X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Flov%2Flov_qos.c;h=b5ba7224e6b9c7f74dc9ef9e509439e442868bb8;hb=7728c26736f4f49400b3143620e15f5ecac9c588;hp=b292fc367868c8d4c5a74efbbab6d413a4d3f1b4;hpb=dbaf6928620fc8ee9e6f6bf6413764571524be8a;p=fs%2Flustre-release.git diff --git a/lustre/lov/lov_qos.c b/lustre/lov/lov_qos.c index b292fc3..b5ba722 100644 --- a/lustre/lov/lov_qos.c +++ b/lustre/lov/lov_qos.c @@ -47,6 +47,7 @@ #include #include +#include #include "lov_internal.h" /* #define QOS_DEBUG 1 */ @@ -121,19 +122,16 @@ out: 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); - oss = lov->lov_tgts[index]->ltd_qos.ltq_oss; + oss = tgt->ltd_qos.ltq_oss; if (!oss) GOTO(out, rc = -ENOENT); @@ -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) @@ -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; @@ -226,8 +235,17 @@ static int qos_calc_ppo(struct obd_device *obd) 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 * @@ -490,12 +509,13 @@ int qos_remedy_create(struct lov_request_set *set, struct lov_request *req) continue; /* check if objects has been created on this ost */ for (stripe = 0; stripe < lsm->lsm_stripe_count; stripe++) { + /* we try send create to this ost but he is failed */ if (stripe == req->rq_stripe) continue; + /* 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, @@ -536,7 +556,7 @@ 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)); + down_read(&pool_tgt_rw_sem(pool)); osts = &(pool->pool_obds); lqr = &(pool->pool_rr); } @@ -613,8 +633,12 @@ repeat_find: *stripe_cnt = idx_pos - idx_arr; out: - if (pool != NULL) - read_unlock(&pool_tgt_rwlock(pool)); + if (pool != NULL) { + up_read(&pool_tgt_rw_sem(pool)); + /* put back ref got by lov_find_pool() */ + lov_pool_putref(pool); + } + RETURN(rc); } @@ -633,7 +657,7 @@ 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)); + down_read(&pool_tgt_rw_sem(pool)); osts = &(pool->pool_obds); } @@ -669,9 +693,13 @@ 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 +725,12 @@ repeat_find: lsm->lsm_stripe_count); rc = -EFBIG; out: - if (pool != NULL) - read_unlock(&pool_tgt_rwlock(pool)); + if (pool != NULL) { + up_read(&pool_tgt_rw_sem(pool)); + /* put back ref got by lov_find_pool() */ + lov_pool_putref(pool); + } + RETURN(rc); } @@ -728,12 +760,17 @@ 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)); + 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) @@ -888,13 +925,16 @@ out: up_write(&lov->lov_qos.lq_rw_sem); out_nolock: - if (pool != NULL) - read_unlock(&pool_tgt_rwlock(pool)); + if (pool != NULL) { + 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); } @@ -1032,6 +1072,7 @@ 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 +1090,10 @@ 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 +1115,110 @@ 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); + 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); + 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; + 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); + 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; + + down_write(&lov->lov_qos.lq_rw_sem); + if (lov->lov_qos.lq_statfs_in_progress) { + 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; + 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 || 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: + 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); + 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); + } +}