sizeof_type = 0;
-        GF_UNUSED void             *pool = NULL;
-        GF_UNUSED int               i = 0;
-        int               ret = 0;
-        GF_UNUSED struct list_head *list = NULL;
-        glusterfs_ctx_t  *ctx = NULL;
-
-        if (!sizeof_type || !count) {
-                gf_msg_callingfn ("mem-pool", GF_LOG_ERROR, EINVAL,
-                                  LG_MSG_INVALID_ARG, "invalid argument");
-                return NULL;
+        unsigned int            i;
+        per_thread_pool_t       *pt_pool;
+
+        if (pool_list->poison) {
+                list_del (&pool_list->thr_list);
+                list_add (&pool_list->thr_list, &state->death_row);
+                return;
         }
-        padded_sizeof_type = sizeof_type + GF_MEM_POOL_PAD_BOUNDARY;
 
-        mem_pool = GF_CALLOC (sizeof (*mem_pool), 1, gf_common_mt_mem_pool);
-        if (!mem_pool)
-                return NULL;
+        if (state->n_cold_lists >= N_COLD_LISTS) {
+                return;
+        }
 
-        ret = gf_asprintf (&mem_pool->name, "%s:%s", THIS->name, name);
-        if (ret < 0)
-                return NULL;
+        (void) pthread_spin_lock (&pool_list->lock);
+        for (i = 0; i < NPOOLS; ++i) {
+                pt_pool = &pool_list->pools[i];
+                if (pt_pool->cold_list) {
+                        state->cold_lists[state->n_cold_lists++]
+                                = pt_pool->cold_list;
+                }
+                pt_pool->cold_list = pt_pool->hot_list;
+                pt_pool->hot_list = NULL;
+                if (state->n_cold_lists >= N_COLD_LISTS) {
+                        /* We'll just catch up on a future pass. */
+                        break;
+                }
+        }
+        (void) pthread_spin_unlock (&pool_list->lock);
+}
 
-        if (!mem_pool->name) {
-                GF_FREE (mem_pool);
-                return NULL;
+
+void
+free_obj_list (pooled_obj_hdr_t *victim)
+{
+        pooled_obj_hdr_t        *next;
+
+        while (victim) {
+                next = victim->next;
+                free (victim);
+                victim = next;
+                ++frees_to_system;
         }
+}
+
+void *
+pool_sweeper (void *arg)
+{
+        sweep_state_t           state;
+        per_thread_pool_list_t  *pool_list;
+        per_thread_pool_list_t  *next_pl;
+        per_thread_pool_t       *pt_pool;
+        unsigned int            i;
+        struct timeval          begin_time;
+        struct timeval          end_time;
+        struct timeval          elapsed;
 
+        /*
+         * This is all a bit inelegant, but the point is to avoid doing
+         * expensive things (like freeing thousands of objects) while holding a
+         * global lock.  Thus, we split each iteration into three passes, with
+         * only the first and fastest holding the lock.
+         */
+
+        for (;;) {
+                sleep (POOL_SWEEP_SECS);
+                INIT_LIST_HEAD (&state.death_row);
+                state.n_cold_lists = 0;
+
+                /* First pass: collect stuff that needs our attention. */
+                (void) gettimeofday (&begin_time, NULL);
+                (void) pthread_mutex_lock (&pool_lock);
+                list_for_each_entry_safe (pool_list, next_pl,
+                                          &pool_threads, thr_list) {
+                        collect_garbage (&state, pool_list);
+                }
+                (void) pthread_mutex_unlock (&pool_lock);
+                (void) gettimeofday (&end_time, NULL);
+                timersub (&end_time, &begin_time, &elapsed);
+                sweep_usecs += elapsed.tv_sec * 1000000 + elapsed.tv_usec;
+                sweep_times += 1;
+
+                /* Second pass: free dead pools. */
+                (void) pthread_mutex_lock (&pool_free_lock);
+                list_for_each_entry_safe (pool_list, next_pl,
+                                          &state.death_row, thr_list) {
+                        for (i = 0; i < NPOOLS; ++i) {
+                                pt_pool = &pool_list->pools[i];
+                                free_obj_list (pt_pool->cold_list);
+                                free_obj_list (pt_pool->hot_list);
+                                pt_pool->hot_list = pt_pool->cold_list = NULL;
+                        }
+                        list_del (&pool_list->thr_list);
+                        list_add (&pool_list->thr_list, &pool_free_threads);
+                }
+                (void) pthread_mutex_unlock (&pool_free_lock);
+
+                /* Third pass: free cold objects from live pools. */
+                for (i = 0; i < state.n_cold_lists; ++i) {
+                        free_obj_list (state.cold_lists[i]);
+                }
+        }
+}
+
+
+void
+pool_destructor (void *arg)
+{
+        per_thread_pool_list_t  *pool_list      = arg;
+
+        /* The pool-sweeper thread will take it from here. */
+        pool_list->poison = 1;
+}
+
+
+static __attribute__((constructor)) void
+mem_pools_preinit (void)
+{
 #if !defined(GF_DISABLE_MEMPOOL)
-        LOCK_INIT (&mem_pool->lock);
-        INIT_LIST_HEAD (&mem_pool->list);
-        INIT_LIST_HEAD (&mem_pool->global_list);
-
-        mem_pool->padded_sizeof_type = padded_sizeof_type;
-        mem_pool->real_sizeof_type = sizeof_type;
-
-#ifndef DEBUG
-        mem_pool->cold_count = count;
-        pool = GF_CALLOC (count, padded_sizeof_type, gf_common_mt_long);
-        if (!pool) {
-                GF_FREE (mem_pool->name);
-                GF_FREE (mem_pool);
-                return NULL;
+        unsigned int    i;
+
+        /* Use a pthread_key destructor to clean up when a thread exits. */
+        if (pthread_key_create (&pool_key, pool_destructor) != 0) {
+                gf_log ("mem-pool", GF_LOG_CRITICAL,
+                        "failed to initialize mem-pool key");
         }
 
-        for (i = 0; i < count; i++) {
-                list = pool + (i * (padded_sizeof_type));
-                INIT_LIST_HEAD (list);
-                list_add_tail (list, &mem_pool->list);
+        INIT_LIST_HEAD (&pool_threads);
+        INIT_LIST_HEAD (&pool_free_threads);
+
+        for (i = 0; i < NPOOLS; ++i) {
+                pools[i].power_of_two = POOL_SMALLEST + i;
         }
 
-        mem_pool->pool = pool;
-        mem_pool->pool_end = pool + (count * (padded_sizeof_type));
+        pool_list_size = sizeof (per_thread_pool_list_t)
+                       + sizeof (per_thread_pool_t) * (NPOOLS - 1);
 #endif
+}
 
-        /* add this pool to the global list */
-        ctx = THIS->ctx;
-        if (!ctx)
-                goto out;
+void
+mem_pools_init (void)
+{
+        pthread_t       kid;
 
-        LOCK (&ctx->lock);
-        {
-                list_add (&mem_pool->global_list, &ctx->mempool_list);
+        (void) pthread_create (&kid, NULL, pool_sweeper, NULL);
+        (void) pthread_detach (kid);
+}
+ 
+struct mem_pool *
+mem_pool_new_fn (unsigned long sizeof_type,
+                 unsigned long count, char *name)
+{
+        unsigned int            i;
+
+        if (!sizeof_type) {
+                gf_msg_callingfn ("mem-pool", GF_LOG_ERROR, EINVAL,
+                                  LG_MSG_INVALID_ARG, "invalid argument");
+                return NULL;
         }
-        UNLOCK (&ctx->lock);
 
-out:
-#endif /* GF_DISABLE_MEMPOOL */
-        return mem_pool;
+        for (i = 0; i < NPOOLS; ++i) {
+                if (sizeof_type <= AVAILABLE_SIZE(pools[i].power_of_two)) {
+                        return &pools[i];
+                }
+        }
+
+        gf_msg_callingfn ("mem-pool", GF_LOG_ERROR, EINVAL,
+                          LG_MSG_INVALID_ARG, "invalid argument");
+        return NULL;
 }
 
 void*
@@ -447,128 +580,133 @@ mem_get0 (struct mem_pool *mem_pool)
 
         ptr = mem_get(mem_pool);
 
-        if (ptr)
-                memset(ptr, 0, mem_pool->real_sizeof_type);
+        if (ptr) {
+                memset (ptr, 0, AVAILABLE_SIZE(mem_pool->power_of_two));
+        }
 
         return ptr;
 }
 
-void *
-mem_get (struct mem_pool *mem_pool)
+
+per_thread_pool_list_t *
+mem_get_pool_list (void)
 {
-#ifdef GF_DISABLE_MEMPOOL
-          return GF_CALLOC (1, mem_pool->real_sizeof_type,
-                                gf_common_mt_mem_pool);
-#else
-        struct list_head *list = NULL;
-        void             *ptr = NULL;
-        int             *in_use = NULL;
-        struct mem_pool **pool_ptr = NULL;
+        per_thread_pool_list_t  *pool_list;
+        unsigned int            i;
 
-        if (!mem_pool) {
-                gf_msg_callingfn ("mem-pool", GF_LOG_ERROR, EINVAL,
-                                  LG_MSG_INVALID_ARG, "invalid argument");
-                return NULL;
+        pool_list = pthread_getspecific (pool_key);
+        if (pool_list) {
+                return pool_list;
         }
 
-        LOCK (&mem_pool->lock);
-        {
-                mem_pool->alloc_count++;
-                if (mem_pool->cold_count) {
-                        list = mem_pool->list.next;
-                        list_del (list);
+        (void) pthread_mutex_lock (&pool_free_lock);
+        if (!list_empty (&pool_free_threads)) {
+                pool_list = list_entry (pool_free_threads.next,
+                                        per_thread_pool_list_t, thr_list);
+                list_del (&pool_list->thr_list);
+        }
+        (void) pthread_mutex_unlock (&pool_free_lock);
 
-                        mem_pool->hot_count++;
-                        mem_pool->cold_count--;
+        if (!pool_list) {
+                pool_list = GF_CALLOC (pool_list_size, 1,
+                                       gf_common_mt_mem_pool);
+                if (!pool_list) {
+                        return NULL;
+                }
 
-                        if (mem_pool->max_alloc < mem_pool->hot_count)
-                                mem_pool->max_alloc = mem_pool->hot_count;
+                INIT_LIST_HEAD (&pool_list->thr_list);
+                (void) pthread_spin_init (&pool_list->lock,
+                                          PTHREAD_PROCESS_PRIVATE);
+                for (i = 0; i < NPOOLS; ++i) {
+                        pool_list->pools[i].parent = &pools[i];
+                        pool_list->pools[i].hot_list = NULL;
+                        pool_list->pools[i].cold_list = NULL;
+                }
+        }
 
-                        ptr = list;
-                        in_use = (ptr + GF_MEM_POOL_LIST_BOUNDARY +
-                                  GF_MEM_POOL_PTR);
-                        *in_use = 1;
+        (void) pthread_mutex_lock (&pool_lock);
+        pool_list->poison = 0;
+        list_add (&pool_list->thr_list, &pool_threads);
+        (void) pthread_mutex_unlock (&pool_lock);
 
-                        goto fwd_addr_out;
-                }
+        (void) pthread_setspecific (pool_key, pool_list);
+        return pool_list;
+}
+
+pooled_obj_hdr_t *
+mem_get_from_pool (per_thread_pool_t *pt_pool)
+{
+        pooled_obj_hdr_t        *retval;
 
-                /* This is a problem area. If we've run out of
-                 * chunks in our slab above, we need to allocate
-                 * enough memory to service this request.
-                 * The problem is, these individual chunks will fail
-                 * the first address range check in __is_member. Now, since
-                 * we're not allocating a full second slab, we wont have
-                 * enough info perform the range check in __is_member.
-                 *
-                 * I am working around this by performing a regular allocation
-                 * , just the way the caller would've done when not using the
-                 * mem-pool. That also means, we're not padding the size with
-                 * the list_head structure because, this will not be added to
-                 * the list of chunks that belong to the mem-pool allocated
-                 * initially.
-                 *
-                 * This is the best we can do without adding functionality for
-                 * managing multiple slabs. That does not interest us at present
-                 * because it is too much work knowing that a better slab
-                 * allocator is coming RSN.
-                 */
-                mem_pool->pool_misses++;
-                mem_pool->curr_stdalloc++;
-                if (mem_pool->max_stdalloc < mem_pool->curr_stdalloc)
-                        mem_pool->max_stdalloc = mem_pool->curr_stdalloc;
-                ptr = GF_CALLOC (1, mem_pool->padded_sizeof_type,
-                                 gf_common_mt_mem_pool);
-
-                /* Memory coming from the heap need not be transformed from a
-                 * chunkhead to a usable pointer since it is not coming from
-                 * the pool.
-                 */
+        retval = pt_pool->hot_list;
+        if (retval) {
+                (void) __sync_fetch_and_add (&pt_pool->parent->allocs_hot, 1);
+                pt_pool->hot_list = retval->next;
+                return retval;
         }
-fwd_addr_out:
-        pool_ptr = mem_pool_from_ptr (ptr);
-        *pool_ptr = (struct mem_pool *)mem_pool;
-        ptr = mem_pool_chunkhead2ptr (ptr);
-        UNLOCK (&mem_pool->lock);
 
-        return ptr;
-#endif /* GF_DISABLE_MEMPOOL */
+        retval = pt_pool->cold_list;
+        if (retval) {
+                (void) __sync_fetch_and_add (&pt_pool->parent->allocs_cold, 1);
+                pt_pool->cold_list = retval->next;
+                return retval;
+        }
+
+        (void) __sync_fetch_and_add (&pt_pool->parent->allocs_stdc, 1);
+        return malloc (1 << pt_pool->parent->power_of_two);
 }
 
 
-#if !defined(GF_DISABLE_MEMPOOL)
-static int
-__is_member (struct mem_pool *pool, void *ptr)
+void *
+mem_get (struct mem_pool *mem_pool)
 {
-        if (!pool || !ptr) {
+#if defined(GF_DISABLE_MEMPOOL)
+        return GF_CALLOC (1, mem_pool->real_sizeof_type,
+                          gf_common_mt_mem_pool);
+#else
+        per_thread_pool_list_t  *pool_list;
+        per_thread_pool_t       *pt_pool;
+        pooled_obj_hdr_t        *retval;
+
+        if (!mem_pool) {
                 gf_msg_callingfn ("mem-pool", GF_LOG_ERROR, EINVAL,
                                   LG_MSG_INVALID_ARG, "invalid argument");
-                return -1;
+                return NULL;
         }
 
-        if (ptr < pool->pool || ptr >= pool->pool_end)
-                return 0;
+        pool_list = mem_get_pool_list ();
+        if (!pool_list || pool_list->poison) {
+                return NULL;
+        }
 
-        if ((mem_pool_ptr2chunkhead (ptr) - pool->pool)
-            % pool->padded_sizeof_type)
-                return -1;
+        (void) pthread_spin_lock (&pool_list->lock);
+        pt_pool = &pool_list->pools[mem_pool->power_of_two-POOL_SMALLEST];
+        retval = mem_get_from_pool (pt_pool);
+        (void) pthread_spin_unlock (&pool_list->lock);
+
+        if (!retval) {
+                return NULL;
+        }
 
-        return 1;
+        retval->magic = GF_MEM_HEADER_MAGIC;
+        retval->next = NULL;
+        retval->pool_list = pool_list;;
+        retval->power_of_two = mem_pool->power_of_two;
+
+        return retval + 1;
 }
-#endif
+#endif /* GF_DISABLE_MEMPOOL */
 
 
 void
 mem_put (void *ptr)
 {
-#ifdef GF_DISABLE_MEMPOOL
+#if defined(GF_DISABLE_MEMPOOL)
         GF_FREE (ptr);
-        return;
 #else
-        struct list_head *list = NULL;
-        int    *in_use = NULL;
-        void   *head = NULL;
-        struct mem_pool **tmp = NULL;
-        struct mem_pool *pool = NULL;
+        pooled_obj_hdr_t        *hdr;
+        per_thread_pool_list_t  *pool_list;
+        per_thread_pool_t       *pt_pool;
 
         if (!ptr) {
                 gf_msg_callingfn ("mem-pool", GF_LOG_ERROR, EINVAL,
@@ -576,71 +714,20 @@ mem_put (void *ptr)
                 return;
         }
 
-        list = head = mem_pool_ptr2chunkhead (ptr);
-        tmp = mem_pool_from_ptr (head);
-        if (!tmp) {
-                gf_msg_callingfn ("mem-pool", GF_LOG_ERROR, 0,
-                                  LG_MSG_PTR_HEADER_CORRUPTED,
-                                  "ptr header is corrupted");
-                return;
-        }
-
-        pool = *tmp;
-        if (!pool) {
-                gf_msg_callingfn ("mem-pool", GF_LOG_ERROR, 0,
-                                  LG_MSG_MEMPOOL_PTR_NULL,
-                                  "mem-pool ptr is NULL");
+        hdr = ((pooled_obj_hdr_t *)ptr) - 1;
+        if (hdr->magic != GF_MEM_HEADER_MAGIC) {
+                /* Not one of ours; don't touch it. */
                 return;
         }
-        LOCK (&pool->lock);
-        {
-
-                switch (__is_member (pool, ptr))
-                {
-                case 1:
-                        in_use = (head + GF_MEM_POOL_LIST_BOUNDARY +
-                                  GF_MEM_POOL_PTR);
-                        if (!is_mem_chunk_in_use(in_use)) {
-                                gf_msg_callingfn ("mem-pool", GF_LOG_CRITICAL,
-                                                  0,
-                                                  LG_MSG_MEMPOOL_INVALID_FREE,
-                                                  "mem_put called on freed ptr"
-                                                  " %p of mem pool %p", ptr,
-                                                  pool);
-                                break;
-                        }
-                        pool->hot_count--;
-                        pool->cold_count++;
-                        *in_use = 0;
-                        list_add (list, &pool->list);
-                        break;
-                case -1:
-                        /* For some reason, the address given is within
-                         * the address range of the mem-pool but does not align
-                         * with the expected start of a chunk that includes
-                         * the list headers also. Sounds like a problem in
-                         * layers of clouds up above us. ;)
-                         */
-                        abort ();
-                        break;
-                case 0:
-                        /* The address is outside the range of the mem-pool. We
-                         * assume here that this address was allocated at a
-                         * point when the mem-pool was out of chunks in mem_get
-                         * or the programmer has made a mistake by calling the
-                         * wrong de-allocation interface. We do
-                         * not have enough info to distinguish between the two
-                         * situations.
-                         */
-                        pool->curr_stdalloc--;
-                        GF_FREE (list);
-                        break;
-                default:
-                        /* log error */
-                        break;
-                }
-        }
-        UNLOCK (&pool->lock);
+        pool_list = hdr->pool_list;
+        pt_pool = &pool_list->pools[hdr->power_of_two-POOL_SMALLEST];
+
+        (void) pthread_spin_lock (&pool_list->lock);
+        hdr->magic = GF_MEM_INVALID_MAGIC;
+        hdr->next = pt_pool->hot_list;
+        pt_pool->hot_list = hdr;
+        (void) __sync_fetch_and_add (&pt_pool->parent->frees_to_list, 1);
+        (void) pthread_spin_unlock (&pool_list->lock);
 #endif /* GF_DISABLE_MEMPOOL */
 }
 
@@ -650,16 +737,11 @@ mem_pool_destroy (struct mem_pool *pool)
         if (!pool)
                 return;
 
-        gf_msg (THIS->name, GF_LOG_INFO, 0, LG_MSG_MEM_POOL_DESTROY, "size=%lu "
-                "max=%d total=%"PRIu64, pool->padded_sizeof_type,
-                pool->max_alloc, pool->alloc_count);
-
-#if !defined(GF_DISABLE_MEMPOOL)
-        list_del (&pool->global_list);
-
-        LOCK_DESTROY (&pool->lock);
-        GF_FREE (pool->name);
-        GF_FREE (pool->pool);
-#endif
-        GF_FREE (pool);
+        /*
+         * Pools are now permanent, so this does nothing.  Yes, this means we
+         * can keep allocating from a pool after calling mem_destroy on it, but
+         * that's kind of OK.  All of the objects *in* the pool will eventually
+         * be freed via the pool-sweeper thread, and this way we don't have to
+         * add a lot of reference-counting complexity.
+         */
 }
diff --git a/libglusterfs/src/mem-pool.h b/libglusterfs/src/mem-pool.h
index 6cff7be94f4..0dc186341b2 100644
--- a/libglusterfs/src/mem-pool.h
+++ b/libglusterfs/src/mem-pool.h
@@ -209,24 +209,61 @@ out:
         return dup_mem;
 }
 
+typedef struct pooled_obj_hdr {
+        unsigned long                   magic;
+        struct pooled_obj_hdr           *next;
+        struct per_thread_pool_list     *pool_list;
+        unsigned int                    power_of_two;
+} pooled_obj_hdr_t;
+
+#define AVAILABLE_SIZE(p2)      ((1 << (p2)) - sizeof(pooled_obj_hdr_t))
+
+typedef struct per_thread_pool {
+        /* This never changes, so doesn't need a lock. */
+        struct mem_pool         *parent;
+        /* Everything else is protected by our own lock. */
+        pooled_obj_hdr_t        *hot_list;
+        pooled_obj_hdr_t        *cold_list;
+} per_thread_pool_t;
+
+typedef struct per_thread_pool_list {
+        /*
+         * These first two members are protected by the global pool lock.  When
+         * a thread first tries to use any pool, we create one of these.  We
+         * link it into the global list using thr_list so the pool-sweeper
+         * thread can find it, and use pthread_setspecific so this thread can
+         * find it.  When the per-thread destructor runs, we "poison" the pool
+         * list to prevent further allocations.  This also signals to the
+         * pool-sweeper thread that the list should be detached and freed after
+         * the next time it's swept.
+         */
+        struct list_head        thr_list;
+        unsigned int            poison;
+        /*
+         * There's really more than one pool, but the actual number is hidden
+         * in the implementation code so we just make it a single-element array
+         * here.
+         */
+        pthread_spinlock_t      lock;
+        per_thread_pool_t       pools[1];
+} per_thread_pool_list_t;
+
 struct mem_pool {
-        struct list_head  list;
-        int               hot_count;
-        int               cold_count;
-        gf_lock_t         lock;
-        unsigned long     padded_sizeof_type;
-        void             *pool;
-        void             *pool_end;
-        int               real_sizeof_type;
-        uint64_t          alloc_count;
-        uint64_t          pool_misses;
-        int               max_alloc;
-        int               curr_stdalloc;
-        int               max_stdalloc;
-        char             *name;
-        struct list_head  global_list;
+        unsigned int            power_of_two;
+        /*
+         * Updates to these are *not* protected by a global lock, so races
+         * could occur and the numbers might be slightly off.  Don't expect
+         * them to line up exactly.  It's the general trends that matter, and
+         * it's not worth the locked-bus-cycle overhead to make these precise.
+         */
+        unsigned long           allocs_hot;
+        unsigned long           allocs_cold;
+        unsigned long           allocs_stdc;
+        unsigned long           frees_to_list;
 };
 
+void mem_pools_init (void);
+
 struct mem_pool *
 mem_pool_new_fn (unsigned long sizeof_type, unsigned long count, char *name);
 
diff --git a/libglusterfs/src/statedump.c b/libglusterfs/src/statedump.c
index a292857fd10..bb8043a869f 100644
--- a/libglusterfs/src/statedump.c
+++ b/libglusterfs/src/statedump.c
@@ -376,6 +376,7 @@ gf_proc_dump_mem_info_to_dict (dict_t *dict)
 void
 gf_proc_dump_mempool_info (glusterfs_ctx_t *ctx)
 {
+#if defined(OLD_MEM_POOLS)
         struct mem_pool *pool = NULL;
 
         gf_proc_dump_add_section ("mempool");
@@ -394,11 +395,13 @@ gf_proc_dump_mempool_info (glusterfs_ctx_t *ctx)
                 gf_proc_dump_write ("cur-stdalloc", "%d", pool->curr_stdalloc);
                 gf_proc_dump_write ("max-stdalloc", "%d", pool->max_stdalloc);
         }
+#endif
 }
 
 void
 gf_proc_dump_mempool_info_to_dict (glusterfs_ctx_t *ctx, dict_t *dict)
 {
+#if defined(OLD_MEM_POOLS)
         struct mem_pool *pool = NULL;
         char            key[GF_DUMP_MAX_BUF_LEN] = {0,};
         int             count = 0;
@@ -458,8 +461,7 @@ gf_proc_dump_mempool_info_to_dict (glusterfs_ctx_t *ctx, dict_t *dict)
                 count++;
         }
         ret = dict_set_int32 (dict, "mempool-count", count);
-
-        return;
+#endif
 }
 
 void gf_proc_dump_latency_info (xlator_t *xl);
diff --git a/tests/basic/quota-anon-fd-nfs.t b/tests/basic/quota-anon-fd-nfs.t
index c6b01553b02..ea07b529c5a 100755
--- a/tests/basic/quota-anon-fd-nfs.t
+++ b/tests/basic/quota-anon-fd-nfs.t
@@ -97,6 +97,15 @@ $CLI volume statedump $V0 all
 
 EXPECT_WITHIN $UMOUNT_TIMEOUT "Y" force_umount $N0
 
+# This is ugly, but there seems to be a latent race between other actions and
+# stopping the volume.  The visible symptom is that "umount -l" (run from
+# gf_umount_lazy in glusterd) hangs.  This happens pretty consistently with the
+# new mem-pool code, though it's not really anything to do with memory pools -
+# just with changed timing.  Adding the sleep here makes it work consistently.
+#
+# If anyone else wants to debug the race condition, feel free.
+sleep 3
+
 TEST $CLI volume stop $V0
 EXPECT "1" get_aux