X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_pool.c;h=bb3b11ca8c6b66245ebdc0103feb5b61b7544e65;hb=813555770487d5f0504145e9cae1c35bedb26231;hp=c5f303a0cb075c7774a037e64edc060fd849757d;hpb=6869932b552ac705f411de3362f01bd50c1f6f7d;p=fs%2Flustre-release.git diff --git a/lustre/ldlm/ldlm_pool.c b/lustre/ldlm/ldlm_pool.c index c5f303a..bb3b11c 100644 --- a/lustre/ldlm/ldlm_pool.c +++ b/lustre/ldlm/ldlm_pool.c @@ -38,7 +38,7 @@ * Author: Yury Umanets */ -/* +/* * Idea of this code is rather simple. Each second, for each server namespace * we have SLV - server lock volume which is calculated on current number of * granted locks, grant speed for past period, etc - that is, locking load. @@ -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 @@ -107,6 +103,8 @@ # include #endif +#include + #include #include #include "ldlm_internal.h" @@ -114,22 +112,33 @@ #ifdef HAVE_LRU_RESIZE_SUPPORT /* - * 50 ldlm locks for 1MB of RAM. + * 50 ldlm locks for 1MB of RAM. + */ +#define LDLM_POOL_HOST_L ((CFS_NUM_CACHEPAGES >> (20 - CFS_PAGE_SHIFT)) * 50) + +/* + * Maximal possible grant step plan in %. */ -#define LDLM_POOL_HOST_L ((num_physpages >> (20 - CFS_PAGE_SHIFT)) * 50) +#define LDLM_POOL_MAX_GSP (30) /* - * Default step in % for grant plan. + * Minimal possible grant step plan in %. */ -#define LDLM_POOL_GSP (10) +#define LDLM_POOL_MIN_GSP (1) -/* - * LDLM_POOL_GSP% of all locks is default GP. +/* + * This controls the speed of reaching LDLM_POOL_MAX_GSP + * with increasing thread period. */ -#define LDLM_POOL_GP(L) (((L) * LDLM_POOL_GSP) / 100) +#define LDLM_POOL_GSP_STEP (4) -/* - * Max age for locks on clients. +/* + * LDLM_POOL_GSP% of all locks is default GP. + */ +#define LDLM_POOL_GP(L) (((L) * LDLM_POOL_MAX_GSP) / 100) + +/* + * Max age for locks on clients. */ #define LDLM_POOL_MAX_AGE (36000) @@ -151,7 +160,7 @@ static inline __u64 ldlm_pool_slv_max(__u32 L) { /* * Allow to have all locks for 1 client for 10 hrs. - * Formula is the following: limit * 10h / 1 client. + * Formula is the following: limit * 10h / 1 client. */ __u64 lim = L * LDLM_POOL_MAX_AGE / 1; return lim; @@ -184,25 +193,57 @@ static inline struct ldlm_namespace *ldlm_pl2ns(struct ldlm_pool *pl) } /** + * 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 yields 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 period 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 in short 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. * - * \pre ->pl_lock is locked. + * \pre ->pl_lock is locked. */ static inline void ldlm_pool_recalc_grant_plan(struct ldlm_pool *pl) { int granted, grant_step, limit; - + limit = ldlm_pool_get_limit(pl); - granted = atomic_read(&pl->pl_granted); + granted = cfs_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; } /** * Recalculates next SLV on passed \a pl. * - * \pre ->pl_lock is locked. + * \pre ->pl_lock is locked. */ static inline void ldlm_pool_recalc_slv(struct ldlm_pool *pl) { @@ -213,19 +254,19 @@ static inline void ldlm_pool_recalc_slv(struct ldlm_pool *pl) slv = pl->pl_server_lock_volume; grant_plan = pl->pl_grant_plan; limit = ldlm_pool_get_limit(pl); - granted = atomic_read(&pl->pl_granted); + granted = cfs_atomic_read(&pl->pl_granted); grant_usage = limit - (granted - grant_plan); if (grant_usage <= 0) grant_usage = 1; - /* - * Find out SLV change factor which is the ratio of grant usage - * from limit. SLV changes as fast as the ratio of grant plan - * consumtion. The more locks from grant plan are not consumed - * by clients in last interval (idle time), the faster grows + /* + * Find out SLV change factor which is the ratio of grant usage + * from limit. SLV changes as fast as the ratio of grant plan + * consumption. The more locks from grant plan are not consumed + * by clients in last interval (idle time), the faster grows * SLV. And the opposite, the more grant plan is over-consumed - * (load time) the faster drops SLV. + * (load time) the faster drops SLV. */ slv_factor = (grant_usage * 100) / limit; if (2 * abs(granted - limit) > limit) { @@ -247,17 +288,17 @@ static inline void ldlm_pool_recalc_slv(struct ldlm_pool *pl) /** * Recalculates next stats on passed \a pl. * - * \pre ->pl_lock is locked. + * \pre ->pl_lock is locked. */ static inline void ldlm_pool_recalc_stats(struct ldlm_pool *pl) { int grant_plan = pl->pl_grant_plan; __u64 slv = pl->pl_server_lock_volume; - int granted = atomic_read(&pl->pl_granted); - int grant_rate = atomic_read(&pl->pl_grant_rate); - int cancel_rate = atomic_read(&pl->pl_cancel_rate); + int granted = cfs_atomic_read(&pl->pl_granted); + int grant_rate = cfs_atomic_read(&pl->pl_grant_rate); + int cancel_rate = cfs_atomic_read(&pl->pl_cancel_rate); - lprocfs_counter_add(pl->pl_stats, LDLM_POOL_SLV_STAT, + lprocfs_counter_add(pl->pl_stats, LDLM_POOL_SLV_STAT, slv); lprocfs_counter_add(pl->pl_stats, LDLM_POOL_GRANTED_STAT, granted); @@ -276,108 +317,97 @@ static void ldlm_srv_pool_push_slv(struct ldlm_pool *pl) { struct obd_device *obd; - /* + /* * Set new SLV in obd field for using it later without accessing the * pool. This is required to avoid race between sending reply to client * with new SLV and cleanup server stack in which we can't guarantee * that namespace is still alive. We know only that obd is alive as - * long as valid export is alive. + * long as valid export is alive. */ obd = ldlm_pl2ns(pl)->ns_obd; LASSERT(obd != NULL); - write_lock(&obd->obd_pool_lock); + cfs_write_lock(&obd->obd_pool_lock); obd->obd_pool_slv = pl->pl_server_lock_volume; - write_unlock(&obd->obd_pool_lock); + cfs_write_unlock(&obd->obd_pool_lock); } /** * Recalculates all pool fields on passed \a pl. * - * \pre ->pl_lock is not locked. + * \pre ->pl_lock is not locked. */ static int ldlm_srv_pool_recalc(struct ldlm_pool *pl) { time_t recalc_interval_sec; ENTRY; - spin_lock(&pl->pl_lock); + cfs_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. + * _before_ recalculating new grant plan. */ ldlm_pool_recalc_slv(pl); - - /* - * Make sure that pool informed obd of last SLV changes. + + /* + * Make sure that pool informed obd of last SLV changes. */ ldlm_srv_pool_push_slv(pl); - /* - * Update grant_plan for new period. + /* + * 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, + lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT, recalc_interval_sec); } - spin_unlock(&pl->pl_lock); + + cfs_spin_unlock(&pl->pl_lock); RETURN(0); } /** * This function is used on server side as main entry point for memory - * preasure handling. It decreases SLV on \a pl according to passed + * pressure handling. It decreases SLV on \a pl according to passed * \a nr and \a gfp_mask. - * + * * Our goal here is to decrease SLV such a way that clients hold \a nr - * locks smaller in next 10h. + * locks smaller in next 10h. */ static int ldlm_srv_pool_shrink(struct ldlm_pool *pl, int nr, unsigned int gfp_mask) { __u32 limit; - ENTRY; - /* - * VM is asking how many entries may be potentially freed. + /* + * VM is asking how many entries may be potentially freed. */ if (nr == 0) - RETURN(atomic_read(&pl->pl_granted)); + return cfs_atomic_read(&pl->pl_granted); - /* + /* * Client already canceled locks but server is already in shrinker - * and can't cancel anything. Let's catch this race. + * and can't cancel anything. Let's catch this race. */ - if (atomic_read(&pl->pl_granted) == 0) + if (cfs_atomic_read(&pl->pl_granted) == 0) RETURN(0); - spin_lock(&pl->pl_lock); + cfs_spin_lock(&pl->pl_lock); - /* - * We want shrinker to possibly cause cancelation of @nr locks from + /* + * We want shrinker to possibly cause cancellation of @nr locks from * clients or grant approximately @nr locks smaller next intervals. * - * This is why we decresed SLV by @nr. This effect will only be as + * This is why we decreased SLV by @nr. This effect will only be as * long as one re-calc interval (1s these days) and this should be * enough to pass this decreased SLV to all clients. On next recalc * interval pool will either increase SLV if locks load is not high * or will keep on same level or even decrease again, thus, shrinker * decreased SLV will affect next recalc intervals and this way will - * make locking load lower. + * make locking load lower. */ if (nr < pl->pl_server_lock_volume) { pl->pl_server_lock_volume = pl->pl_server_lock_volume - nr; @@ -386,17 +416,17 @@ static int ldlm_srv_pool_shrink(struct ldlm_pool *pl, pl->pl_server_lock_volume = ldlm_pool_slv_min(limit); } - /* - * Make sure that pool informed obd of last SLV changes. + /* + * Make sure that pool informed obd of last SLV changes. */ ldlm_srv_pool_push_slv(pl); - spin_unlock(&pl->pl_lock); + cfs_spin_unlock(&pl->pl_lock); - /* + /* * We did not really free any memory here so far, it only will be - * freed later may be, so that we return 0 to not confuse VM. + * freed later may be, so that we return 0 to not confuse VM. */ - RETURN(0); + return 0; } /** @@ -406,13 +436,13 @@ static int ldlm_srv_pool_setup(struct ldlm_pool *pl, int limit) { struct obd_device *obd; ENTRY; - + obd = ldlm_pl2ns(pl)->ns_obd; LASSERT(obd != NULL && obd != LP_POISON); LASSERT(obd->obd_type != LP_POISON); - write_lock(&obd->obd_pool_lock); + cfs_write_lock(&obd->obd_pool_lock); obd->obd_pool_limit = limit; - write_unlock(&obd->obd_pool_lock); + cfs_write_unlock(&obd->obd_pool_lock); ldlm_pool_set_limit(pl, limit); RETURN(0); @@ -425,102 +455,102 @@ static void ldlm_cli_pool_pop_slv(struct ldlm_pool *pl) { struct obd_device *obd; - /* - * Get new SLV and Limit from obd which is updated with comming - * RPCs. + /* + * Get new SLV and Limit from obd which is updated with coming + * RPCs. */ obd = ldlm_pl2ns(pl)->ns_obd; LASSERT(obd != NULL); - read_lock(&obd->obd_pool_lock); + cfs_read_lock(&obd->obd_pool_lock); pl->pl_server_lock_volume = obd->obd_pool_slv; ldlm_pool_set_limit(pl, obd->obd_pool_limit); - read_unlock(&obd->obd_pool_lock); + cfs_read_unlock(&obd->obd_pool_lock); } /** - * Recalculates client sise pool \a pl according to current SLV and Limit. + * Recalculates client size pool \a pl according to current SLV and Limit. */ static int ldlm_cli_pool_recalc(struct ldlm_pool *pl) { time_t recalc_interval_sec; ENTRY; - spin_lock(&pl->pl_lock); + cfs_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) { + cfs_spin_unlock(&pl->pl_lock); + RETURN(0); + } - /* - * Make sure that pool knows last SLV and Limit from obd. + /* + * 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); - } - spin_unlock(&pl->pl_lock); + pl->pl_recalc_time = cfs_time_current_sec(); + lprocfs_counter_add(pl->pl_stats, LDLM_POOL_TIMING_STAT, + recalc_interval_sec); + cfs_spin_unlock(&pl->pl_lock); - /* - * Do not cancel locks in case lru resize is disabled for this ns. + /* + * Do not cancel locks in case lru resize is disabled for this ns. */ if (!ns_connect_lru_resize(ldlm_pl2ns(pl))) RETURN(0); - /* + /* * In the time of canceling locks on client we do not need to maintain * sharp timing, we only want to cancel locks asap according to new SLV. * It may be called when SLV has changed much, this is why we do not - * take into account pl->pl_recalc_time here. + * take into account pl->pl_recalc_time here. */ - RETURN(ldlm_cancel_lru(ldlm_pl2ns(pl), 0, LDLM_ASYNC, + RETURN(ldlm_cancel_lru(ldlm_pl2ns(pl), 0, LDLM_SYNC, LDLM_CANCEL_LRUR)); } /** - * This function is main entry point for memory preasure handling on client side. - * Main goal of this function is to cancel some number of locks on passed \a pl - * according to \a nr and \a gfp_mask. + * This function is main entry point for memory pressure handling on client + * side. Main goal of this function is to cancel some number of locks on + * passed \a pl according to \a nr and \a gfp_mask. */ static int ldlm_cli_pool_shrink(struct ldlm_pool *pl, int nr, unsigned int gfp_mask) { - ENTRY; - - /* - * Do not cancel locks in case lru resize is disabled for this ns. + struct ldlm_namespace *ns; + int canceled = 0, unused; + + ns = ldlm_pl2ns(pl); + + /* + * Do not cancel locks in case lru resize is disabled for this ns. */ - if (!ns_connect_lru_resize(ldlm_pl2ns(pl))) + if (!ns_connect_lru_resize(ns)) RETURN(0); - /* - * Make sure that pool knows last SLV and Limit from obd. + /* + * Make sure that pool knows last SLV and Limit from obd. */ ldlm_cli_pool_pop_slv(pl); - /* - * Find out how many locks may be released according to shrink - * policy. - */ - if (nr == 0) - RETURN(ldlm_cancel_lru_estimate(ldlm_pl2ns(pl), 0, 0, - LDLM_CANCEL_SHRINK)); - - /* - * Cancel @nr locks accoding to shrink policy. + cfs_spin_lock(&ns->ns_unused_lock); + unused = ns->ns_nr_unused; + cfs_spin_unlock(&ns->ns_unused_lock); + + if (nr) { + canceled = ldlm_cancel_lru(ns, nr, LDLM_SYNC, + LDLM_CANCEL_SHRINK); + } +#ifdef __KERNEL__ + /* + * Return the number of potentially reclaimable locks. */ - RETURN(ldlm_cancel_lru(ldlm_pl2ns(pl), nr, LDLM_SYNC, - LDLM_CANCEL_SHRINK)); + return ((unused - canceled) / 100) * sysctl_vfs_cache_pressure; +#else + return unused - canceled; +#endif } struct ldlm_pool_ops ldlm_srv_pool_ops = { @@ -540,14 +570,33 @@ struct ldlm_pool_ops ldlm_cli_pool_ops = { */ int ldlm_pool_recalc(struct ldlm_pool *pl) { + time_t recalc_interval_sec; int count; + cfs_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. + */ + cfs_atomic_set(&pl->pl_grant_rate, 0); + cfs_atomic_set(&pl->pl_cancel_rate, 0); + cfs_atomic_set(&pl->pl_grant_speed, 0); + } + cfs_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, + lprocfs_counter_add(pl->pl_stats, LDLM_POOL_RECALC_STAT, count); return count; } + return 0; } EXPORT_SYMBOL(ldlm_pool_recalc); @@ -560,14 +609,14 @@ int ldlm_pool_shrink(struct ldlm_pool *pl, int nr, unsigned int gfp_mask) { int cancel = 0; - + if (pl->pl_ops->po_shrink != NULL) { cancel = pl->pl_ops->po_shrink(pl, nr, gfp_mask); if (nr > 0) { - lprocfs_counter_add(pl->pl_stats, + lprocfs_counter_add(pl->pl_stats, LDLM_POOL_SHRINK_REQTD_STAT, nr); - lprocfs_counter_add(pl->pl_stats, + lprocfs_counter_add(pl->pl_stats, LDLM_POOL_SHRINK_FREED_STAT, cancel); CDEBUG(D_DLMTRACE, "%s: request to shrink %d locks, " @@ -598,35 +647,36 @@ 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; - spin_lock(&pl->pl_lock); + cfs_spin_lock(&pl->pl_lock); slv = pl->pl_server_lock_volume; 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); - grant_speed = atomic_read(&pl->pl_grant_speed); - cancel_rate = atomic_read(&pl->pl_cancel_rate); - spin_unlock(&pl->pl_lock); + granted = cfs_atomic_read(&pl->pl_granted); + grant_rate = cfs_atomic_read(&pl->pl_grant_rate); + lvf = cfs_atomic_read(&pl->pl_lock_volume_factor); + grant_speed = cfs_atomic_read(&pl->pl_grant_speed); + cancel_rate = cfs_atomic_read(&pl->pl_cancel_rate); + grant_step = ldlm_pool_t2gsp(pl->pl_recalc_period); + cfs_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", @@ -641,8 +691,8 @@ 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) { @@ -711,11 +761,10 @@ 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"); @@ -737,10 +786,10 @@ static int ldlm_pool_proc_init(struct ldlm_pool *pl) lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANTED_STAT, LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, "granted", "locks"); - lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_STAT, + lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_STAT, LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, "grant", "locks"); - lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_STAT, + lprocfs_counter_init(pl->pl_stats, LDLM_POOL_CANCEL_STAT, LPROCFS_CNTR_AVGMINMAX | LPROCFS_CNTR_STDDEV, "cancel", "locks"); lprocfs_counter_init(pl->pl_stats, LDLM_POOL_GRANT_RATE_STAT, @@ -797,15 +846,14 @@ int ldlm_pool_init(struct ldlm_pool *pl, struct ldlm_namespace *ns, int rc; ENTRY; - spin_lock_init(&pl->pl_lock); - atomic_set(&pl->pl_granted, 0); + cfs_spin_lock_init(&pl->pl_lock); + cfs_atomic_set(&pl->pl_granted, 0); pl->pl_recalc_time = cfs_time_current_sec(); - atomic_set(&pl->pl_lock_volume_factor, 1); + cfs_atomic_set(&pl->pl_lock_volume_factor, 1); - 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; + cfs_atomic_set(&pl->pl_grant_rate, 0); + cfs_atomic_set(&pl->pl_cancel_rate, 0); + cfs_atomic_set(&pl->pl_grant_speed, 0); pl->pl_grant_plan = LDLM_POOL_GP(LDLM_POOL_HOST_L); snprintf(pl->pl_name, sizeof(pl->pl_name), "ldlm-pool-%s-%d", @@ -814,11 +862,13 @@ 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); @@ -835,8 +885,8 @@ void ldlm_pool_fini(struct ldlm_pool *pl) { ENTRY; ldlm_pool_proc_fini(pl); - - /* + + /* * Pool should not be used after this point. We can't free it here as * it lives in struct ldlm_namespace, but still interested in catching * any abnormal using cases. @@ -851,28 +901,27 @@ EXPORT_SYMBOL(ldlm_pool_fini); */ void ldlm_pool_add(struct ldlm_pool *pl, struct ldlm_lock *lock) { - /* + /* * FLOCK locks are special in a sense that they are almost never * cancelled, instead special kind of lock is used to drop them. * also there is no LRU for flock locks, so no point in tracking - * them anyway. + * them anyway. */ if (lock->l_resource->lr_type == LDLM_FLOCK) return; - ENTRY; - - atomic_inc(&pl->pl_granted); - atomic_inc(&pl->pl_grant_rate); - atomic_inc(&pl->pl_grant_speed); + + LDLM_DEBUG(lock, "add lock to pool"); + cfs_atomic_inc(&pl->pl_granted); + cfs_atomic_inc(&pl->pl_grant_rate); + cfs_atomic_inc(&pl->pl_grant_speed); lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_GRANT_STAT); - - /* + /* * Do not do pool recalc for client side as all locks which - * potentially may be canceled has already been packed into + * potentially may be canceled has already been packed into * enqueue/cancel rpc. Also we do not want to run out of stack - * with too long call paths. + * with too long call paths. */ if (ns_is_server(ldlm_pl2ns(pl))) ldlm_pool_recalc(pl); @@ -891,11 +940,13 @@ 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); - atomic_dec(&pl->pl_grant_speed); - + + LDLM_DEBUG(lock, "del lock from pool"); + LASSERT(cfs_atomic_read(&pl->pl_granted) > 0); + cfs_atomic_dec(&pl->pl_granted); + cfs_atomic_inc(&pl->pl_cancel_rate); + cfs_atomic_dec(&pl->pl_grant_speed); + lprocfs_counter_incr(pl->pl_stats, LDLM_POOL_CANCEL_STAT); if (ns_is_server(ldlm_pl2ns(pl))) @@ -907,14 +958,14 @@ EXPORT_SYMBOL(ldlm_pool_del); /** * Returns current \a pl SLV. * - * \pre ->pl_lock is not locked. + * \pre ->pl_lock is not locked. */ __u64 ldlm_pool_get_slv(struct ldlm_pool *pl) { __u64 slv; - spin_lock(&pl->pl_lock); + cfs_spin_lock(&pl->pl_lock); slv = pl->pl_server_lock_volume; - spin_unlock(&pl->pl_lock); + cfs_spin_unlock(&pl->pl_lock); return slv; } EXPORT_SYMBOL(ldlm_pool_get_slv); @@ -922,27 +973,27 @@ EXPORT_SYMBOL(ldlm_pool_get_slv); /** * Sets passed \a slv to \a pl. * - * \pre ->pl_lock is not locked. + * \pre ->pl_lock is not locked. */ void ldlm_pool_set_slv(struct ldlm_pool *pl, __u64 slv) { - spin_lock(&pl->pl_lock); + cfs_spin_lock(&pl->pl_lock); pl->pl_server_lock_volume = slv; - spin_unlock(&pl->pl_lock); + cfs_spin_unlock(&pl->pl_lock); } EXPORT_SYMBOL(ldlm_pool_set_slv); /** * Returns current \a pl CLV. * - * \pre ->pl_lock is not locked. + * \pre ->pl_lock is not locked. */ __u64 ldlm_pool_get_clv(struct ldlm_pool *pl) { __u64 slv; - spin_lock(&pl->pl_lock); + cfs_spin_lock(&pl->pl_lock); slv = pl->pl_client_lock_volume; - spin_unlock(&pl->pl_lock); + cfs_spin_unlock(&pl->pl_lock); return slv; } EXPORT_SYMBOL(ldlm_pool_get_clv); @@ -950,13 +1001,13 @@ EXPORT_SYMBOL(ldlm_pool_get_clv); /** * Sets passed \a clv to \a pl. * - * \pre ->pl_lock is not locked. + * \pre ->pl_lock is not locked. */ void ldlm_pool_set_clv(struct ldlm_pool *pl, __u64 clv) { - spin_lock(&pl->pl_lock); + cfs_spin_lock(&pl->pl_lock); pl->pl_client_lock_volume = clv; - spin_unlock(&pl->pl_lock); + cfs_spin_unlock(&pl->pl_lock); } EXPORT_SYMBOL(ldlm_pool_set_clv); @@ -965,7 +1016,7 @@ EXPORT_SYMBOL(ldlm_pool_set_clv); */ __u32 ldlm_pool_get_limit(struct ldlm_pool *pl) { - return atomic_read(&pl->pl_limit); + return cfs_atomic_read(&pl->pl_limit); } EXPORT_SYMBOL(ldlm_pool_get_limit); @@ -974,7 +1025,7 @@ EXPORT_SYMBOL(ldlm_pool_get_limit); */ void ldlm_pool_set_limit(struct ldlm_pool *pl, __u32 limit) { - atomic_set(&pl->pl_limit, limit); + cfs_atomic_set(&pl->pl_limit, limit); } EXPORT_SYMBOL(ldlm_pool_set_limit); @@ -983,42 +1034,32 @@ EXPORT_SYMBOL(ldlm_pool_set_limit); */ __u32 ldlm_pool_get_lvf(struct ldlm_pool *pl) { - return atomic_read(&pl->pl_lock_volume_factor); + return cfs_atomic_read(&pl->pl_lock_volume_factor); } EXPORT_SYMBOL(ldlm_pool_get_lvf); #ifdef __KERNEL__ static int ldlm_pool_granted(struct ldlm_pool *pl) { - return atomic_read(&pl->pl_granted); + return cfs_atomic_read(&pl->pl_granted); } 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; - -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); +static struct cfs_shrinker *ldlm_pools_srv_shrinker; +static struct cfs_shrinker *ldlm_pools_cli_shrinker; +static cfs_completion_t 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. */ -static int ldlm_pools_shrink(ldlm_side_t client, int nr, +static int ldlm_pools_shrink(ldlm_side_t client, int nr, unsigned int gfp_mask) { int total = 0, cached = 0, nr_ns; struct ldlm_namespace *ns; + void *cookie; if (nr != 0 && !(gfp_mask & __GFP_FS)) return -1; @@ -1026,46 +1067,51 @@ static int ldlm_pools_shrink(ldlm_side_t client, int nr, CDEBUG(D_DLMTRACE, "Request to shrink %d %s locks from all pools\n", nr, client == LDLM_NAMESPACE_CLIENT ? "client" : "server"); - /* - * Find out how many resources we may release. + 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 = cfs_atomic_read(ldlm_namespace_nr(client)); + nr_ns > 0; nr_ns--) { - mutex_down(ldlm_namespace_lock(client)); - if (list_empty(ldlm_namespace_list(client))) { - mutex_up(ldlm_namespace_lock(client)); + cfs_mutex_down(ldlm_namespace_lock(client)); + if (cfs_list_empty(ldlm_namespace_list(client))) { + cfs_mutex_up(ldlm_namespace_lock(client)); + cl_env_reexit(cookie); return 0; } ns = ldlm_namespace_first_locked(client); ldlm_namespace_get(ns); ldlm_namespace_move_locked(ns, client); - mutex_up(ldlm_namespace_lock(client)); + cfs_mutex_up(ldlm_namespace_lock(client)); total += ldlm_pool_shrink(&ns->ns_pool, 0, gfp_mask); ldlm_namespace_put(ns, 1); } - - if (nr == 0 || total == 0) + + if (nr == 0 || total == 0) { + cl_env_reexit(cookie); return total; + } - /* - * Shrink at least ldlm_namespace_nr(client) namespaces. + /* + * Shrink at least ldlm_namespace_nr(client) namespaces. */ - for (nr_ns = atomic_read(ldlm_namespace_nr(client)); - nr_ns > 0; nr_ns--) + for (nr_ns = cfs_atomic_read(ldlm_namespace_nr(client)); + nr_ns > 0; nr_ns--) { int cancel, nr_locks; - /* - * Do not call shrink under ldlm_namespace_lock(client) + /* + * Do not call shrink under ldlm_namespace_lock(client) */ - mutex_down(ldlm_namespace_lock(client)); - if (list_empty(ldlm_namespace_list(client))) { - mutex_up(ldlm_namespace_lock(client)); - /* + cfs_mutex_down(ldlm_namespace_lock(client)); + if (cfs_list_empty(ldlm_namespace_list(client))) { + cfs_mutex_up(ldlm_namespace_lock(client)); + /* * If list is empty, we can't return any @cached > 0, * that probably would cause needless shrinker - * call. + * call. */ cached = 0; break; @@ -1073,14 +1119,15 @@ static int ldlm_pools_shrink(ldlm_side_t client, int nr, ns = ldlm_namespace_first_locked(client); ldlm_namespace_get(ns); ldlm_namespace_move_locked(ns, client); - mutex_up(ldlm_namespace_lock(client)); - + cfs_mutex_up(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); ldlm_namespace_put(ns, 1); } + cl_env_reexit(cookie); return cached; } @@ -1100,16 +1147,16 @@ void ldlm_pools_recalc(ldlm_side_t client) struct ldlm_namespace *ns; int nr, equal = 0; - /* + /* * No need to setup pool limit for client pools. */ if (client == LDLM_NAMESPACE_SERVER) { - /* - * Check all modest namespaces first. + /* + * Check all modest namespaces first. */ - mutex_down(ldlm_namespace_lock(client)); - list_for_each_entry(ns, ldlm_namespace_list(client), - ns_list_chain) + cfs_mutex_down(ldlm_namespace_lock(client)); + cfs_list_for_each_entry(ns, ldlm_namespace_list(client), + ns_list_chain) { if (ns->ns_appetite != LDLM_NAMESPACE_MODEST) continue; @@ -1118,9 +1165,9 @@ void ldlm_pools_recalc(ldlm_side_t client) if (l == 0) l = 1; - /* + /* * Set the modest pools limit equal to their avg granted - * locks + 5%. + * locks + 5%. */ l += dru(l * LDLM_POOLS_MODEST_MARGIN, 100); ldlm_pool_setup(&ns->ns_pool, l); @@ -1128,9 +1175,9 @@ void ldlm_pools_recalc(ldlm_side_t client) nr_p++; } - /* - * Make sure that modest namespaces did not eat more that 2/3 - * of limit. + /* + * Make sure that modest namespaces did not eat more that 2/3 + * of limit. */ if (nr_l >= 2 * (LDLM_POOL_HOST_L / 3)) { CWARN("\"Modest\" pools eat out 2/3 of server locks " @@ -1140,63 +1187,81 @@ void ldlm_pools_recalc(ldlm_side_t client) equal = 1; } - /* - * The rest is given to greedy namespaces. + /* + * The rest is given to greedy namespaces. */ - list_for_each_entry(ns, ldlm_namespace_list(client), - ns_list_chain) + cfs_list_for_each_entry(ns, ldlm_namespace_list(client), + ns_list_chain) { if (!equal && ns->ns_appetite != LDLM_NAMESPACE_GREEDY) continue; if (equal) { - /* + /* * In the case 2/3 locks are eaten out by * modest pools, we re-setup equal limit - * for _all_ pools. + * for _all_ pools. */ l = LDLM_POOL_HOST_L / - atomic_read(ldlm_namespace_nr(client)); + cfs_atomic_read( + ldlm_namespace_nr(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)) - + (cfs_atomic_read( + ldlm_namespace_nr(client)) - nr_p); } ldlm_pool_setup(&ns->ns_pool, l); } - mutex_up(ldlm_namespace_lock(client)); + cfs_mutex_up(ldlm_namespace_lock(client)); } - /* - * Recalc at least ldlm_namespace_nr(client) namespaces. + /* + * Recalc at least ldlm_namespace_nr(client) namespaces. */ - for (nr = atomic_read(ldlm_namespace_nr(client)); nr > 0; nr--) { - /* + for (nr = cfs_atomic_read(ldlm_namespace_nr(client)); nr > 0; nr--) { + int skip; + /* * Lock the list, get first @ns in the list, getref, move it * to the tail, unlock and call pool recalc. This way we avoid * calling recalc under @ns lock what is really good as we get * rid of potential deadlock on client nodes when canceling - * locks synchronously. + * locks synchronously. */ - mutex_down(ldlm_namespace_lock(client)); - if (list_empty(ldlm_namespace_list(client))) { - mutex_up(ldlm_namespace_lock(client)); + cfs_mutex_down(ldlm_namespace_lock(client)); + if (cfs_list_empty(ldlm_namespace_list(client))) { + cfs_mutex_up(ldlm_namespace_lock(client)); break; } ns = ldlm_namespace_first_locked(client); - ldlm_namespace_get(ns); + + cfs_spin_lock(&ns->ns_hash_lock); + /* + * skip ns which is being freed, and we don't want to increase + * its refcount again, not even temporarily. bz21519. + */ + if (ns->ns_refcount == 0) { + skip = 1; + } else { + skip = 0; + ldlm_namespace_get_locked(ns); + } + cfs_spin_unlock(&ns->ns_hash_lock); + ldlm_namespace_move_locked(ns, client); - mutex_up(ldlm_namespace_lock(client)); + cfs_mutex_up(ldlm_namespace_lock(client)); - /* - * After setup is done - recalc the pool. + /* + * After setup is done - recalc the pool. */ - ldlm_pool_recalc(&ns->ns_pool); - ldlm_namespace_put(ns, 1); + if (!skip) { + ldlm_pool_recalc(&ns->ns_pool); + ldlm_namespace_put(ns, 1); + } } } EXPORT_SYMBOL(ldlm_pools_recalc); @@ -1218,14 +1283,14 @@ static int ldlm_pools_thread_main(void *arg) struct l_wait_info lwi; /* - * Recal all pools on this tick. + * Recal all pools on this tick. */ ldlm_pools_recalc(LDLM_NAMESPACE_SERVER); ldlm_pools_recalc(LDLM_NAMESPACE_CLIENT); - + /* * Wait until the next check time, or until we're - * stopped. + * stopped. */ lwi = LWI_TIMEOUT(cfs_time_seconds(LDLM_POOLS_THREAD_PERIOD), NULL, NULL); @@ -1247,7 +1312,7 @@ static int ldlm_pools_thread_main(void *arg) CDEBUG(D_DLMTRACE, "%s: pool thread exiting, process %d\n", t_name, cfs_curproc_pid()); - complete_and_exit(&ldlm_pools_comp, 0); + cfs_complete_and_exit(&ldlm_pools_comp, 0); } static int ldlm_pools_thread_start(void) @@ -1263,12 +1328,12 @@ static int ldlm_pools_thread_start(void) if (ldlm_pools_thread == NULL) RETURN(-ENOMEM); - init_completion(&ldlm_pools_comp); + cfs_init_completion(&ldlm_pools_comp); cfs_waitq_init(&ldlm_pools_thread->t_ctl_waitq); - /* + /* * CLONE_VM and CLONE_FILES just avoid a needless copy, because we - * just drop the VM and FILES in ptlrpc_daemonize() right away. + * just drop the VM and FILES in cfs_daemonize() right away. */ rc = cfs_kernel_thread(ldlm_pools_thread_main, ldlm_pools_thread, CLONE_VM | CLONE_FILES); @@ -1296,12 +1361,12 @@ static void ldlm_pools_thread_stop(void) ldlm_pools_thread->t_flags = SVC_STOPPING; cfs_waitq_signal(&ldlm_pools_thread->t_ctl_waitq); - /* + /* * Make sure that pools thread is finished before freeing @thread. * This fixes possible race and oops due to accessing freed memory - * in pools thread. + * in pools thread. */ - wait_for_completion(&ldlm_pools_comp); + cfs_wait_for_completion(&ldlm_pools_comp); OBD_FREE_PTR(ldlm_pools_thread); ldlm_pools_thread = NULL; EXIT; @@ -1314,10 +1379,12 @@ int ldlm_pools_init(void) 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); + ldlm_pools_srv_shrinker = + cfs_set_shrinker(CFS_DEFAULT_SEEKS, + ldlm_pools_srv_shrink); + ldlm_pools_cli_shrinker = + cfs_set_shrinker(CFS_DEFAULT_SEEKS, + ldlm_pools_cli_shrink); } RETURN(rc); } @@ -1326,11 +1393,11 @@ EXPORT_SYMBOL(ldlm_pools_init); void ldlm_pools_fini(void) { if (ldlm_pools_srv_shrinker != NULL) { - remove_shrinker(ldlm_pools_srv_shrinker); + cfs_remove_shrinker(ldlm_pools_srv_shrinker); ldlm_pools_srv_shrinker = NULL; } if (ldlm_pools_cli_shrinker != NULL) { - remove_shrinker(ldlm_pools_cli_shrinker); + cfs_remove_shrinker(ldlm_pools_cli_shrinker); ldlm_pools_cli_shrinker = NULL; } ldlm_pools_thread_stop(); @@ -1437,12 +1504,6 @@ 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;