X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Flov%2Flov_qos.c;h=2262c503af6bc56e2d048422cb6321da5fba1698;hp=b292fc367868c8d4c5a74efbbab6d413a4d3f1b4;hb=eb7c28ff977f4e0a280558aa74e23f2a9ab0ea0c;hpb=dbaf6928620fc8ee9e6f6bf6413764571524be8a diff --git a/lustre/lov/lov_qos.c b/lustre/lov/lov_qos.c index b292fc3..2262c50 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++; @@ -99,13 +99,14 @@ 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_rr.lqr_dirty = 1; @@ -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_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; @@ -285,6 +300,10 @@ static int qos_used(struct lov_obd *lov, struct ost_pool *osts, 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 * @@ -293,7 +312,7 @@ static int qos_used(struct lov_obd *lov, struct ost_pool *osts, 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 @@ -354,7 +373,7 @@ static int qos_calc_rr(struct lov_obd *lov, struct ost_pool *src_pool, } /* 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 @@ -362,7 +381,7 @@ static int qos_calc_rr(struct lov_obd *lov, struct ost_pool *src_pool, */ if (!lqr->lqr_dirty) { LASSERT(lqr->lqr_pool.op_size); - up_write(&lov->lov_qos.lq_rw_sem); + cfs_up_write(&lov->lov_qos.lq_rw_sem); RETURN(0); } @@ -375,7 +394,7 @@ static int qos_calc_rr(struct lov_obd *lov, struct ost_pool *src_pool, lqr->lqr_pool.op_count = real_count; rc = lov_ost_pool_extend(&lqr->lqr_pool, real_count); if (rc) { - up_write(&lov->lov_qos.lq_rw_sem); + cfs_up_write(&lov->lov_qos.lq_rw_sem); RETURN(rc); } for (i = 0; i < lqr->lqr_pool.op_count; i++) @@ -383,7 +402,7 @@ static int qos_calc_rr(struct lov_obd *lov, struct ost_pool *src_pool, /* 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 < lqr->lqr_pool.op_count; i++) { if (lov->lov_tgts[src_pool->op_array[i]] && @@ -401,7 +420,7 @@ static int qos_calc_rr(struct lov_obd *lov, struct ost_pool *src_pool, } lqr->lqr_dirty = 0; - up_write(&lov->lov_qos.lq_rw_sem); + cfs_up_write(&lov->lov_qos.lq_rw_sem); if (placed != real_count) { /* This should never happen */ @@ -475,36 +494,80 @@ void qos_shrink_lsm(struct lov_request_set *set) } } +/** + * 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); } @@ -536,7 +599,7 @@ static int alloc_rr(struct lov_obd *lov, int *idx_arr, int *stripe_cnt, osts = &(lov->lov_packed); lqr = &(lov->lov_qos.lq_rr); } else { - read_lock(&pool_tgt_rwlock(pool)); + cfs_down_read(&pool_tgt_rw_sem(pool)); osts = &(pool->pool_obds); lqr = &(pool->pool_rr); } @@ -559,7 +622,7 @@ static int alloc_rr(struct lov_obd *lov, int *idx_arr, int *stripe_cnt, if (*stripe_cnt > 1 && (osts->op_count % (*stripe_cnt)) != 1) ++lqr->lqr_offset_idx; } - down_read(&lov->lov_qos.lq_rw_sem); + cfs_down_read(&lov->lov_qos.lq_rw_sem); ost_start_idx_temp = lqr->lqr_start_idx; repeat_find: @@ -609,12 +672,16 @@ repeat_find: 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; out: - if (pool != NULL) - read_unlock(&pool_tgt_rwlock(pool)); + 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); } @@ -633,7 +700,7 @@ static int alloc_specific(struct lov_obd *lov, struct lov_stripe_md *lsm, if (pool == NULL) { osts = &(lov->lov_packed); } else { - read_lock(&pool_tgt_rwlock(pool)); + cfs_down_read(&pool_tgt_rw_sem(pool)); osts = &(pool->pool_obds); } @@ -669,9 +736,13 @@ 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; @@ -697,8 +768,12 @@ repeat_find: lsm->lsm_stripe_count); rc = -EFBIG; out: - if (pool != NULL) - read_unlock(&pool_tgt_rwlock(pool)); + 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); } @@ -710,10 +785,8 @@ static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt, 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; - 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; @@ -728,19 +801,24 @@ static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt, osts = &(lov->lov_packed); lqr = &(lov->lov_qos.lq_rr); } else { - read_lock(&pool_tgt_rwlock(pool)); + cfs_down_read(&pool_tgt_rw_sem(pool)); osts = &(pool->pool_obds); lqr = &(pool->pool_rr); } - lov_getref(exp->exp_obd); + 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 @@ -756,35 +834,12 @@ 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 < osts->op_count; i++) { - __u64 bavail; - if (!lov->lov_tgts[osts->op_array[i]] || !lov->lov_tgts[osts->op_array[i]]->ltd_active) continue; - bavail = TGT_BAVAIL(osts->op_array[i]); - if (!bavail) { - if (warn) { - CDEBUG(D_QOS, "no free space on %s\n", - obd_uuid2str(&lov->lov_tgts[osts->op_array[i]]->ltd_uuid)); - last_warn = now; - } - continue; - } - if (!TGT_FFREE(osts->op_array[i])) { - if (warn) { - CDEBUG(D_QOS, "no free inodes on %s\n", - obd_uuid2str(&lov->lov_tgts[osts->op_array[i]]->ltd_uuid)); - last_warn = now; - } - continue; - } /* Fail Check before osc_precreate() is called so we can only 'fail' single OSC. */ @@ -796,7 +851,6 @@ static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt, lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable = 1; qos_calc_weight(lov, osts->op_array[i]); - total_bavail += bavail; total_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight; good_osts++; @@ -809,9 +863,6 @@ static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt, 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; @@ -885,16 +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) - read_unlock(&pool_tgt_rwlock(pool)); + 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, poolname, flags); - lov_putref(exp->exp_obd); + obd_putref(exp->exp_obd); RETURN(rc); } @@ -994,7 +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; @@ -1032,6 +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 @@ -1049,6 +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); @@ -1070,3 +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); + } +}