* Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*
- * Copyright (c) 2011, 2014, Intel Corporation.
+ * Copyright (c) 2011, 2015, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
#define LU_SITE_BITS_MIN 12
#define LU_SITE_BITS_MAX 24
+#define LU_SITE_BITS_MAX_CL 19
/**
* total 256 buckets, we don't want too many buckets because:
* - consume too much memory
"Maximum number of objects in lu_object cache");
static void lu_object_free(const struct lu_env *env, struct lu_object *o);
+static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx);
/**
* Decrease reference counter on object. If last reference is freed, return
struct lu_object_header *top;
struct lu_site *site;
struct lu_object *orig;
- cfs_hash_bd_t bd;
+ struct cfs_hash_bd bd;
const struct lu_fid *fid;
top = o->lo_header;
LASSERT(list_empty(&top->loh_lru));
list_add_tail(&top->loh_lru, &bkt->lsb_lru);
bkt->lsb_lru_len++;
+ lprocfs_counter_incr(site->ls_stats, LU_SS_LRU_LEN);
+ CDEBUG(D_INODE, "Add %p to site lru. hash: %p, bkt: %p, "
+ "lru_len: %ld\n",
+ o, site->ls_obj_hash, bkt, bkt->lsb_lru_len);
cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
return;
}
/*
- * If object is dying (will not be cached), removed it
+ * If object is dying (will not be cached) then remove it
* from hash table and LRU.
*
* This is done with hash table and LRU lists locked. As the only
top = o->lo_header;
set_bit(LU_OBJECT_HEARD_BANSHEE, &top->loh_flags);
if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags)) {
- cfs_hash_t *obj_hash = o->lo_dev->ld_site->ls_obj_hash;
- cfs_hash_bd_t bd;
+ struct lu_site *site = o->lo_dev->ld_site;
+ struct cfs_hash *obj_hash = site->ls_obj_hash;
+ struct cfs_hash_bd bd;
cfs_hash_bd_get_and_lock(obj_hash, &top->loh_fid, &bd, 1);
if (!list_empty(&top->loh_lru)) {
list_del_init(&top->loh_lru);
bkt = cfs_hash_bd_extra_get(obj_hash, &bd);
bkt->lsb_lru_len--;
+ lprocfs_counter_decr(site->ls_stats, LU_SS_LRU_LEN);
}
cfs_hash_bd_del_locked(obj_hash, &bd, &top->loh_hash);
cfs_hash_bd_unlock(obj_hash, &bd, 1);
struct lu_object_header *h;
struct lu_object_header *temp;
struct lu_site_bkt_data *bkt;
- cfs_hash_bd_t bd;
- cfs_hash_bd_t bd2;
+ struct cfs_hash_bd bd;
+ struct cfs_hash_bd bd2;
struct list_head dispose;
int did_sth;
- unsigned int start;
+ unsigned int start = 0;
int count;
int bnr;
unsigned int i;
* Under LRU list lock, scan LRU list and move unreferenced objects to
* the dispose list, removing them from LRU and hash table.
*/
- start = s->ls_purge_start;
+ if (nr != ~0)
+ start = s->ls_purge_start;
bnr = (nr == ~0) ? -1 : nr / (int)CFS_HASH_NBKT(s->ls_obj_hash) + 1;
again:
/*
&bd2, &h->loh_hash);
list_move(&h->loh_lru, &dispose);
bkt->lsb_lru_len--;
+ lprocfs_counter_decr(s->ls_stats, LU_SS_LRU_LEN);
if (did_sth == 0)
did_sth = 1;
}
static struct lu_object *htable_lookup(struct lu_site *s,
- cfs_hash_bd_t *bd,
+ struct cfs_hash_bd *bd,
const struct lu_fid *f,
wait_queue_t *waiter,
__u64 *version)
if (!list_empty(&h->loh_lru)) {
list_del_init(&h->loh_lru);
bkt->lsb_lru_len--;
+ lprocfs_counter_decr(s->ls_stats, LU_SS_LRU_LEN);
}
return lu_object_top(h);
}
*/
if (likely(waiter != NULL)) {
- init_waitqueue_entry_current(waiter);
+ init_waitqueue_entry(waiter, current);
add_wait_queue(&bkt->lsb_marche_funebre, waiter);
set_current_state(TASK_UNINTERRUPTIBLE);
lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_DEATH_RACE);
const struct lu_object_conf *conf)
{
struct lu_object *o;
- cfs_hash_t *hs;
- cfs_hash_bd_t bd;
+ struct cfs_hash *hs;
+ struct cfs_hash_bd bd;
o = lu_object_alloc(env, dev, f, conf);
if (unlikely(IS_ERR(o)))
struct lu_object *o;
struct lu_object *shadow;
struct lu_site *s;
- cfs_hash_t *hs;
- cfs_hash_bd_t bd;
+ struct cfs_hash *hs;
+ struct cfs_hash_bd bd;
__u64 version = 0;
/*
* lu_object_find_try() already added waiter into the
* wait queue.
*/
- waitq_wait(&wait, TASK_UNINTERRUPTIBLE);
+ schedule();
bkt = lu_site_bkt_from_fid(dev->ld_site, (void *)f);
remove_wait_queue(&bkt->lsb_marche_funebre, &wait);
}
};
static int
-lu_site_obj_print(cfs_hash_t *hs, cfs_hash_bd_t *bd,
+lu_site_obj_print(struct cfs_hash *hs, struct cfs_hash_bd *bd,
struct hlist_node *hnode, void *data)
{
struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
{
unsigned long cache_size;
unsigned long bits;
+ unsigned long bits_max = LU_SITE_BITS_MAX;
/*
* For ZFS based OSDs the cache should be disabled by default. This
return LU_SITE_BITS_MIN;
}
+ if (strcmp(top->ld_type->ldt_name, LUSTRE_VVP_NAME) == 0)
+ bits_max = LU_SITE_BITS_MAX_CL;
+
/*
* Calculate hash table size, assuming that we want reasonable
* performance when 20% of total memory is occupied by cache of
for (bits = 1; (1 << bits) < cache_size; ++bits) {
;
}
- return bits;
+
+ return clamp_t(typeof(bits), bits, LU_SITE_BITS_MIN, bits_max);
}
-static unsigned lu_obj_hop_hash(cfs_hash_t *hs,
+static unsigned lu_obj_hop_hash(struct cfs_hash *hs,
const void *key, unsigned mask)
{
struct lu_fid *fid = (struct lu_fid *)key;
return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
}
-static void lu_obj_hop_get(cfs_hash_t *hs, struct hlist_node *hnode)
+static void lu_obj_hop_get(struct cfs_hash *hs, struct hlist_node *hnode)
{
struct lu_object_header *h;
atomic_inc(&h->loh_ref);
}
-static void lu_obj_hop_put_locked(cfs_hash_t *hs, struct hlist_node *hnode)
+static void lu_obj_hop_put_locked(struct cfs_hash *hs, struct hlist_node *hnode)
{
LBUG(); /* we should never called it */
}
-static cfs_hash_ops_t lu_site_hash_ops = {
+static struct cfs_hash_ops lu_site_hash_ops = {
.hs_hash = lu_obj_hop_hash,
.hs_key = lu_obj_hop_key,
.hs_keycmp = lu_obj_hop_keycmp,
int lu_site_init(struct lu_site *s, struct lu_device *top)
{
struct lu_site_bkt_data *bkt;
- cfs_hash_bd_t bd;
+ struct cfs_hash_bd bd;
char name[16];
unsigned long bits;
unsigned int i;
memset(s, 0, sizeof *s);
mutex_init(&s->ls_purge_mutex);
- bits = lu_htable_order(top);
snprintf(name, sizeof(name), "lu_site_%s", top->ld_type->ldt_name);
- for (bits = clamp_t(typeof(bits), bits,
- LU_SITE_BITS_MIN, LU_SITE_BITS_MAX);
+ for (bits = lu_htable_order(top);
bits >= LU_SITE_BITS_MIN; bits--) {
s->ls_obj_hash = cfs_hash_create(name, bits, bits,
bits - LU_SITE_BKT_BITS,
0, "cache_death_race", "cache_death_race");
lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
0, "lru_purged", "lru_purged");
+ /*
+ * Unlike other counters, lru_len can be decremented so
+ * need lc_sum instead of just lc_count
+ */
+ lprocfs_counter_init(s->ls_stats, LU_SS_LRU_LEN,
+ LPROCFS_CNTR_AVGMINMAX, "lru_len", "lru_len");
INIT_LIST_HEAD(&s->ls_linkage);
s->ls_top_dev = top;
static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
-static DEFINE_SPINLOCK(lu_keys_guard);
+DEFINE_RWLOCK(lu_keys_guard);
+static atomic_t lu_key_initing_cnt = ATOMIC_INIT(0);
/**
* Global counter incremented whenever key is registered, unregistered,
LASSERT(key->lct_owner != NULL);
result = -ENFILE;
- spin_lock(&lu_keys_guard);
+ write_lock(&lu_keys_guard);
for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
if (lu_keys[i] == NULL) {
key->lct_index = i;
break;
}
}
- spin_unlock(&lu_keys_guard);
+ write_unlock(&lu_keys_guard);
return result;
}
EXPORT_SYMBOL(lu_context_key_register);
lu_context_key_quiesce(key);
++key_set_version;
- spin_lock(&lu_keys_guard);
+ write_lock(&lu_keys_guard);
key_fini(&lu_shrink_env.le_ctx, key->lct_index);
+
+ /**
+ * Wait until all transient contexts referencing this key have
+ * run lu_context_key::lct_fini() method.
+ */
+ while (atomic_read(&key->lct_used) > 1) {
+ write_unlock(&lu_keys_guard);
+ CDEBUG(D_INFO, "lu_context_key_degister: \"%s\" %p, %d\n",
+ key->lct_owner ? key->lct_owner->name : "", key,
+ atomic_read(&key->lct_used));
+ schedule();
+ write_lock(&lu_keys_guard);
+ }
if (lu_keys[key->lct_index]) {
lu_keys[key->lct_index] = NULL;
lu_ref_fini(&key->lct_reference);
}
- spin_unlock(&lu_keys_guard);
+ write_unlock(&lu_keys_guard);
LASSERTF(atomic_read(&key->lct_used) == 1,
"key has instances: %d\n",
* XXX layering violation.
*/
cl_env_cache_purge(~0);
- key->lct_tags |= LCT_QUIESCENT;
/*
* XXX memory barrier has to go here.
*/
- spin_lock(&lu_keys_guard);
+ write_lock(&lu_keys_guard);
+ key->lct_tags |= LCT_QUIESCENT;
+
+ /**
+ * Wait until all lu_context_key::lct_init() methods
+ * have completed.
+ */
+ while (atomic_read(&lu_key_initing_cnt) > 0) {
+ write_unlock(&lu_keys_guard);
+ CDEBUG(D_INFO, "lu_context_key_quiesce: \"%s\""
+ " %p, %d (%d)\n",
+ key->lct_owner ? key->lct_owner->name : "",
+ key, atomic_read(&key->lct_used),
+ atomic_read(&lu_key_initing_cnt));
+ schedule();
+ write_lock(&lu_keys_guard);
+ }
+
list_for_each_entry(ctx, &lu_context_remembered,
lc_remember)
key_fini(ctx, key->lct_index);
- spin_unlock(&lu_keys_guard);
+ write_unlock(&lu_keys_guard);
++key_set_version;
}
}
{
unsigned int i;
+ /*
+ * A serialisation with lu_context_key_quiesce() is needed, but some
+ * "key->lct_init()" are calling kernel memory allocation routine and
+ * can't be called while holding a spin_lock.
+ * "lu_keys_guard" is held while incrementing "lu_key_initing_cnt"
+ * to ensure the start of the serialisation.
+ * An atomic_t variable is still used, in order not to reacquire the
+ * lock when decrementing the counter.
+ */
+ read_lock(&lu_keys_guard);
+ atomic_inc(&lu_key_initing_cnt);
+ read_unlock(&lu_keys_guard);
+
LINVRNT(ctx->lc_value != NULL);
for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
struct lu_context_key *key;
LINVRNT(key->lct_init != NULL);
LINVRNT(key->lct_index == i);
- value = key->lct_init(ctx, key);
- if (unlikely(IS_ERR(value)))
- return PTR_ERR(value);
-
LASSERT(key->lct_owner != NULL);
- if (!(ctx->lc_tags & LCT_NOREF))
- try_module_get(key->lct_owner);
+ if (!(ctx->lc_tags & LCT_NOREF) &&
+ try_module_get(key->lct_owner) == 0) {
+ /* module is unloading, skip this key */
+ continue;
+ }
+
+ value = key->lct_init(ctx, key);
+ if (unlikely(IS_ERR(value))) {
+ atomic_dec(&lu_key_initing_cnt);
+ return PTR_ERR(value);
+ }
+
lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
atomic_inc(&key->lct_used);
/*
}
ctx->lc_version = key_set_version;
}
+ atomic_dec(&lu_key_initing_cnt);
return 0;
}
ctx->lc_state = LCS_INITIALIZED;
ctx->lc_tags = tags;
if (tags & LCT_REMEMBER) {
- spin_lock(&lu_keys_guard);
+ write_lock(&lu_keys_guard);
list_add(&ctx->lc_remember, &lu_context_remembered);
- spin_unlock(&lu_keys_guard);
+ write_unlock(&lu_keys_guard);
} else {
INIT_LIST_HEAD(&ctx->lc_remember);
}
keys_fini(ctx);
} else { /* could race with key degister */
- spin_lock(&lu_keys_guard);
+ write_lock(&lu_keys_guard);
keys_fini(ctx);
list_del_init(&ctx->lc_remember);
- spin_unlock(&lu_keys_guard);
+ write_unlock(&lu_keys_guard);
}
}
EXPORT_SYMBOL(lu_context_fini);
for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
/* could race with key quiescency */
if (ctx->lc_tags & LCT_REMEMBER)
- spin_lock(&lu_keys_guard);
+ read_lock(&lu_keys_guard);
if (ctx->lc_value[i] != NULL) {
struct lu_context_key *key;
key, ctx->lc_value[i]);
}
if (ctx->lc_tags & LCT_REMEMBER)
- spin_unlock(&lu_keys_guard);
+ read_unlock(&lu_keys_guard);
}
}
}
void lu_context_tags_update(__u32 tags)
{
- spin_lock(&lu_keys_guard);
+ write_lock(&lu_keys_guard);
lu_context_tags_default |= tags;
key_set_version++;
- spin_unlock(&lu_keys_guard);
+ write_unlock(&lu_keys_guard);
}
EXPORT_SYMBOL(lu_context_tags_update);
void lu_context_tags_clear(__u32 tags)
{
- spin_lock(&lu_keys_guard);
+ write_lock(&lu_keys_guard);
lu_context_tags_default &= ~tags;
key_set_version++;
- spin_unlock(&lu_keys_guard);
+ write_unlock(&lu_keys_guard);
}
EXPORT_SYMBOL(lu_context_tags_clear);
void lu_session_tags_update(__u32 tags)
{
- spin_lock(&lu_keys_guard);
+ write_lock(&lu_keys_guard);
lu_session_tags_default |= tags;
key_set_version++;
- spin_unlock(&lu_keys_guard);
+ write_unlock(&lu_keys_guard);
}
EXPORT_SYMBOL(lu_session_tags_update);
void lu_session_tags_clear(__u32 tags)
{
- spin_lock(&lu_keys_guard);
+ write_lock(&lu_keys_guard);
lu_session_tags_default &= ~tags;
key_set_version++;
- spin_unlock(&lu_keys_guard);
+ write_unlock(&lu_keys_guard);
}
EXPORT_SYMBOL(lu_session_tags_clear);
unsigned lss_busy;
} lu_site_stats_t;
-static void lu_site_stats_get(cfs_hash_t *hs,
+static void lu_site_stats_get(struct cfs_hash *hs,
lu_site_stats_t *stats, int populated)
{
- cfs_hash_bd_t bd;
+ struct cfs_hash_bd bd;
unsigned int i;
cfs_hash_for_each_bucket(hs, &bd, i) {
}
+/*
+ * lu_cache_shrink_count returns the number of cached objects that are
+ * candidates to be freed by shrink_slab(). A counter, which tracks
+ * the number of items in the site's lru, is maintained in the per cpu
+ * stats of each site. The counter is incremented when an object is added
+ * to a site's lru and decremented when one is removed. The number of
+ * free-able objects is the sum of all per cpu counters for all sites.
+ *
+ * Using a per cpu counter is a compromise solution to concurrent access:
+ * lu_object_put() can update the counter without locking the site and
+ * lu_cache_shrink_count can sum the counters without locking each
+ * ls_obj_hash bucket.
+ */
static unsigned long lu_cache_shrink_count(struct shrinker *sk,
struct shrink_control *sc)
{
- lu_site_stats_t stats;
struct lu_site *s;
struct lu_site *tmp;
unsigned long cached = 0;
mutex_lock(&lu_sites_guard);
list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
- memset(&stats, 0, sizeof(stats));
- lu_site_stats_get(s->ls_obj_hash, &stats, 0);
- cached += stats.lss_total - stats.lss_busy;
+ cached += ls_stats_read(s->ls_stats, LU_SS_LRU_LEN);
}
mutex_unlock(&lu_sites_guard);
cached = (cached / 100) * sysctl_vfs_cache_pressure;
- CDEBUG(D_INODE, "%ld objects cached\n", cached);
+ CDEBUG(D_INODE, "%ld objects cached, cache pressure %d\n",
+ cached, sysctl_vfs_cache_pressure);
+
return cached;
}
* is safe to take the lu_sites_guard lock.
*
* Ideally we should accurately return the remaining number of cached
- * objects without taking the lu_sites_guard lock, but this is not
+ * objects without taking the lu_sites_guard lock, but this is not
* possible in the current implementation.
*/
static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
CDEBUG(D_INODE, "Shrink %lu objects\n", scv.nr_to_scan);
- lu_cache_shrink_scan(shrinker, &scv);
+ if (scv.nr_to_scan != 0)
+ lu_cache_shrink_scan(shrinker, &scv);
cached = lu_cache_shrink_count(shrinker, &scv);
- if (scv.nr_to_scan == 0)
- CDEBUG(D_INODE, "%d objects cached\n", cached);
return cached;
}
static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
{
#ifdef CONFIG_PROC_FS
- struct lprocfs_counter ret;
+ struct lprocfs_counter ret;
- lprocfs_stats_collect(stats, idx, &ret);
- return (__u32)ret.lc_count;
+ lprocfs_stats_collect(stats, idx, &ret);
+ if (idx == LU_SS_LRU_LEN)
+ /*
+ * protect against counter on cpu A being decremented
+ * before counter is incremented on cpu B; unlikely
+ */
+ return (__u32)((ret.lc_sum > 0) ? ret.lc_sum : 0);
+ else
+ return (__u32)ret.lc_count;
#else
- return 0;
+ return 0;
#endif
}
memset(&stats, 0, sizeof(stats));
lu_site_stats_get(s->ls_obj_hash, &stats, 1);
- return seq_printf(m, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
+ return seq_printf(m, "%d/%d %d/%d %d %d %d %d %d %d %d %d\n",
stats.lss_busy,
stats.lss_total,
stats.lss_populated,
ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
- ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
+ ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED),
+ ls_stats_read(s->ls_stats, LU_SS_LRU_LEN));
}
EXPORT_SYMBOL(lu_site_stats_seq_print);
-int lu_site_stats_print(const struct lu_site *s, char *page, int count)
-{
- lu_site_stats_t stats;
-
- memset(&stats, 0, sizeof(stats));
- lu_site_stats_get(s->ls_obj_hash, &stats, 1);
-
- return snprintf(page, count, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
- stats.lss_busy,
- stats.lss_total,
- stats.lss_populated,
- CFS_HASH_NHLIST(s->ls_obj_hash),
- stats.lss_max_search,
- ls_stats_read(s->ls_stats, LU_SS_CREATED),
- ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
- ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
- ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
- ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
- ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
-}
-
/**
* Helper function to initialize a number of kmem slab caches at once.
*/
struct lu_fid *old = &o->lo_header->loh_fid;
struct lu_object *shadow;
wait_queue_t waiter;
- cfs_hash_t *hs;
- cfs_hash_bd_t bd;
+ struct cfs_hash *hs;
+ struct cfs_hash_bd bd;
__u64 version = 0;
LASSERT(fid_is_zero(old));
git://git.whamcloud.com / fs / lustre-release.git / blobdiff