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.
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)
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 cfs_down_write(&lov->lov_qos.lq_rw_sem);
76 cfs_mutex_down(&lov->lov_lock);
77 cfs_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 cfs_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 cfs_list_for_each_entry(temposs, &lov->lov_qos.lq_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 cfs_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 cfs_mutex_up(&lov->lov_lock);
121 cfs_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 cfs_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 cfs_list_del(&oss->lqo_oss_list);
146 lov->lov_qos.lq_dirty = 1;
147 lov->lov_qos.lq_rr.lqr_dirty = 1;
149 cfs_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 cfs_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] == NULL)
229 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj <<= 1;
233 /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
234 cfs_list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
235 temp = oss->lqo_bavail >> 1;
236 do_div(temp, oss->lqo_ost_count * num_active);
237 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
239 age = (now - oss->lqo_used) >> 3;
240 if (lov->lov_qos.lq_reset || age > 32 * lov->desc.ld_qos_maxage)
241 oss->lqo_penalty = 0;
242 else if (age > lov->desc.ld_qos_maxage)
243 /* Decay the penalty by half for every 8x the update
244 * interval that the device has been idle. That gives
245 * lots of time for the statfs information to be
246 * updated (which the penalty is only a proxy for),
247 * and avoids penalizing OSS/OSTs under light load. */
248 oss->lqo_penalty >>= (age / lov->desc.ld_qos_maxage);
251 lov->lov_qos.lq_dirty = 0;
252 lov->lov_qos.lq_reset = 0;
254 /* If each ost has almost same free space,
255 * do rr allocation for better creation performance */
256 lov->lov_qos.lq_same_space = 0;
257 if ((ba_max * (256 - lov->lov_qos.lq_threshold_rr)) >> 8 < ba_min) {
258 lov->lov_qos.lq_same_space = 1;
259 /* Reset weights for the next time we enter qos mode */
260 lov->lov_qos.lq_reset = 1;
265 if (!rc && lov->lov_qos.lq_same_space)
270 static int qos_calc_weight(struct lov_obd *lov, int i)
274 /* Final ost weight = TGT_BAVAIL - ost_penalty - oss_penalty */
275 temp = TGT_BAVAIL(i);
276 temp2 = lov->lov_tgts[i]->ltd_qos.ltq_penalty +
277 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty;
279 lov->lov_tgts[i]->ltd_qos.ltq_weight = 0;
281 lov->lov_tgts[i]->ltd_qos.ltq_weight = temp - temp2;
285 /* We just used this index for a stripe; adjust everyone's weights */
286 static int qos_used(struct lov_obd *lov, struct ost_pool *osts,
287 __u32 index, __u64 *total_wt)
289 struct lov_qos_oss *oss;
293 /* Don't allocate from this stripe anymore, until the next alloc_qos */
294 lov->lov_tgts[index]->ltd_qos.ltq_usable = 0;
296 oss = lov->lov_tgts[index]->ltd_qos.ltq_oss;
298 /* Decay old penalty by half (we're adding max penalty, and don't
299 want it to run away.) */
300 lov->lov_tgts[index]->ltd_qos.ltq_penalty >>= 1;
301 oss->lqo_penalty >>= 1;
303 /* mark the OSS and OST as recently used */
304 lov->lov_tgts[index]->ltd_qos.ltq_used =
305 oss->lqo_used = cfs_time_current_sec();
307 /* Set max penalties for this OST and OSS */
308 lov->lov_tgts[index]->ltd_qos.ltq_penalty +=
309 lov->lov_tgts[index]->ltd_qos.ltq_penalty_per_obj *
310 lov->desc.ld_active_tgt_count;
311 oss->lqo_penalty += oss->lqo_penalty_per_obj *
312 lov->lov_qos.lq_active_oss_count;
314 /* Decrease all OSS penalties */
315 cfs_list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
316 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
317 oss->lqo_penalty = 0;
319 oss->lqo_penalty -= oss->lqo_penalty_per_obj;
323 /* Decrease all OST penalties */
324 for (j = 0; j < osts->op_count; j++) {
327 i = osts->op_array[j];
328 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
330 if (lov->lov_tgts[i]->ltd_qos.ltq_penalty <
331 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj)
332 lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
334 lov->lov_tgts[i]->ltd_qos.ltq_penalty -=
335 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj;
337 qos_calc_weight(lov, i);
339 /* Recalc the total weight of usable osts */
340 if (lov->lov_tgts[i]->ltd_qos.ltq_usable)
341 *total_wt += lov->lov_tgts[i]->ltd_qos.ltq_weight;
344 CDEBUG(D_QOS, "recalc tgt %d usable=%d avail="LPU64
345 " ostppo="LPU64" ostp="LPU64" ossppo="LPU64
346 " ossp="LPU64" wt="LPU64"\n",
347 i, lov->lov_tgts[i]->ltd_qos.ltq_usable,
349 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj >> 10,
350 lov->lov_tgts[i]->ltd_qos.ltq_penalty >> 10,
351 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty_per_obj>>10,
352 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty >> 10,
353 lov->lov_tgts[i]->ltd_qos.ltq_weight >> 10);
360 #define LOV_QOS_EMPTY ((__u32)-1)
361 /* compute optimal round-robin order, based on OSTs per OSS */
362 static int qos_calc_rr(struct lov_obd *lov, struct ost_pool *src_pool,
363 struct lov_qos_rr *lqr)
365 struct lov_qos_oss *oss;
366 unsigned placed, real_count;
370 if (!lqr->lqr_dirty) {
371 LASSERT(lqr->lqr_pool.op_size);
375 /* Do actual allocation. */
376 cfs_down_write(&lov->lov_qos.lq_rw_sem);
379 * Check again. While we were sleeping on @lq_rw_sem something could
382 if (!lqr->lqr_dirty) {
383 LASSERT(lqr->lqr_pool.op_size);
384 cfs_up_write(&lov->lov_qos.lq_rw_sem);
388 real_count = src_pool->op_count;
390 /* Zero the pool array */
391 /* alloc_rr is holding a read lock on the pool, so nobody is adding/
392 deleting from the pool. The lq_rw_sem insures that nobody else
394 lqr->lqr_pool.op_count = real_count;
395 rc = lov_ost_pool_extend(&lqr->lqr_pool, real_count);
397 cfs_up_write(&lov->lov_qos.lq_rw_sem);
400 for (i = 0; i < lqr->lqr_pool.op_count; i++)
401 lqr->lqr_pool.op_array[i] = LOV_QOS_EMPTY;
403 /* Place all the OSTs from 1 OSS at the same time. */
405 cfs_list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
407 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
408 if (lov->lov_tgts[src_pool->op_array[i]] &&
409 (lov->lov_tgts[src_pool->op_array[i]]->ltd_qos.ltq_oss == oss)) {
410 /* Evenly space these OSTs across arrayspace */
411 int next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count;
412 while (lqr->lqr_pool.op_array[next] !=
414 next = (next + 1) % lqr->lqr_pool.op_count;
415 lqr->lqr_pool.op_array[next] = src_pool->op_array[i];
423 cfs_up_write(&lov->lov_qos.lq_rw_sem);
425 if (placed != real_count) {
426 /* This should never happen */
427 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
428 "round-robin list (%d of %d).\n",
430 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
431 LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
432 lqr->lqr_pool.op_array[i]);
439 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
440 LCONSOLE(D_QOS, "rr #%d ost idx=%d\n", i,
441 lqr->lqr_pool.op_array[i]);
449 void qos_shrink_lsm(struct lov_request_set *set)
451 struct lov_stripe_md *lsm = set->set_oi->oi_md, *lsm_new;
452 /* XXX LOV STACKING call into osc for sizes */
453 unsigned oldsize, newsize;
455 if (set->set_oti && set->set_cookies && set->set_cookie_sent) {
456 struct llog_cookie *cookies;
457 oldsize = lsm->lsm_stripe_count * sizeof(*cookies);
458 newsize = set->set_count * sizeof(*cookies);
460 cookies = set->set_cookies;
461 oti_alloc_cookies(set->set_oti, set->set_count);
462 if (set->set_oti->oti_logcookies) {
463 memcpy(set->set_oti->oti_logcookies, cookies, newsize);
464 OBD_FREE_LARGE(cookies, oldsize);
465 set->set_cookies = set->set_oti->oti_logcookies;
467 CWARN("'leaking' %d bytes\n", oldsize - newsize);
471 CWARN("using fewer stripes for object "LPU64": old %u new %u\n",
472 lsm->lsm_object_id, lsm->lsm_stripe_count, set->set_count);
473 LASSERT(lsm->lsm_stripe_count >= set->set_count);
475 newsize = lov_stripe_md_size(set->set_count);
476 OBD_ALLOC_LARGE(lsm_new, newsize);
477 if (lsm_new != NULL) {
479 memcpy(lsm_new, lsm, sizeof(*lsm));
480 for (i = 0; i < lsm->lsm_stripe_count; i++) {
481 if (i < set->set_count) {
482 lsm_new->lsm_oinfo[i] = lsm->lsm_oinfo[i];
485 OBD_SLAB_FREE(lsm->lsm_oinfo[i], lov_oinfo_slab,
486 sizeof(struct lov_oinfo));
488 lsm_new->lsm_stripe_count = set->set_count;
489 OBD_FREE_LARGE(lsm, sizeof(struct lov_stripe_md) +
490 lsm->lsm_stripe_count*sizeof(struct lov_oinfo*));
491 set->set_oi->oi_md = lsm_new;
493 CWARN("'leaking' few bytes\n");
498 * Check whether we can create the object on the OST(refered by ost_idx)
500 * 0: create the object.
501 * other value: did not create the object.
503 static int lov_check_and_create_object(struct lov_obd *lov, int ost_idx,
504 struct lov_stripe_md *lsm,
505 struct lov_request *req,
506 struct obd_trans_info *oti)
512 CDEBUG(D_QOS, "Check and create on idx %d \n", ost_idx);
513 if (!lov->lov_tgts[ost_idx] ||
514 !lov->lov_tgts[ost_idx]->ltd_active)
517 /* check if objects has been created on this ost */
518 for (stripe = 0; stripe < lsm->lsm_stripe_count; stripe++) {
519 /* already have object at this stripe */
520 if (ost_idx == lsm->lsm_oinfo[stripe]->loi_ost_idx)
524 if (stripe >= lsm->lsm_stripe_count) {
525 req->rq_idx = ost_idx;
526 rc = obd_create(lov->lov_tgts[ost_idx]->ltd_exp,
527 req->rq_oi.oi_oa, &req->rq_oi.oi_md,
533 int qos_remedy_create(struct lov_request_set *set, struct lov_request *req)
535 struct lov_stripe_md *lsm = set->set_oi->oi_md;
536 struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
537 unsigned ost_idx, ost_count;
538 struct pool_desc *pool;
539 struct ost_pool *osts = NULL;
543 /* First check whether we can create the objects on the pool */
544 pool = lov_find_pool(lov, lsm->lsm_pool_name);
546 cfs_down_read(&pool_tgt_rw_sem(pool));
547 osts = &(pool->pool_obds);
548 ost_count = osts->op_count;
549 for (i = 0, ost_idx = osts->op_array[0]; i < ost_count;
550 i++, ost_idx = osts->op_array[i]) {
551 rc = lov_check_and_create_object(lov, ost_idx, lsm, req,
556 cfs_up_read(&pool_tgt_rw_sem(pool));
557 lov_pool_putref(pool);
561 ost_count = lov->desc.ld_tgt_count;
562 /* Then check whether we can create the objects on other OSTs */
563 ost_idx = (req->rq_idx + lsm->lsm_stripe_count) % ost_count;
564 for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
565 rc = lov_check_and_create_object(lov, ost_idx, lsm, req,
575 static int min_stripe_count(int stripe_cnt, int flags)
577 return (flags & LOV_USES_DEFAULT_STRIPE ?
578 stripe_cnt - (stripe_cnt / 4) : stripe_cnt);
581 #define LOV_CREATE_RESEED_MULT 30
582 #define LOV_CREATE_RESEED_MIN 2000
583 /* Allocate objects on osts with round-robin algorithm */
584 static int alloc_rr(struct lov_obd *lov, int *idx_arr, int *stripe_cnt,
585 char *poolname, int flags)
590 int ost_start_idx_temp;
592 int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
593 struct pool_desc *pool;
594 struct ost_pool *osts;
595 struct lov_qos_rr *lqr;
598 pool = lov_find_pool(lov, poolname);
600 osts = &(lov->lov_packed);
601 lqr = &(lov->lov_qos.lq_rr);
603 cfs_down_read(&pool_tgt_rw_sem(pool));
604 osts = &(pool->pool_obds);
605 lqr = &(pool->pool_rr);
608 rc = qos_calc_rr(lov, osts, lqr);
612 if (--lqr->lqr_start_count <= 0) {
613 lqr->lqr_start_idx = cfs_rand() % osts->op_count;
614 lqr->lqr_start_count =
615 (LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) +
616 LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U);
617 } else if (stripe_cnt_min >= osts->op_count ||
618 lqr->lqr_start_idx > osts->op_count) {
619 /* If we have allocated from all of the OSTs, slowly
620 * precess the next start if the OST/stripe count isn't
621 * already doing this for us. */
622 lqr->lqr_start_idx %= osts->op_count;
623 if (*stripe_cnt > 1 && (osts->op_count % (*stripe_cnt)) != 1)
624 ++lqr->lqr_offset_idx;
626 cfs_down_read(&lov->lov_qos.lq_rw_sem);
627 ost_start_idx_temp = lqr->lqr_start_idx;
630 array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) % osts->op_count;
633 CDEBUG(D_QOS, "pool '%s' want %d startidx %d startcnt %d offset %d "
634 "active %d count %d arrayidx %d\n", poolname,
635 *stripe_cnt, lqr->lqr_start_idx, lqr->lqr_start_count,
636 lqr->lqr_offset_idx, osts->op_count, osts->op_count, array_idx);
639 for (i = 0; i < osts->op_count;
640 i++, array_idx=(array_idx + 1) % osts->op_count) {
641 ++lqr->lqr_start_idx;
642 ost_idx = lqr->lqr_pool.op_array[array_idx];
644 CDEBUG(D_QOS, "#%d strt %d act %d strp %d ary %d idx %d\n",
645 i, lqr->lqr_start_idx,
646 ((ost_idx != LOV_QOS_EMPTY) && lov->lov_tgts[ost_idx]) ?
647 lov->lov_tgts[ost_idx]->ltd_active : 0,
648 idx_pos - idx_arr, array_idx, ost_idx);
650 if ((ost_idx == LOV_QOS_EMPTY) || !lov->lov_tgts[ost_idx] ||
651 !lov->lov_tgts[ost_idx]->ltd_active)
654 /* Fail Check before osc_precreate() is called
655 so we can only 'fail' single OSC. */
656 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
659 /* Drop slow OSCs if we can */
660 if (obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed)
665 /* We have enough stripes */
666 if (idx_pos - idx_arr == *stripe_cnt)
669 if ((speed < 2) && (idx_pos - idx_arr < stripe_cnt_min)) {
670 /* Try again, allowing slower OSCs */
672 lqr->lqr_start_idx = ost_start_idx_temp;
676 cfs_up_read(&lov->lov_qos.lq_rw_sem);
678 *stripe_cnt = idx_pos - idx_arr;
681 cfs_up_read(&pool_tgt_rw_sem(pool));
682 /* put back ref got by lov_find_pool() */
683 lov_pool_putref(pool);
689 /* alloc objects on osts with specific stripe offset */
690 static int alloc_specific(struct lov_obd *lov, struct lov_stripe_md *lsm,
693 unsigned ost_idx, array_idx, ost_count;
696 struct pool_desc *pool;
697 struct ost_pool *osts;
700 pool = lov_find_pool(lov, lsm->lsm_pool_name);
702 osts = &(lov->lov_packed);
704 cfs_down_read(&pool_tgt_rw_sem(pool));
705 osts = &(pool->pool_obds);
708 ost_count = osts->op_count;
711 /* search loi_ost_idx in ost array */
713 for (i = 0; i < ost_count; i++) {
714 if (osts->op_array[i] == lsm->lsm_oinfo[0]->loi_ost_idx) {
719 if (i == ost_count) {
720 CERROR("Start index %d not found in pool '%s'\n",
721 lsm->lsm_oinfo[0]->loi_ost_idx, lsm->lsm_pool_name);
722 GOTO(out, rc = -EINVAL);
726 for (i = 0; i < ost_count;
727 i++, array_idx = (array_idx + 1) % ost_count) {
728 ost_idx = osts->op_array[array_idx];
730 if (!lov->lov_tgts[ost_idx] ||
731 !lov->lov_tgts[ost_idx]->ltd_active) {
735 /* Fail Check before osc_precreate() is called
736 so we can only 'fail' single OSC. */
737 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
740 /* Drop slow OSCs if we can, but not for requested start idx.
742 * This means "if OSC is slow and it is not the requested
743 * start OST, then it can be skipped, otherwise skip it only
744 * if it is inactive/recovering/out-of-space." */
745 if ((obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed) &&
746 (i != 0 || speed >= 2))
751 /* We have enough stripes */
752 if (idx_pos - idx_arr == lsm->lsm_stripe_count)
756 /* Try again, allowing slower OSCs */
761 /* If we were passed specific striping params, then a failure to
762 * meet those requirements is an error, since we can't reallocate
763 * that memory (it might be part of a larger array or something).
765 * We can only get here if lsm_stripe_count was originally > 1.
767 CERROR("can't lstripe objid "LPX64": have %d want %u\n",
768 lsm->lsm_object_id, (int)(idx_pos - idx_arr),
769 lsm->lsm_stripe_count);
773 cfs_up_read(&pool_tgt_rw_sem(pool));
774 /* put back ref got by lov_find_pool() */
775 lov_pool_putref(pool);
781 /* Alloc objects on osts with optimization based on:
783 - network resources (shared OSS's)
785 static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt,
786 char *poolname, int flags)
788 struct lov_obd *lov = &exp->exp_obd->u.lov;
789 __u64 total_weight = 0;
790 int nfound, good_osts, i, rc = 0;
791 int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
792 struct pool_desc *pool;
793 struct ost_pool *osts;
794 struct lov_qos_rr *lqr;
797 if (stripe_cnt_min < 1)
800 pool = lov_find_pool(lov, poolname);
802 osts = &(lov->lov_packed);
803 lqr = &(lov->lov_qos.lq_rr);
805 cfs_down_read(&pool_tgt_rw_sem(pool));
806 osts = &(pool->pool_obds);
807 lqr = &(pool->pool_rr);
810 obd_getref(exp->exp_obd);
812 /* wait for fresh statfs info if needed, the rpcs are sent in
814 qos_statfs_update(exp->exp_obd,
815 cfs_time_shift_64(-2 * lov->desc.ld_qos_maxage), 1);
817 /* Detect -EAGAIN early, before expensive lock is taken. */
818 if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space)
819 GOTO(out_nolock, rc = -EAGAIN);
821 /* Do actual allocation, use write lock here. */
822 cfs_down_write(&lov->lov_qos.lq_rw_sem);
825 * Check again, while we were sleeping on @lq_rw_sem things could
828 if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space)
829 GOTO(out, rc = -EAGAIN);
831 if (lov->desc.ld_active_tgt_count < 2)
832 GOTO(out, rc = -EAGAIN);
834 rc = qos_calc_ppo(exp->exp_obd);
839 /* Find all the OSTs that are valid stripe candidates */
840 for (i = 0; i < osts->op_count; i++) {
841 if (!lov->lov_tgts[osts->op_array[i]] ||
842 !lov->lov_tgts[osts->op_array[i]]->ltd_active)
845 /* Fail Check before osc_precreate() is called
846 so we can only 'fail' single OSC. */
847 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && osts->op_array[i] == 0)
850 if (obd_precreate(lov->lov_tgts[osts->op_array[i]]->ltd_exp) > 2)
853 lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable = 1;
854 qos_calc_weight(lov, osts->op_array[i]);
855 total_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight;
861 CDEBUG(D_QOS, "found %d good osts\n", good_osts);
864 if (good_osts < stripe_cnt_min)
865 GOTO(out, rc = -EAGAIN);
867 /* We have enough osts */
868 if (good_osts < *stripe_cnt)
869 *stripe_cnt = good_osts;
872 GOTO(out, rc = -EAGAIN);
874 /* Find enough OSTs with weighted random allocation. */
876 while (nfound < *stripe_cnt) {
877 __u64 rand, cur_weight;
883 #if BITS_PER_LONG == 32
884 rand = cfs_rand() % (unsigned)total_weight;
885 /* If total_weight > 32-bit, first generate the high
886 * 32 bits of the random number, then add in the low
887 * 32 bits (truncated to the upper limit, if needed) */
888 if (total_weight > 0xffffffffULL)
889 rand = (__u64)(cfs_rand() %
890 (unsigned)(total_weight >> 32)) << 32;
894 if (rand == (total_weight & 0xffffffff00000000ULL))
895 rand |= cfs_rand() % (unsigned)total_weight;
900 rand = ((__u64)cfs_rand() << 32 | cfs_rand()) %
907 /* On average, this will hit larger-weighted osts more often.
908 0-weight osts will always get used last (only when rand=0).*/
909 for (i = 0; i < osts->op_count; i++) {
910 if (!lov->lov_tgts[osts->op_array[i]] ||
911 !lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable)
914 cur_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight;
916 CDEBUG(D_QOS, "stripe_cnt=%d nfound=%d cur_weight="LPU64
917 " rand="LPU64" total_weight="LPU64"\n",
918 *stripe_cnt, nfound, cur_weight, rand, total_weight);
920 if (cur_weight >= rand) {
922 CDEBUG(D_QOS, "assigned stripe=%d to idx=%d\n",
923 nfound, osts->op_array[i]);
925 idx_arr[nfound++] = osts->op_array[i];
926 qos_used(lov, osts, osts->op_array[i], &total_weight);
931 /* should never satisfy below condition */
933 CERROR("Didn't find any OSTs?\n");
937 LASSERT(nfound == *stripe_cnt);
940 cfs_up_write(&lov->lov_qos.lq_rw_sem);
944 cfs_up_read(&pool_tgt_rw_sem(pool));
945 /* put back ref got by lov_find_pool() */
946 lov_pool_putref(pool);
950 rc = alloc_rr(lov, idx_arr, stripe_cnt, poolname, flags);
952 obd_putref(exp->exp_obd);
956 /* return new alloced stripe count on success */
957 static int alloc_idx_array(struct obd_export *exp, struct lov_stripe_md *lsm,
958 int newea, int **idx_arr, int *arr_cnt, int flags)
960 struct lov_obd *lov = &exp->exp_obd->u.lov;
961 int stripe_cnt = lsm->lsm_stripe_count;
966 *arr_cnt = stripe_cnt;
967 OBD_ALLOC(tmp_arr, *arr_cnt * sizeof(int));
970 for (i = 0; i < *arr_cnt; i++)
974 lsm->lsm_oinfo[0]->loi_ost_idx >= lov->desc.ld_tgt_count)
975 rc = alloc_qos(exp, tmp_arr, &stripe_cnt,
976 lsm->lsm_pool_name, flags);
978 rc = alloc_specific(lov, lsm, tmp_arr);
986 OBD_FREE(tmp_arr, *arr_cnt * sizeof(int));
991 static void free_idx_array(int *idx_arr, int arr_cnt)
994 OBD_FREE(idx_arr, arr_cnt * sizeof(int));
997 int qos_prep_create(struct obd_export *exp, struct lov_request_set *set)
999 struct lov_obd *lov = &exp->exp_obd->u.lov;
1000 struct lov_stripe_md *lsm;
1001 struct obdo *src_oa = set->set_oi->oi_oa;
1002 struct obd_trans_info *oti = set->set_oti;
1003 int i, stripes, rc = 0, newea = 0;
1004 int flag = LOV_USES_ASSIGNED_STRIPE;
1005 int *idx_arr = NULL, idx_cnt = 0;
1008 LASSERT(src_oa->o_valid & OBD_MD_FLID);
1009 LASSERT(src_oa->o_valid & OBD_MD_FLGROUP);
1011 if (set->set_oi->oi_md == NULL) {
1012 int stripes_def = lov_get_stripecnt(lov, 0);
1014 /* If the MDS file was truncated up to some size, stripe over
1015 * enough OSTs to allow the file to be created at that size.
1016 * This may mean we use more than the default # of stripes. */
1017 if (src_oa->o_valid & OBD_MD_FLSIZE) {
1018 obd_size min_bavail = LUSTRE_STRIPE_MAXBYTES;
1020 /* Find a small number of stripes we can use
1021 (up to # of active osts). */
1023 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1024 if (!lov->lov_tgts[i] ||
1025 !lov->lov_tgts[i]->ltd_active)
1027 min_bavail = min(min_bavail, TGT_BAVAIL(i));
1028 if (min_bavail * stripes > src_oa->o_size)
1033 if (stripes < stripes_def)
1034 stripes = stripes_def;
1036 flag = LOV_USES_DEFAULT_STRIPE;
1037 stripes = stripes_def;
1040 rc = lov_alloc_memmd(&set->set_oi->oi_md, stripes,
1041 lov->desc.ld_pattern ?
1042 lov->desc.ld_pattern : LOV_PATTERN_RAID0,
1050 lsm = set->set_oi->oi_md;
1051 lsm->lsm_object_id = src_oa->o_id;
1052 lsm->lsm_object_seq = src_oa->o_seq;
1054 if (!lsm->lsm_stripe_size)
1055 lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size;
1056 if (!lsm->lsm_pattern) {
1057 LASSERT(lov->desc.ld_pattern);
1058 lsm->lsm_pattern = lov->desc.ld_pattern;
1061 stripes = alloc_idx_array(exp, lsm, newea, &idx_arr, &idx_cnt, flag);
1063 GOTO(out_err, rc = stripes ? stripes : -EIO);
1064 LASSERTF(stripes <= lsm->lsm_stripe_count,"requested %d allocated %d\n",
1065 lsm->lsm_stripe_count, stripes);
1067 for (i = 0; i < stripes; i++) {
1068 struct lov_request *req;
1069 int ost_idx = idx_arr[i];
1070 LASSERT(ost_idx >= 0);
1072 OBD_ALLOC(req, sizeof(*req));
1074 GOTO(out_err, rc = -ENOMEM);
1075 lov_set_add_req(req, set);
1077 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
1078 OBD_ALLOC_LARGE(req->rq_oi.oi_md, req->rq_buflen);
1079 if (req->rq_oi.oi_md == NULL)
1080 GOTO(out_err, rc = -ENOMEM);
1082 OBDO_ALLOC(req->rq_oi.oi_oa);
1083 if (req->rq_oi.oi_oa == NULL)
1084 GOTO(out_err, rc = -ENOMEM);
1086 req->rq_idx = ost_idx;
1088 /* create data objects with "parent" OA */
1089 memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
1090 req->rq_oi.oi_cb_up = cb_create_update;
1092 /* XXX When we start creating objects on demand, we need to
1093 * make sure that we always create the object on the
1094 * stripe which holds the existing file size.
1096 if (src_oa->o_valid & OBD_MD_FLSIZE) {
1097 req->rq_oi.oi_oa->o_size =
1098 lov_size_to_stripe(lsm, src_oa->o_size, i);
1100 CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
1101 i, req->rq_oi.oi_oa->o_size, src_oa->o_size);
1104 LASSERT(set->set_count == stripes);
1106 if (stripes < lsm->lsm_stripe_count)
1107 qos_shrink_lsm(set);
1108 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_LOV_PREP_CREATE)) {
1109 qos_shrink_lsm(set);
1113 if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) {
1114 oti_alloc_cookies(oti, set->set_count);
1115 if (!oti->oti_logcookies)
1116 GOTO(out_err, rc = -ENOMEM);
1117 set->set_cookies = oti->oti_logcookies;
1121 obd_free_memmd(exp, &set->set_oi->oi_md);
1123 free_idx_array(idx_arr, idx_cnt);
1128 void qos_update(struct lov_obd *lov)
1131 lov->lov_qos.lq_dirty = 1;
1134 void qos_statfs_done(struct lov_obd *lov)
1136 LASSERT(lov->lov_qos.lq_statfs_in_progress);
1137 cfs_down_write(&lov->lov_qos.lq_rw_sem);
1138 lov->lov_qos.lq_statfs_in_progress = 0;
1139 /* wake up any threads waiting for the statfs rpcs to complete */
1140 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);