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, 2013, 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
36 #define DEBUG_SUBSYSTEM S_LOV
38 #include <asm/div64.h>
39 #include <libcfs/libcfs.h>
40 #include <obd_class.h>
41 #include <lustre/lustre_idl.h>
42 #include "lod_internal.h"
45 * force QoS policy (not RR) to be used for testing purposes
52 #define QOS_DEBUG(fmt, ...) CDEBUG(D_OTHER, fmt, ## __VA_ARGS__)
53 #define QOS_CONSOLE(fmt, ...) LCONSOLE(D_OTHER, fmt, ## __VA_ARGS__)
55 #define QOS_DEBUG(fmt, ...)
56 #define QOS_CONSOLE(fmt, ...)
59 #define TGT_BAVAIL(i) (OST_TGT(lod,i)->ltd_statfs.os_bavail * \
60 OST_TGT(lod,i)->ltd_statfs.os_bsize)
62 int qos_add_tgt(struct lod_device *lod, struct lod_tgt_desc *ost_desc)
64 struct lod_qos_oss *oss = NULL, *temposs;
65 struct obd_export *exp = ost_desc->ltd_exp;
66 int rc = 0, found = 0;
70 down_write(&lod->lod_qos.lq_rw_sem);
72 * a bit hacky approach to learn NID of corresponding connection
73 * but there is no official API to access information like this
76 cfs_list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
77 if (obd_uuid_equals(&oss->lqo_uuid,
78 &exp->exp_connection->c_remote_uuid)) {
87 GOTO(out, rc = -ENOMEM);
88 memcpy(&oss->lqo_uuid, &exp->exp_connection->c_remote_uuid,
89 sizeof(oss->lqo_uuid));
91 /* Assume we have to move this one */
92 cfs_list_del(&oss->lqo_oss_list);
96 ost_desc->ltd_qos.ltq_oss = oss;
98 CDEBUG(D_QOS, "add tgt %s to OSS %s (%d OSTs)\n",
99 obd_uuid2str(&ost_desc->ltd_uuid), obd_uuid2str(&oss->lqo_uuid),
102 /* Add sorted by # of OSTs. Find the first entry that we're
104 list = &lod->lod_qos.lq_oss_list;
105 cfs_list_for_each_entry(temposs, list, lqo_oss_list) {
106 if (oss->lqo_ost_count > temposs->lqo_ost_count)
109 /* ...and add before it. If we're the first or smallest, temposs
110 points to the list head, and we add to the end. */
111 cfs_list_add_tail(&oss->lqo_oss_list, &temposs->lqo_oss_list);
113 lod->lod_qos.lq_dirty = 1;
114 lod->lod_qos.lq_rr.lqr_dirty = 1;
117 up_write(&lod->lod_qos.lq_rw_sem);
121 int qos_del_tgt(struct lod_device *lod, struct lod_tgt_desc *ost_desc)
123 struct lod_qos_oss *oss;
127 down_write(&lod->lod_qos.lq_rw_sem);
128 oss = ost_desc->ltd_qos.ltq_oss;
130 GOTO(out, rc = -ENOENT);
132 oss->lqo_ost_count--;
133 if (oss->lqo_ost_count == 0) {
134 CDEBUG(D_QOS, "removing OSS %s\n",
135 obd_uuid2str(&oss->lqo_uuid));
136 cfs_list_del(&oss->lqo_oss_list);
137 ost_desc->ltd_qos.ltq_oss = NULL;
141 lod->lod_qos.lq_dirty = 1;
142 lod->lod_qos.lq_rr.lqr_dirty = 1;
144 up_write(&lod->lod_qos.lq_rw_sem);
148 static int lod_statfs_and_check(const struct lu_env *env, struct lod_device *d,
149 int index, struct obd_statfs *sfs)
151 struct lod_tgt_desc *ost;
155 ost = OST_TGT(d,index);
158 rc = dt_statfs(env, ost->ltd_ost, sfs);
159 if (rc && rc != -ENOTCONN)
160 CERROR("%s: statfs: rc = %d\n", lod2obd(d)->obd_name, rc);
162 /* If the OST is readonly then we can't allocate objects there */
163 if (sfs->os_state & OS_STATE_READONLY)
166 /* check whether device has changed state (active, inactive) */
167 if (rc != 0 && ost->ltd_active) {
168 /* turned inactive? */
169 spin_lock(&d->lod_desc_lock);
170 if (ost->ltd_active) {
172 LASSERT(d->lod_desc.ld_active_tgt_count > 0);
173 d->lod_desc.ld_active_tgt_count--;
174 d->lod_qos.lq_dirty = 1;
175 d->lod_qos.lq_rr.lqr_dirty = 1;
176 CDEBUG(D_CONFIG, "%s: turns inactive\n",
177 ost->ltd_exp->exp_obd->obd_name);
179 spin_unlock(&d->lod_desc_lock);
180 } else if (rc == 0 && ost->ltd_active == 0) {
182 LASSERTF(d->lod_desc.ld_active_tgt_count < d->lod_ostnr,
183 "active tgt count %d, ost nr %d\n",
184 d->lod_desc.ld_active_tgt_count, d->lod_ostnr);
185 spin_lock(&d->lod_desc_lock);
186 if (ost->ltd_active == 0) {
188 d->lod_desc.ld_active_tgt_count++;
189 d->lod_qos.lq_dirty = 1;
190 d->lod_qos.lq_rr.lqr_dirty = 1;
191 CDEBUG(D_CONFIG, "%s: turns active\n",
192 ost->ltd_exp->exp_obd->obd_name);
194 spin_unlock(&d->lod_desc_lock);
200 static void lod_qos_statfs_update(const struct lu_env *env,
201 struct lod_device *lod)
203 struct obd_device *obd = lod2obd(lod);
204 struct ost_pool *osts = &(lod->lod_pool_info);
206 __u64 max_age, avail;
209 max_age = cfs_time_shift_64(-2 * lod->lod_desc.ld_qos_maxage);
211 if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
212 /* statfs data are quite recent, don't need to refresh it */
215 down_write(&lod->lod_qos.lq_rw_sem);
216 if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
219 for (i = 0; i < osts->op_count; i++) {
220 idx = osts->op_array[i];
221 avail = OST_TGT(lod,idx)->ltd_statfs.os_bavail;
222 rc = lod_statfs_and_check(env, lod, idx,
223 &OST_TGT(lod,idx)->ltd_statfs);
226 if (OST_TGT(lod,idx)->ltd_statfs.os_bavail != avail)
227 /* recalculate weigths */
228 lod->lod_qos.lq_dirty = 1;
230 obd->obd_osfs_age = cfs_time_current_64();
233 up_write(&lod->lod_qos.lq_rw_sem);
237 /* Recalculate per-object penalties for OSSs and OSTs,
238 depends on size of each ost in an oss */
239 static int lod_qos_calc_ppo(struct lod_device *lod)
241 struct lod_qos_oss *oss;
242 __u64 ba_max, ba_min, temp;
244 int rc, i, prio_wide;
248 if (!lod->lod_qos.lq_dirty)
251 num_active = lod->lod_desc.ld_active_tgt_count - 1;
253 GOTO(out, rc = -EAGAIN);
255 /* find bavail on each OSS */
256 cfs_list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list)
258 lod->lod_qos.lq_active_oss_count = 0;
261 * How badly user wants to select OSTs "widely" (not recently chosen
262 * and not on recent OSS's). As opposed to "freely" (free space
265 prio_wide = 256 - lod->lod_qos.lq_prio_free;
267 ba_min = (__u64)(-1);
269 now = cfs_time_current_sec();
270 /* Calculate OST penalty per object
271 * (lod ref taken in lod_qos_prep_create()) */
272 cfs_foreach_bit(lod->lod_ost_bitmap, i) {
273 LASSERT(OST_TGT(lod,i));
274 temp = TGT_BAVAIL(i);
277 ba_min = min(temp, ba_min);
278 ba_max = max(temp, ba_max);
280 /* Count the number of usable OSS's */
281 if (OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail == 0)
282 lod->lod_qos.lq_active_oss_count++;
283 OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail += temp;
285 /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
287 do_div(temp, num_active);
288 OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj =
289 (temp * prio_wide) >> 8;
291 age = (now - OST_TGT(lod,i)->ltd_qos.ltq_used) >> 3;
292 if (lod->lod_qos.lq_reset ||
293 age > 32 * lod->lod_desc.ld_qos_maxage)
294 OST_TGT(lod,i)->ltd_qos.ltq_penalty = 0;
295 else if (age > lod->lod_desc.ld_qos_maxage)
296 /* Decay the penalty by half for every 8x the update
297 * interval that the device has been idle. That gives
298 * lots of time for the statfs information to be
299 * updated (which the penalty is only a proxy for),
300 * and avoids penalizing OSS/OSTs under light load. */
301 OST_TGT(lod,i)->ltd_qos.ltq_penalty >>=
302 (age / lod->lod_desc.ld_qos_maxage);
305 num_active = lod->lod_qos.lq_active_oss_count - 1;
306 if (num_active < 1) {
307 /* If there's only 1 OSS, we can't penalize it, so instead
308 we have to double the OST penalty */
310 cfs_foreach_bit(lod->lod_ost_bitmap, i)
311 OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj <<= 1;
314 /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
315 cfs_list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
316 temp = oss->lqo_bavail >> 1;
317 do_div(temp, oss->lqo_ost_count * num_active);
318 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
320 age = (now - oss->lqo_used) >> 3;
321 if (lod->lod_qos.lq_reset ||
322 age > 32 * lod->lod_desc.ld_qos_maxage)
323 oss->lqo_penalty = 0;
324 else if (age > lod->lod_desc.ld_qos_maxage)
325 /* Decay the penalty by half for every 8x the update
326 * interval that the device has been idle. That gives
327 * lots of time for the statfs information to be
328 * updated (which the penalty is only a proxy for),
329 * and avoids penalizing OSS/OSTs under light load. */
330 oss->lqo_penalty >>= age / lod->lod_desc.ld_qos_maxage;
333 lod->lod_qos.lq_dirty = 0;
334 lod->lod_qos.lq_reset = 0;
336 /* If each ost has almost same free space,
337 * do rr allocation for better creation performance */
338 lod->lod_qos.lq_same_space = 0;
339 if ((ba_max * (256 - lod->lod_qos.lq_threshold_rr)) >> 8 < ba_min) {
340 lod->lod_qos.lq_same_space = 1;
341 /* Reset weights for the next time we enter qos mode */
342 lod->lod_qos.lq_reset = 1;
348 if (!rc && lod->lod_qos.lq_same_space)
354 static int lod_qos_calc_weight(struct lod_device *lod, int i)
358 /* Final ost weight = TGT_BAVAIL - ost_penalty - oss_penalty */
359 temp = TGT_BAVAIL(i);
360 temp2 = OST_TGT(lod,i)->ltd_qos.ltq_penalty +
361 OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_penalty;
363 OST_TGT(lod,i)->ltd_qos.ltq_weight = 0;
365 OST_TGT(lod,i)->ltd_qos.ltq_weight = temp - temp2;
369 /* We just used this index for a stripe; adjust everyone's weights */
370 static int lod_qos_used(struct lod_device *lod, struct ost_pool *osts,
371 __u32 index, __u64 *total_wt)
373 struct lod_tgt_desc *ost;
374 struct lod_qos_oss *oss;
378 ost = OST_TGT(lod,index);
381 /* Don't allocate on this devuce anymore, until the next alloc_qos */
382 ost->ltd_qos.ltq_usable = 0;
384 oss = ost->ltd_qos.ltq_oss;
386 /* Decay old penalty by half (we're adding max penalty, and don't
387 want it to run away.) */
388 ost->ltd_qos.ltq_penalty >>= 1;
389 oss->lqo_penalty >>= 1;
391 /* mark the OSS and OST as recently used */
392 ost->ltd_qos.ltq_used = oss->lqo_used = cfs_time_current_sec();
394 /* Set max penalties for this OST and OSS */
395 ost->ltd_qos.ltq_penalty +=
396 ost->ltd_qos.ltq_penalty_per_obj * lod->lod_ostnr;
397 oss->lqo_penalty += oss->lqo_penalty_per_obj *
398 lod->lod_qos.lq_active_oss_count;
400 /* Decrease all OSS penalties */
401 cfs_list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
402 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
403 oss->lqo_penalty = 0;
405 oss->lqo_penalty -= oss->lqo_penalty_per_obj;
409 /* Decrease all OST penalties */
410 for (j = 0; j < osts->op_count; j++) {
413 i = osts->op_array[j];
414 if (!cfs_bitmap_check(lod->lod_ost_bitmap, i))
417 ost = OST_TGT(lod,i);
420 if (ost->ltd_qos.ltq_penalty <
421 ost->ltd_qos.ltq_penalty_per_obj)
422 ost->ltd_qos.ltq_penalty = 0;
424 ost->ltd_qos.ltq_penalty -=
425 ost->ltd_qos.ltq_penalty_per_obj;
427 lod_qos_calc_weight(lod, i);
429 /* Recalc the total weight of usable osts */
430 if (ost->ltd_qos.ltq_usable)
431 *total_wt += ost->ltd_qos.ltq_weight;
433 QOS_DEBUG("recalc tgt %d usable=%d avail="LPU64
434 " ostppo="LPU64" ostp="LPU64" ossppo="LPU64
435 " ossp="LPU64" wt="LPU64"\n",
436 i, ost->ltd_qos.ltq_usable, TGT_BAVAIL(i) >> 10,
437 ost->ltd_qos.ltq_penalty_per_obj >> 10,
438 ost->ltd_qos.ltq_penalty >> 10,
439 ost->ltd_qos.ltq_oss->lqo_penalty_per_obj >> 10,
440 ost->ltd_qos.ltq_oss->lqo_penalty >> 10,
441 ost->ltd_qos.ltq_weight >> 10);
447 #define LOV_QOS_EMPTY ((__u32)-1)
448 /* compute optimal round-robin order, based on OSTs per OSS */
449 static int lod_qos_calc_rr(struct lod_device *lod, struct ost_pool *src_pool,
450 struct lod_qos_rr *lqr)
452 struct lod_qos_oss *oss;
453 struct lod_tgt_desc *ost;
454 unsigned placed, real_count;
458 if (!lqr->lqr_dirty) {
459 LASSERT(lqr->lqr_pool.op_size);
463 /* Do actual allocation. */
464 down_write(&lod->lod_qos.lq_rw_sem);
467 * Check again. While we were sleeping on @lq_rw_sem something could
470 if (!lqr->lqr_dirty) {
471 LASSERT(lqr->lqr_pool.op_size);
472 up_write(&lod->lod_qos.lq_rw_sem);
476 real_count = src_pool->op_count;
478 /* Zero the pool array */
479 /* alloc_rr is holding a read lock on the pool, so nobody is adding/
480 deleting from the pool. The lq_rw_sem insures that nobody else
482 lqr->lqr_pool.op_count = real_count;
483 rc = lod_ost_pool_extend(&lqr->lqr_pool, real_count);
485 up_write(&lod->lod_qos.lq_rw_sem);
488 for (i = 0; i < lqr->lqr_pool.op_count; i++)
489 lqr->lqr_pool.op_array[i] = LOV_QOS_EMPTY;
491 /* Place all the OSTs from 1 OSS at the same time. */
493 cfs_list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
496 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
499 if (!cfs_bitmap_check(lod->lod_ost_bitmap,
500 src_pool->op_array[i]))
503 ost = OST_TGT(lod,src_pool->op_array[i]);
504 LASSERT(ost && ost->ltd_ost);
505 if (ost->ltd_qos.ltq_oss != oss)
508 /* Evenly space these OSTs across arrayspace */
509 next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count;
510 while (lqr->lqr_pool.op_array[next] != LOV_QOS_EMPTY)
511 next = (next + 1) % lqr->lqr_pool.op_count;
513 lqr->lqr_pool.op_array[next] = src_pool->op_array[i];
520 up_write(&lod->lod_qos.lq_rw_sem);
522 if (placed != real_count) {
523 /* This should never happen */
524 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
525 "round-robin list (%d of %d).\n",
527 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
528 LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
529 lqr->lqr_pool.op_array[i]);
536 for (i = 0; i < lqr->lqr_pool.op_count; i++)
537 QOS_CONSOLE("rr #%d ost idx=%d\n", i, lqr->lqr_pool.op_array[i]);
544 * A helper function to:
545 * create in-core lu object on the specified OSP
546 * declare creation of the object
547 * IMPORTANT: at this stage object is anonymouos - it has no fid assigned
548 * this is a workaround till we have natural FIDs on OST
550 * at this point we want to declare (reserve) object for us as
551 * we can't block at execution (when create method is called).
552 * otherwise we'd block whole transaction batch
554 static struct dt_object *lod_qos_declare_object_on(const struct lu_env *env,
555 struct lod_device *d,
559 struct lod_tgt_desc *ost;
560 struct lu_object *o, *n;
561 struct lu_device *nd;
562 struct dt_object *dt;
567 LASSERT(ost_idx >= 0);
568 LASSERT(ost_idx < d->lod_osts_size);
569 ost = OST_TGT(d,ost_idx);
571 LASSERT(ost->ltd_ost);
573 nd = &ost->ltd_ost->dd_lu_dev;
576 * allocate anonymous object with zero fid, real fid
577 * will be assigned by OSP within transaction
578 * XXX: to be fixed with fully-functional OST fids
580 o = lu_object_anon(env, nd, NULL);
582 GOTO(out, dt = ERR_PTR(PTR_ERR(o)));
584 n = lu_object_locate(o->lo_header, nd->ld_type);
585 if (unlikely(n == NULL)) {
586 CERROR("can't find slice\n");
587 lu_object_put(env, o);
588 GOTO(out, dt = ERR_PTR(-EINVAL));
591 dt = container_of(n, struct dt_object, do_lu);
593 rc = dt_declare_create(env, dt, NULL, NULL, NULL, th);
595 CDEBUG(D_OTHER, "can't declare creation on #%u: %d\n",
597 lu_object_put(env, o);
605 static int min_stripe_count(int stripe_cnt, int flags)
607 return (flags & LOV_USES_DEFAULT_STRIPE ?
608 stripe_cnt - (stripe_cnt / 4) : stripe_cnt);
611 #define LOV_CREATE_RESEED_MULT 30
612 #define LOV_CREATE_RESEED_MIN 2000
614 static int inline lod_qos_dev_is_full(struct obd_statfs *msfs)
617 int bs = msfs->os_bsize;
619 LASSERT(((bs - 1) & bs) == 0);
621 /* the minimum of 0.1% used blocks and 1GB bytes. */
622 used = min_t(__u64, (msfs->os_blocks - msfs->os_bfree) >> 10,
623 1 << (31 - ffs(bs)));
624 return (msfs->os_bavail < used);
627 static inline int lod_qos_ost_in_use_clear(const struct lu_env *env, int stripes)
629 struct lod_thread_info *info = lod_env_info(env);
631 if (info->lti_ea_store_size < sizeof(int) * stripes)
632 lod_ea_store_resize(info, stripes * sizeof(int));
633 if (info->lti_ea_store_size < sizeof(int) * stripes) {
634 CERROR("can't allocate memory for ost-in-use array\n");
637 memset(info->lti_ea_store, -1, sizeof(int) * stripes);
641 static inline void lod_qos_ost_in_use(const struct lu_env *env, int idx, int ost)
643 struct lod_thread_info *info = lod_env_info(env);
644 int *osts = info->lti_ea_store;
646 LASSERT(info->lti_ea_store_size >= idx * sizeof(int));
650 static int lod_qos_is_ost_used(const struct lu_env *env, int ost, int stripes)
652 struct lod_thread_info *info = lod_env_info(env);
653 int *osts = info->lti_ea_store;
656 for (j = 0; j < stripes; j++) {
663 /* Allocate objects on OSTs with round-robin algorithm */
664 static int lod_alloc_rr(const struct lu_env *env, struct lod_object *lo,
665 struct dt_object **stripe, int flags,
668 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
669 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
670 struct pool_desc *pool = NULL;
671 struct ost_pool *osts;
672 struct lod_qos_rr *lqr;
676 int ost_start_idx_temp;
679 int stripe_cnt = lo->ldo_stripenr;
680 int stripe_cnt_min = min_stripe_count(stripe_cnt, flags);
685 pool = lod_find_pool(m, lo->ldo_pool);
688 down_read(&pool_tgt_rw_sem(pool));
689 osts = &(pool->pool_obds);
690 lqr = &(pool->pool_rr);
692 osts = &(m->lod_pool_info);
693 lqr = &(m->lod_qos.lq_rr);
696 rc = lod_qos_calc_rr(m, osts, lqr);
700 rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
704 if (--lqr->lqr_start_count <= 0) {
705 lqr->lqr_start_idx = cfs_rand() % osts->op_count;
706 lqr->lqr_start_count =
707 (LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) +
708 LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U);
709 } else if (stripe_cnt_min >= osts->op_count ||
710 lqr->lqr_start_idx > osts->op_count) {
711 /* If we have allocated from all of the OSTs, slowly
712 * precess the next start if the OST/stripe count isn't
713 * already doing this for us. */
714 lqr->lqr_start_idx %= osts->op_count;
715 if (stripe_cnt > 1 && (osts->op_count % stripe_cnt) != 1)
716 ++lqr->lqr_offset_idx;
718 down_read(&m->lod_qos.lq_rw_sem);
719 ost_start_idx_temp = lqr->lqr_start_idx;
722 array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) %
725 QOS_DEBUG("pool '%s' want %d startidx %d startcnt %d offset %d "
726 "active %d count %d arrayidx %d\n",
727 lo->ldo_pool ? lo->ldo_pool : "",
728 stripe_cnt, lqr->lqr_start_idx, lqr->lqr_start_count,
729 lqr->lqr_offset_idx, osts->op_count, osts->op_count,
732 for (i = 0; i < osts->op_count && stripe_idx < lo->ldo_stripenr;
733 i++, array_idx = (array_idx + 1) % osts->op_count) {
734 ++lqr->lqr_start_idx;
735 ost_idx = lqr->lqr_pool.op_array[array_idx];
737 QOS_DEBUG("#%d strt %d act %d strp %d ary %d idx %d\n",
738 i, lqr->lqr_start_idx, /* XXX: active*/ 0,
739 stripe_idx, array_idx, ost_idx);
741 if ((ost_idx == LOV_QOS_EMPTY) ||
742 !cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
745 /* Fail Check before osc_precreate() is called
746 so we can only 'fail' single OSC. */
747 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
750 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
752 /* this OSP doesn't feel well */
759 if (lod_qos_dev_is_full(sfs)) {
760 QOS_DEBUG("#%d is full\n", ost_idx);
765 * We expect number of precreated objects in f_ffree at
766 * the first iteration, skip OSPs with no objects ready
768 if (sfs->os_fprecreated == 0 && speed == 0) {
769 QOS_DEBUG("#%d: precreation is empty\n", ost_idx);
774 * try to use another OSP if this one is degraded
776 if (sfs->os_state & OS_STATE_DEGRADED && speed < 2) {
777 QOS_DEBUG("#%d: degraded\n", ost_idx);
782 * do not put >1 objects on a single OST
784 if (speed && lod_qos_is_ost_used(env, ost_idx, stripe_idx))
787 o = lod_qos_declare_object_on(env, m, ost_idx, th);
789 CDEBUG(D_OTHER, "can't declare new object on #%u: %d\n",
790 ost_idx, (int) PTR_ERR(o));
796 * We've successfuly declared (reserved) an object
798 lod_qos_ost_in_use(env, stripe_idx, ost_idx);
799 stripe[stripe_idx] = o;
803 if ((speed < 2) && (stripe_idx < stripe_cnt_min)) {
804 /* Try again, allowing slower OSCs */
806 lqr->lqr_start_idx = ost_start_idx_temp;
810 up_read(&m->lod_qos.lq_rw_sem);
813 lo->ldo_stripenr = stripe_idx;
814 /* at least one stripe is allocated */
817 /* nobody provided us with a single object */
823 up_read(&pool_tgt_rw_sem(pool));
824 /* put back ref got by lod_find_pool() */
825 lod_pool_putref(pool);
831 /* alloc objects on osts with specific stripe offset */
832 static int lod_alloc_specific(const struct lu_env *env, struct lod_object *lo,
833 struct dt_object **stripe, int flags,
836 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
837 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
839 unsigned ost_idx, array_idx, ost_count;
840 int i, rc, stripe_num = 0;
842 struct pool_desc *pool = NULL;
843 struct ost_pool *osts;
846 rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
851 pool = lod_find_pool(m, lo->ldo_pool);
854 down_read(&pool_tgt_rw_sem(pool));
855 osts = &(pool->pool_obds);
857 osts = &(m->lod_pool_info);
860 ost_count = osts->op_count;
863 /* search loi_ost_idx in ost array */
865 for (i = 0; i < ost_count; i++) {
866 if (osts->op_array[i] == lo->ldo_def_stripe_offset) {
871 if (i == ost_count) {
872 CERROR("Start index %d not found in pool '%s'\n",
873 lo->ldo_def_stripe_offset,
874 lo->ldo_pool ? lo->ldo_pool : "");
875 GOTO(out, rc = -EINVAL);
878 for (i = 0; i < ost_count;
879 i++, array_idx = (array_idx + 1) % ost_count) {
880 ost_idx = osts->op_array[array_idx];
882 if (!cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
885 /* Fail Check before osc_precreate() is called
886 so we can only 'fail' single OSC. */
887 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
891 * do not put >1 objects on a single OST
893 if (lod_qos_is_ost_used(env, ost_idx, stripe_num))
896 /* Drop slow OSCs if we can, but not for requested start idx.
898 * This means "if OSC is slow and it is not the requested
899 * start OST, then it can be skipped, otherwise skip it only
900 * if it is inactive/recovering/out-of-space." */
902 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
904 /* this OSP doesn't feel well */
909 * We expect number of precreated objects in f_ffree at
910 * the first iteration, skip OSPs with no objects ready
911 * don't apply this logic to OST specified with stripe_offset
913 if (i != 0 && sfs->os_fprecreated == 0 && speed == 0)
916 o = lod_qos_declare_object_on(env, m, ost_idx, th);
918 CDEBUG(D_OTHER, "can't declare new object on #%u: %d\n",
919 ost_idx, (int) PTR_ERR(o));
924 * We've successfuly declared (reserved) an object
926 lod_qos_ost_in_use(env, stripe_num, ost_idx);
927 stripe[stripe_num] = o;
930 /* We have enough stripes */
931 if (stripe_num == lo->ldo_stripenr)
935 /* Try again, allowing slower OSCs */
940 /* If we were passed specific striping params, then a failure to
941 * meet those requirements is an error, since we can't reallocate
942 * that memory (it might be part of a larger array or something).
944 * We can only get here if lsm_stripe_count was originally > 1.
946 CERROR("can't lstripe objid "DFID": have %d want %u\n",
947 PFID(lu_object_fid(lod2lu_obj(lo))), stripe_num,
952 up_read(&pool_tgt_rw_sem(pool));
953 /* put back ref got by lod_find_pool() */
954 lod_pool_putref(pool);
960 static inline int lod_qos_is_usable(struct lod_device *lod)
963 /* to be able to debug QoS code */
967 /* Detect -EAGAIN early, before expensive lock is taken. */
968 if (!lod->lod_qos.lq_dirty && lod->lod_qos.lq_same_space)
971 if (lod->lod_desc.ld_active_tgt_count < 2)
977 /* Alloc objects on OSTs with optimization based on:
979 - network resources (shared OSS's)
981 static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo,
982 struct dt_object **stripe, int flags,
985 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
986 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
987 struct lod_tgt_desc *ost;
989 __u64 total_weight = 0;
990 int nfound, good_osts, i, rc = 0;
991 int stripe_cnt = lo->ldo_stripenr;
993 struct pool_desc *pool = NULL;
994 struct ost_pool *osts;
997 stripe_cnt_min = min_stripe_count(stripe_cnt, flags);
998 if (stripe_cnt_min < 1)
1002 pool = lod_find_pool(m, lo->ldo_pool);
1005 down_read(&pool_tgt_rw_sem(pool));
1006 osts = &(pool->pool_obds);
1008 osts = &(m->lod_pool_info);
1011 /* Detect -EAGAIN early, before expensive lock is taken. */
1012 if (!lod_qos_is_usable(m))
1013 GOTO(out_nolock, rc = -EAGAIN);
1015 /* Do actual allocation, use write lock here. */
1016 down_write(&m->lod_qos.lq_rw_sem);
1019 * Check again, while we were sleeping on @lq_rw_sem things could
1022 if (!lod_qos_is_usable(m))
1023 GOTO(out, rc = -EAGAIN);
1025 rc = lod_qos_calc_ppo(m);
1029 rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
1034 /* Find all the OSTs that are valid stripe candidates */
1035 for (i = 0; i < osts->op_count; i++) {
1036 if (!cfs_bitmap_check(m->lod_ost_bitmap, osts->op_array[i]))
1039 rc = lod_statfs_and_check(env, m, osts->op_array[i], sfs);
1041 /* this OSP doesn't feel well */
1048 if (lod_qos_dev_is_full(sfs))
1051 /* Fail Check before osc_precreate() is called
1052 so we can only 'fail' single OSC. */
1053 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) &&
1054 osts->op_array[i] == 0)
1057 ost = OST_TGT(m,osts->op_array[i]);
1058 ost->ltd_qos.ltq_usable = 1;
1059 lod_qos_calc_weight(m, osts->op_array[i]);
1060 total_weight += ost->ltd_qos.ltq_weight;
1065 QOS_DEBUG("found %d good osts\n", good_osts);
1067 if (good_osts < stripe_cnt_min)
1068 GOTO(out, rc = -EAGAIN);
1070 /* We have enough osts */
1071 if (good_osts < stripe_cnt)
1072 stripe_cnt = good_osts;
1074 /* Find enough OSTs with weighted random allocation. */
1076 while (nfound < stripe_cnt) {
1077 __u64 rand, cur_weight;
1083 #if BITS_PER_LONG == 32
1084 rand = cfs_rand() % (unsigned)total_weight;
1085 /* If total_weight > 32-bit, first generate the high
1086 * 32 bits of the random number, then add in the low
1087 * 32 bits (truncated to the upper limit, if needed) */
1088 if (total_weight > 0xffffffffULL)
1089 rand = (__u64)(cfs_rand() %
1090 (unsigned)(total_weight >> 32)) << 32;
1094 if (rand == (total_weight & 0xffffffff00000000ULL))
1095 rand |= cfs_rand() % (unsigned)total_weight;
1100 rand = ((__u64)cfs_rand() << 32 | cfs_rand()) %
1107 /* On average, this will hit larger-weighted osts more often.
1108 0-weight osts will always get used last (only when rand=0) */
1109 for (i = 0; i < osts->op_count; i++) {
1110 int idx = osts->op_array[i];
1112 if (!cfs_bitmap_check(m->lod_ost_bitmap, idx))
1115 ost = OST_TGT(m,idx);
1117 if (!ost->ltd_qos.ltq_usable)
1120 cur_weight += ost->ltd_qos.ltq_weight;
1121 QOS_DEBUG("stripe_cnt=%d nfound=%d cur_weight="LPU64
1122 " rand="LPU64" total_weight="LPU64"\n",
1123 stripe_cnt, nfound, cur_weight, rand,
1126 if (cur_weight < rand)
1129 QOS_DEBUG("stripe=%d to idx=%d\n", nfound, idx);
1132 * do not put >1 objects on a single OST
1134 if (lod_qos_is_ost_used(env, idx, nfound))
1136 lod_qos_ost_in_use(env, nfound, idx);
1138 o = lod_qos_declare_object_on(env, m, idx, th);
1140 QOS_DEBUG("can't declare object on #%u: %d\n",
1141 idx, (int) PTR_ERR(o));
1144 stripe[nfound++] = o;
1145 lod_qos_used(m, osts, idx, &total_weight);
1151 /* no OST found on this iteration, give up */
1156 if (unlikely(nfound != stripe_cnt)) {
1158 * when the decision to use weighted algorithm was made
1159 * we had enough appropriate OSPs, but this state can
1160 * change anytime (no space on OST, broken connection, etc)
1161 * so it's possible OSP won't be able to provide us with
1162 * an object due to just changed state
1164 LCONSOLE_INFO("wanted %d, found %d\n", stripe_cnt, nfound);
1165 for (i = 0; i < nfound; i++) {
1166 LASSERT(stripe[i] != NULL);
1167 lu_object_put(env, &stripe[i]->do_lu);
1171 /* makes sense to rebalance next time */
1172 m->lod_qos.lq_dirty = 1;
1173 m->lod_qos.lq_same_space = 0;
1179 up_write(&m->lod_qos.lq_rw_sem);
1183 up_read(&pool_tgt_rw_sem(pool));
1184 /* put back ref got by lod_find_pool() */
1185 lod_pool_putref(pool);
1191 /* Find the max stripecount we should use */
1192 static __u16 lod_get_stripecnt(struct lod_device *lod, __u32 magic,
1195 __u32 max_stripes = LOV_MAX_STRIPE_COUNT_OLD;
1198 stripe_count = lod->lod_desc.ld_default_stripe_count;
1199 if (stripe_count > lod->lod_desc.ld_active_tgt_count)
1200 stripe_count = lod->lod_desc.ld_active_tgt_count;
1204 /* stripe count is based on whether OSD can handle larger EA sizes */
1205 if (lod->lod_osd_max_easize > 0)
1206 max_stripes = lov_mds_md_max_stripe_count(
1207 lod->lod_osd_max_easize, magic);
1209 return (stripe_count < max_stripes) ? stripe_count : max_stripes;
1212 static int lod_use_defined_striping(const struct lu_env *env,
1213 struct lod_object *mo,
1214 const struct lu_buf *buf)
1216 struct lov_mds_md_v1 *v1 = buf->lb_buf;
1217 struct lov_mds_md_v3 *v3 = buf->lb_buf;
1218 struct lov_ost_data_v1 *objs;
1223 magic = le32_to_cpu(v1->lmm_magic);
1224 if (magic == LOV_MAGIC_V1_DEF) {
1225 magic = LOV_MAGIC_V1;
1226 objs = &v1->lmm_objects[0];
1227 } else if (magic == LOV_MAGIC_V3_DEF) {
1228 magic = LOV_MAGIC_V3;
1229 objs = &v3->lmm_objects[0];
1230 lod_object_set_pool(mo, v3->lmm_pool_name);
1232 GOTO(out, rc = -EINVAL);
1235 mo->ldo_pattern = le32_to_cpu(v1->lmm_pattern);
1236 mo->ldo_stripe_size = le32_to_cpu(v1->lmm_stripe_size);
1237 mo->ldo_stripenr = le16_to_cpu(v1->lmm_stripe_count);
1238 mo->ldo_layout_gen = le16_to_cpu(v1->lmm_layout_gen);
1240 /* fixup for released file before object initialization */
1241 if (mo->ldo_pattern & LOV_PATTERN_F_RELEASED) {
1242 mo->ldo_released_stripenr = mo->ldo_stripenr;
1243 mo->ldo_stripenr = 0;
1246 LASSERT(buf->lb_len >= lov_mds_md_size(mo->ldo_stripenr, magic));
1248 if (mo->ldo_stripenr > 0)
1249 rc = lod_initialize_objects(env, mo, objs);
1255 static int lod_qos_parse_config(const struct lu_env *env,
1256 struct lod_object *lo,
1257 const struct lu_buf *buf)
1259 struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
1260 struct lov_user_md_v1 *v1 = NULL;
1261 struct lov_user_md_v3 *v3 = NULL;
1262 struct pool_desc *pool;
1267 if (buf == NULL || buf->lb_buf == NULL || buf->lb_len == 0)
1271 magic = v1->lmm_magic;
1273 if (magic == __swab32(LOV_USER_MAGIC_V1)) {
1274 lustre_swab_lov_user_md_v1(v1);
1275 magic = v1->lmm_magic;
1276 } else if (magic == __swab32(LOV_USER_MAGIC_V3)) {
1278 lustre_swab_lov_user_md_v3(v3);
1279 magic = v3->lmm_magic;
1282 if (unlikely(magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3)) {
1283 /* try to use as fully defined striping */
1284 rc = lod_use_defined_striping(env, lo, buf);
1288 if (unlikely(buf->lb_len < sizeof(*v1))) {
1289 CERROR("wrong size: %u\n", (unsigned) buf->lb_len);
1293 v1->lmm_magic = magic;
1294 if (v1->lmm_pattern == 0)
1295 v1->lmm_pattern = LOV_PATTERN_RAID0;
1296 if (lov_pattern(v1->lmm_pattern) != LOV_PATTERN_RAID0) {
1297 CERROR("invalid pattern: %x\n", v1->lmm_pattern);
1300 lo->ldo_pattern = v1->lmm_pattern;
1302 if (v1->lmm_stripe_size)
1303 lo->ldo_stripe_size = v1->lmm_stripe_size;
1304 if (lo->ldo_stripe_size & (LOV_MIN_STRIPE_SIZE - 1))
1305 lo->ldo_stripe_size = LOV_MIN_STRIPE_SIZE;
1307 if (v1->lmm_stripe_count)
1308 lo->ldo_stripenr = v1->lmm_stripe_count;
1310 if ((v1->lmm_stripe_offset >= d->lod_desc.ld_tgt_count) &&
1311 (v1->lmm_stripe_offset != (typeof(v1->lmm_stripe_offset))(-1))) {
1312 CERROR("invalid offset: %x\n", v1->lmm_stripe_offset);
1315 lo->ldo_def_stripe_offset = v1->lmm_stripe_offset;
1317 CDEBUG(D_OTHER, "lsm: %u size, %u stripes, %u offset\n",
1318 v1->lmm_stripe_size, v1->lmm_stripe_count,
1319 v1->lmm_stripe_offset);
1321 if (v1->lmm_magic == LOV_MAGIC_V3) {
1322 if (buf->lb_len < sizeof(*v3)) {
1323 CERROR("wrong size: %u\n", (unsigned) buf->lb_len);
1328 lod_object_set_pool(lo, v3->lmm_pool_name);
1330 /* In the function below, .hs_keycmp resolves to
1331 * pool_hashkey_keycmp() */
1332 /* coverity[overrun-buffer-val] */
1333 pool = lod_find_pool(d, v3->lmm_pool_name);
1335 if (lo->ldo_def_stripe_offset !=
1336 (typeof(v1->lmm_stripe_offset))(-1)) {
1337 rc = lo->ldo_def_stripe_offset;
1338 rc = lod_check_index_in_pool(rc, pool);
1340 lod_pool_putref(pool);
1341 CERROR("invalid offset\n");
1346 if (lo->ldo_stripenr > pool_tgt_count(pool))
1347 lo->ldo_stripenr= pool_tgt_count(pool);
1349 lod_pool_putref(pool);
1352 lod_object_set_pool(lo, NULL);
1354 /* fixup for released file */
1355 if (lo->ldo_pattern & LOV_PATTERN_F_RELEASED) {
1356 lo->ldo_released_stripenr = lo->ldo_stripenr;
1357 lo->ldo_stripenr = 0;
1364 * buf should be NULL or contain striping settings
1366 int lod_qos_prep_create(const struct lu_env *env, struct lod_object *lo,
1367 struct lu_attr *attr, const struct lu_buf *buf,
1370 struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
1371 struct dt_object **stripe;
1373 int flag = LOV_USES_ASSIGNED_STRIPE;
1379 /* no OST available */
1380 /* XXX: should we be waiting a bit to prevent failures during
1381 * cluster initialization? */
1382 if (d->lod_ostnr == 0)
1383 GOTO(out, rc = -EIO);
1386 * by this time, the object's ldo_stripenr and ldo_stripe_size
1387 * contain default value for striping: taken from the parent
1388 * or from filesystem defaults
1390 * in case the caller is passing lovea with new striping config,
1391 * we may need to parse lovea and apply new configuration
1393 rc = lod_qos_parse_config(env, lo, buf);
1397 /* A released file is being created */
1398 if (lo->ldo_stripenr == 0)
1401 if (likely(lo->ldo_stripe == NULL)) {
1403 * no striping has been created so far
1405 LASSERT(lo->ldo_stripenr > 0);
1407 * statfs and check OST targets now, since ld_active_tgt_count
1408 * could be changed if some OSTs are [de]activated manually.
1410 lod_qos_statfs_update(env, d);
1411 lo->ldo_stripenr = lod_get_stripecnt(d, LOV_MAGIC,
1414 stripe_len = lo->ldo_stripenr;
1415 OBD_ALLOC(stripe, sizeof(stripe[0]) * stripe_len);
1417 GOTO(out, rc = -ENOMEM);
1419 lod_getref(&d->lod_ost_descs);
1420 /* XXX: support for non-0 files w/o objects */
1421 CDEBUG(D_OTHER, "tgt_count %d stripenr %d\n",
1422 d->lod_desc.ld_tgt_count, stripe_len);
1423 if (lo->ldo_def_stripe_offset >= d->lod_desc.ld_tgt_count) {
1424 rc = lod_alloc_qos(env, lo, stripe, flag, th);
1426 rc = lod_alloc_rr(env, lo, stripe, flag, th);
1428 rc = lod_alloc_specific(env, lo, stripe, flag, th);
1430 lod_putref(d, &d->lod_ost_descs);
1433 for (i = 0; i < stripe_len; i++)
1434 if (stripe[i] != NULL)
1435 lu_object_put(env, &stripe[i]->do_lu);
1437 OBD_FREE(stripe, sizeof(stripe[0]) * stripe_len);
1438 lo->ldo_stripenr = 0;
1440 lo->ldo_stripe = stripe;
1441 lo->ldo_stripes_allocated = stripe_len;
1445 * lod_qos_parse_config() found supplied buf as a predefined
1446 * striping (not a hint), so it allocated all the object
1447 * now we need to create them
1449 for (i = 0; i < lo->ldo_stripenr; i++) {
1450 struct dt_object *o;
1452 o = lo->ldo_stripe[i];
1455 rc = dt_declare_create(env, o, attr, NULL, NULL, th);
1457 CERROR("can't declare create: %d\n", rc);