X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;ds=sidebyside;f=lustre%2Flod%2Flod_qos.c;h=739720f64d1585073b96edb3451bb6b488e2a4d4;hb=f843facff59226d3788d855d1d6948523ab8d944;hp=4930dfa1e2461eb16b4a5b8e407e642256dae3f9;hpb=89693927f0b065d44fdc496f6b49539118570104;p=fs%2Flustre-release.git diff --git a/lustre/lod/lod_qos.c b/lustre/lod/lod_qos.c index 4930dfa..739720f 100644 --- a/lustre/lod/lod_qos.c +++ b/lustre/lod/lod_qos.c @@ -23,7 +23,7 @@ * Copyright 2009 Sun Microsystems, Inc. All rights reserved * Use is subject to license terms. * - * Copyright (c) 2012, 2016, Intel Corporation. + * Copyright (c) 2012, 2017, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -38,8 +38,10 @@ #define DEBUG_SUBSYSTEM S_LOV #include +#include + #include -#include +#include #include #include @@ -59,117 +61,6 @@ OST_TGT(lod,i)->ltd_statfs.os_bsize) /** - * Add a new target to Quality of Service (QoS) target table. - * - * Add a new OST target to the structure representing an OSS. Resort the list - * of known OSSs by the number of OSTs attached to each OSS. The OSS list is - * protected internally and no external locking is required. - * - * \param[in] lod LOD device - * \param[in] ost_desc OST description - * - * \retval 0 on success - * \retval -ENOMEM on error - */ -int qos_add_tgt(struct lod_device *lod, struct lod_tgt_desc *ost_desc) -{ - struct lod_qos_oss *oss = NULL, *temposs; - struct obd_export *exp = ost_desc->ltd_exp; - int rc = 0, found = 0; - struct list_head *list; - ENTRY; - - down_write(&lod->lod_qos.lq_rw_sem); - /* - * a bit hacky approach to learn NID of corresponding connection - * but there is no official API to access information like this - * with OSD API. - */ - list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) { - if (obd_uuid_equals(&oss->lqo_uuid, - &exp->exp_connection->c_remote_uuid)) { - found++; - break; - } - } - - if (!found) { - OBD_ALLOC_PTR(oss); - if (!oss) - GOTO(out, rc = -ENOMEM); - memcpy(&oss->lqo_uuid, &exp->exp_connection->c_remote_uuid, - sizeof(oss->lqo_uuid)); - } else { - /* Assume we have to move this one */ - list_del(&oss->lqo_oss_list); - } - - oss->lqo_ost_count++; - ost_desc->ltd_qos.ltq_oss = oss; - - CDEBUG(D_QOS, "add tgt %s to OSS %s (%d OSTs)\n", - obd_uuid2str(&ost_desc->ltd_uuid), obd_uuid2str(&oss->lqo_uuid), - oss->lqo_ost_count); - - /* Add sorted by # of OSTs. Find the first entry that we're - bigger than... */ - list = &lod->lod_qos.lq_oss_list; - list_for_each_entry(temposs, list, lqo_oss_list) { - if (oss->lqo_ost_count > temposs->lqo_ost_count) - break; - } - /* ...and add before it. If we're the first or smallest, temposs - points to the list head, and we add to the end. */ - list_add_tail(&oss->lqo_oss_list, &temposs->lqo_oss_list); - - lod->lod_qos.lq_dirty = 1; - lod->lod_qos.lq_rr.lqr_dirty = 1; - -out: - up_write(&lod->lod_qos.lq_rw_sem); - RETURN(rc); -} - -/** - * Remove OST target from QoS table. - * - * Removes given OST target from QoS table and releases related OSS structure - * if no OSTs remain on the OSS. - * - * \param[in] lod LOD device - * \param[in] ost_desc OST description - * - * \retval 0 on success - * \retval -ENOENT if no OSS was found - */ -int qos_del_tgt(struct lod_device *lod, struct lod_tgt_desc *ost_desc) -{ - struct lod_qos_oss *oss; - int rc = 0; - ENTRY; - - down_write(&lod->lod_qos.lq_rw_sem); - oss = ost_desc->ltd_qos.ltq_oss; - if (!oss) - GOTO(out, rc = -ENOENT); - - oss->lqo_ost_count--; - if (oss->lqo_ost_count == 0) { - CDEBUG(D_QOS, "removing OSS %s\n", - obd_uuid2str(&oss->lqo_uuid)); - list_del(&oss->lqo_oss_list); - ost_desc->ltd_qos.ltq_oss = NULL; - OBD_FREE_PTR(oss); - } - - lod->lod_qos.lq_dirty = 1; - lod->lod_qos.lq_rr.lqr_dirty = 1; -out: - up_write(&lod->lod_qos.lq_rw_sem); - RETURN(rc); -} - -/** * Check whether the target is available for new OST objects. * * Request statfs data from the given target and verify it's active and not @@ -211,6 +102,10 @@ static int lod_statfs_and_check(const struct lu_env *env, struct lod_device *d, if (sfs->os_state & OS_STATE_READONLY) rc = -EROFS; + /* object precreation is skipped on the OST with max_create_count=0 */ + if (sfs->os_state & OS_STATE_NOPRECREATE) + rc = -ENOBUFS; + /* check whether device has changed state (active, inactive) */ if (rc != 0 && ost->ltd_active) { /* turned inactive? */ @@ -259,24 +154,25 @@ static int lod_statfs_and_check(const struct lu_env *env, struct lod_device *d, * \param[in] env execution environment for this thread * \param[in] lod LOD device */ -static void lod_qos_statfs_update(const struct lu_env *env, - struct lod_device *lod) +void lod_qos_statfs_update(const struct lu_env *env, struct lod_device *lod) { struct obd_device *obd = lod2obd(lod); - struct ost_pool *osts = &(lod->lod_pool_info); - unsigned int i; - int idx; - __u64 max_age, avail; + struct ost_pool *osts = &(lod->lod_pool_info); + time64_t max_age; + unsigned int i; + u64 avail; + int idx; ENTRY; - max_age = cfs_time_shift_64(-2 * lod->lod_desc.ld_qos_maxage); + max_age = ktime_get_seconds() - 2 * lod->lod_desc.ld_qos_maxage; - if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age)) + if (obd->obd_osfs_age > max_age) /* statfs data are quite recent, don't need to refresh it */ RETURN_EXIT; down_write(&lod->lod_qos.lq_rw_sem); - if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age)) + + if (obd->obd_osfs_age > max_age) goto out; for (i = 0; i < osts->op_count; i++) { @@ -289,7 +185,7 @@ static void lod_qos_statfs_update(const struct lu_env *env, /* recalculate weigths */ lod->lod_qos.lq_dirty = 1; } - obd->obd_osfs_age = cfs_time_current_64(); + obd->obd_osfs_age = ktime_get_seconds(); out: up_write(&lod->lod_qos.lq_rw_sem); @@ -314,12 +210,13 @@ out: */ static int lod_qos_calc_ppo(struct lod_device *lod) { - struct lod_qos_oss *oss; - __u64 ba_max, ba_min, temp; - __u32 num_active; - unsigned int i; - int rc, prio_wide; - time_t now, age; + struct lu_svr_qos *oss; + __u64 ba_max, ba_min, temp; + __u32 num_active; + unsigned int i; + int rc, prio_wide; + time64_t now, age; + ENTRY; if (!lod->lod_qos.lq_dirty) @@ -330,9 +227,9 @@ static int lod_qos_calc_ppo(struct lod_device *lod) GOTO(out, rc = -EAGAIN); /* find bavail on each OSS */ - list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) - oss->lqo_bavail = 0; - lod->lod_qos.lq_active_oss_count = 0; + list_for_each_entry(oss, &lod->lod_qos.lq_svr_list, lsq_svr_list) + oss->lsq_bavail = 0; + lod->lod_qos.lq_active_svr_count = 0; /* * How badly user wants to select OSTs "widely" (not recently chosen @@ -343,9 +240,10 @@ static int lod_qos_calc_ppo(struct lod_device *lod) ba_min = (__u64)(-1); ba_max = 0; - now = cfs_time_current_sec(); + now = ktime_get_real_seconds(); /* Calculate OST penalty per object - * (lod ref taken in lod_qos_prep_create()) */ + * (lod ref taken in lod_qos_prep_create()) + */ cfs_foreach_bit(lod->lod_ost_bitmap, i) { LASSERT(OST_TGT(lod,i)); temp = TGT_BAVAIL(i); @@ -355,9 +253,9 @@ static int lod_qos_calc_ppo(struct lod_device *lod) ba_max = max(temp, ba_max); /* Count the number of usable OSS's */ - if (OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail == 0) - lod->lod_qos.lq_active_oss_count++; - OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail += temp; + if (OST_TGT(lod, i)->ltd_qos.ltq_svr->lsq_bavail == 0) + lod->lod_qos.lq_active_svr_count++; + OST_TGT(lod, i)->ltd_qos.ltq_svr->lsq_bavail += temp; /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */ temp >>= 1; @@ -375,7 +273,7 @@ static int lod_qos_calc_ppo(struct lod_device *lod) (age / lod->lod_desc.ld_qos_maxage); } - num_active = lod->lod_qos.lq_active_oss_count - 1; + num_active = lod->lod_qos.lq_active_svr_count - 1; if (num_active < 1) { /* If there's only 1 OSS, we can't penalize it, so instead we have to double the OST penalty */ @@ -385,18 +283,18 @@ static int lod_qos_calc_ppo(struct lod_device *lod) } /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */ - list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) { - temp = oss->lqo_bavail >> 1; - do_div(temp, oss->lqo_ost_count * num_active); - oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8; + list_for_each_entry(oss, &lod->lod_qos.lq_svr_list, lsq_svr_list) { + temp = oss->lsq_bavail >> 1; + do_div(temp, oss->lsq_tgt_count * num_active); + oss->lsq_penalty_per_obj = (temp * prio_wide) >> 8; - age = (now - oss->lqo_used) >> 3; + age = (now - oss->lsq_used) >> 3; if (lod->lod_qos.lq_reset || age > 32 * lod->lod_desc.ld_qos_maxage) - oss->lqo_penalty = 0; + oss->lsq_penalty = 0; else if (age > lod->lod_desc.ld_qos_maxage) /* Decay OSS penalty. */ - oss->lqo_penalty >>= age / lod->lod_desc.ld_qos_maxage; + oss->lsq_penalty >>= age / lod->lod_desc.ld_qos_maxage; } lod->lod_qos.lq_dirty = 0; @@ -436,12 +334,12 @@ static int lod_qos_calc_weight(struct lod_device *lod, int i) __u64 temp, temp2; temp = TGT_BAVAIL(i); - temp2 = OST_TGT(lod,i)->ltd_qos.ltq_penalty + - OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_penalty; + temp2 = OST_TGT(lod, i)->ltd_qos.ltq_penalty + + OST_TGT(lod, i)->ltd_qos.ltq_svr->lsq_penalty; if (temp < temp2) - OST_TGT(lod,i)->ltd_qos.ltq_weight = 0; + OST_TGT(lod, i)->ltd_qos.ltq_weight = 0; else - OST_TGT(lod,i)->ltd_qos.ltq_weight = temp - temp2; + OST_TGT(lod, i)->ltd_qos.ltq_weight = temp - temp2; return 0; } @@ -463,7 +361,7 @@ static int lod_qos_used(struct lod_device *lod, struct ost_pool *osts, __u32 index, __u64 *total_wt) { struct lod_tgt_desc *ost; - struct lod_qos_oss *oss; + struct lu_svr_qos *oss; unsigned int j; ENTRY; @@ -473,28 +371,28 @@ static int lod_qos_used(struct lod_device *lod, struct ost_pool *osts, /* Don't allocate on this devuce anymore, until the next alloc_qos */ ost->ltd_qos.ltq_usable = 0; - oss = ost->ltd_qos.ltq_oss; + oss = ost->ltd_qos.ltq_svr; /* Decay old penalty by half (we're adding max penalty, and don't want it to run away.) */ ost->ltd_qos.ltq_penalty >>= 1; - oss->lqo_penalty >>= 1; + oss->lsq_penalty >>= 1; /* mark the OSS and OST as recently used */ - ost->ltd_qos.ltq_used = oss->lqo_used = cfs_time_current_sec(); + ost->ltd_qos.ltq_used = oss->lsq_used = ktime_get_real_seconds(); /* Set max penalties for this OST and OSS */ ost->ltd_qos.ltq_penalty += ost->ltd_qos.ltq_penalty_per_obj * lod->lod_ostnr; - oss->lqo_penalty += oss->lqo_penalty_per_obj * - lod->lod_qos.lq_active_oss_count; + oss->lsq_penalty += oss->lsq_penalty_per_obj * + lod->lod_qos.lq_active_svr_count; /* Decrease all OSS penalties */ - list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) { - if (oss->lqo_penalty < oss->lqo_penalty_per_obj) - oss->lqo_penalty = 0; + list_for_each_entry(oss, &lod->lod_qos.lq_svr_list, lsq_svr_list) { + if (oss->lsq_penalty < oss->lsq_penalty_per_obj) + oss->lsq_penalty = 0; else - oss->lqo_penalty -= oss->lqo_penalty_per_obj; + oss->lsq_penalty -= oss->lsq_penalty_per_obj; } *total_wt = 0; @@ -528,21 +426,14 @@ static int lod_qos_used(struct lod_device *lod, struct ost_pool *osts, i, ost->ltd_qos.ltq_usable, TGT_BAVAIL(i) >> 10, ost->ltd_qos.ltq_penalty_per_obj >> 10, ost->ltd_qos.ltq_penalty >> 10, - ost->ltd_qos.ltq_oss->lqo_penalty_per_obj >> 10, - ost->ltd_qos.ltq_oss->lqo_penalty >> 10, + ost->ltd_qos.ltq_svr->lsq_penalty_per_obj >> 10, + ost->ltd_qos.ltq_svr->lsq_penalty >> 10, ost->ltd_qos.ltq_weight >> 10); } RETURN(0); } -void lod_qos_rr_init(struct lod_qos_rr *lqr) -{ - spin_lock_init(&lqr->lqr_alloc); - lqr->lqr_dirty = 1; -} - - #define LOV_QOS_EMPTY ((__u32)-1) /** @@ -562,9 +453,9 @@ void lod_qos_rr_init(struct lod_qos_rr *lqr) * \retval -ENOMEM fails to allocate the array */ static int lod_qos_calc_rr(struct lod_device *lod, struct ost_pool *src_pool, - struct lod_qos_rr *lqr) + struct lu_qos_rr *lqr) { - struct lod_qos_oss *oss; + struct lu_svr_qos *oss; struct lod_tgt_desc *ost; unsigned placed, real_count; unsigned int i; @@ -606,7 +497,7 @@ static int lod_qos_calc_rr(struct lod_device *lod, struct ost_pool *src_pool, /* Place all the OSTs from 1 OSS at the same time. */ placed = 0; - list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) { + list_for_each_entry(oss, &lod->lod_qos.lq_svr_list, lsq_svr_list) { int j = 0; for (i = 0; i < lqr->lqr_pool.op_count; i++) { @@ -618,11 +509,11 @@ static int lod_qos_calc_rr(struct lod_device *lod, struct ost_pool *src_pool, ost = OST_TGT(lod,src_pool->op_array[i]); LASSERT(ost && ost->ltd_ost); - if (ost->ltd_qos.ltq_oss != oss) + if (ost->ltd_qos.ltq_svr != oss) continue; /* Evenly space these OSTs across arrayspace */ - next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count; + next = j * lqr->lqr_pool.op_count / oss->lsq_tgt_count; while (lqr->lqr_pool.op_array[next] != LOV_QOS_EMPTY) next = (next + 1) % lqr->lqr_pool.op_count; @@ -714,7 +605,7 @@ static struct dt_object *lod_qos_declare_object_on(const struct lu_env *env, dt = container_of(n, struct dt_object, do_lu); - rc = lod_sub_object_declare_create(env, dt, NULL, NULL, NULL, th); + rc = lod_sub_declare_create(env, dt, NULL, NULL, NULL, th); if (rc < 0) { CDEBUG(D_OTHER, "can't declare creation on #%u: %d\n", ost_idx, rc); @@ -732,15 +623,15 @@ out: * Return an acceptable stripe count depending on flag LOV_USES_DEFAULT_STRIPE: * all stripes or 3/4 of stripes. * - * \param[in] stripe_cnt number of stripes requested + * \param[in] stripe_count number of stripes requested * \param[in] flags 0 or LOV_USES_DEFAULT_STRIPE * * \retval acceptable stripecount */ -static int min_stripe_count(__u32 stripe_cnt, int flags) +static int min_stripe_count(__u32 stripe_count, int flags) { return (flags & LOV_USES_DEFAULT_STRIPE ? - stripe_cnt - (stripe_cnt / 4) : stripe_cnt); + stripe_count - (stripe_count / 4) : stripe_count); } #define LOV_CREATE_RESEED_MULT 30 @@ -819,64 +710,130 @@ static int lod_qos_is_ost_used(const struct lu_env *env, int ost, __u32 stripes) return 0; } +static inline bool +lod_obj_is_ost_use_skip_cb(const struct lu_env *env, struct lod_object *lo, + int comp_idx, struct lod_obj_stripe_cb_data *data) +{ + struct lod_layout_component *comp = &lo->ldo_comp_entries[comp_idx]; + + return comp->llc_ost_indices == NULL; +} + +static inline int +lod_obj_is_ost_use_cb(const struct lu_env *env, struct lod_object *lo, + int comp_idx, struct lod_obj_stripe_cb_data *data) +{ + struct lod_layout_component *comp = &lo->ldo_comp_entries[comp_idx]; + int i; + + for (i = 0; i < comp->llc_stripe_count; i++) { + if (comp->llc_ost_indices[i] == data->locd_ost_index) { + data->locd_ost_index = -1; + return -EEXIST; + } + } + + return 0; +} + /** * Check is OST used in a composite layout * - * \param[in] inuse all inuse ost indexs + * \param[in] lo lod object * \param[in] ost OST target index to check * - * \retval 0 not used - * \retval 1 used + * \retval false not used + * \retval true used */ -static inline int lod_comp_is_ost_used(struct ost_pool *inuse, int ost) +static inline bool lod_comp_is_ost_used(const struct lu_env *env, + struct lod_object *lo, int ost) { - __u32 j; - LASSERT(inuse != NULL); + struct lod_obj_stripe_cb_data data = { { 0 } }; - if (inuse->op_size == 0) - return 0; + data.locd_ost_index = ost; + data.locd_comp_skip_cb = lod_obj_is_ost_use_skip_cb; + data.locd_comp_cb = lod_obj_is_ost_use_cb; - LASSERT(inuse->op_count <= inuse->op_size); - for (j = 0; j < inuse->op_count; j++) { - if (inuse->op_array[j] == ost) - return 1; - } - return 0; + (void)lod_obj_for_each_stripe(env, lo, NULL, &data); + + return data.locd_ost_index == -1; } -/** - * Mark the given target as used for a composite layout - * - * \param[in] inuse inuse ost index array - * \param[in] idx index in the array - */ -static inline void lod_comp_ost_in_use(struct ost_pool *inuse, int ost) +static inline void lod_avoid_update(struct lod_object *lo, + struct lod_avoid_guide *lag) { - LASSERT(inuse != NULL); - if (inuse->op_size && !lod_comp_is_ost_used(inuse, ost)) { - LASSERT(inuse->op_count < inuse->op_size); - inuse->op_array[inuse->op_count] = ost; - inuse->op_count++; + if (!lod_is_flr(lo)) + return; + + lag->lag_ost_avail--; +} + +static inline bool lod_should_avoid_ost(struct lod_object *lo, + struct lod_avoid_guide *lag, + __u32 index) +{ + struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev); + struct lod_tgt_desc *ost = OST_TGT(lod, index); + struct lu_svr_qos *lsq = ost->ltd_qos.ltq_svr; + bool used = false; + int i; + + if (!cfs_bitmap_check(lod->lod_ost_bitmap, index)) { + QOS_DEBUG("OST%d: been used in conflicting mirror component\n", + index); + return true; + } + + /** + * we've tried our best, all available OSTs have been used in + * overlapped components in the other mirror + */ + if (lag->lag_ost_avail == 0) + return false; + + /* check OSS use */ + for (i = 0; i < lag->lag_oaa_count; i++) { + if (lag->lag_oss_avoid_array[i] == lsq->lsq_id) { + used = true; + break; + } } + /** + * if the OSS which OST[index] resides has not been used, we'd like to + * use it + */ + if (!used) + return false; + + /* if the OSS has been used, check whether the OST has been used */ + if (!cfs_bitmap_check(lag->lag_ost_avoid_bitmap, index)) + used = false; + else + QOS_DEBUG("OST%d: been used in conflicting mirror component\n", + index); + return used; } static int lod_check_and_reserve_ost(const struct lu_env *env, - struct lod_device *m, + struct lod_object *lo, + struct lod_layout_component *lod_comp, struct obd_statfs *sfs, __u32 ost_idx, __u32 speed, __u32 *s_idx, struct dt_object **stripe, + __u32 *ost_indices, struct thandle *th, - struct ost_pool *inuse) + bool *overstriped) { + struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev); + struct lod_avoid_guide *lag = &lod_env_info(env)->lti_avoid; struct dt_object *o; __u32 stripe_idx = *s_idx; int rc; + ENTRY; - rc = lod_statfs_and_check(env, m, ost_idx, sfs); - if (rc) { - /* this OSP doesn't feel well */ - goto out_return; - } + rc = lod_statfs_and_check(env, lod, ost_idx, sfs); + if (rc) + RETURN(rc); /* * We expect number of precreated objects in f_ffree at @@ -884,7 +841,7 @@ static int lod_check_and_reserve_ost(const struct lu_env *env, */ if (sfs->os_fprecreated == 0 && speed == 0) { QOS_DEBUG("#%d: precreation is empty\n", ost_idx); - goto out_return; + RETURN(rc); } /* @@ -892,44 +849,57 @@ static int lod_check_and_reserve_ost(const struct lu_env *env, */ if (sfs->os_state & OS_STATE_DEGRADED && speed < 2) { QOS_DEBUG("#%d: degraded\n", ost_idx); - goto out_return; + RETURN(rc); } /* * try not allocate on OST which has been used by other * component */ - if (speed == 0 && lod_comp_is_ost_used(inuse, ost_idx)) { - QOS_DEBUG("#%d: used by other component\n", ost_idx); - goto out_return; + if (speed == 0 && lod_comp_is_ost_used(env, lo, ost_idx)) { + QOS_DEBUG("iter %d: OST%d used by other component\n", + speed, ost_idx); + RETURN(rc); } - /* - * do not put >1 objects on a single OST + /** + * try not allocate OSTs used by conflicting component of other mirrors + * for the first and second time. */ - if (lod_qos_is_ost_used(env, ost_idx, stripe_idx)) - goto out_return; + if (speed < 2 && lod_should_avoid_ost(lo, lag, ost_idx)) { + QOS_DEBUG("iter %d: OST%d used by conflicting mirror " + "component\n", speed, ost_idx); + RETURN(rc); + } - o = lod_qos_declare_object_on(env, m, ost_idx, th); + /* do not put >1 objects on a single OST, except for overstriping */ + if (lod_qos_is_ost_used(env, ost_idx, stripe_idx)) { + if (lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING) + *overstriped = true; + else + RETURN(rc); + } + + o = lod_qos_declare_object_on(env, lod, ost_idx, th); if (IS_ERR(o)) { CDEBUG(D_OTHER, "can't declare new object on #%u: %d\n", ost_idx, (int) PTR_ERR(o)); rc = PTR_ERR(o); - goto out_return; + RETURN(rc); } /* * We've successfully declared (reserved) an object */ + lod_avoid_update(lo, lag); lod_qos_ost_in_use(env, stripe_idx, ost_idx); - lod_comp_ost_in_use(inuse, ost_idx); stripe[stripe_idx] = o; + ost_indices[stripe_idx] = ost_idx; OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LOV_CREATE_RACE, 2); stripe_idx++; *s_idx = stripe_idx; -out_return: - return rc; + RETURN(rc); } /** @@ -946,40 +916,41 @@ out_return: * time we give priority to targets which already have objects precreated. * Full OSTs are skipped (see lod_qos_dev_is_full() for the details). * - * \param[in] env execution environment for this thread - * \param[in] lo LOD object - * \param[out] stripe striping created - * \param[in] flags allocation flags (0 or LOV_USES_DEFAULT_STRIPE) - * \param[in] th transaction handle - * \param[in] comp_idx index of ldo_comp_entries - * \param[in|out] inuse array of inuse ost index + * \param[in] env execution environment for this thread + * \param[in] lo LOD object + * \param[out] stripe striping created + * \param[out] ost_indices ost indices of striping created + * \param[in] flags allocation flags (0 or LOV_USES_DEFAULT_STRIPE) + * \param[in] th transaction handle + * \param[in] comp_idx index of ldo_comp_entries * * \retval 0 on success * \retval -ENOSPC if not enough OSTs are found * \retval negative negated errno for other failures */ static int lod_alloc_rr(const struct lu_env *env, struct lod_object *lo, - struct dt_object **stripe, int flags, - struct thandle *th, int comp_idx, - struct ost_pool *inuse) + struct dt_object **stripe, __u32 *ost_indices, + int flags, struct thandle *th, int comp_idx) { struct lod_layout_component *lod_comp; struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev); struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs; struct pool_desc *pool = NULL; struct ost_pool *osts; - struct lod_qos_rr *lqr; + struct lu_qos_rr *lqr; unsigned int i, array_idx; - __u32 ost_start_idx_temp; - __u32 stripe_idx = 0; - __u32 stripe_cnt, stripe_cnt_min, ost_idx; - int rc, speed = 0, ost_connecting = 0; + __u32 ost_start_idx_temp; + __u32 stripe_idx = 0; + __u32 stripe_count, stripe_count_min, ost_idx; + int rc, speed = 0, ost_connecting = 0; + int stripes_per_ost = 1; + bool overstriped = false; ENTRY; LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL); lod_comp = &lo->ldo_comp_entries[comp_idx]; - stripe_cnt = lod_comp->llc_stripenr; - stripe_cnt_min = min_stripe_count(stripe_cnt, flags); + stripe_count = lod_comp->llc_stripe_count; + stripe_count_min = min_stripe_count(stripe_count, flags); if (lod_comp->llc_pool != NULL) pool = lod_find_pool(m, lod_comp->llc_pool); @@ -997,37 +968,42 @@ static int lod_alloc_rr(const struct lu_env *env, struct lod_object *lo, if (rc) GOTO(out, rc); - rc = lod_qos_ost_in_use_clear(env, stripe_cnt); + rc = lod_qos_ost_in_use_clear(env, stripe_count); if (rc) GOTO(out, rc); down_read(&m->lod_qos.lq_rw_sem); spin_lock(&lqr->lqr_alloc); if (--lqr->lqr_start_count <= 0) { - lqr->lqr_start_idx = cfs_rand() % osts->op_count; + lqr->lqr_start_idx = prandom_u32_max(osts->op_count); lqr->lqr_start_count = (LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) + LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U); - } else if (stripe_cnt_min >= osts->op_count || + } else if (stripe_count_min >= osts->op_count || lqr->lqr_start_idx > osts->op_count) { /* If we have allocated from all of the OSTs, slowly * precess the next start if the OST/stripe count isn't * already doing this for us. */ lqr->lqr_start_idx %= osts->op_count; - if (stripe_cnt > 1 && (osts->op_count % stripe_cnt) != 1) + if (stripe_count > 1 && (osts->op_count % stripe_count) != 1) ++lqr->lqr_offset_idx; } ost_start_idx_temp = lqr->lqr_start_idx; repeat_find: - QOS_DEBUG("pool '%s' want %d startidx %d startcnt %d offset %d " + QOS_DEBUG("pool '%s' want %d start_idx %d start_count %d offset %d " "active %d count %d\n", lod_comp->llc_pool ? lod_comp->llc_pool : "", - stripe_cnt, lqr->lqr_start_idx, lqr->lqr_start_count, + stripe_count, lqr->lqr_start_idx, lqr->lqr_start_count, lqr->lqr_offset_idx, osts->op_count, osts->op_count); - for (i = 0; i < osts->op_count && stripe_idx < stripe_cnt; i++) { + if (lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING) + stripes_per_ost = + (lod_comp->llc_stripe_count - 1)/osts->op_count + 1; + + for (i = 0; i < osts->op_count * stripes_per_ost + && stripe_idx < stripe_count; i++) { array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) % osts->op_count; ++lqr->lqr_start_idx; @@ -1047,14 +1023,15 @@ repeat_find: continue; spin_unlock(&lqr->lqr_alloc); - rc = lod_check_and_reserve_ost(env, m, sfs, ost_idx, speed, - &stripe_idx, stripe, th, inuse); + rc = lod_check_and_reserve_ost(env, lo, lod_comp, sfs, ost_idx, + speed, &stripe_idx, stripe, + ost_indices, th, &overstriped); spin_lock(&lqr->lqr_alloc); if (rc != 0 && OST_TGT(m, ost_idx)->ltd_connecting) ost_connecting = 1; } - if ((speed < 2) && (stripe_idx < stripe_cnt_min)) { + if ((speed < 2) && (stripe_idx < stripe_count_min)) { /* Try again, allowing slower OSCs */ speed++; lqr->lqr_start_idx = ost_start_idx_temp; @@ -1066,8 +1043,14 @@ repeat_find: spin_unlock(&lqr->lqr_alloc); up_read(&m->lod_qos.lq_rw_sem); + /* If there are enough OSTs, a component with overstriping requested + * will not actually end up overstriped. The comp should reflect this. + */ + if (!overstriped) + lod_comp->llc_pattern &= ~LOV_PATTERN_OVERSTRIPING; + if (stripe_idx) { - lod_comp->llc_stripenr = stripe_idx; + lod_comp->llc_stripe_count = stripe_idx; /* at least one stripe is allocated */ rc = 0; } else { @@ -1101,29 +1084,26 @@ out: * structures are protected, but no concurrent allocation is allowed on the * same objects. * - * \param[in] env execution environment for this thread - * \param[in] lo LOD object - * \param[out] stripe striping created - * \param[in] lum stripe md to specify list of OSTs - * \param[in] th transaction handle - * \param[in] comp_idx index of ldo_comp_entries - * \param[in|out] inuse array of inuse ost index + * \param[in] env execution environment for this thread + * \param[in] lo LOD object + * \param[out] stripe striping created + * \param[out] ost_indices ost indices of striping created + * \param[in] th transaction handle + * \param[in] comp_idx index of ldo_comp_entries * * \retval 0 on success * \retval -ENODEV OST index does not exist on file system * \retval -EINVAL requested OST index is invalid * \retval negative negated errno on error */ -static int lod_alloc_ost_list(const struct lu_env *env, - struct lod_object *lo, struct dt_object **stripe, - struct lov_user_md *lum, struct thandle *th, - int comp_idx, struct ost_pool *inuse) +static int lod_alloc_ost_list(const struct lu_env *env, struct lod_object *lo, + struct dt_object **stripe, __u32 *ost_indices, + struct thandle *th, int comp_idx) { struct lod_layout_component *lod_comp; struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev); struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs; struct dt_object *o; - struct lov_user_md_v3 *v3; unsigned int array_idx = 0; int stripe_count = 0; int i; @@ -1131,44 +1111,47 @@ static int lod_alloc_ost_list(const struct lu_env *env, ENTRY; /* for specific OSTs layout */ - LASSERT(lum != NULL && lum->lmm_magic == LOV_USER_MAGIC_SPECIFIC); - lustre_print_user_md(D_OTHER, lum, __func__); - LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL); lod_comp = &lo->ldo_comp_entries[comp_idx]; + LASSERT(lod_comp->llc_ostlist.op_array); + LASSERT(lod_comp->llc_ostlist.op_count); - rc = lod_qos_ost_in_use_clear(env, lod_comp->llc_stripenr); + rc = lod_qos_ost_in_use_clear(env, lod_comp->llc_stripe_count); if (rc < 0) RETURN(rc); - v3 = (struct lov_user_md_v3 *)lum; - for (i = 0; i < lod_comp->llc_stripenr; i++) { - if (v3->lmm_objects[i].l_ost_idx == - lod_comp->llc_stripe_offset) { + if (lod_comp->llc_stripe_offset == LOV_OFFSET_DEFAULT) + lod_comp->llc_stripe_offset = + lod_comp->llc_ostlist.op_array[0]; + + for (i = 0; i < lod_comp->llc_stripe_count; i++) { + if (lod_comp->llc_ostlist.op_array[i] == + lod_comp->llc_stripe_offset) { array_idx = i; break; } } - if (i == lod_comp->llc_stripenr) { + if (i == lod_comp->llc_stripe_count) { CDEBUG(D_OTHER, "%s: start index %d not in the specified list of OSTs\n", lod2obd(m)->obd_name, lod_comp->llc_stripe_offset); RETURN(-EINVAL); } - for (i = 0; i < lod_comp->llc_stripenr; - i++, array_idx = (array_idx + 1) % lod_comp->llc_stripenr) { - __u32 ost_idx = v3->lmm_objects[array_idx].l_ost_idx; + for (i = 0; i < lod_comp->llc_stripe_count; + i++, array_idx = (array_idx + 1) % lod_comp->llc_stripe_count) { + __u32 ost_idx = lod_comp->llc_ostlist.op_array[array_idx]; if (!cfs_bitmap_check(m->lod_ost_bitmap, ost_idx)) { rc = -ENODEV; break; } - /* - * do not put >1 objects on a single OST + /* do not put >1 objects on a single OST, except for + * overstriping */ - if (lod_qos_is_ost_used(env, ost_idx, stripe_count)) { + if (lod_qos_is_ost_used(env, ost_idx, stripe_count) && + !(lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING)) { rc = -EINVAL; break; } @@ -1190,8 +1173,8 @@ static int lod_alloc_ost_list(const struct lu_env *env, * We've successfully declared (reserved) an object */ lod_qos_ost_in_use(env, stripe_count, ost_idx); - lod_comp_ost_in_use(inuse, ost_idx); stripe[stripe_count] = o; + ost_indices[stripe_count] = ost_idx; stripe_count++; } @@ -1210,13 +1193,13 @@ static int lod_alloc_ost_list(const struct lu_env *env, * release the stripes allocated. All the internal structures are protected, * but no concurrent allocation is allowed on the same objects. * - * \param[in] env execution environment for this thread - * \param[in] lo LOD object - * \param[out] stripe striping created - * \param[in] flags not used - * \param[in] th transaction handle - * \param[in] comp_idx index of ldo_comp_entries - * \param[in|out]inuse array of inuse ost index + * \param[in] env execution environment for this thread + * \param[in] lo LOD object + * \param[out] stripe striping created + * \param[out] ost_indices ost indices of striping created + * \param[in] flags not used + * \param[in] th transaction handle + * \param[in] comp_idx index of ldo_comp_entries * * \retval 0 on success * \retval -ENOSPC if no OST objects are available at all @@ -1225,26 +1208,27 @@ static int lod_alloc_ost_list(const struct lu_env *env, * \retval negative errno on failure */ static int lod_alloc_specific(const struct lu_env *env, struct lod_object *lo, - struct dt_object **stripe, int flags, - struct thandle *th, int comp_idx, - struct ost_pool *inuse) + struct dt_object **stripe, __u32 *ost_indices, + int flags, struct thandle *th, int comp_idx) { struct lod_layout_component *lod_comp; struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev); struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs; - struct dt_object *o; - __u32 ost_idx; - unsigned int i, array_idx, ost_count; - int rc, stripe_num = 0; - int speed = 0; + struct dt_object *o; + __u32 ost_idx; + unsigned int i, array_idx, ost_count; + int rc, stripe_num = 0; + int speed = 0; struct pool_desc *pool = NULL; struct ost_pool *osts; + int stripes_per_ost = 1; + bool overstriped = false; ENTRY; LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL); lod_comp = &lo->ldo_comp_entries[comp_idx]; - rc = lod_qos_ost_in_use_clear(env, lod_comp->llc_stripenr); + rc = lod_qos_ost_in_use_clear(env, lod_comp->llc_stripe_count); if (rc) GOTO(out, rc); @@ -1276,7 +1260,11 @@ repeat_find: GOTO(out, rc = -EINVAL); } - for (i = 0; i < ost_count; + if (lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING) + stripes_per_ost = + (lod_comp->llc_stripe_count - 1)/ost_count + 1; + + for (i = 0; i < ost_count * stripes_per_ost; i++, array_idx = (array_idx + 1) % ost_count) { ost_idx = osts->op_array[array_idx]; @@ -1289,16 +1277,21 @@ repeat_find: continue; /* - * do not put >1 objects on a single OST + * do not put >1 objects on a single OST, except for + * overstriping, where it is intended */ - if (lod_qos_is_ost_used(env, ost_idx, stripe_num)) - continue; + if (lod_qos_is_ost_used(env, ost_idx, stripe_num)) { + if (lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING) + overstriped = true; + else + continue; + } /* * try not allocate on the OST used by other component */ if (speed == 0 && i != 0 && - lod_comp_is_ost_used(inuse, ost_idx)) + lod_comp_is_ost_used(env, lo, ost_idx)) continue; /* Drop slow OSCs if we can, but not for requested start idx. @@ -1314,9 +1307,9 @@ repeat_find: } /* - * We expect number of precreated objects in f_ffree at - * the first iteration, skip OSPs with no objects ready - * don't apply this logic to OST specified with stripe_offset + * We expect number of precreated objects at the first + * iteration. Skip OSPs with no objects ready. Don't apply + * this logic to OST specified with stripe_offset. */ if (i != 0 && sfs->os_fprecreated == 0 && speed == 0) continue; @@ -1332,12 +1325,12 @@ repeat_find: * We've successfully declared (reserved) an object */ lod_qos_ost_in_use(env, stripe_num, ost_idx); - lod_comp_ost_in_use(inuse, ost_idx); stripe[stripe_num] = o; + ost_indices[stripe_num] = ost_idx; stripe_num++; /* We have enough stripes */ - if (stripe_num == lod_comp->llc_stripenr) + if (stripe_num == lod_comp->llc_stripe_count) GOTO(out, rc = 0); } if (speed < 2) { @@ -1352,8 +1345,15 @@ repeat_find: */ CERROR("can't lstripe objid "DFID": have %d want %u\n", PFID(lu_object_fid(lod2lu_obj(lo))), stripe_num, - lod_comp->llc_stripenr); + lod_comp->llc_stripe_count); rc = stripe_num == 0 ? -ENOSPC : -EFBIG; + + /* If there are enough OSTs, a component with overstriping requessted + * will not actually end up overstriped. The comp should reflect this. + */ + if (rc == 0 && !overstriped) + lod_comp->llc_pattern &= ~LOV_PATTERN_OVERSTRIPING; + out: if (pool != NULL) { up_read(&pool_tgt_rw_sem(pool)); @@ -1416,13 +1416,13 @@ static inline int lod_qos_is_usable(struct lod_device *lod) * An OST with a higher weight is proportionately more likely to be selected * than one with a lower weight. * - * \param[in] env execution environment for this thread - * \param[in] lo LOD object - * \param[out] stripe striping created - * \param[in] flags 0 or LOV_USES_DEFAULT_STRIPE - * \param[in] th transaction handle - * \param[in] comp_idx index of ldo_comp_entries - * \param[in|out]inuse array of inuse ost index + * \param[in] env execution environment for this thread + * \param[in] lo LOD object + * \param[out] stripe striping created + * \param[out] ost_indices ost indices of striping created + * \param[in] flags 0 or LOV_USES_DEFAULT_STRIPE + * \param[in] th transaction handle + * \param[in] comp_idx index of ldo_comp_entries * * \retval 0 on success * \retval -EAGAIN not enough OSTs are found for specified stripe count @@ -1430,28 +1430,30 @@ static inline int lod_qos_is_usable(struct lod_device *lod) * \retval negative errno on failure */ static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo, - struct dt_object **stripe, int flags, - struct thandle *th, int comp_idx, - struct ost_pool *inuse) + struct dt_object **stripe, __u32 *ost_indices, + int flags, struct thandle *th, int comp_idx) { struct lod_layout_component *lod_comp; struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev); struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs; + struct lod_avoid_guide *lag = &lod_env_info(env)->lti_avoid; struct lod_tgt_desc *ost; struct dt_object *o; __u64 total_weight = 0; struct pool_desc *pool = NULL; struct ost_pool *osts; unsigned int i; - __u32 nfound, good_osts, stripe_cnt, stripe_cnt_min; + __u32 nfound, good_osts, stripe_count, stripe_count_min; + bool overstriped = false; + int stripes_per_ost = 1; int rc = 0; ENTRY; LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL); lod_comp = &lo->ldo_comp_entries[comp_idx]; - stripe_cnt = lod_comp->llc_stripenr; - stripe_cnt_min = min_stripe_count(stripe_cnt, flags); - if (stripe_cnt_min < 1) + stripe_count = lod_comp->llc_stripe_count; + stripe_count_min = min_stripe_count(stripe_count, flags); + if (stripe_count_min < 1) RETURN(-EINVAL); if (lod_comp->llc_pool != NULL) @@ -1468,6 +1470,10 @@ static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo, if (!lod_qos_is_usable(lod)) GOTO(out_nolock, rc = -EAGAIN); + if (lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING) + stripes_per_ost = + (lod_comp->llc_stripe_count - 1)/osts->op_count + 1; + /* Do actual allocation, use write lock here. */ down_write(&lod->lod_qos.lq_rw_sem); @@ -1482,7 +1488,7 @@ static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo, if (rc) GOTO(out, rc); - rc = lod_qos_ost_in_use_clear(env, lod_comp->llc_stripenr); + rc = lod_qos_ost_in_use_clear(env, lod_comp->llc_stripe_count); if (rc) GOTO(out, rc); @@ -1519,52 +1525,31 @@ static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo, QOS_DEBUG("found %d good osts\n", good_osts); - if (good_osts < stripe_cnt_min) + if (good_osts < stripe_count_min) GOTO(out, rc = -EAGAIN); - /* We have enough osts */ - if (good_osts < stripe_cnt) - stripe_cnt = good_osts; + /* If we do not have enough OSTs for the requested stripe count, do not + * put more stripes per OST than requested. + */ + if (stripe_count / stripes_per_ost > good_osts) + stripe_count = good_osts * stripes_per_ost; /* Find enough OSTs with weighted random allocation. */ nfound = 0; - while (nfound < stripe_cnt) { - __u64 rand, cur_weight; + while (nfound < stripe_count) { + u64 rand, cur_weight; cur_weight = 0; rc = -ENOSPC; - if (total_weight) { -#if BITS_PER_LONG == 32 - rand = cfs_rand() % (unsigned)total_weight; - /* If total_weight > 32-bit, first generate the high - * 32 bits of the random number, then add in the low - * 32 bits (truncated to the upper limit, if needed) */ - if (total_weight > 0xffffffffULL) - rand = (__u64)(cfs_rand() % - (unsigned)(total_weight >> 32)) << 32; - else - rand = 0; - - if (rand == (total_weight & 0xffffffff00000000ULL)) - rand |= cfs_rand() % (unsigned)total_weight; - else - rand |= cfs_rand(); - -#else - rand = ((__u64)cfs_rand() << 32 | cfs_rand()) % - total_weight; -#endif - } else { - rand = 0; - } + rand = lu_prandom_u64_max(total_weight); /* On average, this will hit larger-weighted OSTs more often. * 0-weight OSTs will always get used last (only when rand=0) */ for (i = 0; i < osts->op_count; i++) { __u32 idx = osts->op_array[i]; - if (!cfs_bitmap_check(lod->lod_ost_bitmap, idx)) + if (lod_should_avoid_ost(lo, lag, idx)) continue; ost = OST_TGT(lod, idx); @@ -1573,24 +1558,30 @@ static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo, continue; cur_weight += ost->ltd_qos.ltq_weight; - QOS_DEBUG("stripe_cnt=%d nfound=%d cur_weight=%llu" - " rand=%llu total_weight=%llu\n", - stripe_cnt, nfound, cur_weight, rand, + QOS_DEBUG("stripe_count=%d nfound=%d cur_weight=%llu " + "rand=%llu total_weight=%llu\n", + stripe_count, nfound, cur_weight, rand, total_weight); if (cur_weight < rand) continue; QOS_DEBUG("stripe=%d to idx=%d\n", nfound, idx); - /* - * do not put >1 objects on a single OST + * do not put >1 objects on a single OST, except for + * overstriping */ - if (lod_qos_is_ost_used(env, idx, nfound) || - lod_comp_is_ost_used(inuse, idx)) + if ((lod_comp_is_ost_used(env, lo, idx)) && + !(lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING)) continue; - lod_qos_ost_in_use(env, nfound, idx); - lod_comp_ost_in_use(inuse, idx); + + if (lod_qos_is_ost_used(env, idx, nfound)) { + if (lod_comp->llc_pattern & + LOV_PATTERN_OVERSTRIPING) + overstriped = true; + else + continue; + } o = lod_qos_declare_object_on(env, lod, idx, th); if (IS_ERR(o)) { @@ -1598,8 +1589,13 @@ static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo, idx, (int) PTR_ERR(o)); continue; } - stripe[nfound++] = o; + + lod_avoid_update(lo, lag); + lod_qos_ost_in_use(env, nfound, idx); + stripe[nfound] = o; + ost_indices[nfound] = idx; lod_qos_used(lod, osts, idx, &total_weight); + nfound++; rc = 0; break; } @@ -1610,7 +1606,7 @@ static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo, } } - if (unlikely(nfound != stripe_cnt)) { + if (unlikely(nfound != stripe_count)) { /* * when the decision to use weighted algorithm was made * we had enough appropriate OSPs, but this state can @@ -1619,15 +1615,12 @@ static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo, * an object due to just changed state */ QOS_DEBUG("%s: wanted %d objects, found only %d\n", - lod2obd(lod)->obd_name, stripe_cnt, nfound); + lod2obd(lod)->obd_name, stripe_count, nfound); for (i = 0; i < nfound; i++) { LASSERT(stripe[i] != NULL); dt_object_put(env, stripe[i]); stripe[i] = NULL; } - LASSERTF(nfound <= inuse->op_count, - "nfound:%d, op_count:%u\n", nfound, inuse->op_count); - inuse->op_count -= nfound; /* makes sense to rebalance next time */ lod->lod_qos.lq_dirty = 1; @@ -1636,6 +1629,12 @@ static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo, rc = -EAGAIN; } + /* If there are enough OSTs, a component with overstriping requessted + * will not actually end up overstriped. The comp should reflect this. + */ + if (rc == 0 && !overstriped) + lod_comp->llc_pattern &= ~LOV_PATTERN_OVERSTRIPING; + out: up_write(&lod->lod_qos.lq_rw_sem); @@ -1650,12 +1649,15 @@ out_nolock: } /** - * Find largest stripe count the caller can use. + * Check stripe count the caller can use. + * + * For new layouts (no initialized components), check the total size of the + * layout against the maximum EA size from the backing file system. This + * stops us from creating a layout which will be too large once initialized. * + * For existing layouts (with initialized components): * Find the maximal possible stripe count not greater than \a stripe_count. - * Sometimes suggested stripecount can't be reached for a number of reasons: - * lack of enough active OSTs or the backend does not support EAs that large. - * If the passed one is 0, then the filesystem's default one is used. + * If the provided stripe count is 0, then the filesystem's default is used. * * \param[in] lod LOD device * \param[in] lo The lod_object @@ -1663,44 +1665,55 @@ out_nolock: * * \retval the maximum usable stripe count */ -static __u16 lod_get_stripecnt(struct lod_device *lod, struct lod_object *lo, - __u16 stripe_count) +__u16 lod_get_stripe_count(struct lod_device *lod, struct lod_object *lo, + __u16 stripe_count, bool overstriping) { __u32 max_stripes = LOV_MAX_STRIPE_COUNT_OLD; + /* max stripe count is based on OSD ea size */ + unsigned int easize = lod->lod_osd_max_easize; + int i; + if (!stripe_count) stripe_count = lod->lod_desc.ld_default_stripe_count; - if (stripe_count > lod->lod_desc.ld_active_tgt_count) - stripe_count = lod->lod_desc.ld_active_tgt_count; if (!stripe_count) stripe_count = 1; + /* Overstriping allows more stripes than targets */ + if (stripe_count > lod->lod_desc.ld_active_tgt_count && !overstriping) + stripe_count = lod->lod_desc.ld_active_tgt_count; - /* stripe count is based on whether OSD can handle larger EA sizes */ - if (lod->lod_osd_max_easize > 0) { - unsigned int easize = lod->lod_osd_max_easize; - int i; - - if (lo->ldo_is_composite) { - struct lod_layout_component *lod_comp; - unsigned int header_sz = sizeof(struct lov_comp_md_v1); - - header_sz += sizeof(struct lov_comp_md_entry_v1) * - lo->ldo_comp_cnt; - for (i = 0; i < lo->ldo_comp_cnt; i++) { - lod_comp = &lo->ldo_comp_entries[i]; - if (lod_comp->llc_flags & LCME_FL_INIT) - header_sz += lov_mds_md_size( - lod_comp->llc_stripenr, LOV_MAGIC_V3); - } - if (easize > header_sz) - easize -= header_sz; - else - easize = 0; + if (lo->ldo_is_composite) { + struct lod_layout_component *lod_comp; + unsigned int header_sz = sizeof(struct lov_comp_md_v1); + unsigned int init_comp_sz = 0; + unsigned int total_comp_sz = 0; + unsigned int comp_sz; + + header_sz += sizeof(struct lov_comp_md_entry_v1) * + lo->ldo_comp_cnt; + + for (i = 0; i < lo->ldo_comp_cnt; i++) { + lod_comp = &lo->ldo_comp_entries[i]; + comp_sz = lov_mds_md_size(lod_comp->llc_stripe_count, + LOV_MAGIC_V3); + total_comp_sz += comp_sz; + if (lod_comp->llc_flags & LCME_FL_INIT) + init_comp_sz += comp_sz; } - max_stripes = lov_mds_md_max_stripe_count(easize, LOV_MAGIC_V3); + if (init_comp_sz > 0) + total_comp_sz = init_comp_sz; + + header_sz += total_comp_sz; + + if (easize > header_sz) + easize -= header_sz; + else + easize = 0; } + max_stripes = lov_mds_md_max_stripe_count(easize, LOV_MAGIC_V3); + return (stripe_count < max_stripes) ? stripe_count : max_stripes; } @@ -1730,29 +1743,64 @@ int lod_use_defined_striping(const struct lu_env *env, struct lov_ost_data_v1 *objs; __u32 magic; __u16 comp_cnt; + __u16 mirror_cnt; int rc = 0, i; ENTRY; - magic = le32_to_cpu(v1->lmm_magic) & ~LOV_MAGIC_DEF; + mutex_lock(&mo->ldo_layout_mutex); + lod_striping_free_nolock(env, mo); + + magic = le32_to_cpu(v1->lmm_magic) & ~LOV_MAGIC_DEFINED; if (magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3 && - magic != LOV_MAGIC_COMP_V1) - RETURN(-EINVAL); + magic != LOV_MAGIC_COMP_V1 && magic != LOV_MAGIC_FOREIGN) + GOTO(unlock, rc = -EINVAL); if (magic == LOV_MAGIC_COMP_V1) { comp_v1 = buf->lb_buf; comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count); if (comp_cnt == 0) - RETURN(-EINVAL); + GOTO(unlock, rc = -EINVAL); + mirror_cnt = le16_to_cpu(comp_v1->lcm_mirror_count) + 1; + mo->ldo_flr_state = le16_to_cpu(comp_v1->lcm_flags) & + LCM_FL_FLR_MASK; mo->ldo_is_composite = 1; + } else if (magic == LOV_MAGIC_FOREIGN) { + struct lov_foreign_md *foreign; + size_t length; + + if (buf->lb_len < offsetof(typeof(*foreign), lfm_value)) { + CDEBUG(D_LAYOUT, + "buf len %zu < min lov_foreign_md size (%zu)\n", + buf->lb_len, + offsetof(typeof(*foreign), lfm_value)); + GOTO(out, rc = -EINVAL); + } + foreign = (struct lov_foreign_md *)buf->lb_buf; + length = foreign_size_le(foreign); + if (buf->lb_len < length) { + CDEBUG(D_LAYOUT, + "buf len %zu < this lov_foreign_md size (%zu)\n", + buf->lb_len, length); + GOTO(out, rc = -EINVAL); + } + + /* just cache foreign LOV EA raw */ + rc = lod_alloc_foreign_lov(mo, length); + if (rc) + GOTO(out, rc); + memcpy(mo->ldo_foreign_lov, buf->lb_buf, length); + GOTO(out, rc); } else { mo->ldo_is_composite = 0; comp_cnt = 1; + mirror_cnt = 0; } + mo->ldo_layout_gen = le16_to_cpu(v1->lmm_layout_gen); - rc = lod_alloc_comp_entries(mo, comp_cnt); + rc = lod_alloc_comp_entries(mo, mirror_cnt, comp_cnt); if (rc) - RETURN(rc); + GOTO(unlock, rc); for (i = 0; i < comp_cnt; i++) { struct lu_extent *ext; @@ -1764,6 +1812,7 @@ int lod_use_defined_striping(const struct lu_env *env, if (mo->ldo_is_composite) { offs = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset); v1 = (struct lov_mds_md_v1 *)((char *)comp_v1 + offs); + v3 = (struct lov_mds_md_v3 *)v1; magic = le32_to_cpu(v1->lmm_magic); ext = &comp_v1->lcm_entries[i].lcme_extent; @@ -1772,6 +1821,9 @@ int lod_use_defined_striping(const struct lu_env *env, lod_comp->llc_extent.e_end = le64_to_cpu(ext->e_end); lod_comp->llc_flags = le32_to_cpu(comp_v1->lcm_entries[i].lcme_flags); + if (lod_comp->llc_flags & LCME_FL_NOSYNC) + lod_comp->llc_timestamp = le64_to_cpu( + comp_v1->lcm_entries[i].lcme_timestamp); lod_comp->llc_id = le32_to_cpu(comp_v1->lcm_entries[i].lcme_id); if (lod_comp->llc_id == LCME_ID_INVAL) @@ -1792,19 +1844,32 @@ int lod_use_defined_striping(const struct lu_env *env, lod_comp->llc_pattern = le32_to_cpu(v1->lmm_pattern); lod_comp->llc_stripe_size = le32_to_cpu(v1->lmm_stripe_size); - lod_comp->llc_stripenr = le16_to_cpu(v1->lmm_stripe_count); + lod_comp->llc_stripe_count = le16_to_cpu(v1->lmm_stripe_count); lod_comp->llc_layout_gen = le16_to_cpu(v1->lmm_layout_gen); + /** + * The stripe_offset of an uninit-ed component is stored in + * the lmm_layout_gen + */ + if (mo->ldo_is_composite && !lod_comp_inited(lod_comp)) + lod_comp->llc_stripe_offset = lod_comp->llc_layout_gen; lod_obj_set_pool(mo, i, pool_name); - if (!(lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED)) { + if ((!mo->ldo_is_composite || lod_comp_inited(lod_comp)) && + !(lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED) && + !(lod_comp->llc_pattern & LOV_PATTERN_MDT)) { rc = lod_initialize_objects(env, mo, objs, i); if (rc) GOTO(out, rc); } } + + rc = lod_fill_mirrors(mo); + GOTO(out, rc); out: if (rc) - lod_object_free_striping(env, mo); + lod_striping_free_nolock(env, mo); +unlock: + mutex_unlock(&mo->ldo_layout_mutex); RETURN(rc); } @@ -1826,36 +1891,48 @@ out: * \retval 0 on success * \retval negative negated errno on error */ -static int lod_qos_parse_config(const struct lu_env *env, - struct lod_object *lo, - const struct lu_buf *buf) +int lod_qos_parse_config(const struct lu_env *env, struct lod_object *lo, + const struct lu_buf *buf) { struct lod_layout_component *lod_comp; - struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev); - struct lov_desc *desc = &d->lod_desc; - struct lov_user_md_v1 *v1 = NULL; - struct lov_user_md_v3 *v3 = NULL; - struct lov_comp_md_v1 *comp_v1 = NULL; - __u32 magic; - __u16 comp_cnt; - int i, rc; + struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev); + struct lov_desc *desc = &d->lod_desc; + struct lov_user_md_v1 *v1 = NULL; + struct lov_user_md_v3 *v3 = NULL; + struct lov_comp_md_v1 *comp_v1 = NULL; + struct lov_foreign_md *lfm = NULL; + char def_pool[LOV_MAXPOOLNAME + 1]; + __u32 magic; + __u16 comp_cnt; + __u16 mirror_cnt; + int i, rc; ENTRY; if (buf == NULL || buf->lb_buf == NULL || buf->lb_len == 0) RETURN(0); - rc = lod_verify_striping(d, buf, false, 0); + memset(def_pool, 0, sizeof(def_pool)); + if (lo->ldo_comp_entries != NULL) + lod_layout_get_pool(lo->ldo_comp_entries, lo->ldo_comp_cnt, + def_pool, sizeof(def_pool)); + + /* free default striping info */ + if (lo->ldo_is_foreign) + lod_free_foreign_lov(lo); + else + lod_free_comp_entries(lo); + + rc = lod_verify_striping(d, lo, buf, false); if (rc) RETURN(-EINVAL); - lod_free_comp_entries(lo); - v3 = buf->lb_buf; v1 = buf->lb_buf; comp_v1 = buf->lb_buf; + /* {lmm,lfm}_magic position/length work for all LOV formats */ magic = v1->lmm_magic; - if (unlikely(le32_to_cpu(magic) & LOV_MAGIC_DEF)) { + if (unlikely(le32_to_cpu(magic) & LOV_MAGIC_DEFINED)) { /* try to use as fully defined striping */ rc = lod_use_defined_striping(env, lo, buf); RETURN(rc); @@ -1888,6 +1965,22 @@ static int lod_qos_parse_config(const struct lu_env *env, /* fall trhough */ case LOV_USER_MAGIC_COMP_V1: break; + case __swab32(LOV_USER_MAGIC_FOREIGN): + lfm = buf->lb_buf; + __swab32s(&lfm->lfm_magic); + __swab32s(&lfm->lfm_length); + __swab32s(&lfm->lfm_type); + __swab32s(&lfm->lfm_flags); + magic = lfm->lfm_magic; + /* fall through */ + case LOV_USER_MAGIC_FOREIGN: + if (!lfm) + lfm = buf->lb_buf; + rc = lod_alloc_foreign_lov(lo, foreign_size(lfm)); + if (rc) + RETURN(rc); + memcpy(lo->ldo_foreign_lov, buf->lb_buf, foreign_size(lfm)); + RETURN(0); default: CERROR("%s: unrecognized magic %X\n", lod2obd(d)->obd_name, magic); @@ -1900,16 +1993,22 @@ static int lod_qos_parse_config(const struct lu_env *env, comp_cnt = comp_v1->lcm_entry_count; if (comp_cnt == 0) RETURN(-EINVAL); + mirror_cnt = comp_v1->lcm_mirror_count + 1; + if (mirror_cnt > 1) + lo->ldo_flr_state = LCM_FL_RDONLY; lo->ldo_is_composite = 1; } else { comp_cnt = 1; + mirror_cnt = 0; lo->ldo_is_composite = 0; } - rc = lod_alloc_comp_entries(lo, comp_cnt); + rc = lod_alloc_comp_entries(lo, mirror_cnt, comp_cnt); if (rc) RETURN(rc); + LASSERT(lo->ldo_comp_entries); + for (i = 0; i < comp_cnt; i++) { struct pool_desc *pool; struct lu_extent *ext; @@ -1922,44 +2021,55 @@ static int lod_qos_parse_config(const struct lu_env *env, comp_v1->lcm_entries[i].lcme_offset); ext = &comp_v1->lcm_entries[i].lcme_extent; lod_comp->llc_extent = *ext; + lod_comp->llc_flags = + comp_v1->lcm_entries[i].lcme_flags & + LCME_CL_COMP_FLAGS; } pool_name = NULL; if (v1->lmm_magic == LOV_USER_MAGIC_V3 || v1->lmm_magic == LOV_USER_MAGIC_SPECIFIC) { v3 = (struct lov_user_md_v3 *)v1; - if (v3->lmm_pool_name[0] != '\0') pool_name = v3->lmm_pool_name; - if (v3->lmm_magic == LOV_USER_MAGIC_SPECIFIC && - v3->lmm_stripe_offset == LOV_OFFSET_DEFAULT) - v3->lmm_stripe_offset = - v3->lmm_objects[0].l_ost_idx; + if (v3->lmm_magic == LOV_USER_MAGIC_SPECIFIC) { + rc = lod_comp_copy_ost_lists(lod_comp, v3); + if (rc) + GOTO(free_comp, rc); + } } + if (pool_name == NULL && def_pool[0] != '\0') + pool_name = def_pool; + if (v1->lmm_pattern == 0) v1->lmm_pattern = LOV_PATTERN_RAID0; - if (lov_pattern(v1->lmm_pattern) != LOV_PATTERN_RAID0) { + if (lov_pattern(v1->lmm_pattern) != LOV_PATTERN_RAID0 && + lov_pattern(v1->lmm_pattern) != LOV_PATTERN_MDT && + lov_pattern(v1->lmm_pattern) != + (LOV_PATTERN_RAID0 | LOV_PATTERN_OVERSTRIPING)) { CDEBUG(D_LAYOUT, "%s: invalid pattern: %x\n", lod2obd(d)->obd_name, v1->lmm_pattern); - lod_free_comp_entries(lo); - RETURN(-EINVAL); + GOTO(free_comp, rc = -EINVAL); } lod_comp->llc_pattern = v1->lmm_pattern; - lod_comp->llc_stripe_size = desc->ld_default_stripe_size; - if (v1->lmm_stripe_size > 0) + if (v1->lmm_stripe_size) lod_comp->llc_stripe_size = v1->lmm_stripe_size; - lod_comp->llc_stripenr = desc->ld_default_stripe_count; - if (v1->lmm_stripe_count > 0) - lod_comp->llc_stripenr = v1->lmm_stripe_count; + lod_comp->llc_stripe_count = desc->ld_default_stripe_count; + if (v1->lmm_stripe_count || + lov_pattern(v1->lmm_pattern) == LOV_PATTERN_MDT) + lod_comp->llc_stripe_count = v1->lmm_stripe_count; lod_comp->llc_stripe_offset = v1->lmm_stripe_offset; lod_obj_set_pool(lo, i, pool_name); + LASSERT(ergo(lov_pattern(lod_comp->llc_pattern) == + LOV_PATTERN_MDT, lod_comp->llc_stripe_count == 0)); + if (pool_name == NULL) continue; @@ -1978,25 +2088,169 @@ static int lod_qos_parse_config(const struct lu_env *env, CDEBUG(D_LAYOUT, "%s: invalid offset, %u\n", lod2obd(d)->obd_name, lod_comp->llc_stripe_offset); - lod_free_comp_entries(lo); - RETURN(-EINVAL); + GOTO(free_comp, rc = -EINVAL); } } - if (lod_comp->llc_stripenr > pool_tgt_count(pool)) - lod_comp->llc_stripenr = pool_tgt_count(pool); + if (lod_comp->llc_stripe_count > pool_tgt_count(pool) && + !(lod_comp->llc_pattern & LOV_PATTERN_OVERSTRIPING)) + lod_comp->llc_stripe_count = pool_tgt_count(pool); lod_pool_putref(pool); } RETURN(0); + +free_comp: + lod_free_comp_entries(lo); + RETURN(rc); +} + +/** + * prepare enough OST avoidance bitmap space + */ +int lod_prepare_avoidance(const struct lu_env *env, struct lod_object *lo) +{ + struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev); + struct lod_tgt_descs *ltds = &lod->lod_ost_descs; + struct lod_avoid_guide *lag = &lod_env_info(env)->lti_avoid; + struct cfs_bitmap *bitmap = NULL; + __u32 *new_oss = NULL; + + lag->lag_ost_avail = ltds->ltd_tgtnr; + + /* reset OSS avoid guide array */ + lag->lag_oaa_count = 0; + if (lag->lag_oss_avoid_array && lag->lag_oaa_size < ltds->ltd_tgtnr) { + OBD_FREE(lag->lag_oss_avoid_array, + sizeof(__u32) * lag->lag_oaa_size); + lag->lag_oss_avoid_array = NULL; + lag->lag_oaa_size = 0; + } + + /* init OST avoid guide bitmap */ + if (lag->lag_ost_avoid_bitmap) { + if (ltds->ltd_tgtnr <= lag->lag_ost_avoid_bitmap->size) { + CFS_RESET_BITMAP(lag->lag_ost_avoid_bitmap); + } else { + CFS_FREE_BITMAP(lag->lag_ost_avoid_bitmap); + lag->lag_ost_avoid_bitmap = NULL; + } + } + + if (!lag->lag_ost_avoid_bitmap) { + bitmap = CFS_ALLOCATE_BITMAP(ltds->ltd_tgtnr); + if (!bitmap) + return -ENOMEM; + } + + if (!lag->lag_oss_avoid_array) { + /** + * usually there are multiple OSTs in one OSS, but we don't + * know the exact OSS number, so we choose a safe option, + * using OST count to allocate the array to store the OSS + * id. + */ + OBD_ALLOC(new_oss, sizeof(*new_oss) * ltds->ltd_tgtnr); + if (!new_oss) { + CFS_FREE_BITMAP(bitmap); + return -ENOMEM; + } + } + + if (new_oss) { + lag->lag_oss_avoid_array = new_oss; + lag->lag_oaa_size = ltds->ltd_tgtnr; + } + if (bitmap) + lag->lag_ost_avoid_bitmap = bitmap; + + return 0; +} + +/** + * Collect information of used OSTs and OSSs in the overlapped components + * of other mirrors + */ +void lod_collect_avoidance(struct lod_object *lo, struct lod_avoid_guide *lag, + int comp_idx) +{ + struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev); + struct lod_layout_component *lod_comp = &lo->ldo_comp_entries[comp_idx]; + struct cfs_bitmap *bitmap = lag->lag_ost_avoid_bitmap; + int i, j; + + /* iterate mirrors */ + for (i = 0; i < lo->ldo_mirror_count; i++) { + struct lod_layout_component *comp; + + /** + * skip mirror containing component[comp_idx], we only + * collect OSTs info of conflicting component in other mirrors, + * so that during read, if OSTs of a mirror's component are + * not available, we still have other mirror with different + * OSTs to read the data. + */ + comp = &lo->ldo_comp_entries[lo->ldo_mirrors[i].lme_start]; + if (comp->llc_id != LCME_ID_INVAL && + mirror_id_of(comp->llc_id) == + mirror_id_of(lod_comp->llc_id)) + continue; + + /* iterate components of a mirror */ + lod_foreach_mirror_comp(comp, lo, i) { + /** + * skip non-overlapped or un-instantiated components, + * NOTE: don't use lod_comp_inited(comp) to judge + * whether @comp has been inited, since during + * declare phase, comp->llc_stripe has been allocated + * while it's init flag not been set until the exec + * phase. + */ + if (!lu_extent_is_overlapped(&comp->llc_extent, + &lod_comp->llc_extent) || + !comp->llc_stripe) + continue; + + /** + * collect used OSTs index and OSS info from a + * component + */ + for (j = 0; j < comp->llc_stripe_count; j++) { + struct lod_tgt_desc *ost; + struct lu_svr_qos *lsq; + int k; + + ost = OST_TGT(lod, comp->llc_ost_indices[j]); + lsq = ost->ltd_qos.ltq_svr; + + if (cfs_bitmap_check(bitmap, ost->ltd_index)) + continue; + + QOS_DEBUG("OST%d used in conflicting mirror " + "component\n", ost->ltd_index); + cfs_bitmap_set(bitmap, ost->ltd_index); + lag->lag_ost_avail--; + + for (k = 0; k < lag->lag_oaa_count; k++) { + if (lag->lag_oss_avoid_array[k] == + lsq->lsq_id) + break; + } + if (k == lag->lag_oaa_count) { + lag->lag_oss_avoid_array[k] = + lsq->lsq_id; + lag->lag_oaa_count++; + } + } + } + } } /** * Create a striping for an obejct. * - * The function creates a new striping for the object. A buffer containing - * configuration hints can be provided optionally. The function tries QoS + * The function creates a new striping for the object. The function tries QoS * algorithm first unless free space is distributed evenly among OSTs, but * by default RR algorithm is preferred due to internal concurrency (QoS is * serialized). The caller must ensure no concurrent calls to the function @@ -2005,142 +2259,143 @@ static int lod_qos_parse_config(const struct lu_env *env, * \param[in] env execution environment for this thread * \param[in] lo LOD object * \param[in] attr attributes OST objects will be declared with - * \param[in] buf suggested striping configuration or NULL * \param[in] th transaction handle * \param[in] comp_idx index of ldo_comp_entries - * \param[in|out]inuse array of inuse ost index * * \retval 0 on success * \retval negative negated errno on error */ -static int lod_qos_prep_create(const struct lu_env *env, struct lod_object *lo, - struct lu_attr *attr, const struct lu_buf *buf, - struct thandle *th, int comp_idx, - struct ost_pool *inuse) +int lod_qos_prep_create(const struct lu_env *env, struct lod_object *lo, + struct lu_attr *attr, struct thandle *th, + int comp_idx) { struct lod_layout_component *lod_comp; struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev); - struct dt_object **stripe; int stripe_len; int flag = LOV_USES_ASSIGNED_STRIPE; int i, rc = 0; + struct lod_avoid_guide *lag = &lod_env_info(env)->lti_avoid; + struct dt_object **stripe = NULL; + __u32 *ost_indices = NULL; ENTRY; LASSERT(lo); LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL); lod_comp = &lo->ldo_comp_entries[comp_idx]; + LASSERT(!(lod_comp->llc_flags & LCME_FL_EXTENSION)); /* A released component is being created */ if (lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED) RETURN(0); - if (likely(lod_comp->llc_stripe == NULL)) { - struct lov_user_md *lum = NULL; + /* A Data-on-MDT component is being created */ + if (lov_pattern(lod_comp->llc_pattern) == LOV_PATTERN_MDT) + RETURN(0); + if (likely(lod_comp->llc_stripe == NULL)) { /* * no striping has been created so far */ - LASSERT(lod_comp->llc_stripenr > 0); + LASSERT(lod_comp->llc_stripe_count); /* * statfs and check OST targets now, since ld_active_tgt_count * could be changed if some OSTs are [de]activated manually. */ lod_qos_statfs_update(env, d); - stripe_len = lod_get_stripecnt(d, lo, lod_comp->llc_stripenr); + stripe_len = lod_get_stripe_count(d, lo, + lod_comp->llc_stripe_count, + lod_comp->llc_pattern & + LOV_PATTERN_OVERSTRIPING); + if (stripe_len == 0) GOTO(out, rc = -ERANGE); - lod_comp->llc_stripenr = stripe_len; + lod_comp->llc_stripe_count = stripe_len; OBD_ALLOC(stripe, sizeof(stripe[0]) * stripe_len); if (stripe == NULL) GOTO(out, rc = -ENOMEM); + OBD_ALLOC(ost_indices, sizeof(*ost_indices) * stripe_len); + if (!ost_indices) + GOTO(out, rc = -ENOMEM); lod_getref(&d->lod_ost_descs); /* XXX: support for non-0 files w/o objects */ - CDEBUG(D_OTHER, "tgt_count %d stripenr %d\n", + CDEBUG(D_OTHER, "tgt_count %d stripe_count %d\n", d->lod_desc.ld_tgt_count, stripe_len); - if (buf != NULL && buf->lb_buf != NULL) { - lum = buf->lb_buf; - if (lum->lmm_magic == LOV_USER_MAGIC_COMP_V1) { - struct lov_comp_md_v1 *comp_v1; + if (lod_comp->llc_ostlist.op_array && + lod_comp->llc_ostlist.op_count) { + rc = lod_alloc_ost_list(env, lo, stripe, ost_indices, + th, comp_idx); + } else if (lod_comp->llc_stripe_offset == LOV_OFFSET_DEFAULT) { + /** + * collect OSTs and OSSs used in other mirrors whose + * components cross the ldo_comp_entries[comp_idx] + */ + rc = lod_prepare_avoidance(env, lo); + if (rc) + GOTO(put_ldts, rc); - comp_v1 = (struct lov_comp_md_v1 *)lum; - lum = (struct lov_user_md *)((char *)comp_v1 + - comp_v1->lcm_entries[comp_idx].lcme_offset); - } - } + QOS_DEBUG("collecting conflict osts for comp[%d]\n", + comp_idx); + lod_collect_avoidance(lo, lag, comp_idx); - if (lum != NULL && lum->lmm_magic == LOV_USER_MAGIC_SPECIFIC) { - rc = lod_alloc_ost_list(env, lo, stripe, lum, th, - comp_idx, inuse); - } else if (lod_comp->llc_stripe_offset == LOV_OFFSET_DEFAULT) { - rc = lod_alloc_qos(env, lo, stripe, flag, th, - comp_idx, inuse); + rc = lod_alloc_qos(env, lo, stripe, ost_indices, flag, + th, comp_idx); if (rc == -EAGAIN) - rc = lod_alloc_rr(env, lo, stripe, flag, th, - comp_idx, inuse); + rc = lod_alloc_rr(env, lo, stripe, ost_indices, + flag, th, comp_idx); } else { - rc = lod_alloc_specific(env, lo, stripe, flag, th, - comp_idx, inuse); + rc = lod_alloc_specific(env, lo, stripe, ost_indices, + flag, th, comp_idx); } +put_ldts: lod_putref(d, &d->lod_ost_descs); - if (rc < 0) { for (i = 0; i < stripe_len; i++) if (stripe[i] != NULL) dt_object_put(env, stripe[i]); - - OBD_FREE(stripe, sizeof(stripe[0]) * stripe_len); - lod_comp->llc_stripenr = 0; + lod_comp->llc_stripe_count = 0; } else { lod_comp->llc_stripe = stripe; + lod_comp->llc_ost_indices = ost_indices; lod_comp->llc_stripes_allocated = stripe_len; } - } else if (!(lod_comp->llc_flags & LCME_FL_INIT)) { + } else { /* * lod_qos_parse_config() found supplied buf as a predefined * striping (not a hint), so it allocated all the object * now we need to create them */ - for (i = 0; i < lod_comp->llc_stripenr; i++) { + for (i = 0; i < lod_comp->llc_stripe_count; i++) { struct dt_object *o; o = lod_comp->llc_stripe[i]; LASSERT(o); - rc = lod_sub_object_declare_create(env, o, attr, NULL, - NULL, th); + rc = lod_sub_declare_create(env, o, attr, NULL, + NULL, th); if (rc < 0) { CERROR("can't declare create: %d\n", rc); break; } } + /** + * Clear LCME_FL_INIT for the component so that + * lod_striping_create() can create the striping objects + * in replay. + */ + lod_comp_unset_init(lod_comp); } out: - RETURN(rc); -} - -static int -lod_obj_stripe_set_inuse_cb(const struct lu_env *env, struct lod_object *lo, - struct dt_object *dt, struct thandle *th, - int stripe_idx, struct lod_obj_stripe_cb_data *data) -{ - struct lod_thread_info *info = lod_env_info(env); - struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev); - struct lu_fid *fid = &info->lti_fid; - __u32 index; - int rc, type = LU_SEQ_RANGE_OST; - - *fid = *lu_object_fid(&dt->do_lu); - rc = lod_fld_lookup(env, d, fid, &index, &type); if (rc < 0) { - CERROR("%s: fail to locate "DFID": rc = %d\n", - lod2obd(d)->obd_name, PFID(fid), rc); - return rc; + if (stripe) + OBD_FREE(stripe, sizeof(stripe[0]) * stripe_len); + if (ost_indices) + OBD_FREE(ost_indices, + sizeof(*ost_indices) * stripe_len); } - lod_comp_ost_in_use(data->locd_inuse, index); - return 0; + RETURN(rc); } int lod_prepare_create(const struct lu_env *env, struct lod_object *lo, @@ -2148,9 +2403,10 @@ int lod_prepare_create(const struct lu_env *env, struct lod_object *lo, struct thandle *th) { - struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev); - struct ost_pool inuse; - int i, rc, comp_cnt; + struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev); + uint64_t size = 0; + int i; + int rc; ENTRY; LASSERT(lo); @@ -2162,7 +2418,7 @@ int lod_prepare_create(const struct lu_env *env, struct lod_object *lo, RETURN(-EIO); /* - * by this time, the object's ldo_stripenr and ldo_stripe_size + * by this time, the object's ldo_stripe_count and ldo_stripe_size * contain default value for striping: taken from the parent * or from filesystem defaults * @@ -2173,40 +2429,27 @@ int lod_prepare_create(const struct lu_env *env, struct lod_object *lo, if (rc) RETURN(rc); - memset(&inuse, 0, sizeof(inuse)); - init_rwsem(&inuse.op_rw_sem); - comp_cnt = lo->ldo_comp_cnt; + if (attr->la_valid & LA_SIZE) + size = attr->la_size; - /* Prepare inuse array for composite file */ - if (lo->ldo_is_composite) { - struct lod_obj_stripe_cb_data data; - - inuse.op_size = comp_cnt * LOV_MAX_STRIPE_COUNT_OLD * - sizeof(__u32); - if (d->lod_osd_max_easize > 0 && - inuse.op_size > d->lod_osd_max_easize) - inuse.op_size = d->lod_osd_max_easize; - OBD_ALLOC(inuse.op_array, inuse.op_size); - if (inuse.op_array == NULL) - RETURN(-ENOMEM); - - data.locd_inuse = &inuse; - rc = lod_obj_for_each_stripe(env, lo, NULL, - lod_obj_stripe_set_inuse_cb, &data); - if (rc) { - OBD_FREE(inuse.op_array, inuse.op_size); - RETURN(rc); - } - } + /** + * prepare OST object creation for the component covering file's + * size, the 1st component (including plain layout file) is always + * instantiated. + */ + for (i = 0; i < lo->ldo_comp_cnt; i++) { + struct lod_layout_component *lod_comp; + struct lu_extent *extent; - /* prepare OST object creation */ - for (i = 0; i < comp_cnt; i++) { - rc = lod_qos_prep_create(env, lo, attr, buf, th, i, &inuse); - if (rc) - break; + lod_comp = &lo->ldo_comp_entries[i]; + extent = &lod_comp->llc_extent; + QOS_DEBUG("comp[%d] %lld "DEXT"\n", i, size, PEXT(extent)); + if (!lo->ldo_is_composite || size >= extent->e_start) { + rc = lod_qos_prep_create(env, lo, attr, th, i); + if (rc) + break; + } } - if (inuse.op_size) - OBD_FREE(inuse.op_array, inuse.op_size); RETURN(rc); }