X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Flov%2Flov_qos.c;h=7d8ae114008aeb999cb0fdaff176daff79a8fdaa;hp=053ef6c74528f8b859107b060b2f66a5ce2122c1;hb=cec72a356891eaa729314a7bc89c4b2aaef0a31b;hpb=6869932b552ac705f411de3362f01bd50c1f6f7d diff --git a/lustre/lov/lov_qos.c b/lustre/lov/lov_qos.c index 053ef6c..7d8ae11 100644 --- a/lustre/lov/lov_qos.c +++ b/lustre/lov/lov_qos.c @@ -26,7 +26,7 @@ * GPL HEADER END */ /* - * Copyright 2008 Sun Microsystems, Inc. All rights reserved + * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. */ /* @@ -47,6 +47,7 @@ #include #include +#include #include "lov_internal.h" /* #define QOS_DEBUG 1 */ @@ -54,7 +55,6 @@ #define TGT_BAVAIL(i) (lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bavail *\ lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bsize) -#define TGT_FFREE(i) (lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_ffree) int qos_add_tgt(struct obd_device *obd, __u32 index) @@ -72,9 +72,9 @@ int qos_add_tgt(struct obd_device *obd, __u32 index) RETURN(-ENOTCONN); } - down_write(&lov->lov_qos.lq_rw_sem); - mutex_down(&lov->lov_lock); - list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) { + cfs_down_write(&lov->lov_qos.lq_rw_sem); + cfs_mutex_down(&lov->lov_lock); + cfs_list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) { if (obd_uuid_equals(&oss->lqo_uuid, &exp->exp_connection->c_remote_uuid)) { found++; @@ -91,7 +91,7 @@ int qos_add_tgt(struct obd_device *obd, __u32 index) sizeof(oss->lqo_uuid)); } else { /* Assume we have to move this one */ - list_del(&oss->lqo_oss_list); + cfs_list_del(&oss->lqo_oss_list); } oss->lqo_ost_count++; @@ -99,16 +99,17 @@ int qos_add_tgt(struct obd_device *obd, __u32 index) /* Add sorted by # of OSTs. Find the first entry that we're bigger than... */ - list_for_each_entry(temposs, &lov->lov_qos.lq_oss_list, lqo_oss_list) { + cfs_list_for_each_entry(temposs, &lov->lov_qos.lq_oss_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); + cfs_list_add_tail(&oss->lqo_oss_list, &temposs->lqo_oss_list); lov->lov_qos.lq_dirty = 1; - lov->lov_qos.lq_dirty_rr = 1; + lov->lov_qos.lq_rr.lqr_dirty = 1; CDEBUG(D_QOS, "add tgt %s to OSS %s (%d OSTs)\n", obd_uuid2str(&lov->lov_tgts[index]->ltd_uuid), @@ -116,24 +117,21 @@ int qos_add_tgt(struct obd_device *obd, __u32 index) oss->lqo_ost_count); out: - mutex_up(&lov->lov_lock); - up_write(&lov->lov_qos.lq_rw_sem); + cfs_mutex_up(&lov->lov_lock); + cfs_up_write(&lov->lov_qos.lq_rw_sem); RETURN(rc); } -int qos_del_tgt(struct obd_device *obd, __u32 index) +int qos_del_tgt(struct obd_device *obd, struct lov_tgt_desc *tgt) { struct lov_obd *lov = &obd->u.lov; struct lov_qos_oss *oss; int rc = 0; ENTRY; - if (!lov->lov_tgts[index]) - RETURN(0); - - down_write(&lov->lov_qos.lq_rw_sem); + cfs_down_write(&lov->lov_qos.lq_rw_sem); - oss = lov->lov_tgts[index]->ltd_qos.ltq_oss; + oss = tgt->ltd_qos.ltq_oss; if (!oss) GOTO(out, rc = -ENOENT); @@ -141,14 +139,14 @@ int qos_del_tgt(struct obd_device *obd, __u32 index) if (oss->lqo_ost_count == 0) { CDEBUG(D_QOS, "removing OSS %s\n", obd_uuid2str(&oss->lqo_uuid)); - list_del(&oss->lqo_oss_list); + cfs_list_del(&oss->lqo_oss_list); OBD_FREE_PTR(oss); } lov->lov_qos.lq_dirty = 1; - lov->lov_qos.lq_dirty_rr = 1; + lov->lov_qos.lq_rr.lqr_dirty = 1; out: - up_write(&lov->lov_qos.lq_rw_sem); + cfs_up_write(&lov->lov_qos.lq_rw_sem); RETURN(rc); } @@ -161,6 +159,7 @@ static int qos_calc_ppo(struct obd_device *obd) __u64 ba_max, ba_min, temp; __u32 num_active; int rc, i, prio_wide; + time_t now, age; ENTRY; if (!lov->lov_qos.lq_dirty) @@ -171,7 +170,7 @@ static int qos_calc_ppo(struct obd_device *obd) GOTO(out, rc = -EAGAIN); /* find bavail on each OSS */ - list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) { + cfs_list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) { oss->lqo_bavail = 0; } lov->lov_qos.lq_active_oss_count = 0; @@ -183,6 +182,7 @@ static int qos_calc_ppo(struct obd_device *obd) ba_min = (__u64)(-1); ba_max = 0; + now = cfs_time_current_sec(); /* Calculate OST penalty per object */ /* (lov ref taken in alloc_qos) */ for (i = 0; i < lov->desc.ld_tgt_count; i++) { @@ -205,8 +205,17 @@ static int qos_calc_ppo(struct obd_device *obd) lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj = (temp * prio_wide) >> 8; - if (lov->lov_qos.lq_reset == 0) + age = (now - lov->lov_tgts[i]->ltd_qos.ltq_used) >> 3; + if (lov->lov_qos.lq_reset || age > 32 * lov->desc.ld_qos_maxage) lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0; + else if (age > lov->desc.ld_qos_maxage) + /* Decay the penalty by half for every 8x the update + * interval that the device has been idle. That gives + * lots of time for the statfs information to be + * updated (which the penalty is only a proxy for), + * and avoids penalizing OSS/OSTs under light load. */ + lov->lov_tgts[i]->ltd_qos.ltq_penalty >>= + (age / lov->desc.ld_qos_maxage); } num_active = lov->lov_qos.lq_active_oss_count - 1; @@ -222,12 +231,21 @@ static int qos_calc_ppo(struct obd_device *obd) } /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */ - list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) { + cfs_list_for_each_entry(oss, &lov->lov_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; - if (lov->lov_qos.lq_reset == 0) + + age = (now - oss->lqo_used) >> 3; + if (lov->lov_qos.lq_reset || age > 32 * lov->desc.ld_qos_maxage) oss->lqo_penalty = 0; + else if (age > lov->desc.ld_qos_maxage) + /* Decay the penalty by half for every 8x the update + * interval that the device has been idle. That gives + * lots of time for the statfs information to be + * updated (which the penalty is only a proxy for), + * and avoids penalizing OSS/OSTs under light load. */ + oss->lqo_penalty >>= (age / lov->desc.ld_qos_maxage); } lov->lov_qos.lq_dirty = 0; @@ -236,13 +254,10 @@ static int qos_calc_ppo(struct obd_device *obd) /* If each ost has almost same free space, * do rr allocation for better creation performance */ lov->lov_qos.lq_same_space = 0; - temp = ba_max - ba_min; - ba_min = (ba_min * 51) >> 8; /* 51/256 = .20 */ - if (temp < ba_min) { - /* Difference is less than 20% */ + if ((ba_max * (256 - lov->lov_qos.lq_threshold_rr)) >> 8 < ba_min) { lov->lov_qos.lq_same_space = 1; /* Reset weights for the next time we enter qos mode */ - lov->lov_qos.lq_reset = 0; + lov->lov_qos.lq_reset = 1; } rc = 0; @@ -268,10 +283,11 @@ static int qos_calc_weight(struct lov_obd *lov, int i) } /* We just used this index for a stripe; adjust everyone's weights */ -static int qos_used(struct lov_obd *lov, __u32 index, __u64 *total_wt) +static int qos_used(struct lov_obd *lov, struct ost_pool *osts, + __u32 index, __u64 *total_wt) { struct lov_qos_oss *oss; - int i; + int j; ENTRY; /* Don't allocate from this stripe anymore, until the next alloc_qos */ @@ -284,6 +300,10 @@ static int qos_used(struct lov_obd *lov, __u32 index, __u64 *total_wt) lov->lov_tgts[index]->ltd_qos.ltq_penalty >>= 1; oss->lqo_penalty >>= 1; + /* mark the OSS and OST as recently used */ + lov->lov_tgts[index]->ltd_qos.ltq_used = + oss->lqo_used = cfs_time_current_sec(); + /* Set max penalties for this OST and OSS */ lov->lov_tgts[index]->ltd_qos.ltq_penalty += lov->lov_tgts[index]->ltd_qos.ltq_penalty_per_obj * @@ -292,7 +312,7 @@ static int qos_used(struct lov_obd *lov, __u32 index, __u64 *total_wt) lov->lov_qos.lq_active_oss_count; /* Decrease all OSS penalties */ - list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) { + cfs_list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) { if (oss->lqo_penalty < oss->lqo_penalty_per_obj) oss->lqo_penalty = 0; else @@ -301,7 +321,10 @@ static int qos_used(struct lov_obd *lov, __u32 index, __u64 *total_wt) *total_wt = 0; /* Decrease all OST penalties */ - for (i = 0; i < lov->desc.ld_tgt_count; i++) { + for (j = 0; j < osts->op_count; j++) { + int i; + + i = osts->op_array[j]; if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active) continue; if (lov->lov_tgts[i]->ltd_qos.ltq_penalty < @@ -318,10 +341,11 @@ static int qos_used(struct lov_obd *lov, __u32 index, __u64 *total_wt) *total_wt += lov->lov_tgts[i]->ltd_qos.ltq_weight; #ifdef QOS_DEBUG - CDEBUG(D_QOS, "recalc tgt %d avail="LPU64 + CDEBUG(D_QOS, "recalc tgt %d usable=%d avail="LPU64 " ostppo="LPU64" ostp="LPU64" ossppo="LPU64 " ossp="LPU64" wt="LPU64"\n", - i, TGT_BAVAIL(i) >> 10, + i, lov->lov_tgts[i]->ltd_qos.ltq_usable, + TGT_BAVAIL(i) >> 10, lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj >> 10, lov->lov_tgts[i]->ltd_qos.ltq_penalty >> 10, lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty_per_obj>>10, @@ -335,92 +359,86 @@ static int qos_used(struct lov_obd *lov, __u32 index, __u64 *total_wt) #define LOV_QOS_EMPTY ((__u32)-1) /* compute optimal round-robin order, based on OSTs per OSS */ -static int qos_calc_rr(struct lov_obd *lov) +static int qos_calc_rr(struct lov_obd *lov, struct ost_pool *src_pool, + struct lov_qos_rr *lqr) { struct lov_qos_oss *oss; - unsigned ost_count, placed, real_count; - int i; + unsigned placed, real_count; + int i, rc; ENTRY; - if (!lov->lov_qos.lq_dirty_rr) { - LASSERT(lov->lov_qos.lq_rr_size); + if (!lqr->lqr_dirty) { + LASSERT(lqr->lqr_pool.op_size); RETURN(0); } /* Do actual allocation. */ - down_write(&lov->lov_qos.lq_rw_sem); + cfs_down_write(&lov->lov_qos.lq_rw_sem); /* * Check again. While we were sleeping on @lq_rw_sem something could * change. */ - if (!lov->lov_qos.lq_dirty_rr) { - LASSERT(lov->lov_qos.lq_rr_size); - up_write(&lov->lov_qos.lq_rw_sem); + if (!lqr->lqr_dirty) { + LASSERT(lqr->lqr_pool.op_size); + cfs_up_write(&lov->lov_qos.lq_rw_sem); RETURN(0); } - ost_count = lov->desc.ld_tgt_count; - - if (lov->lov_qos.lq_rr_size) - OBD_FREE(lov->lov_qos.lq_rr_array, lov->lov_qos.lq_rr_size); - lov->lov_qos.lq_rr_size = ost_count * - sizeof(lov->lov_qos.lq_rr_array[0]); - OBD_ALLOC(lov->lov_qos.lq_rr_array, lov->lov_qos.lq_rr_size); - if (!lov->lov_qos.lq_rr_array) { - lov->lov_qos.lq_rr_size = 0; - up_write(&lov->lov_qos.lq_rw_sem); - RETURN(-ENOMEM); - } - - real_count = 0; - for (i = 0; i < ost_count; i++) { - lov->lov_qos.lq_rr_array[i] = LOV_QOS_EMPTY; - if (lov->lov_tgts[i]) - real_count++; + real_count = src_pool->op_count; + + /* Zero the pool array */ + /* alloc_rr is holding a read lock on the pool, so nobody is adding/ + deleting from the pool. The lq_rw_sem insures that nobody else + is reading. */ + lqr->lqr_pool.op_count = real_count; + rc = lov_ost_pool_extend(&lqr->lqr_pool, real_count); + if (rc) { + cfs_up_write(&lov->lov_qos.lq_rw_sem); + RETURN(rc); } + for (i = 0; i < lqr->lqr_pool.op_count; i++) + lqr->lqr_pool.op_array[i] = LOV_QOS_EMPTY; /* Place all the OSTs from 1 OSS at the same time. */ placed = 0; - list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) { + cfs_list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) { int j = 0; - for (i = 0; i < ost_count; i++) { - if (lov->lov_tgts[i] && - lov->lov_tgts[i]->ltd_qos.ltq_oss == oss) { + for (i = 0; i < lqr->lqr_pool.op_count; i++) { + if (lov->lov_tgts[src_pool->op_array[i]] && + (lov->lov_tgts[src_pool->op_array[i]]->ltd_qos.ltq_oss == oss)) { /* Evenly space these OSTs across arrayspace */ - int next = j * ost_count / oss->lqo_ost_count; - LASSERT(next < ost_count); - while (lov->lov_qos.lq_rr_array[next] != + int next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count; + while (lqr->lqr_pool.op_array[next] != LOV_QOS_EMPTY) - next = (next + 1) % ost_count; - lov->lov_qos.lq_rr_array[next] = i; + next = (next + 1) % lqr->lqr_pool.op_count; + lqr->lqr_pool.op_array[next] = src_pool->op_array[i]; j++; placed++; } } - LASSERT(j == oss->lqo_ost_count); } - lov->lov_qos.lq_dirty_rr = 0; - up_write(&lov->lov_qos.lq_rw_sem); + lqr->lqr_dirty = 0; + cfs_up_write(&lov->lov_qos.lq_rw_sem); if (placed != real_count) { /* This should never happen */ LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the " "round-robin list (%d of %d).\n", placed, real_count); - for (i = 0; i < ost_count; i++) { + for (i = 0; i < lqr->lqr_pool.op_count; i++) { LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i, - lov->lov_qos.lq_rr_array[i]); + lqr->lqr_pool.op_array[i]); } - lov->lov_qos.lq_dirty_rr = 1; + lqr->lqr_dirty = 1; RETURN(-EAGAIN); } #ifdef QOS_DEBUG - for (i = 0; i < ost_count; i++) { + for (i = 0; i < lqr->lqr_pool.op_count; i++) { LCONSOLE(D_QOS, "rr #%d ost idx=%d\n", i, - lov->lov_qos.lq_rr_array[i]); + lqr->lqr_pool.op_array[i]); } #endif @@ -443,7 +461,7 @@ void qos_shrink_lsm(struct lov_request_set *set) oti_alloc_cookies(set->set_oti, set->set_count); if (set->set_oti->oti_logcookies) { memcpy(set->set_oti->oti_logcookies, cookies, newsize); - OBD_FREE(cookies, oldsize); + OBD_FREE_LARGE(cookies, oldsize); set->set_cookies = set->set_oti->oti_logcookies; } else { CWARN("'leaking' %d bytes\n", oldsize - newsize); @@ -455,7 +473,7 @@ void qos_shrink_lsm(struct lov_request_set *set) LASSERT(lsm->lsm_stripe_count >= set->set_count); newsize = lov_stripe_md_size(set->set_count); - OBD_ALLOC(lsm_new, newsize); + OBD_ALLOC_LARGE(lsm_new, newsize); if (lsm_new != NULL) { int i; memcpy(lsm_new, lsm, sizeof(*lsm)); @@ -468,44 +486,88 @@ void qos_shrink_lsm(struct lov_request_set *set) sizeof(struct lov_oinfo)); } lsm_new->lsm_stripe_count = set->set_count; - OBD_FREE(lsm, sizeof(struct lov_stripe_md) + - lsm->lsm_stripe_count * sizeof(struct lov_oinfo *)); + OBD_FREE_LARGE(lsm, sizeof(struct lov_stripe_md) + + lsm->lsm_stripe_count*sizeof(struct lov_oinfo*)); set->set_oi->oi_md = lsm_new; } else { CWARN("'leaking' few bytes\n"); } } +/** + * Check whether we can create the object on the OST(refered by ost_idx) + * \retval: + * 0: create the object. + * other value: did not create the object. + */ +static int lov_check_and_create_object(struct lov_obd *lov, int ost_idx, + struct lov_stripe_md *lsm, + struct lov_request *req, + struct obd_trans_info *oti) +{ + int stripe; + int rc = -EIO; + ENTRY; + + CDEBUG(D_QOS, "Check and create on idx %d \n", ost_idx); + if (!lov->lov_tgts[ost_idx] || + !lov->lov_tgts[ost_idx]->ltd_active) + RETURN(rc); + + /* check if objects has been created on this ost */ + for (stripe = 0; stripe < lsm->lsm_stripe_count; stripe++) { + /* already have object at this stripe */ + if (ost_idx == lsm->lsm_oinfo[stripe]->loi_ost_idx) + break; + } + + if (stripe >= lsm->lsm_stripe_count) { + req->rq_idx = ost_idx; + rc = obd_create(lov->lov_tgts[ost_idx]->ltd_exp, + req->rq_oi.oi_oa, &req->rq_oi.oi_md, + oti); + } + RETURN(rc); +} + int qos_remedy_create(struct lov_request_set *set, struct lov_request *req) { struct lov_stripe_md *lsm = set->set_oi->oi_md; struct lov_obd *lov = &set->set_exp->exp_obd->u.lov; - unsigned ost_idx, ost_count = lov->desc.ld_tgt_count; - int stripe, i, rc = -EIO; + unsigned ost_idx = 0, ost_count; + struct pool_desc *pool; + struct ost_pool *osts = NULL; + int i, rc = -EIO; ENTRY; - ost_idx = (req->rq_idx + lsm->lsm_stripe_count) % ost_count; - for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) { - if (!lov->lov_tgts[ost_idx] || - !lov->lov_tgts[ost_idx]->ltd_active) - continue; - /* check if objects has been created on this ost */ - for (stripe = 0; stripe < lsm->lsm_stripe_count; stripe++) { - if (stripe == req->rq_stripe) - continue; - if (ost_idx == lsm->lsm_oinfo[stripe]->loi_ost_idx) + /* First check whether we can create the objects on the pool */ + pool = lov_find_pool(lov, lsm->lsm_pool_name); + if (pool != NULL) { + cfs_down_read(&pool_tgt_rw_sem(pool)); + osts = &(pool->pool_obds); + ost_count = osts->op_count; + for (i = 0; i < ost_count; i++, ost_idx = osts->op_array[i]) { + rc = lov_check_and_create_object(lov, ost_idx, lsm, req, + set->set_oti); + if (rc == 0) break; } + cfs_up_read(&pool_tgt_rw_sem(pool)); + lov_pool_putref(pool); + RETURN(rc); + } - if (stripe >= lsm->lsm_stripe_count) { - req->rq_idx = ost_idx; - rc = obd_create(lov->lov_tgts[ost_idx]->ltd_exp, - req->rq_oi.oi_oa, &req->rq_oi.oi_md, - set->set_oti); - if (!rc) - break; - } + ost_count = lov->desc.ld_tgt_count; + /* Then check whether we can create the objects on other OSTs */ + ost_idx = (req->rq_idx + lsm->lsm_stripe_count) % ost_count; + for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) { + rc = lov_check_and_create_object(lov, ost_idx, lsm, req, + set->set_oti); + + if (rc == 0) + break; } + RETURN(rc); } @@ -519,54 +581,67 @@ static int min_stripe_count(int stripe_cnt, int flags) #define LOV_CREATE_RESEED_MIN 1000 /* Allocate objects on osts with round-robin algorithm */ static int alloc_rr(struct lov_obd *lov, int *idx_arr, int *stripe_cnt, - int flags) + char *poolname, int flags) { - unsigned array_idx, ost_count = lov->desc.ld_tgt_count; - unsigned ost_active_count = lov->desc.ld_active_tgt_count; - int i, *idx_pos; + unsigned array_idx; + int i, rc, *idx_pos; __u32 ost_idx; int ost_start_idx_temp; int speed = 0; int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags); + struct pool_desc *pool; + struct ost_pool *osts; + struct lov_qos_rr *lqr; ENTRY; - i = qos_calc_rr(lov); - if (i) - RETURN(i); - - if (--lov->lov_start_count <= 0) { - lov->lov_start_idx = ll_rand() % ost_count; - lov->lov_start_count = - (LOV_CREATE_RESEED_MIN / max(ost_active_count, 1U) + - LOV_CREATE_RESEED_MULT) * max(ost_active_count, 1U); - } else if (stripe_cnt_min >= ost_active_count || - lov->lov_start_idx > ost_count) { + pool = lov_find_pool(lov, poolname); + if (pool == NULL) { + osts = &(lov->lov_packed); + lqr = &(lov->lov_qos.lq_rr); + } else { + cfs_down_read(&pool_tgt_rw_sem(pool)); + osts = &(pool->pool_obds); + lqr = &(pool->pool_rr); + } + + rc = qos_calc_rr(lov, osts, lqr); + if (rc) + GOTO(out, rc); + + 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 || + 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. */ - lov->lov_start_idx %= ost_count; - if (*stripe_cnt > 1 && (ost_active_count % (*stripe_cnt)) != 1) - ++lov->lov_offset_idx; + lqr->lqr_start_idx %= osts->op_count; + if (*stripe_cnt > 1 && (osts->op_count % (*stripe_cnt)) != 1) + ++lqr->lqr_offset_idx; } - down_read(&lov->lov_qos.lq_rw_sem); - ost_start_idx_temp = lov->lov_start_idx; + cfs_down_read(&lov->lov_qos.lq_rw_sem); + ost_start_idx_temp = lqr->lqr_start_idx; repeat_find: - array_idx = (lov->lov_start_idx + lov->lov_offset_idx) % ost_count; + array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) % osts->op_count; idx_pos = idx_arr; #ifdef QOS_DEBUG - CDEBUG(D_QOS, "want %d startidx %d startcnt %d offset %d active %d " - "count %d arrayidx %d\n", - stripe_cnt, lov->lov_start_idx, lov->lov_start_count, - lov->lov_offset_idx, ost_active_count, ost_count, array_idx); + CDEBUG(D_QOS, "pool '%s' want %d startidx %d startcnt %d offset %d " + "active %d count %d arrayidx %d\n", poolname, + *stripe_cnt, lqr->lqr_start_idx, lqr->lqr_start_count, + lqr->lqr_offset_idx, osts->op_count, osts->op_count, array_idx); #endif - for (i = 0; i < ost_count; i++, array_idx=(array_idx + 1) % ost_count) { - ++lov->lov_start_idx; - ost_idx = lov->lov_qos.lq_rr_array[array_idx]; + for (i = 0; i < osts->op_count; + i++, array_idx=(array_idx + 1) % osts->op_count) { + ++lqr->lqr_start_idx; + ost_idx = lqr->lqr_pool.op_array[array_idx]; #ifdef QOS_DEBUG CDEBUG(D_QOS, "#%d strt %d act %d strp %d ary %d idx %d\n", - i, lov->lov_start_idx, + i, lqr->lqr_start_idx, ((ost_idx != LOV_QOS_EMPTY) && lov->lov_tgts[ost_idx]) ? lov->lov_tgts[ost_idx]->ltd_active : 0, idx_pos - idx_arr, array_idx, ost_idx); @@ -593,29 +668,64 @@ repeat_find: if ((speed < 2) && (idx_pos - idx_arr < stripe_cnt_min)) { /* Try again, allowing slower OSCs */ speed++; - lov->lov_start_idx = ost_start_idx_temp; + lqr->lqr_start_idx = ost_start_idx_temp; goto repeat_find; } - up_read(&lov->lov_qos.lq_rw_sem); + cfs_up_read(&lov->lov_qos.lq_rw_sem); *stripe_cnt = idx_pos - idx_arr; - RETURN(0); +out: + if (pool != NULL) { + cfs_up_read(&pool_tgt_rw_sem(pool)); + /* put back ref got by lov_find_pool() */ + lov_pool_putref(pool); + } + + RETURN(rc); } /* alloc objects on osts with specific stripe offset */ static int alloc_specific(struct lov_obd *lov, struct lov_stripe_md *lsm, int *idx_arr) { - unsigned ost_idx, ost_count = lov->desc.ld_tgt_count; - int i, *idx_pos; + unsigned ost_idx, array_idx, ost_count; + int i, rc, *idx_pos; int speed = 0; + struct pool_desc *pool; + struct ost_pool *osts; ENTRY; + pool = lov_find_pool(lov, lsm->lsm_pool_name); + if (pool == NULL) { + osts = &(lov->lov_packed); + } else { + cfs_down_read(&pool_tgt_rw_sem(pool)); + osts = &(pool->pool_obds); + } + + ost_count = osts->op_count; + repeat_find: - ost_idx = lsm->lsm_oinfo[0]->loi_ost_idx; + /* search loi_ost_idx in ost array */ + array_idx = 0; + for (i = 0; i < ost_count; i++) { + if (osts->op_array[i] == lsm->lsm_oinfo[0]->loi_ost_idx) { + array_idx = i; + break; + } + } + if (i == ost_count) { + CERROR("Start index %d not found in pool '%s'\n", + lsm->lsm_oinfo[0]->loi_ost_idx, lsm->lsm_pool_name); + GOTO(out, rc = -EINVAL); + } + idx_pos = idx_arr; - for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) { + for (i = 0; i < ost_count; + i++, array_idx = (array_idx + 1) % ost_count) { + ost_idx = osts->op_array[array_idx]; + if (!lov->lov_tgts[ost_idx] || !lov->lov_tgts[ost_idx]->ltd_active) { continue; @@ -626,16 +736,20 @@ repeat_find: if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0) continue; - /* Drop slow OSCs if we can, but not for requested start idx */ + /* Drop slow OSCs if we can, but not for requested start idx. + * + * This means "if OSC is slow and it is not the requested + * start OST, then it can be skipped, otherwise skip it only + * if it is inactive/recovering/out-of-space." */ if ((obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed) && - (i != 0 || speed < 2)) + (i != 0 || speed >= 2)) continue; *idx_pos = ost_idx; idx_pos++; /* We have enough stripes */ if (idx_pos - idx_arr == lsm->lsm_stripe_count) - RETURN(0); + GOTO(out, rc = 0); } if (speed < 2) { /* Try again, allowing slower OSCs */ @@ -652,7 +766,15 @@ repeat_find: CERROR("can't lstripe objid "LPX64": have %d want %u\n", lsm->lsm_object_id, (int)(idx_pos - idx_arr), lsm->lsm_stripe_count); - RETURN(-EFBIG); + rc = -EFBIG; +out: + if (pool != NULL) { + cfs_up_read(&pool_tgt_rw_sem(pool)); + /* put back ref got by lov_find_pool() */ + lov_pool_putref(pool); + } + + RETURN(rc); } /* Alloc objects on osts with optimization based on: @@ -660,28 +782,43 @@ repeat_find: - network resources (shared OSS's) */ static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt, - int flags) + char *poolname, int flags) { struct lov_obd *lov = &exp->exp_obd->u.lov; - static time_t last_warn = 0; - time_t now = cfs_time_current_sec(); - __u64 total_bavail, total_weight = 0; - __u32 ost_count; - int nfound, good_osts, i, warn = 0, rc = 0; + __u64 total_weight = 0; + int nfound, good_osts, i, rc = 0; int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags); + struct pool_desc *pool; + struct ost_pool *osts; + struct lov_qos_rr *lqr; ENTRY; if (stripe_cnt_min < 1) - GOTO(out_nolock, rc = -EINVAL); + RETURN(-EINVAL); - lov_getref(exp->exp_obd); + pool = lov_find_pool(lov, poolname); + if (pool == NULL) { + osts = &(lov->lov_packed); + lqr = &(lov->lov_qos.lq_rr); + } else { + cfs_down_read(&pool_tgt_rw_sem(pool)); + osts = &(pool->pool_obds); + lqr = &(pool->pool_rr); + } + + obd_getref(exp->exp_obd); + + /* wait for fresh statfs info if needed, the rpcs are sent in + * lov_create() */ + qos_statfs_update(exp->exp_obd, + cfs_time_shift_64(-2 * lov->desc.ld_qos_maxage), 1); /* Detect -EAGAIN early, before expensive lock is taken. */ if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space) GOTO(out_nolock, rc = -EAGAIN); /* Do actual allocation, use write lock here. */ - down_write(&lov->lov_qos.lq_rw_sem); + cfs_down_write(&lov->lov_qos.lq_rw_sem); /* * Check again, while we were sleeping on @lq_rw_sem things could @@ -690,8 +827,6 @@ static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt, if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space) GOTO(out, rc = -EAGAIN); - ost_count = lov->desc.ld_tgt_count; - if (lov->desc.ld_active_tgt_count < 2) GOTO(out, rc = -EAGAIN); @@ -699,57 +834,35 @@ static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt, if (rc) GOTO(out, rc); - total_bavail = 0; good_osts = 0; - /* Warn users about zero available space/inode every 30 min */ - if (cfs_time_sub(now, last_warn) > 60 * 30) - warn = 1; /* Find all the OSTs that are valid stripe candidates */ - for (i = 0; i < ost_count; i++) { - __u64 bavail; - - if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active) + for (i = 0; i < osts->op_count; i++) { + if (!lov->lov_tgts[osts->op_array[i]] || + !lov->lov_tgts[osts->op_array[i]]->ltd_active) continue; - bavail = TGT_BAVAIL(i); - if (!bavail) { - if (warn) { - CDEBUG(D_QOS, "no free space on %s\n", - obd_uuid2str(&lov->lov_tgts[i]->ltd_uuid)); - last_warn = now; - } - continue; - } - if (!TGT_FFREE(i)) { - if (warn) { - CDEBUG(D_QOS, "no free inodes on %s\n", - obd_uuid2str(&lov->lov_tgts[i]->ltd_uuid)); - last_warn = now; - } - continue; - } /* Fail Check before osc_precreate() is called so we can only 'fail' single OSC. */ - if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && i == 0) + if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && osts->op_array[i] == 0) continue; - if (obd_precreate(lov->lov_tgts[i]->ltd_exp) > 2) + if (obd_precreate(lov->lov_tgts[osts->op_array[i]]->ltd_exp) > 2) continue; - lov->lov_tgts[i]->ltd_qos.ltq_usable = 1; - qos_calc_weight(lov, i); - total_bavail += bavail; - total_weight += lov->lov_tgts[i]->ltd_qos.ltq_weight; + lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable = 1; + qos_calc_weight(lov, osts->op_array[i]); + total_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight; good_osts++; } +#ifdef QOS_DEBUG + CDEBUG(D_QOS, "found %d good osts\n", good_osts); +#endif + if (good_osts < stripe_cnt_min) GOTO(out, rc = -EAGAIN); - if (!total_bavail) - GOTO(out, rc = -ENOSPC); - /* We have enough osts */ if (good_osts < *stripe_cnt) *stripe_cnt = good_osts; @@ -767,23 +880,23 @@ static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt, if (total_weight) { #if BITS_PER_LONG == 32 - rand = ll_rand() % (unsigned)total_weight; + 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)(ll_rand() % + rand = (__u64)(cfs_rand() % (unsigned)(total_weight >> 32)) << 32; else rand = 0; if (rand == (total_weight & 0xffffffff00000000ULL)) - rand |= ll_rand() % (unsigned)total_weight; + rand |= cfs_rand() % (unsigned)total_weight; else - rand |= ll_rand(); + rand |= cfs_rand(); #else - rand = ((__u64)ll_rand() << 32 | ll_rand()) % + rand = ((__u64)cfs_rand() << 32 | cfs_rand()) % total_weight; #endif } else { @@ -792,19 +905,24 @@ static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt, /* 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 < ost_count; i++) { - if (!lov->lov_tgts[i] || - !lov->lov_tgts[i]->ltd_qos.ltq_usable) + for (i = 0; i < osts->op_count; i++) { + if (!lov->lov_tgts[osts->op_array[i]] || + !lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable) continue; - cur_weight += lov->lov_tgts[i]->ltd_qos.ltq_weight; + cur_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight; +#ifdef QOS_DEBUG + CDEBUG(D_QOS, "stripe_cnt=%d nfound=%d cur_weight="LPU64 + " rand="LPU64" total_weight="LPU64"\n", + *stripe_cnt, nfound, cur_weight, rand, total_weight); +#endif if (cur_weight >= rand) { #ifdef QOS_DEBUG CDEBUG(D_QOS, "assigned stripe=%d to idx=%d\n", - nfound, i); + nfound, osts->op_array[i]); #endif - idx_arr[nfound++] = i; - qos_used(lov, i, &total_weight); + idx_arr[nfound++] = osts->op_array[i]; + qos_used(lov, osts, osts->op_array[i], &total_weight); rc = 0; break; } @@ -818,13 +936,19 @@ static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt, LASSERT(nfound == *stripe_cnt); out: - up_write(&lov->lov_qos.lq_rw_sem); + cfs_up_write(&lov->lov_qos.lq_rw_sem); out_nolock: + if (pool != NULL) { + cfs_up_read(&pool_tgt_rw_sem(pool)); + /* put back ref got by lov_find_pool() */ + lov_pool_putref(pool); + } + if (rc == -EAGAIN) - rc = alloc_rr(lov, idx_arr, stripe_cnt, flags); + rc = alloc_rr(lov, idx_arr, stripe_cnt, poolname, flags); - lov_putref(exp->exp_obd); + obd_putref(exp->exp_obd); RETURN(rc); } @@ -847,7 +971,8 @@ static int alloc_idx_array(struct obd_export *exp, struct lov_stripe_md *lsm, if (newea || lsm->lsm_oinfo[0]->loi_ost_idx >= lov->desc.ld_tgt_count) - rc = alloc_qos(exp, tmp_arr, &stripe_cnt, flags); + rc = alloc_qos(exp, tmp_arr, &stripe_cnt, + lsm->lsm_pool_name, flags); else rc = alloc_specific(lov, lsm, tmp_arr); @@ -881,7 +1006,7 @@ int qos_prep_create(struct obd_export *exp, struct lov_request_set *set) LASSERT(src_oa->o_valid & OBD_MD_FLID); LASSERT(src_oa->o_valid & OBD_MD_FLGROUP); - + if (set->set_oi->oi_md == NULL) { int stripes_def = lov_get_stripecnt(lov, 0); @@ -894,7 +1019,6 @@ int qos_prep_create(struct obd_export *exp, struct lov_request_set *set) /* Find a small number of stripes we can use (up to # of active osts). */ stripes = 1; - lov_getref(exp->exp_obd); for (i = 0; i < lov->desc.ld_tgt_count; i++) { if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active) @@ -904,7 +1028,6 @@ int qos_prep_create(struct obd_export *exp, struct lov_request_set *set) break; stripes++; } - lov_putref(exp->exp_obd); if (stripes < stripes_def) stripes = stripes_def; @@ -925,7 +1048,7 @@ int qos_prep_create(struct obd_export *exp, struct lov_request_set *set) lsm = set->set_oi->oi_md; lsm->lsm_object_id = src_oa->o_id; - lsm->lsm_object_gr = src_oa->o_gr; + lsm->lsm_object_seq = src_oa->o_seq; if (!lsm->lsm_stripe_size) lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size; @@ -951,7 +1074,7 @@ int qos_prep_create(struct obd_export *exp, struct lov_request_set *set) lov_set_add_req(req, set); req->rq_buflen = sizeof(*req->rq_oi.oi_md); - OBD_ALLOC(req->rq_oi.oi_md, req->rq_buflen); + OBD_ALLOC_LARGE(req->rq_oi.oi_md, req->rq_buflen); if (req->rq_oi.oi_md == NULL) GOTO(out_err, rc = -ENOMEM); @@ -963,6 +1086,7 @@ int qos_prep_create(struct obd_export *exp, struct lov_request_set *set) req->rq_stripe = i; /* create data objects with "parent" OA */ memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa)); + req->rq_oi.oi_cb_up = cb_create_update; /* XXX When we start creating objects on demand, we need to * make sure that we always create the object on the @@ -980,6 +1104,10 @@ int qos_prep_create(struct obd_export *exp, struct lov_request_set *set) if (stripes < lsm->lsm_stripe_count) qos_shrink_lsm(set); + if (OBD_FAIL_CHECK(OBD_FAIL_MDS_LOV_PREP_CREATE)) { + qos_shrink_lsm(set); + rc = -EIO; + } if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) { oti_alloc_cookies(oti, set->set_count); @@ -1001,3 +1129,110 @@ void qos_update(struct lov_obd *lov) ENTRY; lov->lov_qos.lq_dirty = 1; } + +void qos_statfs_done(struct lov_obd *lov) +{ + LASSERT(lov->lov_qos.lq_statfs_in_progress); + cfs_down_write(&lov->lov_qos.lq_rw_sem); + lov->lov_qos.lq_statfs_in_progress = 0; + /* wake up any threads waiting for the statfs rpcs to complete */ + cfs_waitq_signal(&lov->lov_qos.lq_statfs_waitq); + cfs_up_write(&lov->lov_qos.lq_rw_sem); +} + +static int qos_statfs_ready(struct obd_device *obd, __u64 max_age) +{ + struct lov_obd *lov = &obd->u.lov; + int rc; + ENTRY; + cfs_down_read(&lov->lov_qos.lq_rw_sem); + rc = lov->lov_qos.lq_statfs_in_progress == 0 || + cfs_time_beforeq_64(max_age, obd->obd_osfs_age); + cfs_up_read(&lov->lov_qos.lq_rw_sem); + RETURN(rc); +} + +/* + * Update statfs data if the current osfs age is older than max_age. + * If wait is not set, it means that we are called from lov_create() + * and we should just issue the rpcs without waiting for them to complete. + * If wait is set, we are called from alloc_qos() and we just have + * to wait for the request set to complete. + */ +void qos_statfs_update(struct obd_device *obd, __u64 max_age, int wait) +{ + struct lov_obd *lov = &obd->u.lov; + struct obd_info *oinfo; + int rc = 0; + struct ptlrpc_request_set *set = NULL; + ENTRY; + + if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age)) + /* statfs data are quite recent, don't need to refresh it */ + RETURN_EXIT; + + if (!wait && lov->lov_qos.lq_statfs_in_progress) + /* statfs already in progress */ + RETURN_EXIT; + + cfs_down_write(&lov->lov_qos.lq_rw_sem); + if (lov->lov_qos.lq_statfs_in_progress) { + cfs_up_write(&lov->lov_qos.lq_rw_sem); + GOTO(out, rc = 0); + } + /* no statfs in flight, send rpcs */ + lov->lov_qos.lq_statfs_in_progress = 1; + cfs_up_write(&lov->lov_qos.lq_rw_sem); + + if (wait) + CDEBUG(D_QOS, "%s: did not manage to get fresh statfs data " + "in a timely manner (osfs age "LPU64", max age "LPU64")" + ", sending new statfs rpcs\n", + obd_uuid2str(&lov->desc.ld_uuid), obd->obd_osfs_age, + max_age); + + /* need to send statfs rpcs */ + CDEBUG(D_QOS, "sending new statfs requests\n"); + memset(lov->lov_qos.lq_statfs_data, 0, + sizeof(*lov->lov_qos.lq_statfs_data)); + oinfo = &lov->lov_qos.lq_statfs_data->lsd_oi; + oinfo->oi_osfs = &lov->lov_qos.lq_statfs_data->lsd_statfs; + oinfo->oi_flags = OBD_STATFS_NODELAY; + set = ptlrpc_prep_set(); + if (!set) + GOTO(out_failed, rc = -ENOMEM); + + rc = obd_statfs_async(obd, oinfo, max_age, set); + if (rc || cfs_list_empty(&set->set_requests)) { + if (rc) + CWARN("statfs failed with %d\n", rc); + GOTO(out_failed, rc); + } + /* send requests via ptlrpcd */ + oinfo->oi_flags |= OBD_STATFS_PTLRPCD; + ptlrpcd_add_rqset(set); + GOTO(out, rc); + +out_failed: + cfs_down_write(&lov->lov_qos.lq_rw_sem); + lov->lov_qos.lq_statfs_in_progress = 0; + /* wake up any threads waiting for the statfs rpcs to complete */ + cfs_waitq_signal(&lov->lov_qos.lq_statfs_waitq); + cfs_up_write(&lov->lov_qos.lq_rw_sem); + wait = 0; +out: + if (set) + ptlrpc_set_destroy(set); + if (wait) { + struct l_wait_info lwi = { 0 }; + CDEBUG(D_QOS, "waiting for statfs requests to complete\n"); + l_wait_event(lov->lov_qos.lq_statfs_waitq, + qos_statfs_ready(obd, max_age), &lwi); + if (cfs_time_before_64(obd->obd_osfs_age, max_age)) + CDEBUG(D_QOS, "%s: still no fresh statfs data after " + "waiting (osfs age "LPU64", max age " + LPU64")\n", + obd_uuid2str(&lov->desc.ld_uuid), + obd->obd_osfs_age, max_age); + } +}