4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License version 2 for more details. A copy is
14 * included in the COPYING file that accompanied this code.
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, write to the Free Software
18 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
23 * Copyright 2009 Sun Microsystems, Inc. All rights reserved
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/lod/lod_qos.c
34 * Implementation of different allocation algorithm used
35 * to distribute objects and data among OSTs.
38 #define DEBUG_SUBSYSTEM S_LOV
40 #include <asm/div64.h>
41 #include <libcfs/libcfs.h>
42 #include <uapi/linux/lustre/lustre_idl.h>
43 #include <lustre_swab.h>
44 #include <obd_class.h>
46 #include "lod_internal.h"
49 * force QoS policy (not RR) to be used for testing purposes
55 #define QOS_DEBUG(fmt, ...) CDEBUG(D_QOS, fmt, ## __VA_ARGS__)
56 #define QOS_CONSOLE(fmt, ...) LCONSOLE(D_QOS, fmt, ## __VA_ARGS__)
58 #define TGT_BAVAIL(i) (OST_TGT(lod,i)->ltd_statfs.os_bavail * \
59 OST_TGT(lod,i)->ltd_statfs.os_bsize)
62 * Add a new target to Quality of Service (QoS) target table.
64 * Add a new OST target to the structure representing an OSS. Resort the list
65 * of known OSSs by the number of OSTs attached to each OSS. The OSS list is
66 * protected internally and no external locking is required.
68 * \param[in] lod LOD device
69 * \param[in] ost_desc OST description
71 * \retval 0 on success
72 * \retval -ENOMEM on error
74 int qos_add_tgt(struct lod_device *lod, struct lod_tgt_desc *ost_desc)
76 struct lod_qos_oss *oss = NULL, *temposs;
77 struct obd_export *exp = ost_desc->ltd_exp;
78 int rc = 0, found = 0;
79 struct list_head *list;
82 down_write(&lod->lod_qos.lq_rw_sem);
84 * a bit hacky approach to learn NID of corresponding connection
85 * but there is no official API to access information like this
88 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
89 if (obd_uuid_equals(&oss->lqo_uuid,
90 &exp->exp_connection->c_remote_uuid)) {
99 GOTO(out, rc = -ENOMEM);
100 memcpy(&oss->lqo_uuid, &exp->exp_connection->c_remote_uuid,
101 sizeof(oss->lqo_uuid));
103 /* Assume we have to move this one */
104 list_del(&oss->lqo_oss_list);
107 oss->lqo_ost_count++;
108 ost_desc->ltd_qos.ltq_oss = oss;
110 CDEBUG(D_QOS, "add tgt %s to OSS %s (%d OSTs)\n",
111 obd_uuid2str(&ost_desc->ltd_uuid), obd_uuid2str(&oss->lqo_uuid),
114 /* Add sorted by # of OSTs. Find the first entry that we're
116 list = &lod->lod_qos.lq_oss_list;
117 list_for_each_entry(temposs, list, lqo_oss_list) {
118 if (oss->lqo_ost_count > temposs->lqo_ost_count)
121 /* ...and add before it. If we're the first or smallest, temposs
122 points to the list head, and we add to the end. */
123 list_add_tail(&oss->lqo_oss_list, &temposs->lqo_oss_list);
125 lod->lod_qos.lq_dirty = 1;
126 lod->lod_qos.lq_rr.lqr_dirty = 1;
129 up_write(&lod->lod_qos.lq_rw_sem);
134 * Remove OST target from QoS table.
136 * Removes given OST target from QoS table and releases related OSS structure
137 * if no OSTs remain on the OSS.
139 * \param[in] lod LOD device
140 * \param[in] ost_desc OST description
142 * \retval 0 on success
143 * \retval -ENOENT if no OSS was found
145 int qos_del_tgt(struct lod_device *lod, struct lod_tgt_desc *ost_desc)
147 struct lod_qos_oss *oss;
151 down_write(&lod->lod_qos.lq_rw_sem);
152 oss = ost_desc->ltd_qos.ltq_oss;
154 GOTO(out, rc = -ENOENT);
156 oss->lqo_ost_count--;
157 if (oss->lqo_ost_count == 0) {
158 CDEBUG(D_QOS, "removing OSS %s\n",
159 obd_uuid2str(&oss->lqo_uuid));
160 list_del(&oss->lqo_oss_list);
161 ost_desc->ltd_qos.ltq_oss = NULL;
165 lod->lod_qos.lq_dirty = 1;
166 lod->lod_qos.lq_rr.lqr_dirty = 1;
168 up_write(&lod->lod_qos.lq_rw_sem);
173 * Check whether the target is available for new OST objects.
175 * Request statfs data from the given target and verify it's active and not
176 * read-only. If so, then it can be used to place new OST objects. This
177 * function also maintains the number of active/inactive targets and sets
178 * dirty flags if those numbers change so others can run re-balance procedures.
179 * No external locking is required.
181 * \param[in] env execution environment for this thread
182 * \param[in] d LOD device
183 * \param[in] index index of OST target to check
184 * \param[out] sfs buffer for statfs data
186 * \retval 0 if the target is good
187 * \retval negative negated errno on error
190 static int lod_statfs_and_check(const struct lu_env *env, struct lod_device *d,
191 int index, struct obd_statfs *sfs)
193 struct lod_tgt_desc *ost;
198 ost = OST_TGT(d,index);
201 rc = dt_statfs(env, ost->ltd_ost, sfs);
203 if (rc == 0 && ((sfs->os_state & OS_STATE_ENOSPC) ||
204 (sfs->os_state & OS_STATE_ENOINO && sfs->os_fprecreated == 0)))
207 if (rc && rc != -ENOTCONN)
208 CERROR("%s: statfs: rc = %d\n", lod2obd(d)->obd_name, rc);
210 /* If the OST is readonly then we can't allocate objects there */
211 if (sfs->os_state & OS_STATE_READONLY)
214 /* check whether device has changed state (active, inactive) */
215 if (rc != 0 && ost->ltd_active) {
216 /* turned inactive? */
217 spin_lock(&d->lod_lock);
218 if (ost->ltd_active) {
221 ost->ltd_connecting = 1;
223 LASSERT(d->lod_desc.ld_active_tgt_count > 0);
224 d->lod_desc.ld_active_tgt_count--;
225 d->lod_qos.lq_dirty = 1;
226 d->lod_qos.lq_rr.lqr_dirty = 1;
227 CDEBUG(D_CONFIG, "%s: turns inactive\n",
228 ost->ltd_exp->exp_obd->obd_name);
230 spin_unlock(&d->lod_lock);
231 } else if (rc == 0 && ost->ltd_active == 0) {
233 LASSERTF(d->lod_desc.ld_active_tgt_count < d->lod_ostnr,
234 "active tgt count %d, ost nr %d\n",
235 d->lod_desc.ld_active_tgt_count, d->lod_ostnr);
236 spin_lock(&d->lod_lock);
237 if (ost->ltd_active == 0) {
239 ost->ltd_connecting = 0;
240 d->lod_desc.ld_active_tgt_count++;
241 d->lod_qos.lq_dirty = 1;
242 d->lod_qos.lq_rr.lqr_dirty = 1;
243 CDEBUG(D_CONFIG, "%s: turns active\n",
244 ost->ltd_exp->exp_obd->obd_name);
246 spin_unlock(&d->lod_lock);
253 * Maintain per-target statfs data.
255 * The function refreshes statfs data for all the targets every N seconds.
256 * The actual N is controlled via procfs and set to LOV_DESC_QOS_MAXAGE_DEFAULT
259 * \param[in] env execution environment for this thread
260 * \param[in] lod LOD device
262 void lod_qos_statfs_update(const struct lu_env *env, struct lod_device *lod)
264 struct obd_device *obd = lod2obd(lod);
265 struct ost_pool *osts = &(lod->lod_pool_info);
272 max_age = ktime_get_seconds() - 2 * lod->lod_desc.ld_qos_maxage;
274 if (obd->obd_osfs_age > max_age)
275 /* statfs data are quite recent, don't need to refresh it */
278 down_write(&lod->lod_qos.lq_rw_sem);
280 if (obd->obd_osfs_age > max_age)
283 for (i = 0; i < osts->op_count; i++) {
284 idx = osts->op_array[i];
285 avail = OST_TGT(lod,idx)->ltd_statfs.os_bavail;
286 if (lod_statfs_and_check(env, lod, idx,
287 &OST_TGT(lod, idx)->ltd_statfs))
289 if (OST_TGT(lod,idx)->ltd_statfs.os_bavail != avail)
290 /* recalculate weigths */
291 lod->lod_qos.lq_dirty = 1;
293 obd->obd_osfs_age = ktime_get_seconds();
296 up_write(&lod->lod_qos.lq_rw_sem);
301 * Calculate per-OST and per-OSS penalties
303 * Re-calculate penalties when the configuration changes, active targets
304 * change and after statfs refresh (all these are reflected by lq_dirty flag).
305 * On every OST and OSS: decay the penalty by half for every 8x the update
306 * interval that the device has been idle. That gives lots of time for the
307 * statfs information to be updated (which the penalty is only a proxy for),
308 * and avoids penalizing OSS/OSTs under light load.
309 * See lod_qos_calc_weight() for how penalties are factored into the weight.
311 * \param[in] lod LOD device
313 * \retval 0 on success
314 * \retval -EAGAIN the number of OSTs isn't enough
316 static int lod_qos_calc_ppo(struct lod_device *lod)
318 struct lod_qos_oss *oss;
319 __u64 ba_max, ba_min, temp;
326 if (!lod->lod_qos.lq_dirty)
329 num_active = lod->lod_desc.ld_active_tgt_count - 1;
331 GOTO(out, rc = -EAGAIN);
333 /* find bavail on each OSS */
334 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list)
336 lod->lod_qos.lq_active_oss_count = 0;
339 * How badly user wants to select OSTs "widely" (not recently chosen
340 * and not on recent OSS's). As opposed to "freely" (free space
343 prio_wide = 256 - lod->lod_qos.lq_prio_free;
345 ba_min = (__u64)(-1);
347 now = cfs_time_current_sec();
348 /* Calculate OST penalty per object
349 * (lod ref taken in lod_qos_prep_create()) */
350 cfs_foreach_bit(lod->lod_ost_bitmap, i) {
351 LASSERT(OST_TGT(lod,i));
352 temp = TGT_BAVAIL(i);
355 ba_min = min(temp, ba_min);
356 ba_max = max(temp, ba_max);
358 /* Count the number of usable OSS's */
359 if (OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail == 0)
360 lod->lod_qos.lq_active_oss_count++;
361 OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail += temp;
363 /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
365 do_div(temp, num_active);
366 OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj =
367 (temp * prio_wide) >> 8;
369 age = (now - OST_TGT(lod,i)->ltd_qos.ltq_used) >> 3;
370 if (lod->lod_qos.lq_reset ||
371 age > 32 * lod->lod_desc.ld_qos_maxage)
372 OST_TGT(lod,i)->ltd_qos.ltq_penalty = 0;
373 else if (age > lod->lod_desc.ld_qos_maxage)
374 /* Decay OST penalty. */
375 OST_TGT(lod,i)->ltd_qos.ltq_penalty >>=
376 (age / lod->lod_desc.ld_qos_maxage);
379 num_active = lod->lod_qos.lq_active_oss_count - 1;
380 if (num_active < 1) {
381 /* If there's only 1 OSS, we can't penalize it, so instead
382 we have to double the OST penalty */
384 cfs_foreach_bit(lod->lod_ost_bitmap, i)
385 OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj <<= 1;
388 /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
389 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
390 temp = oss->lqo_bavail >> 1;
391 do_div(temp, oss->lqo_ost_count * num_active);
392 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
394 age = (now - oss->lqo_used) >> 3;
395 if (lod->lod_qos.lq_reset ||
396 age > 32 * lod->lod_desc.ld_qos_maxage)
397 oss->lqo_penalty = 0;
398 else if (age > lod->lod_desc.ld_qos_maxage)
399 /* Decay OSS penalty. */
400 oss->lqo_penalty >>= age / lod->lod_desc.ld_qos_maxage;
403 lod->lod_qos.lq_dirty = 0;
404 lod->lod_qos.lq_reset = 0;
406 /* If each ost has almost same free space,
407 * do rr allocation for better creation performance */
408 lod->lod_qos.lq_same_space = 0;
409 if ((ba_max * (256 - lod->lod_qos.lq_threshold_rr)) >> 8 < ba_min) {
410 lod->lod_qos.lq_same_space = 1;
411 /* Reset weights for the next time we enter qos mode */
412 lod->lod_qos.lq_reset = 1;
418 if (!rc && lod->lod_qos.lq_same_space)
425 * Calculate weight for a given OST target.
427 * The final OST weight is the number of bytes available minus the OST and
428 * OSS penalties. See lod_qos_calc_ppo() for how penalties are calculated.
430 * \param[in] lod LOD device, where OST targets are listed
431 * \param[in] i OST target index
435 static int lod_qos_calc_weight(struct lod_device *lod, int i)
439 temp = TGT_BAVAIL(i);
440 temp2 = OST_TGT(lod,i)->ltd_qos.ltq_penalty +
441 OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_penalty;
443 OST_TGT(lod,i)->ltd_qos.ltq_weight = 0;
445 OST_TGT(lod,i)->ltd_qos.ltq_weight = temp - temp2;
450 * Re-calculate weights.
452 * The function is called when some OST target was used for a new object. In
453 * this case we should re-calculate all the weights to keep new allocations
456 * \param[in] lod LOD device
457 * \param[in] osts OST pool where a new object was placed
458 * \param[in] index OST target where a new object was placed
459 * \param[out] total_wt new total weight for the pool
463 static int lod_qos_used(struct lod_device *lod, struct ost_pool *osts,
464 __u32 index, __u64 *total_wt)
466 struct lod_tgt_desc *ost;
467 struct lod_qos_oss *oss;
471 ost = OST_TGT(lod,index);
474 /* Don't allocate on this devuce anymore, until the next alloc_qos */
475 ost->ltd_qos.ltq_usable = 0;
477 oss = ost->ltd_qos.ltq_oss;
479 /* Decay old penalty by half (we're adding max penalty, and don't
480 want it to run away.) */
481 ost->ltd_qos.ltq_penalty >>= 1;
482 oss->lqo_penalty >>= 1;
484 /* mark the OSS and OST as recently used */
485 ost->ltd_qos.ltq_used = oss->lqo_used = cfs_time_current_sec();
487 /* Set max penalties for this OST and OSS */
488 ost->ltd_qos.ltq_penalty +=
489 ost->ltd_qos.ltq_penalty_per_obj * lod->lod_ostnr;
490 oss->lqo_penalty += oss->lqo_penalty_per_obj *
491 lod->lod_qos.lq_active_oss_count;
493 /* Decrease all OSS penalties */
494 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
495 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
496 oss->lqo_penalty = 0;
498 oss->lqo_penalty -= oss->lqo_penalty_per_obj;
502 /* Decrease all OST penalties */
503 for (j = 0; j < osts->op_count; j++) {
506 i = osts->op_array[j];
507 if (!cfs_bitmap_check(lod->lod_ost_bitmap, i))
510 ost = OST_TGT(lod,i);
513 if (ost->ltd_qos.ltq_penalty <
514 ost->ltd_qos.ltq_penalty_per_obj)
515 ost->ltd_qos.ltq_penalty = 0;
517 ost->ltd_qos.ltq_penalty -=
518 ost->ltd_qos.ltq_penalty_per_obj;
520 lod_qos_calc_weight(lod, i);
522 /* Recalc the total weight of usable osts */
523 if (ost->ltd_qos.ltq_usable)
524 *total_wt += ost->ltd_qos.ltq_weight;
526 QOS_DEBUG("recalc tgt %d usable=%d avail=%llu"
527 " ostppo=%llu ostp=%llu ossppo=%llu"
528 " ossp=%llu wt=%llu\n",
529 i, ost->ltd_qos.ltq_usable, TGT_BAVAIL(i) >> 10,
530 ost->ltd_qos.ltq_penalty_per_obj >> 10,
531 ost->ltd_qos.ltq_penalty >> 10,
532 ost->ltd_qos.ltq_oss->lqo_penalty_per_obj >> 10,
533 ost->ltd_qos.ltq_oss->lqo_penalty >> 10,
534 ost->ltd_qos.ltq_weight >> 10);
540 void lod_qos_rr_init(struct lod_qos_rr *lqr)
542 spin_lock_init(&lqr->lqr_alloc);
547 #define LOV_QOS_EMPTY ((__u32)-1)
550 * Calculate optimal round-robin order with regard to OSSes.
552 * Place all the OSTs from pool \a src_pool in a special array to be used for
553 * round-robin (RR) stripe allocation. The placement algorithm interleaves
554 * OSTs from the different OSSs so that RR allocation can balance OSSs evenly.
555 * Resorts the targets when the number of active targets changes (because of
556 * a new target or activation/deactivation).
558 * \param[in] lod LOD device
559 * \param[in] src_pool OST pool
560 * \param[in] lqr round-robin list
562 * \retval 0 on success
563 * \retval -ENOMEM fails to allocate the array
565 static int lod_qos_calc_rr(struct lod_device *lod, struct ost_pool *src_pool,
566 struct lod_qos_rr *lqr)
568 struct lod_qos_oss *oss;
569 struct lod_tgt_desc *ost;
570 unsigned placed, real_count;
575 if (!lqr->lqr_dirty) {
576 LASSERT(lqr->lqr_pool.op_size);
580 /* Do actual allocation. */
581 down_write(&lod->lod_qos.lq_rw_sem);
584 * Check again. While we were sleeping on @lq_rw_sem something could
587 if (!lqr->lqr_dirty) {
588 LASSERT(lqr->lqr_pool.op_size);
589 up_write(&lod->lod_qos.lq_rw_sem);
593 real_count = src_pool->op_count;
595 /* Zero the pool array */
596 /* alloc_rr is holding a read lock on the pool, so nobody is adding/
597 deleting from the pool. The lq_rw_sem insures that nobody else
599 lqr->lqr_pool.op_count = real_count;
600 rc = lod_ost_pool_extend(&lqr->lqr_pool, real_count);
602 up_write(&lod->lod_qos.lq_rw_sem);
605 for (i = 0; i < lqr->lqr_pool.op_count; i++)
606 lqr->lqr_pool.op_array[i] = LOV_QOS_EMPTY;
608 /* Place all the OSTs from 1 OSS at the same time. */
610 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
613 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
616 if (!cfs_bitmap_check(lod->lod_ost_bitmap,
617 src_pool->op_array[i]))
620 ost = OST_TGT(lod,src_pool->op_array[i]);
621 LASSERT(ost && ost->ltd_ost);
622 if (ost->ltd_qos.ltq_oss != oss)
625 /* Evenly space these OSTs across arrayspace */
626 next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count;
627 while (lqr->lqr_pool.op_array[next] != LOV_QOS_EMPTY)
628 next = (next + 1) % lqr->lqr_pool.op_count;
630 lqr->lqr_pool.op_array[next] = src_pool->op_array[i];
637 up_write(&lod->lod_qos.lq_rw_sem);
639 if (placed != real_count) {
640 /* This should never happen */
641 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
642 "round-robin list (%d of %d).\n",
644 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
645 LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
646 lqr->lqr_pool.op_array[i]);
653 for (i = 0; i < lqr->lqr_pool.op_count; i++)
654 QOS_CONSOLE("rr #%d ost idx=%d\n", i, lqr->lqr_pool.op_array[i]);
661 * Instantiate and declare creation of a new object.
663 * The function instantiates LU representation for a new object on the
664 * specified device. Also it declares an intention to create that
665 * object on the storage target.
667 * Note lu_object_anon() is used which is a trick with regard to LU/OSD
668 * infrastructure - in the existing precreation framework we can't assign FID
669 * at this moment, we do this later once a transaction is started. So the
670 * special method instantiates FID-less object in the cache and later it
671 * will get a FID and proper placement in LU cache.
673 * \param[in] env execution environment for this thread
674 * \param[in] d LOD device
675 * \param[in] ost_idx OST target index where the object is being created
676 * \param[in] th transaction handle
678 * \retval object ptr on success, ERR_PTR() otherwise
680 static struct dt_object *lod_qos_declare_object_on(const struct lu_env *env,
681 struct lod_device *d,
685 struct lod_tgt_desc *ost;
686 struct lu_object *o, *n;
687 struct lu_device *nd;
688 struct dt_object *dt;
693 LASSERT(ost_idx < d->lod_osts_size);
694 ost = OST_TGT(d,ost_idx);
696 LASSERT(ost->ltd_ost);
698 nd = &ost->ltd_ost->dd_lu_dev;
701 * allocate anonymous object with zero fid, real fid
702 * will be assigned by OSP within transaction
703 * XXX: to be fixed with fully-functional OST fids
705 o = lu_object_anon(env, nd, NULL);
707 GOTO(out, dt = ERR_PTR(PTR_ERR(o)));
709 n = lu_object_locate(o->lo_header, nd->ld_type);
710 if (unlikely(n == NULL)) {
711 CERROR("can't find slice\n");
712 lu_object_put(env, o);
713 GOTO(out, dt = ERR_PTR(-EINVAL));
716 dt = container_of(n, struct dt_object, do_lu);
718 rc = lod_sub_declare_create(env, dt, NULL, NULL, NULL, th);
720 CDEBUG(D_OTHER, "can't declare creation on #%u: %d\n",
722 lu_object_put(env, o);
731 * Calculate a minimum acceptable stripe count.
733 * Return an acceptable stripe count depending on flag LOV_USES_DEFAULT_STRIPE:
734 * all stripes or 3/4 of stripes.
736 * \param[in] stripe_count number of stripes requested
737 * \param[in] flags 0 or LOV_USES_DEFAULT_STRIPE
739 * \retval acceptable stripecount
741 static int min_stripe_count(__u32 stripe_count, int flags)
743 return (flags & LOV_USES_DEFAULT_STRIPE ?
744 stripe_count - (stripe_count / 4) : stripe_count);
747 #define LOV_CREATE_RESEED_MULT 30
748 #define LOV_CREATE_RESEED_MIN 2000
751 * Initialize temporary OST-in-use array.
753 * Allocate or extend the array used to mark targets already assigned to a new
754 * striping so they are not used more than once.
756 * \param[in] env execution environment for this thread
757 * \param[in] stripes number of items needed in the array
759 * \retval 0 on success
760 * \retval -ENOMEM on error
762 static inline int lod_qos_ost_in_use_clear(const struct lu_env *env,
765 struct lod_thread_info *info = lod_env_info(env);
767 if (info->lti_ea_store_size < sizeof(int) * stripes)
768 lod_ea_store_resize(info, stripes * sizeof(int));
769 if (info->lti_ea_store_size < sizeof(int) * stripes) {
770 CERROR("can't allocate memory for ost-in-use array\n");
773 memset(info->lti_ea_store, -1, sizeof(int) * stripes);
778 * Remember a target in the array of used targets.
780 * Mark the given target as used for a new striping being created. The status
781 * of an OST in a striping can be checked with lod_qos_is_ost_used().
783 * \param[in] env execution environment for this thread
784 * \param[in] idx index in the array
785 * \param[in] ost OST target index to mark as used
787 static inline void lod_qos_ost_in_use(const struct lu_env *env,
790 struct lod_thread_info *info = lod_env_info(env);
791 int *osts = info->lti_ea_store;
793 LASSERT(info->lti_ea_store_size >= idx * sizeof(int));
798 * Check is OST used in a striping.
800 * Checks whether OST with the given index is marked as used in the temporary
801 * array (see lod_qos_ost_in_use()).
803 * \param[in] env execution environment for this thread
804 * \param[in] ost OST target index to check
805 * \param[in] stripes the number of items used in the array already
810 static int lod_qos_is_ost_used(const struct lu_env *env, int ost, __u32 stripes)
812 struct lod_thread_info *info = lod_env_info(env);
813 int *osts = info->lti_ea_store;
816 for (j = 0; j < stripes; j++) {
824 * Check is OST used in a composite layout
826 * \param[in] inuse all inuse ost indexs
827 * \param[in] ost OST target index to check
832 static inline int lod_comp_is_ost_used(struct ost_pool *inuse, int ost)
835 LASSERT(inuse != NULL);
837 if (inuse->op_size == 0)
840 LASSERT(inuse->op_count * sizeof(inuse->op_array[0]) <= inuse->op_size);
841 for (j = 0; j < inuse->op_count; j++) {
842 if (inuse->op_array[j] == ost)
849 * Mark the given target as used for a composite layout
851 * \param[in] inuse inuse ost index array
852 * \param[in] idx index in the array
854 static inline void lod_comp_ost_in_use(struct ost_pool *inuse, int ost)
856 LASSERT(inuse != NULL);
857 if (inuse->op_size && !lod_comp_is_ost_used(inuse, ost)) {
858 LASSERTF(inuse->op_count * sizeof(inuse->op_array[0]) <
860 "count %d size %u", inuse->op_count, inuse->op_size);
861 inuse->op_array[inuse->op_count] = ost;
866 static int lod_check_and_reserve_ost(const struct lu_env *env,
867 struct lod_device *m,
868 struct obd_statfs *sfs, __u32 ost_idx,
869 __u32 speed, __u32 *s_idx,
870 struct dt_object **stripe,
872 struct ost_pool *inuse)
875 __u32 stripe_idx = *s_idx;
878 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
880 /* this OSP doesn't feel well */
885 * We expect number of precreated objects in f_ffree at
886 * the first iteration, skip OSPs with no objects ready
888 if (sfs->os_fprecreated == 0 && speed == 0) {
889 QOS_DEBUG("#%d: precreation is empty\n", ost_idx);
894 * try to use another OSP if this one is degraded
896 if (sfs->os_state & OS_STATE_DEGRADED && speed < 2) {
897 QOS_DEBUG("#%d: degraded\n", ost_idx);
902 * try not allocate on OST which has been used by other
905 if (speed == 0 && lod_comp_is_ost_used(inuse, ost_idx)) {
906 QOS_DEBUG("#%d: used by other component\n", ost_idx);
911 * do not put >1 objects on a single OST
913 if (lod_qos_is_ost_used(env, ost_idx, stripe_idx))
916 o = lod_qos_declare_object_on(env, m, ost_idx, th);
918 CDEBUG(D_OTHER, "can't declare new object on #%u: %d\n",
919 ost_idx, (int) PTR_ERR(o));
925 * We've successfully declared (reserved) an object
927 lod_qos_ost_in_use(env, stripe_idx, ost_idx);
928 lod_comp_ost_in_use(inuse, ost_idx);
929 stripe[stripe_idx] = o;
930 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LOV_CREATE_RACE, 2);
939 * Allocate a striping using round-robin algorithm.
941 * Allocates a new striping using round-robin algorithm. The function refreshes
942 * all the internal structures (statfs cache, array of available OSTs sorted
943 * with regard to OSS, etc). The number of stripes required is taken from the
944 * object (must be prepared by the caller), but can change if the flag
945 * LOV_USES_DEFAULT_STRIPE is supplied. The caller should ensure nobody else
946 * is trying to create a striping on the object in parallel. All the internal
947 * structures (like pools, etc) are protected and no additional locking is
948 * required. The function succeeds even if a single stripe is allocated. To save
949 * time we give priority to targets which already have objects precreated.
950 * Full OSTs are skipped (see lod_qos_dev_is_full() for the details).
952 * \param[in] env execution environment for this thread
953 * \param[in] lo LOD object
954 * \param[out] stripe striping created
955 * \param[in] flags allocation flags (0 or LOV_USES_DEFAULT_STRIPE)
956 * \param[in] th transaction handle
957 * \param[in] comp_idx index of ldo_comp_entries
958 * \param[in|out] inuse array of inuse ost index
960 * \retval 0 on success
961 * \retval -ENOSPC if not enough OSTs are found
962 * \retval negative negated errno for other failures
964 static int lod_alloc_rr(const struct lu_env *env, struct lod_object *lo,
965 struct dt_object **stripe, int flags,
966 struct thandle *th, int comp_idx,
967 struct ost_pool *inuse)
969 struct lod_layout_component *lod_comp;
970 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
971 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
972 struct pool_desc *pool = NULL;
973 struct ost_pool *osts;
974 struct lod_qos_rr *lqr;
975 unsigned int i, array_idx;
976 __u32 ost_start_idx_temp;
977 __u32 stripe_idx = 0;
978 __u32 stripe_count, stripe_count_min, ost_idx;
979 int rc, speed = 0, ost_connecting = 0;
982 LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
983 lod_comp = &lo->ldo_comp_entries[comp_idx];
984 stripe_count = lod_comp->llc_stripe_count;
985 stripe_count_min = min_stripe_count(stripe_count, flags);
987 if (lod_comp->llc_pool != NULL)
988 pool = lod_find_pool(m, lod_comp->llc_pool);
991 down_read(&pool_tgt_rw_sem(pool));
992 osts = &(pool->pool_obds);
993 lqr = &(pool->pool_rr);
995 osts = &(m->lod_pool_info);
996 lqr = &(m->lod_qos.lq_rr);
999 rc = lod_qos_calc_rr(m, osts, lqr);
1003 rc = lod_qos_ost_in_use_clear(env, stripe_count);
1007 down_read(&m->lod_qos.lq_rw_sem);
1008 spin_lock(&lqr->lqr_alloc);
1009 if (--lqr->lqr_start_count <= 0) {
1010 lqr->lqr_start_idx = cfs_rand() % osts->op_count;
1011 lqr->lqr_start_count =
1012 (LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) +
1013 LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U);
1014 } else if (stripe_count_min >= osts->op_count ||
1015 lqr->lqr_start_idx > osts->op_count) {
1016 /* If we have allocated from all of the OSTs, slowly
1017 * precess the next start if the OST/stripe count isn't
1018 * already doing this for us. */
1019 lqr->lqr_start_idx %= osts->op_count;
1020 if (stripe_count > 1 && (osts->op_count % stripe_count) != 1)
1021 ++lqr->lqr_offset_idx;
1023 ost_start_idx_temp = lqr->lqr_start_idx;
1027 QOS_DEBUG("pool '%s' want %d start_idx %d start_count %d offset %d "
1028 "active %d count %d\n",
1029 lod_comp->llc_pool ? lod_comp->llc_pool : "",
1030 stripe_count, lqr->lqr_start_idx, lqr->lqr_start_count,
1031 lqr->lqr_offset_idx, osts->op_count, osts->op_count);
1033 for (i = 0; i < osts->op_count && stripe_idx < stripe_count; i++) {
1034 array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) %
1036 ++lqr->lqr_start_idx;
1037 ost_idx = lqr->lqr_pool.op_array[array_idx];
1039 QOS_DEBUG("#%d strt %d act %d strp %d ary %d idx %d\n",
1040 i, lqr->lqr_start_idx, /* XXX: active*/ 0,
1041 stripe_idx, array_idx, ost_idx);
1043 if ((ost_idx == LOV_QOS_EMPTY) ||
1044 !cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
1047 /* Fail Check before osc_precreate() is called
1048 so we can only 'fail' single OSC. */
1049 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
1052 spin_unlock(&lqr->lqr_alloc);
1053 rc = lod_check_and_reserve_ost(env, m, sfs, ost_idx, speed,
1054 &stripe_idx, stripe, th, inuse);
1055 spin_lock(&lqr->lqr_alloc);
1057 if (rc != 0 && OST_TGT(m, ost_idx)->ltd_connecting)
1060 if ((speed < 2) && (stripe_idx < stripe_count_min)) {
1061 /* Try again, allowing slower OSCs */
1063 lqr->lqr_start_idx = ost_start_idx_temp;
1069 spin_unlock(&lqr->lqr_alloc);
1070 up_read(&m->lod_qos.lq_rw_sem);
1073 lod_comp->llc_stripe_count = stripe_idx;
1074 /* at least one stripe is allocated */
1077 /* nobody provided us with a single object */
1086 up_read(&pool_tgt_rw_sem(pool));
1087 /* put back ref got by lod_find_pool() */
1088 lod_pool_putref(pool);
1095 * Allocate a specific striping layout on a user defined set of OSTs.
1097 * Allocates new striping using the OST index range provided by the data from
1098 * the lmm_obejcts contained in the lov_user_md passed to this method. Full
1099 * OSTs are not considered. The exact order of OSTs requested by the user
1100 * is respected as much as possible depending on OST status. The number of
1101 * stripes needed and stripe offset are taken from the object. If that number
1102 * can not be met, then the function returns a failure and then it's the
1103 * caller's responsibility to release the stripes allocated. All the internal
1104 * structures are protected, but no concurrent allocation is allowed on the
1107 * \param[in] env execution environment for this thread
1108 * \param[in] lo LOD object
1109 * \param[out] stripe striping created
1110 * \param[in] th transaction handle
1111 * \param[in] comp_idx index of ldo_comp_entries
1112 * \param[in|out] inuse array of inuse ost index
1114 * \retval 0 on success
1115 * \retval -ENODEV OST index does not exist on file system
1116 * \retval -EINVAL requested OST index is invalid
1117 * \retval negative negated errno on error
1119 static int lod_alloc_ost_list(const struct lu_env *env, struct lod_object *lo,
1120 struct dt_object **stripe, struct thandle *th,
1121 int comp_idx, struct ost_pool *inuse)
1123 struct lod_layout_component *lod_comp;
1124 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1125 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
1126 struct dt_object *o;
1127 unsigned int array_idx = 0;
1128 int stripe_count = 0;
1133 /* for specific OSTs layout */
1134 LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
1135 lod_comp = &lo->ldo_comp_entries[comp_idx];
1136 LASSERT(lod_comp->llc_ostlist.op_array);
1138 rc = lod_qos_ost_in_use_clear(env, lod_comp->llc_stripe_count);
1142 for (i = 0; i < lod_comp->llc_stripe_count; i++) {
1143 if (lod_comp->llc_ostlist.op_array[i] ==
1144 lod_comp->llc_stripe_offset) {
1149 if (i == lod_comp->llc_stripe_count) {
1151 "%s: start index %d not in the specified list of OSTs\n",
1152 lod2obd(m)->obd_name, lod_comp->llc_stripe_offset);
1156 for (i = 0; i < lod_comp->llc_stripe_count;
1157 i++, array_idx = (array_idx + 1) % lod_comp->llc_stripe_count) {
1158 __u32 ost_idx = lod_comp->llc_ostlist.op_array[array_idx];
1160 if (!cfs_bitmap_check(m->lod_ost_bitmap, ost_idx)) {
1166 * do not put >1 objects on a single OST
1168 if (lod_qos_is_ost_used(env, ost_idx, stripe_count)) {
1173 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
1174 if (rc < 0) /* this OSP doesn't feel well */
1177 o = lod_qos_declare_object_on(env, m, ost_idx, th);
1181 "%s: can't declare new object on #%u: %d\n",
1182 lod2obd(m)->obd_name, ost_idx, rc);
1187 * We've successfully declared (reserved) an object
1189 lod_qos_ost_in_use(env, stripe_count, ost_idx);
1190 lod_comp_ost_in_use(inuse, ost_idx);
1191 stripe[stripe_count] = o;
1199 * Allocate a striping on a predefined set of OSTs.
1201 * Allocates new layout starting from OST index in lo->ldo_stripe_offset.
1202 * Full OSTs are not considered. The exact order of OSTs is not important and
1203 * varies depending on OST status. The allocation procedure prefers the targets
1204 * with precreated objects ready. The number of stripes needed and stripe
1205 * offset are taken from the object. If that number cannot be met, then the
1206 * function returns an error and then it's the caller's responsibility to
1207 * release the stripes allocated. All the internal structures are protected,
1208 * but no concurrent allocation is allowed on the same objects.
1210 * \param[in] env execution environment for this thread
1211 * \param[in] lo LOD object
1212 * \param[out] stripe striping created
1213 * \param[in] flags not used
1214 * \param[in] th transaction handle
1215 * \param[in] comp_idx index of ldo_comp_entries
1216 * \param[in|out]inuse array of inuse ost index
1218 * \retval 0 on success
1219 * \retval -ENOSPC if no OST objects are available at all
1220 * \retval -EFBIG if not enough OST objects are found
1221 * \retval -EINVAL requested offset is invalid
1222 * \retval negative errno on failure
1224 static int lod_alloc_specific(const struct lu_env *env, struct lod_object *lo,
1225 struct dt_object **stripe, int flags,
1226 struct thandle *th, int comp_idx,
1227 struct ost_pool *inuse)
1229 struct lod_layout_component *lod_comp;
1230 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1231 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
1232 struct dt_object *o;
1234 unsigned int i, array_idx, ost_count;
1235 int rc, stripe_num = 0;
1237 struct pool_desc *pool = NULL;
1238 struct ost_pool *osts;
1241 LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
1242 lod_comp = &lo->ldo_comp_entries[comp_idx];
1244 rc = lod_qos_ost_in_use_clear(env, lod_comp->llc_stripe_count);
1248 if (lod_comp->llc_pool != NULL)
1249 pool = lod_find_pool(m, lod_comp->llc_pool);
1252 down_read(&pool_tgt_rw_sem(pool));
1253 osts = &(pool->pool_obds);
1255 osts = &(m->lod_pool_info);
1258 ost_count = osts->op_count;
1261 /* search loi_ost_idx in ost array */
1263 for (i = 0; i < ost_count; i++) {
1264 if (osts->op_array[i] == lod_comp->llc_stripe_offset) {
1269 if (i == ost_count) {
1270 CERROR("Start index %d not found in pool '%s'\n",
1271 lod_comp->llc_stripe_offset,
1272 lod_comp->llc_pool ? lod_comp->llc_pool : "");
1273 GOTO(out, rc = -EINVAL);
1276 for (i = 0; i < ost_count;
1277 i++, array_idx = (array_idx + 1) % ost_count) {
1278 ost_idx = osts->op_array[array_idx];
1280 if (!cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
1283 /* Fail Check before osc_precreate() is called
1284 so we can only 'fail' single OSC. */
1285 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
1289 * do not put >1 objects on a single OST
1291 if (lod_qos_is_ost_used(env, ost_idx, stripe_num))
1295 * try not allocate on the OST used by other component
1297 if (speed == 0 && i != 0 &&
1298 lod_comp_is_ost_used(inuse, ost_idx))
1301 /* Drop slow OSCs if we can, but not for requested start idx.
1303 * This means "if OSC is slow and it is not the requested
1304 * start OST, then it can be skipped, otherwise skip it only
1305 * if it is inactive/recovering/out-of-space." */
1307 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
1309 /* this OSP doesn't feel well */
1314 * We expect number of precreated objects at the first
1315 * iteration. Skip OSPs with no objects ready. Don't apply
1316 * this logic to OST specified with stripe_offset.
1318 if (i != 0 && sfs->os_fprecreated == 0 && speed == 0)
1321 o = lod_qos_declare_object_on(env, m, ost_idx, th);
1323 CDEBUG(D_OTHER, "can't declare new object on #%u: %d\n",
1324 ost_idx, (int) PTR_ERR(o));
1329 * We've successfully declared (reserved) an object
1331 lod_qos_ost_in_use(env, stripe_num, ost_idx);
1332 lod_comp_ost_in_use(inuse, ost_idx);
1333 stripe[stripe_num] = o;
1336 /* We have enough stripes */
1337 if (stripe_num == lod_comp->llc_stripe_count)
1341 /* Try again, allowing slower OSCs */
1346 /* If we were passed specific striping params, then a failure to
1347 * meet those requirements is an error, since we can't reallocate
1348 * that memory (it might be part of a larger array or something).
1350 CERROR("can't lstripe objid "DFID": have %d want %u\n",
1351 PFID(lu_object_fid(lod2lu_obj(lo))), stripe_num,
1352 lod_comp->llc_stripe_count);
1353 rc = stripe_num == 0 ? -ENOSPC : -EFBIG;
1356 up_read(&pool_tgt_rw_sem(pool));
1357 /* put back ref got by lod_find_pool() */
1358 lod_pool_putref(pool);
1365 * Check whether QoS allocation should be used.
1367 * A simple helper to decide when QoS allocation should be used:
1368 * if it's just a single available target or the used space is
1369 * evenly distributed among the targets at the moment, then QoS
1370 * allocation algorithm should not be used.
1372 * \param[in] lod LOD device
1374 * \retval 0 should not be used
1375 * \retval 1 should be used
1377 static inline int lod_qos_is_usable(struct lod_device *lod)
1380 /* to be able to debug QoS code */
1384 /* Detect -EAGAIN early, before expensive lock is taken. */
1385 if (!lod->lod_qos.lq_dirty && lod->lod_qos.lq_same_space)
1388 if (lod->lod_desc.ld_active_tgt_count < 2)
1395 * Allocate a striping using an algorithm with weights.
1397 * The function allocates OST objects to create a striping. The algorithm
1398 * used is based on weights (currently only using the free space), and it's
1399 * trying to ensure the space is used evenly by OSTs and OSSs. The striping
1400 * configuration (# of stripes, offset, pool) is taken from the object and
1401 * is prepared by the caller.
1403 * If LOV_USES_DEFAULT_STRIPE is not passed and prepared configuration can't
1404 * be met due to too few OSTs, then allocation fails. If the flag is passed
1405 * fewer than 3/4 of the requested number of stripes can be allocated, then
1408 * No concurrent allocation is allowed on the object and this must be ensured
1409 * by the caller. All the internal structures are protected by the function.
1411 * The algorithm has two steps: find available OSTs and calculate their
1412 * weights, then select the OSTs with their weights used as the probability.
1413 * An OST with a higher weight is proportionately more likely to be selected
1414 * than one with a lower weight.
1416 * \param[in] env execution environment for this thread
1417 * \param[in] lo LOD object
1418 * \param[out] stripe striping created
1419 * \param[in] flags 0 or LOV_USES_DEFAULT_STRIPE
1420 * \param[in] th transaction handle
1421 * \param[in] comp_idx index of ldo_comp_entries
1422 * \param[in|out]inuse array of inuse ost index
1424 * \retval 0 on success
1425 * \retval -EAGAIN not enough OSTs are found for specified stripe count
1426 * \retval -EINVAL requested OST index is invalid
1427 * \retval negative errno on failure
1429 static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo,
1430 struct dt_object **stripe, int flags,
1431 struct thandle *th, int comp_idx,
1432 struct ost_pool *inuse)
1434 struct lod_layout_component *lod_comp;
1435 struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1436 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
1437 struct lod_tgt_desc *ost;
1438 struct dt_object *o;
1439 __u64 total_weight = 0;
1440 struct pool_desc *pool = NULL;
1441 struct ost_pool *osts;
1443 __u32 nfound, good_osts, stripe_count, stripe_count_min;
1444 __u32 inuse_old_count = inuse->op_count;
1448 LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
1449 lod_comp = &lo->ldo_comp_entries[comp_idx];
1450 stripe_count = lod_comp->llc_stripe_count;
1451 stripe_count_min = min_stripe_count(stripe_count, flags);
1452 if (stripe_count_min < 1)
1455 if (lod_comp->llc_pool != NULL)
1456 pool = lod_find_pool(lod, lod_comp->llc_pool);
1459 down_read(&pool_tgt_rw_sem(pool));
1460 osts = &(pool->pool_obds);
1462 osts = &(lod->lod_pool_info);
1465 /* Detect -EAGAIN early, before expensive lock is taken. */
1466 if (!lod_qos_is_usable(lod))
1467 GOTO(out_nolock, rc = -EAGAIN);
1469 /* Do actual allocation, use write lock here. */
1470 down_write(&lod->lod_qos.lq_rw_sem);
1473 * Check again, while we were sleeping on @lq_rw_sem things could
1476 if (!lod_qos_is_usable(lod))
1477 GOTO(out, rc = -EAGAIN);
1479 rc = lod_qos_calc_ppo(lod);
1483 rc = lod_qos_ost_in_use_clear(env, lod_comp->llc_stripe_count);
1488 /* Find all the OSTs that are valid stripe candidates */
1489 for (i = 0; i < osts->op_count; i++) {
1490 if (!cfs_bitmap_check(lod->lod_ost_bitmap, osts->op_array[i]))
1493 ost = OST_TGT(lod, osts->op_array[i]);
1494 ost->ltd_qos.ltq_usable = 0;
1496 rc = lod_statfs_and_check(env, lod, osts->op_array[i], sfs);
1498 /* this OSP doesn't feel well */
1502 if (sfs->os_state & OS_STATE_DEGRADED)
1505 /* Fail Check before osc_precreate() is called
1506 so we can only 'fail' single OSC. */
1507 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) &&
1508 osts->op_array[i] == 0)
1511 ost->ltd_qos.ltq_usable = 1;
1512 lod_qos_calc_weight(lod, osts->op_array[i]);
1513 total_weight += ost->ltd_qos.ltq_weight;
1518 QOS_DEBUG("found %d good osts\n", good_osts);
1520 if (good_osts < stripe_count_min)
1521 GOTO(out, rc = -EAGAIN);
1523 /* We have enough osts */
1524 if (good_osts < stripe_count)
1525 stripe_count = good_osts;
1527 /* Find enough OSTs with weighted random allocation. */
1529 while (nfound < stripe_count) {
1530 __u64 rand, cur_weight;
1536 #if BITS_PER_LONG == 32
1537 rand = cfs_rand() % (unsigned)total_weight;
1538 /* If total_weight > 32-bit, first generate the high
1539 * 32 bits of the random number, then add in the low
1540 * 32 bits (truncated to the upper limit, if needed) */
1541 if (total_weight > 0xffffffffULL)
1542 rand = (__u64)(cfs_rand() %
1543 (unsigned)(total_weight >> 32)) << 32;
1547 if (rand == (total_weight & 0xffffffff00000000ULL))
1548 rand |= cfs_rand() % (unsigned)total_weight;
1553 rand = ((__u64)cfs_rand() << 32 | cfs_rand()) %
1560 /* On average, this will hit larger-weighted OSTs more often.
1561 * 0-weight OSTs will always get used last (only when rand=0) */
1562 for (i = 0; i < osts->op_count; i++) {
1563 __u32 idx = osts->op_array[i];
1565 if (!cfs_bitmap_check(lod->lod_ost_bitmap, idx))
1568 ost = OST_TGT(lod, idx);
1570 if (!ost->ltd_qos.ltq_usable)
1573 cur_weight += ost->ltd_qos.ltq_weight;
1574 QOS_DEBUG("stripe_count=%d nfound=%d cur_weight=%llu "
1575 "rand=%llu total_weight=%llu\n",
1576 stripe_count, nfound, cur_weight, rand,
1579 if (cur_weight < rand)
1582 QOS_DEBUG("stripe=%d to idx=%d\n", nfound, idx);
1584 * do not put >1 objects on a single OST
1586 if (lod_qos_is_ost_used(env, idx, nfound) ||
1587 lod_comp_is_ost_used(inuse, idx))
1590 o = lod_qos_declare_object_on(env, lod, idx, th);
1592 QOS_DEBUG("can't declare object on #%u: %d\n",
1593 idx, (int) PTR_ERR(o));
1597 lod_qos_ost_in_use(env, nfound, idx);
1598 lod_comp_ost_in_use(inuse, idx);
1599 stripe[nfound++] = o;
1600 lod_qos_used(lod, osts, idx, &total_weight);
1606 /* no OST found on this iteration, give up */
1611 if (unlikely(nfound != stripe_count)) {
1613 * when the decision to use weighted algorithm was made
1614 * we had enough appropriate OSPs, but this state can
1615 * change anytime (no space on OST, broken connection, etc)
1616 * so it's possible OSP won't be able to provide us with
1617 * an object due to just changed state
1619 QOS_DEBUG("%s: wanted %d objects, found only %d\n",
1620 lod2obd(lod)->obd_name, stripe_count, nfound);
1621 for (i = 0; i < nfound; i++) {
1622 LASSERT(stripe[i] != NULL);
1623 dt_object_put(env, stripe[i]);
1626 LASSERTF(nfound <= inuse->op_count,
1627 "nfound:%d, op_count:%u\n", nfound, inuse->op_count);
1628 inuse->op_count = inuse_old_count;
1630 /* makes sense to rebalance next time */
1631 lod->lod_qos.lq_dirty = 1;
1632 lod->lod_qos.lq_same_space = 0;
1638 up_write(&lod->lod_qos.lq_rw_sem);
1642 up_read(&pool_tgt_rw_sem(pool));
1643 /* put back ref got by lod_find_pool() */
1644 lod_pool_putref(pool);
1651 * Find largest stripe count the caller can use.
1653 * Find the maximal possible stripe count not greater than \a stripe_count.
1654 * Sometimes suggested stripecount can't be reached for a number of reasons:
1655 * lack of enough active OSTs or the backend does not support EAs that large.
1656 * If the passed one is 0, then the filesystem's default one is used.
1658 * \param[in] lod LOD device
1659 * \param[in] lo The lod_object
1660 * \param[in] stripe_count count the caller would like to use
1662 * \retval the maximum usable stripe count
1664 __u16 lod_get_stripe_count(struct lod_device *lod, struct lod_object *lo,
1667 __u32 max_stripes = LOV_MAX_STRIPE_COUNT_OLD;
1670 stripe_count = lod->lod_desc.ld_default_stripe_count;
1671 if (stripe_count > lod->lod_desc.ld_active_tgt_count)
1672 stripe_count = lod->lod_desc.ld_active_tgt_count;
1676 /* stripe count is based on whether OSD can handle larger EA sizes */
1677 if (lod->lod_osd_max_easize > 0) {
1678 unsigned int easize = lod->lod_osd_max_easize;
1681 if (lo->ldo_is_composite) {
1682 struct lod_layout_component *lod_comp;
1683 unsigned int header_sz = sizeof(struct lov_comp_md_v1);
1685 header_sz += sizeof(struct lov_comp_md_entry_v1) *
1687 for (i = 0; i < lo->ldo_comp_cnt; i++) {
1688 lod_comp = &lo->ldo_comp_entries[i];
1689 if (lod_comp->llc_flags & LCME_FL_INIT)
1690 header_sz += lov_mds_md_size(
1691 lod_comp->llc_stripe_count,
1694 if (easize > header_sz)
1695 easize -= header_sz;
1700 max_stripes = lov_mds_md_max_stripe_count(easize, LOV_MAGIC_V3);
1703 return (stripe_count < max_stripes) ? stripe_count : max_stripes;
1707 * Create in-core respresentation for a fully-defined striping
1709 * When the caller passes a fully-defined striping (i.e. everything including
1710 * OST object FIDs are defined), then we still need to instantiate LU-cache
1711 * with the objects representing the stripes defined. This function completes
1714 * \param[in] env execution environment for this thread
1715 * \param[in] mo LOD object
1716 * \param[in] buf buffer containing the striping
1718 * \retval 0 on success
1719 * \retval negative negated errno on error
1721 int lod_use_defined_striping(const struct lu_env *env,
1722 struct lod_object *mo,
1723 const struct lu_buf *buf)
1725 struct lod_layout_component *lod_comp;
1726 struct lov_mds_md_v1 *v1 = buf->lb_buf;
1727 struct lov_mds_md_v3 *v3 = buf->lb_buf;
1728 struct lov_comp_md_v1 *comp_v1 = NULL;
1729 struct lov_ost_data_v1 *objs;
1736 magic = le32_to_cpu(v1->lmm_magic) & ~LOV_MAGIC_DEFINED;
1738 if (magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3 &&
1739 magic != LOV_MAGIC_COMP_V1)
1742 if (magic == LOV_MAGIC_COMP_V1) {
1743 comp_v1 = buf->lb_buf;
1744 comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count);
1747 mirror_cnt = le16_to_cpu(comp_v1->lcm_mirror_count) + 1;
1748 mo->ldo_flr_state = le16_to_cpu(comp_v1->lcm_flags) &
1750 mo->ldo_is_composite = 1;
1752 mo->ldo_is_composite = 0;
1756 mo->ldo_layout_gen = le16_to_cpu(v1->lmm_layout_gen);
1758 rc = lod_alloc_comp_entries(mo, mirror_cnt, comp_cnt);
1762 for (i = 0; i < comp_cnt; i++) {
1763 struct lu_extent *ext;
1767 lod_comp = &mo->ldo_comp_entries[i];
1769 if (mo->ldo_is_composite) {
1770 offs = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
1771 v1 = (struct lov_mds_md_v1 *)((char *)comp_v1 + offs);
1772 magic = le32_to_cpu(v1->lmm_magic);
1774 ext = &comp_v1->lcm_entries[i].lcme_extent;
1775 lod_comp->llc_extent.e_start =
1776 le64_to_cpu(ext->e_start);
1777 lod_comp->llc_extent.e_end = le64_to_cpu(ext->e_end);
1778 lod_comp->llc_flags =
1779 le32_to_cpu(comp_v1->lcm_entries[i].lcme_flags);
1781 le32_to_cpu(comp_v1->lcm_entries[i].lcme_id);
1782 if (lod_comp->llc_id == LCME_ID_INVAL)
1783 GOTO(out, rc = -EINVAL);
1787 if (magic == LOV_MAGIC_V1) {
1788 objs = &v1->lmm_objects[0];
1789 } else if (magic == LOV_MAGIC_V3) {
1790 objs = &v3->lmm_objects[0];
1791 if (v3->lmm_pool_name[0] != '\0')
1792 pool_name = v3->lmm_pool_name;
1794 CDEBUG(D_LAYOUT, "Invalid magic %x\n", magic);
1795 GOTO(out, rc = -EINVAL);
1798 lod_comp->llc_pattern = le32_to_cpu(v1->lmm_pattern);
1799 lod_comp->llc_stripe_size = le32_to_cpu(v1->lmm_stripe_size);
1800 lod_comp->llc_stripe_count = le16_to_cpu(v1->lmm_stripe_count);
1801 lod_comp->llc_layout_gen = le16_to_cpu(v1->lmm_layout_gen);
1803 * The stripe_offset of an uninit-ed component is stored in
1804 * the lmm_layout_gen
1806 if (mo->ldo_is_composite && !lod_comp_inited(lod_comp))
1807 lod_comp->llc_stripe_offset = lod_comp->llc_layout_gen;
1808 lod_obj_set_pool(mo, i, pool_name);
1810 if ((!mo->ldo_is_composite || lod_comp_inited(lod_comp)) &&
1811 !(lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED) &&
1812 !(lod_comp->llc_pattern & LOV_PATTERN_MDT)) {
1813 rc = lod_initialize_objects(env, mo, objs, i);
1819 rc = lod_fill_mirrors(mo);
1824 lod_object_free_striping(env, mo);
1830 * Parse suggested striping configuration.
1832 * The caller gets a suggested striping configuration from a number of sources
1833 * including per-directory default and applications. Then it needs to verify
1834 * the suggested striping is valid, apply missing bits and store the resulting
1835 * configuration in the object to be used by the allocator later. Must not be
1836 * called concurrently against the same object. It's OK to provide a
1837 * fully-defined striping.
1839 * \param[in] env execution environment for this thread
1840 * \param[in] lo LOD object
1841 * \param[in] buf buffer containing the striping
1843 * \retval 0 on success
1844 * \retval negative negated errno on error
1846 int lod_qos_parse_config(const struct lu_env *env, struct lod_object *lo,
1847 const struct lu_buf *buf)
1849 struct lod_layout_component *lod_comp;
1850 struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
1851 struct lov_desc *desc = &d->lod_desc;
1852 struct lov_user_md_v1 *v1 = NULL;
1853 struct lov_user_md_v3 *v3 = NULL;
1854 struct lov_comp_md_v1 *comp_v1 = NULL;
1855 char def_pool[LOV_MAXPOOLNAME + 1];
1862 if (buf == NULL || buf->lb_buf == NULL || buf->lb_len == 0)
1865 memset(def_pool, 0, sizeof(def_pool));
1866 if (lo->ldo_comp_entries != NULL)
1867 lod_layout_get_pool(lo->ldo_comp_entries, lo->ldo_comp_cnt,
1868 def_pool, sizeof(def_pool));
1870 /* free default striping info */
1871 lod_free_comp_entries(lo);
1873 rc = lod_verify_striping(d, lo, buf, false);
1879 comp_v1 = buf->lb_buf;
1880 magic = v1->lmm_magic;
1882 if (unlikely(le32_to_cpu(magic) & LOV_MAGIC_DEFINED)) {
1883 /* try to use as fully defined striping */
1884 rc = lod_use_defined_striping(env, lo, buf);
1889 case __swab32(LOV_USER_MAGIC_V1):
1890 lustre_swab_lov_user_md_v1(v1);
1891 magic = v1->lmm_magic;
1893 case LOV_USER_MAGIC_V1:
1895 case __swab32(LOV_USER_MAGIC_V3):
1896 lustre_swab_lov_user_md_v3(v3);
1897 magic = v3->lmm_magic;
1899 case LOV_USER_MAGIC_V3:
1901 case __swab32(LOV_USER_MAGIC_SPECIFIC):
1902 lustre_swab_lov_user_md_v3(v3);
1903 lustre_swab_lov_user_md_objects(v3->lmm_objects,
1904 v3->lmm_stripe_count);
1905 magic = v3->lmm_magic;
1907 case LOV_USER_MAGIC_SPECIFIC:
1909 case __swab32(LOV_USER_MAGIC_COMP_V1):
1910 lustre_swab_lov_comp_md_v1(comp_v1);
1911 magic = comp_v1->lcm_magic;
1913 case LOV_USER_MAGIC_COMP_V1:
1916 CERROR("%s: unrecognized magic %X\n",
1917 lod2obd(d)->obd_name, magic);
1921 lustre_print_user_md(D_OTHER, v1, "parse config");
1923 if (magic == LOV_USER_MAGIC_COMP_V1) {
1924 comp_cnt = comp_v1->lcm_entry_count;
1927 mirror_cnt = comp_v1->lcm_mirror_count + 1;
1929 lo->ldo_flr_state = LCM_FL_RDONLY;
1930 lo->ldo_is_composite = 1;
1934 lo->ldo_is_composite = 0;
1937 rc = lod_alloc_comp_entries(lo, mirror_cnt, comp_cnt);
1941 LASSERT(lo->ldo_comp_entries);
1943 for (i = 0; i < comp_cnt; i++) {
1944 struct pool_desc *pool;
1945 struct lu_extent *ext;
1948 lod_comp = &lo->ldo_comp_entries[i];
1950 if (lo->ldo_is_composite) {
1951 v1 = (struct lov_user_md *)((char *)comp_v1 +
1952 comp_v1->lcm_entries[i].lcme_offset);
1953 ext = &comp_v1->lcm_entries[i].lcme_extent;
1954 lod_comp->llc_extent = *ext;
1955 lod_comp->llc_flags =
1956 comp_v1->lcm_entries[i].lcme_flags &
1961 if (v1->lmm_magic == LOV_USER_MAGIC_V3 ||
1962 v1->lmm_magic == LOV_USER_MAGIC_SPECIFIC) {
1965 v3 = (struct lov_user_md_v3 *)v1;
1966 if (v3->lmm_pool_name[0] != '\0')
1967 pool_name = v3->lmm_pool_name;
1969 if (v3->lmm_magic == LOV_USER_MAGIC_SPECIFIC) {
1970 if (v3->lmm_stripe_offset == LOV_OFFSET_DEFAULT)
1971 v3->lmm_stripe_offset =
1972 v3->lmm_objects[0].l_ost_idx;
1974 /* copy ost list from lmm */
1975 lod_comp->llc_ostlist.op_count =
1976 v3->lmm_stripe_count;
1977 lod_comp->llc_ostlist.op_size =
1978 v3->lmm_stripe_count * sizeof(__u32);
1979 OBD_ALLOC(lod_comp->llc_ostlist.op_array,
1980 lod_comp->llc_ostlist.op_size);
1981 if (!lod_comp->llc_ostlist.op_array)
1982 GOTO(free_comp, rc = -ENOMEM);
1984 for (j = 0; j < v3->lmm_stripe_count; j++)
1985 lod_comp->llc_ostlist.op_array[j] =
1986 v3->lmm_objects[j].l_ost_idx;
1990 if (pool_name == NULL && def_pool[0] != '\0')
1991 pool_name = def_pool;
1993 if (v1->lmm_pattern == 0)
1994 v1->lmm_pattern = LOV_PATTERN_RAID0;
1995 if (lov_pattern(v1->lmm_pattern) != LOV_PATTERN_RAID0 &&
1996 lov_pattern(v1->lmm_pattern) != LOV_PATTERN_MDT) {
1997 CDEBUG(D_LAYOUT, "%s: invalid pattern: %x\n",
1998 lod2obd(d)->obd_name, v1->lmm_pattern);
1999 GOTO(free_comp, rc = -EINVAL);
2002 lod_comp->llc_pattern = v1->lmm_pattern;
2003 lod_comp->llc_stripe_size = desc->ld_default_stripe_size;
2004 if (v1->lmm_stripe_size)
2005 lod_comp->llc_stripe_size = v1->lmm_stripe_size;
2007 lod_comp->llc_stripe_count = desc->ld_default_stripe_count;
2008 if (v1->lmm_stripe_count ||
2009 lov_pattern(v1->lmm_pattern) == LOV_PATTERN_MDT)
2010 lod_comp->llc_stripe_count = v1->lmm_stripe_count;
2012 lod_comp->llc_stripe_offset = v1->lmm_stripe_offset;
2013 lod_obj_set_pool(lo, i, pool_name);
2015 LASSERT(ergo(lov_pattern(lod_comp->llc_pattern) ==
2016 LOV_PATTERN_MDT, lod_comp->llc_stripe_count == 0));
2018 if (pool_name == NULL)
2021 /* In the function below, .hs_keycmp resolves to
2022 * pool_hashkey_keycmp() */
2023 /* coverity[overrun-buffer-val] */
2024 pool = lod_find_pool(d, pool_name);
2028 if (lod_comp->llc_stripe_offset != LOV_OFFSET_DEFAULT) {
2029 rc = lod_check_index_in_pool(
2030 lod_comp->llc_stripe_offset, pool);
2032 lod_pool_putref(pool);
2033 CDEBUG(D_LAYOUT, "%s: invalid offset, %u\n",
2034 lod2obd(d)->obd_name,
2035 lod_comp->llc_stripe_offset);
2036 GOTO(free_comp, rc = -EINVAL);
2040 if (lod_comp->llc_stripe_count > pool_tgt_count(pool))
2041 lod_comp->llc_stripe_count = pool_tgt_count(pool);
2043 lod_pool_putref(pool);
2049 lod_free_comp_entries(lo);
2054 * Create a striping for an obejct.
2056 * The function creates a new striping for the object. The function tries QoS
2057 * algorithm first unless free space is distributed evenly among OSTs, but
2058 * by default RR algorithm is preferred due to internal concurrency (QoS is
2059 * serialized). The caller must ensure no concurrent calls to the function
2060 * are made against the same object.
2062 * \param[in] env execution environment for this thread
2063 * \param[in] lo LOD object
2064 * \param[in] attr attributes OST objects will be declared with
2065 * \param[in] th transaction handle
2066 * \param[in] comp_idx index of ldo_comp_entries
2067 * \param[in|out] inuse array of inuse ost index
2069 * \retval 0 on success
2070 * \retval negative negated errno on error
2072 int lod_qos_prep_create(const struct lu_env *env, struct lod_object *lo,
2073 struct lu_attr *attr, struct thandle *th,
2074 int comp_idx, struct ost_pool *inuse)
2076 struct lod_layout_component *lod_comp;
2077 struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
2078 struct dt_object **stripe;
2080 int flag = LOV_USES_ASSIGNED_STRIPE;
2085 LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
2086 lod_comp = &lo->ldo_comp_entries[comp_idx];
2088 /* A released component is being created */
2089 if (lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED)
2092 /* A Data-on-MDT component is being created */
2093 if (lov_pattern(lod_comp->llc_pattern) == LOV_PATTERN_MDT)
2096 if (likely(lod_comp->llc_stripe == NULL)) {
2098 * no striping has been created so far
2100 LASSERT(lod_comp->llc_stripe_count);
2102 * statfs and check OST targets now, since ld_active_tgt_count
2103 * could be changed if some OSTs are [de]activated manually.
2105 lod_qos_statfs_update(env, d);
2106 stripe_len = lod_get_stripe_count(d, lo,
2107 lod_comp->llc_stripe_count);
2108 if (stripe_len == 0)
2109 GOTO(out, rc = -ERANGE);
2110 lod_comp->llc_stripe_count = stripe_len;
2111 OBD_ALLOC(stripe, sizeof(stripe[0]) * stripe_len);
2113 GOTO(out, rc = -ENOMEM);
2115 lod_getref(&d->lod_ost_descs);
2116 /* XXX: support for non-0 files w/o objects */
2117 CDEBUG(D_OTHER, "tgt_count %d stripe_count %d\n",
2118 d->lod_desc.ld_tgt_count, stripe_len);
2120 if (lod_comp->llc_ostlist.op_array) {
2121 rc = lod_alloc_ost_list(env, lo, stripe, th, comp_idx,
2123 } else if (lod_comp->llc_stripe_offset == LOV_OFFSET_DEFAULT) {
2124 rc = lod_alloc_qos(env, lo, stripe, flag, th,
2127 rc = lod_alloc_rr(env, lo, stripe, flag, th,
2130 rc = lod_alloc_specific(env, lo, stripe, flag, th,
2133 lod_putref(d, &d->lod_ost_descs);
2136 for (i = 0; i < stripe_len; i++)
2137 if (stripe[i] != NULL)
2138 dt_object_put(env, stripe[i]);
2140 OBD_FREE(stripe, sizeof(stripe[0]) * stripe_len);
2141 lod_comp->llc_stripe_count = 0;
2143 lod_comp->llc_stripe = stripe;
2144 lod_comp->llc_stripes_allocated = stripe_len;
2148 * lod_qos_parse_config() found supplied buf as a predefined
2149 * striping (not a hint), so it allocated all the object
2150 * now we need to create them
2152 for (i = 0; i < lod_comp->llc_stripe_count; i++) {
2153 struct dt_object *o;
2155 o = lod_comp->llc_stripe[i];
2158 rc = lod_sub_declare_create(env, o, attr, NULL,
2161 CERROR("can't declare create: %d\n", rc);
2166 * Clear LCME_FL_INIT for the component so that
2167 * lod_striping_create() can create the striping objects
2170 lod_comp_unset_init(lod_comp);
2177 int lod_obj_stripe_set_inuse_cb(const struct lu_env *env,
2178 struct lod_object *lo,
2179 struct dt_object *dt, struct thandle *th,
2180 int comp_idx, int stripe_idx,
2181 struct lod_obj_stripe_cb_data *data)
2183 struct lod_thread_info *info = lod_env_info(env);
2184 struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
2185 struct lu_fid *fid = &info->lti_fid;
2187 int rc, type = LU_SEQ_RANGE_OST;
2189 *fid = *lu_object_fid(&dt->do_lu);
2190 rc = lod_fld_lookup(env, d, fid, &index, &type);
2192 CERROR("%s: fail to locate "DFID": rc = %d\n",
2193 lod2obd(d)->obd_name, PFID(fid), rc);
2196 lod_comp_ost_in_use(data->locd_inuse, index);
2201 * Resize per-thread ost list to hold OST target index list already used.
2203 * \param[in,out] inuse structure contains ost list array
2204 * \param[in] cnt total stripe count of all components
2205 * \param[in] max array's max size if @max > 0
2207 * \retval 0 on success
2208 * \retval -ENOMEM reallocation failed
2210 static int lod_inuse_resize(struct ost_pool *inuse, __u16 cnt, __u16 max)
2213 __u32 new = cnt * sizeof(inuse->op_array[0]);
2215 inuse->op_count = 0;
2217 if (new <= inuse->op_size)
2221 new = min_t(__u32, new, max);
2223 OBD_ALLOC(array, new);
2227 if (inuse->op_array)
2228 OBD_FREE(inuse->op_array, inuse->op_size);
2230 inuse->op_array = array;
2231 inuse->op_size = new;
2236 int lod_prepare_inuse(const struct lu_env *env, struct lod_object *lo)
2238 struct lod_thread_info *info = lod_env_info(env);
2239 struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
2240 struct ost_pool *inuse = &info->lti_inuse_osts;
2241 struct lod_obj_stripe_cb_data data = { { 0 } };
2242 __u32 stripe_count = 0;
2246 for (i = 0; i < lo->ldo_comp_cnt; i++)
2247 stripe_count += lod_comp_entry_stripe_count(lo,
2248 &lo->ldo_comp_entries[i], false);
2249 rc = lod_inuse_resize(inuse, stripe_count, d->lod_osd_max_easize);
2253 data.locd_inuse = inuse;
2254 data.locd_stripe_cb = lod_obj_stripe_set_inuse_cb;
2255 return lod_obj_for_each_stripe(env, lo, NULL, &data);
2258 int lod_prepare_create(const struct lu_env *env, struct lod_object *lo,
2259 struct lu_attr *attr, const struct lu_buf *buf,
2263 struct lod_thread_info *info = lod_env_info(env);
2264 struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
2265 struct ost_pool *inuse = &info->lti_inuse_osts;
2273 /* no OST available */
2274 /* XXX: should we be waiting a bit to prevent failures during
2275 * cluster initialization? */
2276 if (d->lod_ostnr == 0)
2280 * by this time, the object's ldo_stripe_count and ldo_stripe_size
2281 * contain default value for striping: taken from the parent
2282 * or from filesystem defaults
2284 * in case the caller is passing lovea with new striping config,
2285 * we may need to parse lovea and apply new configuration
2287 rc = lod_qos_parse_config(env, lo, buf);
2291 if (attr->la_valid & LA_SIZE)
2292 size = attr->la_size;
2295 rc = lod_prepare_inuse(env, lo);
2300 * prepare OST object creation for the component covering file's
2301 * size, the 1st component (including plain layout file) is always
2304 for (i = 0; i < lo->ldo_comp_cnt; i++) {
2305 struct lod_layout_component *lod_comp;
2306 struct lu_extent *extent;
2308 lod_comp = &lo->ldo_comp_entries[i];
2309 extent = &lod_comp->llc_extent;
2310 CDEBUG(D_QOS, "%lld [%lld, %lld)\n",
2311 size, extent->e_start, extent->e_end);
2312 if (!lo->ldo_is_composite || size >= extent->e_start) {
2313 rc = lod_qos_prep_create(env, lo, attr, th, i, inuse);