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)
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 if ((ba_max * (256 - lov->lov_qos.lq_threshold_rr)) >> 8 < ba_min) {
257 lov->lov_qos.lq_same_space = 1;
258 /* Reset weights for the next time we enter qos mode */
259 lov->lov_qos.lq_reset = 1;
264 if (!rc && lov->lov_qos.lq_same_space)
269 static int qos_calc_weight(struct lov_obd *lov, int i)
273 /* Final ost weight = TGT_BAVAIL - ost_penalty - oss_penalty */
274 temp = TGT_BAVAIL(i);
275 temp2 = lov->lov_tgts[i]->ltd_qos.ltq_penalty +
276 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty;
278 lov->lov_tgts[i]->ltd_qos.ltq_weight = 0;
280 lov->lov_tgts[i]->ltd_qos.ltq_weight = temp - temp2;
284 /* We just used this index for a stripe; adjust everyone's weights */
285 static int qos_used(struct lov_obd *lov, struct ost_pool *osts,
286 __u32 index, __u64 *total_wt)
288 struct lov_qos_oss *oss;
292 /* Don't allocate from this stripe anymore, until the next alloc_qos */
293 lov->lov_tgts[index]->ltd_qos.ltq_usable = 0;
295 oss = lov->lov_tgts[index]->ltd_qos.ltq_oss;
297 /* Decay old penalty by half (we're adding max penalty, and don't
298 want it to run away.) */
299 lov->lov_tgts[index]->ltd_qos.ltq_penalty >>= 1;
300 oss->lqo_penalty >>= 1;
302 /* mark the OSS and OST as recently used */
303 lov->lov_tgts[index]->ltd_qos.ltq_used =
304 oss->lqo_used = cfs_time_current_sec();
306 /* Set max penalties for this OST and OSS */
307 lov->lov_tgts[index]->ltd_qos.ltq_penalty +=
308 lov->lov_tgts[index]->ltd_qos.ltq_penalty_per_obj *
309 lov->desc.ld_active_tgt_count;
310 oss->lqo_penalty += oss->lqo_penalty_per_obj *
311 lov->lov_qos.lq_active_oss_count;
313 /* Decrease all OSS penalties */
314 list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
315 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
316 oss->lqo_penalty = 0;
318 oss->lqo_penalty -= oss->lqo_penalty_per_obj;
322 /* Decrease all OST penalties */
323 for (j = 0; j < osts->op_count; j++) {
326 i = osts->op_array[j];
327 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
329 if (lov->lov_tgts[i]->ltd_qos.ltq_penalty <
330 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj)
331 lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
333 lov->lov_tgts[i]->ltd_qos.ltq_penalty -=
334 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj;
336 qos_calc_weight(lov, i);
338 /* Recalc the total weight of usable osts */
339 if (lov->lov_tgts[i]->ltd_qos.ltq_usable)
340 *total_wt += lov->lov_tgts[i]->ltd_qos.ltq_weight;
343 CDEBUG(D_QOS, "recalc tgt %d usable=%d avail="LPU64
344 " ostppo="LPU64" ostp="LPU64" ossppo="LPU64
345 " ossp="LPU64" wt="LPU64"\n",
346 i, lov->lov_tgts[i]->ltd_qos.ltq_usable,
348 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj >> 10,
349 lov->lov_tgts[i]->ltd_qos.ltq_penalty >> 10,
350 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty_per_obj>>10,
351 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty >> 10,
352 lov->lov_tgts[i]->ltd_qos.ltq_weight >> 10);
359 #define LOV_QOS_EMPTY ((__u32)-1)
360 /* 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);
378 * Check again. While we were sleeping on @lq_rw_sem something could
381 if (!lqr->lqr_dirty) {
382 LASSERT(lqr->lqr_pool.op_size);
383 up_write(&lov->lov_qos.lq_rw_sem);
387 real_count = src_pool->op_count;
389 /* Zero the pool array */
390 /* alloc_rr is holding a read lock on the pool, so nobody is adding/
391 deleting from the pool. The lq_rw_sem insures that nobody else
393 lqr->lqr_pool.op_count = real_count;
394 rc = lov_ost_pool_extend(&lqr->lqr_pool, real_count);
396 up_write(&lov->lov_qos.lq_rw_sem);
399 for (i = 0; i < lqr->lqr_pool.op_count; i++)
400 lqr->lqr_pool.op_array[i] = LOV_QOS_EMPTY;
402 /* Place all the OSTs from 1 OSS at the same time. */
404 list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
406 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
407 if (lov->lov_tgts[src_pool->op_array[i]] &&
408 (lov->lov_tgts[src_pool->op_array[i]]->ltd_qos.ltq_oss == oss)) {
409 /* Evenly space these OSTs across arrayspace */
410 int next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count;
411 while (lqr->lqr_pool.op_array[next] !=
413 next = (next + 1) % lqr->lqr_pool.op_count;
414 lqr->lqr_pool.op_array[next] = src_pool->op_array[i];
422 up_write(&lov->lov_qos.lq_rw_sem);
424 if (placed != real_count) {
425 /* This should never happen */
426 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
427 "round-robin list (%d of %d).\n",
429 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
430 LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
431 lqr->lqr_pool.op_array[i]);
438 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
439 LCONSOLE(D_QOS, "rr #%d ost idx=%d\n", i,
440 lqr->lqr_pool.op_array[i]);
448 void qos_shrink_lsm(struct lov_request_set *set)
450 struct lov_stripe_md *lsm = set->set_oi->oi_md, *lsm_new;
451 /* XXX LOV STACKING call into osc for sizes */
452 unsigned oldsize, newsize;
454 if (set->set_oti && set->set_cookies && set->set_cookie_sent) {
455 struct llog_cookie *cookies;
456 oldsize = lsm->lsm_stripe_count * sizeof(*cookies);
457 newsize = set->set_count * sizeof(*cookies);
459 cookies = set->set_cookies;
460 oti_alloc_cookies(set->set_oti, set->set_count);
461 if (set->set_oti->oti_logcookies) {
462 memcpy(set->set_oti->oti_logcookies, cookies, newsize);
463 OBD_FREE(cookies, oldsize);
464 set->set_cookies = set->set_oti->oti_logcookies;
466 CWARN("'leaking' %d bytes\n", oldsize - newsize);
470 CWARN("using fewer stripes for object "LPU64": old %u new %u\n",
471 lsm->lsm_object_id, lsm->lsm_stripe_count, set->set_count);
472 LASSERT(lsm->lsm_stripe_count >= set->set_count);
474 newsize = lov_stripe_md_size(set->set_count);
475 OBD_ALLOC(lsm_new, newsize);
476 if (lsm_new != NULL) {
478 memcpy(lsm_new, lsm, sizeof(*lsm));
479 for (i = 0; i < lsm->lsm_stripe_count; i++) {
480 if (i < set->set_count) {
481 lsm_new->lsm_oinfo[i] = lsm->lsm_oinfo[i];
484 OBD_SLAB_FREE(lsm->lsm_oinfo[i], lov_oinfo_slab,
485 sizeof(struct lov_oinfo));
487 lsm_new->lsm_stripe_count = set->set_count;
488 OBD_FREE(lsm, sizeof(struct lov_stripe_md) +
489 lsm->lsm_stripe_count * sizeof(struct lov_oinfo *));
490 set->set_oi->oi_md = lsm_new;
492 CWARN("'leaking' few bytes\n");
496 int qos_remedy_create(struct lov_request_set *set, struct lov_request *req)
498 struct lov_stripe_md *lsm = set->set_oi->oi_md;
499 struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
500 unsigned ost_idx, ost_count = lov->desc.ld_tgt_count;
501 int stripe, i, rc = -EIO;
504 ost_idx = (req->rq_idx + lsm->lsm_stripe_count) % ost_count;
505 for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
506 if (!lov->lov_tgts[ost_idx] ||
507 !lov->lov_tgts[ost_idx]->ltd_active)
509 /* check if objects has been created on this ost */
510 for (stripe = 0; stripe < lsm->lsm_stripe_count; stripe++) {
511 /* we try send create to this ost but he is failed */
512 if (stripe == req->rq_stripe)
514 /* already have object at this stripe */
515 if (ost_idx == lsm->lsm_oinfo[stripe]->loi_ost_idx)
518 if (stripe >= lsm->lsm_stripe_count) {
519 req->rq_idx = ost_idx;
520 rc = obd_create(lov->lov_tgts[ost_idx]->ltd_exp,
521 req->rq_oi.oi_oa, &req->rq_oi.oi_md,
530 static int min_stripe_count(int stripe_cnt, int flags)
532 return (flags & LOV_USES_DEFAULT_STRIPE ?
533 stripe_cnt - (stripe_cnt / 4) : stripe_cnt);
536 #define LOV_CREATE_RESEED_MULT 4
537 #define LOV_CREATE_RESEED_MIN 1000
538 /* Allocate objects on osts with round-robin algorithm */
539 static int alloc_rr(struct lov_obd *lov, int *idx_arr, int *stripe_cnt,
540 char *poolname, int flags)
545 int ost_start_idx_temp;
547 int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
548 struct pool_desc *pool;
549 struct ost_pool *osts;
550 struct lov_qos_rr *lqr;
553 pool = lov_find_pool(lov, poolname);
555 osts = &(lov->lov_packed);
556 lqr = &(lov->lov_qos.lq_rr);
558 down_read(&pool_tgt_rw_sem(pool));
559 osts = &(pool->pool_obds);
560 lqr = &(pool->pool_rr);
563 rc = qos_calc_rr(lov, osts, lqr);
567 if (--lqr->lqr_start_count <= 0) {
568 lqr->lqr_start_idx = ll_rand() % osts->op_count;
569 lqr->lqr_start_count =
570 (LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) +
571 LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U);
572 } else if (stripe_cnt_min >= osts->op_count ||
573 lqr->lqr_start_idx > osts->op_count) {
574 /* If we have allocated from all of the OSTs, slowly
575 * precess the next start if the OST/stripe count isn't
576 * already doing this for us. */
577 lqr->lqr_start_idx %= osts->op_count;
578 if (*stripe_cnt > 1 && (osts->op_count % (*stripe_cnt)) != 1)
579 ++lqr->lqr_offset_idx;
581 down_read(&lov->lov_qos.lq_rw_sem);
582 ost_start_idx_temp = lqr->lqr_start_idx;
585 array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) % osts->op_count;
588 CDEBUG(D_QOS, "pool '%s' want %d startidx %d startcnt %d offset %d "
589 "active %d count %d arrayidx %d\n", poolname,
590 *stripe_cnt, lqr->lqr_start_idx, lqr->lqr_start_count,
591 lqr->lqr_offset_idx, osts->op_count, osts->op_count, array_idx);
594 for (i = 0; i < osts->op_count;
595 i++, array_idx=(array_idx + 1) % osts->op_count) {
596 ++lqr->lqr_start_idx;
597 ost_idx = lqr->lqr_pool.op_array[array_idx];
599 CDEBUG(D_QOS, "#%d strt %d act %d strp %d ary %d idx %d\n",
600 i, lqr->lqr_start_idx,
601 ((ost_idx != LOV_QOS_EMPTY) && lov->lov_tgts[ost_idx]) ?
602 lov->lov_tgts[ost_idx]->ltd_active : 0,
603 idx_pos - idx_arr, array_idx, ost_idx);
605 if ((ost_idx == LOV_QOS_EMPTY) || !lov->lov_tgts[ost_idx] ||
606 !lov->lov_tgts[ost_idx]->ltd_active)
609 /* Fail Check before osc_precreate() is called
610 so we can only 'fail' single OSC. */
611 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
614 /* Drop slow OSCs if we can */
615 if (obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed)
620 /* We have enough stripes */
621 if (idx_pos - idx_arr == *stripe_cnt)
624 if ((speed < 2) && (idx_pos - idx_arr < stripe_cnt_min)) {
625 /* Try again, allowing slower OSCs */
627 lqr->lqr_start_idx = ost_start_idx_temp;
631 up_read(&lov->lov_qos.lq_rw_sem);
633 *stripe_cnt = idx_pos - idx_arr;
636 up_read(&pool_tgt_rw_sem(pool));
637 /* put back ref got by lov_find_pool() */
638 lov_pool_putref(pool);
644 /* alloc objects on osts with specific stripe offset */
645 static int alloc_specific(struct lov_obd *lov, struct lov_stripe_md *lsm,
648 unsigned ost_idx, array_idx, ost_count;
651 struct pool_desc *pool;
652 struct ost_pool *osts;
655 pool = lov_find_pool(lov, lsm->lsm_pool_name);
657 osts = &(lov->lov_packed);
659 down_read(&pool_tgt_rw_sem(pool));
660 osts = &(pool->pool_obds);
663 ost_count = osts->op_count;
666 /* search loi_ost_idx in ost array */
668 for (i = 0; i < ost_count; i++) {
669 if (osts->op_array[i] == lsm->lsm_oinfo[0]->loi_ost_idx) {
674 if (i == ost_count) {
675 CERROR("Start index %d not found in pool '%s'\n",
676 lsm->lsm_oinfo[0]->loi_ost_idx, lsm->lsm_pool_name);
677 GOTO(out, rc = -EINVAL);
681 for (i = 0; i < ost_count;
682 i++, array_idx = (array_idx + 1) % ost_count) {
683 ost_idx = osts->op_array[array_idx];
685 if (!lov->lov_tgts[ost_idx] ||
686 !lov->lov_tgts[ost_idx]->ltd_active) {
690 /* Fail Check before osc_precreate() is called
691 so we can only 'fail' single OSC. */
692 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
695 /* Drop slow OSCs if we can, but not for requested start idx.
697 * This means "if OSC is slow and it is not the requested
698 * start OST, then it can be skipped, otherwise skip it only
699 * if it is inactive/recovering/out-of-space." */
700 if ((obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed) &&
701 (i != 0 || speed >= 2))
706 /* We have enough stripes */
707 if (idx_pos - idx_arr == lsm->lsm_stripe_count)
711 /* Try again, allowing slower OSCs */
716 /* If we were passed specific striping params, then a failure to
717 * meet those requirements is an error, since we can't reallocate
718 * that memory (it might be part of a larger array or something).
720 * We can only get here if lsm_stripe_count was originally > 1.
722 CERROR("can't lstripe objid "LPX64": have %d want %u\n",
723 lsm->lsm_object_id, (int)(idx_pos - idx_arr),
724 lsm->lsm_stripe_count);
728 up_read(&pool_tgt_rw_sem(pool));
729 /* put back ref got by lov_find_pool() */
730 lov_pool_putref(pool);
736 /* Alloc objects on osts with optimization based on:
738 - network resources (shared OSS's)
740 static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt,
741 char *poolname, int flags)
743 struct lov_obd *lov = &exp->exp_obd->u.lov;
744 __u64 total_weight = 0;
745 int nfound, good_osts, i, rc = 0;
746 int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
747 struct pool_desc *pool;
748 struct ost_pool *osts;
749 struct lov_qos_rr *lqr;
752 if (stripe_cnt_min < 1)
755 pool = lov_find_pool(lov, poolname);
757 osts = &(lov->lov_packed);
758 lqr = &(lov->lov_qos.lq_rr);
760 down_read(&pool_tgt_rw_sem(pool));
761 osts = &(pool->pool_obds);
762 lqr = &(pool->pool_rr);
765 obd_getref(exp->exp_obd);
767 /* wait for fresh statfs info if needed, the rpcs are sent in
769 qos_statfs_update(exp->exp_obd,
770 cfs_time_shift_64(-2 * lov->desc.ld_qos_maxage), 1);
772 /* Detect -EAGAIN early, before expensive lock is taken. */
773 if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space)
774 GOTO(out_nolock, rc = -EAGAIN);
776 /* Do actual allocation, use write lock here. */
777 down_write(&lov->lov_qos.lq_rw_sem);
780 * Check again, while we were sleeping on @lq_rw_sem things could
783 if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space)
784 GOTO(out, rc = -EAGAIN);
786 if (lov->desc.ld_active_tgt_count < 2)
787 GOTO(out, rc = -EAGAIN);
789 rc = qos_calc_ppo(exp->exp_obd);
794 /* Find all the OSTs that are valid stripe candidates */
795 for (i = 0; i < osts->op_count; i++) {
796 if (!lov->lov_tgts[osts->op_array[i]] ||
797 !lov->lov_tgts[osts->op_array[i]]->ltd_active)
800 /* Fail Check before osc_precreate() is called
801 so we can only 'fail' single OSC. */
802 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && osts->op_array[i] == 0)
805 if (obd_precreate(lov->lov_tgts[osts->op_array[i]]->ltd_exp) > 2)
808 lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable = 1;
809 qos_calc_weight(lov, osts->op_array[i]);
810 total_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight;
816 CDEBUG(D_QOS, "found %d good osts\n", good_osts);
819 if (good_osts < stripe_cnt_min)
820 GOTO(out, rc = -EAGAIN);
822 /* We have enough osts */
823 if (good_osts < *stripe_cnt)
824 *stripe_cnt = good_osts;
827 GOTO(out, rc = -EAGAIN);
829 /* Find enough OSTs with weighted random allocation. */
831 while (nfound < *stripe_cnt) {
832 __u64 rand, cur_weight;
838 #if BITS_PER_LONG == 32
839 rand = ll_rand() % (unsigned)total_weight;
840 /* If total_weight > 32-bit, first generate the high
841 * 32 bits of the random number, then add in the low
842 * 32 bits (truncated to the upper limit, if needed) */
843 if (total_weight > 0xffffffffULL)
844 rand = (__u64)(ll_rand() %
845 (unsigned)(total_weight >> 32)) << 32;
849 if (rand == (total_weight & 0xffffffff00000000ULL))
850 rand |= ll_rand() % (unsigned)total_weight;
855 rand = ((__u64)ll_rand() << 32 | ll_rand()) %
862 /* On average, this will hit larger-weighted osts more often.
863 0-weight osts will always get used last (only when rand=0).*/
864 for (i = 0; i < osts->op_count; i++) {
865 if (!lov->lov_tgts[osts->op_array[i]] ||
866 !lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable)
869 cur_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight;
871 CDEBUG(D_QOS, "stripe_cnt=%d nfound=%d cur_weight="LPU64
872 " rand="LPU64" total_weight="LPU64"\n",
873 *stripe_cnt, nfound, cur_weight, rand, total_weight);
875 if (cur_weight >= rand) {
877 CDEBUG(D_QOS, "assigned stripe=%d to idx=%d\n",
878 nfound, osts->op_array[i]);
880 idx_arr[nfound++] = osts->op_array[i];
881 qos_used(lov, osts, osts->op_array[i], &total_weight);
886 /* should never satisfy below condition */
888 CERROR("Didn't find any OSTs?\n");
892 LASSERT(nfound == *stripe_cnt);
895 up_write(&lov->lov_qos.lq_rw_sem);
899 up_read(&pool_tgt_rw_sem(pool));
900 /* put back ref got by lov_find_pool() */
901 lov_pool_putref(pool);
905 rc = alloc_rr(lov, idx_arr, stripe_cnt, poolname, flags);
907 obd_putref(exp->exp_obd);
911 /* return new alloced stripe count on success */
912 static int alloc_idx_array(struct obd_export *exp, struct lov_stripe_md *lsm,
913 int newea, int **idx_arr, int *arr_cnt, int flags)
915 struct lov_obd *lov = &exp->exp_obd->u.lov;
916 int stripe_cnt = lsm->lsm_stripe_count;
921 *arr_cnt = stripe_cnt;
922 OBD_ALLOC(tmp_arr, *arr_cnt * sizeof(int));
925 for (i = 0; i < *arr_cnt; i++)
929 lsm->lsm_oinfo[0]->loi_ost_idx >= lov->desc.ld_tgt_count)
930 rc = alloc_qos(exp, tmp_arr, &stripe_cnt,
931 lsm->lsm_pool_name, flags);
933 rc = alloc_specific(lov, lsm, tmp_arr);
941 OBD_FREE(tmp_arr, *arr_cnt * sizeof(int));
946 static void free_idx_array(int *idx_arr, int arr_cnt)
949 OBD_FREE(idx_arr, arr_cnt * sizeof(int));
952 int qos_prep_create(struct obd_export *exp, struct lov_request_set *set)
954 struct lov_obd *lov = &exp->exp_obd->u.lov;
955 struct lov_stripe_md *lsm;
956 struct obdo *src_oa = set->set_oi->oi_oa;
957 struct obd_trans_info *oti = set->set_oti;
958 int i, stripes, rc = 0, newea = 0;
959 int flag = LOV_USES_ASSIGNED_STRIPE;
960 int *idx_arr = NULL, idx_cnt = 0;
963 LASSERT(src_oa->o_valid & OBD_MD_FLID);
964 LASSERT(src_oa->o_valid & OBD_MD_FLGROUP);
966 if (set->set_oi->oi_md == NULL) {
967 int stripes_def = lov_get_stripecnt(lov, 0);
969 /* If the MDS file was truncated up to some size, stripe over
970 * enough OSTs to allow the file to be created at that size.
971 * This may mean we use more than the default # of stripes. */
972 if (src_oa->o_valid & OBD_MD_FLSIZE) {
973 obd_size min_bavail = LUSTRE_STRIPE_MAXBYTES;
975 /* Find a small number of stripes we can use
976 (up to # of active osts). */
978 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
979 if (!lov->lov_tgts[i] ||
980 !lov->lov_tgts[i]->ltd_active)
982 min_bavail = min(min_bavail, TGT_BAVAIL(i));
983 if (min_bavail * stripes > src_oa->o_size)
988 if (stripes < stripes_def)
989 stripes = stripes_def;
991 flag = LOV_USES_DEFAULT_STRIPE;
992 stripes = stripes_def;
995 rc = lov_alloc_memmd(&set->set_oi->oi_md, stripes,
996 lov->desc.ld_pattern ?
997 lov->desc.ld_pattern : LOV_PATTERN_RAID0,
1005 lsm = set->set_oi->oi_md;
1006 lsm->lsm_object_id = src_oa->o_id;
1007 lsm->lsm_object_gr = src_oa->o_gr;
1009 if (!lsm->lsm_stripe_size)
1010 lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size;
1011 if (!lsm->lsm_pattern) {
1012 LASSERT(lov->desc.ld_pattern);
1013 lsm->lsm_pattern = lov->desc.ld_pattern;
1016 stripes = alloc_idx_array(exp, lsm, newea, &idx_arr, &idx_cnt, flag);
1018 GOTO(out_err, rc = stripes ? stripes : -EIO);
1019 LASSERTF(stripes <= lsm->lsm_stripe_count,"requested %d allocated %d\n",
1020 lsm->lsm_stripe_count, stripes);
1022 for (i = 0; i < stripes; i++) {
1023 struct lov_request *req;
1024 int ost_idx = idx_arr[i];
1025 LASSERT(ost_idx >= 0);
1027 OBD_ALLOC(req, sizeof(*req));
1029 GOTO(out_err, rc = -ENOMEM);
1030 lov_set_add_req(req, set);
1032 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
1033 OBD_ALLOC(req->rq_oi.oi_md, req->rq_buflen);
1034 if (req->rq_oi.oi_md == NULL)
1035 GOTO(out_err, rc = -ENOMEM);
1037 OBDO_ALLOC(req->rq_oi.oi_oa);
1038 if (req->rq_oi.oi_oa == NULL)
1039 GOTO(out_err, rc = -ENOMEM);
1041 req->rq_idx = ost_idx;
1043 /* create data objects with "parent" OA */
1044 memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
1045 req->rq_oi.oi_cb_up = cb_create_update;
1047 /* XXX When we start creating objects on demand, we need to
1048 * make sure that we always create the object on the
1049 * stripe which holds the existing file size.
1051 if (src_oa->o_valid & OBD_MD_FLSIZE) {
1052 req->rq_oi.oi_oa->o_size =
1053 lov_size_to_stripe(lsm, src_oa->o_size, i);
1055 CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
1056 i, req->rq_oi.oi_oa->o_size, src_oa->o_size);
1059 LASSERT(set->set_count == stripes);
1061 if (stripes < lsm->lsm_stripe_count)
1062 qos_shrink_lsm(set);
1063 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_LOV_PREP_CREATE)) {
1064 qos_shrink_lsm(set);
1068 if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) {
1069 oti_alloc_cookies(oti, set->set_count);
1070 if (!oti->oti_logcookies)
1071 GOTO(out_err, rc = -ENOMEM);
1072 set->set_cookies = oti->oti_logcookies;
1076 obd_free_memmd(exp, &set->set_oi->oi_md);
1078 free_idx_array(idx_arr, idx_cnt);
1083 void qos_update(struct lov_obd *lov)
1086 lov->lov_qos.lq_dirty = 1;
1089 void qos_statfs_done(struct lov_obd *lov)
1091 LASSERT(lov->lov_qos.lq_statfs_in_progress);
1092 down_write(&lov->lov_qos.lq_rw_sem);
1093 lov->lov_qos.lq_statfs_in_progress = 0;
1094 /* wake up any threads waiting for the statfs rpcs to complete */
1095 cfs_waitq_signal(&lov->lov_qos.lq_statfs_waitq);
1096 up_write(&lov->lov_qos.lq_rw_sem);
1099 static int qos_statfs_ready(struct obd_device *obd, __u64 max_age)
1101 struct lov_obd *lov = &obd->u.lov;
1104 down_read(&lov->lov_qos.lq_rw_sem);
1105 rc = lov->lov_qos.lq_statfs_in_progress == 0 ||
1106 cfs_time_beforeq_64(max_age, obd->obd_osfs_age);
1107 up_read(&lov->lov_qos.lq_rw_sem);
1112 * Update statfs data if the current osfs age is older than max_age.
1113 * If wait is not set, it means that we are called from lov_create()
1114 * and we should just issue the rpcs without waiting for them to complete.
1115 * If wait is set, we are called from alloc_qos() and we just have
1116 * to wait for the request set to complete.
1118 void qos_statfs_update(struct obd_device *obd, __u64 max_age, int wait)
1120 struct lov_obd *lov = &obd->u.lov;
1121 struct obd_info *oinfo;
1123 struct ptlrpc_request_set *set = NULL;
1126 if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
1127 /* statfs data are quite recent, don't need to refresh it */
1130 if (!wait && lov->lov_qos.lq_statfs_in_progress)
1131 /* statfs already in progress */
1134 down_write(&lov->lov_qos.lq_rw_sem);
1135 if (lov->lov_qos.lq_statfs_in_progress) {
1136 up_write(&lov->lov_qos.lq_rw_sem);
1139 /* no statfs in flight, send rpcs */
1140 lov->lov_qos.lq_statfs_in_progress = 1;
1141 up_write(&lov->lov_qos.lq_rw_sem);
1144 CDEBUG(D_QOS, "%s: did not manage to get fresh statfs data "
1145 "in a timely manner (osfs age "LPU64", max age "LPU64")"
1146 ", sending new statfs rpcs\n",
1147 obd_uuid2str(&lov->desc.ld_uuid), obd->obd_osfs_age,
1150 /* need to send statfs rpcs */
1151 CDEBUG(D_QOS, "sending new statfs requests\n");
1152 memset(lov->lov_qos.lq_statfs_data, 0,
1153 sizeof(*lov->lov_qos.lq_statfs_data));
1154 oinfo = &lov->lov_qos.lq_statfs_data->lsd_oi;
1155 oinfo->oi_osfs = &lov->lov_qos.lq_statfs_data->lsd_statfs;
1156 oinfo->oi_flags = OBD_STATFS_NODELAY;
1157 set = ptlrpc_prep_set();
1159 GOTO(out_failed, rc = -ENOMEM);
1161 rc = obd_statfs_async(obd, oinfo, max_age, set);
1162 if (rc || list_empty(&set->set_requests)) {
1164 CWARN("statfs failed with %d\n", rc);
1165 GOTO(out_failed, rc);
1167 /* send requests via ptlrpcd */
1168 oinfo->oi_flags |= OBD_STATFS_PTLRPCD;
1169 ptlrpcd_add_rqset(set);
1173 down_write(&lov->lov_qos.lq_rw_sem);
1174 lov->lov_qos.lq_statfs_in_progress = 0;
1175 /* wake up any threads waiting for the statfs rpcs to complete */
1176 cfs_waitq_signal(&lov->lov_qos.lq_statfs_waitq);
1177 up_write(&lov->lov_qos.lq_rw_sem);
1181 ptlrpc_set_destroy(set);
1183 struct l_wait_info lwi = { 0 };
1184 CDEBUG(D_QOS, "waiting for statfs requests to complete\n");
1185 l_wait_event(lov->lov_qos.lq_statfs_waitq,
1186 qos_statfs_ready(obd, max_age), &lwi);
1187 if (cfs_time_before_64(obd->obd_osfs_age, max_age))
1188 CDEBUG(D_QOS, "%s: still no fresh statfs data after "
1189 "waiting (osfs age "LPU64", max age "
1191 obd_uuid2str(&lov->desc.ld_uuid),
1192 obd->obd_osfs_age, max_age);