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);
159 if (rc && rc != -ENOTCONN)
160 CERROR("%s: statfs: rc = %d\n", lod2obd(d)->obd_name, rc);
162 /* check whether device has changed state (active, inactive) */
163 if (rc != 0 && ost->ltd_active) {
164 /* turned inactive? */
165 cfs_spin_lock(&d->lod_desc_lock);
166 if (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 (rc == 0 && 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 (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);
194 static void lod_qos_statfs_update(const struct lu_env *env,
195 struct lod_device *lod)
197 struct obd_device *obd = lod2obd(lod);
198 struct ost_pool *osts = &(lod->lod_pool_info);
200 __u64 max_age, avail;
203 max_age = cfs_time_shift_64(-2 * lod->lod_desc.ld_qos_maxage);
205 if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
206 /* statfs data are quite recent, don't need to refresh it */
209 cfs_down_write(&lod->lod_qos.lq_rw_sem);
210 if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
213 for (i = 0; i < osts->op_count; i++) {
214 idx = osts->op_array[i];
215 avail = OST_TGT(lod,idx)->ltd_statfs.os_bavail;
216 rc = lod_statfs_and_check(env, lod, idx,
217 &OST_TGT(lod,idx)->ltd_statfs);
220 if (OST_TGT(lod,idx)->ltd_statfs.os_bavail != avail)
221 /* recalculate weigths */
222 lod->lod_qos.lq_dirty = 1;
224 obd->obd_osfs_age = cfs_time_current_64();
227 cfs_up_write(&lod->lod_qos.lq_rw_sem);
230 /* Recalculate per-object penalties for OSSs and OSTs,
231 depends on size of each ost in an oss */
232 static int lod_qos_calc_ppo(struct lod_device *lod)
234 struct lov_qos_oss *oss;
235 __u64 ba_max, ba_min, temp;
237 int rc, i, prio_wide;
241 if (!lod->lod_qos.lq_dirty)
244 num_active = lod->lod_desc.ld_active_tgt_count - 1;
246 GOTO(out, rc = -EAGAIN);
248 /* find bavail on each OSS */
249 cfs_list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list)
251 lod->lod_qos.lq_active_oss_count = 0;
254 * How badly user wants to select OSTs "widely" (not recently chosen
255 * and not on recent OSS's). As opposed to "freely" (free space
258 prio_wide = 256 - lod->lod_qos.lq_prio_free;
260 ba_min = (__u64)(-1);
262 now = cfs_time_current_sec();
263 /* Calculate OST penalty per object
264 * (lod ref taken in lod_qos_prep_create()) */
265 cfs_foreach_bit(lod->lod_ost_bitmap, i) {
266 LASSERT(OST_TGT(lod,i));
267 temp = TGT_BAVAIL(i);
270 ba_min = min(temp, ba_min);
271 ba_max = max(temp, ba_max);
273 /* Count the number of usable OSS's */
274 if (OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail == 0)
275 lod->lod_qos.lq_active_oss_count++;
276 OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail += temp;
278 /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
280 lov_do_div64(temp, num_active);
281 OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj =
282 (temp * prio_wide) >> 8;
284 age = (now - OST_TGT(lod,i)->ltd_qos.ltq_used) >> 3;
285 if (lod->lod_qos.lq_reset ||
286 age > 32 * lod->lod_desc.ld_qos_maxage)
287 OST_TGT(lod,i)->ltd_qos.ltq_penalty = 0;
288 else if (age > lod->lod_desc.ld_qos_maxage)
289 /* Decay the penalty by half for every 8x the update
290 * interval that the device has been idle. That gives
291 * lots of time for the statfs information to be
292 * updated (which the penalty is only a proxy for),
293 * and avoids penalizing OSS/OSTs under light load. */
294 OST_TGT(lod,i)->ltd_qos.ltq_penalty >>=
295 (age / lod->lod_desc.ld_qos_maxage);
298 num_active = lod->lod_qos.lq_active_oss_count - 1;
299 if (num_active < 1) {
300 /* If there's only 1 OSS, we can't penalize it, so instead
301 we have to double the OST penalty */
303 cfs_foreach_bit(lod->lod_ost_bitmap, i)
304 OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj <<= 1;
307 /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
308 cfs_list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
309 temp = oss->lqo_bavail >> 1;
310 lov_do_div64(temp, oss->lqo_ost_count * num_active);
311 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
313 age = (now - oss->lqo_used) >> 3;
314 if (lod->lod_qos.lq_reset ||
315 age > 32 * lod->lod_desc.ld_qos_maxage)
316 oss->lqo_penalty = 0;
317 else if (age > lod->lod_desc.ld_qos_maxage)
318 /* Decay the penalty by half for every 8x the update
319 * interval that the device has been idle. That gives
320 * lots of time for the statfs information to be
321 * updated (which the penalty is only a proxy for),
322 * and avoids penalizing OSS/OSTs under light load. */
323 oss->lqo_penalty >>= age / lod->lod_desc.ld_qos_maxage;
326 lod->lod_qos.lq_dirty = 0;
327 lod->lod_qos.lq_reset = 0;
329 /* If each ost has almost same free space,
330 * do rr allocation for better creation performance */
331 lod->lod_qos.lq_same_space = 0;
332 if ((ba_max * (256 - lod->lod_qos.lq_threshold_rr)) >> 8 < ba_min) {
333 lod->lod_qos.lq_same_space = 1;
334 /* Reset weights for the next time we enter qos mode */
335 lod->lod_qos.lq_reset = 1;
341 if (!rc && lod->lod_qos.lq_same_space)
347 static int lod_qos_calc_weight(struct lod_device *lod, int i)
351 /* Final ost weight = TGT_BAVAIL - ost_penalty - oss_penalty */
352 temp = TGT_BAVAIL(i);
353 temp2 = OST_TGT(lod,i)->ltd_qos.ltq_penalty +
354 OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_penalty;
356 OST_TGT(lod,i)->ltd_qos.ltq_weight = 0;
358 OST_TGT(lod,i)->ltd_qos.ltq_weight = temp - temp2;
362 /* We just used this index for a stripe; adjust everyone's weights */
363 static int lod_qos_used(struct lod_device *lod, struct ost_pool *osts,
364 __u32 index, __u64 *total_wt)
366 struct lod_ost_desc *ost;
367 struct lov_qos_oss *oss;
371 ost = OST_TGT(lod,index);
374 /* Don't allocate on this devuce anymore, until the next alloc_qos */
375 ost->ltd_qos.ltq_usable = 0;
377 oss = ost->ltd_qos.ltq_oss;
379 /* Decay old penalty by half (we're adding max penalty, and don't
380 want it to run away.) */
381 ost->ltd_qos.ltq_penalty >>= 1;
382 oss->lqo_penalty >>= 1;
384 /* mark the OSS and OST as recently used */
385 ost->ltd_qos.ltq_used = oss->lqo_used = cfs_time_current_sec();
387 /* Set max penalties for this OST and OSS */
388 ost->ltd_qos.ltq_penalty +=
389 ost->ltd_qos.ltq_penalty_per_obj * lod->lod_ostnr;
390 oss->lqo_penalty += oss->lqo_penalty_per_obj *
391 lod->lod_qos.lq_active_oss_count;
393 /* Decrease all OSS penalties */
394 cfs_list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
395 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
396 oss->lqo_penalty = 0;
398 oss->lqo_penalty -= oss->lqo_penalty_per_obj;
402 /* Decrease all OST penalties */
403 for (j = 0; j < osts->op_count; j++) {
406 i = osts->op_array[j];
407 if (!cfs_bitmap_check(lod->lod_ost_bitmap, i))
410 ost = OST_TGT(lod,i);
413 if (ost->ltd_qos.ltq_penalty <
414 ost->ltd_qos.ltq_penalty_per_obj)
415 ost->ltd_qos.ltq_penalty = 0;
417 ost->ltd_qos.ltq_penalty -=
418 ost->ltd_qos.ltq_penalty_per_obj;
420 lod_qos_calc_weight(lod, i);
422 /* Recalc the total weight of usable osts */
423 if (ost->ltd_qos.ltq_usable)
424 *total_wt += ost->ltd_qos.ltq_weight;
426 QOS_DEBUG("recalc tgt %d usable=%d avail="LPU64
427 " ostppo="LPU64" ostp="LPU64" ossppo="LPU64
428 " ossp="LPU64" wt="LPU64"\n",
429 i, ost->ltd_qos.ltq_usable, TGT_BAVAIL(i) >> 10,
430 ost->ltd_qos.ltq_penalty_per_obj >> 10,
431 ost->ltd_qos.ltq_penalty >> 10,
432 ost->ltd_qos.ltq_oss->lqo_penalty_per_obj >> 10,
433 ost->ltd_qos.ltq_oss->lqo_penalty >> 10,
434 ost->ltd_qos.ltq_weight >> 10);
440 #define LOV_QOS_EMPTY ((__u32)-1)
441 /* compute optimal round-robin order, based on OSTs per OSS */
442 static int lod_qos_calc_rr(struct lod_device *lod, struct ost_pool *src_pool,
443 struct lov_qos_rr *lqr)
445 struct lov_qos_oss *oss;
446 struct lod_ost_desc *ost;
447 unsigned placed, real_count;
451 if (!lqr->lqr_dirty) {
452 LASSERT(lqr->lqr_pool.op_size);
456 /* Do actual allocation. */
457 cfs_down_write(&lod->lod_qos.lq_rw_sem);
460 * Check again. While we were sleeping on @lq_rw_sem something could
463 if (!lqr->lqr_dirty) {
464 LASSERT(lqr->lqr_pool.op_size);
465 cfs_up_write(&lod->lod_qos.lq_rw_sem);
469 real_count = src_pool->op_count;
471 /* Zero the pool array */
472 /* alloc_rr is holding a read lock on the pool, so nobody is adding/
473 deleting from the pool. The lq_rw_sem insures that nobody else
475 lqr->lqr_pool.op_count = real_count;
476 rc = lod_ost_pool_extend(&lqr->lqr_pool, real_count);
478 cfs_up_write(&lod->lod_qos.lq_rw_sem);
481 for (i = 0; i < lqr->lqr_pool.op_count; i++)
482 lqr->lqr_pool.op_array[i] = LOV_QOS_EMPTY;
484 /* Place all the OSTs from 1 OSS at the same time. */
486 cfs_list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
489 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
492 if (!cfs_bitmap_check(lod->lod_ost_bitmap,
493 src_pool->op_array[i]))
496 ost = OST_TGT(lod,src_pool->op_array[i]);
497 LASSERT(ost && ost->ltd_ost);
498 if (ost->ltd_qos.ltq_oss != oss)
501 /* Evenly space these OSTs across arrayspace */
502 next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count;
503 while (lqr->lqr_pool.op_array[next] != LOV_QOS_EMPTY)
504 next = (next + 1) % lqr->lqr_pool.op_count;
506 lqr->lqr_pool.op_array[next] = src_pool->op_array[i];
513 cfs_up_write(&lod->lod_qos.lq_rw_sem);
515 if (placed != real_count) {
516 /* This should never happen */
517 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
518 "round-robin list (%d of %d).\n",
520 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
521 LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
522 lqr->lqr_pool.op_array[i]);
529 for (i = 0; i < lqr->lqr_pool.op_count; i++)
530 QOS_CONSOLE("rr #%d ost idx=%d\n", i, lqr->lqr_pool.op_array[i]);
537 * A helper function to:
538 * create in-core lu object on the specified OSP
539 * declare creation of the object
540 * IMPORTANT: at this stage object is anonymouos - it has no fid assigned
541 * this is a workaround till we have natural FIDs on OST
543 * at this point we want to declare (reserve) object for us as
544 * we can't block at execution (when create method is called).
545 * otherwise we'd block whole transaction batch
547 static struct dt_object *lod_qos_declare_object_on(const struct lu_env *env,
548 struct lod_device *d,
552 struct lod_ost_desc *ost;
553 struct lu_object *o, *n;
554 struct lu_device *nd;
555 struct dt_object *dt;
560 LASSERT(ost_idx >= 0);
561 LASSERT(ost_idx < d->lod_osts_size);
562 ost = OST_TGT(d,ost_idx);
564 LASSERT(ost->ltd_ost);
566 nd = &ost->ltd_ost->dd_lu_dev;
569 * allocate anonymous object with zero fid, real fid
570 * will be assigned by OSP within transaction
571 * XXX: to be fixed with fully-functional OST fids
573 o = lu_object_anon(env, nd, NULL);
575 GOTO(out, dt = ERR_PTR(PTR_ERR(o)));
577 n = lu_object_locate(o->lo_header, nd->ld_type);
578 if (unlikely(n == NULL)) {
579 CERROR("can't find slice\n");
580 lu_object_put(env, o);
581 GOTO(out, dt = ERR_PTR(-EINVAL));
584 dt = container_of(n, struct dt_object, do_lu);
586 rc = dt_declare_create(env, dt, NULL, NULL, NULL, th);
588 CDEBUG(D_OTHER, "can't declare creation on #%u: %d\n",
590 lu_object_put(env, o);
598 static int min_stripe_count(int stripe_cnt, int flags)
600 return (flags & LOV_USES_DEFAULT_STRIPE ?
601 stripe_cnt - (stripe_cnt / 4) : stripe_cnt);
604 #define LOV_CREATE_RESEED_MULT 30
605 #define LOV_CREATE_RESEED_MIN 2000
607 static int inline lod_qos_dev_is_full(struct obd_statfs *msfs)
610 int bs = msfs->os_bsize;
612 LASSERT(((bs - 1) & bs) == 0);
614 /* the minimum of 0.1% used blocks and 1GB bytes. */
615 used = min_t(__u64, (msfs->os_blocks - msfs->os_bfree) >> 10,
616 1 << (31 - cfs_ffs(bs)));
617 return (msfs->os_bavail < used);
620 int lod_ea_store_resize(struct lod_thread_info *info, int size);
622 static inline int lod_qos_ost_in_use_clear(const struct lu_env *env, int stripes)
624 struct lod_thread_info *info = lod_env_info(env);
626 if (info->lti_ea_store_size < sizeof(int) * stripes)
627 lod_ea_store_resize(info, stripes * sizeof(int));
628 if (info->lti_ea_store_size < sizeof(int) * stripes) {
629 CERROR("can't allocate memory for ost-in-use array\n");
632 memset(info->lti_ea_store, 0, sizeof(int) * stripes);
636 static inline void lod_qos_ost_in_use(const struct lu_env *env, int idx, int ost)
638 struct lod_thread_info *info = lod_env_info(env);
639 int *osts = info->lti_ea_store;
641 LASSERT(info->lti_ea_store_size >= idx * sizeof(int));
645 static int lod_qos_is_ost_used(const struct lu_env *env, int ost, int stripes)
647 struct lod_thread_info *info = lod_env_info(env);
648 int *osts = info->lti_ea_store;
651 for (j = 0; j < stripes; j++) {
658 /* Allocate objects on OSTs with round-robin algorithm */
659 static int lod_alloc_rr(const struct lu_env *env, struct lod_object *lo,
660 int flags, struct thandle *th)
662 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
663 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
664 struct pool_desc *pool = NULL;
665 struct ost_pool *osts;
666 struct lov_qos_rr *lqr;
670 int ost_start_idx_temp;
673 int stripe_cnt = lo->ldo_stripenr;
674 int stripe_cnt_min = min_stripe_count(stripe_cnt, flags);
679 pool = lod_find_pool(m, lo->ldo_pool);
682 cfs_down_read(&pool_tgt_rw_sem(pool));
683 osts = &(pool->pool_obds);
684 lqr = &(pool->pool_rr);
686 osts = &(m->lod_pool_info);
687 lqr = &(m->lod_qos.lq_rr);
690 rc = lod_qos_calc_rr(m, osts, lqr);
694 rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
698 if (--lqr->lqr_start_count <= 0) {
699 lqr->lqr_start_idx = cfs_rand() % osts->op_count;
700 lqr->lqr_start_count =
701 (LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) +
702 LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U);
703 } else if (stripe_cnt_min >= osts->op_count ||
704 lqr->lqr_start_idx > osts->op_count) {
705 /* If we have allocated from all of the OSTs, slowly
706 * precess the next start if the OST/stripe count isn't
707 * already doing this for us. */
708 lqr->lqr_start_idx %= osts->op_count;
709 if (stripe_cnt > 1 && (osts->op_count % stripe_cnt) != 1)
710 ++lqr->lqr_offset_idx;
712 cfs_down_read(&m->lod_qos.lq_rw_sem);
713 ost_start_idx_temp = lqr->lqr_start_idx;
716 array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) %
719 QOS_DEBUG("pool '%s' want %d startidx %d startcnt %d offset %d "
720 "active %d count %d arrayidx %d\n",
721 lo->ldo_pool ? lo->ldo_pool : "",
722 stripe_cnt, lqr->lqr_start_idx, lqr->lqr_start_count,
723 lqr->lqr_offset_idx, osts->op_count, osts->op_count,
726 for (i = 0; i < osts->op_count && stripe_idx < lo->ldo_stripenr;
727 i++, array_idx = (array_idx + 1) % osts->op_count) {
728 ++lqr->lqr_start_idx;
729 ost_idx = lqr->lqr_pool.op_array[array_idx];
731 QOS_DEBUG("#%d strt %d act %d strp %d ary %d idx %d\n",
732 i, lqr->lqr_start_idx, /* XXX: active*/ 0,
733 stripe_idx, array_idx, ost_idx);
735 if ((ost_idx == LOV_QOS_EMPTY) ||
736 !cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
739 /* Fail Check before osc_precreate() is called
740 so we can only 'fail' single OSC. */
741 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
744 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
746 /* this OSP doesn't feel well */
753 if (lod_qos_dev_is_full(sfs)) {
754 QOS_DEBUG("#%d is full\n", ost_idx);
759 * We expect number of precreated objects in f_ffree at
760 * the first iteration, skip OSPs with no objects ready
762 if (sfs->os_fprecreated == 0 && speed == 0) {
763 QOS_DEBUG("#%d: precreation is empty\n", ost_idx);
768 * try to use another OSP if this one is degraded
770 if (sfs->os_state == OS_STATE_DEGRADED && speed < 2) {
771 QOS_DEBUG("#%d: degraded\n", ost_idx);
776 * do not put >1 objects on a single OST
778 if (speed && lod_qos_is_ost_used(env, ost_idx, stripe_idx))
781 o = lod_qos_declare_object_on(env, m, ost_idx, th);
783 CDEBUG(D_OTHER, "can't declare new object on #%u: %d\n",
784 ost_idx, (int) PTR_ERR(o));
790 * We've successfuly declared (reserved) an object
792 lod_qos_ost_in_use(env, stripe_idx, ost_idx);
793 lo->ldo_stripe[stripe_idx] = o;
797 if ((speed < 2) && (stripe_idx < stripe_cnt_min)) {
798 /* Try again, allowing slower OSCs */
800 lqr->lqr_start_idx = ost_start_idx_temp;
804 cfs_up_read(&m->lod_qos.lq_rw_sem);
807 lo->ldo_stripenr = stripe_idx;
808 /* at least one stripe is allocated */
811 /* nobody provided us with a single object */
817 cfs_up_read(&pool_tgt_rw_sem(pool));
818 /* put back ref got by lod_find_pool() */
819 lod_pool_putref(pool);
825 /* alloc objects on osts with specific stripe offset */
826 static int lod_alloc_specific(const struct lu_env *env, struct lod_object *lo,
827 int flags, struct thandle *th)
829 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
830 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
832 unsigned ost_idx, array_idx, ost_count;
833 int i, rc, stripe_num = 0;
835 struct pool_desc *pool = NULL;
836 struct ost_pool *osts;
839 rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
844 pool = lod_find_pool(m, lo->ldo_pool);
847 cfs_down_read(&pool_tgt_rw_sem(pool));
848 osts = &(pool->pool_obds);
850 osts = &(m->lod_pool_info);
853 ost_count = osts->op_count;
856 /* search loi_ost_idx in ost array */
858 for (i = 0; i < ost_count; i++) {
859 if (osts->op_array[i] == lo->ldo_def_stripe_offset) {
864 if (i == ost_count) {
865 CERROR("Start index %d not found in pool '%s'\n",
866 lo->ldo_def_stripe_offset,
867 lo->ldo_pool ? lo->ldo_pool : "");
868 GOTO(out, rc = -EINVAL);
871 for (i = 0; i < ost_count;
872 i++, array_idx = (array_idx + 1) % ost_count) {
873 ost_idx = osts->op_array[array_idx];
875 if (!cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
878 /* Fail Check before osc_precreate() is called
879 so we can only 'fail' single OSC. */
880 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
884 * do not put >1 objects on a single OST
886 if (lod_qos_is_ost_used(env, ost_idx, stripe_num))
889 /* Drop slow OSCs if we can, but not for requested start idx.
891 * This means "if OSC is slow and it is not the requested
892 * start OST, then it can be skipped, otherwise skip it only
893 * if it is inactive/recovering/out-of-space." */
895 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
897 /* this OSP doesn't feel well */
902 * We expect number of precreated objects in f_ffree at
903 * the first iteration, skip OSPs with no objects ready
904 * don't apply this logic to OST specified with stripe_offset
906 if (i != 0 && sfs->os_fprecreated == 0 && speed == 0)
909 o = lod_qos_declare_object_on(env, m, ost_idx, th);
911 CDEBUG(D_OTHER, "can't declare new object on #%u: %d\n",
912 ost_idx, (int) PTR_ERR(o));
917 * We've successfuly declared (reserved) an object
919 lo->ldo_stripe[stripe_num] = o;
922 /* We have enough stripes */
923 if (stripe_num == lo->ldo_stripenr)
927 /* Try again, allowing slower OSCs */
932 /* If we were passed specific striping params, then a failure to
933 * meet those requirements is an error, since we can't reallocate
934 * that memory (it might be part of a larger array or something).
936 * We can only get here if lsm_stripe_count was originally > 1.
938 CERROR("can't lstripe objid "DFID": have %d want %u\n",
939 PFID(lu_object_fid(lod2lu_obj(lo))), stripe_num,
944 cfs_up_read(&pool_tgt_rw_sem(pool));
945 /* put back ref got by lod_find_pool() */
946 lod_pool_putref(pool);
952 static inline int lod_qos_is_usable(struct lod_device *lod)
955 /* to be able to debug QoS code */
959 /* Detect -EAGAIN early, before expensive lock is taken. */
960 if (!lod->lod_qos.lq_dirty && lod->lod_qos.lq_same_space)
963 if (lod->lod_desc.ld_active_tgt_count < 2)
969 /* Alloc objects on OSTs with optimization based on:
971 - network resources (shared OSS's)
973 static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo,
974 int flags, struct thandle *th)
976 struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
977 struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
978 struct lod_ost_desc *ost;
980 __u64 total_weight = 0;
981 int nfound, good_osts, i, rc = 0;
982 int stripe_cnt = lo->ldo_stripenr;
984 struct pool_desc *pool = NULL;
985 struct ost_pool *osts;
988 stripe_cnt_min = min_stripe_count(stripe_cnt, flags);
989 if (stripe_cnt_min < 1)
993 pool = lod_find_pool(m, lo->ldo_pool);
996 cfs_down_read(&pool_tgt_rw_sem(pool));
997 osts = &(pool->pool_obds);
999 osts = &(m->lod_pool_info);
1002 /* Detect -EAGAIN early, before expensive lock is taken. */
1003 if (!lod_qos_is_usable(m))
1004 GOTO(out_nolock, rc = -EAGAIN);
1006 /* Do actual allocation, use write lock here. */
1007 cfs_down_write(&m->lod_qos.lq_rw_sem);
1010 * Check again, while we were sleeping on @lq_rw_sem things could
1013 if (!lod_qos_is_usable(m))
1014 GOTO(out, rc = -EAGAIN);
1016 rc = lod_qos_calc_ppo(m);
1020 rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
1025 /* Find all the OSTs that are valid stripe candidates */
1026 for (i = 0; i < osts->op_count; i++) {
1027 if (!cfs_bitmap_check(m->lod_ost_bitmap, osts->op_array[i]))
1030 rc = lod_statfs_and_check(env, m, osts->op_array[i], sfs);
1032 /* this OSP doesn't feel well */
1039 if (lod_qos_dev_is_full(sfs))
1042 /* Fail Check before osc_precreate() is called
1043 so we can only 'fail' single OSC. */
1044 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) &&
1045 osts->op_array[i] == 0)
1048 ost = OST_TGT(m,osts->op_array[i]);
1049 ost->ltd_qos.ltq_usable = 1;
1050 lod_qos_calc_weight(m, osts->op_array[i]);
1051 total_weight += ost->ltd_qos.ltq_weight;
1056 QOS_DEBUG("found %d good osts\n", good_osts);
1058 if (good_osts < stripe_cnt_min)
1059 GOTO(out, rc = -EAGAIN);
1061 /* We have enough osts */
1062 if (good_osts < stripe_cnt)
1063 stripe_cnt = good_osts;
1065 /* Find enough OSTs with weighted random allocation. */
1067 while (nfound < stripe_cnt) {
1068 __u64 rand, cur_weight;
1074 #if BITS_PER_LONG == 32
1075 rand = cfs_rand() % (unsigned)total_weight;
1076 /* If total_weight > 32-bit, first generate the high
1077 * 32 bits of the random number, then add in the low
1078 * 32 bits (truncated to the upper limit, if needed) */
1079 if (total_weight > 0xffffffffULL)
1080 rand = (__u64)(cfs_rand() %
1081 (unsigned)(total_weight >> 32)) << 32;
1085 if (rand == (total_weight & 0xffffffff00000000ULL))
1086 rand |= cfs_rand() % (unsigned)total_weight;
1091 rand = ((__u64)cfs_rand() << 32 | cfs_rand()) %
1098 /* On average, this will hit larger-weighted osts more often.
1099 0-weight osts will always get used last (only when rand=0) */
1100 for (i = 0; i < osts->op_count; i++) {
1101 int idx = osts->op_array[i];
1103 if (!cfs_bitmap_check(m->lod_ost_bitmap, idx))
1106 ost = OST_TGT(m,idx);
1108 if (!ost->ltd_qos.ltq_usable)
1111 cur_weight += ost->ltd_qos.ltq_weight;
1112 QOS_DEBUG("stripe_cnt=%d nfound=%d cur_weight="LPU64
1113 " rand="LPU64" total_weight="LPU64"\n",
1114 stripe_cnt, nfound, cur_weight, rand,
1117 if (cur_weight < rand)
1120 QOS_DEBUG("stripe=%d to idx=%d\n", nfound, idx);
1123 * do not put >1 objects on a single OST
1125 if (lod_qos_is_ost_used(env, idx, nfound))
1127 lod_qos_ost_in_use(env, nfound, idx);
1129 o = lod_qos_declare_object_on(env, m, idx, th);
1131 QOS_DEBUG("can't declare object on #%u: %d\n",
1132 idx, (int) PTR_ERR(o));
1135 lo->ldo_stripe[nfound++] = o;
1136 lod_qos_used(m, osts, idx, &total_weight);
1142 /* no OST found on this iteration, give up */
1147 if (unlikely(nfound != stripe_cnt)) {
1149 * when the decision to use weighted algorithm was made
1150 * we had enough appropriate OSPs, but this state can
1151 * change anytime (no space on OST, broken connection, etc)
1152 * so it's possible OSP won't be able to provide us with
1153 * an object due to just changed state
1155 LCONSOLE_INFO("wanted %d, found %d\n", stripe_cnt, nfound);
1156 for (i = 0; i < nfound; i++) {
1157 LASSERT(lo->ldo_stripe[i]);
1158 lu_object_put(env, &lo->ldo_stripe[i]->do_lu);
1159 lo->ldo_stripe[i] = NULL;
1162 /* makes sense to rebalance next time */
1163 m->lod_qos.lq_dirty = 1;
1164 m->lod_qos.lq_same_space = 0;
1170 cfs_up_write(&m->lod_qos.lq_rw_sem);
1174 cfs_up_read(&pool_tgt_rw_sem(pool));
1175 /* put back ref got by lod_find_pool() */
1176 lod_pool_putref(pool);
1182 /* Find the max stripecount we should use */
1183 static __u16 lod_get_stripecnt(struct lod_device *lod, __u32 magic,
1186 __u32 max_stripes = LOV_MAX_STRIPE_COUNT_OLD;
1189 stripe_count = lod->lod_desc.ld_default_stripe_count;
1190 if (stripe_count > lod->lod_desc.ld_active_tgt_count)
1191 stripe_count = lod->lod_desc.ld_active_tgt_count;
1195 /* stripe count is based on whether OSD can handle larger EA sizes */
1196 if (lod->lod_osd_max_easize > 0)
1197 max_stripes = lov_mds_md_stripecnt(lod->lod_osd_max_easize,
1200 return (stripe_count < max_stripes) ? stripe_count : max_stripes;
1203 static int lod_use_defined_striping(const struct lu_env *env,
1204 struct lod_object *mo,
1205 const struct lu_buf *buf)
1207 struct lod_device *d = lu2lod_dev(lod2lu_obj(mo)->lo_dev);
1208 struct lov_mds_md_v1 *v1 = buf->lb_buf;
1209 struct lov_mds_md_v3 *v3 = buf->lb_buf;
1210 struct lov_ost_data_v1 *objs;
1215 rc = lod_verify_striping(d, buf, 1);
1219 magic = le32_to_cpu(v1->lmm_magic);
1220 if (magic == LOV_MAGIC_V1_DEF) {
1221 objs = &v1->lmm_objects[0];
1222 } else if (magic == LOV_MAGIC_V3_DEF) {
1223 objs = &v3->lmm_objects[0];
1224 lod_object_set_pool(mo, v3->lmm_pool_name);
1226 GOTO(out, rc = -EINVAL);
1230 * LOD shouldn't be aware of recovery at all,
1231 * but this track recovery status (to some extent)
1232 * to be do additional checks like this one
1234 LASSERT(d->lod_recovery_completed == 0);
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);
1239 LASSERT(buf->lb_len >= lov_mds_md_size(mo->ldo_stripenr, magic));
1241 rc = lod_initialize_objects(env, mo, objs);
1247 static int lod_qos_parse_config(const struct lu_env *env,
1248 struct lod_object *lo,
1249 const struct lu_buf *buf)
1251 struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
1252 struct lov_user_md_v1 *v1 = NULL;
1253 struct lov_user_md_v3 *v3 = NULL;
1254 struct pool_desc *pool;
1259 if (buf == NULL || buf->lb_buf == NULL || buf->lb_len == 0)
1263 magic = v1->lmm_magic;
1265 if (magic == __swab32(LOV_USER_MAGIC_V1))
1266 lustre_swab_lov_user_md_v1(v1);
1267 else if (magic == __swab32(LOV_USER_MAGIC_V3))
1268 lustre_swab_lov_user_md_v3(v3);
1270 if (unlikely(magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3)) {
1271 /* try to use as fully defined striping */
1272 rc = lod_use_defined_striping(env, lo, buf);
1276 if (unlikely(buf->lb_len < sizeof(*v1))) {
1277 CERROR("wrong size: %u\n", (unsigned) buf->lb_len);
1281 if (v1->lmm_pattern != 0 && v1->lmm_pattern != LOV_PATTERN_RAID0) {
1282 CERROR("invalid pattern: %x\n", v1->lmm_pattern);
1286 if (v1->lmm_stripe_size)
1287 lo->ldo_stripe_size = v1->lmm_stripe_size;
1288 if (lo->ldo_stripe_size & (LOV_MIN_STRIPE_SIZE - 1))
1289 lo->ldo_stripe_size = LOV_MIN_STRIPE_SIZE;
1291 if (v1->lmm_stripe_count)
1292 lo->ldo_stripenr = v1->lmm_stripe_count;
1294 if ((v1->lmm_stripe_offset >= d->lod_desc.ld_tgt_count) &&
1295 (v1->lmm_stripe_offset != (typeof(v1->lmm_stripe_offset))(-1))) {
1296 CERROR("invalid offset: %x\n", v1->lmm_stripe_offset);
1299 lo->ldo_def_stripe_offset = v1->lmm_stripe_offset;
1301 CDEBUG(D_OTHER, "lsm: %u size, %u stripes, %u offset\n",
1302 v1->lmm_stripe_size, v1->lmm_stripe_count,
1303 v1->lmm_stripe_offset);
1305 if (v1->lmm_magic == LOV_MAGIC_V3) {
1306 if (buf->lb_len < sizeof(*v3)) {
1307 CERROR("wrong size: %u\n", (unsigned) buf->lb_len);
1312 lod_object_set_pool(lo, v3->lmm_pool_name);
1314 pool = lod_find_pool(d, v3->lmm_pool_name);
1316 if (lo->ldo_def_stripe_offset !=
1317 (typeof(v1->lmm_stripe_offset))(-1)) {
1318 rc = lo->ldo_def_stripe_offset;
1319 rc = lod_check_index_in_pool(rc, pool);
1321 lod_pool_putref(pool);
1322 CERROR("invalid offset\n");
1327 if (lo->ldo_stripenr > pool_tgt_count(pool))
1328 lo->ldo_stripenr= pool_tgt_count(pool);
1330 lod_pool_putref(pool);
1333 lod_object_set_pool(lo, NULL);
1339 * buf should be NULL or contain striping settings
1341 int lod_qos_prep_create(const struct lu_env *env, struct lod_object *lo,
1342 struct lu_attr *attr, const struct lu_buf *buf,
1345 struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
1346 int flag = LOV_USES_ASSIGNED_STRIPE;
1352 /* no OST available */
1353 /* XXX: should we be waiting a bit to prevent failures during
1354 * cluster initialization? */
1355 if (d->lod_ostnr == 0)
1356 GOTO(out, rc = -EIO);
1359 * by this time, the object's ldo_stripenr and ldo_stripe_size
1360 * contain default value for striping: taken from the parent
1361 * or from filesystem defaults
1363 * in case the caller is passing lovea with new striping config,
1364 * we may need to parse lovea and apply new configuration
1366 rc = lod_qos_parse_config(env, lo, buf);
1370 if (likely(lo->ldo_stripe == NULL)) {
1372 * no striping has been created so far
1374 LASSERT(lo->ldo_stripenr > 0);
1375 lo->ldo_stripenr = lod_get_stripecnt(d, LOV_MAGIC,
1377 i = sizeof(struct dt_object *) * lo->ldo_stripenr;
1378 OBD_ALLOC(lo->ldo_stripe, i);
1379 if (lo->ldo_stripe == NULL)
1380 GOTO(out, rc = -ENOMEM);
1381 lo->ldo_stripes_allocated = lo->ldo_stripenr;
1384 /* XXX: support for non-0 files w/o objects */
1385 if (lo->ldo_def_stripe_offset >= d->lod_desc.ld_tgt_count) {
1386 lod_qos_statfs_update(env, d);
1387 rc = lod_alloc_qos(env, lo, flag, th);
1389 rc = lod_alloc_rr(env, lo, flag, th);
1391 rc = lod_alloc_specific(env, lo, flag, th);
1395 * lod_qos_parse_config() found supplied buf as a predefined
1396 * striping (not a hint), so it allocated all the object
1397 * now we need to create them
1399 for (i = 0; i < lo->ldo_stripenr; i++) {
1400 struct dt_object *o;
1402 o = lo->ldo_stripe[i];
1405 rc = dt_declare_create(env, o, attr, NULL, NULL, th);
1407 CERROR("can't declare create: %d\n", rc);