1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
32 * Copyright (c) 2011, 2012, Whamcloud, Inc.
35 * This file is part of Lustre, http://www.lustre.org/
36 * Lustre is a trademark of Sun Microsystems, Inc.
40 # define EXPORT_SYMTAB
42 #define DEBUG_SUBSYSTEM S_LOV
45 #include <libcfs/libcfs.h>
47 #include <liblustre.h>
50 #include <obd_class.h>
52 #include <lustre/lustre_idl.h>
53 #include "lov_internal.h"
55 /* #define QOS_DEBUG 1 */
58 #define TGT_BAVAIL(i) (lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bavail *\
59 lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bsize)
62 int qos_add_tgt(struct obd_device *obd, __u32 index)
64 struct lov_obd *lov = &obd->u.lov;
65 struct lov_qos_oss *oss, *temposs;
66 struct obd_export *exp = lov->lov_tgts[index]->ltd_exp;
67 int rc = 0, found = 0;
70 /* We only need this QOS struct on MDT, not clients - but we may not
71 * have registered the LOV's observer yet, so there's no way to know */
72 if (!exp || !exp->exp_connection) {
73 CERROR("Missing connection\n");
77 cfs_down_write(&lov->lov_qos.lq_rw_sem);
78 cfs_mutex_lock(&lov->lov_lock);
79 cfs_list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
80 if (obd_uuid_equals(&oss->lqo_uuid,
81 &exp->exp_connection->c_remote_uuid)) {
90 GOTO(out, rc = -ENOMEM);
91 memcpy(&oss->lqo_uuid,
92 &exp->exp_connection->c_remote_uuid,
93 sizeof(oss->lqo_uuid));
95 /* Assume we have to move this one */
96 cfs_list_del(&oss->lqo_oss_list);
100 lov->lov_tgts[index]->ltd_qos.ltq_oss = oss;
102 /* Add sorted by # of OSTs. Find the first entry that we're
104 cfs_list_for_each_entry(temposs, &lov->lov_qos.lq_oss_list,
106 if (oss->lqo_ost_count > temposs->lqo_ost_count)
109 /* ...and add before it. If we're the first or smallest, temposs
110 points to the list head, and we add to the end. */
111 cfs_list_add_tail(&oss->lqo_oss_list, &temposs->lqo_oss_list);
113 lov->lov_qos.lq_dirty = 1;
114 lov->lov_qos.lq_rr.lqr_dirty = 1;
116 CDEBUG(D_QOS, "add tgt %s to OSS %s (%d OSTs)\n",
117 obd_uuid2str(&lov->lov_tgts[index]->ltd_uuid),
118 obd_uuid2str(&oss->lqo_uuid),
122 cfs_mutex_unlock(&lov->lov_lock);
123 cfs_up_write(&lov->lov_qos.lq_rw_sem);
127 int qos_del_tgt(struct obd_device *obd, struct lov_tgt_desc *tgt)
129 struct lov_obd *lov = &obd->u.lov;
130 struct lov_qos_oss *oss;
134 cfs_down_write(&lov->lov_qos.lq_rw_sem);
136 oss = tgt->ltd_qos.ltq_oss;
138 GOTO(out, rc = -ENOENT);
140 oss->lqo_ost_count--;
141 if (oss->lqo_ost_count == 0) {
142 CDEBUG(D_QOS, "removing OSS %s\n",
143 obd_uuid2str(&oss->lqo_uuid));
144 cfs_list_del(&oss->lqo_oss_list);
148 lov->lov_qos.lq_dirty = 1;
149 lov->lov_qos.lq_rr.lqr_dirty = 1;
151 cfs_up_write(&lov->lov_qos.lq_rw_sem);
155 /* Recalculate per-object penalties for OSSs and OSTs,
156 depends on size of each ost in an oss */
157 static int qos_calc_ppo(struct obd_device *obd)
159 struct lov_obd *lov = &obd->u.lov;
160 struct lov_qos_oss *oss;
161 __u64 ba_max, ba_min, temp;
163 int rc, i, prio_wide;
167 if (!lov->lov_qos.lq_dirty)
170 num_active = lov->desc.ld_active_tgt_count - 1;
172 GOTO(out, rc = -EAGAIN);
174 /* find bavail on each OSS */
175 cfs_list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
178 lov->lov_qos.lq_active_oss_count = 0;
180 /* How badly user wants to select osts "widely" (not recently chosen
181 and not on recent oss's). As opposed to "freely" (free space
183 prio_wide = 256 - lov->lov_qos.lq_prio_free;
185 ba_min = (__u64)(-1);
187 now = cfs_time_current_sec();
188 /* Calculate OST penalty per object */
189 /* (lov ref taken in alloc_qos) */
190 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
191 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
193 temp = TGT_BAVAIL(i);
196 ba_min = min(temp, ba_min);
197 ba_max = max(temp, ba_max);
199 /* Count the number of usable OSS's */
200 if (lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_bavail == 0)
201 lov->lov_qos.lq_active_oss_count++;
202 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_bavail += temp;
204 /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
206 do_div(temp, num_active);
207 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj =
208 (temp * prio_wide) >> 8;
210 age = (now - lov->lov_tgts[i]->ltd_qos.ltq_used) >> 3;
211 if (lov->lov_qos.lq_reset || age > 32 * lov->desc.ld_qos_maxage)
212 lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
213 else if (age > lov->desc.ld_qos_maxage)
214 /* Decay the penalty by half for every 8x the update
215 * interval that the device has been idle. That gives
216 * lots of time for the statfs information to be
217 * updated (which the penalty is only a proxy for),
218 * and avoids penalizing OSS/OSTs under light load. */
219 lov->lov_tgts[i]->ltd_qos.ltq_penalty >>=
220 (age / lov->desc.ld_qos_maxage);
223 num_active = lov->lov_qos.lq_active_oss_count - 1;
224 if (num_active < 1) {
225 /* If there's only 1 OSS, we can't penalize it, so instead
226 we have to double the OST penalty */
228 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
229 if (lov->lov_tgts[i] == NULL)
231 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj <<= 1;
235 /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
236 cfs_list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
237 temp = oss->lqo_bavail >> 1;
238 do_div(temp, oss->lqo_ost_count * num_active);
239 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
241 age = (now - oss->lqo_used) >> 3;
242 if (lov->lov_qos.lq_reset || age > 32 * lov->desc.ld_qos_maxage)
243 oss->lqo_penalty = 0;
244 else if (age > lov->desc.ld_qos_maxage)
245 /* Decay the penalty by half for every 8x the update
246 * interval that the device has been idle. That gives
247 * lots of time for the statfs information to be
248 * updated (which the penalty is only a proxy for),
249 * and avoids penalizing OSS/OSTs under light load. */
250 oss->lqo_penalty >>= (age / lov->desc.ld_qos_maxage);
253 lov->lov_qos.lq_dirty = 0;
254 lov->lov_qos.lq_reset = 0;
256 /* If each ost has almost same free space,
257 * do rr allocation for better creation performance */
258 lov->lov_qos.lq_same_space = 0;
259 if ((ba_max * (256 - lov->lov_qos.lq_threshold_rr)) >> 8 < ba_min) {
260 lov->lov_qos.lq_same_space = 1;
261 /* Reset weights for the next time we enter qos mode */
262 lov->lov_qos.lq_reset = 1;
267 if (!rc && lov->lov_qos.lq_same_space)
272 static int qos_calc_weight(struct lov_obd *lov, int i)
276 /* Final ost weight = TGT_BAVAIL - ost_penalty - oss_penalty */
277 temp = TGT_BAVAIL(i);
278 temp2 = lov->lov_tgts[i]->ltd_qos.ltq_penalty +
279 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty;
281 lov->lov_tgts[i]->ltd_qos.ltq_weight = 0;
283 lov->lov_tgts[i]->ltd_qos.ltq_weight = temp - temp2;
287 /* We just used this index for a stripe; adjust everyone's weights */
288 static int qos_used(struct lov_obd *lov, struct ost_pool *osts,
289 __u32 index, __u64 *total_wt)
291 struct lov_qos_oss *oss;
295 /* Don't allocate from this stripe anymore, until the next alloc_qos */
296 lov->lov_tgts[index]->ltd_qos.ltq_usable = 0;
298 oss = lov->lov_tgts[index]->ltd_qos.ltq_oss;
300 /* Decay old penalty by half (we're adding max penalty, and don't
301 want it to run away.) */
302 lov->lov_tgts[index]->ltd_qos.ltq_penalty >>= 1;
303 oss->lqo_penalty >>= 1;
305 /* mark the OSS and OST as recently used */
306 lov->lov_tgts[index]->ltd_qos.ltq_used =
307 oss->lqo_used = cfs_time_current_sec();
309 /* Set max penalties for this OST and OSS */
310 lov->lov_tgts[index]->ltd_qos.ltq_penalty +=
311 lov->lov_tgts[index]->ltd_qos.ltq_penalty_per_obj *
312 lov->desc.ld_active_tgt_count;
313 oss->lqo_penalty += oss->lqo_penalty_per_obj *
314 lov->lov_qos.lq_active_oss_count;
316 /* Decrease all OSS penalties */
317 cfs_list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
318 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
319 oss->lqo_penalty = 0;
321 oss->lqo_penalty -= oss->lqo_penalty_per_obj;
325 /* Decrease all OST penalties */
326 for (j = 0; j < osts->op_count; j++) {
329 i = osts->op_array[j];
330 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
332 if (lov->lov_tgts[i]->ltd_qos.ltq_penalty <
333 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj)
334 lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
336 lov->lov_tgts[i]->ltd_qos.ltq_penalty -=
337 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj;
339 qos_calc_weight(lov, i);
341 /* Recalc the total weight of usable osts */
342 if (lov->lov_tgts[i]->ltd_qos.ltq_usable)
343 *total_wt += lov->lov_tgts[i]->ltd_qos.ltq_weight;
346 CDEBUG(D_QOS, "recalc tgt %d usable=%d avail="LPU64
347 " ostppo="LPU64" ostp="LPU64" ossppo="LPU64
348 " ossp="LPU64" wt="LPU64"\n",
349 i, lov->lov_tgts[i]->ltd_qos.ltq_usable,
351 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj >> 10,
352 lov->lov_tgts[i]->ltd_qos.ltq_penalty >> 10,
353 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty_per_obj>>10,
354 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty >> 10,
355 lov->lov_tgts[i]->ltd_qos.ltq_weight >> 10);
362 #define LOV_QOS_EMPTY ((__u32)-1)
363 /* compute optimal round-robin order, based on OSTs per OSS */
364 static int qos_calc_rr(struct lov_obd *lov, struct ost_pool *src_pool,
365 struct lov_qos_rr *lqr)
367 struct lov_qos_oss *oss;
368 unsigned placed, real_count;
372 if (!lqr->lqr_dirty) {
373 LASSERT(lqr->lqr_pool.op_size);
377 /* Do actual allocation. */
378 cfs_down_write(&lov->lov_qos.lq_rw_sem);
381 * Check again. While we were sleeping on @lq_rw_sem something could
384 if (!lqr->lqr_dirty) {
385 LASSERT(lqr->lqr_pool.op_size);
386 cfs_up_write(&lov->lov_qos.lq_rw_sem);
390 real_count = src_pool->op_count;
392 /* Zero the pool array */
393 /* alloc_rr is holding a read lock on the pool, so nobody is adding/
394 deleting from the pool. The lq_rw_sem insures that nobody else
396 lqr->lqr_pool.op_count = real_count;
397 rc = lov_ost_pool_extend(&lqr->lqr_pool, real_count);
399 cfs_up_write(&lov->lov_qos.lq_rw_sem);
402 for (i = 0; i < lqr->lqr_pool.op_count; i++)
403 lqr->lqr_pool.op_array[i] = LOV_QOS_EMPTY;
405 /* Place all the OSTs from 1 OSS at the same time. */
407 cfs_list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
409 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
410 if (lov->lov_tgts[src_pool->op_array[i]] &&
411 (lov->lov_tgts[src_pool->op_array[i]]->ltd_qos.ltq_oss == oss)) {
412 /* Evenly space these OSTs across arrayspace */
413 int next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count;
414 while (lqr->lqr_pool.op_array[next] !=
416 next = (next + 1) % lqr->lqr_pool.op_count;
417 lqr->lqr_pool.op_array[next] = src_pool->op_array[i];
425 cfs_up_write(&lov->lov_qos.lq_rw_sem);
427 if (placed != real_count) {
428 /* This should never happen */
429 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
430 "round-robin list (%d of %d).\n",
432 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
433 LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
434 lqr->lqr_pool.op_array[i]);
441 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
442 LCONSOLE(D_QOS, "rr #%d ost idx=%d\n", i,
443 lqr->lqr_pool.op_array[i]);
451 void qos_shrink_lsm(struct lov_request_set *set)
453 struct lov_stripe_md *lsm = set->set_oi->oi_md, *lsm_new;
454 /* XXX LOV STACKING call into osc for sizes */
455 unsigned oldsize, newsize;
457 if (set->set_oti && set->set_cookies && set->set_cookie_sent) {
458 struct llog_cookie *cookies;
459 oldsize = lsm->lsm_stripe_count * sizeof(*cookies);
460 newsize = set->set_count * sizeof(*cookies);
462 cookies = set->set_cookies;
463 oti_alloc_cookies(set->set_oti, set->set_count);
464 if (set->set_oti->oti_logcookies) {
465 memcpy(set->set_oti->oti_logcookies, cookies, newsize);
466 OBD_FREE_LARGE(cookies, oldsize);
467 set->set_cookies = set->set_oti->oti_logcookies;
469 CWARN("'leaking' %d bytes\n", oldsize - newsize);
473 CWARN("using fewer stripes for object "LPU64": old %u new %u\n",
474 lsm->lsm_object_id, lsm->lsm_stripe_count, set->set_count);
475 LASSERT(lsm->lsm_stripe_count >= set->set_count);
477 newsize = lov_stripe_md_size(set->set_count);
478 OBD_ALLOC_LARGE(lsm_new, newsize);
479 if (lsm_new != NULL) {
481 memcpy(lsm_new, lsm, sizeof(*lsm));
482 for (i = 0; i < lsm->lsm_stripe_count; i++) {
483 if (i < set->set_count) {
484 lsm_new->lsm_oinfo[i] = lsm->lsm_oinfo[i];
487 OBD_SLAB_FREE(lsm->lsm_oinfo[i], lov_oinfo_slab,
488 sizeof(struct lov_oinfo));
490 lsm_new->lsm_stripe_count = set->set_count;
491 OBD_FREE_LARGE(lsm, sizeof(struct lov_stripe_md) +
492 lsm->lsm_stripe_count*sizeof(struct lov_oinfo*));
493 set->set_oi->oi_md = lsm_new;
495 CWARN("'leaking' few bytes\n");
500 * Check whether we can create the object on the OST(refered by ost_idx)
502 * 0: create the object.
503 * other value: did not create the object.
505 static int lov_check_and_create_object(struct lov_obd *lov, int ost_idx,
506 struct lov_stripe_md *lsm,
507 struct lov_request *req,
508 struct obd_trans_info *oti)
514 CDEBUG(D_QOS, "Check and create on idx %d \n", ost_idx);
515 if (!lov->lov_tgts[ost_idx] ||
516 !lov->lov_tgts[ost_idx]->ltd_active)
519 /* check if objects has been created on this ost */
520 for (stripe = 0; stripe < lsm->lsm_stripe_count; stripe++) {
521 /* already have object at this stripe */
522 if (ost_idx == lsm->lsm_oinfo[stripe]->loi_ost_idx)
526 if (stripe >= lsm->lsm_stripe_count) {
527 req->rq_idx = ost_idx;
528 rc = obd_create(lov->lov_tgts[ost_idx]->ltd_exp,
529 req->rq_oi.oi_oa, &req->rq_oi.oi_md,
535 int qos_remedy_create(struct lov_request_set *set, struct lov_request *req)
537 struct lov_stripe_md *lsm = set->set_oi->oi_md;
538 struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
539 unsigned ost_idx = 0, ost_count;
540 struct pool_desc *pool;
541 struct ost_pool *osts = NULL;
545 /* First check whether we can create the objects on the pool */
546 pool = lov_find_pool(lov, lsm->lsm_pool_name);
548 cfs_down_read(&pool_tgt_rw_sem(pool));
549 osts = &(pool->pool_obds);
550 ost_count = osts->op_count;
551 for (i = 0; i < ost_count; i++, ost_idx = osts->op_array[i]) {
552 rc = lov_check_and_create_object(lov, ost_idx, lsm, req,
557 cfs_up_read(&pool_tgt_rw_sem(pool));
558 lov_pool_putref(pool);
562 ost_count = lov->desc.ld_tgt_count;
563 /* Then check whether we can create the objects on other OSTs */
564 ost_idx = (req->rq_idx + lsm->lsm_stripe_count) % ost_count;
565 for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
566 rc = lov_check_and_create_object(lov, ost_idx, lsm, req,
576 static int min_stripe_count(int stripe_cnt, int flags)
578 return (flags & LOV_USES_DEFAULT_STRIPE ?
579 stripe_cnt - (stripe_cnt / 4) : stripe_cnt);
582 #define LOV_CREATE_RESEED_MULT 30
583 #define LOV_CREATE_RESEED_MIN 2000
584 /* Allocate objects on osts with round-robin algorithm */
585 static int alloc_rr(struct lov_obd *lov, int *idx_arr, int *stripe_cnt,
586 char *poolname, int flags)
591 int ost_start_idx_temp;
593 int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
594 struct pool_desc *pool;
595 struct ost_pool *osts;
596 struct lov_qos_rr *lqr;
599 pool = lov_find_pool(lov, poolname);
601 osts = &(lov->lov_packed);
602 lqr = &(lov->lov_qos.lq_rr);
604 cfs_down_read(&pool_tgt_rw_sem(pool));
605 osts = &(pool->pool_obds);
606 lqr = &(pool->pool_rr);
609 rc = qos_calc_rr(lov, osts, lqr);
613 if (--lqr->lqr_start_count <= 0) {
614 lqr->lqr_start_idx = cfs_rand() % osts->op_count;
615 lqr->lqr_start_count =
616 (LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) +
617 LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U);
618 } else if (stripe_cnt_min >= osts->op_count ||
619 lqr->lqr_start_idx > osts->op_count) {
620 /* If we have allocated from all of the OSTs, slowly
621 * precess the next start if the OST/stripe count isn't
622 * already doing this for us. */
623 lqr->lqr_start_idx %= osts->op_count;
624 if (*stripe_cnt > 1 && (osts->op_count % (*stripe_cnt)) != 1)
625 ++lqr->lqr_offset_idx;
627 cfs_down_read(&lov->lov_qos.lq_rw_sem);
628 ost_start_idx_temp = lqr->lqr_start_idx;
631 array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) % osts->op_count;
634 CDEBUG(D_QOS, "pool '%s' want %d startidx %d startcnt %d offset %d "
635 "active %d count %d arrayidx %d\n", poolname,
636 *stripe_cnt, lqr->lqr_start_idx, lqr->lqr_start_count,
637 lqr->lqr_offset_idx, osts->op_count, osts->op_count, array_idx);
640 for (i = 0; i < osts->op_count;
641 i++, array_idx=(array_idx + 1) % osts->op_count) {
642 ++lqr->lqr_start_idx;
643 ost_idx = lqr->lqr_pool.op_array[array_idx];
645 CDEBUG(D_QOS, "#%d strt %d act %d strp %d ary %d idx %d\n",
646 i, lqr->lqr_start_idx,
647 ((ost_idx != LOV_QOS_EMPTY) && lov->lov_tgts[ost_idx]) ?
648 lov->lov_tgts[ost_idx]->ltd_active : 0,
649 idx_pos - idx_arr, array_idx, ost_idx);
651 if ((ost_idx == LOV_QOS_EMPTY) || !lov->lov_tgts[ost_idx] ||
652 !lov->lov_tgts[ost_idx]->ltd_active)
655 /* Fail Check before osc_precreate() is called
656 so we can only 'fail' single OSC. */
657 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
660 /* Drop slow OSCs if we can */
661 if (obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed)
666 /* We have enough stripes */
667 if (idx_pos - idx_arr == *stripe_cnt)
670 if ((speed < 2) && (idx_pos - idx_arr < stripe_cnt_min)) {
671 /* Try again, allowing slower OSCs */
673 lqr->lqr_start_idx = ost_start_idx_temp;
677 cfs_up_read(&lov->lov_qos.lq_rw_sem);
679 *stripe_cnt = idx_pos - idx_arr;
682 cfs_up_read(&pool_tgt_rw_sem(pool));
683 /* put back ref got by lov_find_pool() */
684 lov_pool_putref(pool);
690 /* alloc objects on osts with specific stripe offset */
691 static int alloc_specific(struct lov_obd *lov, struct lov_stripe_md *lsm,
694 unsigned ost_idx, array_idx, ost_count;
697 struct pool_desc *pool;
698 struct ost_pool *osts;
701 pool = lov_find_pool(lov, lsm->lsm_pool_name);
703 osts = &(lov->lov_packed);
705 cfs_down_read(&pool_tgt_rw_sem(pool));
706 osts = &(pool->pool_obds);
709 ost_count = osts->op_count;
712 /* search loi_ost_idx in ost array */
714 for (i = 0; i < ost_count; i++) {
715 if (osts->op_array[i] == lsm->lsm_oinfo[0]->loi_ost_idx) {
720 if (i == ost_count) {
721 CERROR("Start index %d not found in pool '%s'\n",
722 lsm->lsm_oinfo[0]->loi_ost_idx, lsm->lsm_pool_name);
723 GOTO(out, rc = -EINVAL);
727 for (i = 0; i < ost_count;
728 i++, array_idx = (array_idx + 1) % ost_count) {
729 ost_idx = osts->op_array[array_idx];
731 if (!lov->lov_tgts[ost_idx] ||
732 !lov->lov_tgts[ost_idx]->ltd_active) {
736 /* Fail Check before osc_precreate() is called
737 so we can only 'fail' single OSC. */
738 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
741 /* Drop slow OSCs if we can, but not for requested start idx.
743 * This means "if OSC is slow and it is not the requested
744 * start OST, then it can be skipped, otherwise skip it only
745 * if it is inactive/recovering/out-of-space." */
746 if ((obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed) &&
747 (i != 0 || speed >= 2))
752 /* We have enough stripes */
753 if (idx_pos - idx_arr == lsm->lsm_stripe_count)
757 /* Try again, allowing slower OSCs */
762 /* If we were passed specific striping params, then a failure to
763 * meet those requirements is an error, since we can't reallocate
764 * that memory (it might be part of a larger array or something).
766 * We can only get here if lsm_stripe_count was originally > 1.
768 CERROR("can't lstripe objid "LPX64": have %d want %u\n",
769 lsm->lsm_object_id, (int)(idx_pos - idx_arr),
770 lsm->lsm_stripe_count);
774 cfs_up_read(&pool_tgt_rw_sem(pool));
775 /* put back ref got by lov_find_pool() */
776 lov_pool_putref(pool);
782 /* Alloc objects on osts with optimization based on:
784 - network resources (shared OSS's)
786 static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt,
787 char *poolname, int flags)
789 struct lov_obd *lov = &exp->exp_obd->u.lov;
790 __u64 total_weight = 0;
791 int nfound, good_osts, i, rc = 0;
792 int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
793 struct pool_desc *pool;
794 struct ost_pool *osts;
797 if (stripe_cnt_min < 1)
800 pool = lov_find_pool(lov, poolname);
802 osts = &(lov->lov_packed);
804 cfs_down_read(&pool_tgt_rw_sem(pool));
805 osts = &(pool->pool_obds);
808 obd_getref(exp->exp_obd);
810 /* wait for fresh statfs info if needed, the rpcs are sent in
812 qos_statfs_update(exp->exp_obd,
813 cfs_time_shift_64(-2 * lov->desc.ld_qos_maxage), 1);
815 /* Detect -EAGAIN early, before expensive lock is taken. */
816 if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space)
817 GOTO(out_nolock, rc = -EAGAIN);
819 /* Do actual allocation, use write lock here. */
820 cfs_down_write(&lov->lov_qos.lq_rw_sem);
823 * Check again, while we were sleeping on @lq_rw_sem things could
826 if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space)
827 GOTO(out, rc = -EAGAIN);
829 if (lov->desc.ld_active_tgt_count < 2)
830 GOTO(out, rc = -EAGAIN);
832 rc = qos_calc_ppo(exp->exp_obd);
837 /* Find all the OSTs that are valid stripe candidates */
838 for (i = 0; i < osts->op_count; i++) {
839 if (!lov->lov_tgts[osts->op_array[i]] ||
840 !lov->lov_tgts[osts->op_array[i]]->ltd_active)
843 /* Fail Check before osc_precreate() is called
844 so we can only 'fail' single OSC. */
845 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && osts->op_array[i] == 0)
848 if (obd_precreate(lov->lov_tgts[osts->op_array[i]]->ltd_exp) > 2)
851 lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable = 1;
852 qos_calc_weight(lov, osts->op_array[i]);
853 total_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight;
859 CDEBUG(D_QOS, "found %d good osts\n", good_osts);
862 if (good_osts < stripe_cnt_min)
863 GOTO(out, rc = -EAGAIN);
865 /* We have enough osts */
866 if (good_osts < *stripe_cnt)
867 *stripe_cnt = good_osts;
870 GOTO(out, rc = -EAGAIN);
872 /* Find enough OSTs with weighted random allocation. */
874 while (nfound < *stripe_cnt) {
875 __u64 rand, cur_weight;
881 #if BITS_PER_LONG == 32
882 rand = cfs_rand() % (unsigned)total_weight;
883 /* If total_weight > 32-bit, first generate the high
884 * 32 bits of the random number, then add in the low
885 * 32 bits (truncated to the upper limit, if needed) */
886 if (total_weight > 0xffffffffULL)
887 rand = (__u64)(cfs_rand() %
888 (unsigned)(total_weight >> 32)) << 32;
892 if (rand == (total_weight & 0xffffffff00000000ULL))
893 rand |= cfs_rand() % (unsigned)total_weight;
898 rand = ((__u64)cfs_rand() << 32 | cfs_rand()) %
905 /* On average, this will hit larger-weighted osts more often.
906 0-weight osts will always get used last (only when rand=0).*/
907 for (i = 0; i < osts->op_count; i++) {
908 if (!lov->lov_tgts[osts->op_array[i]] ||
909 !lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable)
912 cur_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight;
914 CDEBUG(D_QOS, "stripe_cnt=%d nfound=%d cur_weight="LPU64
915 " rand="LPU64" total_weight="LPU64"\n",
916 *stripe_cnt, nfound, cur_weight, rand, total_weight);
918 if (cur_weight >= rand) {
920 CDEBUG(D_QOS, "assigned stripe=%d to idx=%d\n",
921 nfound, osts->op_array[i]);
923 idx_arr[nfound++] = osts->op_array[i];
924 qos_used(lov, osts, osts->op_array[i], &total_weight);
929 /* should never satisfy below condition */
931 CERROR("Didn't find any OSTs?\n");
935 LASSERT(nfound == *stripe_cnt);
938 cfs_up_write(&lov->lov_qos.lq_rw_sem);
942 cfs_up_read(&pool_tgt_rw_sem(pool));
943 /* put back ref got by lov_find_pool() */
944 lov_pool_putref(pool);
948 rc = alloc_rr(lov, idx_arr, stripe_cnt, poolname, flags);
950 obd_putref(exp->exp_obd);
954 /* return new alloced stripe count on success */
955 static int alloc_idx_array(struct obd_export *exp, struct lov_stripe_md *lsm,
956 int newea, int **idx_arr, int *arr_cnt, int flags)
958 struct lov_obd *lov = &exp->exp_obd->u.lov;
959 int stripe_cnt = lsm->lsm_stripe_count;
964 *arr_cnt = stripe_cnt;
965 OBD_ALLOC(tmp_arr, *arr_cnt * sizeof(int));
968 for (i = 0; i < *arr_cnt; i++)
972 lsm->lsm_oinfo[0]->loi_ost_idx >= lov->desc.ld_tgt_count)
973 rc = alloc_qos(exp, tmp_arr, &stripe_cnt,
974 lsm->lsm_pool_name, flags);
976 rc = alloc_specific(lov, lsm, tmp_arr);
984 OBD_FREE(tmp_arr, *arr_cnt * sizeof(int));
989 static void free_idx_array(int *idx_arr, int arr_cnt)
992 OBD_FREE(idx_arr, arr_cnt * sizeof(int));
995 int qos_prep_create(struct obd_export *exp, struct lov_request_set *set)
997 struct lov_obd *lov = &exp->exp_obd->u.lov;
998 struct lov_stripe_md *lsm;
999 struct obdo *src_oa = set->set_oi->oi_oa;
1000 struct obd_trans_info *oti = set->set_oti;
1001 int i, stripes, rc = 0, newea = 0;
1002 int flag = LOV_USES_ASSIGNED_STRIPE;
1003 int *idx_arr = NULL, idx_cnt = 0;
1006 LASSERT(src_oa->o_valid & OBD_MD_FLID);
1007 LASSERT(src_oa->o_valid & OBD_MD_FLGROUP);
1009 if (set->set_oi->oi_md == NULL) {
1010 __u16 stripes_def = lov_get_stripecnt(lov, LOV_MAGIC, 0);
1012 /* If the MDS file was truncated up to some size, stripe over
1013 * enough OSTs to allow the file to be created at that size.
1014 * This may mean we use more than the default # of stripes. */
1015 if (src_oa->o_valid & OBD_MD_FLSIZE) {
1016 obd_size min_bavail = LUSTRE_STRIPE_MAXBYTES;
1018 /* Find a small number of stripes we can use
1019 (up to # of active osts). */
1021 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1022 if (!lov->lov_tgts[i] ||
1023 !lov->lov_tgts[i]->ltd_active)
1025 min_bavail = min(min_bavail, TGT_BAVAIL(i));
1026 if (min_bavail * stripes > src_oa->o_size)
1031 if (stripes < stripes_def)
1032 stripes = stripes_def;
1034 flag = LOV_USES_DEFAULT_STRIPE;
1035 stripes = stripes_def;
1038 rc = lov_alloc_memmd(&set->set_oi->oi_md, stripes,
1039 lov->desc.ld_pattern ?
1040 lov->desc.ld_pattern : LOV_PATTERN_RAID0,
1048 lsm = set->set_oi->oi_md;
1049 lsm->lsm_object_id = src_oa->o_id;
1050 lsm->lsm_object_seq = src_oa->o_seq;
1051 lsm->lsm_layout_gen = 0; /* actual generation set in mdd_lov_create() */
1053 if (!lsm->lsm_stripe_size)
1054 lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size;
1055 if (!lsm->lsm_pattern) {
1056 LASSERT(lov->desc.ld_pattern);
1057 lsm->lsm_pattern = lov->desc.ld_pattern;
1060 stripes = alloc_idx_array(exp, lsm, newea, &idx_arr, &idx_cnt, flag);
1062 GOTO(out_err, rc = stripes ? stripes : -EIO);
1063 LASSERTF(stripes <= lsm->lsm_stripe_count,"requested %d allocated %d\n",
1064 lsm->lsm_stripe_count, stripes);
1066 for (i = 0; i < stripes; i++) {
1067 struct lov_request *req;
1068 int ost_idx = idx_arr[i];
1069 LASSERT(ost_idx >= 0);
1071 OBD_ALLOC(req, sizeof(*req));
1073 GOTO(out_err, rc = -ENOMEM);
1074 lov_set_add_req(req, set);
1076 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
1077 OBD_ALLOC_LARGE(req->rq_oi.oi_md, req->rq_buflen);
1078 if (req->rq_oi.oi_md == NULL)
1079 GOTO(out_err, rc = -ENOMEM);
1081 OBDO_ALLOC(req->rq_oi.oi_oa);
1082 if (req->rq_oi.oi_oa == NULL)
1083 GOTO(out_err, rc = -ENOMEM);
1085 req->rq_idx = ost_idx;
1087 /* create data objects with "parent" OA */
1088 memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
1089 req->rq_oi.oi_cb_up = cb_create_update;
1091 /* XXX When we start creating objects on demand, we need to
1092 * make sure that we always create the object on the
1093 * stripe which holds the existing file size.
1095 if (src_oa->o_valid & OBD_MD_FLSIZE) {
1096 req->rq_oi.oi_oa->o_size =
1097 lov_size_to_stripe(lsm, src_oa->o_size, i);
1099 CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
1100 i, req->rq_oi.oi_oa->o_size, src_oa->o_size);
1103 LASSERT(set->set_count == stripes);
1105 if (stripes < lsm->lsm_stripe_count)
1106 qos_shrink_lsm(set);
1107 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_LOV_PREP_CREATE)) {
1108 qos_shrink_lsm(set);
1112 if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) {
1113 oti_alloc_cookies(oti, set->set_count);
1114 if (!oti->oti_logcookies)
1115 GOTO(out_err, rc = -ENOMEM);
1116 set->set_cookies = oti->oti_logcookies;
1120 obd_free_memmd(exp, &set->set_oi->oi_md);
1122 free_idx_array(idx_arr, idx_cnt);
1127 void qos_update(struct lov_obd *lov)
1130 lov->lov_qos.lq_dirty = 1;
1133 void qos_statfs_done(struct lov_obd *lov)
1135 cfs_down_write(&lov->lov_qos.lq_rw_sem);
1136 if (lov->lov_qos.lq_statfs_in_progress) {
1137 lov->lov_qos.lq_statfs_in_progress = 0;
1138 /* wake up any threads waiting for the statfs rpcs to complete*/
1139 cfs_waitq_signal(&lov->lov_qos.lq_statfs_waitq);
1141 cfs_up_write(&lov->lov_qos.lq_rw_sem);
1144 static int qos_statfs_ready(struct obd_device *obd, __u64 max_age)
1146 struct lov_obd *lov = &obd->u.lov;
1149 cfs_down_read(&lov->lov_qos.lq_rw_sem);
1150 rc = lov->lov_qos.lq_statfs_in_progress == 0 ||
1151 cfs_time_beforeq_64(max_age, obd->obd_osfs_age);
1152 cfs_up_read(&lov->lov_qos.lq_rw_sem);
1157 * Update statfs data if the current osfs age is older than max_age.
1158 * If wait is not set, it means that we are called from lov_create()
1159 * and we should just issue the rpcs without waiting for them to complete.
1160 * If wait is set, we are called from alloc_qos() and we just have
1161 * to wait for the request set to complete.
1163 void qos_statfs_update(struct obd_device *obd, __u64 max_age, int wait)
1165 struct lov_obd *lov = &obd->u.lov;
1166 struct obd_info *oinfo;
1168 struct ptlrpc_request_set *set = NULL;
1171 if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
1172 /* statfs data are quite recent, don't need to refresh it */
1175 if (!wait && lov->lov_qos.lq_statfs_in_progress)
1176 /* statfs already in progress */
1179 cfs_down_write(&lov->lov_qos.lq_rw_sem);
1180 if (lov->lov_qos.lq_statfs_in_progress) {
1181 cfs_up_write(&lov->lov_qos.lq_rw_sem);
1184 /* no statfs in flight, send rpcs */
1185 lov->lov_qos.lq_statfs_in_progress = 1;
1186 cfs_up_write(&lov->lov_qos.lq_rw_sem);
1189 CDEBUG(D_QOS, "%s: did not manage to get fresh statfs data "
1190 "in a timely manner (osfs age "LPU64", max age "LPU64")"
1191 ", sending new statfs rpcs\n",
1192 obd_uuid2str(&lov->desc.ld_uuid), obd->obd_osfs_age,
1195 /* need to send statfs rpcs */
1196 CDEBUG(D_QOS, "sending new statfs requests\n");
1197 memset(lov->lov_qos.lq_statfs_data, 0,
1198 sizeof(*lov->lov_qos.lq_statfs_data));
1199 oinfo = &lov->lov_qos.lq_statfs_data->lsd_oi;
1200 oinfo->oi_osfs = &lov->lov_qos.lq_statfs_data->lsd_statfs;
1201 oinfo->oi_flags = OBD_STATFS_NODELAY;
1202 set = ptlrpc_prep_set();
1204 GOTO(out_failed, rc = -ENOMEM);
1206 rc = obd_statfs_async(obd, oinfo, max_age, set);
1207 if (rc || cfs_list_empty(&set->set_requests)) {
1209 CWARN("statfs failed with %d\n", rc);
1210 GOTO(out_failed, rc);
1212 /* send requests via ptlrpcd */
1213 oinfo->oi_flags |= OBD_STATFS_PTLRPCD;
1214 ptlrpcd_add_rqset(set);
1218 cfs_down_write(&lov->lov_qos.lq_rw_sem);
1219 lov->lov_qos.lq_statfs_in_progress = 0;
1220 /* wake up any threads waiting for the statfs rpcs to complete */
1221 cfs_waitq_signal(&lov->lov_qos.lq_statfs_waitq);
1222 cfs_up_write(&lov->lov_qos.lq_rw_sem);
1226 ptlrpc_set_destroy(set);
1228 struct l_wait_info lwi = { 0 };
1229 CDEBUG(D_QOS, "waiting for statfs requests to complete\n");
1230 l_wait_event(lov->lov_qos.lq_statfs_waitq,
1231 qos_statfs_ready(obd, max_age), &lwi);
1232 if (cfs_time_before_64(obd->obd_osfs_age, max_age))
1233 CDEBUG(D_QOS, "%s: still no fresh statfs data after "
1234 "waiting (osfs age "LPU64", max age "
1236 obd_uuid2str(&lov->desc.ld_uuid),
1237 obd->obd_osfs_age, max_age);