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) 2004, 2010, Oracle and/or its affiliates. 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)
59 int qos_add_tgt(struct obd_device *obd, __u32 index)
61 struct lov_obd *lov = &obd->u.lov;
62 struct lov_qos_oss *oss, *temposs;
63 struct obd_export *exp = lov->lov_tgts[index]->ltd_exp;
64 int rc = 0, found = 0;
67 /* We only need this QOS struct on MDT, not clients - but we may not
68 * have registered the LOV's observer yet, so there's no way to know */
69 if (!exp || !exp->exp_connection) {
70 CERROR("Missing connection\n");
74 down_write(&lov->lov_qos.lq_rw_sem);
75 mutex_down(&lov->lov_lock);
76 list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
77 if (obd_uuid_equals(&oss->lqo_uuid,
78 &exp->exp_connection->c_remote_uuid)) {
87 GOTO(out, rc = -ENOMEM);
88 memcpy(&oss->lqo_uuid,
89 &exp->exp_connection->c_remote_uuid,
90 sizeof(oss->lqo_uuid));
92 /* Assume we have to move this one */
93 list_del(&oss->lqo_oss_list);
97 lov->lov_tgts[index]->ltd_qos.ltq_oss = oss;
99 /* Add sorted by # of OSTs. Find the first entry that we're
101 list_for_each_entry(temposs, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
102 if (oss->lqo_ost_count > temposs->lqo_ost_count)
105 /* ...and add before it. If we're the first or smallest, temposs
106 points to the list head, and we add to the end. */
107 list_add_tail(&oss->lqo_oss_list, &temposs->lqo_oss_list);
109 lov->lov_qos.lq_dirty = 1;
110 lov->lov_qos.lq_rr.lqr_dirty = 1;
112 CDEBUG(D_QOS, "add tgt %s to OSS %s (%d OSTs)\n",
113 obd_uuid2str(&lov->lov_tgts[index]->ltd_uuid),
114 obd_uuid2str(&oss->lqo_uuid),
118 mutex_up(&lov->lov_lock);
119 up_write(&lov->lov_qos.lq_rw_sem);
123 int qos_del_tgt(struct obd_device *obd, struct lov_tgt_desc *tgt)
125 struct lov_obd *lov = &obd->u.lov;
126 struct lov_qos_oss *oss;
130 down_write(&lov->lov_qos.lq_rw_sem);
132 oss = tgt->ltd_qos.ltq_oss;
134 GOTO(out, rc = -ENOENT);
136 oss->lqo_ost_count--;
137 if (oss->lqo_ost_count == 0) {
138 CDEBUG(D_QOS, "removing OSS %s\n",
139 obd_uuid2str(&oss->lqo_uuid));
140 list_del(&oss->lqo_oss_list);
144 lov->lov_qos.lq_dirty = 1;
145 lov->lov_qos.lq_rr.lqr_dirty = 1;
147 up_write(&lov->lov_qos.lq_rw_sem);
151 /* Recalculate per-object penalties for OSSs and OSTs,
152 depends on size of each ost in an oss */
153 static int qos_calc_ppo(struct obd_device *obd)
155 struct lov_obd *lov = &obd->u.lov;
156 struct lov_qos_oss *oss;
157 __u64 ba_max, ba_min, temp;
159 int rc, i, prio_wide;
163 if (!lov->lov_qos.lq_dirty)
166 num_active = lov->desc.ld_active_tgt_count - 1;
168 GOTO(out, rc = -EAGAIN);
170 /* find bavail on each OSS */
171 list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
174 lov->lov_qos.lq_active_oss_count = 0;
176 /* How badly user wants to select osts "widely" (not recently chosen
177 and not on recent oss's). As opposed to "freely" (free space
179 prio_wide = 256 - lov->lov_qos.lq_prio_free;
181 ba_min = (__u64)(-1);
183 now = cfs_time_current_sec();
184 /* Calculate OST penalty per object */
185 /* (lov ref taken in alloc_qos) */
186 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
187 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
189 temp = TGT_BAVAIL(i);
192 ba_min = min(temp, ba_min);
193 ba_max = max(temp, ba_max);
195 /* Count the number of usable OSS's */
196 if (lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_bavail == 0)
197 lov->lov_qos.lq_active_oss_count++;
198 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_bavail += temp;
200 /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
202 do_div(temp, num_active);
203 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj =
204 (temp * prio_wide) >> 8;
206 age = (now - lov->lov_tgts[i]->ltd_qos.ltq_used) >> 3;
207 if (lov->lov_qos.lq_reset || age > 32 * lov->desc.ld_qos_maxage)
208 lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
209 else if (age > lov->desc.ld_qos_maxage)
210 /* Decay the penalty by half for every 8x the update
211 * interval that the device has been idle. That gives
212 * lots of time for the statfs information to be
213 * updated (which the penalty is only a proxy for),
214 * and avoids penalizing OSS/OSTs under light load. */
215 lov->lov_tgts[i]->ltd_qos.ltq_penalty >>=
216 (age / lov->desc.ld_qos_maxage);
219 num_active = lov->lov_qos.lq_active_oss_count - 1;
220 if (num_active < 1) {
221 /* If there's only 1 OSS, we can't penalize it, so instead
222 we have to double the OST penalty */
224 for (i = 0; i < lov->desc.ld_tgt_count; i++)
225 if (lov->lov_tgts[i])
226 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj <<= 1;
229 /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
230 list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
231 temp = oss->lqo_bavail >> 1;
232 do_div(temp, oss->lqo_ost_count * num_active);
233 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
235 age = (now - oss->lqo_used) >> 3;
236 if (lov->lov_qos.lq_reset || age > 32 * lov->desc.ld_qos_maxage)
237 oss->lqo_penalty = 0;
238 else if (age > lov->desc.ld_qos_maxage)
239 /* Decay the penalty by half for every 8x the update
240 * interval that the device has been idle. That gives
241 * lots of time for the statfs information to be
242 * updated (which the penalty is only a proxy for),
243 * and avoids penalizing OSS/OSTs under light load. */
244 oss->lqo_penalty >>= (age / lov->desc.ld_qos_maxage);
247 lov->lov_qos.lq_dirty = 0;
248 lov->lov_qos.lq_reset = 0;
250 /* If each ost has almost same free space,
251 * do rr allocation for better creation performance */
252 lov->lov_qos.lq_same_space = 0;
253 if ((ba_max * (256 - lov->lov_qos.lq_threshold_rr)) >> 8 < ba_min) {
254 lov->lov_qos.lq_same_space = 1;
255 /* Reset weights for the next time we enter qos mode */
256 lov->lov_qos.lq_reset = 1;
261 if (!rc && lov->lov_qos.lq_same_space)
266 static int qos_calc_weight(struct lov_obd *lov, int i)
270 /* Final ost weight = TGT_BAVAIL - ost_penalty - oss_penalty */
271 temp = TGT_BAVAIL(i);
272 temp2 = lov->lov_tgts[i]->ltd_qos.ltq_penalty +
273 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty;
275 lov->lov_tgts[i]->ltd_qos.ltq_weight = 0;
277 lov->lov_tgts[i]->ltd_qos.ltq_weight = temp - temp2;
281 /* We just used this index for a stripe; adjust everyone's weights */
282 static int qos_used(struct lov_obd *lov, struct ost_pool *osts,
283 __u32 index, __u64 *total_wt)
285 struct lov_qos_oss *oss;
289 /* Don't allocate from this stripe anymore, until the next alloc_qos */
290 lov->lov_tgts[index]->ltd_qos.ltq_usable = 0;
292 oss = lov->lov_tgts[index]->ltd_qos.ltq_oss;
294 /* Decay old penalty by half (we're adding max penalty, and don't
295 want it to run away.) */
296 lov->lov_tgts[index]->ltd_qos.ltq_penalty >>= 1;
297 oss->lqo_penalty >>= 1;
299 /* mark the OSS and OST as recently used */
300 lov->lov_tgts[index]->ltd_qos.ltq_used =
301 oss->lqo_used = cfs_time_current_sec();
303 /* Set max penalties for this OST and OSS */
304 lov->lov_tgts[index]->ltd_qos.ltq_penalty +=
305 lov->lov_tgts[index]->ltd_qos.ltq_penalty_per_obj *
306 lov->desc.ld_active_tgt_count;
307 oss->lqo_penalty += oss->lqo_penalty_per_obj *
308 lov->lov_qos.lq_active_oss_count;
310 /* Decrease all OSS penalties */
311 list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
312 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
313 oss->lqo_penalty = 0;
315 oss->lqo_penalty -= oss->lqo_penalty_per_obj;
319 /* Decrease all OST penalties */
320 for (j = 0; j < osts->op_count; j++) {
323 i = osts->op_array[j];
324 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
326 if (lov->lov_tgts[i]->ltd_qos.ltq_penalty <
327 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj)
328 lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
330 lov->lov_tgts[i]->ltd_qos.ltq_penalty -=
331 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj;
333 qos_calc_weight(lov, i);
335 /* Recalc the total weight of usable osts */
336 if (lov->lov_tgts[i]->ltd_qos.ltq_usable)
337 *total_wt += lov->lov_tgts[i]->ltd_qos.ltq_weight;
340 CDEBUG(D_QOS, "recalc tgt %d usable=%d avail="LPU64
341 " ostppo="LPU64" ostp="LPU64" ossppo="LPU64
342 " ossp="LPU64" wt="LPU64"\n",
343 i, lov->lov_tgts[i]->ltd_qos.ltq_usable,
345 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj >> 10,
346 lov->lov_tgts[i]->ltd_qos.ltq_penalty >> 10,
347 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty_per_obj>>10,
348 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty >> 10,
349 lov->lov_tgts[i]->ltd_qos.ltq_weight >> 10);
356 #define LOV_QOS_EMPTY ((__u32)-1)
357 /* compute optimal round-robin order, based on OSTs per OSS
359 static int qos_calc_rr(struct lov_obd *lov, struct ost_pool *src_pool,
360 struct lov_qos_rr *lqr)
362 struct lov_qos_oss *oss;
363 unsigned placed, real_count;
367 if (!lqr->lqr_dirty) {
368 LASSERT(lqr->lqr_pool.op_size);
372 /* Do actual allocation. */
373 down_write(&lov->lov_qos.lq_rw_sem);
375 real_count = src_pool->op_count;
377 /* Zero the pool array */
378 /* alloc_rr is holding a read lock on the pool, so nobody is adding/
379 deleting from the pool. The lq_rw_sem insures that nobody else
381 lqr->lqr_pool.op_count = real_count;
382 rc = lov_ost_pool_extend(&lqr->lqr_pool, real_count);
384 up_write(&lov->lov_qos.lq_rw_sem);
387 for (i = 0; i < lqr->lqr_pool.op_count; i++)
388 lqr->lqr_pool.op_array[i] = LOV_QOS_EMPTY;
390 /* Place all the OSTs from 1 OSS at the same time. */
392 list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
394 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
395 if (lov->lov_tgts[src_pool->op_array[i]] &&
396 (lov->lov_tgts[src_pool->op_array[i]]->ltd_qos.ltq_oss == oss)) {
397 /* Evenly space these OSTs across arrayspace */
398 int next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count;
399 while (lqr->lqr_pool.op_array[next] !=
401 next = (next + 1) % lqr->lqr_pool.op_count;
402 lqr->lqr_pool.op_array[next] = src_pool->op_array[i];
410 up_write(&lov->lov_qos.lq_rw_sem);
412 if (placed != real_count) {
413 /* This should never happen */
414 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
415 "round-robin list (%d of %d).\n",
417 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
418 LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
419 lqr->lqr_pool.op_array[i]);
426 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
427 LCONSOLE(D_QOS, "rr #%d ost idx=%d\n", i,
428 lqr->lqr_pool.op_array[i]);
436 void qos_shrink_lsm(struct lov_request_set *set)
438 struct lov_stripe_md *lsm = set->set_oi->oi_md, *lsm_new;
439 /* XXX LOV STACKING call into osc for sizes */
440 unsigned oldsize, newsize;
442 if (set->set_oti && set->set_cookies && set->set_cookie_sent) {
443 struct llog_cookie *cookies;
444 oldsize = lsm->lsm_stripe_count * sizeof(*cookies);
445 newsize = set->set_count * sizeof(*cookies);
447 cookies = set->set_cookies;
448 oti_alloc_cookies(set->set_oti, set->set_count);
449 if (set->set_oti->oti_logcookies) {
450 memcpy(set->set_oti->oti_logcookies, cookies, newsize);
451 OBD_FREE(cookies, oldsize);
452 set->set_cookies = set->set_oti->oti_logcookies;
454 CWARN("'leaking' %d bytes\n", oldsize - newsize);
458 CWARN("using fewer stripes for object "LPU64": old %u new %u\n",
459 lsm->lsm_object_id, lsm->lsm_stripe_count, set->set_count);
460 LASSERT(lsm->lsm_stripe_count >= set->set_count);
462 newsize = lov_stripe_md_size(set->set_count);
463 OBD_ALLOC(lsm_new, newsize);
464 if (lsm_new != NULL) {
466 memcpy(lsm_new, lsm, sizeof(*lsm));
467 for (i = 0; i < lsm->lsm_stripe_count; i++) {
468 if (i < set->set_count) {
469 lsm_new->lsm_oinfo[i] = lsm->lsm_oinfo[i];
472 OBD_SLAB_FREE(lsm->lsm_oinfo[i], lov_oinfo_slab,
473 sizeof(struct lov_oinfo));
475 lsm_new->lsm_stripe_count = set->set_count;
476 OBD_FREE(lsm, sizeof(struct lov_stripe_md) +
477 lsm->lsm_stripe_count * sizeof(struct lov_oinfo *));
478 set->set_oi->oi_md = lsm_new;
480 CWARN("'leaking' few bytes\n");
484 int qos_remedy_create(struct lov_request_set *set, struct lov_request *req)
486 struct lov_stripe_md *lsm = set->set_oi->oi_md;
487 struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
488 unsigned ost_idx, ost_count = lov->desc.ld_tgt_count;
489 int stripe, i, rc = -EIO;
492 ost_idx = (req->rq_idx + lsm->lsm_stripe_count) % ost_count;
493 for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
494 if (!lov->lov_tgts[ost_idx] ||
495 !lov->lov_tgts[ost_idx]->ltd_active)
497 /* check if objects has been created on this ost */
498 for (stripe = 0; stripe < lsm->lsm_stripe_count; stripe++) {
499 /* we try send create to this ost but he is failed */
500 if (stripe == req->rq_stripe)
502 /* already have object at this stripe */
503 if (ost_idx == lsm->lsm_oinfo[stripe]->loi_ost_idx)
506 if (stripe >= lsm->lsm_stripe_count) {
507 req->rq_idx = ost_idx;
508 rc = obd_create(lov->lov_tgts[ost_idx]->ltd_exp,
509 req->rq_oi.oi_oa, &req->rq_oi.oi_md,
518 static int min_stripe_count(int stripe_cnt, int flags)
520 return (flags & LOV_USES_DEFAULT_STRIPE ?
521 stripe_cnt - (stripe_cnt / 4) : stripe_cnt);
524 #define LOV_CREATE_RESEED_MULT 4
525 #define LOV_CREATE_RESEED_MIN 1000
526 /* Allocate objects on osts with round-robin algorithm */
527 static int alloc_rr(struct lov_obd *lov, int *idx_arr, int *stripe_cnt,
528 char *poolname, int flags)
533 int ost_start_idx_temp;
535 int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
536 struct pool_desc *pool;
537 struct ost_pool *osts;
538 struct lov_qos_rr *lqr;
541 pool = lov_find_pool(lov, poolname);
543 osts = &(lov->lov_packed);
544 lqr = &(lov->lov_qos.lq_rr);
546 down_read(&pool_tgt_rw_sem(pool));
547 osts = &(pool->pool_obds);
548 lqr = &(pool->pool_rr);
551 rc = qos_calc_rr(lov, osts, lqr);
555 if (--lqr->lqr_start_count <= 0) {
556 lqr->lqr_start_idx = ll_rand() % osts->op_count;
557 lqr->lqr_start_count =
558 (LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) +
559 LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U);
560 } else if (stripe_cnt_min >= osts->op_count ||
561 lqr->lqr_start_idx > osts->op_count) {
562 /* If we have allocated from all of the OSTs, slowly
563 * precess the next start if the OST/stripe count isn't
564 * already doing this for us. */
565 lqr->lqr_start_idx %= osts->op_count;
566 if (*stripe_cnt > 1 && (osts->op_count % (*stripe_cnt)) != 1)
567 ++lqr->lqr_offset_idx;
569 down_read(&lov->lov_qos.lq_rw_sem);
570 ost_start_idx_temp = lqr->lqr_start_idx;
573 array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) % osts->op_count;
576 CDEBUG(D_QOS, "pool '%s' want %d startidx %d startcnt %d offset %d "
577 "active %d count %d arrayidx %d\n", poolname,
578 *stripe_cnt, lqr->lqr_start_idx, lqr->lqr_start_count,
579 lqr->lqr_offset_idx, osts->op_count, osts->op_count, array_idx);
582 for (i = 0; i < osts->op_count;
583 i++, array_idx=(array_idx + 1) % osts->op_count) {
584 ++lqr->lqr_start_idx;
585 ost_idx = lqr->lqr_pool.op_array[array_idx];
587 CDEBUG(D_QOS, "#%d strt %d act %d strp %d ary %d idx %d\n",
588 i, lqr->lqr_start_idx,
589 ((ost_idx != LOV_QOS_EMPTY) && lov->lov_tgts[ost_idx]) ?
590 lov->lov_tgts[ost_idx]->ltd_active : 0,
591 idx_pos - idx_arr, array_idx, ost_idx);
593 if ((ost_idx == LOV_QOS_EMPTY) || !lov->lov_tgts[ost_idx] ||
594 !lov->lov_tgts[ost_idx]->ltd_active)
597 /* Fail Check before osc_precreate() is called
598 so we can only 'fail' single OSC. */
599 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
602 /* Drop slow OSCs if we can */
603 if (obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed)
608 /* We have enough stripes */
609 if (idx_pos - idx_arr == *stripe_cnt)
612 if ((speed < 2) && (idx_pos - idx_arr < stripe_cnt_min)) {
613 /* Try again, allowing slower OSCs */
615 lqr->lqr_start_idx = ost_start_idx_temp;
619 up_read(&lov->lov_qos.lq_rw_sem);
621 *stripe_cnt = idx_pos - idx_arr;
624 up_read(&pool_tgt_rw_sem(pool));
625 /* put back ref got by lov_find_pool() */
626 lov_pool_putref(pool);
632 /* alloc objects on osts with specific stripe offset */
633 static int alloc_specific(struct lov_obd *lov, struct lov_stripe_md *lsm,
636 unsigned ost_idx, array_idx, ost_count;
639 struct pool_desc *pool;
640 struct ost_pool *osts;
643 pool = lov_find_pool(lov, lsm->lsm_pool_name);
645 osts = &(lov->lov_packed);
647 down_read(&pool_tgt_rw_sem(pool));
648 osts = &(pool->pool_obds);
651 ost_count = osts->op_count;
654 /* search loi_ost_idx in ost array */
656 for (i = 0; i < ost_count; i++) {
657 if (osts->op_array[i] == lsm->lsm_oinfo[0]->loi_ost_idx) {
662 if (i == ost_count) {
663 CERROR("Start index %d not found in pool '%s'\n",
664 lsm->lsm_oinfo[0]->loi_ost_idx, lsm->lsm_pool_name);
665 GOTO(out, rc = -EINVAL);
669 for (i = 0; i < ost_count;
670 i++, array_idx = (array_idx + 1) % ost_count) {
671 ost_idx = osts->op_array[array_idx];
673 if (!lov->lov_tgts[ost_idx] ||
674 !lov->lov_tgts[ost_idx]->ltd_active) {
678 /* Fail Check before osc_precreate() is called
679 so we can only 'fail' single OSC. */
680 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
683 /* Drop slow OSCs if we can, but not for requested start idx.
685 * This means "if OSC is slow and it is not the requested
686 * start OST, then it can be skipped, otherwise skip it only
687 * if it is inactive/recovering/out-of-space." */
688 if ((obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed) &&
689 (i != 0 || speed >= 2))
694 /* We have enough stripes */
695 if (idx_pos - idx_arr == lsm->lsm_stripe_count)
699 /* Try again, allowing slower OSCs */
704 /* If we were passed specific striping params, then a failure to
705 * meet those requirements is an error, since we can't reallocate
706 * that memory (it might be part of a larger array or something).
708 * We can only get here if lsm_stripe_count was originally > 1.
710 CERROR("can't lstripe objid "LPX64": have %d want %u\n",
711 lsm->lsm_object_id, (int)(idx_pos - idx_arr),
712 lsm->lsm_stripe_count);
716 up_read(&pool_tgt_rw_sem(pool));
717 /* put back ref got by lov_find_pool() */
718 lov_pool_putref(pool);
723 /* Alloc objects on osts with optimization based on:
725 - network resources (shared OSS's)
727 static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt,
728 char *poolname, int flags)
730 struct lov_obd *lov = &exp->exp_obd->u.lov;
731 __u64 total_weight = 0;
732 int nfound, good_osts, i, rc = 0;
733 int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
734 struct pool_desc *pool;
735 struct ost_pool *osts;
736 struct lov_qos_rr *lqr;
739 if (stripe_cnt_min < 1)
742 pool = lov_find_pool(lov, poolname);
744 osts = &(lov->lov_packed);
745 lqr = &(lov->lov_qos.lq_rr);
747 down_read(&pool_tgt_rw_sem(pool));
748 osts = &(pool->pool_obds);
749 lqr = &(pool->pool_rr);
752 obd_getref(exp->exp_obd);
753 /* wait for fresh statfs info if needed, the rpcs are sent in
755 qos_statfs_update(exp->exp_obd,
756 cfs_time_shift_64(-2 * lov->desc.ld_qos_maxage), 1);
758 down_write(&lov->lov_qos.lq_rw_sem);
760 if (lov->desc.ld_active_tgt_count < 2)
761 GOTO(out, rc = -EAGAIN);
763 rc = qos_calc_ppo(exp->exp_obd);
768 /* Find all the OSTs that are valid stripe candidates */
769 for (i = 0; i < osts->op_count; i++) {
770 if (!lov->lov_tgts[osts->op_array[i]] ||
771 !lov->lov_tgts[osts->op_array[i]]->ltd_active)
774 /* Fail Check before osc_precreate() is called
775 so we can only 'fail' single OSC. */
776 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && osts->op_array[i] == 0)
779 if (obd_precreate(lov->lov_tgts[osts->op_array[i]]->ltd_exp) > 2)
782 lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable = 1;
783 qos_calc_weight(lov, osts->op_array[i]);
784 total_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight;
790 CDEBUG(D_QOS, "found %d good osts\n", good_osts);
793 if (good_osts < stripe_cnt_min)
794 GOTO(out, rc = -EAGAIN);
796 /* We have enough osts */
797 if (good_osts < *stripe_cnt)
798 *stripe_cnt = good_osts;
800 /* Find enough OSTs with weighted random allocation. */
802 while (nfound < *stripe_cnt) {
803 __u64 rand, cur_weight;
809 #if BITS_PER_LONG == 32
810 rand = ll_rand() % (unsigned)total_weight;
811 /* If total_weight > 32-bit, first generate the high
812 * 32 bits of the random number, then add in the low
813 * 32 bits (truncated to the upper limit, if needed) */
814 if (total_weight > 0xffffffffULL)
815 rand = (__u64)(ll_rand() %
816 (unsigned)(total_weight >> 32)) << 32;
820 if (rand == (total_weight & 0xffffffff00000000ULL))
821 rand |= ll_rand() % (unsigned)total_weight;
826 rand = ((__u64)ll_rand() << 32 | ll_rand()) %
833 /* On average, this will hit larger-weighted osts more often.
834 0-weight osts will always get used last (only when rand=0).*/
835 for (i = 0; i < osts->op_count; i++) {
836 if (!lov->lov_tgts[osts->op_array[i]] ||
837 !lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable)
840 cur_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight;
842 CDEBUG(D_QOS, "stripe_cnt=%d nfound=%d cur_weight="LPU64
843 " rand="LPU64" total_weight="LPU64"\n",
844 *stripe_cnt, nfound, cur_weight, rand, total_weight);
846 if (cur_weight >= rand) {
848 CDEBUG(D_QOS, "assigned stripe=%d to idx=%d\n",
849 nfound, osts->op_array[i]);
851 idx_arr[nfound++] = osts->op_array[i];
852 qos_used(lov, osts, osts->op_array[i], &total_weight);
858 CDEBUG(D_QOS, "Didn't find any OSTs? Reduce total weight\n");
859 if (total_weight == 0)
866 LASSERT(nfound == *stripe_cnt);
869 up_write(&lov->lov_qos.lq_rw_sem);
872 up_read(&pool_tgt_rw_sem(pool));
873 /* put back ref got by lov_find_pool() */
874 lov_pool_putref(pool);
878 rc = alloc_rr(lov, idx_arr, stripe_cnt, poolname, flags);
880 obd_putref(exp->exp_obd);
884 /* return new alloced stripe count on success */
885 static int alloc_idx_array(struct obd_export *exp, struct lov_stripe_md *lsm,
886 int newea, int **idx_arr, int *arr_cnt, int flags)
888 struct lov_obd *lov = &exp->exp_obd->u.lov;
889 int stripe_cnt = lsm->lsm_stripe_count;
894 *arr_cnt = stripe_cnt;
895 OBD_ALLOC(tmp_arr, *arr_cnt * sizeof(int));
898 for (i = 0; i < *arr_cnt; i++)
902 lsm->lsm_oinfo[0]->loi_ost_idx >= lov->desc.ld_tgt_count)
903 rc = alloc_qos(exp, tmp_arr, &stripe_cnt,
904 lsm->lsm_pool_name, flags);
906 rc = alloc_specific(lov, lsm, tmp_arr);
914 OBD_FREE(tmp_arr, *arr_cnt * sizeof(int));
919 static void free_idx_array(int *idx_arr, int arr_cnt)
922 OBD_FREE(idx_arr, arr_cnt * sizeof(int));
925 int qos_prep_create(struct obd_export *exp, struct lov_request_set *set)
927 struct lov_obd *lov = &exp->exp_obd->u.lov;
928 struct lov_stripe_md *lsm;
929 struct obdo *src_oa = set->set_oi->oi_oa;
930 struct obd_trans_info *oti = set->set_oti;
931 int i, stripes, rc = 0, newea = 0;
932 int flag = LOV_USES_ASSIGNED_STRIPE;
933 int *idx_arr = NULL, idx_cnt = 0;
936 LASSERT(src_oa->o_valid & OBD_MD_FLID);
938 if (set->set_oi->oi_md == NULL) {
939 int stripes_def = lov_get_stripecnt(lov, 0);
941 /* If the MDS file was truncated up to some size, stripe over
942 * enough OSTs to allow the file to be created at that size.
943 * This may mean we use more than the default # of stripes. */
944 if (src_oa->o_valid & OBD_MD_FLSIZE) {
945 obd_size min_bavail = LUSTRE_STRIPE_MAXBYTES;
947 /* Find a small number of stripes we can use
948 (up to # of active osts). */
950 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
951 if (!lov->lov_tgts[i] ||
952 !lov->lov_tgts[i]->ltd_active)
954 min_bavail = min(min_bavail, TGT_BAVAIL(i));
955 if (min_bavail * stripes > src_oa->o_size)
960 if (stripes < stripes_def)
961 stripes = stripes_def;
963 flag = LOV_USES_DEFAULT_STRIPE;
964 stripes = stripes_def;
967 rc = lov_alloc_memmd(&set->set_oi->oi_md, stripes,
968 lov->desc.ld_pattern ?
969 lov->desc.ld_pattern : LOV_PATTERN_RAID0,
977 lsm = set->set_oi->oi_md;
978 lsm->lsm_object_id = src_oa->o_id;
979 lsm->lsm_object_gr = src_oa->o_gr;
981 if (!lsm->lsm_stripe_size)
982 lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size;
983 if (!lsm->lsm_pattern) {
984 LASSERT(lov->desc.ld_pattern);
985 lsm->lsm_pattern = lov->desc.ld_pattern;
988 stripes = alloc_idx_array(exp, lsm, newea, &idx_arr, &idx_cnt, flag);
990 GOTO(out_err, rc = stripes ? stripes : -EIO);
991 LASSERTF(stripes <= lsm->lsm_stripe_count,"requested %d allocated %d\n",
992 lsm->lsm_stripe_count, stripes);
994 for (i = 0; i < stripes; i++) {
995 struct lov_request *req;
996 int ost_idx = idx_arr[i];
997 LASSERT(ost_idx >= 0);
999 OBD_ALLOC(req, sizeof(*req));
1001 GOTO(out_err, rc = -ENOMEM);
1002 lov_set_add_req(req, set);
1004 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
1005 OBD_ALLOC(req->rq_oi.oi_md, req->rq_buflen);
1006 if (req->rq_oi.oi_md == NULL)
1007 GOTO(out_err, rc = -ENOMEM);
1009 OBDO_ALLOC(req->rq_oi.oi_oa);
1010 if (req->rq_oi.oi_oa == NULL)
1011 GOTO(out_err, rc = -ENOMEM);
1013 req->rq_idx = ost_idx;
1015 /* create data objects with "parent" OA */
1016 memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
1017 req->rq_oi.oi_cb_up = cb_create_update;
1019 /* XXX When we start creating objects on demand, we need to
1020 * make sure that we always create the object on the
1021 * stripe which holds the existing file size.
1023 if (src_oa->o_valid & OBD_MD_FLSIZE) {
1024 req->rq_oi.oi_oa->o_size =
1025 lov_size_to_stripe(lsm, src_oa->o_size, i);
1027 CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
1028 i, req->rq_oi.oi_oa->o_size, src_oa->o_size);
1031 LASSERT(set->set_count == stripes);
1033 if (stripes < lsm->lsm_stripe_count)
1034 qos_shrink_lsm(set);
1035 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_LOV_PREP_CREATE)) {
1036 qos_shrink_lsm(set);
1040 if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) {
1041 oti_alloc_cookies(oti, set->set_count);
1042 if (!oti->oti_logcookies)
1043 GOTO(out_err, rc = -ENOMEM);
1044 set->set_cookies = oti->oti_logcookies;
1048 obd_free_memmd(exp, &set->set_oi->oi_md);
1050 free_idx_array(idx_arr, idx_cnt);
1055 void qos_update(struct lov_obd *lov)
1058 lov->lov_qos.lq_dirty = 1;
1061 void qos_statfs_done(struct lov_obd *lov)
1063 LASSERT(lov->lov_qos.lq_statfs_in_progress);
1064 down_write(&lov->lov_qos.lq_rw_sem);
1065 lov->lov_qos.lq_statfs_in_progress = 0;
1066 /* wake up any threads waiting for the statfs rpcs to complete */
1067 cfs_waitq_signal(&lov->lov_qos.lq_statfs_waitq);
1068 up_write(&lov->lov_qos.lq_rw_sem);
1071 static int qos_statfs_ready(struct obd_device *obd, __u64 max_age)
1073 struct lov_obd *lov = &obd->u.lov;
1076 down_read(&lov->lov_qos.lq_rw_sem);
1077 rc = lov->lov_qos.lq_statfs_in_progress == 0 ||
1078 cfs_time_beforeq_64(max_age, obd->obd_osfs_age);
1079 up_read(&lov->lov_qos.lq_rw_sem);
1084 * Update statfs data if the current osfs age is older than max_age.
1085 * If wait is not set, it means that we are called from lov_create()
1086 * and we should just issue the rpcs without waiting for them to complete.
1087 * If wait is set, we are called from alloc_qos() and we just have
1088 * to wait for the request set to complete.
1090 void qos_statfs_update(struct obd_device *obd, __u64 max_age, int wait)
1092 struct lov_obd *lov = &obd->u.lov;
1093 struct obd_info *oinfo;
1095 struct ptlrpc_request_set *set = NULL;
1098 if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
1099 /* statfs data are quite recent, don't need to refresh it */
1102 if (!wait && lov->lov_qos.lq_statfs_in_progress)
1103 /* statfs already in progress */
1106 down_write(&lov->lov_qos.lq_rw_sem);
1107 if (lov->lov_qos.lq_statfs_in_progress) {
1108 up_write(&lov->lov_qos.lq_rw_sem);
1111 /* no statfs in flight, send rpcs */
1112 lov->lov_qos.lq_statfs_in_progress = 1;
1113 up_write(&lov->lov_qos.lq_rw_sem);
1116 CDEBUG(D_QOS, "%s: did not manage to get fresh statfs data "
1117 "in a timely manner (osfs age "LPU64", max age "LPU64")"
1118 ", sending new statfs rpcs\n",
1119 obd_uuid2str(&lov->desc.ld_uuid), obd->obd_osfs_age,
1122 /* need to send statfs rpcs */
1123 CDEBUG(D_QOS, "sending new statfs requests\n");
1124 memset(lov->lov_qos.lq_statfs_data, 0,
1125 sizeof(*lov->lov_qos.lq_statfs_data));
1126 oinfo = &lov->lov_qos.lq_statfs_data->lsd_oi;
1127 oinfo->oi_osfs = &lov->lov_qos.lq_statfs_data->lsd_statfs;
1128 oinfo->oi_flags = OBD_STATFS_NODELAY;
1129 set = ptlrpc_prep_set();
1131 GOTO(out_failed, rc = -ENOMEM);
1133 rc = obd_statfs_async(obd, oinfo, max_age, set);
1134 if (rc || list_empty(&set->set_requests)) {
1136 CWARN("statfs failed with %d\n", rc);
1137 GOTO(out_failed, rc);
1139 /* send requests via ptlrpcd */
1140 oinfo->oi_flags |= OBD_STATFS_PTLRPCD;
1141 ptlrpcd_add_rqset(set);
1145 down_write(&lov->lov_qos.lq_rw_sem);
1146 lov->lov_qos.lq_statfs_in_progress = 0;
1147 /* wake up any threads waiting for the statfs rpcs to complete */
1148 cfs_waitq_signal(&lov->lov_qos.lq_statfs_waitq);
1149 up_write(&lov->lov_qos.lq_rw_sem);
1153 ptlrpc_set_destroy(set);
1155 struct l_wait_info lwi = { 0 };
1156 CDEBUG(D_QOS, "waiting for statfs requests to complete\n");
1157 l_wait_event(lov->lov_qos.lq_statfs_waitq,
1158 qos_statfs_ready(obd, max_age), &lwi);
1159 if (cfs_time_before_64(obd->obd_osfs_age, max_age))
1160 CDEBUG(D_QOS, "%s: still no fresh statfs data after "
1161 "waiting (osfs age "LPU64", max age "
1163 obd_uuid2str(&lov->desc.ld_uuid),
1164 obd->obd_osfs_age, max_age);