1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2002, 2003 Cluster File Systems, Inc.
6 * This file is part of the Lustre file system, http://www.lustre.org
7 * Lustre is a trademark of Cluster File Systems, Inc.
9 * You may have signed or agreed to another license before downloading
10 * this software. If so, you are bound by the terms and conditions
11 * of that agreement, and the following does not apply to you. See the
12 * LICENSE file included with this distribution for more information.
14 * If you did not agree to a different license, then this copy of Lustre
15 * is open source software; you can redistribute it and/or modify it
16 * under the terms of version 2 of the GNU General Public License as
17 * published by the Free Software Foundation.
19 * In either case, Lustre is distributed in the hope that it will be
20 * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
21 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
22 * license text for more details.
26 # define EXPORT_SYMTAB
28 #define DEBUG_SUBSYSTEM S_LOV
31 #include <libcfs/libcfs.h>
33 #include <liblustre.h>
36 #include <obd_class.h>
38 #include "lov_internal.h"
40 /* #define QOS_DEBUG 1 */
43 #define TGT_BAVAIL(i) (lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bavail *\
44 lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bsize)
45 #define TGT_FFREE(i) (lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_ffree)
48 int qos_add_tgt(struct obd_device *obd, __u32 index)
50 struct lov_obd *lov = &obd->u.lov;
51 struct lov_qos_oss *oss, *temposs;
52 struct obd_export *exp = lov->lov_tgts[index]->ltd_exp;
53 int rc = 0, found = 0;
56 /* We only need this QOS struct on MDT, not clients - but we may not
57 * have registered the LOV's observer yet, so there's no way to know */
58 if (!exp || !exp->exp_connection) {
59 CERROR("Missing connection\n");
63 down_write(&lov->lov_qos.lq_rw_sem);
64 mutex_down(&lov->lov_lock);
65 list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
66 if (obd_uuid_equals(&oss->lqo_uuid,
67 &exp->exp_connection->c_remote_uuid)) {
76 GOTO(out, rc = -ENOMEM);
77 memcpy(&oss->lqo_uuid,
78 &exp->exp_connection->c_remote_uuid,
79 sizeof(oss->lqo_uuid));
81 /* Assume we have to move this one */
82 list_del(&oss->lqo_oss_list);
86 lov->lov_tgts[index]->ltd_qos.ltq_oss = oss;
88 /* Add sorted by # of OSTs. Find the first entry that we're
90 list_for_each_entry(temposs, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
91 if (oss->lqo_ost_count > temposs->lqo_ost_count)
94 /* ...and add before it. If we're the first or smallest, temposs
95 points to the list head, and we add to the end. */
96 list_add_tail(&oss->lqo_oss_list, &temposs->lqo_oss_list);
98 lov->lov_qos.lq_dirty = 1;
99 lov->lov_qos.lq_dirty_rr = 1;
101 CDEBUG(D_QOS, "add tgt %s to OSS %s (%d OSTs)\n",
102 obd_uuid2str(&lov->lov_tgts[index]->ltd_uuid),
103 obd_uuid2str(&oss->lqo_uuid),
107 mutex_up(&lov->lov_lock);
108 up_write(&lov->lov_qos.lq_rw_sem);
112 int qos_del_tgt(struct obd_device *obd, __u32 index)
114 struct lov_obd *lov = &obd->u.lov;
115 struct lov_qos_oss *oss;
119 if (!lov->lov_tgts[index])
122 down_write(&lov->lov_qos.lq_rw_sem);
124 oss = lov->lov_tgts[index]->ltd_qos.ltq_oss;
126 GOTO(out, rc = -ENOENT);
128 oss->lqo_ost_count--;
129 if (oss->lqo_ost_count == 0) {
130 CDEBUG(D_QOS, "removing OSS %s\n",
131 obd_uuid2str(&oss->lqo_uuid));
132 list_del(&oss->lqo_oss_list);
136 lov->lov_qos.lq_dirty = 1;
137 lov->lov_qos.lq_dirty_rr = 1;
139 up_write(&lov->lov_qos.lq_rw_sem);
143 /* Recalculate per-object penalties for OSSs and OSTs,
144 depends on size of each ost in an oss */
145 static int qos_calc_ppo(struct obd_device *obd)
147 struct lov_obd *lov = &obd->u.lov;
148 struct lov_qos_oss *oss;
149 __u64 ba_max, ba_min, temp;
151 int rc, i, prio_wide;
154 if (!lov->lov_qos.lq_dirty)
157 num_active = lov->desc.ld_active_tgt_count - 1;
159 GOTO(out, rc = -EAGAIN);
161 /* find bavail on each OSS */
162 list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
165 lov->lov_qos.lq_active_oss_count = 0;
167 /* How badly user wants to select osts "widely" (not recently chosen
168 and not on recent oss's). As opposed to "freely" (free space
170 prio_wide = 256 - lov->lov_qos.lq_prio_free;
172 ba_min = (__u64)(-1);
174 /* Calculate OST penalty per object */
175 /* (lov ref taken in alloc_qos) */
176 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
177 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
179 temp = TGT_BAVAIL(i);
182 ba_min = min(temp, ba_min);
183 ba_max = max(temp, ba_max);
185 /* Count the number of usable OSS's */
186 if (lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_bavail == 0)
187 lov->lov_qos.lq_active_oss_count++;
188 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_bavail += temp;
190 /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
192 do_div(temp, num_active);
193 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj =
194 (temp * prio_wide) >> 8;
196 if (lov->lov_qos.lq_reset == 0)
197 lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
200 num_active = lov->lov_qos.lq_active_oss_count - 1;
201 if (num_active < 1) {
202 /* If there's only 1 OSS, we can't penalize it, so instead
203 we have to double the OST penalty */
205 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
206 if (lov->lov_tgts[i] == NULL)
208 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj <<= 1;
212 /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
213 list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
214 temp = oss->lqo_bavail >> 1;
215 do_div(temp, oss->lqo_ost_count * num_active);
216 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
217 if (lov->lov_qos.lq_reset == 0)
218 oss->lqo_penalty = 0;
221 lov->lov_qos.lq_dirty = 0;
222 lov->lov_qos.lq_reset = 0;
224 /* If each ost has almost same free space,
225 * do rr allocation for better creation performance */
226 lov->lov_qos.lq_same_space = 0;
227 temp = ba_max - ba_min;
228 ba_min = (ba_min * 51) >> 8; /* 51/256 = .20 */
230 /* Difference is less than 20% */
231 lov->lov_qos.lq_same_space = 1;
232 /* Reset weights for the next time we enter qos mode */
233 lov->lov_qos.lq_reset = 0;
238 if (!rc && lov->lov_qos.lq_same_space)
243 static int qos_calc_weight(struct lov_obd *lov, int i)
247 /* Final ost weight = TGT_BAVAIL - ost_penalty - oss_penalty */
248 temp = TGT_BAVAIL(i);
249 temp2 = lov->lov_tgts[i]->ltd_qos.ltq_penalty +
250 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty;
252 lov->lov_tgts[i]->ltd_qos.ltq_weight = 0;
254 lov->lov_tgts[i]->ltd_qos.ltq_weight = temp - temp2;
258 /* We just used this index for a stripe; adjust everyone's weights */
259 static int qos_used(struct lov_obd *lov, __u32 index, __u64 *total_wt)
261 struct lov_qos_oss *oss;
265 /* Don't allocate from this stripe anymore, until the next alloc_qos */
266 lov->lov_tgts[index]->ltd_qos.ltq_usable = 0;
268 oss = lov->lov_tgts[index]->ltd_qos.ltq_oss;
270 /* Decay old penalty by half (we're adding max penalty, and don't
271 want it to run away.) */
272 lov->lov_tgts[index]->ltd_qos.ltq_penalty >>= 1;
273 oss->lqo_penalty >>= 1;
275 /* Set max penalties for this OST and OSS */
276 lov->lov_tgts[index]->ltd_qos.ltq_penalty +=
277 lov->lov_tgts[index]->ltd_qos.ltq_penalty_per_obj *
278 lov->desc.ld_active_tgt_count;
279 oss->lqo_penalty += oss->lqo_penalty_per_obj *
280 lov->lov_qos.lq_active_oss_count;
282 /* Decrease all OSS penalties */
283 list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
284 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
285 oss->lqo_penalty = 0;
287 oss->lqo_penalty -= oss->lqo_penalty_per_obj;
291 /* Decrease all OST penalties */
292 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
293 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
295 if (lov->lov_tgts[i]->ltd_qos.ltq_penalty <
296 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj)
297 lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
299 lov->lov_tgts[i]->ltd_qos.ltq_penalty -=
300 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj;
302 qos_calc_weight(lov, i);
304 /* Recalc the total weight of usable osts */
305 if (lov->lov_tgts[i]->ltd_qos.ltq_usable)
306 *total_wt += lov->lov_tgts[i]->ltd_qos.ltq_weight;
309 CDEBUG(D_QOS, "recalc tgt %d avail="LPU64
310 " ostppo="LPU64" ostp="LPU64" ossppo="LPU64
311 " ossp="LPU64" wt="LPU64"\n",
312 i, TGT_BAVAIL(i) >> 10,
313 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj >> 10,
314 lov->lov_tgts[i]->ltd_qos.ltq_penalty >> 10,
315 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty_per_obj>>10,
316 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty >> 10,
317 lov->lov_tgts[i]->ltd_qos.ltq_weight >> 10);
324 #define LOV_QOS_EMPTY ((__u32)-1)
325 /* compute optimal round-robin order, based on OSTs per OSS */
326 static int qos_calc_rr(struct lov_obd *lov)
328 struct lov_qos_oss *oss;
329 unsigned ost_count, placed, real_count;
333 if (!lov->lov_qos.lq_dirty_rr) {
334 LASSERT(lov->lov_qos.lq_rr_size);
338 /* Do actual allocation. */
339 down_write(&lov->lov_qos.lq_rw_sem);
342 * Check again. While we were sleeping on @lq_rw_sem something could
345 if (!lov->lov_qos.lq_dirty_rr) {
346 LASSERT(lov->lov_qos.lq_rr_size);
347 up_write(&lov->lov_qos.lq_rw_sem);
351 ost_count = lov->desc.ld_tgt_count;
353 if (lov->lov_qos.lq_rr_size)
354 OBD_FREE(lov->lov_qos.lq_rr_array, lov->lov_qos.lq_rr_size);
355 lov->lov_qos.lq_rr_size = ost_count *
356 sizeof(lov->lov_qos.lq_rr_array[0]);
357 OBD_ALLOC(lov->lov_qos.lq_rr_array, lov->lov_qos.lq_rr_size);
358 if (!lov->lov_qos.lq_rr_array) {
359 lov->lov_qos.lq_rr_size = 0;
360 up_write(&lov->lov_qos.lq_rw_sem);
365 for (i = 0; i < ost_count; i++) {
366 lov->lov_qos.lq_rr_array[i] = LOV_QOS_EMPTY;
367 if (lov->lov_tgts[i])
371 /* Place all the OSTs from 1 OSS at the same time. */
373 list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
375 for (i = 0; i < ost_count; i++) {
376 if (lov->lov_tgts[i] &&
377 lov->lov_tgts[i]->ltd_qos.ltq_oss == oss) {
378 /* Evenly space these OSTs across arrayspace */
379 int next = j * ost_count / oss->lqo_ost_count;
380 LASSERT(next < ost_count);
381 while (lov->lov_qos.lq_rr_array[next] !=
383 next = (next + 1) % ost_count;
384 lov->lov_qos.lq_rr_array[next] = i;
389 LASSERT(j == oss->lqo_ost_count);
392 lov->lov_qos.lq_dirty_rr = 0;
393 up_write(&lov->lov_qos.lq_rw_sem);
395 if (placed != real_count) {
396 /* This should never happen */
397 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
398 "round-robin list (%d of %d).\n",
400 for (i = 0; i < ost_count; i++) {
401 LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
402 lov->lov_qos.lq_rr_array[i]);
404 lov->lov_qos.lq_dirty_rr = 1;
409 for (i = 0; i < ost_count; i++) {
410 LCONSOLE(D_QOS, "rr #%d ost idx=%d\n", i,
411 lov->lov_qos.lq_rr_array[i]);
419 void qos_shrink_lsm(struct lov_request_set *set)
421 struct lov_stripe_md *lsm = set->set_oi->oi_md, *lsm_new;
422 /* XXX LOV STACKING call into osc for sizes */
423 unsigned oldsize, newsize;
425 if (set->set_oti && set->set_cookies && set->set_cookie_sent) {
426 struct llog_cookie *cookies;
427 oldsize = lsm->lsm_stripe_count * sizeof(*cookies);
428 newsize = set->set_count * sizeof(*cookies);
430 cookies = set->set_cookies;
431 oti_alloc_cookies(set->set_oti, set->set_count);
432 if (set->set_oti->oti_logcookies) {
433 memcpy(set->set_oti->oti_logcookies, cookies, newsize);
434 OBD_FREE(cookies, oldsize);
435 set->set_cookies = set->set_oti->oti_logcookies;
437 CWARN("'leaking' %d bytes\n", oldsize - newsize);
441 CWARN("using fewer stripes for object "LPU64": old %u new %u\n",
442 lsm->lsm_object_id, lsm->lsm_stripe_count, set->set_count);
443 LASSERT(lsm->lsm_stripe_count >= set->set_count);
445 newsize = lov_stripe_md_size(set->set_count);
446 OBD_ALLOC(lsm_new, newsize);
447 if (lsm_new != NULL) {
449 memcpy(lsm_new, lsm, sizeof(*lsm));
450 for (i = 0; i < lsm->lsm_stripe_count; i++) {
451 if (i < set->set_count) {
452 lsm_new->lsm_oinfo[i] = lsm->lsm_oinfo[i];
455 OBD_SLAB_FREE(lsm->lsm_oinfo[i], lov_oinfo_slab,
456 sizeof(struct lov_oinfo));
458 lsm_new->lsm_stripe_count = set->set_count;
459 OBD_FREE(lsm, sizeof(struct lov_stripe_md) +
460 lsm->lsm_stripe_count * sizeof(struct lov_oinfo *));
461 set->set_oi->oi_md = lsm_new;
463 CWARN("'leaking' few bytes\n");
467 int qos_remedy_create(struct lov_request_set *set, struct lov_request *req)
469 struct lov_stripe_md *lsm = set->set_oi->oi_md;
470 struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
471 unsigned ost_idx, ost_count = lov->desc.ld_tgt_count;
472 int stripe, i, rc = -EIO;
475 ost_idx = (req->rq_idx + lsm->lsm_stripe_count) % ost_count;
476 for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
477 if (!lov->lov_tgts[ost_idx] ||
478 !lov->lov_tgts[ost_idx]->ltd_active)
480 /* check if objects has been created on this ost */
481 for (stripe = 0; stripe < lsm->lsm_stripe_count; stripe++) {
482 if (stripe == req->rq_stripe)
484 if (ost_idx == lsm->lsm_oinfo[stripe]->loi_ost_idx)
488 if (stripe >= lsm->lsm_stripe_count) {
489 req->rq_idx = ost_idx;
490 rc = obd_create(lov->lov_tgts[ost_idx]->ltd_exp,
491 req->rq_oi.oi_oa, &req->rq_oi.oi_md,
500 static int min_stripe_count(int stripe_cnt, int flags)
502 return (flags & LOV_USES_DEFAULT_STRIPE ?
503 stripe_cnt - (stripe_cnt / 4) : stripe_cnt);
506 #define LOV_CREATE_RESEED_MULT 4
507 #define LOV_CREATE_RESEED_MIN 1000
508 /* Allocate objects on osts with round-robin algorithm */
509 static int alloc_rr(struct lov_obd *lov, int *idx_arr, int *stripe_cnt,
512 unsigned array_idx, ost_count = lov->desc.ld_tgt_count;
513 unsigned ost_active_count = lov->desc.ld_active_tgt_count;
516 int ost_start_idx_temp;
518 int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
521 i = qos_calc_rr(lov);
525 if (--lov->lov_start_count <= 0) {
526 lov->lov_start_idx = ll_rand() % ost_count;
527 lov->lov_start_count =
528 (LOV_CREATE_RESEED_MIN / max(ost_active_count, 1U) +
529 LOV_CREATE_RESEED_MULT) * max(ost_active_count, 1U);
530 } else if (stripe_cnt_min >= ost_active_count ||
531 lov->lov_start_idx > ost_count) {
532 /* If we have allocated from all of the OSTs, slowly
533 * precess the next start if the OST/stripe count isn't
534 * already doing this for us. */
535 lov->lov_start_idx %= ost_count;
536 if (*stripe_cnt > 1 && (ost_active_count % (*stripe_cnt)) != 1)
537 ++lov->lov_offset_idx;
539 down_read(&lov->lov_qos.lq_rw_sem);
540 ost_start_idx_temp = lov->lov_start_idx;
543 array_idx = (lov->lov_start_idx + lov->lov_offset_idx) % ost_count;
546 CDEBUG(D_QOS, "want %d startidx %d startcnt %d offset %d active %d "
547 "count %d arrayidx %d\n",
548 stripe_cnt, lov->lov_start_idx, lov->lov_start_count,
549 lov->lov_offset_idx, ost_active_count, ost_count, array_idx);
552 for (i = 0; i < ost_count; i++, array_idx=(array_idx + 1) % ost_count) {
553 ++lov->lov_start_idx;
554 ost_idx = lov->lov_qos.lq_rr_array[array_idx];
556 CDEBUG(D_QOS, "#%d strt %d act %d strp %d ary %d idx %d\n",
557 i, lov->lov_start_idx,
558 ((ost_idx != LOV_QOS_EMPTY) && lov->lov_tgts[ost_idx]) ?
559 lov->lov_tgts[ost_idx]->ltd_active : 0,
560 idx_pos - idx_arr, array_idx, ost_idx);
562 if ((ost_idx == LOV_QOS_EMPTY) || !lov->lov_tgts[ost_idx] ||
563 !lov->lov_tgts[ost_idx]->ltd_active)
566 /* Fail Check before osc_precreate() is called
567 so we can only 'fail' single OSC. */
568 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
571 /* Drop slow OSCs if we can */
572 if (obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed)
577 /* We have enough stripes */
578 if (idx_pos - idx_arr == *stripe_cnt)
581 if ((speed < 2) && (idx_pos - idx_arr < stripe_cnt_min)) {
582 /* Try again, allowing slower OSCs */
584 lov->lov_start_idx = ost_start_idx_temp;
588 up_read(&lov->lov_qos.lq_rw_sem);
590 *stripe_cnt = idx_pos - idx_arr;
594 /* alloc objects on osts with specific stripe offset */
595 static int alloc_specific(struct lov_obd *lov, struct lov_stripe_md *lsm,
598 unsigned ost_idx, ost_count = lov->desc.ld_tgt_count;
604 ost_idx = lsm->lsm_oinfo[0]->loi_ost_idx;
606 for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
607 if (!lov->lov_tgts[ost_idx] ||
608 !lov->lov_tgts[ost_idx]->ltd_active) {
612 /* Fail Check before osc_precreate() is called
613 so we can only 'fail' single OSC. */
614 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
617 /* Drop slow OSCs if we can, but not for requested start idx */
618 if ((obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed) &&
619 (i != 0 || speed < 2))
624 /* We have enough stripes */
625 if (idx_pos - idx_arr == lsm->lsm_stripe_count)
629 /* Try again, allowing slower OSCs */
634 /* If we were passed specific striping params, then a failure to
635 * meet those requirements is an error, since we can't reallocate
636 * that memory (it might be part of a larger array or something).
638 * We can only get here if lsm_stripe_count was originally > 1.
640 CERROR("can't lstripe objid "LPX64": have %d want %u\n",
641 lsm->lsm_object_id, (int)(idx_pos - idx_arr),
642 lsm->lsm_stripe_count);
646 /* Alloc objects on osts with optimization based on:
648 - network resources (shared OSS's)
650 static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt,
653 struct lov_obd *lov = &exp->exp_obd->u.lov;
654 static time_t last_warn = 0;
655 time_t now = cfs_time_current_sec();
656 __u64 total_bavail, total_weight = 0;
658 int nfound, good_osts, i, warn = 0, rc = 0;
659 int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
662 if (stripe_cnt_min < 1)
663 GOTO(out_nolock, rc = -EINVAL);
665 lov_getref(exp->exp_obd);
667 /* Detect -EAGAIN early, before expensive lock is taken. */
668 if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space)
669 GOTO(out_nolock, rc = -EAGAIN);
671 /* Do actual allocation, use write lock here. */
672 down_write(&lov->lov_qos.lq_rw_sem);
675 * Check again, while we were sleeping on @lq_rw_sem things could
678 if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space)
679 GOTO(out, rc = -EAGAIN);
681 ost_count = lov->desc.ld_tgt_count;
683 if (lov->desc.ld_active_tgt_count < 2)
684 GOTO(out, rc = -EAGAIN);
686 rc = qos_calc_ppo(exp->exp_obd);
692 /* Warn users about zero available space/inode every 30 min */
693 if (cfs_time_sub(now, last_warn) > 60 * 30)
695 /* Find all the OSTs that are valid stripe candidates */
696 for (i = 0; i < ost_count; i++) {
699 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
701 bavail = TGT_BAVAIL(i);
704 CDEBUG(D_QOS, "no free space on %s\n",
705 obd_uuid2str(&lov->lov_tgts[i]->ltd_uuid));
712 CDEBUG(D_QOS, "no free inodes on %s\n",
713 obd_uuid2str(&lov->lov_tgts[i]->ltd_uuid));
719 /* Fail Check before osc_precreate() is called
720 so we can only 'fail' single OSC. */
721 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && i == 0)
724 if (obd_precreate(lov->lov_tgts[i]->ltd_exp) > 2)
727 lov->lov_tgts[i]->ltd_qos.ltq_usable = 1;
728 qos_calc_weight(lov, i);
729 total_bavail += bavail;
730 total_weight += lov->lov_tgts[i]->ltd_qos.ltq_weight;
735 if (good_osts < stripe_cnt_min)
736 GOTO(out, rc = -EAGAIN);
739 GOTO(out, rc = -ENOSPC);
741 /* We have enough osts */
742 if (good_osts < *stripe_cnt)
743 *stripe_cnt = good_osts;
746 GOTO(out, rc = -EAGAIN);
748 /* Find enough OSTs with weighted random allocation. */
750 while (nfound < *stripe_cnt) {
751 __u64 rand, cur_weight;
757 #if BITS_PER_LONG == 32
758 rand = ll_rand() % (unsigned)total_weight;
759 /* If total_weight > 32-bit, first generate the high
760 * 32 bits of the random number, then add in the low
761 * 32 bits (truncated to the upper limit, if needed) */
762 if (total_weight > 0xffffffffULL)
763 rand = (__u64)(ll_rand() %
764 (unsigned)(total_weight >> 32)) << 32;
768 if (rand == (total_weight & 0xffffffff00000000ULL))
769 rand |= ll_rand() % (unsigned)total_weight;
774 rand = ((__u64)ll_rand() << 32 | ll_rand()) %
781 /* On average, this will hit larger-weighted osts more often.
782 0-weight osts will always get used last (only when rand=0).*/
783 for (i = 0; i < ost_count; i++) {
784 if (!lov->lov_tgts[i] ||
785 !lov->lov_tgts[i]->ltd_qos.ltq_usable)
788 cur_weight += lov->lov_tgts[i]->ltd_qos.ltq_weight;
789 if (cur_weight >= rand) {
791 CDEBUG(D_QOS, "assigned stripe=%d to idx=%d\n",
794 idx_arr[nfound++] = i;
795 qos_used(lov, i, &total_weight);
800 /* should never satisfy below condition */
802 CERROR("Didn't find any OSTs?\n");
806 LASSERT(nfound == *stripe_cnt);
809 up_write(&lov->lov_qos.lq_rw_sem);
813 rc = alloc_rr(lov, idx_arr, stripe_cnt, flags);
815 lov_putref(exp->exp_obd);
819 /* return new alloced stripe count on success */
820 static int alloc_idx_array(struct obd_export *exp, struct lov_stripe_md *lsm,
821 int newea, int **idx_arr, int *arr_cnt, int flags)
823 struct lov_obd *lov = &exp->exp_obd->u.lov;
824 int stripe_cnt = lsm->lsm_stripe_count;
829 *arr_cnt = stripe_cnt;
830 OBD_ALLOC(tmp_arr, *arr_cnt * sizeof(int));
833 for (i = 0; i < *arr_cnt; i++)
837 lsm->lsm_oinfo[0]->loi_ost_idx >= lov->desc.ld_tgt_count)
838 rc = alloc_qos(exp, tmp_arr, &stripe_cnt, flags);
840 rc = alloc_specific(lov, lsm, tmp_arr);
848 OBD_FREE(tmp_arr, *arr_cnt * sizeof(int));
853 static void free_idx_array(int *idx_arr, int arr_cnt)
856 OBD_FREE(idx_arr, arr_cnt * sizeof(int));
859 int qos_prep_create(struct obd_export *exp, struct lov_request_set *set)
861 struct lov_obd *lov = &exp->exp_obd->u.lov;
862 struct lov_stripe_md *lsm;
863 struct obdo *src_oa = set->set_oi->oi_oa;
864 struct obd_trans_info *oti = set->set_oti;
865 int i, stripes, rc = 0, newea = 0;
866 int flag = LOV_USES_ASSIGNED_STRIPE;
867 int *idx_arr = NULL, idx_cnt = 0;
870 LASSERT(src_oa->o_valid & OBD_MD_FLID);
871 LASSERT(src_oa->o_valid & OBD_MD_FLGROUP);
873 if (set->set_oi->oi_md == NULL) {
874 int stripes_def = lov_get_stripecnt(lov, 0);
876 /* If the MDS file was truncated up to some size, stripe over
877 * enough OSTs to allow the file to be created at that size.
878 * This may mean we use more than the default # of stripes. */
879 if (src_oa->o_valid & OBD_MD_FLSIZE) {
880 obd_size min_bavail = LUSTRE_STRIPE_MAXBYTES;
882 /* Find a small number of stripes we can use
883 (up to # of active osts). */
885 lov_getref(exp->exp_obd);
886 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
887 if (!lov->lov_tgts[i] ||
888 !lov->lov_tgts[i]->ltd_active)
890 min_bavail = min(min_bavail, TGT_BAVAIL(i));
891 if (min_bavail * stripes > src_oa->o_size)
895 lov_putref(exp->exp_obd);
897 if (stripes < stripes_def)
898 stripes = stripes_def;
900 flag = LOV_USES_DEFAULT_STRIPE;
901 stripes = stripes_def;
904 rc = lov_alloc_memmd(&set->set_oi->oi_md, stripes,
905 lov->desc.ld_pattern ?
906 lov->desc.ld_pattern : LOV_PATTERN_RAID0,
914 lsm = set->set_oi->oi_md;
915 lsm->lsm_object_id = src_oa->o_id;
916 lsm->lsm_object_gr = src_oa->o_gr;
918 if (!lsm->lsm_stripe_size)
919 lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size;
920 if (!lsm->lsm_pattern) {
921 LASSERT(lov->desc.ld_pattern);
922 lsm->lsm_pattern = lov->desc.ld_pattern;
925 stripes = alloc_idx_array(exp, lsm, newea, &idx_arr, &idx_cnt, flag);
927 GOTO(out_err, rc = stripes ? stripes : -EIO);
928 LASSERTF(stripes <= lsm->lsm_stripe_count,"requested %d allocated %d\n",
929 lsm->lsm_stripe_count, stripes);
931 for (i = 0; i < stripes; i++) {
932 struct lov_request *req;
933 int ost_idx = idx_arr[i];
934 LASSERT(ost_idx >= 0);
936 OBD_ALLOC(req, sizeof(*req));
938 GOTO(out_err, rc = -ENOMEM);
939 lov_set_add_req(req, set);
941 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
942 OBD_ALLOC(req->rq_oi.oi_md, req->rq_buflen);
943 if (req->rq_oi.oi_md == NULL)
944 GOTO(out_err, rc = -ENOMEM);
946 OBDO_ALLOC(req->rq_oi.oi_oa);
947 if (req->rq_oi.oi_oa == NULL)
948 GOTO(out_err, rc = -ENOMEM);
950 req->rq_idx = ost_idx;
952 /* create data objects with "parent" OA */
953 memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
955 /* XXX When we start creating objects on demand, we need to
956 * make sure that we always create the object on the
957 * stripe which holds the existing file size.
959 if (src_oa->o_valid & OBD_MD_FLSIZE) {
960 req->rq_oi.oi_oa->o_size =
961 lov_size_to_stripe(lsm, src_oa->o_size, i);
963 CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
964 i, req->rq_oi.oi_oa->o_size, src_oa->o_size);
967 LASSERT(set->set_count == stripes);
969 if (stripes < lsm->lsm_stripe_count)
972 if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) {
973 oti_alloc_cookies(oti, set->set_count);
974 if (!oti->oti_logcookies)
975 GOTO(out_err, rc = -ENOMEM);
976 set->set_cookies = oti->oti_logcookies;
980 obd_free_memmd(exp, &set->set_oi->oi_md);
982 free_idx_array(idx_arr, idx_cnt);
987 void qos_update(struct lov_obd *lov)
990 lov->lov_qos.lq_dirty = 1;