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 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 # define EXPORT_SYMTAB
40 #define DEBUG_SUBSYSTEM S_LOV
43 #include <libcfs/libcfs.h>
45 #include <liblustre.h>
48 #include <obd_class.h>
50 #include <lustre/lustre_idl.h>
51 #include "lov_internal.h"
53 /* #define QOS_DEBUG 1 */
56 #define TGT_BAVAIL(i) (lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bavail*\
57 lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bsize)
58 #define TGT_FFREE(i) (lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_ffree)
61 int qos_add_tgt(struct obd_device *obd, __u32 index)
63 struct lov_obd *lov = &obd->u.lov;
64 struct lov_qos_oss *oss, *temposs;
65 struct obd_export *exp = lov->lov_tgts[index]->ltd_exp;
66 int rc = 0, found = 0;
69 /* We only need this QOS struct on MDT, not clients - but we may not
70 * have registered the LOV's observer yet, so there's no way to know */
71 if (!exp || !exp->exp_connection) {
72 CERROR("Missing connection\n");
76 down_write(&lov->lov_qos.lq_rw_sem);
77 mutex_down(&lov->lov_lock);
78 list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
79 if (obd_uuid_equals(&oss->lqo_uuid,
80 &exp->exp_connection->c_remote_uuid)) {
89 GOTO(out, rc = -ENOMEM);
90 memcpy(&oss->lqo_uuid,
91 &exp->exp_connection->c_remote_uuid,
92 sizeof(oss->lqo_uuid));
94 /* Assume we have to move this one */
95 list_del(&oss->lqo_oss_list);
99 lov->lov_tgts[index]->ltd_qos.ltq_oss = oss;
101 /* Add sorted by # of OSTs. Find the first entry that we're
103 list_for_each_entry(temposs, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
104 if (oss->lqo_ost_count > temposs->lqo_ost_count)
107 /* ...and add before it. If we're the first or smallest, temposs
108 points to the list head, and we add to the end. */
109 list_add_tail(&oss->lqo_oss_list, &temposs->lqo_oss_list);
111 lov->lov_qos.lq_dirty = 1;
112 lov->lov_qos.lq_rr.lqr_dirty = 1;
114 CDEBUG(D_QOS, "add tgt %s to OSS %s (%d OSTs)\n",
115 obd_uuid2str(&lov->lov_tgts[index]->ltd_uuid),
116 obd_uuid2str(&oss->lqo_uuid),
120 mutex_up(&lov->lov_lock);
121 up_write(&lov->lov_qos.lq_rw_sem);
125 int qos_del_tgt(struct obd_device *obd, struct lov_tgt_desc *tgt)
127 struct lov_obd *lov = &obd->u.lov;
128 struct lov_qos_oss *oss;
132 down_write(&lov->lov_qos.lq_rw_sem);
134 oss = tgt->ltd_qos.ltq_oss;
136 GOTO(out, rc = -ENOENT);
138 oss->lqo_ost_count--;
139 if (oss->lqo_ost_count == 0) {
140 CDEBUG(D_QOS, "removing OSS %s\n",
141 obd_uuid2str(&oss->lqo_uuid));
142 list_del(&oss->lqo_oss_list);
146 lov->lov_qos.lq_dirty = 1;
147 lov->lov_qos.lq_rr.lqr_dirty = 1;
149 up_write(&lov->lov_qos.lq_rw_sem);
153 /* Recalculate per-object penalties for OSSs and OSTs,
154 depends on size of each ost in an oss */
155 static int qos_calc_ppo(struct obd_device *obd)
157 struct lov_obd *lov = &obd->u.lov;
158 struct lov_qos_oss *oss;
159 __u64 ba_max, ba_min, temp;
161 int rc, i, prio_wide;
165 if (!lov->lov_qos.lq_dirty)
168 num_active = lov->desc.ld_active_tgt_count - 1;
170 GOTO(out, rc = -EAGAIN);
172 /* find bavail on each OSS */
173 list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
176 lov->lov_qos.lq_active_oss_count = 0;
178 /* How badly user wants to select osts "widely" (not recently chosen
179 and not on recent oss's). As opposed to "freely" (free space
181 prio_wide = 256 - lov->lov_qos.lq_prio_free;
183 ba_min = (__u64)(-1);
185 now = cfs_time_current_sec();
186 /* Calculate OST penalty per object */
187 /* (lov ref taken in alloc_qos) */
188 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
189 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
191 temp = TGT_BAVAIL(i);
194 ba_min = min(temp, ba_min);
195 ba_max = max(temp, ba_max);
197 /* Count the number of usable OSS's */
198 if (lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_bavail == 0)
199 lov->lov_qos.lq_active_oss_count++;
200 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_bavail += temp;
202 /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
204 do_div(temp, num_active);
205 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj =
206 (temp * prio_wide) >> 8;
208 age = (now - lov->lov_tgts[i]->ltd_qos.ltq_used) >> 3;
209 if (lov->lov_qos.lq_reset || age > 32 * lov->desc.ld_qos_maxage)
210 lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
211 else if (age > lov->desc.ld_qos_maxage)
212 /* Decay the penalty by half for every 8x the update
213 * interval that the device has been idle. That gives
214 * lots of time for the statfs information to be
215 * updated (which the penalty is only a proxy for),
216 * and avoids penalizing OSS/OSTs under light load. */
217 lov->lov_tgts[i]->ltd_qos.ltq_penalty >>=
218 (age / lov->desc.ld_qos_maxage);
221 num_active = lov->lov_qos.lq_active_oss_count - 1;
222 if (num_active < 1) {
223 /* If there's only 1 OSS, we can't penalize it, so instead
224 we have to double the OST penalty */
226 for (i = 0; i < lov->desc.ld_tgt_count; i++)
227 if (lov->lov_tgts[i])
228 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj <<= 1;
231 /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
232 list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
233 temp = oss->lqo_bavail >> 1;
234 do_div(temp, oss->lqo_ost_count * num_active);
235 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
237 age = (now - oss->lqo_used) >> 3;
238 if (lov->lov_qos.lq_reset || age > 32 * lov->desc.ld_qos_maxage)
239 oss->lqo_penalty = 0;
240 else if (age > lov->desc.ld_qos_maxage)
241 /* Decay the penalty by half for every 8x the update
242 * interval that the device has been idle. That gives
243 * lots of time for the statfs information to be
244 * updated (which the penalty is only a proxy for),
245 * and avoids penalizing OSS/OSTs under light load. */
246 oss->lqo_penalty >>= (age / lov->desc.ld_qos_maxage);
249 lov->lov_qos.lq_dirty = 0;
250 lov->lov_qos.lq_reset = 0;
252 /* If each ost has almost same free space,
253 * do rr allocation for better creation performance */
254 lov->lov_qos.lq_same_space = 0;
255 if ((ba_max * (256 - lov->lov_qos.lq_threshold_rr)) >> 8 < ba_min) {
256 lov->lov_qos.lq_same_space = 1;
257 /* Reset weights for the next time we enter qos mode */
258 lov->lov_qos.lq_reset = 1;
263 if (!rc && lov->lov_qos.lq_same_space)
268 static int qos_calc_weight(struct lov_obd *lov, int i)
272 /* Final ost weight = TGT_BAVAIL - ost_penalty - oss_penalty */
273 temp = TGT_BAVAIL(i);
274 temp2 = lov->lov_tgts[i]->ltd_qos.ltq_penalty +
275 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty;
277 lov->lov_tgts[i]->ltd_qos.ltq_weight = 0;
279 lov->lov_tgts[i]->ltd_qos.ltq_weight = temp - temp2;
283 /* We just used this index for a stripe; adjust everyone's weights */
284 static int qos_used(struct lov_obd *lov, struct ost_pool *osts,
285 __u32 index, __u64 *total_wt)
287 struct lov_qos_oss *oss;
291 /* Don't allocate from this stripe anymore, until the next alloc_qos */
292 lov->lov_tgts[index]->ltd_qos.ltq_usable = 0;
294 oss = lov->lov_tgts[index]->ltd_qos.ltq_oss;
296 /* Decay old penalty by half (we're adding max penalty, and don't
297 want it to run away.) */
298 lov->lov_tgts[index]->ltd_qos.ltq_penalty >>= 1;
299 oss->lqo_penalty >>= 1;
301 /* mark the OSS and OST as recently used */
302 lov->lov_tgts[index]->ltd_qos.ltq_used =
303 oss->lqo_used = cfs_time_current_sec();
305 /* Set max penalties for this OST and OSS */
306 lov->lov_tgts[index]->ltd_qos.ltq_penalty +=
307 lov->lov_tgts[index]->ltd_qos.ltq_penalty_per_obj *
308 lov->desc.ld_active_tgt_count;
309 oss->lqo_penalty += oss->lqo_penalty_per_obj *
310 lov->lov_qos.lq_active_oss_count;
312 /* Decrease all OSS penalties */
313 list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
314 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
315 oss->lqo_penalty = 0;
317 oss->lqo_penalty -= oss->lqo_penalty_per_obj;
321 /* Decrease all OST penalties */
322 for (j = 0; j < osts->op_count; j++) {
325 i = osts->op_array[j];
326 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
328 if (lov->lov_tgts[i]->ltd_qos.ltq_penalty <
329 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj)
330 lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
332 lov->lov_tgts[i]->ltd_qos.ltq_penalty -=
333 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj;
335 qos_calc_weight(lov, i);
337 /* Recalc the total weight of usable osts */
338 if (lov->lov_tgts[i]->ltd_qos.ltq_usable)
339 *total_wt += lov->lov_tgts[i]->ltd_qos.ltq_weight;
342 CDEBUG(D_QOS, "recalc tgt %d usable=%d avail="LPU64
343 " ostppo="LPU64" ostp="LPU64" ossppo="LPU64
344 " ossp="LPU64" wt="LPU64"\n",
345 i, lov->lov_tgts[i]->ltd_qos.ltq_usable,
347 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj >> 10,
348 lov->lov_tgts[i]->ltd_qos.ltq_penalty >> 10,
349 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty_per_obj>>10,
350 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty >> 10,
351 lov->lov_tgts[i]->ltd_qos.ltq_weight >> 10);
358 #define LOV_QOS_EMPTY ((__u32)-1)
359 /* compute optimal round-robin order, based on OSTs per OSS
361 static int qos_calc_rr(struct lov_obd *lov, struct ost_pool *src_pool,
362 struct lov_qos_rr *lqr)
364 struct lov_qos_oss *oss;
365 unsigned placed, real_count;
369 if (!lqr->lqr_dirty) {
370 LASSERT(lqr->lqr_pool.op_size);
374 /* Do actual allocation. */
375 down_write(&lov->lov_qos.lq_rw_sem);
377 real_count = src_pool->op_count;
379 /* Zero the pool array */
380 /* alloc_rr is holding a read lock on the pool, so nobody is adding/
381 deleting from the pool. The lq_rw_sem insures that nobody else
383 lqr->lqr_pool.op_count = real_count;
384 rc = lov_ost_pool_extend(&lqr->lqr_pool, real_count);
386 up_write(&lov->lov_qos.lq_rw_sem);
389 for (i = 0; i < lqr->lqr_pool.op_count; i++)
390 lqr->lqr_pool.op_array[i] = LOV_QOS_EMPTY;
392 /* Place all the OSTs from 1 OSS at the same time. */
394 list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
396 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
397 if (lov->lov_tgts[src_pool->op_array[i]] &&
398 (lov->lov_tgts[src_pool->op_array[i]]->ltd_qos.ltq_oss == oss)) {
399 /* Evenly space these OSTs across arrayspace */
400 int next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count;
401 while (lqr->lqr_pool.op_array[next] !=
403 next = (next + 1) % lqr->lqr_pool.op_count;
404 lqr->lqr_pool.op_array[next] = src_pool->op_array[i];
412 up_write(&lov->lov_qos.lq_rw_sem);
414 if (placed != real_count) {
415 /* This should never happen */
416 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
417 "round-robin list (%d of %d).\n",
419 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
420 LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
421 lqr->lqr_pool.op_array[i]);
428 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
429 LCONSOLE(D_QOS, "rr #%d ost idx=%d\n", i,
430 lqr->lqr_pool.op_array[i]);
438 void qos_shrink_lsm(struct lov_request_set *set)
440 struct lov_stripe_md *lsm = set->set_oi->oi_md, *lsm_new;
441 /* XXX LOV STACKING call into osc for sizes */
442 unsigned oldsize, newsize;
444 if (set->set_oti && set->set_cookies && set->set_cookie_sent) {
445 struct llog_cookie *cookies;
446 oldsize = lsm->lsm_stripe_count * sizeof(*cookies);
447 newsize = set->set_count * sizeof(*cookies);
449 cookies = set->set_cookies;
450 oti_alloc_cookies(set->set_oti, set->set_count);
451 if (set->set_oti->oti_logcookies) {
452 memcpy(set->set_oti->oti_logcookies, cookies, newsize);
453 OBD_FREE(cookies, oldsize);
454 set->set_cookies = set->set_oti->oti_logcookies;
456 CWARN("'leaking' %d bytes\n", oldsize - newsize);
460 CWARN("using fewer stripes for object "LPU64": old %u new %u\n",
461 lsm->lsm_object_id, lsm->lsm_stripe_count, set->set_count);
462 LASSERT(lsm->lsm_stripe_count >= set->set_count);
464 newsize = lov_stripe_md_size(set->set_count);
465 OBD_ALLOC(lsm_new, newsize);
466 if (lsm_new != NULL) {
468 memcpy(lsm_new, lsm, sizeof(*lsm));
469 for (i = 0; i < lsm->lsm_stripe_count; i++) {
470 if (i < set->set_count) {
471 lsm_new->lsm_oinfo[i] = lsm->lsm_oinfo[i];
474 OBD_SLAB_FREE(lsm->lsm_oinfo[i], lov_oinfo_slab,
475 sizeof(struct lov_oinfo));
477 lsm_new->lsm_stripe_count = set->set_count;
478 OBD_FREE(lsm, sizeof(struct lov_stripe_md) +
479 lsm->lsm_stripe_count * sizeof(struct lov_oinfo *));
480 set->set_oi->oi_md = lsm_new;
482 CWARN("'leaking' few bytes\n");
486 int qos_remedy_create(struct lov_request_set *set, struct lov_request *req)
488 struct lov_stripe_md *lsm = set->set_oi->oi_md;
489 struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
490 unsigned ost_idx, ost_count = lov->desc.ld_tgt_count;
491 int stripe, i, rc = -EIO;
494 ost_idx = (req->rq_idx + lsm->lsm_stripe_count) % ost_count;
495 for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
496 if (!lov->lov_tgts[ost_idx] ||
497 !lov->lov_tgts[ost_idx]->ltd_active)
499 /* check if objects has been created on this ost */
500 for (stripe = 0; stripe < lsm->lsm_stripe_count; stripe++) {
501 /* we try send create to this ost but he is failed */
502 if (stripe == req->rq_stripe)
504 /* already have object at this stripe */
505 if (ost_idx == lsm->lsm_oinfo[stripe]->loi_ost_idx)
508 if (stripe >= lsm->lsm_stripe_count) {
509 req->rq_idx = ost_idx;
510 rc = obd_create(lov->lov_tgts[ost_idx]->ltd_exp,
511 req->rq_oi.oi_oa, &req->rq_oi.oi_md,
520 static int min_stripe_count(int stripe_cnt, int flags)
522 return (flags & LOV_USES_DEFAULT_STRIPE ?
523 stripe_cnt - (stripe_cnt / 4) : stripe_cnt);
526 #define LOV_CREATE_RESEED_MULT 4
527 #define LOV_CREATE_RESEED_MIN 1000
528 /* Allocate objects on osts with round-robin algorithm */
529 static int alloc_rr(struct lov_obd *lov, int *idx_arr, int *stripe_cnt,
530 char *poolname, int flags)
535 int ost_start_idx_temp;
537 int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
538 struct pool_desc *pool;
539 struct ost_pool *osts;
540 struct lov_qos_rr *lqr;
543 pool = lov_find_pool(lov, poolname);
545 osts = &(lov->lov_packed);
546 lqr = &(lov->lov_qos.lq_rr);
548 down_read(&pool_tgt_rw_sem(pool));
549 osts = &(pool->pool_obds);
550 lqr = &(pool->pool_rr);
553 rc = qos_calc_rr(lov, osts, lqr);
557 if (--lqr->lqr_start_count <= 0) {
558 lqr->lqr_start_idx = ll_rand() % osts->op_count;
559 lqr->lqr_start_count =
560 (LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) +
561 LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U);
562 } else if (stripe_cnt_min >= osts->op_count ||
563 lqr->lqr_start_idx > osts->op_count) {
564 /* If we have allocated from all of the OSTs, slowly
565 * precess the next start if the OST/stripe count isn't
566 * already doing this for us. */
567 lqr->lqr_start_idx %= osts->op_count;
568 if (*stripe_cnt > 1 && (osts->op_count % (*stripe_cnt)) != 1)
569 ++lqr->lqr_offset_idx;
571 down_read(&lov->lov_qos.lq_rw_sem);
572 ost_start_idx_temp = lqr->lqr_start_idx;
575 array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) % osts->op_count;
578 CDEBUG(D_QOS, "pool '%s' want %d startidx %d startcnt %d offset %d "
579 "active %d count %d arrayidx %d\n", poolname,
580 *stripe_cnt, lqr->lqr_start_idx, lqr->lqr_start_count,
581 lqr->lqr_offset_idx, osts->op_count, osts->op_count, array_idx);
584 for (i = 0; i < osts->op_count;
585 i++, array_idx=(array_idx + 1) % osts->op_count) {
586 ++lqr->lqr_start_idx;
587 ost_idx = lqr->lqr_pool.op_array[array_idx];
589 CDEBUG(D_QOS, "#%d strt %d act %d strp %d ary %d idx %d\n",
590 i, lqr->lqr_start_idx,
591 ((ost_idx != LOV_QOS_EMPTY) && lov->lov_tgts[ost_idx]) ?
592 lov->lov_tgts[ost_idx]->ltd_active : 0,
593 idx_pos - idx_arr, array_idx, ost_idx);
595 if ((ost_idx == LOV_QOS_EMPTY) || !lov->lov_tgts[ost_idx] ||
596 !lov->lov_tgts[ost_idx]->ltd_active)
599 /* Fail Check before osc_precreate() is called
600 so we can only 'fail' single OSC. */
601 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
604 /* Drop slow OSCs if we can */
605 if (obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed)
610 /* We have enough stripes */
611 if (idx_pos - idx_arr == *stripe_cnt)
614 if ((speed < 2) && (idx_pos - idx_arr < stripe_cnt_min)) {
615 /* Try again, allowing slower OSCs */
617 lqr->lqr_start_idx = ost_start_idx_temp;
621 up_read(&lov->lov_qos.lq_rw_sem);
623 *stripe_cnt = idx_pos - idx_arr;
626 up_read(&pool_tgt_rw_sem(pool));
627 /* put back ref got by lov_find_pool() */
628 lov_pool_putref(pool);
634 /* alloc objects on osts with specific stripe offset */
635 static int alloc_specific(struct lov_obd *lov, struct lov_stripe_md *lsm,
638 unsigned ost_idx, array_idx, ost_count;
641 struct pool_desc *pool;
642 struct ost_pool *osts;
645 pool = lov_find_pool(lov, lsm->lsm_pool_name);
647 osts = &(lov->lov_packed);
649 down_read(&pool_tgt_rw_sem(pool));
650 osts = &(pool->pool_obds);
653 ost_count = osts->op_count;
656 /* search loi_ost_idx in ost array */
658 for (i = 0; i < ost_count; i++) {
659 if (osts->op_array[i] == lsm->lsm_oinfo[0]->loi_ost_idx) {
664 if (i == ost_count) {
665 CERROR("Start index %d not found in pool '%s'\n",
666 lsm->lsm_oinfo[0]->loi_ost_idx, lsm->lsm_pool_name);
667 GOTO(out, rc = -EINVAL);
671 for (i = 0; i < ost_count;
672 i++, array_idx = (array_idx + 1) % ost_count) {
673 ost_idx = osts->op_array[array_idx];
675 if (!lov->lov_tgts[ost_idx] ||
676 !lov->lov_tgts[ost_idx]->ltd_active) {
680 /* Fail Check before osc_precreate() is called
681 so we can only 'fail' single OSC. */
682 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
685 /* Drop slow OSCs if we can, but not for requested start idx.
687 * This means "if OSC is slow and it is not the requested
688 * start OST, then it can be skipped, otherwise skip it only
689 * if it is inactive/recovering/out-of-space." */
690 if ((obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed) &&
691 (i != 0 || speed >= 2))
696 /* We have enough stripes */
697 if (idx_pos - idx_arr == lsm->lsm_stripe_count)
701 /* Try again, allowing slower OSCs */
706 /* If we were passed specific striping params, then a failure to
707 * meet those requirements is an error, since we can't reallocate
708 * that memory (it might be part of a larger array or something).
710 * We can only get here if lsm_stripe_count was originally > 1.
712 CERROR("can't lstripe objid "LPX64": have %d want %u\n",
713 lsm->lsm_object_id, (int)(idx_pos - idx_arr),
714 lsm->lsm_stripe_count);
718 up_read(&pool_tgt_rw_sem(pool));
719 /* put back ref got by lov_find_pool() */
720 lov_pool_putref(pool);
725 /* Alloc objects on osts with optimization based on:
727 - network resources (shared OSS's)
729 static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt,
730 char *poolname, int flags)
732 struct lov_obd *lov = &exp->exp_obd->u.lov;
733 static time_t last_warn = 0;
734 time_t now = cfs_time_current_sec();
735 __u64 total_bavail, total_weight = 0;
736 int nfound, good_osts, i, warn = 0, rc = 0;
737 int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
738 struct pool_desc *pool;
739 struct ost_pool *osts;
740 struct lov_qos_rr *lqr;
743 if (stripe_cnt_min < 1)
746 pool = lov_find_pool(lov, poolname);
748 osts = &(lov->lov_packed);
749 lqr = &(lov->lov_qos.lq_rr);
751 down_read(&pool_tgt_rw_sem(pool));
752 osts = &(pool->pool_obds);
753 lqr = &(pool->pool_rr);
756 obd_getref(exp->exp_obd);
757 /* wait for fresh statfs info if needed, the rpcs are sent in
759 qos_statfs_update(exp->exp_obd,
760 cfs_time_shift_64(-2 * lov->desc.ld_qos_maxage), 1);
762 down_write(&lov->lov_qos.lq_rw_sem);
764 if (lov->desc.ld_active_tgt_count < 2)
765 GOTO(out, rc = -EAGAIN);
767 rc = qos_calc_ppo(exp->exp_obd);
773 /* Warn users about zero available space/inode every 30 min */
774 if (cfs_time_sub(now, last_warn) > 60 * 30)
776 /* Find all the OSTs that are valid stripe candidates */
777 for (i = 0; i < osts->op_count; i++) {
780 if (!lov->lov_tgts[osts->op_array[i]] ||
781 !lov->lov_tgts[osts->op_array[i]]->ltd_active)
783 bavail = TGT_BAVAIL(osts->op_array[i]);
786 CDEBUG(D_QOS, "no free space on %s\n",
787 obd_uuid2str(&lov->lov_tgts[osts->op_array[i]]->ltd_uuid));
792 if (!TGT_FFREE(osts->op_array[i])) {
794 CDEBUG(D_QOS, "no free inodes on %s\n",
795 obd_uuid2str(&lov->lov_tgts[osts->op_array[i]]->ltd_uuid));
801 /* Fail Check before osc_precreate() is called
802 so we can only 'fail' single OSC. */
803 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && osts->op_array[i] == 0)
806 if (obd_precreate(lov->lov_tgts[osts->op_array[i]]->ltd_exp) > 2)
809 lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable = 1;
810 qos_calc_weight(lov, osts->op_array[i]);
811 total_bavail += bavail;
812 total_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight;
818 CDEBUG(D_QOS, "found %d good osts\n", good_osts);
821 if (good_osts < stripe_cnt_min)
822 GOTO(out, rc = -EAGAIN);
825 GOTO(out, rc = -ENOSPC);
827 /* We have enough osts */
828 if (good_osts < *stripe_cnt)
829 *stripe_cnt = good_osts;
831 /* Find enough OSTs with weighted random allocation. */
833 while (nfound < *stripe_cnt) {
834 __u64 rand, cur_weight;
840 #if BITS_PER_LONG == 32
841 rand = ll_rand() % (unsigned)total_weight;
842 /* If total_weight > 32-bit, first generate the high
843 * 32 bits of the random number, then add in the low
844 * 32 bits (truncated to the upper limit, if needed) */
845 if (total_weight > 0xffffffffULL)
846 rand = (__u64)(ll_rand() %
847 (unsigned)(total_weight >> 32)) << 32;
851 if (rand == (total_weight & 0xffffffff00000000ULL))
852 rand |= ll_rand() % (unsigned)total_weight;
857 rand = ((__u64)ll_rand() << 32 | ll_rand()) %
864 /* On average, this will hit larger-weighted osts more often.
865 0-weight osts will always get used last (only when rand=0).*/
866 for (i = 0; i < osts->op_count; i++) {
867 if (!lov->lov_tgts[osts->op_array[i]] ||
868 !lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable)
871 cur_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight;
873 CDEBUG(D_QOS, "stripe_cnt=%d nfound=%d cur_weight="LPU64
874 " rand="LPU64" total_weight="LPU64"\n",
875 *stripe_cnt, nfound, cur_weight, rand, total_weight);
877 if (cur_weight >= rand) {
879 CDEBUG(D_QOS, "assigned stripe=%d to idx=%d\n",
880 nfound, osts->op_array[i]);
882 idx_arr[nfound++] = osts->op_array[i];
883 qos_used(lov, osts, osts->op_array[i], &total_weight);
889 CDEBUG(D_QOS, "Didn't find any OSTs? Reduce total weight\n");
890 if (total_weight == 0)
897 LASSERT(nfound == *stripe_cnt);
900 up_write(&lov->lov_qos.lq_rw_sem);
903 up_read(&pool_tgt_rw_sem(pool));
904 /* put back ref got by lov_find_pool() */
905 lov_pool_putref(pool);
909 rc = alloc_rr(lov, idx_arr, stripe_cnt, poolname, flags);
911 obd_putref(exp->exp_obd);
915 /* return new alloced stripe count on success */
916 static int alloc_idx_array(struct obd_export *exp, struct lov_stripe_md *lsm,
917 int newea, int **idx_arr, int *arr_cnt, int flags)
919 struct lov_obd *lov = &exp->exp_obd->u.lov;
920 int stripe_cnt = lsm->lsm_stripe_count;
925 *arr_cnt = stripe_cnt;
926 OBD_ALLOC(tmp_arr, *arr_cnt * sizeof(int));
929 for (i = 0; i < *arr_cnt; i++)
933 lsm->lsm_oinfo[0]->loi_ost_idx >= lov->desc.ld_tgt_count)
934 rc = alloc_qos(exp, tmp_arr, &stripe_cnt,
935 lsm->lsm_pool_name, flags);
937 rc = alloc_specific(lov, lsm, tmp_arr);
945 OBD_FREE(tmp_arr, *arr_cnt * sizeof(int));
950 static void free_idx_array(int *idx_arr, int arr_cnt)
953 OBD_FREE(idx_arr, arr_cnt * sizeof(int));
956 int qos_prep_create(struct obd_export *exp, struct lov_request_set *set)
958 struct lov_obd *lov = &exp->exp_obd->u.lov;
959 struct lov_stripe_md *lsm;
960 struct obdo *src_oa = set->set_oi->oi_oa;
961 struct obd_trans_info *oti = set->set_oti;
962 int i, stripes, rc = 0, newea = 0;
963 int flag = LOV_USES_ASSIGNED_STRIPE;
964 int *idx_arr = NULL, idx_cnt = 0;
967 LASSERT(src_oa->o_valid & OBD_MD_FLID);
969 if (set->set_oi->oi_md == NULL) {
970 int stripes_def = lov_get_stripecnt(lov, 0);
972 /* If the MDS file was truncated up to some size, stripe over
973 * enough OSTs to allow the file to be created at that size.
974 * This may mean we use more than the default # of stripes. */
975 if (src_oa->o_valid & OBD_MD_FLSIZE) {
976 obd_size min_bavail = LUSTRE_STRIPE_MAXBYTES;
978 /* Find a small number of stripes we can use
979 (up to # of active osts). */
981 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
982 if (!lov->lov_tgts[i] ||
983 !lov->lov_tgts[i]->ltd_active)
985 min_bavail = min(min_bavail, TGT_BAVAIL(i));
986 if (min_bavail * stripes > src_oa->o_size)
991 if (stripes < stripes_def)
992 stripes = stripes_def;
994 flag = LOV_USES_DEFAULT_STRIPE;
995 stripes = stripes_def;
998 rc = lov_alloc_memmd(&set->set_oi->oi_md, stripes,
999 lov->desc.ld_pattern ?
1000 lov->desc.ld_pattern : LOV_PATTERN_RAID0,
1008 lsm = set->set_oi->oi_md;
1009 lsm->lsm_object_id = src_oa->o_id;
1010 if (!lsm->lsm_stripe_size)
1011 lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size;
1012 if (!lsm->lsm_pattern) {
1013 LASSERT(lov->desc.ld_pattern);
1014 lsm->lsm_pattern = lov->desc.ld_pattern;
1017 stripes = alloc_idx_array(exp, lsm, newea, &idx_arr, &idx_cnt, flag);
1019 GOTO(out_err, rc = stripes ? stripes : -EIO);
1020 LASSERTF(stripes <= lsm->lsm_stripe_count,"requested %d allocated %d\n",
1021 lsm->lsm_stripe_count, stripes);
1023 for (i = 0; i < stripes; i++) {
1024 struct lov_request *req;
1025 int ost_idx = idx_arr[i];
1026 LASSERT(ost_idx >= 0);
1028 OBD_ALLOC(req, sizeof(*req));
1030 GOTO(out_err, rc = -ENOMEM);
1031 lov_set_add_req(req, set);
1033 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
1034 OBD_ALLOC(req->rq_oi.oi_md, req->rq_buflen);
1035 if (req->rq_oi.oi_md == NULL)
1036 GOTO(out_err, rc = -ENOMEM);
1038 OBDO_ALLOC(req->rq_oi.oi_oa);
1039 if (req->rq_oi.oi_oa == NULL)
1040 GOTO(out_err, rc = -ENOMEM);
1042 req->rq_idx = ost_idx;
1044 /* create data objects with "parent" OA */
1045 memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
1046 req->rq_oi.oi_cb_up = cb_create_update;
1048 /* XXX When we start creating objects on demand, we need to
1049 * make sure that we always create the object on the
1050 * stripe which holds the existing file size.
1052 if (src_oa->o_valid & OBD_MD_FLSIZE) {
1053 req->rq_oi.oi_oa->o_size =
1054 lov_size_to_stripe(lsm, src_oa->o_size, i);
1056 CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
1057 i, req->rq_oi.oi_oa->o_size, src_oa->o_size);
1060 LASSERT(set->set_count == stripes);
1062 if (stripes < lsm->lsm_stripe_count)
1063 qos_shrink_lsm(set);
1064 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_LOV_PREP_CREATE)) {
1065 qos_shrink_lsm(set);
1069 if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) {
1070 oti_alloc_cookies(oti, set->set_count);
1071 if (!oti->oti_logcookies)
1072 GOTO(out_err, rc = -ENOMEM);
1073 set->set_cookies = oti->oti_logcookies;
1077 obd_free_memmd(exp, &set->set_oi->oi_md);
1079 free_idx_array(idx_arr, idx_cnt);
1084 void qos_update(struct lov_obd *lov)
1087 lov->lov_qos.lq_dirty = 1;
1090 void qos_statfs_done(struct lov_obd *lov)
1092 LASSERT(lov->lov_qos.lq_statfs_in_progress);
1093 down_write(&lov->lov_qos.lq_rw_sem);
1094 lov->lov_qos.lq_statfs_in_progress = 0;
1095 /* wake up any threads waiting for the statfs rpcs to complete */
1096 cfs_waitq_signal(&lov->lov_qos.lq_statfs_waitq);
1097 up_write(&lov->lov_qos.lq_rw_sem);
1100 static int qos_statfs_ready(struct obd_device *obd, __u64 max_age)
1102 struct lov_obd *lov = &obd->u.lov;
1105 down_read(&lov->lov_qos.lq_rw_sem);
1106 rc = lov->lov_qos.lq_statfs_in_progress == 0 ||
1107 cfs_time_beforeq_64(max_age, obd->obd_osfs_age);
1108 up_read(&lov->lov_qos.lq_rw_sem);
1113 * Update statfs data if the current osfs age is older than max_age.
1114 * If wait is not set, it means that we are called from lov_create()
1115 * and we should just issue the rpcs without waiting for them to complete.
1116 * If wait is set, we are called from alloc_qos() and we just have
1117 * to wait for the request set to complete.
1119 void qos_statfs_update(struct obd_device *obd, __u64 max_age, int wait)
1121 struct lov_obd *lov = &obd->u.lov;
1122 struct obd_info *oinfo;
1124 struct ptlrpc_request_set *set = NULL;
1127 if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
1128 /* statfs data are quite recent, don't need to refresh it */
1131 if (!wait && lov->lov_qos.lq_statfs_in_progress)
1132 /* statfs already in progress */
1135 down_write(&lov->lov_qos.lq_rw_sem);
1136 if (lov->lov_qos.lq_statfs_in_progress) {
1137 up_write(&lov->lov_qos.lq_rw_sem);
1140 /* no statfs in flight, send rpcs */
1141 lov->lov_qos.lq_statfs_in_progress = 1;
1142 up_write(&lov->lov_qos.lq_rw_sem);
1145 CDEBUG(D_QOS, "%s: did not manage to get fresh statfs data "
1146 "in a timely manner (osfs age "LPU64", max age "LPU64")"
1147 ", sending new statfs rpcs\n",
1148 obd_uuid2str(&lov->desc.ld_uuid), obd->obd_osfs_age,
1151 /* need to send statfs rpcs */
1152 CDEBUG(D_QOS, "sending new statfs requests\n");
1153 memset(lov->lov_qos.lq_statfs_data, 0,
1154 sizeof(*lov->lov_qos.lq_statfs_data));
1155 oinfo = &lov->lov_qos.lq_statfs_data->lsd_oi;
1156 oinfo->oi_osfs = &lov->lov_qos.lq_statfs_data->lsd_statfs;
1157 oinfo->oi_flags = OBD_STATFS_NODELAY;
1158 set = ptlrpc_prep_set();
1160 GOTO(out_failed, rc = -ENOMEM);
1162 rc = obd_statfs_async(obd, oinfo, max_age, set);
1163 if (rc || list_empty(&set->set_requests)) {
1165 CWARN("statfs failed with %d\n", rc);
1166 GOTO(out_failed, rc);
1168 /* send requests via ptlrpcd */
1169 oinfo->oi_flags |= OBD_STATFS_PTLRPCD;
1170 ptlrpcd_add_rqset(set);
1174 down_write(&lov->lov_qos.lq_rw_sem);
1175 lov->lov_qos.lq_statfs_in_progress = 0;
1176 /* wake up any threads waiting for the statfs rpcs to complete */
1177 cfs_waitq_signal(&lov->lov_qos.lq_statfs_waitq);
1178 up_write(&lov->lov_qos.lq_rw_sem);
1182 ptlrpc_set_destroy(set);
1184 struct l_wait_info lwi = { 0 };
1185 CDEBUG(D_QOS, "waiting for statfs requests to complete\n");
1186 l_wait_event(lov->lov_qos.lq_statfs_waitq,
1187 qos_statfs_ready(obd, max_age), &lwi);
1188 if (cfs_time_before_64(obd->obd_osfs_age, max_age))
1189 CDEBUG(D_QOS, "%s: still no fresh statfs data after "
1190 "waiting (osfs age "LPU64", max age "
1192 obd_uuid2str(&lov->desc.ld_uuid),
1193 obd->obd_osfs_age, max_age);