lustre/lov/lov_qos.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #ifndef EXPORT_SYMTAB
  38 # define EXPORT_SYMTAB
  39 #endif
  40 #define DEBUG_SUBSYSTEM S_LOV
  41
  42 #ifdef __KERNEL__
  43 #include <libcfs/libcfs.h>
  44 #else
  45 #include <liblustre.h>
  46 #endif
  47
  48 #include <obd_class.h>
  49 #include <obd_lov.h>
  50 #include "lov_internal.h"
  51
  52 /* #define QOS_DEBUG 1 */
  53 #define D_QOS D_OTHER
  54
  55 #define TGT_BAVAIL(i)  (lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bavail*\
  56                         lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bsize)
  57 #define TGT_FFREE(i)   (lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_ffree)
  58
  59
  60 int qos_add_tgt(struct obd_device *obd, __u32 index)
  61 {
  62         struct lov_obd *lov = &obd->u.lov;
  63         struct lov_qos_oss *oss, *temposs;
  64         struct obd_export *exp = lov->lov_tgts[index]->ltd_exp;
  65         int rc = 0, found = 0;
  66         ENTRY;
  67
  68         /* We only need this QOS struct on MDT, not clients - but we may not
  69          * have registered the LOV's observer yet, so there's no way to know */
  70         if (!exp || !exp->exp_connection) {
  71                 CERROR("Missing connection\n");
  72                 RETURN(-ENOTCONN);
  73         }
  74
  75         down_write(&lov->lov_qos.lq_rw_sem);
  76         mutex_down(&lov->lov_lock);
  77         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
  78                 if (obd_uuid_equals(&oss->lqo_uuid,
  79                                     &exp->exp_connection->c_remote_uuid)) {
  80                         found++;
  81                         break;
  82                 }
  83         }
  84
  85         if (!found) {
  86                 OBD_ALLOC_PTR(oss);
  87                 if (!oss)
  88                         GOTO(out, rc = -ENOMEM);
  89                 memcpy(&oss->lqo_uuid,
  90                        &exp->exp_connection->c_remote_uuid,
  91                        sizeof(oss->lqo_uuid));
  92         } else {
  93                 /* Assume we have to move this one */
  94                 list_del(&oss->lqo_oss_list);
  95         }
  96
  97         oss->lqo_ost_count++;
  98         lov->lov_tgts[index]->ltd_qos.ltq_oss = oss;
  99
 100         /* Add sorted by # of OSTs.  Find the first entry that we're
 101            bigger than... */
 102         list_for_each_entry(temposs, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 103                 if (oss->lqo_ost_count > temposs->lqo_ost_count)
 104                         break;
 105         }
 106         /* ...and add before it.  If we're the first or smallest, temposs
 107            points to the list head, and we add to the end. */
 108         list_add_tail(&oss->lqo_oss_list, &temposs->lqo_oss_list);
 109
 110         lov->lov_qos.lq_dirty = 1;
 111         lov->lov_qos.lq_rr.lqr_dirty = 1;
 112
 113         CDEBUG(D_QOS, "add tgt %s to OSS %s (%d OSTs)\n",
 114                obd_uuid2str(&lov->lov_tgts[index]->ltd_uuid),
 115                obd_uuid2str(&oss->lqo_uuid),
 116                oss->lqo_ost_count);
 117
 118 out:
 119         mutex_up(&lov->lov_lock);
 120         up_write(&lov->lov_qos.lq_rw_sem);
 121         RETURN(rc);
 122 }
 123
 124 int qos_del_tgt(struct obd_device *obd, struct lov_tgt_desc *tgt)
 125 {
 126         struct lov_obd *lov = &obd->u.lov;
 127         struct lov_qos_oss *oss;
 128         int rc = 0;
 129         ENTRY;
 130
 131         down_write(&lov->lov_qos.lq_rw_sem);
 132
 133         oss = tgt->ltd_qos.ltq_oss;
 134         if (!oss)
 135                 GOTO(out, rc = -ENOENT);
 136
 137         oss->lqo_ost_count--;
 138         if (oss->lqo_ost_count == 0) {
 139                 CDEBUG(D_QOS, "removing OSS %s\n",
 140                        obd_uuid2str(&oss->lqo_uuid));
 141                 list_del(&oss->lqo_oss_list);
 142                 OBD_FREE_PTR(oss);
 143         }
 144
 145         lov->lov_qos.lq_dirty = 1;
 146         lov->lov_qos.lq_rr.lqr_dirty = 1;
 147 out:
 148         up_write(&lov->lov_qos.lq_rw_sem);
 149         RETURN(rc);
 150 }
 151
 152 /* Recalculate per-object penalties for OSSs and OSTs,
 153    depends on size of each ost in an oss */
 154 static int qos_calc_ppo(struct obd_device *obd)
 155 {
 156         struct lov_obd *lov = &obd->u.lov;
 157         struct lov_qos_oss *oss;
 158         __u64 ba_max, ba_min, temp;
 159         __u32 num_active;
 160         int rc, i, prio_wide;
 161         time_t now, age;
 162         ENTRY;
 163
 164         if (!lov->lov_qos.lq_dirty)
 165                 GOTO(out, rc = 0);
 166
 167         num_active = lov->desc.ld_active_tgt_count - 1;
 168         if (num_active < 1)
 169                 GOTO(out, rc = -EAGAIN);
 170
 171         /* find bavail on each OSS */
 172         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 173                 oss->lqo_bavail = 0;
 174         }
 175         lov->lov_qos.lq_active_oss_count = 0;
 176
 177         /* How badly user wants to select osts "widely" (not recently chosen
 178            and not on recent oss's).  As opposed to "freely" (free space
 179            avail.) 0-256. */
 180         prio_wide = 256 - lov->lov_qos.lq_prio_free;
 181
 182         ba_min = (__u64)(-1);
 183         ba_max = 0;
 184         now = cfs_time_current_sec();
 185         /* Calculate OST penalty per object */
 186         /* (lov ref taken in alloc_qos) */
 187         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 188                 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
 189                         continue;
 190                 temp = TGT_BAVAIL(i);
 191                 if (!temp)
 192                         continue;
 193                 ba_min = min(temp, ba_min);
 194                 ba_max = max(temp, ba_max);
 195
 196                 /* Count the number of usable OSS's */
 197                 if (lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_bavail == 0)
 198                         lov->lov_qos.lq_active_oss_count++;
 199                 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_bavail += temp;
 200
 201                 /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
 202                 temp >>= 1;
 203                 do_div(temp, num_active);
 204                 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj =
 205                         (temp * prio_wide) >> 8;
 206
 207                 age = (now - lov->lov_tgts[i]->ltd_qos.ltq_used) >> 3;
 208                 if (lov->lov_qos.lq_reset || age > 32 * lov->desc.ld_qos_maxage)
 209                         lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
 210                 else if (age > lov->desc.ld_qos_maxage)
 211                         /* Decay the penalty by half for every 8x the update
 212                          * interval that the device has been idle.  That gives
 213                          * lots of time for the statfs information to be
 214                          * updated (which the penalty is only a proxy for),
 215                          * and avoids penalizing OSS/OSTs under light load. */
 216                         lov->lov_tgts[i]->ltd_qos.ltq_penalty >>=
 217                                 (age / lov->desc.ld_qos_maxage);
 218         }
 219
 220         num_active = lov->lov_qos.lq_active_oss_count - 1;
 221         if (num_active < 1) {
 222                 /* If there's only 1 OSS, we can't penalize it, so instead
 223                    we have to double the OST penalty */
 224                 num_active = 1;
 225                 for (i = 0; i < lov->desc.ld_tgt_count; i++)
 226                         if (lov->lov_tgts[i])
 227                             lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj <<= 1;
 228         }
 229
 230         /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
 231         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 232                 temp = oss->lqo_bavail >> 1;
 233                 do_div(temp, oss->lqo_ost_count * num_active);
 234                 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
 235
 236                 age = (now - oss->lqo_used) >> 3;
 237                 if (lov->lov_qos.lq_reset || age > 32 * lov->desc.ld_qos_maxage)
 238                         oss->lqo_penalty = 0;
 239                 else if (age > lov->desc.ld_qos_maxage)
 240                         /* Decay the penalty by half for every 8x the update
 241                          * interval that the device has been idle.  That gives
 242                          * lots of time for the statfs information to be
 243                          * updated (which the penalty is only a proxy for),
 244                          * and avoids penalizing OSS/OSTs under light load. */
 245                         oss->lqo_penalty >>= (age / lov->desc.ld_qos_maxage);
 246         }
 247
 248         lov->lov_qos.lq_dirty = 0;
 249         lov->lov_qos.lq_reset = 0;
 250
 251         /* If each ost has almost same free space,
 252          * do rr allocation for better creation performance */
 253         lov->lov_qos.lq_same_space = 0;
 254         if ((ba_max * (256 - lov->lov_qos.lq_threshold_rr)) >> 8 < ba_min) {
 255                 lov->lov_qos.lq_same_space = 1;
 256                 /* Reset weights for the next time we enter qos mode */
 257                 lov->lov_qos.lq_reset = 1;
 258         }
 259         rc = 0;
 260
 261 out:
 262         if (!rc && lov->lov_qos.lq_same_space)
 263                 RETURN(-EAGAIN);
 264         RETURN(rc);
 265 }
 266
 267 static int qos_calc_weight(struct lov_obd *lov, int i)
 268 {
 269         __u64 temp, temp2;
 270
 271         /* Final ost weight = TGT_BAVAIL - ost_penalty - oss_penalty */
 272         temp = TGT_BAVAIL(i);
 273         temp2 = lov->lov_tgts[i]->ltd_qos.ltq_penalty +
 274                 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty;
 275         if (temp < temp2)
 276                 lov->lov_tgts[i]->ltd_qos.ltq_weight = 0;
 277         else
 278                 lov->lov_tgts[i]->ltd_qos.ltq_weight = temp - temp2;
 279         return 0;
 280 }
 281
 282 /* We just used this index for a stripe; adjust everyone's weights */
 283 static int qos_used(struct lov_obd *lov, struct ost_pool *osts,
 284                     __u32 index, __u64 *total_wt)
 285 {
 286         struct lov_qos_oss *oss;
 287         int j;
 288         ENTRY;
 289
 290         /* Don't allocate from this stripe anymore, until the next alloc_qos */
 291         lov->lov_tgts[index]->ltd_qos.ltq_usable = 0;
 292
 293         oss = lov->lov_tgts[index]->ltd_qos.ltq_oss;
 294
 295         /* Decay old penalty by half (we're adding max penalty, and don't
 296            want it to run away.) */
 297         lov->lov_tgts[index]->ltd_qos.ltq_penalty >>= 1;
 298         oss->lqo_penalty >>= 1;
 299
 300         /* mark the OSS and OST as recently used */
 301         lov->lov_tgts[index]->ltd_qos.ltq_used =
 302                 oss->lqo_used = cfs_time_current_sec();
 303
 304         /* Set max penalties for this OST and OSS */
 305         lov->lov_tgts[index]->ltd_qos.ltq_penalty +=
 306                 lov->lov_tgts[index]->ltd_qos.ltq_penalty_per_obj *
 307                 lov->desc.ld_active_tgt_count;
 308         oss->lqo_penalty += oss->lqo_penalty_per_obj *
 309                 lov->lov_qos.lq_active_oss_count;
 310
 311         /* Decrease all OSS penalties */
 312         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 313                 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
 314                         oss->lqo_penalty = 0;
 315                 else
 316                         oss->lqo_penalty -= oss->lqo_penalty_per_obj;
 317         }
 318
 319         *total_wt = 0;
 320         /* Decrease all OST penalties */
 321         for (j = 0; j < osts->op_count; j++) {
 322                 int i;
 323
 324                 i = osts->op_array[j];
 325                 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
 326                         continue;
 327                 if (lov->lov_tgts[i]->ltd_qos.ltq_penalty <
 328                     lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj)
 329                         lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
 330                 else
 331                         lov->lov_tgts[i]->ltd_qos.ltq_penalty -=
 332                         lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj;
 333
 334                 qos_calc_weight(lov, i);
 335
 336                 /* Recalc the total weight of usable osts */
 337                 if (lov->lov_tgts[i]->ltd_qos.ltq_usable)
 338                         *total_wt += lov->lov_tgts[i]->ltd_qos.ltq_weight;
 339
 340 #ifdef QOS_DEBUG
 341                 CDEBUG(D_QOS, "recalc tgt %d usable=%d avail="LPU64
 342                        " ostppo="LPU64" ostp="LPU64" ossppo="LPU64
 343                        " ossp="LPU64" wt="LPU64"\n",
 344                        i, lov->lov_tgts[i]->ltd_qos.ltq_usable,
 345                        TGT_BAVAIL(i) >> 10,
 346                        lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj >> 10,
 347                        lov->lov_tgts[i]->ltd_qos.ltq_penalty >> 10,
 348                        lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty_per_obj>>10,
 349                        lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty >> 10,
 350                        lov->lov_tgts[i]->ltd_qos.ltq_weight >> 10);
 351 #endif
 352         }
 353
 354         RETURN(0);
 355 }
 356
 357 #define LOV_QOS_EMPTY ((__u32)-1)
 358 /* compute optimal round-robin order, based on OSTs per OSS
 359  */
 360 static int qos_calc_rr(struct lov_obd *lov, struct ost_pool *src_pool,
 361                        struct lov_qos_rr *lqr)
 362 {
 363         struct lov_qos_oss *oss;
 364         unsigned placed, real_count;
 365         int i, rc;
 366         ENTRY;
 367
 368         if (!lqr->lqr_dirty) {
 369                 LASSERT(lqr->lqr_pool.op_size);
 370                 RETURN(0);
 371         }
 372
 373         /* Do actual allocation. */
 374         down_write(&lov->lov_qos.lq_rw_sem);
 375
 376         real_count = src_pool->op_count;
 377
 378         /* Zero the pool array */
 379         /* alloc_rr is holding a read lock on the pool, so nobody is adding/
 380            deleting from the pool. The lq_rw_sem insures that nobody else
 381            is reading. */
 382         lqr->lqr_pool.op_count = real_count;
 383         rc = lov_ost_pool_extend(&lqr->lqr_pool, real_count);
 384         if (rc) {
 385                 up_write(&lov->lov_qos.lq_rw_sem);
 386                 RETURN(rc);
 387         }
 388         for (i = 0; i < lqr->lqr_pool.op_count; i++)
 389                 lqr->lqr_pool.op_array[i] = LOV_QOS_EMPTY;
 390
 391         /* Place all the OSTs from 1 OSS at the same time. */
 392         placed = 0;
 393         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 394                 int j = 0;
 395                 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
 396                         if (lov->lov_tgts[src_pool->op_array[i]] &&
 397                             (lov->lov_tgts[src_pool->op_array[i]]->ltd_qos.ltq_oss == oss)) {
 398                               /* Evenly space these OSTs across arrayspace */
 399                               int next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count;
 400                               while (lqr->lqr_pool.op_array[next] !=
 401                                      LOV_QOS_EMPTY)
 402                                       next = (next + 1) % lqr->lqr_pool.op_count;
 403                               lqr->lqr_pool.op_array[next] = src_pool->op_array[i];
 404                               j++;
 405                               placed++;
 406                         }
 407                 }
 408         }
 409
 410         lqr->lqr_dirty = 0;
 411         up_write(&lov->lov_qos.lq_rw_sem);
 412
 413         if (placed != real_count) {
 414                 /* This should never happen */
 415                 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
 416                                    "round-robin list (%d of %d).\n",
 417                                    placed, real_count);
 418                 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
 419                         LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
 420                                  lqr->lqr_pool.op_array[i]);
 421                 }
 422                 lqr->lqr_dirty = 1;
 423                 RETURN(-EAGAIN);
 424         }
 425
 426 #ifdef QOS_DEBUG
 427         for (i = 0; i < lqr->lqr_pool.op_count; i++) {
 428                 LCONSOLE(D_QOS, "rr #%d ost idx=%d\n", i,
 429                          lqr->lqr_pool.op_array[i]);
 430         }
 431 #endif
 432
 433         RETURN(0);
 434 }
 435
 436
 437 void qos_shrink_lsm(struct lov_request_set *set)
 438 {
 439         struct lov_stripe_md *lsm = set->set_oi->oi_md, *lsm_new;
 440         /* XXX LOV STACKING call into osc for sizes */
 441         unsigned oldsize, newsize;
 442
 443         if (set->set_oti && set->set_cookies && set->set_cookie_sent) {
 444                 struct llog_cookie *cookies;
 445                 oldsize = lsm->lsm_stripe_count * sizeof(*cookies);
 446                 newsize = set->set_count * sizeof(*cookies);
 447
 448                 cookies = set->set_cookies;
 449                 oti_alloc_cookies(set->set_oti, set->set_count);
 450                 if (set->set_oti->oti_logcookies) {
 451                         memcpy(set->set_oti->oti_logcookies, cookies, newsize);
 452                         OBD_FREE(cookies, oldsize);
 453                         set->set_cookies = set->set_oti->oti_logcookies;
 454                 } else {
 455                         CWARN("'leaking' %d bytes\n", oldsize - newsize);
 456                 }
 457         }
 458
 459         CWARN("using fewer stripes for object "LPU64": old %u new %u\n",
 460               lsm->lsm_object_id, lsm->lsm_stripe_count, set->set_count);
 461         LASSERT(lsm->lsm_stripe_count >= set->set_count);
 462
 463         newsize = lov_stripe_md_size(set->set_count);
 464         OBD_ALLOC(lsm_new, newsize);
 465         if (lsm_new != NULL) {
 466                 int i;
 467                 memcpy(lsm_new, lsm, sizeof(*lsm));
 468                 for (i = 0; i < lsm->lsm_stripe_count; i++) {
 469                         if (i < set->set_count) {
 470                                 lsm_new->lsm_oinfo[i] = lsm->lsm_oinfo[i];
 471                                 continue;
 472                         }
 473                         OBD_SLAB_FREE(lsm->lsm_oinfo[i], lov_oinfo_slab,
 474                                       sizeof(struct lov_oinfo));
 475                 }
 476                 lsm_new->lsm_stripe_count = set->set_count;
 477                 OBD_FREE(lsm, sizeof(struct lov_stripe_md) +
 478                          lsm->lsm_stripe_count * sizeof(struct lov_oinfo *));
 479                 set->set_oi->oi_md = lsm_new;
 480         } else {
 481                 CWARN("'leaking' few bytes\n");
 482         }
 483 }
 484
 485 int qos_remedy_create(struct lov_request_set *set, struct lov_request *req)
 486 {
 487         struct lov_stripe_md *lsm = set->set_oi->oi_md;
 488         struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
 489         unsigned ost_idx, ost_count = lov->desc.ld_tgt_count;
 490         int stripe, i, rc = -EIO;
 491         ENTRY;
 492
 493         ost_idx = (req->rq_idx + lsm->lsm_stripe_count) % ost_count;
 494         for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
 495                 if (!lov->lov_tgts[ost_idx] ||
 496                     !lov->lov_tgts[ost_idx]->ltd_active)
 497                         continue;
 498                 /* check if objects has been created on this ost */
 499                 for (stripe = 0; stripe < lsm->lsm_stripe_count; stripe++) {
 500                         if (stripe == req->rq_stripe)
 501                                 continue;
 502                         if (ost_idx == lsm->lsm_oinfo[stripe]->loi_ost_idx)
 503                                 break;
 504                 }
 505
 506                 if (stripe >= lsm->lsm_stripe_count) {
 507                         req->rq_idx = ost_idx;
 508                         rc = obd_create(lov->lov_tgts[ost_idx]->ltd_exp,
 509                                         req->rq_oi.oi_oa, &req->rq_oi.oi_md,
 510                                         set->set_oti);
 511                         if (!rc)
 512                                 break;
 513                 }
 514         }
 515         RETURN(rc);
 516 }
 517
 518 static int min_stripe_count(int stripe_cnt, int flags)
 519 {
 520         return (flags & LOV_USES_DEFAULT_STRIPE ?
 521                 stripe_cnt - (stripe_cnt / 4) : stripe_cnt);
 522 }
 523
 524 #define LOV_CREATE_RESEED_MULT 4
 525 #define LOV_CREATE_RESEED_MIN  1000
 526 /* Allocate objects on osts with round-robin algorithm */
 527 static int alloc_rr(struct lov_obd *lov, int *idx_arr, int *stripe_cnt,
 528                     char *poolname, int flags)
 529 {
 530         unsigned array_idx;
 531         int i, rc, *idx_pos;
 532         __u32 ost_idx;
 533         int ost_start_idx_temp;
 534         int speed = 0;
 535         int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
 536         struct pool_desc *pool;
 537         struct ost_pool *osts;
 538         struct lov_qos_rr *lqr;
 539         ENTRY;
 540
 541         pool = lov_find_pool(lov, poolname);
 542         if (pool == NULL) {
 543                 osts = &(lov->lov_packed);
 544                 lqr = &(lov->lov_qos.lq_rr);
 545         } else {
 546                 down_read(&pool_tgt_rw_sem(pool));
 547                 osts = &(pool->pool_obds);
 548                 lqr = &(pool->pool_rr);
 549         }
 550
 551         rc = qos_calc_rr(lov, osts, lqr);
 552         if (rc)
 553                 GOTO(out, rc);
 554
 555         if (--lqr->lqr_start_count <= 0) {
 556                 lqr->lqr_start_idx = ll_rand() % osts->op_count;
 557                 lqr->lqr_start_count =
 558                         (LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) +
 559                          LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U);
 560         } else if (stripe_cnt_min >= osts->op_count ||
 561                    lqr->lqr_start_idx > osts->op_count) {
 562                 /* If we have allocated from all of the OSTs, slowly
 563                  * precess the next start if the OST/stripe count isn't
 564                  * already doing this for us. */
 565                 lqr->lqr_start_idx %= osts->op_count;
 566                 if (*stripe_cnt > 1 && (osts->op_count % (*stripe_cnt)) != 1)
 567                         ++lqr->lqr_offset_idx;
 568         }
 569         down_read(&lov->lov_qos.lq_rw_sem);
 570         ost_start_idx_temp = lqr->lqr_start_idx;
 571
 572 repeat_find:
 573         array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) % osts->op_count;
 574         idx_pos = idx_arr;
 575 #ifdef QOS_DEBUG
 576         CDEBUG(D_QOS, "pool '%s' want %d startidx %d startcnt %d offset %d "
 577                "active %d count %d arrayidx %d\n", poolname,
 578                *stripe_cnt, lqr->lqr_start_idx, lqr->lqr_start_count,
 579                lqr->lqr_offset_idx, osts->op_count, osts->op_count, array_idx);
 580 #endif
 581
 582         for (i = 0; i < osts->op_count;
 583                     i++, array_idx=(array_idx + 1) % osts->op_count) {
 584                 ++lqr->lqr_start_idx;
 585                 ost_idx = lqr->lqr_pool.op_array[array_idx];
 586 #ifdef QOS_DEBUG
 587                 CDEBUG(D_QOS, "#%d strt %d act %d strp %d ary %d idx %d\n",
 588                        i, lqr->lqr_start_idx,
 589                        ((ost_idx != LOV_QOS_EMPTY) && lov->lov_tgts[ost_idx]) ?
 590                        lov->lov_tgts[ost_idx]->ltd_active : 0,
 591                        idx_pos - idx_arr, array_idx, ost_idx);
 592 #endif
 593                 if ((ost_idx == LOV_QOS_EMPTY) || !lov->lov_tgts[ost_idx] ||
 594                     !lov->lov_tgts[ost_idx]->ltd_active)
 595                         continue;
 596
 597                 /* Fail Check before osc_precreate() is called
 598                    so we can only 'fail' single OSC. */
 599                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
 600                         continue;
 601
 602                 /* Drop slow OSCs if we can */
 603                 if (obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed)
 604                         continue;
 605
 606                 *idx_pos = ost_idx;
 607                 idx_pos++;
 608                 /* We have enough stripes */
 609                 if (idx_pos - idx_arr == *stripe_cnt)
 610                         break;
 611         }
 612         if ((speed < 2) && (idx_pos - idx_arr < stripe_cnt_min)) {
 613                 /* Try again, allowing slower OSCs */
 614                 speed++;
 615                 lqr->lqr_start_idx = ost_start_idx_temp;
 616                 goto repeat_find;
 617         }
 618
 619         up_read(&lov->lov_qos.lq_rw_sem);
 620
 621         *stripe_cnt = idx_pos - idx_arr;
 622 out:
 623         if (pool != NULL) {
 624                 up_read(&pool_tgt_rw_sem(pool));
 625                 /* put back ref got by lov_find_pool() */
 626                 lov_pool_putref(pool);
 627         }
 628
 629         RETURN(rc);
 630 }
 631
 632 /* alloc objects on osts with specific stripe offset */
 633 static int alloc_specific(struct lov_obd *lov, struct lov_stripe_md *lsm,
 634                           int *idx_arr)
 635 {
 636         unsigned ost_idx, array_idx, ost_count;
 637         int i, rc, *idx_pos;
 638         int speed = 0;
 639         struct pool_desc *pool;
 640         struct ost_pool *osts;
 641         ENTRY;
 642
 643         pool = lov_find_pool(lov, lsm->lsm_pool_name);
 644         if (pool == NULL) {
 645                 osts = &(lov->lov_packed);
 646         } else {
 647                 down_read(&pool_tgt_rw_sem(pool));
 648                 osts = &(pool->pool_obds);
 649         }
 650
 651         ost_count = osts->op_count;
 652
 653 repeat_find:
 654         /* search loi_ost_idx in ost array */
 655         array_idx = 0;
 656         for (i = 0; i < ost_count; i++) {
 657                 if (osts->op_array[i] == lsm->lsm_oinfo[0]->loi_ost_idx) {
 658                         array_idx = i;
 659                         break;
 660                 }
 661         }
 662         if (i == ost_count) {
 663                 CERROR("Start index %d not found in pool '%s'\n",
 664                        lsm->lsm_oinfo[0]->loi_ost_idx, lsm->lsm_pool_name);
 665                 GOTO(out, rc = -EINVAL);
 666         }
 667
 668         idx_pos = idx_arr;
 669         for (i = 0; i < ost_count;
 670              i++, array_idx = (array_idx + 1) % ost_count) {
 671                 ost_idx = osts->op_array[array_idx];
 672
 673                 if (!lov->lov_tgts[ost_idx] ||
 674                     !lov->lov_tgts[ost_idx]->ltd_active) {
 675                         continue;
 676                 }
 677
 678                 /* Fail Check before osc_precreate() is called
 679                    so we can only 'fail' single OSC. */
 680                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
 681                         continue;
 682
 683                 /* Drop slow OSCs if we can, but not for requested start idx.
 684                  *
 685                  * This means "if OSC is slow and it is not the requested
 686                  * start OST, then it can be skipped, otherwise skip it only
 687                  * if it is inactive/recovering/out-of-space." */
 688                 if ((obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed) &&
 689                     (i != 0 || speed >= 2))
 690                         continue;
 691
 692                 *idx_pos = ost_idx;
 693                 idx_pos++;
 694                 /* We have enough stripes */
 695                 if (idx_pos - idx_arr == lsm->lsm_stripe_count)
 696                         GOTO(out, rc = 0);
 697         }
 698         if (speed < 2) {
 699                 /* Try again, allowing slower OSCs */
 700                 speed++;
 701                 goto repeat_find;
 702         }
 703
 704         /* If we were passed specific striping params, then a failure to
 705          * meet those requirements is an error, since we can't reallocate
 706          * that memory (it might be part of a larger array or something).
 707          *
 708          * We can only get here if lsm_stripe_count was originally > 1.
 709          */
 710         CERROR("can't lstripe objid "LPX64": have %d want %u\n",
 711                lsm->lsm_object_id, (int)(idx_pos - idx_arr),
 712                lsm->lsm_stripe_count);
 713         rc = -EFBIG;
 714 out:
 715         if (pool != NULL) {
 716                 up_read(&pool_tgt_rw_sem(pool));
 717                 /* put back ref got by lov_find_pool() */
 718                 lov_pool_putref(pool);
 719         }
 720         RETURN(rc);
 721 }
 722
 723 /* Alloc objects on osts with optimization based on:
 724    - free space
 725    - network resources (shared OSS's)
 726 */
 727 static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt,
 728                      char *poolname, int flags)
 729 {
 730         struct lov_obd *lov = &exp->exp_obd->u.lov;
 731         static time_t last_warn = 0;
 732         time_t now = cfs_time_current_sec();
 733         __u64 total_bavail, total_weight = 0;
 734         int nfound, good_osts, i, warn = 0, rc = 0;
 735         int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
 736         struct pool_desc *pool;
 737         struct ost_pool *osts;
 738         struct lov_qos_rr *lqr;
 739         ENTRY;
 740
 741         if (stripe_cnt_min < 1)
 742                 RETURN(-EINVAL);
 743
 744         pool = lov_find_pool(lov, poolname);
 745         if (pool == NULL) {
 746                 osts = &(lov->lov_packed);
 747                 lqr = &(lov->lov_qos.lq_rr);
 748         } else {
 749                 down_read(&pool_tgt_rw_sem(pool));
 750                 osts = &(pool->pool_obds);
 751                 lqr = &(pool->pool_rr);
 752         }
 753
 754         lov_getref(exp->exp_obd);
 755         down_write(&lov->lov_qos.lq_rw_sem);
 756
 757         if (lov->desc.ld_active_tgt_count < 2)
 758                 GOTO(out, rc = -EAGAIN);
 759
 760         rc = qos_calc_ppo(exp->exp_obd);
 761         if (rc)
 762                 GOTO(out, rc);
 763
 764         total_bavail = 0;
 765         good_osts = 0;
 766         /* Warn users about zero available space/inode every 30 min */
 767         if (cfs_time_sub(now, last_warn) > 60 * 30)
 768                 warn = 1;
 769         /* Find all the OSTs that are valid stripe candidates */
 770         for (i = 0; i < osts->op_count; i++) {
 771                 __u64 bavail;
 772
 773                 if (!lov->lov_tgts[osts->op_array[i]] ||
 774                     !lov->lov_tgts[osts->op_array[i]]->ltd_active)
 775                         continue;
 776                 bavail = TGT_BAVAIL(osts->op_array[i]);
 777                 if (!bavail) {
 778                         if (warn) {
 779                                 CDEBUG(D_QOS, "no free space on %s\n",
 780                                      obd_uuid2str(&lov->lov_tgts[osts->op_array[i]]->ltd_uuid));
 781                                 last_warn = now;
 782                         }
 783                         continue;
 784                 }
 785                 if (!TGT_FFREE(osts->op_array[i])) {
 786                         if (warn) {
 787                                 CDEBUG(D_QOS, "no free inodes on %s\n",
 788                                      obd_uuid2str(&lov->lov_tgts[osts->op_array[i]]->ltd_uuid));
 789                                 last_warn = now;
 790                         }
 791                         continue;
 792                 }
 793
 794                 /* Fail Check before osc_precreate() is called
 795                    so we can only 'fail' single OSC. */
 796                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && osts->op_array[i] == 0)
 797                         continue;
 798
 799                 if (obd_precreate(lov->lov_tgts[osts->op_array[i]]->ltd_exp) > 2)
 800                         continue;
 801
 802                 lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable = 1;
 803                 qos_calc_weight(lov, osts->op_array[i]);
 804                 total_bavail += bavail;
 805                 total_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight;
 806
 807                 good_osts++;
 808         }
 809
 810 #ifdef QOS_DEBUG
 811         CDEBUG(D_QOS, "found %d good osts\n", good_osts);
 812 #endif
 813
 814         if (good_osts < stripe_cnt_min)
 815                 GOTO(out, rc = -EAGAIN);
 816
 817         if (!total_bavail)
 818                 GOTO(out, rc = -ENOSPC);
 819
 820         /* We have enough osts */
 821         if (good_osts < *stripe_cnt)
 822                 *stripe_cnt = good_osts;
 823
 824         /* Find enough OSTs with weighted random allocation. */
 825         nfound = 0;
 826         while (nfound < *stripe_cnt) {
 827                 __u64 rand, cur_weight;
 828
 829                 cur_weight = 0;
 830                 rc = -ENODEV;
 831
 832                 if (total_weight) {
 833 #if BITS_PER_LONG == 32
 834                         rand = ll_rand() % (unsigned)total_weight;
 835                         /* If total_weight > 32-bit, first generate the high
 836                          * 32 bits of the random number, then add in the low
 837                          * 32 bits (truncated to the upper limit, if needed) */
 838                         if (total_weight > 0xffffffffULL)
 839                                 rand = (__u64)(ll_rand() %
 840                                           (unsigned)(total_weight >> 32)) << 32;
 841                         else
 842                                 rand = 0;
 843
 844                         if (rand == (total_weight & 0xffffffff00000000ULL))
 845                                 rand |= ll_rand() % (unsigned)total_weight;
 846                         else
 847                                 rand |= ll_rand();
 848
 849 #else
 850                         rand = ((__u64)ll_rand() << 32 | ll_rand()) %
 851                                 total_weight;
 852 #endif
 853                 } else {
 854                         rand = 0;
 855                 }
 856
 857                 /* On average, this will hit larger-weighted osts more often.
 858                    0-weight osts will always get used last (only when rand=0).*/
 859                 for (i = 0; i < osts->op_count; i++) {
 860                         if (!lov->lov_tgts[osts->op_array[i]] ||
 861                             !lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable)
 862                                 continue;
 863
 864                         cur_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight;
 865 #ifdef QOS_DEBUG
 866                         CDEBUG(D_QOS, "stripe_cnt=%d nfound=%d cur_weight="LPU64
 867                                       " rand="LPU64" total_weight="LPU64"\n",
 868                                *stripe_cnt, nfound, cur_weight, rand, total_weight);
 869 #endif
 870                         if (cur_weight >= rand) {
 871 #ifdef QOS_DEBUG
 872                                 CDEBUG(D_QOS, "assigned stripe=%d to idx=%d\n",
 873                                        nfound, osts->op_array[i]);
 874 #endif
 875                                 idx_arr[nfound++] = osts->op_array[i];
 876                                 qos_used(lov, osts, osts->op_array[i], &total_weight);
 877                                 rc = 0;
 878                                 break;
 879                         }
 880                 }
 881                 if (rc) {
 882                         CDEBUG(D_QOS, "Didn't find any OSTs? Reduce total weight\n");
 883                         if (total_weight == 0)
 884                                 break;
 885                         else
 886                                 total_weight = 0;
 887                 }
 888         }
 889
 890         LASSERT(nfound == *stripe_cnt);
 891
 892 out:
 893         up_write(&lov->lov_qos.lq_rw_sem);
 894
 895         if (pool != NULL) {
 896                 up_read(&pool_tgt_rw_sem(pool));
 897                 /* put back ref got by lov_find_pool() */
 898                 lov_pool_putref(pool);
 899         }
 900
 901         if (rc == -EAGAIN)
 902                 rc = alloc_rr(lov, idx_arr, stripe_cnt, poolname, flags);
 903
 904         lov_putref(exp->exp_obd);
 905         RETURN(rc);
 906 }
 907
 908 /* return new alloced stripe count on success */
 909 static int alloc_idx_array(struct obd_export *exp, struct lov_stripe_md *lsm,
 910                            int newea, int **idx_arr, int *arr_cnt, int flags)
 911 {
 912         struct lov_obd *lov = &exp->exp_obd->u.lov;
 913         int stripe_cnt = lsm->lsm_stripe_count;
 914         int i, rc = 0;
 915         int *tmp_arr = NULL;
 916         ENTRY;
 917
 918         *arr_cnt = stripe_cnt;
 919         OBD_ALLOC(tmp_arr, *arr_cnt * sizeof(int));
 920         if (tmp_arr == NULL)
 921                 RETURN(-ENOMEM);
 922         for (i = 0; i < *arr_cnt; i++)
 923                 tmp_arr[i] = -1;
 924
 925         if (newea ||
 926             lsm->lsm_oinfo[0]->loi_ost_idx >= lov->desc.ld_tgt_count)
 927                 rc = alloc_qos(exp, tmp_arr, &stripe_cnt,
 928                                lsm->lsm_pool_name, flags);
 929         else
 930                 rc = alloc_specific(lov, lsm, tmp_arr);
 931
 932         if (rc)
 933                 GOTO(out_arr, rc);
 934
 935         *idx_arr = tmp_arr;
 936         RETURN(stripe_cnt);
 937 out_arr:
 938         OBD_FREE(tmp_arr, *arr_cnt * sizeof(int));
 939         *arr_cnt = 0;
 940         RETURN(rc);
 941 }
 942
 943 static void free_idx_array(int *idx_arr, int arr_cnt)
 944 {
 945         if (arr_cnt)
 946                 OBD_FREE(idx_arr, arr_cnt * sizeof(int));
 947 }
 948
 949 int qos_prep_create(struct obd_export *exp, struct lov_request_set *set)
 950 {
 951         struct lov_obd *lov = &exp->exp_obd->u.lov;
 952         struct lov_stripe_md *lsm;
 953         struct obdo *src_oa = set->set_oi->oi_oa;
 954         struct obd_trans_info *oti = set->set_oti;
 955         int i, stripes, rc = 0, newea = 0;
 956         int flag = LOV_USES_ASSIGNED_STRIPE;
 957         int *idx_arr = NULL, idx_cnt = 0;
 958         ENTRY;
 959
 960         LASSERT(src_oa->o_valid & OBD_MD_FLID);
 961
 962         if (set->set_oi->oi_md == NULL) {
 963                 int stripes_def = lov_get_stripecnt(lov, 0);
 964
 965                 /* If the MDS file was truncated up to some size, stripe over
 966                  * enough OSTs to allow the file to be created at that size.
 967                  * This may mean we use more than the default # of stripes. */
 968                 if (src_oa->o_valid & OBD_MD_FLSIZE) {
 969                         obd_size min_bavail = LUSTRE_STRIPE_MAXBYTES;
 970
 971                         /* Find a small number of stripes we can use
 972                            (up to # of active osts). */
 973                         stripes = 1;
 974                         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 975                                 if (!lov->lov_tgts[i] ||
 976                                     !lov->lov_tgts[i]->ltd_active)
 977                                         continue;
 978                                 min_bavail = min(min_bavail, TGT_BAVAIL(i));
 979                                 if (min_bavail * stripes > src_oa->o_size)
 980                                         break;
 981                                 stripes++;
 982                         }
 983
 984                         if (stripes < stripes_def)
 985                                 stripes = stripes_def;
 986                 } else {
 987                          flag = LOV_USES_DEFAULT_STRIPE;
 988                          stripes = stripes_def;
 989                 }
 990
 991                 rc = lov_alloc_memmd(&set->set_oi->oi_md, stripes,
 992                                      lov->desc.ld_pattern ?
 993                                      lov->desc.ld_pattern : LOV_PATTERN_RAID0,
 994                                      LOV_MAGIC);
 995                 if (rc < 0)
 996                         GOTO(out_err, rc);
 997                 newea = 1;
 998                 rc = 0;
 999         }
1000
1001         lsm = set->set_oi->oi_md;
1002         lsm->lsm_object_id = src_oa->o_id;
1003         if (!lsm->lsm_stripe_size)
1004                 lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size;
1005         if (!lsm->lsm_pattern) {
1006                 LASSERT(lov->desc.ld_pattern);
1007                 lsm->lsm_pattern = lov->desc.ld_pattern;
1008         }
1009
1010         stripes = alloc_idx_array(exp, lsm, newea, &idx_arr, &idx_cnt, flag);
1011         if (stripes <= 0)
1012                 GOTO(out_err, rc = stripes ? stripes : -EIO);
1013         LASSERTF(stripes <= lsm->lsm_stripe_count,"requested %d allocated %d\n",
1014                  lsm->lsm_stripe_count, stripes);
1015
1016         for (i = 0; i < stripes; i++) {
1017                 struct lov_request *req;
1018                 int ost_idx = idx_arr[i];
1019                 LASSERT(ost_idx >= 0);
1020
1021                 OBD_ALLOC(req, sizeof(*req));
1022                 if (req == NULL)
1023                         GOTO(out_err, rc = -ENOMEM);
1024                 lov_set_add_req(req, set);
1025
1026                 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
1027                 OBD_ALLOC(req->rq_oi.oi_md, req->rq_buflen);
1028                 if (req->rq_oi.oi_md == NULL)
1029                         GOTO(out_err, rc = -ENOMEM);
1030
1031                 OBDO_ALLOC(req->rq_oi.oi_oa);
1032                 if (req->rq_oi.oi_oa == NULL)
1033                         GOTO(out_err, rc = -ENOMEM);
1034
1035                 req->rq_idx = ost_idx;
1036                 req->rq_stripe = i;
1037                 /* create data objects with "parent" OA */
1038                 memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
1039
1040                 /* XXX When we start creating objects on demand, we need to
1041                  *     make sure that we always create the object on the
1042                  *     stripe which holds the existing file size.
1043                  */
1044                 if (src_oa->o_valid & OBD_MD_FLSIZE) {
1045                         req->rq_oi.oi_oa->o_size =
1046                                 lov_size_to_stripe(lsm, src_oa->o_size, i);
1047
1048                         CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
1049                                i, req->rq_oi.oi_oa->o_size, src_oa->o_size);
1050                 }
1051         }
1052         LASSERT(set->set_count == stripes);
1053
1054         if (stripes < lsm->lsm_stripe_count)
1055                 qos_shrink_lsm(set);
1056
1057         if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) {
1058                 oti_alloc_cookies(oti, set->set_count);
1059                 if (!oti->oti_logcookies)
1060                         GOTO(out_err, rc = -ENOMEM);
1061                 set->set_cookies = oti->oti_logcookies;
1062         }
1063 out_err:
1064         if (newea && rc)
1065                 obd_free_memmd(exp, &set->set_oi->oi_md);
1066         if (idx_arr)
1067                 free_idx_array(idx_arr, idx_cnt);
1068         EXIT;
1069         return rc;
1070 }
1071
1072 void qos_update(struct lov_obd *lov)
1073 {
1074         ENTRY;
1075         lov->lov_qos.lq_dirty = 1;
1076 }