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