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) {
215 LASSERT(d->lod_desc.ld_active_tgt_count > 0);
216 d->lod_desc.ld_active_tgt_count--;
217 d->lod_qos.lq_dirty = 1;
218 d->lod_qos.lq_rr.lqr_dirty = 1;
219 CDEBUG(D_CONFIG, "%s: turns inactive\n",
220 ost->ltd_exp->exp_obd->obd_name);
222 spin_unlock(&d->lod_desc_lock);
223 } else if (rc == 0 && ost->ltd_active == 0) {
225 LASSERTF(d->lod_desc.ld_active_tgt_count < d->lod_ostnr,
226 "active tgt count %d, ost nr %d\n",
227 d->lod_desc.ld_active_tgt_count, d->lod_ostnr);
228 spin_lock(&d->lod_desc_lock);
229 if (ost->ltd_active == 0) {
231 d->lod_desc.ld_active_tgt_count++;
232 d->lod_qos.lq_dirty = 1;
233 d->lod_qos.lq_rr.lqr_dirty = 1;
234 CDEBUG(D_CONFIG, "%s: turns active\n",
235 ost->ltd_exp->exp_obd->obd_name);
237 spin_unlock(&d->lod_desc_lock);
244 * Maintain per-target statfs data.
246 * The function refreshes statfs data for all the targets every N seconds.
247 * The actual N is controlled via procfs and set to LOV_DESC_QOS_MAXAGE_DEFAULT
250 * \param[in] env execution environment for this thread
251 * \param[in] lod LOD device
253 static void lod_qos_statfs_update(const struct lu_env *env,
254 struct lod_device *lod)
256 struct obd_device *obd = lod2obd(lod);
257 struct ost_pool *osts = &(lod->lod_pool_info);
260 __u64 max_age, avail;
263 max_age = cfs_time_shift_64(-2 * lod->lod_desc.ld_qos_maxage);
265 if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
266 /* statfs data are quite recent, don't need to refresh it */
269 down_write(&lod->lod_qos.lq_rw_sem);
270 if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
273 for (i = 0; i < osts->op_count; i++) {
274 idx = osts->op_array[i];
275 avail = OST_TGT(lod,idx)->ltd_statfs.os_bavail;
276 if (lod_statfs_and_check(env, lod, idx,
277 &OST_TGT(lod, idx)->ltd_statfs))
279 if (OST_TGT(lod,idx)->ltd_statfs.os_bavail != avail)
280 /* recalculate weigths */
281 lod->lod_qos.lq_dirty = 1;
283 obd->obd_osfs_age = cfs_time_current_64();
286 up_write(&lod->lod_qos.lq_rw_sem);
291 * Calculate per-OST and per-OSS penalties
293 * Re-calculate penalties when the configuration changes, active targets
294 * change and after statfs refresh (all these are reflected by lq_dirty flag).
295 * On every OST and OSS: decay the penalty by half for every 8x the update
296 * interval that the device has been idle. That gives lots of time for the
297 * statfs information to be updated (which the penalty is only a proxy for),
298 * and avoids penalizing OSS/OSTs under light load.
299 * See lod_qos_calc_weight() for how penalties are factored into the weight.
301 * \param[in] lod LOD device
303 * \retval 0 on success
304 * \retval -EAGAIN the number of OSTs isn't enough
306 static int lod_qos_calc_ppo(struct lod_device *lod)
308 struct lod_qos_oss *oss;
309 __u64 ba_max, ba_min, temp;
316 if (!lod->lod_qos.lq_dirty)
319 num_active = lod->lod_desc.ld_active_tgt_count - 1;
321 GOTO(out, rc = -EAGAIN);
323 /* find bavail on each OSS */
324 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list)
326 lod->lod_qos.lq_active_oss_count = 0;
329 * How badly user wants to select OSTs "widely" (not recently chosen
330 * and not on recent OSS's). As opposed to "freely" (free space
333 prio_wide = 256 - lod->lod_qos.lq_prio_free;
335 ba_min = (__u64)(-1);
337 now = cfs_time_current_sec();
338 /* Calculate OST penalty per object
339 * (lod ref taken in lod_qos_prep_create()) */
340 cfs_foreach_bit(lod->lod_ost_bitmap, i) {
341 LASSERT(OST_TGT(lod,i));
342 temp = TGT_BAVAIL(i);
345 ba_min = min(temp, ba_min);
346 ba_max = max(temp, ba_max);
348 /* Count the number of usable OSS's */
349 if (OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail == 0)
350 lod->lod_qos.lq_active_oss_count++;
351 OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail += temp;
353 /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
355 do_div(temp, num_active);
356 OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj =
357 (temp * prio_wide) >> 8;
359 age = (now - OST_TGT(lod,i)->ltd_qos.ltq_used) >> 3;
360 if (lod->lod_qos.lq_reset ||
361 age > 32 * lod->lod_desc.ld_qos_maxage)
362 OST_TGT(lod,i)->ltd_qos.ltq_penalty = 0;
363 else if (age > lod->lod_desc.ld_qos_maxage)
364 /* Decay OST penalty. */
365 OST_TGT(lod,i)->ltd_qos.ltq_penalty >>=
366 (age / lod->lod_desc.ld_qos_maxage);
369 num_active = lod->lod_qos.lq_active_oss_count - 1;
370 if (num_active < 1) {
371 /* If there's only 1 OSS, we can't penalize it, so instead
372 we have to double the OST penalty */
374 cfs_foreach_bit(lod->lod_ost_bitmap, i)
375 OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj <<= 1;
378 /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
379 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
380 temp = oss->lqo_bavail >> 1;
381 do_div(temp, oss->lqo_ost_count * num_active);
382 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
384 age = (now - oss->lqo_used) >> 3;
385 if (lod->lod_qos.lq_reset ||
386 age > 32 * lod->lod_desc.ld_qos_maxage)
387 oss->lqo_penalty = 0;
388 else if (age > lod->lod_desc.ld_qos_maxage)
389 /* Decay OSS penalty. */
390 oss->lqo_penalty >>= age / lod->lod_desc.ld_qos_maxage;
393 lod->lod_qos.lq_dirty = 0;
394 lod->lod_qos.lq_reset = 0;
396 /* If each ost has almost same free space,
397 * do rr allocation for better creation performance */
398 lod->lod_qos.lq_same_space = 0;
399 if ((ba_max * (256 - lod->lod_qos.lq_threshold_rr)) >> 8 < ba_min) {
400 lod->lod_qos.lq_same_space = 1;
401 /* Reset weights for the next time we enter qos mode */
402 lod->lod_qos.lq_reset = 1;
408 if (!rc && lod->lod_qos.lq_same_space)
415 * Calculate weight for a given OST target.
417 * The final OST weight is the number of bytes available minus the OST and
418 * OSS penalties. See lod_qos_calc_ppo() for how penalties are calculated.
420 * \param[in] lod LOD device, where OST targets are listed
421 * \param[in] i OST target index
425 static int lod_qos_calc_weight(struct lod_device *lod, int i)
429 temp = TGT_BAVAIL(i);
430 temp2 = OST_TGT(lod,i)->ltd_qos.ltq_penalty +
431 OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_penalty;
433 OST_TGT(lod,i)->ltd_qos.ltq_weight = 0;
435 OST_TGT(lod,i)->ltd_qos.ltq_weight = temp - temp2;
440 * Re-calculate weights.
442 * The function is called when some OST target was used for a new object. In
443 * this case we should re-calculate all the weights to keep new allocations
446 * \param[in] lod LOD device
447 * \param[in] osts OST pool where a new object was placed
448 * \param[in] index OST target where a new object was placed
449 * \param[out] total_wt new total weight for the pool
453 static int lod_qos_used(struct lod_device *lod, struct ost_pool *osts,
454 __u32 index, __u64 *total_wt)
456 struct lod_tgt_desc *ost;
457 struct lod_qos_oss *oss;
461 ost = OST_TGT(lod,index);
464 /* Don't allocate on this devuce anymore, until the next alloc_qos */
465 ost->ltd_qos.ltq_usable = 0;
467 oss = ost->ltd_qos.ltq_oss;
469 /* Decay old penalty by half (we're adding max penalty, and don't
470 want it to run away.) */
471 ost->ltd_qos.ltq_penalty >>= 1;
472 oss->lqo_penalty >>= 1;
474 /* mark the OSS and OST as recently used */
475 ost->ltd_qos.ltq_used = oss->lqo_used = cfs_time_current_sec();
477 /* Set max penalties for this OST and OSS */
478 ost->ltd_qos.ltq_penalty +=
479 ost->ltd_qos.ltq_penalty_per_obj * lod->lod_ostnr;
480 oss->lqo_penalty += oss->lqo_penalty_per_obj *
481 lod->lod_qos.lq_active_oss_count;
483 /* Decrease all OSS penalties */
484 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
485 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
486 oss->lqo_penalty = 0;
488 oss->lqo_penalty -= oss->lqo_penalty_per_obj;
492 /* Decrease all OST penalties */
493 for (j = 0; j < osts->op_count; j++) {
496 i = osts->op_array[j];
497 if (!cfs_bitmap_check(lod->lod_ost_bitmap, i))
500 ost = OST_TGT(lod,i);
503 if (ost->ltd_qos.ltq_penalty <
504 ost->ltd_qos.ltq_penalty_per_obj)
505 ost->ltd_qos.ltq_penalty = 0;
507 ost->ltd_qos.ltq_penalty -=
508 ost->ltd_qos.ltq_penalty_per_obj;
510 lod_qos_calc_weight(lod, i);
512 /* Recalc the total weight of usable osts */
513 if (ost->ltd_qos.ltq_usable)
514 *total_wt += ost->ltd_qos.ltq_weight;
516 QOS_DEBUG("recalc tgt %d usable=%d avail="LPU64
517 " ostppo="LPU64" ostp="LPU64" ossppo="LPU64
518 " ossp="LPU64" wt="LPU64"\n",
519 i, ost->ltd_qos.ltq_usable, TGT_BAVAIL(i) >> 10,
520 ost->ltd_qos.ltq_penalty_per_obj >> 10,
521 ost->ltd_qos.ltq_penalty >> 10,
522 ost->ltd_qos.ltq_oss->lqo_penalty_per_obj >> 10,
523 ost->ltd_qos.ltq_oss->lqo_penalty >> 10,
524 ost->ltd_qos.ltq_weight >> 10);
530 void lod_qos_rr_init(struct lod_qos_rr *lqr)
532 spin_lock_init(&lqr->lqr_alloc);
537 #define LOV_QOS_EMPTY ((__u32)-1)
540 * Calculate optimal round-robin order with regard to OSSes.
542 * Place all the OSTs from pool \a src_pool in a special array to be used for
543 * round-robin (RR) stripe allocation. The placement algorithm interleaves
544 * OSTs from the different OSSs so that RR allocation can balance OSSs evenly.
545 * Resorts the targets when the number of active targets changes (because of
546 * a new target or activation/deactivation).
548 * \param[in] lod LOD device
549 * \param[in] src_pool OST pool
550 * \param[in] lqr round-robin list
552 * \retval 0 on success
553 * \retval -ENOMEM fails to allocate the array
555 static int lod_qos_calc_rr(struct lod_device *lod, struct ost_pool *src_pool,
556 struct lod_qos_rr *lqr)
558 struct lod_qos_oss *oss;
559 struct lod_tgt_desc *ost;
560 unsigned placed, real_count;
565 if (!lqr->lqr_dirty) {
566 LASSERT(lqr->lqr_pool.op_size);
570 /* Do actual allocation. */
571 down_write(&lod->lod_qos.lq_rw_sem);
574 * Check again. While we were sleeping on @lq_rw_sem something could
577 if (!lqr->lqr_dirty) {
578 LASSERT(lqr->lqr_pool.op_size);
579 up_write(&lod->lod_qos.lq_rw_sem);
583 real_count = src_pool->op_count;
585 /* Zero the pool array */
586 /* alloc_rr is holding a read lock on the pool, so nobody is adding/
587 deleting from the pool. The lq_rw_sem insures that nobody else
589 lqr->lqr_pool.op_count = real_count;
590 rc = lod_ost_pool_extend(&lqr->lqr_pool, real_count);
592 up_write(&lod->lod_qos.lq_rw_sem);
595 for (i = 0; i < lqr->lqr_pool.op_count; i++)
596 lqr->lqr_pool.op_array[i] = LOV_QOS_EMPTY;
598 /* Place all the OSTs from 1 OSS at the same time. */
600 list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
603 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
606 if (!cfs_bitmap_check(lod->lod_ost_bitmap,
607 src_pool->op_array[i]))
610 ost = OST_TGT(lod,src_pool->op_array[i]);
611 LASSERT(ost && ost->ltd_ost);
612 if (ost->ltd_qos.ltq_oss != oss)
615 /* Evenly space these OSTs across arrayspace */
616 next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count;
617 while (lqr->lqr_pool.op_array[next] != LOV_QOS_EMPTY)
618 next = (next + 1) % lqr->lqr_pool.op_count;
620 lqr->lqr_pool.op_array[next] = src_pool->op_array[i];
627 up_write(&lod->lod_qos.lq_rw_sem);
629 if (placed != real_count) {
630 /* This should never happen */
631 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
632 "round-robin list (%d of %d).\n",
634 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
635 LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
636 lqr->lqr_pool.op_array[i]);
643 for (i = 0; i < lqr->lqr_pool.op_count; i++)
644 QOS_CONSOLE("rr #%d ost idx=%d\n", i, lqr->lqr_pool.op_array[i]);
651 * Instantiate and declare creation of a new object.
653 * The function instantiates LU representation for a new object on the
654 * specified device. Also it declares an intention to create that
655 * object on the storage target.
657 * Note lu_object_anon() is used which is a trick with regard to LU/OSD
658 * infrastructure - in the existing precreation framework we can't assign FID
659 * at this moment, we do this later once a transaction is started. So the
660 * special method instantiates FID-less object in the cache and later it
661 * will get a FID and proper placement in LU cache.
663 * \param[in] env execution environment for this thread
664 * \param[in] d LOD device
665 * \param[in] ost_idx OST target index where the object is being created
666 * \param[in] th transaction handle
668 * \retval object ptr on success, ERR_PTR() otherwise
670 static struct dt_object *lod_qos_declare_object_on(const struct lu_env *env,
671 struct lod_device *d,
675 struct lod_tgt_desc *ost;
676 struct lu_object *o, *n;
677 struct lu_device *nd;
678 struct dt_object *dt;
683 LASSERT(ost_idx < d->lod_osts_size);
684 ost = OST_TGT(d,ost_idx);
686 LASSERT(ost->ltd_ost);
688 nd = &ost->ltd_ost->dd_lu_dev;
691 * allocate anonymous object with zero fid, real fid
692 * will be assigned by OSP within transaction
693 * XXX: to be fixed with fully-functional OST fids
695 o = lu_object_anon(env, nd, NULL);
697 GOTO(out, dt = ERR_PTR(PTR_ERR(o)));
699 n = lu_object_locate(o->lo_header, nd->ld_type);
700 if (unlikely(n == NULL)) {
701 CERROR("can't find slice\n");
702 lu_object_put(env, o);
703 GOTO(out, dt = ERR_PTR(-EINVAL));
706 dt = container_of(n, struct dt_object, do_lu);
708 rc = lod_sub_object_declare_create(env, dt, NULL, NULL, NULL, th);
710 CDEBUG(D_OTHER, "can't declare creation on #%u: %d\n",
712 lu_object_put(env, o);
721 * Calculate a minimum acceptable stripe count.
723 * Return an acceptable stripe count depending on flag LOV_USES_DEFAULT_STRIPE:
724 * all stripes or 3/4 of stripes.
726 * \param[in] stripe_cnt number of stripes requested
727 * \param[in] flags 0 or LOV_USES_DEFAULT_STRIPE
729 * \retval acceptable stripecount
731 static int min_stripe_count(__u32 stripe_cnt, int flags)
733 return (flags & LOV_USES_DEFAULT_STRIPE ?
734 stripe_cnt - (stripe_cnt / 4) : stripe_cnt);
737 #define LOV_CREATE_RESEED_MULT 30
738 #define LOV_CREATE_RESEED_MIN 2000
741 * Check if an OST is full.
743 * Check whether an OST should be considered full based
744 * on the given statfs data.
746 * \param[in] msfs statfs data
748 * \retval false not full
751 static int inline lod_qos_dev_is_full(struct obd_statfs *msfs)
754 int bs = msfs->os_bsize;
756 LASSERT(((bs - 1) & bs) == 0);
758 /* the minimum of 0.1% used blocks and 1GB bytes. */
759 used = min_t(__u64, (msfs->os_blocks - msfs->os_bfree) >> 10,
760 1 << (31 - ffs(bs)));
761 return (msfs->os_bavail < used);
765 * Initialize temporary OST-in-use array.
767 * Allocate or extend the array used to mark targets already assigned to a new
768 * striping so they are not used more than once.
770 * \param[in] env execution environment for this thread
771 * \param[in] stripes number of items needed in the array
773 * \retval 0 on success
774 * \retval -ENOMEM on error
776 static inline int lod_qos_ost_in_use_clear(const struct lu_env *env,
779 struct lod_thread_info *info = lod_env_info(env);
781 if (info->lti_ea_store_size < sizeof(int) * stripes)
782 lod_ea_store_resize(info, stripes * sizeof(int));
783 if (info->lti_ea_store_size < sizeof(int) * stripes) {
784 CERROR("can't allocate memory for ost-in-use array\n");
787 memset(info->lti_ea_store, -1, sizeof(int) * stripes);
792 * Remember a target in the array of used targets.
794 * Mark the given target as used for a new striping being created. The status
795 * of an OST in a striping can be checked with lod_qos_is_ost_used().
797 * \param[in] env execution environment for this thread
798 * \param[in] idx index in the array
799 * \param[in] ost OST target index to mark as used
801 static inline void lod_qos_ost_in_use(const struct lu_env *env,
804 struct lod_thread_info *info = lod_env_info(env);
805 int *osts = info->lti_ea_store;
807 LASSERT(info->lti_ea_store_size >= idx * sizeof(int));
812 * Check is OST used in a striping.
814 * Checks whether OST with the given index is marked as used in the temporary
815 * array (see lod_qos_ost_in_use()).
817 * \param[in] env execution environment for this thread
818 * \param[in] ost OST target index to check
819 * \param[in] stripes the number of items used in the array already
824 static int lod_qos_is_ost_used(const struct lu_env *env, int ost, __u32 stripes)
826 struct lod_thread_info *info = lod_env_info(env);
827 int *osts = info->lti_ea_store;
830 for (j = 0; j < stripes; j++) {
837 static int lod_check_and_reserve_ost(const struct lu_env *env,
838 struct lod_device *m,
839 struct obd_statfs *sfs, __u32 ost_idx,
840 __u32 speed, __u32 *s_idx,
841 struct dt_object **stripe,
845 __u32 stripe_idx = *s_idx;
848 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
850 /* this OSP doesn't feel well */
857 if (lod_qos_dev_is_full(sfs)) {
858 QOS_DEBUG("#%d is full\n", ost_idx);
863 * We expect number of precreated objects in f_ffree at
864 * the first iteration, skip OSPs with no objects ready
866 if (sfs->os_fprecreated == 0 && speed == 0) {
867 QOS_DEBUG("#%d: precreation is empty\n", ost_idx);
872 * try to use another OSP if this one is degraded
874 if (sfs->os_state & OS_STATE_DEGRADED && speed < 2) {
875 QOS_DEBUG("#%d: degraded\n", ost_idx);
880 * do not put >1 objects on a single OST
882 if (speed && lod_qos_is_ost_used(env, ost_idx, stripe_idx))
885 o = lod_qos_declare_object_on(env, m, ost_idx, th);
887 CDEBUG(D_OTHER, "can't declare new object on #%u: %d\n",
888 ost_idx, (int) PTR_ERR(o));
894 * We've successfully declared (reserved) an object
896 lod_qos_ost_in_use(env, stripe_idx, ost_idx);
897 stripe[stripe_idx] = o;
906 * Allocate a striping using round-robin algorithm.
908 * Allocates a new striping using round-robin algorithm. The function refreshes
909 * all the internal structures (statfs cache, array of available OSTs sorted
910 * with regard to OSS, etc). The number of stripes required is taken from the
911 * object (must be prepared by the caller), but can change if the flag
912 * LOV_USES_DEFAULT_STRIPE is supplied. The caller should ensure nobody else
913 * is trying to create a striping on the object in parallel. All the internal
914 * structures (like pools, etc) are protected and no additional locking is
915 * required. The function succeeds even if a single stripe is allocated. To save
916 * time we give priority to targets which already have objects precreated.
917 * Full OSTs are skipped (see lod_qos_dev_is_full() for the details).
919 * \param[in] env execution environment for this thread
920 * \param[in] lo LOD object
921 * \param[out] stripe striping created
922 * \param[in] flags allocation flags (0 or LOV_USES_DEFAULT_STRIPE)
923 * \param[in] th transaction handle
925 * \retval 0 on success
926 * \retval -ENOSPC if not enough OSTs are found
927 * \retval negative negated errno for other failures
929 static int lod_alloc_rr(const struct lu_env *env, struct lod_object *lo,
930 struct dt_object **stripe, int flags,
933 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
934 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
935 struct pool_desc *pool = NULL;
936 struct ost_pool *osts;
937 struct lod_qos_rr *lqr;
938 unsigned int i, array_idx;
940 __u32 ost_start_idx_temp;
942 __u32 stripe_idx = 0;
943 __u32 stripe_cnt = lo->ldo_stripenr;
944 __u32 stripe_cnt_min = min_stripe_count(stripe_cnt, flags);
949 pool = lod_find_pool(m, lo->ldo_pool);
952 down_read(&pool_tgt_rw_sem(pool));
953 osts = &(pool->pool_obds);
954 lqr = &(pool->pool_rr);
956 osts = &(m->lod_pool_info);
957 lqr = &(m->lod_qos.lq_rr);
960 rc = lod_qos_calc_rr(m, osts, lqr);
964 rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
968 down_read(&m->lod_qos.lq_rw_sem);
969 spin_lock(&lqr->lqr_alloc);
970 if (--lqr->lqr_start_count <= 0) {
971 lqr->lqr_start_idx = cfs_rand() % osts->op_count;
972 lqr->lqr_start_count =
973 (LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) +
974 LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U);
975 } else if (stripe_cnt_min >= osts->op_count ||
976 lqr->lqr_start_idx > osts->op_count) {
977 /* If we have allocated from all of the OSTs, slowly
978 * precess the next start if the OST/stripe count isn't
979 * already doing this for us. */
980 lqr->lqr_start_idx %= osts->op_count;
981 if (stripe_cnt > 1 && (osts->op_count % stripe_cnt) != 1)
982 ++lqr->lqr_offset_idx;
984 ost_start_idx_temp = lqr->lqr_start_idx;
988 QOS_DEBUG("pool '%s' want %d startidx %d startcnt %d offset %d "
989 "active %d count %d\n",
990 lo->ldo_pool ? lo->ldo_pool : "",
991 stripe_cnt, lqr->lqr_start_idx, lqr->lqr_start_count,
992 lqr->lqr_offset_idx, osts->op_count, osts->op_count);
994 for (i = 0; i < osts->op_count && stripe_idx < lo->ldo_stripenr; i++) {
995 array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) %
997 ++lqr->lqr_start_idx;
998 ost_idx = lqr->lqr_pool.op_array[array_idx];
1000 QOS_DEBUG("#%d strt %d act %d strp %d ary %d idx %d\n",
1001 i, lqr->lqr_start_idx, /* XXX: active*/ 0,
1002 stripe_idx, array_idx, ost_idx);
1004 if ((ost_idx == LOV_QOS_EMPTY) ||
1005 !cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
1008 /* Fail Check before osc_precreate() is called
1009 so we can only 'fail' single OSC. */
1010 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
1013 spin_unlock(&lqr->lqr_alloc);
1014 rc = lod_check_and_reserve_ost(env, m, sfs, ost_idx, speed,
1015 &stripe_idx, stripe, th);
1016 spin_lock(&lqr->lqr_alloc);
1018 if ((speed < 2) && (stripe_idx < stripe_cnt_min)) {
1019 /* Try again, allowing slower OSCs */
1021 lqr->lqr_start_idx = ost_start_idx_temp;
1025 spin_unlock(&lqr->lqr_alloc);
1026 up_read(&m->lod_qos.lq_rw_sem);
1029 lo->ldo_stripenr = stripe_idx;
1030 /* at least one stripe is allocated */
1033 /* nobody provided us with a single object */
1039 up_read(&pool_tgt_rw_sem(pool));
1040 /* put back ref got by lod_find_pool() */
1041 lod_pool_putref(pool);
1048 * Allocate a specific striping layout on a user defined set of OSTs.
1050 * Allocates new striping using the OST index range provided by the data from
1051 * the lmm_obejcts contained in the lov_user_md passed to this method. Full
1052 * OSTs are not considered. The exact order of OSTs requested by the user
1053 * is respected as much as possible depending on OST status. The number of
1054 * stripes needed and stripe offset are taken from the object. If that number
1055 * can not be met, then the function returns a failure and then it's the
1056 * caller's responsibility to release the stripes allocated. All the internal
1057 * structures are protected, but no concurrent allocation is allowed on the
1060 * \param[in] env execution environment for this thread
1061 * \param[in] lo LOD object
1062 * \param[out] stripe striping created
1063 * \param[in] lum stripe md to specify list of OSTs
1064 * \param[in] th transaction handle
1066 * \retval 0 on success
1067 * \retval -ENODEV OST index does not exist on file system
1068 * \retval -EINVAL requested OST index is invalid
1069 * \retval negative negated errno on error
1071 static int lod_alloc_ost_list(const struct lu_env *env,
1072 struct lod_object *lo, struct dt_object **stripe,
1073 struct lov_user_md *lum, struct thandle *th)
1075 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1076 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
1077 struct dt_object *o;
1078 struct lov_user_md_v3 *v3;
1079 unsigned int array_idx = 0;
1080 int stripe_count = 0;
1085 /* for specific OSTs layout */
1086 LASSERT(lum != NULL && lum->lmm_magic == LOV_USER_MAGIC_SPECIFIC);
1087 lustre_print_user_md(D_OTHER, lum, __func__);
1089 rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
1093 v3 = (struct lov_user_md_v3 *)lum;
1094 for (i = 0; i < lo->ldo_stripenr; i++) {
1095 if (v3->lmm_objects[i].l_ost_idx == lo->ldo_def_stripe_offset) {
1100 if (i == lo->ldo_stripenr) {
1102 "%s: start index %d not in the specified list of OSTs\n",
1103 lod2obd(m)->obd_name, lo->ldo_def_stripe_offset);
1107 for (i = 0; i < lo->ldo_stripenr;
1108 i++, array_idx = (array_idx + 1) % lo->ldo_stripenr) {
1109 __u32 ost_idx = v3->lmm_objects[array_idx].l_ost_idx;
1111 if (!cfs_bitmap_check(m->lod_ost_bitmap, ost_idx)) {
1117 * do not put >1 objects on a single OST
1119 if (lod_qos_is_ost_used(env, ost_idx, stripe_count)) {
1124 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
1125 if (rc < 0) /* this OSP doesn't feel well */
1128 o = lod_qos_declare_object_on(env, m, ost_idx, th);
1132 "%s: can't declare new object on #%u: %d\n",
1133 lod2obd(m)->obd_name, ost_idx, rc);
1138 * We've successfully declared (reserved) an object
1140 lod_qos_ost_in_use(env, stripe_count, ost_idx);
1141 stripe[stripe_count] = o;
1149 * Allocate a striping on a predefined set of OSTs.
1151 * Allocates new layout starting from OST index in lo->ldo_def_stripe_offset.
1152 * Full OSTs are not considered. The exact order of OSTs is not important and
1153 * varies depending on OST status. The allocation procedure prefers the targets
1154 * with precreated objects ready. The number of stripes needed and stripe
1155 * offset are taken from the object. If that number cannot be met, then the
1156 * function returns an error and then it's the caller's responsibility to
1157 * release the stripes allocated. All the internal structures are protected,
1158 * but no concurrent allocation is allowed on the same objects.
1160 * \param[in] env execution environment for this thread
1161 * \param[in] lo LOD object
1162 * \param[out] stripe striping created
1163 * \param[in] flags not used
1164 * \param[in] th transaction handle
1166 * \retval 0 on success
1167 * \retval -ENOSPC if no OST objects are available at all
1168 * \retval -EFBIG if not enough OST objects are found
1169 * \retval -EINVAL requested offset is invalid
1170 * \retval negative errno on failure
1172 static int lod_alloc_specific(const struct lu_env *env, struct lod_object *lo,
1173 struct dt_object **stripe, int flags,
1176 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1177 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
1178 struct dt_object *o;
1180 unsigned int i, array_idx, ost_count;
1181 int rc, stripe_num = 0;
1183 struct pool_desc *pool = NULL;
1184 struct ost_pool *osts;
1187 rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
1192 pool = lod_find_pool(m, lo->ldo_pool);
1195 down_read(&pool_tgt_rw_sem(pool));
1196 osts = &(pool->pool_obds);
1198 osts = &(m->lod_pool_info);
1201 ost_count = osts->op_count;
1204 /* search loi_ost_idx in ost array */
1206 for (i = 0; i < ost_count; i++) {
1207 if (osts->op_array[i] == lo->ldo_def_stripe_offset) {
1212 if (i == ost_count) {
1213 CERROR("Start index %d not found in pool '%s'\n",
1214 lo->ldo_def_stripe_offset,
1215 lo->ldo_pool ? lo->ldo_pool : "");
1216 GOTO(out, rc = -EINVAL);
1219 for (i = 0; i < ost_count;
1220 i++, array_idx = (array_idx + 1) % ost_count) {
1221 ost_idx = osts->op_array[array_idx];
1223 if (!cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
1226 /* Fail Check before osc_precreate() is called
1227 so we can only 'fail' single OSC. */
1228 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
1232 * do not put >1 objects on a single OST
1234 if (lod_qos_is_ost_used(env, ost_idx, stripe_num))
1237 /* Drop slow OSCs if we can, but not for requested start idx.
1239 * This means "if OSC is slow and it is not the requested
1240 * start OST, then it can be skipped, otherwise skip it only
1241 * if it is inactive/recovering/out-of-space." */
1243 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
1245 /* this OSP doesn't feel well */
1250 * We expect number of precreated objects in f_ffree at
1251 * the first iteration, skip OSPs with no objects ready
1252 * don't apply this logic to OST specified with stripe_offset
1254 if (i != 0 && sfs->os_fprecreated == 0 && speed == 0)
1257 o = lod_qos_declare_object_on(env, m, ost_idx, th);
1259 CDEBUG(D_OTHER, "can't declare new object on #%u: %d\n",
1260 ost_idx, (int) PTR_ERR(o));
1265 * We've successfully declared (reserved) an object
1267 lod_qos_ost_in_use(env, stripe_num, ost_idx);
1268 stripe[stripe_num] = o;
1271 /* We have enough stripes */
1272 if (stripe_num == lo->ldo_stripenr)
1276 /* Try again, allowing slower OSCs */
1281 /* If we were passed specific striping params, then a failure to
1282 * meet those requirements is an error, since we can't reallocate
1283 * that memory (it might be part of a larger array or something).
1285 CERROR("can't lstripe objid "DFID": have %d want %u\n",
1286 PFID(lu_object_fid(lod2lu_obj(lo))), stripe_num,
1288 rc = stripe_num == 0 ? -ENOSPC : -EFBIG;
1291 up_read(&pool_tgt_rw_sem(pool));
1292 /* put back ref got by lod_find_pool() */
1293 lod_pool_putref(pool);
1300 * Check whether QoS allocation should be used.
1302 * A simple helper to decide when QoS allocation should be used:
1303 * if it's just a single available target or the used space is
1304 * evenly distributed among the targets at the moment, then QoS
1305 * allocation algorithm should not be used.
1307 * \param[in] lod LOD device
1309 * \retval 0 should not be used
1310 * \retval 1 should be used
1312 static inline int lod_qos_is_usable(struct lod_device *lod)
1315 /* to be able to debug QoS code */
1319 /* Detect -EAGAIN early, before expensive lock is taken. */
1320 if (!lod->lod_qos.lq_dirty && lod->lod_qos.lq_same_space)
1323 if (lod->lod_desc.ld_active_tgt_count < 2)
1330 * Allocate a striping using an algorithm with weights.
1332 * The function allocates OST objects to create a striping. The algorithm
1333 * used is based on weights (currently only using the free space), and it's
1334 * trying to ensure the space is used evenly by OSTs and OSSs. The striping
1335 * configuration (# of stripes, offset, pool) is taken from the object and
1336 * is prepared by the caller.
1338 * If LOV_USES_DEFAULT_STRIPE is not passed and prepared configuration can't
1339 * be met due to too few OSTs, then allocation fails. If the flag is passed
1340 * fewer than 3/4 of the requested number of stripes can be allocated, then
1343 * No concurrent allocation is allowed on the object and this must be ensured
1344 * by the caller. All the internal structures are protected by the function.
1346 * The algorithm has two steps: find available OSTs and calculate their
1347 * weights, then select the OSTs with their weights used as the probability.
1348 * An OST with a higher weight is proportionately more likely to be selected
1349 * than one with a lower weight.
1351 * \param[in] env execution environment for this thread
1352 * \param[in] lo LOD object
1353 * \param[out] stripe striping created
1354 * \param[in] flags 0 or LOV_USES_DEFAULT_STRIPE
1355 * \param[in] th transaction handle
1357 * \retval 0 on success
1358 * \retval -EAGAIN not enough OSTs are found for specified stripe count
1359 * \retval -EINVAL requested OST index is invalid
1360 * \retval negative errno on failure
1362 static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo,
1363 struct dt_object **stripe, int flags,
1366 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1367 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
1368 struct lod_tgt_desc *ost;
1369 struct dt_object *o;
1370 __u64 total_weight = 0;
1373 __u32 nfound, good_osts;
1374 __u32 stripe_cnt = lo->ldo_stripenr;
1375 __u32 stripe_cnt_min;
1376 struct pool_desc *pool = NULL;
1377 struct ost_pool *osts;
1380 stripe_cnt_min = min_stripe_count(stripe_cnt, flags);
1381 if (stripe_cnt_min < 1)
1385 pool = lod_find_pool(m, lo->ldo_pool);
1388 down_read(&pool_tgt_rw_sem(pool));
1389 osts = &(pool->pool_obds);
1391 osts = &(m->lod_pool_info);
1394 /* Detect -EAGAIN early, before expensive lock is taken. */
1395 if (!lod_qos_is_usable(m))
1396 GOTO(out_nolock, rc = -EAGAIN);
1398 /* Do actual allocation, use write lock here. */
1399 down_write(&m->lod_qos.lq_rw_sem);
1402 * Check again, while we were sleeping on @lq_rw_sem things could
1405 if (!lod_qos_is_usable(m))
1406 GOTO(out, rc = -EAGAIN);
1408 rc = lod_qos_calc_ppo(m);
1412 rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
1417 /* Find all the OSTs that are valid stripe candidates */
1418 for (i = 0; i < osts->op_count; i++) {
1419 if (!cfs_bitmap_check(m->lod_ost_bitmap, osts->op_array[i]))
1422 rc = lod_statfs_and_check(env, m, osts->op_array[i], sfs);
1424 /* this OSP doesn't feel well */
1431 if (lod_qos_dev_is_full(sfs))
1434 /* Fail Check before osc_precreate() is called
1435 so we can only 'fail' single OSC. */
1436 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) &&
1437 osts->op_array[i] == 0)
1440 ost = OST_TGT(m,osts->op_array[i]);
1441 ost->ltd_qos.ltq_usable = 1;
1442 lod_qos_calc_weight(m, osts->op_array[i]);
1443 total_weight += ost->ltd_qos.ltq_weight;
1448 QOS_DEBUG("found %d good osts\n", good_osts);
1450 if (good_osts < stripe_cnt_min)
1451 GOTO(out, rc = -EAGAIN);
1453 /* We have enough osts */
1454 if (good_osts < stripe_cnt)
1455 stripe_cnt = good_osts;
1457 /* Find enough OSTs with weighted random allocation. */
1459 while (nfound < stripe_cnt) {
1460 __u64 rand, cur_weight;
1466 #if BITS_PER_LONG == 32
1467 rand = cfs_rand() % (unsigned)total_weight;
1468 /* If total_weight > 32-bit, first generate the high
1469 * 32 bits of the random number, then add in the low
1470 * 32 bits (truncated to the upper limit, if needed) */
1471 if (total_weight > 0xffffffffULL)
1472 rand = (__u64)(cfs_rand() %
1473 (unsigned)(total_weight >> 32)) << 32;
1477 if (rand == (total_weight & 0xffffffff00000000ULL))
1478 rand |= cfs_rand() % (unsigned)total_weight;
1483 rand = ((__u64)cfs_rand() << 32 | cfs_rand()) %
1490 /* On average, this will hit larger-weighted OSTs more often.
1491 * 0-weight OSTs will always get used last (only when rand=0) */
1492 for (i = 0; i < osts->op_count; i++) {
1493 __u32 idx = osts->op_array[i];
1495 if (!cfs_bitmap_check(m->lod_ost_bitmap, idx))
1498 ost = OST_TGT(m,idx);
1500 if (!ost->ltd_qos.ltq_usable)
1503 cur_weight += ost->ltd_qos.ltq_weight;
1504 QOS_DEBUG("stripe_cnt=%d nfound=%d cur_weight="LPU64
1505 " rand="LPU64" total_weight="LPU64"\n",
1506 stripe_cnt, nfound, cur_weight, rand,
1509 if (cur_weight < rand)
1512 QOS_DEBUG("stripe=%d to idx=%d\n", nfound, idx);
1515 * do not put >1 objects on a single OST
1517 if (lod_qos_is_ost_used(env, idx, nfound))
1519 lod_qos_ost_in_use(env, nfound, idx);
1521 o = lod_qos_declare_object_on(env, m, idx, th);
1523 QOS_DEBUG("can't declare object on #%u: %d\n",
1524 idx, (int) PTR_ERR(o));
1527 stripe[nfound++] = o;
1528 lod_qos_used(m, osts, idx, &total_weight);
1534 /* no OST found on this iteration, give up */
1539 if (unlikely(nfound != stripe_cnt)) {
1541 * when the decision to use weighted algorithm was made
1542 * we had enough appropriate OSPs, but this state can
1543 * change anytime (no space on OST, broken connection, etc)
1544 * so it's possible OSP won't be able to provide us with
1545 * an object due to just changed state
1547 LCONSOLE_INFO("wanted %d, found %d\n", stripe_cnt, nfound);
1548 for (i = 0; i < nfound; i++) {
1549 LASSERT(stripe[i] != NULL);
1550 lu_object_put(env, &stripe[i]->do_lu);
1554 /* makes sense to rebalance next time */
1555 m->lod_qos.lq_dirty = 1;
1556 m->lod_qos.lq_same_space = 0;
1562 up_write(&m->lod_qos.lq_rw_sem);
1566 up_read(&pool_tgt_rw_sem(pool));
1567 /* put back ref got by lod_find_pool() */
1568 lod_pool_putref(pool);
1575 * Find largest stripe count the caller can use.
1577 * Find the maximal possible stripe count not greater than \a stripe_count.
1578 * Sometimes suggested stripecount can't be reached for a number of reasons:
1579 * lack of enough active OSTs or the backend does not support EAs that large.
1580 * If the passed one is 0, then the filesystem's default one is used.
1582 * \param[in] lod LOD device
1583 * \param[in] magic the format if striping
1584 * \param[in] stripe_count count the caller would like to use
1586 * \retval the maximum usable stripe count
1588 static __u16 lod_get_stripecnt(struct lod_device *lod, __u32 magic,
1591 __u32 max_stripes = LOV_MAX_STRIPE_COUNT_OLD;
1594 stripe_count = lod->lod_desc.ld_default_stripe_count;
1595 if (stripe_count > lod->lod_desc.ld_active_tgt_count)
1596 stripe_count = lod->lod_desc.ld_active_tgt_count;
1600 /* stripe count is based on whether OSD can handle larger EA sizes */
1601 if (lod->lod_osd_max_easize > 0)
1602 max_stripes = lov_mds_md_max_stripe_count(
1603 lod->lod_osd_max_easize, magic);
1605 return (stripe_count < max_stripes) ? stripe_count : max_stripes;
1609 * Create in-core respresentation for a fully-defined striping
1611 * When the caller passes a fully-defined striping (i.e. everything including
1612 * OST object FIDs are defined), then we still need to instantiate LU-cache
1613 * with the objects representing the stripes defined. This function completes
1616 * \param[in] env execution environment for this thread
1617 * \param[in] mo LOD object
1618 * \param[in] buf buffer containing the striping
1620 * \retval 0 on success
1621 * \retval negative negated errno on error
1623 static int lod_use_defined_striping(const struct lu_env *env,
1624 struct lod_object *mo,
1625 const struct lu_buf *buf)
1627 struct lov_mds_md_v1 *v1 = buf->lb_buf;
1628 struct lov_mds_md_v3 *v3 = buf->lb_buf;
1629 struct lov_ost_data_v1 *objs;
1634 magic = le32_to_cpu(v1->lmm_magic);
1635 if (magic == LOV_MAGIC_V1_DEF) {
1636 magic = LOV_MAGIC_V1;
1637 objs = &v1->lmm_objects[0];
1638 } else if (magic == LOV_MAGIC_V3_DEF) {
1639 magic = LOV_MAGIC_V3;
1640 objs = &v3->lmm_objects[0];
1641 lod_object_set_pool(mo, v3->lmm_pool_name);
1643 GOTO(out, rc = -EINVAL);
1646 mo->ldo_pattern = le32_to_cpu(v1->lmm_pattern);
1647 mo->ldo_stripe_size = le32_to_cpu(v1->lmm_stripe_size);
1648 mo->ldo_stripenr = le16_to_cpu(v1->lmm_stripe_count);
1649 mo->ldo_layout_gen = le16_to_cpu(v1->lmm_layout_gen);
1651 /* fixup for released file before object initialization */
1652 if (mo->ldo_pattern & LOV_PATTERN_F_RELEASED) {
1653 mo->ldo_released_stripenr = mo->ldo_stripenr;
1654 mo->ldo_stripenr = 0;
1657 LASSERT(buf->lb_len >= lov_mds_md_size(mo->ldo_stripenr, magic));
1659 if (mo->ldo_stripenr > 0)
1660 rc = lod_initialize_objects(env, mo, objs);
1667 * Parse suggested striping configuration.
1669 * The caller gets a suggested striping configuration from a number of sources
1670 * including per-directory default and applications. Then it needs to verify
1671 * the suggested striping is valid, apply missing bits and store the resulting
1672 * configuration in the object to be used by the allocator later. Must not be
1673 * called concurrently against the same object. It's OK to provide a
1674 * fully-defined striping.
1676 * \param[in] env execution environment for this thread
1677 * \param[in] lo LOD object
1678 * \param[in] buf buffer containing the striping
1680 * \retval 0 on success
1681 * \retval negative negated errno on error
1683 static int lod_qos_parse_config(const struct lu_env *env,
1684 struct lod_object *lo,
1685 const struct lu_buf *buf)
1687 struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
1688 struct lov_user_md_v1 *v1 = NULL;
1689 struct lov_user_md_v3 *v3 = NULL;
1690 char *pool_name = NULL;
1696 if (buf == NULL || buf->lb_buf == NULL || buf->lb_len == 0)
1701 magic = v1->lmm_magic;
1703 if (unlikely(magic == LOV_MAGIC_V1_DEF || magic == LOV_MAGIC_V3_DEF)) {
1704 /* try to use as fully defined striping */
1705 rc = lod_use_defined_striping(env, lo, buf);
1710 case __swab32(LOV_USER_MAGIC_V1):
1711 lustre_swab_lov_user_md_v1(v1);
1712 magic = v1->lmm_magic;
1714 case LOV_USER_MAGIC_V1:
1718 case __swab32(LOV_USER_MAGIC_V3):
1719 lustre_swab_lov_user_md_v3(v3);
1720 magic = v3->lmm_magic;
1722 case LOV_USER_MAGIC_V3:
1724 pool_name = v3->lmm_pool_name;
1727 case __swab32(LOV_USER_MAGIC_SPECIFIC):
1728 lustre_swab_lov_user_md_v3(v3);
1729 lustre_swab_lov_user_md_objects(v3->lmm_objects,
1730 v3->lmm_stripe_count);
1731 magic = v3->lmm_magic;
1733 case LOV_USER_MAGIC_SPECIFIC:
1734 if (v3->lmm_stripe_offset == LOV_OFFSET_DEFAULT)
1735 v3->lmm_stripe_offset = v3->lmm_objects[0].l_ost_idx;
1736 if (v3->lmm_pool_name[0] != '\0')
1737 pool_name = v3->lmm_pool_name;
1738 size = lov_user_md_size(v3->lmm_stripe_count,
1739 LOV_USER_MAGIC_SPECIFIC);
1743 CERROR("%s: unrecognized magic %X\n",
1744 lod2obd(d)->obd_name, magic);
1748 if (unlikely(buf->lb_len < size)) {
1749 CERROR("%s: wrong size: %zd, expect: %u\n",
1750 lod2obd(d)->obd_name, buf->lb_len, size);
1754 lustre_print_user_md(D_OTHER, v1, "parse config");
1756 v1->lmm_magic = magic;
1757 if (v1->lmm_pattern == 0)
1758 v1->lmm_pattern = LOV_PATTERN_RAID0;
1759 if (lov_pattern(v1->lmm_pattern) != LOV_PATTERN_RAID0) {
1760 CERROR("%s: invalid pattern: %x\n",
1761 lod2obd(d)->obd_name, v1->lmm_pattern);
1764 lo->ldo_pattern = v1->lmm_pattern;
1766 if (v1->lmm_stripe_size > 0)
1767 lo->ldo_stripe_size = v1->lmm_stripe_size;
1769 if (lo->ldo_stripe_size & (LOV_MIN_STRIPE_SIZE - 1))
1770 lo->ldo_stripe_size = LOV_MIN_STRIPE_SIZE;
1772 if (v1->lmm_stripe_count > 0)
1773 lo->ldo_stripenr = v1->lmm_stripe_count;
1775 lo->ldo_def_stripe_offset = v1->lmm_stripe_offset;
1777 lod_object_set_pool(lo, NULL);
1778 if (pool_name != NULL) {
1779 struct pool_desc *pool;
1781 /* In the function below, .hs_keycmp resolves to
1782 * pool_hashkey_keycmp() */
1783 /* coverity[overrun-buffer-val] */
1784 pool = lod_find_pool(d, pool_name);
1786 if (lo->ldo_def_stripe_offset != LOV_OFFSET_DEFAULT) {
1787 rc = lod_check_index_in_pool(
1788 lo->ldo_def_stripe_offset, pool);
1790 lod_pool_putref(pool);
1791 CERROR("%s: invalid offset, %u\n",
1792 lod2obd(d)->obd_name,
1793 lo->ldo_def_stripe_offset);
1798 if (lo->ldo_stripenr > pool_tgt_count(pool))
1799 lo->ldo_stripenr = pool_tgt_count(pool);
1801 lod_pool_putref(pool);
1804 lod_object_set_pool(lo, pool_name);
1807 /* fixup for released file */
1808 if (lo->ldo_pattern & LOV_PATTERN_F_RELEASED) {
1809 lo->ldo_released_stripenr = lo->ldo_stripenr;
1810 lo->ldo_stripenr = 0;
1817 * Create a striping for an obejct.
1819 * The function creates a new striping for the object. A buffer containing
1820 * configuration hints can be provided optionally. The function tries QoS
1821 * algorithm first unless free space is distributed evenly among OSTs, but
1822 * by default RR algorithm is preferred due to internal concurrency (QoS is
1823 * serialized). The caller must ensure no concurrent calls to the function
1824 * are made against the same object.
1826 * \param[in] env execution environment for this thread
1827 * \param[in] lo LOD object
1828 * \param[in] attr attributes OST objects will be declared with
1829 * \param[in] buf suggested striping configuration or NULL
1830 * \param[in] th transaction handle
1832 * \retval 0 on success
1833 * \retval negative negated errno on error
1835 int lod_qos_prep_create(const struct lu_env *env, struct lod_object *lo,
1836 struct lu_attr *attr, const struct lu_buf *buf,
1839 struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
1840 struct dt_object **stripe;
1842 int flag = LOV_USES_ASSIGNED_STRIPE;
1848 /* no OST available */
1849 /* XXX: should we be waiting a bit to prevent failures during
1850 * cluster initialization? */
1851 if (d->lod_ostnr == 0)
1852 GOTO(out, rc = -EIO);
1855 * by this time, the object's ldo_stripenr and ldo_stripe_size
1856 * contain default value for striping: taken from the parent
1857 * or from filesystem defaults
1859 * in case the caller is passing lovea with new striping config,
1860 * we may need to parse lovea and apply new configuration
1862 rc = lod_qos_parse_config(env, lo, buf);
1866 /* A released file is being created */
1867 if (lo->ldo_stripenr == 0)
1870 if (likely(lo->ldo_stripe == NULL)) {
1871 struct lov_user_md *lum = NULL;
1874 * no striping has been created so far
1876 LASSERT(lo->ldo_stripenr > 0);
1878 * statfs and check OST targets now, since ld_active_tgt_count
1879 * could be changed if some OSTs are [de]activated manually.
1881 lod_qos_statfs_update(env, d);
1882 lo->ldo_stripenr = lod_get_stripecnt(d, LOV_MAGIC,
1885 stripe_len = lo->ldo_stripenr;
1886 OBD_ALLOC(stripe, sizeof(stripe[0]) * stripe_len);
1888 GOTO(out, rc = -ENOMEM);
1890 lod_getref(&d->lod_ost_descs);
1891 /* XXX: support for non-0 files w/o objects */
1892 CDEBUG(D_OTHER, "tgt_count %d stripenr %d\n",
1893 d->lod_desc.ld_tgt_count, stripe_len);
1895 if (buf != NULL && buf->lb_buf != NULL)
1898 if (lum != NULL && lum->lmm_magic == LOV_USER_MAGIC_SPECIFIC) {
1899 rc = lod_alloc_ost_list(env, lo, stripe, lum, th);
1900 } else if (lo->ldo_def_stripe_offset == LOV_OFFSET_DEFAULT) {
1901 rc = lod_alloc_qos(env, lo, stripe, flag, th);
1903 rc = lod_alloc_rr(env, lo, stripe, flag, th);
1905 rc = lod_alloc_specific(env, lo, stripe, flag, th);
1907 lod_putref(d, &d->lod_ost_descs);
1910 for (i = 0; i < stripe_len; i++)
1911 if (stripe[i] != NULL)
1912 lu_object_put(env, &stripe[i]->do_lu);
1914 OBD_FREE(stripe, sizeof(stripe[0]) * stripe_len);
1915 lo->ldo_stripenr = 0;
1917 lo->ldo_stripe = stripe;
1918 lo->ldo_stripes_allocated = stripe_len;
1922 * lod_qos_parse_config() found supplied buf as a predefined
1923 * striping (not a hint), so it allocated all the object
1924 * now we need to create them
1926 for (i = 0; i < lo->ldo_stripenr; i++) {
1927 struct dt_object *o;
1929 o = lo->ldo_stripe[i];
1932 rc = lod_sub_object_declare_create(env, o, attr, NULL,
1935 CERROR("can't declare create: %d\n", rc);