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 "lov_internal.h"
52 /* #define QOS_DEBUG 1 */
55 #define TGT_BAVAIL(i) (lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bavail *\
56 lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bsize)
57 #define TGT_FFREE(i) (lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_ffree)
60 int qos_add_tgt(struct obd_device *obd, __u32 index)
62 struct lov_obd *lov = &obd->u.lov;
63 struct lov_qos_oss *oss, *temposs;
64 struct obd_export *exp = lov->lov_tgts[index]->ltd_exp;
65 int rc = 0, found = 0;
68 /* We only need this QOS struct on MDT, not clients - but we may not
69 * have registered the LOV's observer yet, so there's no way to know */
70 if (!exp || !exp->exp_connection) {
71 CERROR("Missing connection\n");
75 down_write(&lov->lov_qos.lq_rw_sem);
76 mutex_down(&lov->lov_lock);
77 list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
78 if (obd_uuid_equals(&oss->lqo_uuid,
79 &exp->exp_connection->c_remote_uuid)) {
88 GOTO(out, rc = -ENOMEM);
89 memcpy(&oss->lqo_uuid,
90 &exp->exp_connection->c_remote_uuid,
91 sizeof(oss->lqo_uuid));
93 /* Assume we have to move this one */
94 list_del(&oss->lqo_oss_list);
98 lov->lov_tgts[index]->ltd_qos.ltq_oss = oss;
100 /* Add sorted by # of OSTs. Find the first entry that we're
102 list_for_each_entry(temposs, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
103 if (oss->lqo_ost_count > temposs->lqo_ost_count)
106 /* ...and add before it. If we're the first or smallest, temposs
107 points to the list head, and we add to the end. */
108 list_add_tail(&oss->lqo_oss_list, &temposs->lqo_oss_list);
110 lov->lov_qos.lq_dirty = 1;
111 lov->lov_qos.lq_rr.lqr_dirty = 1;
113 CDEBUG(D_QOS, "add tgt %s to OSS %s (%d OSTs)\n",
114 obd_uuid2str(&lov->lov_tgts[index]->ltd_uuid),
115 obd_uuid2str(&oss->lqo_uuid),
119 mutex_up(&lov->lov_lock);
120 up_write(&lov->lov_qos.lq_rw_sem);
124 int qos_del_tgt(struct obd_device *obd, struct lov_tgt_desc *tgt)
126 struct lov_obd *lov = &obd->u.lov;
127 struct lov_qos_oss *oss;
131 down_write(&lov->lov_qos.lq_rw_sem);
133 oss = tgt->ltd_qos.ltq_oss;
135 GOTO(out, rc = -ENOENT);
137 oss->lqo_ost_count--;
138 if (oss->lqo_ost_count == 0) {
139 CDEBUG(D_QOS, "removing OSS %s\n",
140 obd_uuid2str(&oss->lqo_uuid));
141 list_del(&oss->lqo_oss_list);
145 lov->lov_qos.lq_dirty = 1;
146 lov->lov_qos.lq_rr.lqr_dirty = 1;
148 up_write(&lov->lov_qos.lq_rw_sem);
152 /* Recalculate per-object penalties for OSSs and OSTs,
153 depends on size of each ost in an oss */
154 static int qos_calc_ppo(struct obd_device *obd)
156 struct lov_obd *lov = &obd->u.lov;
157 struct lov_qos_oss *oss;
158 __u64 ba_max, ba_min, temp;
160 int rc, i, prio_wide;
164 if (!lov->lov_qos.lq_dirty)
167 num_active = lov->desc.ld_active_tgt_count - 1;
169 GOTO(out, rc = -EAGAIN);
171 /* find bavail on each OSS */
172 list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
175 lov->lov_qos.lq_active_oss_count = 0;
177 /* How badly user wants to select osts "widely" (not recently chosen
178 and not on recent oss's). As opposed to "freely" (free space
180 prio_wide = 256 - lov->lov_qos.lq_prio_free;
182 ba_min = (__u64)(-1);
184 now = cfs_time_current_sec();
185 /* Calculate OST penalty per object */
186 /* (lov ref taken in alloc_qos) */
187 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
188 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
190 temp = TGT_BAVAIL(i);
193 ba_min = min(temp, ba_min);
194 ba_max = max(temp, ba_max);
196 /* Count the number of usable OSS's */
197 if (lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_bavail == 0)
198 lov->lov_qos.lq_active_oss_count++;
199 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_bavail += temp;
201 /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
203 do_div(temp, num_active);
204 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj =
205 (temp * prio_wide) >> 8;
207 age = (now - lov->lov_tgts[i]->ltd_qos.ltq_used) >> 3;
208 if (lov->lov_qos.lq_reset || age > 32 * lov->desc.ld_qos_maxage)
209 lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
210 else if (age > lov->desc.ld_qos_maxage)
211 /* Decay the penalty by half for every 8x the update
212 * interval that the device has been idle. That gives
213 * lots of time for the statfs information to be
214 * updated (which the penalty is only a proxy for),
215 * and avoids penalizing OSS/OSTs under light load. */
216 lov->lov_tgts[i]->ltd_qos.ltq_penalty >>=
217 (age / lov->desc.ld_qos_maxage);
220 num_active = lov->lov_qos.lq_active_oss_count - 1;
221 if (num_active < 1) {
222 /* If there's only 1 OSS, we can't penalize it, so instead
223 we have to double the OST penalty */
225 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
226 if (lov->lov_tgts[i] == NULL)
228 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj <<= 1;
232 /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
233 list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
234 temp = oss->lqo_bavail >> 1;
235 do_div(temp, oss->lqo_ost_count * num_active);
236 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
238 age = (now - oss->lqo_used) >> 3;
239 if (lov->lov_qos.lq_reset || age > 32 * lov->desc.ld_qos_maxage)
240 oss->lqo_penalty = 0;
241 else if (age > lov->desc.ld_qos_maxage)
242 /* Decay the penalty by half for every 8x the update
243 * interval that the device has been idle. That gives
244 * lots of time for the statfs information to be
245 * updated (which the penalty is only a proxy for),
246 * and avoids penalizing OSS/OSTs under light load. */
247 oss->lqo_penalty >>= (age / lov->desc.ld_qos_maxage);
250 lov->lov_qos.lq_dirty = 0;
251 lov->lov_qos.lq_reset = 0;
253 /* If each ost has almost same free space,
254 * do rr allocation for better creation performance */
255 lov->lov_qos.lq_same_space = 0;
256 temp = ba_max - ba_min;
257 ba_min = (ba_min * 51) >> 8; /* 51/256 = .20 */
259 /* Difference is less than 20% */
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 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 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 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 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 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 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(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(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(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");
499 int qos_remedy_create(struct lov_request_set *set, struct lov_request *req)
501 struct lov_stripe_md *lsm = set->set_oi->oi_md;
502 struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
503 unsigned ost_idx, ost_count = lov->desc.ld_tgt_count;
504 int stripe, i, rc = -EIO;
507 ost_idx = (req->rq_idx + lsm->lsm_stripe_count) % ost_count;
508 for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
509 if (!lov->lov_tgts[ost_idx] ||
510 !lov->lov_tgts[ost_idx]->ltd_active)
512 /* check if objects has been created on this ost */
513 for (stripe = 0; stripe < lsm->lsm_stripe_count; stripe++) {
514 if (stripe == req->rq_stripe)
516 if (ost_idx == lsm->lsm_oinfo[stripe]->loi_ost_idx)
520 if (stripe >= lsm->lsm_stripe_count) {
521 req->rq_idx = ost_idx;
522 rc = obd_create(lov->lov_tgts[ost_idx]->ltd_exp,
523 req->rq_oi.oi_oa, &req->rq_oi.oi_md,
532 static int min_stripe_count(int stripe_cnt, int flags)
534 return (flags & LOV_USES_DEFAULT_STRIPE ?
535 stripe_cnt - (stripe_cnt / 4) : stripe_cnt);
538 #define LOV_CREATE_RESEED_MULT 4
539 #define LOV_CREATE_RESEED_MIN 1000
540 /* Allocate objects on osts with round-robin algorithm */
541 static int alloc_rr(struct lov_obd *lov, int *idx_arr, int *stripe_cnt,
542 char *poolname, int flags)
547 int ost_start_idx_temp;
549 int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
550 struct pool_desc *pool;
551 struct ost_pool *osts;
552 struct lov_qos_rr *lqr;
555 pool = lov_find_pool(lov, poolname);
557 osts = &(lov->lov_packed);
558 lqr = &(lov->lov_qos.lq_rr);
560 down_read(&pool_tgt_rw_sem(pool));
561 osts = &(pool->pool_obds);
562 lqr = &(pool->pool_rr);
565 rc = qos_calc_rr(lov, osts, lqr);
569 if (--lqr->lqr_start_count <= 0) {
570 lqr->lqr_start_idx = ll_rand() % osts->op_count;
571 lqr->lqr_start_count =
572 (LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) +
573 LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U);
574 } else if (stripe_cnt_min >= osts->op_count ||
575 lqr->lqr_start_idx > osts->op_count) {
576 /* If we have allocated from all of the OSTs, slowly
577 * precess the next start if the OST/stripe count isn't
578 * already doing this for us. */
579 lqr->lqr_start_idx %= osts->op_count;
580 if (*stripe_cnt > 1 && (osts->op_count % (*stripe_cnt)) != 1)
581 ++lqr->lqr_offset_idx;
583 down_read(&lov->lov_qos.lq_rw_sem);
584 ost_start_idx_temp = lqr->lqr_start_idx;
587 array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) % osts->op_count;
590 CDEBUG(D_QOS, "pool '%s' want %d startidx %d startcnt %d offset %d "
591 "active %d count %d arrayidx %d\n", poolname,
592 *stripe_cnt, lqr->lqr_start_idx, lqr->lqr_start_count,
593 lqr->lqr_offset_idx, osts->op_count, osts->op_count, array_idx);
596 for (i = 0; i < osts->op_count;
597 i++, array_idx=(array_idx + 1) % osts->op_count) {
598 ++lqr->lqr_start_idx;
599 ost_idx = lqr->lqr_pool.op_array[array_idx];
601 CDEBUG(D_QOS, "#%d strt %d act %d strp %d ary %d idx %d\n",
602 i, lqr->lqr_start_idx,
603 ((ost_idx != LOV_QOS_EMPTY) && lov->lov_tgts[ost_idx]) ?
604 lov->lov_tgts[ost_idx]->ltd_active : 0,
605 idx_pos - idx_arr, array_idx, ost_idx);
607 if ((ost_idx == LOV_QOS_EMPTY) || !lov->lov_tgts[ost_idx] ||
608 !lov->lov_tgts[ost_idx]->ltd_active)
611 /* Fail Check before osc_precreate() is called
612 so we can only 'fail' single OSC. */
613 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
616 /* Drop slow OSCs if we can */
617 if (obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed)
622 /* We have enough stripes */
623 if (idx_pos - idx_arr == *stripe_cnt)
626 if ((speed < 2) && (idx_pos - idx_arr < stripe_cnt_min)) {
627 /* Try again, allowing slower OSCs */
629 lqr->lqr_start_idx = ost_start_idx_temp;
633 up_read(&lov->lov_qos.lq_rw_sem);
635 *stripe_cnt = idx_pos - idx_arr;
638 up_read(&pool_tgt_rw_sem(pool));
639 /* put back ref got by lov_find_pool() */
640 lh_put(lov->lov_pools_hash_body, &pool->pool_hash);
646 /* alloc objects on osts with specific stripe offset */
647 static int alloc_specific(struct lov_obd *lov, struct lov_stripe_md *lsm,
650 unsigned ost_idx, array_idx, ost_count;
653 struct pool_desc *pool;
654 struct ost_pool *osts;
657 pool = lov_find_pool(lov, lsm->lsm_pool_name);
659 osts = &(lov->lov_packed);
661 down_read(&pool_tgt_rw_sem(pool));
662 osts = &(pool->pool_obds);
665 ost_count = osts->op_count;
668 /* search loi_ost_idx in ost array */
670 for (i = 0; i < ost_count; i++) {
671 if (osts->op_array[i] == lsm->lsm_oinfo[0]->loi_ost_idx) {
676 if (i == ost_count) {
677 CERROR("Start index %d not found in pool '%s'\n",
678 lsm->lsm_oinfo[0]->loi_ost_idx, lsm->lsm_pool_name);
679 GOTO(out, rc = -EINVAL);
683 for (i = 0; i < ost_count;
684 i++, array_idx = (array_idx + 1) % ost_count) {
685 ost_idx = osts->op_array[array_idx];
687 if (!lov->lov_tgts[ost_idx] ||
688 !lov->lov_tgts[ost_idx]->ltd_active) {
692 /* Fail Check before osc_precreate() is called
693 so we can only 'fail' single OSC. */
694 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
697 /* Drop slow OSCs if we can, but not for requested start idx.
699 * This means "if OSC is slow and it is not the requested
700 * start OST, then it can be skipped, otherwise skip it only
701 * if it is inactive/recovering/out-of-space." */
702 if ((obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed) &&
703 (i != 0 || speed >= 2))
708 /* We have enough stripes */
709 if (idx_pos - idx_arr == lsm->lsm_stripe_count)
713 /* Try again, allowing slower OSCs */
718 /* If we were passed specific striping params, then a failure to
719 * meet those requirements is an error, since we can't reallocate
720 * that memory (it might be part of a larger array or something).
722 * We can only get here if lsm_stripe_count was originally > 1.
724 CERROR("can't lstripe objid "LPX64": have %d want %u\n",
725 lsm->lsm_object_id, (int)(idx_pos - idx_arr),
726 lsm->lsm_stripe_count);
730 up_read(&pool_tgt_rw_sem(pool));
731 /* put back ref got by lov_find_pool() */
732 lh_put(lov->lov_pools_hash_body, &pool->pool_hash);
738 /* Alloc objects on osts with optimization based on:
740 - network resources (shared OSS's)
742 static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt,
743 char *poolname, int flags)
745 struct lov_obd *lov = &exp->exp_obd->u.lov;
746 static time_t last_warn = 0;
747 time_t now = cfs_time_current_sec();
748 __u64 total_bavail, total_weight = 0;
749 int nfound, good_osts, i, warn = 0, rc = 0;
750 int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
751 struct pool_desc *pool;
752 struct ost_pool *osts;
753 struct lov_qos_rr *lqr;
756 if (stripe_cnt_min < 1)
759 pool = lov_find_pool(lov, poolname);
761 osts = &(lov->lov_packed);
762 lqr = &(lov->lov_qos.lq_rr);
764 down_read(&pool_tgt_rw_sem(pool));
765 osts = &(pool->pool_obds);
766 lqr = &(pool->pool_rr);
769 lov_getref(exp->exp_obd);
771 /* Detect -EAGAIN early, before expensive lock is taken. */
772 if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space)
773 GOTO(out_nolock, rc = -EAGAIN);
775 /* Do actual allocation, use write lock here. */
776 down_write(&lov->lov_qos.lq_rw_sem);
779 * Check again, while we were sleeping on @lq_rw_sem things could
782 if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space)
783 GOTO(out, rc = -EAGAIN);
785 if (lov->desc.ld_active_tgt_count < 2)
786 GOTO(out, rc = -EAGAIN);
788 rc = qos_calc_ppo(exp->exp_obd);
794 /* Warn users about zero available space/inode every 30 min */
795 if (cfs_time_sub(now, last_warn) > 60 * 30)
797 /* Find all the OSTs that are valid stripe candidates */
798 for (i = 0; i < osts->op_count; i++) {
801 if (!lov->lov_tgts[osts->op_array[i]] ||
802 !lov->lov_tgts[osts->op_array[i]]->ltd_active)
804 bavail = TGT_BAVAIL(osts->op_array[i]);
807 CDEBUG(D_QOS, "no free space on %s\n",
808 obd_uuid2str(&lov->lov_tgts[osts->op_array[i]]->ltd_uuid));
813 if (!TGT_FFREE(osts->op_array[i])) {
815 CDEBUG(D_QOS, "no free inodes on %s\n",
816 obd_uuid2str(&lov->lov_tgts[osts->op_array[i]]->ltd_uuid));
822 /* Fail Check before osc_precreate() is called
823 so we can only 'fail' single OSC. */
824 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && osts->op_array[i] == 0)
827 if (obd_precreate(lov->lov_tgts[osts->op_array[i]]->ltd_exp) > 2)
830 lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable = 1;
831 qos_calc_weight(lov, osts->op_array[i]);
832 total_bavail += bavail;
833 total_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight;
839 CDEBUG(D_QOS, "found %d good osts\n", good_osts);
842 if (good_osts < stripe_cnt_min)
843 GOTO(out, rc = -EAGAIN);
846 GOTO(out, rc = -ENOSPC);
848 /* We have enough osts */
849 if (good_osts < *stripe_cnt)
850 *stripe_cnt = good_osts;
853 GOTO(out, rc = -EAGAIN);
855 /* Find enough OSTs with weighted random allocation. */
857 while (nfound < *stripe_cnt) {
858 __u64 rand, cur_weight;
864 #if BITS_PER_LONG == 32
865 rand = ll_rand() % (unsigned)total_weight;
866 /* If total_weight > 32-bit, first generate the high
867 * 32 bits of the random number, then add in the low
868 * 32 bits (truncated to the upper limit, if needed) */
869 if (total_weight > 0xffffffffULL)
870 rand = (__u64)(ll_rand() %
871 (unsigned)(total_weight >> 32)) << 32;
875 if (rand == (total_weight & 0xffffffff00000000ULL))
876 rand |= ll_rand() % (unsigned)total_weight;
881 rand = ((__u64)ll_rand() << 32 | ll_rand()) %
888 /* On average, this will hit larger-weighted osts more often.
889 0-weight osts will always get used last (only when rand=0).*/
890 for (i = 0; i < osts->op_count; i++) {
891 if (!lov->lov_tgts[osts->op_array[i]] ||
892 !lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable)
895 cur_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight;
897 CDEBUG(D_QOS, "stripe_cnt=%d nfound=%d cur_weight="LPU64
898 " rand="LPU64" total_weight="LPU64"\n",
899 *stripe_cnt, nfound, cur_weight, rand, total_weight);
901 if (cur_weight >= rand) {
903 CDEBUG(D_QOS, "assigned stripe=%d to idx=%d\n",
904 nfound, osts->op_array[i]);
906 idx_arr[nfound++] = osts->op_array[i];
907 qos_used(lov, osts, osts->op_array[i], &total_weight);
912 /* should never satisfy below condition */
914 CERROR("Didn't find any OSTs?\n");
918 LASSERT(nfound == *stripe_cnt);
921 up_write(&lov->lov_qos.lq_rw_sem);
925 up_read(&pool_tgt_rw_sem(pool));
926 /* put back ref got by lov_find_pool() */
927 lh_put(lov->lov_pools_hash_body, &pool->pool_hash);
931 rc = alloc_rr(lov, idx_arr, stripe_cnt, poolname, flags);
933 lov_putref(exp->exp_obd);
937 /* return new alloced stripe count on success */
938 static int alloc_idx_array(struct obd_export *exp, struct lov_stripe_md *lsm,
939 int newea, int **idx_arr, int *arr_cnt, int flags)
941 struct lov_obd *lov = &exp->exp_obd->u.lov;
942 int stripe_cnt = lsm->lsm_stripe_count;
947 *arr_cnt = stripe_cnt;
948 OBD_ALLOC(tmp_arr, *arr_cnt * sizeof(int));
951 for (i = 0; i < *arr_cnt; i++)
955 lsm->lsm_oinfo[0]->loi_ost_idx >= lov->desc.ld_tgt_count)
956 rc = alloc_qos(exp, tmp_arr, &stripe_cnt,
957 lsm->lsm_pool_name, flags);
959 rc = alloc_specific(lov, lsm, tmp_arr);
967 OBD_FREE(tmp_arr, *arr_cnt * sizeof(int));
972 static void free_idx_array(int *idx_arr, int arr_cnt)
975 OBD_FREE(idx_arr, arr_cnt * sizeof(int));
978 int qos_prep_create(struct obd_export *exp, struct lov_request_set *set)
980 struct lov_obd *lov = &exp->exp_obd->u.lov;
981 struct lov_stripe_md *lsm;
982 struct obdo *src_oa = set->set_oi->oi_oa;
983 struct obd_trans_info *oti = set->set_oti;
984 int i, stripes, rc = 0, newea = 0;
985 int flag = LOV_USES_ASSIGNED_STRIPE;
986 int *idx_arr = NULL, idx_cnt = 0;
989 LASSERT(src_oa->o_valid & OBD_MD_FLID);
990 LASSERT(src_oa->o_valid & OBD_MD_FLGROUP);
992 if (set->set_oi->oi_md == NULL) {
993 int stripes_def = lov_get_stripecnt(lov, 0);
995 /* If the MDS file was truncated up to some size, stripe over
996 * enough OSTs to allow the file to be created at that size.
997 * This may mean we use more than the default # of stripes. */
998 if (src_oa->o_valid & OBD_MD_FLSIZE) {
999 obd_size min_bavail = LUSTRE_STRIPE_MAXBYTES;
1001 /* Find a small number of stripes we can use
1002 (up to # of active osts). */
1004 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1005 if (!lov->lov_tgts[i] ||
1006 !lov->lov_tgts[i]->ltd_active)
1008 min_bavail = min(min_bavail, TGT_BAVAIL(i));
1009 if (min_bavail * stripes > src_oa->o_size)
1014 if (stripes < stripes_def)
1015 stripes = stripes_def;
1017 flag = LOV_USES_DEFAULT_STRIPE;
1018 stripes = stripes_def;
1021 rc = lov_alloc_memmd(&set->set_oi->oi_md, stripes,
1022 lov->desc.ld_pattern ?
1023 lov->desc.ld_pattern : LOV_PATTERN_RAID0,
1031 lsm = set->set_oi->oi_md;
1032 lsm->lsm_object_id = src_oa->o_id;
1033 lsm->lsm_object_gr = src_oa->o_gr;
1035 if (!lsm->lsm_stripe_size)
1036 lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size;
1037 if (!lsm->lsm_pattern) {
1038 LASSERT(lov->desc.ld_pattern);
1039 lsm->lsm_pattern = lov->desc.ld_pattern;
1042 stripes = alloc_idx_array(exp, lsm, newea, &idx_arr, &idx_cnt, flag);
1044 GOTO(out_err, rc = stripes ? stripes : -EIO);
1045 LASSERTF(stripes <= lsm->lsm_stripe_count,"requested %d allocated %d\n",
1046 lsm->lsm_stripe_count, stripes);
1048 for (i = 0; i < stripes; i++) {
1049 struct lov_request *req;
1050 int ost_idx = idx_arr[i];
1051 LASSERT(ost_idx >= 0);
1053 OBD_ALLOC(req, sizeof(*req));
1055 GOTO(out_err, rc = -ENOMEM);
1056 lov_set_add_req(req, set);
1058 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
1059 OBD_ALLOC(req->rq_oi.oi_md, req->rq_buflen);
1060 if (req->rq_oi.oi_md == NULL)
1061 GOTO(out_err, rc = -ENOMEM);
1063 OBDO_ALLOC(req->rq_oi.oi_oa);
1064 if (req->rq_oi.oi_oa == NULL)
1065 GOTO(out_err, rc = -ENOMEM);
1067 req->rq_idx = ost_idx;
1069 /* create data objects with "parent" OA */
1070 memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
1072 /* XXX When we start creating objects on demand, we need to
1073 * make sure that we always create the object on the
1074 * stripe which holds the existing file size.
1076 if (src_oa->o_valid & OBD_MD_FLSIZE) {
1077 req->rq_oi.oi_oa->o_size =
1078 lov_size_to_stripe(lsm, src_oa->o_size, i);
1080 CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
1081 i, req->rq_oi.oi_oa->o_size, src_oa->o_size);
1084 LASSERT(set->set_count == stripes);
1086 if (stripes < lsm->lsm_stripe_count)
1087 qos_shrink_lsm(set);
1089 if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) {
1090 oti_alloc_cookies(oti, set->set_count);
1091 if (!oti->oti_logcookies)
1092 GOTO(out_err, rc = -ENOMEM);
1093 set->set_cookies = oti->oti_logcookies;
1097 obd_free_memmd(exp, &set->set_oi->oi_md);
1099 free_idx_array(idx_arr, idx_cnt);
1104 void qos_update(struct lov_obd *lov)
1107 lov->lov_qos.lq_dirty = 1;