lustre/lov/lov_qos.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  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 #ifndef EXPORT_SYMTAB
  38 # define EXPORT_SYMTAB
  39 #endif
  40 #define DEBUG_SUBSYSTEM S_LOV
  41
  42 #ifdef __KERNEL__
  43 #include <libcfs/libcfs.h>
  44 #else
  45 #include <liblustre.h>
  46 #endif
  47
  48 #include <obd_class.h>
  49 #include <obd_lov.h>
  50 #include <lustre/lustre_idl.h>
  51 #include "lov_internal.h"
  52
  53 /* #define QOS_DEBUG 1 */
  54 #define D_QOS D_OTHER
  55
  56 #define TGT_BAVAIL(i)  (lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bavail*\
  57                         lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bsize)
  58
  59 int qos_add_tgt(struct obd_device *obd, __u32 index)
  60 {
  61         struct lov_obd *lov = &obd->u.lov;
  62         struct lov_qos_oss *oss, *temposs;
  63         struct obd_export *exp = lov->lov_tgts[index]->ltd_exp;
  64         int rc = 0, found = 0;
  65         ENTRY;
  66
  67         /* We only need this QOS struct on MDT, not clients - but we may not
  68          * have registered the LOV's observer yet, so there's no way to know */
  69         if (!exp || !exp->exp_connection) {
  70                 CERROR("Missing connection\n");
  71                 RETURN(-ENOTCONN);
  72         }
  73
  74         down_write(&lov->lov_qos.lq_rw_sem);
  75         mutex_down(&lov->lov_lock);
  76         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
  77                 if (obd_uuid_equals(&oss->lqo_uuid,
  78                                     &exp->exp_connection->c_remote_uuid)) {
  79                         found++;
  80                         break;
  81                 }
  82         }
  83
  84         if (!found) {
  85                 OBD_ALLOC_PTR(oss);
  86                 if (!oss)
  87                         GOTO(out, rc = -ENOMEM);
  88                 memcpy(&oss->lqo_uuid,
  89                        &exp->exp_connection->c_remote_uuid,
  90                        sizeof(oss->lqo_uuid));
  91         } else {
  92                 /* Assume we have to move this one */
  93                 list_del(&oss->lqo_oss_list);
  94         }
  95
  96         oss->lqo_ost_count++;
  97         lov->lov_tgts[index]->ltd_qos.ltq_oss = oss;
  98
  99         /* Add sorted by # of OSTs.  Find the first entry that we're
 100            bigger than... */
 101         list_for_each_entry(temposs, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 102                 if (oss->lqo_ost_count > temposs->lqo_ost_count)
 103                         break;
 104         }
 105         /* ...and add before it.  If we're the first or smallest, temposs
 106            points to the list head, and we add to the end. */
 107         list_add_tail(&oss->lqo_oss_list, &temposs->lqo_oss_list);
 108
 109         lov->lov_qos.lq_dirty = 1;
 110         lov->lov_qos.lq_rr.lqr_dirty = 1;
 111
 112         CDEBUG(D_QOS, "add tgt %s to OSS %s (%d OSTs)\n",
 113                obd_uuid2str(&lov->lov_tgts[index]->ltd_uuid),
 114                obd_uuid2str(&oss->lqo_uuid),
 115                oss->lqo_ost_count);
 116
 117 out:
 118         mutex_up(&lov->lov_lock);
 119         up_write(&lov->lov_qos.lq_rw_sem);
 120         RETURN(rc);
 121 }
 122
 123 int qos_del_tgt(struct obd_device *obd, struct lov_tgt_desc *tgt)
 124 {
 125         struct lov_obd *lov = &obd->u.lov;
 126         struct lov_qos_oss *oss;
 127         int rc = 0;
 128         ENTRY;
 129
 130         down_write(&lov->lov_qos.lq_rw_sem);
 131
 132         oss = tgt->ltd_qos.ltq_oss;
 133         if (!oss)
 134                 GOTO(out, rc = -ENOENT);
 135
 136         oss->lqo_ost_count--;
 137         if (oss->lqo_ost_count == 0) {
 138                 CDEBUG(D_QOS, "removing OSS %s\n",
 139                        obd_uuid2str(&oss->lqo_uuid));
 140                 list_del(&oss->lqo_oss_list);
 141                 OBD_FREE_PTR(oss);
 142         }
 143
 144         lov->lov_qos.lq_dirty = 1;
 145         lov->lov_qos.lq_rr.lqr_dirty = 1;
 146 out:
 147         up_write(&lov->lov_qos.lq_rw_sem);
 148         RETURN(rc);
 149 }
 150
 151 /* Recalculate per-object penalties for OSSs and OSTs,
 152    depends on size of each ost in an oss */
 153 static int qos_calc_ppo(struct obd_device *obd)
 154 {
 155         struct lov_obd *lov = &obd->u.lov;
 156         struct lov_qos_oss *oss;
 157         __u64 ba_max, ba_min, temp;
 158         __u32 num_active;
 159         int rc, i, prio_wide;
 160         time_t now, age;
 161         ENTRY;
 162
 163         if (!lov->lov_qos.lq_dirty)
 164                 GOTO(out, rc = 0);
 165
 166         num_active = lov->desc.ld_active_tgt_count - 1;
 167         if (num_active < 1)
 168                 GOTO(out, rc = -EAGAIN);
 169
 170         /* find bavail on each OSS */
 171         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 172                 oss->lqo_bavail = 0;
 173         }
 174         lov->lov_qos.lq_active_oss_count = 0;
 175
 176         /* How badly user wants to select osts "widely" (not recently chosen
 177            and not on recent oss's).  As opposed to "freely" (free space
 178            avail.) 0-256. */
 179         prio_wide = 256 - lov->lov_qos.lq_prio_free;
 180
 181         ba_min = (__u64)(-1);
 182         ba_max = 0;
 183         now = cfs_time_current_sec();
 184         /* Calculate OST penalty per object */
 185         /* (lov ref taken in alloc_qos) */
 186         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 187                 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
 188                         continue;
 189                 temp = TGT_BAVAIL(i);
 190                 if (!temp)
 191                         continue;
 192                 ba_min = min(temp, ba_min);
 193                 ba_max = max(temp, ba_max);
 194
 195                 /* Count the number of usable OSS's */
 196                 if (lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_bavail == 0)
 197                         lov->lov_qos.lq_active_oss_count++;
 198                 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_bavail += temp;
 199
 200                 /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
 201                 temp >>= 1;
 202                 do_div(temp, num_active);
 203                 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj =
 204                         (temp * prio_wide) >> 8;
 205
 206                 age = (now - lov->lov_tgts[i]->ltd_qos.ltq_used) >> 3;
 207                 if (lov->lov_qos.lq_reset || age > 32 * lov->desc.ld_qos_maxage)
 208                         lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
 209                 else if (age > lov->desc.ld_qos_maxage)
 210                         /* Decay the penalty by half for every 8x the update
 211                          * interval that the device has been idle.  That gives
 212                          * lots of time for the statfs information to be
 213                          * updated (which the penalty is only a proxy for),
 214                          * and avoids penalizing OSS/OSTs under light load. */
 215                         lov->lov_tgts[i]->ltd_qos.ltq_penalty >>=
 216                                 (age / lov->desc.ld_qos_maxage);
 217         }
 218
 219         num_active = lov->lov_qos.lq_active_oss_count - 1;
 220         if (num_active < 1) {
 221                 /* If there's only 1 OSS, we can't penalize it, so instead
 222                    we have to double the OST penalty */
 223                 num_active = 1;
 224                 for (i = 0; i < lov->desc.ld_tgt_count; i++)
 225                         if (lov->lov_tgts[i])
 226                             lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj <<= 1;
 227         }
 228
 229         /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
 230         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 231                 temp = oss->lqo_bavail >> 1;
 232                 do_div(temp, oss->lqo_ost_count * num_active);
 233                 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
 234
 235                 age = (now - oss->lqo_used) >> 3;
 236                 if (lov->lov_qos.lq_reset || age > 32 * lov->desc.ld_qos_maxage)
 237                         oss->lqo_penalty = 0;
 238                 else if (age > lov->desc.ld_qos_maxage)
 239                         /* Decay the penalty by half for every 8x the update
 240                          * interval that the device has been idle.  That gives
 241                          * lots of time for the statfs information to be
 242                          * updated (which the penalty is only a proxy for),
 243                          * and avoids penalizing OSS/OSTs under light load. */
 244                         oss->lqo_penalty >>= (age / lov->desc.ld_qos_maxage);
 245         }
 246
 247         lov->lov_qos.lq_dirty = 0;
 248         lov->lov_qos.lq_reset = 0;
 249
 250         /* If each ost has almost same free space,
 251          * do rr allocation for better creation performance */
 252         lov->lov_qos.lq_same_space = 0;
 253         if ((ba_max * (256 - lov->lov_qos.lq_threshold_rr)) >> 8 < ba_min) {
 254                 lov->lov_qos.lq_same_space = 1;
 255                 /* Reset weights for the next time we enter qos mode */
 256                 lov->lov_qos.lq_reset = 1;
 257         }
 258         rc = 0;
 259
 260 out:
 261         if (!rc && lov->lov_qos.lq_same_space)
 262                 RETURN(-EAGAIN);
 263         RETURN(rc);
 264 }
 265
 266 static int qos_calc_weight(struct lov_obd *lov, int i)
 267 {
 268         __u64 temp, temp2;
 269
 270         /* Final ost weight = TGT_BAVAIL - ost_penalty - oss_penalty */
 271         temp = TGT_BAVAIL(i);
 272         temp2 = lov->lov_tgts[i]->ltd_qos.ltq_penalty +
 273                 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty;
 274         if (temp < temp2)
 275                 lov->lov_tgts[i]->ltd_qos.ltq_weight = 0;
 276         else
 277                 lov->lov_tgts[i]->ltd_qos.ltq_weight = temp - temp2;
 278         return 0;
 279 }
 280
 281 /* We just used this index for a stripe; adjust everyone's weights */
 282 static int qos_used(struct lov_obd *lov, struct ost_pool *osts,
 283                     __u32 index, __u64 *total_wt)
 284 {
 285         struct lov_qos_oss *oss;
 286         int j;
 287         ENTRY;
 288
 289         /* Don't allocate from this stripe anymore, until the next alloc_qos */
 290         lov->lov_tgts[index]->ltd_qos.ltq_usable = 0;
 291
 292         oss = lov->lov_tgts[index]->ltd_qos.ltq_oss;
 293
 294         /* Decay old penalty by half (we're adding max penalty, and don't
 295            want it to run away.) */
 296         lov->lov_tgts[index]->ltd_qos.ltq_penalty >>= 1;
 297         oss->lqo_penalty >>= 1;
 298
 299         /* mark the OSS and OST as recently used */
 300         lov->lov_tgts[index]->ltd_qos.ltq_used =
 301                 oss->lqo_used = cfs_time_current_sec();
 302
 303         /* Set max penalties for this OST and OSS */
 304         lov->lov_tgts[index]->ltd_qos.ltq_penalty +=
 305                 lov->lov_tgts[index]->ltd_qos.ltq_penalty_per_obj *
 306                 lov->desc.ld_active_tgt_count;
 307         oss->lqo_penalty += oss->lqo_penalty_per_obj *
 308                 lov->lov_qos.lq_active_oss_count;
 309
 310         /* Decrease all OSS penalties */
 311         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 312                 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
 313                         oss->lqo_penalty = 0;
 314                 else
 315                         oss->lqo_penalty -= oss->lqo_penalty_per_obj;
 316         }
 317
 318         *total_wt = 0;
 319         /* Decrease all OST penalties */
 320         for (j = 0; j < osts->op_count; j++) {
 321                 int i;
 322
 323                 i = osts->op_array[j];
 324                 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
 325                         continue;
 326                 if (lov->lov_tgts[i]->ltd_qos.ltq_penalty <
 327                     lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj)
 328                         lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
 329                 else
 330                         lov->lov_tgts[i]->ltd_qos.ltq_penalty -=
 331                         lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj;
 332
 333                 qos_calc_weight(lov, i);
 334
 335                 /* Recalc the total weight of usable osts */
 336                 if (lov->lov_tgts[i]->ltd_qos.ltq_usable)
 337                         *total_wt += lov->lov_tgts[i]->ltd_qos.ltq_weight;
 338
 339 #ifdef QOS_DEBUG
 340                 CDEBUG(D_QOS, "recalc tgt %d usable=%d avail="LPU64
 341                        " ostppo="LPU64" ostp="LPU64" ossppo="LPU64
 342                        " ossp="LPU64" wt="LPU64"\n",
 343                        i, lov->lov_tgts[i]->ltd_qos.ltq_usable,
 344                        TGT_BAVAIL(i) >> 10,
 345                        lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj >> 10,
 346                        lov->lov_tgts[i]->ltd_qos.ltq_penalty >> 10,
 347                        lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty_per_obj>>10,
 348                        lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty >> 10,
 349                        lov->lov_tgts[i]->ltd_qos.ltq_weight >> 10);
 350 #endif
 351         }
 352
 353         RETURN(0);
 354 }
 355
 356 #define LOV_QOS_EMPTY ((__u32)-1)
 357 /* compute optimal round-robin order, based on OSTs per OSS
 358  */
 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         down_write(&lov->lov_qos.lq_rw_sem);
 374
 375         real_count = src_pool->op_count;
 376
 377         /* Zero the pool array */
 378         /* alloc_rr is holding a read lock on the pool, so nobody is adding/
 379            deleting from the pool. The lq_rw_sem insures that nobody else
 380            is reading. */
 381         lqr->lqr_pool.op_count = real_count;
 382         rc = lov_ost_pool_extend(&lqr->lqr_pool, real_count);
 383         if (rc) {
 384                 up_write(&lov->lov_qos.lq_rw_sem);
 385                 RETURN(rc);
 386         }
 387         for (i = 0; i < lqr->lqr_pool.op_count; i++)
 388                 lqr->lqr_pool.op_array[i] = LOV_QOS_EMPTY;
 389
 390         /* Place all the OSTs from 1 OSS at the same time. */
 391         placed = 0;
 392         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 393                 int j = 0;
 394                 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
 395                         if (lov->lov_tgts[src_pool->op_array[i]] &&
 396                             (lov->lov_tgts[src_pool->op_array[i]]->ltd_qos.ltq_oss == oss)) {
 397                               /* Evenly space these OSTs across arrayspace */
 398                               int next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count;
 399                               while (lqr->lqr_pool.op_array[next] !=
 400                                      LOV_QOS_EMPTY)
 401                                       next = (next + 1) % lqr->lqr_pool.op_count;
 402                               lqr->lqr_pool.op_array[next] = src_pool->op_array[i];
 403                               j++;
 404                               placed++;
 405                         }
 406                 }
 407         }
 408
 409         lqr->lqr_dirty = 0;
 410         up_write(&lov->lov_qos.lq_rw_sem);
 411
 412         if (placed != real_count) {
 413                 /* This should never happen */
 414                 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
 415                                    "round-robin list (%d of %d).\n",
 416                                    placed, real_count);
 417                 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
 418                         LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
 419                                  lqr->lqr_pool.op_array[i]);
 420                 }
 421                 lqr->lqr_dirty = 1;
 422                 RETURN(-EAGAIN);
 423         }
 424
 425 #ifdef QOS_DEBUG
 426         for (i = 0; i < lqr->lqr_pool.op_count; i++) {
 427                 LCONSOLE(D_QOS, "rr #%d ost idx=%d\n", i,
 428                          lqr->lqr_pool.op_array[i]);
 429         }
 430 #endif
 431
 432         RETURN(0);
 433 }
 434
 435
 436 void qos_shrink_lsm(struct lov_request_set *set)
 437 {
 438         struct lov_stripe_md *lsm = set->set_oi->oi_md, *lsm_new;
 439         /* XXX LOV STACKING call into osc for sizes */
 440         unsigned oldsize, newsize;
 441
 442         if (set->set_oti && set->set_cookies && set->set_cookie_sent) {
 443                 struct llog_cookie *cookies;
 444                 oldsize = lsm->lsm_stripe_count * sizeof(*cookies);
 445                 newsize = set->set_count * sizeof(*cookies);
 446
 447                 cookies = set->set_cookies;
 448                 oti_alloc_cookies(set->set_oti, set->set_count);
 449                 if (set->set_oti->oti_logcookies) {
 450                         memcpy(set->set_oti->oti_logcookies, cookies, newsize);
 451                         OBD_FREE(cookies, oldsize);
 452                         set->set_cookies = set->set_oti->oti_logcookies;
 453                 } else {
 454                         CWARN("'leaking' %d bytes\n", oldsize - newsize);
 455                 }
 456         }
 457
 458         CWARN("using fewer stripes for object "LPU64": old %u new %u\n",
 459               lsm->lsm_object_id, lsm->lsm_stripe_count, set->set_count);
 460         LASSERT(lsm->lsm_stripe_count >= set->set_count);
 461
 462         newsize = lov_stripe_md_size(set->set_count);
 463         OBD_ALLOC(lsm_new, newsize);
 464         if (lsm_new != NULL) {
 465                 int i;
 466                 memcpy(lsm_new, lsm, sizeof(*lsm));
 467                 for (i = 0; i < lsm->lsm_stripe_count; i++) {
 468                         if (i < set->set_count) {
 469                                 lsm_new->lsm_oinfo[i] = lsm->lsm_oinfo[i];
 470                                 continue;
 471                         }
 472                         OBD_SLAB_FREE(lsm->lsm_oinfo[i], lov_oinfo_slab,
 473                                       sizeof(struct lov_oinfo));
 474                 }
 475                 lsm_new->lsm_stripe_count = set->set_count;
 476                 OBD_FREE(lsm, sizeof(struct lov_stripe_md) +
 477                          lsm->lsm_stripe_count * sizeof(struct lov_oinfo *));
 478                 set->set_oi->oi_md = lsm_new;
 479         } else {
 480                 CWARN("'leaking' few bytes\n");
 481         }
 482 }
 483
 484 int qos_remedy_create(struct lov_request_set *set, struct lov_request *req)
 485 {
 486         struct lov_stripe_md *lsm = set->set_oi->oi_md;
 487         struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
 488         unsigned ost_idx, ost_count = lov->desc.ld_tgt_count;
 489         int stripe, i, rc = -EIO;
 490         ENTRY;
 491
 492         ost_idx = (req->rq_idx + lsm->lsm_stripe_count) % ost_count;
 493         for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
 494                 if (!lov->lov_tgts[ost_idx] ||
 495                     !lov->lov_tgts[ost_idx]->ltd_active)
 496                         continue;
 497                 /* check if objects has been created on this ost */
 498                 for (stripe = 0; stripe < lsm->lsm_stripe_count; stripe++) {
 499                         /* we try send create to this ost but he is failed */
 500                         if (stripe == req->rq_stripe)
 501                                 continue;
 502                         /* already have object at this stripe */
 503                         if (ost_idx == lsm->lsm_oinfo[stripe]->loi_ost_idx)
 504                                 break;
 505                 }
 506                 if (stripe >= lsm->lsm_stripe_count) {
 507                         req->rq_idx = ost_idx;
 508                         rc = obd_create(lov->lov_tgts[ost_idx]->ltd_exp,
 509                                         req->rq_oi.oi_oa, &req->rq_oi.oi_md,
 510                                         set->set_oti);
 511                         if (!rc)
 512                                 break;
 513                 }
 514         }
 515         RETURN(rc);
 516 }
 517
 518 static int min_stripe_count(int stripe_cnt, int flags)
 519 {
 520         return (flags & LOV_USES_DEFAULT_STRIPE ?
 521                 stripe_cnt - (stripe_cnt / 4) : stripe_cnt);
 522 }
 523
 524 #define LOV_CREATE_RESEED_MULT 4
 525 #define LOV_CREATE_RESEED_MIN  1000
 526 /* Allocate objects on osts with round-robin algorithm */
 527 static int alloc_rr(struct lov_obd *lov, int *idx_arr, int *stripe_cnt,
 528                     char *poolname, int flags)
 529 {
 530         unsigned array_idx;
 531         int i, rc, *idx_pos;
 532         __u32 ost_idx;
 533         int ost_start_idx_temp;
 534         int speed = 0;
 535         int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
 536         struct pool_desc *pool;
 537         struct ost_pool *osts;
 538         struct lov_qos_rr *lqr;
 539         ENTRY;
 540
 541         pool = lov_find_pool(lov, poolname);
 542         if (pool == NULL) {
 543                 osts = &(lov->lov_packed);
 544                 lqr = &(lov->lov_qos.lq_rr);
 545         } else {
 546                 down_read(&pool_tgt_rw_sem(pool));
 547                 osts = &(pool->pool_obds);
 548                 lqr = &(pool->pool_rr);
 549         }
 550
 551         rc = qos_calc_rr(lov, osts, lqr);
 552         if (rc)
 553                 GOTO(out, rc);
 554
 555         if (--lqr->lqr_start_count <= 0) {
 556                 lqr->lqr_start_idx = ll_rand() % osts->op_count;
 557                 lqr->lqr_start_count =
 558                         (LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) +
 559                          LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U);
 560         } else if (stripe_cnt_min >= osts->op_count ||
 561                    lqr->lqr_start_idx > osts->op_count) {
 562                 /* If we have allocated from all of the OSTs, slowly
 563                  * precess the next start if the OST/stripe count isn't
 564                  * already doing this for us. */
 565                 lqr->lqr_start_idx %= osts->op_count;
 566                 if (*stripe_cnt > 1 && (osts->op_count % (*stripe_cnt)) != 1)
 567                         ++lqr->lqr_offset_idx;
 568         }
 569         down_read(&lov->lov_qos.lq_rw_sem);
 570         ost_start_idx_temp = lqr->lqr_start_idx;
 571
 572 repeat_find:
 573         array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) % osts->op_count;
 574         idx_pos = idx_arr;
 575 #ifdef QOS_DEBUG
 576         CDEBUG(D_QOS, "pool '%s' want %d startidx %d startcnt %d offset %d "
 577                "active %d count %d arrayidx %d\n", poolname,
 578                *stripe_cnt, lqr->lqr_start_idx, lqr->lqr_start_count,
 579                lqr->lqr_offset_idx, osts->op_count, osts->op_count, array_idx);
 580 #endif
 581
 582         for (i = 0; i < osts->op_count;
 583                     i++, array_idx=(array_idx + 1) % osts->op_count) {
 584                 ++lqr->lqr_start_idx;
 585                 ost_idx = lqr->lqr_pool.op_array[array_idx];
 586 #ifdef QOS_DEBUG
 587                 CDEBUG(D_QOS, "#%d strt %d act %d strp %d ary %d idx %d\n",
 588                        i, lqr->lqr_start_idx,
 589                        ((ost_idx != LOV_QOS_EMPTY) && lov->lov_tgts[ost_idx]) ?
 590                        lov->lov_tgts[ost_idx]->ltd_active : 0,
 591                        idx_pos - idx_arr, array_idx, ost_idx);
 592 #endif
 593                 if ((ost_idx == LOV_QOS_EMPTY) || !lov->lov_tgts[ost_idx] ||
 594                     !lov->lov_tgts[ost_idx]->ltd_active)
 595                         continue;
 596
 597                 /* Fail Check before osc_precreate() is called
 598                    so we can only 'fail' single OSC. */
 599                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
 600                         continue;
 601
 602                 /* Drop slow OSCs if we can */
 603                 if (obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed)
 604                         continue;
 605
 606                 *idx_pos = ost_idx;
 607                 idx_pos++;
 608                 /* We have enough stripes */
 609                 if (idx_pos - idx_arr == *stripe_cnt)
 610                         break;
 611         }
 612         if ((speed < 2) && (idx_pos - idx_arr < stripe_cnt_min)) {
 613                 /* Try again, allowing slower OSCs */
 614                 speed++;
 615                 lqr->lqr_start_idx = ost_start_idx_temp;
 616                 goto repeat_find;
 617         }
 618
 619         up_read(&lov->lov_qos.lq_rw_sem);
 620
 621         *stripe_cnt = idx_pos - idx_arr;
 622 out:
 623         if (pool != NULL) {
 624                 up_read(&pool_tgt_rw_sem(pool));
 625                 /* put back ref got by lov_find_pool() */
 626                 lov_pool_putref(pool);
 627         }
 628
 629         RETURN(rc);
 630 }
 631
 632 /* alloc objects on osts with specific stripe offset */
 633 static int alloc_specific(struct lov_obd *lov, struct lov_stripe_md *lsm,
 634                           int *idx_arr)
 635 {
 636         unsigned ost_idx, array_idx, ost_count;
 637         int i, rc, *idx_pos;
 638         int speed = 0;
 639         struct pool_desc *pool;
 640         struct ost_pool *osts;
 641         ENTRY;
 642
 643         pool = lov_find_pool(lov, lsm->lsm_pool_name);
 644         if (pool == NULL) {
 645                 osts = &(lov->lov_packed);
 646         } else {
 647                 down_read(&pool_tgt_rw_sem(pool));
 648                 osts = &(pool->pool_obds);
 649         }
 650
 651         ost_count = osts->op_count;
 652
 653 repeat_find:
 654         /* search loi_ost_idx in ost array */
 655         array_idx = 0;
 656         for (i = 0; i < ost_count; i++) {
 657                 if (osts->op_array[i] == lsm->lsm_oinfo[0]->loi_ost_idx) {
 658                         array_idx = i;
 659                         break;
 660                 }
 661         }
 662         if (i == ost_count) {
 663                 CERROR("Start index %d not found in pool '%s'\n",
 664                        lsm->lsm_oinfo[0]->loi_ost_idx, lsm->lsm_pool_name);
 665                 GOTO(out, rc = -EINVAL);
 666         }
 667
 668         idx_pos = idx_arr;
 669         for (i = 0; i < ost_count;
 670              i++, array_idx = (array_idx + 1) % ost_count) {
 671                 ost_idx = osts->op_array[array_idx];
 672
 673                 if (!lov->lov_tgts[ost_idx] ||
 674                     !lov->lov_tgts[ost_idx]->ltd_active) {
 675                         continue;
 676                 }
 677
 678                 /* Fail Check before osc_precreate() is called
 679                    so we can only 'fail' single OSC. */
 680                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
 681                         continue;
 682
 683                 /* Drop slow OSCs if we can, but not for requested start idx.
 684                  *
 685                  * This means "if OSC is slow and it is not the requested
 686                  * start OST, then it can be skipped, otherwise skip it only
 687                  * if it is inactive/recovering/out-of-space." */
 688                 if ((obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed) &&
 689                     (i != 0 || speed >= 2))
 690                         continue;
 691
 692                 *idx_pos = ost_idx;
 693                 idx_pos++;
 694                 /* We have enough stripes */
 695                 if (idx_pos - idx_arr == lsm->lsm_stripe_count)
 696                         GOTO(out, rc = 0);
 697         }
 698         if (speed < 2) {
 699                 /* Try again, allowing slower OSCs */
 700                 speed++;
 701                 goto repeat_find;
 702         }
 703
 704         /* If we were passed specific striping params, then a failure to
 705          * meet those requirements is an error, since we can't reallocate
 706          * that memory (it might be part of a larger array or something).
 707          *
 708          * We can only get here if lsm_stripe_count was originally > 1.
 709          */
 710         CERROR("can't lstripe objid "LPX64": have %d want %u\n",
 711                lsm->lsm_object_id, (int)(idx_pos - idx_arr),
 712                lsm->lsm_stripe_count);
 713         rc = -EFBIG;
 714 out:
 715         if (pool != NULL) {
 716                 up_read(&pool_tgt_rw_sem(pool));
 717                 /* put back ref got by lov_find_pool() */
 718                 lov_pool_putref(pool);
 719         }
 720         RETURN(rc);
 721 }
 722
 723 /* Alloc objects on osts with optimization based on:
 724    - free space
 725    - network resources (shared OSS's)
 726 */
 727 static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt,
 728                      char *poolname, int flags)
 729 {
 730         struct lov_obd *lov = &exp->exp_obd->u.lov;
 731         __u64 total_weight = 0;
 732         int nfound, good_osts, i, rc = 0;
 733         int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
 734         struct pool_desc *pool;
 735         struct ost_pool *osts;
 736         struct lov_qos_rr *lqr;
 737         ENTRY;
 738
 739         if (stripe_cnt_min < 1)
 740                 RETURN(-EINVAL);
 741
 742         pool = lov_find_pool(lov, poolname);
 743         if (pool == NULL) {
 744                 osts = &(lov->lov_packed);
 745                 lqr = &(lov->lov_qos.lq_rr);
 746         } else {
 747                 down_read(&pool_tgt_rw_sem(pool));
 748                 osts = &(pool->pool_obds);
 749                 lqr = &(pool->pool_rr);
 750         }
 751
 752         obd_getref(exp->exp_obd);
 753         /* wait for fresh statfs info if needed, the rpcs are sent in
 754          * lov_create() */
 755         qos_statfs_update(exp->exp_obd,
 756                           cfs_time_shift_64(-2 * lov->desc.ld_qos_maxage), 1);
 757
 758         down_write(&lov->lov_qos.lq_rw_sem);
 759
 760         if (lov->desc.ld_active_tgt_count < 2)
 761                 GOTO(out, rc = -EAGAIN);
 762
 763         rc = qos_calc_ppo(exp->exp_obd);
 764         if (rc)
 765                 GOTO(out, rc);
 766
 767         good_osts = 0;
 768         /* Find all the OSTs that are valid stripe candidates */
 769         for (i = 0; i < osts->op_count; i++) {
 770                 if (!lov->lov_tgts[osts->op_array[i]] ||
 771                     !lov->lov_tgts[osts->op_array[i]]->ltd_active)
 772                         continue;
 773
 774                 /* Fail Check before osc_precreate() is called
 775                    so we can only 'fail' single OSC. */
 776                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && osts->op_array[i] == 0)
 777                         continue;
 778
 779                 if (obd_precreate(lov->lov_tgts[osts->op_array[i]]->ltd_exp) > 2)
 780                         continue;
 781
 782                 lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable = 1;
 783                 qos_calc_weight(lov, osts->op_array[i]);
 784                 total_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight;
 785
 786                 good_osts++;
 787         }
 788
 789 #ifdef QOS_DEBUG
 790         CDEBUG(D_QOS, "found %d good osts\n", good_osts);
 791 #endif
 792
 793         if (good_osts < stripe_cnt_min)
 794                 GOTO(out, rc = -EAGAIN);
 795
 796         /* We have enough osts */
 797         if (good_osts < *stripe_cnt)
 798                 *stripe_cnt = good_osts;
 799
 800         /* Find enough OSTs with weighted random allocation. */
 801         nfound = 0;
 802         while (nfound < *stripe_cnt) {
 803                 __u64 rand, cur_weight;
 804
 805                 cur_weight = 0;
 806                 rc = -ENODEV;
 807
 808                 if (total_weight) {
 809 #if BITS_PER_LONG == 32
 810                         rand = ll_rand() % (unsigned)total_weight;
 811                         /* If total_weight > 32-bit, first generate the high
 812                          * 32 bits of the random number, then add in the low
 813                          * 32 bits (truncated to the upper limit, if needed) */
 814                         if (total_weight > 0xffffffffULL)
 815                                 rand = (__u64)(ll_rand() %
 816                                           (unsigned)(total_weight >> 32)) << 32;
 817                         else
 818                                 rand = 0;
 819
 820                         if (rand == (total_weight & 0xffffffff00000000ULL))
 821                                 rand |= ll_rand() % (unsigned)total_weight;
 822                         else
 823                                 rand |= ll_rand();
 824
 825 #else
 826                         rand = ((__u64)ll_rand() << 32 | ll_rand()) %
 827                                 total_weight;
 828 #endif
 829                 } else {
 830                         rand = 0;
 831                 }
 832
 833                 /* On average, this will hit larger-weighted osts more often.
 834                    0-weight osts will always get used last (only when rand=0).*/
 835                 for (i = 0; i < osts->op_count; i++) {
 836                         if (!lov->lov_tgts[osts->op_array[i]] ||
 837                             !lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable)
 838                                 continue;
 839
 840                         cur_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight;
 841 #ifdef QOS_DEBUG
 842                         CDEBUG(D_QOS, "stripe_cnt=%d nfound=%d cur_weight="LPU64
 843                                       " rand="LPU64" total_weight="LPU64"\n",
 844                                *stripe_cnt, nfound, cur_weight, rand, total_weight);
 845 #endif
 846                         if (cur_weight >= rand) {
 847 #ifdef QOS_DEBUG
 848                                 CDEBUG(D_QOS, "assigned stripe=%d to idx=%d\n",
 849                                        nfound, osts->op_array[i]);
 850 #endif
 851                                 idx_arr[nfound++] = osts->op_array[i];
 852                                 qos_used(lov, osts, osts->op_array[i], &total_weight);
 853                                 rc = 0;
 854                                 break;
 855                         }
 856                 }
 857                 if (rc) {
 858                         CDEBUG(D_QOS, "Didn't find any OSTs? Reduce total weight\n");
 859                         if (total_weight == 0)
 860                                 break;
 861                         else
 862                                 total_weight = 0;
 863                 }
 864         }
 865
 866         LASSERT(nfound == *stripe_cnt);
 867
 868 out:
 869         up_write(&lov->lov_qos.lq_rw_sem);
 870
 871         if (pool != NULL) {
 872                 up_read(&pool_tgt_rw_sem(pool));
 873                 /* put back ref got by lov_find_pool() */
 874                 lov_pool_putref(pool);
 875         }
 876
 877         if (rc == -EAGAIN)
 878                 rc = alloc_rr(lov, idx_arr, stripe_cnt, poolname, flags);
 879
 880         obd_putref(exp->exp_obd);
 881         RETURN(rc);
 882 }
 883
 884 /* return new alloced stripe count on success */
 885 static int alloc_idx_array(struct obd_export *exp, struct lov_stripe_md *lsm,
 886                            int newea, int **idx_arr, int *arr_cnt, int flags)
 887 {
 888         struct lov_obd *lov = &exp->exp_obd->u.lov;
 889         int stripe_cnt = lsm->lsm_stripe_count;
 890         int i, rc = 0;
 891         int *tmp_arr = NULL;
 892         ENTRY;
 893
 894         *arr_cnt = stripe_cnt;
 895         OBD_ALLOC(tmp_arr, *arr_cnt * sizeof(int));
 896         if (tmp_arr == NULL)
 897                 RETURN(-ENOMEM);
 898         for (i = 0; i < *arr_cnt; i++)
 899                 tmp_arr[i] = -1;
 900
 901         if (newea ||
 902             lsm->lsm_oinfo[0]->loi_ost_idx >= lov->desc.ld_tgt_count)
 903                 rc = alloc_qos(exp, tmp_arr, &stripe_cnt,
 904                                lsm->lsm_pool_name, flags);
 905         else
 906                 rc = alloc_specific(lov, lsm, tmp_arr);
 907
 908         if (rc)
 909                 GOTO(out_arr, rc);
 910
 911         *idx_arr = tmp_arr;
 912         RETURN(stripe_cnt);
 913 out_arr:
 914         OBD_FREE(tmp_arr, *arr_cnt * sizeof(int));
 915         *arr_cnt = 0;
 916         RETURN(rc);
 917 }
 918
 919 static void free_idx_array(int *idx_arr, int arr_cnt)
 920 {
 921         if (arr_cnt)
 922                 OBD_FREE(idx_arr, arr_cnt * sizeof(int));
 923 }
 924
 925 int qos_prep_create(struct obd_export *exp, struct lov_request_set *set)
 926 {
 927         struct lov_obd *lov = &exp->exp_obd->u.lov;
 928         struct lov_stripe_md *lsm;
 929         struct obdo *src_oa = set->set_oi->oi_oa;
 930         struct obd_trans_info *oti = set->set_oti;
 931         int i, stripes, rc = 0, newea = 0;
 932         int flag = LOV_USES_ASSIGNED_STRIPE;
 933         int *idx_arr = NULL, idx_cnt = 0;
 934         ENTRY;
 935
 936         LASSERT(src_oa->o_valid & OBD_MD_FLID);
 937
 938         if (set->set_oi->oi_md == NULL) {
 939                 int stripes_def = lov_get_stripecnt(lov, 0);
 940
 941                 /* If the MDS file was truncated up to some size, stripe over
 942                  * enough OSTs to allow the file to be created at that size.
 943                  * This may mean we use more than the default # of stripes. */
 944                 if (src_oa->o_valid & OBD_MD_FLSIZE) {
 945                         obd_size min_bavail = LUSTRE_STRIPE_MAXBYTES;
 946
 947                         /* Find a small number of stripes we can use
 948                            (up to # of active osts). */
 949                         stripes = 1;
 950                         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 951                                 if (!lov->lov_tgts[i] ||
 952                                     !lov->lov_tgts[i]->ltd_active)
 953                                         continue;
 954                                 min_bavail = min(min_bavail, TGT_BAVAIL(i));
 955                                 if (min_bavail * stripes > src_oa->o_size)
 956                                         break;
 957                                 stripes++;
 958                         }
 959
 960                         if (stripes < stripes_def)
 961                                 stripes = stripes_def;
 962                 } else {
 963                          flag = LOV_USES_DEFAULT_STRIPE;
 964                          stripes = stripes_def;
 965                 }
 966
 967                 rc = lov_alloc_memmd(&set->set_oi->oi_md, stripes,
 968                                      lov->desc.ld_pattern ?
 969                                      lov->desc.ld_pattern : LOV_PATTERN_RAID0,
 970                                      LOV_MAGIC);
 971                 if (rc < 0)
 972                         GOTO(out_err, rc);
 973                 newea = 1;
 974                 rc = 0;
 975         }
 976
 977         lsm = set->set_oi->oi_md;
 978         lsm->lsm_object_id = src_oa->o_id;
 979         if (!lsm->lsm_stripe_size)
 980                 lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size;
 981         if (!lsm->lsm_pattern) {
 982                 LASSERT(lov->desc.ld_pattern);
 983                 lsm->lsm_pattern = lov->desc.ld_pattern;
 984         }
 985
 986         stripes = alloc_idx_array(exp, lsm, newea, &idx_arr, &idx_cnt, flag);
 987         if (stripes <= 0)
 988                 GOTO(out_err, rc = stripes ? stripes : -EIO);
 989         LASSERTF(stripes <= lsm->lsm_stripe_count,"requested %d allocated %d\n",
 990                  lsm->lsm_stripe_count, stripes);
 991
 992         for (i = 0; i < stripes; i++) {
 993                 struct lov_request *req;
 994                 int ost_idx = idx_arr[i];
 995                 LASSERT(ost_idx >= 0);
 996
 997                 OBD_ALLOC(req, sizeof(*req));
 998                 if (req == NULL)
 999                         GOTO(out_err, rc = -ENOMEM);
1000                 lov_set_add_req(req, set);
1001
1002                 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
1003                 OBD_ALLOC(req->rq_oi.oi_md, req->rq_buflen);
1004                 if (req->rq_oi.oi_md == NULL)
1005                         GOTO(out_err, rc = -ENOMEM);
1006
1007                 OBDO_ALLOC(req->rq_oi.oi_oa);
1008                 if (req->rq_oi.oi_oa == NULL)
1009                         GOTO(out_err, rc = -ENOMEM);
1010
1011                 req->rq_idx = ost_idx;
1012                 req->rq_stripe = i;
1013                 /* create data objects with "parent" OA */
1014                 memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
1015                 req->rq_oi.oi_cb_up = cb_create_update;
1016
1017                 /* XXX When we start creating objects on demand, we need to
1018                  *     make sure that we always create the object on the
1019                  *     stripe which holds the existing file size.
1020                  */
1021                 if (src_oa->o_valid & OBD_MD_FLSIZE) {
1022                         req->rq_oi.oi_oa->o_size =
1023                                 lov_size_to_stripe(lsm, src_oa->o_size, i);
1024
1025                         CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
1026                                i, req->rq_oi.oi_oa->o_size, src_oa->o_size);
1027                 }
1028         }
1029         LASSERT(set->set_count == stripes);
1030
1031         if (stripes < lsm->lsm_stripe_count)
1032                 qos_shrink_lsm(set);
1033         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_LOV_PREP_CREATE)) {
1034                 qos_shrink_lsm(set);
1035                 rc = -EIO;
1036         }
1037
1038         if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) {
1039                 oti_alloc_cookies(oti, set->set_count);
1040                 if (!oti->oti_logcookies)
1041                         GOTO(out_err, rc = -ENOMEM);
1042                 set->set_cookies = oti->oti_logcookies;
1043         }
1044 out_err:
1045         if (newea && rc)
1046                 obd_free_memmd(exp, &set->set_oi->oi_md);
1047         if (idx_arr)
1048                 free_idx_array(idx_arr, idx_cnt);
1049         EXIT;
1050         return rc;
1051 }
1052
1053 void qos_update(struct lov_obd *lov)
1054 {
1055         ENTRY;
1056         lov->lov_qos.lq_dirty = 1;
1057 }
1058
1059 void qos_statfs_done(struct lov_obd *lov)
1060 {
1061         LASSERT(lov->lov_qos.lq_statfs_in_progress);
1062         down_write(&lov->lov_qos.lq_rw_sem);
1063         lov->lov_qos.lq_statfs_in_progress = 0;
1064         /* wake up any threads waiting for the statfs rpcs to complete */
1065         cfs_waitq_signal(&lov->lov_qos.lq_statfs_waitq);
1066         up_write(&lov->lov_qos.lq_rw_sem);
1067 }
1068
1069 static int qos_statfs_ready(struct obd_device *obd, __u64 max_age)
1070 {
1071         struct lov_obd         *lov = &obd->u.lov;
1072         int rc;
1073         ENTRY;
1074         down_read(&lov->lov_qos.lq_rw_sem);
1075         rc = lov->lov_qos.lq_statfs_in_progress == 0 ||
1076              cfs_time_beforeq_64(max_age, obd->obd_osfs_age);
1077         up_read(&lov->lov_qos.lq_rw_sem);
1078         RETURN(rc);
1079 }
1080
1081 /*
1082  * Update statfs data if the current osfs age is older than max_age.
1083  * If wait is not set, it means that we are called from lov_create()
1084  * and we should just issue the rpcs without waiting for them to complete.
1085  * If wait is set, we are called from alloc_qos() and we just have
1086  * to wait for the request set to complete.
1087  */
1088 void qos_statfs_update(struct obd_device *obd, __u64 max_age, int wait)
1089 {
1090         struct lov_obd         *lov = &obd->u.lov;
1091         struct obd_info        *oinfo;
1092         int                     rc = 0;
1093         struct ptlrpc_request_set *set = NULL;
1094         ENTRY;
1095
1096         if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
1097                 /* statfs data are quite recent, don't need to refresh it */
1098                 RETURN_EXIT;
1099
1100         if (!wait && lov->lov_qos.lq_statfs_in_progress)
1101                 /* statfs already in progress */
1102                 RETURN_EXIT;
1103
1104         down_write(&lov->lov_qos.lq_rw_sem);
1105         if (lov->lov_qos.lq_statfs_in_progress) {
1106                 up_write(&lov->lov_qos.lq_rw_sem);
1107                 GOTO(out, rc = 0);
1108         }
1109         /* no statfs in flight, send rpcs */
1110         lov->lov_qos.lq_statfs_in_progress = 1;
1111         up_write(&lov->lov_qos.lq_rw_sem);
1112
1113         if (wait)
1114                 CDEBUG(D_QOS, "%s: did not manage to get fresh statfs data "
1115                        "in a timely manner (osfs age "LPU64", max age "LPU64")"
1116                        ", sending new statfs rpcs\n",
1117                        obd_uuid2str(&lov->desc.ld_uuid), obd->obd_osfs_age,
1118                        max_age);
1119
1120         /* need to send statfs rpcs */
1121         CDEBUG(D_QOS, "sending new statfs requests\n");
1122         memset(lov->lov_qos.lq_statfs_data, 0,
1123                sizeof(*lov->lov_qos.lq_statfs_data));
1124         oinfo = &lov->lov_qos.lq_statfs_data->lsd_oi;
1125         oinfo->oi_osfs = &lov->lov_qos.lq_statfs_data->lsd_statfs;
1126         oinfo->oi_flags = OBD_STATFS_NODELAY;
1127         set = ptlrpc_prep_set();
1128         if (!set)
1129                 GOTO(out_failed, rc = -ENOMEM);
1130
1131         rc = obd_statfs_async(obd, oinfo, max_age, set);
1132         if (rc || list_empty(&set->set_requests)) {
1133                 if (rc)
1134                         CWARN("statfs failed with %d\n", rc);
1135                 GOTO(out_failed, rc);
1136         }
1137         /* send requests via ptlrpcd */
1138         oinfo->oi_flags |= OBD_STATFS_PTLRPCD;
1139         ptlrpcd_add_rqset(set);
1140         GOTO(out, rc);
1141
1142 out_failed:
1143         down_write(&lov->lov_qos.lq_rw_sem);
1144         lov->lov_qos.lq_statfs_in_progress = 0;
1145         /* wake up any threads waiting for the statfs rpcs to complete */
1146         cfs_waitq_signal(&lov->lov_qos.lq_statfs_waitq);
1147         up_write(&lov->lov_qos.lq_rw_sem);
1148         wait = 0;
1149 out:
1150         if (set)
1151                 ptlrpc_set_destroy(set);
1152         if (wait) {
1153                 struct l_wait_info lwi = { 0 };
1154                 CDEBUG(D_QOS, "waiting for statfs requests to complete\n");
1155                 l_wait_event(lov->lov_qos.lq_statfs_waitq,
1156                              qos_statfs_ready(obd, max_age), &lwi);
1157                 if (cfs_time_before_64(obd->obd_osfs_age, max_age))
1158                         CDEBUG(D_QOS, "%s: still no fresh statfs data after "
1159                                       "waiting (osfs age "LPU64", max age "
1160                                       LPU64")\n",
1161                                       obd_uuid2str(&lov->desc.ld_uuid),
1162                                       obd->obd_osfs_age, max_age);
1163         }
1164 }