/* refcount == 0 is safe since nobody can incr while cache_lock is held.
* refcount != 0 is impossible since flags/etc can be modified in other
* threads. instead, note we found a busy one and bail. logic in do_item_get
* will prevent busy items from continuing to be busy
*/
static int slab_rebalance_move(void) {
slabclass_t *s_cls;
int x;
int was_busy = 0;
int refcount = 0;
enum move_status status = MOVE_PASS;
pthread_mutex_lock(&cache_lock);
pthread_mutex_lock(&slabs_lock);
s_cls = &slabclass[slab_rebal.s_clsid];
for (x = 0; x < slab_bulk_check; x++) {
item *it = slab_rebal.slab_pos;
status = MOVE_PASS;
if (it->slabs_clsid != 255) {
void *hold_lock = NULL;
uint32_t hv = hash(ITEM_key(it), it->nkey, 0);
if ((hold_lock = item_trylock(hv)) == NULL) {
status = MOVE_LOCKED;
} else {
refcount = refcount_incr(&it->refcount);
if (refcount == 1) { /* item is unlinked, unused */
if (it->it_flags & ITEM_SLABBED) {
/* remove from slab freelist */
if (s_cls->slots == it) {
s_cls->slots = it->next;
}
if (it->next) it->next->prev = it->prev;
if (it->prev) it->prev->next = it->next;
s_cls->sl_curr--;
status = MOVE_DONE;
} else {
status = MOVE_BUSY;
}
} else if (refcount == 2) { /* item is linked but not busy */
if ((it->it_flags & ITEM_LINKED) != 0) {
do_item_unlink_nolock(it, hv);
status = MOVE_DONE;
} else {
/* refcount == 1 + !ITEM_LINKED means the item is being
* uploaded to, or was just unlinked but hasn't been freed
* yet. Let it bleed off on its own and try again later */
status = MOVE_BUSY;
}
} else {
if (settings.verbose > 2) {
fprintf(stderr, "Slab reassign hit a busy item: refcount: %d (%d -> %d)\n",
it->refcount, slab_rebal.s_clsid, slab_rebal.d_clsid);
}
status = MOVE_BUSY;
}
item_trylock_unlock(hold_lock);
}
}
switch (status) {
case MOVE_DONE:
it->refcount = 0;
it->it_flags = 0;
it->slabs_clsid = 255;
break;
case MOVE_BUSY:
refcount_decr(&it->refcount);
case MOVE_LOCKED:
slab_rebal.busy_items++;
was_busy++;
break;
case MOVE_PASS:
break;
}
slab_rebal.slab_pos = (char *)slab_rebal.slab_pos + s_cls->size;
if (slab_rebal.slab_pos >= slab_rebal.slab_end)
break;
}
if (slab_rebal.slab_pos >= slab_rebal.slab_end) {
/* Some items were busy, start again from the top */
if (slab_rebal.busy_items) {
slab_rebal.slab_pos = slab_rebal.slab_start;
slab_rebal.busy_items = 0;
} else {
slab_rebal.done++;
}
}
pthread_mutex_unlock(&slabs_lock);
pthread_mutex_unlock(&cache_lock);
return was_busy;
}
static void *assoc_maintenance_thread(void *arg) {
mutex_lock(&maintenance_lock);
while (do_run_maintenance_thread) {
int ii = 0;
/* There is only one expansion thread, so no need to global lock. */
for (ii = 0; ii < hash_bulk_move && expanding; ++ii) {
item *it, *next;
int bucket;
void *item_lock = NULL;
/* bucket = hv & hashmask(hashpower) =>the bucket of hash table
* is the lowest N bits of the hv, and the bucket of item_locks is
* also the lowest M bits of hv, and N is greater than M.
* So we can process expanding with only one item_lock. cool! */
if ((item_lock = item_trylock(expand_bucket))) {
for (it = old_hashtable[expand_bucket]; NULL != it; it = next) {
next = it->h_next;
bucket = hash(ITEM_key(it), it->nkey) & hashmask(hashpower);
it->h_next = primary_hashtable[bucket];
primary_hashtable[bucket] = it;
}
old_hashtable[expand_bucket] = NULL;
expand_bucket++;
if (expand_bucket == hashsize(hashpower - 1)) {
expanding = false;
free(old_hashtable);
STATS_LOCK();
stats.hash_bytes -= hashsize(hashpower - 1) * sizeof(void *);
stats.hash_is_expanding = 0;
STATS_UNLOCK();
if (settings.verbose > 1)
fprintf(stderr, "Hash table expansion done\n");
}
} else {
usleep(10*1000);
}
if (item_lock) {
item_trylock_unlock(item_lock);
item_lock = NULL;
}
}
if (!expanding) {
/* We are done expanding.. just wait for next invocation */
started_expanding = false;
pthread_cond_wait(&maintenance_cond, &maintenance_lock);
/* assoc_expand() swaps out the hash table entirely, so we need
* all threads to not hold any references related to the hash
* table while this happens.
* This is instead of a more complex, possibly slower algorithm to
* allow dynamic hash table expansion without causing significant
* wait times.
*/
pause_threads(PAUSE_ALL_THREADS);
assoc_expand();
pause_threads(RESUME_ALL_THREADS);
}
}
return NULL;
}
/*@null@*/
item *do_item_alloc(char *key, const size_t nkey, const int flags,
const rel_time_t exptime, const int nbytes,
const uint32_t cur_hv) {
uint8_t nsuffix;
item *it = NULL;
char suffix[40];
//计算这个item的空间
size_t ntotal = item_make_header(nkey + 1, flags, nbytes, suffix, &nsuffix);
if (settings.use_cas) {
ntotal += sizeof(uint64_t);
}
//根据大小判断从属于哪个slab
unsigned int id = slabs_clsid(ntotal);
if (id == 0)
return 0;
mutex_lock(&cache_lock);
/* do a quick check if we have any expired items in the tail.. */
//在LRU中尝试5次还没合适的空间，则执行申请空间的操作
int tries = 5;
int tried_alloc = 0;
item *search;
void *hold_lock = NULL;
rel_time_t oldest_live = settings.oldest_live;
search = tails[id];
/* We walk up *only* for locked items. Never searching for expired.
* Waste of CPU for almost all deployments */
for (; tries > 0 && search != NULL; tries--, search=search->prev) {
uint32_t hv = hash(ITEM_key(search), search->nkey, 0);
/* Attempt to hash item lock the "search" item. If locked, no
* other callers can incr the refcount
*/
/* FIXME: I think we need to mask the hv here for comparison? */
if (hv != cur_hv && (hold_lock = item_trylock(hv)) == NULL)
continue;
/* Now see if the item is refcount locked */
if (refcount_incr(&search->refcount) != 2) {
refcount_decr(&search->refcount);
/* Old rare bug could cause a refcount leak. We haven't seen
* it in years, but we leave this code in to prevent failures
* just in case */
if (search->time + TAIL_REPAIR_TIME < current_time) {
itemstats[id].tailrepairs++;
search->refcount = 1;
do_item_unlink_nolock(search, hv);
}
if (hold_lock)
item_trylock_unlock(hold_lock);
continue;
}
/* Expired or flushed */
// search指向的item过期了，则直接复用这块内存
if ((search->exptime != 0 && search->exptime < current_time)
|| (search->time <= oldest_live && oldest_live <= current_time)) {
itemstats[id].reclaimed++;
if ((search->it_flags & ITEM_FETCHED) == 0) {
itemstats[id].expired_unfetched++;
}
it = search;
slabs_adjust_mem_requested(it->slabs_clsid, ITEM_ntotal(it), ntotal);
do_item_unlink_nolock(it, hv);
/* Initialize the item block: */
it->slabs_clsid = 0;
}
//此刻，过期失效的item没有找到，申请内存又失败了。看来只能使用
//LRU淘汰一个item(即使这个item并没有过期失效)
else if ((it = slabs_alloc(ntotal, id)) == NULL) {
tried_alloc = 1;
if (settings.evict_to_free == 0) {
itemstats[id].outofmemory++;
} else {
itemstats[id].evicted++;
itemstats[id].evicted_time = current_time - search->time;
if (search->exptime != 0)
itemstats[id].evicted_nonzero++;
if ((search->it_flags & ITEM_FETCHED) == 0) {
itemstats[id].evicted_unfetched++;
}
it = search;
slabs_adjust_mem_requested(it->slabs_clsid, ITEM_ntotal(it), ntotal);
do_item_unlink_nolock(it, hv);
/* Initialize the item block: */
it->slabs_clsid = 0;
/* If we've just evicted an item, and the automover is set to
* angry bird mode, attempt to rip memory into this slab class.
* TODO: Move valid object detection into a function, and on a
* "successful" memory pull, look behind and see if the next alloc
* would be an eviction. Then kick off the slab mover before the
* eviction happens.
*/
if (settings.slab_automove == 2)
slabs_reassign(-1, id);
}
}
refcount_decr(&search->refcount);
/* If hash values were equal, we don't grab a second lock */
if (hold_lock)
item_trylock_unlock(hold_lock);
break;
}
//从slab分配器中申请内存
if (!tried_alloc && (tries == 0 || search == NULL))
it = slabs_alloc(ntotal, id);
if (it == NULL) {
itemstats[id].outofmemory++;
mutex_unlock(&cache_lock);
return NULL;
}
assert(it->slabs_clsid == 0);
assert(it != heads[id]);
/* Item initialization can happen outside of the lock; the item's already
* been removed from the slab LRU.
*/
it->refcount = 1; /* the caller will have a reference */
mutex_unlock(&cache_lock);
it->next = it->prev = it->h_next = 0;
it->slabs_clsid = id;
DEBUG_REFCNT(it, '*');
it->it_flags = settings.use_cas ? ITEM_CAS : 0;
it->nkey = nkey;
it->nbytes = nbytes;
memcpy(ITEM_key(it), key, nkey);
it->exptime = exptime;
memcpy(ITEM_suffix(it), suffix, (size_t)nsuffix);
it->nsuffix = nsuffix;
return it;
}
static void *item_crawler_thread(void *arg) {
int i;
pthread_mutex_lock(&lru_crawler_lock);
if (settings.verbose > 2)
fprintf(stderr, "Starting LRU crawler background thread\n");
while (do_run_lru_crawler_thread) {
pthread_cond_wait(&lru_crawler_cond, &lru_crawler_lock);
while (crawler_count) {
item *search = NULL;
void *hold_lock = NULL;
for (i = 0; i < LARGEST_ID; i++) {
if (crawlers[i].it_flags != 1) {
continue;
}
pthread_mutex_lock(&cache_lock);
search = crawler_crawl_q((item *)&crawlers[i]);
if (search == NULL ||
(crawlers[i].remaining && --crawlers[i].remaining < 1)) {
if (settings.verbose > 2)
fprintf(stderr, "Nothing left to crawl for %d\n", i);
crawlers[i].it_flags = 0;
crawler_count--;
crawler_unlink_q((item *)&crawlers[i]);
pthread_mutex_unlock(&cache_lock);
continue;
}
uint32_t hv = hash(ITEM_key(search), search->nkey);
/* Attempt to hash item lock the "search" item. If locked, no
* other callers can incr the refcount
*/
if ((hold_lock = item_trylock(hv)) == NULL) {
pthread_mutex_unlock(&cache_lock);
continue;
}
/* Now see if the item is refcount locked */
if (refcount_incr(&search->refcount) != 2) {
refcount_decr(&search->refcount);
if (hold_lock)
item_trylock_unlock(hold_lock);
pthread_mutex_unlock(&cache_lock);
continue;
}
/* Frees the item or decrements the refcount. */
/* Interface for this could improve: do the free/decr here
* instead? */
item_crawler_evaluate(search, hv, i);
if (hold_lock)
item_trylock_unlock(hold_lock);
pthread_mutex_unlock(&cache_lock);
if (settings.lru_crawler_sleep)
usleep(settings.lru_crawler_sleep);
}
}
if (settings.verbose > 2)
fprintf(stderr, "LRU crawler thread sleeping\n");
STATS_LOCK();
stats.lru_crawler_running = false;
STATS_UNLOCK();
}
pthread_mutex_unlock(&lru_crawler_lock);
if (settings.verbose > 2)
fprintf(stderr, "LRU crawler thread stopping\n");
return NULL;
}
/* Returns number of items remove, expired, or evicted.
* Callable from worker threads or the LRU maintainer thread */
static int lru_pull_tail(const int orig_id, const int cur_lru,
const uint64_t total_bytes, uint8_t flags) {
item *it = NULL;
int id = orig_id;
int removed = 0;
if (id == 0)
return 0;
int tries = 5;
item *search;
item *next_it;
void *hold_lock = NULL;
unsigned int move_to_lru = 0;
uint64_t limit = 0;
id |= cur_lru;
pthread_mutex_lock(&lru_locks[id]);
search = tails[id];
/* We walk up *only* for locked items, and if bottom is expired. */
for (; tries > 0 && search != NULL; tries--, search=next_it) {
/* we might relink search mid-loop, so search->prev isn't reliable */
next_it = search->prev;
if (search->nbytes == 0 && search->nkey == 0 && search->it_flags == 1) {
/* We are a crawler, ignore it. */
if (flags & LRU_PULL_CRAWL_BLOCKS) {
pthread_mutex_unlock(&lru_locks[id]);
return 0;
}
tries++;
continue;
}
uint32_t hv = hash(ITEM_key(search), search->nkey);
/* Attempt to hash item lock the "search" item. If locked, no
* other callers can incr the refcount. Also skip ourselves. */
if ((hold_lock = item_trylock(hv)) == NULL)
continue;
/* Now see if the item is refcount locked */
if (refcount_incr(&search->refcount) != 2) {
/* Note pathological case with ref'ed items in tail.
* Can still unlink the item, but it won't be reusable yet */
itemstats[id].lrutail_reflocked++;
/* In case of refcount leaks, enable for quick workaround. */
/* WARNING: This can cause terrible corruption */
if (settings.tail_repair_time &&
search->time + settings.tail_repair_time < current_time) {
itemstats[id].tailrepairs++;
search->refcount = 1;
/* This will call item_remove -> item_free since refcnt is 1 */
do_item_unlink_nolock(search, hv);
item_trylock_unlock(hold_lock);
continue;
}
}
/* Expired or flushed */
if ((search->exptime != 0 && search->exptime < current_time)
|| item_is_flushed(search)) {
itemstats[id].reclaimed++;
if ((search->it_flags & ITEM_FETCHED) == 0) {
itemstats[id].expired_unfetched++;
}
/* refcnt 2 -> 1 */
do_item_unlink_nolock(search, hv);
/* refcnt 1 -> 0 -> item_free */
do_item_remove(search);
item_trylock_unlock(hold_lock);
removed++;
/* If all we're finding are expired, can keep going */
continue;
}
/* If we're HOT_LRU or WARM_LRU and over size limit, send to COLD_LRU.
* If we're COLD_LRU, send to WARM_LRU unless we need to evict
*/
switch (cur_lru) {
case HOT_LRU:
limit = total_bytes * settings.hot_lru_pct / 100;
case WARM_LRU:
if (limit == 0)
limit = total_bytes * settings.warm_lru_pct / 100;
if (sizes_bytes[id] > limit) {
itemstats[id].moves_to_cold++;
move_to_lru = COLD_LRU;
do_item_unlink_q(search);
it = search;
removed++;
break;
} else if ((search->it_flags & ITEM_ACTIVE) != 0) {
/* Only allow ACTIVE relinking if we're not too large. */
itemstats[id].moves_within_lru++;
search->it_flags &= ~ITEM_ACTIVE;
do_item_update_nolock(search);
do_item_remove(search);
item_trylock_unlock(hold_lock);
} else {
/* Don't want to move to COLD, not active, bail out */
it = search;
}
break;
case COLD_LRU:
it = search; /* No matter what, we're stopping */
if (flags & LRU_PULL_EVICT) {
if (settings.evict_to_free == 0) {
/* Don't think we need a counter for this. It'll OOM. */
break;
}
itemstats[id].evicted++;
itemstats[id].evicted_time = current_time - search->time;
if (search->exptime != 0)
itemstats[id].evicted_nonzero++;
if ((search->it_flags & ITEM_FETCHED) == 0) {
itemstats[id].evicted_unfetched++;
}
LOGGER_LOG(NULL, LOG_EVICTIONS, LOGGER_EVICTION, search);
do_item_unlink_nolock(search, hv);
removed++;
if (settings.slab_automove == 2) {
slabs_reassign(-1, orig_id);
}
} else if ((search->it_flags & ITEM_ACTIVE) != 0
&& settings.lru_maintainer_thread) {
itemstats[id].moves_to_warm++;
search->it_flags &= ~ITEM_ACTIVE;
move_to_lru = WARM_LRU;
do_item_unlink_q(search);
removed++;
}
break;
}
if (it != NULL)
break;
}
pthread_mutex_unlock(&lru_locks[id]);
if (it != NULL) {
if (move_to_lru) {
it->slabs_clsid = ITEM_clsid(it);
it->slabs_clsid |= move_to_lru;
item_link_q(it);
}
do_item_remove(it);
item_trylock_unlock(hold_lock);
}
return removed;
}
/*@null@*/
item *do_item_alloc(char *key, const size_t nkey, const int flags,
const rel_time_t exptime, const int nbytes,
const uint32_t cur_hv) {
uint8_t nsuffix;
item *it = NULL;
char suffix[40];
size_t ntotal = item_make_header(nkey + 1, flags, nbytes, suffix, &nsuffix);
if (settings.use_cas) {
ntotal += sizeof(uint64_t);
}
unsigned int id = slabs_clsid(ntotal);
if (id == 0)
return 0;
mutex_lock(&cache_lock);
/* do a quick check if we have any expired items in the tail.. */
int tries = 5;
/* Avoid hangs if a slab has nothing but refcounted stuff in it. */
int tries_lrutail_reflocked = 1000;
int tried_alloc = 0;
item *search;
item *next_it;
void *hold_lock = NULL;
rel_time_t oldest_live = settings.oldest_live;
search = tails[id];
/* We walk up *only* for locked items. Never searching for expired.
* Waste of CPU for almost all deployments */
for (; tries > 0 && search != NULL; tries--, search=next_it) {
/* we might relink search mid-loop, so search->prev isn't reliable */
next_it = search->prev;
if (search->nbytes == 0 && search->nkey == 0 && search->it_flags == 1) {
/* We are a crawler, ignore it. */
tries++;
continue;
}
uint32_t hv = hash(ITEM_key(search), search->nkey);
/* Attempt to hash item lock the "search" item. If locked, no
* other callers can incr the refcount
*/
/* Don't accidentally grab ourselves, or bail if we can't quicklock */
if (hv == cur_hv || (hold_lock = item_trylock(hv)) == NULL)
continue;
/* Now see if the item is refcount locked */
if (refcount_incr(&search->refcount) != 2) {
/* Avoid pathological case with ref'ed items in tail */
do_item_update_nolock(search);
tries_lrutail_reflocked--;
tries++;
refcount_decr(&search->refcount);
itemstats[id].lrutail_reflocked++;
/* Old rare bug could cause a refcount leak. We haven't seen
* it in years, but we leave this code in to prevent failures
* just in case */
if (settings.tail_repair_time &&
search->time + settings.tail_repair_time < current_time) {
itemstats[id].tailrepairs++;
search->refcount = 1;
do_item_unlink_nolock(search, hv);
}
if (hold_lock)
item_trylock_unlock(hold_lock);
if (tries_lrutail_reflocked < 1)
break;
continue;
}
/* Expired or flushed */
if ((search->exptime != 0 && search->exptime < current_time)
|| (search->time <= oldest_live && oldest_live <= current_time)) {
itemstats[id].reclaimed++;
if ((search->it_flags & ITEM_FETCHED) == 0) {
itemstats[id].expired_unfetched++;
}
it = search;
slabs_adjust_mem_requested(it->slabs_clsid, ITEM_ntotal(it), ntotal);
do_item_unlink_nolock(it, hv);
/* Initialize the item block: */
it->slabs_clsid = 0;
} else if ((it = slabs_alloc(ntotal, id)) == NULL) {
tried_alloc = 1;
if (settings.evict_to_free == 0) {
itemstats[id].outofmemory++;
} else {
itemstats[id].evicted++;
itemstats[id].evicted_time = current_time - search->time;
if (search->exptime != 0)
itemstats[id].evicted_nonzero++;
if ((search->it_flags & ITEM_FETCHED) == 0) {
itemstats[id].evicted_unfetched++;
}
shadow_item* new_shadow_it = create_shadow_item(search);
hv = hash(new_shadow_it->key, new_shadow_it->nkey);
shadow_assoc_insert(new_shadow_it, hv);
insert_shadowq_item(new_shadow_it,new_shadow_it->slabs_clsid);
it = search;
slabs_adjust_mem_requested(it->slabs_clsid, ITEM_ntotal(it), ntotal);
do_item_unlink_nolock(it, hv);
/* Initialize the item block: */
it->slabs_clsid = 0;
/* If we've just evicted an item, and the automover is set to
* angry bird mode, attempt to rip memory into this slab class.
* TODO: Move valid object detection into a function, and on a
* "successful" memory pull, look behind and see if the next alloc
* would be an eviction. Then kick off the slab mover before the
* eviction happens.
*/
if (settings.slab_automove == 2)
slabs_reassign(-1, id);
}
}
refcount_decr(&search->refcount);
/* If hash values were equal, we don't grab a second lock */
if (hold_lock)
item_trylock_unlock(hold_lock);
break;
}
if (!tried_alloc && (tries == 0 || search == NULL))
it = slabs_alloc(ntotal, id);
if (it == NULL) {
itemstats[id].outofmemory++;
mutex_unlock(&cache_lock);
return NULL;
}
assert(it->slabs_clsid == 0);
assert(it != heads[id]);
/* Item initialization can happen outside of the lock; the item's already
* been removed from the slab LRU.
*/
it->refcount = 1; /* the caller will have a reference */
mutex_unlock(&cache_lock);
it->next = it->prev = it->h_next = 0;
it->slabs_clsid = id;
DEBUG_REFCNT(it, '*');
it->it_flags = settings.use_cas ? ITEM_CAS : 0;
it->nkey = nkey;
it->nbytes = nbytes;
memcpy(ITEM_key(it), key, nkey);
it->exptime = exptime;
memcpy(ITEM_suffix(it), suffix, (size_t)nsuffix);
it->nsuffix = nsuffix;
return it;
}
/* refcount == 0 is safe since nobody can incr while cache_lock is held.
* refcount != 0 is impossible since flags/etc can be modified in other
* threads. instead, note we found a busy one and bail. logic in do_item_get
* will prevent busy items from continuing to be busy
*/
static int slab_rebalance_move(void) {
slabclass_t *s_cls;
int x;
int was_busy = 0;
int refcount = 0;
enum move_status status = MOVE_PASS;
pthread_mutex_lock(&cache_lock);
pthread_mutex_lock(&slabs_lock);
s_cls = &slabclass[slab_rebal.s_clsid];
//会在start_slab_maintenance_thread函数中读取环境变量设置slab_bulk_check
//默认值为1.同样这里也是采用分期处理的方案处理一个页上的多个item
for (x = 0; x < slab_bulk_check; x++) {
item *it = slab_rebal.slab_pos;
status = MOVE_PASS;
if (it->slabs_clsid != 255) {
void *hold_lock = NULL;
uint32_t hv = hash(ITEM_key(it), it->nkey);
if ((hold_lock = item_trylock(hv)) == NULL) {
status = MOVE_LOCKED;
} else {
refcount = refcount_incr(&it->refcount);
if (refcount == 1) { /* item is unlinked, unused */
//如果it_flags&ITEM_SLABBED为真，那么就说明这个item
//根本就没有分配出去。如果为假，那么说明这个item被分配
//出去了，但处于归还途中。参考do_item_get函数里面的
//判断语句，有slab_rebalance_signal作为判断条件的那个。
if (it->it_flags & ITEM_SLABBED) {//没有分配出去
/* remove from slab freelist */
if (s_cls->slots == it) {
s_cls->slots = it->next;
}
if (it->next) it->next->prev = it->prev;
if (it->prev) it->prev->next = it->next;
s_cls->sl_curr--;
status = MOVE_DONE;//这个item处理成功
} else {//此时还有另外一个worker线程在归还这个item
status = MOVE_BUSY;
}
} else if (refcount == 2) { /* item is linked but not busy */
//没有worker线程引用这个item
if ((it->it_flags & ITEM_LINKED) != 0) {
//直接把这个item从哈希表和LRU队列中删除
do_item_unlink_nolock(it, hv);
status = MOVE_DONE;
} else {
/* refcount == 1 + !ITEM_LINKED means the item is being
* uploaded to, or was just unlinked but hasn't been freed
* yet. Let it bleed off on its own and try again later */
//现在有worker线程正在引用这个item
status = MOVE_BUSY;
}
} else {
if (settings.verbose > 2) {
fprintf(stderr, "Slab reassign hit a busy item: refcount: %d (%d -> %d)\n",
it->refcount, slab_rebal.s_clsid, slab_rebal.d_clsid);
}
status = MOVE_BUSY;
}
item_trylock_unlock(hold_lock);
}
}
switch (status) {
case MOVE_DONE:
it->refcount = 0;//引用计数清零
it->it_flags = 0;//清零所有属性
it->slabs_clsid = 255;
break;
case MOVE_BUSY:
refcount_decr(&it->refcount);//注意这里没有break
case MOVE_LOCKED:
slab_rebal.busy_items++;
was_busy++;//记录是否有不能马上处理的item
break;
case MOVE_PASS:
break;
}
//处理这个页的下一个item
slab_rebal.slab_pos = (char *)slab_rebal.slab_pos + s_cls->size;
if (slab_rebal.slab_pos >= slab_rebal.slab_end)//遍历完了这个页
break;
}
//遍历完了这个页的所有item
if (slab_rebal.slab_pos >= slab_rebal.slab_end) {
/* Some items were busy, start again from the top */
//在处理的时候，跳过了一些item(因为有worker线程在引用)
if (slab_rebal.busy_items) {//此时需要从头再扫描一次这个页
slab_rebal.slab_pos = slab_rebal.slab_start;
slab_rebal.busy_items = 0;
} else {
slab_rebal.done++;//标志已经处理完这个页的所有item
}
}
pthread_mutex_unlock(&slabs_lock);
pthread_mutex_unlock(&cache_lock);
return was_busy;//返回记录
}
//分配一个item，这个函数包含了memcached具体item分配的逻辑
item *do_item_alloc(char *key, const size_t nkey, const int flags,
const rel_time_t exptime, const int nbytes,
const uint32_t cur_hv) {
uint8_t nsuffix;
item *it = NULL;
char suffix[40];
size_t ntotal = item_make_header(nkey + 1, flags, nbytes, suffix, &nsuffix); //item总大小
if (settings.use_cas) {
ntotal += sizeof(uint64_t); //如果有用到cas，那么item大小还要加上unit64_t的size
}
unsigned int id = slabs_clsid(ntotal); //根据item的大小，找到适合的slabclass
if (id == 0)
return 0;
mutex_lock(&cache_lock); //cache锁
/* do a quick check if we have any expired items in the tail.. */
int tries = 5;
int tried_alloc = 0;
item *search;
void *hold_lock = NULL;
rel_time_t oldest_live = settings.oldest_live;
search = tails[id]; //全局变量，tails[x]是id为x的slabclass lru链表的尾部
/* We walk up *only* for locked items. Never searching for expired.
* Waste of CPU for almost all deployments */
//首先从lru链表尾部查找有没有过期的item，tries = 5，最多循环5次
//注意这里是最多查找5次，只要找到一个没有被其他地方引用的item，那么就不再继续查找，如果这个item过期，就使用这个item的空间，否则创建新的slab
for (; tries > 0 && search != NULL; tries--, search=search->prev) {
if (search->nbytes == 0 && search->nkey == 0 && search->it_flags == 1) {
/* We are a crawler, ignore it. */
//这里只是搜索过期的item，对于异常的item，直接忽略继续查找
tries++;
continue;
}
//计算item的hash值，hv有两个作用：1.用于hash表保存item 2.用于item lock表中锁住item，通过hv计算出item_lock中哪个锁对当前item加锁
//不同item的hash值可能相同，hash表中用链表的方式解决冲突；item lock中多个item共享一个锁
uint32_t hv = hash(ITEM_key(search), search->nkey);
/* Attempt to hash item lock the "search" item. If locked, no
* other callers can incr the refcount
*/
/* Don't accidentally grab ourselves, or bail if we can't quicklock */
//锁住当前item
if (hv == cur_hv || (hold_lock = item_trylock(hv)) == NULL)
continue;
/* Now see if the item is refcount locked */
//检查这个指向的这个item是否被其他地方引用，如果是的话，继续向前查找
if (refcount_incr(&search->refcount) != 2) {
refcount_decr(&search->refcount);
/* Old rare bug could cause a refcount leak. We haven't seen
* it in years, but we leave this code in to prevent failures
* just in case */
if (settings.tail_repair_time &&
search->time + settings.tail_repair_time < current_time) {
itemstats[id].tailrepairs++;
search->refcount = 1;
do_item_unlink_nolock(search, hv);
}
if (hold_lock)
item_trylock_unlock(hold_lock);
continue;
}
/* Expired or flushed */
//如果找到过期的item
if ((search->exptime != 0 && search->exptime < current_time)
|| (search->time <= oldest_live && oldest_live <= current_time)) {
itemstats[id].reclaimed++;
if ((search->it_flags & ITEM_FETCHED) == 0) {
itemstats[id].expired_unfetched++;
}
it = search;
//更新统计数据
slabs_adjust_mem_requested(it->slabs_clsid, ITEM_ntotal(it), ntotal);
//把旧的item从hash表和LRU链表中移除
do_item_unlink_nolock(it, hv);
/* Initialize the item block: */
it->slabs_clsid = 0;
}
//如果没有找到过期的item，则调用slabs_alloc分配空间
//如果slabs_alloc返回null，表示分配失败，内存空间已满
//需要按LRU进行淘汰
else if ((it = slabs_alloc(ntotal, id)) == NULL) {
tried_alloc = 1; //标记一下，表示有尝试调用slabs_alloc分配空间
//记录被淘汰item的信息, 使用memcached经常会查看的evicted_time就是在这里赋值的
if (settings.evict_to_free == 0) {
itemstats[id].outofmemory++;
} else {
itemstats[id].evicted++;
itemstats[id].evicted_time = current_time - search->time; //被淘汰item距离上次使用的时间
if (search->exptime != 0)
itemstats[id].evicted_nonzero++;
if ((search->it_flags & ITEM_FETCHED) == 0) {
itemstats[id].evicted_unfetched++;
}
it = search;
slabs_adjust_mem_requested(it->slabs_clsid, ITEM_ntotal(it), ntotal); //更新统计数据
do_item_unlink_nolock(it, hv); //从hash表和LRU链表中移除
/* Initialize the item block: */
it->slabs_clsid = 0;
/* If we've just evicted an item, and the automover is set to
* angry bird mode, attempt to rip memory into this slab class.
* TODO: Move valid object detection into a function, and on a
* "successful" memory pull, look behind and see if the next alloc
* would be an eviction. Then kick off the slab mover before the
* eviction happens.
*/
//默认情况下，slab_automove=1，会合理地更具淘汰统计数据来分析怎么进行slabclass空间的分配
//如果slab_automove=2，只要分配失败了，马上进行slabclass空间的重分配
if (settings.slab_automove == 2)
slabs_reassign(-1, id);
}
}
refcount_decr(&search->refcount);
/* If hash values were equal, we don't grab a second lock */
if (hold_lock)
item_trylock_unlock(hold_lock);
break;
}
//查找5次过期的item都失败，并且也没有淘汰可用且没有过期的item
//分配新的内存空间
if (!tried_alloc && (tries == 0 || search == NULL))
it = slabs_alloc(ntotal, id);
//分配失败，返回null
if (it == NULL) {
itemstats[id].outofmemory++;
mutex_unlock(&cache_lock);
return NULL;
}
assert(it->slabs_clsid == 0);
assert(it != heads[id]);
/* Item initialization can happen outside of the lock; the item's already
* been removed from the slab LRU.
*/
//item内存空间分配成功，做一些初始化工作
it->refcount = 1; /* the caller will have a reference */
mutex_unlock(&cache_lock);
it->next = it->prev = it->h_next = 0;
it->slabs_clsid = id;
DEBUG_REFCNT(it, '*');
it->it_flags = settings.use_cas ? ITEM_CAS : 0;
it->nkey = nkey;
it->nbytes = nbytes;
memcpy(ITEM_key(it), key, nkey);
it->exptime = exptime;
memcpy(ITEM_suffix(it), suffix, (size_t)nsuffix);
it->nsuffix = nsuffix;
return it;
}
/* slawek - reclaim patch */
char *do_item_cacheremove(const unsigned int slabs_clsid, const unsigned int limit, const unsigned int limit_remove, unsigned int *bytes) {
uint32_t hv;
void *hold_lock = NULL;
const short clean_stats_size = 255;
unsigned short stats_times[clean_stats_size];
unsigned short stats_expired[clean_stats_size];
memset(&stats_times, 0, sizeof(short) * clean_stats_size);
memset(&stats_expired, 0, sizeof(short) * clean_stats_size);
char* buffer = malloc((size_t) 16 * 1024);
unsigned int bufcurr = 0;
char temp[1024];
slabclass_t* slab = &slabclass[slabs_clsid];
if (slab->list_size == 0 || slabs_clsid > power_largest /*slabclass[power_largest] is defined*/ )
{
// return as slab is currently empty (0 items, nothing even allocated!)
int len = snprintf(temp, sizeof(temp), "Slab %d is currently empty (not allocated), no cleaning done\r\n", slabs_clsid);
memcpy(buffer + bufcurr, temp, len);
bufcurr += len;
memcpy(buffer + bufcurr, "END\r\n", 6);
bufcurr += 5;
*bytes = bufcurr;
return buffer;
}
slab->reclaimed_slab_item_num++;
bool slab_changed = false;
if (slab->reclaimed_slab_item_num >= slab->perslab)
{
// we're switching slabs
slab->reclaimed_slab_num = (slab->reclaimed_slab_num + 1) % (slab->slabs);
slab->reclaimed_slab_item_num = 0;
slab_changed = true;
}
unsigned int _ipos_start = slab->reclaimed_slab_item_num;
item *it_first = NULL;
int i=0;
for (; i < slab->slabs; i++)
{
it_first = (void *)(slab->slab_list[slab->reclaimed_slab_num]);
if (it_first != NULL)
break;
// we're switching slabs
slab->reclaimed_slab_num = (slab->reclaimed_slab_num + 1) % slab->slabs;
slab->reclaimed_slab_item_num = 0;
}
if (it_first == NULL)
{
// oh crap, no items found in whole slab class, return the good news!
int len = snprintf(temp, sizeof(temp), "Slab %d is currently empty (no items), no cleaning done\r\n", slabs_clsid);
memcpy(buffer + bufcurr, temp, len);
bufcurr += len;
memcpy(buffer + bufcurr, "END\r\n", 6);
bufcurr += 5;
*bytes = bufcurr;
return buffer;
}
int start = slab->reclaimed_slab_item_num;
if (start >= slab->perslab)
start = 0;
int end = start + limit;
if (end >= slab->perslab || limit == 0)
end = slab->perslab;
// run the cleaning, taking limits into account!
uint64_t items_removed = 0;
uint64_t bytes_removed = 0;
item *it = (void*) ((char*)it_first + (slab->size*start));
i=start;
int items_found = 0;
int _ttl_left = 0;
for (; i<end; i++)
{
// iterate over items at start, so we won't have to iterate near each continue
if (i!=start) {
it = (void*) (((char*)it) + slab->size);
}
// dirty reads, prevent locking!
if ( (it->it_flags & ITEM_SLABBED) == 0 && it->refcount > 0 /* from remove items */
&& it->slabs_clsid > 0 )
{
items_found ++;
if (it->exptime == 0)
continue;
}
else
continue;
// check if we can expire the item
_ttl_left = it->exptime - current_time;
if (_ttl_left > 0)
{
if (_ttl_left >= clean_stats_size)
_ttl_left = clean_stats_size-1;
stats_times[_ttl_left] ++;
continue;
}
else
{
_ttl_left = 0-_ttl_left;
if (_ttl_left >= clean_stats_size)
_ttl_left = clean_stats_size-1;
stats_expired[_ttl_left]++;
}
// dirty reads ends here, proceed with locking!
hv = hash(ITEM_key(it), it->nkey, 0);
/* Attempt to hash item lock the item. If locked, no
* other callers can incr the refcount */
if ((hold_lock = item_trylock(hv)) == NULL) {
continue;
}
/* Now see if the item is refcount locked */
if (refcount_incr(&it->refcount) != 2) {
refcount_decr(&it->refcount);
if (hold_lock) {
item_trylock_unlock(hold_lock);
}
continue;
}
// LOCKED, proceed with removing!
assert(it->nkey <= KEY_MAX_LENGTH);
_ttl_left = it->exptime - current_time;
if ( (it->it_flags & ITEM_SLABBED) == 0 && it->refcount > 0 /* from remove items */
&& it->slabs_clsid > 0 && _ttl_left <= 0)
{
//do_item_remove(it); item will get removed automatically in unlink function!
do_item_unlink_nolock_nostat(it, hv, &items_removed, &bytes_removed);
/* Initialize the item block: */
it->slabs_clsid = 0;
slabs_free(it, ITEM_ntotal(it), slabs_clsid);
}
else
{
refcount_decr(&it->refcount);
}
if (hold_lock) {
item_trylock_unlock(hold_lock);
}
}
// let's update position!
slab->reclaimed_slab_item_num = i;
// update stats
STATS_LOCK();
stats.curr_bytes -= bytes_removed;
stats.curr_items -= items_removed;
stats.reclaimed_fast += items_removed;
stats.reclaimed_fast_bytes += bytes_removed;
stats.reclaim_item_passes += (i - start);
stats.reclaim_item_found += items_found;
stats.reclaim_slab_memory_passes += (slab_changed ? 1 : 0);
STATS_UNLOCK();
int len = snprintf(temp, sizeof(temp),
"Expiring items in SLAB: %d, Memory region: %d / %d (max: %d), Item pos: S:%d / E:%d\r\n"
"Scanned items: %d (Found: %d) / Expired: %llu (%llu KB)\r\n"
"Limits ... Items: %u, Max Remove: %u (0 - none)\r\n"
,slabs_clsid, slab->reclaimed_slab_num, slab->slabs, slab->list_size, _ipos_start, slab->reclaimed_slab_item_num,
(i - start), items_found, (unsigned long long) items_removed, (unsigned long long) ( (bytes_removed+1023) / 1024),
limit, limit_remove);
memcpy(buffer + bufcurr, temp, len);
bufcurr += len;
len = snprintf(temp, sizeof(temp), "TTLs: ");
memcpy(buffer + bufcurr, temp, len);
bufcurr += len;
for (i=0; i<clean_stats_size; i++)
{
int len = snprintf(temp, sizeof(temp), "%d,", stats_times[i]);
memcpy(buffer + bufcurr, temp, len);
bufcurr += len;
}
bufcurr --;
*(buffer + bufcurr) = 0;
len = snprintf(temp, sizeof(temp), "\r\nExpired Age: ");
memcpy(buffer + bufcurr, temp, len);
bufcurr += len;
for (i=0; i<clean_stats_size; i++)
{
int len = snprintf(temp, sizeof(temp), "%d,", stats_expired[i]);
memcpy(buffer + bufcurr, temp, len);
bufcurr += len;
}
bufcurr --;
*(buffer + bufcurr) = 0;
memcpy(buffer + bufcurr, "\r\nEND\r\n", 8);
bufcurr += 7;
*bytes = bufcurr;
return buffer;
}
/* refcount == 0 is safe since nobody can incr while item_lock is held.
* refcount != 0 is impossible since flags/etc can be modified in other
* threads. instead, note we found a busy one and bail. logic in do_item_get
* will prevent busy items from continuing to be busy
* NOTE: This is checking it_flags outside of an item lock. I believe this
* works since it_flags is 8 bits, and we're only ever comparing a single bit
* regardless. ITEM_SLABBED bit will always be correct since we're holding the
* lock which modifies that bit. ITEM_LINKED won't exist if we're between an
* item having ITEM_SLABBED removed, and the key hasn't been added to the item
* yet. The memory barrier from the slabs lock should order the key write and the
* flags to the item?
* If ITEM_LINKED did exist and was just removed, but we still see it, that's
* still safe since it will have a valid key, which we then lock, and then
* recheck everything.
* This may not be safe on all platforms; If not, slabs_alloc() will need to
* seed the item key while holding slabs_lock.
*/
static int slab_rebalance_move(void) {
slabclass_t *s_cls;
int x;
int was_busy = 0;
int refcount = 0;
uint32_t hv;
void *hold_lock;
enum move_status status = MOVE_PASS;
pthread_mutex_lock(&slabs_lock);
s_cls = &slabclass[slab_rebal.s_clsid];
for (x = 0; x < slab_bulk_check; x++) {
hv = 0;
hold_lock = NULL;
item *it = slab_rebal.slab_pos;
status = MOVE_PASS;
/* ITEM_FETCHED when ITEM_SLABBED is overloaded to mean we've cleared
* the chunk for move. Only these two flags should exist.
*/
if (it->it_flags != (ITEM_SLABBED|ITEM_FETCHED)) {
/* ITEM_SLABBED can only be added/removed under the slabs_lock */
if (it->it_flags & ITEM_SLABBED) {
/* remove from slab freelist */
if (s_cls->slots == it) {
s_cls->slots = it->next;
}
if (it->next) it->next->prev = it->prev;
if (it->prev) it->prev->next = it->next;
s_cls->sl_curr--;
status = MOVE_FROM_SLAB;
} else if ((it->it_flags & ITEM_LINKED) != 0) {
/* If it doesn't have ITEM_SLABBED, the item could be in any
* state on its way to being freed or written to. If no
* ITEM_SLABBED, but it's had ITEM_LINKED, it must be active
* and have the key written to it already.
*/
hv = hash(ITEM_key(it), it->nkey);
if ((hold_lock = item_trylock(hv)) == NULL) {
status = MOVE_LOCKED;
} else {
refcount = refcount_incr(&it->refcount);
if (refcount == 2) { /* item is linked but not busy */
/* Double check ITEM_LINKED flag here, since we're
* past a memory barrier from the mutex. */
if ((it->it_flags & ITEM_LINKED) != 0) {
status = MOVE_FROM_LRU;
} else {
/* refcount == 1 + !ITEM_LINKED means the item is being
* uploaded to, or was just unlinked but hasn't been freed
* yet. Let it bleed off on its own and try again later */
status = MOVE_BUSY;
}
} else {
if (settings.verbose > 2) {
fprintf(stderr, "Slab reassign hit a busy item: refcount: %d (%d -> %d)\n",
it->refcount, slab_rebal.s_clsid, slab_rebal.d_clsid);
}
status = MOVE_BUSY;
}
/* Item lock must be held while modifying refcount */
if (status == MOVE_BUSY) {
refcount_decr(&it->refcount);
item_trylock_unlock(hold_lock);
}
}
} else {
/* See above comment. No ITEM_SLABBED or ITEM_LINKED. Mark
* busy and wait for item to complete its upload. */
status = MOVE_BUSY;
}
}
int save_item = 0;
item *new_it = NULL;
size_t ntotal = 0;
switch (status) {
case MOVE_FROM_LRU:
/* Lock order is LRU locks -> slabs_lock. unlink uses LRU lock.
* We only need to hold the slabs_lock while initially looking
* at an item, and at this point we have an exclusive refcount
* (2) + the item is locked. Drop slabs lock, drop item to
* refcount 1 (just our own, then fall through and wipe it
*/
/* Check if expired or flushed */
ntotal = ITEM_ntotal(it);
/* REQUIRES slabs_lock: CHECK FOR cls->sl_curr > 0 */
if ((it->exptime != 0 && it->exptime < current_time)
|| item_is_flushed(it)) {
/* TODO: maybe we only want to save if item is in HOT or
* WARM LRU?
*/
save_item = 0;
} else if ((new_it = slab_rebalance_alloc(ntotal, slab_rebal.s_clsid)) == NULL) {
save_item = 0;
slab_rebal.evictions_nomem++;
} else {
save_item = 1;
}
pthread_mutex_unlock(&slabs_lock);
if (save_item) {
/* if free memory, memcpy. clear prev/next/h_bucket */
memcpy(new_it, it, ntotal);
new_it->prev = 0;
new_it->next = 0;
new_it->h_next = 0;
/* These are definitely required. else fails assert */
new_it->it_flags &= ~ITEM_LINKED;
new_it->refcount = 0;
do_item_replace(it, new_it, hv);
slab_rebal.rescues++;
} else {
do_item_unlink(it, hv);
}
item_trylock_unlock(hold_lock);
pthread_mutex_lock(&slabs_lock);
/* Always remove the ntotal, as we added it in during
* do_slabs_alloc() when copying the item.
*/
s_cls->requested -= ntotal;
case MOVE_FROM_SLAB:
it->refcount = 0;
it->it_flags = ITEM_SLABBED|ITEM_FETCHED;
#ifdef DEBUG_SLAB_MOVER
memcpy(ITEM_key(it), "deadbeef", 8);
#endif
break;
case MOVE_BUSY:
case MOVE_LOCKED:
slab_rebal.busy_items++;
was_busy++;
break;
case MOVE_PASS:
break;
}
slab_rebal.slab_pos = (char *)slab_rebal.slab_pos + s_cls->size;
if (slab_rebal.slab_pos >= slab_rebal.slab_end)
break;
}
if (slab_rebal.slab_pos >= slab_rebal.slab_end) {
/* Some items were busy, start again from the top */
if (slab_rebal.busy_items) {
slab_rebal.slab_pos = slab_rebal.slab_start;
STATS_LOCK();
stats.slab_reassign_busy_items += slab_rebal.busy_items;
STATS_UNLOCK();
slab_rebal.busy_items = 0;
} else {
slab_rebal.done++;
}
}
pthread_mutex_unlock(&slabs_lock);
return was_busy;
}
/* refcount == 0 is safe since nobody can incr while item_lock is held.
* refcount != 0 is impossible since flags/etc can be modified in other
* threads. instead, note we found a busy one and bail. logic in do_item_get
* will prevent busy items from continuing to be busy
* NOTE: This is checking it_flags outside of an item lock. I believe this
* works since it_flags is 8 bits, and we're only ever comparing a single bit
* regardless. ITEM_SLABBED bit will always be correct since we're holding the
* lock which modifies that bit. ITEM_LINKED won't exist if we're between an
* item having ITEM_SLABBED removed, and the key hasn't been added to the item
* yet. The memory barrier from the slabs lock should order the key write and the
* flags to the item?
* If ITEM_LINKED did exist and was just removed, but we still see it, that's
* still safe since it will have a valid key, which we then lock, and then
* recheck everything.
* This may not be safe on all platforms; If not, slabs_alloc() will need to
* seed the item key while holding slabs_lock.
*/
static int slab_rebalance_move(void) {
slabclass_t *s_cls;
int x;
int was_busy = 0;
int refcount = 0;
uint32_t hv;
void *hold_lock;
enum move_status status = MOVE_PASS;
pthread_mutex_lock(&slabs_lock);
s_cls = &slabclass[slab_rebal.s_clsid];
for (x = 0; x < slab_bulk_check; x++) {
hv = 0;
hold_lock = NULL;
item *it = slab_rebal.slab_pos;
item_chunk *ch = NULL;
status = MOVE_PASS;
if (it->it_flags & ITEM_CHUNK) {
/* This chunk is a chained part of a larger item. */
ch = (item_chunk *) it;
/* Instead, we use the head chunk to find the item and effectively
* lock the entire structure. If a chunk has ITEM_CHUNK flag, its
* head cannot be slabbed, so the normal routine is safe. */
it = ch->head;
assert(it->it_flags & ITEM_CHUNKED);
}
/* ITEM_FETCHED when ITEM_SLABBED is overloaded to mean we've cleared
* the chunk for move. Only these two flags should exist.
*/
if (it->it_flags != (ITEM_SLABBED|ITEM_FETCHED)) {
/* ITEM_SLABBED can only be added/removed under the slabs_lock */
if (it->it_flags & ITEM_SLABBED) {
assert(ch == NULL);
slab_rebalance_cut_free(s_cls, it);
status = MOVE_FROM_SLAB;
} else if ((it->it_flags & ITEM_LINKED) != 0) {
/* If it doesn't have ITEM_SLABBED, the item could be in any
* state on its way to being freed or written to. If no
* ITEM_SLABBED, but it's had ITEM_LINKED, it must be active
* and have the key written to it already.
*/
hv = hash(ITEM_key(it), it->nkey);
if ((hold_lock = item_trylock(hv)) == NULL) {
status = MOVE_LOCKED;
} else {
refcount = refcount_incr(it);
if (refcount == 2) { /* item is linked but not busy */
/* Double check ITEM_LINKED flag here, since we're
* past a memory barrier from the mutex. */
if ((it->it_flags & ITEM_LINKED) != 0) {
status = MOVE_FROM_LRU;
} else {
/* refcount == 1 + !ITEM_LINKED means the item is being
* uploaded to, or was just unlinked but hasn't been freed
* yet. Let it bleed off on its own and try again later */
status = MOVE_BUSY;
}
} else {
if (settings.verbose > 2) {
fprintf(stderr, "Slab reassign hit a busy item: refcount: %d (%d -> %d)\n",
it->refcount, slab_rebal.s_clsid, slab_rebal.d_clsid);
}
status = MOVE_BUSY;
}
/* Item lock must be held while modifying refcount */
if (status == MOVE_BUSY) {
refcount_decr(it);
item_trylock_unlock(hold_lock);
}
}
} else {
/* See above comment. No ITEM_SLABBED or ITEM_LINKED. Mark
* busy and wait for item to complete its upload. */
status = MOVE_BUSY;
}
}
int save_item = 0;
item *new_it = NULL;
size_t ntotal = 0;
switch (status) {
case MOVE_FROM_LRU:
/* Lock order is LRU locks -> slabs_lock. unlink uses LRU lock.
* We only need to hold the slabs_lock while initially looking
* at an item, and at this point we have an exclusive refcount
* (2) + the item is locked. Drop slabs lock, drop item to
* refcount 1 (just our own, then fall through and wipe it
*/
/* Check if expired or flushed */
ntotal = ITEM_ntotal(it);
/* REQUIRES slabs_lock: CHECK FOR cls->sl_curr > 0 */
if (ch == NULL && (it->it_flags & ITEM_CHUNKED)) {
/* Chunked should be identical to non-chunked, except we need
* to swap out ntotal for the head-chunk-total. */
ntotal = s_cls->size;
}
if ((it->exptime != 0 && it->exptime < current_time)
|| item_is_flushed(it)) {
/* Expired, don't save. */
save_item = 0;
} else if (ch == NULL &&
(new_it = slab_rebalance_alloc(ntotal, slab_rebal.s_clsid)) == NULL) {
/* Not a chunk of an item, and nomem. */
save_item = 0;
slab_rebal.evictions_nomem++;
} else if (ch != NULL &&
(new_it = slab_rebalance_alloc(s_cls->size, slab_rebal.s_clsid)) == NULL) {
/* Is a chunk of an item, and nomem. */
save_item = 0;
slab_rebal.evictions_nomem++;
} else {
/* Was whatever it was, and we have memory for it. */
save_item = 1;
}
pthread_mutex_unlock(&slabs_lock);
unsigned int requested_adjust = 0;
if (save_item) {
if (ch == NULL) {
assert((new_it->it_flags & ITEM_CHUNKED) == 0);
/* if free memory, memcpy. clear prev/next/h_bucket */
memcpy(new_it, it, ntotal);
new_it->prev = 0;
new_it->next = 0;
new_it->h_next = 0;
/* These are definitely required. else fails assert */
new_it->it_flags &= ~ITEM_LINKED;
new_it->refcount = 0;
do_item_replace(it, new_it, hv);
/* Need to walk the chunks and repoint head */
if (new_it->it_flags & ITEM_CHUNKED) {
item_chunk *fch = (item_chunk *) ITEM_data(new_it);
fch->next->prev = fch;
while (fch) {
fch->head = new_it;
fch = fch->next;
}
}
it->refcount = 0;
it->it_flags = ITEM_SLABBED|ITEM_FETCHED;
#ifdef DEBUG_SLAB_MOVER
memcpy(ITEM_key(it), "deadbeef", 8);
#endif
slab_rebal.rescues++;
requested_adjust = ntotal;
} else {
item_chunk *nch = (item_chunk *) new_it;
/* Chunks always have head chunk (the main it) */
ch->prev->next = nch;
if (ch->next)
ch->next->prev = nch;
memcpy(nch, ch, ch->used + sizeof(item_chunk));
ch->refcount = 0;
ch->it_flags = ITEM_SLABBED|ITEM_FETCHED;
slab_rebal.chunk_rescues++;
#ifdef DEBUG_SLAB_MOVER
memcpy(ITEM_key((item *)ch), "deadbeef", 8);
#endif
refcount_decr(it);
requested_adjust = s_cls->size;
}
} else {
/* restore ntotal in case we tried saving a head chunk. */
ntotal = ITEM_ntotal(it);
do_item_unlink(it, hv);
slabs_free(it, ntotal, slab_rebal.s_clsid);
/* Swing around again later to remove it from the freelist. */
slab_rebal.busy_items++;
was_busy++;
}
item_trylock_unlock(hold_lock);
pthread_mutex_lock(&slabs_lock);
/* Always remove the ntotal, as we added it in during
* do_slabs_alloc() when copying the item.
*/
s_cls->requested -= requested_adjust;
break;
case MOVE_FROM_SLAB:
it->refcount = 0;
it->it_flags = ITEM_SLABBED|ITEM_FETCHED;
#ifdef DEBUG_SLAB_MOVER
memcpy(ITEM_key(it), "deadbeef", 8);
#endif
break;
case MOVE_BUSY:
case MOVE_LOCKED:
slab_rebal.busy_items++;
was_busy++;
break;
case MOVE_PASS:
break;
}
slab_rebal.slab_pos = (char *)slab_rebal.slab_pos + s_cls->size;
if (slab_rebal.slab_pos >= slab_rebal.slab_end)
break;
}
if (slab_rebal.slab_pos >= slab_rebal.slab_end) {
/* Some items were busy, start again from the top */
if (slab_rebal.busy_items) {
slab_rebal.slab_pos = slab_rebal.slab_start;
STATS_LOCK();
stats.slab_reassign_busy_items += slab_rebal.busy_items;
STATS_UNLOCK();
slab_rebal.busy_items = 0;
} else {
slab_rebal.done++;
}
}
pthread_mutex_unlock(&slabs_lock);
return was_busy;
}
/* 扩容线程函数，扩容策略如下:
* 扩容线程在main函数中创建，在assoc_insert后发现item数目大于哈希表容量1.5倍，唤醒扩容线程。
* 扩容线程先创建一个2倍容量的新哈希表，然后进行把数据从旧哈希表迁移到新哈希表。
* 迁移从旧表索引0开始，每次迁移一个桶(可以增加迁移粒度，但由于迁移需要加锁，可能导致work线程获取锁的等待时间增加)，
* 迁移完成后释放旧表。
*/
static void *assoc_maintenance_thread(void *arg) {
mutex_lock(&maintenance_lock);
while (do_run_maintenance_thread) {
int ii = 0;
/* There is only one expansion thread, so no need to global lock. */
//如果expanding为true才会进入循环体，所以迁移线程刚创建的时候，并不会进入循环体
for (ii = 0; ii < hash_bulk_move && expanding; ++ii) {
item *it, *next;
int bucket;
void *item_lock = NULL;
/* bucket = hv & hashmask(hashpower) =>the bucket of hash table
* is the lowest N bits of the hv, and the bucket of item_locks is
* also the lowest M bits of hv, and N is greater than M.
* So we can process expanding with only one item_lock. cool! */
//获取单个桶锁
if ((item_lock = item_trylock(expand_bucket))) {
//迁移一个桶中所有item
for (it = old_hashtable[expand_bucket]; NULL != it; it = next) {
next = it->h_next;
//重新计算哈希值
bucket = hash(ITEM_key(it), it->nkey) & hashmask(hashpower);
it->h_next = primary_hashtable[bucket];
primary_hashtable[bucket] = it;
}
old_hashtable[expand_bucket] = NULL;
expand_bucket++;
//迁移完成
if (expand_bucket == hashsize(hashpower - 1)) {
expanding = false; //将迁移标志设置0
free(old_hashtable); //释放旧表
STATS_LOCK();
stats.hash_bytes -= hashsize(hashpower - 1) * sizeof(void *);
stats.hash_is_expanding = 0;
STATS_UNLOCK();
if (settings.verbose > 1)
fprintf(stderr, "Hash table expansion done\n");
}
} else {
usleep(10*1000);
}
if (item_lock) {
item_trylock_unlock(item_lock);
item_lock = NULL;
}
}
if (!expanding) {
/* We are done expanding.. just wait for next invocation */
//不需要迁移，挂起迁移线程，直到worker线程插入数据后发现item数量已经到了1.5倍哈希表大小，
//此时调用worker线程调用assoc_start_expand函数，该函数会调用pthread_cond_signal
//唤醒迁移线程
started_expanding = false;
pthread_cond_wait(&maintenance_cond, &maintenance_lock);
/* assoc_expand() swaps out the hash table entirely, so we need
* all threads to not hold any references related to the hash
* table while this happens.
* This is instead of a more complex, possibly slower algorithm to
* allow dynamic hash table expansion without causing significant
* wait times.
*/
pause_threads(PAUSE_ALL_THREADS);
assoc_expand(); //申请更大的哈希表,并将expanding设置为true
pause_threads(RESUME_ALL_THREADS);
}
}
return NULL;
}
/* 从 slab 系统分配一个空闲 item */
item *do_item_alloc(char *key, const size_t nkey, const int flags, const rel_time_t exptime, const int nbytes,
const uint32_t cur_hv)
{
uint8_t nsuffix;
item *it = NULL;
char suffix[40];
size_t ntotal = item_make_header(nkey + 1, flags, nbytes, suffix, &nsuffix);
if (settings.use_cas) {
ntotal += sizeof(uint64_t);
}
unsigned int id = slabs_clsid(ntotal);
if (id == 0)
return 0;
mutex_lock(&cache_lock);
/* do a quick check if we have any expired items in the tail.. */
int tries = 5;
int tried_alloc = 0;
item *search;
void *hold_lock = NULL;
rel_time_t oldest_live = settings.oldest_live;
search = tails[id];
/* We walk up *only* for locked items. Never searching for expired.
* Waste of CPU for almost all deployments */
for (; tries > 0 && search != NULL; tries--, search=search->prev)
{
uint32_t hv = hash(ITEM_key(search), search->nkey, 0);
/* Attempt to hash item lock the "search" item. If locked, no
* other callers can incr the refcount
*/
/* FIXME: I think we need to mask the hv here for comparison? */
if (hv != cur_hv && (hold_lock = item_trylock(hv)) == NULL)
continue;
/* Now see if the item is refcount locked */
if (refcount_incr(&search->refcount) != 2)
{
refcount_decr(&search->refcount);
/* Old rare bug could cause a refcount leak. We haven't seen
* it in years, but we leave this code in to prevent failures
* just in case */
if (search->time + TAIL_REPAIR_TIME < current_time)
{
itemstats[id].tailrepairs++;
search->refcount = 1;
do_item_unlink_nolock(search, hv);
}
if (hold_lock)
item_trylock_unlock(hold_lock);
continue;
}
/* 先检查 LRU 队列最后一个 item 是否超时, 超时的话就把这个 item 分配给用户 */
if ((search->exptime != 0 && search->exptime < current_time)
|| (search->time <= oldest_live && oldest_live <= current_time))
{
itemstats[id].reclaimed++;
if ((search->it_flags & ITEM_FETCHED) == 0)
{
itemstats[id].expired_unfetched++;
}
it = search;
slabs_adjust_mem_requested(it->slabs_clsid, ITEM_ntotal(it), ntotal);
/* 把这个 item 从 LRU 队列和哈希表中移除 */
do_item_unlink_nolock(it, hv);
/* Initialize the item block: */
it->slabs_clsid = 0;
}
/* 没有超时的item, 那就尝试从slabclass分配, 运气不好的话, 分配失败, 那就把 LRU 队列最后一个 item 剔除, 然后分配给用户 */
else if ((it = slabs_alloc(ntotal, id)) == NULL)
{
tried_alloc = 1;
if (settings.evict_to_free == 0)
{
itemstats[id].outofmemory++;//显示出的统计信息
}
else
{
itemstats[id].evicted++;//这个slab的分配失败次数加1，后面的分析统计信息的线程会用到这个统计信息
itemstats[id].evicted_time = current_time - search->time;//显示的统计信息
if (search->exptime != 0)
itemstats[id].evicted_nonzero++;
if ((search->it_flags & ITEM_FETCHED) == 0)
{
itemstats[id].evicted_unfetched++;
}
it = search;
slabs_adjust_mem_requested(it->slabs_clsid, ITEM_ntotal(it), ntotal);//不用请求新的item了，减少相关的统计信息
/* 把老的item从hash表和lru队列中删除 */
do_item_unlink_nolock(it, hv);
/* Initialize the item block: */
it->slabs_clsid = 0;
/* If we've just 回收 an item, and the automover is set to angry bird mode, attempt to rip memory into this
slab class. TODO: Move valid object detection into a function, and on a "successful" memory pull, look
behind and see if the next alloc would be an eviction. Then kick off the slab mover before the eviction
happens.可以看到如果slab_automove=2(默认是1)，这样会导致angry模式，就是只要分配失败了，马上就选择一个slab(旧的slagclass
释放的)，把这个slab移动到当前slab-class中（不会有通过统计信息有选择的移动slab）*/
if (settings.slab_automove == 2)
slabs_reassign(-1, id);
}
}
refcount_decr(&search->refcount);
/* If hash values were equal, we don't grab a second lock */
if (hold_lock)
item_trylock_unlock(hold_lock);
break;
}
if (!tried_alloc && (tries == 0 || search == NULL)) it = slabs_alloc(ntotal, id);
if (it == NULL)
{
itemstats[id].outofmemory++;
mutex_unlock(&cache_lock);
return NULL;
}
assert(it->slabs_clsid == 0);
assert(it != heads[id]);
/* Item initialization can happen outside of the lock; the item's already been removed from the slab LRU. */
it->refcount = 1; /* the caller will have a reference */
mutex_unlock(&cache_lock);
it->next = it->prev = it->h_next = 0;
it->slabs_clsid = id;
DEBUG_REFCNT(it, '*');
it->it_flags = settings.use_cas ? ITEM_CAS : 0;
it->nkey = nkey;
it->nbytes = nbytes;
memcpy(ITEM_key(it), key, nkey);
it->exptime = exptime;
memcpy(ITEM_suffix(it), suffix, (size_t)nsuffix);
it->nsuffix = nsuffix;
return it;
}
/* refcount == 0 is safe since nobody can incr while item_lock is held.
* refcount != 0 is impossible since flags/etc can be modified in other
* threads. instead, note we found a busy one and bail. logic in do_item_get
* will prevent busy items from continuing to be busy
* NOTE: This is checking it_flags outside of an item lock. I believe this
* works since it_flags is 8 bits, and we're only ever comparing a single bit
* regardless. ITEM_SLABBED bit will always be correct since we're holding the
* lock which modifies that bit. ITEM_LINKED won't exist if we're between an
* item having ITEM_SLABBED removed, and the key hasn't been added to the item
* yet. The memory barrier from the slabs lock should order the key write and the
* flags to the item?
* If ITEM_LINKED did exist and was just removed, but we still see it, that's
* still safe since it will have a valid key, which we then lock, and then
* recheck everything.
* This may not be safe on all platforms; If not, slabs_alloc() will need to
* seed the item key while holding slabs_lock.
*/
static int slab_rebalance_move(void) {
slabclass_t *s_cls;
int x;
int was_busy = 0;
int refcount = 0;
uint32_t hv;
void *hold_lock;
enum move_status status = MOVE_PASS;
pthread_mutex_lock(&slabs_lock);
s_cls = &slabclass[slab_rebal.s_clsid];
for (x = 0; x < slab_bulk_check; x++) {
hv = 0;
hold_lock = NULL;
item *it = slab_rebal.slab_pos;
status = MOVE_PASS;
if (it->slabs_clsid != 255) {
/* ITEM_SLABBED can only be added/removed under the slabs_lock */
if (it->it_flags & ITEM_SLABBED) {
/* remove from slab freelist */
if (s_cls->slots == it) {
s_cls->slots = it->next;
}
if (it->next) it->next->prev = it->prev;
if (it->prev) it->prev->next = it->next;
s_cls->sl_curr--;
status = MOVE_FROM_SLAB;
} else if ((it->it_flags & ITEM_LINKED) != 0) {
/* If it doesn't have ITEM_SLABBED, the item could be in any
* state on its way to being freed or written to. If no
* ITEM_SLABBED, but it's had ITEM_LINKED, it must be active
* and have the key written to it already.
*/
hv = hash(ITEM_key(it), it->nkey);
if ((hold_lock = item_trylock(hv)) == NULL) {
status = MOVE_LOCKED;
} else {
refcount = refcount_incr(&it->refcount);
if (refcount == 2) { /* item is linked but not busy */
/* Double check ITEM_LINKED flag here, since we're
* past a memory barrier from the mutex. */
if ((it->it_flags & ITEM_LINKED) != 0) {
status = MOVE_FROM_LRU;
} else {
/* refcount == 1 + !ITEM_LINKED means the item is being
* uploaded to, or was just unlinked but hasn't been freed
* yet. Let it bleed off on its own and try again later */
status = MOVE_BUSY;
}
} else {
if (settings.verbose > 2) {
fprintf(stderr, "Slab reassign hit a busy item: refcount: %d (%d -> %d)\n",
it->refcount, slab_rebal.s_clsid, slab_rebal.d_clsid);
}
status = MOVE_BUSY;
}
/* Item lock must be held while modifying refcount */
if (status == MOVE_BUSY) {
refcount_decr(&it->refcount);
item_trylock_unlock(hold_lock);
}
}
}
}
switch (status) {
case MOVE_FROM_LRU:
/* Lock order is LRU locks -> slabs_lock. unlink uses LRU lock.
* We only need to hold the slabs_lock while initially looking
* at an item, and at this point we have an exclusive refcount
* (2) + the item is locked. Drop slabs lock, drop item to
* refcount 1 (just our own, then fall through and wipe it
*/
pthread_mutex_unlock(&slabs_lock);
do_item_unlink(it, hv);
item_trylock_unlock(hold_lock);
pthread_mutex_lock(&slabs_lock);
case MOVE_FROM_SLAB:
it->refcount = 0;
it->it_flags = 0;
it->slabs_clsid = 255;
break;
case MOVE_BUSY:
case MOVE_LOCKED:
slab_rebal.busy_items++;
was_busy++;
break;
case MOVE_PASS:
break;
}
slab_rebal.slab_pos = (char *)slab_rebal.slab_pos + s_cls->size;
if (slab_rebal.slab_pos >= slab_rebal.slab_end)
break;
}
if (slab_rebal.slab_pos >= slab_rebal.slab_end) {
/* Some items were busy, start again from the top */
if (slab_rebal.busy_items) {
slab_rebal.slab_pos = slab_rebal.slab_start;
slab_rebal.busy_items = 0;
} else {
slab_rebal.done++;
}
}
pthread_mutex_unlock(&slabs_lock);
return was_busy;
}
static void *assoc_maintenance_thread(void *arg) {
mutex_lock(&maintenance_lock);
//do_run_maintenance_thread是全局变量，初始值为1，在stop_assoc_maintenance_thread
//函数中会被赋值0，终止迁移线程
while (do_run_maintenance_thread) {
int ii = 0;
/* There is only one expansion thread, so no need to global lock. */
for (ii = 0; ii < hash_bulk_move && expanding; ++ii) {
item *it, *next;
int bucket;
void *item_lock = NULL;
/* bucket = hv & hashmask(hashpower) =>the bucket of hash table
* is the lowest N bits of the hv, and the bucket of item_locks is
* also the lowest M bits of hv, and N is greater than M.
* So we can process expanding with only one item_lock. cool! */
//hash_bulk_move用来控制每次迁移，移动多少个桶的item。默认是一个.
//如果expanding为true才会进入循环体，所以迁移线程刚创建的时候，并不会进入循环体
if ((item_lock = item_trylock(expand_bucket))) {
//在assoc_expand函数中expand_bucket会被赋值0
//遍历旧哈希表中由expand_bucket指明的桶,将该桶的所有item
//迁移到新哈希表中。
for (it = old_hashtable[expand_bucket]; NULL != it; it = next) {
next = it->h_next;
//重新计算新的哈希值,得到其在新哈希表的位置
bucket = hash(ITEM_key(it), it->nkey) & hashmask(hashpower);
//将这个item插入到新哈希表中
it->h_next = primary_hashtable[bucket];
primary_hashtable[bucket] = it;
}
//不需要清空旧桶。直接将冲突链的链头赋值为NULL即可
old_hashtable[expand_bucket] = NULL;
//迁移完一个桶，接着把expand_bucket指向下一个待迁移的桶
expand_bucket++;
if (expand_bucket == hashsize(hashpower - 1)) {//全部数据迁移完毕
expanding = false;//将扩展标志设置为false
free(old_hashtable);
STATS_LOCK();
stats.hash_bytes -= hashsize(hashpower - 1) * sizeof(void *);
stats.hash_is_expanding = 0;
STATS_UNLOCK();
if (settings.verbose > 1)
fprintf(stderr, "Hash table expansion done\n");
}
} else {
usleep(10*1000);
}
if (item_lock) {
//遍历完hash_bulk_move个桶的所有item后，就释放锁
item_trylock_unlock(item_lock);
item_lock = NULL;
}
}
if (!expanding) {//不需要迁移数据(了)
/* We are done expanding.. just wait for next invocation */
started_expanding = false;//重置
//挂起迁移线程，直到worker线程插入数据后发现item数量已经到了1.5倍哈希表大小，
//此时调用worker线程调用assoc_start_expand函数，该函数会调用pthread_cond_signal
//唤醒迁移线程
pthread_cond_wait(&maintenance_cond, &maintenance_lock);
/* assoc_expand() swaps out the hash table entirely, so we need
* all threads to not hold any references related to the hash
* table while this happens.
* This is instead of a more complex, possibly slower algorithm to
* allow dynamic hash table expansion without causing significant
* wait times.
*/
pause_threads(PAUSE_ALL_THREADS);
assoc_expand();//申请更大的哈希表,并将expanding设置为true
pause_threads(RESUME_ALL_THREADS);
}
}
return NULL;
}
