lustre/lov/lov_qos.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   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.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_LOV
  38
  39 #ifdef __KERNEL__
  40 #include <libcfs/libcfs.h>
  41 #else
  42 #include <liblustre.h>
  43 #endif
  44
  45 #include <obd_class.h>
  46 #include <obd_lov.h>
  47 #include <lustre/lustre_idl.h>
  48 #include "lov_internal.h"
  49
  50 /* #define QOS_DEBUG 1 */
  51 #define D_QOS D_OTHER
  52
  53 #define TGT_BAVAIL(i) (lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bavail *\
  54                        lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bsize)
  55
  56
  57 int qos_add_tgt(struct obd_device *obd, __u32 index)
  58 {
  59         struct lov_obd *lov = &obd->u.lov;
  60         struct lov_qos_oss *oss, *temposs;
  61         struct obd_export *exp = lov->lov_tgts[index]->ltd_exp;
  62         int rc = 0, found = 0;
  63         ENTRY;
  64
  65         /* We only need this QOS struct on MDT, not clients - but we may not
  66          * have registered the LOV's observer yet, so there's no way to know */
  67         if (!exp || !exp->exp_connection) {
  68                 CERROR("Missing connection\n");
  69                 RETURN(-ENOTCONN);
  70         }
  71
  72         cfs_down_write(&lov->lov_qos.lq_rw_sem);
  73         cfs_mutex_lock(&lov->lov_lock);
  74         cfs_list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
  75                 if (obd_uuid_equals(&oss->lqo_uuid,
  76                                     &exp->exp_connection->c_remote_uuid)) {
  77                         found++;
  78                         break;
  79                 }
  80         }
  81
  82         if (!found) {
  83                 OBD_ALLOC_PTR(oss);
  84                 if (!oss)
  85                         GOTO(out, rc = -ENOMEM);
  86                 memcpy(&oss->lqo_uuid,
  87                        &exp->exp_connection->c_remote_uuid,
  88                        sizeof(oss->lqo_uuid));
  89         } else {
  90                 /* Assume we have to move this one */
  91                 cfs_list_del(&oss->lqo_oss_list);
  92         }
  93
  94         oss->lqo_ost_count++;
  95         lov->lov_tgts[index]->ltd_qos.ltq_oss = oss;
  96
  97         /* Add sorted by # of OSTs.  Find the first entry that we're
  98            bigger than... */
  99         cfs_list_for_each_entry(temposs, &lov->lov_qos.lq_oss_list,
 100                                 lqo_oss_list) {
 101                 if (oss->lqo_ost_count > temposs->lqo_ost_count)
 102                         break;
 103         }
 104         /* ...and add before it.  If we're the first or smallest, temposs
 105            points to the list head, and we add to the end. */
 106         cfs_list_add_tail(&oss->lqo_oss_list, &temposs->lqo_oss_list);
 107
 108         lov->lov_qos.lq_dirty = 1;
 109         lov->lov_qos.lq_rr.lqr_dirty = 1;
 110
 111         CDEBUG(D_QOS, "add tgt %s to OSS %s (%d OSTs)\n",
 112                obd_uuid2str(&lov->lov_tgts[index]->ltd_uuid),
 113                obd_uuid2str(&oss->lqo_uuid),
 114                oss->lqo_ost_count);
 115
 116 out:
 117         cfs_mutex_unlock(&lov->lov_lock);
 118         cfs_up_write(&lov->lov_qos.lq_rw_sem);
 119         RETURN(rc);
 120 }
 121
 122 int qos_del_tgt(struct obd_device *obd, struct lov_tgt_desc *tgt)
 123 {
 124         struct lov_obd *lov = &obd->u.lov;
 125         struct lov_qos_oss *oss;
 126         int rc = 0;
 127         ENTRY;
 128
 129         cfs_down_write(&lov->lov_qos.lq_rw_sem);
 130
 131         oss = tgt->ltd_qos.ltq_oss;
 132         if (!oss)
 133                 GOTO(out, rc = -ENOENT);
 134
 135         oss->lqo_ost_count--;
 136         if (oss->lqo_ost_count == 0) {
 137                 CDEBUG(D_QOS, "removing OSS %s\n",
 138                        obd_uuid2str(&oss->lqo_uuid));
 139                 cfs_list_del(&oss->lqo_oss_list);
 140                 OBD_FREE_PTR(oss);
 141         }
 142
 143         lov->lov_qos.lq_dirty = 1;
 144         lov->lov_qos.lq_rr.lqr_dirty = 1;
 145 out:
 146         cfs_up_write(&lov->lov_qos.lq_rw_sem);
 147         RETURN(rc);
 148 }
 149
 150 /* Recalculate per-object penalties for OSSs and OSTs,
 151    depends on size of each ost in an oss */
 152 static int qos_calc_ppo(struct obd_device *obd)
 153 {
 154         struct lov_obd *lov = &obd->u.lov;
 155         struct lov_qos_oss *oss;
 156         __u64 ba_max, ba_min, temp;
 157         __u32 num_active;
 158         int rc, i, prio_wide;
 159         time_t now, age;
 160         ENTRY;
 161
 162         if (!lov->lov_qos.lq_dirty)
 163                 GOTO(out, rc = 0);
 164
 165         num_active = lov->desc.ld_active_tgt_count - 1;
 166         if (num_active < 1)
 167                 GOTO(out, rc = -EAGAIN);
 168
 169         /* find bavail on each OSS */
 170         cfs_list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 171                 oss->lqo_bavail = 0;
 172         }
 173         lov->lov_qos.lq_active_oss_count = 0;
 174
 175         /* How badly user wants to select osts "widely" (not recently chosen
 176            and not on recent oss's).  As opposed to "freely" (free space
 177            avail.) 0-256. */
 178         prio_wide = 256 - lov->lov_qos.lq_prio_free;
 179
 180         ba_min = (__u64)(-1);
 181         ba_max = 0;
 182         now = cfs_time_current_sec();
 183         /* Calculate OST penalty per object */
 184         /* (lov ref taken in alloc_qos) */
 185         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 186                 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
 187                         continue;
 188                 temp = TGT_BAVAIL(i);
 189                 if (!temp)
 190                         continue;
 191                 ba_min = min(temp, ba_min);
 192                 ba_max = max(temp, ba_max);
 193
 194                 /* Count the number of usable OSS's */
 195                 if (lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_bavail == 0)
 196                         lov->lov_qos.lq_active_oss_count++;
 197                 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_bavail += temp;
 198
 199                 /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
 200                 temp >>= 1;
 201                 lov_do_div64(temp, num_active);
 202                 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj =
 203                         (temp * prio_wide) >> 8;
 204
 205                 age = (now - lov->lov_tgts[i]->ltd_qos.ltq_used) >> 3;
 206                 if (lov->lov_qos.lq_reset || age > 32 * lov->desc.ld_qos_maxage)
 207                         lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
 208                 else if (age > lov->desc.ld_qos_maxage)
 209                         /* Decay the penalty by half for every 8x the update
 210                          * interval that the device has been idle.  That gives
 211                          * lots of time for the statfs information to be
 212                          * updated (which the penalty is only a proxy for),
 213                          * and avoids penalizing OSS/OSTs under light load. */
 214                         lov->lov_tgts[i]->ltd_qos.ltq_penalty >>=
 215                                 (age / lov->desc.ld_qos_maxage);
 216         }
 217
 218         num_active = lov->lov_qos.lq_active_oss_count - 1;
 219         if (num_active < 1) {
 220                 /* If there's only 1 OSS, we can't penalize it, so instead
 221                    we have to double the OST penalty */
 222                 num_active = 1;
 223                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 224                         if (lov->lov_tgts[i] == NULL)
 225                                 continue;
 226                         lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj <<= 1;
 227                 }
 228         }
 229
 230         /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
 231         cfs_list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 232                 temp = oss->lqo_bavail >> 1;
 233                 lov_do_div64(temp, oss->lqo_ost_count * num_active);
 234                 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
 235
 236                 age = (now - oss->lqo_used) >> 3;
 237                 if (lov->lov_qos.lq_reset || age > 32 * lov->desc.ld_qos_maxage)
 238                         oss->lqo_penalty = 0;
 239                 else if (age > lov->desc.ld_qos_maxage)
 240                         /* Decay the penalty by half for every 8x the update
 241                          * interval that the device has been idle.  That gives
 242                          * lots of time for the statfs information to be
 243                          * updated (which the penalty is only a proxy for),
 244                          * and avoids penalizing OSS/OSTs under light load. */
 245                         oss->lqo_penalty >>= (age / lov->desc.ld_qos_maxage);
 246         }
 247
 248         lov->lov_qos.lq_dirty = 0;
 249         lov->lov_qos.lq_reset = 0;
 250
 251         /* If each ost has almost same free space,
 252          * do rr allocation for better creation performance */
 253         lov->lov_qos.lq_same_space = 0;
 254         if ((ba_max * (256 - lov->lov_qos.lq_threshold_rr)) >> 8 < ba_min) {
 255                 lov->lov_qos.lq_same_space = 1;
 256                 /* Reset weights for the next time we enter qos mode */
 257                 lov->lov_qos.lq_reset = 1;
 258         }
 259         rc = 0;
 260
 261 out:
 262         if (!rc && lov->lov_qos.lq_same_space)
 263                 RETURN(-EAGAIN);
 264         RETURN(rc);
 265 }
 266
 267 static int qos_calc_weight(struct lov_obd *lov, int i)
 268 {
 269         __u64 temp, temp2;
 270
 271         /* Final ost weight = TGT_BAVAIL - ost_penalty - oss_penalty */
 272         temp = TGT_BAVAIL(i);
 273         temp2 = lov->lov_tgts[i]->ltd_qos.ltq_penalty +
 274                 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty;
 275         if (temp < temp2)
 276                 lov->lov_tgts[i]->ltd_qos.ltq_weight = 0;
 277         else
 278                 lov->lov_tgts[i]->ltd_qos.ltq_weight = temp - temp2;
 279         return 0;
 280 }
 281
 282 /* We just used this index for a stripe; adjust everyone's weights */
 283 static int qos_used(struct lov_obd *lov, struct ost_pool *osts,
 284                     __u32 index, __u64 *total_wt)
 285 {
 286         struct lov_qos_oss *oss;
 287         int j;
 288         ENTRY;
 289
 290         /* Don't allocate from this stripe anymore, until the next alloc_qos */
 291         lov->lov_tgts[index]->ltd_qos.ltq_usable = 0;
 292
 293         oss = lov->lov_tgts[index]->ltd_qos.ltq_oss;
 294
 295         /* Decay old penalty by half (we're adding max penalty, and don't
 296            want it to run away.) */
 297         lov->lov_tgts[index]->ltd_qos.ltq_penalty >>= 1;
 298         oss->lqo_penalty >>= 1;
 299
 300         /* mark the OSS and OST as recently used */
 301         lov->lov_tgts[index]->ltd_qos.ltq_used =
 302                 oss->lqo_used = cfs_time_current_sec();
 303
 304         /* Set max penalties for this OST and OSS */
 305         lov->lov_tgts[index]->ltd_qos.ltq_penalty +=
 306                 lov->lov_tgts[index]->ltd_qos.ltq_penalty_per_obj *
 307                 lov->desc.ld_active_tgt_count;
 308         oss->lqo_penalty += oss->lqo_penalty_per_obj *
 309                 lov->lov_qos.lq_active_oss_count;
 310
 311         /* Decrease all OSS penalties */
 312         cfs_list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 313                 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
 314                         oss->lqo_penalty = 0;
 315                 else
 316                         oss->lqo_penalty -= oss->lqo_penalty_per_obj;
 317         }
 318
 319         *total_wt = 0;
 320         /* Decrease all OST penalties */
 321         for (j = 0; j < osts->op_count; j++) {
 322                 int i;
 323
 324                 i = osts->op_array[j];
 325                 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
 326                         continue;
 327                 if (lov->lov_tgts[i]->ltd_qos.ltq_penalty <
 328                     lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj)
 329                         lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
 330                 else
 331                         lov->lov_tgts[i]->ltd_qos.ltq_penalty -=
 332                         lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj;
 333
 334                 qos_calc_weight(lov, i);
 335
 336                 /* Recalc the total weight of usable osts */
 337                 if (lov->lov_tgts[i]->ltd_qos.ltq_usable)
 338                         *total_wt += lov->lov_tgts[i]->ltd_qos.ltq_weight;
 339
 340 #ifdef QOS_DEBUG
 341                 CDEBUG(D_QOS, "recalc tgt %d usable=%d avail="LPU64
 342                        " ostppo="LPU64" ostp="LPU64" ossppo="LPU64
 343                        " ossp="LPU64" wt="LPU64"\n",
 344                        i, lov->lov_tgts[i]->ltd_qos.ltq_usable,
 345                        TGT_BAVAIL(i) >> 10,
 346                        lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj >> 10,
 347                        lov->lov_tgts[i]->ltd_qos.ltq_penalty >> 10,
 348                        lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty_per_obj>>10,
 349                        lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty >> 10,
 350                        lov->lov_tgts[i]->ltd_qos.ltq_weight >> 10);
 351 #endif
 352         }
 353
 354         RETURN(0);
 355 }
 356
 357 #define LOV_QOS_EMPTY ((__u32)-1)
 358 /* compute optimal round-robin order, based on OSTs per OSS */
 359 static int qos_calc_rr(struct lov_obd *lov, struct ost_pool *src_pool,
 360                        struct lov_qos_rr *lqr)
 361 {
 362         struct lov_qos_oss *oss;
 363         unsigned placed, real_count;
 364         int i, rc;
 365         ENTRY;
 366
 367         if (!lqr->lqr_dirty) {
 368                 LASSERT(lqr->lqr_pool.op_size);
 369                 RETURN(0);
 370         }
 371
 372         /* Do actual allocation. */
 373         cfs_down_write(&lov->lov_qos.lq_rw_sem);
 374
 375         /*
 376          * Check again. While we were sleeping on @lq_rw_sem something could
 377          * change.
 378          */
 379         if (!lqr->lqr_dirty) {
 380                 LASSERT(lqr->lqr_pool.op_size);
 381                 cfs_up_write(&lov->lov_qos.lq_rw_sem);
 382                 RETURN(0);
 383         }
 384
 385         real_count = src_pool->op_count;
 386
 387         /* Zero the pool array */
 388         /* alloc_rr is holding a read lock on the pool, so nobody is adding/
 389            deleting from the pool. The lq_rw_sem insures that nobody else
 390            is reading. */
 391         lqr->lqr_pool.op_count = real_count;
 392         rc = lov_ost_pool_extend(&lqr->lqr_pool, real_count);
 393         if (rc) {
 394                 cfs_up_write(&lov->lov_qos.lq_rw_sem);
 395                 RETURN(rc);
 396         }
 397         for (i = 0; i < lqr->lqr_pool.op_count; i++)
 398                 lqr->lqr_pool.op_array[i] = LOV_QOS_EMPTY;
 399
 400         /* Place all the OSTs from 1 OSS at the same time. */
 401         placed = 0;
 402         cfs_list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 403                 int j = 0;
 404                 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
 405                         if (lov->lov_tgts[src_pool->op_array[i]] &&
 406                             (lov->lov_tgts[src_pool->op_array[i]]->ltd_qos.ltq_oss == oss)) {
 407                               /* Evenly space these OSTs across arrayspace */
 408                               int next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count;
 409                               while (lqr->lqr_pool.op_array[next] !=
 410                                      LOV_QOS_EMPTY)
 411                                         next = (next + 1) % lqr->lqr_pool.op_count;
 412                               lqr->lqr_pool.op_array[next] = src_pool->op_array[i];
 413                               j++;
 414                               placed++;
 415                         }
 416                 }
 417         }
 418
 419         lqr->lqr_dirty = 0;
 420         cfs_up_write(&lov->lov_qos.lq_rw_sem);
 421
 422         if (placed != real_count) {
 423                 /* This should never happen */
 424                 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
 425                                    "round-robin list (%d of %d).\n",
 426                                    placed, real_count);
 427                 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
 428                         LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
 429                                  lqr->lqr_pool.op_array[i]);
 430                 }
 431                 lqr->lqr_dirty = 1;
 432                 RETURN(-EAGAIN);
 433         }
 434
 435 #ifdef QOS_DEBUG
 436         for (i = 0; i < lqr->lqr_pool.op_count; i++) {
 437                 LCONSOLE(D_QOS, "rr #%d ost idx=%d\n", i,
 438                          lqr->lqr_pool.op_array[i]);
 439         }
 440 #endif
 441
 442         RETURN(0);
 443 }
 444
 445
 446 void qos_shrink_lsm(struct lov_request_set *set)
 447 {
 448         struct lov_stripe_md *lsm = set->set_oi->oi_md, *lsm_new;
 449         /* XXX LOV STACKING call into osc for sizes */
 450         unsigned oldsize, newsize;
 451
 452         if (set->set_oti && set->set_cookies && set->set_cookie_sent) {
 453                 struct llog_cookie *cookies;
 454                 oldsize = lsm->lsm_stripe_count * sizeof(*cookies);
 455                 newsize = set->set_count * sizeof(*cookies);
 456
 457                 cookies = set->set_cookies;
 458                 oti_alloc_cookies(set->set_oti, set->set_count);
 459                 if (set->set_oti->oti_logcookies) {
 460                         memcpy(set->set_oti->oti_logcookies, cookies, newsize);
 461                         OBD_FREE_LARGE(cookies, oldsize);
 462                         set->set_cookies = set->set_oti->oti_logcookies;
 463                 } else {
 464                         CWARN("'leaking' %d bytes\n", oldsize - newsize);
 465                 }
 466         }
 467
 468         CWARN("using fewer stripes for object "LPU64": old %u new %u\n",
 469               lsm->lsm_object_id, lsm->lsm_stripe_count, set->set_count);
 470         LASSERT(lsm->lsm_stripe_count >= set->set_count);
 471
 472         newsize = lov_stripe_md_size(set->set_count);
 473         OBD_ALLOC_LARGE(lsm_new, newsize);
 474         if (lsm_new != NULL) {
 475                 int i;
 476                 memcpy(lsm_new, lsm, sizeof(*lsm));
 477                 for (i = 0; i < lsm->lsm_stripe_count; i++) {
 478                         if (i < set->set_count) {
 479                                 lsm_new->lsm_oinfo[i] = lsm->lsm_oinfo[i];
 480                                 continue;
 481                         }
 482                         OBD_SLAB_FREE(lsm->lsm_oinfo[i], lov_oinfo_slab,
 483                                       sizeof(struct lov_oinfo));
 484                 }
 485                 lsm_new->lsm_stripe_count = set->set_count;
 486                 OBD_FREE_LARGE(lsm, sizeof(struct lov_stripe_md) +
 487                                lsm->lsm_stripe_count*sizeof(struct lov_oinfo*));
 488                 set->set_oi->oi_md = lsm_new;
 489         } else {
 490                 CWARN("'leaking' few bytes\n");
 491         }
 492 }
 493
 494 /**
 495  * Check whether we can create the object on the OST(refered by ost_idx)
 496  * \retval:
 497  *          0: create the object.
 498  *          other value: did not create the object.
 499  */
 500 static int lov_check_and_create_object(struct lov_obd *lov, int ost_idx,
 501                                        struct lov_stripe_md *lsm,
 502                                        struct lov_request *req,
 503                                        struct obd_trans_info *oti)
 504 {
 505         __u16 stripe;
 506         int rc = -EIO;
 507         ENTRY;
 508
 509         CDEBUG(D_QOS, "Check and create on idx %d \n", ost_idx);
 510         if (!lov->lov_tgts[ost_idx] ||
 511             !lov->lov_tgts[ost_idx]->ltd_active)
 512                 RETURN(rc);
 513
 514         /* check if objects has been created on this ost */
 515         for (stripe = 0; stripe < lsm->lsm_stripe_count; stripe++) {
 516                 /* already have object at this stripe */
 517                 if (ost_idx == lsm->lsm_oinfo[stripe]->loi_ost_idx)
 518                         break;
 519         }
 520
 521         if (stripe >= lsm->lsm_stripe_count) {
 522                 req->rq_idx = ost_idx;
 523                 rc = obd_create(NULL, lov->lov_tgts[ost_idx]->ltd_exp,
 524                                 req->rq_oi.oi_oa, &req->rq_oi.oi_md,
 525                                 oti);
 526         }
 527         RETURN(rc);
 528 }
 529
 530 int qos_remedy_create(struct lov_request_set *set, struct lov_request *req)
 531 {
 532         struct lov_stripe_md *lsm = set->set_oi->oi_md;
 533         struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
 534         unsigned ost_idx = 0, ost_count;
 535         struct pool_desc *pool;
 536         struct ost_pool *osts = NULL;
 537         int i, rc = -EIO;
 538         ENTRY;
 539
 540         /* First check whether we can create the objects on the pool */
 541         pool = lov_find_pool(lov, lsm->lsm_pool_name);
 542         if (pool != NULL) {
 543                 cfs_down_read(&pool_tgt_rw_sem(pool));
 544                 osts = &(pool->pool_obds);
 545                 ost_count = osts->op_count;
 546                 for (i = 0; i < ost_count; i++, ost_idx = osts->op_array[i]) {
 547                         rc = lov_check_and_create_object(lov, ost_idx, lsm, req,
 548                                                          set->set_oti);
 549                         if (rc == 0)
 550                                 break;
 551                 }
 552                 cfs_up_read(&pool_tgt_rw_sem(pool));
 553                 lov_pool_putref(pool);
 554                 RETURN(rc);
 555         }
 556
 557         ost_count = lov->desc.ld_tgt_count;
 558         /* Then check whether we can create the objects on other OSTs */
 559         ost_idx = (req->rq_idx + lsm->lsm_stripe_count) % ost_count;
 560         for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
 561                 rc = lov_check_and_create_object(lov, ost_idx, lsm, req,
 562                                                  set->set_oti);
 563
 564                 if (rc == 0)
 565                         break;
 566         }
 567
 568         RETURN(rc);
 569 }
 570
 571 static int min_stripe_count(int stripe_cnt, int flags)
 572 {
 573         return (flags & LOV_USES_DEFAULT_STRIPE ?
 574                 stripe_cnt - (stripe_cnt / 4) : stripe_cnt);
 575 }
 576
 577 #define LOV_CREATE_RESEED_MULT 30
 578 #define LOV_CREATE_RESEED_MIN  2000
 579 /* Allocate objects on osts with round-robin algorithm */
 580 static int alloc_rr(struct lov_obd *lov, int *idx_arr, int *stripe_cnt,
 581                     char *poolname, int flags)
 582 {
 583         unsigned array_idx;
 584         int i, rc, *idx_pos;
 585         __u32 ost_idx;
 586         int ost_start_idx_temp;
 587         int speed = 0;
 588         int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
 589         struct pool_desc *pool;
 590         struct ost_pool *osts;
 591         struct lov_qos_rr *lqr;
 592         ENTRY;
 593
 594         pool = lov_find_pool(lov, poolname);
 595         if (pool == NULL) {
 596                 osts = &(lov->lov_packed);
 597                 lqr = &(lov->lov_qos.lq_rr);
 598         } else {
 599                 cfs_down_read(&pool_tgt_rw_sem(pool));
 600                 osts = &(pool->pool_obds);
 601                 lqr = &(pool->pool_rr);
 602         }
 603
 604         rc = qos_calc_rr(lov, osts, lqr);
 605         if (rc)
 606                 GOTO(out, rc);
 607
 608         if (--lqr->lqr_start_count <= 0) {
 609                 lqr->lqr_start_idx = cfs_rand() % osts->op_count;
 610                 lqr->lqr_start_count =
 611                         (LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) +
 612                          LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U);
 613         } else if (stripe_cnt_min >= osts->op_count ||
 614                    lqr->lqr_start_idx > osts->op_count) {
 615                 /* If we have allocated from all of the OSTs, slowly
 616                  * precess the next start if the OST/stripe count isn't
 617                  * already doing this for us. */
 618                 lqr->lqr_start_idx %= osts->op_count;
 619                 if (*stripe_cnt > 1 && (osts->op_count % (*stripe_cnt)) != 1)
 620                         ++lqr->lqr_offset_idx;
 621         }
 622         cfs_down_read(&lov->lov_qos.lq_rw_sem);
 623         ost_start_idx_temp = lqr->lqr_start_idx;
 624
 625 repeat_find:
 626         array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) % osts->op_count;
 627         idx_pos = idx_arr;
 628 #ifdef QOS_DEBUG
 629         CDEBUG(D_QOS, "pool '%s' want %d startidx %d startcnt %d offset %d "
 630                "active %d count %d arrayidx %d\n", poolname,
 631                *stripe_cnt, lqr->lqr_start_idx, lqr->lqr_start_count,
 632                lqr->lqr_offset_idx, osts->op_count, osts->op_count, array_idx);
 633 #endif
 634
 635         for (i = 0; i < osts->op_count;
 636                     i++, array_idx=(array_idx + 1) % osts->op_count) {
 637                 ++lqr->lqr_start_idx;
 638                 ost_idx = lqr->lqr_pool.op_array[array_idx];
 639 #ifdef QOS_DEBUG
 640                 CDEBUG(D_QOS, "#%d strt %d act %d strp %d ary %d idx %d\n",
 641                        i, lqr->lqr_start_idx,
 642                        ((ost_idx != LOV_QOS_EMPTY) && lov->lov_tgts[ost_idx]) ?
 643                        lov->lov_tgts[ost_idx]->ltd_active : 0,
 644                        idx_pos - idx_arr, array_idx, ost_idx);
 645 #endif
 646                 if ((ost_idx == LOV_QOS_EMPTY) || !lov->lov_tgts[ost_idx] ||
 647                     !lov->lov_tgts[ost_idx]->ltd_active)
 648                         continue;
 649
 650                 /* Fail Check before osc_precreate() is called
 651                    so we can only 'fail' single OSC. */
 652                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
 653                         continue;
 654
 655                 /* Drop slow OSCs if we can */
 656                 if (obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed)
 657                         continue;
 658
 659                 *idx_pos = ost_idx;
 660                 idx_pos++;
 661                 /* We have enough stripes */
 662                 if (idx_pos - idx_arr == *stripe_cnt)
 663                         break;
 664         }
 665         if ((speed < 2) && (idx_pos - idx_arr < stripe_cnt_min)) {
 666                 /* Try again, allowing slower OSCs */
 667                 speed++;
 668                 lqr->lqr_start_idx = ost_start_idx_temp;
 669                 goto repeat_find;
 670         }
 671
 672         cfs_up_read(&lov->lov_qos.lq_rw_sem);
 673
 674         *stripe_cnt = idx_pos - idx_arr;
 675 out:
 676         if (pool != NULL) {
 677                 cfs_up_read(&pool_tgt_rw_sem(pool));
 678                 /* put back ref got by lov_find_pool() */
 679                 lov_pool_putref(pool);
 680         }
 681
 682         RETURN(rc);
 683 }
 684
 685 /* alloc objects on osts with specific stripe offset */
 686 static int alloc_specific(struct lov_obd *lov, struct lov_stripe_md *lsm,
 687                           int *idx_arr)
 688 {
 689         unsigned ost_idx, array_idx, ost_count;
 690         int i, rc, *idx_pos;
 691         int speed = 0;
 692         struct pool_desc *pool;
 693         struct ost_pool *osts;
 694         ENTRY;
 695
 696         pool = lov_find_pool(lov, lsm->lsm_pool_name);
 697         if (pool == NULL) {
 698                 osts = &(lov->lov_packed);
 699         } else {
 700                 cfs_down_read(&pool_tgt_rw_sem(pool));
 701                 osts = &(pool->pool_obds);
 702         }
 703
 704         ost_count = osts->op_count;
 705
 706 repeat_find:
 707         /* search loi_ost_idx in ost array */
 708         array_idx = 0;
 709         for (i = 0; i < ost_count; i++) {
 710                 if (osts->op_array[i] == lsm->lsm_oinfo[0]->loi_ost_idx) {
 711                         array_idx = i;
 712                         break;
 713                 }
 714         }
 715         if (i == ost_count) {
 716                 CERROR("Start index %d not found in pool '%s'\n",
 717                        lsm->lsm_oinfo[0]->loi_ost_idx, lsm->lsm_pool_name);
 718                 GOTO(out, rc = -EINVAL);
 719         }
 720
 721         idx_pos = idx_arr;
 722         for (i = 0; i < ost_count;
 723              i++, array_idx = (array_idx + 1) % ost_count) {
 724                 ost_idx = osts->op_array[array_idx];
 725
 726                 if (!lov->lov_tgts[ost_idx] ||
 727                     !lov->lov_tgts[ost_idx]->ltd_active) {
 728                         continue;
 729                 }
 730
 731                 /* Fail Check before osc_precreate() is called
 732                    so we can only 'fail' single OSC. */
 733                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
 734                         continue;
 735
 736                 /* Drop slow OSCs if we can, but not for requested start idx.
 737                  *
 738                  * This means "if OSC is slow and it is not the requested
 739                  * start OST, then it can be skipped, otherwise skip it only
 740                  * if it is inactive/recovering/out-of-space." */
 741                 if ((obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed) &&
 742                     (i != 0 || speed >= 2))
 743                         continue;
 744
 745                 *idx_pos = ost_idx;
 746                 idx_pos++;
 747                 /* We have enough stripes */
 748                 if (idx_pos - idx_arr == lsm->lsm_stripe_count)
 749                         GOTO(out, rc = 0);
 750         }
 751         if (speed < 2) {
 752                 /* Try again, allowing slower OSCs */
 753                 speed++;
 754                 goto repeat_find;
 755         }
 756
 757         /* If we were passed specific striping params, then a failure to
 758          * meet those requirements is an error, since we can't reallocate
 759          * that memory (it might be part of a larger array or something).
 760          *
 761          * We can only get here if lsm_stripe_count was originally > 1.
 762          */
 763         CERROR("can't lstripe objid "LPX64": have %d want %u\n",
 764                lsm->lsm_object_id, (int)(idx_pos - idx_arr),
 765                lsm->lsm_stripe_count);
 766         rc = -EFBIG;
 767 out:
 768         if (pool != NULL) {
 769                 cfs_up_read(&pool_tgt_rw_sem(pool));
 770                 /* put back ref got by lov_find_pool() */
 771                 lov_pool_putref(pool);
 772         }
 773
 774         RETURN(rc);
 775 }
 776
 777 /* Alloc objects on osts with optimization based on:
 778    - free space
 779    - network resources (shared OSS's)
 780 */
 781 static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt,
 782                      char *poolname, int flags)
 783 {
 784         struct lov_obd *lov = &exp->exp_obd->u.lov;
 785         __u64 total_weight = 0;
 786         int nfound, good_osts, i, rc = 0;
 787         int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
 788         struct pool_desc *pool;
 789         struct ost_pool *osts;
 790         ENTRY;
 791
 792         if (stripe_cnt_min < 1)
 793                 RETURN(-EINVAL);
 794
 795         pool = lov_find_pool(lov, poolname);
 796         if (pool == NULL) {
 797                 osts = &(lov->lov_packed);
 798         } else {
 799                 cfs_down_read(&pool_tgt_rw_sem(pool));
 800                 osts = &(pool->pool_obds);
 801         }
 802
 803         obd_getref(exp->exp_obd);
 804
 805         /* wait for fresh statfs info if needed, the rpcs are sent in
 806          * lov_create() */
 807         qos_statfs_update(exp->exp_obd,
 808                           cfs_time_shift_64(-2 * lov->desc.ld_qos_maxage), 1);
 809
 810         /* Detect -EAGAIN early, before expensive lock is taken. */
 811         if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space)
 812                 GOTO(out_nolock, rc = -EAGAIN);
 813
 814         /* Do actual allocation, use write lock here. */
 815         cfs_down_write(&lov->lov_qos.lq_rw_sem);
 816
 817         /*
 818          * Check again, while we were sleeping on @lq_rw_sem things could
 819          * change.
 820          */
 821         if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space)
 822                 GOTO(out, rc = -EAGAIN);
 823
 824         if (lov->desc.ld_active_tgt_count < 2)
 825                 GOTO(out, rc = -EAGAIN);
 826
 827         rc = qos_calc_ppo(exp->exp_obd);
 828         if (rc)
 829                 GOTO(out, rc);
 830
 831         good_osts = 0;
 832         /* Find all the OSTs that are valid stripe candidates */
 833         for (i = 0; i < osts->op_count; i++) {
 834                 if (!lov->lov_tgts[osts->op_array[i]] ||
 835                     !lov->lov_tgts[osts->op_array[i]]->ltd_active)
 836                         continue;
 837
 838                 /* Fail Check before osc_precreate() is called
 839                    so we can only 'fail' single OSC. */
 840                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && osts->op_array[i] == 0)
 841                         continue;
 842
 843                 if (obd_precreate(lov->lov_tgts[osts->op_array[i]]->ltd_exp) > 2)
 844                         continue;
 845
 846                 lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable = 1;
 847                 qos_calc_weight(lov, osts->op_array[i]);
 848                 total_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight;
 849
 850                 good_osts++;
 851         }
 852
 853 #ifdef QOS_DEBUG
 854         CDEBUG(D_QOS, "found %d good osts\n", good_osts);
 855 #endif
 856
 857         if (good_osts < stripe_cnt_min)
 858                 GOTO(out, rc = -EAGAIN);
 859
 860         /* We have enough osts */
 861         if (good_osts < *stripe_cnt)
 862                 *stripe_cnt = good_osts;
 863
 864         if (!*stripe_cnt)
 865                 GOTO(out, rc = -EAGAIN);
 866
 867         /* Find enough OSTs with weighted random allocation. */
 868         nfound = 0;
 869         while (nfound < *stripe_cnt) {
 870                 __u64 rand, cur_weight;
 871
 872                 cur_weight = 0;
 873                 rc = -ENODEV;
 874
 875                 if (total_weight) {
 876 #if BITS_PER_LONG == 32
 877                         rand = cfs_rand() % (unsigned)total_weight;
 878                         /* If total_weight > 32-bit, first generate the high
 879                          * 32 bits of the random number, then add in the low
 880                          * 32 bits (truncated to the upper limit, if needed) */
 881                         if (total_weight > 0xffffffffULL)
 882                                 rand = (__u64)(cfs_rand() %
 883                                           (unsigned)(total_weight >> 32)) << 32;
 884                         else
 885                                 rand = 0;
 886
 887                         if (rand == (total_weight & 0xffffffff00000000ULL))
 888                                 rand |= cfs_rand() % (unsigned)total_weight;
 889                         else
 890                                 rand |= cfs_rand();
 891
 892 #else
 893                         rand = ((__u64)cfs_rand() << 32 | cfs_rand()) %
 894                                 total_weight;
 895 #endif
 896                 } else {
 897                         rand = 0;
 898                 }
 899
 900                 /* On average, this will hit larger-weighted osts more often.
 901                    0-weight osts will always get used last (only when rand=0).*/
 902                 for (i = 0; i < osts->op_count; i++) {
 903                         if (!lov->lov_tgts[osts->op_array[i]] ||
 904                             !lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable)
 905                                 continue;
 906
 907                         cur_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight;
 908 #ifdef QOS_DEBUG
 909                         CDEBUG(D_QOS, "stripe_cnt=%d nfound=%d cur_weight="LPU64
 910                                       " rand="LPU64" total_weight="LPU64"\n",
 911                                *stripe_cnt, nfound, cur_weight, rand, total_weight);
 912 #endif
 913                         if (cur_weight >= rand) {
 914 #ifdef QOS_DEBUG
 915                                 CDEBUG(D_QOS, "assigned stripe=%d to idx=%d\n",
 916                                        nfound, osts->op_array[i]);
 917 #endif
 918                                 idx_arr[nfound++] = osts->op_array[i];
 919                                 qos_used(lov, osts, osts->op_array[i], &total_weight);
 920                                 rc = 0;
 921                                 break;
 922                         }
 923                 }
 924                 /* should never satisfy below condition */
 925                 if (rc) {
 926                         CERROR("Didn't find any OSTs?\n");
 927                         break;
 928                 }
 929         }
 930         LASSERT(nfound == *stripe_cnt);
 931
 932 out:
 933         cfs_up_write(&lov->lov_qos.lq_rw_sem);
 934
 935 out_nolock:
 936         if (pool != NULL) {
 937                 cfs_up_read(&pool_tgt_rw_sem(pool));
 938                 /* put back ref got by lov_find_pool() */
 939                 lov_pool_putref(pool);
 940         }
 941
 942         if (rc == -EAGAIN)
 943                 rc = alloc_rr(lov, idx_arr, stripe_cnt, poolname, flags);
 944
 945         obd_putref(exp->exp_obd);
 946         RETURN(rc);
 947 }
 948
 949 /* return new alloced stripe count on success */
 950 static int alloc_idx_array(struct obd_export *exp, struct lov_stripe_md *lsm,
 951                            int newea, int **idx_arr, int *arr_cnt, int flags)
 952 {
 953         struct lov_obd *lov = &exp->exp_obd->u.lov;
 954         int stripe_cnt = lsm->lsm_stripe_count;
 955         int i, rc = 0;
 956         int *tmp_arr = NULL;
 957         ENTRY;
 958
 959         *arr_cnt = stripe_cnt;
 960         OBD_ALLOC(tmp_arr, *arr_cnt * sizeof(int));
 961         if (tmp_arr == NULL)
 962                 RETURN(-ENOMEM);
 963         for (i = 0; i < *arr_cnt; i++)
 964                 tmp_arr[i] = -1;
 965
 966         if (newea ||
 967             lsm->lsm_oinfo[0]->loi_ost_idx >= lov->desc.ld_tgt_count)
 968                 rc = alloc_qos(exp, tmp_arr, &stripe_cnt,
 969                                lsm->lsm_pool_name, flags);
 970         else
 971                 rc = alloc_specific(lov, lsm, tmp_arr);
 972
 973         if (rc)
 974                 GOTO(out_arr, rc);
 975
 976         *idx_arr = tmp_arr;
 977         RETURN(stripe_cnt);
 978 out_arr:
 979         OBD_FREE(tmp_arr, *arr_cnt * sizeof(int));
 980         *arr_cnt = 0;
 981         RETURN(rc);
 982 }
 983
 984 static void free_idx_array(int *idx_arr, int arr_cnt)
 985 {
 986         if (arr_cnt)
 987                 OBD_FREE(idx_arr, arr_cnt * sizeof(int));
 988 }
 989
 990 int qos_prep_create(struct obd_export *exp, struct lov_request_set *set)
 991 {
 992         struct lov_obd *lov = &exp->exp_obd->u.lov;
 993         struct lov_stripe_md *lsm;
 994         struct obdo *src_oa = set->set_oi->oi_oa;
 995         struct obd_trans_info *oti = set->set_oti;
 996         int i, stripes, rc = 0, newea = 0;
 997         int flag = LOV_USES_ASSIGNED_STRIPE;
 998         int *idx_arr = NULL, idx_cnt = 0;
 999         ENTRY;
1000
1001         LASSERT(src_oa->o_valid & OBD_MD_FLID);
1002         LASSERT(src_oa->o_valid & OBD_MD_FLGROUP);
1003
1004         if (set->set_oi->oi_md == NULL) {
1005                 __u16 stripes_def = lov_get_stripecnt(lov, LOV_MAGIC, 0);
1006
1007                 /* If the MDS file was truncated up to some size, stripe over
1008                  * enough OSTs to allow the file to be created at that size.
1009                  * This may mean we use more than the default # of stripes. */
1010                 if (src_oa->o_valid & OBD_MD_FLSIZE) {
1011                         obd_size min_bavail = LUSTRE_STRIPE_MAXBYTES;
1012
1013                         /* Find a small number of stripes we can use
1014                            (up to # of active osts). */
1015                         stripes = 1;
1016                         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1017                                 if (!lov->lov_tgts[i] ||
1018                                     !lov->lov_tgts[i]->ltd_active)
1019                                         continue;
1020                                 min_bavail = min(min_bavail, TGT_BAVAIL(i));
1021                                 if (min_bavail * stripes > src_oa->o_size)
1022                                         break;
1023                                 stripes++;
1024                         }
1025
1026                         if (stripes < stripes_def)
1027                                 stripes = stripes_def;
1028                 } else {
1029                         flag = LOV_USES_DEFAULT_STRIPE;
1030                         stripes = stripes_def;
1031                 }
1032
1033                 rc = lov_alloc_memmd(&set->set_oi->oi_md, stripes,
1034                                      lov->desc.ld_pattern ?
1035                                      lov->desc.ld_pattern : LOV_PATTERN_RAID0,
1036                                      LOV_MAGIC);
1037                 if (rc < 0)
1038                         GOTO(out_err, rc);
1039                 newea = 1;
1040                 rc = 0;
1041         }
1042
1043         lsm = set->set_oi->oi_md;
1044         lsm->lsm_object_id = src_oa->o_id;
1045         lsm->lsm_object_seq = src_oa->o_seq;
1046         lsm->lsm_layout_gen = 0; /* actual generation set in mdd_lov_create() */
1047
1048         if (!lsm->lsm_stripe_size)
1049                 lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size;
1050         if (!lsm->lsm_pattern) {
1051                 LASSERT(lov->desc.ld_pattern);
1052                 lsm->lsm_pattern = lov->desc.ld_pattern;
1053         }
1054
1055         stripes = alloc_idx_array(exp, lsm, newea, &idx_arr, &idx_cnt, flag);
1056         if (stripes <= 0)
1057                 GOTO(out_err, rc = stripes ? stripes : -EIO);
1058         LASSERTF(stripes <= lsm->lsm_stripe_count,"requested %d allocated %d\n",
1059                  lsm->lsm_stripe_count, stripes);
1060
1061         for (i = 0; i < stripes; i++) {
1062                 struct lov_request *req;
1063                 int ost_idx = idx_arr[i];
1064                 LASSERT(ost_idx >= 0);
1065
1066                 OBD_ALLOC(req, sizeof(*req));
1067                 if (req == NULL)
1068                         GOTO(out_err, rc = -ENOMEM);
1069                 lov_set_add_req(req, set);
1070
1071                 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
1072                 OBD_ALLOC_LARGE(req->rq_oi.oi_md, req->rq_buflen);
1073                 if (req->rq_oi.oi_md == NULL)
1074                         GOTO(out_err, rc = -ENOMEM);
1075
1076                 OBDO_ALLOC(req->rq_oi.oi_oa);
1077                 if (req->rq_oi.oi_oa == NULL)
1078                         GOTO(out_err, rc = -ENOMEM);
1079
1080                 req->rq_idx = ost_idx;
1081                 req->rq_stripe = i;
1082                 /* create data objects with "parent" OA */
1083                 memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
1084                 req->rq_oi.oi_cb_up = cb_create_update;
1085
1086                 /* XXX When we start creating objects on demand, we need to
1087                  *     make sure that we always create the object on the
1088                  *     stripe which holds the existing file size.
1089                  */
1090                 if (src_oa->o_valid & OBD_MD_FLSIZE) {
1091                         req->rq_oi.oi_oa->o_size =
1092                                 lov_size_to_stripe(lsm, src_oa->o_size, i);
1093
1094                         CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
1095                                i, req->rq_oi.oi_oa->o_size, src_oa->o_size);
1096                 }
1097         }
1098         LASSERT(set->set_count == stripes);
1099
1100         if (stripes < lsm->lsm_stripe_count)
1101                 qos_shrink_lsm(set);
1102         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_LOV_PREP_CREATE)) {
1103                 qos_shrink_lsm(set);
1104                 rc = -EIO;
1105         }
1106
1107         if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) {
1108                 oti_alloc_cookies(oti, set->set_count);
1109                 if (!oti->oti_logcookies)
1110                         GOTO(out_err, rc = -ENOMEM);
1111                 set->set_cookies = oti->oti_logcookies;
1112         }
1113 out_err:
1114         if (newea && rc)
1115                 obd_free_memmd(exp, &set->set_oi->oi_md);
1116         if (idx_arr)
1117                 free_idx_array(idx_arr, idx_cnt);
1118         EXIT;
1119         return rc;
1120 }
1121
1122 void qos_update(struct lov_obd *lov)
1123 {
1124         ENTRY;
1125         lov->lov_qos.lq_dirty = 1;
1126 }
1127
1128 void qos_statfs_done(struct lov_obd *lov)
1129 {
1130         cfs_down_write(&lov->lov_qos.lq_rw_sem);
1131         if (lov->lov_qos.lq_statfs_in_progress) {
1132                 lov->lov_qos.lq_statfs_in_progress = 0;
1133                 /* wake up any threads waiting for the statfs rpcs to complete*/
1134                 cfs_waitq_signal(&lov->lov_qos.lq_statfs_waitq);
1135         }
1136         cfs_up_write(&lov->lov_qos.lq_rw_sem);
1137 }
1138
1139 static int qos_statfs_ready(struct obd_device *obd, __u64 max_age)
1140 {
1141         struct lov_obd         *lov = &obd->u.lov;
1142         int rc;
1143         ENTRY;
1144         cfs_down_read(&lov->lov_qos.lq_rw_sem);
1145         rc = lov->lov_qos.lq_statfs_in_progress == 0 ||
1146              cfs_time_beforeq_64(max_age, obd->obd_osfs_age);
1147         cfs_up_read(&lov->lov_qos.lq_rw_sem);
1148         RETURN(rc);
1149 }
1150
1151 /*
1152  * Update statfs data if the current osfs age is older than max_age.
1153  * If wait is not set, it means that we are called from lov_create()
1154  * and we should just issue the rpcs without waiting for them to complete.
1155  * If wait is set, we are called from alloc_qos() and we just have
1156  * to wait for the request set to complete.
1157  */
1158 void qos_statfs_update(struct obd_device *obd, __u64 max_age, int wait)
1159 {
1160         struct lov_obd         *lov = &obd->u.lov;
1161         struct obd_info        *oinfo;
1162         int                     rc = 0;
1163         struct ptlrpc_request_set *set = NULL;
1164         ENTRY;
1165
1166         if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
1167                 /* statfs data are quite recent, don't need to refresh it */
1168                 RETURN_EXIT;
1169
1170         if (!wait && lov->lov_qos.lq_statfs_in_progress)
1171                 /* statfs already in progress */
1172                 RETURN_EXIT;
1173
1174         cfs_down_write(&lov->lov_qos.lq_rw_sem);
1175         if (lov->lov_qos.lq_statfs_in_progress) {
1176                 cfs_up_write(&lov->lov_qos.lq_rw_sem);
1177                 GOTO(out, rc = 0);
1178         }
1179         /* no statfs in flight, send rpcs */
1180         lov->lov_qos.lq_statfs_in_progress = 1;
1181         cfs_up_write(&lov->lov_qos.lq_rw_sem);
1182
1183         if (wait)
1184                 CDEBUG(D_QOS, "%s: did not manage to get fresh statfs data "
1185                        "in a timely manner (osfs age "LPU64", max age "LPU64")"
1186                        ", sending new statfs rpcs\n",
1187                        obd_uuid2str(&lov->desc.ld_uuid), obd->obd_osfs_age,
1188                        max_age);
1189
1190         /* need to send statfs rpcs */
1191         CDEBUG(D_QOS, "sending new statfs requests\n");
1192         memset(lov->lov_qos.lq_statfs_data, 0,
1193                sizeof(*lov->lov_qos.lq_statfs_data));
1194         oinfo = &lov->lov_qos.lq_statfs_data->lsd_oi;
1195         oinfo->oi_osfs = &lov->lov_qos.lq_statfs_data->lsd_statfs;
1196         oinfo->oi_flags = OBD_STATFS_NODELAY;
1197         set = ptlrpc_prep_set();
1198         if (!set)
1199                 GOTO(out_failed, rc = -ENOMEM);
1200
1201         rc = obd_statfs_async(obd->obd_self_export, oinfo, max_age, set);
1202         if (rc || cfs_list_empty(&set->set_requests)) {
1203                 if (rc)
1204                         CWARN("statfs failed with %d\n", rc);
1205                 GOTO(out_failed, rc);
1206         }
1207         /* send requests via ptlrpcd */
1208         oinfo->oi_flags |= OBD_STATFS_PTLRPCD;
1209         ptlrpcd_add_rqset(set);
1210         GOTO(out, rc);
1211
1212 out_failed:
1213         cfs_down_write(&lov->lov_qos.lq_rw_sem);
1214         lov->lov_qos.lq_statfs_in_progress = 0;
1215         /* wake up any threads waiting for the statfs rpcs to complete */
1216         cfs_waitq_signal(&lov->lov_qos.lq_statfs_waitq);
1217         cfs_up_write(&lov->lov_qos.lq_rw_sem);
1218         wait = 0;
1219 out:
1220         if (set)
1221                 ptlrpc_set_destroy(set);
1222         if (wait) {
1223                 struct l_wait_info lwi = { 0 };
1224                 CDEBUG(D_QOS, "waiting for statfs requests to complete\n");
1225                 l_wait_event(lov->lov_qos.lq_statfs_waitq,
1226                              qos_statfs_ready(obd, max_age), &lwi);
1227                 if (cfs_time_before_64(obd->obd_osfs_age, max_age))
1228                         CDEBUG(D_QOS, "%s: still no fresh statfs data after "
1229                                       "waiting (osfs age "LPU64", max age "
1230                                       LPU64")\n",
1231                                       obd_uuid2str(&lov->desc.ld_uuid),
1232                                       obd->obd_osfs_age, max_age);
1233         }
1234 }