From 4c97c0e5ed6c631d9be052a67752a96c15ffc4a7 Mon Sep 17 00:00:00 2001
From: yury <yury>
Date: Tue, 12 Aug 2008 12:26:12 +0000
Subject: [PATCH] b=16367 r=adilger,nikita
- fixes pools thread often wake up;
- cleanups, comments.
---
lustre/include/lustre_dlm.h | 114 ++++++++++++++++-------
lustre/ldlm/ldlm_pool.c | 178 +++++++++++++++++++++---------------
lustre/ldlm/ldlm_request.c | 20 ----
3 files changed, 183 insertions(+), 129 deletions(-)
diff --git a/lustre/include/lustre_dlm.h b/lustre/include/lustre_dlm.h
index b10d8b27e0..65f2a58bf8 100644
--- a/lustre/include/lustre_dlm.h
+++ b/lustre/include/lustre_dlm.h
@@ -255,45 +255,92 @@ struct ldlm_pool_ops {
int (*po_setup)(struct ldlm_pool *pl, int limit);
};
-/* One second for pools thread check interval. */
+/**
+ * One second for pools thread check interval. Each pool has own period.
+ */
#define LDLM_POOLS_THREAD_PERIOD (1)
-/* 5% margin for modest pools. See ldlm_pool.c for details. */
+/**
+ * 5% margin for modest pools. See ldlm_pool.c for details.
+ */
#define LDLM_POOLS_MODEST_MARGIN (5)
-/* A change to SLV in % after which we want to wake up pools thread asap. */
-#define LDLM_POOLS_FAST_SLV_CHANGE (50)
+/**
+ * Default recalc period for server side pools in sec.
+ */
+#define LDLM_POOL_SRV_DEF_RECALC_PERIOD (1)
+
+/**
+ * Default recalc period for client side pools in sec.
+ */
+#define LDLM_POOL_CLI_DEF_RECALC_PERIOD (10)
struct ldlm_pool {
- /* Common pool fields */
- cfs_proc_dir_entry_t *pl_proc_dir; /* Pool proc directory. */
- char pl_name[100]; /* Pool name, should be long
- * enough to contain complex
- * proc entry name. */
- spinlock_t pl_lock; /* Lock for protecting slv/clv
- * updates. */
- atomic_t pl_limit; /* Number of allowed locks in
- * in pool, both, client and
- * server side. */
- atomic_t pl_granted; /* Number of granted locks. */
- atomic_t pl_grant_rate; /* Grant rate per T. */
- atomic_t pl_cancel_rate; /* Cancel rate per T. */
- atomic_t pl_grant_speed; /* Grant speed (GR-CR) per T. */
- __u64 pl_server_lock_volume; /* Server lock volume.
- * Protected by pl_lock */
- __u64 pl_client_lock_volume; /* Client lock volue. */
- atomic_t pl_lock_volume_factor; /* Lock volume factor. */
-
- time_t pl_recalc_time; /* Time when last slv from
- * server was obtained. */
- struct ldlm_pool_ops *pl_ops; /* Recalc and shrink ops. */
-
- int pl_grant_plan; /* Planned number of granted
- * locks for next T. */
- int pl_grant_step; /* Grant plan step for next
- * T. */
-
- struct lprocfs_stats *pl_stats; /* Pool statistics. */
+ /**
+ * Pool proc directory.
+ */
+ cfs_proc_dir_entry_t *pl_proc_dir;
+ /**
+ * Pool name, should be long enough to contain compound proc entry name.
+ */
+ char pl_name[100];
+ /**
+ * Lock for protecting slv/clv updates.
+ */
+ spinlock_t pl_lock;
+ /**
+ * Number of allowed locks in in pool, both, client and server side.
+ */
+ atomic_t pl_limit;
+ /**
+ * Number of granted locks in
+ */
+ atomic_t pl_granted;
+ /**
+ * Grant rate per T.
+ */
+ atomic_t pl_grant_rate;
+ /**
+ * Cancel rate per T.
+ */
+ atomic_t pl_cancel_rate;
+ /**
+ * Grant speed (GR-CR) per T.
+ */
+ atomic_t pl_grant_speed;
+ /**
+ * Server lock volume. Protected by pl_lock.
+ */
+ __u64 pl_server_lock_volume;
+ /**
+ * Current biggest client lock volume. Protected by pl_lock.
+ */
+ __u64 pl_client_lock_volume;
+ /**
+ * Lock volume factor. SLV on client is calculated as following:
+ * server_slv * lock_volume_factor.
+ */
+ atomic_t pl_lock_volume_factor;
+ /**
+ * Time when last slv from server was obtained.
+ */
+ time_t pl_recalc_time;
+ /**
+ * Recalc period for pool.
+ */
+ time_t pl_recalc_period;
+ /**
+ * Recalc and shrink ops.
+ */
+ struct ldlm_pool_ops *pl_ops;
+ /**
+ * Number of planned locks for next period.
+ */
+ int pl_grant_plan;
+ /**
+ * Pool statistics.
+ */
+ struct lprocfs_stats *pl_stats;
};
typedef int (*ldlm_res_policy)(struct ldlm_namespace *, struct ldlm_lock **,
@@ -856,7 +903,6 @@ void unlock_res_and_lock(struct ldlm_lock *lock);
void ldlm_pools_recalc(ldlm_side_t client);
int ldlm_pools_init(void);
void ldlm_pools_fini(void);
-void ldlm_pools_wakeup(void);
int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
int idx, ldlm_side_t client);
diff --git a/lustre/ldlm/ldlm_pool.c b/lustre/ldlm/ldlm_pool.c
index a884ca10d4..da67a4cd3b 100644
--- a/lustre/ldlm/ldlm_pool.c
+++ b/lustre/ldlm/ldlm_pool.c
@@ -86,10 +86,6 @@
* pl_cancel_rate - Number of canceled locks for last T (calculated);
* pl_grant_speed - Grant speed (GR - CR) for last T (calculated);
* pl_grant_plan - Planned number of granted locks for next T (calculated);
- *
- * pl_grant_step - Grant plan step, that is how ->pl_grant_plan
- * will change in next T (tunable);
- *
* pl_server_lock_volume - Current server lock volume (calculated);
*
* As it may be seen from list above, we have few possible tunables which may
@@ -120,14 +116,27 @@
#define LDLM_POOL_HOST_L ((num_physpages >> (20 - CFS_PAGE_SHIFT)) * 50)
/*
- * Default step in % for grant plan.
+ * Maximal possible grant step plan in %.
+ */
+#define LDLM_POOL_MAX_GSP (30)
+
+/*
+ * Minimal possible grant step plan in %.
+ */
+#define LDLM_POOL_MIN_GSP (1)
+
+/*
+ * This controls the speed of reaching LDLM_POOL_MAX_GSP
+ * with increasing thread period. This is 4s which means
+ * that for 10s thread period we will have 2 steps by 4s
+ * each.
*/
-#define LDLM_POOL_GSP (10)
+#define LDLM_POOL_GSP_STEP (4)
/*
* LDLM_POOL_GSP% of all locks is default GP.
*/
-#define LDLM_POOL_GP(L) (((L) * LDLM_POOL_GSP) / 100)
+#define LDLM_POOL_GP(L) (((L) * LDLM_POOL_MAX_GSP) / 100)
/*
* Max age for locks on clients.
@@ -184,6 +193,37 @@ static inline struct ldlm_namespace *ldlm_pl2ns(struct ldlm_pool *pl)
return container_of(pl, struct ldlm_namespace, ns_pool);
}
+/**
+ * Calculates suggested grant_step in % of available locks for passed
+ * \a period. This is later used in grant_plan calculations.
+ */
+static inline int ldlm_pool_t2gsp(int t)
+{
+ /*
+ * This yeilds 1% grant step for anything below LDLM_POOL_GSP_STEP
+ * and up to 30% for anything higher than LDLM_POOL_GSP_STEP.
+ *
+ * How this will affect execution is the following:
+ *
+ * - for thread peroid 1s we will have grant_step 1% which good from
+ * pov of taking some load off from server and push it out to clients.
+ * This is like that because 1% for grant_step means that server will
+ * not allow clients to get lots of locks inshort period of time and
+ * keep all old locks in their caches. Clients will always have to
+ * get some locks back if they want to take some new;
+ *
+ * - for thread period 10s (which is default) we will have 23% which
+ * means that clients will have enough of room to take some new locks
+ * without getting some back. All locks from this 23% which were not
+ * taken by clients in current period will contribute in SLV growing.
+ * SLV growing means more locks cached on clients until limit or grant
+ * plan is reached.
+ */
+ return LDLM_POOL_MAX_GSP -
+ (LDLM_POOL_MAX_GSP - LDLM_POOL_MIN_GSP) /
+ (1 << (t / LDLM_POOL_GSP_STEP));
+}
+
/**
* Recalculates next grant limit on passed \a pl.
*
@@ -196,7 +236,8 @@ static inline void ldlm_pool_recalc_grant_plan(struct ldlm_pool *pl)
limit = ldlm_pool_get_limit(pl);
granted = atomic_read(&pl->pl_granted);
- grant_step = ((limit - granted) * pl->pl_grant_step) / 100;
+ grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
+ grant_step = ((limit - granted) * grant_step) / 100;
pl->pl_grant_plan = granted + grant_step;
}
@@ -303,38 +344,28 @@ static int ldlm_srv_pool_recalc(struct ldlm_pool *pl)
spin_lock(&pl->pl_lock);
recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
- if (recalc_interval_sec > 0) {
- /*
- * Update statistics.
- */
- ldlm_pool_recalc_stats(pl);
-
- /*
+ if (recalc_interval_sec >= pl->pl_recalc_period) {
+ /*
* Recalc SLV after last period. This should be done
* _before_ recalculating new grant plan.
*/
ldlm_pool_recalc_slv(pl);
- /*
+ /*
* Make sure that pool informed obd of last SLV changes.
*/
ldlm_srv_pool_push_slv(pl);
- /*
+ /*
* Update grant_plan for new period.
*/
ldlm_pool_recalc_grant_plan(pl);
- /*
- * Zero out all rates and speed for the last period.
- */
- atomic_set(&pl->pl_grant_rate, 0);
- atomic_set(&pl->pl_cancel_rate, 0);
- atomic_set(&pl->pl_grant_speed, 0);
pl->pl_recalc_time = cfs_time_current_sec();
lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
recalc_interval_sec);
}
+
spin_unlock(&pl->pl_lock);
RETURN(0);
}
@@ -447,29 +478,23 @@ static int ldlm_cli_pool_recalc(struct ldlm_pool *pl)
ENTRY;
spin_lock(&pl->pl_lock);
+ /*
+ * Check if we need to recalc lists now.
+ */
+ 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);
+ }
/*
* Make sure that pool knows last SLV and Limit from obd.
*/
ldlm_cli_pool_pop_slv(pl);
- recalc_interval_sec = cfs_time_current_sec() - pl->pl_recalc_time;
- if (recalc_interval_sec > 0) {
- /*
- * Update statistics only every T.
- */
- ldlm_pool_recalc_stats(pl);
-
- /*
- * Zero out grant/cancel rates and speed for last period.
- */
- atomic_set(&pl->pl_grant_rate, 0);
- atomic_set(&pl->pl_cancel_rate, 0);
- atomic_set(&pl->pl_grant_speed, 0);
- pl->pl_recalc_time = cfs_time_current_sec();
- lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
- recalc_interval_sec);
- }
+ pl->pl_recalc_time = cfs_time_current_sec();
+ lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT,
+ recalc_interval_sec);
spin_unlock(&pl->pl_lock);
/*
@@ -541,14 +566,33 @@ struct ldlm_pool_ops ldlm_cli_pool_ops = {
*/
int ldlm_pool_recalc(struct ldlm_pool *pl)
{
+ time_t recalc_interval_sec;
int count;
+ 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.
+ */
+ ldlm_pool_recalc_stats(pl);
+
+ /*
+ * Zero out all rates and speed for the last period.
+ */
+ atomic_set(&pl->pl_grant_rate, 0);
+ atomic_set(&pl->pl_cancel_rate, 0);
+ atomic_set(&pl->pl_grant_speed, 0);
+ }
+ spin_unlock(&pl->pl_lock);
+
if (pl->pl_ops->po_recalc != NULL) {
count = pl->pl_ops->po_recalc(pl);
lprocfs_counter_add(pl->pl_stats, LDLM_POOL_RECALC_STAT,
count);
return count;
}
+
return 0;
}
EXPORT_SYMBOL(ldlm_pool_recalc);
@@ -599,7 +643,7 @@ static int lprocfs_rd_pool_state(char *page, char **start, off_t off,
int count, int *eof, void *data)
{
int granted, grant_rate, cancel_rate, grant_step;
- int nr = 0, grant_speed, grant_plan;
+ int nr = 0, grant_speed, grant_plan, lvf;
struct ldlm_pool *pl = data;
__u64 slv, clv;
__u32 limit;
@@ -609,25 +653,26 @@ static int lprocfs_rd_pool_state(char *page, char **start, off_t off,
clv = pl->pl_client_lock_volume;
limit = ldlm_pool_get_limit(pl);
grant_plan = pl->pl_grant_plan;
- grant_step = pl->pl_grant_step;
granted = atomic_read(&pl->pl_granted);
grant_rate = atomic_read(&pl->pl_grant_rate);
+ lvf = atomic_read(&pl->pl_lock_volume_factor);
grant_speed = atomic_read(&pl->pl_grant_speed);
cancel_rate = atomic_read(&pl->pl_cancel_rate);
+ grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period);
spin_unlock(&pl->pl_lock);
nr += snprintf(page + nr, count - nr, "LDLM pool state (%s):\n",
pl->pl_name);
nr += snprintf(page + nr, count - nr, " SLV: "LPU64"\n", slv);
nr += snprintf(page + nr, count - nr, " CLV: "LPU64"\n", clv);
+ nr += snprintf(page + nr, count - nr, " LVF: %d\n", lvf);
- nr += snprintf(page + nr, count - nr, " LVF: %d\n",
- atomic_read(&pl->pl_lock_volume_factor));
-
- nr += snprintf(page + nr, count - nr, " GSP: %d%%\n",
- grant_step);
- nr += snprintf(page + nr, count - nr, " GP: %d\n",
- grant_plan);
+ if (ns_is_server(ldlm_pl2ns(pl))) {
+ nr += snprintf(page + nr, count - nr, " GSP: %d%%\n",
+ grant_step);
+ nr += snprintf(page + nr, count - nr, " GP: %d\n",
+ grant_plan);
+ }
nr += snprintf(page + nr, count - nr, " GR: %d\n",
grant_rate);
nr += snprintf(page + nr, count - nr, " CR: %d\n",
@@ -642,8 +687,8 @@ static int lprocfs_rd_pool_state(char *page, char **start, off_t off,
}
LDLM_POOL_PROC_READER(grant_plan, int);
-LDLM_POOL_PROC_READER(grant_step, int);
-LDLM_POOL_PROC_WRITER(grant_step, int);
+LDLM_POOL_PROC_READER(recalc_period, int);
+LDLM_POOL_PROC_WRITER(recalc_period, int);
static int ldlm_pool_proc_init(struct ldlm_pool *pl)
{
@@ -712,11 +757,10 @@ static int ldlm_pool_proc_init(struct ldlm_pool *pl)
pool_vars[0].read_fptr = lprocfs_rd_grant_plan;
lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0);
- snprintf(var_name, MAX_STRING_SIZE, "grant_step");
+ snprintf(var_name, MAX_STRING_SIZE, "recalc_period");
pool_vars[0].data = pl;
- pool_vars[0].read_fptr = lprocfs_rd_grant_step;
- if (ns_is_server(ns))
- pool_vars[0].write_fptr = lprocfs_wr_grant_step;
+ pool_vars[0].read_fptr = lprocfs_rd_recalc_period;
+ pool_vars[0].write_fptr = lprocfs_wr_recalc_period;
lprocfs_add_vars(pl->pl_proc_dir, pool_vars, 0);
snprintf(var_name, MAX_STRING_SIZE, "lock_volume_factor");
@@ -806,7 +850,6 @@ int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
atomic_set(&pl->pl_grant_rate, 0);
atomic_set(&pl->pl_cancel_rate, 0);
atomic_set(&pl->pl_grant_speed, 0);
- pl->pl_grant_step = LDLM_POOL_GSP;
pl->pl_grant_plan = LDLM_POOL_GP(LDLM_POOL_HOST_L);
snprintf(pl->pl_name, sizeof(pl->pl_name), "ldlm-pool-%s-%d",
@@ -815,11 +858,13 @@ int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns,
if (client == LDLM_NAMESPACE_SERVER) {
pl->pl_ops = &ldlm_srv_pool_ops;
ldlm_pool_set_limit(pl, LDLM_POOL_HOST_L);
+ pl->pl_recalc_period = LDLM_POOL_SRV_DEF_RECALC_PERIOD;
pl->pl_server_lock_volume = ldlm_pool_slv_max(LDLM_POOL_HOST_L);
} else {
- pl->pl_server_lock_volume = 1;
ldlm_pool_set_limit(pl, 1);
+ pl->pl_server_lock_volume = 1;
pl->pl_ops = &ldlm_cli_pool_ops;
+ pl->pl_recalc_period = LDLM_POOL_CLI_DEF_RECALC_PERIOD;
}
pl->pl_client_lock_volume = 0;
rc = ldlm_pool_proc_init(pl);
@@ -860,7 +905,6 @@ void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock)
*/
if (lock->l_resource->lr_type == LDLM_FLOCK)
return;
-
ENTRY;
atomic_inc(&pl->pl_granted);
@@ -892,6 +936,7 @@ void ldlm_pool_del(struct ldlm_pool *pl, struct ldlm_lock *lock)
if (lock->l_resource->lr_type == LDLM_FLOCK)
return;
ENTRY;
+
LASSERT(atomic_read(&pl->pl_granted) > 0);
atomic_dec(&pl->pl_granted);
atomic_inc(&pl->pl_cancel_rate);
@@ -999,17 +1044,6 @@ static struct shrinker *ldlm_pools_srv_shrinker;
static struct shrinker *ldlm_pools_cli_shrinker;
static struct completion ldlm_pools_comp;
-void ldlm_pools_wakeup(void)
-{
- ENTRY;
- if (ldlm_pools_thread == NULL)
- return;
- ldlm_pools_thread->t_flags |= SVC_EVENT;
- cfs_waitq_signal(&ldlm_pools_thread->t_ctl_waitq);
- EXIT;
-}
-EXPORT_SYMBOL(ldlm_pools_wakeup);
-
/*
* Cancel \a nr locks from all namespaces (if possible). Returns number of
* cached locks after shrink is finished. All namespaces are asked to
@@ -1438,12 +1472,6 @@ void ldlm_pools_fini(void)
}
EXPORT_SYMBOL(ldlm_pools_fini);
-void ldlm_pools_wakeup(void)
-{
- return;
-}
-EXPORT_SYMBOL(ldlm_pools_wakeup);
-
void ldlm_pools_recalc(ldlm_side_t client)
{
return;
diff --git a/lustre/ldlm/ldlm_request.c b/lustre/ldlm/ldlm_request.c
index c002494502..68c24dce3b 100644
--- a/lustre/ldlm/ldlm_request.c
+++ b/lustre/ldlm/ldlm_request.c
@@ -1061,26 +1061,6 @@ int ldlm_cli_update_pool(struct ptlrpc_request *req)
obd->obd_pool_limit = new_limit;
write_unlock(&obd->obd_pool_lock);
- /*
- * Check if we need to wakeup pools thread for fast SLV change.
- * This is only done when threads period is noticably long like
- * 10s or more.
- */
-#if defined(__KERNEL__) && (LDLM_POOLS_THREAD_PERIOD >= 10)
- if (old_slv > 0) {
- __u64 fast_change = old_slv * LDLM_POOLS_FAST_SLV_CHANGE;
- do_div(fast_change, 100);
-
- /*
- * Wake up pools thread only if SLV has changed more than
- * 50% since last update. In this case we want to react asap.
- * Otherwise it is no sense to wake up pools as they are
- * re-calculated every LDLM_POOLS_THREAD_PERIOD anyways.
- */
- if (old_slv > new_slv && old_slv - new_slv > fast_change)
- ldlm_pools_wakeup();
- }
-#endif
RETURN(0);
}
EXPORT_SYMBOL(ldlm_cli_update_pool);
--
GitLab