From 6459a2bf47b46b84910b2d837be9e43451f3a918 Mon Sep 17 00:00:00 2001
From: Konstantin Osipov <kostja@tarantool.org>
Date: Sun, 1 Feb 2015 19:41:45 +0300
Subject: [PATCH] gh-695: slab_alloc_maximal and variable slab sizes
Avoid overwriting tuple data with information necessary for
smfree(). Only use a single pointer to store delayed free context.
Add a test case.
Minor cleanups/comments/renames.
---
src/box/tuple.cc | 41 +++++++++--------
src/box/tuple.h | 6 +++
src/lib/small/lifo.h | 6 +++
src/lib/small/mempool.c | 2 +
src/lib/small/mempool.h | 29 +++++++-----
src/lib/small/slab_cache.c | 12 +++--
src/lib/small/small.c | 91 ++++++++++++++++++--------------------
src/lib/small/small.h | 2 -
test/app/snapshot.result | 1 +
test/app/snapshot.test.lua | 62 ++++++++++++++++++++++++++
10 files changed, 169 insertions(+), 83 deletions(-)
create mode 100644 test/app/snapshot.result
create mode 100755 test/app/snapshot.test.lua
diff --git a/src/box/tuple.cc b/src/box/tuple.cc
index 6fc4b2b00f..ab63e89271 100644
--- a/src/box/tuple.cc
+++ b/src/box/tuple.cc
@@ -50,12 +50,14 @@ struct slab_arena memtx_arena;
static struct slab_cache memtx_slab_cache;
struct small_alloc memtx_alloc;
-/* Lowest allowed objsize_min setting of tuple arena */
-const uint32_t MIN_TUPLE_SIZE_NOT_LESS_THAN = 16;
-/* Lowest allowed objsize_max setting of tuple arena */
-const uint32_t MAX_TUPLE_SIZE_NOT_LESS_THAN = 16 * 1024;
-/* Lowest allowed slab size for tuple arena */
-const size_t SLAB_SIZE_NOT_LESS_THAN = 1024 * 1024;
+enum {
+ /** Lowest allowed slab_alloc_minimal */
+ OBJSIZE_MIN = 16,
+ /** Lowest allowed slab_alloc_maximal */
+ OBJSIZE_MAX_MIN = 16 * 1024,
+ /** Lowest allowed slab size, for mmapped slabs */
+ SLAB_SIZE_MIN = 1024 * 1024
+};
/** Extract all available type info from keys. */
void
@@ -534,18 +536,19 @@ tuple_init(float tuple_arena_max_size, uint32_t objsize_min,
tuple_format_ref(tuple_format_ber, 1);
/* Apply lowest allowed objsize bounds */
- if (objsize_min < MIN_TUPLE_SIZE_NOT_LESS_THAN)
- objsize_min = MIN_TUPLE_SIZE_NOT_LESS_THAN;
- if (objsize_max < MAX_TUPLE_SIZE_NOT_LESS_THAN)
- objsize_max = MAX_TUPLE_SIZE_NOT_LESS_THAN;
+ if (objsize_min < OBJSIZE_MIN)
+ objsize_min = OBJSIZE_MIN;
+ if (objsize_max < OBJSIZE_MAX_MIN)
+ objsize_max = OBJSIZE_MAX_MIN;
/* Calculate slab size for tuple arena */
- size_t max_slab_size = small_round(objsize_max * 4);
- if (max_slab_size < SLAB_SIZE_NOT_LESS_THAN)
- max_slab_size = SLAB_SIZE_NOT_LESS_THAN;
+ size_t slab_size = small_round(objsize_max * 4);
+ if (slab_size < SLAB_SIZE_MIN)
+ slab_size = SLAB_SIZE_MIN;
- size_t max_size = tuple_arena_max_size * 1024 * 1024 * 1024;
- quota_init(&memtx_quota, max_size);
+ /** Preallocate entire quota. */
+ size_t prealloc = tuple_arena_max_size * 1024 * 1024 * 1024;
+ quota_init(&memtx_quota, prealloc);
int flags;
if (access("/proc/user_beancounters", F_OK) == 0) {
@@ -554,20 +557,20 @@ tuple_init(float tuple_arena_max_size, uint32_t objsize_min,
flags = MAP_PRIVATE;
} else {
say_info("mapping %zu bytes for a shared arena...",
- max_size);
+ prealloc);
flags = MAP_SHARED;
}
if (slab_arena_create(&memtx_arena, &memtx_quota,
- max_size, max_slab_size, flags)) {
+ prealloc, slab_size, flags)) {
if (ENOMEM == errno) {
panic("failed to preallocate %zu bytes: "
"Cannot allocate memory, check option "
"'slab_alloc_arena' in box.cfg(..)",
- max_size);
+ prealloc);
} else {
panic_syserror("failed to preallocate %zu bytes",
- max_size);
+ prealloc);
}
}
slab_cache_create(&memtx_slab_cache, &memtx_arena);
diff --git a/src/box/tuple.h b/src/box/tuple.h
index 815e191874..666f190a70 100644
--- a/src/box/tuple.h
+++ b/src/box/tuple.h
@@ -139,6 +139,12 @@ tuple_format_ref(struct tuple_format *format, int count)
*/
struct tuple
{
+ /*
+ * sic: the header of the tuple is used
+ * to store a free list pointer in smfree_delayed.
+ * Please don't change it without understanding
+ * how smfree_delayed and snapshotting COW works.
+ */
/** snapshot generation version */
uint32_t version;
/** reference counter */
diff --git a/src/lib/small/lifo.h b/src/lib/small/lifo.h
index 9c322b6c47..6641e396ec 100644
--- a/src/lib/small/lifo.h
+++ b/src/lib/small/lifo.h
@@ -65,6 +65,12 @@ lifo_pop(struct lifo *head)
return elem;
}
+static inline void *
+lifo_peek(struct lifo *head)
+{
+ return head->next;
+}
+
static inline bool
lifo_is_empty(struct lifo *head)
{
diff --git a/src/lib/small/mempool.c b/src/lib/small/mempool.c
index 83a1dc3cd7..5570c674f2 100644
--- a/src/lib/small/mempool.c
+++ b/src/lib/small/mempool.c
@@ -125,6 +125,8 @@ mempool_create_with_order(struct mempool *pool, struct slab_cache *cache,
uint32_t objsize, uint8_t order)
{
assert(order <= cache->order_max);
+ lifo_init(&pool->link);
+ lifo_init(&pool->delayed);
pool->cache = cache;
slab_list_create(&pool->slabs);
mslab_tree_new(&pool->free_slabs);
diff --git a/src/lib/small/mempool.h b/src/lib/small/mempool.h
index dab88d42fd..5f2737957e 100644
--- a/src/lib/small/mempool.h
+++ b/src/lib/small/mempool.h
@@ -33,6 +33,7 @@
#include <inttypes.h>
#include <string.h>
#include "small/slab_cache.h"
+#include "small/lifo.h"
#define RB_COMPACT 1
#include "third_party/rb.h"
@@ -86,10 +87,11 @@ struct mslab {
};
/**
- * Mempool will try to allocate blocks large enough to have overhead
- * less than specified below
+ * Mempool will try to allocate blocks large enough to ensure
+ * the overhead from internal fragmentation is less than the
+ * specified below.
*/
-static const double expected_overhead_max = 0.01;
+static const double OVERHEAD_RATIO = 0.01;
static inline uint32_t
mslab_sizeof()
@@ -114,6 +116,14 @@ typedef rb_tree(struct mslab) mslab_tree_t;
/** A memory pool. */
struct mempool
{
+ /**
+ * A link in delayed free list of pools. Must be the first
+ * member in the struct.
+ * @sa smfree_delayed().
+ */
+ struct lifo link;
+ /** List of pointers for delayed free. */
+ struct lifo delayed;
/** The source of empty slabs. */
struct slab_cache *cache;
/** All slabs. */
@@ -191,18 +201,17 @@ static inline void
mempool_create(struct mempool *pool, struct slab_cache *cache,
uint32_t objsize)
{
- size_t expected_loss = objsize > sizeof(struct mslab)
- ? objsize : sizeof(struct mslab);
- size_t slab_size_min = (size_t)(expected_loss / expected_overhead_max);
- if (slab_size_min > cache->arena->slab_size)
- slab_size_min = cache->arena->slab_size;
-
+ size_t overhead = (objsize > sizeof(struct mslab) ?
+ objsize : sizeof(struct mslab));
+ size_t slab_size = (size_t) (overhead / OVERHEAD_RATIO);
+ if (slab_size > cache->arena->slab_size)
+ slab_size = cache->arena->slab_size;
/*
* Calculate the amount of usable space in a slab.
* @note: this asserts that slab_size_min is less than
* SLAB_ORDER_MAX.
*/
- uint8_t order = slab_order(cache, slab_size_min);
+ uint8_t order = slab_order(cache, slab_size);
assert(order <= cache->order_max);
return mempool_create_with_order(pool, cache, objsize, order);
}
diff --git a/src/lib/small/slab_cache.c b/src/lib/small/slab_cache.c
index d837fe462d..c64eb99e5d 100644
--- a/src/lib/small/slab_cache.c
+++ b/src/lib/small/slab_cache.c
@@ -162,10 +162,14 @@ void
slab_cache_create(struct slab_cache *cache, struct slab_arena *arena)
{
cache->arena = arena;
-
- long lowest_order0_size = small_round(sysconf(_SC_PAGESIZE));
- assert(arena->slab_size >= lowest_order0_size);
- cache->order_max = small_lb(arena->slab_size / lowest_order0_size);
+ /*
+ * We have a fixed number of orders (ORDER_MAX); calculate
+ * the size of buddies in the smallest order, given the size
+ * of the slab size in the slab arena.
+ */
+ long min_order0_size = sysconf(_SC_PAGESIZE);
+ assert(arena->slab_size >= min_order0_size);
+ cache->order_max = small_lb(arena->slab_size / min_order0_size);
if (cache->order_max > ORDER_MAX - 1)
cache->order_max = ORDER_MAX - 1;
diff --git a/src/lib/small/small.c b/src/lib/small/small.c
index c665020e2b..b065d52470 100644
--- a/src/lib/small/small.c
+++ b/src/lib/small/small.c
@@ -42,15 +42,6 @@ enum {
STEP_SIZE_LB = 3,
};
-/**
- * Extended lifo struct for store object size in addition to pointer
- * to the next object in the list
- */
-struct small_lifo_ext {
- struct lifo base;
- size_t size;
-};
-
rb_proto(, factor_tree_, factor_tree_t, struct factor_pool)
/** Used for search in the tree. */
@@ -109,15 +100,10 @@ small_alloc_create(struct small_alloc *alloc, struct slab_cache *cache,
{
alloc->cache = cache;
/* Align sizes. */
- objsize_min = small_align(objsize_min, sizeof(STEP_SIZE));
- /* An object must be large enough to contain struct small_lifo_ext */
- if (objsize_min < sizeof(struct small_lifo_ext))
- objsize_min = small_align(sizeof(struct small_lifo_ext),
- sizeof(STEP_SIZE));
- alloc->slab_order = cache->order_max;
+ objsize_min = small_align(objsize_min, STEP_SIZE);
/* Make sure at least 4 largest objects can fit in a slab. */
alloc->objsize_max =
- mempool_objsize_max(slab_order_size(cache, alloc->slab_order));
+ mempool_objsize_max(slab_order_size(cache, cache->order_max));
assert(alloc->objsize_max > objsize_min + STEP_POOL_MAX * STEP_SIZE);
struct mempool *step_pool;
@@ -180,14 +166,18 @@ smfree_batch(struct small_alloc *alloc)
return;
const int BATCH = 100;
+ struct mempool *pool = lifo_peek(&alloc->delayed);
for (int i = 0; i < BATCH; i++) {
- void *item = lifo_pop(&alloc->delayed);
- if (item == NULL)
- break;
- struct small_lifo_ext *ext = (struct small_lifo_ext *) item;
- size_t size = ext->size;
- smfree(alloc, item, size);
+ void *item = lifo_pop(&pool->delayed);
+ if (item == NULL) {
+ (void) lifo_pop(&alloc->delayed);
+ pool = lifo_peek(&alloc->delayed);
+ if (pool == NULL)
+ break;
+ continue;
+ }
+ mempool_free(pool, item);
}
}
@@ -256,33 +246,20 @@ small_recycle_pool(struct small_alloc *alloc, struct mempool *pool)
}
}
-/** Free memory chunk allocated by the small allocator. */
-/**
- * Free a small objects.
- *
- * This boils down to finding the object's mempool and delegating
- * to mempool_free().
- *
- * If the pool becomes completely empty, and it's a factored pool,
- * and the factored pool's cache is empty, put back the empty
- * factored pool into the factored pool cache.
- */
-void
-smfree(struct small_alloc *alloc, void *ptr, size_t size)
+static inline struct mempool *
+mempool_find(struct small_alloc *alloc, size_t size)
{
struct mempool *pool;
if (size <= alloc->step_pool_objsize_max) {
/* Allocated in a stepped pool. */
- int idx;
- if (size <= alloc->step_pools[0].objsize)
- idx = 0;
- else {
- idx = (size - alloc->step_pools[0].objsize
- + STEP_SIZE - 1) >> STEP_SIZE_LB;
+ if (size <= alloc->step_pools[0].objsize) {
+ pool = &alloc->step_pools[0];
+ } else {
+ int idx = (size - alloc->step_pools[0].objsize
+ + STEP_SIZE - 1) >> STEP_SIZE_LB;
+ pool = &alloc->step_pools[idx];
+ assert(size + STEP_SIZE > pool->objsize);
}
- pool = &alloc->step_pools[idx];
- assert(size <= pool->objsize &&
- (size + STEP_SIZE > pool->objsize || idx == 0));
} else {
/* Allocated in a factor pool. */
struct factor_pool pattern;
@@ -294,6 +271,24 @@ smfree(struct small_alloc *alloc, void *ptr, size_t size)
pool = &upper_bound->pool;
}
assert(size <= pool->objsize);
+ return pool;
+}
+
+/** Free memory chunk allocated by the small allocator. */
+/**
+ * Free a small object.
+ *
+ * This boils down to finding the object's mempool and delegating
+ * to mempool_free().
+ *
+ * If the pool becomes completely empty, and it's a factored pool,
+ * and the factored pool's cache is empty, put back the empty
+ * factored pool into the factored pool cache.
+ */
+void
+smfree(struct small_alloc *alloc, void *ptr, size_t size)
+{
+ struct mempool *pool = mempool_find(alloc, size);
mempool_free(pool, ptr);
if (mempool_used(pool) == 0)
@@ -308,11 +303,11 @@ smfree(struct small_alloc *alloc, void *ptr, size_t size)
void
smfree_delayed(struct small_alloc *alloc, void *ptr, size_t size)
{
- assert(size >= sizeof(struct small_lifo_ext));
if (alloc->is_delayed_free_mode && ptr) {
- struct small_lifo_ext *ext = (struct small_lifo_ext *)ptr;
- ext->size = size;
- lifo_push(&alloc->delayed, ptr);
+ struct mempool *pool = mempool_find(alloc, size);
+ if (lifo_is_empty(&pool->delayed))
+ lifo_push(&alloc->delayed, &pool->link);
+ lifo_push(&pool->delayed, ptr);
} else {
smfree(alloc, ptr, size);
}
diff --git a/src/lib/small/small.h b/src/lib/small/small.h
index e400b613bf..1b8e5ffd5d 100644
--- a/src/lib/small/small.h
+++ b/src/lib/small/small.h
@@ -154,8 +154,6 @@ struct small_alloc {
*/
float factor;
uint32_t objsize_max;
- /** All slabs in all mempools have the same order. */
- uint8_t slab_order;
/**
* If true, smfree_delayed puts items to delayed list.
*/
diff --git a/test/app/snapshot.result b/test/app/snapshot.result
new file mode 100644
index 0000000000..9766475a41
--- /dev/null
+++ b/test/app/snapshot.result
@@ -0,0 +1 @@
+ok
diff --git a/test/app/snapshot.test.lua b/test/app/snapshot.test.lua
new file mode 100755
index 0000000000..d904b6b7c1
--- /dev/null
+++ b/test/app/snapshot.test.lua
@@ -0,0 +1,62 @@
+#!/usr/bin/env tarantool
+math = require('math')
+fiber = require('fiber')
+
+--
+-- Check that Tarantool creates ADMIN session for #! script
+--
+function noise()
+ fiber.name('noise-'..fiber.id())
+ while true do
+ if box.space.test:len() < 300000 then
+ local value = string.rep('a', math.random(255)+1)
+ box.space.test:auto_increment{fiber.time64(), value}
+ end
+ fiber.sleep(0)
+ end
+end
+
+function purge()
+ fiber.name('purge-'..fiber.id())
+ while true do
+ local min = box.space.test.index.primary:min()
+ if min ~= nil then
+ box.space.test:delete{min[1]}
+ end
+ fiber.sleep(0)
+ end
+end
+
+function snapshot(lsn)
+ fiber.name('snapshot')
+ while true do
+ local new_lsn = box.info.server.lsn
+ if new_lsn ~= lsn then
+ lsn = new_lsn;
+ box.snapshot()
+ end
+ fiber.sleep(0.001)
+ end
+end
+box.cfg{logger="tarantool.log", slab_alloc_arena=0.1, rows_per_wal=5000}
+
+if box.space.test == nil then
+ box.schema.space.create('test')
+ box.space.test:create_index('primary')
+end
+
+require('console').listen(3303)
+
+fiber.create(noise)
+fiber.create(purge)
+fiber.create(noise)
+fiber.create(purge)
+fiber.create(noise)
+fiber.create(purge)
+fiber.create(noise)
+fiber.create(purge)
+fiber.create(snapshot, box.info.server.lsn)
+
+fiber.sleep(0.3)
+print('ok')
+os.exit(0)
--
GitLab