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, 2015, 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 <obd_class.h>
43 #include <lustre/lustre_idl.h>
44 #include "lod_internal.h"
47 * force QoS policy (not RR) to be used for testing purposes
53 #define QOS_DEBUG(fmt, ...) CDEBUG(D_QOS, fmt, ## __VA_ARGS__)
54 #define QOS_CONSOLE(fmt, ...) LCONSOLE(D_QOS, fmt, ## __VA_ARGS__)
56 #define TGT_BAVAIL(i) (OST_TGT(lod,i)->ltd_statfs.os_bavail * \
57 OST_TGT(lod,i)->ltd_statfs.os_bsize)
60 * Add a new target to Quality of Service (QoS) target table.
62 * Add a new OST target to the structure representing an OSS. Resort the list
63 * of known OSSs by the number of OSTs attached to each OSS. The OSS list is
64 * protected internally and no external locking is required.
66 * \param[in] lod LOD device
67 * \param[in] ost_desc OST description
69 * \retval 0 on success
70 * \retval -ENOMEM on error
72 int qos_add_tgt(struct lod_device *lod, struct lod_tgt_desc *ost_desc)
74 struct lod_qos_oss *oss = NULL, *temposs;
75 struct obd_export *exp = ost_desc->ltd_exp;
76 int rc = 0, found = 0;
77 struct list_head *list;
80 down_write(&lod->lod_qos.lq_rw_sem);
82 * a bit hacky approach to learn NID of corresponding connection
83 * but there is no official API to access information like this
86 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
87 if (obd_uuid_equals(&oss->lqo_uuid,
88 &exp->exp_connection->c_remote_uuid)) {
97 GOTO(out, rc = -ENOMEM);
98 memcpy(&oss->lqo_uuid, &exp->exp_connection->c_remote_uuid,
99 sizeof(oss->lqo_uuid));
101 /* Assume we have to move this one */
102 list_del(&oss->lqo_oss_list);
105 oss->lqo_ost_count++;
106 ost_desc->ltd_qos.ltq_oss = oss;
108 CDEBUG(D_QOS, "add tgt %s to OSS %s (%d OSTs)\n",
109 obd_uuid2str(&ost_desc->ltd_uuid), obd_uuid2str(&oss->lqo_uuid),
112 /* Add sorted by # of OSTs. Find the first entry that we're
114 list = &lod->lod_qos.lq_oss_list;
115 list_for_each_entry(temposs, list, lqo_oss_list) {
116 if (oss->lqo_ost_count > temposs->lqo_ost_count)
119 /* ...and add before it. If we're the first or smallest, temposs
120 points to the list head, and we add to the end. */
121 list_add_tail(&oss->lqo_oss_list, &temposs->lqo_oss_list);
123 lod->lod_qos.lq_dirty = 1;
124 lod->lod_qos.lq_rr.lqr_dirty = 1;
127 up_write(&lod->lod_qos.lq_rw_sem);
132 * Remove OST target from QoS table.
134 * Removes given OST target from QoS table and releases related OSS structure
135 * if no OSTs remain on the OSS.
137 * \param[in] lod LOD device
138 * \param[in] ost_desc OST description
140 * \retval 0 on success
141 * \retval -ENOENT if no OSS was found
143 int qos_del_tgt(struct lod_device *lod, struct lod_tgt_desc *ost_desc)
145 struct lod_qos_oss *oss;
149 down_write(&lod->lod_qos.lq_rw_sem);
150 oss = ost_desc->ltd_qos.ltq_oss;
152 GOTO(out, rc = -ENOENT);
154 oss->lqo_ost_count--;
155 if (oss->lqo_ost_count == 0) {
156 CDEBUG(D_QOS, "removing OSS %s\n",
157 obd_uuid2str(&oss->lqo_uuid));
158 list_del(&oss->lqo_oss_list);
159 ost_desc->ltd_qos.ltq_oss = NULL;
163 lod->lod_qos.lq_dirty = 1;
164 lod->lod_qos.lq_rr.lqr_dirty = 1;
166 up_write(&lod->lod_qos.lq_rw_sem);
171 * Check whether the target is available for new OST objects.
173 * Request statfs data from the given target and verify it's active and not
174 * read-only. If so, then it can be used to place new OST objects. This
175 * function also maintains the number of active/inactive targets and sets
176 * dirty flags if those numbers change so others can run re-balance procedures.
177 * No external locking is required.
179 * \param[in] env execution environment for this thread
180 * \param[in] d LOD device
181 * \param[in] index index of OST target to check
182 * \param[out] sfs buffer for statfs data
184 * \retval 0 if the target is good
185 * \retval negative negated errno on error
188 static int lod_statfs_and_check(const struct lu_env *env, struct lod_device *d,
189 int index, struct obd_statfs *sfs)
191 struct lod_tgt_desc *ost;
196 ost = OST_TGT(d,index);
199 rc = dt_statfs(env, ost->ltd_ost, sfs);
204 if (rc && rc != -ENOTCONN)
205 CERROR("%s: statfs: rc = %d\n", lod2obd(d)->obd_name, rc);
207 /* If the OST is readonly then we can't allocate objects there */
208 if (sfs->os_state & OS_STATE_READONLY)
211 /* check whether device has changed state (active, inactive) */
212 if (rc != 0 && ost->ltd_active) {
213 /* turned inactive? */
214 spin_lock(&d->lod_desc_lock);
215 if (ost->ltd_active) {
217 LASSERT(d->lod_desc.ld_active_tgt_count > 0);
218 d->lod_desc.ld_active_tgt_count--;
219 d->lod_qos.lq_dirty = 1;
220 d->lod_qos.lq_rr.lqr_dirty = 1;
221 CDEBUG(D_CONFIG, "%s: turns inactive\n",
222 ost->ltd_exp->exp_obd->obd_name);
224 spin_unlock(&d->lod_desc_lock);
225 } else if (rc == 0 && ost->ltd_active == 0) {
227 LASSERTF(d->lod_desc.ld_active_tgt_count < d->lod_ostnr,
228 "active tgt count %d, ost nr %d\n",
229 d->lod_desc.ld_active_tgt_count, d->lod_ostnr);
230 spin_lock(&d->lod_desc_lock);
231 if (ost->ltd_active == 0) {
233 d->lod_desc.ld_active_tgt_count++;
234 d->lod_qos.lq_dirty = 1;
235 d->lod_qos.lq_rr.lqr_dirty = 1;
236 CDEBUG(D_CONFIG, "%s: turns active\n",
237 ost->ltd_exp->exp_obd->obd_name);
239 spin_unlock(&d->lod_desc_lock);
246 * Maintain per-target statfs data.
248 * The function refreshes statfs data for all the targets every N seconds.
249 * The actual N is controlled via procfs and set to LOV_DESC_QOS_MAXAGE_DEFAULT
252 * \param[in] env execution environment for this thread
253 * \param[in] lod LOD device
255 static void lod_qos_statfs_update(const struct lu_env *env,
256 struct lod_device *lod)
258 struct obd_device *obd = lod2obd(lod);
259 struct ost_pool *osts = &(lod->lod_pool_info);
262 __u64 max_age, avail;
265 max_age = cfs_time_shift_64(-2 * lod->lod_desc.ld_qos_maxage);
267 if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
268 /* statfs data are quite recent, don't need to refresh it */
271 down_write(&lod->lod_qos.lq_rw_sem);
272 if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
275 for (i = 0; i < osts->op_count; i++) {
276 idx = osts->op_array[i];
277 avail = OST_TGT(lod,idx)->ltd_statfs.os_bavail;
278 if (lod_statfs_and_check(env, lod, idx,
279 &OST_TGT(lod, idx)->ltd_statfs))
281 if (OST_TGT(lod,idx)->ltd_statfs.os_bavail != avail)
282 /* recalculate weigths */
283 lod->lod_qos.lq_dirty = 1;
285 obd->obd_osfs_age = cfs_time_current_64();
288 up_write(&lod->lod_qos.lq_rw_sem);
293 * Calculate per-OST and per-OSS penalties
295 * Re-calculate penalties when the configuration changes, active targets
296 * change and after statfs refresh (all these are reflected by lq_dirty flag).
297 * On every OST and OSS: decay the penalty by half for every 8x the update
298 * interval that the device has been idle. That gives lots of time for the
299 * statfs information to be updated (which the penalty is only a proxy for),
300 * and avoids penalizing OSS/OSTs under light load.
301 * See lod_qos_calc_weight() for how penalties are factored into the weight.
303 * \param[in] lod LOD device
305 * \retval 0 on success
306 * \retval -EAGAIN the number of OSTs isn't enough
308 static int lod_qos_calc_ppo(struct lod_device *lod)
310 struct lod_qos_oss *oss;
311 __u64 ba_max, ba_min, temp;
318 if (!lod->lod_qos.lq_dirty)
321 num_active = lod->lod_desc.ld_active_tgt_count - 1;
323 GOTO(out, rc = -EAGAIN);
325 /* find bavail on each OSS */
326 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list)
328 lod->lod_qos.lq_active_oss_count = 0;
331 * How badly user wants to select OSTs "widely" (not recently chosen
332 * and not on recent OSS's). As opposed to "freely" (free space
335 prio_wide = 256 - lod->lod_qos.lq_prio_free;
337 ba_min = (__u64)(-1);
339 now = cfs_time_current_sec();
340 /* Calculate OST penalty per object
341 * (lod ref taken in lod_qos_prep_create()) */
342 cfs_foreach_bit(lod->lod_ost_bitmap, i) {
343 LASSERT(OST_TGT(lod,i));
344 temp = TGT_BAVAIL(i);
347 ba_min = min(temp, ba_min);
348 ba_max = max(temp, ba_max);
350 /* Count the number of usable OSS's */
351 if (OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail == 0)
352 lod->lod_qos.lq_active_oss_count++;
353 OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail += temp;
355 /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
357 do_div(temp, num_active);
358 OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj =
359 (temp * prio_wide) >> 8;
361 age = (now - OST_TGT(lod,i)->ltd_qos.ltq_used) >> 3;
362 if (lod->lod_qos.lq_reset ||
363 age > 32 * lod->lod_desc.ld_qos_maxage)
364 OST_TGT(lod,i)->ltd_qos.ltq_penalty = 0;
365 else if (age > lod->lod_desc.ld_qos_maxage)
366 /* Decay OST penalty. */
367 OST_TGT(lod,i)->ltd_qos.ltq_penalty >>=
368 (age / lod->lod_desc.ld_qos_maxage);
371 num_active = lod->lod_qos.lq_active_oss_count - 1;
372 if (num_active < 1) {
373 /* If there's only 1 OSS, we can't penalize it, so instead
374 we have to double the OST penalty */
376 cfs_foreach_bit(lod->lod_ost_bitmap, i)
377 OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj <<= 1;
380 /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
381 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
382 temp = oss->lqo_bavail >> 1;
383 do_div(temp, oss->lqo_ost_count * num_active);
384 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
386 age = (now - oss->lqo_used) >> 3;
387 if (lod->lod_qos.lq_reset ||
388 age > 32 * lod->lod_desc.ld_qos_maxage)
389 oss->lqo_penalty = 0;
390 else if (age > lod->lod_desc.ld_qos_maxage)
391 /* Decay OSS penalty. */
392 oss->lqo_penalty >>= age / lod->lod_desc.ld_qos_maxage;
395 lod->lod_qos.lq_dirty = 0;
396 lod->lod_qos.lq_reset = 0;
398 /* If each ost has almost same free space,
399 * do rr allocation for better creation performance */
400 lod->lod_qos.lq_same_space = 0;
401 if ((ba_max * (256 - lod->lod_qos.lq_threshold_rr)) >> 8 < ba_min) {
402 lod->lod_qos.lq_same_space = 1;
403 /* Reset weights for the next time we enter qos mode */
404 lod->lod_qos.lq_reset = 1;
410 if (!rc && lod->lod_qos.lq_same_space)
417 * Calculate weight for a given OST target.
419 * The final OST weight is the number of bytes available minus the OST and
420 * OSS penalties. See lod_qos_calc_ppo() for how penalties are calculated.
422 * \param[in] lod LOD device, where OST targets are listed
423 * \param[in] i OST target index
427 static int lod_qos_calc_weight(struct lod_device *lod, int i)
431 temp = TGT_BAVAIL(i);
432 temp2 = OST_TGT(lod,i)->ltd_qos.ltq_penalty +
433 OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_penalty;
435 OST_TGT(lod,i)->ltd_qos.ltq_weight = 0;
437 OST_TGT(lod,i)->ltd_qos.ltq_weight = temp - temp2;
442 * Re-calculate weights.
444 * The function is called when some OST target was used for a new object. In
445 * this case we should re-calculate all the weights to keep new allocations
448 * \param[in] lod LOD device
449 * \param[in] osts OST pool where a new object was placed
450 * \param[in] index OST target where a new object was placed
451 * \param[out] total_wt new total weight for the pool
455 static int lod_qos_used(struct lod_device *lod, struct ost_pool *osts,
456 __u32 index, __u64 *total_wt)
458 struct lod_tgt_desc *ost;
459 struct lod_qos_oss *oss;
463 ost = OST_TGT(lod,index);
466 /* Don't allocate on this devuce anymore, until the next alloc_qos */
467 ost->ltd_qos.ltq_usable = 0;
469 oss = ost->ltd_qos.ltq_oss;
471 /* Decay old penalty by half (we're adding max penalty, and don't
472 want it to run away.) */
473 ost->ltd_qos.ltq_penalty >>= 1;
474 oss->lqo_penalty >>= 1;
476 /* mark the OSS and OST as recently used */
477 ost->ltd_qos.ltq_used = oss->lqo_used = cfs_time_current_sec();
479 /* Set max penalties for this OST and OSS */
480 ost->ltd_qos.ltq_penalty +=
481 ost->ltd_qos.ltq_penalty_per_obj * lod->lod_ostnr;
482 oss->lqo_penalty += oss->lqo_penalty_per_obj *
483 lod->lod_qos.lq_active_oss_count;
485 /* Decrease all OSS penalties */
486 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
487 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
488 oss->lqo_penalty = 0;
490 oss->lqo_penalty -= oss->lqo_penalty_per_obj;
494 /* Decrease all OST penalties */
495 for (j = 0; j < osts->op_count; j++) {
498 i = osts->op_array[j];
499 if (!cfs_bitmap_check(lod->lod_ost_bitmap, i))
502 ost = OST_TGT(lod,i);
505 if (ost->ltd_qos.ltq_penalty <
506 ost->ltd_qos.ltq_penalty_per_obj)
507 ost->ltd_qos.ltq_penalty = 0;
509 ost->ltd_qos.ltq_penalty -=
510 ost->ltd_qos.ltq_penalty_per_obj;
512 lod_qos_calc_weight(lod, i);
514 /* Recalc the total weight of usable osts */
515 if (ost->ltd_qos.ltq_usable)
516 *total_wt += ost->ltd_qos.ltq_weight;
518 QOS_DEBUG("recalc tgt %d usable=%d avail="LPU64
519 " ostppo="LPU64" ostp="LPU64" ossppo="LPU64
520 " ossp="LPU64" wt="LPU64"\n",
521 i, ost->ltd_qos.ltq_usable, TGT_BAVAIL(i) >> 10,
522 ost->ltd_qos.ltq_penalty_per_obj >> 10,
523 ost->ltd_qos.ltq_penalty >> 10,
524 ost->ltd_qos.ltq_oss->lqo_penalty_per_obj >> 10,
525 ost->ltd_qos.ltq_oss->lqo_penalty >> 10,
526 ost->ltd_qos.ltq_weight >> 10);
532 void lod_qos_rr_init(struct lod_qos_rr *lqr)
534 spin_lock_init(&lqr->lqr_alloc);
539 #define LOV_QOS_EMPTY ((__u32)-1)
542 * Calculate optimal round-robin order with regard to OSSes.
544 * Place all the OSTs from pool \a src_pool in a special array to be used for
545 * round-robin (RR) stripe allocation. The placement algorithm interleaves
546 * OSTs from the different OSSs so that RR allocation can balance OSSs evenly.
547 * Resorts the targets when the number of active targets changes (because of
548 * a new target or activation/deactivation).
550 * \param[in] lod LOD device
551 * \param[in] src_pool OST pool
552 * \param[in] lqr round-robin list
554 * \retval 0 on success
555 * \retval -ENOMEM fails to allocate the array
557 static int lod_qos_calc_rr(struct lod_device *lod, struct ost_pool *src_pool,
558 struct lod_qos_rr *lqr)
560 struct lod_qos_oss *oss;
561 struct lod_tgt_desc *ost;
562 unsigned placed, real_count;
567 if (!lqr->lqr_dirty) {
568 LASSERT(lqr->lqr_pool.op_size);
572 /* Do actual allocation. */
573 down_write(&lod->lod_qos.lq_rw_sem);
576 * Check again. While we were sleeping on @lq_rw_sem something could
579 if (!lqr->lqr_dirty) {
580 LASSERT(lqr->lqr_pool.op_size);
581 up_write(&lod->lod_qos.lq_rw_sem);
585 real_count = src_pool->op_count;
587 /* Zero the pool array */
588 /* alloc_rr is holding a read lock on the pool, so nobody is adding/
589 deleting from the pool. The lq_rw_sem insures that nobody else
591 lqr->lqr_pool.op_count = real_count;
592 rc = lod_ost_pool_extend(&lqr->lqr_pool, real_count);
594 up_write(&lod->lod_qos.lq_rw_sem);
597 for (i = 0; i < lqr->lqr_pool.op_count; i++)
598 lqr->lqr_pool.op_array[i] = LOV_QOS_EMPTY;
600 /* Place all the OSTs from 1 OSS at the same time. */
602 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
605 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
608 if (!cfs_bitmap_check(lod->lod_ost_bitmap,
609 src_pool->op_array[i]))
612 ost = OST_TGT(lod,src_pool->op_array[i]);
613 LASSERT(ost && ost->ltd_ost);
614 if (ost->ltd_qos.ltq_oss != oss)
617 /* Evenly space these OSTs across arrayspace */
618 next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count;
619 while (lqr->lqr_pool.op_array[next] != LOV_QOS_EMPTY)
620 next = (next + 1) % lqr->lqr_pool.op_count;
622 lqr->lqr_pool.op_array[next] = src_pool->op_array[i];
629 up_write(&lod->lod_qos.lq_rw_sem);
631 if (placed != real_count) {
632 /* This should never happen */
633 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
634 "round-robin list (%d of %d).\n",
636 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
637 LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
638 lqr->lqr_pool.op_array[i]);
645 for (i = 0; i < lqr->lqr_pool.op_count; i++)
646 QOS_CONSOLE("rr #%d ost idx=%d\n", i, lqr->lqr_pool.op_array[i]);
653 * Instantiate and declare creation of a new object.
655 * The function instantiates LU representation for a new object on the
656 * specified device. Also it declares an intention to create that
657 * object on the storage target.
659 * Note lu_object_anon() is used which is a trick with regard to LU/OSD
660 * infrastructure - in the existing precreation framework we can't assign FID
661 * at this moment, we do this later once a transaction is started. So the
662 * special method instantiates FID-less object in the cache and later it
663 * will get a FID and proper placement in LU cache.
665 * \param[in] env execution environment for this thread
666 * \param[in] d LOD device
667 * \param[in] ost_idx OST target index where the object is being created
668 * \param[in] th transaction handle
670 * \retval object ptr on success, ERR_PTR() otherwise
672 static struct dt_object *lod_qos_declare_object_on(const struct lu_env *env,
673 struct lod_device *d,
677 struct lod_tgt_desc *ost;
678 struct lu_object *o, *n;
679 struct lu_device *nd;
680 struct dt_object *dt;
685 LASSERT(ost_idx < d->lod_osts_size);
686 ost = OST_TGT(d,ost_idx);
688 LASSERT(ost->ltd_ost);
690 nd = &ost->ltd_ost->dd_lu_dev;
693 * allocate anonymous object with zero fid, real fid
694 * will be assigned by OSP within transaction
695 * XXX: to be fixed with fully-functional OST fids
697 o = lu_object_anon(env, nd, NULL);
699 GOTO(out, dt = ERR_PTR(PTR_ERR(o)));
701 n = lu_object_locate(o->lo_header, nd->ld_type);
702 if (unlikely(n == NULL)) {
703 CERROR("can't find slice\n");
704 lu_object_put(env, o);
705 GOTO(out, dt = ERR_PTR(-EINVAL));
708 dt = container_of(n, struct dt_object, do_lu);
710 rc = lod_sub_object_declare_create(env, dt, NULL, NULL, NULL, th);
712 CDEBUG(D_OTHER, "can't declare creation on #%u: %d\n",
714 lu_object_put(env, o);
723 * Calculate a minimum acceptable stripe count.
725 * Return an acceptable stripe count depending on flag LOV_USES_DEFAULT_STRIPE:
726 * all stripes or 3/4 of stripes.
728 * \param[in] stripe_cnt number of stripes requested
729 * \param[in] flags 0 or LOV_USES_DEFAULT_STRIPE
731 * \retval acceptable stripecount
733 static int min_stripe_count(__u32 stripe_cnt, int flags)
735 return (flags & LOV_USES_DEFAULT_STRIPE ?
736 stripe_cnt - (stripe_cnt / 4) : stripe_cnt);
739 #define LOV_CREATE_RESEED_MULT 30
740 #define LOV_CREATE_RESEED_MIN 2000
743 * Initialize temporary OST-in-use array.
745 * Allocate or extend the array used to mark targets already assigned to a new
746 * striping so they are not used more than once.
748 * \param[in] env execution environment for this thread
749 * \param[in] stripes number of items needed in the array
751 * \retval 0 on success
752 * \retval -ENOMEM on error
754 static inline int lod_qos_ost_in_use_clear(const struct lu_env *env,
757 struct lod_thread_info *info = lod_env_info(env);
759 if (info->lti_ea_store_size < sizeof(int) * stripes)
760 lod_ea_store_resize(info, stripes * sizeof(int));
761 if (info->lti_ea_store_size < sizeof(int) * stripes) {
762 CERROR("can't allocate memory for ost-in-use array\n");
765 memset(info->lti_ea_store, -1, sizeof(int) * stripes);
770 * Remember a target in the array of used targets.
772 * Mark the given target as used for a new striping being created. The status
773 * of an OST in a striping can be checked with lod_qos_is_ost_used().
775 * \param[in] env execution environment for this thread
776 * \param[in] idx index in the array
777 * \param[in] ost OST target index to mark as used
779 static inline void lod_qos_ost_in_use(const struct lu_env *env,
782 struct lod_thread_info *info = lod_env_info(env);
783 int *osts = info->lti_ea_store;
785 LASSERT(info->lti_ea_store_size >= idx * sizeof(int));
790 * Check is OST used in a striping.
792 * Checks whether OST with the given index is marked as used in the temporary
793 * array (see lod_qos_ost_in_use()).
795 * \param[in] env execution environment for this thread
796 * \param[in] ost OST target index to check
797 * \param[in] stripes the number of items used in the array already
802 static int lod_qos_is_ost_used(const struct lu_env *env, int ost, __u32 stripes)
804 struct lod_thread_info *info = lod_env_info(env);
805 int *osts = info->lti_ea_store;
808 for (j = 0; j < stripes; j++) {
815 static int lod_check_and_reserve_ost(const struct lu_env *env,
816 struct lod_device *m,
817 struct obd_statfs *sfs, __u32 ost_idx,
818 __u32 speed, __u32 *s_idx,
819 struct dt_object **stripe,
823 __u32 stripe_idx = *s_idx;
826 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
828 /* this OSP doesn't feel well */
833 * We expect number of precreated objects in f_ffree at
834 * the first iteration, skip OSPs with no objects ready
836 if (sfs->os_fprecreated == 0 && speed == 0) {
837 QOS_DEBUG("#%d: precreation is empty\n", ost_idx);
842 * try to use another OSP if this one is degraded
844 if (sfs->os_state & OS_STATE_DEGRADED && speed < 2) {
845 QOS_DEBUG("#%d: degraded\n", ost_idx);
850 * do not put >1 objects on a single OST
852 if (speed && lod_qos_is_ost_used(env, ost_idx, stripe_idx))
855 o = lod_qos_declare_object_on(env, m, ost_idx, th);
857 CDEBUG(D_OTHER, "can't declare new object on #%u: %d\n",
858 ost_idx, (int) PTR_ERR(o));
864 * We've successfully declared (reserved) an object
866 lod_qos_ost_in_use(env, stripe_idx, ost_idx);
867 stripe[stripe_idx] = o;
876 * Allocate a striping using round-robin algorithm.
878 * Allocates a new striping using round-robin algorithm. The function refreshes
879 * all the internal structures (statfs cache, array of available OSTs sorted
880 * with regard to OSS, etc). The number of stripes required is taken from the
881 * object (must be prepared by the caller), but can change if the flag
882 * LOV_USES_DEFAULT_STRIPE is supplied. The caller should ensure nobody else
883 * is trying to create a striping on the object in parallel. All the internal
884 * structures (like pools, etc) are protected and no additional locking is
885 * required. The function succeeds even if a single stripe is allocated. To save
886 * time we give priority to targets which already have objects precreated.
887 * Full OSTs are skipped (see lod_qos_dev_is_full() for the details).
889 * \param[in] env execution environment for this thread
890 * \param[in] lo LOD object
891 * \param[out] stripe striping created
892 * \param[in] flags allocation flags (0 or LOV_USES_DEFAULT_STRIPE)
893 * \param[in] th transaction handle
895 * \retval 0 on success
896 * \retval -ENOSPC if not enough OSTs are found
897 * \retval negative negated errno for other failures
899 static int lod_alloc_rr(const struct lu_env *env, struct lod_object *lo,
900 struct dt_object **stripe, int flags,
903 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
904 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
905 struct pool_desc *pool = NULL;
906 struct ost_pool *osts;
907 struct lod_qos_rr *lqr;
908 unsigned int i, array_idx;
910 __u32 ost_start_idx_temp;
912 __u32 stripe_idx = 0;
913 __u32 stripe_cnt = lo->ldo_stripenr;
914 __u32 stripe_cnt_min = min_stripe_count(stripe_cnt, flags);
919 pool = lod_find_pool(m, lo->ldo_pool);
922 down_read(&pool_tgt_rw_sem(pool));
923 osts = &(pool->pool_obds);
924 lqr = &(pool->pool_rr);
926 osts = &(m->lod_pool_info);
927 lqr = &(m->lod_qos.lq_rr);
930 rc = lod_qos_calc_rr(m, osts, lqr);
934 rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
938 down_read(&m->lod_qos.lq_rw_sem);
939 spin_lock(&lqr->lqr_alloc);
940 if (--lqr->lqr_start_count <= 0) {
941 lqr->lqr_start_idx = cfs_rand() % osts->op_count;
942 lqr->lqr_start_count =
943 (LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) +
944 LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U);
945 } else if (stripe_cnt_min >= osts->op_count ||
946 lqr->lqr_start_idx > osts->op_count) {
947 /* If we have allocated from all of the OSTs, slowly
948 * precess the next start if the OST/stripe count isn't
949 * already doing this for us. */
950 lqr->lqr_start_idx %= osts->op_count;
951 if (stripe_cnt > 1 && (osts->op_count % stripe_cnt) != 1)
952 ++lqr->lqr_offset_idx;
954 ost_start_idx_temp = lqr->lqr_start_idx;
958 QOS_DEBUG("pool '%s' want %d startidx %d startcnt %d offset %d "
959 "active %d count %d\n",
960 lo->ldo_pool ? lo->ldo_pool : "",
961 stripe_cnt, lqr->lqr_start_idx, lqr->lqr_start_count,
962 lqr->lqr_offset_idx, osts->op_count, osts->op_count);
964 for (i = 0; i < osts->op_count && stripe_idx < lo->ldo_stripenr; i++) {
965 array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) %
967 ++lqr->lqr_start_idx;
968 ost_idx = lqr->lqr_pool.op_array[array_idx];
970 QOS_DEBUG("#%d strt %d act %d strp %d ary %d idx %d\n",
971 i, lqr->lqr_start_idx, /* XXX: active*/ 0,
972 stripe_idx, array_idx, ost_idx);
974 if ((ost_idx == LOV_QOS_EMPTY) ||
975 !cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
978 /* Fail Check before osc_precreate() is called
979 so we can only 'fail' single OSC. */
980 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
983 spin_unlock(&lqr->lqr_alloc);
984 rc = lod_check_and_reserve_ost(env, m, sfs, ost_idx, speed,
985 &stripe_idx, stripe, th);
986 spin_lock(&lqr->lqr_alloc);
988 if ((speed < 2) && (stripe_idx < stripe_cnt_min)) {
989 /* Try again, allowing slower OSCs */
991 lqr->lqr_start_idx = ost_start_idx_temp;
995 spin_unlock(&lqr->lqr_alloc);
996 up_read(&m->lod_qos.lq_rw_sem);
999 lo->ldo_stripenr = stripe_idx;
1000 /* at least one stripe is allocated */
1003 /* nobody provided us with a single object */
1009 up_read(&pool_tgt_rw_sem(pool));
1010 /* put back ref got by lod_find_pool() */
1011 lod_pool_putref(pool);
1018 * Allocate a specific striping layout on a user defined set of OSTs.
1020 * Allocates new striping using the OST index range provided by the data from
1021 * the lmm_obejcts contained in the lov_user_md passed to this method. Full
1022 * OSTs are not considered. The exact order of OSTs requested by the user
1023 * is respected as much as possible depending on OST status. The number of
1024 * stripes needed and stripe offset are taken from the object. If that number
1025 * can not be met, then the function returns a failure and then it's the
1026 * caller's responsibility to release the stripes allocated. All the internal
1027 * structures are protected, but no concurrent allocation is allowed on the
1030 * \param[in] env execution environment for this thread
1031 * \param[in] lo LOD object
1032 * \param[out] stripe striping created
1033 * \param[in] lum stripe md to specify list of OSTs
1034 * \param[in] th transaction handle
1036 * \retval 0 on success
1037 * \retval -ENODEV OST index does not exist on file system
1038 * \retval -EINVAL requested OST index is invalid
1039 * \retval negative negated errno on error
1041 static int lod_alloc_ost_list(const struct lu_env *env,
1042 struct lod_object *lo, struct dt_object **stripe,
1043 struct lov_user_md *lum, struct thandle *th)
1045 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1046 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
1047 struct dt_object *o;
1048 struct lov_user_md_v3 *v3;
1049 unsigned int array_idx = 0;
1050 int stripe_count = 0;
1055 /* for specific OSTs layout */
1056 LASSERT(lum != NULL && lum->lmm_magic == LOV_USER_MAGIC_SPECIFIC);
1057 lustre_print_user_md(D_OTHER, lum, __func__);
1059 rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
1063 v3 = (struct lov_user_md_v3 *)lum;
1064 for (i = 0; i < lo->ldo_stripenr; i++) {
1065 if (v3->lmm_objects[i].l_ost_idx == lo->ldo_def_stripe_offset) {
1070 if (i == lo->ldo_stripenr) {
1072 "%s: start index %d not in the specified list of OSTs\n",
1073 lod2obd(m)->obd_name, lo->ldo_def_stripe_offset);
1077 for (i = 0; i < lo->ldo_stripenr;
1078 i++, array_idx = (array_idx + 1) % lo->ldo_stripenr) {
1079 __u32 ost_idx = v3->lmm_objects[array_idx].l_ost_idx;
1081 if (!cfs_bitmap_check(m->lod_ost_bitmap, ost_idx)) {
1087 * do not put >1 objects on a single OST
1089 if (lod_qos_is_ost_used(env, ost_idx, stripe_count)) {
1094 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
1095 if (rc < 0) /* this OSP doesn't feel well */
1098 o = lod_qos_declare_object_on(env, m, ost_idx, th);
1102 "%s: can't declare new object on #%u: %d\n",
1103 lod2obd(m)->obd_name, ost_idx, rc);
1108 * We've successfully declared (reserved) an object
1110 lod_qos_ost_in_use(env, stripe_count, ost_idx);
1111 stripe[stripe_count] = o;
1119 * Allocate a striping on a predefined set of OSTs.
1121 * Allocates new layout starting from OST index in lo->ldo_def_stripe_offset.
1122 * Full OSTs are not considered. The exact order of OSTs is not important and
1123 * varies depending on OST status. The allocation procedure prefers the targets
1124 * with precreated objects ready. The number of stripes needed and stripe
1125 * offset are taken from the object. If that number cannot be met, then the
1126 * function returns an error and then it's the caller's responsibility to
1127 * release the stripes allocated. All the internal structures are protected,
1128 * but no concurrent allocation is allowed on the same objects.
1130 * \param[in] env execution environment for this thread
1131 * \param[in] lo LOD object
1132 * \param[out] stripe striping created
1133 * \param[in] flags not used
1134 * \param[in] th transaction handle
1136 * \retval 0 on success
1137 * \retval -ENOSPC if no OST objects are available at all
1138 * \retval -EFBIG if not enough OST objects are found
1139 * \retval -EINVAL requested offset is invalid
1140 * \retval negative errno on failure
1142 static int lod_alloc_specific(const struct lu_env *env, struct lod_object *lo,
1143 struct dt_object **stripe, int flags,
1146 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1147 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
1148 struct dt_object *o;
1150 unsigned int i, array_idx, ost_count;
1151 int rc, stripe_num = 0;
1153 struct pool_desc *pool = NULL;
1154 struct ost_pool *osts;
1157 rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
1162 pool = lod_find_pool(m, lo->ldo_pool);
1165 down_read(&pool_tgt_rw_sem(pool));
1166 osts = &(pool->pool_obds);
1168 osts = &(m->lod_pool_info);
1171 ost_count = osts->op_count;
1174 /* search loi_ost_idx in ost array */
1176 for (i = 0; i < ost_count; i++) {
1177 if (osts->op_array[i] == lo->ldo_def_stripe_offset) {
1182 if (i == ost_count) {
1183 CERROR("Start index %d not found in pool '%s'\n",
1184 lo->ldo_def_stripe_offset,
1185 lo->ldo_pool ? lo->ldo_pool : "");
1186 GOTO(out, rc = -EINVAL);
1189 for (i = 0; i < ost_count;
1190 i++, array_idx = (array_idx + 1) % ost_count) {
1191 ost_idx = osts->op_array[array_idx];
1193 if (!cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
1196 /* Fail Check before osc_precreate() is called
1197 so we can only 'fail' single OSC. */
1198 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
1202 * do not put >1 objects on a single OST
1204 if (lod_qos_is_ost_used(env, ost_idx, stripe_num))
1207 /* Drop slow OSCs if we can, but not for requested start idx.
1209 * This means "if OSC is slow and it is not the requested
1210 * start OST, then it can be skipped, otherwise skip it only
1211 * if it is inactive/recovering/out-of-space." */
1213 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
1215 /* this OSP doesn't feel well */
1220 * We expect number of precreated objects in f_ffree at
1221 * the first iteration, skip OSPs with no objects ready
1222 * don't apply this logic to OST specified with stripe_offset
1224 if (i != 0 && sfs->os_fprecreated == 0 && speed == 0)
1227 o = lod_qos_declare_object_on(env, m, ost_idx, th);
1229 CDEBUG(D_OTHER, "can't declare new object on #%u: %d\n",
1230 ost_idx, (int) PTR_ERR(o));
1235 * We've successfully declared (reserved) an object
1237 lod_qos_ost_in_use(env, stripe_num, ost_idx);
1238 stripe[stripe_num] = o;
1241 /* We have enough stripes */
1242 if (stripe_num == lo->ldo_stripenr)
1246 /* Try again, allowing slower OSCs */
1251 /* If we were passed specific striping params, then a failure to
1252 * meet those requirements is an error, since we can't reallocate
1253 * that memory (it might be part of a larger array or something).
1255 CERROR("can't lstripe objid "DFID": have %d want %u\n",
1256 PFID(lu_object_fid(lod2lu_obj(lo))), stripe_num,
1258 rc = stripe_num == 0 ? -ENOSPC : -EFBIG;
1261 up_read(&pool_tgt_rw_sem(pool));
1262 /* put back ref got by lod_find_pool() */
1263 lod_pool_putref(pool);
1270 * Check whether QoS allocation should be used.
1272 * A simple helper to decide when QoS allocation should be used:
1273 * if it's just a single available target or the used space is
1274 * evenly distributed among the targets at the moment, then QoS
1275 * allocation algorithm should not be used.
1277 * \param[in] lod LOD device
1279 * \retval 0 should not be used
1280 * \retval 1 should be used
1282 static inline int lod_qos_is_usable(struct lod_device *lod)
1285 /* to be able to debug QoS code */
1289 /* Detect -EAGAIN early, before expensive lock is taken. */
1290 if (!lod->lod_qos.lq_dirty && lod->lod_qos.lq_same_space)
1293 if (lod->lod_desc.ld_active_tgt_count < 2)
1300 * Allocate a striping using an algorithm with weights.
1302 * The function allocates OST objects to create a striping. The algorithm
1303 * used is based on weights (currently only using the free space), and it's
1304 * trying to ensure the space is used evenly by OSTs and OSSs. The striping
1305 * configuration (# of stripes, offset, pool) is taken from the object and
1306 * is prepared by the caller.
1308 * If LOV_USES_DEFAULT_STRIPE is not passed and prepared configuration can't
1309 * be met due to too few OSTs, then allocation fails. If the flag is passed
1310 * fewer than 3/4 of the requested number of stripes can be allocated, then
1313 * No concurrent allocation is allowed on the object and this must be ensured
1314 * by the caller. All the internal structures are protected by the function.
1316 * The algorithm has two steps: find available OSTs and calculate their
1317 * weights, then select the OSTs with their weights used as the probability.
1318 * An OST with a higher weight is proportionately more likely to be selected
1319 * than one with a lower weight.
1321 * \param[in] env execution environment for this thread
1322 * \param[in] lo LOD object
1323 * \param[out] stripe striping created
1324 * \param[in] flags 0 or LOV_USES_DEFAULT_STRIPE
1325 * \param[in] th transaction handle
1327 * \retval 0 on success
1328 * \retval -EAGAIN not enough OSTs are found for specified stripe count
1329 * \retval -EINVAL requested OST index is invalid
1330 * \retval negative errno on failure
1332 static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo,
1333 struct dt_object **stripe, int flags,
1336 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1337 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
1338 struct lod_tgt_desc *ost;
1339 struct dt_object *o;
1340 __u64 total_weight = 0;
1343 __u32 nfound, good_osts;
1344 __u32 stripe_cnt = lo->ldo_stripenr;
1345 __u32 stripe_cnt_min;
1346 struct pool_desc *pool = NULL;
1347 struct ost_pool *osts;
1350 stripe_cnt_min = min_stripe_count(stripe_cnt, flags);
1351 if (stripe_cnt_min < 1)
1355 pool = lod_find_pool(m, lo->ldo_pool);
1358 down_read(&pool_tgt_rw_sem(pool));
1359 osts = &(pool->pool_obds);
1361 osts = &(m->lod_pool_info);
1364 /* Detect -EAGAIN early, before expensive lock is taken. */
1365 if (!lod_qos_is_usable(m))
1366 GOTO(out_nolock, rc = -EAGAIN);
1368 /* Do actual allocation, use write lock here. */
1369 down_write(&m->lod_qos.lq_rw_sem);
1372 * Check again, while we were sleeping on @lq_rw_sem things could
1375 if (!lod_qos_is_usable(m))
1376 GOTO(out, rc = -EAGAIN);
1378 rc = lod_qos_calc_ppo(m);
1382 rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
1387 /* Find all the OSTs that are valid stripe candidates */
1388 for (i = 0; i < osts->op_count; i++) {
1389 if (!cfs_bitmap_check(m->lod_ost_bitmap, osts->op_array[i]))
1392 rc = lod_statfs_and_check(env, m, osts->op_array[i], sfs);
1394 /* this OSP doesn't feel well */
1398 /* Fail Check before osc_precreate() is called
1399 so we can only 'fail' single OSC. */
1400 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) &&
1401 osts->op_array[i] == 0)
1404 ost = OST_TGT(m,osts->op_array[i]);
1405 ost->ltd_qos.ltq_usable = 1;
1406 lod_qos_calc_weight(m, osts->op_array[i]);
1407 total_weight += ost->ltd_qos.ltq_weight;
1412 QOS_DEBUG("found %d good osts\n", good_osts);
1414 if (good_osts < stripe_cnt_min)
1415 GOTO(out, rc = -EAGAIN);
1417 /* We have enough osts */
1418 if (good_osts < stripe_cnt)
1419 stripe_cnt = good_osts;
1421 /* Find enough OSTs with weighted random allocation. */
1423 while (nfound < stripe_cnt) {
1424 __u64 rand, cur_weight;
1430 #if BITS_PER_LONG == 32
1431 rand = cfs_rand() % (unsigned)total_weight;
1432 /* If total_weight > 32-bit, first generate the high
1433 * 32 bits of the random number, then add in the low
1434 * 32 bits (truncated to the upper limit, if needed) */
1435 if (total_weight > 0xffffffffULL)
1436 rand = (__u64)(cfs_rand() %
1437 (unsigned)(total_weight >> 32)) << 32;
1441 if (rand == (total_weight & 0xffffffff00000000ULL))
1442 rand |= cfs_rand() % (unsigned)total_weight;
1447 rand = ((__u64)cfs_rand() << 32 | cfs_rand()) %
1454 /* On average, this will hit larger-weighted OSTs more often.
1455 * 0-weight OSTs will always get used last (only when rand=0) */
1456 for (i = 0; i < osts->op_count; i++) {
1457 __u32 idx = osts->op_array[i];
1459 if (!cfs_bitmap_check(m->lod_ost_bitmap, idx))
1462 ost = OST_TGT(m,idx);
1464 if (!ost->ltd_qos.ltq_usable)
1467 cur_weight += ost->ltd_qos.ltq_weight;
1468 QOS_DEBUG("stripe_cnt=%d nfound=%d cur_weight="LPU64
1469 " rand="LPU64" total_weight="LPU64"\n",
1470 stripe_cnt, nfound, cur_weight, rand,
1473 if (cur_weight < rand)
1476 QOS_DEBUG("stripe=%d to idx=%d\n", nfound, idx);
1479 * do not put >1 objects on a single OST
1481 if (lod_qos_is_ost_used(env, idx, nfound))
1483 lod_qos_ost_in_use(env, nfound, idx);
1485 o = lod_qos_declare_object_on(env, m, idx, th);
1487 QOS_DEBUG("can't declare object on #%u: %d\n",
1488 idx, (int) PTR_ERR(o));
1491 stripe[nfound++] = o;
1492 lod_qos_used(m, osts, idx, &total_weight);
1498 /* no OST found on this iteration, give up */
1503 if (unlikely(nfound != stripe_cnt)) {
1505 * when the decision to use weighted algorithm was made
1506 * we had enough appropriate OSPs, but this state can
1507 * change anytime (no space on OST, broken connection, etc)
1508 * so it's possible OSP won't be able to provide us with
1509 * an object due to just changed state
1511 LCONSOLE_INFO("wanted %d, found %d\n", stripe_cnt, nfound);
1512 for (i = 0; i < nfound; i++) {
1513 LASSERT(stripe[i] != NULL);
1514 lu_object_put(env, &stripe[i]->do_lu);
1518 /* makes sense to rebalance next time */
1519 m->lod_qos.lq_dirty = 1;
1520 m->lod_qos.lq_same_space = 0;
1526 up_write(&m->lod_qos.lq_rw_sem);
1530 up_read(&pool_tgt_rw_sem(pool));
1531 /* put back ref got by lod_find_pool() */
1532 lod_pool_putref(pool);
1539 * Find largest stripe count the caller can use.
1541 * Find the maximal possible stripe count not greater than \a stripe_count.
1542 * Sometimes suggested stripecount can't be reached for a number of reasons:
1543 * lack of enough active OSTs or the backend does not support EAs that large.
1544 * If the passed one is 0, then the filesystem's default one is used.
1546 * \param[in] lod LOD device
1547 * \param[in] magic the format if striping
1548 * \param[in] stripe_count count the caller would like to use
1550 * \retval the maximum usable stripe count
1552 static __u16 lod_get_stripecnt(struct lod_device *lod, __u32 magic,
1555 __u32 max_stripes = LOV_MAX_STRIPE_COUNT_OLD;
1558 stripe_count = lod->lod_desc.ld_default_stripe_count;
1559 if (stripe_count > lod->lod_desc.ld_active_tgt_count)
1560 stripe_count = lod->lod_desc.ld_active_tgt_count;
1564 /* stripe count is based on whether OSD can handle larger EA sizes */
1565 if (lod->lod_osd_max_easize > 0)
1566 max_stripes = lov_mds_md_max_stripe_count(
1567 lod->lod_osd_max_easize, magic);
1569 return (stripe_count < max_stripes) ? stripe_count : max_stripes;
1573 * Create in-core respresentation for a fully-defined striping
1575 * When the caller passes a fully-defined striping (i.e. everything including
1576 * OST object FIDs are defined), then we still need to instantiate LU-cache
1577 * with the objects representing the stripes defined. This function completes
1580 * \param[in] env execution environment for this thread
1581 * \param[in] mo LOD object
1582 * \param[in] buf buffer containing the striping
1584 * \retval 0 on success
1585 * \retval negative negated errno on error
1587 static int lod_use_defined_striping(const struct lu_env *env,
1588 struct lod_object *mo,
1589 const struct lu_buf *buf)
1591 struct lov_mds_md_v1 *v1 = buf->lb_buf;
1592 struct lov_mds_md_v3 *v3 = buf->lb_buf;
1593 struct lov_ost_data_v1 *objs;
1598 magic = le32_to_cpu(v1->lmm_magic);
1599 if (magic == LOV_MAGIC_V1_DEF) {
1600 magic = LOV_MAGIC_V1;
1601 objs = &v1->lmm_objects[0];
1602 } else if (magic == LOV_MAGIC_V3_DEF) {
1603 magic = LOV_MAGIC_V3;
1604 objs = &v3->lmm_objects[0];
1605 lod_object_set_pool(mo, v3->lmm_pool_name);
1607 GOTO(out, rc = -EINVAL);
1610 mo->ldo_pattern = le32_to_cpu(v1->lmm_pattern);
1611 mo->ldo_stripe_size = le32_to_cpu(v1->lmm_stripe_size);
1612 mo->ldo_stripenr = le16_to_cpu(v1->lmm_stripe_count);
1613 mo->ldo_layout_gen = le16_to_cpu(v1->lmm_layout_gen);
1615 /* fixup for released file before object initialization */
1616 if (mo->ldo_pattern & LOV_PATTERN_F_RELEASED) {
1617 mo->ldo_released_stripenr = mo->ldo_stripenr;
1618 mo->ldo_stripenr = 0;
1621 LASSERT(buf->lb_len >= lov_mds_md_size(mo->ldo_stripenr, magic));
1623 if (mo->ldo_stripenr > 0)
1624 rc = lod_initialize_objects(env, mo, objs);
1631 * Parse suggested striping configuration.
1633 * The caller gets a suggested striping configuration from a number of sources
1634 * including per-directory default and applications. Then it needs to verify
1635 * the suggested striping is valid, apply missing bits and store the resulting
1636 * configuration in the object to be used by the allocator later. Must not be
1637 * called concurrently against the same object. It's OK to provide a
1638 * fully-defined striping.
1640 * \param[in] env execution environment for this thread
1641 * \param[in] lo LOD object
1642 * \param[in] buf buffer containing the striping
1644 * \retval 0 on success
1645 * \retval negative negated errno on error
1647 static int lod_qos_parse_config(const struct lu_env *env,
1648 struct lod_object *lo,
1649 const struct lu_buf *buf)
1651 struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
1652 struct lov_user_md_v1 *v1 = NULL;
1653 struct lov_user_md_v3 *v3 = NULL;
1654 char *pool_name = NULL;
1660 if (buf == NULL || buf->lb_buf == NULL || buf->lb_len == 0)
1665 magic = v1->lmm_magic;
1667 if (unlikely(magic == LOV_MAGIC_V1_DEF || magic == LOV_MAGIC_V3_DEF)) {
1668 /* try to use as fully defined striping */
1669 rc = lod_use_defined_striping(env, lo, buf);
1674 case __swab32(LOV_USER_MAGIC_V1):
1675 lustre_swab_lov_user_md_v1(v1);
1676 magic = v1->lmm_magic;
1678 case LOV_USER_MAGIC_V1:
1682 case __swab32(LOV_USER_MAGIC_V3):
1683 lustre_swab_lov_user_md_v3(v3);
1684 magic = v3->lmm_magic;
1686 case LOV_USER_MAGIC_V3:
1688 pool_name = v3->lmm_pool_name;
1691 case __swab32(LOV_USER_MAGIC_SPECIFIC):
1692 lustre_swab_lov_user_md_v3(v3);
1693 lustre_swab_lov_user_md_objects(v3->lmm_objects,
1694 v3->lmm_stripe_count);
1695 magic = v3->lmm_magic;
1697 case LOV_USER_MAGIC_SPECIFIC:
1698 if (v3->lmm_stripe_offset == LOV_OFFSET_DEFAULT)
1699 v3->lmm_stripe_offset = v3->lmm_objects[0].l_ost_idx;
1700 if (v3->lmm_pool_name[0] != '\0')
1701 pool_name = v3->lmm_pool_name;
1702 size = lov_user_md_size(v3->lmm_stripe_count,
1703 LOV_USER_MAGIC_SPECIFIC);
1707 CERROR("%s: unrecognized magic %X\n",
1708 lod2obd(d)->obd_name, magic);
1712 if (unlikely(buf->lb_len < size)) {
1713 CERROR("%s: wrong size: %zd, expect: %u\n",
1714 lod2obd(d)->obd_name, buf->lb_len, size);
1718 lustre_print_user_md(D_OTHER, v1, "parse config");
1720 v1->lmm_magic = magic;
1721 if (v1->lmm_pattern == 0)
1722 v1->lmm_pattern = LOV_PATTERN_RAID0;
1723 if (lov_pattern(v1->lmm_pattern) != LOV_PATTERN_RAID0) {
1724 CERROR("%s: invalid pattern: %x\n",
1725 lod2obd(d)->obd_name, v1->lmm_pattern);
1728 lo->ldo_pattern = v1->lmm_pattern;
1730 if (v1->lmm_stripe_size > 0)
1731 lo->ldo_stripe_size = v1->lmm_stripe_size;
1733 if (lo->ldo_stripe_size & (LOV_MIN_STRIPE_SIZE - 1))
1734 lo->ldo_stripe_size = LOV_MIN_STRIPE_SIZE;
1736 if (v1->lmm_stripe_count > 0)
1737 lo->ldo_stripenr = v1->lmm_stripe_count;
1739 lo->ldo_def_stripe_offset = v1->lmm_stripe_offset;
1741 lod_object_set_pool(lo, NULL);
1742 if (pool_name != NULL) {
1743 struct pool_desc *pool;
1745 /* In the function below, .hs_keycmp resolves to
1746 * pool_hashkey_keycmp() */
1747 /* coverity[overrun-buffer-val] */
1748 pool = lod_find_pool(d, pool_name);
1750 if (lo->ldo_def_stripe_offset != LOV_OFFSET_DEFAULT) {
1751 rc = lod_check_index_in_pool(
1752 lo->ldo_def_stripe_offset, pool);
1754 lod_pool_putref(pool);
1755 CERROR("%s: invalid offset, %u\n",
1756 lod2obd(d)->obd_name,
1757 lo->ldo_def_stripe_offset);
1762 if (lo->ldo_stripenr > pool_tgt_count(pool))
1763 lo->ldo_stripenr = pool_tgt_count(pool);
1765 lod_pool_putref(pool);
1768 lod_object_set_pool(lo, pool_name);
1771 /* fixup for released file */
1772 if (lo->ldo_pattern & LOV_PATTERN_F_RELEASED) {
1773 lo->ldo_released_stripenr = lo->ldo_stripenr;
1774 lo->ldo_stripenr = 0;
1781 * Create a striping for an obejct.
1783 * The function creates a new striping for the object. A buffer containing
1784 * configuration hints can be provided optionally. The function tries QoS
1785 * algorithm first unless free space is distributed evenly among OSTs, but
1786 * by default RR algorithm is preferred due to internal concurrency (QoS is
1787 * serialized). The caller must ensure no concurrent calls to the function
1788 * are made against the same object.
1790 * \param[in] env execution environment for this thread
1791 * \param[in] lo LOD object
1792 * \param[in] attr attributes OST objects will be declared with
1793 * \param[in] buf suggested striping configuration or NULL
1794 * \param[in] th transaction handle
1796 * \retval 0 on success
1797 * \retval negative negated errno on error
1799 int lod_qos_prep_create(const struct lu_env *env, struct lod_object *lo,
1800 struct lu_attr *attr, const struct lu_buf *buf,
1803 struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
1804 struct dt_object **stripe;
1806 int flag = LOV_USES_ASSIGNED_STRIPE;
1812 /* no OST available */
1813 /* XXX: should we be waiting a bit to prevent failures during
1814 * cluster initialization? */
1815 if (d->lod_ostnr == 0)
1816 GOTO(out, rc = -EIO);
1819 * by this time, the object's ldo_stripenr and ldo_stripe_size
1820 * contain default value for striping: taken from the parent
1821 * or from filesystem defaults
1823 * in case the caller is passing lovea with new striping config,
1824 * we may need to parse lovea and apply new configuration
1826 rc = lod_qos_parse_config(env, lo, buf);
1830 /* A released file is being created */
1831 if (lo->ldo_stripenr == 0)
1834 if (likely(lo->ldo_stripe == NULL)) {
1835 struct lov_user_md *lum = NULL;
1838 * no striping has been created so far
1840 LASSERT(lo->ldo_stripenr > 0);
1842 * statfs and check OST targets now, since ld_active_tgt_count
1843 * could be changed if some OSTs are [de]activated manually.
1845 lod_qos_statfs_update(env, d);
1846 lo->ldo_stripenr = lod_get_stripecnt(d, LOV_MAGIC,
1849 stripe_len = lo->ldo_stripenr;
1850 OBD_ALLOC(stripe, sizeof(stripe[0]) * stripe_len);
1852 GOTO(out, rc = -ENOMEM);
1854 lod_getref(&d->lod_ost_descs);
1855 /* XXX: support for non-0 files w/o objects */
1856 CDEBUG(D_OTHER, "tgt_count %d stripenr %d\n",
1857 d->lod_desc.ld_tgt_count, stripe_len);
1859 if (buf != NULL && buf->lb_buf != NULL)
1862 if (lum != NULL && lum->lmm_magic == LOV_USER_MAGIC_SPECIFIC) {
1863 rc = lod_alloc_ost_list(env, lo, stripe, lum, th);
1864 } else if (lo->ldo_def_stripe_offset == LOV_OFFSET_DEFAULT) {
1865 rc = lod_alloc_qos(env, lo, stripe, flag, th);
1867 rc = lod_alloc_rr(env, lo, stripe, flag, th);
1869 rc = lod_alloc_specific(env, lo, stripe, flag, th);
1871 lod_putref(d, &d->lod_ost_descs);
1874 for (i = 0; i < stripe_len; i++)
1875 if (stripe[i] != NULL)
1876 lu_object_put(env, &stripe[i]->do_lu);
1878 OBD_FREE(stripe, sizeof(stripe[0]) * stripe_len);
1879 lo->ldo_stripenr = 0;
1881 lo->ldo_stripe = stripe;
1882 lo->ldo_stripes_allocated = stripe_len;
1886 * lod_qos_parse_config() found supplied buf as a predefined
1887 * striping (not a hint), so it allocated all the object
1888 * now we need to create them
1890 for (i = 0; i < lo->ldo_stripenr; i++) {
1891 struct dt_object *o;
1893 o = lo->ldo_stripe[i];
1896 rc = lod_sub_object_declare_create(env, o, attr, NULL,
1899 CERROR("can't declare create: %d\n", rc);