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) 2012, 2013, Intel Corporation.
  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_tgt_desc *ost_desc)
  63 {
  64         struct lod_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_tgt_desc *ost_desc)
 122 {
 123         struct lod_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_tgt_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         /* If the OST is readonly then we can't allocate objects there */
 163         if (sfs->os_state & OS_STATE_READONLY)
 164                 rc = -EROFS;
 165
 166         /* check whether device has changed state (active, inactive) */
 167         if (rc != 0 && ost->ltd_active) {
 168                 /* turned inactive? */
 169                 spin_lock(&d->lod_desc_lock);
 170                 if (ost->ltd_active) {
 171                         ost->ltd_active = 0;
 172                         LASSERT(d->lod_desc.ld_active_tgt_count > 0);
 173                         d->lod_desc.ld_active_tgt_count--;
 174                         d->lod_qos.lq_dirty = 1;
 175                         d->lod_qos.lq_rr.lqr_dirty = 1;
 176                         CDEBUG(D_CONFIG, "%s: turns inactive\n",
 177                                ost->ltd_exp->exp_obd->obd_name);
 178                 }
 179                 spin_unlock(&d->lod_desc_lock);
 180         } else if (rc == 0 && ost->ltd_active == 0) {
 181                 /* turned active? */
 182                 LASSERTF(d->lod_desc.ld_active_tgt_count < d->lod_ostnr,
 183                          "active tgt count %d, ost nr %d\n",
 184                          d->lod_desc.ld_active_tgt_count, d->lod_ostnr);
 185                 spin_lock(&d->lod_desc_lock);
 186                 if (ost->ltd_active == 0) {
 187                         ost->ltd_active = 1;
 188                         d->lod_desc.ld_active_tgt_count++;
 189                         d->lod_qos.lq_dirty = 1;
 190                         d->lod_qos.lq_rr.lqr_dirty = 1;
 191                         CDEBUG(D_CONFIG, "%s: turns active\n",
 192                                ost->ltd_exp->exp_obd->obd_name);
 193                 }
 194                 spin_unlock(&d->lod_desc_lock);
 195         }
 196
 197         RETURN(rc);
 198 }
 199
 200 static void lod_qos_statfs_update(const struct lu_env *env,
 201                                   struct lod_device *lod)
 202 {
 203         struct obd_device *obd = lod2obd(lod);
 204         struct ost_pool   *osts = &(lod->lod_pool_info);
 205         int                i, idx, rc = 0;
 206         __u64              max_age, avail;
 207         ENTRY;
 208
 209         max_age = cfs_time_shift_64(-2 * lod->lod_desc.ld_qos_maxage);
 210
 211         if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
 212                 /* statfs data are quite recent, don't need to refresh it */
 213                 RETURN_EXIT;
 214
 215         down_write(&lod->lod_qos.lq_rw_sem);
 216         if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
 217                 GOTO(out, rc = 0);
 218
 219         for (i = 0; i < osts->op_count; i++) {
 220                 idx = osts->op_array[i];
 221                 avail = OST_TGT(lod,idx)->ltd_statfs.os_bavail;
 222                 rc = lod_statfs_and_check(env, lod, idx,
 223                                           &OST_TGT(lod,idx)->ltd_statfs);
 224                 if (rc)
 225                         break;
 226                 if (OST_TGT(lod,idx)->ltd_statfs.os_bavail != avail)
 227                         /* recalculate weigths */
 228                         lod->lod_qos.lq_dirty = 1;
 229         }
 230         obd->obd_osfs_age = cfs_time_current_64();
 231
 232 out:
 233         up_write(&lod->lod_qos.lq_rw_sem);
 234         EXIT;
 235 }
 236
 237 /* Recalculate per-object penalties for OSSs and OSTs,
 238    depends on size of each ost in an oss */
 239 static int lod_qos_calc_ppo(struct lod_device *lod)
 240 {
 241         struct lod_qos_oss *oss;
 242         __u64               ba_max, ba_min, temp;
 243         __u32               num_active;
 244         int                 rc, i, prio_wide;
 245         time_t              now, age;
 246         ENTRY;
 247
 248         if (!lod->lod_qos.lq_dirty)
 249                 GOTO(out, rc = 0);
 250
 251         num_active = lod->lod_desc.ld_active_tgt_count - 1;
 252         if (num_active < 1)
 253                 GOTO(out, rc = -EAGAIN);
 254
 255         /* find bavail on each OSS */
 256         cfs_list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list)
 257                                 oss->lqo_bavail = 0;
 258         lod->lod_qos.lq_active_oss_count = 0;
 259
 260         /*
 261          * How badly user wants to select OSTs "widely" (not recently chosen
 262          * and not on recent OSS's).  As opposed to "freely" (free space
 263          * avail.) 0-256
 264          */
 265         prio_wide = 256 - lod->lod_qos.lq_prio_free;
 266
 267         ba_min = (__u64)(-1);
 268         ba_max = 0;
 269         now = cfs_time_current_sec();
 270         /* Calculate OST penalty per object
 271          * (lod ref taken in lod_qos_prep_create()) */
 272         cfs_foreach_bit(lod->lod_ost_bitmap, i) {
 273                 LASSERT(OST_TGT(lod,i));
 274                 temp = TGT_BAVAIL(i);
 275                 if (!temp)
 276                         continue;
 277                 ba_min = min(temp, ba_min);
 278                 ba_max = max(temp, ba_max);
 279
 280                 /* Count the number of usable OSS's */
 281                 if (OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail == 0)
 282                         lod->lod_qos.lq_active_oss_count++;
 283                 OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_bavail += temp;
 284
 285                 /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
 286                 temp >>= 1;
 287                 lov_do_div64(temp, num_active);
 288                 OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj =
 289                         (temp * prio_wide) >> 8;
 290
 291                 age = (now - OST_TGT(lod,i)->ltd_qos.ltq_used) >> 3;
 292                 if (lod->lod_qos.lq_reset ||
 293                     age > 32 * lod->lod_desc.ld_qos_maxage)
 294                         OST_TGT(lod,i)->ltd_qos.ltq_penalty = 0;
 295                 else if (age > lod->lod_desc.ld_qos_maxage)
 296                         /* Decay the penalty by half for every 8x the update
 297                          * interval that the device has been idle.  That gives
 298                          * lots of time for the statfs information to be
 299                          * updated (which the penalty is only a proxy for),
 300                          * and avoids penalizing OSS/OSTs under light load. */
 301                         OST_TGT(lod,i)->ltd_qos.ltq_penalty >>=
 302                                 (age / lod->lod_desc.ld_qos_maxage);
 303         }
 304
 305         num_active = lod->lod_qos.lq_active_oss_count - 1;
 306         if (num_active < 1) {
 307                 /* If there's only 1 OSS, we can't penalize it, so instead
 308                    we have to double the OST penalty */
 309                 num_active = 1;
 310                 cfs_foreach_bit(lod->lod_ost_bitmap, i)
 311                         OST_TGT(lod,i)->ltd_qos.ltq_penalty_per_obj <<= 1;
 312         }
 313
 314         /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
 315         cfs_list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
 316                 temp = oss->lqo_bavail >> 1;
 317                 lov_do_div64(temp, oss->lqo_ost_count * num_active);
 318                 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
 319
 320                 age = (now - oss->lqo_used) >> 3;
 321                 if (lod->lod_qos.lq_reset ||
 322                     age > 32 * lod->lod_desc.ld_qos_maxage)
 323                         oss->lqo_penalty = 0;
 324                 else if (age > lod->lod_desc.ld_qos_maxage)
 325                         /* Decay the penalty by half for every 8x the update
 326                          * interval that the device has been idle.  That gives
 327                          * lots of time for the statfs information to be
 328                          * updated (which the penalty is only a proxy for),
 329                          * and avoids penalizing OSS/OSTs under light load. */
 330                         oss->lqo_penalty >>= age / lod->lod_desc.ld_qos_maxage;
 331         }
 332
 333         lod->lod_qos.lq_dirty = 0;
 334         lod->lod_qos.lq_reset = 0;
 335
 336         /* If each ost has almost same free space,
 337          * do rr allocation for better creation performance */
 338         lod->lod_qos.lq_same_space = 0;
 339         if ((ba_max * (256 - lod->lod_qos.lq_threshold_rr)) >> 8 < ba_min) {
 340                 lod->lod_qos.lq_same_space = 1;
 341                 /* Reset weights for the next time we enter qos mode */
 342                 lod->lod_qos.lq_reset = 1;
 343         }
 344         rc = 0;
 345
 346 out:
 347 #ifndef FORCE_QOS
 348         if (!rc && lod->lod_qos.lq_same_space)
 349                 RETURN(-EAGAIN);
 350 #endif
 351         RETURN(rc);
 352 }
 353
 354 static int lod_qos_calc_weight(struct lod_device *lod, int i)
 355 {
 356         __u64 temp, temp2;
 357
 358         /* Final ost weight = TGT_BAVAIL - ost_penalty - oss_penalty */
 359         temp = TGT_BAVAIL(i);
 360         temp2 = OST_TGT(lod,i)->ltd_qos.ltq_penalty +
 361                 OST_TGT(lod,i)->ltd_qos.ltq_oss->lqo_penalty;
 362         if (temp < temp2)
 363                 OST_TGT(lod,i)->ltd_qos.ltq_weight = 0;
 364         else
 365                 OST_TGT(lod,i)->ltd_qos.ltq_weight = temp - temp2;
 366         return 0;
 367 }
 368
 369 /* We just used this index for a stripe; adjust everyone's weights */
 370 static int lod_qos_used(struct lod_device *lod, struct ost_pool *osts,
 371                         __u32 index, __u64 *total_wt)
 372 {
 373         struct lod_tgt_desc *ost;
 374         struct lod_qos_oss  *oss;
 375         int j;
 376         ENTRY;
 377
 378         ost = OST_TGT(lod,index);
 379         LASSERT(ost);
 380
 381         /* Don't allocate on this devuce anymore, until the next alloc_qos */
 382         ost->ltd_qos.ltq_usable = 0;
 383
 384         oss = ost->ltd_qos.ltq_oss;
 385
 386         /* Decay old penalty by half (we're adding max penalty, and don't
 387            want it to run away.) */
 388         ost->ltd_qos.ltq_penalty >>= 1;
 389         oss->lqo_penalty >>= 1;
 390
 391         /* mark the OSS and OST as recently used */
 392         ost->ltd_qos.ltq_used = oss->lqo_used = cfs_time_current_sec();
 393
 394         /* Set max penalties for this OST and OSS */
 395         ost->ltd_qos.ltq_penalty +=
 396                 ost->ltd_qos.ltq_penalty_per_obj * lod->lod_ostnr;
 397         oss->lqo_penalty += oss->lqo_penalty_per_obj *
 398                 lod->lod_qos.lq_active_oss_count;
 399
 400         /* Decrease all OSS penalties */
 401         cfs_list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
 402                 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
 403                         oss->lqo_penalty = 0;
 404                 else
 405                         oss->lqo_penalty -= oss->lqo_penalty_per_obj;
 406         }
 407
 408         *total_wt = 0;
 409         /* Decrease all OST penalties */
 410         for (j = 0; j < osts->op_count; j++) {
 411                 int i;
 412
 413                 i = osts->op_array[j];
 414                 if (!cfs_bitmap_check(lod->lod_ost_bitmap, i))
 415                         continue;
 416
 417                 ost = OST_TGT(lod,i);
 418                 LASSERT(ost);
 419
 420                 if (ost->ltd_qos.ltq_penalty <
 421                                 ost->ltd_qos.ltq_penalty_per_obj)
 422                         ost->ltd_qos.ltq_penalty = 0;
 423                 else
 424                         ost->ltd_qos.ltq_penalty -=
 425                                 ost->ltd_qos.ltq_penalty_per_obj;
 426
 427                 lod_qos_calc_weight(lod, i);
 428
 429                 /* Recalc the total weight of usable osts */
 430                 if (ost->ltd_qos.ltq_usable)
 431                         *total_wt += ost->ltd_qos.ltq_weight;
 432
 433                 QOS_DEBUG("recalc tgt %d usable=%d avail="LPU64
 434                           " ostppo="LPU64" ostp="LPU64" ossppo="LPU64
 435                           " ossp="LPU64" wt="LPU64"\n",
 436                           i, ost->ltd_qos.ltq_usable, TGT_BAVAIL(i) >> 10,
 437                           ost->ltd_qos.ltq_penalty_per_obj >> 10,
 438                           ost->ltd_qos.ltq_penalty >> 10,
 439                           ost->ltd_qos.ltq_oss->lqo_penalty_per_obj >> 10,
 440                           ost->ltd_qos.ltq_oss->lqo_penalty >> 10,
 441                           ost->ltd_qos.ltq_weight >> 10);
 442         }
 443
 444         RETURN(0);
 445 }
 446
 447 #define LOV_QOS_EMPTY ((__u32)-1)
 448 /* compute optimal round-robin order, based on OSTs per OSS */
 449 static int lod_qos_calc_rr(struct lod_device *lod, struct ost_pool *src_pool,
 450                            struct lod_qos_rr *lqr)
 451 {
 452         struct lod_qos_oss  *oss;
 453         struct lod_tgt_desc *ost;
 454         unsigned placed, real_count;
 455         int i, rc;
 456         ENTRY;
 457
 458         if (!lqr->lqr_dirty) {
 459                 LASSERT(lqr->lqr_pool.op_size);
 460                 RETURN(0);
 461         }
 462
 463         /* Do actual allocation. */
 464         down_write(&lod->lod_qos.lq_rw_sem);
 465
 466         /*
 467          * Check again. While we were sleeping on @lq_rw_sem something could
 468          * change.
 469          */
 470         if (!lqr->lqr_dirty) {
 471                 LASSERT(lqr->lqr_pool.op_size);
 472                 up_write(&lod->lod_qos.lq_rw_sem);
 473                 RETURN(0);
 474         }
 475
 476         real_count = src_pool->op_count;
 477
 478         /* Zero the pool array */
 479         /* alloc_rr is holding a read lock on the pool, so nobody is adding/
 480            deleting from the pool. The lq_rw_sem insures that nobody else
 481            is reading. */
 482         lqr->lqr_pool.op_count = real_count;
 483         rc = lod_ost_pool_extend(&lqr->lqr_pool, real_count);
 484         if (rc) {
 485                 up_write(&lod->lod_qos.lq_rw_sem);
 486                 RETURN(rc);
 487         }
 488         for (i = 0; i < lqr->lqr_pool.op_count; i++)
 489                 lqr->lqr_pool.op_array[i] = LOV_QOS_EMPTY;
 490
 491         /* Place all the OSTs from 1 OSS at the same time. */
 492         placed = 0;
 493         cfs_list_for_each_entry(oss, &lod->lod_qos.lq_oss_list, lqo_oss_list) {
 494                 int j = 0;
 495
 496                 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
 497                         int next;
 498
 499                         if (!cfs_bitmap_check(lod->lod_ost_bitmap,
 500                                                 src_pool->op_array[i]))
 501                                 continue;
 502
 503                         ost = OST_TGT(lod,src_pool->op_array[i]);
 504                         LASSERT(ost && ost->ltd_ost);
 505                         if (ost->ltd_qos.ltq_oss != oss)
 506                                 continue;
 507
 508                         /* Evenly space these OSTs across arrayspace */
 509                         next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count;
 510                         while (lqr->lqr_pool.op_array[next] != LOV_QOS_EMPTY)
 511                                 next = (next + 1) % lqr->lqr_pool.op_count;
 512
 513                         lqr->lqr_pool.op_array[next] = src_pool->op_array[i];
 514                         j++;
 515                         placed++;
 516                 }
 517         }
 518
 519         lqr->lqr_dirty = 0;
 520         up_write(&lod->lod_qos.lq_rw_sem);
 521
 522         if (placed != real_count) {
 523                 /* This should never happen */
 524                 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
 525                                    "round-robin list (%d of %d).\n",
 526                                    placed, real_count);
 527                 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
 528                         LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
 529                                  lqr->lqr_pool.op_array[i]);
 530                 }
 531                 lqr->lqr_dirty = 1;
 532                 RETURN(-EAGAIN);
 533         }
 534
 535 #if 0
 536         for (i = 0; i < lqr->lqr_pool.op_count; i++)
 537                 QOS_CONSOLE("rr #%d ost idx=%d\n", i, lqr->lqr_pool.op_array[i]);
 538 #endif
 539
 540         RETURN(0);
 541 }
 542
 543 /**
 544  * A helper function to:
 545  *   create in-core lu object on the specified OSP
 546  *   declare creation of the object
 547  * IMPORTANT: at this stage object is anonymouos - it has no fid assigned
 548  *            this is a workaround till we have natural FIDs on OST
 549  *
 550  *            at this point we want to declare (reserve) object for us as
 551  *            we can't block at execution (when create method is called).
 552  *            otherwise we'd block whole transaction batch
 553  */
 554 static struct dt_object *lod_qos_declare_object_on(const struct lu_env *env,
 555                                                    struct lod_device *d,
 556                                                    int ost_idx,
 557                                                    struct thandle *th)
 558 {
 559         struct lod_tgt_desc *ost;
 560         struct lu_object *o, *n;
 561         struct lu_device *nd;
 562         struct dt_object *dt;
 563         int               rc;
 564         ENTRY;
 565
 566         LASSERT(d);
 567         LASSERT(ost_idx >= 0);
 568         LASSERT(ost_idx < d->lod_osts_size);
 569         ost = OST_TGT(d,ost_idx);
 570         LASSERT(ost);
 571         LASSERT(ost->ltd_ost);
 572
 573         nd = &ost->ltd_ost->dd_lu_dev;
 574
 575         /*
 576          * allocate anonymous object with zero fid, real fid
 577          * will be assigned by OSP within transaction
 578          * XXX: to be fixed with fully-functional OST fids
 579          */
 580         o = lu_object_anon(env, nd, NULL);
 581         if (IS_ERR(o))
 582                 GOTO(out, dt = ERR_PTR(PTR_ERR(o)));
 583
 584         n = lu_object_locate(o->lo_header, nd->ld_type);
 585         if (unlikely(n == NULL)) {
 586                 CERROR("can't find slice\n");
 587                 lu_object_put(env, o);
 588                 GOTO(out, dt = ERR_PTR(-EINVAL));
 589         }
 590
 591         dt = container_of(n, struct dt_object, do_lu);
 592
 593         rc = dt_declare_create(env, dt, NULL, NULL, NULL, th);
 594         if (rc) {
 595                 CDEBUG(D_OTHER, "can't declare creation on #%u: %d\n",
 596                        ost_idx, rc);
 597                 lu_object_put(env, o);
 598                 dt = ERR_PTR(rc);
 599         }
 600
 601 out:
 602         RETURN(dt);
 603 }
 604
 605 static int min_stripe_count(int stripe_cnt, int flags)
 606 {
 607         return (flags & LOV_USES_DEFAULT_STRIPE ?
 608                         stripe_cnt - (stripe_cnt / 4) : stripe_cnt);
 609 }
 610
 611 #define LOV_CREATE_RESEED_MULT 30
 612 #define LOV_CREATE_RESEED_MIN  2000
 613
 614 static int inline lod_qos_dev_is_full(struct obd_statfs *msfs)
 615 {
 616         __u64 used;
 617         int   bs = msfs->os_bsize;
 618
 619         LASSERT(((bs - 1) & bs) == 0);
 620
 621         /* the minimum of 0.1% used blocks and 1GB bytes. */
 622         used = min_t(__u64, (msfs->os_blocks - msfs->os_bfree) >> 10,
 623                         1 << (31 - ffs(bs)));
 624         return (msfs->os_bavail < used);
 625 }
 626
 627 int lod_ea_store_resize(struct lod_thread_info *info, int size);
 628
 629 static inline int lod_qos_ost_in_use_clear(const struct lu_env *env, int stripes)
 630 {
 631         struct lod_thread_info *info = lod_env_info(env);
 632
 633         if (info->lti_ea_store_size < sizeof(int) * stripes)
 634                 lod_ea_store_resize(info, stripes * sizeof(int));
 635         if (info->lti_ea_store_size < sizeof(int) * stripes) {
 636                 CERROR("can't allocate memory for ost-in-use array\n");
 637                 return -ENOMEM;
 638         }
 639         memset(info->lti_ea_store, -1, sizeof(int) * stripes);
 640         return 0;
 641 }
 642
 643 static inline void lod_qos_ost_in_use(const struct lu_env *env, int idx, int ost)
 644 {
 645         struct lod_thread_info *info = lod_env_info(env);
 646         int *osts = info->lti_ea_store;
 647
 648         LASSERT(info->lti_ea_store_size >= idx * sizeof(int));
 649         osts[idx] = ost;
 650 }
 651
 652 static int lod_qos_is_ost_used(const struct lu_env *env, int ost, int stripes)
 653 {
 654         struct lod_thread_info *info = lod_env_info(env);
 655         int *osts = info->lti_ea_store;
 656         int j;
 657
 658         for (j = 0; j < stripes; j++) {
 659                 if (osts[j] == ost)
 660                         return 1;
 661         }
 662         return 0;
 663 }
 664
 665 /* Allocate objects on OSTs with round-robin algorithm */
 666 static int lod_alloc_rr(const struct lu_env *env, struct lod_object *lo,
 667                         struct dt_object **stripe, int flags,
 668                         struct thandle *th)
 669 {
 670         struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
 671         struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
 672         struct pool_desc  *pool = NULL;
 673         struct ost_pool   *osts;
 674         struct lod_qos_rr *lqr;
 675         struct dt_object  *o;
 676         unsigned           array_idx;
 677         int                i, rc;
 678         int                ost_start_idx_temp;
 679         int                speed = 0;
 680         int                stripe_idx = 0;
 681         int                stripe_cnt = lo->ldo_stripenr;
 682         int                stripe_cnt_min = min_stripe_count(stripe_cnt, flags);
 683         __u32              ost_idx;
 684         ENTRY;
 685
 686         if (lo->ldo_pool)
 687                 pool = lod_find_pool(m, lo->ldo_pool);
 688
 689         if (pool != NULL) {
 690                 down_read(&pool_tgt_rw_sem(pool));
 691                 osts = &(pool->pool_obds);
 692                 lqr = &(pool->pool_rr);
 693         } else {
 694                 osts = &(m->lod_pool_info);
 695                 lqr = &(m->lod_qos.lq_rr);
 696         }
 697
 698         rc = lod_qos_calc_rr(m, osts, lqr);
 699         if (rc)
 700                 GOTO(out, rc);
 701
 702         rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
 703         if (rc)
 704                 GOTO(out, rc);
 705
 706         if (--lqr->lqr_start_count <= 0) {
 707                 lqr->lqr_start_idx = cfs_rand() % osts->op_count;
 708                 lqr->lqr_start_count =
 709                         (LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) +
 710                          LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U);
 711         } else if (stripe_cnt_min >= osts->op_count ||
 712                         lqr->lqr_start_idx > osts->op_count) {
 713                 /* If we have allocated from all of the OSTs, slowly
 714                  * precess the next start if the OST/stripe count isn't
 715                  * already doing this for us. */
 716                 lqr->lqr_start_idx %= osts->op_count;
 717                 if (stripe_cnt > 1 && (osts->op_count % stripe_cnt) != 1)
 718                         ++lqr->lqr_offset_idx;
 719         }
 720         down_read(&m->lod_qos.lq_rw_sem);
 721         ost_start_idx_temp = lqr->lqr_start_idx;
 722
 723 repeat_find:
 724         array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) %
 725                         osts->op_count;
 726
 727         QOS_DEBUG("pool '%s' want %d startidx %d startcnt %d offset %d "
 728                   "active %d count %d arrayidx %d\n",
 729                   lo->ldo_pool ? lo->ldo_pool : "",
 730                   stripe_cnt, lqr->lqr_start_idx, lqr->lqr_start_count,
 731                   lqr->lqr_offset_idx, osts->op_count, osts->op_count,
 732                   array_idx);
 733
 734         for (i = 0; i < osts->op_count && stripe_idx < lo->ldo_stripenr;
 735              i++, array_idx = (array_idx + 1) % osts->op_count) {
 736                 ++lqr->lqr_start_idx;
 737                 ost_idx = lqr->lqr_pool.op_array[array_idx];
 738
 739                 QOS_DEBUG("#%d strt %d act %d strp %d ary %d idx %d\n",
 740                           i, lqr->lqr_start_idx, /* XXX: active*/ 0,
 741                           stripe_idx, array_idx, ost_idx);
 742
 743                 if ((ost_idx == LOV_QOS_EMPTY) ||
 744                     !cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
 745                         continue;
 746
 747                 /* Fail Check before osc_precreate() is called
 748                    so we can only 'fail' single OSC. */
 749                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
 750                         continue;
 751
 752                 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
 753                 if (rc) {
 754                         /* this OSP doesn't feel well */
 755                         continue;
 756                 }
 757
 758                 /*
 759                  * skip full devices
 760                  */
 761                 if (lod_qos_dev_is_full(sfs)) {
 762                         QOS_DEBUG("#%d is full\n", ost_idx);
 763                         continue;
 764                 }
 765
 766                 /*
 767                  * We expect number of precreated objects in f_ffree at
 768                  * the first iteration, skip OSPs with no objects ready
 769                  */
 770                 if (sfs->os_fprecreated == 0 && speed == 0) {
 771                         QOS_DEBUG("#%d: precreation is empty\n", ost_idx);
 772                         continue;
 773                 }
 774
 775                 /*
 776                  * try to use another OSP if this one is degraded
 777                  */
 778                 if (sfs->os_state & OS_STATE_DEGRADED && speed < 2) {
 779                         QOS_DEBUG("#%d: degraded\n", ost_idx);
 780                         continue;
 781                 }
 782
 783                 /*
 784                  * do not put >1 objects on a single OST
 785                  */
 786                 if (speed && lod_qos_is_ost_used(env, ost_idx, stripe_idx))
 787                         continue;
 788
 789                 o = lod_qos_declare_object_on(env, m, ost_idx, th);
 790                 if (IS_ERR(o)) {
 791                         CDEBUG(D_OTHER, "can't declare new object on #%u: %d\n",
 792                                ost_idx, (int) PTR_ERR(o));
 793                         rc = PTR_ERR(o);
 794                         continue;
 795                 }
 796
 797                 /*
 798                  * We've successfuly declared (reserved) an object
 799                  */
 800                 lod_qos_ost_in_use(env, stripe_idx, ost_idx);
 801                 stripe[stripe_idx] = o;
 802                 stripe_idx++;
 803
 804         }
 805         if ((speed < 2) && (stripe_idx < stripe_cnt_min)) {
 806                 /* Try again, allowing slower OSCs */
 807                 speed++;
 808                 lqr->lqr_start_idx = ost_start_idx_temp;
 809                 goto repeat_find;
 810         }
 811
 812         up_read(&m->lod_qos.lq_rw_sem);
 813
 814         if (stripe_idx) {
 815                 lo->ldo_stripenr = stripe_idx;
 816                 /* at least one stripe is allocated */
 817                 rc = 0;
 818         } else {
 819                 /* nobody provided us with a single object */
 820                 rc = -ENOSPC;
 821         }
 822
 823 out:
 824         if (pool != NULL) {
 825                 up_read(&pool_tgt_rw_sem(pool));
 826                 /* put back ref got by lod_find_pool() */
 827                 lod_pool_putref(pool);
 828         }
 829
 830         RETURN(rc);
 831 }
 832
 833 /* alloc objects on osts with specific stripe offset */
 834 static int lod_alloc_specific(const struct lu_env *env, struct lod_object *lo,
 835                               struct dt_object **stripe, int flags,
 836                               struct thandle *th)
 837 {
 838         struct lod_device *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
 839         struct obd_statfs *sfs = &lod_env_info(env)->lti_osfs;
 840         struct dt_object  *o;
 841         unsigned           ost_idx, array_idx, ost_count;
 842         int                i, rc, stripe_num = 0;
 843         int                speed = 0;
 844         struct pool_desc  *pool = NULL;
 845         struct ost_pool   *osts;
 846         ENTRY;
 847
 848         rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
 849         if (rc)
 850                 GOTO(out, rc);
 851
 852         if (lo->ldo_pool)
 853                 pool = lod_find_pool(m, lo->ldo_pool);
 854
 855         if (pool != NULL) {
 856                 down_read(&pool_tgt_rw_sem(pool));
 857                 osts = &(pool->pool_obds);
 858         } else {
 859                 osts = &(m->lod_pool_info);
 860         }
 861
 862         ost_count = osts->op_count;
 863
 864 repeat_find:
 865         /* search loi_ost_idx in ost array */
 866         array_idx = 0;
 867         for (i = 0; i < ost_count; i++) {
 868                 if (osts->op_array[i] == lo->ldo_def_stripe_offset) {
 869                         array_idx = i;
 870                         break;
 871                 }
 872         }
 873         if (i == ost_count) {
 874                 CERROR("Start index %d not found in pool '%s'\n",
 875                        lo->ldo_def_stripe_offset,
 876                        lo->ldo_pool ? lo->ldo_pool : "");
 877                 GOTO(out, rc = -EINVAL);
 878         }
 879
 880         for (i = 0; i < ost_count;
 881                         i++, array_idx = (array_idx + 1) % ost_count) {
 882                 ost_idx = osts->op_array[array_idx];
 883
 884                 if (!cfs_bitmap_check(m->lod_ost_bitmap, ost_idx))
 885                         continue;
 886
 887                 /* Fail Check before osc_precreate() is called
 888                    so we can only 'fail' single OSC. */
 889                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
 890                         continue;
 891
 892                 /*
 893                  * do not put >1 objects on a single OST
 894                  */
 895                 if (lod_qos_is_ost_used(env, ost_idx, stripe_num))
 896                         continue;
 897
 898                 /* Drop slow OSCs if we can, but not for requested start idx.
 899                  *
 900                  * This means "if OSC is slow and it is not the requested
 901                  * start OST, then it can be skipped, otherwise skip it only
 902                  * if it is inactive/recovering/out-of-space." */
 903
 904                 rc = lod_statfs_and_check(env, m, ost_idx, sfs);
 905                 if (rc) {
 906                         /* this OSP doesn't feel well */
 907                         continue;
 908                 }
 909
 910                 /*
 911                  * We expect number of precreated objects in f_ffree at
 912                  * the first iteration, skip OSPs with no objects ready
 913                  * don't apply this logic to OST specified with stripe_offset
 914                  */
 915                 if (i != 0 && sfs->os_fprecreated == 0 && speed == 0)
 916                         continue;
 917
 918                 o = lod_qos_declare_object_on(env, m, ost_idx, th);
 919                 if (IS_ERR(o)) {
 920                         CDEBUG(D_OTHER, "can't declare new object on #%u: %d\n",
 921                                ost_idx, (int) PTR_ERR(o));
 922                         continue;
 923                 }
 924
 925                 /*
 926                  * We've successfuly declared (reserved) an object
 927                  */
 928                 lod_qos_ost_in_use(env, stripe_num, ost_idx);
 929                 stripe[stripe_num] = o;
 930                 stripe_num++;
 931
 932                 /* We have enough stripes */
 933                 if (stripe_num == lo->ldo_stripenr)
 934                         GOTO(out, rc = 0);
 935         }
 936         if (speed < 2) {
 937                 /* Try again, allowing slower OSCs */
 938                 speed++;
 939                 goto repeat_find;
 940         }
 941
 942         /* If we were passed specific striping params, then a failure to
 943          * meet those requirements is an error, since we can't reallocate
 944          * that memory (it might be part of a larger array or something).
 945          *
 946          * We can only get here if lsm_stripe_count was originally > 1.
 947          */
 948         CERROR("can't lstripe objid "DFID": have %d want %u\n",
 949                PFID(lu_object_fid(lod2lu_obj(lo))), stripe_num,
 950                lo->ldo_stripenr);
 951         rc = -EFBIG;
 952 out:
 953         if (pool != NULL) {
 954                 up_read(&pool_tgt_rw_sem(pool));
 955                 /* put back ref got by lod_find_pool() */
 956                 lod_pool_putref(pool);
 957         }
 958
 959         RETURN(rc);
 960 }
 961
 962 static inline int lod_qos_is_usable(struct lod_device *lod)
 963 {
 964 #ifdef FORCE_QOS
 965         /* to be able to debug QoS code */
 966         return 1;
 967 #endif
 968
 969         /* Detect -EAGAIN early, before expensive lock is taken. */
 970         if (!lod->lod_qos.lq_dirty && lod->lod_qos.lq_same_space)
 971                 return 0;
 972
 973         if (lod->lod_desc.ld_active_tgt_count < 2)
 974                 return 0;
 975
 976         return 1;
 977 }
 978
 979 /* Alloc objects on OSTs with optimization based on:
 980    - free space
 981    - network resources (shared OSS's)
 982  */
 983 static int lod_alloc_qos(const struct lu_env *env, struct lod_object *lo,
 984                          struct dt_object **stripe, int flags,
 985                          struct thandle *th)
 986 {
 987         struct lod_device   *m = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
 988         struct obd_statfs   *sfs = &lod_env_info(env)->lti_osfs;
 989         struct lod_tgt_desc *ost;
 990         struct dt_object    *o;
 991         __u64                total_weight = 0;
 992         int                  nfound, good_osts, i, rc = 0;
 993         int                  stripe_cnt = lo->ldo_stripenr;
 994         int                  stripe_cnt_min;
 995         struct pool_desc    *pool = NULL;
 996         struct ost_pool    *osts;
 997         ENTRY;
 998
 999         stripe_cnt_min = min_stripe_count(stripe_cnt, flags);
1000         if (stripe_cnt_min < 1)
1001                 RETURN(-EINVAL);
1002
1003         if (lo->ldo_pool)
1004                 pool = lod_find_pool(m, lo->ldo_pool);
1005
1006         if (pool != NULL) {
1007                 down_read(&pool_tgt_rw_sem(pool));
1008                 osts = &(pool->pool_obds);
1009         } else {
1010                 osts = &(m->lod_pool_info);
1011         }
1012
1013         /* Detect -EAGAIN early, before expensive lock is taken. */
1014         if (!lod_qos_is_usable(m))
1015                 GOTO(out_nolock, rc = -EAGAIN);
1016
1017         /* Do actual allocation, use write lock here. */
1018         down_write(&m->lod_qos.lq_rw_sem);
1019
1020         /*
1021          * Check again, while we were sleeping on @lq_rw_sem things could
1022          * change.
1023          */
1024         if (!lod_qos_is_usable(m))
1025                 GOTO(out, rc = -EAGAIN);
1026
1027         rc = lod_qos_calc_ppo(m);
1028         if (rc)
1029                 GOTO(out, rc);
1030
1031         rc = lod_qos_ost_in_use_clear(env, lo->ldo_stripenr);
1032         if (rc)
1033                 GOTO(out, rc);
1034
1035         good_osts = 0;
1036         /* Find all the OSTs that are valid stripe candidates */
1037         for (i = 0; i < osts->op_count; i++) {
1038                 if (!cfs_bitmap_check(m->lod_ost_bitmap, osts->op_array[i]))
1039                         continue;
1040
1041                 rc = lod_statfs_and_check(env, m, osts->op_array[i], sfs);
1042                 if (rc) {
1043                         /* this OSP doesn't feel well */
1044                         continue;
1045                 }
1046
1047                 /*
1048                  * skip full devices
1049                  */
1050                 if (lod_qos_dev_is_full(sfs))
1051                         continue;
1052
1053                 /* Fail Check before osc_precreate() is called
1054                    so we can only 'fail' single OSC. */
1055                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) &&
1056                                    osts->op_array[i] == 0)
1057                         continue;
1058
1059                 ost = OST_TGT(m,osts->op_array[i]);
1060                 ost->ltd_qos.ltq_usable = 1;
1061                 lod_qos_calc_weight(m, osts->op_array[i]);
1062                 total_weight += ost->ltd_qos.ltq_weight;
1063
1064                 good_osts++;
1065         }
1066
1067         QOS_DEBUG("found %d good osts\n", good_osts);
1068
1069         if (good_osts < stripe_cnt_min)
1070                 GOTO(out, rc = -EAGAIN);
1071
1072         /* We have enough osts */
1073         if (good_osts < stripe_cnt)
1074                 stripe_cnt = good_osts;
1075
1076         /* Find enough OSTs with weighted random allocation. */
1077         nfound = 0;
1078         while (nfound < stripe_cnt) {
1079                 __u64 rand, cur_weight;
1080
1081                 cur_weight = 0;
1082                 rc = -ENOSPC;
1083
1084                 if (total_weight) {
1085 #if BITS_PER_LONG == 32
1086                         rand = cfs_rand() % (unsigned)total_weight;
1087                         /* If total_weight > 32-bit, first generate the high
1088                          * 32 bits of the random number, then add in the low
1089                          * 32 bits (truncated to the upper limit, if needed) */
1090                         if (total_weight > 0xffffffffULL)
1091                                 rand = (__u64)(cfs_rand() %
1092                                         (unsigned)(total_weight >> 32)) << 32;
1093                         else
1094                                 rand = 0;
1095
1096                         if (rand == (total_weight & 0xffffffff00000000ULL))
1097                                 rand |= cfs_rand() % (unsigned)total_weight;
1098                         else
1099                                 rand |= cfs_rand();
1100
1101 #else
1102                         rand = ((__u64)cfs_rand() << 32 | cfs_rand()) %
1103                                 total_weight;
1104 #endif
1105                 } else {
1106                         rand = 0;
1107                 }
1108
1109                 /* On average, this will hit larger-weighted osts more often.
1110                    0-weight osts will always get used last (only when rand=0) */
1111                 for (i = 0; i < osts->op_count; i++) {
1112                         int idx = osts->op_array[i];
1113
1114                         if (!cfs_bitmap_check(m->lod_ost_bitmap, idx))
1115                                 continue;
1116
1117                         ost = OST_TGT(m,idx);
1118
1119                         if (!ost->ltd_qos.ltq_usable)
1120                                 continue;
1121
1122                         cur_weight += ost->ltd_qos.ltq_weight;
1123                         QOS_DEBUG("stripe_cnt=%d nfound=%d cur_weight="LPU64
1124                                   " rand="LPU64" total_weight="LPU64"\n",
1125                                   stripe_cnt, nfound, cur_weight, rand,
1126                                   total_weight);
1127
1128                         if (cur_weight < rand)
1129                                 continue;
1130
1131                         QOS_DEBUG("stripe=%d to idx=%d\n", nfound, idx);
1132
1133                         /*
1134                          * do not put >1 objects on a single OST
1135                          */
1136                         if (lod_qos_is_ost_used(env, idx, nfound))
1137                                 continue;
1138                         lod_qos_ost_in_use(env, nfound, idx);
1139
1140                         o = lod_qos_declare_object_on(env, m, idx, th);
1141                         if (IS_ERR(o)) {
1142                                 QOS_DEBUG("can't declare object on #%u: %d\n",
1143                                           idx, (int) PTR_ERR(o));
1144                                 continue;
1145                         }
1146                         stripe[nfound++] = o;
1147                         lod_qos_used(m, osts, idx, &total_weight);
1148                         rc = 0;
1149                         break;
1150                 }
1151
1152                 if (rc) {
1153                         /* no OST found on this iteration, give up */
1154                         break;
1155                 }
1156         }
1157
1158         if (unlikely(nfound != stripe_cnt)) {
1159                 /*
1160                  * when the decision to use weighted algorithm was made
1161                  * we had enough appropriate OSPs, but this state can
1162                  * change anytime (no space on OST, broken connection, etc)
1163                  * so it's possible OSP won't be able to provide us with
1164                  * an object due to just changed state
1165                  */
1166                 LCONSOLE_INFO("wanted %d, found %d\n", stripe_cnt, nfound);
1167                 for (i = 0; i < nfound; i++) {
1168                         LASSERT(stripe[i] != NULL);
1169                         lu_object_put(env, &stripe[i]->do_lu);
1170                         stripe[i] = NULL;
1171                 }
1172
1173                 /* makes sense to rebalance next time */
1174                 m->lod_qos.lq_dirty = 1;
1175                 m->lod_qos.lq_same_space = 0;
1176
1177                 rc = -EAGAIN;
1178         }
1179
1180 out:
1181         up_write(&m->lod_qos.lq_rw_sem);
1182
1183 out_nolock:
1184         if (pool != NULL) {
1185                 up_read(&pool_tgt_rw_sem(pool));
1186                 /* put back ref got by lod_find_pool() */
1187                 lod_pool_putref(pool);
1188         }
1189
1190         RETURN(rc);
1191 }
1192
1193 /* Find the max stripecount we should use */
1194 static __u16 lod_get_stripecnt(struct lod_device *lod, __u32 magic,
1195                                __u16 stripe_count)
1196 {
1197         __u32 max_stripes = LOV_MAX_STRIPE_COUNT_OLD;
1198
1199         if (!stripe_count)
1200                 stripe_count = lod->lod_desc.ld_default_stripe_count;
1201         if (stripe_count > lod->lod_desc.ld_active_tgt_count)
1202                 stripe_count = lod->lod_desc.ld_active_tgt_count;
1203         if (!stripe_count)
1204                 stripe_count = 1;
1205
1206         /* stripe count is based on whether OSD can handle larger EA sizes */
1207         if (lod->lod_osd_max_easize > 0)
1208                 max_stripes = lov_mds_md_stripecnt(lod->lod_osd_max_easize,
1209                                                    magic);
1210
1211         return (stripe_count < max_stripes) ? stripe_count : max_stripes;
1212 }
1213
1214 static int lod_use_defined_striping(const struct lu_env *env,
1215                                     struct lod_object *mo,
1216                                     const struct lu_buf *buf)
1217 {
1218         struct lov_mds_md_v1   *v1 = buf->lb_buf;
1219         struct lov_mds_md_v3   *v3 = buf->lb_buf;
1220         struct lov_ost_data_v1 *objs;
1221         __u32                   magic;
1222         int                     rc = 0;
1223         ENTRY;
1224
1225         magic = le32_to_cpu(v1->lmm_magic);
1226         if (magic == LOV_MAGIC_V1_DEF) {
1227                 magic = LOV_MAGIC_V1;
1228                 objs = &v1->lmm_objects[0];
1229         } else if (magic == LOV_MAGIC_V3_DEF) {
1230                 magic = LOV_MAGIC_V3;
1231                 objs = &v3->lmm_objects[0];
1232                 lod_object_set_pool(mo, v3->lmm_pool_name);
1233         } else {
1234                 GOTO(out, rc = -EINVAL);
1235         }
1236
1237         mo->ldo_pattern = le32_to_cpu(v1->lmm_pattern);
1238         mo->ldo_stripe_size = le32_to_cpu(v1->lmm_stripe_size);
1239         mo->ldo_stripenr = le16_to_cpu(v1->lmm_stripe_count);
1240         mo->ldo_layout_gen = le16_to_cpu(v1->lmm_layout_gen);
1241
1242         /* fixup for released file before object initialization */
1243         if (mo->ldo_pattern & LOV_PATTERN_F_RELEASED) {
1244                 mo->ldo_released_stripenr = mo->ldo_stripenr;
1245                 mo->ldo_stripenr = 0;
1246         }
1247
1248         LASSERT(buf->lb_len >= lov_mds_md_size(mo->ldo_stripenr, magic));
1249
1250         if (mo->ldo_stripenr > 0)
1251                 rc = lod_initialize_objects(env, mo, objs);
1252
1253 out:
1254         RETURN(rc);
1255 }
1256
1257 static int lod_qos_parse_config(const struct lu_env *env,
1258                                 struct lod_object *lo,
1259                                 const struct lu_buf *buf)
1260 {
1261         struct lod_device     *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
1262         struct lov_user_md_v1 *v1 = NULL;
1263         struct lov_user_md_v3 *v3 = NULL;
1264         struct pool_desc      *pool;
1265         __u32                  magic;
1266         int                    rc;
1267         ENTRY;
1268
1269         if (buf == NULL || buf->lb_buf == NULL || buf->lb_len == 0)
1270                 RETURN(0);
1271
1272         v1 = buf->lb_buf;
1273         magic = v1->lmm_magic;
1274
1275         if (magic == __swab32(LOV_USER_MAGIC_V1)) {
1276                 lustre_swab_lov_user_md_v1(v1);
1277                 magic = v1->lmm_magic;
1278         } else if (magic == __swab32(LOV_USER_MAGIC_V3)) {
1279                 v3 = buf->lb_buf;
1280                 lustre_swab_lov_user_md_v3(v3);
1281                 magic = v3->lmm_magic;
1282         }
1283
1284         if (unlikely(magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3)) {
1285                 /* try to use as fully defined striping */
1286                 rc = lod_use_defined_striping(env, lo, buf);
1287                 RETURN(rc);
1288         }
1289
1290         if (unlikely(buf->lb_len < sizeof(*v1))) {
1291                 CERROR("wrong size: %u\n", (unsigned) buf->lb_len);
1292                 RETURN(-EINVAL);
1293         }
1294
1295         v1->lmm_magic = magic;
1296         if (v1->lmm_pattern == 0)
1297                 v1->lmm_pattern = LOV_PATTERN_RAID0;
1298         if (lov_pattern(v1->lmm_pattern) != LOV_PATTERN_RAID0) {
1299                 CERROR("invalid pattern: %x\n", v1->lmm_pattern);
1300                 RETURN(-EINVAL);
1301         }
1302         lo->ldo_pattern = v1->lmm_pattern;
1303
1304         if (v1->lmm_stripe_size)
1305                 lo->ldo_stripe_size = v1->lmm_stripe_size;
1306         if (lo->ldo_stripe_size & (LOV_MIN_STRIPE_SIZE - 1))
1307                 lo->ldo_stripe_size = LOV_MIN_STRIPE_SIZE;
1308
1309         if (v1->lmm_stripe_count)
1310                 lo->ldo_stripenr = v1->lmm_stripe_count;
1311
1312         if ((v1->lmm_stripe_offset >= d->lod_desc.ld_tgt_count) &&
1313             (v1->lmm_stripe_offset != (typeof(v1->lmm_stripe_offset))(-1))) {
1314                 CERROR("invalid offset: %x\n", v1->lmm_stripe_offset);
1315                 RETURN(-EINVAL);
1316         }
1317         lo->ldo_def_stripe_offset = v1->lmm_stripe_offset;
1318
1319         CDEBUG(D_OTHER, "lsm: %u size, %u stripes, %u offset\n",
1320                v1->lmm_stripe_size, v1->lmm_stripe_count,
1321                v1->lmm_stripe_offset);
1322
1323         if (v1->lmm_magic == LOV_MAGIC_V3) {
1324                 if (buf->lb_len < sizeof(*v3)) {
1325                         CERROR("wrong size: %u\n", (unsigned) buf->lb_len);
1326                         RETURN(-EINVAL);
1327                 }
1328
1329                 v3 = buf->lb_buf;
1330                 lod_object_set_pool(lo, v3->lmm_pool_name);
1331
1332                 /* In the function below, .hs_keycmp resolves to
1333                  * pool_hashkey_keycmp() */
1334                 /* coverity[overrun-buffer-val] */
1335                 pool = lod_find_pool(d, v3->lmm_pool_name);
1336                 if (pool != NULL) {
1337                         if (lo->ldo_def_stripe_offset !=
1338                             (typeof(v1->lmm_stripe_offset))(-1)) {
1339                                 rc = lo->ldo_def_stripe_offset;
1340                                 rc = lod_check_index_in_pool(rc, pool);
1341                                 if (rc < 0) {
1342                                         lod_pool_putref(pool);
1343                                         CERROR("invalid offset\n");
1344                                         RETURN(-EINVAL);
1345                                 }
1346                         }
1347
1348                         if (lo->ldo_stripenr > pool_tgt_count(pool))
1349                                 lo->ldo_stripenr= pool_tgt_count(pool);
1350
1351                         lod_pool_putref(pool);
1352                 }
1353         } else
1354                 lod_object_set_pool(lo, NULL);
1355
1356         /* fixup for released file */
1357         if (lo->ldo_pattern & LOV_PATTERN_F_RELEASED) {
1358                 lo->ldo_released_stripenr = lo->ldo_stripenr;
1359                 lo->ldo_stripenr = 0;
1360         }
1361
1362         RETURN(0);
1363 }
1364
1365 /*
1366  * buf should be NULL or contain striping settings
1367  */
1368 int lod_qos_prep_create(const struct lu_env *env, struct lod_object *lo,
1369                         struct lu_attr *attr, const struct lu_buf *buf,
1370                         struct thandle *th)
1371 {
1372         struct lod_device      *d = lu2lod_dev(lod2lu_obj(lo)->lo_dev);
1373         struct dt_object      **stripe;
1374         int                     stripe_len;
1375         int                     flag = LOV_USES_ASSIGNED_STRIPE;
1376         int                     i, rc;
1377         ENTRY;
1378
1379         LASSERT(lo);
1380
1381         /* no OST available */
1382         /* XXX: should we be waiting a bit to prevent failures during
1383          * cluster initialization? */
1384         if (d->lod_ostnr == 0)
1385                 GOTO(out, rc = -EIO);
1386
1387         /*
1388          * by this time, the object's ldo_stripenr and ldo_stripe_size
1389          * contain default value for striping: taken from the parent
1390          * or from filesystem defaults
1391          *
1392          * in case the caller is passing lovea with new striping config,
1393          * we may need to parse lovea and apply new configuration
1394          */
1395         rc = lod_qos_parse_config(env, lo, buf);
1396         if (rc)
1397                 GOTO(out, rc);
1398
1399         /* A released file is being created */
1400         if (lo->ldo_stripenr == 0)
1401                 GOTO(out, rc = 0);
1402
1403         if (likely(lo->ldo_stripe == NULL)) {
1404                 /*
1405                  * no striping has been created so far
1406                  */
1407                 LASSERT(lo->ldo_stripenr > 0);
1408                 /*
1409                  * statfs and check OST targets now, since ld_active_tgt_count
1410                  * could be changed if some OSTs are [de]activated manually.
1411                  */
1412                 lod_qos_statfs_update(env, d);
1413                 lo->ldo_stripenr = lod_get_stripecnt(d, LOV_MAGIC,
1414                                 lo->ldo_stripenr);
1415
1416                 stripe_len = lo->ldo_stripenr;
1417                 OBD_ALLOC(stripe, sizeof(stripe[0]) * stripe_len);
1418                 if (stripe == NULL)
1419                         GOTO(out, rc = -ENOMEM);
1420
1421                 lod_getref(&d->lod_ost_descs);
1422                 /* XXX: support for non-0 files w/o objects */
1423                 CDEBUG(D_OTHER, "tgt_count %d stripenr %d\n",
1424                                 d->lod_desc.ld_tgt_count, stripe_len);
1425                 if (lo->ldo_def_stripe_offset >= d->lod_desc.ld_tgt_count) {
1426                         rc = lod_alloc_qos(env, lo, stripe, flag, th);
1427                         if (rc == -EAGAIN)
1428                                 rc = lod_alloc_rr(env, lo, stripe, flag, th);
1429                 } else {
1430                         rc = lod_alloc_specific(env, lo, stripe, flag, th);
1431                 }
1432                 lod_putref(d, &d->lod_ost_descs);
1433
1434                 if (rc < 0) {
1435                         for (i = 0; i < stripe_len; i++)
1436                                 if (stripe[i] != NULL)
1437                                         lu_object_put(env, &stripe[i]->do_lu);
1438
1439                         OBD_FREE(stripe, sizeof(stripe[0]) * stripe_len);
1440                 } else {
1441                         lo->ldo_stripe = stripe;
1442                         lo->ldo_stripes_allocated = stripe_len;
1443                 }
1444         } else {
1445                 /*
1446                  * lod_qos_parse_config() found supplied buf as a predefined
1447                  * striping (not a hint), so it allocated all the object
1448                  * now we need to create them
1449                  */
1450                 for (i = 0; i < lo->ldo_stripenr; i++) {
1451                         struct dt_object  *o;
1452
1453                         o = lo->ldo_stripe[i];
1454                         LASSERT(o);
1455
1456                         rc = dt_declare_create(env, o, attr, NULL, NULL, th);
1457                         if (rc) {
1458                                 CERROR("can't declare create: %d\n", rc);
1459                                 break;
1460                         }
1461                 }
1462         }
1463
1464 out:
1465         RETURN(rc);
1466 }
1467