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 <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);
202 if (rc && rc != -ENOTCONN)
203 CERROR("%s: statfs: rc = %d\n", lod2obd(d)->obd_name, rc);
205 /* If the OST is readonly then we can't allocate objects there */
206 if (sfs->os_state & OS_STATE_READONLY)
209 /* check whether device has changed state (active, inactive) */
210 if (rc != 0 && ost->ltd_active) {
211 /* turned inactive? */
212 spin_lock(&d->lod_desc_lock);
213 if (ost->ltd_active) {
216 ost->ltd_connecting = 1;
218 LASSERT(d->lod_desc.ld_active_tgt_count > 0);
219 d->lod_desc.ld_active_tgt_count--;
220 d->lod_qos.lq_dirty = 1;
221 d->lod_qos.lq_rr.lqr_dirty = 1;
222 CDEBUG(D_CONFIG, "%s: turns inactive\n",
223 ost->ltd_exp->exp_obd->obd_name);
225 spin_unlock(&d->lod_desc_lock);
226 } else if (rc == 0 && ost->ltd_active == 0) {
228 LASSERTF(d->lod_desc.ld_active_tgt_count < d->lod_ostnr,
229 "active tgt count %d, ost nr %d\n",
230 d->lod_desc.ld_active_tgt_count, d->lod_ostnr);
231 spin_lock(&d->lod_desc_lock);
232 if (ost->ltd_active == 0) {
234 ost->ltd_connecting = 0;
235 d->lod_desc.ld_active_tgt_count++;
236 d->lod_qos.lq_dirty = 1;
237 d->lod_qos.lq_rr.lqr_dirty = 1;
238 CDEBUG(D_CONFIG, "%s: turns active\n",
239 ost->ltd_exp->exp_obd->obd_name);
241 spin_unlock(&d->lod_desc_lock);
248 * Maintain per-target statfs data.
250 * The function refreshes statfs data for all the targets every N seconds.
251 * The actual N is controlled via procfs and set to LOV_DESC_QOS_MAXAGE_DEFAULT
254 * \param[in] env execution environment for this thread
255 * \param[in] lod LOD device
257 static void lod_qos_statfs_update(const struct lu_env *env,
258 struct lod_device *lod)
260 struct obd_device *obd = lod2obd(lod);
261 struct ost_pool *osts = &(lod->lod_pool_info);
264 __u64 max_age, avail;
267 max_age = cfs_time_shift_64(-2 * lod->lod_desc.ld_qos_maxage);
269 if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
270 /* statfs data are quite recent, don't need to refresh it */
273 down_write(&lod->lod_qos.lq_rw_sem);
274 if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
277 for (i = 0; i < osts->op_count; i++) {
278 idx = osts->op_array[i];
279 avail = OST_TGT(lod,idx)->ltd_statfs.os_bavail;
280 if (lod_statfs_and_check(env, lod, idx,
281 &OST_TGT(lod, idx)->ltd_statfs))
283 if (OST_TGT(lod,idx)->ltd_statfs.os_bavail != avail)
284 /* recalculate weigths */
285 lod->lod_qos.lq_dirty = 1;
287 obd->obd_osfs_age = cfs_time_current_64();
290 up_write(&lod->lod_qos.lq_rw_sem);
295 * Calculate per-OST and per-OSS penalties
297 * Re-calculate penalties when the configuration changes, active targets
298 * change and after statfs refresh (all these are reflected by lq_dirty flag).
299 * On every OST and OSS: decay the penalty by half for every 8x the update
300 * interval that the device has been idle. That gives lots of time for the
301 * statfs information to be updated (which the penalty is only a proxy for),
302 * and avoids penalizing OSS/OSTs under light load.
303 * See lod_qos_calc_weight() for how penalties are factored into the weight.
305 * \param[in] lod LOD device
307 * \retval 0 on success
308 * \retval -EAGAIN the number of OSTs isn't enough
310 static int lod_qos_calc_ppo(struct lod_device *lod)
312 struct lod_qos_oss *oss;
313 __u64 ba_max, ba_min, temp;
320 if (!lod->lod_qos.lq_dirty)
323 num_active = lod->lod_desc.ld_active_tgt_count - 1;
325 GOTO(out, rc = -EAGAIN);
327 /* find bavail on each OSS */
328 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list)
330 lod->lod_qos.lq_active_oss_count = 0;
333 * How badly user wants to select OSTs "widely" (not recently chosen
334 * and not on recent OSS's). As opposed to "freely" (free space
337 prio_wide = 256 - lod->lod_qos.lq_prio_free;
339 ba_min = (__u64)(-1);
341 now = cfs_time_current_sec();
342 /* Calculate OST penalty per object
343 * (lod ref taken in lod_qos_prep_create()) */
344 cfs_foreach_bit(lod->lod_ost_bitmap, i) {
345 LASSERT(OST_TGT(lod,i));
346 temp = TGT_BAVAIL(i);
349 ba_min = min(temp, ba_min);
350 ba_max = max(temp, ba_max);
352 /* Count the number of usable OSS's */
353 if (OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail == 0)
354 lod->lod_qos.lq_active_oss_count++;
355 OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail += temp;
357 /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
359 do_div(temp, num_active);
360 OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj =
361 (temp * prio_wide) >> 8;
363 age = (now - OST_TGT(lod,i)->ltd_qos.ltq_used) >> 3;
364 if (lod->lod_qos.lq_reset ||
365 age > 32 * lod->lod_desc.ld_qos_maxage)
366 OST_TGT(lod,i)->ltd_qos.ltq_penalty = 0;
367 else if (age > lod->lod_desc.ld_qos_maxage)
368 /* Decay OST penalty. */
369 OST_TGT(lod,i)->ltd_qos.ltq_penalty >>=
370 (age / lod->lod_desc.ld_qos_maxage);
373 num_active = lod->lod_qos.lq_active_oss_count - 1;
374 if (num_active < 1) {
375 /* If there's only 1 OSS, we can't penalize it, so instead
376 we have to double the OST penalty */
378 cfs_foreach_bit(lod->lod_ost_bitmap, i)
379 OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj <<= 1;
382 /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
383 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
384 temp = oss->lqo_bavail >> 1;
385 do_div(temp, oss->lqo_ost_count * num_active);
386 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
388 age = (now - oss->lqo_used) >> 3;
389 if (lod->lod_qos.lq_reset ||
390 age > 32 * lod->lod_desc.ld_qos_maxage)
391 oss->lqo_penalty = 0;
392 else if (age > lod->lod_desc.ld_qos_maxage)
393 /* Decay OSS penalty. */
394 oss->lqo_penalty >>= age / lod->lod_desc.ld_qos_maxage;
397 lod->lod_qos.lq_dirty = 0;
398 lod->lod_qos.lq_reset = 0;
400 /* If each ost has almost same free space,
401 * do rr allocation for better creation performance */
402 lod->lod_qos.lq_same_space = 0;
403 if ((ba_max * (256 - lod->lod_qos.lq_threshold_rr)) >> 8 < ba_min) {
404 lod->lod_qos.lq_same_space = 1;
405 /* Reset weights for the next time we enter qos mode */
406 lod->lod_qos.lq_reset = 1;
412 if (!rc && lod->lod_qos.lq_same_space)
419 * Calculate weight for a given OST target.
421 * The final OST weight is the number of bytes available minus the OST and
422 * OSS penalties. See lod_qos_calc_ppo() for how penalties are calculated.
424 * \param[in] lod LOD device, where OST targets are listed
425 * \param[in] i OST target index
429 static int lod_qos_calc_weight(struct lod_device *lod, int i)
433 temp = TGT_BAVAIL(i);
434 temp2 = OST_TGT(lod,i)->ltd_qos.ltq_penalty +
435 OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_penalty;
437 OST_TGT(lod,i)->ltd_qos.ltq_weight = 0;
439 OST_TGT(lod,i)->ltd_qos.ltq_weight = temp - temp2;
444 * Re-calculate weights.
446 * The function is called when some OST target was used for a new object. In
447 * this case we should re-calculate all the weights to keep new allocations
450 * \param[in] lod LOD device
451 * \param[in] osts OST pool where a new object was placed
452 * \param[in] index OST target where a new object was placed
453 * \param[out] total_wt new total weight for the pool
457 static int lod_qos_used(struct lod_device *lod, struct ost_pool *osts,
458 __u32 index, __u64 *total_wt)
460 struct lod_tgt_desc *ost;
461 struct lod_qos_oss *oss;
465 ost = OST_TGT(lod,index);
468 /* Don't allocate on this devuce anymore, until the next alloc_qos */
469 ost->ltd_qos.ltq_usable = 0;
471 oss = ost->ltd_qos.ltq_oss;
473 /* Decay old penalty by half (we're adding max penalty, and don't
474 want it to run away.) */
475 ost->ltd_qos.ltq_penalty >>= 1;
476 oss->lqo_penalty >>= 1;
478 /* mark the OSS and OST as recently used */
479 ost->ltd_qos.ltq_used = oss->lqo_used = cfs_time_current_sec();
481 /* Set max penalties for this OST and OSS */
482 ost->ltd_qos.ltq_penalty +=
483 ost->ltd_qos.ltq_penalty_per_obj * lod->lod_ostnr;
484 oss->lqo_penalty += oss->lqo_penalty_per_obj *
485 lod->lod_qos.lq_active_oss_count;
487 /* Decrease all OSS penalties */
488 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
489 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
490 oss->lqo_penalty = 0;
492 oss->lqo_penalty -= oss->lqo_penalty_per_obj;
496 /* Decrease all OST penalties */
497 for (j = 0; j < osts->op_count; j++) {
500 i = osts->op_array[j];
501 if (!cfs_bitmap_check(lod->lod_ost_bitmap, i))
504 ost = OST_TGT(lod,i);
507 if (ost->ltd_qos.ltq_penalty <
508 ost->ltd_qos.ltq_penalty_per_obj)
509 ost->ltd_qos.ltq_penalty = 0;
511 ost->ltd_qos.ltq_penalty -=
512 ost->ltd_qos.ltq_penalty_per_obj;
514 lod_qos_calc_weight(lod, i);
516 /* Recalc the total weight of usable osts */
517 if (ost->ltd_qos.ltq_usable)
518 *total_wt += ost->ltd_qos.ltq_weight;
520 QOS_DEBUG("recalc tgt %d usable=%d avail="LPU64
521 " ostppo="LPU64" ostp="LPU64" ossppo="LPU64
522 " ossp="LPU64" wt="LPU64"\n",
523 i, ost->ltd_qos.ltq_usable, TGT_BAVAIL(i) >> 10,
524 ost->ltd_qos.ltq_penalty_per_obj >> 10,
525 ost->ltd_qos.ltq_penalty >> 10,
526 ost->ltd_qos.ltq_oss->lqo_penalty_per_obj >> 10,
527 ost->ltd_qos.ltq_oss->lqo_penalty >> 10,
528 ost->ltd_qos.ltq_weight >> 10);
534 void lod_qos_rr_init(struct lod_qos_rr *lqr)
536 spin_lock_init(&lqr->lqr_alloc);
541 #define LOV_QOS_EMPTY ((__u32)-1)
544 * Calculate optimal round-robin order with regard to OSSes.
546 * Place all the OSTs from pool \a src_pool in a special array to be used for
547 * round-robin (RR) stripe allocation. The placement algorithm interleaves
548 * OSTs from the different OSSs so that RR allocation can balance OSSs evenly.
549 * Resorts the targets when the number of active targets changes (because of
550 * a new target or activation/deactivation).
552 * \param[in] lod LOD device
553 * \param[in] src_pool OST pool
554 * \param[in] lqr round-robin list
556 * \retval 0 on success
557 * \retval -ENOMEM fails to allocate the array
559 static int lod_qos_calc_rr(struct lod_device *lod, struct ost_pool *src_pool,
560 struct lod_qos_rr *lqr)
562 struct lod_qos_oss *oss;
563 struct lod_tgt_desc *ost;
564 unsigned placed, real_count;
569 if (!lqr->lqr_dirty) {
570 LASSERT(lqr->lqr_pool.op_size);
574 /* Do actual allocation. */
575 down_write(&lod->lod_qos.lq_rw_sem);
578 * Check again. While we were sleeping on @lq_rw_sem something could
581 if (!lqr->lqr_dirty) {
582 LASSERT(lqr->lqr_pool.op_size);
583 up_write(&lod->lod_qos.lq_rw_sem);
587 real_count = src_pool->op_count;
589 /* Zero the pool array */
590 /* alloc_rr is holding a read lock on the pool, so nobody is adding/
591 deleting from the pool. The lq_rw_sem insures that nobody else
593 lqr->lqr_pool.op_count = real_count;
594 rc = lod_ost_pool_extend(&lqr->lqr_pool, real_count);
596 up_write(&lod->lod_qos.lq_rw_sem);
599 for (i = 0; i < lqr->lqr_pool.op_count; i++)
600 lqr->lqr_pool.op_array[i] = LOV_QOS_EMPTY;
602 /* Place all the OSTs from 1 OSS at the same time. */
604 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
607 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
610 if (!cfs_bitmap_check(lod->lod_ost_bitmap,
611 src_pool->op_array[i]))
614 ost = OST_TGT(lod,src_pool->op_array[i]);
615 LASSERT(ost && ost->ltd_ost);
616 if (ost->ltd_qos.ltq_oss != oss)
619 /* Evenly space these OSTs across arrayspace */
620 next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count;
621 while (lqr->lqr_pool.op_array[next] != LOV_QOS_EMPTY)
622 next = (next + 1) % lqr->lqr_pool.op_count;
624 lqr->lqr_pool.op_array[next] = src_pool->op_array[i];
631 up_write(&lod->lod_qos.lq_rw_sem);
633 if (placed != real_count) {
634 /* This should never happen */
635 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
636 "round-robin list (%d of %d).\n",
638 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
639 LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
640 lqr->lqr_pool.op_array[i]);
647 for (i = 0; i < lqr->lqr_pool.op_count; i++)
648 QOS_CONSOLE("rr #%d ost idx=%d\n", i, lqr->lqr_pool.op_array[i]);
655 * Instantiate and declare creation of a new object.
657 * The function instantiates LU representation for a new object on the
658 * specified device. Also it declares an intention to create that
659 * object on the storage target.
661 * Note lu_object_anon() is used which is a trick with regard to LU/OSD
662 * infrastructure - in the existing precreation framework we can't assign FID
663 * at this moment, we do this later once a transaction is started. So the
664 * special method instantiates FID-less object in the cache and later it
665 * will get a FID and proper placement in LU cache.
667 * \param[in] env execution environment for this thread
668 * \param[in] d LOD device
669 * \param[in] ost_idx OST target index where the object is being created
670 * \param[in] th transaction handle
672 * \retval object ptr on success, ERR_PTR() otherwise
674 static struct dt_object *lod_qos_declare_object_on(const struct lu_env *env,
675 struct lod_device *d,
679 struct lod_tgt_desc *ost;
680 struct lu_object *o, *n;
681 struct lu_device *nd;
682 struct dt_object *dt;
687 LASSERT(ost_idx < d->lod_osts_size);
688 ost = OST_TGT(d,ost_idx);
690 LASSERT(ost->ltd_ost);
692 nd = &ost->ltd_ost->dd_lu_dev;
695 * allocate anonymous object with zero fid, real fid
696 * will be assigned by OSP within transaction
697 * XXX: to be fixed with fully-functional OST fids
699 o = lu_object_anon(env, nd, NULL);
701 GOTO(out, dt = ERR_PTR(PTR_ERR(o)));
703 n = lu_object_locate(o->lo_header, nd->ld_type);
704 if (unlikely(n == NULL)) {
705 CERROR("can't find slice\n");
706 lu_object_put(env, o);
707 GOTO(out, dt = ERR_PTR(-EINVAL));
710 dt = container_of(n, struct dt_object, do_lu);
712 rc = lod_sub_object_declare_create(env, dt, NULL, NULL, NULL, th);
714 CDEBUG(D_OTHER, "can't declare creation on #%u: %d\n",
716 lu_object_put(env, o);
725 * Calculate a minimum acceptable stripe count.
727 * Return an acceptable stripe count depending on flag LOV_USES_DEFAULT_STRIPE:
728 * all stripes or 3/4 of stripes.
730 * \param[in] stripe_cnt number of stripes requested
731 * \param[in] flags 0 or LOV_USES_DEFAULT_STRIPE
733 * \retval acceptable stripecount
735 static int min_stripe_count(__u32 stripe_cnt, int flags)
737 return (flags & LOV_USES_DEFAULT_STRIPE ?
738 stripe_cnt - (stripe_cnt / 4) : stripe_cnt);
741 #define LOV_CREATE_RESEED_MULT 30
742 #define LOV_CREATE_RESEED_MIN 2000
745 * Check if an OST is full.
747 * Check whether an OST should be considered full based
748 * on the given statfs data.
750 * \param[in] msfs statfs data
752 * \retval false not full
755 static int inline lod_qos_dev_is_full(struct obd_statfs *msfs)
758 int bs = msfs->os_bsize;
760 LASSERT(((bs - 1) & bs) == 0);
762 /* the minimum of 0.1% used blocks and 1GB bytes. */
763 used = min_t(__u64, (msfs->os_blocks - msfs->os_bfree) >> 10,
764 1 << (31 - ffs(bs)));
765 return (msfs->os_bavail < used);
769 * Initialize temporary OST-in-use array.
771 * Allocate or extend the array used to mark targets already assigned to a new
772 * striping so they are not used more than once.
774 * \param[in] env execution environment for this thread
775 * \param[in] stripes number of items needed in the array
777 * \retval 0 on success
778 * \retval -ENOMEM on error
780 static inline int lod_qos_ost_in_use_clear(const struct lu_env *env,
783 struct lod_thread_info *info = lod_env_info(env);
785 if (info->lti_ea_store_size < sizeof(int) * stripes)
786 lod_ea_store_resize(info, stripes * sizeof(int));
787 if (info->lti_ea_store_size < sizeof(int) * stripes) {
788 CERROR("can't allocate memory for ost-in-use array\n");
791 memset(info->lti_ea_store, -1, sizeof(int) * stripes);
796 * Remember a target in the array of used targets.
798 * Mark the given target as used for a new striping being created. The status
799 * of an OST in a striping can be checked with lod_qos_is_ost_used().
801 * \param[in] env execution environment for this thread
802 * \param[in] idx index in the array
803 * \param[in] ost OST target index to mark as used
805 static inline void lod_qos_ost_in_use(const struct lu_env *env,
808 struct lod_thread_info *info = lod_env_info(env);
809 int *osts = info->lti_ea_store;
811 LASSERT(info->lti_ea_store_size >= idx * sizeof(int));
816 * Check is OST used in a striping.
818 * Checks whether OST with the given index is marked as used in the temporary
819 * array (see lod_qos_ost_in_use()).
821 * \param[in] env execution environment for this thread
822 * \param[in] ost OST target index to check
823 * \param[in] stripes the number of items used in the array already
828 static int lod_qos_is_ost_used(const struct lu_env *env, int ost, __u32 stripes)
830 struct lod_thread_info *info = lod_env_info(env);
831 int *osts = info->lti_ea_store;
834 for (j = 0; j < stripes; j++) {
841 static int lod_check_and_reserve_ost(const struct lu_env *env,
842 struct lod_device *m,
843 struct obd_statfs *sfs, __u32 ost_idx,
844 __u32 speed, __u32 *s_idx,
845 struct dt_object **stripe,
849 __u32 stripe_idx = *s_idx;
852 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
854 /* this OSP doesn't feel well */
861 if (lod_qos_dev_is_full(sfs)) {
862 QOS_DEBUG("#%d is full\n", ost_idx);
867 * We expect number of precreated objects in f_ffree at
868 * the first iteration, skip OSPs with no objects ready
870 if (sfs->os_fprecreated == 0 && speed == 0) {
871 QOS_DEBUG("#%d: precreation is empty\n", ost_idx);
876 * try to use another OSP if this one is degraded
878 if (sfs->os_state & OS_STATE_DEGRADED && speed < 2) {
879 QOS_DEBUG("#%d: degraded\n", ost_idx);
884 * do not put >1 objects on a single OST
886 if (speed && lod_qos_is_ost_used(env, ost_idx, stripe_idx))
889 o = lod_qos_declare_object_on(env, m, ost_idx, th);
891 CDEBUG(D_OTHER, "can't declare new object on #%u: %d\n",
892 ost_idx, (int) PTR_ERR(o));
898 * We've successfully declared (reserved) an object
900 lod_qos_ost_in_use(env, stripe_idx, ost_idx);
901 stripe[stripe_idx] = o;
910 * Allocate a striping using round-robin algorithm.
912 * Allocates a new striping using round-robin algorithm. The function refreshes
913 * all the internal structures (statfs cache, array of available OSTs sorted
914 * with regard to OSS, etc). The number of stripes required is taken from the
915 * object (must be prepared by the caller), but can change if the flag
916 * LOV_USES_DEFAULT_STRIPE is supplied. The caller should ensure nobody else
917 * is trying to create a striping on the object in parallel. All the internal
918 * structures (like pools, etc) are protected and no additional locking is
919 * required. The function succeeds even if a single stripe is allocated. To save
920 * time we give priority to targets which already have objects precreated.
921 * Full OSTs are skipped (see lod_qos_dev_is_full() for the details).
923 * \param[in] env execution environment for this thread
924 * \param[in] lo LOD object
925 * \param[out] stripe striping created
926 * \param[in] flags allocation flags (0 or LOV_USES_DEFAULT_STRIPE)
927 * \param[in] th transaction handle
929 * \retval 0 on success
930 * \retval -ENOSPC if not enough OSTs are found
931 * \retval negative negated errno for other failures
933 static int lod_alloc_rr(const struct lu_env *env, struct lod_object *lo,
934 struct dt_object **stripe, int flags,
937 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
938 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
939 struct pool_desc *pool = NULL;
940 struct ost_pool *osts;
941 struct lod_qos_rr *lqr;
942 unsigned int i, array_idx;
944 __u32 ost_start_idx_temp;
946 int ost_connecting = 0;
947 __u32 stripe_idx = 0;
948 __u32 stripe_cnt = lo->ldo_stripenr;
949 __u32 stripe_cnt_min = min_stripe_count(stripe_cnt, flags);
954 pool = lod_find_pool(m, lo->ldo_pool);
957 down_read(&pool_tgt_rw_sem(pool));
958 osts = &(pool->pool_obds);
959 lqr = &(pool->pool_rr);
961 osts = &(m->lod_pool_info);
962 lqr = &(m->lod_qos.lq_rr);
965 rc = lod_qos_calc_rr(m, osts, lqr);
969 rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
973 down_read(&m->lod_qos.lq_rw_sem);
974 spin_lock(&lqr->lqr_alloc);
975 if (--lqr->lqr_start_count <= 0) {
976 lqr->lqr_start_idx = cfs_rand() % osts->op_count;
977 lqr->lqr_start_count =
978 (LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) +
979 LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U);
980 } else if (stripe_cnt_min >= osts->op_count ||
981 lqr->lqr_start_idx > osts->op_count) {
982 /* If we have allocated from all of the OSTs, slowly
983 * precess the next start if the OST/stripe count isn't
984 * already doing this for us. */
985 lqr->lqr_start_idx %= osts->op_count;
986 if (stripe_cnt > 1 && (osts->op_count % stripe_cnt) != 1)
987 ++lqr->lqr_offset_idx;
989 ost_start_idx_temp = lqr->lqr_start_idx;
993 QOS_DEBUG("pool '%s' want %d startidx %d startcnt %d offset %d "
994 "active %d count %d\n",
995 lo->ldo_pool ? lo->ldo_pool : "",
996 stripe_cnt, lqr->lqr_start_idx, lqr->lqr_start_count,
997 lqr->lqr_offset_idx, osts->op_count, osts->op_count);
999 for (i = 0; i < osts->op_count && stripe_idx < lo->ldo_stripenr; i++) {
1000 array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) %
1002 ++lqr->lqr_start_idx;
1003 ost_idx = lqr->lqr_pool.op_array[array_idx];
1005 QOS_DEBUG("#%d strt %d act %d strp %d ary %d idx %d\n",
1006 i, lqr->lqr_start_idx, /* XXX: active*/ 0,
1007 stripe_idx, array_idx, ost_idx);
1009 if ((ost_idx == LOV_QOS_EMPTY) ||
1010 !cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
1013 /* Fail Check before osc_precreate() is called
1014 so we can only 'fail' single OSC. */
1015 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
1018 spin_unlock(&lqr->lqr_alloc);
1019 rc = lod_check_and_reserve_ost(env, m, sfs, ost_idx, speed,
1020 &stripe_idx, stripe, th);
1021 spin_lock(&lqr->lqr_alloc);
1023 if (rc != 0 && OST_TGT(m, ost_idx)->ltd_connecting)
1026 if ((speed < 2) && (stripe_idx < stripe_cnt_min)) {
1027 /* Try again, allowing slower OSCs */
1029 lqr->lqr_start_idx = ost_start_idx_temp;
1035 spin_unlock(&lqr->lqr_alloc);
1036 up_read(&m->lod_qos.lq_rw_sem);
1039 lo->ldo_stripenr = stripe_idx;
1040 /* at least one stripe is allocated */
1043 /* nobody provided us with a single object */
1052 up_read(&pool_tgt_rw_sem(pool));
1053 /* put back ref got by lod_find_pool() */
1054 lod_pool_putref(pool);
1061 * Allocate a specific striping layout on a user defined set of OSTs.
1063 * Allocates new striping using the OST index range provided by the data from
1064 * the lmm_obejcts contained in the lov_user_md passed to this method. Full
1065 * OSTs are not considered. The exact order of OSTs requested by the user
1066 * is respected as much as possible depending on OST status. The number of
1067 * stripes needed and stripe offset are taken from the object. If that number
1068 * can not be met, then the function returns a failure and then it's the
1069 * caller's responsibility to release the stripes allocated. All the internal
1070 * structures are protected, but no concurrent allocation is allowed on the
1073 * \param[in] env execution environment for this thread
1074 * \param[in] lo LOD object
1075 * \param[out] stripe striping created
1076 * \param[in] lum stripe md to specify list of OSTs
1077 * \param[in] th transaction handle
1079 * \retval 0 on success
1080 * \retval -ENODEV OST index does not exist on file system
1081 * \retval -EINVAL requested OST index is invalid
1082 * \retval negative negated errno on error
1084 static int lod_alloc_ost_list(const struct lu_env *env,
1085 struct lod_object *lo, struct dt_object **stripe,
1086 struct lov_user_md *lum, struct thandle *th)
1088 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1089 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
1090 struct dt_object *o;
1091 struct lov_user_md_v3 *v3;
1092 unsigned int array_idx = 0;
1093 int stripe_count = 0;
1098 /* for specific OSTs layout */
1099 LASSERT(lum != NULL && lum->lmm_magic == LOV_USER_MAGIC_SPECIFIC);
1100 lustre_print_user_md(D_OTHER, lum, __func__);
1102 rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
1106 v3 = (struct lov_user_md_v3 *)lum;
1107 for (i = 0; i < lo->ldo_stripenr; i++) {
1108 if (v3->lmm_objects[i].l_ost_idx == lo->ldo_def_stripe_offset) {
1113 if (i == lo->ldo_stripenr) {
1115 "%s: start index %d not in the specified list of OSTs\n",
1116 lod2obd(m)->obd_name, lo->ldo_def_stripe_offset);
1120 for (i = 0; i < lo->ldo_stripenr;
1121 i++, array_idx = (array_idx + 1) % lo->ldo_stripenr) {
1122 __u32 ost_idx = v3->lmm_objects[array_idx].l_ost_idx;
1124 if (!cfs_bitmap_check(m->lod_ost_bitmap, ost_idx)) {
1130 * do not put >1 objects on a single OST
1132 if (lod_qos_is_ost_used(env, ost_idx, stripe_count)) {
1137 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
1138 if (rc < 0) /* this OSP doesn't feel well */
1141 o = lod_qos_declare_object_on(env, m, ost_idx, th);
1145 "%s: can't declare new object on #%u: %d\n",
1146 lod2obd(m)->obd_name, ost_idx, rc);
1151 * We've successfully declared (reserved) an object
1153 lod_qos_ost_in_use(env, stripe_count, ost_idx);
1154 stripe[stripe_count] = o;
1162 * Allocate a striping on a predefined set of OSTs.
1164 * Allocates new layout starting from OST index in lo->ldo_def_stripe_offset.
1165 * Full OSTs are not considered. The exact order of OSTs is not important and
1166 * varies depending on OST status. The allocation procedure prefers the targets
1167 * with precreated objects ready. The number of stripes needed and stripe
1168 * offset are taken from the object. If that number cannot be met, then the
1169 * function returns an error and then it's the caller's responsibility to
1170 * release the stripes allocated. All the internal structures are protected,
1171 * but no concurrent allocation is allowed on the same objects.
1173 * \param[in] env execution environment for this thread
1174 * \param[in] lo LOD object
1175 * \param[out] stripe striping created
1176 * \param[in] flags not used
1177 * \param[in] th transaction handle
1179 * \retval 0 on success
1180 * \retval -ENOSPC if no OST objects are available at all
1181 * \retval -EFBIG if not enough OST objects are found
1182 * \retval -EINVAL requested offset is invalid
1183 * \retval negative errno on failure
1185 static int lod_alloc_specific(const struct lu_env *env, struct lod_object *lo,
1186 struct dt_object **stripe, int flags,
1189 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1190 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
1191 struct dt_object *o;
1193 unsigned int i, array_idx, ost_count;
1194 int rc, stripe_num = 0;
1196 struct pool_desc *pool = NULL;
1197 struct ost_pool *osts;
1200 rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
1205 pool = lod_find_pool(m, lo->ldo_pool);
1208 down_read(&pool_tgt_rw_sem(pool));
1209 osts = &(pool->pool_obds);
1211 osts = &(m->lod_pool_info);
1214 ost_count = osts->op_count;
1217 /* search loi_ost_idx in ost array */
1219 for (i = 0; i < ost_count; i++) {
1220 if (osts->op_array[i] == lo->ldo_def_stripe_offset) {
1225 if (i == ost_count) {
1226 CERROR("Start index %d not found in pool '%s'\n",
1227 lo->ldo_def_stripe_offset,
1228 lo->ldo_pool ? lo->ldo_pool : "");
1229 GOTO(out, rc = -EINVAL);
1232 for (i = 0; i < ost_count;
1233 i++, array_idx = (array_idx + 1) % ost_count) {
1234 ost_idx = osts->op_array[array_idx];
1236 if (!cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
1239 /* Fail Check before osc_precreate() is called
1240 so we can only 'fail' single OSC. */
1241 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
1245 * do not put >1 objects on a single OST
1247 if (lod_qos_is_ost_used(env, ost_idx, stripe_num))
1250 /* Drop slow OSCs if we can, but not for requested start idx.
1252 * This means "if OSC is slow and it is not the requested
1253 * start OST, then it can be skipped, otherwise skip it only
1254 * if it is inactive/recovering/out-of-space." */
1256 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
1258 /* this OSP doesn't feel well */
1263 * We expect number of precreated objects in f_ffree at
1264 * the first iteration, skip OSPs with no objects ready
1265 * don't apply this logic to OST specified with stripe_offset
1267 if (i != 0 && sfs->os_fprecreated == 0 && speed == 0)
1270 o = lod_qos_declare_object_on(env, m, ost_idx, th);
1272 CDEBUG(D_OTHER, "can't declare new object on #%u: %d\n",
1273 ost_idx, (int) PTR_ERR(o));
1278 * We've successfully declared (reserved) an object
1280 lod_qos_ost_in_use(env, stripe_num, ost_idx);
1281 stripe[stripe_num] = o;
1284 /* We have enough stripes */
1285 if (stripe_num == lo->ldo_stripenr)
1289 /* Try again, allowing slower OSCs */
1294 /* If we were passed specific striping params, then a failure to
1295 * meet those requirements is an error, since we can't reallocate
1296 * that memory (it might be part of a larger array or something).
1298 CERROR("can't lstripe objid "DFID": have %d want %u\n",
1299 PFID(lu_object_fid(lod2lu_obj(lo))), stripe_num,
1301 rc = stripe_num == 0 ? -ENOSPC : -EFBIG;
1304 up_read(&pool_tgt_rw_sem(pool));
1305 /* put back ref got by lod_find_pool() */
1306 lod_pool_putref(pool);
1313 * Check whether QoS allocation should be used.
1315 * A simple helper to decide when QoS allocation should be used:
1316 * if it's just a single available target or the used space is
1317 * evenly distributed among the targets at the moment, then QoS
1318 * allocation algorithm should not be used.
1320 * \param[in] lod LOD device
1322 * \retval 0 should not be used
1323 * \retval 1 should be used
1325 static inline int lod_qos_is_usable(struct lod_device *lod)
1328 /* to be able to debug QoS code */
1332 /* Detect -EAGAIN early, before expensive lock is taken. */
1333 if (!lod->lod_qos.lq_dirty && lod->lod_qos.lq_same_space)
1336 if (lod->lod_desc.ld_active_tgt_count < 2)
1343 * Allocate a striping using an algorithm with weights.
1345 * The function allocates OST objects to create a striping. The algorithm
1346 * used is based on weights (currently only using the free space), and it's
1347 * trying to ensure the space is used evenly by OSTs and OSSs. The striping
1348 * configuration (# of stripes, offset, pool) is taken from the object and
1349 * is prepared by the caller.
1351 * If LOV_USES_DEFAULT_STRIPE is not passed and prepared configuration can't
1352 * be met due to too few OSTs, then allocation fails. If the flag is passed
1353 * fewer than 3/4 of the requested number of stripes can be allocated, then
1356 * No concurrent allocation is allowed on the object and this must be ensured
1357 * by the caller. All the internal structures are protected by the function.
1359 * The algorithm has two steps: find available OSTs and calculate their
1360 * weights, then select the OSTs with their weights used as the probability.
1361 * An OST with a higher weight is proportionately more likely to be selected
1362 * than one with a lower weight.
1364 * \param[in] env execution environment for this thread
1365 * \param[in] lo LOD object
1366 * \param[out] stripe striping created
1367 * \param[in] flags 0 or LOV_USES_DEFAULT_STRIPE
1368 * \param[in] th transaction handle
1370 * \retval 0 on success
1371 * \retval -EAGAIN not enough OSTs are found for specified stripe count
1372 * \retval -EINVAL requested OST index is invalid
1373 * \retval negative errno on failure
1375 static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo,
1376 struct dt_object **stripe, int flags,
1379 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1380 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
1381 struct lod_tgt_desc *ost;
1382 struct dt_object *o;
1383 __u64 total_weight = 0;
1386 __u32 nfound, good_osts;
1387 __u32 stripe_cnt = lo->ldo_stripenr;
1388 __u32 stripe_cnt_min;
1389 struct pool_desc *pool = NULL;
1390 struct ost_pool *osts;
1393 stripe_cnt_min = min_stripe_count(stripe_cnt, flags);
1394 if (stripe_cnt_min < 1)
1398 pool = lod_find_pool(m, lo->ldo_pool);
1401 down_read(&pool_tgt_rw_sem(pool));
1402 osts = &(pool->pool_obds);
1404 osts = &(m->lod_pool_info);
1407 /* Detect -EAGAIN early, before expensive lock is taken. */
1408 if (!lod_qos_is_usable(m))
1409 GOTO(out_nolock, rc = -EAGAIN);
1411 /* Do actual allocation, use write lock here. */
1412 down_write(&m->lod_qos.lq_rw_sem);
1415 * Check again, while we were sleeping on @lq_rw_sem things could
1418 if (!lod_qos_is_usable(m))
1419 GOTO(out, rc = -EAGAIN);
1421 rc = lod_qos_calc_ppo(m);
1425 rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
1430 /* Find all the OSTs that are valid stripe candidates */
1431 for (i = 0; i < osts->op_count; i++) {
1432 if (!cfs_bitmap_check(m->lod_ost_bitmap, osts->op_array[i]))
1435 rc = lod_statfs_and_check(env, m, osts->op_array[i], sfs);
1437 /* this OSP doesn't feel well */
1444 if (lod_qos_dev_is_full(sfs))
1447 /* Fail Check before osc_precreate() is called
1448 so we can only 'fail' single OSC. */
1449 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) &&
1450 osts->op_array[i] == 0)
1453 ost = OST_TGT(m,osts->op_array[i]);
1454 ost->ltd_qos.ltq_usable = 1;
1455 lod_qos_calc_weight(m, osts->op_array[i]);
1456 total_weight += ost->ltd_qos.ltq_weight;
1461 QOS_DEBUG("found %d good osts\n", good_osts);
1463 if (good_osts < stripe_cnt_min)
1464 GOTO(out, rc = -EAGAIN);
1466 /* We have enough osts */
1467 if (good_osts < stripe_cnt)
1468 stripe_cnt = good_osts;
1470 /* Find enough OSTs with weighted random allocation. */
1472 while (nfound < stripe_cnt) {
1473 __u64 rand, cur_weight;
1479 #if BITS_PER_LONG == 32
1480 rand = cfs_rand() % (unsigned)total_weight;
1481 /* If total_weight > 32-bit, first generate the high
1482 * 32 bits of the random number, then add in the low
1483 * 32 bits (truncated to the upper limit, if needed) */
1484 if (total_weight > 0xffffffffULL)
1485 rand = (__u64)(cfs_rand() %
1486 (unsigned)(total_weight >> 32)) << 32;
1490 if (rand == (total_weight & 0xffffffff00000000ULL))
1491 rand |= cfs_rand() % (unsigned)total_weight;
1496 rand = ((__u64)cfs_rand() << 32 | cfs_rand()) %
1503 /* On average, this will hit larger-weighted OSTs more often.
1504 * 0-weight OSTs will always get used last (only when rand=0) */
1505 for (i = 0; i < osts->op_count; i++) {
1506 __u32 idx = osts->op_array[i];
1508 if (!cfs_bitmap_check(m->lod_ost_bitmap, idx))
1511 ost = OST_TGT(m,idx);
1513 if (!ost->ltd_qos.ltq_usable)
1516 cur_weight += ost->ltd_qos.ltq_weight;
1517 QOS_DEBUG("stripe_cnt=%d nfound=%d cur_weight="LPU64
1518 " rand="LPU64" total_weight="LPU64"\n",
1519 stripe_cnt, nfound, cur_weight, rand,
1522 if (cur_weight < rand)
1525 QOS_DEBUG("stripe=%d to idx=%d\n", nfound, idx);
1528 * do not put >1 objects on a single OST
1530 if (lod_qos_is_ost_used(env, idx, nfound))
1532 lod_qos_ost_in_use(env, nfound, idx);
1534 o = lod_qos_declare_object_on(env, m, idx, th);
1536 QOS_DEBUG("can't declare object on #%u: %d\n",
1537 idx, (int) PTR_ERR(o));
1540 stripe[nfound++] = o;
1541 lod_qos_used(m, osts, idx, &total_weight);
1547 /* no OST found on this iteration, give up */
1552 if (unlikely(nfound != stripe_cnt)) {
1554 * when the decision to use weighted algorithm was made
1555 * we had enough appropriate OSPs, but this state can
1556 * change anytime (no space on OST, broken connection, etc)
1557 * so it's possible OSP won't be able to provide us with
1558 * an object due to just changed state
1560 LCONSOLE_INFO("wanted %d, found %d\n", stripe_cnt, nfound);
1561 for (i = 0; i < nfound; i++) {
1562 LASSERT(stripe[i] != NULL);
1563 lu_object_put(env, &stripe[i]->do_lu);
1567 /* makes sense to rebalance next time */
1568 m->lod_qos.lq_dirty = 1;
1569 m->lod_qos.lq_same_space = 0;
1575 up_write(&m->lod_qos.lq_rw_sem);
1579 up_read(&pool_tgt_rw_sem(pool));
1580 /* put back ref got by lod_find_pool() */
1581 lod_pool_putref(pool);
1588 * Find largest stripe count the caller can use.
1590 * Find the maximal possible stripe count not greater than \a stripe_count.
1591 * Sometimes suggested stripecount can't be reached for a number of reasons:
1592 * lack of enough active OSTs or the backend does not support EAs that large.
1593 * If the passed one is 0, then the filesystem's default one is used.
1595 * \param[in] lod LOD device
1596 * \param[in] magic the format if striping
1597 * \param[in] stripe_count count the caller would like to use
1599 * \retval the maximum usable stripe count
1601 static __u16 lod_get_stripecnt(struct lod_device *lod, __u32 magic,
1604 __u32 max_stripes = LOV_MAX_STRIPE_COUNT_OLD;
1607 stripe_count = lod->lod_desc.ld_default_stripe_count;
1608 if (stripe_count > lod->lod_desc.ld_active_tgt_count)
1609 stripe_count = lod->lod_desc.ld_active_tgt_count;
1613 /* stripe count is based on whether OSD can handle larger EA sizes */
1614 if (lod->lod_osd_max_easize > 0)
1615 max_stripes = lov_mds_md_max_stripe_count(
1616 lod->lod_osd_max_easize, magic);
1618 return (stripe_count < max_stripes) ? stripe_count : max_stripes;
1622 * Create in-core respresentation for a fully-defined striping
1624 * When the caller passes a fully-defined striping (i.e. everything including
1625 * OST object FIDs are defined), then we still need to instantiate LU-cache
1626 * with the objects representing the stripes defined. This function completes
1629 * \param[in] env execution environment for this thread
1630 * \param[in] mo LOD object
1631 * \param[in] buf buffer containing the striping
1633 * \retval 0 on success
1634 * \retval negative negated errno on error
1636 static int lod_use_defined_striping(const struct lu_env *env,
1637 struct lod_object *mo,
1638 const struct lu_buf *buf)
1640 struct lov_mds_md_v1 *v1 = buf->lb_buf;
1641 struct lov_mds_md_v3 *v3 = buf->lb_buf;
1642 struct lov_ost_data_v1 *objs;
1647 magic = le32_to_cpu(v1->lmm_magic);
1648 if (magic == LOV_MAGIC_V1_DEF) {
1649 magic = LOV_MAGIC_V1;
1650 objs = &v1->lmm_objects[0];
1651 } else if (magic == LOV_MAGIC_V3_DEF) {
1652 magic = LOV_MAGIC_V3;
1653 objs = &v3->lmm_objects[0];
1654 lod_object_set_pool(mo, v3->lmm_pool_name);
1656 GOTO(out, rc = -EINVAL);
1659 mo->ldo_pattern = le32_to_cpu(v1->lmm_pattern);
1660 mo->ldo_stripe_size = le32_to_cpu(v1->lmm_stripe_size);
1661 mo->ldo_stripenr = le16_to_cpu(v1->lmm_stripe_count);
1662 mo->ldo_layout_gen = le16_to_cpu(v1->lmm_layout_gen);
1664 /* fixup for released file before object initialization */
1665 if (mo->ldo_pattern & LOV_PATTERN_F_RELEASED) {
1666 mo->ldo_released_stripenr = mo->ldo_stripenr;
1667 mo->ldo_stripenr = 0;
1670 LASSERT(buf->lb_len >= lov_mds_md_size(mo->ldo_stripenr, magic));
1672 if (mo->ldo_stripenr > 0)
1673 rc = lod_initialize_objects(env, mo, objs);
1680 * Parse suggested striping configuration.
1682 * The caller gets a suggested striping configuration from a number of sources
1683 * including per-directory default and applications. Then it needs to verify
1684 * the suggested striping is valid, apply missing bits and store the resulting
1685 * configuration in the object to be used by the allocator later. Must not be
1686 * called concurrently against the same object. It's OK to provide a
1687 * fully-defined striping.
1689 * \param[in] env execution environment for this thread
1690 * \param[in] lo LOD object
1691 * \param[in] buf buffer containing the striping
1693 * \retval 0 on success
1694 * \retval negative negated errno on error
1696 static int lod_qos_parse_config(const struct lu_env *env,
1697 struct lod_object *lo,
1698 const struct lu_buf *buf)
1700 struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
1701 struct lov_user_md_v1 *v1 = NULL;
1702 struct lov_user_md_v3 *v3 = NULL;
1703 char *pool_name = NULL;
1709 if (buf == NULL || buf->lb_buf == NULL || buf->lb_len == 0)
1714 magic = v1->lmm_magic;
1716 if (unlikely(magic == LOV_MAGIC_V1_DEF || magic == LOV_MAGIC_V3_DEF)) {
1717 /* try to use as fully defined striping */
1718 rc = lod_use_defined_striping(env, lo, buf);
1723 case __swab32(LOV_USER_MAGIC_V1):
1724 lustre_swab_lov_user_md_v1(v1);
1725 magic = v1->lmm_magic;
1727 case LOV_USER_MAGIC_V1:
1731 case __swab32(LOV_USER_MAGIC_V3):
1732 lustre_swab_lov_user_md_v3(v3);
1733 magic = v3->lmm_magic;
1735 case LOV_USER_MAGIC_V3:
1737 pool_name = v3->lmm_pool_name;
1740 case __swab32(LOV_USER_MAGIC_SPECIFIC):
1741 lustre_swab_lov_user_md_v3(v3);
1742 lustre_swab_lov_user_md_objects(v3->lmm_objects,
1743 v3->lmm_stripe_count);
1744 magic = v3->lmm_magic;
1746 case LOV_USER_MAGIC_SPECIFIC:
1747 if (v3->lmm_stripe_offset == LOV_OFFSET_DEFAULT)
1748 v3->lmm_stripe_offset = v3->lmm_objects[0].l_ost_idx;
1749 if (v3->lmm_pool_name[0] != '\0')
1750 pool_name = v3->lmm_pool_name;
1751 size = lov_user_md_size(v3->lmm_stripe_count,
1752 LOV_USER_MAGIC_SPECIFIC);
1756 CERROR("%s: unrecognized magic %X\n",
1757 lod2obd(d)->obd_name, magic);
1761 if (unlikely(buf->lb_len < size)) {
1762 CERROR("%s: wrong size: %zd, expect: %u\n",
1763 lod2obd(d)->obd_name, buf->lb_len, size);
1767 lustre_print_user_md(D_OTHER, v1, "parse config");
1769 v1->lmm_magic = magic;
1770 if (v1->lmm_pattern == 0)
1771 v1->lmm_pattern = LOV_PATTERN_RAID0;
1772 if (lov_pattern(v1->lmm_pattern) != LOV_PATTERN_RAID0) {
1773 CERROR("%s: invalid pattern: %x\n",
1774 lod2obd(d)->obd_name, v1->lmm_pattern);
1777 lo->ldo_pattern = v1->lmm_pattern;
1779 if (v1->lmm_stripe_size > 0)
1780 lo->ldo_stripe_size = v1->lmm_stripe_size;
1782 if (lo->ldo_stripe_size & (LOV_MIN_STRIPE_SIZE - 1))
1783 lo->ldo_stripe_size = LOV_MIN_STRIPE_SIZE;
1785 if (v1->lmm_stripe_count > 0)
1786 lo->ldo_stripenr = v1->lmm_stripe_count;
1788 lo->ldo_def_stripe_offset = v1->lmm_stripe_offset;
1790 lod_object_set_pool(lo, NULL);
1791 if (pool_name != NULL) {
1792 struct pool_desc *pool;
1794 /* In the function below, .hs_keycmp resolves to
1795 * pool_hashkey_keycmp() */
1796 /* coverity[overrun-buffer-val] */
1797 pool = lod_find_pool(d, pool_name);
1799 if (lo->ldo_def_stripe_offset != LOV_OFFSET_DEFAULT) {
1800 rc = lod_check_index_in_pool(
1801 lo->ldo_def_stripe_offset, pool);
1803 lod_pool_putref(pool);
1804 CERROR("%s: invalid offset, %u\n",
1805 lod2obd(d)->obd_name,
1806 lo->ldo_def_stripe_offset);
1811 if (lo->ldo_stripenr > pool_tgt_count(pool))
1812 lo->ldo_stripenr = pool_tgt_count(pool);
1814 lod_pool_putref(pool);
1817 lod_object_set_pool(lo, pool_name);
1820 /* fixup for released file */
1821 if (lo->ldo_pattern & LOV_PATTERN_F_RELEASED) {
1822 lo->ldo_released_stripenr = lo->ldo_stripenr;
1823 lo->ldo_stripenr = 0;
1830 * Create a striping for an obejct.
1832 * The function creates a new striping for the object. A buffer containing
1833 * configuration hints can be provided optionally. The function tries QoS
1834 * algorithm first unless free space is distributed evenly among OSTs, but
1835 * by default RR algorithm is preferred due to internal concurrency (QoS is
1836 * serialized). The caller must ensure no concurrent calls to the function
1837 * are made against the same object.
1839 * \param[in] env execution environment for this thread
1840 * \param[in] lo LOD object
1841 * \param[in] attr attributes OST objects will be declared with
1842 * \param[in] buf suggested striping configuration or NULL
1843 * \param[in] th transaction handle
1845 * \retval 0 on success
1846 * \retval negative negated errno on error
1848 int lod_qos_prep_create(const struct lu_env *env, struct lod_object *lo,
1849 struct lu_attr *attr, const struct lu_buf *buf,
1852 struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
1853 struct dt_object **stripe;
1855 int flag = LOV_USES_ASSIGNED_STRIPE;
1861 /* no OST available */
1862 /* XXX: should we be waiting a bit to prevent failures during
1863 * cluster initialization? */
1864 if (d->lod_ostnr == 0)
1865 GOTO(out, rc = -EIO);
1868 * by this time, the object's ldo_stripenr and ldo_stripe_size
1869 * contain default value for striping: taken from the parent
1870 * or from filesystem defaults
1872 * in case the caller is passing lovea with new striping config,
1873 * we may need to parse lovea and apply new configuration
1875 rc = lod_qos_parse_config(env, lo, buf);
1879 /* A released file is being created */
1880 if (lo->ldo_stripenr == 0)
1883 if (likely(lo->ldo_stripe == NULL)) {
1884 struct lov_user_md *lum = NULL;
1887 * no striping has been created so far
1889 LASSERT(lo->ldo_stripenr > 0);
1891 * statfs and check OST targets now, since ld_active_tgt_count
1892 * could be changed if some OSTs are [de]activated manually.
1894 lod_qos_statfs_update(env, d);
1895 lo->ldo_stripenr = lod_get_stripecnt(d, LOV_MAGIC,
1898 stripe_len = lo->ldo_stripenr;
1899 OBD_ALLOC(stripe, sizeof(stripe[0]) * stripe_len);
1901 GOTO(out, rc = -ENOMEM);
1903 lod_getref(&d->lod_ost_descs);
1904 /* XXX: support for non-0 files w/o objects */
1905 CDEBUG(D_OTHER, "tgt_count %d stripenr %d\n",
1906 d->lod_desc.ld_tgt_count, stripe_len);
1908 if (buf != NULL && buf->lb_buf != NULL)
1911 if (lum != NULL && lum->lmm_magic == LOV_USER_MAGIC_SPECIFIC) {
1912 rc = lod_alloc_ost_list(env, lo, stripe, lum, th);
1913 } else if (lo->ldo_def_stripe_offset == LOV_OFFSET_DEFAULT) {
1914 rc = lod_alloc_qos(env, lo, stripe, flag, th);
1916 rc = lod_alloc_rr(env, lo, stripe, flag, th);
1918 rc = lod_alloc_specific(env, lo, stripe, flag, th);
1920 lod_putref(d, &d->lod_ost_descs);
1923 for (i = 0; i < stripe_len; i++)
1924 if (stripe[i] != NULL)
1925 lu_object_put(env, &stripe[i]->do_lu);
1927 OBD_FREE(stripe, sizeof(stripe[0]) * stripe_len);
1928 lo->ldo_stripenr = 0;
1930 lo->ldo_stripe = stripe;
1931 lo->ldo_stripes_allocated = stripe_len;
1935 * lod_qos_parse_config() found supplied buf as a predefined
1936 * striping (not a hint), so it allocated all the object
1937 * now we need to create them
1939 for (i = 0; i < lo->ldo_stripenr; i++) {
1940 struct dt_object *o;
1942 o = lo->ldo_stripe[i];
1945 rc = lod_sub_object_declare_create(env, o, attr, NULL,
1948 CERROR("can't declare create: %d\n", rc);