lustre/lov/lov_qos.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #ifndef EXPORT_SYMTAB
  38 # define EXPORT_SYMTAB
  39 #endif
  40 #define DEBUG_SUBSYSTEM S_LOV
  41
  42 #ifdef __KERNEL__
  43 #include <libcfs/libcfs.h>
  44 #else
  45 #include <liblustre.h>
  46 #endif
  47
  48 #include <obd_class.h>
  49 #include <obd_lov.h>
  50 #include <lustre/lustre_idl.h>
  51 #include "lov_internal.h"
  52
  53 /* #define QOS_DEBUG 1 */
  54 #define D_QOS D_OTHER
  55
  56 #define TGT_BAVAIL(i) (lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bavail *\
  57                        lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bsize)
  58 #define TGT_FFREE(i)  (lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_ffree)
  59
  60
  61 int qos_add_tgt(struct obd_device *obd, __u32 index)
  62 {
  63         struct lov_obd *lov = &obd->u.lov;
  64         struct lov_qos_oss *oss, *temposs;
  65         struct obd_export *exp = lov->lov_tgts[index]->ltd_exp;
  66         int rc = 0, found = 0;
  67         ENTRY;
  68
  69         /* We only need this QOS struct on MDT, not clients - but we may not
  70          * have registered the LOV's observer yet, so there's no way to know */
  71         if (!exp || !exp->exp_connection) {
  72                 CERROR("Missing connection\n");
  73                 RETURN(-ENOTCONN);
  74         }
  75
  76         down_write(&lov->lov_qos.lq_rw_sem);
  77         mutex_down(&lov->lov_lock);
  78         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
  79                 if (obd_uuid_equals(&oss->lqo_uuid,
  80                                     &exp->exp_connection->c_remote_uuid)) {
  81                         found++;
  82                         break;
  83                 }
  84         }
  85
  86         if (!found) {
  87                 OBD_ALLOC_PTR(oss);
  88                 if (!oss)
  89                         GOTO(out, rc = -ENOMEM);
  90                 memcpy(&oss->lqo_uuid,
  91                        &exp->exp_connection->c_remote_uuid,
  92                        sizeof(oss->lqo_uuid));
  93         } else {
  94                 /* Assume we have to move this one */
  95                 list_del(&oss->lqo_oss_list);
  96         }
  97
  98         oss->lqo_ost_count++;
  99         lov->lov_tgts[index]->ltd_qos.ltq_oss = oss;
 100
 101         /* Add sorted by # of OSTs.  Find the first entry that we're
 102            bigger than... */
 103         list_for_each_entry(temposs, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 104                 if (oss->lqo_ost_count > temposs->lqo_ost_count)
 105                         break;
 106         }
 107         /* ...and add before it.  If we're the first or smallest, temposs
 108            points to the list head, and we add to the end. */
 109         list_add_tail(&oss->lqo_oss_list, &temposs->lqo_oss_list);
 110
 111         lov->lov_qos.lq_dirty = 1;
 112         lov->lov_qos.lq_rr.lqr_dirty = 1;
 113
 114         CDEBUG(D_QOS, "add tgt %s to OSS %s (%d OSTs)\n",
 115                obd_uuid2str(&lov->lov_tgts[index]->ltd_uuid),
 116                obd_uuid2str(&oss->lqo_uuid),
 117                oss->lqo_ost_count);
 118
 119 out:
 120         mutex_up(&lov->lov_lock);
 121         up_write(&lov->lov_qos.lq_rw_sem);
 122         RETURN(rc);
 123 }
 124
 125 int qos_del_tgt(struct obd_device *obd, struct lov_tgt_desc *tgt)
 126 {
 127         struct lov_obd *lov = &obd->u.lov;
 128         struct lov_qos_oss *oss;
 129         int rc = 0;
 130         ENTRY;
 131
 132         down_write(&lov->lov_qos.lq_rw_sem);
 133
 134         oss = tgt->ltd_qos.ltq_oss;
 135         if (!oss)
 136                 GOTO(out, rc = -ENOENT);
 137
 138         oss->lqo_ost_count--;
 139         if (oss->lqo_ost_count == 0) {
 140                 CDEBUG(D_QOS, "removing OSS %s\n",
 141                        obd_uuid2str(&oss->lqo_uuid));
 142                 list_del(&oss->lqo_oss_list);
 143                 OBD_FREE_PTR(oss);
 144         }
 145
 146         lov->lov_qos.lq_dirty = 1;
 147         lov->lov_qos.lq_rr.lqr_dirty = 1;
 148 out:
 149         up_write(&lov->lov_qos.lq_rw_sem);
 150         RETURN(rc);
 151 }
 152
 153 /* Recalculate per-object penalties for OSSs and OSTs,
 154    depends on size of each ost in an oss */
 155 static int qos_calc_ppo(struct obd_device *obd)
 156 {
 157         struct lov_obd *lov = &obd->u.lov;
 158         struct lov_qos_oss *oss;
 159         __u64 ba_max, ba_min, temp;
 160         __u32 num_active;
 161         int rc, i, prio_wide;
 162         time_t now, age;
 163         ENTRY;
 164
 165         if (!lov->lov_qos.lq_dirty)
 166                 GOTO(out, rc = 0);
 167
 168         num_active = lov->desc.ld_active_tgt_count - 1;
 169         if (num_active < 1)
 170                 GOTO(out, rc = -EAGAIN);
 171
 172         /* find bavail on each OSS */
 173         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 174                 oss->lqo_bavail = 0;
 175         }
 176         lov->lov_qos.lq_active_oss_count = 0;
 177
 178         /* How badly user wants to select osts "widely" (not recently chosen
 179            and not on recent oss's).  As opposed to "freely" (free space
 180            avail.) 0-256. */
 181         prio_wide = 256 - lov->lov_qos.lq_prio_free;
 182
 183         ba_min = (__u64)(-1);
 184         ba_max = 0;
 185         now = cfs_time_current_sec();
 186         /* Calculate OST penalty per object */
 187         /* (lov ref taken in alloc_qos) */
 188         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 189                 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
 190                         continue;
 191                 temp = TGT_BAVAIL(i);
 192                 if (!temp)
 193                         continue;
 194                 ba_min = min(temp, ba_min);
 195                 ba_max = max(temp, ba_max);
 196
 197                 /* Count the number of usable OSS's */
 198                 if (lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_bavail == 0)
 199                         lov->lov_qos.lq_active_oss_count++;
 200                 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_bavail += temp;
 201
 202                 /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
 203                 temp >>= 1;
 204                 do_div(temp, num_active);
 205                 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj =
 206                         (temp * prio_wide) >> 8;
 207
 208                 age = (now - lov->lov_tgts[i]->ltd_qos.ltq_used) >> 3;
 209                 if (lov->lov_qos.lq_reset || age > 32 * lov->desc.ld_qos_maxage)
 210                         lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
 211                 else if (age > lov->desc.ld_qos_maxage)
 212                         /* Decay the penalty by half for every 8x the update
 213                          * interval that the device has been idle.  That gives
 214                          * lots of time for the statfs information to be
 215                          * updated (which the penalty is only a proxy for),
 216                          * and avoids penalizing OSS/OSTs under light load. */
 217                         lov->lov_tgts[i]->ltd_qos.ltq_penalty >>=
 218                                 (age / lov->desc.ld_qos_maxage);
 219         }
 220
 221         num_active = lov->lov_qos.lq_active_oss_count - 1;
 222         if (num_active < 1) {
 223                 /* If there's only 1 OSS, we can't penalize it, so instead
 224                    we have to double the OST penalty */
 225                 num_active = 1;
 226                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 227                         if (lov->lov_tgts[i] == NULL)
 228                                 continue;
 229                         lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj <<= 1;
 230                 }
 231         }
 232
 233         /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
 234         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 235                 temp = oss->lqo_bavail >> 1;
 236                 do_div(temp, oss->lqo_ost_count * num_active);
 237                 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
 238
 239                 age = (now - oss->lqo_used) >> 3;
 240                 if (lov->lov_qos.lq_reset || age > 32 * lov->desc.ld_qos_maxage)
 241                         oss->lqo_penalty = 0;
 242                 else if (age > lov->desc.ld_qos_maxage)
 243                         /* Decay the penalty by half for every 8x the update
 244                          * interval that the device has been idle.  That gives
 245                          * lots of time for the statfs information to be
 246                          * updated (which the penalty is only a proxy for),
 247                          * and avoids penalizing OSS/OSTs under light load. */
 248                         oss->lqo_penalty >>= (age / lov->desc.ld_qos_maxage);
 249         }
 250
 251         lov->lov_qos.lq_dirty = 0;
 252         lov->lov_qos.lq_reset = 0;
 253
 254         /* If each ost has almost same free space,
 255          * do rr allocation for better creation performance */
 256         lov->lov_qos.lq_same_space = 0;
 257         if ((ba_max * (256 - lov->lov_qos.lq_threshold_rr)) >> 8 < ba_min) {
 258                 lov->lov_qos.lq_same_space = 1;
 259                 /* Reset weights for the next time we enter qos mode */
 260                 lov->lov_qos.lq_reset = 1;
 261         }
 262         rc = 0;
 263
 264 out:
 265         if (!rc && lov->lov_qos.lq_same_space)
 266                 RETURN(-EAGAIN);
 267         RETURN(rc);
 268 }
 269
 270 static int qos_calc_weight(struct lov_obd *lov, int i)
 271 {
 272         __u64 temp, temp2;
 273
 274         /* Final ost weight = TGT_BAVAIL - ost_penalty - oss_penalty */
 275         temp = TGT_BAVAIL(i);
 276         temp2 = lov->lov_tgts[i]->ltd_qos.ltq_penalty +
 277                 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty;
 278         if (temp < temp2)
 279                 lov->lov_tgts[i]->ltd_qos.ltq_weight = 0;
 280         else
 281                 lov->lov_tgts[i]->ltd_qos.ltq_weight = temp - temp2;
 282         return 0;
 283 }
 284
 285 /* We just used this index for a stripe; adjust everyone's weights */
 286 static int qos_used(struct lov_obd *lov, struct ost_pool *osts,
 287                     __u32 index, __u64 *total_wt)
 288 {
 289         struct lov_qos_oss *oss;
 290         int j;
 291         ENTRY;
 292
 293         /* Don't allocate from this stripe anymore, until the next alloc_qos */
 294         lov->lov_tgts[index]->ltd_qos.ltq_usable = 0;
 295
 296         oss = lov->lov_tgts[index]->ltd_qos.ltq_oss;
 297
 298         /* Decay old penalty by half (we're adding max penalty, and don't
 299            want it to run away.) */
 300         lov->lov_tgts[index]->ltd_qos.ltq_penalty >>= 1;
 301         oss->lqo_penalty >>= 1;
 302
 303         /* mark the OSS and OST as recently used */
 304         lov->lov_tgts[index]->ltd_qos.ltq_used =
 305                 oss->lqo_used = cfs_time_current_sec();
 306
 307         /* Set max penalties for this OST and OSS */
 308         lov->lov_tgts[index]->ltd_qos.ltq_penalty +=
 309                 lov->lov_tgts[index]->ltd_qos.ltq_penalty_per_obj *
 310                 lov->desc.ld_active_tgt_count;
 311         oss->lqo_penalty += oss->lqo_penalty_per_obj *
 312                 lov->lov_qos.lq_active_oss_count;
 313
 314         /* Decrease all OSS penalties */
 315         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 316                 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
 317                         oss->lqo_penalty = 0;
 318                 else
 319                         oss->lqo_penalty -= oss->lqo_penalty_per_obj;
 320         }
 321
 322         *total_wt = 0;
 323         /* Decrease all OST penalties */
 324         for (j = 0; j < osts->op_count; j++) {
 325                 int i;
 326
 327                 i = osts->op_array[j];
 328                 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
 329                         continue;
 330                 if (lov->lov_tgts[i]->ltd_qos.ltq_penalty <
 331                     lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj)
 332                         lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
 333                 else
 334                         lov->lov_tgts[i]->ltd_qos.ltq_penalty -=
 335                         lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj;
 336
 337                 qos_calc_weight(lov, i);
 338
 339                 /* Recalc the total weight of usable osts */
 340                 if (lov->lov_tgts[i]->ltd_qos.ltq_usable)
 341                         *total_wt += lov->lov_tgts[i]->ltd_qos.ltq_weight;
 342
 343 #ifdef QOS_DEBUG
 344                 CDEBUG(D_QOS, "recalc tgt %d usable=%d avail="LPU64
 345                        " ostppo="LPU64" ostp="LPU64" ossppo="LPU64
 346                        " ossp="LPU64" wt="LPU64"\n",
 347                        i, lov->lov_tgts[i]->ltd_qos.ltq_usable,
 348                        TGT_BAVAIL(i) >> 10,
 349                        lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj >> 10,
 350                        lov->lov_tgts[i]->ltd_qos.ltq_penalty >> 10,
 351                        lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty_per_obj>>10,
 352                        lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty >> 10,
 353                        lov->lov_tgts[i]->ltd_qos.ltq_weight >> 10);
 354 #endif
 355         }
 356
 357         RETURN(0);
 358 }
 359
 360 #define LOV_QOS_EMPTY ((__u32)-1)
 361 /* compute optimal round-robin order, based on OSTs per OSS */
 362 static int qos_calc_rr(struct lov_obd *lov, struct ost_pool *src_pool,
 363                        struct lov_qos_rr *lqr)
 364 {
 365         struct lov_qos_oss *oss;
 366         unsigned placed, real_count;
 367         int i, rc;
 368         ENTRY;
 369
 370         if (!lqr->lqr_dirty) {
 371                 LASSERT(lqr->lqr_pool.op_size);
 372                 RETURN(0);
 373         }
 374
 375         /* Do actual allocation. */
 376         down_write(&lov->lov_qos.lq_rw_sem);
 377
 378         /*
 379          * Check again. While we were sleeping on @lq_rw_sem something could
 380          * change.
 381          */
 382         if (!lqr->lqr_dirty) {
 383                 LASSERT(lqr->lqr_pool.op_size);
 384                 up_write(&lov->lov_qos.lq_rw_sem);
 385                 RETURN(0);
 386         }
 387
 388         real_count = src_pool->op_count;
 389
 390         /* Zero the pool array */
 391         /* alloc_rr is holding a read lock on the pool, so nobody is adding/
 392            deleting from the pool. The lq_rw_sem insures that nobody else
 393            is reading. */
 394         lqr->lqr_pool.op_count = real_count;
 395         rc = lov_ost_pool_extend(&lqr->lqr_pool, real_count);
 396         if (rc) {
 397                 up_write(&lov->lov_qos.lq_rw_sem);
 398                 RETURN(rc);
 399         }
 400         for (i = 0; i < lqr->lqr_pool.op_count; i++)
 401                 lqr->lqr_pool.op_array[i] = LOV_QOS_EMPTY;
 402
 403         /* Place all the OSTs from 1 OSS at the same time. */
 404         placed = 0;
 405         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 406                 int j = 0;
 407                 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
 408                         if (lov->lov_tgts[src_pool->op_array[i]] &&
 409                             (lov->lov_tgts[src_pool->op_array[i]]->ltd_qos.ltq_oss == oss)) {
 410                               /* Evenly space these OSTs across arrayspace */
 411                               int next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count;
 412                               while (lqr->lqr_pool.op_array[next] !=
 413                                      LOV_QOS_EMPTY)
 414                                         next = (next + 1) % lqr->lqr_pool.op_count;
 415                               lqr->lqr_pool.op_array[next] = src_pool->op_array[i];
 416                               j++;
 417                               placed++;
 418                         }
 419                 }
 420         }
 421
 422         lqr->lqr_dirty = 0;
 423         up_write(&lov->lov_qos.lq_rw_sem);
 424
 425         if (placed != real_count) {
 426                 /* This should never happen */
 427                 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
 428                                    "round-robin list (%d of %d).\n",
 429                                    placed, real_count);
 430                 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
 431                         LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
 432                                  lqr->lqr_pool.op_array[i]);
 433                 }
 434                 lqr->lqr_dirty = 1;
 435                 RETURN(-EAGAIN);
 436         }
 437
 438 #ifdef QOS_DEBUG
 439         for (i = 0; i < lqr->lqr_pool.op_count; i++) {
 440                 LCONSOLE(D_QOS, "rr #%d ost idx=%d\n", i,
 441                          lqr->lqr_pool.op_array[i]);
 442         }
 443 #endif
 444
 445         RETURN(0);
 446 }
 447
 448
 449 void qos_shrink_lsm(struct lov_request_set *set)
 450 {
 451         struct lov_stripe_md *lsm = set->set_oi->oi_md, *lsm_new;
 452         /* XXX LOV STACKING call into osc for sizes */
 453         unsigned oldsize, newsize;
 454
 455         if (set->set_oti && set->set_cookies && set->set_cookie_sent) {
 456                 struct llog_cookie *cookies;
 457                 oldsize = lsm->lsm_stripe_count * sizeof(*cookies);
 458                 newsize = set->set_count * sizeof(*cookies);
 459
 460                 cookies = set->set_cookies;
 461                 oti_alloc_cookies(set->set_oti, set->set_count);
 462                 if (set->set_oti->oti_logcookies) {
 463                         memcpy(set->set_oti->oti_logcookies, cookies, newsize);
 464                         OBD_FREE(cookies, oldsize);
 465                         set->set_cookies = set->set_oti->oti_logcookies;
 466                 } else {
 467                         CWARN("'leaking' %d bytes\n", oldsize - newsize);
 468                 }
 469         }
 470
 471         CWARN("using fewer stripes for object "LPU64": old %u new %u\n",
 472               lsm->lsm_object_id, lsm->lsm_stripe_count, set->set_count);
 473         LASSERT(lsm->lsm_stripe_count >= set->set_count);
 474
 475         newsize = lov_stripe_md_size(set->set_count);
 476         OBD_ALLOC(lsm_new, newsize);
 477         if (lsm_new != NULL) {
 478                 int i;
 479                 memcpy(lsm_new, lsm, sizeof(*lsm));
 480                 for (i = 0; i < lsm->lsm_stripe_count; i++) {
 481                         if (i < set->set_count) {
 482                                 lsm_new->lsm_oinfo[i] = lsm->lsm_oinfo[i];
 483                                 continue;
 484                         }
 485                         OBD_SLAB_FREE(lsm->lsm_oinfo[i], lov_oinfo_slab,
 486                                       sizeof(struct lov_oinfo));
 487                 }
 488                 lsm_new->lsm_stripe_count = set->set_count;
 489                 OBD_FREE(lsm, sizeof(struct lov_stripe_md) +
 490                          lsm->lsm_stripe_count * sizeof(struct lov_oinfo *));
 491                 set->set_oi->oi_md = lsm_new;
 492         } else {
 493                 CWARN("'leaking' few bytes\n");
 494         }
 495 }
 496
 497 int qos_remedy_create(struct lov_request_set *set, struct lov_request *req)
 498 {
 499         struct lov_stripe_md *lsm = set->set_oi->oi_md;
 500         struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
 501         unsigned ost_idx, ost_count = lov->desc.ld_tgt_count;
 502         int stripe, i, rc = -EIO;
 503         ENTRY;
 504
 505         ost_idx = (req->rq_idx + lsm->lsm_stripe_count) % ost_count;
 506         for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
 507                 if (!lov->lov_tgts[ost_idx] ||
 508                     !lov->lov_tgts[ost_idx]->ltd_active)
 509                         continue;
 510                 /* check if objects has been created on this ost */
 511                 for (stripe = 0; stripe < lsm->lsm_stripe_count; stripe++) {
 512                         /* we try send create to this ost but he is failed */
 513                         if (stripe == req->rq_stripe)
 514                                 continue;
 515                         /* already have object at this stripe */
 516                         if (ost_idx == lsm->lsm_oinfo[stripe]->loi_ost_idx)
 517                                 break;
 518                 }
 519                 if (stripe >= lsm->lsm_stripe_count) {
 520                         req->rq_idx = ost_idx;
 521                         rc = obd_create(lov->lov_tgts[ost_idx]->ltd_exp,
 522                                         req->rq_oi.oi_oa, &req->rq_oi.oi_md,
 523                                         set->set_oti);
 524                         if (!rc)
 525                                 break;
 526                 }
 527         }
 528         RETURN(rc);
 529 }
 530
 531 static int min_stripe_count(int stripe_cnt, int flags)
 532 {
 533         return (flags & LOV_USES_DEFAULT_STRIPE ?
 534                 stripe_cnt - (stripe_cnt / 4) : stripe_cnt);
 535 }
 536
 537 #define LOV_CREATE_RESEED_MULT 4
 538 #define LOV_CREATE_RESEED_MIN  1000
 539 /* Allocate objects on osts with round-robin algorithm */
 540 static int alloc_rr(struct lov_obd *lov, int *idx_arr, int *stripe_cnt,
 541                     char *poolname, int flags)
 542 {
 543         unsigned array_idx;
 544         int i, rc, *idx_pos;
 545         __u32 ost_idx;
 546         int ost_start_idx_temp;
 547         int speed = 0;
 548         int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
 549         struct pool_desc *pool;
 550         struct ost_pool *osts;
 551         struct lov_qos_rr *lqr;
 552         ENTRY;
 553
 554         pool = lov_find_pool(lov, poolname);
 555         if (pool == NULL) {
 556                 osts = &(lov->lov_packed);
 557                 lqr = &(lov->lov_qos.lq_rr);
 558         } else {
 559                 down_read(&pool_tgt_rw_sem(pool));
 560                 osts = &(pool->pool_obds);
 561                 lqr = &(pool->pool_rr);
 562         }
 563
 564         rc = qos_calc_rr(lov, osts, lqr);
 565         if (rc)
 566                 GOTO(out, rc);
 567
 568         if (--lqr->lqr_start_count <= 0) {
 569                 lqr->lqr_start_idx = ll_rand() % osts->op_count;
 570                 lqr->lqr_start_count =
 571                         (LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) +
 572                          LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U);
 573         } else if (stripe_cnt_min >= osts->op_count ||
 574                    lqr->lqr_start_idx > osts->op_count) {
 575                 /* If we have allocated from all of the OSTs, slowly
 576                  * precess the next start if the OST/stripe count isn't
 577                  * already doing this for us. */
 578                 lqr->lqr_start_idx %= osts->op_count;
 579                 if (*stripe_cnt > 1 && (osts->op_count % (*stripe_cnt)) != 1)
 580                         ++lqr->lqr_offset_idx;
 581         }
 582         down_read(&lov->lov_qos.lq_rw_sem);
 583         ost_start_idx_temp = lqr->lqr_start_idx;
 584
 585 repeat_find:
 586         array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) % osts->op_count;
 587         idx_pos = idx_arr;
 588 #ifdef QOS_DEBUG
 589         CDEBUG(D_QOS, "pool '%s' want %d startidx %d startcnt %d offset %d "
 590                "active %d count %d arrayidx %d\n", poolname,
 591                *stripe_cnt, lqr->lqr_start_idx, lqr->lqr_start_count,
 592                lqr->lqr_offset_idx, osts->op_count, osts->op_count, array_idx);
 593 #endif
 594
 595         for (i = 0; i < osts->op_count;
 596                     i++, array_idx=(array_idx + 1) % osts->op_count) {
 597                 ++lqr->lqr_start_idx;
 598                 ost_idx = lqr->lqr_pool.op_array[array_idx];
 599 #ifdef QOS_DEBUG
 600                 CDEBUG(D_QOS, "#%d strt %d act %d strp %d ary %d idx %d\n",
 601                        i, lqr->lqr_start_idx,
 602                        ((ost_idx != LOV_QOS_EMPTY) && lov->lov_tgts[ost_idx]) ?
 603                        lov->lov_tgts[ost_idx]->ltd_active : 0,
 604                        idx_pos - idx_arr, array_idx, ost_idx);
 605 #endif
 606                 if ((ost_idx == LOV_QOS_EMPTY) || !lov->lov_tgts[ost_idx] ||
 607                     !lov->lov_tgts[ost_idx]->ltd_active)
 608                         continue;
 609
 610                 /* Fail Check before osc_precreate() is called
 611                    so we can only 'fail' single OSC. */
 612                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
 613                         continue;
 614
 615                 /* Drop slow OSCs if we can */
 616                 if (obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed)
 617                         continue;
 618
 619                 *idx_pos = ost_idx;
 620                 idx_pos++;
 621                 /* We have enough stripes */
 622                 if (idx_pos - idx_arr == *stripe_cnt)
 623                         break;
 624         }
 625         if ((speed < 2) && (idx_pos - idx_arr < stripe_cnt_min)) {
 626                 /* Try again, allowing slower OSCs */
 627                 speed++;
 628                 lqr->lqr_start_idx = ost_start_idx_temp;
 629                 goto repeat_find;
 630         }
 631
 632         up_read(&lov->lov_qos.lq_rw_sem);
 633
 634         *stripe_cnt = idx_pos - idx_arr;
 635 out:
 636         if (pool != NULL) {
 637                 up_read(&pool_tgt_rw_sem(pool));
 638                 /* put back ref got by lov_find_pool() */
 639                 lov_pool_putref(pool);
 640         }
 641
 642         RETURN(rc);
 643 }
 644
 645 /* alloc objects on osts with specific stripe offset */
 646 static int alloc_specific(struct lov_obd *lov, struct lov_stripe_md *lsm,
 647                           int *idx_arr)
 648 {
 649         unsigned ost_idx, array_idx, ost_count;
 650         int i, rc, *idx_pos;
 651         int speed = 0;
 652         struct pool_desc *pool;
 653         struct ost_pool *osts;
 654         ENTRY;
 655
 656         pool = lov_find_pool(lov, lsm->lsm_pool_name);
 657         if (pool == NULL) {
 658                 osts = &(lov->lov_packed);
 659         } else {
 660                 down_read(&pool_tgt_rw_sem(pool));
 661                 osts = &(pool->pool_obds);
 662         }
 663
 664         ost_count = osts->op_count;
 665
 666 repeat_find:
 667         /* search loi_ost_idx in ost array */
 668         array_idx = 0;
 669         for (i = 0; i < ost_count; i++) {
 670                 if (osts->op_array[i] == lsm->lsm_oinfo[0]->loi_ost_idx) {
 671                         array_idx = i;
 672                         break;
 673                 }
 674         }
 675         if (i == ost_count) {
 676                 CERROR("Start index %d not found in pool '%s'\n",
 677                        lsm->lsm_oinfo[0]->loi_ost_idx, lsm->lsm_pool_name);
 678                 GOTO(out, rc = -EINVAL);
 679         }
 680
 681         idx_pos = idx_arr;
 682         for (i = 0; i < ost_count;
 683              i++, array_idx = (array_idx + 1) % ost_count) {
 684                 ost_idx = osts->op_array[array_idx];
 685
 686                 if (!lov->lov_tgts[ost_idx] ||
 687                     !lov->lov_tgts[ost_idx]->ltd_active) {
 688                         continue;
 689                 }
 690
 691                 /* Fail Check before osc_precreate() is called
 692                    so we can only 'fail' single OSC. */
 693                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
 694                         continue;
 695
 696                 /* Drop slow OSCs if we can, but not for requested start idx.
 697                  *
 698                  * This means "if OSC is slow and it is not the requested
 699                  * start OST, then it can be skipped, otherwise skip it only
 700                  * if it is inactive/recovering/out-of-space." */
 701                 if ((obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed) &&
 702                     (i != 0 || speed >= 2))
 703                         continue;
 704
 705                 *idx_pos = ost_idx;
 706                 idx_pos++;
 707                 /* We have enough stripes */
 708                 if (idx_pos - idx_arr == lsm->lsm_stripe_count)
 709                         GOTO(out, rc = 0);
 710         }
 711         if (speed < 2) {
 712                 /* Try again, allowing slower OSCs */
 713                 speed++;
 714                 goto repeat_find;
 715         }
 716
 717         /* If we were passed specific striping params, then a failure to
 718          * meet those requirements is an error, since we can't reallocate
 719          * that memory (it might be part of a larger array or something).
 720          *
 721          * We can only get here if lsm_stripe_count was originally > 1.
 722          */
 723         CERROR("can't lstripe objid "LPX64": have %d want %u\n",
 724                lsm->lsm_object_id, (int)(idx_pos - idx_arr),
 725                lsm->lsm_stripe_count);
 726         rc = -EFBIG;
 727 out:
 728         if (pool != NULL) {
 729                 up_read(&pool_tgt_rw_sem(pool));
 730                 /* put back ref got by lov_find_pool() */
 731                 lov_pool_putref(pool);
 732         }
 733
 734         RETURN(rc);
 735 }
 736
 737 /* Alloc objects on osts with optimization based on:
 738    - free space
 739    - network resources (shared OSS's)
 740 */
 741 static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt,
 742                      char *poolname, int flags)
 743 {
 744         struct lov_obd *lov = &exp->exp_obd->u.lov;
 745         static time_t last_warn = 0;
 746         time_t now = cfs_time_current_sec();
 747         __u64 total_bavail, total_weight = 0;
 748         int nfound, good_osts, i, warn = 0, rc = 0;
 749         int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
 750         struct pool_desc *pool;
 751         struct ost_pool *osts;
 752         struct lov_qos_rr *lqr;
 753         ENTRY;
 754
 755         if (stripe_cnt_min < 1)
 756                 RETURN(-EINVAL);
 757
 758         pool = lov_find_pool(lov, poolname);
 759         if (pool == NULL) {
 760                 osts = &(lov->lov_packed);
 761                 lqr = &(lov->lov_qos.lq_rr);
 762         } else {
 763                 down_read(&pool_tgt_rw_sem(pool));
 764                 osts = &(pool->pool_obds);
 765                 lqr = &(pool->pool_rr);
 766         }
 767
 768         obd_getref(exp->exp_obd);
 769
 770         /* wait for fresh statfs info if needed, the rpcs are sent in
 771          * lov_create() */
 772         qos_statfs_update(exp->exp_obd,
 773                           cfs_time_shift_64(-2 * lov->desc.ld_qos_maxage), 1);
 774
 775         /* Detect -EAGAIN early, before expensive lock is taken. */
 776         if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space)
 777                 GOTO(out_nolock, rc = -EAGAIN);
 778
 779         /* Do actual allocation, use write lock here. */
 780         down_write(&lov->lov_qos.lq_rw_sem);
 781
 782         /*
 783          * Check again, while we were sleeping on @lq_rw_sem things could
 784          * change.
 785          */
 786         if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space)
 787                 GOTO(out, rc = -EAGAIN);
 788
 789         if (lov->desc.ld_active_tgt_count < 2)
 790                 GOTO(out, rc = -EAGAIN);
 791
 792         rc = qos_calc_ppo(exp->exp_obd);
 793         if (rc)
 794                 GOTO(out, rc);
 795
 796         total_bavail = 0;
 797         good_osts = 0;
 798         /* Warn users about zero available space/inode every 30 min */
 799         if (cfs_time_sub(now, last_warn) > 60 * 30)
 800                 warn = 1;
 801         /* Find all the OSTs that are valid stripe candidates */
 802         for (i = 0; i < osts->op_count; i++) {
 803                 __u64 bavail;
 804
 805                 if (!lov->lov_tgts[osts->op_array[i]] ||
 806                     !lov->lov_tgts[osts->op_array[i]]->ltd_active)
 807                         continue;
 808                 bavail = TGT_BAVAIL(osts->op_array[i]);
 809                 if (!bavail) {
 810                         if (warn) {
 811                                 CDEBUG(D_QOS, "no free space on %s\n",
 812                                      obd_uuid2str(&lov->lov_tgts[osts->op_array[i]]->ltd_uuid));
 813                                 last_warn = now;
 814                         }
 815                         continue;
 816                 }
 817                 if (!TGT_FFREE(osts->op_array[i])) {
 818                         if (warn) {
 819                                 CDEBUG(D_QOS, "no free inodes on %s\n",
 820                                      obd_uuid2str(&lov->lov_tgts[osts->op_array[i]]->ltd_uuid));
 821                                 last_warn = now;
 822                         }
 823                         continue;
 824                 }
 825
 826                 /* Fail Check before osc_precreate() is called
 827                    so we can only 'fail' single OSC. */
 828                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && osts->op_array[i] == 0)
 829                         continue;
 830
 831                 if (obd_precreate(lov->lov_tgts[osts->op_array[i]]->ltd_exp) > 2)
 832                         continue;
 833
 834                 lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable = 1;
 835                 qos_calc_weight(lov, osts->op_array[i]);
 836                 total_bavail += bavail;
 837                 total_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight;
 838
 839                 good_osts++;
 840         }
 841
 842 #ifdef QOS_DEBUG
 843         CDEBUG(D_QOS, "found %d good osts\n", good_osts);
 844 #endif
 845
 846         if (good_osts < stripe_cnt_min)
 847                 GOTO(out, rc = -EAGAIN);
 848
 849         if (!total_bavail)
 850                 GOTO(out, rc = -ENOSPC);
 851
 852         /* We have enough osts */
 853         if (good_osts < *stripe_cnt)
 854                 *stripe_cnt = good_osts;
 855
 856         if (!*stripe_cnt)
 857                 GOTO(out, rc = -EAGAIN);
 858
 859         /* Find enough OSTs with weighted random allocation. */
 860         nfound = 0;
 861         while (nfound < *stripe_cnt) {
 862                 __u64 rand, cur_weight;
 863
 864                 cur_weight = 0;
 865                 rc = -ENODEV;
 866
 867                 if (total_weight) {
 868 #if BITS_PER_LONG == 32
 869                         rand = ll_rand() % (unsigned)total_weight;
 870                         /* If total_weight > 32-bit, first generate the high
 871                          * 32 bits of the random number, then add in the low
 872                          * 32 bits (truncated to the upper limit, if needed) */
 873                         if (total_weight > 0xffffffffULL)
 874                                 rand = (__u64)(ll_rand() %
 875                                           (unsigned)(total_weight >> 32)) << 32;
 876                         else
 877                                 rand = 0;
 878
 879                         if (rand == (total_weight & 0xffffffff00000000ULL))
 880                                 rand |= ll_rand() % (unsigned)total_weight;
 881                         else
 882                                 rand |= ll_rand();
 883
 884 #else
 885                         rand = ((__u64)ll_rand() << 32 | ll_rand()) %
 886                                 total_weight;
 887 #endif
 888                 } else {
 889                         rand = 0;
 890                 }
 891
 892                 /* On average, this will hit larger-weighted osts more often.
 893                    0-weight osts will always get used last (only when rand=0).*/
 894                 for (i = 0; i < osts->op_count; i++) {
 895                         if (!lov->lov_tgts[osts->op_array[i]] ||
 896                             !lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable)
 897                                 continue;
 898
 899                         cur_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight;
 900 #ifdef QOS_DEBUG
 901                         CDEBUG(D_QOS, "stripe_cnt=%d nfound=%d cur_weight="LPU64
 902                                       " rand="LPU64" total_weight="LPU64"\n",
 903                                *stripe_cnt, nfound, cur_weight, rand, total_weight);
 904 #endif
 905                         if (cur_weight >= rand) {
 906 #ifdef QOS_DEBUG
 907                                 CDEBUG(D_QOS, "assigned stripe=%d to idx=%d\n",
 908                                        nfound, osts->op_array[i]);
 909 #endif
 910                                 idx_arr[nfound++] = osts->op_array[i];
 911                                 qos_used(lov, osts, osts->op_array[i], &total_weight);
 912                                 rc = 0;
 913                                 break;
 914                         }
 915                 }
 916                 /* should never satisfy below condition */
 917                 if (rc) {
 918                         CERROR("Didn't find any OSTs?\n");
 919                         break;
 920                 }
 921         }
 922         LASSERT(nfound == *stripe_cnt);
 923
 924 out:
 925         up_write(&lov->lov_qos.lq_rw_sem);
 926
 927 out_nolock:
 928         if (pool != NULL) {
 929                 up_read(&pool_tgt_rw_sem(pool));
 930                 /* put back ref got by lov_find_pool() */
 931                 lov_pool_putref(pool);
 932         }
 933
 934         if (rc == -EAGAIN)
 935                 rc = alloc_rr(lov, idx_arr, stripe_cnt, poolname, flags);
 936
 937         obd_putref(exp->exp_obd);
 938         RETURN(rc);
 939 }
 940
 941 /* return new alloced stripe count on success */
 942 static int alloc_idx_array(struct obd_export *exp, struct lov_stripe_md *lsm,
 943                            int newea, int **idx_arr, int *arr_cnt, int flags)
 944 {
 945         struct lov_obd *lov = &exp->exp_obd->u.lov;
 946         int stripe_cnt = lsm->lsm_stripe_count;
 947         int i, rc = 0;
 948         int *tmp_arr = NULL;
 949         ENTRY;
 950
 951         *arr_cnt = stripe_cnt;
 952         OBD_ALLOC(tmp_arr, *arr_cnt * sizeof(int));
 953         if (tmp_arr == NULL)
 954                 RETURN(-ENOMEM);
 955         for (i = 0; i < *arr_cnt; i++)
 956                 tmp_arr[i] = -1;
 957
 958         if (newea ||
 959             lsm->lsm_oinfo[0]->loi_ost_idx >= lov->desc.ld_tgt_count)
 960                 rc = alloc_qos(exp, tmp_arr, &stripe_cnt,
 961                                lsm->lsm_pool_name, flags);
 962         else
 963                 rc = alloc_specific(lov, lsm, tmp_arr);
 964
 965         if (rc)
 966                 GOTO(out_arr, rc);
 967
 968         *idx_arr = tmp_arr;
 969         RETURN(stripe_cnt);
 970 out_arr:
 971         OBD_FREE(tmp_arr, *arr_cnt * sizeof(int));
 972         *arr_cnt = 0;
 973         RETURN(rc);
 974 }
 975
 976 static void free_idx_array(int *idx_arr, int arr_cnt)
 977 {
 978         if (arr_cnt)
 979                 OBD_FREE(idx_arr, arr_cnt * sizeof(int));
 980 }
 981
 982 int qos_prep_create(struct obd_export *exp, struct lov_request_set *set)
 983 {
 984         struct lov_obd *lov = &exp->exp_obd->u.lov;
 985         struct lov_stripe_md *lsm;
 986         struct obdo *src_oa = set->set_oi->oi_oa;
 987         struct obd_trans_info *oti = set->set_oti;
 988         int i, stripes, rc = 0, newea = 0;
 989         int flag = LOV_USES_ASSIGNED_STRIPE;
 990         int *idx_arr = NULL, idx_cnt = 0;
 991         ENTRY;
 992
 993         LASSERT(src_oa->o_valid & OBD_MD_FLID);
 994         LASSERT(src_oa->o_valid & OBD_MD_FLGROUP);
 995
 996         if (set->set_oi->oi_md == NULL) {
 997                 int stripes_def = lov_get_stripecnt(lov, 0);
 998
 999                 /* If the MDS file was truncated up to some size, stripe over
1000                  * enough OSTs to allow the file to be created at that size.
1001                  * This may mean we use more than the default # of stripes. */
1002                 if (src_oa->o_valid & OBD_MD_FLSIZE) {
1003                         obd_size min_bavail = LUSTRE_STRIPE_MAXBYTES;
1004
1005                         /* Find a small number of stripes we can use
1006                            (up to # of active osts). */
1007                         stripes = 1;
1008                         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1009                                 if (!lov->lov_tgts[i] ||
1010                                     !lov->lov_tgts[i]->ltd_active)
1011                                         continue;
1012                                 min_bavail = min(min_bavail, TGT_BAVAIL(i));
1013                                 if (min_bavail * stripes > src_oa->o_size)
1014                                         break;
1015                                 stripes++;
1016                         }
1017
1018                         if (stripes < stripes_def)
1019                                 stripes = stripes_def;
1020                 } else {
1021                          flag = LOV_USES_DEFAULT_STRIPE;
1022                          stripes = stripes_def;
1023                 }
1024
1025                 rc = lov_alloc_memmd(&set->set_oi->oi_md, stripes,
1026                                      lov->desc.ld_pattern ?
1027                                      lov->desc.ld_pattern : LOV_PATTERN_RAID0,
1028                                      LOV_MAGIC);
1029                 if (rc < 0)
1030                         GOTO(out_err, rc);
1031                 newea = 1;
1032                 rc = 0;
1033         }
1034
1035         lsm = set->set_oi->oi_md;
1036         lsm->lsm_object_id = src_oa->o_id;
1037         lsm->lsm_object_gr = src_oa->o_gr;
1038
1039         if (!lsm->lsm_stripe_size)
1040                 lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size;
1041         if (!lsm->lsm_pattern) {
1042                 LASSERT(lov->desc.ld_pattern);
1043                 lsm->lsm_pattern = lov->desc.ld_pattern;
1044         }
1045
1046         stripes = alloc_idx_array(exp, lsm, newea, &idx_arr, &idx_cnt, flag);
1047         if (stripes <= 0)
1048                 GOTO(out_err, rc = stripes ? stripes : -EIO);
1049         LASSERTF(stripes <= lsm->lsm_stripe_count,"requested %d allocated %d\n",
1050                  lsm->lsm_stripe_count, stripes);
1051
1052         for (i = 0; i < stripes; i++) {
1053                 struct lov_request *req;
1054                 int ost_idx = idx_arr[i];
1055                 LASSERT(ost_idx >= 0);
1056
1057                 OBD_ALLOC(req, sizeof(*req));
1058                 if (req == NULL)
1059                         GOTO(out_err, rc = -ENOMEM);
1060                 lov_set_add_req(req, set);
1061
1062                 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
1063                 OBD_ALLOC(req->rq_oi.oi_md, req->rq_buflen);
1064                 if (req->rq_oi.oi_md == NULL)
1065                         GOTO(out_err, rc = -ENOMEM);
1066
1067                 OBDO_ALLOC(req->rq_oi.oi_oa);
1068                 if (req->rq_oi.oi_oa == NULL)
1069                         GOTO(out_err, rc = -ENOMEM);
1070
1071                 req->rq_idx = ost_idx;
1072                 req->rq_stripe = i;
1073                 /* create data objects with "parent" OA */
1074                 memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
1075
1076                 /* XXX When we start creating objects on demand, we need to
1077                  *     make sure that we always create the object on the
1078                  *     stripe which holds the existing file size.
1079                  */
1080                 if (src_oa->o_valid & OBD_MD_FLSIZE) {
1081                         req->rq_oi.oi_oa->o_size =
1082                                 lov_size_to_stripe(lsm, src_oa->o_size, i);
1083
1084                         CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
1085                                i, req->rq_oi.oi_oa->o_size, src_oa->o_size);
1086                 }
1087         }
1088         LASSERT(set->set_count == stripes);
1089
1090         if (stripes < lsm->lsm_stripe_count)
1091                 qos_shrink_lsm(set);
1092         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_LOV_PREP_CREATE)) {
1093                 qos_shrink_lsm(set);
1094                 rc = -EIO;
1095         }
1096
1097         if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) {
1098                 oti_alloc_cookies(oti, set->set_count);
1099                 if (!oti->oti_logcookies)
1100                         GOTO(out_err, rc = -ENOMEM);
1101                 set->set_cookies = oti->oti_logcookies;
1102         }
1103 out_err:
1104         if (newea && rc)
1105                 obd_free_memmd(exp, &set->set_oi->oi_md);
1106         if (idx_arr)
1107                 free_idx_array(idx_arr, idx_cnt);
1108         EXIT;
1109         return rc;
1110 }
1111
1112 void qos_update(struct lov_obd *lov)
1113 {
1114         ENTRY;
1115         lov->lov_qos.lq_dirty = 1;
1116 }
1117
1118 void qos_statfs_done(struct lov_obd *lov)
1119 {
1120         LASSERT(lov->lov_qos.lq_statfs_in_progress);
1121         down_write(&lov->lov_qos.lq_rw_sem);
1122         lov->lov_qos.lq_statfs_in_progress = 0;
1123         /* wake up any threads waiting for the statfs rpcs to complete */
1124         cfs_waitq_signal(&lov->lov_qos.lq_statfs_waitq);
1125         up_write(&lov->lov_qos.lq_rw_sem);
1126 }
1127
1128 static int qos_statfs_ready(struct obd_device *obd, __u64 max_age)
1129 {
1130         struct lov_obd         *lov = &obd->u.lov;
1131         int rc;
1132         ENTRY;
1133         down_read(&lov->lov_qos.lq_rw_sem);
1134         rc = lov->lov_qos.lq_statfs_in_progress == 0 ||
1135              cfs_time_beforeq_64(max_age, obd->obd_osfs_age);
1136         up_read(&lov->lov_qos.lq_rw_sem);
1137         RETURN(rc);
1138 }
1139
1140 /*
1141  * Update statfs data if the current osfs age is older than max_age.
1142  * If wait is not set, it means that we are called from lov_create()
1143  * and we should just issue the rpcs without waiting for them to complete.
1144  * If wait is set, we are called from alloc_qos() and we just have
1145  * to wait for the request set to complete.
1146  */
1147 void qos_statfs_update(struct obd_device *obd, __u64 max_age, int wait)
1148 {
1149         struct lov_obd         *lov = &obd->u.lov;
1150         struct obd_info        *oinfo;
1151         int                     rc = 0;
1152         struct ptlrpc_request_set *set = NULL;
1153         ENTRY;
1154
1155         if (cfs_time_beforeq_64(max_age, obd->obd_osfs_age))
1156                 /* statfs data are quite recent, don't need to refresh it */
1157                 RETURN_EXIT;
1158
1159         if (!wait && lov->lov_qos.lq_statfs_in_progress)
1160                 /* statfs already in progress */
1161                 RETURN_EXIT;
1162
1163         down_write(&lov->lov_qos.lq_rw_sem);
1164         if (lov->lov_qos.lq_statfs_in_progress) {
1165                 up_write(&lov->lov_qos.lq_rw_sem);
1166                 GOTO(out, rc = 0);
1167         }
1168         /* no statfs in flight, send rpcs */
1169         lov->lov_qos.lq_statfs_in_progress = 1;
1170         up_write(&lov->lov_qos.lq_rw_sem);
1171
1172         if (wait)
1173                 CDEBUG(D_QOS, "%s: did not manage to get fresh statfs data "
1174                        "in a timely manner (osfs age "LPU64", max age "LPU64")"
1175                        ", sending new statfs rpcs\n",
1176                        obd_uuid2str(&lov->desc.ld_uuid), obd->obd_osfs_age,
1177                        max_age);
1178
1179         /* need to send statfs rpcs */
1180         CDEBUG(D_QOS, "sending new statfs requests\n");
1181         memset(lov->lov_qos.lq_statfs_data, 0,
1182                sizeof(*lov->lov_qos.lq_statfs_data));
1183         oinfo = &lov->lov_qos.lq_statfs_data->lsd_oi;
1184         oinfo->oi_osfs = &lov->lov_qos.lq_statfs_data->lsd_statfs;
1185         oinfo->oi_flags = OBD_STATFS_NODELAY;
1186         set = ptlrpc_prep_set();
1187         if (!set)
1188                 GOTO(out_failed, rc = -ENOMEM);
1189
1190         rc = obd_statfs_async(obd, oinfo, max_age, set);
1191         if (rc || list_empty(&set->set_requests)) {
1192                 if (rc)
1193                         CWARN("statfs failed with %d\n", rc);
1194                 GOTO(out_failed, rc);
1195         }
1196         /* send requests via ptlrpcd */
1197         oinfo->oi_flags |= OBD_STATFS_PTLRPCD;
1198         ptlrpcd_add_rqset(set);
1199         GOTO(out, rc);
1200
1201 out_failed:
1202         down_write(&lov->lov_qos.lq_rw_sem);
1203         lov->lov_qos.lq_statfs_in_progress = 0;
1204         /* wake up any threads waiting for the statfs rpcs to complete */
1205         cfs_waitq_signal(&lov->lov_qos.lq_statfs_waitq);
1206         up_write(&lov->lov_qos.lq_rw_sem);
1207         wait = 0;
1208 out:
1209         if (set)
1210                 ptlrpc_set_destroy(set);
1211         if (wait) {
1212                 struct l_wait_info lwi = { 0 };
1213                 CDEBUG(D_QOS, "waiting for statfs requests to complete\n");
1214                 l_wait_event(lov->lov_qos.lq_statfs_waitq,
1215                              qos_statfs_ready(obd, max_age), &lwi);
1216                 if (cfs_time_before_64(obd->obd_osfs_age, max_age))
1217                         CDEBUG(D_QOS, "%s: still no fresh statfs data after "
1218                                       "waiting (osfs age "LPU64", max age "
1219                                       LPU64")\n",
1220                                       obd_uuid2str(&lov->desc.ld_uuid),
1221                                       obd->obd_osfs_age, max_age);
1222         }
1223 }