summaryrefslogtreecommitdiff
path: root/drivers/staging/lustre/lustre/ldlm/ldlm_pool.c
diff options
context:
space:
mode:
Diffstat (limited to 'drivers/staging/lustre/lustre/ldlm/ldlm_pool.c')
-rw-r--r--drivers/staging/lustre/lustre/ldlm/ldlm_pool.c283
1 files changed, 162 insertions, 121 deletions
diff --git a/drivers/staging/lustre/lustre/ldlm/ldlm_pool.c b/drivers/staging/lustre/lustre/ldlm/ldlm_pool.c
index b3b60288e5f5..0025ee6356da 100644
--- a/drivers/staging/lustre/lustre/ldlm/ldlm_pool.c
+++ b/drivers/staging/lustre/lustre/ldlm/ldlm_pool.c
@@ -142,7 +142,7 @@
*/
#define LDLM_POOL_SLV_SHIFT (10)
-extern proc_dir_entry_t *ldlm_ns_proc_dir;
+extern struct proc_dir_entry *ldlm_ns_proc_dir;
static inline __u64 dru(__u64 val, __u32 shift, int round_up)
{
@@ -335,17 +335,16 @@ static void ldlm_srv_pool_push_slv(struct ldlm_pool *pl)
static int ldlm_srv_pool_recalc(struct ldlm_pool *pl)
{
time_t recalc_interval_sec;
- ENTRY;
recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
if (recalc_interval_sec < pl->pl_recalc_period)
- RETURN(0);
+ return 0;
spin_lock(&pl->pl_lock);
recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
if (recalc_interval_sec < pl->pl_recalc_period) {
spin_unlock(&pl->pl_lock);
- RETURN(0);
+ return 0;
}
/*
* Recalc SLV after last period. This should be done
@@ -367,7 +366,7 @@ static int ldlm_srv_pool_recalc(struct ldlm_pool *pl)
lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
recalc_interval_sec);
spin_unlock(&pl->pl_lock);
- RETURN(0);
+ return 0;
}
/**
@@ -394,7 +393,7 @@ static int ldlm_srv_pool_shrink(struct ldlm_pool *pl,
* and can't cancel anything. Let's catch this race.
*/
if (atomic_read(&pl->pl_granted) == 0)
- RETURN(0);
+ return 0;
spin_lock(&pl->pl_lock);
@@ -473,11 +472,10 @@ static void ldlm_cli_pool_pop_slv(struct ldlm_pool *pl)
static int ldlm_cli_pool_recalc(struct ldlm_pool *pl)
{
time_t recalc_interval_sec;
- ENTRY;
recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
if (recalc_interval_sec < pl->pl_recalc_period)
- RETURN(0);
+ return 0;
spin_lock(&pl->pl_lock);
/*
@@ -486,7 +484,7 @@ static int ldlm_cli_pool_recalc(struct ldlm_pool *pl)
recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
if (recalc_interval_sec < pl->pl_recalc_period) {
spin_unlock(&pl->pl_lock);
- RETURN(0);
+ return 0;
}
/*
@@ -503,7 +501,7 @@ static int ldlm_cli_pool_recalc(struct ldlm_pool *pl)
* Do not cancel locks in case lru resize is disabled for this ns.
*/
if (!ns_connect_lru_resize(ldlm_pl2ns(pl)))
- RETURN(0);
+ return 0;
/*
* In the time of canceling locks on client we do not need to maintain
@@ -511,8 +509,7 @@ static int ldlm_cli_pool_recalc(struct ldlm_pool *pl)
* It may be called when SLV has changed much, this is why we do not
* take into account pl->pl_recalc_time here.
*/
- RETURN(ldlm_cancel_lru(ldlm_pl2ns(pl), 0, LCF_ASYNC,
- LDLM_CANCEL_LRUR));
+ return ldlm_cancel_lru(ldlm_pl2ns(pl), 0, LCF_ASYNC, LDLM_CANCEL_LRUR);
}
/**
@@ -524,7 +521,7 @@ static int ldlm_cli_pool_shrink(struct ldlm_pool *pl,
int nr, unsigned int gfp_mask)
{
struct ldlm_namespace *ns;
- int canceled = 0, unused;
+ int unused;
ns = ldlm_pl2ns(pl);
@@ -532,7 +529,7 @@ static int ldlm_cli_pool_shrink(struct ldlm_pool *pl,
* Do not cancel locks in case lru resize is disabled for this ns.
*/
if (!ns_connect_lru_resize(ns))
- RETURN(0);
+ return 0;
/*
* Make sure that pool knows last SLV and Limit from obd.
@@ -543,14 +540,10 @@ static int ldlm_cli_pool_shrink(struct ldlm_pool *pl,
unused = ns->ns_nr_unused;
spin_unlock(&ns->ns_lock);
- if (nr) {
- canceled = ldlm_cancel_lru(ns, nr, LCF_ASYNC,
- LDLM_CANCEL_SHRINK);
- }
- /*
- * Return the number of potentially reclaimable locks.
- */
- return ((unused - canceled) / 100) * sysctl_vfs_cache_pressure;
+ if (nr == 0)
+ return (unused / 100) * sysctl_vfs_cache_pressure;
+ else
+ return ldlm_cancel_lru(ns, nr, LCF_ASYNC, LDLM_CANCEL_SHRINK);
}
struct ldlm_pool_ops ldlm_srv_pool_ops = {
@@ -578,7 +571,6 @@ int ldlm_pool_recalc(struct ldlm_pool *pl)
goto recalc;
spin_lock(&pl->pl_lock);
- recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
if (recalc_interval_sec > 0) {
/*
* Update pool statistics every 1s.
@@ -598,16 +590,17 @@ int ldlm_pool_recalc(struct ldlm_pool *pl)
count = pl->pl_ops->po_recalc(pl);
lprocfs_counter_add(pl->pl_stats, LDLM_POOL_RECALC_STAT,
count);
- return count;
}
+ recalc_interval_sec = pl->pl_recalc_time - cfs_time_current_sec() +
+ pl->pl_recalc_period;
- return 0;
+ return recalc_interval_sec;
}
-EXPORT_SYMBOL(ldlm_pool_recalc);
-/**
+/*
* Pool shrink wrapper. Will call either client or server pool recalc callback
- * depending what pool \a pl is used.
+ * depending what pool pl is used. When nr == 0, just return the number of
+ * freeable locks. Otherwise, return the number of canceled locks.
*/
int ldlm_pool_shrink(struct ldlm_pool *pl, int nr,
unsigned int gfp_mask)
@@ -734,11 +727,10 @@ static int ldlm_pool_proc_init(struct ldlm_pool *pl)
struct lprocfs_vars pool_vars[2];
char *var_name = NULL;
int rc = 0;
- ENTRY;
OBD_ALLOC(var_name, MAX_STRING_SIZE + 1);
if (!var_name)
- RETURN(-ENOMEM);
+ return -ENOMEM;
parent_ns_proc = ns->ns_proc_dir_entry;
if (parent_ns_proc == NULL) {
@@ -751,6 +743,7 @@ static int ldlm_pool_proc_init(struct ldlm_pool *pl)
if (IS_ERR(pl->pl_proc_dir)) {
CERROR("LProcFS failed in ldlm-pool-init\n");
rc = PTR_ERR(pl->pl_proc_dir);
+ pl->pl_proc_dir = NULL;
GOTO(out_free_name, rc);
}
@@ -813,7 +806,6 @@ static int ldlm_pool_proc_init(struct ldlm_pool *pl)
"recalc_timing", "sec");
rc = lprocfs_register_stats(pl->pl_proc_dir, "stats", pl->pl_stats);
- EXIT;
out_free_name:
OBD_FREE(var_name, MAX_STRING_SIZE + 1);
return rc;
@@ -835,7 +827,6 @@ int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
int idx, ldlm_side_t client)
{
int rc;
- ENTRY;
spin_lock_init(&pl->pl_lock);
atomic_set(&pl->pl_granted, 0);
@@ -863,17 +854,16 @@ int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
pl->pl_client_lock_volume = 0;
rc = ldlm_pool_proc_init(pl);
if (rc)
- RETURN(rc);
+ return rc;
CDEBUG(D_DLMTRACE, "Lock pool %s is initialized\n", pl->pl_name);
- RETURN(rc);
+ return rc;
}
EXPORT_SYMBOL(ldlm_pool_init);
void ldlm_pool_fini(struct ldlm_pool *pl)
{
- ENTRY;
ldlm_pool_proc_fini(pl);
/*
@@ -882,7 +872,6 @@ void ldlm_pool_fini(struct ldlm_pool *pl)
* any abnormal using cases.
*/
POISON(pl, 0x5a, sizeof(*pl));
- EXIT;
}
EXPORT_SYMBOL(ldlm_pool_fini);
@@ -1025,37 +1014,32 @@ static int ldlm_pool_granted(struct ldlm_pool *pl)
}
static struct ptlrpc_thread *ldlm_pools_thread;
-static struct shrinker *ldlm_pools_srv_shrinker;
-static struct shrinker *ldlm_pools_cli_shrinker;
static struct completion ldlm_pools_comp;
/*
- * Cancel \a nr locks from all namespaces (if possible). Returns number of
- * cached locks after shrink is finished. All namespaces are asked to
- * cancel approximately equal amount of locks to keep balancing.
+ * count locks from all namespaces (if possible). Returns number of
+ * cached locks.
*/
-static int ldlm_pools_shrink(ldlm_side_t client, int nr,
- unsigned int gfp_mask)
+static unsigned long ldlm_pools_count(ldlm_side_t client, unsigned int gfp_mask)
{
- int total = 0, cached = 0, nr_ns;
+ int total = 0, nr_ns;
struct ldlm_namespace *ns;
+ struct ldlm_namespace *ns_old = NULL; /* loop detection */
void *cookie;
- if (client == LDLM_NAMESPACE_CLIENT && nr != 0 &&
- !(gfp_mask & __GFP_FS))
- return -1;
+ if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS))
+ return 0;
- CDEBUG(D_DLMTRACE, "Request to shrink %d %s locks from all pools\n",
- nr, client == LDLM_NAMESPACE_CLIENT ? "client" : "server");
+ CDEBUG(D_DLMTRACE, "Request to count %s locks from all pools\n",
+ client == LDLM_NAMESPACE_CLIENT ? "client" : "server");
cookie = cl_env_reenter();
/*
* Find out how many resources we may release.
*/
- for (nr_ns = atomic_read(ldlm_namespace_nr(client));
- nr_ns > 0; nr_ns--)
- {
+ for (nr_ns = ldlm_namespace_nr_read(client);
+ nr_ns > 0; nr_ns--) {
mutex_lock(ldlm_namespace_lock(client));
if (list_empty(ldlm_namespace_list(client))) {
mutex_unlock(ldlm_namespace_lock(client));
@@ -1063,24 +1047,49 @@ static int ldlm_pools_shrink(ldlm_side_t client, int nr,
return 0;
}
ns = ldlm_namespace_first_locked(client);
+
+ if (ns == ns_old) {
+ mutex_unlock(ldlm_namespace_lock(client));
+ break;
+ }
+
+ if (ldlm_ns_empty(ns)) {
+ ldlm_namespace_move_to_inactive_locked(ns, client);
+ mutex_unlock(ldlm_namespace_lock(client));
+ continue;
+ }
+
+ if (ns_old == NULL)
+ ns_old = ns;
+
ldlm_namespace_get(ns);
- ldlm_namespace_move_locked(ns, client);
+ ldlm_namespace_move_to_active_locked(ns, client);
mutex_unlock(ldlm_namespace_lock(client));
total += ldlm_pool_shrink(&ns->ns_pool, 0, gfp_mask);
ldlm_namespace_put(ns);
}
- if (nr == 0 || total == 0) {
- cl_env_reexit(cookie);
- return total;
- }
+ cl_env_reexit(cookie);
+ return total;
+}
+
+static unsigned long ldlm_pools_scan(ldlm_side_t client, int nr, unsigned int gfp_mask)
+{
+ unsigned long freed = 0;
+ int tmp, nr_ns;
+ struct ldlm_namespace *ns;
+ void *cookie;
+
+ if (client == LDLM_NAMESPACE_CLIENT && !(gfp_mask & __GFP_FS))
+ return -1;
+
+ cookie = cl_env_reenter();
/*
- * Shrink at least ldlm_namespace_nr(client) namespaces.
+ * Shrink at least ldlm_namespace_nr_read(client) namespaces.
*/
- for (nr_ns = atomic_read(ldlm_namespace_nr(client));
- nr_ns > 0; nr_ns--)
- {
+ for (tmp = nr_ns = ldlm_namespace_nr_read(client);
+ tmp > 0; tmp--) {
int cancel, nr_locks;
/*
@@ -1089,50 +1098,59 @@ static int ldlm_pools_shrink(ldlm_side_t client, int nr,
mutex_lock(ldlm_namespace_lock(client));
if (list_empty(ldlm_namespace_list(client))) {
mutex_unlock(ldlm_namespace_lock(client));
- /*
- * If list is empty, we can't return any @cached > 0,
- * that probably would cause needless shrinker
- * call.
- */
- cached = 0;
break;
}
ns = ldlm_namespace_first_locked(client);
ldlm_namespace_get(ns);
- ldlm_namespace_move_locked(ns, client);
+ ldlm_namespace_move_to_active_locked(ns, client);
mutex_unlock(ldlm_namespace_lock(client));
nr_locks = ldlm_pool_granted(&ns->ns_pool);
- cancel = 1 + nr_locks * nr / total;
- ldlm_pool_shrink(&ns->ns_pool, cancel, gfp_mask);
- cached += ldlm_pool_granted(&ns->ns_pool);
+ /*
+ * We use to shrink propotionally but with new shrinker API,
+ * we lost the total number of freeable locks.
+ */
+ cancel = 1 + min_t(int, nr_locks, nr / nr_ns);
+ freed += ldlm_pool_shrink(&ns->ns_pool, cancel, gfp_mask);
ldlm_namespace_put(ns);
}
cl_env_reexit(cookie);
- /* we only decrease the SLV in server pools shrinker, return -1 to
- * kernel to avoid needless loop. LU-1128 */
- return (client == LDLM_NAMESPACE_SERVER) ? -1 : cached;
+ /*
+ * we only decrease the SLV in server pools shrinker, return
+ * SHRINK_STOP to kernel to avoid needless loop. LU-1128
+ */
+ return (client == LDLM_NAMESPACE_SERVER) ? SHRINK_STOP : freed;
+}
+
+static unsigned long ldlm_pools_srv_count(struct shrinker *s, struct shrink_control *sc)
+{
+ return ldlm_pools_count(LDLM_NAMESPACE_SERVER, sc->gfp_mask);
+}
+
+static unsigned long ldlm_pools_srv_scan(struct shrinker *s, struct shrink_control *sc)
+{
+ return ldlm_pools_scan(LDLM_NAMESPACE_SERVER, sc->nr_to_scan,
+ sc->gfp_mask);
}
-static int ldlm_pools_srv_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
+static unsigned long ldlm_pools_cli_count(struct shrinker *s, struct shrink_control *sc)
{
- return ldlm_pools_shrink(LDLM_NAMESPACE_SERVER,
- shrink_param(sc, nr_to_scan),
- shrink_param(sc, gfp_mask));
+ return ldlm_pools_count(LDLM_NAMESPACE_CLIENT, sc->gfp_mask);
}
-static int ldlm_pools_cli_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
+static unsigned long ldlm_pools_cli_scan(struct shrinker *s, struct shrink_control *sc)
{
- return ldlm_pools_shrink(LDLM_NAMESPACE_CLIENT,
- shrink_param(sc, nr_to_scan),
- shrink_param(sc, gfp_mask));
+ return ldlm_pools_scan(LDLM_NAMESPACE_CLIENT, sc->nr_to_scan,
+ sc->gfp_mask);
}
-void ldlm_pools_recalc(ldlm_side_t client)
+int ldlm_pools_recalc(ldlm_side_t client)
{
__u32 nr_l = 0, nr_p = 0, l;
struct ldlm_namespace *ns;
+ struct ldlm_namespace *ns_old = NULL;
int nr, equal = 0;
+ int time = 50; /* seconds of sleep if no active namespaces */
/*
* No need to setup pool limit for client pools.
@@ -1190,16 +1208,14 @@ void ldlm_pools_recalc(ldlm_side_t client)
* for _all_ pools.
*/
l = LDLM_POOL_HOST_L /
- atomic_read(
- ldlm_namespace_nr(client));
+ ldlm_namespace_nr_read(client);
} else {
/*
* All the rest of greedy pools will have
* all locks in equal parts.
*/
l = (LDLM_POOL_HOST_L - nr_l) /
- (atomic_read(
- ldlm_namespace_nr(client)) -
+ (ldlm_namespace_nr_read(client) -
nr_p);
}
ldlm_pool_setup(&ns->ns_pool, l);
@@ -1208,9 +1224,9 @@ void ldlm_pools_recalc(ldlm_side_t client)
}
/*
- * Recalc at least ldlm_namespace_nr(client) namespaces.
+ * Recalc at least ldlm_namespace_nr_read(client) namespaces.
*/
- for (nr = atomic_read(ldlm_namespace_nr(client)); nr > 0; nr--) {
+ for (nr = ldlm_namespace_nr_read(client); nr > 0; nr--) {
int skip;
/*
* Lock the list, get first @ns in the list, getref, move it
@@ -1226,6 +1242,30 @@ void ldlm_pools_recalc(ldlm_side_t client)
}
ns = ldlm_namespace_first_locked(client);
+ if (ns_old == ns) { /* Full pass complete */
+ mutex_unlock(ldlm_namespace_lock(client));
+ break;
+ }
+
+ /* We got an empty namespace, need to move it back to inactive
+ * list.
+ * The race with parallel resource creation is fine:
+ * - If they do namespace_get before our check, we fail the
+ * check and they move this item to the end of the list anyway
+ * - If we do the check and then they do namespace_get, then
+ * we move the namespace to inactive and they will move
+ * it back to active (synchronised by the lock, so no clash
+ * there).
+ */
+ if (ldlm_ns_empty(ns)) {
+ ldlm_namespace_move_to_inactive_locked(ns, client);
+ mutex_unlock(ldlm_namespace_lock(client));
+ continue;
+ }
+
+ if (ns_old == NULL)
+ ns_old = ns;
+
spin_lock(&ns->ns_lock);
/*
* skip ns which is being freed, and we don't want to increase
@@ -1239,24 +1279,29 @@ void ldlm_pools_recalc(ldlm_side_t client)
}
spin_unlock(&ns->ns_lock);
- ldlm_namespace_move_locked(ns, client);
+ ldlm_namespace_move_to_active_locked(ns, client);
mutex_unlock(ldlm_namespace_lock(client));
/*
* After setup is done - recalc the pool.
*/
if (!skip) {
- ldlm_pool_recalc(&ns->ns_pool);
+ int ttime = ldlm_pool_recalc(&ns->ns_pool);
+
+ if (ttime < time)
+ time = ttime;
+
ldlm_namespace_put(ns);
}
}
+ return time;
}
EXPORT_SYMBOL(ldlm_pools_recalc);
static int ldlm_pools_thread_main(void *arg)
{
struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
- ENTRY;
+ int s_time, c_time;
thread_set_flags(thread, SVC_RUNNING);
wake_up(&thread->t_ctl_waitq);
@@ -1270,14 +1315,14 @@ static int ldlm_pools_thread_main(void *arg)
/*
* Recal all pools on this tick.
*/
- ldlm_pools_recalc(LDLM_NAMESPACE_SERVER);
- ldlm_pools_recalc(LDLM_NAMESPACE_CLIENT);
+ s_time = ldlm_pools_recalc(LDLM_NAMESPACE_SERVER);
+ c_time = ldlm_pools_recalc(LDLM_NAMESPACE_CLIENT);
/*
* Wait until the next check time, or until we're
* stopped.
*/
- lwi = LWI_TIMEOUT(cfs_time_seconds(LDLM_POOLS_THREAD_PERIOD),
+ lwi = LWI_TIMEOUT(cfs_time_seconds(min(s_time, c_time)),
NULL, NULL);
l_wait_event(thread->t_ctl_waitq,
thread_is_stopping(thread) ||
@@ -1302,15 +1347,14 @@ static int ldlm_pools_thread_main(void *arg)
static int ldlm_pools_thread_start(void)
{
struct l_wait_info lwi = { 0 };
- task_t *task;
- ENTRY;
+ struct task_struct *task;
if (ldlm_pools_thread != NULL)
- RETURN(-EALREADY);
+ return -EALREADY;
OBD_ALLOC_PTR(ldlm_pools_thread);
if (ldlm_pools_thread == NULL)
- RETURN(-ENOMEM);
+ return -ENOMEM;
init_completion(&ldlm_pools_comp);
init_waitqueue_head(&ldlm_pools_thread->t_ctl_waitq);
@@ -1321,19 +1365,16 @@ static int ldlm_pools_thread_start(void)
CERROR("Can't start pool thread, error %ld\n", PTR_ERR(task));
OBD_FREE(ldlm_pools_thread, sizeof(*ldlm_pools_thread));
ldlm_pools_thread = NULL;
- RETURN(PTR_ERR(task));
+ return PTR_ERR(task);
}
l_wait_event(ldlm_pools_thread->t_ctl_waitq,
thread_is_running(ldlm_pools_thread), &lwi);
- RETURN(0);
+ return 0;
}
static void ldlm_pools_thread_stop(void)
{
- ENTRY;
-
if (ldlm_pools_thread == NULL) {
- EXIT;
return;
}
@@ -1348,37 +1389,37 @@ static void ldlm_pools_thread_stop(void)
wait_for_completion(&ldlm_pools_comp);
OBD_FREE_PTR(ldlm_pools_thread);
ldlm_pools_thread = NULL;
- EXIT;
}
+static struct shrinker ldlm_pools_srv_shrinker = {
+ .count_objects = ldlm_pools_srv_count,
+ .scan_objects = ldlm_pools_srv_scan,
+ .seeks = DEFAULT_SEEKS,
+};
+
+static struct shrinker ldlm_pools_cli_shrinker = {
+ .count_objects = ldlm_pools_cli_count,
+ .scan_objects = ldlm_pools_cli_scan,
+ .seeks = DEFAULT_SEEKS,
+};
+
int ldlm_pools_init(void)
{
int rc;
- ENTRY;
rc = ldlm_pools_thread_start();
if (rc == 0) {
- ldlm_pools_srv_shrinker =
- set_shrinker(DEFAULT_SEEKS,
- ldlm_pools_srv_shrink);
- ldlm_pools_cli_shrinker =
- set_shrinker(DEFAULT_SEEKS,
- ldlm_pools_cli_shrink);
+ register_shrinker(&ldlm_pools_srv_shrinker);
+ register_shrinker(&ldlm_pools_cli_shrinker);
}
- RETURN(rc);
+ return rc;
}
EXPORT_SYMBOL(ldlm_pools_init);
void ldlm_pools_fini(void)
{
- if (ldlm_pools_srv_shrinker != NULL) {
- remove_shrinker(ldlm_pools_srv_shrinker);
- ldlm_pools_srv_shrinker = NULL;
- }
- if (ldlm_pools_cli_shrinker != NULL) {
- remove_shrinker(ldlm_pools_cli_shrinker);
- ldlm_pools_cli_shrinker = NULL;
- }
+ unregister_shrinker(&ldlm_pools_srv_shrinker);
+ unregister_shrinker(&ldlm_pools_cli_shrinker);
ldlm_pools_thread_stop();
}
EXPORT_SYMBOL(ldlm_pools_fini);