lustre/lod/lod_qos.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9
  10  * This program is distributed in the hope that it will be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License version 2 for more details.  A copy is
  14  * included in the COPYING file that accompanied this code.
  15
  16  * You should have received a copy of the GNU General Public License
  17  * along with this program; if not, write to the Free Software
  18  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright  2009 Sun Microsystems, Inc. All rights reserved
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/lod/lod_qos.c
  33  *
  34  */
  35
  36 #define DEBUG_SUBSYSTEM S_LOV
  37
  38 #include <libcfs/libcfs.h>
  39 #include <obd_class.h>
  40 #include <obd_lov.h>
  41 #include <lustre/lustre_idl.h>
  42 #include "lod_internal.h"
  43
  44 /*
  45  * force QoS policy (not RR) to be used for testing purposes
  46  */
  47 #define FORCE_QOS_
  48
  49 #define D_QOS   D_OTHER
  50
  51 #if 0
  52 #define QOS_DEBUG(fmt, ...)     CDEBUG(D_OTHER, fmt, ## __VA_ARGS__)
  53 #define QOS_CONSOLE(fmt, ...)   LCONSOLE(D_OTHER, fmt, ## __VA_ARGS__)
  54 #else
  55 #define QOS_DEBUG(fmt, ...)
  56 #define QOS_CONSOLE(fmt, ...)
  57 #endif
  58
  59 #define TGT_BAVAIL(i) (OST_TGT(lod,i)->ltd_statfs.os_bavail * \
  60                        OST_TGT(lod,i)->ltd_statfs.os_bsize)
  61
  62 int qos_add_tgt(struct lod_device *lod, struct lod_ost_desc *ost_desc)
  63 {
  64         struct lov_qos_oss *oss = NULL, *temposs;
  65         struct obd_export  *exp = ost_desc->ltd_exp;
  66         int                 rc = 0, found = 0;
  67         cfs_list_t         *list;
  68         ENTRY;
  69
  70         down_write(&lod->lod_qos.lq_rw_sem);
  71         /*
  72          * a bit hacky approach to learn NID of corresponding connection
  73          * but there is no official API to access information like this
  74          * with OSD API.
  75          */
  76         cfs_list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
  77                 if (obd_uuid_equals(&oss->lqo_uuid,
  78                                     &exp->exp_connection->c_remote_uuid)) {
  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, &exp->exp_connection->c_remote_uuid,
  89                        sizeof(oss->lqo_uuid));
  90         } else {
  91                 /* Assume we have to move this one */
  92                 cfs_list_del(&oss->lqo_oss_list);
  93         }
  94
  95         oss->lqo_ost_count++;
  96         ost_desc->ltd_qos.ltq_oss = oss;
  97
  98         CDEBUG(D_QOS, "add tgt %s to OSS %s (%d OSTs)\n",
  99                obd_uuid2str(&ost_desc->ltd_uuid), obd_uuid2str(&oss->lqo_uuid),
 100                oss->lqo_ost_count);
 101
 102         /* Add sorted by # of OSTs.  Find the first entry that we're
 103            bigger than... */
 104         list = &lod->lod_qos.lq_oss_list;
 105         cfs_list_for_each_entry(temposs, list, lqo_oss_list) {
 106                 if (oss->lqo_ost_count > temposs->lqo_ost_count)
 107                         break;
 108         }
 109         /* ...and add before it.  If we're the first or smallest, temposs
 110            points to the list head, and we add to the end. */
 111         cfs_list_add_tail(&oss->lqo_oss_list, &temposs->lqo_oss_list);
 112
 113         lod->lod_qos.lq_dirty = 1;
 114         lod->lod_qos.lq_rr.lqr_dirty = 1;
 115
 116 out:
 117         up_write(&lod->lod_qos.lq_rw_sem);
 118         RETURN(rc);
 119 }
 120
 121 int qos_del_tgt(struct lod_device *lod, struct lod_ost_desc *ost_desc)
 122 {
 123         struct lov_qos_oss *oss;
 124         int                 rc = 0;
 125         ENTRY;
 126
 127         down_write(&lod->lod_qos.lq_rw_sem);
 128         oss = ost_desc->ltd_qos.ltq_oss;
 129         if (!oss)
 130                 GOTO(out, rc = -ENOENT);
 131
 132         oss->lqo_ost_count--;
 133         if (oss->lqo_ost_count == 0) {
 134                 CDEBUG(D_QOS, "removing OSS %s\n",
 135                        obd_uuid2str(&oss->lqo_uuid));
 136                 cfs_list_del(&oss->lqo_oss_list);
 137                 ost_desc->ltd_qos.ltq_oss = NULL;
 138                 OBD_FREE_PTR(oss);
 139         }
 140
 141         lod->lod_qos.lq_dirty = 1;
 142         lod->lod_qos.lq_rr.lqr_dirty = 1;
 143 out:
 144         up_write(&lod->lod_qos.lq_rw_sem);
 145         RETURN(rc);
 146 }
 147
 148 static int lod_statfs_and_check(const struct lu_env *env, struct lod_device *d,
 149                                 int index, struct obd_statfs *sfs)
 150 {
 151         struct lod_ost_desc *ost;
 152         int                  rc;
 153
 154         LASSERT(d);
 155         ost = OST_TGT(d,index);
 156         LASSERT(ost);
 157
 158         rc = dt_statfs(env, ost->ltd_ost, sfs);
 159         if (rc && rc != -ENOTCONN)
 160                 CERROR("%s: statfs: rc = %d\n", lod2obd(d)->obd_name, rc);
 161
 162         /* check whether device has changed state (active, inactive) */
 163         if (rc != 0 && ost->ltd_active) {
 164                 /* turned inactive? */
 165                 spin_lock(&d->lod_desc_lock);
 166                 if (ost->ltd_active) {
 167                         ost->ltd_active = 0;
 168                         LASSERT(d->lod_desc.ld_active_tgt_count > 0);
 169                         d->lod_desc.ld_active_tgt_count--;
 170                         d->lod_qos.lq_dirty = 1;
 171                         d->lod_qos.lq_rr.lqr_dirty = 1;
 172                         CDEBUG(D_CONFIG, "%s: turns inactive\n",
 173                                ost->ltd_exp->exp_obd->obd_name);
 174                 }
 175                 spin_unlock(&d->lod_desc_lock);
 176         } else if (rc == 0 && ost->ltd_active == 0) {
 177                 /* turned active? */
 178                 LASSERT(d->lod_desc.ld_active_tgt_count < d->lod_ostnr);
 179                 spin_lock(&d->lod_desc_lock);
 180                 if (ost->ltd_active == 0) {
 181                         ost->ltd_active = 1;
 182                         d->lod_desc.ld_active_tgt_count++;
 183                         d->lod_qos.lq_dirty = 1;
 184                         d->lod_qos.lq_rr.lqr_dirty = 1;
 185                         CDEBUG(D_CONFIG, "%s: turns active\n",
 186                                ost->ltd_exp->exp_obd->obd_name);
 187                 }
 188                 spin_unlock(&d->lod_desc_lock);
 189         }
 190
 191         return rc;
 192 }
 193
 194 static void lod_qos_statfs_update(const struct lu_env *env,
 195                                   struct lod_device *lod)
 196 {
 197         struct obd_device *obd = lod2obd(lod);
 198         struct ost_pool   *osts = &(lod->lod_pool_info);
 199         int                i, idx, rc = 0;
 200         __u64              max_age, avail;
 201         ENTRY;
 202
 203         max_age = cfs_time_shift_64(-2 * lod->lod_desc.ld_qos_maxage);
 204
 205         if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
 206                 /* statfs data are quite recent, don't need to refresh it */
 207                 RETURN_EXIT;
 208
 209         down_write(&lod->lod_qos.lq_rw_sem);
 210         if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
 211                 GOTO(out, rc = 0);
 212
 213         for (i = 0; i < osts->op_count; i++) {
 214                 idx = osts->op_array[i];
 215                 avail = OST_TGT(lod,idx)->ltd_statfs.os_bavail;
 216                 rc = lod_statfs_and_check(env, lod, idx,
 217                                           &OST_TGT(lod,idx)->ltd_statfs);
 218                 if (rc)
 219                         break;
 220                 if (OST_TGT(lod,idx)->ltd_statfs.os_bavail != avail)
 221                         /* recalculate weigths */
 222                         lod->lod_qos.lq_dirty = 1;
 223         }
 224         obd->obd_osfs_age = cfs_time_current_64();
 225
 226 out:
 227         up_write(&lod->lod_qos.lq_rw_sem);
 228 }
 229
 230 /* Recalculate per-object penalties for OSSs and OSTs,
 231    depends on size of each ost in an oss */
 232 static int lod_qos_calc_ppo(struct lod_device *lod)
 233 {
 234         struct lov_qos_oss *oss;
 235         __u64               ba_max, ba_min, temp;
 236         __u32               num_active;
 237         int                 rc, i, prio_wide;
 238         time_t              now, age;
 239         ENTRY;
 240
 241         if (!lod->lod_qos.lq_dirty)
 242                 GOTO(out, rc = 0);
 243
 244         num_active = lod->lod_desc.ld_active_tgt_count - 1;
 245         if (num_active < 1)
 246                 GOTO(out, rc = -EAGAIN);
 247
 248         /* find bavail on each OSS */
 249         cfs_list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list)
 250                                 oss->lqo_bavail = 0;
 251         lod->lod_qos.lq_active_oss_count = 0;
 252
 253         /*
 254          * How badly user wants to select OSTs "widely" (not recently chosen
 255          * and not on recent OSS's).  As opposed to "freely" (free space
 256          * avail.) 0-256
 257          */
 258         prio_wide = 256 - lod->lod_qos.lq_prio_free;
 259
 260         ba_min = (__u64)(-1);
 261         ba_max = 0;
 262         now = cfs_time_current_sec();
 263         /* Calculate OST penalty per object
 264          * (lod ref taken in lod_qos_prep_create()) */
 265         cfs_foreach_bit(lod->lod_ost_bitmap, i) {
 266                 LASSERT(OST_TGT(lod,i));
 267                 temp = TGT_BAVAIL(i);
 268                 if (!temp)
 269                         continue;
 270                 ba_min = min(temp, ba_min);
 271                 ba_max = max(temp, ba_max);
 272
 273                 /* Count the number of usable OSS's */
 274                 if (OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail == 0)
 275                         lod->lod_qos.lq_active_oss_count++;
 276                 OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail += temp;
 277
 278                 /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
 279                 temp >>= 1;
 280                 lov_do_div64(temp, num_active);
 281                 OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj =
 282                         (temp * prio_wide) >> 8;
 283
 284                 age = (now - OST_TGT(lod,i)->ltd_qos.ltq_used) >> 3;
 285                 if (lod->lod_qos.lq_reset ||
 286                     age > 32 * lod->lod_desc.ld_qos_maxage)
 287                         OST_TGT(lod,i)->ltd_qos.ltq_penalty = 0;
 288                 else if (age > lod->lod_desc.ld_qos_maxage)
 289                         /* Decay the penalty by half for every 8x the update
 290                          * interval that the device has been idle.  That gives
 291                          * lots of time for the statfs information to be
 292                          * updated (which the penalty is only a proxy for),
 293                          * and avoids penalizing OSS/OSTs under light load. */
 294                         OST_TGT(lod,i)->ltd_qos.ltq_penalty >>=
 295                                 (age / lod->lod_desc.ld_qos_maxage);
 296         }
 297
 298         num_active = lod->lod_qos.lq_active_oss_count - 1;
 299         if (num_active < 1) {
 300                 /* If there's only 1 OSS, we can't penalize it, so instead
 301                    we have to double the OST penalty */
 302                 num_active = 1;
 303                 cfs_foreach_bit(lod->lod_ost_bitmap, i)
 304                         OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj <<= 1;
 305         }
 306
 307         /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
 308         cfs_list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
 309                 temp = oss->lqo_bavail >> 1;
 310                 lov_do_div64(temp, oss->lqo_ost_count * num_active);
 311                 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
 312
 313                 age = (now - oss->lqo_used) >> 3;
 314                 if (lod->lod_qos.lq_reset ||
 315                     age > 32 * lod->lod_desc.ld_qos_maxage)
 316                         oss->lqo_penalty = 0;
 317                 else if (age > lod->lod_desc.ld_qos_maxage)
 318                         /* Decay the penalty by half for every 8x the update
 319                          * interval that the device has been idle.  That gives
 320                          * lots of time for the statfs information to be
 321                          * updated (which the penalty is only a proxy for),
 322                          * and avoids penalizing OSS/OSTs under light load. */
 323                         oss->lqo_penalty >>= age / lod->lod_desc.ld_qos_maxage;
 324         }
 325
 326         lod->lod_qos.lq_dirty = 0;
 327         lod->lod_qos.lq_reset = 0;
 328
 329         /* If each ost has almost same free space,
 330          * do rr allocation for better creation performance */
 331         lod->lod_qos.lq_same_space = 0;
 332         if ((ba_max * (256 - lod->lod_qos.lq_threshold_rr)) >> 8 < ba_min) {
 333                 lod->lod_qos.lq_same_space = 1;
 334                 /* Reset weights for the next time we enter qos mode */
 335                 lod->lod_qos.lq_reset = 1;
 336         }
 337         rc = 0;
 338
 339 out:
 340 #ifndef FORCE_QOS
 341         if (!rc && lod->lod_qos.lq_same_space)
 342                 RETURN(-EAGAIN);
 343 #endif
 344         RETURN(rc);
 345 }
 346
 347 static int lod_qos_calc_weight(struct lod_device *lod, int i)
 348 {
 349         __u64 temp, temp2;
 350
 351         /* Final ost weight = TGT_BAVAIL - ost_penalty - oss_penalty */
 352         temp = TGT_BAVAIL(i);
 353         temp2 = OST_TGT(lod,i)->ltd_qos.ltq_penalty +
 354                 OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_penalty;
 355         if (temp < temp2)
 356                 OST_TGT(lod,i)->ltd_qos.ltq_weight = 0;
 357         else
 358                 OST_TGT(lod,i)->ltd_qos.ltq_weight = temp - temp2;
 359         return 0;
 360 }
 361
 362 /* We just used this index for a stripe; adjust everyone's weights */
 363 static int lod_qos_used(struct lod_device *lod, struct ost_pool *osts,
 364                         __u32 index, __u64 *total_wt)
 365 {
 366         struct lod_ost_desc *ost;
 367         struct lov_qos_oss  *oss;
 368         int j;
 369         ENTRY;
 370
 371         ost = OST_TGT(lod,index);
 372         LASSERT(ost);
 373
 374         /* Don't allocate on this devuce anymore, until the next alloc_qos */
 375         ost->ltd_qos.ltq_usable = 0;
 376
 377         oss = ost->ltd_qos.ltq_oss;
 378
 379         /* Decay old penalty by half (we're adding max penalty, and don't
 380            want it to run away.) */
 381         ost->ltd_qos.ltq_penalty >>= 1;
 382         oss->lqo_penalty >>= 1;
 383
 384         /* mark the OSS and OST as recently used */
 385         ost->ltd_qos.ltq_used = oss->lqo_used = cfs_time_current_sec();
 386
 387         /* Set max penalties for this OST and OSS */
 388         ost->ltd_qos.ltq_penalty +=
 389                 ost->ltd_qos.ltq_penalty_per_obj * lod->lod_ostnr;
 390         oss->lqo_penalty += oss->lqo_penalty_per_obj *
 391                 lod->lod_qos.lq_active_oss_count;
 392
 393         /* Decrease all OSS penalties */
 394         cfs_list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
 395                 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
 396                         oss->lqo_penalty = 0;
 397                 else
 398                         oss->lqo_penalty -= oss->lqo_penalty_per_obj;
 399         }
 400
 401         *total_wt = 0;
 402         /* Decrease all OST penalties */
 403         for (j = 0; j < osts->op_count; j++) {
 404                 int i;
 405
 406                 i = osts->op_array[j];
 407                 if (!cfs_bitmap_check(lod->lod_ost_bitmap, i))
 408                         continue;
 409
 410                 ost = OST_TGT(lod,i);
 411                 LASSERT(ost);
 412
 413                 if (ost->ltd_qos.ltq_penalty <
 414                                 ost->ltd_qos.ltq_penalty_per_obj)
 415                         ost->ltd_qos.ltq_penalty = 0;
 416                 else
 417                         ost->ltd_qos.ltq_penalty -=
 418                                 ost->ltd_qos.ltq_penalty_per_obj;
 419
 420                 lod_qos_calc_weight(lod, i);
 421
 422                 /* Recalc the total weight of usable osts */
 423                 if (ost->ltd_qos.ltq_usable)
 424                         *total_wt += ost->ltd_qos.ltq_weight;
 425
 426                 QOS_DEBUG("recalc tgt %d usable=%d avail="LPU64
 427                           " ostppo="LPU64" ostp="LPU64" ossppo="LPU64
 428                           " ossp="LPU64" wt="LPU64"\n",
 429                           i, ost->ltd_qos.ltq_usable, TGT_BAVAIL(i) >> 10,
 430                           ost->ltd_qos.ltq_penalty_per_obj >> 10,
 431                           ost->ltd_qos.ltq_penalty >> 10,
 432                           ost->ltd_qos.ltq_oss->lqo_penalty_per_obj >> 10,
 433                           ost->ltd_qos.ltq_oss->lqo_penalty >> 10,
 434                           ost->ltd_qos.ltq_weight >> 10);
 435         }
 436
 437         RETURN(0);
 438 }
 439
 440 #define LOV_QOS_EMPTY ((__u32)-1)
 441 /* compute optimal round-robin order, based on OSTs per OSS */
 442 static int lod_qos_calc_rr(struct lod_device *lod, struct ost_pool *src_pool,
 443                            struct lov_qos_rr *lqr)
 444 {
 445         struct lov_qos_oss  *oss;
 446         struct lod_ost_desc *ost;
 447         unsigned placed, real_count;
 448         int i, rc;
 449         ENTRY;
 450
 451         if (!lqr->lqr_dirty) {
 452                 LASSERT(lqr->lqr_pool.op_size);
 453                 RETURN(0);
 454         }
 455
 456         /* Do actual allocation. */
 457         down_write(&lod->lod_qos.lq_rw_sem);
 458
 459         /*
 460          * Check again. While we were sleeping on @lq_rw_sem something could
 461          * change.
 462          */
 463         if (!lqr->lqr_dirty) {
 464                 LASSERT(lqr->lqr_pool.op_size);
 465                 up_write(&lod->lod_qos.lq_rw_sem);
 466                 RETURN(0);
 467         }
 468
 469         real_count = src_pool->op_count;
 470
 471         /* Zero the pool array */
 472         /* alloc_rr is holding a read lock on the pool, so nobody is adding/
 473            deleting from the pool. The lq_rw_sem insures that nobody else
 474            is reading. */
 475         lqr->lqr_pool.op_count = real_count;
 476         rc = lod_ost_pool_extend(&lqr->lqr_pool, real_count);
 477         if (rc) {
 478                 up_write(&lod->lod_qos.lq_rw_sem);
 479                 RETURN(rc);
 480         }
 481         for (i = 0; i < lqr->lqr_pool.op_count; i++)
 482                 lqr->lqr_pool.op_array[i] = LOV_QOS_EMPTY;
 483
 484         /* Place all the OSTs from 1 OSS at the same time. */
 485         placed = 0;
 486         cfs_list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
 487                 int j = 0;
 488
 489                 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
 490                         int next;
 491
 492                         if (!cfs_bitmap_check(lod->lod_ost_bitmap,
 493                                                 src_pool->op_array[i]))
 494                                 continue;
 495
 496                         ost = OST_TGT(lod,src_pool->op_array[i]);
 497                         LASSERT(ost && ost->ltd_ost);
 498                         if (ost->ltd_qos.ltq_oss != oss)
 499                                 continue;
 500
 501                         /* Evenly space these OSTs across arrayspace */
 502                         next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count;
 503                         while (lqr->lqr_pool.op_array[next] != LOV_QOS_EMPTY)
 504                                 next = (next + 1) % lqr->lqr_pool.op_count;
 505
 506                         lqr->lqr_pool.op_array[next] = src_pool->op_array[i];
 507                         j++;
 508                         placed++;
 509                 }
 510         }
 511
 512         lqr->lqr_dirty = 0;
 513         up_write(&lod->lod_qos.lq_rw_sem);
 514
 515         if (placed != real_count) {
 516                 /* This should never happen */
 517                 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
 518                                    "round-robin list (%d of %d).\n",
 519                                    placed, real_count);
 520                 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
 521                         LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
 522                                  lqr->lqr_pool.op_array[i]);
 523                 }
 524                 lqr->lqr_dirty = 1;
 525                 RETURN(-EAGAIN);
 526         }
 527
 528 #if 0
 529         for (i = 0; i < lqr->lqr_pool.op_count; i++)
 530                 QOS_CONSOLE("rr #%d ost idx=%d\n", i, lqr->lqr_pool.op_array[i]);
 531 #endif
 532
 533         RETURN(0);
 534 }
 535
 536 /**
 537  * A helper function to:
 538  *   create in-core lu object on the specified OSP
 539  *   declare creation of the object
 540  * IMPORTANT: at this stage object is anonymouos - it has no fid assigned
 541  *            this is a workaround till we have natural FIDs on OST
 542  *
 543  *            at this point we want to declare (reserve) object for us as
 544  *            we can't block at execution (when create method is called).
 545  *            otherwise we'd block whole transaction batch
 546  */
 547 static struct dt_object *lod_qos_declare_object_on(const struct lu_env *env,
 548                                                    struct lod_device *d,
 549                                                    int ost_idx,
 550                                                    struct thandle *th)
 551 {
 552         struct lod_ost_desc *ost;
 553         struct lu_object *o, *n;
 554         struct lu_device *nd;
 555         struct dt_object *dt;
 556         int               rc;
 557         ENTRY;
 558
 559         LASSERT(d);
 560         LASSERT(ost_idx >= 0);
 561         LASSERT(ost_idx < d->lod_osts_size);
 562         ost = OST_TGT(d,ost_idx);
 563         LASSERT(ost);
 564         LASSERT(ost->ltd_ost);
 565
 566         nd = &ost->ltd_ost->dd_lu_dev;
 567
 568         /*
 569          * allocate anonymous object with zero fid, real fid
 570          * will be assigned by OSP within transaction
 571          * XXX: to be fixed with fully-functional OST fids
 572          */
 573         o = lu_object_anon(env, nd, NULL);
 574         if (IS_ERR(o))
 575                 GOTO(out, dt = ERR_PTR(PTR_ERR(o)));
 576
 577         n = lu_object_locate(o->lo_header, nd->ld_type);
 578         if (unlikely(n == NULL)) {
 579                 CERROR("can't find slice\n");
 580                 lu_object_put(env, o);
 581                 GOTO(out, dt = ERR_PTR(-EINVAL));
 582         }
 583
 584         dt = container_of(n, struct dt_object, do_lu);
 585
 586         rc = dt_declare_create(env, dt, NULL, NULL, NULL, th);
 587         if (rc) {
 588                 CDEBUG(D_OTHER, "can't declare creation on #%u: %d\n",
 589                        ost_idx, rc);
 590                 lu_object_put(env, o);
 591                 dt = ERR_PTR(rc);
 592         }
 593
 594 out:
 595         RETURN(dt);
 596 }
 597
 598 static int min_stripe_count(int stripe_cnt, int flags)
 599 {
 600         return (flags & LOV_USES_DEFAULT_STRIPE ?
 601                         stripe_cnt - (stripe_cnt / 4) : stripe_cnt);
 602 }
 603
 604 #define LOV_CREATE_RESEED_MULT 30
 605 #define LOV_CREATE_RESEED_MIN  2000
 606
 607 static int inline lod_qos_dev_is_full(struct obd_statfs *msfs)
 608 {
 609         __u64 used;
 610         int   bs = msfs->os_bsize;
 611
 612         LASSERT(((bs - 1) & bs) == 0);
 613
 614         /* the minimum of 0.1% used blocks and 1GB bytes. */
 615         used = min_t(__u64, (msfs->os_blocks - msfs->os_bfree) >> 10,
 616                         1 << (31 - ffs(bs)));
 617         return (msfs->os_bavail < used);
 618 }
 619
 620 int lod_ea_store_resize(struct lod_thread_info *info, int size);
 621
 622 static inline int lod_qos_ost_in_use_clear(const struct lu_env *env, int stripes)
 623 {
 624         struct lod_thread_info *info = lod_env_info(env);
 625
 626         if (info->lti_ea_store_size < sizeof(int) * stripes)
 627                 lod_ea_store_resize(info, stripes * sizeof(int));
 628         if (info->lti_ea_store_size < sizeof(int) * stripes) {
 629                 CERROR("can't allocate memory for ost-in-use array\n");
 630                 return -ENOMEM;
 631         }
 632         memset(info->lti_ea_store, 0, sizeof(int) * stripes);
 633         return 0;
 634 }
 635
 636 static inline void lod_qos_ost_in_use(const struct lu_env *env, int idx, int ost)
 637 {
 638         struct lod_thread_info *info = lod_env_info(env);
 639         int *osts = info->lti_ea_store;
 640
 641         LASSERT(info->lti_ea_store_size >= idx * sizeof(int));
 642         osts[idx] = ost;
 643 }
 644
 645 static int lod_qos_is_ost_used(const struct lu_env *env, int ost, int stripes)
 646 {
 647         struct lod_thread_info *info = lod_env_info(env);
 648         int *osts = info->lti_ea_store;
 649         int j;
 650
 651         for (j = 0; j < stripes; j++) {
 652                 if (osts[j] == ost)
 653                         return 1;
 654         }
 655         return 0;
 656 }
 657
 658 /* Allocate objects on OSTs with round-robin algorithm */
 659 static int lod_alloc_rr(const struct lu_env *env, struct lod_object *lo,
 660                         int flags, struct thandle *th)
 661 {
 662         struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
 663         struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
 664         struct pool_desc  *pool = NULL;
 665         struct ost_pool   *osts;
 666         struct lov_qos_rr *lqr;
 667         struct dt_object  *o;
 668         unsigned           array_idx;
 669         int                i, rc;
 670         int                ost_start_idx_temp;
 671         int                speed = 0;
 672         int                stripe_idx = 0;
 673         int                stripe_cnt = lo->ldo_stripenr;
 674         int                stripe_cnt_min = min_stripe_count(stripe_cnt, flags);
 675         __u32              ost_idx;
 676         ENTRY;
 677
 678         if (lo->ldo_pool)
 679                 pool = lod_find_pool(m, lo->ldo_pool);
 680
 681         if (pool != NULL) {
 682                 down_read(&pool_tgt_rw_sem(pool));
 683                 osts = &(pool->pool_obds);
 684                 lqr = &(pool->pool_rr);
 685         } else {
 686                 osts = &(m->lod_pool_info);
 687                 lqr = &(m->lod_qos.lq_rr);
 688         }
 689
 690         rc = lod_qos_calc_rr(m, osts, lqr);
 691         if (rc)
 692                 GOTO(out, rc);
 693
 694         rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
 695         if (rc)
 696                 GOTO(out, rc);
 697
 698         if (--lqr->lqr_start_count <= 0) {
 699                 lqr->lqr_start_idx = cfs_rand() % osts->op_count;
 700                 lqr->lqr_start_count =
 701                         (LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) +
 702                          LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U);
 703         } else if (stripe_cnt_min >= osts->op_count ||
 704                         lqr->lqr_start_idx > osts->op_count) {
 705                 /* If we have allocated from all of the OSTs, slowly
 706                  * precess the next start if the OST/stripe count isn't
 707                  * already doing this for us. */
 708                 lqr->lqr_start_idx %= osts->op_count;
 709                 if (stripe_cnt > 1 && (osts->op_count % stripe_cnt) != 1)
 710                         ++lqr->lqr_offset_idx;
 711         }
 712         down_read(&m->lod_qos.lq_rw_sem);
 713         ost_start_idx_temp = lqr->lqr_start_idx;
 714
 715 repeat_find:
 716         array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) %
 717                         osts->op_count;
 718
 719         QOS_DEBUG("pool '%s' want %d startidx %d startcnt %d offset %d "
 720                   "active %d count %d arrayidx %d\n",
 721                   lo->ldo_pool ? lo->ldo_pool : "",
 722                   stripe_cnt, lqr->lqr_start_idx, lqr->lqr_start_count,
 723                   lqr->lqr_offset_idx, osts->op_count, osts->op_count,
 724                   array_idx);
 725
 726         for (i = 0; i < osts->op_count && stripe_idx < lo->ldo_stripenr;
 727              i++, array_idx = (array_idx + 1) % osts->op_count) {
 728                 ++lqr->lqr_start_idx;
 729                 ost_idx = lqr->lqr_pool.op_array[array_idx];
 730
 731                 QOS_DEBUG("#%d strt %d act %d strp %d ary %d idx %d\n",
 732                           i, lqr->lqr_start_idx, /* XXX: active*/ 0,
 733                           stripe_idx, array_idx, ost_idx);
 734
 735                 if ((ost_idx == LOV_QOS_EMPTY) ||
 736                     !cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
 737                         continue;
 738
 739                 /* Fail Check before osc_precreate() is called
 740                    so we can only 'fail' single OSC. */
 741                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
 742                         continue;
 743
 744                 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
 745                 if (rc) {
 746                         /* this OSP doesn't feel well */
 747                         continue;
 748                 }
 749
 750                 /*
 751                  * skip full devices
 752                  */
 753                 if (lod_qos_dev_is_full(sfs)) {
 754                         QOS_DEBUG("#%d is full\n", ost_idx);
 755                         continue;
 756                 }
 757
 758                 /*
 759                  * We expect number of precreated objects in f_ffree at
 760                  * the first iteration, skip OSPs with no objects ready
 761                  */
 762                 if (sfs->os_fprecreated == 0 && speed == 0) {
 763                         QOS_DEBUG("#%d: precreation is empty\n", ost_idx);
 764                         continue;
 765                 }
 766
 767                 /*
 768                  * try to use another OSP if this one is degraded
 769                  */
 770                 if (sfs->os_state == OS_STATE_DEGRADED && speed < 2) {
 771                         QOS_DEBUG("#%d: degraded\n", ost_idx);
 772                         continue;
 773                 }
 774
 775                 /*
 776                  * do not put >1 objects on a single OST
 777                  */
 778                 if (speed && lod_qos_is_ost_used(env, ost_idx, stripe_idx))
 779                         continue;
 780
 781                 o = lod_qos_declare_object_on(env, m, ost_idx, th);
 782                 if (IS_ERR(o)) {
 783                         CDEBUG(D_OTHER, "can't declare new object on #%u: %d\n",
 784                                ost_idx, (int) PTR_ERR(o));
 785                         rc = PTR_ERR(o);
 786                         continue;
 787                 }
 788
 789                 /*
 790                  * We've successfuly declared (reserved) an object
 791                  */
 792                 lod_qos_ost_in_use(env, stripe_idx, ost_idx);
 793                 lo->ldo_stripe[stripe_idx] = o;
 794                 stripe_idx++;
 795
 796         }
 797         if ((speed < 2) && (stripe_idx < stripe_cnt_min)) {
 798                 /* Try again, allowing slower OSCs */
 799                 speed++;
 800                 lqr->lqr_start_idx = ost_start_idx_temp;
 801                 goto repeat_find;
 802         }
 803
 804         up_read(&m->lod_qos.lq_rw_sem);
 805
 806         if (stripe_idx) {
 807                 lo->ldo_stripenr = stripe_idx;
 808                 /* at least one stripe is allocated */
 809                 rc = 0;
 810         } else {
 811                 /* nobody provided us with a single object */
 812                 rc = -ENOSPC;
 813         }
 814
 815 out:
 816         if (pool != NULL) {
 817                 up_read(&pool_tgt_rw_sem(pool));
 818                 /* put back ref got by lod_find_pool() */
 819                 lod_pool_putref(pool);
 820         }
 821
 822         RETURN(rc);
 823 }
 824
 825 /* alloc objects on osts with specific stripe offset */
 826 static int lod_alloc_specific(const struct lu_env *env, struct lod_object *lo,
 827                               int flags, struct thandle *th)
 828 {
 829         struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
 830         struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
 831         struct dt_object  *o;
 832         unsigned           ost_idx, array_idx, ost_count;
 833         int                i, rc, stripe_num = 0;
 834         int                speed = 0;
 835         struct pool_desc  *pool = NULL;
 836         struct ost_pool   *osts;
 837         ENTRY;
 838
 839         rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
 840         if (rc)
 841                 GOTO(out, rc);
 842
 843         if (lo->ldo_pool)
 844                 pool = lod_find_pool(m, lo->ldo_pool);
 845
 846         if (pool != NULL) {
 847                 down_read(&pool_tgt_rw_sem(pool));
 848                 osts = &(pool->pool_obds);
 849         } else {
 850                 osts = &(m->lod_pool_info);
 851         }
 852
 853         ost_count = osts->op_count;
 854
 855 repeat_find:
 856         /* search loi_ost_idx in ost array */
 857         array_idx = 0;
 858         for (i = 0; i < ost_count; i++) {
 859                 if (osts->op_array[i] == lo->ldo_def_stripe_offset) {
 860                         array_idx = i;
 861                         break;
 862                 }
 863         }
 864         if (i == ost_count) {
 865                 CERROR("Start index %d not found in pool '%s'\n",
 866                        lo->ldo_def_stripe_offset,
 867                        lo->ldo_pool ? lo->ldo_pool : "");
 868                 GOTO(out, rc = -EINVAL);
 869         }
 870
 871         for (i = 0; i < ost_count;
 872                         i++, array_idx = (array_idx + 1) % ost_count) {
 873                 ost_idx = osts->op_array[array_idx];
 874
 875                 if (!cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
 876                         continue;
 877
 878                 /* Fail Check before osc_precreate() is called
 879                    so we can only 'fail' single OSC. */
 880                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
 881                         continue;
 882
 883                 /*
 884                  * do not put >1 objects on a single OST
 885                  */
 886                 if (lod_qos_is_ost_used(env, ost_idx, stripe_num))
 887                         continue;
 888
 889                 /* Drop slow OSCs if we can, but not for requested start idx.
 890                  *
 891                  * This means "if OSC is slow and it is not the requested
 892                  * start OST, then it can be skipped, otherwise skip it only
 893                  * if it is inactive/recovering/out-of-space." */
 894
 895                 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
 896                 if (rc) {
 897                         /* this OSP doesn't feel well */
 898                         continue;
 899                 }
 900
 901                 /*
 902                  * We expect number of precreated objects in f_ffree at
 903                  * the first iteration, skip OSPs with no objects ready
 904                  * don't apply this logic to OST specified with stripe_offset
 905                  */
 906                 if (i != 0 && sfs->os_fprecreated == 0 && speed == 0)
 907                         continue;
 908
 909                 o = lod_qos_declare_object_on(env, m, ost_idx, th);
 910                 if (IS_ERR(o)) {
 911                         CDEBUG(D_OTHER, "can't declare new object on #%u: %d\n",
 912                                ost_idx, (int) PTR_ERR(o));
 913                         continue;
 914                 }
 915
 916                 /*
 917                  * We've successfuly declared (reserved) an object
 918                  */
 919                 lo->ldo_stripe[stripe_num] = o;
 920                 stripe_num++;
 921
 922                 /* We have enough stripes */
 923                 if (stripe_num == lo->ldo_stripenr)
 924                         GOTO(out, rc = 0);
 925         }
 926         if (speed < 2) {
 927                 /* Try again, allowing slower OSCs */
 928                 speed++;
 929                 goto repeat_find;
 930         }
 931
 932         /* If we were passed specific striping params, then a failure to
 933          * meet those requirements is an error, since we can't reallocate
 934          * that memory (it might be part of a larger array or something).
 935          *
 936          * We can only get here if lsm_stripe_count was originally > 1.
 937          */
 938         CERROR("can't lstripe objid "DFID": have %d want %u\n",
 939                PFID(lu_object_fid(lod2lu_obj(lo))), stripe_num,
 940                lo->ldo_stripenr);
 941         rc = -EFBIG;
 942 out:
 943         if (pool != NULL) {
 944                 up_read(&pool_tgt_rw_sem(pool));
 945                 /* put back ref got by lod_find_pool() */
 946                 lod_pool_putref(pool);
 947         }
 948
 949         RETURN(rc);
 950 }
 951
 952 static inline int lod_qos_is_usable(struct lod_device *lod)
 953 {
 954 #ifdef FORCE_QOS
 955         /* to be able to debug QoS code */
 956         return 1;
 957 #endif
 958
 959         /* Detect -EAGAIN early, before expensive lock is taken. */
 960         if (!lod->lod_qos.lq_dirty && lod->lod_qos.lq_same_space)
 961                 return 0;
 962
 963         if (lod->lod_desc.ld_active_tgt_count < 2)
 964                 return 0;
 965
 966         return 1;
 967 }
 968
 969 /* Alloc objects on OSTs with optimization based on:
 970    - free space
 971    - network resources (shared OSS's)
 972  */
 973 static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo,
 974                          int flags, struct thandle *th)
 975 {
 976         struct lod_device   *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
 977         struct obd_statfs   *sfs = &lod_env_info(env)->lti_osfs;
 978         struct lod_ost_desc *ost;
 979         struct dt_object    *o;
 980         __u64                total_weight = 0;
 981         int                  nfound, good_osts, i, rc = 0;
 982         int                  stripe_cnt = lo->ldo_stripenr;
 983         int                  stripe_cnt_min;
 984         struct pool_desc    *pool = NULL;
 985         struct ost_pool    *osts;
 986         ENTRY;
 987
 988         stripe_cnt_min = min_stripe_count(stripe_cnt, flags);
 989         if (stripe_cnt_min < 1)
 990                 RETURN(-EINVAL);
 991
 992         if (lo->ldo_pool)
 993                 pool = lod_find_pool(m, lo->ldo_pool);
 994
 995         if (pool != NULL) {
 996                 down_read(&pool_tgt_rw_sem(pool));
 997                 osts = &(pool->pool_obds);
 998         } else {
 999                 osts = &(m->lod_pool_info);
1000         }
1001
1002         /* Detect -EAGAIN early, before expensive lock is taken. */
1003         if (!lod_qos_is_usable(m))
1004                 GOTO(out_nolock, rc = -EAGAIN);
1005
1006         /* Do actual allocation, use write lock here. */
1007         down_write(&m->lod_qos.lq_rw_sem);
1008
1009         /*
1010          * Check again, while we were sleeping on @lq_rw_sem things could
1011          * change.
1012          */
1013         if (!lod_qos_is_usable(m))
1014                 GOTO(out, rc = -EAGAIN);
1015
1016         rc = lod_qos_calc_ppo(m);
1017         if (rc)
1018                 GOTO(out, rc);
1019
1020         rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
1021         if (rc)
1022                 GOTO(out, rc);
1023
1024         good_osts = 0;
1025         /* Find all the OSTs that are valid stripe candidates */
1026         for (i = 0; i < osts->op_count; i++) {
1027                 if (!cfs_bitmap_check(m->lod_ost_bitmap, osts->op_array[i]))
1028                         continue;
1029
1030                 rc = lod_statfs_and_check(env, m, osts->op_array[i], sfs);
1031                 if (rc) {
1032                         /* this OSP doesn't feel well */
1033                         continue;
1034                 }
1035
1036                 /*
1037                  * skip full devices
1038                  */
1039                 if (lod_qos_dev_is_full(sfs))
1040                         continue;
1041
1042                 /* Fail Check before osc_precreate() is called
1043                    so we can only 'fail' single OSC. */
1044                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) &&
1045                                    osts->op_array[i] == 0)
1046                         continue;
1047
1048                 ost = OST_TGT(m,osts->op_array[i]);
1049                 ost->ltd_qos.ltq_usable = 1;
1050                 lod_qos_calc_weight(m, osts->op_array[i]);
1051                 total_weight += ost->ltd_qos.ltq_weight;
1052
1053                 good_osts++;
1054         }
1055
1056         QOS_DEBUG("found %d good osts\n", good_osts);
1057
1058         if (good_osts < stripe_cnt_min)
1059                 GOTO(out, rc = -EAGAIN);
1060
1061         /* We have enough osts */
1062         if (good_osts < stripe_cnt)
1063                 stripe_cnt = good_osts;
1064
1065         /* Find enough OSTs with weighted random allocation. */
1066         nfound = 0;
1067         while (nfound < stripe_cnt) {
1068                 __u64 rand, cur_weight;
1069
1070                 cur_weight = 0;
1071                 rc = -ENOSPC;
1072
1073                 if (total_weight) {
1074 #if BITS_PER_LONG == 32
1075                         rand = cfs_rand() % (unsigned)total_weight;
1076                         /* If total_weight > 32-bit, first generate the high
1077                          * 32 bits of the random number, then add in the low
1078                          * 32 bits (truncated to the upper limit, if needed) */
1079                         if (total_weight > 0xffffffffULL)
1080                                 rand = (__u64)(cfs_rand() %
1081                                         (unsigned)(total_weight >> 32)) << 32;
1082                         else
1083                                 rand = 0;
1084
1085                         if (rand == (total_weight & 0xffffffff00000000ULL))
1086                                 rand |= cfs_rand() % (unsigned)total_weight;
1087                         else
1088                                 rand |= cfs_rand();
1089
1090 #else
1091                         rand = ((__u64)cfs_rand() << 32 | cfs_rand()) %
1092                                 total_weight;
1093 #endif
1094                 } else {
1095                         rand = 0;
1096                 }
1097
1098                 /* On average, this will hit larger-weighted osts more often.
1099                    0-weight osts will always get used last (only when rand=0) */
1100                 for (i = 0; i < osts->op_count; i++) {
1101                         int idx = osts->op_array[i];
1102
1103                         if (!cfs_bitmap_check(m->lod_ost_bitmap, idx))
1104                                 continue;
1105
1106                         ost = OST_TGT(m,idx);
1107
1108                         if (!ost->ltd_qos.ltq_usable)
1109                                 continue;
1110
1111                         cur_weight += ost->ltd_qos.ltq_weight;
1112                         QOS_DEBUG("stripe_cnt=%d nfound=%d cur_weight="LPU64
1113                                   " rand="LPU64" total_weight="LPU64"\n",
1114                                   stripe_cnt, nfound, cur_weight, rand,
1115                                   total_weight);
1116
1117                         if (cur_weight < rand)
1118                                 continue;
1119
1120                         QOS_DEBUG("stripe=%d to idx=%d\n", nfound, idx);
1121
1122                         /*
1123                          * do not put >1 objects on a single OST
1124                          */
1125                         if (lod_qos_is_ost_used(env, idx, nfound))
1126                                 continue;
1127                         lod_qos_ost_in_use(env, nfound, idx);
1128
1129                         o = lod_qos_declare_object_on(env, m, idx, th);
1130                         if (IS_ERR(o)) {
1131                                 QOS_DEBUG("can't declare object on #%u: %d\n",
1132                                           idx, (int) PTR_ERR(o));
1133                                 continue;
1134                         }
1135                         lo->ldo_stripe[nfound++] = o;
1136                         lod_qos_used(m, osts, idx, &total_weight);
1137                         rc = 0;
1138                         break;
1139                 }
1140
1141                 if (rc) {
1142                         /* no OST found on this iteration, give up */
1143                         break;
1144                 }
1145         }
1146
1147         if (unlikely(nfound != stripe_cnt)) {
1148                 /*
1149                  * when the decision to use weighted algorithm was made
1150                  * we had enough appropriate OSPs, but this state can
1151                  * change anytime (no space on OST, broken connection, etc)
1152                  * so it's possible OSP won't be able to provide us with
1153                  * an object due to just changed state
1154                  */
1155                 LCONSOLE_INFO("wanted %d, found %d\n", stripe_cnt, nfound);
1156                 for (i = 0; i < nfound; i++) {
1157                         LASSERT(lo->ldo_stripe[i]);
1158                         lu_object_put(env, &lo->ldo_stripe[i]->do_lu);
1159                         lo->ldo_stripe[i] = NULL;
1160                 }
1161
1162                 /* makes sense to rebalance next time */
1163                 m->lod_qos.lq_dirty = 1;
1164                 m->lod_qos.lq_same_space = 0;
1165
1166                 rc = -EAGAIN;
1167         }
1168
1169 out:
1170         up_write(&m->lod_qos.lq_rw_sem);
1171
1172 out_nolock:
1173         if (pool != NULL) {
1174                 up_read(&pool_tgt_rw_sem(pool));
1175                 /* put back ref got by lod_find_pool() */
1176                 lod_pool_putref(pool);
1177         }
1178
1179         RETURN(rc);
1180 }
1181
1182 /* Find the max stripecount we should use */
1183 static __u16 lod_get_stripecnt(struct lod_device *lod, __u32 magic,
1184                                __u16 stripe_count)
1185 {
1186         __u32 max_stripes = LOV_MAX_STRIPE_COUNT_OLD;
1187
1188         if (!stripe_count)
1189                 stripe_count = lod->lod_desc.ld_default_stripe_count;
1190         if (stripe_count > lod->lod_desc.ld_active_tgt_count)
1191                 stripe_count = lod->lod_desc.ld_active_tgt_count;
1192         if (!stripe_count)
1193                 stripe_count = 1;
1194
1195         /* stripe count is based on whether OSD can handle larger EA sizes */
1196         if (lod->lod_osd_max_easize > 0)
1197                 max_stripes = lov_mds_md_stripecnt(lod->lod_osd_max_easize,
1198                                                    magic);
1199
1200         return (stripe_count < max_stripes) ? stripe_count : max_stripes;
1201 }
1202
1203 static int lod_use_defined_striping(const struct lu_env *env,
1204                                     struct lod_object *mo,
1205                                     const struct lu_buf *buf)
1206 {
1207         struct lod_device      *d = lu2lod_dev(lod2lu_obj(mo)->lo_dev);
1208         struct lov_mds_md_v1   *v1 = buf->lb_buf;
1209         struct lov_mds_md_v3   *v3 = buf->lb_buf;
1210         struct lov_ost_data_v1 *objs;
1211         __u32                   magic;
1212         int                     rc;
1213         ENTRY;
1214
1215         rc = lod_verify_striping(d, buf, 1);
1216         if (rc)
1217                 RETURN(rc);
1218
1219         magic = le32_to_cpu(v1->lmm_magic);
1220         if (magic == LOV_MAGIC_V1_DEF) {
1221                 objs = &v1->lmm_objects[0];
1222         } else if (magic == LOV_MAGIC_V3_DEF) {
1223                 objs = &v3->lmm_objects[0];
1224                 lod_object_set_pool(mo, v3->lmm_pool_name);
1225         } else {
1226                 GOTO(out, rc = -EINVAL);
1227         }
1228
1229         /*
1230          * LOD shouldn't be aware of recovery at all,
1231          * but this track recovery status (to some extent)
1232          * to be do additional checks like this one
1233          */
1234         LASSERT(d->lod_recovery_completed == 0);
1235
1236         mo->ldo_stripe_size = le32_to_cpu(v1->lmm_stripe_size);
1237         mo->ldo_stripenr = le16_to_cpu(v1->lmm_stripe_count);
1238         mo->ldo_layout_gen = le16_to_cpu(v1->lmm_layout_gen);
1239         LASSERT(buf->lb_len >= lov_mds_md_size(mo->ldo_stripenr, magic));
1240
1241         rc = lod_initialize_objects(env, mo, objs);
1242
1243 out:
1244         RETURN(rc);
1245 }
1246
1247 static int lod_qos_parse_config(const struct lu_env *env,
1248                                 struct lod_object *lo,
1249                                 const struct lu_buf *buf)
1250 {
1251         struct lod_device     *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
1252         struct lov_user_md_v1 *v1 = NULL;
1253         struct lov_user_md_v3 *v3 = NULL;
1254         struct pool_desc      *pool;
1255         __u32                  magic;
1256         int                    rc;
1257         ENTRY;
1258
1259         if (buf == NULL || buf->lb_buf == NULL || buf->lb_len == 0)
1260                 RETURN(0);
1261
1262         v1 = buf->lb_buf;
1263         magic = v1->lmm_magic;
1264
1265         if (magic == __swab32(LOV_USER_MAGIC_V1))
1266                 lustre_swab_lov_user_md_v1(v1);
1267         else if (magic == __swab32(LOV_USER_MAGIC_V3))
1268                 lustre_swab_lov_user_md_v3(v3);
1269
1270         if (unlikely(magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3)) {
1271                 /* try to use as fully defined striping */
1272                 rc = lod_use_defined_striping(env, lo, buf);
1273                 RETURN(rc);
1274         }
1275
1276         if (unlikely(buf->lb_len < sizeof(*v1))) {
1277                 CERROR("wrong size: %u\n", (unsigned) buf->lb_len);
1278                 RETURN(-EINVAL);
1279         }
1280
1281         if (v1->lmm_pattern != 0 && v1->lmm_pattern != LOV_PATTERN_RAID0) {
1282                 CERROR("invalid pattern: %x\n", v1->lmm_pattern);
1283                 RETURN(-EINVAL);
1284         }
1285
1286         if (v1->lmm_stripe_size)
1287                 lo->ldo_stripe_size = v1->lmm_stripe_size;
1288         if (lo->ldo_stripe_size & (LOV_MIN_STRIPE_SIZE - 1))
1289                 lo->ldo_stripe_size = LOV_MIN_STRIPE_SIZE;
1290
1291         if (v1->lmm_stripe_count)
1292                 lo->ldo_stripenr = v1->lmm_stripe_count;
1293
1294         if ((v1->lmm_stripe_offset >= d->lod_desc.ld_tgt_count) &&
1295             (v1->lmm_stripe_offset != (typeof(v1->lmm_stripe_offset))(-1))) {
1296                 CERROR("invalid offset: %x\n", v1->lmm_stripe_offset);
1297                 RETURN(-EINVAL);
1298         }
1299         lo->ldo_def_stripe_offset = v1->lmm_stripe_offset;
1300
1301         CDEBUG(D_OTHER, "lsm: %u size, %u stripes, %u offset\n",
1302                v1->lmm_stripe_size, v1->lmm_stripe_count,
1303                v1->lmm_stripe_offset);
1304
1305         if (v1->lmm_magic == LOV_MAGIC_V3) {
1306                 if (buf->lb_len < sizeof(*v3)) {
1307                         CERROR("wrong size: %u\n", (unsigned) buf->lb_len);
1308                         RETURN(-EINVAL);
1309                 }
1310
1311                 v3 = buf->lb_buf;
1312                 lod_object_set_pool(lo, v3->lmm_pool_name);
1313
1314                 pool = lod_find_pool(d, v3->lmm_pool_name);
1315                 if (pool != NULL) {
1316                         if (lo->ldo_def_stripe_offset !=
1317                             (typeof(v1->lmm_stripe_offset))(-1)) {
1318                                 rc = lo->ldo_def_stripe_offset;
1319                                 rc = lod_check_index_in_pool(rc, pool);
1320                                 if (rc < 0) {
1321                                         lod_pool_putref(pool);
1322                                         CERROR("invalid offset\n");
1323                                         RETURN(-EINVAL);
1324                                 }
1325                         }
1326
1327                         if (lo->ldo_stripenr > pool_tgt_count(pool))
1328                                 lo->ldo_stripenr= pool_tgt_count(pool);
1329
1330                         lod_pool_putref(pool);
1331                 }
1332         } else
1333                 lod_object_set_pool(lo, NULL);
1334
1335         RETURN(0);
1336 }
1337
1338 /*
1339  * buf should be NULL or contain striping settings
1340  */
1341 int lod_qos_prep_create(const struct lu_env *env, struct lod_object *lo,
1342                         struct lu_attr *attr, const struct lu_buf *buf,
1343                         struct thandle *th)
1344 {
1345         struct lod_device *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
1346         int flag = LOV_USES_ASSIGNED_STRIPE;
1347         int i, rc = 0;
1348         ENTRY;
1349
1350         LASSERT(lo);
1351
1352         /* no OST available */
1353         /* XXX: should we be waiting a bit to prevent failures during
1354          * cluster initialization? */
1355         if (d->lod_ostnr == 0)
1356                 GOTO(out, rc = -EIO);
1357
1358         /*
1359          * by this time, the object's ldo_stripenr and ldo_stripe_size
1360          * contain default value for striping: taken from the parent
1361          * or from filesystem defaults
1362          *
1363          * in case the caller is passing lovea with new striping config,
1364          * we may need to parse lovea and apply new configuration
1365          */
1366         rc = lod_qos_parse_config(env, lo, buf);
1367         if (rc)
1368                 GOTO(out, rc);
1369
1370         if (likely(lo->ldo_stripe == NULL)) {
1371                 /*
1372                  * no striping has been created so far
1373                  */
1374                 LASSERT(lo->ldo_stripenr > 0);
1375                 lo->ldo_stripenr = lod_get_stripecnt(d, LOV_MAGIC,
1376                                 lo->ldo_stripenr);
1377                 i = sizeof(struct dt_object *) * lo->ldo_stripenr;
1378                 OBD_ALLOC(lo->ldo_stripe, i);
1379                 if (lo->ldo_stripe == NULL)
1380                         GOTO(out, rc = -ENOMEM);
1381                 lo->ldo_stripes_allocated = lo->ldo_stripenr;
1382
1383                 lod_getref(d);
1384                 /* XXX: support for non-0 files w/o objects */
1385                 if (lo->ldo_def_stripe_offset >= d->lod_desc.ld_tgt_count) {
1386                         lod_qos_statfs_update(env, d);
1387                         rc = lod_alloc_qos(env, lo, flag, th);
1388                         if (rc == -EAGAIN)
1389                                 rc = lod_alloc_rr(env, lo, flag, th);
1390                 } else
1391                         rc = lod_alloc_specific(env, lo, flag, th);
1392                 lod_putref(d);
1393         } else {
1394                 /*
1395                  * lod_qos_parse_config() found supplied buf as a predefined
1396                  * striping (not a hint), so it allocated all the object
1397                  * now we need to create them
1398                  */
1399                 for (i = 0; i < lo->ldo_stripenr; i++) {
1400                         struct dt_object  *o;
1401
1402                         o = lo->ldo_stripe[i];
1403                         LASSERT(o);
1404
1405                         rc = dt_declare_create(env, o, attr, NULL, NULL, th);
1406                         if (rc) {
1407                                 CERROR("can't declare create: %d\n", rc);
1408                                 break;
1409                         }
1410                 }
1411         }
1412
1413 out:
1414         RETURN(rc);
1415 }
1416