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, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/lod/lod_qos.c
34 * Implementation of different allocation algorithm used
35 * to distribute objects and data among OSTs.
38 #define DEBUG_SUBSYSTEM S_LOV
40 #include <asm/div64.h>
41 #include <libcfs/libcfs.h>
42 #include <uapi/linux/lustre/lustre_idl.h>
43 #include <lustre_swab.h>
44 #include <obd_class.h>
46 #include "lod_internal.h"
49 * force QoS policy (not RR) to be used for testing purposes
55 #define QOS_DEBUG(fmt, ...) CDEBUG(D_QOS, fmt, ## __VA_ARGS__)
56 #define QOS_CONSOLE(fmt, ...) LCONSOLE(D_QOS, fmt, ## __VA_ARGS__)
58 #define TGT_BAVAIL(i) (OST_TGT(lod,i)->ltd_statfs.os_bavail * \
59 OST_TGT(lod,i)->ltd_statfs.os_bsize)
62 * Add a new target to Quality of Service (QoS) target table.
64 * Add a new OST target to the structure representing an OSS. Resort the list
65 * of known OSSs by the number of OSTs attached to each OSS. The OSS list is
66 * protected internally and no external locking is required.
68 * \param[in] lod LOD device
69 * \param[in] ost_desc OST description
71 * \retval 0 on success
72 * \retval -ENOMEM on error
74 int qos_add_tgt(struct lod_device *lod, struct lod_tgt_desc *ost_desc)
76 struct lod_qos_oss *oss = NULL, *temposs;
77 struct obd_export *exp = ost_desc->ltd_exp;
78 int rc = 0, found = 0;
79 struct list_head *list;
82 down_write(&lod->lod_qos.lq_rw_sem);
84 * a bit hacky approach to learn NID of corresponding connection
85 * but there is no official API to access information like this
88 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
89 if (obd_uuid_equals(&oss->lqo_uuid,
90 &exp->exp_connection->c_remote_uuid)) {
99 GOTO(out, rc = -ENOMEM);
100 memcpy(&oss->lqo_uuid, &exp->exp_connection->c_remote_uuid,
101 sizeof(oss->lqo_uuid));
103 /* Assume we have to move this one */
104 list_del(&oss->lqo_oss_list);
107 oss->lqo_ost_count++;
108 ost_desc->ltd_qos.ltq_oss = oss;
110 CDEBUG(D_QOS, "add tgt %s to OSS %s (%d OSTs)\n",
111 obd_uuid2str(&ost_desc->ltd_uuid), obd_uuid2str(&oss->lqo_uuid),
114 /* Add sorted by # of OSTs. Find the first entry that we're
116 list = &lod->lod_qos.lq_oss_list;
117 list_for_each_entry(temposs, list, lqo_oss_list) {
118 if (oss->lqo_ost_count > temposs->lqo_ost_count)
121 /* ...and add before it. If we're the first or smallest, temposs
122 points to the list head, and we add to the end. */
123 list_add_tail(&oss->lqo_oss_list, &temposs->lqo_oss_list);
125 lod->lod_qos.lq_dirty = 1;
126 lod->lod_qos.lq_rr.lqr_dirty = 1;
129 up_write(&lod->lod_qos.lq_rw_sem);
134 * Remove OST target from QoS table.
136 * Removes given OST target from QoS table and releases related OSS structure
137 * if no OSTs remain on the OSS.
139 * \param[in] lod LOD device
140 * \param[in] ost_desc OST description
142 * \retval 0 on success
143 * \retval -ENOENT if no OSS was found
145 int qos_del_tgt(struct lod_device *lod, struct lod_tgt_desc *ost_desc)
147 struct lod_qos_oss *oss;
151 down_write(&lod->lod_qos.lq_rw_sem);
152 oss = ost_desc->ltd_qos.ltq_oss;
154 GOTO(out, rc = -ENOENT);
156 oss->lqo_ost_count--;
157 if (oss->lqo_ost_count == 0) {
158 CDEBUG(D_QOS, "removing OSS %s\n",
159 obd_uuid2str(&oss->lqo_uuid));
160 list_del(&oss->lqo_oss_list);
161 ost_desc->ltd_qos.ltq_oss = NULL;
165 lod->lod_qos.lq_dirty = 1;
166 lod->lod_qos.lq_rr.lqr_dirty = 1;
168 up_write(&lod->lod_qos.lq_rw_sem);
173 * Check whether the target is available for new OST objects.
175 * Request statfs data from the given target and verify it's active and not
176 * read-only. If so, then it can be used to place new OST objects. This
177 * function also maintains the number of active/inactive targets and sets
178 * dirty flags if those numbers change so others can run re-balance procedures.
179 * No external locking is required.
181 * \param[in] env execution environment for this thread
182 * \param[in] d LOD device
183 * \param[in] index index of OST target to check
184 * \param[out] sfs buffer for statfs data
186 * \retval 0 if the target is good
187 * \retval negative negated errno on error
190 static int lod_statfs_and_check(const struct lu_env *env, struct lod_device *d,
191 int index, struct obd_statfs *sfs)
193 struct lod_tgt_desc *ost;
198 ost = OST_TGT(d,index);
201 rc = dt_statfs(env, ost->ltd_ost, sfs);
203 if (rc == 0 && ((sfs->os_state & OS_STATE_ENOSPC) ||
204 (sfs->os_state & OS_STATE_ENOINO && sfs->os_fprecreated == 0)))
207 if (rc && rc != -ENOTCONN)
208 CERROR("%s: statfs: rc = %d\n", lod2obd(d)->obd_name, rc);
210 /* If the OST is readonly then we can't allocate objects there */
211 if (sfs->os_state & OS_STATE_READONLY)
214 /* check whether device has changed state (active, inactive) */
215 if (rc != 0 && ost->ltd_active) {
216 /* turned inactive? */
217 spin_lock(&d->lod_lock);
218 if (ost->ltd_active) {
221 ost->ltd_connecting = 1;
223 LASSERT(d->lod_desc.ld_active_tgt_count > 0);
224 d->lod_desc.ld_active_tgt_count--;
225 d->lod_qos.lq_dirty = 1;
226 d->lod_qos.lq_rr.lqr_dirty = 1;
227 CDEBUG(D_CONFIG, "%s: turns inactive\n",
228 ost->ltd_exp->exp_obd->obd_name);
230 spin_unlock(&d->lod_lock);
231 } else if (rc == 0 && ost->ltd_active == 0) {
233 LASSERTF(d->lod_desc.ld_active_tgt_count < d->lod_ostnr,
234 "active tgt count %d, ost nr %d\n",
235 d->lod_desc.ld_active_tgt_count, d->lod_ostnr);
236 spin_lock(&d->lod_lock);
237 if (ost->ltd_active == 0) {
239 ost->ltd_connecting = 0;
240 d->lod_desc.ld_active_tgt_count++;
241 d->lod_qos.lq_dirty = 1;
242 d->lod_qos.lq_rr.lqr_dirty = 1;
243 CDEBUG(D_CONFIG, "%s: turns active\n",
244 ost->ltd_exp->exp_obd->obd_name);
246 spin_unlock(&d->lod_lock);
253 * Maintain per-target statfs data.
255 * The function refreshes statfs data for all the targets every N seconds.
256 * The actual N is controlled via procfs and set to LOV_DESC_QOS_MAXAGE_DEFAULT
259 * \param[in] env execution environment for this thread
260 * \param[in] lod LOD device
262 static void lod_qos_statfs_update(const struct lu_env *env,
263 struct lod_device *lod)
265 struct obd_device *obd = lod2obd(lod);
266 struct ost_pool *osts = &(lod->lod_pool_info);
269 __u64 max_age, avail;
272 max_age = cfs_time_shift_64(-2 * lod->lod_desc.ld_qos_maxage);
274 if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
275 /* statfs data are quite recent, don't need to refresh it */
278 down_write(&lod->lod_qos.lq_rw_sem);
279 if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
282 for (i = 0; i < osts->op_count; i++) {
283 idx = osts->op_array[i];
284 avail = OST_TGT(lod,idx)->ltd_statfs.os_bavail;
285 if (lod_statfs_and_check(env, lod, idx,
286 &OST_TGT(lod, idx)->ltd_statfs))
288 if (OST_TGT(lod,idx)->ltd_statfs.os_bavail != avail)
289 /* recalculate weigths */
290 lod->lod_qos.lq_dirty = 1;
292 obd->obd_osfs_age = cfs_time_current_64();
295 up_write(&lod->lod_qos.lq_rw_sem);
300 * Calculate per-OST and per-OSS penalties
302 * Re-calculate penalties when the configuration changes, active targets
303 * change and after statfs refresh (all these are reflected by lq_dirty flag).
304 * On every OST and OSS: decay the penalty by half for every 8x the update
305 * interval that the device has been idle. That gives lots of time for the
306 * statfs information to be updated (which the penalty is only a proxy for),
307 * and avoids penalizing OSS/OSTs under light load.
308 * See lod_qos_calc_weight() for how penalties are factored into the weight.
310 * \param[in] lod LOD device
312 * \retval 0 on success
313 * \retval -EAGAIN the number of OSTs isn't enough
315 static int lod_qos_calc_ppo(struct lod_device *lod)
317 struct lod_qos_oss *oss;
318 __u64 ba_max, ba_min, temp;
325 if (!lod->lod_qos.lq_dirty)
328 num_active = lod->lod_desc.ld_active_tgt_count - 1;
330 GOTO(out, rc = -EAGAIN);
332 /* find bavail on each OSS */
333 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list)
335 lod->lod_qos.lq_active_oss_count = 0;
338 * How badly user wants to select OSTs "widely" (not recently chosen
339 * and not on recent OSS's). As opposed to "freely" (free space
342 prio_wide = 256 - lod->lod_qos.lq_prio_free;
344 ba_min = (__u64)(-1);
346 now = cfs_time_current_sec();
347 /* Calculate OST penalty per object
348 * (lod ref taken in lod_qos_prep_create()) */
349 cfs_foreach_bit(lod->lod_ost_bitmap, i) {
350 LASSERT(OST_TGT(lod,i));
351 temp = TGT_BAVAIL(i);
354 ba_min = min(temp, ba_min);
355 ba_max = max(temp, ba_max);
357 /* Count the number of usable OSS's */
358 if (OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail == 0)
359 lod->lod_qos.lq_active_oss_count++;
360 OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail += temp;
362 /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
364 do_div(temp, num_active);
365 OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj =
366 (temp * prio_wide) >> 8;
368 age = (now - OST_TGT(lod,i)->ltd_qos.ltq_used) >> 3;
369 if (lod->lod_qos.lq_reset ||
370 age > 32 * lod->lod_desc.ld_qos_maxage)
371 OST_TGT(lod,i)->ltd_qos.ltq_penalty = 0;
372 else if (age > lod->lod_desc.ld_qos_maxage)
373 /* Decay OST penalty. */
374 OST_TGT(lod,i)->ltd_qos.ltq_penalty >>=
375 (age / lod->lod_desc.ld_qos_maxage);
378 num_active = lod->lod_qos.lq_active_oss_count - 1;
379 if (num_active < 1) {
380 /* If there's only 1 OSS, we can't penalize it, so instead
381 we have to double the OST penalty */
383 cfs_foreach_bit(lod->lod_ost_bitmap, i)
384 OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj <<= 1;
387 /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
388 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
389 temp = oss->lqo_bavail >> 1;
390 do_div(temp, oss->lqo_ost_count * num_active);
391 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
393 age = (now - oss->lqo_used) >> 3;
394 if (lod->lod_qos.lq_reset ||
395 age > 32 * lod->lod_desc.ld_qos_maxage)
396 oss->lqo_penalty = 0;
397 else if (age > lod->lod_desc.ld_qos_maxage)
398 /* Decay OSS penalty. */
399 oss->lqo_penalty >>= age / lod->lod_desc.ld_qos_maxage;
402 lod->lod_qos.lq_dirty = 0;
403 lod->lod_qos.lq_reset = 0;
405 /* If each ost has almost same free space,
406 * do rr allocation for better creation performance */
407 lod->lod_qos.lq_same_space = 0;
408 if ((ba_max * (256 - lod->lod_qos.lq_threshold_rr)) >> 8 < ba_min) {
409 lod->lod_qos.lq_same_space = 1;
410 /* Reset weights for the next time we enter qos mode */
411 lod->lod_qos.lq_reset = 1;
417 if (!rc && lod->lod_qos.lq_same_space)
424 * Calculate weight for a given OST target.
426 * The final OST weight is the number of bytes available minus the OST and
427 * OSS penalties. See lod_qos_calc_ppo() for how penalties are calculated.
429 * \param[in] lod LOD device, where OST targets are listed
430 * \param[in] i OST target index
434 static int lod_qos_calc_weight(struct lod_device *lod, int i)
438 temp = TGT_BAVAIL(i);
439 temp2 = OST_TGT(lod,i)->ltd_qos.ltq_penalty +
440 OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_penalty;
442 OST_TGT(lod,i)->ltd_qos.ltq_weight = 0;
444 OST_TGT(lod,i)->ltd_qos.ltq_weight = temp - temp2;
449 * Re-calculate weights.
451 * The function is called when some OST target was used for a new object. In
452 * this case we should re-calculate all the weights to keep new allocations
455 * \param[in] lod LOD device
456 * \param[in] osts OST pool where a new object was placed
457 * \param[in] index OST target where a new object was placed
458 * \param[out] total_wt new total weight for the pool
462 static int lod_qos_used(struct lod_device *lod, struct ost_pool *osts,
463 __u32 index, __u64 *total_wt)
465 struct lod_tgt_desc *ost;
466 struct lod_qos_oss *oss;
470 ost = OST_TGT(lod,index);
473 /* Don't allocate on this devuce anymore, until the next alloc_qos */
474 ost->ltd_qos.ltq_usable = 0;
476 oss = ost->ltd_qos.ltq_oss;
478 /* Decay old penalty by half (we're adding max penalty, and don't
479 want it to run away.) */
480 ost->ltd_qos.ltq_penalty >>= 1;
481 oss->lqo_penalty >>= 1;
483 /* mark the OSS and OST as recently used */
484 ost->ltd_qos.ltq_used = oss->lqo_used = cfs_time_current_sec();
486 /* Set max penalties for this OST and OSS */
487 ost->ltd_qos.ltq_penalty +=
488 ost->ltd_qos.ltq_penalty_per_obj * lod->lod_ostnr;
489 oss->lqo_penalty += oss->lqo_penalty_per_obj *
490 lod->lod_qos.lq_active_oss_count;
492 /* Decrease all OSS penalties */
493 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
494 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
495 oss->lqo_penalty = 0;
497 oss->lqo_penalty -= oss->lqo_penalty_per_obj;
501 /* Decrease all OST penalties */
502 for (j = 0; j < osts->op_count; j++) {
505 i = osts->op_array[j];
506 if (!cfs_bitmap_check(lod->lod_ost_bitmap, i))
509 ost = OST_TGT(lod,i);
512 if (ost->ltd_qos.ltq_penalty <
513 ost->ltd_qos.ltq_penalty_per_obj)
514 ost->ltd_qos.ltq_penalty = 0;
516 ost->ltd_qos.ltq_penalty -=
517 ost->ltd_qos.ltq_penalty_per_obj;
519 lod_qos_calc_weight(lod, i);
521 /* Recalc the total weight of usable osts */
522 if (ost->ltd_qos.ltq_usable)
523 *total_wt += ost->ltd_qos.ltq_weight;
525 QOS_DEBUG("recalc tgt %d usable=%d avail=%llu"
526 " ostppo=%llu ostp=%llu ossppo=%llu"
527 " ossp=%llu wt=%llu\n",
528 i, ost->ltd_qos.ltq_usable, TGT_BAVAIL(i) >> 10,
529 ost->ltd_qos.ltq_penalty_per_obj >> 10,
530 ost->ltd_qos.ltq_penalty >> 10,
531 ost->ltd_qos.ltq_oss->lqo_penalty_per_obj >> 10,
532 ost->ltd_qos.ltq_oss->lqo_penalty >> 10,
533 ost->ltd_qos.ltq_weight >> 10);
539 void lod_qos_rr_init(struct lod_qos_rr *lqr)
541 spin_lock_init(&lqr->lqr_alloc);
546 #define LOV_QOS_EMPTY ((__u32)-1)
549 * Calculate optimal round-robin order with regard to OSSes.
551 * Place all the OSTs from pool \a src_pool in a special array to be used for
552 * round-robin (RR) stripe allocation. The placement algorithm interleaves
553 * OSTs from the different OSSs so that RR allocation can balance OSSs evenly.
554 * Resorts the targets when the number of active targets changes (because of
555 * a new target or activation/deactivation).
557 * \param[in] lod LOD device
558 * \param[in] src_pool OST pool
559 * \param[in] lqr round-robin list
561 * \retval 0 on success
562 * \retval -ENOMEM fails to allocate the array
564 static int lod_qos_calc_rr(struct lod_device *lod, struct ost_pool *src_pool,
565 struct lod_qos_rr *lqr)
567 struct lod_qos_oss *oss;
568 struct lod_tgt_desc *ost;
569 unsigned placed, real_count;
574 if (!lqr->lqr_dirty) {
575 LASSERT(lqr->lqr_pool.op_size);
579 /* Do actual allocation. */
580 down_write(&lod->lod_qos.lq_rw_sem);
583 * Check again. While we were sleeping on @lq_rw_sem something could
586 if (!lqr->lqr_dirty) {
587 LASSERT(lqr->lqr_pool.op_size);
588 up_write(&lod->lod_qos.lq_rw_sem);
592 real_count = src_pool->op_count;
594 /* Zero the pool array */
595 /* alloc_rr is holding a read lock on the pool, so nobody is adding/
596 deleting from the pool. The lq_rw_sem insures that nobody else
598 lqr->lqr_pool.op_count = real_count;
599 rc = lod_ost_pool_extend(&lqr->lqr_pool, real_count);
601 up_write(&lod->lod_qos.lq_rw_sem);
604 for (i = 0; i < lqr->lqr_pool.op_count; i++)
605 lqr->lqr_pool.op_array[i] = LOV_QOS_EMPTY;
607 /* Place all the OSTs from 1 OSS at the same time. */
609 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
612 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
615 if (!cfs_bitmap_check(lod->lod_ost_bitmap,
616 src_pool->op_array[i]))
619 ost = OST_TGT(lod,src_pool->op_array[i]);
620 LASSERT(ost && ost->ltd_ost);
621 if (ost->ltd_qos.ltq_oss != oss)
624 /* Evenly space these OSTs across arrayspace */
625 next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count;
626 while (lqr->lqr_pool.op_array[next] != LOV_QOS_EMPTY)
627 next = (next + 1) % lqr->lqr_pool.op_count;
629 lqr->lqr_pool.op_array[next] = src_pool->op_array[i];
636 up_write(&lod->lod_qos.lq_rw_sem);
638 if (placed != real_count) {
639 /* This should never happen */
640 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
641 "round-robin list (%d of %d).\n",
643 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
644 LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
645 lqr->lqr_pool.op_array[i]);
652 for (i = 0; i < lqr->lqr_pool.op_count; i++)
653 QOS_CONSOLE("rr #%d ost idx=%d\n", i, lqr->lqr_pool.op_array[i]);
660 * Instantiate and declare creation of a new object.
662 * The function instantiates LU representation for a new object on the
663 * specified device. Also it declares an intention to create that
664 * object on the storage target.
666 * Note lu_object_anon() is used which is a trick with regard to LU/OSD
667 * infrastructure - in the existing precreation framework we can't assign FID
668 * at this moment, we do this later once a transaction is started. So the
669 * special method instantiates FID-less object in the cache and later it
670 * will get a FID and proper placement in LU cache.
672 * \param[in] env execution environment for this thread
673 * \param[in] d LOD device
674 * \param[in] ost_idx OST target index where the object is being created
675 * \param[in] th transaction handle
677 * \retval object ptr on success, ERR_PTR() otherwise
679 static struct dt_object *lod_qos_declare_object_on(const struct lu_env *env,
680 struct lod_device *d,
684 struct lod_tgt_desc *ost;
685 struct lu_object *o, *n;
686 struct lu_device *nd;
687 struct dt_object *dt;
692 LASSERT(ost_idx < d->lod_osts_size);
693 ost = OST_TGT(d,ost_idx);
695 LASSERT(ost->ltd_ost);
697 nd = &ost->ltd_ost->dd_lu_dev;
700 * allocate anonymous object with zero fid, real fid
701 * will be assigned by OSP within transaction
702 * XXX: to be fixed with fully-functional OST fids
704 o = lu_object_anon(env, nd, NULL);
706 GOTO(out, dt = ERR_PTR(PTR_ERR(o)));
708 n = lu_object_locate(o->lo_header, nd->ld_type);
709 if (unlikely(n == NULL)) {
710 CERROR("can't find slice\n");
711 lu_object_put(env, o);
712 GOTO(out, dt = ERR_PTR(-EINVAL));
715 dt = container_of(n, struct dt_object, do_lu);
717 rc = lod_sub_declare_create(env, dt, NULL, NULL, NULL, th);
719 CDEBUG(D_OTHER, "can't declare creation on #%u: %d\n",
721 lu_object_put(env, o);
730 * Calculate a minimum acceptable stripe count.
732 * Return an acceptable stripe count depending on flag LOV_USES_DEFAULT_STRIPE:
733 * all stripes or 3/4 of stripes.
735 * \param[in] stripe_count number of stripes requested
736 * \param[in] flags 0 or LOV_USES_DEFAULT_STRIPE
738 * \retval acceptable stripecount
740 static int min_stripe_count(__u32 stripe_count, int flags)
742 return (flags & LOV_USES_DEFAULT_STRIPE ?
743 stripe_count - (stripe_count / 4) : stripe_count);
746 #define LOV_CREATE_RESEED_MULT 30
747 #define LOV_CREATE_RESEED_MIN 2000
750 * Initialize temporary OST-in-use array.
752 * Allocate or extend the array used to mark targets already assigned to a new
753 * striping so they are not used more than once.
755 * \param[in] env execution environment for this thread
756 * \param[in] stripes number of items needed in the array
758 * \retval 0 on success
759 * \retval -ENOMEM on error
761 static inline int lod_qos_ost_in_use_clear(const struct lu_env *env,
764 struct lod_thread_info *info = lod_env_info(env);
766 if (info->lti_ea_store_size < sizeof(int) * stripes)
767 lod_ea_store_resize(info, stripes * sizeof(int));
768 if (info->lti_ea_store_size < sizeof(int) * stripes) {
769 CERROR("can't allocate memory for ost-in-use array\n");
772 memset(info->lti_ea_store, -1, sizeof(int) * stripes);
777 * Remember a target in the array of used targets.
779 * Mark the given target as used for a new striping being created. The status
780 * of an OST in a striping can be checked with lod_qos_is_ost_used().
782 * \param[in] env execution environment for this thread
783 * \param[in] idx index in the array
784 * \param[in] ost OST target index to mark as used
786 static inline void lod_qos_ost_in_use(const struct lu_env *env,
789 struct lod_thread_info *info = lod_env_info(env);
790 int *osts = info->lti_ea_store;
792 LASSERT(info->lti_ea_store_size >= idx * sizeof(int));
797 * Check is OST used in a striping.
799 * Checks whether OST with the given index is marked as used in the temporary
800 * array (see lod_qos_ost_in_use()).
802 * \param[in] env execution environment for this thread
803 * \param[in] ost OST target index to check
804 * \param[in] stripes the number of items used in the array already
809 static int lod_qos_is_ost_used(const struct lu_env *env, int ost, __u32 stripes)
811 struct lod_thread_info *info = lod_env_info(env);
812 int *osts = info->lti_ea_store;
815 for (j = 0; j < stripes; j++) {
823 * Check is OST used in a composite layout
825 * \param[in] inuse all inuse ost indexs
826 * \param[in] ost OST target index to check
831 static inline int lod_comp_is_ost_used(struct ost_pool *inuse, int ost)
834 LASSERT(inuse != NULL);
836 if (inuse->op_size == 0)
839 LASSERT(inuse->op_count * sizeof(inuse->op_array[0]) <= inuse->op_size);
840 for (j = 0; j < inuse->op_count; j++) {
841 if (inuse->op_array[j] == ost)
848 * Mark the given target as used for a composite layout
850 * \param[in] inuse inuse ost index array
851 * \param[in] idx index in the array
853 static inline void lod_comp_ost_in_use(struct ost_pool *inuse, int ost)
855 LASSERT(inuse != NULL);
856 if (inuse->op_size && !lod_comp_is_ost_used(inuse, ost)) {
857 LASSERT(inuse->op_count * sizeof(inuse->op_array[0]) <
859 inuse->op_array[inuse->op_count] = ost;
864 static int lod_check_and_reserve_ost(const struct lu_env *env,
865 struct lod_device *m,
866 struct obd_statfs *sfs, __u32 ost_idx,
867 __u32 speed, __u32 *s_idx,
868 struct dt_object **stripe,
870 struct ost_pool *inuse)
873 __u32 stripe_idx = *s_idx;
876 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
878 /* this OSP doesn't feel well */
883 * We expect number of precreated objects in f_ffree at
884 * the first iteration, skip OSPs with no objects ready
886 if (sfs->os_fprecreated == 0 && speed == 0) {
887 QOS_DEBUG("#%d: precreation is empty\n", ost_idx);
892 * try to use another OSP if this one is degraded
894 if (sfs->os_state & OS_STATE_DEGRADED && speed < 2) {
895 QOS_DEBUG("#%d: degraded\n", ost_idx);
900 * try not allocate on OST which has been used by other
903 if (speed == 0 && lod_comp_is_ost_used(inuse, ost_idx)) {
904 QOS_DEBUG("#%d: used by other component\n", ost_idx);
909 * do not put >1 objects on a single OST
911 if (lod_qos_is_ost_used(env, ost_idx, stripe_idx))
914 o = lod_qos_declare_object_on(env, m, ost_idx, th);
916 CDEBUG(D_OTHER, "can't declare new object on #%u: %d\n",
917 ost_idx, (int) PTR_ERR(o));
923 * We've successfully declared (reserved) an object
925 lod_qos_ost_in_use(env, stripe_idx, ost_idx);
926 lod_comp_ost_in_use(inuse, ost_idx);
927 stripe[stripe_idx] = o;
928 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LOV_CREATE_RACE, 2);
937 * Allocate a striping using round-robin algorithm.
939 * Allocates a new striping using round-robin algorithm. The function refreshes
940 * all the internal structures (statfs cache, array of available OSTs sorted
941 * with regard to OSS, etc). The number of stripes required is taken from the
942 * object (must be prepared by the caller), but can change if the flag
943 * LOV_USES_DEFAULT_STRIPE is supplied. The caller should ensure nobody else
944 * is trying to create a striping on the object in parallel. All the internal
945 * structures (like pools, etc) are protected and no additional locking is
946 * required. The function succeeds even if a single stripe is allocated. To save
947 * time we give priority to targets which already have objects precreated.
948 * Full OSTs are skipped (see lod_qos_dev_is_full() for the details).
950 * \param[in] env execution environment for this thread
951 * \param[in] lo LOD object
952 * \param[out] stripe striping created
953 * \param[in] flags allocation flags (0 or LOV_USES_DEFAULT_STRIPE)
954 * \param[in] th transaction handle
955 * \param[in] comp_idx index of ldo_comp_entries
956 * \param[in|out] inuse array of inuse ost index
958 * \retval 0 on success
959 * \retval -ENOSPC if not enough OSTs are found
960 * \retval negative negated errno for other failures
962 static int lod_alloc_rr(const struct lu_env *env, struct lod_object *lo,
963 struct dt_object **stripe, int flags,
964 struct thandle *th, int comp_idx,
965 struct ost_pool *inuse)
967 struct lod_layout_component *lod_comp;
968 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
969 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
970 struct pool_desc *pool = NULL;
971 struct ost_pool *osts;
972 struct lod_qos_rr *lqr;
973 unsigned int i, array_idx;
974 __u32 ost_start_idx_temp;
975 __u32 stripe_idx = 0;
976 __u32 stripe_count, stripe_count_min, ost_idx;
977 int rc, speed = 0, ost_connecting = 0;
980 LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
981 lod_comp = &lo->ldo_comp_entries[comp_idx];
982 stripe_count = lod_comp->llc_stripe_count;
983 stripe_count_min = min_stripe_count(stripe_count, flags);
985 if (lod_comp->llc_pool != NULL)
986 pool = lod_find_pool(m, lod_comp->llc_pool);
989 down_read(&pool_tgt_rw_sem(pool));
990 osts = &(pool->pool_obds);
991 lqr = &(pool->pool_rr);
993 osts = &(m->lod_pool_info);
994 lqr = &(m->lod_qos.lq_rr);
997 rc = lod_qos_calc_rr(m, osts, lqr);
1001 rc = lod_qos_ost_in_use_clear(env, stripe_count);
1005 down_read(&m->lod_qos.lq_rw_sem);
1006 spin_lock(&lqr->lqr_alloc);
1007 if (--lqr->lqr_start_count <= 0) {
1008 lqr->lqr_start_idx = cfs_rand() % osts->op_count;
1009 lqr->lqr_start_count =
1010 (LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) +
1011 LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U);
1012 } else if (stripe_count_min >= osts->op_count ||
1013 lqr->lqr_start_idx > osts->op_count) {
1014 /* If we have allocated from all of the OSTs, slowly
1015 * precess the next start if the OST/stripe count isn't
1016 * already doing this for us. */
1017 lqr->lqr_start_idx %= osts->op_count;
1018 if (stripe_count > 1 && (osts->op_count % stripe_count) != 1)
1019 ++lqr->lqr_offset_idx;
1021 ost_start_idx_temp = lqr->lqr_start_idx;
1025 QOS_DEBUG("pool '%s' want %d start_idx %d start_count %d offset %d "
1026 "active %d count %d\n",
1027 lod_comp->llc_pool ? lod_comp->llc_pool : "",
1028 stripe_count, lqr->lqr_start_idx, lqr->lqr_start_count,
1029 lqr->lqr_offset_idx, osts->op_count, osts->op_count);
1031 for (i = 0; i < osts->op_count && stripe_idx < stripe_count; i++) {
1032 array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) %
1034 ++lqr->lqr_start_idx;
1035 ost_idx = lqr->lqr_pool.op_array[array_idx];
1037 QOS_DEBUG("#%d strt %d act %d strp %d ary %d idx %d\n",
1038 i, lqr->lqr_start_idx, /* XXX: active*/ 0,
1039 stripe_idx, array_idx, ost_idx);
1041 if ((ost_idx == LOV_QOS_EMPTY) ||
1042 !cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
1045 /* Fail Check before osc_precreate() is called
1046 so we can only 'fail' single OSC. */
1047 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
1050 spin_unlock(&lqr->lqr_alloc);
1051 rc = lod_check_and_reserve_ost(env, m, sfs, ost_idx, speed,
1052 &stripe_idx, stripe, th, inuse);
1053 spin_lock(&lqr->lqr_alloc);
1055 if (rc != 0 && OST_TGT(m, ost_idx)->ltd_connecting)
1058 if ((speed < 2) && (stripe_idx < stripe_count_min)) {
1059 /* Try again, allowing slower OSCs */
1061 lqr->lqr_start_idx = ost_start_idx_temp;
1067 spin_unlock(&lqr->lqr_alloc);
1068 up_read(&m->lod_qos.lq_rw_sem);
1071 lod_comp->llc_stripe_count = stripe_idx;
1072 /* at least one stripe is allocated */
1075 /* nobody provided us with a single object */
1084 up_read(&pool_tgt_rw_sem(pool));
1085 /* put back ref got by lod_find_pool() */
1086 lod_pool_putref(pool);
1093 * Allocate a specific striping layout on a user defined set of OSTs.
1095 * Allocates new striping using the OST index range provided by the data from
1096 * the lmm_obejcts contained in the lov_user_md passed to this method. Full
1097 * OSTs are not considered. The exact order of OSTs requested by the user
1098 * is respected as much as possible depending on OST status. The number of
1099 * stripes needed and stripe offset are taken from the object. If that number
1100 * can not be met, then the function returns a failure and then it's the
1101 * caller's responsibility to release the stripes allocated. All the internal
1102 * structures are protected, but no concurrent allocation is allowed on the
1105 * \param[in] env execution environment for this thread
1106 * \param[in] lo LOD object
1107 * \param[out] stripe striping created
1108 * \param[in] th transaction handle
1109 * \param[in] comp_idx index of ldo_comp_entries
1110 * \param[in|out] inuse array of inuse ost index
1112 * \retval 0 on success
1113 * \retval -ENODEV OST index does not exist on file system
1114 * \retval -EINVAL requested OST index is invalid
1115 * \retval negative negated errno on error
1117 static int lod_alloc_ost_list(const struct lu_env *env, struct lod_object *lo,
1118 struct dt_object **stripe, struct thandle *th,
1119 int comp_idx, struct ost_pool *inuse)
1121 struct lod_layout_component *lod_comp;
1122 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1123 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
1124 struct dt_object *o;
1125 unsigned int array_idx = 0;
1126 int stripe_count = 0;
1131 /* for specific OSTs layout */
1132 LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
1133 lod_comp = &lo->ldo_comp_entries[comp_idx];
1134 LASSERT(lod_comp->llc_ostlist.op_array);
1136 rc = lod_qos_ost_in_use_clear(env, lod_comp->llc_stripe_count);
1140 for (i = 0; i < lod_comp->llc_stripe_count; i++) {
1141 if (lod_comp->llc_ostlist.op_array[i] ==
1142 lod_comp->llc_stripe_offset) {
1147 if (i == lod_comp->llc_stripe_count) {
1149 "%s: start index %d not in the specified list of OSTs\n",
1150 lod2obd(m)->obd_name, lod_comp->llc_stripe_offset);
1154 for (i = 0; i < lod_comp->llc_stripe_count;
1155 i++, array_idx = (array_idx + 1) % lod_comp->llc_stripe_count) {
1156 __u32 ost_idx = lod_comp->llc_ostlist.op_array[array_idx];
1158 if (!cfs_bitmap_check(m->lod_ost_bitmap, ost_idx)) {
1164 * do not put >1 objects on a single OST
1166 if (lod_qos_is_ost_used(env, ost_idx, stripe_count)) {
1171 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
1172 if (rc < 0) /* this OSP doesn't feel well */
1175 o = lod_qos_declare_object_on(env, m, ost_idx, th);
1179 "%s: can't declare new object on #%u: %d\n",
1180 lod2obd(m)->obd_name, ost_idx, rc);
1185 * We've successfully declared (reserved) an object
1187 lod_qos_ost_in_use(env, stripe_count, ost_idx);
1188 lod_comp_ost_in_use(inuse, ost_idx);
1189 stripe[stripe_count] = o;
1197 * Allocate a striping on a predefined set of OSTs.
1199 * Allocates new layout starting from OST index in lo->ldo_stripe_offset.
1200 * Full OSTs are not considered. The exact order of OSTs is not important and
1201 * varies depending on OST status. The allocation procedure prefers the targets
1202 * with precreated objects ready. The number of stripes needed and stripe
1203 * offset are taken from the object. If that number cannot be met, then the
1204 * function returns an error and then it's the caller's responsibility to
1205 * release the stripes allocated. All the internal structures are protected,
1206 * but no concurrent allocation is allowed on the same objects.
1208 * \param[in] env execution environment for this thread
1209 * \param[in] lo LOD object
1210 * \param[out] stripe striping created
1211 * \param[in] flags not used
1212 * \param[in] th transaction handle
1213 * \param[in] comp_idx index of ldo_comp_entries
1214 * \param[in|out]inuse array of inuse ost index
1216 * \retval 0 on success
1217 * \retval -ENOSPC if no OST objects are available at all
1218 * \retval -EFBIG if not enough OST objects are found
1219 * \retval -EINVAL requested offset is invalid
1220 * \retval negative errno on failure
1222 static int lod_alloc_specific(const struct lu_env *env, struct lod_object *lo,
1223 struct dt_object **stripe, int flags,
1224 struct thandle *th, int comp_idx,
1225 struct ost_pool *inuse)
1227 struct lod_layout_component *lod_comp;
1228 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1229 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
1230 struct dt_object *o;
1232 unsigned int i, array_idx, ost_count;
1233 int rc, stripe_num = 0;
1235 struct pool_desc *pool = NULL;
1236 struct ost_pool *osts;
1239 LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
1240 lod_comp = &lo->ldo_comp_entries[comp_idx];
1242 rc = lod_qos_ost_in_use_clear(env, lod_comp->llc_stripe_count);
1246 if (lod_comp->llc_pool != NULL)
1247 pool = lod_find_pool(m, lod_comp->llc_pool);
1250 down_read(&pool_tgt_rw_sem(pool));
1251 osts = &(pool->pool_obds);
1253 osts = &(m->lod_pool_info);
1256 ost_count = osts->op_count;
1259 /* search loi_ost_idx in ost array */
1261 for (i = 0; i < ost_count; i++) {
1262 if (osts->op_array[i] == lod_comp->llc_stripe_offset) {
1267 if (i == ost_count) {
1268 CERROR("Start index %d not found in pool '%s'\n",
1269 lod_comp->llc_stripe_offset,
1270 lod_comp->llc_pool ? lod_comp->llc_pool : "");
1271 GOTO(out, rc = -EINVAL);
1274 for (i = 0; i < ost_count;
1275 i++, array_idx = (array_idx + 1) % ost_count) {
1276 ost_idx = osts->op_array[array_idx];
1278 if (!cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
1281 /* Fail Check before osc_precreate() is called
1282 so we can only 'fail' single OSC. */
1283 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
1287 * do not put >1 objects on a single OST
1289 if (lod_qos_is_ost_used(env, ost_idx, stripe_num))
1293 * try not allocate on the OST used by other component
1295 if (speed == 0 && i != 0 &&
1296 lod_comp_is_ost_used(inuse, ost_idx))
1299 /* Drop slow OSCs if we can, but not for requested start idx.
1301 * This means "if OSC is slow and it is not the requested
1302 * start OST, then it can be skipped, otherwise skip it only
1303 * if it is inactive/recovering/out-of-space." */
1305 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
1307 /* this OSP doesn't feel well */
1312 * We expect number of precreated objects in f_ffree at
1313 * the first iteration, skip OSPs with no objects ready
1314 * don't apply this logic to OST specified with stripe_offset
1316 if (i != 0 && sfs->os_fprecreated == 0 && speed == 0)
1319 o = lod_qos_declare_object_on(env, m, ost_idx, th);
1321 CDEBUG(D_OTHER, "can't declare new object on #%u: %d\n",
1322 ost_idx, (int) PTR_ERR(o));
1327 * We've successfully declared (reserved) an object
1329 lod_qos_ost_in_use(env, stripe_num, ost_idx);
1330 lod_comp_ost_in_use(inuse, ost_idx);
1331 stripe[stripe_num] = o;
1334 /* We have enough stripes */
1335 if (stripe_num == lod_comp->llc_stripe_count)
1339 /* Try again, allowing slower OSCs */
1344 /* If we were passed specific striping params, then a failure to
1345 * meet those requirements is an error, since we can't reallocate
1346 * that memory (it might be part of a larger array or something).
1348 CERROR("can't lstripe objid "DFID": have %d want %u\n",
1349 PFID(lu_object_fid(lod2lu_obj(lo))), stripe_num,
1350 lod_comp->llc_stripe_count);
1351 rc = stripe_num == 0 ? -ENOSPC : -EFBIG;
1354 up_read(&pool_tgt_rw_sem(pool));
1355 /* put back ref got by lod_find_pool() */
1356 lod_pool_putref(pool);
1363 * Check whether QoS allocation should be used.
1365 * A simple helper to decide when QoS allocation should be used:
1366 * if it's just a single available target or the used space is
1367 * evenly distributed among the targets at the moment, then QoS
1368 * allocation algorithm should not be used.
1370 * \param[in] lod LOD device
1372 * \retval 0 should not be used
1373 * \retval 1 should be used
1375 static inline int lod_qos_is_usable(struct lod_device *lod)
1378 /* to be able to debug QoS code */
1382 /* Detect -EAGAIN early, before expensive lock is taken. */
1383 if (!lod->lod_qos.lq_dirty && lod->lod_qos.lq_same_space)
1386 if (lod->lod_desc.ld_active_tgt_count < 2)
1393 * Allocate a striping using an algorithm with weights.
1395 * The function allocates OST objects to create a striping. The algorithm
1396 * used is based on weights (currently only using the free space), and it's
1397 * trying to ensure the space is used evenly by OSTs and OSSs. The striping
1398 * configuration (# of stripes, offset, pool) is taken from the object and
1399 * is prepared by the caller.
1401 * If LOV_USES_DEFAULT_STRIPE is not passed and prepared configuration can't
1402 * be met due to too few OSTs, then allocation fails. If the flag is passed
1403 * fewer than 3/4 of the requested number of stripes can be allocated, then
1406 * No concurrent allocation is allowed on the object and this must be ensured
1407 * by the caller. All the internal structures are protected by the function.
1409 * The algorithm has two steps: find available OSTs and calculate their
1410 * weights, then select the OSTs with their weights used as the probability.
1411 * An OST with a higher weight is proportionately more likely to be selected
1412 * than one with a lower weight.
1414 * \param[in] env execution environment for this thread
1415 * \param[in] lo LOD object
1416 * \param[out] stripe striping created
1417 * \param[in] flags 0 or LOV_USES_DEFAULT_STRIPE
1418 * \param[in] th transaction handle
1419 * \param[in] comp_idx index of ldo_comp_entries
1420 * \param[in|out]inuse array of inuse ost index
1422 * \retval 0 on success
1423 * \retval -EAGAIN not enough OSTs are found for specified stripe count
1424 * \retval -EINVAL requested OST index is invalid
1425 * \retval negative errno on failure
1427 static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo,
1428 struct dt_object **stripe, int flags,
1429 struct thandle *th, int comp_idx,
1430 struct ost_pool *inuse)
1432 struct lod_layout_component *lod_comp;
1433 struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1434 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
1435 struct lod_tgt_desc *ost;
1436 struct dt_object *o;
1437 __u64 total_weight = 0;
1438 struct pool_desc *pool = NULL;
1439 struct ost_pool *osts;
1441 __u32 nfound, good_osts, stripe_count, stripe_count_min;
1445 LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
1446 lod_comp = &lo->ldo_comp_entries[comp_idx];
1447 stripe_count = lod_comp->llc_stripe_count;
1448 stripe_count_min = min_stripe_count(stripe_count, flags);
1449 if (stripe_count_min < 1)
1452 if (lod_comp->llc_pool != NULL)
1453 pool = lod_find_pool(lod, lod_comp->llc_pool);
1456 down_read(&pool_tgt_rw_sem(pool));
1457 osts = &(pool->pool_obds);
1459 osts = &(lod->lod_pool_info);
1462 /* Detect -EAGAIN early, before expensive lock is taken. */
1463 if (!lod_qos_is_usable(lod))
1464 GOTO(out_nolock, rc = -EAGAIN);
1466 /* Do actual allocation, use write lock here. */
1467 down_write(&lod->lod_qos.lq_rw_sem);
1470 * Check again, while we were sleeping on @lq_rw_sem things could
1473 if (!lod_qos_is_usable(lod))
1474 GOTO(out, rc = -EAGAIN);
1476 rc = lod_qos_calc_ppo(lod);
1480 rc = lod_qos_ost_in_use_clear(env, lod_comp->llc_stripe_count);
1485 /* Find all the OSTs that are valid stripe candidates */
1486 for (i = 0; i < osts->op_count; i++) {
1487 if (!cfs_bitmap_check(lod->lod_ost_bitmap, osts->op_array[i]))
1490 ost = OST_TGT(lod, osts->op_array[i]);
1491 ost->ltd_qos.ltq_usable = 0;
1493 rc = lod_statfs_and_check(env, lod, osts->op_array[i], sfs);
1495 /* this OSP doesn't feel well */
1499 if (sfs->os_state & OS_STATE_DEGRADED)
1502 /* Fail Check before osc_precreate() is called
1503 so we can only 'fail' single OSC. */
1504 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) &&
1505 osts->op_array[i] == 0)
1508 ost->ltd_qos.ltq_usable = 1;
1509 lod_qos_calc_weight(lod, osts->op_array[i]);
1510 total_weight += ost->ltd_qos.ltq_weight;
1515 QOS_DEBUG("found %d good osts\n", good_osts);
1517 if (good_osts < stripe_count_min)
1518 GOTO(out, rc = -EAGAIN);
1520 /* We have enough osts */
1521 if (good_osts < stripe_count)
1522 stripe_count = good_osts;
1524 /* Find enough OSTs with weighted random allocation. */
1526 while (nfound < stripe_count) {
1527 __u64 rand, cur_weight;
1533 #if BITS_PER_LONG == 32
1534 rand = cfs_rand() % (unsigned)total_weight;
1535 /* If total_weight > 32-bit, first generate the high
1536 * 32 bits of the random number, then add in the low
1537 * 32 bits (truncated to the upper limit, if needed) */
1538 if (total_weight > 0xffffffffULL)
1539 rand = (__u64)(cfs_rand() %
1540 (unsigned)(total_weight >> 32)) << 32;
1544 if (rand == (total_weight & 0xffffffff00000000ULL))
1545 rand |= cfs_rand() % (unsigned)total_weight;
1550 rand = ((__u64)cfs_rand() << 32 | cfs_rand()) %
1557 /* On average, this will hit larger-weighted OSTs more often.
1558 * 0-weight OSTs will always get used last (only when rand=0) */
1559 for (i = 0; i < osts->op_count; i++) {
1560 __u32 idx = osts->op_array[i];
1562 if (!cfs_bitmap_check(lod->lod_ost_bitmap, idx))
1565 ost = OST_TGT(lod, idx);
1567 if (!ost->ltd_qos.ltq_usable)
1570 cur_weight += ost->ltd_qos.ltq_weight;
1571 QOS_DEBUG("stripe_count=%d nfound=%d cur_weight=%llu "
1572 "rand=%llu total_weight=%llu\n",
1573 stripe_count, nfound, cur_weight, rand,
1576 if (cur_weight < rand)
1579 QOS_DEBUG("stripe=%d to idx=%d\n", nfound, idx);
1581 * do not put >1 objects on a single OST
1583 if (lod_qos_is_ost_used(env, idx, nfound) ||
1584 lod_comp_is_ost_used(inuse, idx))
1587 o = lod_qos_declare_object_on(env, lod, idx, th);
1589 QOS_DEBUG("can't declare object on #%u: %d\n",
1590 idx, (int) PTR_ERR(o));
1594 lod_qos_ost_in_use(env, nfound, idx);
1595 lod_comp_ost_in_use(inuse, idx);
1596 stripe[nfound++] = o;
1597 lod_qos_used(lod, osts, idx, &total_weight);
1603 /* no OST found on this iteration, give up */
1608 if (unlikely(nfound != stripe_count)) {
1610 * when the decision to use weighted algorithm was made
1611 * we had enough appropriate OSPs, but this state can
1612 * change anytime (no space on OST, broken connection, etc)
1613 * so it's possible OSP won't be able to provide us with
1614 * an object due to just changed state
1616 QOS_DEBUG("%s: wanted %d objects, found only %d\n",
1617 lod2obd(lod)->obd_name, stripe_count, nfound);
1618 for (i = 0; i < nfound; i++) {
1619 LASSERT(stripe[i] != NULL);
1620 dt_object_put(env, stripe[i]);
1623 LASSERTF(nfound <= inuse->op_count,
1624 "nfound:%d, op_count:%u\n", nfound, inuse->op_count);
1625 inuse->op_count -= nfound;
1627 /* makes sense to rebalance next time */
1628 lod->lod_qos.lq_dirty = 1;
1629 lod->lod_qos.lq_same_space = 0;
1635 up_write(&lod->lod_qos.lq_rw_sem);
1639 up_read(&pool_tgt_rw_sem(pool));
1640 /* put back ref got by lod_find_pool() */
1641 lod_pool_putref(pool);
1648 * Find largest stripe count the caller can use.
1650 * Find the maximal possible stripe count not greater than \a stripe_count.
1651 * Sometimes suggested stripecount can't be reached for a number of reasons:
1652 * lack of enough active OSTs or the backend does not support EAs that large.
1653 * If the passed one is 0, then the filesystem's default one is used.
1655 * \param[in] lod LOD device
1656 * \param[in] lo The lod_object
1657 * \param[in] stripe_count count the caller would like to use
1659 * \retval the maximum usable stripe count
1661 __u16 lod_get_stripe_count(struct lod_device *lod, struct lod_object *lo,
1664 __u32 max_stripes = LOV_MAX_STRIPE_COUNT_OLD;
1667 stripe_count = lod->lod_desc.ld_default_stripe_count;
1668 if (stripe_count > lod->lod_desc.ld_active_tgt_count)
1669 stripe_count = lod->lod_desc.ld_active_tgt_count;
1673 /* stripe count is based on whether OSD can handle larger EA sizes */
1674 if (lod->lod_osd_max_easize > 0) {
1675 unsigned int easize = lod->lod_osd_max_easize;
1678 if (lo->ldo_is_composite) {
1679 struct lod_layout_component *lod_comp;
1680 unsigned int header_sz = sizeof(struct lov_comp_md_v1);
1682 header_sz += sizeof(struct lov_comp_md_entry_v1) *
1684 for (i = 0; i < lo->ldo_comp_cnt; i++) {
1685 lod_comp = &lo->ldo_comp_entries[i];
1686 if (lod_comp->llc_flags & LCME_FL_INIT)
1687 header_sz += lov_mds_md_size(
1688 lod_comp->llc_stripe_count,
1691 if (easize > header_sz)
1692 easize -= header_sz;
1697 max_stripes = lov_mds_md_max_stripe_count(easize, LOV_MAGIC_V3);
1700 return (stripe_count < max_stripes) ? stripe_count : max_stripes;
1704 * Create in-core respresentation for a fully-defined striping
1706 * When the caller passes a fully-defined striping (i.e. everything including
1707 * OST object FIDs are defined), then we still need to instantiate LU-cache
1708 * with the objects representing the stripes defined. This function completes
1711 * \param[in] env execution environment for this thread
1712 * \param[in] mo LOD object
1713 * \param[in] buf buffer containing the striping
1715 * \retval 0 on success
1716 * \retval negative negated errno on error
1718 int lod_use_defined_striping(const struct lu_env *env,
1719 struct lod_object *mo,
1720 const struct lu_buf *buf)
1722 struct lod_layout_component *lod_comp;
1723 struct lov_mds_md_v1 *v1 = buf->lb_buf;
1724 struct lov_mds_md_v3 *v3 = buf->lb_buf;
1725 struct lov_comp_md_v1 *comp_v1 = NULL;
1726 struct lov_ost_data_v1 *objs;
1732 magic = le32_to_cpu(v1->lmm_magic) & ~LOV_MAGIC_DEF;
1734 if (magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3 &&
1735 magic != LOV_MAGIC_COMP_V1)
1738 if (magic == LOV_MAGIC_COMP_V1) {
1739 comp_v1 = buf->lb_buf;
1740 comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count);
1743 mo->ldo_is_composite = 1;
1745 mo->ldo_is_composite = 0;
1749 rc = lod_alloc_comp_entries(mo, comp_cnt);
1753 for (i = 0; i < comp_cnt; i++) {
1754 struct lu_extent *ext;
1758 lod_comp = &mo->ldo_comp_entries[i];
1760 if (mo->ldo_is_composite) {
1761 offs = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
1762 v1 = (struct lov_mds_md_v1 *)((char *)comp_v1 + offs);
1763 magic = le32_to_cpu(v1->lmm_magic);
1765 ext = &comp_v1->lcm_entries[i].lcme_extent;
1766 lod_comp->llc_extent.e_start =
1767 le64_to_cpu(ext->e_start);
1768 lod_comp->llc_extent.e_end = le64_to_cpu(ext->e_end);
1769 lod_comp->llc_flags =
1770 le32_to_cpu(comp_v1->lcm_entries[i].lcme_flags);
1772 le32_to_cpu(comp_v1->lcm_entries[i].lcme_id);
1773 if (lod_comp->llc_id == LCME_ID_INVAL)
1774 GOTO(out, rc = -EINVAL);
1778 if (magic == LOV_MAGIC_V1) {
1779 objs = &v1->lmm_objects[0];
1780 } else if (magic == LOV_MAGIC_V3) {
1781 objs = &v3->lmm_objects[0];
1782 if (v3->lmm_pool_name[0] != '\0')
1783 pool_name = v3->lmm_pool_name;
1785 CDEBUG(D_LAYOUT, "Invalid magic %x\n", magic);
1786 GOTO(out, rc = -EINVAL);
1789 lod_comp->llc_pattern = le32_to_cpu(v1->lmm_pattern);
1790 lod_comp->llc_stripe_size = le32_to_cpu(v1->lmm_stripe_size);
1791 lod_comp->llc_stripe_count = le16_to_cpu(v1->lmm_stripe_count);
1792 lod_comp->llc_layout_gen = le16_to_cpu(v1->lmm_layout_gen);
1794 * The stripe_offset of an uninit-ed component is stored in
1795 * the lmm_layout_gen
1797 if (mo->ldo_is_composite && !lod_comp_inited(lod_comp))
1798 lod_comp->llc_stripe_offset = lod_comp->llc_layout_gen;
1799 lod_obj_set_pool(mo, i, pool_name);
1801 if ((!mo->ldo_is_composite || lod_comp_inited(lod_comp)) &&
1802 !(lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED)) {
1803 rc = lod_initialize_objects(env, mo, objs, i);
1810 lod_object_free_striping(env, mo);
1816 * Parse suggested striping configuration.
1818 * The caller gets a suggested striping configuration from a number of sources
1819 * including per-directory default and applications. Then it needs to verify
1820 * the suggested striping is valid, apply missing bits and store the resulting
1821 * configuration in the object to be used by the allocator later. Must not be
1822 * called concurrently against the same object. It's OK to provide a
1823 * fully-defined striping.
1825 * \param[in] env execution environment for this thread
1826 * \param[in] lo LOD object
1827 * \param[in] buf buffer containing the striping
1829 * \retval 0 on success
1830 * \retval negative negated errno on error
1832 int lod_qos_parse_config(const struct lu_env *env, struct lod_object *lo,
1833 const struct lu_buf *buf)
1835 struct lod_layout_component *lod_comp;
1836 struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
1837 struct lov_desc *desc = &d->lod_desc;
1838 struct lov_user_md_v1 *v1 = NULL;
1839 struct lov_user_md_v3 *v3 = NULL;
1840 struct lov_comp_md_v1 *comp_v1 = NULL;
1846 if (buf == NULL || buf->lb_buf == NULL || buf->lb_len == 0)
1849 rc = lod_verify_striping(d, buf, false, 0);
1853 lod_free_comp_entries(lo);
1857 comp_v1 = buf->lb_buf;
1858 magic = v1->lmm_magic;
1860 if (unlikely(le32_to_cpu(magic) & LOV_MAGIC_DEF)) {
1861 /* try to use as fully defined striping */
1862 rc = lod_use_defined_striping(env, lo, buf);
1867 case __swab32(LOV_USER_MAGIC_V1):
1868 lustre_swab_lov_user_md_v1(v1);
1869 magic = v1->lmm_magic;
1871 case LOV_USER_MAGIC_V1:
1873 case __swab32(LOV_USER_MAGIC_V3):
1874 lustre_swab_lov_user_md_v3(v3);
1875 magic = v3->lmm_magic;
1877 case LOV_USER_MAGIC_V3:
1879 case __swab32(LOV_USER_MAGIC_SPECIFIC):
1880 lustre_swab_lov_user_md_v3(v3);
1881 lustre_swab_lov_user_md_objects(v3->lmm_objects,
1882 v3->lmm_stripe_count);
1883 magic = v3->lmm_magic;
1885 case LOV_USER_MAGIC_SPECIFIC:
1887 case __swab32(LOV_USER_MAGIC_COMP_V1):
1888 lustre_swab_lov_comp_md_v1(comp_v1);
1889 magic = comp_v1->lcm_magic;
1891 case LOV_USER_MAGIC_COMP_V1:
1894 CERROR("%s: unrecognized magic %X\n",
1895 lod2obd(d)->obd_name, magic);
1899 lustre_print_user_md(D_OTHER, v1, "parse config");
1901 if (magic == LOV_USER_MAGIC_COMP_V1) {
1902 comp_cnt = comp_v1->lcm_entry_count;
1905 lo->ldo_is_composite = 1;
1908 lo->ldo_is_composite = 0;
1911 rc = lod_alloc_comp_entries(lo, comp_cnt);
1915 for (i = 0; i < comp_cnt; i++) {
1916 struct pool_desc *pool;
1917 struct lu_extent *ext;
1920 lod_comp = &lo->ldo_comp_entries[i];
1922 if (lo->ldo_is_composite) {
1923 v1 = (struct lov_user_md *)((char *)comp_v1 +
1924 comp_v1->lcm_entries[i].lcme_offset);
1925 ext = &comp_v1->lcm_entries[i].lcme_extent;
1926 lod_comp->llc_extent = *ext;
1930 if (v1->lmm_magic == LOV_USER_MAGIC_V3 ||
1931 v1->lmm_magic == LOV_USER_MAGIC_SPECIFIC) {
1934 v3 = (struct lov_user_md_v3 *)v1;
1935 if (v3->lmm_pool_name[0] != '\0')
1936 pool_name = v3->lmm_pool_name;
1938 if (v3->lmm_magic == LOV_USER_MAGIC_SPECIFIC) {
1939 if (v3->lmm_stripe_offset == LOV_OFFSET_DEFAULT)
1940 v3->lmm_stripe_offset =
1941 v3->lmm_objects[0].l_ost_idx;
1943 /* copy ost list from lmm */
1944 lod_comp->llc_ostlist.op_count =
1945 v3->lmm_stripe_count;
1946 lod_comp->llc_ostlist.op_size =
1947 v3->lmm_stripe_count * sizeof(__u32);
1948 OBD_ALLOC(lod_comp->llc_ostlist.op_array,
1949 lod_comp->llc_ostlist.op_size);
1950 if (!lod_comp->llc_ostlist.op_array)
1951 GOTO(free_comp, rc = -ENOMEM);
1953 for (j = 0; j < v3->lmm_stripe_count; j++)
1954 lod_comp->llc_ostlist.op_array[j] =
1955 v3->lmm_objects[j].l_ost_idx;
1959 if (v1->lmm_pattern == 0)
1960 v1->lmm_pattern = LOV_PATTERN_RAID0;
1961 if (lov_pattern(v1->lmm_pattern) != LOV_PATTERN_RAID0) {
1962 CDEBUG(D_LAYOUT, "%s: invalid pattern: %x\n",
1963 lod2obd(d)->obd_name, v1->lmm_pattern);
1964 GOTO(free_comp, rc = -EINVAL);
1967 lod_comp->llc_pattern = v1->lmm_pattern;
1969 lod_comp->llc_stripe_size = desc->ld_default_stripe_size;
1970 if (v1->lmm_stripe_size)
1971 lod_comp->llc_stripe_size = v1->lmm_stripe_size;
1973 lod_comp->llc_stripe_count = desc->ld_default_stripe_count;
1974 if (v1->lmm_stripe_count)
1975 lod_comp->llc_stripe_count = v1->lmm_stripe_count;
1977 lod_comp->llc_stripe_offset = v1->lmm_stripe_offset;
1978 lod_obj_set_pool(lo, i, pool_name);
1980 if (pool_name == NULL)
1983 /* In the function below, .hs_keycmp resolves to
1984 * pool_hashkey_keycmp() */
1985 /* coverity[overrun-buffer-val] */
1986 pool = lod_find_pool(d, pool_name);
1990 if (lod_comp->llc_stripe_offset != LOV_OFFSET_DEFAULT) {
1991 rc = lod_check_index_in_pool(
1992 lod_comp->llc_stripe_offset, pool);
1994 lod_pool_putref(pool);
1995 CDEBUG(D_LAYOUT, "%s: invalid offset, %u\n",
1996 lod2obd(d)->obd_name,
1997 lod_comp->llc_stripe_offset);
1998 GOTO(free_comp, rc = -EINVAL);
2002 if (lod_comp->llc_stripe_count > pool_tgt_count(pool))
2003 lod_comp->llc_stripe_count = pool_tgt_count(pool);
2005 lod_pool_putref(pool);
2011 lod_free_comp_entries(lo);
2016 * Create a striping for an obejct.
2018 * The function creates a new striping for the object. The function tries QoS
2019 * algorithm first unless free space is distributed evenly among OSTs, but
2020 * by default RR algorithm is preferred due to internal concurrency (QoS is
2021 * serialized). The caller must ensure no concurrent calls to the function
2022 * are made against the same object.
2024 * \param[in] env execution environment for this thread
2025 * \param[in] lo LOD object
2026 * \param[in] attr attributes OST objects will be declared with
2027 * \param[in] th transaction handle
2028 * \param[in] comp_idx index of ldo_comp_entries
2029 * \param[in|out] inuse array of inuse ost index
2031 * \retval 0 on success
2032 * \retval negative negated errno on error
2034 int lod_qos_prep_create(const struct lu_env *env, struct lod_object *lo,
2035 struct lu_attr *attr, struct thandle *th,
2036 int comp_idx, struct ost_pool *inuse)
2038 struct lod_layout_component *lod_comp;
2039 struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
2040 struct dt_object **stripe;
2042 int flag = LOV_USES_ASSIGNED_STRIPE;
2047 LASSERT(lo->ldo_comp_cnt > comp_idx && lo->ldo_comp_entries != NULL);
2048 lod_comp = &lo->ldo_comp_entries[comp_idx];
2050 /* A released component is being created */
2051 if (lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED)
2054 if (likely(lod_comp->llc_stripe == NULL)) {
2056 * no striping has been created so far
2058 LASSERT(lod_comp->llc_stripe_count);
2060 * statfs and check OST targets now, since ld_active_tgt_count
2061 * could be changed if some OSTs are [de]activated manually.
2063 lod_qos_statfs_update(env, d);
2064 stripe_len = lod_get_stripe_count(d, lo,
2065 lod_comp->llc_stripe_count);
2066 if (stripe_len == 0)
2067 GOTO(out, rc = -ERANGE);
2068 lod_comp->llc_stripe_count = stripe_len;
2069 OBD_ALLOC(stripe, sizeof(stripe[0]) * stripe_len);
2071 GOTO(out, rc = -ENOMEM);
2073 lod_getref(&d->lod_ost_descs);
2074 /* XXX: support for non-0 files w/o objects */
2075 CDEBUG(D_OTHER, "tgt_count %d stripe_count %d\n",
2076 d->lod_desc.ld_tgt_count, stripe_len);
2078 if (lod_comp->llc_ostlist.op_array) {
2079 rc = lod_alloc_ost_list(env, lo, stripe, th, comp_idx,
2081 } else if (lod_comp->llc_stripe_offset == LOV_OFFSET_DEFAULT) {
2082 rc = lod_alloc_qos(env, lo, stripe, flag, th,
2085 rc = lod_alloc_rr(env, lo, stripe, flag, th,
2088 rc = lod_alloc_specific(env, lo, stripe, flag, th,
2091 lod_putref(d, &d->lod_ost_descs);
2094 for (i = 0; i < stripe_len; i++)
2095 if (stripe[i] != NULL)
2096 dt_object_put(env, stripe[i]);
2098 OBD_FREE(stripe, sizeof(stripe[0]) * stripe_len);
2099 lod_comp->llc_stripe_count = 0;
2101 lod_comp->llc_stripe = stripe;
2102 lod_comp->llc_stripes_allocated = stripe_len;
2106 * lod_qos_parse_config() found supplied buf as a predefined
2107 * striping (not a hint), so it allocated all the object
2108 * now we need to create them
2110 for (i = 0; i < lod_comp->llc_stripe_count; i++) {
2111 struct dt_object *o;
2113 o = lod_comp->llc_stripe[i];
2116 rc = lod_sub_declare_create(env, o, attr, NULL,
2119 CERROR("can't declare create: %d\n", rc);
2124 * Clear LCME_FL_INIT for the component so that
2125 * lod_striping_create() can create the striping objects
2128 lod_comp_unset_init(lod_comp);
2135 int lod_obj_stripe_set_inuse_cb(const struct lu_env *env,
2136 struct lod_object *lo,
2137 struct dt_object *dt, struct thandle *th,
2139 struct lod_obj_stripe_cb_data *data)
2141 struct lod_thread_info *info = lod_env_info(env);
2142 struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
2143 struct lu_fid *fid = &info->lti_fid;
2145 int rc, type = LU_SEQ_RANGE_OST;
2147 *fid = *lu_object_fid(&dt->do_lu);
2148 rc = lod_fld_lookup(env, d, fid, &index, &type);
2150 CERROR("%s: fail to locate "DFID": rc = %d\n",
2151 lod2obd(d)->obd_name, PFID(fid), rc);
2154 lod_comp_ost_in_use(data->locd_inuse, index);
2159 * Resize per-thread ost list to hold OST target index list already used.
2161 * \param[in,out] inuse structure contains ost list array
2162 * \param[in] cnt total stripe count of all components
2163 * \param[in] max array's max size if @max > 0
2165 * \retval 0 on success
2166 * \retval -ENOMEM reallocation failed
2168 static int lod_inuse_resize(struct ost_pool *inuse, __u16 cnt, __u16 max)
2171 __u32 new = cnt * sizeof(inuse->op_array[0]);
2173 inuse->op_count = 0;
2175 if (new <= inuse->op_size)
2179 new = min_t(__u32, new, max);
2181 OBD_ALLOC(array, new);
2185 if (inuse->op_array)
2186 OBD_FREE(inuse->op_array, inuse->op_size);
2188 inuse->op_array = array;
2189 inuse->op_size = new;
2194 int lod_prepare_inuse(const struct lu_env *env, struct lod_object *lo)
2196 struct lod_thread_info *info = lod_env_info(env);
2197 struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
2198 struct ost_pool *inuse = &info->lti_inuse_osts;
2199 struct lod_obj_stripe_cb_data data;
2200 __u32 stripe_count = 0;
2204 for (i = 0; i < lo->ldo_comp_cnt; i++)
2205 stripe_count += lod_comp_entry_stripe_count(lo,
2206 &lo->ldo_comp_entries[i], false);
2207 rc = lod_inuse_resize(inuse, stripe_count, d->lod_osd_max_easize);
2211 data.locd_inuse = inuse;
2212 return lod_obj_for_each_stripe(env, lo, NULL,
2213 lod_obj_stripe_set_inuse_cb, &data);
2216 int lod_prepare_create(const struct lu_env *env, struct lod_object *lo,
2217 struct lu_attr *attr, const struct lu_buf *buf,
2221 struct lod_thread_info *info = lod_env_info(env);
2222 struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
2223 struct ost_pool inuse_osts = { 0 };
2224 struct ost_pool *inuse = &inuse_osts;
2232 /* no OST available */
2233 /* XXX: should we be waiting a bit to prevent failures during
2234 * cluster initialization? */
2235 if (d->lod_ostnr == 0)
2239 * by this time, the object's ldo_stripe_count and ldo_stripe_size
2240 * contain default value for striping: taken from the parent
2241 * or from filesystem defaults
2243 * in case the caller is passing lovea with new striping config,
2244 * we may need to parse lovea and apply new configuration
2246 rc = lod_qos_parse_config(env, lo, buf);
2250 if (attr->la_valid & LA_SIZE)
2251 size = attr->la_size;
2253 /* only prepare inuse if multiple components to be created */
2254 if (size && lo->ldo_is_composite) {
2255 rc = lod_prepare_inuse(env, lo);
2258 inuse = &info->lti_inuse_osts;
2262 * prepare OST object creation for the component covering file's
2263 * size, the 1st component (including plain layout file) is always
2266 for (i = 0; i < lo->ldo_comp_cnt; i++) {
2267 struct lod_layout_component *lod_comp;
2268 struct lu_extent *extent;
2270 lod_comp = &lo->ldo_comp_entries[i];
2271 extent = &lod_comp->llc_extent;
2272 CDEBUG(D_QOS, "%lld [%lld, %lld)\n",
2273 size, extent->e_start, extent->e_end);
2274 if (!lo->ldo_is_composite || size >= extent->e_start) {
2275 rc = lod_qos_prep_create(env, lo, attr, th, i, inuse);