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;
83 down_write(&lod->lod_qos.lq_rw_sem);
85 * a bit hacky approach to learn NID of corresponding connection
86 * but there is no official API to access information like this
89 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
90 if (obd_uuid_equals(&oss->lqo_uuid,
91 &exp->exp_connection->c_remote_uuid)) {
102 GOTO(out, rc = -ENOMEM);
103 memcpy(&oss->lqo_uuid, &exp->exp_connection->c_remote_uuid,
104 sizeof(oss->lqo_uuid));
108 /* Assume we have to move this one */
109 list_del(&oss->lqo_oss_list);
112 oss->lqo_ost_count++;
113 ost_desc->ltd_qos.ltq_oss = oss;
115 CDEBUG(D_QOS, "add tgt %s to OSS %s (%d OSTs)\n",
116 obd_uuid2str(&ost_desc->ltd_uuid), obd_uuid2str(&oss->lqo_uuid),
119 /* Add sorted by # of OSTs. Find the first entry that we're
121 list = &lod->lod_qos.lq_oss_list;
122 list_for_each_entry(temposs, list, lqo_oss_list) {
123 if (oss->lqo_ost_count > temposs->lqo_ost_count)
126 /* ...and add before it. If we're the first or smallest, temposs
127 points to the list head, and we add to the end. */
128 list_add_tail(&oss->lqo_oss_list, &temposs->lqo_oss_list);
130 lod->lod_qos.lq_dirty = 1;
131 lod->lod_qos.lq_rr.lqr_dirty = 1;
134 up_write(&lod->lod_qos.lq_rw_sem);
139 * Remove OST target from QoS table.
141 * Removes given OST target from QoS table and releases related OSS structure
142 * if no OSTs remain on the OSS.
144 * \param[in] lod LOD device
145 * \param[in] ost_desc OST description
147 * \retval 0 on success
148 * \retval -ENOENT if no OSS was found
150 int qos_del_tgt(struct lod_device *lod, struct lod_tgt_desc *ost_desc)
152 struct lod_qos_oss *oss;
156 down_write(&lod->lod_qos.lq_rw_sem);
157 oss = ost_desc->ltd_qos.ltq_oss;
159 GOTO(out, rc = -ENOENT);
161 oss->lqo_ost_count--;
162 if (oss->lqo_ost_count == 0) {
163 CDEBUG(D_QOS, "removing OSS %s\n",
164 obd_uuid2str(&oss->lqo_uuid));
165 list_del(&oss->lqo_oss_list);
166 ost_desc->ltd_qos.ltq_oss = NULL;
170 lod->lod_qos.lq_dirty = 1;
171 lod->lod_qos.lq_rr.lqr_dirty = 1;
173 up_write(&lod->lod_qos.lq_rw_sem);
178 * Check whether the target is available for new OST objects.
180 * Request statfs data from the given target and verify it's active and not
181 * read-only. If so, then it can be used to place new OST objects. This
182 * function also maintains the number of active/inactive targets and sets
183 * dirty flags if those numbers change so others can run re-balance procedures.
184 * No external locking is required.
186 * \param[in] env execution environment for this thread
187 * \param[in] d LOD device
188 * \param[in] index index of OST target to check
189 * \param[out] sfs buffer for statfs data
191 * \retval 0 if the target is good
192 * \retval negative negated errno on error
195 static int lod_statfs_and_check(const struct lu_env *env, struct lod_device *d,
196 int index, struct obd_statfs *sfs)
198 struct lod_tgt_desc *ost;
203 ost = OST_TGT(d,index);
206 rc = dt_statfs(env, ost->ltd_ost, sfs);
208 if (rc == 0 && ((sfs->os_state & OS_STATE_ENOSPC) ||
209 (sfs->os_state & OS_STATE_ENOINO && sfs->os_fprecreated == 0)))
212 if (rc && rc != -ENOTCONN)
213 CERROR("%s: statfs: rc = %d\n", lod2obd(d)->obd_name, rc);
215 /* If the OST is readonly then we can't allocate objects there */
216 if (sfs->os_state & OS_STATE_READONLY)
219 /* check whether device has changed state (active, inactive) */
220 if (rc != 0 && ost->ltd_active) {
221 /* turned inactive? */
222 spin_lock(&d->lod_lock);
223 if (ost->ltd_active) {
226 ost->ltd_connecting = 1;
228 LASSERT(d->lod_desc.ld_active_tgt_count > 0);
229 d->lod_desc.ld_active_tgt_count--;
230 d->lod_qos.lq_dirty = 1;
231 d->lod_qos.lq_rr.lqr_dirty = 1;
232 CDEBUG(D_CONFIG, "%s: turns inactive\n",
233 ost->ltd_exp->exp_obd->obd_name);
235 spin_unlock(&d->lod_lock);
236 } else if (rc == 0 && ost->ltd_active == 0) {
238 LASSERTF(d->lod_desc.ld_active_tgt_count < d->lod_ostnr,
239 "active tgt count %d, ost nr %d\n",
240 d->lod_desc.ld_active_tgt_count, d->lod_ostnr);
241 spin_lock(&d->lod_lock);
242 if (ost->ltd_active == 0) {
244 ost->ltd_connecting = 0;
245 d->lod_desc.ld_active_tgt_count++;
246 d->lod_qos.lq_dirty = 1;
247 d->lod_qos.lq_rr.lqr_dirty = 1;
248 CDEBUG(D_CONFIG, "%s: turns active\n",
249 ost->ltd_exp->exp_obd->obd_name);
251 spin_unlock(&d->lod_lock);
258 * Maintain per-target statfs data.
260 * The function refreshes statfs data for all the targets every N seconds.
261 * The actual N is controlled via procfs and set to LOV_DESC_QOS_MAXAGE_DEFAULT
264 * \param[in] env execution environment for this thread
265 * \param[in] lod LOD device
267 void lod_qos_statfs_update(const struct lu_env *env, struct lod_device *lod)
269 struct obd_device *obd = lod2obd(lod);
270 struct ost_pool *osts = &(lod->lod_pool_info);
277 max_age = ktime_get_seconds() - 2 * lod->lod_desc.ld_qos_maxage;
279 if (obd->obd_osfs_age > max_age)
280 /* statfs data are quite recent, don't need to refresh it */
283 down_write(&lod->lod_qos.lq_rw_sem);
285 if (obd->obd_osfs_age > max_age)
288 for (i = 0; i < osts->op_count; i++) {
289 idx = osts->op_array[i];
290 avail = OST_TGT(lod,idx)->ltd_statfs.os_bavail;
291 if (lod_statfs_and_check(env, lod, idx,
292 &OST_TGT(lod, idx)->ltd_statfs))
294 if (OST_TGT(lod,idx)->ltd_statfs.os_bavail != avail)
295 /* recalculate weigths */
296 lod->lod_qos.lq_dirty = 1;
298 obd->obd_osfs_age = ktime_get_seconds();
301 up_write(&lod->lod_qos.lq_rw_sem);
306 * Calculate per-OST and per-OSS penalties
308 * Re-calculate penalties when the configuration changes, active targets
309 * change and after statfs refresh (all these are reflected by lq_dirty flag).
310 * On every OST and OSS: decay the penalty by half for every 8x the update
311 * interval that the device has been idle. That gives lots of time for the
312 * statfs information to be updated (which the penalty is only a proxy for),
313 * and avoids penalizing OSS/OSTs under light load.
314 * See lod_qos_calc_weight() for how penalties are factored into the weight.
316 * \param[in] lod LOD device
318 * \retval 0 on success
319 * \retval -EAGAIN the number of OSTs isn't enough
321 static int lod_qos_calc_ppo(struct lod_device *lod)
323 struct lod_qos_oss *oss;
324 __u64 ba_max, ba_min, temp;
331 if (!lod->lod_qos.lq_dirty)
334 num_active = lod->lod_desc.ld_active_tgt_count - 1;
336 GOTO(out, rc = -EAGAIN);
338 /* find bavail on each OSS */
339 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list)
341 lod->lod_qos.lq_active_oss_count = 0;
344 * How badly user wants to select OSTs "widely" (not recently chosen
345 * and not on recent OSS's). As opposed to "freely" (free space
348 prio_wide = 256 - lod->lod_qos.lq_prio_free;
350 ba_min = (__u64)(-1);
352 now = ktime_get_real_seconds();
353 /* Calculate OST penalty per object
354 * (lod ref taken in lod_qos_prep_create())
356 cfs_foreach_bit(lod->lod_ost_bitmap, i) {
357 LASSERT(OST_TGT(lod,i));
358 temp = TGT_BAVAIL(i);
361 ba_min = min(temp, ba_min);
362 ba_max = max(temp, ba_max);
364 /* Count the number of usable OSS's */
365 if (OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail == 0)
366 lod->lod_qos.lq_active_oss_count++;
367 OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail += temp;
369 /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
371 do_div(temp, num_active);
372 OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj =
373 (temp * prio_wide) >> 8;
375 age = (now - OST_TGT(lod,i)->ltd_qos.ltq_used) >> 3;
376 if (lod->lod_qos.lq_reset ||
377 age > 32 * lod->lod_desc.ld_qos_maxage)
378 OST_TGT(lod,i)->ltd_qos.ltq_penalty = 0;
379 else if (age > lod->lod_desc.ld_qos_maxage)
380 /* Decay OST penalty. */
381 OST_TGT(lod,i)->ltd_qos.ltq_penalty >>=
382 (age / lod->lod_desc.ld_qos_maxage);
385 num_active = lod->lod_qos.lq_active_oss_count - 1;
386 if (num_active < 1) {
387 /* If there's only 1 OSS, we can't penalize it, so instead
388 we have to double the OST penalty */
390 cfs_foreach_bit(lod->lod_ost_bitmap, i)
391 OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj <<= 1;
394 /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
395 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
396 temp = oss->lqo_bavail >> 1;
397 do_div(temp, oss->lqo_ost_count * num_active);
398 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
400 age = (now - oss->lqo_used) >> 3;
401 if (lod->lod_qos.lq_reset ||
402 age > 32 * lod->lod_desc.ld_qos_maxage)
403 oss->lqo_penalty = 0;
404 else if (age > lod->lod_desc.ld_qos_maxage)
405 /* Decay OSS penalty. */
406 oss->lqo_penalty >>= age / lod->lod_desc.ld_qos_maxage;
409 lod->lod_qos.lq_dirty = 0;
410 lod->lod_qos.lq_reset = 0;
412 /* If each ost has almost same free space,
413 * do rr allocation for better creation performance */
414 lod->lod_qos.lq_same_space = 0;
415 if ((ba_max * (256 - lod->lod_qos.lq_threshold_rr)) >> 8 < ba_min) {
416 lod->lod_qos.lq_same_space = 1;
417 /* Reset weights for the next time we enter qos mode */
418 lod->lod_qos.lq_reset = 1;
424 if (!rc && lod->lod_qos.lq_same_space)
431 * Calculate weight for a given OST target.
433 * The final OST weight is the number of bytes available minus the OST and
434 * OSS penalties. See lod_qos_calc_ppo() for how penalties are calculated.
436 * \param[in] lod LOD device, where OST targets are listed
437 * \param[in] i OST target index
441 static int lod_qos_calc_weight(struct lod_device *lod, int i)
445 temp = TGT_BAVAIL(i);
446 temp2 = OST_TGT(lod,i)->ltd_qos.ltq_penalty +
447 OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_penalty;
449 OST_TGT(lod,i)->ltd_qos.ltq_weight = 0;
451 OST_TGT(lod,i)->ltd_qos.ltq_weight = temp - temp2;
456 * Re-calculate weights.
458 * The function is called when some OST target was used for a new object. In
459 * this case we should re-calculate all the weights to keep new allocations
462 * \param[in] lod LOD device
463 * \param[in] osts OST pool where a new object was placed
464 * \param[in] index OST target where a new object was placed
465 * \param[out] total_wt new total weight for the pool
469 static int lod_qos_used(struct lod_device *lod, struct ost_pool *osts,
470 __u32 index, __u64 *total_wt)
472 struct lod_tgt_desc *ost;
473 struct lod_qos_oss *oss;
477 ost = OST_TGT(lod,index);
480 /* Don't allocate on this devuce anymore, until the next alloc_qos */
481 ost->ltd_qos.ltq_usable = 0;
483 oss = ost->ltd_qos.ltq_oss;
485 /* Decay old penalty by half (we're adding max penalty, and don't
486 want it to run away.) */
487 ost->ltd_qos.ltq_penalty >>= 1;
488 oss->lqo_penalty >>= 1;
490 /* mark the OSS and OST as recently used */
491 ost->ltd_qos.ltq_used = oss->lqo_used = ktime_get_real_seconds();
493 /* Set max penalties for this OST and OSS */
494 ost->ltd_qos.ltq_penalty +=
495 ost->ltd_qos.ltq_penalty_per_obj * lod->lod_ostnr;
496 oss->lqo_penalty += oss->lqo_penalty_per_obj *
497 lod->lod_qos.lq_active_oss_count;
499 /* Decrease all OSS penalties */
500 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
501 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
502 oss->lqo_penalty = 0;
504 oss->lqo_penalty -= oss->lqo_penalty_per_obj;
508 /* Decrease all OST penalties */
509 for (j = 0; j < osts->op_count; j++) {
512 i = osts->op_array[j];
513 if (!cfs_bitmap_check(lod->lod_ost_bitmap, i))
516 ost = OST_TGT(lod,i);
519 if (ost->ltd_qos.ltq_penalty <
520 ost->ltd_qos.ltq_penalty_per_obj)
521 ost->ltd_qos.ltq_penalty = 0;
523 ost->ltd_qos.ltq_penalty -=
524 ost->ltd_qos.ltq_penalty_per_obj;
526 lod_qos_calc_weight(lod, i);
528 /* Recalc the total weight of usable osts */
529 if (ost->ltd_qos.ltq_usable)
530 *total_wt += ost->ltd_qos.ltq_weight;
532 QOS_DEBUG("recalc tgt %d usable=%d avail=%llu"
533 " ostppo=%llu ostp=%llu ossppo=%llu"
534 " ossp=%llu wt=%llu\n",
535 i, ost->ltd_qos.ltq_usable, TGT_BAVAIL(i) >> 10,
536 ost->ltd_qos.ltq_penalty_per_obj >> 10,
537 ost->ltd_qos.ltq_penalty >> 10,
538 ost->ltd_qos.ltq_oss->lqo_penalty_per_obj >> 10,
539 ost->ltd_qos.ltq_oss->lqo_penalty >> 10,
540 ost->ltd_qos.ltq_weight >> 10);
546 void lod_qos_rr_init(struct lod_qos_rr *lqr)
548 spin_lock_init(&lqr->lqr_alloc);
553 #define LOV_QOS_EMPTY ((__u32)-1)
556 * Calculate optimal round-robin order with regard to OSSes.
558 * Place all the OSTs from pool \a src_pool in a special array to be used for
559 * round-robin (RR) stripe allocation. The placement algorithm interleaves
560 * OSTs from the different OSSs so that RR allocation can balance OSSs evenly.
561 * Resorts the targets when the number of active targets changes (because of
562 * a new target or activation/deactivation).
564 * \param[in] lod LOD device
565 * \param[in] src_pool OST pool
566 * \param[in] lqr round-robin list
568 * \retval 0 on success
569 * \retval -ENOMEM fails to allocate the array
571 static int lod_qos_calc_rr(struct lod_device *lod, struct ost_pool *src_pool,
572 struct lod_qos_rr *lqr)
574 struct lod_qos_oss *oss;
575 struct lod_tgt_desc *ost;
576 unsigned placed, real_count;
581 if (!lqr->lqr_dirty) {
582 LASSERT(lqr->lqr_pool.op_size);
586 /* Do actual allocation. */
587 down_write(&lod->lod_qos.lq_rw_sem);
590 * Check again. While we were sleeping on @lq_rw_sem something could
593 if (!lqr->lqr_dirty) {
594 LASSERT(lqr->lqr_pool.op_size);
595 up_write(&lod->lod_qos.lq_rw_sem);
599 real_count = src_pool->op_count;
601 /* Zero the pool array */
602 /* alloc_rr is holding a read lock on the pool, so nobody is adding/
603 deleting from the pool. The lq_rw_sem insures that nobody else
605 lqr->lqr_pool.op_count = real_count;
606 rc = lod_ost_pool_extend(&lqr->lqr_pool, real_count);
608 up_write(&lod->lod_qos.lq_rw_sem);
611 for (i = 0; i < lqr->lqr_pool.op_count; i++)
612 lqr->lqr_pool.op_array[i] = LOV_QOS_EMPTY;
614 /* Place all the OSTs from 1 OSS at the same time. */
616 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
619 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
622 if (!cfs_bitmap_check(lod->lod_ost_bitmap,
623 src_pool->op_array[i]))
626 ost = OST_TGT(lod,src_pool->op_array[i]);
627 LASSERT(ost && ost->ltd_ost);
628 if (ost->ltd_qos.ltq_oss != oss)
631 /* Evenly space these OSTs across arrayspace */
632 next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count;
633 while (lqr->lqr_pool.op_array[next] != LOV_QOS_EMPTY)
634 next = (next + 1) % lqr->lqr_pool.op_count;
636 lqr->lqr_pool.op_array[next] = src_pool->op_array[i];
643 up_write(&lod->lod_qos.lq_rw_sem);
645 if (placed != real_count) {
646 /* This should never happen */
647 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
648 "round-robin list (%d of %d).\n",
650 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
651 LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
652 lqr->lqr_pool.op_array[i]);
659 for (i = 0; i < lqr->lqr_pool.op_count; i++)
660 QOS_CONSOLE("rr #%d ost idx=%d\n", i, lqr->lqr_pool.op_array[i]);
667 * Instantiate and declare creation of a new object.
669 * The function instantiates LU representation for a new object on the
670 * specified device. Also it declares an intention to create that
671 * object on the storage target.
673 * Note lu_object_anon() is used which is a trick with regard to LU/OSD
674 * infrastructure - in the existing precreation framework we can't assign FID
675 * at this moment, we do this later once a transaction is started. So the
676 * special method instantiates FID-less object in the cache and later it
677 * will get a FID and proper placement in LU cache.
679 * \param[in] env execution environment for this thread
680 * \param[in] d LOD device
681 * \param[in] ost_idx OST target index where the object is being created
682 * \param[in] th transaction handle
684 * \retval object ptr on success, ERR_PTR() otherwise
686 static struct dt_object *lod_qos_declare_object_on(const struct lu_env *env,
687 struct lod_device *d,
691 struct lod_tgt_desc *ost;
692 struct lu_object *o, *n;
693 struct lu_device *nd;
694 struct dt_object *dt;
699 LASSERT(ost_idx < d->lod_osts_size);
700 ost = OST_TGT(d,ost_idx);
702 LASSERT(ost->ltd_ost);
704 nd = &ost->ltd_ost->dd_lu_dev;
707 * allocate anonymous object with zero fid, real fid
708 * will be assigned by OSP within transaction
709 * XXX: to be fixed with fully-functional OST fids
711 o = lu_object_anon(env, nd, NULL);
713 GOTO(out, dt = ERR_PTR(PTR_ERR(o)));
715 n = lu_object_locate(o->lo_header, nd->ld_type);
716 if (unlikely(n == NULL)) {
717 CERROR("can't find slice\n");
718 lu_object_put(env, o);
719 GOTO(out, dt = ERR_PTR(-EINVAL));
722 dt = container_of(n, struct dt_object, do_lu);
724 rc = lod_sub_declare_create(env, dt, NULL, NULL, NULL, th);
726 CDEBUG(D_OTHER, "can't declare creation on #%u: %d\n",
728 lu_object_put(env, o);
737 * Calculate a minimum acceptable stripe count.
739 * Return an acceptable stripe count depending on flag LOV_USES_DEFAULT_STRIPE:
740 * all stripes or 3/4 of stripes.
742 * \param[in] stripe_count number of stripes requested
743 * \param[in] flags 0 or LOV_USES_DEFAULT_STRIPE
745 * \retval acceptable stripecount
747 static int min_stripe_count(__u32 stripe_count, int flags)
749 return (flags & LOV_USES_DEFAULT_STRIPE ?
750 stripe_count - (stripe_count / 4) : stripe_count);
753 #define LOV_CREATE_RESEED_MULT 30
754 #define LOV_CREATE_RESEED_MIN 2000
757 * Initialize temporary OST-in-use array.
759 * Allocate or extend the array used to mark targets already assigned to a new
760 * striping so they are not used more than once.
762 * \param[in] env execution environment for this thread
763 * \param[in] stripes number of items needed in the array
765 * \retval 0 on success
766 * \retval -ENOMEM on error
768 static inline int lod_qos_ost_in_use_clear(const struct lu_env *env,
771 struct lod_thread_info *info = lod_env_info(env);
773 if (info->lti_ea_store_size < sizeof(int) * stripes)
774 lod_ea_store_resize(info, stripes * sizeof(int));
775 if (info->lti_ea_store_size < sizeof(int) * stripes) {
776 CERROR("can't allocate memory for ost-in-use array\n");
779 memset(info->lti_ea_store, -1, sizeof(int) * stripes);
784 * Remember a target in the array of used targets.
786 * Mark the given target as used for a new striping being created. The status
787 * of an OST in a striping can be checked with lod_qos_is_ost_used().
789 * \param[in] env execution environment for this thread
790 * \param[in] idx index in the array
791 * \param[in] ost OST target index to mark as used
793 static inline void lod_qos_ost_in_use(const struct lu_env *env,
796 struct lod_thread_info *info = lod_env_info(env);
797 int *osts = info->lti_ea_store;
799 LASSERT(info->lti_ea_store_size >= idx * sizeof(int));
804 * Check is OST used in a striping.
806 * Checks whether OST with the given index is marked as used in the temporary
807 * array (see lod_qos_ost_in_use()).
809 * \param[in] env execution environment for this thread
810 * \param[in] ost OST target index to check
811 * \param[in] stripes the number of items used in the array already
816 static int lod_qos_is_ost_used(const struct lu_env *env, int ost, __u32 stripes)
818 struct lod_thread_info *info = lod_env_info(env);
819 int *osts = info->lti_ea_store;
822 for (j = 0; j < stripes; j++) {
830 * Check is OST used in a composite layout
832 * \param[in] inuse all inuse ost indexs
833 * \param[in] ost OST target index to check
838 static inline int lod_comp_is_ost_used(struct ost_pool *inuse, int ost)
841 LASSERT(inuse != NULL);
843 if (inuse->op_size == 0)
846 LASSERT(inuse->op_count * sizeof(inuse->op_array[0]) <= inuse->op_size);
847 for (j = 0; j < inuse->op_count; j++) {
848 if (inuse->op_array[j] == ost)
855 * Mark the given target as used for a composite layout
857 * \param[in] inuse inuse ost index array
858 * \param[in] idx index in the array
860 static inline void lod_comp_ost_in_use(struct ost_pool *inuse, int ost)
862 LASSERT(inuse != NULL);
863 if (inuse->op_size && !lod_comp_is_ost_used(inuse, ost)) {
864 LASSERTF(inuse->op_count * sizeof(inuse->op_array[0]) <
866 "count %d size %u", inuse->op_count, inuse->op_size);
867 inuse->op_array[inuse->op_count] = ost;
872 static inline void lod_avoid_update(struct lod_object *lo,
873 struct lod_avoid_guide *lag)
878 lag->lag_ost_avail--;
881 static inline bool lod_should_avoid_ost(struct lod_object *lo,
882 struct lod_avoid_guide *lag,
885 struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
886 struct lod_tgt_desc *ost = OST_TGT(lod, index);
887 struct lod_qos_oss *lqo = ost->ltd_qos.ltq_oss;
891 if (!cfs_bitmap_check(lod->lod_ost_bitmap, index))
895 * we've tried our best, all available OSTs have been used in
896 * overlapped components in the other mirror
898 if (lag->lag_ost_avail == 0)
902 for (i = 0; i < lag->lag_oaa_count; i++) {
903 if (lag->lag_oss_avoid_array[i] == lqo->lqo_id) {
909 * if the OSS which OST[index] resides has not been used, we'd like to
915 /* if the OSS has been used, check whether the OST has been used */
916 if (!cfs_bitmap_check(lag->lag_ost_avoid_bitmap, index))
919 QOS_DEBUG("OST%d: has been used in overlapped component "
920 "in other mirror\n", index);
924 static int lod_check_and_reserve_ost(const struct lu_env *env,
925 struct lod_object *lo,
926 struct obd_statfs *sfs, __u32 ost_idx,
927 __u32 speed, __u32 *s_idx,
928 struct dt_object **stripe,
931 struct ost_pool *inuse)
933 struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
934 struct lod_avoid_guide *lag = &lod_env_info(env)->lti_avoid;
936 __u32 stripe_idx = *s_idx;
939 rc = lod_statfs_and_check(env, lod, ost_idx, sfs);
941 /* this OSP doesn't feel well */
946 * We expect number of precreated objects in f_ffree at
947 * the first iteration, skip OSPs with no objects ready
949 if (sfs->os_fprecreated == 0 && speed == 0) {
950 QOS_DEBUG("#%d: precreation is empty\n", ost_idx);
955 * try to use another OSP if this one is degraded
957 if (sfs->os_state & OS_STATE_DEGRADED && speed < 2) {
958 QOS_DEBUG("#%d: degraded\n", ost_idx);
963 * try not allocate on OST which has been used by other
966 if (speed == 0 && lod_comp_is_ost_used(inuse, ost_idx)) {
967 QOS_DEBUG("#%d: used by other component\n", ost_idx);
972 * try not allocate OSTs used by conflicting component of other mirrors
973 * for the first and second time.
975 if (speed < 2 && lod_should_avoid_ost(lo, lag, ost_idx)) {
976 QOS_DEBUG("#%d: used by overlapped component of other mirror\n",
981 * do not put >1 objects on a single OST
983 if (lod_qos_is_ost_used(env, ost_idx, stripe_idx))
986 o = lod_qos_declare_object_on(env, lod, ost_idx, th);
988 CDEBUG(D_OTHER, "can't declare new object on #%u: %d\n",
989 ost_idx, (int) PTR_ERR(o));
995 * We've successfully declared (reserved) an object
997 lod_avoid_update(lo, lag);
998 lod_qos_ost_in_use(env, stripe_idx, ost_idx);
999 lod_comp_ost_in_use(inuse, ost_idx);
1000 stripe[stripe_idx] = o;
1001 ost_indices[stripe_idx] = ost_idx;
1002 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LOV_CREATE_RACE, 2);
1004 *s_idx = stripe_idx;
1011 * Allocate a striping using round-robin algorithm.
1013 * Allocates a new striping using round-robin algorithm. The function refreshes
1014 * all the internal structures (statfs cache, array of available OSTs sorted
1015 * with regard to OSS, etc). The number of stripes required is taken from the
1016 * object (must be prepared by the caller), but can change if the flag
1017 * LOV_USES_DEFAULT_STRIPE is supplied. The caller should ensure nobody else
1018 * is trying to create a striping on the object in parallel. All the internal
1019 * structures (like pools, etc) are protected and no additional locking is
1020 * required. The function succeeds even if a single stripe is allocated. To save
1021 * time we give priority to targets which already have objects precreated.
1022 * Full OSTs are skipped (see lod_qos_dev_is_full() for the details).
1024 * \param[in] env execution environment for this thread
1025 * \param[in] lo LOD object
1026 * \param[out] stripe striping created
1027 * \param[out] ost_indices ost indices of striping created
1028 * \param[in] flags allocation flags (0 or LOV_USES_DEFAULT_STRIPE)
1029 * \param[in] th transaction handle
1030 * \param[in] comp_idx index of ldo_comp_entries
1031 * \param[in|out] inuse array of inuse ost index
1033 * \retval 0 on success
1034 * \retval -ENOSPC if not enough OSTs are found
1035 * \retval negative negated errno for other failures
1037 static int lod_alloc_rr(const struct lu_env *env, struct lod_object *lo,
1038 struct dt_object **stripe, __u32 *ost_indices,
1039 int flags, struct thandle *th, int comp_idx,
1040 struct ost_pool *inuse)
1042 struct lod_layout_component *lod_comp;
1043 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1044 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
1045 struct pool_desc *pool = NULL;
1046 struct ost_pool *osts;
1047 struct lod_qos_rr *lqr;
1048 unsigned int i, array_idx;
1049 __u32 ost_start_idx_temp;
1050 __u32 stripe_idx = 0;
1051 __u32 stripe_count, stripe_count_min, ost_idx;
1052 int rc, speed = 0, ost_connecting = 0;
1055 LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
1056 lod_comp = &lo->ldo_comp_entries[comp_idx];
1057 stripe_count = lod_comp->llc_stripe_count;
1058 stripe_count_min = min_stripe_count(stripe_count, flags);
1060 if (lod_comp->llc_pool != NULL)
1061 pool = lod_find_pool(m, lod_comp->llc_pool);
1064 down_read(&pool_tgt_rw_sem(pool));
1065 osts = &(pool->pool_obds);
1066 lqr = &(pool->pool_rr);
1068 osts = &(m->lod_pool_info);
1069 lqr = &(m->lod_qos.lq_rr);
1072 rc = lod_qos_calc_rr(m, osts, lqr);
1076 rc = lod_qos_ost_in_use_clear(env, stripe_count);
1080 down_read(&m->lod_qos.lq_rw_sem);
1081 spin_lock(&lqr->lqr_alloc);
1082 if (--lqr->lqr_start_count <= 0) {
1083 lqr->lqr_start_idx = cfs_rand() % osts->op_count;
1084 lqr->lqr_start_count =
1085 (LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) +
1086 LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U);
1087 } else if (stripe_count_min >= osts->op_count ||
1088 lqr->lqr_start_idx > osts->op_count) {
1089 /* If we have allocated from all of the OSTs, slowly
1090 * precess the next start if the OST/stripe count isn't
1091 * already doing this for us. */
1092 lqr->lqr_start_idx %= osts->op_count;
1093 if (stripe_count > 1 && (osts->op_count % stripe_count) != 1)
1094 ++lqr->lqr_offset_idx;
1096 ost_start_idx_temp = lqr->lqr_start_idx;
1100 QOS_DEBUG("pool '%s' want %d start_idx %d start_count %d offset %d "
1101 "active %d count %d\n",
1102 lod_comp->llc_pool ? lod_comp->llc_pool : "",
1103 stripe_count, lqr->lqr_start_idx, lqr->lqr_start_count,
1104 lqr->lqr_offset_idx, osts->op_count, osts->op_count);
1106 for (i = 0; i < osts->op_count && stripe_idx < stripe_count; i++) {
1107 array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) %
1109 ++lqr->lqr_start_idx;
1110 ost_idx = lqr->lqr_pool.op_array[array_idx];
1112 QOS_DEBUG("#%d strt %d act %d strp %d ary %d idx %d\n",
1113 i, lqr->lqr_start_idx, /* XXX: active*/ 0,
1114 stripe_idx, array_idx, ost_idx);
1116 if ((ost_idx == LOV_QOS_EMPTY) ||
1117 !cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
1120 /* Fail Check before osc_precreate() is called
1121 so we can only 'fail' single OSC. */
1122 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
1125 spin_unlock(&lqr->lqr_alloc);
1126 rc = lod_check_and_reserve_ost(env, lo, sfs, ost_idx, speed,
1127 &stripe_idx, stripe, ost_indices,
1129 spin_lock(&lqr->lqr_alloc);
1131 if (rc != 0 && OST_TGT(m, ost_idx)->ltd_connecting)
1134 if ((speed < 2) && (stripe_idx < stripe_count_min)) {
1135 /* Try again, allowing slower OSCs */
1137 lqr->lqr_start_idx = ost_start_idx_temp;
1143 spin_unlock(&lqr->lqr_alloc);
1144 up_read(&m->lod_qos.lq_rw_sem);
1147 lod_comp->llc_stripe_count = stripe_idx;
1148 /* at least one stripe is allocated */
1151 /* nobody provided us with a single object */
1160 up_read(&pool_tgt_rw_sem(pool));
1161 /* put back ref got by lod_find_pool() */
1162 lod_pool_putref(pool);
1169 * Allocate a specific striping layout on a user defined set of OSTs.
1171 * Allocates new striping using the OST index range provided by the data from
1172 * the lmm_obejcts contained in the lov_user_md passed to this method. Full
1173 * OSTs are not considered. The exact order of OSTs requested by the user
1174 * is respected as much as possible depending on OST status. The number of
1175 * stripes needed and stripe offset are taken from the object. If that number
1176 * can not be met, then the function returns a failure and then it's the
1177 * caller's responsibility to release the stripes allocated. All the internal
1178 * structures are protected, but no concurrent allocation is allowed on the
1181 * \param[in] env execution environment for this thread
1182 * \param[in] lo LOD object
1183 * \param[out] stripe striping created
1184 * \param[out] ost_indices ost indices of striping created
1185 * \param[in] th transaction handle
1186 * \param[in] comp_idx index of ldo_comp_entries
1187 * \param[in|out] inuse array of inuse ost index
1189 * \retval 0 on success
1190 * \retval -ENODEV OST index does not exist on file system
1191 * \retval -EINVAL requested OST index is invalid
1192 * \retval negative negated errno on error
1194 static int lod_alloc_ost_list(const struct lu_env *env, struct lod_object *lo,
1195 struct dt_object **stripe, __u32 *ost_indices,
1196 struct thandle *th, int comp_idx,
1197 struct ost_pool *inuse)
1199 struct lod_layout_component *lod_comp;
1200 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1201 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
1202 struct dt_object *o;
1203 unsigned int array_idx = 0;
1204 int stripe_count = 0;
1209 /* for specific OSTs layout */
1210 LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
1211 lod_comp = &lo->ldo_comp_entries[comp_idx];
1212 LASSERT(lod_comp->llc_ostlist.op_array);
1214 rc = lod_qos_ost_in_use_clear(env, lod_comp->llc_stripe_count);
1218 for (i = 0; i < lod_comp->llc_stripe_count; i++) {
1219 if (lod_comp->llc_ostlist.op_array[i] ==
1220 lod_comp->llc_stripe_offset) {
1225 if (i == lod_comp->llc_stripe_count) {
1227 "%s: start index %d not in the specified list of OSTs\n",
1228 lod2obd(m)->obd_name, lod_comp->llc_stripe_offset);
1232 for (i = 0; i < lod_comp->llc_stripe_count;
1233 i++, array_idx = (array_idx + 1) % lod_comp->llc_stripe_count) {
1234 __u32 ost_idx = lod_comp->llc_ostlist.op_array[array_idx];
1236 if (!cfs_bitmap_check(m->lod_ost_bitmap, ost_idx)) {
1242 * do not put >1 objects on a single OST
1244 if (lod_qos_is_ost_used(env, ost_idx, stripe_count)) {
1249 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
1250 if (rc < 0) /* this OSP doesn't feel well */
1253 o = lod_qos_declare_object_on(env, m, ost_idx, th);
1257 "%s: can't declare new object on #%u: %d\n",
1258 lod2obd(m)->obd_name, ost_idx, rc);
1263 * We've successfully declared (reserved) an object
1265 lod_qos_ost_in_use(env, stripe_count, ost_idx);
1266 lod_comp_ost_in_use(inuse, ost_idx);
1267 stripe[stripe_count] = o;
1268 ost_indices[stripe_count] = ost_idx;
1276 * Allocate a striping on a predefined set of OSTs.
1278 * Allocates new layout starting from OST index in lo->ldo_stripe_offset.
1279 * Full OSTs are not considered. The exact order of OSTs is not important and
1280 * varies depending on OST status. The allocation procedure prefers the targets
1281 * with precreated objects ready. The number of stripes needed and stripe
1282 * offset are taken from the object. If that number cannot be met, then the
1283 * function returns an error and then it's the caller's responsibility to
1284 * release the stripes allocated. All the internal structures are protected,
1285 * but no concurrent allocation is allowed on the same objects.
1287 * \param[in] env execution environment for this thread
1288 * \param[in] lo LOD object
1289 * \param[out] stripe striping created
1290 * \param[out] ost_indices ost indices of striping created
1291 * \param[in] flags not used
1292 * \param[in] th transaction handle
1293 * \param[in] comp_idx index of ldo_comp_entries
1294 * \param[in|out]inuse array of inuse ost index
1296 * \retval 0 on success
1297 * \retval -ENOSPC if no OST objects are available at all
1298 * \retval -EFBIG if not enough OST objects are found
1299 * \retval -EINVAL requested offset is invalid
1300 * \retval negative errno on failure
1302 static int lod_alloc_specific(const struct lu_env *env, struct lod_object *lo,
1303 struct dt_object **stripe, __u32 *ost_indices,
1304 int flags, struct thandle *th, int comp_idx,
1305 struct ost_pool *inuse)
1307 struct lod_layout_component *lod_comp;
1308 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1309 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
1310 struct dt_object *o;
1312 unsigned int i, array_idx, ost_count;
1313 int rc, stripe_num = 0;
1315 struct pool_desc *pool = NULL;
1316 struct ost_pool *osts;
1319 LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
1320 lod_comp = &lo->ldo_comp_entries[comp_idx];
1322 rc = lod_qos_ost_in_use_clear(env, lod_comp->llc_stripe_count);
1326 if (lod_comp->llc_pool != NULL)
1327 pool = lod_find_pool(m, lod_comp->llc_pool);
1330 down_read(&pool_tgt_rw_sem(pool));
1331 osts = &(pool->pool_obds);
1333 osts = &(m->lod_pool_info);
1336 ost_count = osts->op_count;
1339 /* search loi_ost_idx in ost array */
1341 for (i = 0; i < ost_count; i++) {
1342 if (osts->op_array[i] == lod_comp->llc_stripe_offset) {
1347 if (i == ost_count) {
1348 CERROR("Start index %d not found in pool '%s'\n",
1349 lod_comp->llc_stripe_offset,
1350 lod_comp->llc_pool ? lod_comp->llc_pool : "");
1351 GOTO(out, rc = -EINVAL);
1354 for (i = 0; i < ost_count;
1355 i++, array_idx = (array_idx + 1) % ost_count) {
1356 ost_idx = osts->op_array[array_idx];
1358 if (!cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
1361 /* Fail Check before osc_precreate() is called
1362 so we can only 'fail' single OSC. */
1363 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
1367 * do not put >1 objects on a single OST
1369 if (lod_qos_is_ost_used(env, ost_idx, stripe_num))
1373 * try not allocate on the OST used by other component
1375 if (speed == 0 && i != 0 &&
1376 lod_comp_is_ost_used(inuse, ost_idx))
1379 /* Drop slow OSCs if we can, but not for requested start idx.
1381 * This means "if OSC is slow and it is not the requested
1382 * start OST, then it can be skipped, otherwise skip it only
1383 * if it is inactive/recovering/out-of-space." */
1385 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
1387 /* this OSP doesn't feel well */
1392 * We expect number of precreated objects at the first
1393 * iteration. Skip OSPs with no objects ready. Don't apply
1394 * this logic to OST specified with stripe_offset.
1396 if (i != 0 && sfs->os_fprecreated == 0 && speed == 0)
1399 o = lod_qos_declare_object_on(env, m, ost_idx, th);
1401 CDEBUG(D_OTHER, "can't declare new object on #%u: %d\n",
1402 ost_idx, (int) PTR_ERR(o));
1407 * We've successfully declared (reserved) an object
1409 lod_qos_ost_in_use(env, stripe_num, ost_idx);
1410 lod_comp_ost_in_use(inuse, ost_idx);
1411 stripe[stripe_num] = o;
1412 ost_indices[stripe_num] = ost_idx;
1415 /* We have enough stripes */
1416 if (stripe_num == lod_comp->llc_stripe_count)
1420 /* Try again, allowing slower OSCs */
1425 /* If we were passed specific striping params, then a failure to
1426 * meet those requirements is an error, since we can't reallocate
1427 * that memory (it might be part of a larger array or something).
1429 CERROR("can't lstripe objid "DFID": have %d want %u\n",
1430 PFID(lu_object_fid(lod2lu_obj(lo))), stripe_num,
1431 lod_comp->llc_stripe_count);
1432 rc = stripe_num == 0 ? -ENOSPC : -EFBIG;
1435 up_read(&pool_tgt_rw_sem(pool));
1436 /* put back ref got by lod_find_pool() */
1437 lod_pool_putref(pool);
1444 * Check whether QoS allocation should be used.
1446 * A simple helper to decide when QoS allocation should be used:
1447 * if it's just a single available target or the used space is
1448 * evenly distributed among the targets at the moment, then QoS
1449 * allocation algorithm should not be used.
1451 * \param[in] lod LOD device
1453 * \retval 0 should not be used
1454 * \retval 1 should be used
1456 static inline int lod_qos_is_usable(struct lod_device *lod)
1459 /* to be able to debug QoS code */
1463 /* Detect -EAGAIN early, before expensive lock is taken. */
1464 if (!lod->lod_qos.lq_dirty && lod->lod_qos.lq_same_space)
1467 if (lod->lod_desc.ld_active_tgt_count < 2)
1474 * Allocate a striping using an algorithm with weights.
1476 * The function allocates OST objects to create a striping. The algorithm
1477 * used is based on weights (currently only using the free space), and it's
1478 * trying to ensure the space is used evenly by OSTs and OSSs. The striping
1479 * configuration (# of stripes, offset, pool) is taken from the object and
1480 * is prepared by the caller.
1482 * If LOV_USES_DEFAULT_STRIPE is not passed and prepared configuration can't
1483 * be met due to too few OSTs, then allocation fails. If the flag is passed
1484 * fewer than 3/4 of the requested number of stripes can be allocated, then
1487 * No concurrent allocation is allowed on the object and this must be ensured
1488 * by the caller. All the internal structures are protected by the function.
1490 * The algorithm has two steps: find available OSTs and calculate their
1491 * weights, then select the OSTs with their weights used as the probability.
1492 * An OST with a higher weight is proportionately more likely to be selected
1493 * than one with a lower weight.
1495 * \param[in] env execution environment for this thread
1496 * \param[in] lo LOD object
1497 * \param[out] stripe striping created
1498 * \param[out] ost_indices ost indices of striping created
1499 * \param[in] flags 0 or LOV_USES_DEFAULT_STRIPE
1500 * \param[in] th transaction handle
1501 * \param[in] comp_idx index of ldo_comp_entries
1502 * \param[in|out]inuse array of inuse ost index
1504 * \retval 0 on success
1505 * \retval -EAGAIN not enough OSTs are found for specified stripe count
1506 * \retval -EINVAL requested OST index is invalid
1507 * \retval negative errno on failure
1509 static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo,
1510 struct dt_object **stripe, __u32 *ost_indices,
1511 int flags, struct thandle *th, int comp_idx,
1512 struct ost_pool *inuse)
1514 struct lod_layout_component *lod_comp;
1515 struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1516 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
1517 struct lod_avoid_guide *lag = &lod_env_info(env)->lti_avoid;
1518 struct lod_tgt_desc *ost;
1519 struct dt_object *o;
1520 __u64 total_weight = 0;
1521 struct pool_desc *pool = NULL;
1522 struct ost_pool *osts;
1524 __u32 nfound, good_osts, stripe_count, stripe_count_min;
1525 __u32 inuse_old_count = inuse->op_count;
1529 LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
1530 lod_comp = &lo->ldo_comp_entries[comp_idx];
1531 stripe_count = lod_comp->llc_stripe_count;
1532 stripe_count_min = min_stripe_count(stripe_count, flags);
1533 if (stripe_count_min < 1)
1536 if (lod_comp->llc_pool != NULL)
1537 pool = lod_find_pool(lod, lod_comp->llc_pool);
1540 down_read(&pool_tgt_rw_sem(pool));
1541 osts = &(pool->pool_obds);
1543 osts = &(lod->lod_pool_info);
1546 /* Detect -EAGAIN early, before expensive lock is taken. */
1547 if (!lod_qos_is_usable(lod))
1548 GOTO(out_nolock, rc = -EAGAIN);
1550 /* Do actual allocation, use write lock here. */
1551 down_write(&lod->lod_qos.lq_rw_sem);
1554 * Check again, while we were sleeping on @lq_rw_sem things could
1557 if (!lod_qos_is_usable(lod))
1558 GOTO(out, rc = -EAGAIN);
1560 rc = lod_qos_calc_ppo(lod);
1564 rc = lod_qos_ost_in_use_clear(env, lod_comp->llc_stripe_count);
1569 /* Find all the OSTs that are valid stripe candidates */
1570 for (i = 0; i < osts->op_count; i++) {
1571 if (!cfs_bitmap_check(lod->lod_ost_bitmap, osts->op_array[i]))
1574 ost = OST_TGT(lod, osts->op_array[i]);
1575 ost->ltd_qos.ltq_usable = 0;
1577 rc = lod_statfs_and_check(env, lod, osts->op_array[i], sfs);
1579 /* this OSP doesn't feel well */
1583 if (sfs->os_state & OS_STATE_DEGRADED)
1586 /* Fail Check before osc_precreate() is called
1587 so we can only 'fail' single OSC. */
1588 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) &&
1589 osts->op_array[i] == 0)
1592 ost->ltd_qos.ltq_usable = 1;
1593 lod_qos_calc_weight(lod, osts->op_array[i]);
1594 total_weight += ost->ltd_qos.ltq_weight;
1599 QOS_DEBUG("found %d good osts\n", good_osts);
1601 if (good_osts < stripe_count_min)
1602 GOTO(out, rc = -EAGAIN);
1604 /* We have enough osts */
1605 if (good_osts < stripe_count)
1606 stripe_count = good_osts;
1608 /* Find enough OSTs with weighted random allocation. */
1610 while (nfound < stripe_count) {
1611 __u64 rand, cur_weight;
1617 #if BITS_PER_LONG == 32
1618 rand = cfs_rand() % (unsigned)total_weight;
1619 /* If total_weight > 32-bit, first generate the high
1620 * 32 bits of the random number, then add in the low
1621 * 32 bits (truncated to the upper limit, if needed) */
1622 if (total_weight > 0xffffffffULL)
1623 rand = (__u64)(cfs_rand() %
1624 (unsigned)(total_weight >> 32)) << 32;
1628 if (rand == (total_weight & 0xffffffff00000000ULL))
1629 rand |= cfs_rand() % (unsigned)total_weight;
1634 rand = ((__u64)cfs_rand() << 32 | cfs_rand()) %
1641 /* On average, this will hit larger-weighted OSTs more often.
1642 * 0-weight OSTs will always get used last (only when rand=0) */
1643 for (i = 0; i < osts->op_count; i++) {
1644 __u32 idx = osts->op_array[i];
1646 if (lod_should_avoid_ost(lo, lag, idx))
1649 ost = OST_TGT(lod, idx);
1651 if (!ost->ltd_qos.ltq_usable)
1654 cur_weight += ost->ltd_qos.ltq_weight;
1655 QOS_DEBUG("stripe_count=%d nfound=%d cur_weight=%llu "
1656 "rand=%llu total_weight=%llu\n",
1657 stripe_count, nfound, cur_weight, rand,
1660 if (cur_weight < rand)
1663 QOS_DEBUG("stripe=%d to idx=%d\n", nfound, idx);
1665 * do not put >1 objects on a single OST
1667 if (lod_qos_is_ost_used(env, idx, nfound) ||
1668 lod_comp_is_ost_used(inuse, idx))
1671 o = lod_qos_declare_object_on(env, lod, idx, th);
1673 QOS_DEBUG("can't declare object on #%u: %d\n",
1674 idx, (int) PTR_ERR(o));
1678 lod_avoid_update(lo, lag);
1679 lod_qos_ost_in_use(env, nfound, idx);
1680 lod_comp_ost_in_use(inuse, idx);
1682 ost_indices[nfound] = idx;
1683 lod_qos_used(lod, osts, idx, &total_weight);
1690 /* no OST found on this iteration, give up */
1695 if (unlikely(nfound != stripe_count)) {
1697 * when the decision to use weighted algorithm was made
1698 * we had enough appropriate OSPs, but this state can
1699 * change anytime (no space on OST, broken connection, etc)
1700 * so it's possible OSP won't be able to provide us with
1701 * an object due to just changed state
1703 QOS_DEBUG("%s: wanted %d objects, found only %d\n",
1704 lod2obd(lod)->obd_name, stripe_count, nfound);
1705 for (i = 0; i < nfound; i++) {
1706 LASSERT(stripe[i] != NULL);
1707 dt_object_put(env, stripe[i]);
1710 LASSERTF(nfound <= inuse->op_count,
1711 "nfound:%d, op_count:%u\n", nfound, inuse->op_count);
1712 inuse->op_count = inuse_old_count;
1714 /* makes sense to rebalance next time */
1715 lod->lod_qos.lq_dirty = 1;
1716 lod->lod_qos.lq_same_space = 0;
1722 up_write(&lod->lod_qos.lq_rw_sem);
1726 up_read(&pool_tgt_rw_sem(pool));
1727 /* put back ref got by lod_find_pool() */
1728 lod_pool_putref(pool);
1735 * Find largest stripe count the caller can use.
1737 * Find the maximal possible stripe count not greater than \a stripe_count.
1738 * Sometimes suggested stripecount can't be reached for a number of reasons:
1739 * lack of enough active OSTs or the backend does not support EAs that large.
1740 * If the passed one is 0, then the filesystem's default one is used.
1742 * \param[in] lod LOD device
1743 * \param[in] lo The lod_object
1744 * \param[in] stripe_count count the caller would like to use
1746 * \retval the maximum usable stripe count
1748 __u16 lod_get_stripe_count(struct lod_device *lod, struct lod_object *lo,
1751 __u32 max_stripes = LOV_MAX_STRIPE_COUNT_OLD;
1754 stripe_count = lod->lod_desc.ld_default_stripe_count;
1755 if (stripe_count > lod->lod_desc.ld_active_tgt_count)
1756 stripe_count = lod->lod_desc.ld_active_tgt_count;
1760 /* stripe count is based on whether OSD can handle larger EA sizes */
1761 if (lod->lod_osd_max_easize > 0) {
1762 unsigned int easize = lod->lod_osd_max_easize;
1765 if (lo->ldo_is_composite) {
1766 struct lod_layout_component *lod_comp;
1767 unsigned int header_sz = sizeof(struct lov_comp_md_v1);
1769 header_sz += sizeof(struct lov_comp_md_entry_v1) *
1771 for (i = 0; i < lo->ldo_comp_cnt; i++) {
1772 lod_comp = &lo->ldo_comp_entries[i];
1773 if (lod_comp->llc_flags & LCME_FL_INIT)
1774 header_sz += lov_mds_md_size(
1775 lod_comp->llc_stripe_count,
1778 if (easize > header_sz)
1779 easize -= header_sz;
1784 max_stripes = lov_mds_md_max_stripe_count(easize, LOV_MAGIC_V3);
1787 return (stripe_count < max_stripes) ? stripe_count : max_stripes;
1791 * Create in-core respresentation for a fully-defined striping
1793 * When the caller passes a fully-defined striping (i.e. everything including
1794 * OST object FIDs are defined), then we still need to instantiate LU-cache
1795 * with the objects representing the stripes defined. This function completes
1798 * \param[in] env execution environment for this thread
1799 * \param[in] mo LOD object
1800 * \param[in] buf buffer containing the striping
1802 * \retval 0 on success
1803 * \retval negative negated errno on error
1805 int lod_use_defined_striping(const struct lu_env *env,
1806 struct lod_object *mo,
1807 const struct lu_buf *buf)
1809 struct lod_layout_component *lod_comp;
1810 struct lov_mds_md_v1 *v1 = buf->lb_buf;
1811 struct lov_mds_md_v3 *v3 = buf->lb_buf;
1812 struct lov_comp_md_v1 *comp_v1 = NULL;
1813 struct lov_ost_data_v1 *objs;
1820 magic = le32_to_cpu(v1->lmm_magic) & ~LOV_MAGIC_DEFINED;
1822 if (magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3 &&
1823 magic != LOV_MAGIC_COMP_V1)
1826 if (magic == LOV_MAGIC_COMP_V1) {
1827 comp_v1 = buf->lb_buf;
1828 comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count);
1831 mirror_cnt = le16_to_cpu(comp_v1->lcm_mirror_count) + 1;
1832 mo->ldo_flr_state = le16_to_cpu(comp_v1->lcm_flags) &
1834 mo->ldo_is_composite = 1;
1836 mo->ldo_is_composite = 0;
1840 mo->ldo_layout_gen = le16_to_cpu(v1->lmm_layout_gen);
1842 rc = lod_alloc_comp_entries(mo, mirror_cnt, comp_cnt);
1846 for (i = 0; i < comp_cnt; i++) {
1847 struct lu_extent *ext;
1851 lod_comp = &mo->ldo_comp_entries[i];
1853 if (mo->ldo_is_composite) {
1854 offs = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
1855 v1 = (struct lov_mds_md_v1 *)((char *)comp_v1 + offs);
1856 magic = le32_to_cpu(v1->lmm_magic);
1858 ext = &comp_v1->lcm_entries[i].lcme_extent;
1859 lod_comp->llc_extent.e_start =
1860 le64_to_cpu(ext->e_start);
1861 lod_comp->llc_extent.e_end = le64_to_cpu(ext->e_end);
1862 lod_comp->llc_flags =
1863 le32_to_cpu(comp_v1->lcm_entries[i].lcme_flags);
1865 le32_to_cpu(comp_v1->lcm_entries[i].lcme_id);
1866 if (lod_comp->llc_id == LCME_ID_INVAL)
1867 GOTO(out, rc = -EINVAL);
1871 if (magic == LOV_MAGIC_V1) {
1872 objs = &v1->lmm_objects[0];
1873 } else if (magic == LOV_MAGIC_V3) {
1874 objs = &v3->lmm_objects[0];
1875 if (v3->lmm_pool_name[0] != '\0')
1876 pool_name = v3->lmm_pool_name;
1878 CDEBUG(D_LAYOUT, "Invalid magic %x\n", magic);
1879 GOTO(out, rc = -EINVAL);
1882 lod_comp->llc_pattern = le32_to_cpu(v1->lmm_pattern);
1883 lod_comp->llc_stripe_size = le32_to_cpu(v1->lmm_stripe_size);
1884 lod_comp->llc_stripe_count = le16_to_cpu(v1->lmm_stripe_count);
1885 lod_comp->llc_layout_gen = le16_to_cpu(v1->lmm_layout_gen);
1887 * The stripe_offset of an uninit-ed component is stored in
1888 * the lmm_layout_gen
1890 if (mo->ldo_is_composite && !lod_comp_inited(lod_comp))
1891 lod_comp->llc_stripe_offset = lod_comp->llc_layout_gen;
1892 lod_obj_set_pool(mo, i, pool_name);
1894 if ((!mo->ldo_is_composite || lod_comp_inited(lod_comp)) &&
1895 !(lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED) &&
1896 !(lod_comp->llc_pattern & LOV_PATTERN_MDT)) {
1897 rc = lod_initialize_objects(env, mo, objs, i);
1903 rc = lod_fill_mirrors(mo);
1908 lod_object_free_striping(env, mo);
1914 * Parse suggested striping configuration.
1916 * The caller gets a suggested striping configuration from a number of sources
1917 * including per-directory default and applications. Then it needs to verify
1918 * the suggested striping is valid, apply missing bits and store the resulting
1919 * configuration in the object to be used by the allocator later. Must not be
1920 * called concurrently against the same object. It's OK to provide a
1921 * fully-defined striping.
1923 * \param[in] env execution environment for this thread
1924 * \param[in] lo LOD object
1925 * \param[in] buf buffer containing the striping
1927 * \retval 0 on success
1928 * \retval negative negated errno on error
1930 int lod_qos_parse_config(const struct lu_env *env, struct lod_object *lo,
1931 const struct lu_buf *buf)
1933 struct lod_layout_component *lod_comp;
1934 struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
1935 struct lov_desc *desc = &d->lod_desc;
1936 struct lov_user_md_v1 *v1 = NULL;
1937 struct lov_user_md_v3 *v3 = NULL;
1938 struct lov_comp_md_v1 *comp_v1 = NULL;
1939 char def_pool[LOV_MAXPOOLNAME + 1];
1946 if (buf == NULL || buf->lb_buf == NULL || buf->lb_len == 0)
1949 memset(def_pool, 0, sizeof(def_pool));
1950 if (lo->ldo_comp_entries != NULL)
1951 lod_layout_get_pool(lo->ldo_comp_entries, lo->ldo_comp_cnt,
1952 def_pool, sizeof(def_pool));
1954 /* free default striping info */
1955 lod_free_comp_entries(lo);
1957 rc = lod_verify_striping(d, lo, buf, false);
1963 comp_v1 = buf->lb_buf;
1964 magic = v1->lmm_magic;
1966 if (unlikely(le32_to_cpu(magic) & LOV_MAGIC_DEFINED)) {
1967 /* try to use as fully defined striping */
1968 rc = lod_use_defined_striping(env, lo, buf);
1973 case __swab32(LOV_USER_MAGIC_V1):
1974 lustre_swab_lov_user_md_v1(v1);
1975 magic = v1->lmm_magic;
1977 case LOV_USER_MAGIC_V1:
1979 case __swab32(LOV_USER_MAGIC_V3):
1980 lustre_swab_lov_user_md_v3(v3);
1981 magic = v3->lmm_magic;
1983 case LOV_USER_MAGIC_V3:
1985 case __swab32(LOV_USER_MAGIC_SPECIFIC):
1986 lustre_swab_lov_user_md_v3(v3);
1987 lustre_swab_lov_user_md_objects(v3->lmm_objects,
1988 v3->lmm_stripe_count);
1989 magic = v3->lmm_magic;
1991 case LOV_USER_MAGIC_SPECIFIC:
1993 case __swab32(LOV_USER_MAGIC_COMP_V1):
1994 lustre_swab_lov_comp_md_v1(comp_v1);
1995 magic = comp_v1->lcm_magic;
1997 case LOV_USER_MAGIC_COMP_V1:
2000 CERROR("%s: unrecognized magic %X\n",
2001 lod2obd(d)->obd_name, magic);
2005 lustre_print_user_md(D_OTHER, v1, "parse config");
2007 if (magic == LOV_USER_MAGIC_COMP_V1) {
2008 comp_cnt = comp_v1->lcm_entry_count;
2011 mirror_cnt = comp_v1->lcm_mirror_count + 1;
2013 lo->ldo_flr_state = LCM_FL_RDONLY;
2014 lo->ldo_is_composite = 1;
2018 lo->ldo_is_composite = 0;
2021 rc = lod_alloc_comp_entries(lo, mirror_cnt, comp_cnt);
2025 LASSERT(lo->ldo_comp_entries);
2027 for (i = 0; i < comp_cnt; i++) {
2028 struct pool_desc *pool;
2029 struct lu_extent *ext;
2032 lod_comp = &lo->ldo_comp_entries[i];
2034 if (lo->ldo_is_composite) {
2035 v1 = (struct lov_user_md *)((char *)comp_v1 +
2036 comp_v1->lcm_entries[i].lcme_offset);
2037 ext = &comp_v1->lcm_entries[i].lcme_extent;
2038 lod_comp->llc_extent = *ext;
2039 lod_comp->llc_flags =
2040 comp_v1->lcm_entries[i].lcme_flags &
2045 if (v1->lmm_magic == LOV_USER_MAGIC_V3 ||
2046 v1->lmm_magic == LOV_USER_MAGIC_SPECIFIC) {
2049 v3 = (struct lov_user_md_v3 *)v1;
2050 if (v3->lmm_pool_name[0] != '\0')
2051 pool_name = v3->lmm_pool_name;
2053 if (v3->lmm_magic == LOV_USER_MAGIC_SPECIFIC) {
2054 if (v3->lmm_stripe_offset == LOV_OFFSET_DEFAULT)
2055 v3->lmm_stripe_offset =
2056 v3->lmm_objects[0].l_ost_idx;
2058 /* copy ost list from lmm */
2059 lod_comp->llc_ostlist.op_count =
2060 v3->lmm_stripe_count;
2061 lod_comp->llc_ostlist.op_size =
2062 v3->lmm_stripe_count * sizeof(__u32);
2063 OBD_ALLOC(lod_comp->llc_ostlist.op_array,
2064 lod_comp->llc_ostlist.op_size);
2065 if (!lod_comp->llc_ostlist.op_array)
2066 GOTO(free_comp, rc = -ENOMEM);
2068 for (j = 0; j < v3->lmm_stripe_count; j++)
2069 lod_comp->llc_ostlist.op_array[j] =
2070 v3->lmm_objects[j].l_ost_idx;
2074 if (pool_name == NULL && def_pool[0] != '\0')
2075 pool_name = def_pool;
2077 if (v1->lmm_pattern == 0)
2078 v1->lmm_pattern = LOV_PATTERN_RAID0;
2079 if (lov_pattern(v1->lmm_pattern) != LOV_PATTERN_RAID0 &&
2080 lov_pattern(v1->lmm_pattern) != LOV_PATTERN_MDT) {
2081 CDEBUG(D_LAYOUT, "%s: invalid pattern: %x\n",
2082 lod2obd(d)->obd_name, v1->lmm_pattern);
2083 GOTO(free_comp, rc = -EINVAL);
2086 lod_comp->llc_pattern = v1->lmm_pattern;
2087 lod_comp->llc_stripe_size = desc->ld_default_stripe_size;
2088 if (v1->lmm_stripe_size)
2089 lod_comp->llc_stripe_size = v1->lmm_stripe_size;
2091 lod_comp->llc_stripe_count = desc->ld_default_stripe_count;
2092 if (v1->lmm_stripe_count ||
2093 lov_pattern(v1->lmm_pattern) == LOV_PATTERN_MDT)
2094 lod_comp->llc_stripe_count = v1->lmm_stripe_count;
2096 lod_comp->llc_stripe_offset = v1->lmm_stripe_offset;
2097 lod_obj_set_pool(lo, i, pool_name);
2099 LASSERT(ergo(lov_pattern(lod_comp->llc_pattern) ==
2100 LOV_PATTERN_MDT, lod_comp->llc_stripe_count == 0));
2102 if (pool_name == NULL)
2105 /* In the function below, .hs_keycmp resolves to
2106 * pool_hashkey_keycmp() */
2107 /* coverity[overrun-buffer-val] */
2108 pool = lod_find_pool(d, pool_name);
2112 if (lod_comp->llc_stripe_offset != LOV_OFFSET_DEFAULT) {
2113 rc = lod_check_index_in_pool(
2114 lod_comp->llc_stripe_offset, pool);
2116 lod_pool_putref(pool);
2117 CDEBUG(D_LAYOUT, "%s: invalid offset, %u\n",
2118 lod2obd(d)->obd_name,
2119 lod_comp->llc_stripe_offset);
2120 GOTO(free_comp, rc = -EINVAL);
2124 if (lod_comp->llc_stripe_count > pool_tgt_count(pool))
2125 lod_comp->llc_stripe_count = pool_tgt_count(pool);
2127 lod_pool_putref(pool);
2133 lod_free_comp_entries(lo);
2138 * prepare enough OST avoidance bitmap space
2140 int lod_prepare_avoidance(const struct lu_env *env, struct lod_object *lo)
2142 struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
2143 struct lod_tgt_descs *ltds = &lod->lod_ost_descs;
2144 struct lod_avoid_guide *lag = &lod_env_info(env)->lti_avoid;
2145 struct cfs_bitmap *bitmap = NULL;
2146 __u32 *new_oss = NULL;
2148 lag->lag_ost_avail = ltds->ltd_tgtnr;
2150 /* reset OSS avoid guide array */
2151 lag->lag_oaa_count = 0;
2152 if (lag->lag_oss_avoid_array && lag->lag_oaa_size < ltds->ltd_tgtnr) {
2153 OBD_FREE(lag->lag_oss_avoid_array,
2154 sizeof(__u32) * lag->lag_oaa_size);
2155 lag->lag_oss_avoid_array = NULL;
2156 lag->lag_oaa_size = 0;
2159 /* init OST avoid guide bitmap */
2160 if (lag->lag_ost_avoid_bitmap) {
2161 if (ltds->ltd_tgtnr <= lag->lag_ost_avoid_bitmap->size) {
2162 CFS_RESET_BITMAP(lag->lag_ost_avoid_bitmap);
2164 CFS_FREE_BITMAP(lag->lag_ost_avoid_bitmap);
2165 lag->lag_ost_avoid_bitmap = NULL;
2169 if (!lag->lag_ost_avoid_bitmap) {
2170 bitmap = CFS_ALLOCATE_BITMAP(ltds->ltd_tgtnr);
2175 if (!lag->lag_oss_avoid_array) {
2177 * usually there are multiple OSTs in one OSS, but we don't
2178 * know the exact OSS number, so we choose a safe option,
2179 * using OST count to allocate the array to store the OSS
2182 OBD_ALLOC(new_oss, sizeof(*new_oss) * ltds->ltd_tgtnr);
2184 CFS_FREE_BITMAP(bitmap);
2190 lag->lag_oss_avoid_array = new_oss;
2191 lag->lag_oaa_size = ltds->ltd_tgtnr;
2194 lag->lag_ost_avoid_bitmap = bitmap;
2200 * Collect information of used OSTs and OSSs in the overlapped components
2203 void lod_collect_avoidance(struct lod_object *lo, struct lod_avoid_guide *lag,
2206 struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
2207 struct lod_layout_component *lod_comp = &lo->ldo_comp_entries[comp_idx];
2208 struct cfs_bitmap *bitmap = lag->lag_ost_avoid_bitmap;
2211 /* iterate mirrors */
2212 for (i = 0; i < lo->ldo_mirror_count; i++) {
2213 struct lod_layout_component *comp;
2216 * skip mirror containing component[comp_idx], we only
2217 * collect OSTs info of conflicting component in other mirrors,
2218 * so that during read, if OSTs of a mirror's component are
2219 * not available, we still have other mirror with different
2220 * OSTs to read the data.
2222 comp = &lo->ldo_comp_entries[lo->ldo_mirrors[i].lme_start];
2223 if (comp->llc_id == LCME_ID_INVAL ||
2224 mirror_id_of(comp->llc_id) ==
2225 mirror_id_of(lod_comp->llc_id))
2228 /* iterate components of a mirror */
2229 lod_foreach_mirror_comp(comp, lo, i) {
2230 /* skip non-overlapped or un-instantiated components */
2231 if (!lu_extent_is_overlapped(&comp->llc_extent,
2232 &lod_comp->llc_extent) ||
2233 !lod_comp_inited(comp) || !comp->llc_stripe)
2237 * collect used OSTs index and OSS info from a
2240 for (j = 0; j < comp->llc_stripe_count; j++) {
2241 struct lod_tgt_desc *ost;
2242 struct lod_qos_oss *lqo;
2245 ost = OST_TGT(lod, comp->llc_ost_indices[j]);
2246 lqo = ost->ltd_qos.ltq_oss;
2248 if (cfs_bitmap_check(bitmap, ost->ltd_index))
2251 cfs_bitmap_set(bitmap, ost->ltd_index);
2252 lag->lag_ost_avail--;
2254 for (k = 0; k < lag->lag_oaa_count; k++) {
2255 if (lag->lag_oss_avoid_array[k] ==
2259 if (k == lag->lag_oaa_count) {
2260 lag->lag_oss_avoid_array[k] =
2262 lag->lag_oaa_count++;
2270 * Create a striping for an obejct.
2272 * The function creates a new striping for the object. The function tries QoS
2273 * algorithm first unless free space is distributed evenly among OSTs, but
2274 * by default RR algorithm is preferred due to internal concurrency (QoS is
2275 * serialized). The caller must ensure no concurrent calls to the function
2276 * are made against the same object.
2278 * \param[in] env execution environment for this thread
2279 * \param[in] lo LOD object
2280 * \param[in] attr attributes OST objects will be declared with
2281 * \param[in] th transaction handle
2282 * \param[in] comp_idx index of ldo_comp_entries
2283 * \param[in|out] inuse array of inuse ost index
2285 * \retval 0 on success
2286 * \retval negative negated errno on error
2288 int lod_qos_prep_create(const struct lu_env *env, struct lod_object *lo,
2289 struct lu_attr *attr, struct thandle *th,
2290 int comp_idx, struct ost_pool *inuse)
2292 struct lod_layout_component *lod_comp;
2293 struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
2295 int flag = LOV_USES_ASSIGNED_STRIPE;
2297 struct lod_avoid_guide *lag = &lod_env_info(env)->lti_avoid;
2298 struct dt_object **stripe = NULL;
2299 __u32 *ost_indices = NULL;
2303 LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
2304 lod_comp = &lo->ldo_comp_entries[comp_idx];
2306 /* A released component is being created */
2307 if (lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED)
2310 /* A Data-on-MDT component is being created */
2311 if (lov_pattern(lod_comp->llc_pattern) == LOV_PATTERN_MDT)
2314 if (likely(lod_comp->llc_stripe == NULL)) {
2316 * no striping has been created so far
2318 LASSERT(lod_comp->llc_stripe_count);
2320 * statfs and check OST targets now, since ld_active_tgt_count
2321 * could be changed if some OSTs are [de]activated manually.
2323 lod_qos_statfs_update(env, d);
2324 stripe_len = lod_get_stripe_count(d, lo,
2325 lod_comp->llc_stripe_count);
2326 if (stripe_len == 0)
2327 GOTO(out, rc = -ERANGE);
2328 lod_comp->llc_stripe_count = stripe_len;
2329 OBD_ALLOC(stripe, sizeof(stripe[0]) * stripe_len);
2331 GOTO(out, rc = -ENOMEM);
2332 OBD_ALLOC(ost_indices, sizeof(*ost_indices) * stripe_len);
2334 GOTO(out, rc = -ENOMEM);
2336 lod_getref(&d->lod_ost_descs);
2337 /* XXX: support for non-0 files w/o objects */
2338 CDEBUG(D_OTHER, "tgt_count %d stripe_count %d\n",
2339 d->lod_desc.ld_tgt_count, stripe_len);
2341 if (lod_comp->llc_ostlist.op_array) {
2342 rc = lod_alloc_ost_list(env, lo, stripe, ost_indices,
2343 th, comp_idx, inuse);
2344 } else if (lod_comp->llc_stripe_offset == LOV_OFFSET_DEFAULT) {
2346 * collect OSTs and OSSs used in other mirrors whose
2347 * components cross the ldo_comp_entries[comp_idx]
2349 rc = lod_prepare_avoidance(env, lo);
2353 lod_collect_avoidance(lo, lag, comp_idx);
2355 rc = lod_alloc_qos(env, lo, stripe, ost_indices, flag,
2356 th, comp_idx, inuse);
2358 rc = lod_alloc_rr(env, lo, stripe, ost_indices,
2359 flag, th, comp_idx, inuse);
2361 rc = lod_alloc_specific(env, lo, stripe, ost_indices,
2362 flag, th, comp_idx, inuse);
2365 lod_putref(d, &d->lod_ost_descs);
2367 for (i = 0; i < stripe_len; i++)
2368 if (stripe[i] != NULL)
2369 dt_object_put(env, stripe[i]);
2370 lod_comp->llc_stripe_count = 0;
2372 lod_comp->llc_stripe = stripe;
2373 lod_comp->llc_ost_indices = ost_indices;
2374 lod_comp->llc_stripes_allocated = stripe_len;
2378 * lod_qos_parse_config() found supplied buf as a predefined
2379 * striping (not a hint), so it allocated all the object
2380 * now we need to create them
2382 for (i = 0; i < lod_comp->llc_stripe_count; i++) {
2383 struct dt_object *o;
2385 o = lod_comp->llc_stripe[i];
2388 rc = lod_sub_declare_create(env, o, attr, NULL,
2391 CERROR("can't declare create: %d\n", rc);
2396 * Clear LCME_FL_INIT for the component so that
2397 * lod_striping_create() can create the striping objects
2400 lod_comp_unset_init(lod_comp);
2406 OBD_FREE(stripe, sizeof(stripe[0]) * stripe_len);
2408 OBD_FREE(ost_indices,
2409 sizeof(*ost_indices) * stripe_len);
2414 int lod_obj_stripe_set_inuse_cb(const struct lu_env *env,
2415 struct lod_object *lo,
2416 struct dt_object *dt, struct thandle *th,
2417 int comp_idx, int stripe_idx,
2418 struct lod_obj_stripe_cb_data *data)
2420 struct lod_thread_info *info = lod_env_info(env);
2421 struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
2422 struct lu_fid *fid = &info->lti_fid;
2424 int rc, type = LU_SEQ_RANGE_OST;
2426 *fid = *lu_object_fid(&dt->do_lu);
2427 rc = lod_fld_lookup(env, d, fid, &index, &type);
2429 CERROR("%s: fail to locate "DFID": rc = %d\n",
2430 lod2obd(d)->obd_name, PFID(fid), rc);
2433 lod_comp_ost_in_use(data->locd_inuse, index);
2438 * Resize per-thread ost list to hold OST target index list already used.
2440 * \param[in,out] inuse structure contains ost list array
2441 * \param[in] cnt total stripe count of all components
2442 * \param[in] max array's max size if @max > 0
2444 * \retval 0 on success
2445 * \retval -ENOMEM reallocation failed
2447 static int lod_inuse_resize(struct ost_pool *inuse, __u16 cnt, __u16 max)
2450 __u32 new = cnt * sizeof(inuse->op_array[0]);
2452 inuse->op_count = 0;
2454 if (new <= inuse->op_size)
2458 new = min_t(__u32, new, max);
2460 OBD_ALLOC(array, new);
2464 if (inuse->op_array)
2465 OBD_FREE(inuse->op_array, inuse->op_size);
2467 inuse->op_array = array;
2468 inuse->op_size = new;
2473 int lod_prepare_inuse(const struct lu_env *env, struct lod_object *lo)
2475 struct lod_thread_info *info = lod_env_info(env);
2476 struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
2477 struct ost_pool *inuse = &info->lti_inuse_osts;
2478 struct lod_obj_stripe_cb_data data = { { 0 } };
2479 __u32 stripe_count = 0;
2483 for (i = 0; i < lo->ldo_comp_cnt; i++)
2484 stripe_count += lod_comp_entry_stripe_count(lo,
2485 &lo->ldo_comp_entries[i], false);
2486 rc = lod_inuse_resize(inuse, stripe_count, d->lod_osd_max_easize);
2490 data.locd_inuse = inuse;
2491 data.locd_stripe_cb = lod_obj_stripe_set_inuse_cb;
2492 return lod_obj_for_each_stripe(env, lo, NULL, &data);
2495 int lod_prepare_create(const struct lu_env *env, struct lod_object *lo,
2496 struct lu_attr *attr, const struct lu_buf *buf,
2500 struct lod_thread_info *info = lod_env_info(env);
2501 struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
2502 struct ost_pool *inuse = &info->lti_inuse_osts;
2510 /* no OST available */
2511 /* XXX: should we be waiting a bit to prevent failures during
2512 * cluster initialization? */
2513 if (d->lod_ostnr == 0)
2517 * by this time, the object's ldo_stripe_count and ldo_stripe_size
2518 * contain default value for striping: taken from the parent
2519 * or from filesystem defaults
2521 * in case the caller is passing lovea with new striping config,
2522 * we may need to parse lovea and apply new configuration
2524 rc = lod_qos_parse_config(env, lo, buf);
2528 if (attr->la_valid & LA_SIZE)
2529 size = attr->la_size;
2532 rc = lod_prepare_inuse(env, lo);
2537 * prepare OST object creation for the component covering file's
2538 * size, the 1st component (including plain layout file) is always
2541 for (i = 0; i < lo->ldo_comp_cnt; i++) {
2542 struct lod_layout_component *lod_comp;
2543 struct lu_extent *extent;
2545 lod_comp = &lo->ldo_comp_entries[i];
2546 extent = &lod_comp->llc_extent;
2547 CDEBUG(D_QOS, "%lld [%lld, %lld)\n",
2548 size, extent->e_start, extent->e_end);
2549 if (!lo->ldo_is_composite || size >= extent->e_start) {
2550 rc = lod_qos_prep_create(env, lo, attr, th, i, inuse);