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] == NULL)
 227                                 continue;
 228                         lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj <<= 1;
 229                 }
 230         }
 231
 232         /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
 233         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 234                 temp = oss->lqo_bavail >> 1;
 235                 do_div(temp, oss->lqo_ost_count * num_active);
 236                 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
 237
 238                 age = (now - oss->lqo_used) >> 3;
 239                 if (lov->lov_qos.lq_reset || age > 32 * lov->desc.ld_qos_maxage)
 240                         oss->lqo_penalty = 0;
 241                 else if (age > lov->desc.ld_qos_maxage)
 242                         /* Decay the penalty by half for every 8x the update
 243                          * interval that the device has been idle.  That gives
 244                          * lots of time for the statfs information to be
 245                          * updated (which the penalty is only a proxy for),
 246                          * and avoids penalizing OSS/OSTs under light load. */
 247                         oss->lqo_penalty >>= (age / lov->desc.ld_qos_maxage);
 248         }
 249
 250         lov->lov_qos.lq_dirty = 0;
 251         lov->lov_qos.lq_reset = 0;
 252
 253         /* If each ost has almost same free space,
 254          * do rr allocation for better creation performance */
 255         lov->lov_qos.lq_same_space = 0;
 256         if ((ba_max * (256 - lov->lov_qos.lq_threshold_rr)) >> 8 < ba_min) {
 257                 lov->lov_qos.lq_same_space = 1;
 258                 /* Reset weights for the next time we enter qos mode */
 259                 lov->lov_qos.lq_reset = 1;
 260         }
 261         rc = 0;
 262
 263 out:
 264         if (!rc && lov->lov_qos.lq_same_space)
 265                 RETURN(-EAGAIN);
 266         RETURN(rc);
 267 }
 268
 269 static int qos_calc_weight(struct lov_obd *lov, int i)
 270 {
 271         __u64 temp, temp2;
 272
 273         /* Final ost weight = TGT_BAVAIL - ost_penalty - oss_penalty */
 274         temp = TGT_BAVAIL(i);
 275         temp2 = lov->lov_tgts[i]->ltd_qos.ltq_penalty +
 276                 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty;
 277         if (temp < temp2)
 278                 lov->lov_tgts[i]->ltd_qos.ltq_weight = 0;
 279         else
 280                 lov->lov_tgts[i]->ltd_qos.ltq_weight = temp - temp2;
 281         return 0;
 282 }
 283
 284 /* We just used this index for a stripe; adjust everyone's weights */
 285 static int qos_used(struct lov_obd *lov, struct ost_pool *osts,
 286                     __u32 index, __u64 *total_wt)
 287 {
 288         struct lov_qos_oss *oss;
 289         int j;
 290         ENTRY;
 291
 292         /* Don't allocate from this stripe anymore, until the next alloc_qos */
 293         lov->lov_tgts[index]->ltd_qos.ltq_usable = 0;
 294
 295         oss = lov->lov_tgts[index]->ltd_qos.ltq_oss;
 296
 297         /* Decay old penalty by half (we're adding max penalty, and don't
 298            want it to run away.) */
 299         lov->lov_tgts[index]->ltd_qos.ltq_penalty >>= 1;
 300         oss->lqo_penalty >>= 1;
 301
 302         /* mark the OSS and OST as recently used */
 303         lov->lov_tgts[index]->ltd_qos.ltq_used =
 304                 oss->lqo_used = cfs_time_current_sec();
 305
 306         /* Set max penalties for this OST and OSS */
 307         lov->lov_tgts[index]->ltd_qos.ltq_penalty +=
 308                 lov->lov_tgts[index]->ltd_qos.ltq_penalty_per_obj *
 309                 lov->desc.ld_active_tgt_count;
 310         oss->lqo_penalty += oss->lqo_penalty_per_obj *
 311                 lov->lov_qos.lq_active_oss_count;
 312
 313         /* Decrease all OSS penalties */
 314         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 315                 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
 316                         oss->lqo_penalty = 0;
 317                 else
 318                         oss->lqo_penalty -= oss->lqo_penalty_per_obj;
 319         }
 320
 321         *total_wt = 0;
 322         /* Decrease all OST penalties */
 323         for (j = 0; j < osts->op_count; j++) {
 324                 int i;
 325
 326                 i = osts->op_array[j];
 327                 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
 328                         continue;
 329                 if (lov->lov_tgts[i]->ltd_qos.ltq_penalty <
 330                     lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj)
 331                         lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
 332                 else
 333                         lov->lov_tgts[i]->ltd_qos.ltq_penalty -=
 334                         lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj;
 335
 336                 qos_calc_weight(lov, i);
 337
 338                 /* Recalc the total weight of usable osts */
 339                 if (lov->lov_tgts[i]->ltd_qos.ltq_usable)
 340                         *total_wt += lov->lov_tgts[i]->ltd_qos.ltq_weight;
 341
 342 #ifdef QOS_DEBUG
 343                 CDEBUG(D_QOS, "recalc tgt %d usable=%d avail="LPU64
 344                        " ostppo="LPU64" ostp="LPU64" ossppo="LPU64
 345                        " ossp="LPU64" wt="LPU64"\n",
 346                        i, lov->lov_tgts[i]->ltd_qos.ltq_usable,
 347                        TGT_BAVAIL(i) >> 10,
 348                        lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj >> 10,
 349                        lov->lov_tgts[i]->ltd_qos.ltq_penalty >> 10,
 350                        lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty_per_obj>>10,
 351                        lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty >> 10,
 352                        lov->lov_tgts[i]->ltd_qos.ltq_weight >> 10);
 353 #endif
 354         }
 355
 356         RETURN(0);
 357 }
 358
 359 #define LOV_QOS_EMPTY ((__u32)-1)
 360 /* compute optimal round-robin order, based on OSTs per OSS */
 361 static int qos_calc_rr(struct lov_obd *lov, struct ost_pool *src_pool,
 362                        struct lov_qos_rr *lqr)
 363 {
 364         struct lov_qos_oss *oss;
 365         unsigned placed, real_count;
 366         int i, rc;
 367         ENTRY;
 368
 369         if (!lqr->lqr_dirty) {
 370                 LASSERT(lqr->lqr_pool.op_size);
 371                 RETURN(0);
 372         }
 373
 374         /* Do actual allocation. */
 375         down_write(&lov->lov_qos.lq_rw_sem);
 376
 377         /*
 378          * Check again. While we were sleeping on @lq_rw_sem something could
 379          * change.
 380          */
 381         if (!lqr->lqr_dirty) {
 382                 LASSERT(lqr->lqr_pool.op_size);
 383                 up_write(&lov->lov_qos.lq_rw_sem);
 384                 RETURN(0);
 385         }
 386
 387         real_count = src_pool->op_count;
 388
 389         /* Zero the pool array */
 390         /* alloc_rr is holding a read lock on the pool, so nobody is adding/
 391            deleting from the pool. The lq_rw_sem insures that nobody else
 392            is reading. */
 393         lqr->lqr_pool.op_count = real_count;
 394         rc = lov_ost_pool_extend(&lqr->lqr_pool, real_count);
 395         if (rc) {
 396                 up_write(&lov->lov_qos.lq_rw_sem);
 397                 RETURN(rc);
 398         }
 399         for (i = 0; i < lqr->lqr_pool.op_count; i++)
 400                 lqr->lqr_pool.op_array[i] = LOV_QOS_EMPTY;
 401
 402         /* Place all the OSTs from 1 OSS at the same time. */
 403         placed = 0;
 404         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 405                 int j = 0;
 406                 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
 407                         if (lov->lov_tgts[src_pool->op_array[i]] &&
 408                             (lov->lov_tgts[src_pool->op_array[i]]->ltd_qos.ltq_oss == oss)) {
 409                               /* Evenly space these OSTs across arrayspace */
 410                               int next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count;
 411                               while (lqr->lqr_pool.op_array[next] !=
 412                                      LOV_QOS_EMPTY)
 413                                         next = (next + 1) % lqr->lqr_pool.op_count;
 414                               lqr->lqr_pool.op_array[next] = src_pool->op_array[i];
 415                               j++;
 416                               placed++;
 417                         }
 418                 }
 419         }
 420
 421         lqr->lqr_dirty = 0;
 422         up_write(&lov->lov_qos.lq_rw_sem);
 423
 424         if (placed != real_count) {
 425                 /* This should never happen */
 426                 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
 427                                    "round-robin list (%d of %d).\n",
 428                                    placed, real_count);
 429                 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
 430                         LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
 431                                  lqr->lqr_pool.op_array[i]);
 432                 }
 433                 lqr->lqr_dirty = 1;
 434                 RETURN(-EAGAIN);
 435         }
 436
 437 #ifdef QOS_DEBUG
 438         for (i = 0; i < lqr->lqr_pool.op_count; i++) {
 439                 LCONSOLE(D_QOS, "rr #%d ost idx=%d\n", i,
 440                          lqr->lqr_pool.op_array[i]);
 441         }
 442 #endif
 443
 444         RETURN(0);
 445 }
 446
 447
 448 void qos_shrink_lsm(struct lov_request_set *set)
 449 {
 450         struct lov_stripe_md *lsm = set->set_oi->oi_md, *lsm_new;
 451         /* XXX LOV STACKING call into osc for sizes */
 452         unsigned oldsize, newsize;
 453
 454         if (set->set_oti && set->set_cookies && set->set_cookie_sent) {
 455                 struct llog_cookie *cookies;
 456                 oldsize = lsm->lsm_stripe_count * sizeof(*cookies);
 457                 newsize = set->set_count * sizeof(*cookies);
 458
 459                 cookies = set->set_cookies;
 460                 oti_alloc_cookies(set->set_oti, set->set_count);
 461                 if (set->set_oti->oti_logcookies) {
 462                         memcpy(set->set_oti->oti_logcookies, cookies, newsize);
 463                         OBD_FREE(cookies, oldsize);
 464                         set->set_cookies = set->set_oti->oti_logcookies;
 465                 } else {
 466                         CWARN("'leaking' %d bytes\n", oldsize - newsize);
 467                 }
 468         }
 469
 470         CWARN("using fewer stripes for object "LPU64": old %u new %u\n",
 471               lsm->lsm_object_id, lsm->lsm_stripe_count, set->set_count);
 472         LASSERT(lsm->lsm_stripe_count >= set->set_count);
 473
 474         newsize = lov_stripe_md_size(set->set_count);
 475         OBD_ALLOC(lsm_new, newsize);
 476         if (lsm_new != NULL) {
 477                 int i;
 478                 memcpy(lsm_new, lsm, sizeof(*lsm));
 479                 for (i = 0; i < lsm->lsm_stripe_count; i++) {
 480                         if (i < set->set_count) {
 481                                 lsm_new->lsm_oinfo[i] = lsm->lsm_oinfo[i];
 482                                 continue;
 483                         }
 484                         OBD_SLAB_FREE(lsm->lsm_oinfo[i], lov_oinfo_slab,
 485                                       sizeof(struct lov_oinfo));
 486                 }
 487                 lsm_new->lsm_stripe_count = set->set_count;
 488                 OBD_FREE(lsm, sizeof(struct lov_stripe_md) +
 489                          lsm->lsm_stripe_count * sizeof(struct lov_oinfo *));
 490                 set->set_oi->oi_md = lsm_new;
 491         } else {
 492                 CWARN("'leaking' few bytes\n");
 493         }
 494 }
 495
 496 int qos_remedy_create(struct lov_request_set *set, struct lov_request *req)
 497 {
 498         struct lov_stripe_md *lsm = set->set_oi->oi_md;
 499         struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
 500         unsigned ost_idx, ost_count = lov->desc.ld_tgt_count;
 501         int stripe, i, rc = -EIO;
 502         ENTRY;
 503
 504         ost_idx = (req->rq_idx + lsm->lsm_stripe_count) % ost_count;
 505         for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
 506                 if (!lov->lov_tgts[ost_idx] ||
 507                     !lov->lov_tgts[ost_idx]->ltd_active)
 508                         continue;
 509                 /* check if objects has been created on this ost */
 510                 for (stripe = 0; stripe < lsm->lsm_stripe_count; stripe++) {
 511                         if (stripe == req->rq_stripe)
 512                                 continue;
 513                         if (ost_idx == lsm->lsm_oinfo[stripe]->loi_ost_idx)
 514                                 break;
 515                 }
 516
 517                 if (stripe >= lsm->lsm_stripe_count) {
 518                         req->rq_idx = ost_idx;
 519                         rc = obd_create(lov->lov_tgts[ost_idx]->ltd_exp,
 520                                         req->rq_oi.oi_oa, &req->rq_oi.oi_md,
 521                                         set->set_oti);
 522                         if (!rc)
 523                                 break;
 524                 }
 525         }
 526         RETURN(rc);
 527 }
 528
 529 static int min_stripe_count(int stripe_cnt, int flags)
 530 {
 531         return (flags & LOV_USES_DEFAULT_STRIPE ?
 532                 stripe_cnt - (stripe_cnt / 4) : stripe_cnt);
 533 }
 534
 535 #define LOV_CREATE_RESEED_MULT 4
 536 #define LOV_CREATE_RESEED_MIN  1000
 537 /* Allocate objects on osts with round-robin algorithm */
 538 static int alloc_rr(struct lov_obd *lov, int *idx_arr, int *stripe_cnt,
 539                     char *poolname, int flags)
 540 {
 541         unsigned array_idx;
 542         int i, rc, *idx_pos;
 543         __u32 ost_idx;
 544         int ost_start_idx_temp;
 545         int speed = 0;
 546         int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
 547         struct pool_desc *pool;
 548         struct ost_pool *osts;
 549         struct lov_qos_rr *lqr;
 550         ENTRY;
 551
 552         pool = lov_find_pool(lov, poolname);
 553         if (pool == NULL) {
 554                 osts = &(lov->lov_packed);
 555                 lqr = &(lov->lov_qos.lq_rr);
 556         } else {
 557                 down_read(&pool_tgt_rw_sem(pool));
 558                 osts = &(pool->pool_obds);
 559                 lqr = &(pool->pool_rr);
 560         }
 561
 562         rc = qos_calc_rr(lov, osts, lqr);
 563         if (rc)
 564                 GOTO(out, rc);
 565
 566         if (--lqr->lqr_start_count <= 0) {
 567                 lqr->lqr_start_idx = ll_rand() % osts->op_count;
 568                 lqr->lqr_start_count =
 569                         (LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) +
 570                          LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U);
 571         } else if (stripe_cnt_min >= osts->op_count ||
 572                    lqr->lqr_start_idx > osts->op_count) {
 573                 /* If we have allocated from all of the OSTs, slowly
 574                  * precess the next start if the OST/stripe count isn't
 575                  * already doing this for us. */
 576                 lqr->lqr_start_idx %= osts->op_count;
 577                 if (*stripe_cnt > 1 && (osts->op_count % (*stripe_cnt)) != 1)
 578                         ++lqr->lqr_offset_idx;
 579         }
 580         down_read(&lov->lov_qos.lq_rw_sem);
 581         ost_start_idx_temp = lqr->lqr_start_idx;
 582
 583 repeat_find:
 584         array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) % osts->op_count;
 585         idx_pos = idx_arr;
 586 #ifdef QOS_DEBUG
 587         CDEBUG(D_QOS, "pool '%s' want %d startidx %d startcnt %d offset %d "
 588                "active %d count %d arrayidx %d\n", poolname,
 589                *stripe_cnt, lqr->lqr_start_idx, lqr->lqr_start_count,
 590                lqr->lqr_offset_idx, osts->op_count, osts->op_count, array_idx);
 591 #endif
 592
 593         for (i = 0; i < osts->op_count;
 594                     i++, array_idx=(array_idx + 1) % osts->op_count) {
 595                 ++lqr->lqr_start_idx;
 596                 ost_idx = lqr->lqr_pool.op_array[array_idx];
 597 #ifdef QOS_DEBUG
 598                 CDEBUG(D_QOS, "#%d strt %d act %d strp %d ary %d idx %d\n",
 599                        i, lqr->lqr_start_idx,
 600                        ((ost_idx != LOV_QOS_EMPTY) && lov->lov_tgts[ost_idx]) ?
 601                        lov->lov_tgts[ost_idx]->ltd_active : 0,
 602                        idx_pos - idx_arr, array_idx, ost_idx);
 603 #endif
 604                 if ((ost_idx == LOV_QOS_EMPTY) || !lov->lov_tgts[ost_idx] ||
 605                     !lov->lov_tgts[ost_idx]->ltd_active)
 606                         continue;
 607
 608                 /* Fail Check before osc_precreate() is called
 609                    so we can only 'fail' single OSC. */
 610                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
 611                         continue;
 612
 613                 /* Drop slow OSCs if we can */
 614                 if (obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed)
 615                         continue;
 616
 617                 *idx_pos = ost_idx;
 618                 idx_pos++;
 619                 /* We have enough stripes */
 620                 if (idx_pos - idx_arr == *stripe_cnt)
 621                         break;
 622         }
 623         if ((speed < 2) && (idx_pos - idx_arr < stripe_cnt_min)) {
 624                 /* Try again, allowing slower OSCs */
 625                 speed++;
 626                 lqr->lqr_start_idx = ost_start_idx_temp;
 627                 goto repeat_find;
 628         }
 629
 630         up_read(&lov->lov_qos.lq_rw_sem);
 631
 632         *stripe_cnt = idx_pos - idx_arr;
 633 out:
 634         if (pool != NULL) {
 635                 up_read(&pool_tgt_rw_sem(pool));
 636                 /* put back ref got by lov_find_pool() */
 637                 lov_pool_putref(pool);
 638         }
 639
 640         RETURN(rc);
 641 }
 642
 643 /* alloc objects on osts with specific stripe offset */
 644 static int alloc_specific(struct lov_obd *lov, struct lov_stripe_md *lsm,
 645                           int *idx_arr)
 646 {
 647         unsigned ost_idx, array_idx, ost_count;
 648         int i, rc, *idx_pos;
 649         int speed = 0;
 650         struct pool_desc *pool;
 651         struct ost_pool *osts;
 652         ENTRY;
 653
 654         pool = lov_find_pool(lov, lsm->lsm_pool_name);
 655         if (pool == NULL) {
 656                 osts = &(lov->lov_packed);
 657         } else {
 658                 down_read(&pool_tgt_rw_sem(pool));
 659                 osts = &(pool->pool_obds);
 660         }
 661
 662         ost_count = osts->op_count;
 663
 664 repeat_find:
 665         /* search loi_ost_idx in ost array */
 666         array_idx = 0;
 667         for (i = 0; i < ost_count; i++) {
 668                 if (osts->op_array[i] == lsm->lsm_oinfo[0]->loi_ost_idx) {
 669                         array_idx = i;
 670                         break;
 671                 }
 672         }
 673         if (i == ost_count) {
 674                 CERROR("Start index %d not found in pool '%s'\n",
 675                        lsm->lsm_oinfo[0]->loi_ost_idx, lsm->lsm_pool_name);
 676                 GOTO(out, rc = -EINVAL);
 677         }
 678
 679         idx_pos = idx_arr;
 680         for (i = 0; i < ost_count;
 681              i++, array_idx = (array_idx + 1) % ost_count) {
 682                 ost_idx = osts->op_array[array_idx];
 683
 684                 if (!lov->lov_tgts[ost_idx] ||
 685                     !lov->lov_tgts[ost_idx]->ltd_active) {
 686                         continue;
 687                 }
 688
 689                 /* Fail Check before osc_precreate() is called
 690                    so we can only 'fail' single OSC. */
 691                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
 692                         continue;
 693
 694                 /* Drop slow OSCs if we can, but not for requested start idx.
 695                  *
 696                  * This means "if OSC is slow and it is not the requested
 697                  * start OST, then it can be skipped, otherwise skip it only
 698                  * if it is inactive/recovering/out-of-space." */
 699                 if ((obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed) &&
 700                     (i != 0 || speed >= 2))
 701                         continue;
 702
 703                 *idx_pos = ost_idx;
 704                 idx_pos++;
 705                 /* We have enough stripes */
 706                 if (idx_pos - idx_arr == lsm->lsm_stripe_count)
 707                         GOTO(out, rc = 0);
 708         }
 709         if (speed < 2) {
 710                 /* Try again, allowing slower OSCs */
 711                 speed++;
 712                 goto repeat_find;
 713         }
 714
 715         /* If we were passed specific striping params, then a failure to
 716          * meet those requirements is an error, since we can't reallocate
 717          * that memory (it might be part of a larger array or something).
 718          *
 719          * We can only get here if lsm_stripe_count was originally > 1.
 720          */
 721         CERROR("can't lstripe objid "LPX64": have %d want %u\n",
 722                lsm->lsm_object_id, (int)(idx_pos - idx_arr),
 723                lsm->lsm_stripe_count);
 724         rc = -EFBIG;
 725 out:
 726         if (pool != NULL) {
 727                 up_read(&pool_tgt_rw_sem(pool));
 728                 /* put back ref got by lov_find_pool() */
 729                 lov_pool_putref(pool);
 730         }
 731
 732         RETURN(rc);
 733 }
 734
 735 /* Alloc objects on osts with optimization based on:
 736    - free space
 737    - network resources (shared OSS's)
 738 */
 739 static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt,
 740                      char *poolname, int flags)
 741 {
 742         struct lov_obd *lov = &exp->exp_obd->u.lov;
 743         static time_t last_warn = 0;
 744         time_t now = cfs_time_current_sec();
 745         __u64 total_bavail, total_weight = 0;
 746         int nfound, good_osts, i, warn = 0, rc = 0;
 747         int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
 748         struct pool_desc *pool;
 749         struct ost_pool *osts;
 750         struct lov_qos_rr *lqr;
 751         ENTRY;
 752
 753         if (stripe_cnt_min < 1)
 754                 RETURN(-EINVAL);
 755
 756         pool = lov_find_pool(lov, poolname);
 757         if (pool == NULL) {
 758                 osts = &(lov->lov_packed);
 759                 lqr = &(lov->lov_qos.lq_rr);
 760         } else {
 761                 down_read(&pool_tgt_rw_sem(pool));
 762                 osts = &(pool->pool_obds);
 763                 lqr = &(pool->pool_rr);
 764         }
 765
 766         lov_getref(exp->exp_obd);
 767
 768         /* Detect -EAGAIN early, before expensive lock is taken. */
 769         if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space)
 770                 GOTO(out_nolock, rc = -EAGAIN);
 771
 772         /* Do actual allocation, use write lock here. */
 773         down_write(&lov->lov_qos.lq_rw_sem);
 774
 775         /*
 776          * Check again, while we were sleeping on @lq_rw_sem things could
 777          * change.
 778          */
 779         if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space)
 780                 GOTO(out, rc = -EAGAIN);
 781
 782         if (lov->desc.ld_active_tgt_count < 2)
 783                 GOTO(out, rc = -EAGAIN);
 784
 785         rc = qos_calc_ppo(exp->exp_obd);
 786         if (rc)
 787                 GOTO(out, rc);
 788
 789         total_bavail = 0;
 790         good_osts = 0;
 791         /* Warn users about zero available space/inode every 30 min */
 792         if (cfs_time_sub(now, last_warn) > 60 * 30)
 793                 warn = 1;
 794         /* Find all the OSTs that are valid stripe candidates */
 795         for (i = 0; i < osts->op_count; i++) {
 796                 __u64 bavail;
 797
 798                 if (!lov->lov_tgts[osts->op_array[i]] ||
 799                     !lov->lov_tgts[osts->op_array[i]]->ltd_active)
 800                         continue;
 801                 bavail = TGT_BAVAIL(osts->op_array[i]);
 802                 if (!bavail) {
 803                         if (warn) {
 804                                 CDEBUG(D_QOS, "no free space on %s\n",
 805                                      obd_uuid2str(&lov->lov_tgts[osts->op_array[i]]->ltd_uuid));
 806                                 last_warn = now;
 807                         }
 808                         continue;
 809                 }
 810                 if (!TGT_FFREE(osts->op_array[i])) {
 811                         if (warn) {
 812                                 CDEBUG(D_QOS, "no free inodes on %s\n",
 813                                      obd_uuid2str(&lov->lov_tgts[osts->op_array[i]]->ltd_uuid));
 814                                 last_warn = now;
 815                         }
 816                         continue;
 817                 }
 818
 819                 /* Fail Check before osc_precreate() is called
 820                    so we can only 'fail' single OSC. */
 821                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && osts->op_array[i] == 0)
 822                         continue;
 823
 824                 if (obd_precreate(lov->lov_tgts[osts->op_array[i]]->ltd_exp) > 2)
 825                         continue;
 826
 827                 lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable = 1;
 828                 qos_calc_weight(lov, osts->op_array[i]);
 829                 total_bavail += bavail;
 830                 total_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight;
 831
 832                 good_osts++;
 833         }
 834
 835 #ifdef QOS_DEBUG
 836         CDEBUG(D_QOS, "found %d good osts\n", good_osts);
 837 #endif
 838
 839         if (good_osts < stripe_cnt_min)
 840                 GOTO(out, rc = -EAGAIN);
 841
 842         if (!total_bavail)
 843                 GOTO(out, rc = -ENOSPC);
 844
 845         /* We have enough osts */
 846         if (good_osts < *stripe_cnt)
 847                 *stripe_cnt = good_osts;
 848
 849         if (!*stripe_cnt)
 850                 GOTO(out, rc = -EAGAIN);
 851
 852         /* Find enough OSTs with weighted random allocation. */
 853         nfound = 0;
 854         while (nfound < *stripe_cnt) {
 855                 __u64 rand, cur_weight;
 856
 857                 cur_weight = 0;
 858                 rc = -ENODEV;
 859
 860                 if (total_weight) {
 861 #if BITS_PER_LONG == 32
 862                         rand = ll_rand() % (unsigned)total_weight;
 863                         /* If total_weight > 32-bit, first generate the high
 864                          * 32 bits of the random number, then add in the low
 865                          * 32 bits (truncated to the upper limit, if needed) */
 866                         if (total_weight > 0xffffffffULL)
 867                                 rand = (__u64)(ll_rand() %
 868                                           (unsigned)(total_weight >> 32)) << 32;
 869                         else
 870                                 rand = 0;
 871
 872                         if (rand == (total_weight & 0xffffffff00000000ULL))
 873                                 rand |= ll_rand() % (unsigned)total_weight;
 874                         else
 875                                 rand |= ll_rand();
 876
 877 #else
 878                         rand = ((__u64)ll_rand() << 32 | ll_rand()) %
 879                                 total_weight;
 880 #endif
 881                 } else {
 882                         rand = 0;
 883                 }
 884
 885                 /* On average, this will hit larger-weighted osts more often.
 886                    0-weight osts will always get used last (only when rand=0).*/
 887                 for (i = 0; i < osts->op_count; i++) {
 888                         if (!lov->lov_tgts[osts->op_array[i]] ||
 889                             !lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable)
 890                                 continue;
 891
 892                         cur_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight;
 893 #ifdef QOS_DEBUG
 894                         CDEBUG(D_QOS, "stripe_cnt=%d nfound=%d cur_weight="LPU64
 895                                       " rand="LPU64" total_weight="LPU64"\n",
 896                                *stripe_cnt, nfound, cur_weight, rand, total_weight);
 897 #endif
 898                         if (cur_weight >= rand) {
 899 #ifdef QOS_DEBUG
 900                                 CDEBUG(D_QOS, "assigned stripe=%d to idx=%d\n",
 901                                        nfound, osts->op_array[i]);
 902 #endif
 903                                 idx_arr[nfound++] = osts->op_array[i];
 904                                 qos_used(lov, osts, osts->op_array[i], &total_weight);
 905                                 rc = 0;
 906                                 break;
 907                         }
 908                 }
 909                 /* should never satisfy below condition */
 910                 if (rc) {
 911                         CERROR("Didn't find any OSTs?\n");
 912                         break;
 913                 }
 914         }
 915         LASSERT(nfound == *stripe_cnt);
 916
 917 out:
 918         up_write(&lov->lov_qos.lq_rw_sem);
 919
 920 out_nolock:
 921         if (pool != NULL) {
 922                 up_read(&pool_tgt_rw_sem(pool));
 923                 /* put back ref got by lov_find_pool() */
 924                 lov_pool_putref(pool);
 925         }
 926
 927         if (rc == -EAGAIN)
 928                 rc = alloc_rr(lov, idx_arr, stripe_cnt, poolname, flags);
 929
 930         lov_putref(exp->exp_obd);
 931         RETURN(rc);
 932 }
 933
 934 /* return new alloced stripe count on success */
 935 static int alloc_idx_array(struct obd_export *exp, struct lov_stripe_md *lsm,
 936                            int newea, int **idx_arr, int *arr_cnt, int flags)
 937 {
 938         struct lov_obd *lov = &exp->exp_obd->u.lov;
 939         int stripe_cnt = lsm->lsm_stripe_count;
 940         int i, rc = 0;
 941         int *tmp_arr = NULL;
 942         ENTRY;
 943
 944         *arr_cnt = stripe_cnt;
 945         OBD_ALLOC(tmp_arr, *arr_cnt * sizeof(int));
 946         if (tmp_arr == NULL)
 947                 RETURN(-ENOMEM);
 948         for (i = 0; i < *arr_cnt; i++)
 949                 tmp_arr[i] = -1;
 950
 951         if (newea ||
 952             lsm->lsm_oinfo[0]->loi_ost_idx >= lov->desc.ld_tgt_count)
 953                 rc = alloc_qos(exp, tmp_arr, &stripe_cnt,
 954                                lsm->lsm_pool_name, flags);
 955         else
 956                 rc = alloc_specific(lov, lsm, tmp_arr);
 957
 958         if (rc)
 959                 GOTO(out_arr, rc);
 960
 961         *idx_arr = tmp_arr;
 962         RETURN(stripe_cnt);
 963 out_arr:
 964         OBD_FREE(tmp_arr, *arr_cnt * sizeof(int));
 965         *arr_cnt = 0;
 966         RETURN(rc);
 967 }
 968
 969 static void free_idx_array(int *idx_arr, int arr_cnt)
 970 {
 971         if (arr_cnt)
 972                 OBD_FREE(idx_arr, arr_cnt * sizeof(int));
 973 }
 974
 975 int qos_prep_create(struct obd_export *exp, struct lov_request_set *set)
 976 {
 977         struct lov_obd *lov = &exp->exp_obd->u.lov;
 978         struct lov_stripe_md *lsm;
 979         struct obdo *src_oa = set->set_oi->oi_oa;
 980         struct obd_trans_info *oti = set->set_oti;
 981         int i, stripes, rc = 0, newea = 0;
 982         int flag = LOV_USES_ASSIGNED_STRIPE;
 983         int *idx_arr = NULL, idx_cnt = 0;
 984         ENTRY;
 985
 986         LASSERT(src_oa->o_valid & OBD_MD_FLID);
 987         LASSERT(src_oa->o_valid & OBD_MD_FLGROUP);
 988
 989         if (set->set_oi->oi_md == NULL) {
 990                 int stripes_def = lov_get_stripecnt(lov, 0);
 991
 992                 /* If the MDS file was truncated up to some size, stripe over
 993                  * enough OSTs to allow the file to be created at that size.
 994                  * This may mean we use more than the default # of stripes. */
 995                 if (src_oa->o_valid & OBD_MD_FLSIZE) {
 996                         obd_size min_bavail = LUSTRE_STRIPE_MAXBYTES;
 997
 998                         /* Find a small number of stripes we can use
 999                            (up to # of active osts). */
1000                         stripes = 1;
1001                         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1002                                 if (!lov->lov_tgts[i] ||
1003                                     !lov->lov_tgts[i]->ltd_active)
1004                                         continue;
1005                                 min_bavail = min(min_bavail, TGT_BAVAIL(i));
1006                                 if (min_bavail * stripes > src_oa->o_size)
1007                                         break;
1008                                 stripes++;
1009                         }
1010
1011                         if (stripes < stripes_def)
1012                                 stripes = stripes_def;
1013                 } else {
1014                          flag = LOV_USES_DEFAULT_STRIPE;
1015                          stripes = stripes_def;
1016                 }
1017
1018                 rc = lov_alloc_memmd(&set->set_oi->oi_md, stripes,
1019                                      lov->desc.ld_pattern ?
1020                                      lov->desc.ld_pattern : LOV_PATTERN_RAID0,
1021                                      LOV_MAGIC);
1022                 if (rc < 0)
1023                         GOTO(out_err, rc);
1024                 newea = 1;
1025                 rc = 0;
1026         }
1027
1028         lsm = set->set_oi->oi_md;
1029         lsm->lsm_object_id = src_oa->o_id;
1030         lsm->lsm_object_gr = src_oa->o_gr;
1031
1032         if (!lsm->lsm_stripe_size)
1033                 lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size;
1034         if (!lsm->lsm_pattern) {
1035                 LASSERT(lov->desc.ld_pattern);
1036                 lsm->lsm_pattern = lov->desc.ld_pattern;
1037         }
1038
1039         stripes = alloc_idx_array(exp, lsm, newea, &idx_arr, &idx_cnt, flag);
1040         if (stripes <= 0)
1041                 GOTO(out_err, rc = stripes ? stripes : -EIO);
1042         LASSERTF(stripes <= lsm->lsm_stripe_count,"requested %d allocated %d\n",
1043                  lsm->lsm_stripe_count, stripes);
1044
1045         for (i = 0; i < stripes; i++) {
1046                 struct lov_request *req;
1047                 int ost_idx = idx_arr[i];
1048                 LASSERT(ost_idx >= 0);
1049
1050                 OBD_ALLOC(req, sizeof(*req));
1051                 if (req == NULL)
1052                         GOTO(out_err, rc = -ENOMEM);
1053                 lov_set_add_req(req, set);
1054
1055                 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
1056                 OBD_ALLOC(req->rq_oi.oi_md, req->rq_buflen);
1057                 if (req->rq_oi.oi_md == NULL)
1058                         GOTO(out_err, rc = -ENOMEM);
1059
1060                 OBDO_ALLOC(req->rq_oi.oi_oa);
1061                 if (req->rq_oi.oi_oa == NULL)
1062                         GOTO(out_err, rc = -ENOMEM);
1063
1064                 req->rq_idx = ost_idx;
1065                 req->rq_stripe = i;
1066                 /* create data objects with "parent" OA */
1067                 memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
1068
1069                 /* XXX When we start creating objects on demand, we need to
1070                  *     make sure that we always create the object on the
1071                  *     stripe which holds the existing file size.
1072                  */
1073                 if (src_oa->o_valid & OBD_MD_FLSIZE) {
1074                         req->rq_oi.oi_oa->o_size =
1075                                 lov_size_to_stripe(lsm, src_oa->o_size, i);
1076
1077                         CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
1078                                i, req->rq_oi.oi_oa->o_size, src_oa->o_size);
1079                 }
1080         }
1081         LASSERT(set->set_count == stripes);
1082
1083         if (stripes < lsm->lsm_stripe_count)
1084                 qos_shrink_lsm(set);
1085
1086         if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) {
1087                 oti_alloc_cookies(oti, set->set_count);
1088                 if (!oti->oti_logcookies)
1089                         GOTO(out_err, rc = -ENOMEM);
1090                 set->set_cookies = oti->oti_logcookies;
1091         }
1092 out_err:
1093         if (newea && rc)
1094                 obd_free_memmd(exp, &set->set_oi->oi_md);
1095         if (idx_arr)
1096                 free_idx_array(idx_arr, idx_cnt);
1097         EXIT;
1098         return rc;
1099 }
1100
1101 void qos_update(struct lov_obd *lov)
1102 {
1103         ENTRY;
1104         lov->lov_qos.lq_dirty = 1;
1105 }