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) 2011, 2012, Whamcloud, Inc.
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 <libcfs/libcfs.h>
39 #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_ost_desc *ost_desc)
64 struct lov_qos_oss *oss = NULL, *temposs;
65 struct obd_export *exp = ost_desc->ltd_exp;
66 int rc = 0, found = 0;
70 cfs_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 cfs_up_write(&lod->lod_qos.lq_rw_sem);
121 int qos_del_tgt(struct lod_device *lod, struct lod_ost_desc *ost_desc)
123 struct lov_qos_oss *oss;
127 cfs_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 cfs_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_ost_desc *ost;
155 ost = OST_TGT(d,index);
158 rc = dt_statfs(env, ost->ltd_ost, sfs);
162 /* check whether device has changed state (active, inactive) */
163 if (unlikely(sfs->os_blocks == 0 && ost->ltd_active)) {
164 /* turned inactive? */
165 cfs_spin_lock(&d->lod_desc_lock);
166 if (sfs->os_blocks == 0 && ost->ltd_active) {
168 LASSERT(d->lod_desc.ld_active_tgt_count > 0);
169 d->lod_desc.ld_active_tgt_count--;
170 d->lod_qos.lq_dirty = 1;
171 d->lod_qos.lq_rr.lqr_dirty = 1;
172 CDEBUG(D_CONFIG, "%s: turns inactive\n",
173 ost->ltd_exp->exp_obd->obd_name);
175 cfs_spin_unlock(&d->lod_desc_lock);
176 } else if (unlikely(sfs->os_blocks && ost->ltd_active == 0)) {
178 LASSERT(d->lod_desc.ld_active_tgt_count < d->lod_ostnr);
179 cfs_spin_lock(&d->lod_desc_lock);
180 if (sfs->os_blocks && ost->ltd_active == 0) {
182 d->lod_desc.ld_active_tgt_count++;
183 d->lod_qos.lq_dirty = 1;
184 d->lod_qos.lq_rr.lqr_dirty = 1;
185 CDEBUG(D_CONFIG, "%s: turns active\n",
186 ost->ltd_exp->exp_obd->obd_name);
188 cfs_spin_unlock(&d->lod_desc_lock);
195 * Update statfs data if the current osfs age is older than max_age.
196 * If wait is not set, it means that we are called from lov_create()
197 * and we should just issue the rpcs without waiting for them to complete.
198 * If wait is set, we are called from alloc_qos() and we just have
199 * to wait for the request set to complete.
201 static void lod_qos_statfs_update(const struct lu_env *env,
202 struct lod_device *lod)
204 struct obd_device *obd = lod2obd(lod);
205 struct ost_pool *osts = &(lod->lod_pool_info);
207 __u64 max_age, avail;
210 max_age = cfs_time_shift_64(-2*lod->lod_desc.ld_qos_maxage);
212 if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
213 /* statfs data are quite recent, don't need to refresh it */
216 cfs_down_write(&lod->lod_qos.lq_rw_sem);
217 if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
220 for (i = 0; i < osts->op_count; i++) {
221 idx = osts->op_array[i];
222 avail = OST_TGT(lod,idx)->ltd_statfs.os_bavail;
223 rc = lod_statfs_and_check(env, lod, idx,
224 &OST_TGT(lod,idx)->ltd_statfs);
226 /* XXX: disable this OST till next refresh? */
227 CERROR("can't refresh statfs: %d\n", rc);
230 if (OST_TGT(lod,idx)->ltd_statfs.os_bavail != avail)
231 /* recalculate weigths */
232 lod->lod_qos.lq_dirty = 1;
234 obd->obd_osfs_age = cfs_time_current_64();
237 cfs_up_write(&lod->lod_qos.lq_rw_sem);
240 /* Recalculate per-object penalties for OSSs and OSTs,
241 depends on size of each ost in an oss */
242 static int lod_qos_calc_ppo(struct lod_device *lod)
244 struct lov_qos_oss *oss;
245 __u64 ba_max, ba_min, temp;
247 int rc, i, prio_wide;
251 if (!lod->lod_qos.lq_dirty)
254 num_active = lod->lod_desc.ld_active_tgt_count - 1;
256 GOTO(out, rc = -EAGAIN);
258 /* find bavail on each OSS */
259 cfs_list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list)
261 lod->lod_qos.lq_active_oss_count = 0;
263 /* How badly user wants to select osts "widely" (not recently chosen
264 and not on recent oss's). As opposed to "freely" (free space
266 prio_wide = 256 - lod->lod_qos.lq_prio_free;
268 ba_min = (__u64)(-1);
270 now = cfs_time_current_sec();
271 /* Calculate OST penalty per object
272 * (lod ref taken in lod_qos_prep_create()) */
273 cfs_foreach_bit(lod->lod_ost_bitmap, i) {
274 LASSERT(OST_TGT(lod,i));
275 temp = TGT_BAVAIL(i);
278 ba_min = min(temp, ba_min);
279 ba_max = max(temp, ba_max);
281 /* Count the number of usable OSS's */
282 if (OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail == 0)
283 lod->lod_qos.lq_active_oss_count++;
284 OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail += temp;
286 /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
288 lov_do_div64(temp, num_active);
289 OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj =
290 (temp * prio_wide) >> 8;
292 age = (now - OST_TGT(lod,i)->ltd_qos.ltq_used) >> 3;
293 if (lod->lod_qos.lq_reset ||
294 age > 32 * lod->lod_desc.ld_qos_maxage)
295 OST_TGT(lod,i)->ltd_qos.ltq_penalty = 0;
296 else if (age > lod->lod_desc.ld_qos_maxage)
297 /* Decay the penalty by half for every 8x the update
298 * interval that the device has been idle. That gives
299 * lots of time for the statfs information to be
300 * updated (which the penalty is only a proxy for),
301 * and avoids penalizing OSS/OSTs under light load. */
302 OST_TGT(lod,i)->ltd_qos.ltq_penalty >>=
303 (age / lod->lod_desc.ld_qos_maxage);
306 num_active = lod->lod_qos.lq_active_oss_count - 1;
307 if (num_active < 1) {
308 /* If there's only 1 OSS, we can't penalize it, so instead
309 we have to double the OST penalty */
311 cfs_foreach_bit(lod->lod_ost_bitmap, i)
312 OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj <<= 1;
315 /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
316 cfs_list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
317 temp = oss->lqo_bavail >> 1;
318 lov_do_div64(temp, oss->lqo_ost_count * num_active);
319 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
321 age = (now - oss->lqo_used) >> 3;
322 if (lod->lod_qos.lq_reset ||
323 age > 32 * lod->lod_desc.ld_qos_maxage)
324 oss->lqo_penalty = 0;
325 else if (age > lod->lod_desc.ld_qos_maxage)
326 /* Decay the penalty by half for every 8x the update
327 * interval that the device has been idle. That gives
328 * lots of time for the statfs information to be
329 * updated (which the penalty is only a proxy for),
330 * and avoids penalizing OSS/OSTs under light load. */
331 oss->lqo_penalty >>= age / lod->lod_desc.ld_qos_maxage;
334 lod->lod_qos.lq_dirty = 0;
335 lod->lod_qos.lq_reset = 0;
337 /* If each ost has almost same free space,
338 * do rr allocation for better creation performance */
339 lod->lod_qos.lq_same_space = 0;
340 if ((ba_max * (256 - lod->lod_qos.lq_threshold_rr)) >> 8 < ba_min) {
341 lod->lod_qos.lq_same_space = 1;
342 /* Reset weights for the next time we enter qos mode */
343 lod->lod_qos.lq_reset = 1;
349 if (!rc && lod->lod_qos.lq_same_space)
355 static int lod_qos_calc_weight(struct lod_device *lod, int i)
359 /* Final ost weight = TGT_BAVAIL - ost_penalty - oss_penalty */
360 temp = TGT_BAVAIL(i);
361 temp2 = OST_TGT(lod,i)->ltd_qos.ltq_penalty +
362 OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_penalty;
364 OST_TGT(lod,i)->ltd_qos.ltq_weight = 0;
366 OST_TGT(lod,i)->ltd_qos.ltq_weight = temp - temp2;
370 /* We just used this index for a stripe; adjust everyone's weights */
371 static int lod_qos_used(struct lod_device *lod, struct ost_pool *osts,
372 __u32 index, __u64 *total_wt)
374 struct lod_ost_desc *ost;
375 struct lov_qos_oss *oss;
379 ost = OST_TGT(lod,index);
382 /* Don't allocate on this devuce anymore, until the next alloc_qos */
383 ost->ltd_qos.ltq_usable = 0;
385 oss = ost->ltd_qos.ltq_oss;
387 /* Decay old penalty by half (we're adding max penalty, and don't
388 want it to run away.) */
389 ost->ltd_qos.ltq_penalty >>= 1;
390 oss->lqo_penalty >>= 1;
392 /* mark the OSS and OST as recently used */
393 ost->ltd_qos.ltq_used = oss->lqo_used = cfs_time_current_sec();
395 /* Set max penalties for this OST and OSS */
396 ost->ltd_qos.ltq_penalty +=
397 ost->ltd_qos.ltq_penalty_per_obj * lod->lod_ostnr;
398 oss->lqo_penalty += oss->lqo_penalty_per_obj *
399 lod->lod_qos.lq_active_oss_count;
401 /* Decrease all OSS penalties */
402 cfs_list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
403 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
404 oss->lqo_penalty = 0;
406 oss->lqo_penalty -= oss->lqo_penalty_per_obj;
410 /* Decrease all OST penalties */
411 for (j = 0; j < osts->op_count; j++) {
414 i = osts->op_array[j];
415 if (!cfs_bitmap_check(lod->lod_ost_bitmap, i))
418 ost = OST_TGT(lod,i);
421 if (ost->ltd_qos.ltq_penalty <
422 ost->ltd_qos.ltq_penalty_per_obj)
423 ost->ltd_qos.ltq_penalty = 0;
425 ost->ltd_qos.ltq_penalty -=
426 ost->ltd_qos.ltq_penalty_per_obj;
428 lod_qos_calc_weight(lod, i);
430 /* Recalc the total weight of usable osts */
431 if (ost->ltd_qos.ltq_usable)
432 *total_wt += ost->ltd_qos.ltq_weight;
434 QOS_DEBUG("recalc tgt %d usable=%d avail="LPU64
435 " ostppo="LPU64" ostp="LPU64" ossppo="LPU64
436 " ossp="LPU64" wt="LPU64"\n",
437 i, ost->ltd_qos.ltq_usable, TGT_BAVAIL(i) >> 10,
438 ost->ltd_qos.ltq_penalty_per_obj >> 10,
439 ost->ltd_qos.ltq_penalty >> 10,
440 ost->ltd_qos.ltq_oss->lqo_penalty_per_obj >> 10,
441 ost->ltd_qos.ltq_oss->lqo_penalty >> 10,
442 ost->ltd_qos.ltq_weight >> 10);
448 #define LOV_QOS_EMPTY ((__u32)-1)
449 /* compute optimal round-robin order, based on OSTs per OSS */
450 static int lod_qos_calc_rr(struct lod_device *lod, struct ost_pool *src_pool,
451 struct lov_qos_rr *lqr)
453 struct lov_qos_oss *oss;
454 struct lod_ost_desc *ost;
455 unsigned placed, real_count;
459 if (!lqr->lqr_dirty) {
460 LASSERT(lqr->lqr_pool.op_size);
464 /* Do actual allocation. */
465 cfs_down_write(&lod->lod_qos.lq_rw_sem);
468 * Check again. While we were sleeping on @lq_rw_sem something could
471 if (!lqr->lqr_dirty) {
472 LASSERT(lqr->lqr_pool.op_size);
473 cfs_up_write(&lod->lod_qos.lq_rw_sem);
477 real_count = src_pool->op_count;
479 /* Zero the pool array */
480 /* alloc_rr is holding a read lock on the pool, so nobody is adding/
481 deleting from the pool. The lq_rw_sem insures that nobody else
483 lqr->lqr_pool.op_count = real_count;
484 rc = lod_ost_pool_extend(&lqr->lqr_pool, real_count);
486 cfs_up_write(&lod->lod_qos.lq_rw_sem);
489 for (i = 0; i < lqr->lqr_pool.op_count; i++)
490 lqr->lqr_pool.op_array[i] = LOV_QOS_EMPTY;
492 /* Place all the OSTs from 1 OSS at the same time. */
494 cfs_list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
497 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
500 if (!cfs_bitmap_check(lod->lod_ost_bitmap,
501 src_pool->op_array[i]))
504 ost = OST_TGT(lod,src_pool->op_array[i]);
505 LASSERT(ost && ost->ltd_ost);
506 if (ost->ltd_qos.ltq_oss != oss)
509 /* Evenly space these OSTs across arrayspace */
510 next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count;
511 while (lqr->lqr_pool.op_array[next] != LOV_QOS_EMPTY)
512 next = (next + 1) % lqr->lqr_pool.op_count;
514 lqr->lqr_pool.op_array[next] = src_pool->op_array[i];
521 cfs_up_write(&lod->lod_qos.lq_rw_sem);
523 if (placed != real_count) {
524 /* This should never happen */
525 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
526 "round-robin list (%d of %d).\n",
528 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
529 LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
530 lqr->lqr_pool.op_array[i]);
537 for (i = 0; i < lqr->lqr_pool.op_count; i++)
538 QOS_CONSOLE("rr #%d ost idx=%d\n", i, lqr->lqr_pool.op_array[i]);
545 * A helper function to:
546 * create in-core lu object on the specified OSP
547 * declare creation of the object
548 * IMPORTANT: at this stage object is anonymouos - it has no fid assigned
549 * this is a workaround till we have natural FIDs on OST
551 * at this point we want to declare (reserve) object for us as
552 * we can't block at execution (when create method is called).
553 * otherwise we'd block whole transaction batch
555 static struct dt_object *lod_qos_declare_object_on(const struct lu_env *env,
556 struct lod_device *d,
560 struct lod_ost_desc *ost;
561 struct lu_object *o, *n;
562 struct lu_device *nd;
563 struct dt_object *dt;
568 LASSERT(ost_idx >= 0);
569 LASSERT(ost_idx < d->lod_osts_size);
570 ost = OST_TGT(d,ost_idx);
572 LASSERT(ost->ltd_ost);
574 nd = &ost->ltd_ost->dd_lu_dev;
577 * allocate anonymous object with zero fid, real fid
578 * will be assigned by OSP within transaction
579 * XXX: to be fixed with fully-functional OST fids
581 o = lu_object_anon(env, nd, NULL);
583 GOTO(out, dt = ERR_PTR(PTR_ERR(o)));
585 n = lu_object_locate(o->lo_header, nd->ld_type);
586 if (unlikely(n == NULL)) {
587 CERROR("can't find slice\n");
588 lu_object_put(env, o);
589 GOTO(out, dt = ERR_PTR(-EINVAL));
592 dt = container_of(n, struct dt_object, do_lu);
594 rc = dt_declare_create(env, dt, NULL, NULL, NULL, th);
596 CDEBUG(D_OTHER, "can't declare creation on #%u: %d\n",
598 lu_object_put(env, o);
606 static int min_stripe_count(int stripe_cnt, int flags)
608 return (flags & LOV_USES_DEFAULT_STRIPE ?
609 stripe_cnt - (stripe_cnt / 4) : stripe_cnt);
612 #define LOV_CREATE_RESEED_MULT 30
613 #define LOV_CREATE_RESEED_MIN 2000
615 static int inline lod_qos_dev_is_full(struct obd_statfs *msfs)
618 int bs = msfs->os_bsize;
620 LASSERT(((bs - 1) & bs) == 0);
622 /* the minimum of 0.1% used blocks and 1GB bytes. */
623 used = min_t(__u64, (msfs->os_blocks - msfs->os_bfree) >> 10,
624 1 << (31 - cfs_ffs(bs)));
625 return (msfs->os_bavail < used);
628 int lod_ea_store_resize(struct lod_thread_info *info, int size);
630 static inline int lod_qos_ost_in_use_clear(const struct lu_env *env, int stripes)
632 struct lod_thread_info *info = lod_env_info(env);
634 if (info->lti_ea_store_size < sizeof(int) * stripes)
635 lod_ea_store_resize(info, stripes * sizeof(int));
636 if (info->lti_ea_store_size < sizeof(int) * stripes) {
637 CERROR("can't allocate memory for ost-in-use array\n");
640 memset(info->lti_ea_store, 0, sizeof(int) * stripes);
644 static inline void lod_qos_ost_in_use(const struct lu_env *env, int idx, int ost)
646 struct lod_thread_info *info = lod_env_info(env);
647 int *osts = info->lti_ea_store;
649 LASSERT(info->lti_ea_store_size >= idx * sizeof(int));
653 static int lod_qos_is_ost_used(const struct lu_env *env, int ost, int stripes)
655 struct lod_thread_info *info = lod_env_info(env);
656 int *osts = info->lti_ea_store;
659 for (j = 0; j < stripes; j++) {
666 /* Allocate objects on osts with round-robin algorithm */
667 static int lod_alloc_rr(const struct lu_env *env, struct lod_object *lo,
668 int flags, struct thandle *th)
670 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
671 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
672 struct pool_desc *pool = NULL;
673 struct ost_pool *osts;
674 struct lov_qos_rr *lqr;
678 int ost_start_idx_temp;
681 int stripe_cnt = lo->ldo_stripenr;
682 int stripe_cnt_min = min_stripe_count(stripe_cnt, flags);
687 pool = lod_find_pool(m, lo->ldo_pool);
690 cfs_down_read(&pool_tgt_rw_sem(pool));
691 osts = &(pool->pool_obds);
692 lqr = &(pool->pool_rr);
694 osts = &(m->lod_pool_info);
695 lqr = &(m->lod_qos.lq_rr);
698 rc = lod_qos_calc_rr(m, osts, lqr);
702 rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
706 if (--lqr->lqr_start_count <= 0) {
707 lqr->lqr_start_idx = cfs_rand() % osts->op_count;
708 lqr->lqr_start_count =
709 (LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) +
710 LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U);
711 } else if (stripe_cnt_min >= osts->op_count ||
712 lqr->lqr_start_idx > osts->op_count) {
713 /* If we have allocated from all of the OSTs, slowly
714 * precess the next start if the OST/stripe count isn't
715 * already doing this for us. */
716 lqr->lqr_start_idx %= osts->op_count;
717 if (stripe_cnt > 1 && (osts->op_count % stripe_cnt) != 1)
718 ++lqr->lqr_offset_idx;
720 cfs_down_read(&m->lod_qos.lq_rw_sem);
721 ost_start_idx_temp = lqr->lqr_start_idx;
724 array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) %
727 QOS_DEBUG("pool '%s' want %d startidx %d startcnt %d offset %d "
728 "active %d count %d arrayidx %d\n",
729 lo->ldo_pool ? lo->ldo_pool : "",
730 stripe_cnt, lqr->lqr_start_idx, lqr->lqr_start_count,
731 lqr->lqr_offset_idx, osts->op_count, osts->op_count,
734 for (i = 0; i < osts->op_count;
735 i++, array_idx = (array_idx + 1) % osts->op_count) {
736 ++lqr->lqr_start_idx;
737 ost_idx = lqr->lqr_pool.op_array[array_idx];
739 QOS_DEBUG("#%d strt %d act %d strp %d ary %d idx %d\n",
740 i, lqr->lqr_start_idx, /* XXX: active*/ 0,
741 stripe_idx, array_idx, ost_idx);
743 if ((ost_idx == LOV_QOS_EMPTY) ||
744 !cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
747 /* Fail Check before osc_precreate() is called
748 so we can only 'fail' single OSC. */
749 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
752 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
754 /* this OSP doesn't feel well */
755 CERROR("can't statfs #%u: %d\n", ost_idx, rc);
760 * skip empty devices - usually it means inactive device
762 if (sfs->os_blocks == 0) {
763 QOS_DEBUG("#%d: inactive\n", ost_idx);
770 if (lod_qos_dev_is_full(sfs)) {
771 QOS_DEBUG("#%d is full\n", ost_idx);
776 * We expect number of precreated objects in f_ffree at
777 * the first iteration, skip OSPs with no objects ready
779 if (sfs->os_ffree == 0 && speed == 0) {
780 QOS_DEBUG("#%d: precreation is empty\n", ost_idx);
785 * try to use another OSP if this one is degraded
787 if (sfs->os_state == OS_STATE_DEGRADED && speed == 0) {
788 QOS_DEBUG("#%d: degraded\n", ost_idx);
793 * do not put >1 objects on a single OST
795 if (speed && lod_qos_is_ost_used(env, ost_idx, stripe_idx))
798 o = lod_qos_declare_object_on(env, m, ost_idx, th);
800 CDEBUG(D_OTHER, "can't declare new object on #%u: %d\n",
801 ost_idx, (int) PTR_ERR(o));
807 * We've successfuly declared (reserved) an object
809 lod_qos_ost_in_use(env, stripe_idx, ost_idx);
810 lo->ldo_stripe[stripe_idx] = o;
813 /* We have enough stripes */
814 if (stripe_idx == lo->ldo_stripenr)
817 if ((speed < 2) && (stripe_idx < stripe_cnt_min)) {
818 /* Try again, allowing slower OSCs */
820 lqr->lqr_start_idx = ost_start_idx_temp;
824 cfs_up_read(&m->lod_qos.lq_rw_sem);
827 lo->ldo_stripenr = stripe_idx;
828 /* at least one stripe is allocated */
831 /* nobody provided us with a single object */
837 cfs_up_read(&pool_tgt_rw_sem(pool));
838 /* put back ref got by lod_find_pool() */
839 lod_pool_putref(pool);
845 /* alloc objects on osts with specific stripe offset */
846 static int lod_alloc_specific(const struct lu_env *env, struct lod_object *lo,
847 int flags, struct thandle *th)
849 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
850 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
852 unsigned ost_idx, array_idx, ost_count;
853 int i, rc, stripe_num = 0;
855 struct pool_desc *pool = NULL;
856 struct ost_pool *osts;
860 pool = lod_find_pool(m, lo->ldo_pool);
863 cfs_down_read(&pool_tgt_rw_sem(pool));
864 osts = &(pool->pool_obds);
866 osts = &(m->lod_pool_info);
869 ost_count = osts->op_count;
872 /* search loi_ost_idx in ost array */
874 for (i = 0; i < ost_count; i++) {
875 if (osts->op_array[i] == lo->ldo_def_stripe_offset) {
880 if (i == ost_count) {
881 CERROR("Start index %d not found in pool '%s'\n",
882 lo->ldo_def_stripe_offset,
883 lo->ldo_pool ? lo->ldo_pool : "");
884 GOTO(out, rc = -EINVAL);
887 for (i = 0; i < ost_count;
888 i++, array_idx = (array_idx + 1) % ost_count) {
889 ost_idx = osts->op_array[array_idx];
891 if (!cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
894 /* Fail Check before osc_precreate() is called
895 so we can only 'fail' single OSC. */
896 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
899 /* Drop slow OSCs if we can, but not for requested start idx.
901 * This means "if OSC is slow and it is not the requested
902 * start OST, then it can be skipped, otherwise skip it only
903 * if it is inactive/recovering/out-of-space." */
905 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
907 /* this OSP doesn't feel well */
908 CERROR("can't statfs #%u: %d\n", ost_idx, rc);
913 * skip empty devices - usually it means inactive device
915 if (sfs->os_blocks == 0)
919 * We expect number of precreated objects in f_ffree at
920 * the first iteration, skip OSPs with no objects ready
921 * don't apply this logic to OST specified with stripe_offset
923 if (i != 0 && sfs->os_ffree == 0 && speed == 0)
926 o = lod_qos_declare_object_on(env, m, ost_idx, th);
928 CDEBUG(D_OTHER, "can't declare new object on #%u: %d\n",
929 ost_idx, (int) PTR_ERR(o));
934 * We've successfuly declared (reserved) an object
936 lo->ldo_stripe[stripe_num] = o;
939 /* We have enough stripes */
940 if (stripe_num == lo->ldo_stripenr)
944 /* Try again, allowing slower OSCs */
949 /* If we were passed specific striping params, then a failure to
950 * meet those requirements is an error, since we can't reallocate
951 * that memory (it might be part of a larger array or something).
953 * We can only get here if lsm_stripe_count was originally > 1.
955 CERROR("can't lstripe objid "DFID": have %d want %u\n",
956 PFID(lu_object_fid(lod2lu_obj(lo))), stripe_num,
961 cfs_up_read(&pool_tgt_rw_sem(pool));
962 /* put back ref got by lod_find_pool() */
963 lod_pool_putref(pool);
969 static inline int lod_qos_is_usable(struct lod_device *lod)
972 /* to be able to debug QoS code */
976 /* Detect -EAGAIN early, before expensive lock is taken. */
977 if (!lod->lod_qos.lq_dirty && lod->lod_qos.lq_same_space)
980 if (lod->lod_desc.ld_active_tgt_count < 2)
986 /* Alloc objects on osts with optimization based on:
988 - network resources (shared OSS's)
990 static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo,
991 int flags, struct thandle *th)
993 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
994 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
995 struct lod_ost_desc *ost;
997 __u64 total_weight = 0;
998 int nfound, good_osts, i, rc = 0;
999 int stripe_cnt = lo->ldo_stripenr;
1001 struct pool_desc *pool = NULL;
1002 struct ost_pool *osts;
1005 stripe_cnt_min = min_stripe_count(stripe_cnt, flags);
1006 if (stripe_cnt_min < 1)
1010 pool = lod_find_pool(m, lo->ldo_pool);
1013 cfs_down_read(&pool_tgt_rw_sem(pool));
1014 osts = &(pool->pool_obds);
1016 osts = &(m->lod_pool_info);
1019 /* Detect -EAGAIN early, before expensive lock is taken. */
1020 if (!lod_qos_is_usable(m))
1021 GOTO(out_nolock, rc = -EAGAIN);
1023 /* Do actual allocation, use write lock here. */
1024 cfs_down_write(&m->lod_qos.lq_rw_sem);
1027 * Check again, while we were sleeping on @lq_rw_sem things could
1030 if (!lod_qos_is_usable(m))
1031 GOTO(out, rc = -EAGAIN);
1033 rc = lod_qos_calc_ppo(m);
1037 rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
1042 /* Find all the OSTs that are valid stripe candidates */
1043 for (i = 0; i < osts->op_count; i++) {
1044 if (!cfs_bitmap_check(m->lod_ost_bitmap, osts->op_array[i]))
1047 rc = lod_statfs_and_check(env, m, osts->op_array[i], sfs);
1049 /* this OSP doesn't feel well */
1050 CERROR("can't statfs #%u: %d\n", i, rc);
1055 * skip empty devices - usually it means inactive device
1057 if (sfs->os_blocks == 0)
1063 if (lod_qos_dev_is_full(sfs))
1066 /* Fail Check before osc_precreate() is called
1067 so we can only 'fail' single OSC. */
1068 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) &&
1069 osts->op_array[i] == 0)
1072 ost = OST_TGT(m,osts->op_array[i]);
1073 ost->ltd_qos.ltq_usable = 1;
1074 lod_qos_calc_weight(m, osts->op_array[i]);
1075 total_weight += ost->ltd_qos.ltq_weight;
1080 QOS_DEBUG("found %d good osts\n", good_osts);
1082 if (good_osts < stripe_cnt_min)
1083 GOTO(out, rc = -EAGAIN);
1085 /* We have enough osts */
1086 if (good_osts < stripe_cnt)
1087 stripe_cnt = good_osts;
1089 /* Find enough OSTs with weighted random allocation. */
1091 while (nfound < stripe_cnt) {
1092 __u64 rand, cur_weight;
1098 #if BITS_PER_LONG == 32
1099 rand = cfs_rand() % (unsigned)total_weight;
1100 /* If total_weight > 32-bit, first generate the high
1101 * 32 bits of the random number, then add in the low
1102 * 32 bits (truncated to the upper limit, if needed) */
1103 if (total_weight > 0xffffffffULL)
1104 rand = (__u64)(cfs_rand() %
1105 (unsigned)(total_weight >> 32)) << 32;
1109 if (rand == (total_weight & 0xffffffff00000000ULL))
1110 rand |= cfs_rand() % (unsigned)total_weight;
1115 rand = ((__u64)cfs_rand() << 32 | cfs_rand()) %
1122 /* On average, this will hit larger-weighted osts more often.
1123 0-weight osts will always get used last (only when rand=0) */
1124 for (i = 0; i < osts->op_count; i++) {
1125 int idx = osts->op_array[i];
1127 if (!cfs_bitmap_check(m->lod_ost_bitmap, idx))
1130 ost = OST_TGT(m,idx);
1132 if (!ost->ltd_qos.ltq_usable)
1135 cur_weight += ost->ltd_qos.ltq_weight;
1136 QOS_DEBUG("stripe_cnt=%d nfound=%d cur_weight="LPU64
1137 " rand="LPU64" total_weight="LPU64"\n",
1138 stripe_cnt, nfound, cur_weight, rand,
1141 if (cur_weight < rand)
1144 QOS_DEBUG("stripe=%d to idx=%d\n", nfound, idx);
1147 * do not put >1 objects on a single OST
1149 if (lod_qos_is_ost_used(env, idx, nfound))
1151 lod_qos_ost_in_use(env, nfound, idx);
1153 o = lod_qos_declare_object_on(env, m, idx, th);
1155 QOS_DEBUG("can't declare object on #%u: %d\n",
1156 idx, (int) PTR_ERR(o));
1159 lo->ldo_stripe[nfound++] = o;
1160 lod_qos_used(m, osts, idx, &total_weight);
1166 if (unlikely(nfound != stripe_cnt)) {
1168 * when the decision to use weighted algorithm was made
1169 * we had enough appropriate OSPs, but this state can
1170 * change anytime (no space on OST, broken connection, etc)
1171 * so it's possible OSP won't be able to provide us with
1172 * an object due to just changed state
1174 LCONSOLE_INFO("wanted %d, found %d\n", stripe_cnt, nfound);
1175 for (i = 0; i < nfound; i++) {
1176 LASSERT(lo->ldo_stripe[i]);
1177 lu_object_put(env, &lo->ldo_stripe[i]->do_lu);
1178 lo->ldo_stripe[i] = NULL;
1181 /* makes sense to rebalance next time */
1182 m->lod_qos.lq_dirty = 1;
1183 m->lod_qos.lq_same_space = 0;
1189 cfs_up_write(&m->lod_qos.lq_rw_sem);
1193 cfs_up_read(&pool_tgt_rw_sem(pool));
1194 /* put back ref got by lod_find_pool() */
1195 lod_pool_putref(pool);
1201 /* Find the max stripecount we should use */
1202 static __u16 lod_get_stripecnt(struct lod_device *lod, __u32 magic,
1205 __u32 max_stripes = LOV_MAX_STRIPE_COUNT_OLD;
1208 stripe_count = lod->lod_desc.ld_default_stripe_count;
1209 if (stripe_count > lod->lod_desc.ld_active_tgt_count)
1210 stripe_count = lod->lod_desc.ld_active_tgt_count;
1214 /* stripe count is based on whether OSD can handle larger EA sizes */
1215 if (lod->lod_osd_max_easize > 0)
1216 max_stripes = lov_mds_md_stripecnt(lod->lod_osd_max_easize,
1219 return (stripe_count < max_stripes) ? stripe_count : max_stripes;
1222 static int lod_use_defined_striping(const struct lu_env *env,
1223 struct lod_object *mo,
1224 const struct lu_buf *buf)
1226 struct lod_device *d = lu2lod_dev(lod2lu_obj(mo)->lo_dev);
1227 struct lov_mds_md_v1 *v1 = buf->lb_buf;
1228 struct lov_mds_md_v3 *v3 = buf->lb_buf;
1229 struct lov_ost_data_v1 *objs;
1234 rc = lod_verify_striping(d, buf, 1);
1238 magic = le32_to_cpu(v1->lmm_magic);
1239 if (magic == LOV_MAGIC_V1_DEF) {
1240 objs = &v1->lmm_objects[0];
1241 } else if (magic == LOV_MAGIC_V3_DEF) {
1242 objs = &v3->lmm_objects[0];
1243 lod_object_set_pool(mo, v3->lmm_pool_name);
1245 GOTO(out, rc = -EINVAL);
1249 * LOD shouldn't be aware of recovery at all,
1250 * but this track recovery status (to some extent)
1251 * to be do additional checks like this one
1253 LASSERT(d->lod_recovery_completed == 0);
1255 mo->ldo_stripe_size = le32_to_cpu(v1->lmm_stripe_size);
1256 mo->ldo_stripenr = le16_to_cpu(v1->lmm_stripe_count);
1257 mo->ldo_layout_gen = le16_to_cpu(v1->lmm_layout_gen);
1258 LASSERT(buf->lb_len >= lov_mds_md_size(mo->ldo_stripenr, magic));
1260 rc = lod_initialize_objects(env, mo, objs);
1266 static int lod_qos_parse_config(const struct lu_env *env,
1267 struct lod_object *lo,
1268 const struct lu_buf *buf)
1270 struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
1271 struct lov_user_md_v1 *v1 = NULL;
1272 struct lov_user_md_v3 *v3 = NULL;
1273 struct pool_desc *pool;
1278 if (buf == NULL || buf->lb_buf == NULL || buf->lb_len == 0)
1282 magic = v1->lmm_magic;
1284 if (magic == __swab32(LOV_USER_MAGIC_V1))
1285 lustre_swab_lov_user_md_v1(v1);
1286 else if (magic == __swab32(LOV_USER_MAGIC_V3))
1287 lustre_swab_lov_user_md_v3(v3);
1289 if (unlikely(magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3)) {
1290 /* try to use as fully defined striping */
1291 rc = lod_use_defined_striping(env, lo, buf);
1295 if (unlikely(buf->lb_len < sizeof(*v1))) {
1296 CERROR("wrong size: %u\n", (unsigned) buf->lb_len);
1300 if (v1->lmm_pattern != 0 && v1->lmm_pattern != LOV_PATTERN_RAID0) {
1301 CERROR("invalid pattern: %x\n", v1->lmm_pattern);
1305 if (v1->lmm_stripe_size)
1306 lo->ldo_stripe_size = v1->lmm_stripe_size;
1307 if (lo->ldo_stripe_size & (LOV_MIN_STRIPE_SIZE - 1))
1308 lo->ldo_stripe_size = LOV_MIN_STRIPE_SIZE;
1310 if (v1->lmm_stripe_count)
1311 lo->ldo_stripenr = v1->lmm_stripe_count;
1313 if ((v1->lmm_stripe_offset >= d->lod_desc.ld_tgt_count) &&
1314 (v1->lmm_stripe_offset != (typeof(v1->lmm_stripe_offset))(-1))) {
1315 CERROR("invalid offset: %x\n", v1->lmm_stripe_offset);
1318 lo->ldo_def_stripe_offset = v1->lmm_stripe_offset;
1320 CDEBUG(D_OTHER, "lsm: %u size, %u stripes, %u offset\n",
1321 v1->lmm_stripe_size, v1->lmm_stripe_count,
1322 v1->lmm_stripe_offset);
1324 if (v1->lmm_magic == LOV_MAGIC_V3) {
1325 if (buf->lb_len < sizeof(*v3)) {
1326 CERROR("wrong size: %u\n", (unsigned) buf->lb_len);
1331 lod_object_set_pool(lo, v3->lmm_pool_name);
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);
1358 * buf should be NULL or contain striping settings
1360 int lod_qos_prep_create(const struct lu_env *env, struct lod_object *lo,
1361 struct lu_attr *attr, const struct lu_buf *buf,
1364 struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
1365 int flag = LOV_USES_ASSIGNED_STRIPE;
1371 /* no OST available */
1372 /* XXX: should we be waiting a bit to prevent failures during
1373 * cluster initialization? */
1374 if (d->lod_ostnr == 0)
1375 GOTO(out, rc = -EIO);
1378 * by this time, the object's ldo_stripenr and ldo_stripe_size
1379 * contain default value for striping: taken from the parent
1380 * or from filesystem defaults
1382 * in case the caller is passing lovea with new striping config,
1383 * we may need to parse lovea and apply new configuration
1385 rc = lod_qos_parse_config(env, lo, buf);
1389 if (likely(lo->ldo_stripe == NULL)) {
1391 * no striping has been created so far
1393 LASSERT(lo->ldo_stripenr > 0);
1394 lo->ldo_stripenr = lod_get_stripecnt(d, LOV_MAGIC,
1396 i = sizeof(struct dt_object *) * lo->ldo_stripenr;
1397 OBD_ALLOC(lo->ldo_stripe, i);
1398 if (lo->ldo_stripe == NULL)
1399 GOTO(out, rc = -ENOMEM);
1400 lo->ldo_stripes_allocated = lo->ldo_stripenr;
1403 /* XXX: support for non-0 files w/o objects */
1404 if (lo->ldo_def_stripe_offset >= d->lod_desc.ld_tgt_count) {
1405 lod_qos_statfs_update(env, d);
1406 rc = lod_alloc_qos(env, lo, flag, th);
1408 rc = lod_alloc_rr(env, lo, flag, th);
1410 rc = lod_alloc_specific(env, lo, flag, th);
1414 * lod_qos_parse_config() found supplied buf as a predefined
1415 * striping (not a hint), so it allocated all the object
1416 * now we need to create them
1418 for (i = 0; i < lo->ldo_stripenr; i++) {
1419 struct dt_object *o;
1421 o = lo->ldo_stripe[i];
1424 rc = dt_declare_create(env, o, attr, NULL, NULL, th);
1426 CERROR("can't declare create: %d\n", rc);