X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Flod%2Flod_qos.c;h=09a063004b1609431ebb4a32be76c8afaf3d1ded;hp=d447d3141c30091b7d9b17fd57fb8c364680712f;hb=083d62ee6de5ac6cee95c1d2f86b62b75034093b;hpb=59a59cc9d33f94fb857129b255ad42f393d7242e diff --git a/lustre/lod/lod_qos.c b/lustre/lod/lod_qos.c index d447d31..09a0630 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, 2014, Intel Corporation. + * Copyright (c) 2012, 2017, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -39,8 +39,10 @@ #include #include +#include +#include #include -#include + #include "lod_internal.h" /* @@ -75,6 +77,7 @@ int qos_add_tgt(struct lod_device *lod, struct lod_tgt_desc *ost_desc) struct obd_export *exp = ost_desc->ltd_exp; int rc = 0, found = 0; struct list_head *list; + __u32 id = 0; ENTRY; down_write(&lod->lod_qos.lq_rw_sem); @@ -89,6 +92,8 @@ int qos_add_tgt(struct lod_device *lod, struct lod_tgt_desc *ost_desc) found++; break; } + if (oss->lqo_id > id) + id = oss->lqo_id; } if (!found) { @@ -97,6 +102,8 @@ int qos_add_tgt(struct lod_device *lod, struct lod_tgt_desc *ost_desc) GOTO(out, rc = -ENOMEM); memcpy(&oss->lqo_uuid, &exp->exp_connection->c_remote_uuid, sizeof(oss->lqo_uuid)); + ++id; + oss->lqo_id = id; } else { /* Assume we have to move this one */ list_del(&oss->lqo_oss_list); @@ -197,6 +204,11 @@ static int lod_statfs_and_check(const struct lu_env *env, struct lod_device *d, LASSERT(ost); rc = dt_statfs(env, ost->ltd_ost, sfs); + + if (rc == 0 && ((sfs->os_state & OS_STATE_ENOSPC) || + (sfs->os_state & OS_STATE_ENOINO && sfs->os_fprecreated == 0))) + RETURN(-ENOSPC); + if (rc && rc != -ENOTCONN) CERROR("%s: statfs: rc = %d\n", lod2obd(d)->obd_name, rc); @@ -204,12 +216,19 @@ 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? */ - spin_lock(&d->lod_desc_lock); + spin_lock(&d->lod_lock); if (ost->ltd_active) { ost->ltd_active = 0; + if (rc == -ENOTCONN) + ost->ltd_connecting = 1; + LASSERT(d->lod_desc.ld_active_tgt_count > 0); d->lod_desc.ld_active_tgt_count--; d->lod_qos.lq_dirty = 1; @@ -217,22 +236,23 @@ static int lod_statfs_and_check(const struct lu_env *env, struct lod_device *d, CDEBUG(D_CONFIG, "%s: turns inactive\n", ost->ltd_exp->exp_obd->obd_name); } - spin_unlock(&d->lod_desc_lock); + spin_unlock(&d->lod_lock); } else if (rc == 0 && ost->ltd_active == 0) { /* turned active? */ LASSERTF(d->lod_desc.ld_active_tgt_count < d->lod_ostnr, "active tgt count %d, ost nr %d\n", d->lod_desc.ld_active_tgt_count, d->lod_ostnr); - spin_lock(&d->lod_desc_lock); + spin_lock(&d->lod_lock); if (ost->ltd_active == 0) { ost->ltd_active = 1; + ost->ltd_connecting = 0; d->lod_desc.ld_active_tgt_count++; d->lod_qos.lq_dirty = 1; d->lod_qos.lq_rr.lqr_dirty = 1; CDEBUG(D_CONFIG, "%s: turns active\n", ost->ltd_exp->exp_obd->obd_name); } - spin_unlock(&d->lod_desc_lock); + spin_unlock(&d->lod_lock); } RETURN(rc); @@ -248,24 +268,25 @@ 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++) { @@ -278,7 +299,7 @@ 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); @@ -308,7 +329,7 @@ static int lod_qos_calc_ppo(struct lod_device *lod) __u32 num_active; unsigned int i; int rc, prio_wide; - time_t now, age; + time64_t now, age; ENTRY; if (!lod->lod_qos.lq_dirty) @@ -332,9 +353,10 @@ 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); @@ -470,7 +492,7 @@ static int lod_qos_used(struct lod_device *lod, struct ost_pool *osts, oss->lqo_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->lqo_used = ktime_get_real_seconds(); /* Set max penalties for this OST and OSS */ ost->ltd_qos.ltq_penalty += @@ -511,9 +533,9 @@ static int lod_qos_used(struct lod_device *lod, struct ost_pool *osts, if (ost->ltd_qos.ltq_usable) *total_wt += ost->ltd_qos.ltq_weight; - QOS_DEBUG("recalc tgt %d usable=%d avail="LPU64 - " ostppo="LPU64" ostp="LPU64" ossppo="LPU64 - " ossp="LPU64" wt="LPU64"\n", + QOS_DEBUG("recalc tgt %d usable=%d avail=%llu" + " ostppo=%llu ostp=%llu ossppo=%llu" + " ossp=%llu wt=%llu\n", i, ost->ltd_qos.ltq_usable, TGT_BAVAIL(i) >> 10, ost->ltd_qos.ltq_penalty_per_obj >> 10, ost->ltd_qos.ltq_penalty >> 10, @@ -525,6 +547,13 @@ static int lod_qos_used(struct lod_device *lod, struct ost_pool *osts, 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) /** @@ -696,7 +725,7 @@ 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); @@ -714,45 +743,21 @@ 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 #define LOV_CREATE_RESEED_MIN 2000 /** - * Check if an OST is full. - * - * Check whether an OST should be considered full based - * on the given statfs data. - * - * \param[in] msfs statfs data - * - * \retval false not full - * \retval true full - */ -static int inline lod_qos_dev_is_full(struct obd_statfs *msfs) -{ - __u64 used; - int bs = msfs->os_bsize; - - LASSERT(((bs - 1) & bs) == 0); - - /* the minimum of 0.1% used blocks and 1GB bytes. */ - used = min_t(__u64, (msfs->os_blocks - msfs->os_bfree) >> 10, - 1 << (31 - ffs(bs))); - return (msfs->os_bavail < used); -} - -/** * Initialize temporary OST-in-use array. * * Allocate or extend the array used to mark targets already assigned to a new @@ -825,6 +830,191 @@ 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] lo lod object + * \param[in] ost OST target index to check + * + * \retval false not used + * \retval true used + */ +static inline bool lod_comp_is_ost_used(const struct lu_env *env, + struct lod_object *lo, int ost) +{ + struct lod_obj_stripe_cb_data data = { { 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; + + (void)lod_obj_for_each_stripe(env, lo, NULL, &data); + + return data.locd_ost_index == -1; +} + +static inline void lod_avoid_update(struct lod_object *lo, + struct lod_avoid_guide *lag) +{ + 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 lod_qos_oss *lqo = ost->ltd_qos.ltq_oss; + bool used = false; + int i; + + if (!cfs_bitmap_check(lod->lod_ost_bitmap, 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] == lqo->lqo_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: has been used in overlapped component " + "in other mirror\n", index); + return used; +} + +static int lod_check_and_reserve_ost(const struct lu_env *env, + struct lod_object *lo, + struct obd_statfs *sfs, __u32 ost_idx, + __u32 speed, __u32 *s_idx, + struct dt_object **stripe, + __u32 *ost_indices, + struct thandle *th) +{ + 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; + + rc = lod_statfs_and_check(env, lod, ost_idx, sfs); + if (rc) { + /* this OSP doesn't feel well */ + goto out_return; + } + + /* + * We expect number of precreated objects in f_ffree at + * the first iteration, skip OSPs with no objects ready + */ + if (sfs->os_fprecreated == 0 && speed == 0) { + QOS_DEBUG("#%d: precreation is empty\n", ost_idx); + goto out_return; + } + + /* + * try to use another OSP if this one is degraded + */ + if (sfs->os_state & OS_STATE_DEGRADED && speed < 2) { + QOS_DEBUG("#%d: degraded\n", ost_idx); + goto out_return; + } + + /* + * try not allocate on OST which has been used by other + * component + */ + if (speed == 0 && lod_comp_is_ost_used(env, lo, ost_idx)) { + QOS_DEBUG("#%d: used by other component\n", ost_idx); + goto out_return; + } + + /** + * try not allocate OSTs used by conflicting component of other mirrors + * for the first and second time. + */ + if (speed < 2 && lod_should_avoid_ost(lo, lag, ost_idx)) { + QOS_DEBUG("#%d: used by overlapped component of other mirror\n", + ost_idx); + goto out_return; + } + /* + * do not put >1 objects on a single OST + */ + if (lod_qos_is_ost_used(env, ost_idx, stripe_idx)) + goto out_return; + + 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; + } + + /* + * We've successfully declared (reserved) an object + */ + lod_avoid_update(lo, lag); + lod_qos_ost_in_use(env, stripe_idx, 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; +} + /** * Allocate a striping using round-robin algorithm. * @@ -839,38 +1029,42 @@ static int lod_qos_is_ost_used(const struct lu_env *env, int ost, __u32 stripes) * 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] 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) + 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 dt_object *o; - unsigned int i, array_idx; - int rc; - __u32 ost_start_idx_temp; - int speed = 0; - __u32 stripe_idx = 0; - __u32 stripe_cnt = lo->ldo_stripenr; - __u32 stripe_cnt_min = min_stripe_count(stripe_cnt, flags); - __u32 ost_idx; + unsigned int i, array_idx; + __u32 ost_start_idx_temp; + __u32 stripe_idx = 0; + __u32 stripe_count, stripe_count_min, ost_idx; + int rc, speed = 0, ost_connecting = 0; ENTRY; - if (lo->ldo_pool) - pool = lod_find_pool(m, lo->ldo_pool); + LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL); + lod_comp = &lo->ldo_comp_entries[comp_idx]; + 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); if (pool != NULL) { down_read(&pool_tgt_rw_sem(pool)); @@ -885,40 +1079,39 @@ 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, lo->ldo_stripenr); + 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_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; } - down_read(&m->lod_qos.lq_rw_sem); ost_start_idx_temp = lqr->lqr_start_idx; repeat_find: - array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) % - osts->op_count; - - QOS_DEBUG("pool '%s' want %d startidx %d startcnt %d offset %d " - "active %d count %d arrayidx %d\n", - lo->ldo_pool ? lo->ldo_pool : "", - stripe_cnt, lqr->lqr_start_idx, lqr->lqr_start_count, - lqr->lqr_offset_idx, osts->op_count, osts->op_count, - array_idx); - - for (i = 0; i < osts->op_count && stripe_idx < lo->ldo_stripenr; - i++, array_idx = (array_idx + 1) % osts->op_count) { + + 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_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_count; i++) { + array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) % + osts->op_count; ++lqr->lqr_start_idx; ost_idx = lqr->lqr_pool.op_array[array_idx]; @@ -935,75 +1128,37 @@ repeat_find: if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0) continue; - rc = lod_statfs_and_check(env, m, ost_idx, sfs); - if (rc) { - /* this OSP doesn't feel well */ - continue; - } - - /* - * skip full devices - */ - if (lod_qos_dev_is_full(sfs)) { - QOS_DEBUG("#%d is full\n", ost_idx); - continue; - } - - /* - * We expect number of precreated objects in f_ffree at - * the first iteration, skip OSPs with no objects ready - */ - if (sfs->os_fprecreated == 0 && speed == 0) { - QOS_DEBUG("#%d: precreation is empty\n", ost_idx); - continue; - } - - /* - * try to use another OSP if this one is degraded - */ - if (sfs->os_state & OS_STATE_DEGRADED && speed < 2) { - QOS_DEBUG("#%d: degraded\n", ost_idx); - continue; - } - - /* - * do not put >1 objects on a single OST - */ - if (speed && lod_qos_is_ost_used(env, ost_idx, stripe_idx)) - continue; - - o = lod_qos_declare_object_on(env, m, 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); - continue; - } - - /* - * We've successfully declared (reserved) an object - */ - lod_qos_ost_in_use(env, stripe_idx, ost_idx); - stripe[stripe_idx] = o; - stripe_idx++; + spin_unlock(&lqr->lqr_alloc); + rc = lod_check_and_reserve_ost(env, lo, sfs, ost_idx, speed, + &stripe_idx, stripe, ost_indices, + th); + 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; + + ost_connecting = 0; goto repeat_find; } + spin_unlock(&lqr->lqr_alloc); up_read(&m->lod_qos.lq_rw_sem); if (stripe_idx) { - lo->ldo_stripenr = stripe_idx; + lod_comp->llc_stripe_count = stripe_idx; /* at least one stripe is allocated */ rc = 0; } else { /* nobody provided us with a single object */ - rc = -ENOSPC; + if (ost_connecting) + rc = -EINPROGRESS; + else + rc = -ENOSPC; } out: @@ -1029,56 +1184,62 @@ 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] 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) +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; - int rc; + int rc = -EINVAL; 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); - rc = lod_qos_ost_in_use_clear(env, lo->ldo_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 < lo->ldo_stripenr; i++) { - if (v3->lmm_objects[i].l_ost_idx == lo->ldo_def_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 == lo->ldo_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, lo->ldo_def_stripe_offset); + lod2obd(m)->obd_name, lod_comp->llc_stripe_offset); RETURN(-EINVAL); } - for (i = 0; i < lo->ldo_stripenr; - i++, array_idx = (array_idx + 1) % lo->ldo_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; @@ -1111,6 +1272,7 @@ static int lod_alloc_ost_list(const struct lu_env *env, */ lod_qos_ost_in_use(env, stripe_count, ost_idx); stripe[stripe_count] = o; + ost_indices[stripe_count] = ost_idx; stripe_count++; } @@ -1120,30 +1282,34 @@ static int lod_alloc_ost_list(const struct lu_env *env, /** * Allocate a striping on a predefined set of OSTs. * - * Allocates new striping starting from OST provided lo->ldo_def_stripe_offset. + * Allocates new layout starting from OST index in lo->ldo_stripe_offset. * Full OSTs are not considered. The exact order of OSTs is not important and * varies depending on OST status. The allocation procedure prefers the targets * with precreated objects ready. The number of stripes needed and stripe - * offset are taken from the object. If that number can not be met, then the - * function returns a failure and then it's the caller's responsibility to + * offset are taken from the object. If that number cannot be met, then the + * function returns an error and then it's the caller's responsibility to * 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] 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 -E2BIG if no enough OSTs are found + * \retval -ENOSPC if no OST objects are available at all + * \retval -EFBIG if not enough OST objects are found * \retval -EINVAL requested offset is invalid - * \retval negative negated errno on error + * \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) + 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; @@ -1155,12 +1321,15 @@ static int lod_alloc_specific(const struct lu_env *env, struct lod_object *lo, struct ost_pool *osts; ENTRY; - rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr); + 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_stripe_count); if (rc) GOTO(out, rc); - if (lo->ldo_pool) - pool = lod_find_pool(m, lo->ldo_pool); + if (lod_comp->llc_pool != NULL) + pool = lod_find_pool(m, lod_comp->llc_pool); if (pool != NULL) { down_read(&pool_tgt_rw_sem(pool)); @@ -1175,15 +1344,15 @@ repeat_find: /* search loi_ost_idx in ost array */ array_idx = 0; for (i = 0; i < ost_count; i++) { - if (osts->op_array[i] == lo->ldo_def_stripe_offset) { + if (osts->op_array[i] == lod_comp->llc_stripe_offset) { array_idx = i; break; } } if (i == ost_count) { CERROR("Start index %d not found in pool '%s'\n", - lo->ldo_def_stripe_offset, - lo->ldo_pool ? lo->ldo_pool : ""); + lod_comp->llc_stripe_offset, + lod_comp->llc_pool ? lod_comp->llc_pool : ""); GOTO(out, rc = -EINVAL); } @@ -1205,6 +1374,13 @@ repeat_find: if (lod_qos_is_ost_used(env, ost_idx, stripe_num)) continue; + /* + * try not allocate on the OST used by other component + */ + if (speed == 0 && i != 0 && + lod_comp_is_ost_used(env, lo, ost_idx)) + continue; + /* Drop slow OSCs if we can, but not for requested start idx. * * This means "if OSC is slow and it is not the requested @@ -1218,9 +1394,9 @@ 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; @@ -1237,10 +1413,11 @@ repeat_find: */ lod_qos_ost_in_use(env, stripe_num, ost_idx); stripe[stripe_num] = o; + ost_indices[stripe_num] = ost_idx; stripe_num++; /* We have enough stripes */ - if (stripe_num == lo->ldo_stripenr) + if (stripe_num == lod_comp->llc_stripe_count) GOTO(out, rc = 0); } if (speed < 2) { @@ -1252,13 +1429,11 @@ repeat_find: /* If we were passed specific striping params, then a failure to * meet those requirements is an error, since we can't reallocate * that memory (it might be part of a larger array or something). - * - * We can only get here if lsm_stripe_count was originally > 1. */ CERROR("can't lstripe objid "DFID": have %d want %u\n", PFID(lu_object_fid(lod2lu_obj(lo))), stripe_num, - lo->ldo_stripenr); - rc = -EFBIG; + lod_comp->llc_stripe_count); + rc = stripe_num == 0 ? -ENOSPC : -EFBIG; out: if (pool != NULL) { up_read(&pool_tgt_rw_sem(pool)); @@ -1305,101 +1480,108 @@ static inline int lod_qos_is_usable(struct lod_device *lod) * The function allocates OST objects to create a striping. The algorithm * used is based on weights (currently only using the free space), and it's * trying to ensure the space is used evenly by OSTs and OSSs. The striping - * configuration (# of stripes, offset, - * pool) is taken from the object and is prepared by the caller. - * If LOV_USES_DEFAULT_STRIPE is not passed and prepared configuration can't - * be met due to too few OSTs, then allocation fails. If the flag is - * passed and less than 75% of the requested number of stripes can be - * allocated, then allocation fails. - * No concurrent allocation is allowed on the object and this must be - * ensured by the caller. All the internal structures are protected by the - * function. - * The algorithm has two steps: find available OSTs and calucate their weights, - * then select the OSTs the weights used as the probability. An OST with a - * higher weight is proportionately more likely to be selected than one with - * a lower weight. + * configuration (# of stripes, offset, pool) is taken from the object and + * is prepared by the caller. * - * \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 + * If LOV_USES_DEFAULT_STRIPE is not passed and prepared configuration can't + * be met due to too few OSTs, then allocation fails. If the flag is passed + * fewer than 3/4 of the requested number of stripes can be allocated, then + * allocation fails. + * + * No concurrent allocation is allowed on the object and this must be ensured + * by the caller. All the internal structures are protected by the function. + * + * The algorithm has two steps: find available OSTs and calculate their + * weights, then select the OSTs with their weights used as the probability. + * 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[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 -E2BIG if no enough OSTs are found - * \retval -EINVAL requested offset is invalid - * \retval negative negated errno on error + * \retval -EAGAIN not enough OSTs are found for specified stripe count + * \retval -EINVAL requested OST index is invalid + * \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) + struct dt_object **stripe, __u32 *ost_indices, + int flags, struct thandle *th, int comp_idx) { - struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev); - struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs; + 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; - unsigned int i; - int rc = 0; - __u32 nfound, good_osts; - __u32 stripe_cnt = lo->ldo_stripenr; - __u32 stripe_cnt_min; - struct pool_desc *pool = NULL; - struct ost_pool *osts; + 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_count, stripe_count_min; + int rc = 0; ENTRY; - stripe_cnt_min = min_stripe_count(stripe_cnt, flags); - if (stripe_cnt_min < 1) + LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL); + lod_comp = &lo->ldo_comp_entries[comp_idx]; + stripe_count = lod_comp->llc_stripe_count; + stripe_count_min = min_stripe_count(stripe_count, flags); + if (stripe_count_min < 1) RETURN(-EINVAL); - if (lo->ldo_pool) - pool = lod_find_pool(m, lo->ldo_pool); + if (lod_comp->llc_pool != NULL) + pool = lod_find_pool(lod, lod_comp->llc_pool); if (pool != NULL) { down_read(&pool_tgt_rw_sem(pool)); osts = &(pool->pool_obds); } else { - osts = &(m->lod_pool_info); + osts = &(lod->lod_pool_info); } /* Detect -EAGAIN early, before expensive lock is taken. */ - if (!lod_qos_is_usable(m)) + if (!lod_qos_is_usable(lod)) GOTO(out_nolock, rc = -EAGAIN); /* Do actual allocation, use write lock here. */ - down_write(&m->lod_qos.lq_rw_sem); + down_write(&lod->lod_qos.lq_rw_sem); /* * Check again, while we were sleeping on @lq_rw_sem things could * change. */ - if (!lod_qos_is_usable(m)) + if (!lod_qos_is_usable(lod)) GOTO(out, rc = -EAGAIN); - rc = lod_qos_calc_ppo(m); + rc = lod_qos_calc_ppo(lod); if (rc) GOTO(out, rc); - rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr); + rc = lod_qos_ost_in_use_clear(env, lod_comp->llc_stripe_count); if (rc) GOTO(out, rc); good_osts = 0; /* Find all the OSTs that are valid stripe candidates */ for (i = 0; i < osts->op_count; i++) { - if (!cfs_bitmap_check(m->lod_ost_bitmap, osts->op_array[i])) + if (!cfs_bitmap_check(lod->lod_ost_bitmap, osts->op_array[i])) continue; - rc = lod_statfs_and_check(env, m, osts->op_array[i], sfs); + ost = OST_TGT(lod, osts->op_array[i]); + ost->ltd_qos.ltq_usable = 0; + + rc = lod_statfs_and_check(env, lod, osts->op_array[i], sfs); if (rc) { /* this OSP doesn't feel well */ continue; } - /* - * skip full devices - */ - if (lod_qos_dev_is_full(sfs)) + if (sfs->os_state & OS_STATE_DEGRADED) continue; /* Fail Check before osc_precreate() is called @@ -1408,9 +1590,8 @@ static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo, osts->op_array[i] == 0) continue; - ost = OST_TGT(m,osts->op_array[i]); ost->ltd_qos.ltq_usable = 1; - lod_qos_calc_weight(m, osts->op_array[i]); + lod_qos_calc_weight(lod, osts->op_array[i]); total_weight += ost->ltd_qos.ltq_weight; good_osts++; @@ -1418,16 +1599,16 @@ 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 (good_osts < stripe_count) + stripe_count = good_osts; /* Find enough OSTs with weighted random allocation. */ nfound = 0; - while (nfound < stripe_cnt) { + while (nfound < stripe_count) { __u64 rand, cur_weight; cur_weight = 0; @@ -1458,45 +1639,49 @@ static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo, rand = 0; } - /* On average, this will hit larger-weighted osts more often. - 0-weight osts will always get used last (only when rand=0) */ + /* 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(m->lod_ost_bitmap, idx)) + if (lod_should_avoid_ost(lo, lag, idx)) continue; - ost = OST_TGT(m,idx); + ost = OST_TGT(lod, idx); if (!ost->ltd_qos.ltq_usable) continue; cur_weight += ost->ltd_qos.ltq_weight; - QOS_DEBUG("stripe_cnt=%d nfound=%d cur_weight="LPU64 - " rand="LPU64" total_weight="LPU64"\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 */ - if (lod_qos_is_ost_used(env, idx, nfound)) + if (lod_qos_is_ost_used(env, idx, nfound) || + lod_comp_is_ost_used(env, lo, idx)) continue; - lod_qos_ost_in_use(env, nfound, idx); - o = lod_qos_declare_object_on(env, m, idx, th); + o = lod_qos_declare_object_on(env, lod, idx, th); if (IS_ERR(o)) { QOS_DEBUG("can't declare object on #%u: %d\n", idx, (int) PTR_ERR(o)); continue; } - stripe[nfound++] = o; - lod_qos_used(m, osts, idx, &total_weight); + + 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; } @@ -1507,7 +1692,7 @@ 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 @@ -1515,22 +1700,23 @@ static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo, * so it's possible OSP won't be able to provide us with * an object due to just changed state */ - LCONSOLE_INFO("wanted %d, found %d\n", stripe_cnt, nfound); + QOS_DEBUG("%s: wanted %d objects, found only %d\n", + lod2obd(lod)->obd_name, stripe_count, nfound); for (i = 0; i < nfound; i++) { LASSERT(stripe[i] != NULL); - lu_object_put(env, &stripe[i]->do_lu); + dt_object_put(env, stripe[i]); stripe[i] = NULL; } /* makes sense to rebalance next time */ - m->lod_qos.lq_dirty = 1; - m->lod_qos.lq_same_space = 0; + lod->lod_qos.lq_dirty = 1; + lod->lod_qos.lq_same_space = 0; rc = -EAGAIN; } out: - up_write(&m->lod_qos.lq_rw_sem); + up_write(&lod->lod_qos.lq_rw_sem); out_nolock: if (pool != NULL) { @@ -1551,13 +1737,13 @@ out_nolock: * If the passed one is 0, then the filesystem's default one is used. * * \param[in] lod LOD device - * \param[in] magic the format if striping + * \param[in] lo The lod_object * \param[in] stripe_count count the caller would like to use * * \retval the maximum usable stripe count */ -static __u16 lod_get_stripecnt(struct lod_device *lod, __u32 magic, - __u16 stripe_count) +__u16 lod_get_stripe_count(struct lod_device *lod, struct lod_object *lo, + __u16 stripe_count) { __u32 max_stripes = LOV_MAX_STRIPE_COUNT_OLD; @@ -1569,9 +1755,31 @@ static __u16 lod_get_stripecnt(struct lod_device *lod, __u32 magic, stripe_count = 1; /* stripe count is based on whether OSD can handle larger EA sizes */ - if (lod->lod_osd_max_easize > 0) - max_stripes = lov_mds_md_max_stripe_count( - lod->lod_osd_max_easize, magic); + 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_stripe_count, + LOV_MAGIC_V3); + } + 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; } @@ -1591,46 +1799,115 @@ static __u16 lod_get_stripecnt(struct lod_device *lod, __u32 magic, * \retval 0 on success * \retval negative negated errno on error */ -static int lod_use_defined_striping(const struct lu_env *env, - struct lod_object *mo, - const struct lu_buf *buf) +int lod_use_defined_striping(const struct lu_env *env, + struct lod_object *mo, + const struct lu_buf *buf) { + struct lod_layout_component *lod_comp; struct lov_mds_md_v1 *v1 = buf->lb_buf; struct lov_mds_md_v3 *v3 = buf->lb_buf; + struct lov_comp_md_v1 *comp_v1 = NULL; struct lov_ost_data_v1 *objs; - __u32 magic; - int rc = 0; + __u32 magic; + __u16 comp_cnt; + __u16 mirror_cnt; + int rc = 0, i; ENTRY; - magic = le32_to_cpu(v1->lmm_magic); - if (magic == LOV_MAGIC_V1_DEF) { - magic = LOV_MAGIC_V1; - objs = &v1->lmm_objects[0]; - } else if (magic == LOV_MAGIC_V3_DEF) { - magic = LOV_MAGIC_V3; - objs = &v3->lmm_objects[0]; - lod_object_set_pool(mo, v3->lmm_pool_name); + 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) + 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) + 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 { - GOTO(out, rc = -EINVAL); + mo->ldo_is_composite = 0; + comp_cnt = 1; + mirror_cnt = 0; } - - mo->ldo_pattern = le32_to_cpu(v1->lmm_pattern); - mo->ldo_stripe_size = le32_to_cpu(v1->lmm_stripe_size); - mo->ldo_stripenr = le16_to_cpu(v1->lmm_stripe_count); mo->ldo_layout_gen = le16_to_cpu(v1->lmm_layout_gen); - /* fixup for released file before object initialization */ - if (mo->ldo_pattern & LOV_PATTERN_F_RELEASED) { - mo->ldo_released_stripenr = mo->ldo_stripenr; - mo->ldo_stripenr = 0; - } + rc = lod_alloc_comp_entries(mo, mirror_cnt, comp_cnt); + if (rc) + GOTO(unlock, rc); + + for (i = 0; i < comp_cnt; i++) { + struct lu_extent *ext; + char *pool_name; + __u32 offs; + + lod_comp = &mo->ldo_comp_entries[i]; + + 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); + magic = le32_to_cpu(v1->lmm_magic); + + ext = &comp_v1->lcm_entries[i].lcme_extent; + lod_comp->llc_extent.e_start = + le64_to_cpu(ext->e_start); + 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); + lod_comp->llc_id = + le32_to_cpu(comp_v1->lcm_entries[i].lcme_id); + if (lod_comp->llc_id == LCME_ID_INVAL) + GOTO(out, rc = -EINVAL); + } - LASSERT(buf->lb_len >= lov_mds_md_size(mo->ldo_stripenr, magic)); + pool_name = NULL; + if (magic == LOV_MAGIC_V1) { + objs = &v1->lmm_objects[0]; + } else if (magic == LOV_MAGIC_V3) { + objs = &v3->lmm_objects[0]; + if (v3->lmm_pool_name[0] != '\0') + pool_name = v3->lmm_pool_name; + } else { + CDEBUG(D_LAYOUT, "Invalid magic %x\n", magic); + GOTO(out, rc = -EINVAL); + } - if (mo->ldo_stripenr > 0) - rc = lod_initialize_objects(env, mo, objs); + 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_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 ((!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_striping_free_nolock(env, mo); +unlock: + mutex_unlock(&mo->ldo_layout_mutex); + RETURN(rc); } @@ -1651,27 +1928,43 @@ 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_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev); - struct lov_user_md_v1 *v1 = NULL; - struct lov_user_md_v3 *v3 = NULL; - char *pool_name = NULL; - __u32 magic; - int rc; - unsigned int size; + 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; + 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); + 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 */ + lod_free_comp_entries(lo); + + rc = lod_verify_striping(d, lo, buf, false); + if (rc) + RETURN(-EINVAL); + v3 = buf->lb_buf; v1 = buf->lb_buf; + comp_v1 = buf->lb_buf; magic = v1->lmm_magic; - if (unlikely(magic == LOV_MAGIC_V1_DEF || magic == LOV_MAGIC_V3_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); @@ -1683,18 +1976,13 @@ static int lod_qos_parse_config(const struct lu_env *env, magic = v1->lmm_magic; /* fall through */ case LOV_USER_MAGIC_V1: - size = sizeof(*v1); break; - case __swab32(LOV_USER_MAGIC_V3): lustre_swab_lov_user_md_v3(v3); magic = v3->lmm_magic; /* fall through */ case LOV_USER_MAGIC_V3: - size = sizeof(*v3); - pool_name = v3->lmm_pool_name; break; - case __swab32(LOV_USER_MAGIC_SPECIFIC): lustre_swab_lov_user_md_v3(v3); lustre_swab_lov_user_md_objects(v3->lmm_objects, @@ -1702,93 +1990,271 @@ static int lod_qos_parse_config(const struct lu_env *env, magic = v3->lmm_magic; /* fall through */ case LOV_USER_MAGIC_SPECIFIC: - if (v3->lmm_stripe_offset == LOV_OFFSET_DEFAULT) - v3->lmm_stripe_offset = v3->lmm_objects[0].l_ost_idx; - if (v3->lmm_pool_name[0] != '\0') - pool_name = v3->lmm_pool_name; - size = lov_user_md_size(v3->lmm_stripe_count, - LOV_USER_MAGIC_SPECIFIC); break; - + case __swab32(LOV_USER_MAGIC_COMP_V1): + lustre_swab_lov_comp_md_v1(comp_v1); + magic = comp_v1->lcm_magic; + /* fall trhough */ + case LOV_USER_MAGIC_COMP_V1: + break; default: CERROR("%s: unrecognized magic %X\n", lod2obd(d)->obd_name, magic); RETURN(-EINVAL); } - if (unlikely(buf->lb_len < size)) { - CERROR("%s: wrong size: %zd, expect: %u\n", - lod2obd(d)->obd_name, buf->lb_len, size); - RETURN(-EINVAL); - } - lustre_print_user_md(D_OTHER, v1, "parse config"); - v1->lmm_magic = magic; - if (v1->lmm_pattern == 0) - v1->lmm_pattern = LOV_PATTERN_RAID0; - if (lov_pattern(v1->lmm_pattern) != LOV_PATTERN_RAID0) { - CERROR("%s: invalid pattern: %x\n", - lod2obd(d)->obd_name, v1->lmm_pattern); - RETURN(-EINVAL); + if (magic == LOV_USER_MAGIC_COMP_V1) { + 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; } - lo->ldo_pattern = v1->lmm_pattern; - if (v1->lmm_stripe_size > 0) - lo->ldo_stripe_size = v1->lmm_stripe_size; + rc = lod_alloc_comp_entries(lo, mirror_cnt, comp_cnt); + if (rc) + RETURN(rc); + + LASSERT(lo->ldo_comp_entries); - if (lo->ldo_stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) - lo->ldo_stripe_size = LOV_MIN_STRIPE_SIZE; + for (i = 0; i < comp_cnt; i++) { + struct pool_desc *pool; + struct lu_extent *ext; + char *pool_name; - if (v1->lmm_stripe_count > 0) - lo->ldo_stripenr = v1->lmm_stripe_count; + lod_comp = &lo->ldo_comp_entries[i]; - lo->ldo_def_stripe_offset = v1->lmm_stripe_offset; + if (lo->ldo_is_composite) { + v1 = (struct lov_user_md *)((char *)comp_v1 + + 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_USER_FLAGS; + } - lod_object_set_pool(lo, NULL); - if (pool_name != NULL) { - struct pool_desc *pool; + 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) { + 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 && + lov_pattern(v1->lmm_pattern) != LOV_PATTERN_MDT) { + CDEBUG(D_LAYOUT, "%s: invalid pattern: %x\n", + lod2obd(d)->obd_name, v1->lmm_pattern); + 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) + lod_comp->llc_stripe_size = v1->lmm_stripe_size; + + 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; /* In the function below, .hs_keycmp resolves to * pool_hashkey_keycmp() */ /* coverity[overrun-buffer-val] */ pool = lod_find_pool(d, pool_name); - if (pool != NULL) { - if (lo->ldo_def_stripe_offset != LOV_OFFSET_DEFAULT) { - rc = lod_check_index_in_pool( - lo->ldo_def_stripe_offset, pool); - if (rc < 0) { - lod_pool_putref(pool); - CERROR("%s: invalid offset, %u\n", - lod2obd(d)->obd_name, - lo->ldo_def_stripe_offset); - RETURN(-EINVAL); - } + if (pool == NULL) + continue; + + if (lod_comp->llc_stripe_offset != LOV_OFFSET_DEFAULT) { + rc = lod_check_index_in_pool( + lod_comp->llc_stripe_offset, pool); + if (rc < 0) { + lod_pool_putref(pool); + CDEBUG(D_LAYOUT, "%s: invalid offset, %u\n", + lod2obd(d)->obd_name, + lod_comp->llc_stripe_offset); + GOTO(free_comp, rc = -EINVAL); } + } + + if (lod_comp->llc_stripe_count > pool_tgt_count(pool)) + lod_comp->llc_stripe_count = pool_tgt_count(pool); + + lod_pool_putref(pool); + } + + RETURN(0); + +free_comp: + lod_free_comp_entries(lo); + RETURN(rc); +} - if (lo->ldo_stripenr > pool_tgt_count(pool)) - lo->ldo_stripenr = pool_tgt_count(pool); +/** + * 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; + } - lod_pool_putref(pool); + /* 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; } + } - lod_object_set_pool(lo, pool_name); + if (!lag->lag_ost_avoid_bitmap) { + bitmap = CFS_ALLOCATE_BITMAP(ltds->ltd_tgtnr); + if (!bitmap) + return -ENOMEM; } - /* fixup for released file */ - if (lo->ldo_pattern & LOV_PATTERN_F_RELEASED) { - lo->ldo_released_stripenr = lo->ldo_stripenr; - lo->ldo_stripenr = 0; + 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; + } } - RETURN(0); + 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 */ + if (!lu_extent_is_overlapped(&comp->llc_extent, + &lod_comp->llc_extent) || + !lod_comp_inited(comp) || !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 lod_qos_oss *lqo; + int k; + + ost = OST_TGT(lod, comp->llc_ost_indices[j]); + lqo = ost->ltd_qos.ltq_oss; + + if (cfs_bitmap_check(bitmap, ost->ltd_index)) + continue; + + 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] == + lqo->lqo_id) + break; + } + if (k == lag->lag_oaa_count) { + lag->lag_oss_avoid_array[k] = + lqo->lqo_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 @@ -1797,96 +2263,99 @@ 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 * * \retval 0 on success * \retval negative negated errno on error */ 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) + 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; + 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]; - /* no OST available */ - /* XXX: should we be waiting a bit to prevent failures during - * cluster initialization? */ - if (d->lod_ostnr == 0) - GOTO(out, rc = -EIO); - - /* - * by this time, the object's ldo_stripenr and ldo_stripe_size - * contain default value for striping: taken from the parent - * or from filesystem defaults - * - * in case the caller is passing lovea with new striping config, - * we may need to parse lovea and apply new configuration - */ - rc = lod_qos_parse_config(env, lo, buf); - if (rc) - GOTO(out, rc); - - /* A released file is being created */ - if (lo->ldo_stripenr == 0) - GOTO(out, rc = 0); + /* A released component is being created */ + if (lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED) + RETURN(0); - if (likely(lo->ldo_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(lo->ldo_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); - lo->ldo_stripenr = lod_get_stripecnt(d, LOV_MAGIC, - lo->ldo_stripenr); - - stripe_len = lo->ldo_stripenr; + stripe_len = lod_get_stripe_count(d, lo, + lod_comp->llc_stripe_count); + if (stripe_len == 0) + GOTO(out, rc = -ERANGE); + 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 (lod_comp->llc_ostlist.op_array) { + 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); + + 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); - } else if (lo->ldo_def_stripe_offset == LOV_OFFSET_DEFAULT) { - rc = lod_alloc_qos(env, lo, stripe, flag, th); + 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); + rc = lod_alloc_rr(env, lo, stripe, ost_indices, + flag, th, comp_idx); } else { - rc = lod_alloc_specific(env, lo, stripe, flag, th); + 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) - lu_object_put(env, &stripe[i]->do_lu); - - OBD_FREE(stripe, sizeof(stripe[0]) * stripe_len); - lo->ldo_stripenr = 0; + dt_object_put(env, stripe[i]); + lod_comp->llc_stripe_count = 0; } else { - lo->ldo_stripe = stripe; - lo->ldo_stripes_allocated = stripe_len; + lod_comp->llc_stripe = stripe; + lod_comp->llc_ost_indices = ost_indices; + lod_comp->llc_stripes_allocated = stripe_len; } } else { /* @@ -1894,22 +2363,91 @@ int lod_qos_prep_create(const struct lu_env *env, struct lod_object *lo, * striping (not a hint), so it allocated all the object * now we need to create them */ - for (i = 0; i < lo->ldo_stripenr; i++) { + for (i = 0; i < lod_comp->llc_stripe_count; i++) { struct dt_object *o; - o = lo->ldo_stripe[i]; + 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: + if (rc < 0) { + if (stripe) + OBD_FREE(stripe, sizeof(stripe[0]) * stripe_len); + if (ost_indices) + OBD_FREE(ost_indices, + sizeof(*ost_indices) * stripe_len); + } RETURN(rc); } +int lod_prepare_create(const struct lu_env *env, struct lod_object *lo, + struct lu_attr *attr, const struct lu_buf *buf, + struct thandle *th) + +{ + struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev); + uint64_t size = 0; + int i; + int rc; + ENTRY; + + LASSERT(lo); + + /* no OST available */ + /* XXX: should we be waiting a bit to prevent failures during + * cluster initialization? */ + if (d->lod_ostnr == 0) + RETURN(-EIO); + + /* + * 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 + * + * in case the caller is passing lovea with new striping config, + * we may need to parse lovea and apply new configuration + */ + rc = lod_qos_parse_config(env, lo, buf); + if (rc) + RETURN(rc); + + if (attr->la_valid & LA_SIZE) + size = attr->la_size; + + /** + * 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; + + lod_comp = &lo->ldo_comp_entries[i]; + extent = &lod_comp->llc_extent; + CDEBUG(D_QOS, "%lld [%lld, %lld)\n", + size, extent->e_start, extent->e_end); + if (!lo->ldo_is_composite || size >= extent->e_start) { + rc = lod_qos_prep_create(env, lo, attr, th, i); + if (rc) + break; + } + } + + RETURN(rc); +}