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])
207 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj <<= 1;
210 /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
211 list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
212 temp = oss->lqo_bavail >> 1;
213 do_div(temp, oss->lqo_ost_count * num_active);
214 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
215 if (lov->lov_qos.lq_reset == 0)
216 oss->lqo_penalty = 0;
219 lov->lov_qos.lq_dirty = 0;
220 lov->lov_qos.lq_reset = 0;
222 /* If each ost has almost same free space,
223 * do rr allocation for better creation performance */
224 lov->lov_qos.lq_same_space = 0;
225 temp = ba_max - ba_min;
226 ba_min = (ba_min * 51) >> 8; /* 51/256 = .20 */
228 /* Difference is less than 20% */
229 lov->lov_qos.lq_same_space = 1;
230 /* Reset weights for the next time we enter qos mode */
231 lov->lov_qos.lq_reset = 0;
236 if (!rc && lov->lov_qos.lq_same_space)
241 static int qos_calc_weight(struct lov_obd *lov, int i)
245 /* Final ost weight = TGT_BAVAIL - ost_penalty - oss_penalty */
246 temp = TGT_BAVAIL(i);
247 temp2 = lov->lov_tgts[i]->ltd_qos.ltq_penalty +
248 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty;
250 lov->lov_tgts[i]->ltd_qos.ltq_weight = 0;
252 lov->lov_tgts[i]->ltd_qos.ltq_weight = temp - temp2;
256 /* We just used this index for a stripe; adjust everyone's weights */
257 static int qos_used(struct lov_obd *lov, __u32 index, __u64 *total_wt)
259 struct lov_qos_oss *oss;
263 /* Don't allocate from this stripe anymore, until the next alloc_qos */
264 lov->lov_tgts[index]->ltd_qos.ltq_usable = 0;
266 oss = lov->lov_tgts[index]->ltd_qos.ltq_oss;
268 /* Decay old penalty by half (we're adding max penalty, and don't
269 want it to run away.) */
270 lov->lov_tgts[index]->ltd_qos.ltq_penalty >>= 1;
271 oss->lqo_penalty >>= 1;
273 /* Set max penalties for this OST and OSS */
274 lov->lov_tgts[index]->ltd_qos.ltq_penalty +=
275 lov->lov_tgts[index]->ltd_qos.ltq_penalty_per_obj *
276 lov->desc.ld_active_tgt_count;
277 oss->lqo_penalty += oss->lqo_penalty_per_obj *
278 lov->lov_qos.lq_active_oss_count;
280 /* Decrease all OSS penalties */
281 list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
282 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
283 oss->lqo_penalty = 0;
285 oss->lqo_penalty -= oss->lqo_penalty_per_obj;
289 /* Decrease all OST penalties */
290 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
291 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
293 if (lov->lov_tgts[i]->ltd_qos.ltq_penalty <
294 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj)
295 lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
297 lov->lov_tgts[i]->ltd_qos.ltq_penalty -=
298 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj;
300 qos_calc_weight(lov, i);
302 /* Recalc the total weight of usable osts */
303 if (lov->lov_tgts[i]->ltd_qos.ltq_usable)
304 *total_wt += lov->lov_tgts[i]->ltd_qos.ltq_weight;
307 CDEBUG(D_QOS, "recalc tgt %d avail="LPU64
308 " ostppo="LPU64" ostp="LPU64" ossppo="LPU64
309 " ossp="LPU64" wt="LPU64"\n",
310 i, TGT_BAVAIL(i) >> 10,
311 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj >> 10,
312 lov->lov_tgts[i]->ltd_qos.ltq_penalty >> 10,
313 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty_per_obj>>10,
314 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty>>10,
315 lov->lov_tgts[i]->ltd_qos.ltq_weight>>10);
322 #define LOV_QOS_EMPTY ((__u32)-1)
323 /* compute optimal round-robin order, based on OSTs per OSS */
324 static int qos_calc_rr(struct lov_obd *lov)
326 struct lov_qos_oss *oss;
327 unsigned ost_count, placed, real_count;
331 if (!lov->lov_qos.lq_dirty_rr) {
332 LASSERT(lov->lov_qos.lq_rr_size);
336 down_write(&lov->lov_qos.lq_rw_sem);
337 ost_count = lov->desc.ld_tgt_count;
339 if (lov->lov_qos.lq_rr_size)
340 OBD_FREE(lov->lov_qos.lq_rr_array, lov->lov_qos.lq_rr_size);
341 lov->lov_qos.lq_rr_size = ost_count *
342 sizeof(lov->lov_qos.lq_rr_array[0]);
343 OBD_ALLOC(lov->lov_qos.lq_rr_array, lov->lov_qos.lq_rr_size);
344 if (!lov->lov_qos.lq_rr_array) {
345 lov->lov_qos.lq_rr_size = 0;
346 up_write(&lov->lov_qos.lq_rw_sem);
351 for (i = 0; i < ost_count; i++) {
352 lov->lov_qos.lq_rr_array[i] = LOV_QOS_EMPTY;
353 if (lov->lov_tgts[i])
357 /* Place all the OSTs from 1 OSS at the same time. */
359 list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
361 for (i = 0; i < ost_count; i++) {
362 if (lov->lov_tgts[i] &&
363 (lov->lov_tgts[i]->ltd_qos.ltq_oss == oss)) {
364 /* Evenly space these OSTs across arrayspace */
365 int next = j * ost_count / oss->lqo_ost_count;
366 while (lov->lov_qos.lq_rr_array[next] !=
368 next = (next + 1) % ost_count;
369 lov->lov_qos.lq_rr_array[next] = i;
374 LASSERT(j == oss->lqo_ost_count);
377 lov->lov_qos.lq_dirty_rr = 0;
378 up_write(&lov->lov_qos.lq_rw_sem);
380 if (placed != real_count) {
381 /* This should never happen */
382 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
383 "round-robin list (%d of %d).\n",
385 for (i = 0; i < ost_count; i++) {
386 LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
387 lov->lov_qos.lq_rr_array[i]);
389 lov->lov_qos.lq_dirty_rr = 1;
394 for (i = 0; i < ost_count; i++) {
395 LCONSOLE(D_QOS, "rr #%d ost idx=%d\n", i,
396 lov->lov_qos.lq_rr_array[i]);
404 void qos_shrink_lsm(struct lov_request_set *set)
406 struct lov_stripe_md *lsm = set->set_oi->oi_md, *lsm_new;
407 /* XXX LOV STACKING call into osc for sizes */
408 unsigned oldsize, newsize;
410 if (set->set_oti && set->set_cookies && set->set_cookie_sent) {
411 struct llog_cookie *cookies;
412 oldsize = lsm->lsm_stripe_count * sizeof(*cookies);
413 newsize = set->set_count * sizeof(*cookies);
415 cookies = set->set_cookies;
416 oti_alloc_cookies(set->set_oti, set->set_count);
417 if (set->set_oti->oti_logcookies) {
418 memcpy(set->set_oti->oti_logcookies, cookies, newsize);
419 OBD_FREE(cookies, oldsize);
420 set->set_cookies = set->set_oti->oti_logcookies;
422 CWARN("'leaking' %d bytes\n", oldsize - newsize);
426 CWARN("using fewer stripes for object "LPU64": old %u new %u\n",
427 lsm->lsm_object_id, lsm->lsm_stripe_count, set->set_count);
428 LASSERT(lsm->lsm_stripe_count >= set->set_count);
430 newsize = lov_stripe_md_size(set->set_count);
431 OBD_ALLOC(lsm_new, newsize);
432 if (lsm_new != NULL) {
434 memcpy(lsm_new, lsm, sizeof(*lsm));
435 for (i = 0; i < lsm->lsm_stripe_count; i++) {
436 if (i < set->set_count) {
437 lsm_new->lsm_oinfo[i] = lsm->lsm_oinfo[i];
440 OBD_SLAB_FREE(lsm->lsm_oinfo[i], lov_oinfo_slab,
441 sizeof(struct lov_oinfo));
443 lsm_new->lsm_stripe_count = set->set_count;
444 OBD_FREE(lsm, sizeof(struct lov_stripe_md) +
445 lsm->lsm_stripe_count * sizeof(struct lov_oinfo *));
446 set->set_oi->oi_md = lsm_new;
448 CWARN("'leaking' few bytes\n");
452 int qos_remedy_create(struct lov_request_set *set, struct lov_request *req)
454 struct lov_stripe_md *lsm = set->set_oi->oi_md;
455 struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
456 unsigned ost_idx, ost_count = lov->desc.ld_tgt_count;
457 int stripe, i, rc = -EIO;
460 ost_idx = (req->rq_idx + lsm->lsm_stripe_count) % ost_count;
461 for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
462 if (!lov->lov_tgts[ost_idx] ||
463 !lov->lov_tgts[ost_idx]->ltd_active)
465 /* check if objects has been created on this ost */
466 for (stripe = 0; stripe < lsm->lsm_stripe_count; stripe++) {
467 if (stripe == req->rq_stripe)
469 if (ost_idx == lsm->lsm_oinfo[stripe]->loi_ost_idx)
473 if (stripe >= lsm->lsm_stripe_count) {
474 req->rq_idx = ost_idx;
475 rc = obd_create(lov->lov_tgts[ost_idx]->ltd_exp,
476 req->rq_oi.oi_oa, &req->rq_oi.oi_md,
485 static int min_stripe_count(int stripe_cnt, int flags)
487 return (flags & LOV_USES_DEFAULT_STRIPE ?
488 stripe_cnt - (stripe_cnt / 4) : stripe_cnt);
491 #define LOV_CREATE_RESEED_MULT 4
492 #define LOV_CREATE_RESEED_MIN 1000
493 /* Allocate objects on osts with round-robin algorithm */
494 static int alloc_rr(struct lov_obd *lov, int *idx_arr, int *stripe_cnt,
497 unsigned array_idx, ost_count = lov->desc.ld_tgt_count;
498 unsigned ost_active_count = lov->desc.ld_active_tgt_count;
501 int ost_start_idx_temp;
503 int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
506 i = qos_calc_rr(lov);
510 if (--lov->lov_start_count <= 0) {
511 lov->lov_start_idx = ll_rand() % ost_count;
512 lov->lov_start_count =
513 (LOV_CREATE_RESEED_MIN / max(ost_active_count, 1U) +
514 LOV_CREATE_RESEED_MULT) * max(ost_active_count, 1U);
515 } else if (stripe_cnt_min >= ost_active_count ||
516 lov->lov_start_idx > ost_count) {
517 /* If we have allocated from all of the OSTs, slowly
518 precess the next start */
519 lov->lov_start_idx %= ost_count;
520 if (*stripe_cnt > 1 && (ost_active_count % (*stripe_cnt) != 1))
521 ++lov->lov_offset_idx;
523 down_read(&lov->lov_qos.lq_rw_sem);
524 ost_start_idx_temp = lov->lov_start_idx;
527 array_idx = (lov->lov_start_idx + lov->lov_offset_idx) % ost_count;
530 CDEBUG(D_QOS, "want %d startidx %d startcnt %d offset %d arrayidx %d\n",
531 stripe_cnt_min, lov->lov_start_idx, lov->lov_start_count,
532 lov->lov_offset_idx, array_idx);
535 for (i = 0; i < ost_count; i++, array_idx=(array_idx + 1) % ost_count) {
536 ++lov->lov_start_idx;
537 ost_idx = lov->lov_qos.lq_rr_array[array_idx];
539 CDEBUG(D_QOS, "#%d strt %d act %d strp %d ary %d idx %d\n",
540 i, lov->lov_start_idx,
541 ((ost_idx != LOV_QOS_EMPTY) && lov->lov_tgts[ost_idx]) ?
542 lov->lov_tgts[ost_idx]->ltd_active : 0,
543 idx_pos - idx_arr, array_idx, ost_idx);
545 if ((ost_idx == LOV_QOS_EMPTY) || !lov->lov_tgts[ost_idx] ||
546 !lov->lov_tgts[ost_idx]->ltd_active)
549 /* Fail Check before osc_precreate() is called
550 so we can only 'fail' single OSC. */
551 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
554 /* Drop slow OSCs if we can */
555 if (obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp, speed == 0) >
561 /* We have enough stripes */
562 if (idx_pos - idx_arr == *stripe_cnt)
565 if ((speed < 2) && (idx_pos - idx_arr < stripe_cnt_min)) {
566 /* Try again, allowing slower OSCs */
568 lov->lov_start_idx = ost_start_idx_temp;
572 up_read(&lov->lov_qos.lq_rw_sem);
574 *stripe_cnt = idx_pos - idx_arr;
578 /* alloc objects on osts with specific stripe offset */
579 static int alloc_specific(struct lov_obd *lov, struct lov_stripe_md *lsm,
582 unsigned ost_idx, ost_count = lov->desc.ld_tgt_count;
588 ost_idx = lsm->lsm_oinfo[0]->loi_ost_idx;
590 for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
591 if (!lov->lov_tgts[ost_idx] ||
592 !lov->lov_tgts[ost_idx]->ltd_active) {
596 /* Fail Check before osc_precreate() is called
597 so we can only 'fail' single OSC. */
598 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
601 /* Drop slow OSCs if we can */
602 if (obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp, speed == 0) >
608 /* We have enough stripes */
609 if (idx_pos - idx_arr == lsm->lsm_stripe_count)
613 /* Try again, allowing slower OSCs */
618 /* If we were passed specific striping params, then a failure to
619 * meet those requirements is an error, since we can't reallocate
620 * that memory (it might be part of a larger array or something).
622 * We can only get here if lsm_stripe_count was originally > 1.
624 CERROR("can't lstripe objid "LPX64": have "LPSZ" want %u\n",
625 lsm->lsm_object_id, idx_pos - idx_arr, lsm->lsm_stripe_count);
629 /* Alloc objects on osts with optimization based on:
631 - network resources (shared OSS's)
633 static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt,
636 struct lov_obd *lov = &exp->exp_obd->u.lov;
637 static time_t last_warn = 0;
638 time_t now = cfs_time_current_sec();
639 __u64 total_bavail, total_weight = 0;
641 int nfound, good_osts, i, warn = 0, rc = 0;
642 int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
645 if (stripe_cnt_min < 1)
646 GOTO(out, rc = -EINVAL);
648 lov_getref(exp->exp_obd);
649 down_write(&lov->lov_qos.lq_rw_sem);
651 ost_count = lov->desc.ld_tgt_count;
653 if (lov->desc.ld_active_tgt_count < 2)
654 GOTO(out, rc = -EAGAIN);
656 rc = qos_calc_ppo(exp->exp_obd);
662 /* Warn users about zero available space/inode every 30 min */
663 if (cfs_time_sub(now, last_warn) > 60 * 30)
665 /* Find all the OSTs that are valid stripe candidates */
666 for (i = 0; i < ost_count; i++) {
669 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
671 bavail = TGT_BAVAIL(i);
674 CDEBUG(D_QOS, "no free space on %s\n",
675 obd_uuid2str(&lov->lov_tgts[i]->ltd_uuid));
682 CDEBUG(D_QOS, "no free inodes on %s\n",
683 obd_uuid2str(&lov->lov_tgts[i]->ltd_uuid));
689 /* Fail Check before osc_precreate() is called
690 so we can only 'fail' single OSC. */
691 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && i == 0)
694 if (obd_precreate(lov->lov_tgts[i]->ltd_exp, 1) >= 2)
697 lov->lov_tgts[i]->ltd_qos.ltq_usable = 1;
698 qos_calc_weight(lov, i);
699 total_bavail += bavail;
700 total_weight += lov->lov_tgts[i]->ltd_qos.ltq_weight;
706 GOTO(out, rc = -ENOSPC);
708 if (good_osts < stripe_cnt_min)
709 GOTO(out, rc = -EAGAIN);
711 /* We have enough osts */
712 if (good_osts < *stripe_cnt)
713 *stripe_cnt = good_osts;
715 /* Find enough OSTs with weighted random allocation. */
717 while (nfound < *stripe_cnt) {
718 __u64 rand, cur_weight;
724 #if BITS_PER_LONG == 32
725 rand = ll_rand() % (unsigned)total_weight;
726 /* If total_weight > 32-bit, first generate the high
727 * 32 bits of the random number, then add in the low
728 * 32 bits (truncated to the upper limit, if needed) */
729 if (total_weight > 0xffffffffULL)
730 rand = (__u64)(ll_rand() %
731 (unsigned)(total_weight >> 32)) << 32;
735 if (rand == (total_weight & 0xffffffff00000000ULL))
736 rand |= ll_rand() % (unsigned)total_weight;
741 rand = ((__u64)ll_rand() << 32 | ll_rand()) %
748 /* On average, this will hit larger-weighted osts more often.
749 0-weight osts will always get used last (only when rand=0).*/
750 for (i = 0; i < ost_count; i++) {
751 if (!lov->lov_tgts[i] ||
752 !lov->lov_tgts[i]->ltd_qos.ltq_usable)
755 cur_weight += lov->lov_tgts[i]->ltd_qos.ltq_weight;
756 if (cur_weight >= rand) {
758 CDEBUG(D_QOS, "assigned stripe=%d to idx=%d\n",
761 idx_arr[nfound++] = i;
762 qos_used(lov, i, &total_weight);
767 /* should never satisfy below condition */
769 CERROR("Didn't find any OSTs?\n");
773 LASSERT(nfound == *stripe_cnt);
776 up_write(&lov->lov_qos.lq_rw_sem);
779 rc = alloc_rr(lov, idx_arr, stripe_cnt, flags);
781 lov_putref(exp->exp_obd);
785 /* return new alloced stripe count on success */
786 static int alloc_idx_array(struct obd_export *exp, struct lov_stripe_md *lsm,
787 int newea, int **idx_arr, int *arr_cnt, int flags)
789 struct lov_obd *lov = &exp->exp_obd->u.lov;
790 int stripe_cnt = lsm->lsm_stripe_count;
795 *arr_cnt = stripe_cnt;
796 OBD_ALLOC(tmp_arr, *arr_cnt * sizeof(int));
799 for (i = 0; i < *arr_cnt; i++)
803 lsm->lsm_oinfo[0]->loi_ost_idx >= lov->desc.ld_tgt_count)
804 rc = alloc_qos(exp, tmp_arr, &stripe_cnt, flags);
806 rc = alloc_specific(lov, lsm, tmp_arr);
814 OBD_FREE(tmp_arr, *arr_cnt * sizeof(int));
819 static void free_idx_array(int *idx_arr, int arr_cnt)
822 OBD_FREE(idx_arr, arr_cnt * sizeof(int));
825 int qos_prep_create(struct obd_export *exp, struct lov_request_set *set)
827 struct lov_obd *lov = &exp->exp_obd->u.lov;
828 struct lov_stripe_md *lsm;
829 struct obdo *src_oa = set->set_oi->oi_oa;
830 struct obd_trans_info *oti = set->set_oti;
831 int i, stripes, rc = 0, newea = 0;
832 int flag = LOV_USES_ASSIGNED_STRIPE;
833 int *idx_arr = NULL, idx_cnt = 0;
836 LASSERT(src_oa->o_valid & OBD_MD_FLID);
838 if (set->set_oi->oi_md == NULL) {
839 int stripes_def = lov_get_stripecnt(lov, 0);
841 /* If the MDS file was truncated up to some size, stripe over
842 * enough OSTs to allow the file to be created at that size.
843 * This may mean we use more than the default # of stripes. */
844 if (src_oa->o_valid & OBD_MD_FLSIZE) {
845 obd_size min_bavail = LUSTRE_STRIPE_MAXBYTES;
847 /* Find a small number of stripes we can use
848 (up to # of active osts). */
850 lov_getref(exp->exp_obd);
851 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
852 if (!lov->lov_tgts[i] ||
853 !lov->lov_tgts[i]->ltd_active)
855 min_bavail = min(min_bavail, TGT_BAVAIL(i));
856 if (min_bavail * stripes > src_oa->o_size)
860 lov_putref(exp->exp_obd);
862 if (stripes < stripes_def)
863 stripes = stripes_def;
865 flag = LOV_USES_DEFAULT_STRIPE;
866 stripes = stripes_def;
869 rc = lov_alloc_memmd(&set->set_oi->oi_md, stripes,
870 lov->desc.ld_pattern ?
871 lov->desc.ld_pattern : LOV_PATTERN_RAID0,
879 lsm = set->set_oi->oi_md;
880 lsm->lsm_object_id = src_oa->o_id;
881 if (!lsm->lsm_stripe_size)
882 lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size;
883 if (!lsm->lsm_pattern) {
884 LASSERT(lov->desc.ld_pattern);
885 lsm->lsm_pattern = lov->desc.ld_pattern;
888 stripes = alloc_idx_array(exp, lsm, newea, &idx_arr, &idx_cnt, flag);
890 GOTO(out_err, rc = stripes ? stripes : -EIO);
891 LASSERTF(stripes <= lsm->lsm_stripe_count,"requested %d allocated %d\n",
892 lsm->lsm_stripe_count, stripes);
894 for (i = 0; i < stripes; i++) {
895 struct lov_request *req;
896 int ost_idx = idx_arr[i];
897 LASSERT(ost_idx >= 0);
899 OBD_ALLOC(req, sizeof(*req));
901 GOTO(out_err, rc = -ENOMEM);
902 lov_set_add_req(req, set);
904 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
905 OBD_ALLOC(req->rq_oi.oi_md, req->rq_buflen);
906 if (req->rq_oi.oi_md == NULL)
907 GOTO(out_err, rc = -ENOMEM);
909 OBDO_ALLOC(req->rq_oi.oi_oa);
910 if (req->rq_oi.oi_oa == NULL)
911 GOTO(out_err, rc = -ENOMEM);
913 req->rq_idx = ost_idx;
915 /* create data objects with "parent" OA */
916 memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
918 /* XXX When we start creating objects on demand, we need to
919 * make sure that we always create the object on the
920 * stripe which holds the existing file size.
922 if (src_oa->o_valid & OBD_MD_FLSIZE) {
923 req->rq_oi.oi_oa->o_size =
924 lov_size_to_stripe(lsm, src_oa->o_size, i);
926 CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
927 i, req->rq_oi.oi_oa->o_size, src_oa->o_size);
931 LASSERT(set->set_count == stripes);
933 if (stripes < lsm->lsm_stripe_count)
936 if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) {
937 oti_alloc_cookies(oti, set->set_count);
938 if (!oti->oti_logcookies)
939 GOTO(out_err, rc = -ENOMEM);
940 set->set_cookies = oti->oti_logcookies;
944 obd_free_memmd(exp, &set->set_oi->oi_md);
945 free_idx_array(idx_arr, idx_cnt);
950 void qos_update(struct lov_obd *lov)
953 lov->lov_qos.lq_dirty = 1;