/* * GPL HEADER START * * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. * * This program is free software; you can redistribute it and/or modify * it under the terms of the GNU General Public License version 2 only, * as published by the Free Software Foundation. * * This program is distributed in the hope that it will be useful, but * WITHOUT ANY WARRANTY; without even the implied warranty of * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU * General Public License version 2 for more details (a copy is included * in the LICENSE file that accompanied this code). * * You should have received a copy of the GNU General Public License * version 2 along with this program; If not, see * http://www.gnu.org/licenses/gpl-2.0.html * * GPL HEADER END */ /* * This file is part of Lustre, http://www.lustre.org/ * * lustre/obdclass/lu_qos.c * * Lustre QoS. * These are the only exported functions, they provide some generic * infrastructure for object allocation QoS * */ #define DEBUG_SUBSYSTEM S_CLASS #include #include #include #include #include /* hash_long() */ #include #include #include #include #include #include void lu_qos_rr_init(struct lu_qos_rr *lqr) { spin_lock_init(&lqr->lqr_alloc); lqr->lqr_dirty = 1; } EXPORT_SYMBOL(lu_qos_rr_init); /** * Add a new target to Quality of Service (QoS) target table. * * Add a new MDT/OST target to the structure representing an OSS. Resort the * list of known MDSs/OSSs by the number of MDTs/OSTs attached to each MDS/OSS. * The MDS/OSS list is protected internally and no external locking is required. * * \param[in] qos lu_qos data * \param[in] ltd target description * * \retval 0 on success * \retval -ENOMEM on error */ int lqos_add_tgt(struct lu_qos *qos, struct lu_tgt_desc *ltd) { struct lu_svr_qos *svr = NULL; struct lu_svr_qos *tempsvr; struct obd_export *exp = ltd->ltd_exp; int found = 0; __u32 id = 0; int rc = 0; ENTRY; down_write(&qos->lq_rw_sem); /* * a bit hacky approach to learn NID of corresponding connection * but there is no official API to access information like this * with OSD API. */ list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) { if (obd_uuid_equals(&svr->lsq_uuid, &exp->exp_connection->c_remote_uuid)) { found++; break; } if (svr->lsq_id > id) id = svr->lsq_id; } if (!found) { OBD_ALLOC_PTR(svr); if (!svr) GOTO(out, rc = -ENOMEM); memcpy(&svr->lsq_uuid, &exp->exp_connection->c_remote_uuid, sizeof(svr->lsq_uuid)); ++id; svr->lsq_id = id; } else { /* Assume we have to move this one */ list_del(&svr->lsq_svr_list); } svr->lsq_tgt_count++; ltd->ltd_qos.ltq_svr = svr; CDEBUG(D_OTHER, "add tgt %s to server %s (%d targets)\n", obd_uuid2str(<d->ltd_uuid), obd_uuid2str(&svr->lsq_uuid), svr->lsq_tgt_count); /* * Add sorted by # of tgts. Find the first entry that we're * bigger than... */ list_for_each_entry(tempsvr, &qos->lq_svr_list, lsq_svr_list) { if (svr->lsq_tgt_count > tempsvr->lsq_tgt_count) break; } /* * ...and add before it. If we're the first or smallest, tempsvr * points to the list head, and we add to the end. */ list_add_tail(&svr->lsq_svr_list, &tempsvr->lsq_svr_list); qos->lq_dirty = 1; qos->lq_rr.lqr_dirty = 1; out: up_write(&qos->lq_rw_sem); RETURN(rc); } EXPORT_SYMBOL(lqos_add_tgt); /** * Remove MDT/OST target from QoS table. * * Removes given MDT/OST target from QoS table and releases related * MDS/OSS structure if no target remain on the MDS/OSS. * * \param[in] qos lu_qos data * \param[in] ltd target description * * \retval 0 on success * \retval -ENOENT if no server was found */ int lqos_del_tgt(struct lu_qos *qos, struct lu_tgt_desc *ltd) { struct lu_svr_qos *svr; int rc = 0; ENTRY; down_write(&qos->lq_rw_sem); svr = ltd->ltd_qos.ltq_svr; if (!svr) GOTO(out, rc = -ENOENT); svr->lsq_tgt_count--; if (svr->lsq_tgt_count == 0) { CDEBUG(D_OTHER, "removing server %s\n", obd_uuid2str(&svr->lsq_uuid)); list_del(&svr->lsq_svr_list); ltd->ltd_qos.ltq_svr = NULL; OBD_FREE_PTR(svr); } qos->lq_dirty = 1; qos->lq_rr.lqr_dirty = 1; out: up_write(&qos->lq_rw_sem); RETURN(rc); } EXPORT_SYMBOL(lqos_del_tgt); /** * lu_prandom_u64_max - returns a pseudo-random u64 number in interval * [0, ep_ro) * * \param[in] ep_ro right open interval endpoint * * \retval a pseudo-random 64-bit number that is in interval [0, ep_ro). */ u64 lu_prandom_u64_max(u64 ep_ro) { u64 rand = 0; if (ep_ro) { #if BITS_PER_LONG == 32 /* * If ep_ro > 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 (ep_ro > 0xffffffffULL) rand = prandom_u32_max((u32)(ep_ro >> 32)) << 32; if (rand == (ep_ro & 0xffffffff00000000ULL)) rand |= prandom_u32_max((u32)ep_ro); else rand |= prandom_u32(); #else rand = ((u64)prandom_u32() << 32 | prandom_u32()) % ep_ro; #endif } return rand; } EXPORT_SYMBOL(lu_prandom_u64_max); static inline __u64 tgt_statfs_bavail(struct lu_tgt_desc *tgt) { struct obd_statfs *statfs = &tgt->ltd_statfs; return statfs->os_bavail * statfs->os_bsize; } static inline __u64 tgt_statfs_iavail(struct lu_tgt_desc *tgt) { return tgt->ltd_statfs.os_ffree; } /** * Calculate penalties per-tgt and per-server * * Re-calculate penalties when the configuration changes, active targets * change and after statfs refresh (all these are reflected by lq_dirty flag). * On every tgt and server: 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 server/tgt under light load. * See lqos_calc_weight() for how penalties are factored into the weight. * * \param[in] qos lu_qos * \param[in] ltd lu_tgt_descs * \param[in] active_tgt_nr active tgt number * \param[in] maxage qos max age * \param[in] is_mdt MDT will count inode usage * * \retval 0 on success * \retval -EAGAIN the number of tgt isn't enough or all tgt spaces are * almost the same */ int lqos_calc_penalties(struct lu_qos *qos, struct lu_tgt_descs *ltd, __u32 active_tgt_nr, __u32 maxage, bool is_mdt) { struct lu_tgt_desc *tgt; struct lu_svr_qos *svr; __u64 ba_max, ba_min, ba; __u64 ia_max, ia_min, ia = 1; __u32 num_active; int prio_wide; time64_t now, age; int rc; ENTRY; if (!qos->lq_dirty) GOTO(out, rc = 0); num_active = active_tgt_nr - 1; if (num_active < 1) GOTO(out, rc = -EAGAIN); /* find bavail on each server */ list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) { svr->lsq_bavail = 0; /* if inode is not counted, set to 1 to ignore */ svr->lsq_iavail = is_mdt ? 0 : 1; } qos->lq_active_svr_count = 0; /* * How badly user wants to select targets "widely" (not recently chosen * and not on recent MDS's). As opposed to "freely" (free space avail.) * 0-256 */ prio_wide = 256 - qos->lq_prio_free; ba_min = (__u64)(-1); ba_max = 0; ia_min = (__u64)(-1); ia_max = 0; now = ktime_get_real_seconds(); /* Calculate server penalty per object */ ltd_foreach_tgt(ltd, tgt) { if (!tgt->ltd_active) continue; /* when inode is counted, bavail >> 16 to avoid overflow */ ba = tgt_statfs_bavail(tgt); if (is_mdt) ba >>= 16; else ba >>= 8; if (!ba) continue; ba_min = min(ba, ba_min); ba_max = max(ba, ba_max); /* Count the number of usable servers */ if (tgt->ltd_qos.ltq_svr->lsq_bavail == 0) qos->lq_active_svr_count++; tgt->ltd_qos.ltq_svr->lsq_bavail += ba; if (is_mdt) { /* iavail >> 8 to avoid overflow */ ia = tgt_statfs_iavail(tgt) >> 8; if (!ia) continue; ia_min = min(ia, ia_min); ia_max = max(ia, ia_max); tgt->ltd_qos.ltq_svr->lsq_iavail += ia; } /* * per-tgt penalty is * prio * bavail * iavail / (num_tgt - 1) / 2 */ tgt->ltd_qos.ltq_penalty_per_obj = prio_wide * ba * ia >> 8; do_div(tgt->ltd_qos.ltq_penalty_per_obj, num_active); tgt->ltd_qos.ltq_penalty_per_obj >>= 1; age = (now - tgt->ltd_qos.ltq_used) >> 3; if (qos->lq_reset || age > 32 * maxage) tgt->ltd_qos.ltq_penalty = 0; else if (age > maxage) /* Decay tgt penalty. */ tgt->ltd_qos.ltq_penalty >>= (age / maxage); } num_active = qos->lq_active_svr_count - 1; if (num_active < 1) { /* * If there's only 1 server, we can't penalize it, so instead * we have to double the tgt penalty */ num_active = 1; ltd_foreach_tgt(ltd, tgt) { if (!tgt->ltd_active) continue; tgt->ltd_qos.ltq_penalty_per_obj <<= 1; } } /* * Per-server penalty is * prio * bavail * iavail / server_tgts / (num_svr - 1) / 2 */ list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) { ba = svr->lsq_bavail; ia = svr->lsq_iavail; svr->lsq_penalty_per_obj = prio_wide * ba * ia >> 8; do_div(ba, svr->lsq_tgt_count * num_active); svr->lsq_penalty_per_obj >>= 1; age = (now - svr->lsq_used) >> 3; if (qos->lq_reset || age > 32 * maxage) svr->lsq_penalty = 0; else if (age > maxage) /* Decay server penalty. */ svr->lsq_penalty >>= age / maxage; } qos->lq_dirty = 0; qos->lq_reset = 0; /* * If each tgt has almost same free space, do rr allocation for better * creation performance */ qos->lq_same_space = 0; if ((ba_max * (256 - qos->lq_threshold_rr)) >> 8 < ba_min && (ia_max * (256 - qos->lq_threshold_rr)) >> 8 < ia_min) { qos->lq_same_space = 1; /* Reset weights for the next time we enter qos mode */ qos->lq_reset = 1; } rc = 0; out: if (!rc && qos->lq_same_space) RETURN(-EAGAIN); RETURN(rc); } EXPORT_SYMBOL(lqos_calc_penalties); bool lqos_is_usable(struct lu_qos *qos, __u32 active_tgt_nr) { if (!qos->lq_dirty && qos->lq_same_space) return false; if (active_tgt_nr < 2) return false; return true; } EXPORT_SYMBOL(lqos_is_usable); /** * Calculate weight for a given tgt. * * The final tgt weight is bavail >> 16 * iavail >> 8 minus the tgt and server * penalties. See lqos_calc_ppts() for how penalties are calculated. * * \param[in] tgt target descriptor */ void lqos_calc_weight(struct lu_tgt_desc *tgt) { struct lu_tgt_qos *ltq = &tgt->ltd_qos; __u64 temp, temp2; temp = (tgt_statfs_bavail(tgt) >> 16) * (tgt_statfs_iavail(tgt) >> 8); temp2 = ltq->ltq_penalty + ltq->ltq_svr->lsq_penalty; if (temp < temp2) ltq->ltq_weight = 0; else ltq->ltq_weight = temp - temp2; } EXPORT_SYMBOL(lqos_calc_weight); /** * Re-calculate weights. * * The function is called when some target was used for a new object. In * this case we should re-calculate all the weights to keep new allocations * balanced well. * * \param[in] qos lu_qos * \param[in] ltd lu_tgt_descs * \param[in] tgt target where a new object was placed * \param[in] active_tgt_nr active tgt number * \param[out] total_wt new total weight for the pool * * \retval 0 */ int lqos_recalc_weight(struct lu_qos *qos, struct lu_tgt_descs *ltd, struct lu_tgt_desc *tgt, __u32 active_tgt_nr, __u64 *total_wt) { struct lu_tgt_qos *ltq; struct lu_svr_qos *svr; ENTRY; ltq = &tgt->ltd_qos; LASSERT(ltq); /* Don't allocate on this device anymore, until the next alloc_qos */ ltq->ltq_usable = 0; svr = ltq->ltq_svr; /* * Decay old penalty by half (we're adding max penalty, and don't * want it to run away.) */ ltq->ltq_penalty >>= 1; svr->lsq_penalty >>= 1; /* mark the server and tgt as recently used */ ltq->ltq_used = svr->lsq_used = ktime_get_real_seconds(); /* Set max penalties for this tgt and server */ ltq->ltq_penalty += ltq->ltq_penalty_per_obj * active_tgt_nr; svr->lsq_penalty += svr->lsq_penalty_per_obj * active_tgt_nr; /* Decrease all MDS penalties */ list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) { if (svr->lsq_penalty < svr->lsq_penalty_per_obj) svr->lsq_penalty = 0; else svr->lsq_penalty -= svr->lsq_penalty_per_obj; } *total_wt = 0; /* Decrease all tgt penalties */ ltd_foreach_tgt(ltd, tgt) { if (!tgt->ltd_active) continue; if (ltq->ltq_penalty < ltq->ltq_penalty_per_obj) ltq->ltq_penalty = 0; else ltq->ltq_penalty -= ltq->ltq_penalty_per_obj; lqos_calc_weight(tgt); /* Recalc the total weight of usable osts */ if (ltq->ltq_usable) *total_wt += ltq->ltq_weight; CDEBUG(D_OTHER, "recalc tgt %d usable=%d avail=%llu" " tgtppo=%llu tgtp=%llu svrppo=%llu" " svrp=%llu wt=%llu\n", tgt->ltd_index, ltq->ltq_usable, tgt_statfs_bavail(tgt) >> 10, ltq->ltq_penalty_per_obj >> 10, ltq->ltq_penalty >> 10, ltq->ltq_svr->lsq_penalty_per_obj >> 10, ltq->ltq_svr->lsq_penalty >> 10, ltq->ltq_weight >> 10); } RETURN(0); } EXPORT_SYMBOL(lqos_recalc_weight);