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  * Copyright (C) 2002, 2003 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of the Lustre file system, http://www.lustre.org
   7  *   Lustre is a trademark of Cluster File Systems, Inc.
   8  *
   9  *   You may have signed or agreed to another license before downloading
  10  *   this software.  If so, you are bound by the terms and conditions
  11  *   of that agreement, and the following does not apply to you.  See the
  12  *   LICENSE file included with this distribution for more information.
  13  *
  14  *   If you did not agree to a different license, then this copy of Lustre
  15  *   is open source software; you can redistribute it and/or modify it
  16  *   under the terms of version 2 of the GNU General Public License as
  17  *   published by the Free Software Foundation.
  18  *
  19  *   In either case, Lustre is distributed in the hope that it will be
  20  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  21  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  22  *   license text for more details.
  23  */
  24
  25 #ifndef EXPORT_SYMTAB
  26 # define EXPORT_SYMTAB
  27 #endif
  28 #define DEBUG_SUBSYSTEM S_LOV
  29
  30 #ifdef __KERNEL__
  31 #include <libcfs/libcfs.h>
  32 #else
  33 #include <liblustre.h>
  34 #endif
  35
  36 #include <obd_class.h>
  37 #include <obd_lov.h>
  38 #include "lov_internal.h"
  39
  40 /* #define QOS_DEBUG 1 */
  41 #define D_QOS D_OTHER
  42
  43 #define TGT_BAVAIL(i)  (lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bavail*\
  44                         lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bsize)
  45 #define TGT_FFREE(i)   (lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_ffree)
  46
  47
  48 int qos_add_tgt(struct obd_device *obd, __u32 index)
  49 {
  50         struct lov_obd *lov = &obd->u.lov;
  51         struct lov_qos_oss *oss, *temposs;
  52         struct obd_export *exp = lov->lov_tgts[index]->ltd_exp;
  53         int rc = 0, found = 0;
  54         ENTRY;
  55
  56         /* We only need this QOS struct on MDT, not clients - but we may not
  57          * have registered the LOV's observer yet, so there's no way to know */
  58         if (!exp || !exp->exp_connection) {
  59                 CERROR("Missing connection\n");
  60                 RETURN(-ENOTCONN);
  61         }
  62
  63         down_write(&lov->lov_qos.lq_rw_sem);
  64         mutex_down(&lov->lov_lock);
  65         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
  66                 if (obd_uuid_equals(&oss->lqo_uuid,
  67                                     &exp->exp_connection->c_remote_uuid)) {
  68                         found++;
  69                         break;
  70                 }
  71         }
  72
  73         if (!found) {
  74                 OBD_ALLOC_PTR(oss);
  75                 if (!oss)
  76                         GOTO(out, rc = -ENOMEM);
  77                 memcpy(&oss->lqo_uuid,
  78                        &exp->exp_connection->c_remote_uuid,
  79                        sizeof(oss->lqo_uuid));
  80         } else {
  81                 /* Assume we have to move this one */
  82                 list_del(&oss->lqo_oss_list);
  83         }
  84
  85         oss->lqo_ost_count++;
  86         lov->lov_tgts[index]->ltd_qos.ltq_oss = oss;
  87
  88         /* Add sorted by # of OSTs.  Find the first entry that we're
  89            bigger than... */
  90         list_for_each_entry(temposs, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
  91                 if (oss->lqo_ost_count > temposs->lqo_ost_count)
  92                         break;
  93         }
  94         /* ...and add before it.  If we're the first or smallest, temposs
  95            points to the list head, and we add to the end. */
  96         list_add_tail(&oss->lqo_oss_list, &temposs->lqo_oss_list);
  97
  98         lov->lov_qos.lq_dirty = 1;
  99         lov->lov_qos.lq_dirty_rr = 1;
 100
 101         CDEBUG(D_QOS, "add tgt %s to OSS %s (%d OSTs)\n",
 102                obd_uuid2str(&lov->lov_tgts[index]->ltd_uuid),
 103                obd_uuid2str(&oss->lqo_uuid),
 104                oss->lqo_ost_count);
 105
 106 out:
 107         mutex_up(&lov->lov_lock);
 108         up_write(&lov->lov_qos.lq_rw_sem);
 109         RETURN(rc);
 110 }
 111
 112 int qos_del_tgt(struct obd_device *obd, __u32 index)
 113 {
 114         struct lov_obd *lov = &obd->u.lov;
 115         struct lov_qos_oss *oss;
 116         int rc = 0;
 117         ENTRY;
 118
 119         if (!lov->lov_tgts[index])
 120                 RETURN(0);
 121
 122         down_write(&lov->lov_qos.lq_rw_sem);
 123
 124         oss = lov->lov_tgts[index]->ltd_qos.ltq_oss;
 125         if (!oss)
 126                 GOTO(out, rc = -ENOENT);
 127
 128         oss->lqo_ost_count--;
 129         if (oss->lqo_ost_count == 0) {
 130                 CDEBUG(D_QOS, "removing OSS %s\n",
 131                        obd_uuid2str(&oss->lqo_uuid));
 132                 list_del(&oss->lqo_oss_list);
 133                 OBD_FREE_PTR(oss);
 134         }
 135
 136         lov->lov_qos.lq_dirty = 1;
 137         lov->lov_qos.lq_dirty_rr = 1;
 138 out:
 139         up_write(&lov->lov_qos.lq_rw_sem);
 140         RETURN(rc);
 141 }
 142
 143 /* Recalculate per-object penalties for OSSs and OSTs,
 144    depends on size of each ost in an oss */
 145 static int qos_calc_ppo(struct obd_device *obd)
 146 {
 147         struct lov_obd *lov = &obd->u.lov;
 148         struct lov_qos_oss *oss;
 149         __u64 ba_max, ba_min, temp;
 150         __u32 num_active;
 151         int rc, i, prio_wide;
 152         ENTRY;
 153
 154         if (!lov->lov_qos.lq_dirty)
 155                 GOTO(out, rc = 0);
 156
 157         num_active = lov->desc.ld_active_tgt_count - 1;
 158         if (num_active < 1)
 159                 GOTO(out, rc = -EAGAIN);
 160
 161         /* find bavail on each OSS */
 162         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 163                 oss->lqo_bavail = 0;
 164         }
 165         lov->lov_qos.lq_active_oss_count = 0;
 166
 167         /* How badly user wants to select osts "widely" (not recently chosen
 168            and not on recent oss's).  As opposed to "freely" (free space
 169            avail.) 0-256. */
 170         prio_wide = 256 - lov->lov_qos.lq_prio_free;
 171
 172         ba_min = (__u64)(-1);
 173         ba_max = 0;
 174         /* Calculate OST penalty per object */
 175         /* (lov ref taken in alloc_qos) */
 176         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 177                 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
 178                         continue;
 179                 temp = TGT_BAVAIL(i);
 180                 if (!temp)
 181                         continue;
 182                 ba_min = min(temp, ba_min);
 183                 ba_max = max(temp, ba_max);
 184
 185                 /* Count the number of usable OSS's */
 186                 if (lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_bavail == 0)
 187                         lov->lov_qos.lq_active_oss_count++;
 188                 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_bavail += temp;
 189
 190                 /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
 191                 temp >>= 1;
 192                 do_div(temp, num_active);
 193                 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj =
 194                         (temp * prio_wide) >> 8;
 195
 196                 if (lov->lov_qos.lq_reset == 0)
 197                         lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
 198         }
 199
 200         num_active = lov->lov_qos.lq_active_oss_count - 1;
 201         if (num_active < 1) {
 202                 /* If there's only 1 OSS, we can't penalize it, so instead
 203                    we have to double the OST penalty */
 204                 num_active = 1;
 205                 for (i = 0; i < lov->desc.ld_tgt_count; i++)
 206                         if (lov->lov_tgts[i])
 207                             lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj <<= 1;
 208         }
 209
 210         /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
 211         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 212                 temp = oss->lqo_bavail >> 1;
 213                 do_div(temp, oss->lqo_ost_count * num_active);
 214                 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
 215                 if (lov->lov_qos.lq_reset == 0)
 216                         oss->lqo_penalty = 0;
 217         }
 218
 219         lov->lov_qos.lq_dirty = 0;
 220         lov->lov_qos.lq_reset = 0;
 221
 222         /* If each ost has almost same free space,
 223          * do rr allocation for better creation performance */
 224         lov->lov_qos.lq_same_space = 0;
 225         temp = ba_max - ba_min;
 226         ba_min = (ba_min * 51) >> 8;     /* 51/256 = .20 */
 227         if (temp < ba_min) {
 228                 /* Difference is less than 20% */
 229                 lov->lov_qos.lq_same_space = 1;
 230                 /* Reset weights for the next time we enter qos mode */
 231                 lov->lov_qos.lq_reset = 0;
 232         }
 233         rc = 0;
 234
 235 out:
 236         if (!rc && lov->lov_qos.lq_same_space)
 237                 RETURN(-EAGAIN);
 238         RETURN(rc);
 239 }
 240
 241 static int qos_calc_weight(struct lov_obd *lov, int i)
 242 {
 243         __u64 temp, temp2;
 244
 245         /* Final ost weight = TGT_BAVAIL - ost_penalty - oss_penalty */
 246         temp = TGT_BAVAIL(i);
 247         temp2 = lov->lov_tgts[i]->ltd_qos.ltq_penalty +
 248                 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty;
 249         if (temp < temp2)
 250                 lov->lov_tgts[i]->ltd_qos.ltq_weight = 0;
 251         else
 252                 lov->lov_tgts[i]->ltd_qos.ltq_weight = temp - temp2;
 253         return 0;
 254 }
 255
 256 /* We just used this index for a stripe; adjust everyone's weights */
 257 static int qos_used(struct lov_obd *lov, __u32 index, __u64 *total_wt)
 258 {
 259         struct lov_qos_oss *oss;
 260         int i;
 261         ENTRY;
 262
 263         /* Don't allocate from this stripe anymore, until the next alloc_qos */
 264         lov->lov_tgts[index]->ltd_qos.ltq_usable = 0;
 265
 266         oss = lov->lov_tgts[index]->ltd_qos.ltq_oss;
 267
 268         /* Decay old penalty by half (we're adding max penalty, and don't
 269            want it to run away.) */
 270         lov->lov_tgts[index]->ltd_qos.ltq_penalty >>= 1;
 271         oss->lqo_penalty >>= 1;
 272
 273         /* Set max penalties for this OST and OSS */
 274         lov->lov_tgts[index]->ltd_qos.ltq_penalty +=
 275                 lov->lov_tgts[index]->ltd_qos.ltq_penalty_per_obj *
 276                 lov->desc.ld_active_tgt_count;
 277         oss->lqo_penalty += oss->lqo_penalty_per_obj *
 278                 lov->lov_qos.lq_active_oss_count;
 279
 280         /* Decrease all OSS penalties */
 281         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 282                 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
 283                         oss->lqo_penalty = 0;
 284                 else
 285                         oss->lqo_penalty -= oss->lqo_penalty_per_obj;
 286         }
 287
 288         *total_wt = 0;
 289         /* Decrease all OST penalties */
 290         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 291                 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
 292                         continue;
 293                 if (lov->lov_tgts[i]->ltd_qos.ltq_penalty <
 294                     lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj)
 295                         lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
 296                 else
 297                         lov->lov_tgts[i]->ltd_qos.ltq_penalty -=
 298                         lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj;
 299
 300                 qos_calc_weight(lov, i);
 301
 302                 /* Recalc the total weight of usable osts */
 303                 if (lov->lov_tgts[i]->ltd_qos.ltq_usable)
 304                         *total_wt += lov->lov_tgts[i]->ltd_qos.ltq_weight;
 305
 306 #ifdef QOS_DEBUG
 307                 CDEBUG(D_QOS, "recalc tgt %d avail="LPU64
 308                        " ostppo="LPU64" ostp="LPU64" ossppo="LPU64
 309                        " ossp="LPU64" wt="LPU64"\n",
 310                        i, TGT_BAVAIL(i) >> 10,
 311                        lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj >> 10,
 312                        lov->lov_tgts[i]->ltd_qos.ltq_penalty >> 10,
 313                        lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty_per_obj>>10,
 314                        lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty>>10,
 315                        lov->lov_tgts[i]->ltd_qos.ltq_weight>>10);
 316 #endif
 317         }
 318
 319         RETURN(0);
 320 }
 321
 322 #define LOV_QOS_EMPTY ((__u32)-1)
 323 /* compute optimal round-robin order, based on OSTs per OSS */
 324 static int qos_calc_rr(struct lov_obd *lov)
 325 {
 326         struct lov_qos_oss *oss;
 327         unsigned ost_count, placed, real_count;
 328         int i;
 329         ENTRY;
 330
 331         if (!lov->lov_qos.lq_dirty_rr) {
 332                 LASSERT(lov->lov_qos.lq_rr_size);
 333                 RETURN(0);
 334         }
 335
 336         down_write(&lov->lov_qos.lq_rw_sem);
 337         ost_count = lov->desc.ld_tgt_count;
 338
 339         if (lov->lov_qos.lq_rr_size)
 340                 OBD_FREE(lov->lov_qos.lq_rr_array, lov->lov_qos.lq_rr_size);
 341         lov->lov_qos.lq_rr_size = ost_count *
 342                 sizeof(lov->lov_qos.lq_rr_array[0]);
 343         OBD_ALLOC(lov->lov_qos.lq_rr_array, lov->lov_qos.lq_rr_size);
 344         if (!lov->lov_qos.lq_rr_array) {
 345                 lov->lov_qos.lq_rr_size = 0;
 346                 up_write(&lov->lov_qos.lq_rw_sem);
 347                 RETURN(-ENOMEM);
 348         }
 349
 350         real_count = 0;
 351         for (i = 0; i < ost_count; i++) {
 352                 lov->lov_qos.lq_rr_array[i] = LOV_QOS_EMPTY;
 353                 if (lov->lov_tgts[i])
 354                         real_count++;
 355         }
 356
 357         /* Place all the OSTs from 1 OSS at the same time. */
 358         placed = 0;
 359         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 360                 int j = 0;
 361                 for (i = 0; i < ost_count; i++) {
 362                         if (lov->lov_tgts[i] &&
 363                             (lov->lov_tgts[i]->ltd_qos.ltq_oss == oss)) {
 364                               /* Evenly space these OSTs across arrayspace */
 365                               int next = j * ost_count / oss->lqo_ost_count;
 366                               while (lov->lov_qos.lq_rr_array[next] !=
 367                                      LOV_QOS_EMPTY)
 368                                       next = (next + 1) % ost_count;
 369                               lov->lov_qos.lq_rr_array[next] = i;
 370                               j++;
 371                               placed++;
 372                       }
 373                 }
 374                 LASSERT(j == oss->lqo_ost_count);
 375         }
 376
 377         lov->lov_qos.lq_dirty_rr = 0;
 378         up_write(&lov->lov_qos.lq_rw_sem);
 379
 380         if (placed != real_count) {
 381                 /* This should never happen */
 382                 LCONSOLE_ERROR("Failed to place all OSTs in the round-robin "
 383                                "list (%d of %d).\n", placed, real_count);
 384                 for (i = 0; i < ost_count; i++) {
 385                         LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
 386                                  lov->lov_qos.lq_rr_array[i]);
 387                 }
 388                 lov->lov_qos.lq_dirty_rr = 1;
 389                 RETURN(-EAGAIN);
 390         }
 391
 392 #ifdef QOS_DEBUG
 393         for (i = 0; i < ost_count; i++) {
 394                 LCONSOLE(D_QOS, "rr #%d ost idx=%d\n", i,
 395                          lov->lov_qos.lq_rr_array[i]);
 396         }
 397 #endif
 398
 399         RETURN(0);
 400 }
 401
 402
 403 void qos_shrink_lsm(struct lov_request_set *set)
 404 {
 405         struct lov_stripe_md *lsm = set->set_oi->oi_md, *lsm_new;
 406         /* XXX LOV STACKING call into osc for sizes */
 407         unsigned oldsize, newsize;
 408
 409         if (set->set_oti && set->set_cookies && set->set_cookie_sent) {
 410                 struct llog_cookie *cookies;
 411                 oldsize = lsm->lsm_stripe_count * sizeof(*cookies);
 412                 newsize = set->set_count * sizeof(*cookies);
 413
 414                 cookies = set->set_cookies;
 415                 oti_alloc_cookies(set->set_oti, set->set_count);
 416                 if (set->set_oti->oti_logcookies) {
 417                         memcpy(set->set_oti->oti_logcookies, cookies, newsize);
 418                         OBD_FREE(cookies, oldsize);
 419                         set->set_cookies = set->set_oti->oti_logcookies;
 420                 } else {
 421                         CWARN("'leaking' %d bytes\n", oldsize - newsize);
 422                 }
 423         }
 424
 425         CWARN("using fewer stripes for object "LPU64": old %u new %u\n",
 426               lsm->lsm_object_id, lsm->lsm_stripe_count, set->set_count);
 427         LASSERT(lsm->lsm_stripe_count >= set->set_count);
 428
 429         newsize = lov_stripe_md_size(set->set_count);
 430         OBD_ALLOC(lsm_new, newsize);
 431         if (lsm_new != NULL) {
 432                 int i;
 433                 memcpy(lsm_new, lsm, sizeof(*lsm));
 434                 for (i = 0; i < lsm->lsm_stripe_count; i++) {
 435                         if (i < set->set_count) {
 436                                 lsm_new->lsm_oinfo[i] = lsm->lsm_oinfo[i];
 437                                 continue;
 438                         }
 439                         OBD_SLAB_FREE(lsm->lsm_oinfo[i], lov_oinfo_slab,
 440                                       sizeof(struct lov_oinfo));
 441                 }
 442                 lsm_new->lsm_stripe_count = set->set_count;
 443                 OBD_FREE(lsm, sizeof(struct lov_stripe_md) +
 444                          lsm->lsm_stripe_count * sizeof(struct lov_oinfo *));
 445                 set->set_oi->oi_md = lsm_new;
 446         } else {
 447                 CWARN("'leaking' few bytes\n");
 448         }
 449 }
 450
 451 int qos_remedy_create(struct lov_request_set *set, struct lov_request *req)
 452 {
 453         struct lov_stripe_md *lsm = set->set_oi->oi_md;
 454         struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
 455         unsigned ost_idx, ost_count = lov->desc.ld_tgt_count;
 456         int stripe, i, rc = -EIO;
 457         ENTRY;
 458
 459         ost_idx = (req->rq_idx + lsm->lsm_stripe_count) % ost_count;
 460         for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
 461                 if (!lov->lov_tgts[ost_idx] ||
 462                     !lov->lov_tgts[ost_idx]->ltd_active)
 463                         continue;
 464                 /* check if objects has been created on this ost */
 465                 for (stripe = 0; stripe < lsm->lsm_stripe_count; stripe++) {
 466                         if (stripe == req->rq_stripe)
 467                                 continue;
 468                         if (ost_idx == lsm->lsm_oinfo[stripe]->loi_ost_idx)
 469                                 break;
 470                 }
 471
 472                 if (stripe >= lsm->lsm_stripe_count) {
 473                         req->rq_idx = ost_idx;
 474                         rc = obd_create(lov->lov_tgts[ost_idx]->ltd_exp,
 475                                         req->rq_oi.oi_oa, &req->rq_oi.oi_md,
 476                                         set->set_oti);
 477                         if (!rc)
 478                                 break;
 479                 }
 480         }
 481         RETURN(rc);
 482 }
 483
 484 #define LOV_CREATE_RESEED_MULT 4
 485 #define LOV_CREATE_RESEED_MIN  1000
 486 /* Allocate objects on osts with round-robin algorithm */
 487 static int alloc_rr(struct lov_obd *lov, int *idx_arr, int *stripe_cnt)
 488 {
 489         unsigned array_idx, ost_count = lov->desc.ld_tgt_count;
 490         unsigned ost_active_count = lov->desc.ld_active_tgt_count;
 491         int i, *idx_pos = idx_arr;
 492         __u32 ost_idx;
 493         ENTRY;
 494
 495         i = qos_calc_rr(lov);
 496         if (i)
 497                 RETURN(i);
 498
 499         if (--lov->lov_start_count <= 0) {
 500                 lov->lov_start_idx = ll_rand() % ost_count;
 501                 lov->lov_start_count =
 502                         (LOV_CREATE_RESEED_MIN / max(ost_active_count, 1U) +
 503                          LOV_CREATE_RESEED_MULT) * max(ost_active_count, 1U);
 504         } else if (*stripe_cnt >= ost_active_count ||
 505                    lov->lov_start_idx > ost_count) {
 506                 /* If we have allocated from all of the OSTs, slowly
 507                    precess the next start */
 508                 lov->lov_start_idx %= ost_count;
 509                 ++lov->lov_offset_idx;
 510         }
 511         array_idx = (lov->lov_start_idx + lov->lov_offset_idx) % ost_count;
 512 #ifdef QOS_DEBUG
 513         CDEBUG(D_QOS, "want %d startidx %d startcnt %d offset %d arrayidx %d\n",
 514                *stripe_cnt, lov->lov_start_idx, lov->lov_start_count,
 515                lov->lov_offset_idx, array_idx);
 516 #endif
 517
 518         down_read(&lov->lov_qos.lq_rw_sem);
 519         for (i = 0; i < ost_count; i++, array_idx=(array_idx + 1) % ost_count) {
 520                 ++lov->lov_start_idx;
 521                 ost_idx = lov->lov_qos.lq_rr_array[array_idx];
 522 #ifdef QOS_DEBUG
 523                 CDEBUG(D_QOS, "#%d strt %d act %d strp %d ary %d idx %d\n",
 524                        i, lov->lov_start_idx,
 525                        ((ost_idx != LOV_QOS_EMPTY) && lov->lov_tgts[ost_idx]) ?
 526                        lov->lov_tgts[ost_idx]->ltd_active : 0,
 527                        idx_pos - idx_arr, array_idx, ost_idx);
 528 #endif
 529                 if ((ost_idx == LOV_QOS_EMPTY) || !lov->lov_tgts[ost_idx] ||
 530                     !lov->lov_tgts[ost_idx]->ltd_active)
 531                         continue;
 532                 *idx_pos = ost_idx;
 533                 idx_pos++;
 534                 /* We have enough stripes */
 535                 if (idx_pos - idx_arr == *stripe_cnt)
 536                         break;
 537         }
 538         up_read(&lov->lov_qos.lq_rw_sem);
 539
 540         *stripe_cnt = idx_pos - idx_arr;
 541         RETURN(0);
 542 }
 543
 544 /* alloc objects on osts with specific stripe offset */
 545 static int alloc_specific(struct lov_obd *lov, struct lov_stripe_md *lsm,
 546                           int *idx_arr)
 547 {
 548         unsigned ost_idx, ost_count = lov->desc.ld_tgt_count;
 549         int i, *idx_pos = idx_arr;
 550         ENTRY;
 551
 552         ost_idx = lsm->lsm_oinfo[0]->loi_ost_idx;
 553         for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
 554                 if (!lov->lov_tgts[ost_idx] ||
 555                     !lov->lov_tgts[ost_idx]->ltd_active) {
 556                         continue;
 557                 }
 558                 *idx_pos = ost_idx;
 559                 idx_pos++;
 560                 /* got enough ost */
 561                 if (idx_pos - idx_arr == lsm->lsm_stripe_count)
 562                         RETURN(0);
 563         }
 564         /* If we were passed specific striping params, then a failure to
 565          * meet those requirements is an error, since we can't reallocate
 566          * that memory (it might be part of a larger array or something).
 567          *
 568          * We can only get here if lsm_stripe_count was originally > 1.
 569          */
 570         CERROR("can't lstripe objid "LPX64": have "LPSZ" want %u\n",
 571                lsm->lsm_object_id, idx_pos - idx_arr, lsm->lsm_stripe_count);
 572         RETURN(-EFBIG);
 573 }
 574
 575 /* Alloc objects on osts with optimization based on:
 576    - free space
 577    - network resources (shared OSS's)
 578 */
 579 static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt)
 580 {
 581         struct lov_obd *lov = &exp->exp_obd->u.lov;
 582         static time_t last_warn = 0;
 583         time_t now = cfs_time_current_sec();
 584         __u64 total_bavail, total_weight = 0;
 585         __u32 ost_count;
 586         int nfound, good_osts, i, warn = 0, rc = 0;
 587         ENTRY;
 588
 589         lov_getref(exp->exp_obd);
 590         down_write(&lov->lov_qos.lq_rw_sem);
 591
 592         ost_count = lov->desc.ld_tgt_count;
 593
 594         if (lov->desc.ld_active_tgt_count < 2)
 595                 GOTO(out, rc = -EAGAIN);
 596
 597         rc = qos_calc_ppo(exp->exp_obd);
 598         if (rc)
 599                 GOTO(out, rc);
 600
 601         total_bavail = 0;
 602         good_osts = 0;
 603         /* Warn users about zero available space/inode every 30 min */
 604         if (cfs_time_sub(now, last_warn) > 60 * 30)
 605                 warn = 1;
 606         /* Find all the OSTs that are valid stripe candidates */
 607         for (i = 0; i < ost_count; i++) {
 608                 __u64 bavail;
 609
 610                 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
 611                         continue;
 612                 bavail = TGT_BAVAIL(i);
 613                 if (!bavail) {
 614                         if (warn) {
 615                                 CDEBUG(D_QOS, "no free space on %s\n",
 616                                      obd_uuid2str(&lov->lov_tgts[i]->ltd_uuid));
 617                                 last_warn = now;
 618                         }
 619                         continue;
 620                 }
 621                 if (!TGT_FFREE(i)) {
 622                         if (warn) {
 623                                 CDEBUG(D_QOS, "no free inodes on %s\n",
 624                                      obd_uuid2str(&lov->lov_tgts[i]->ltd_uuid));
 625                                 last_warn = now;
 626                         }
 627                         continue;
 628                 }
 629
 630                 lov->lov_tgts[i]->ltd_qos.ltq_usable = 1;
 631                 qos_calc_weight(lov, i);
 632                 total_bavail += bavail;
 633                 total_weight += lov->lov_tgts[i]->ltd_qos.ltq_weight;
 634
 635                 good_osts++;
 636         }
 637
 638         if (!total_bavail)
 639                 GOTO(out, rc = -ENOSPC);
 640
 641         /* if we don't have enough good OSTs, we reduce the stripe count. */
 642         if (good_osts < *stripe_cnt)
 643                 *stripe_cnt = good_osts;
 644
 645         if (!*stripe_cnt)
 646                 GOTO(out, rc = -EAGAIN);
 647
 648         /* Find enough OSTs with weighted random allocation. */
 649         nfound = 0;
 650         while (nfound < *stripe_cnt) {
 651                 __u64 rand, cur_weight;
 652
 653                 cur_weight = 0;
 654                 rc = -ENODEV;
 655
 656                 if (total_weight) {
 657 #if BITS_PER_LONG == 32
 658                         rand = ll_rand() % (unsigned)total_weight;
 659                         /* If total_weight > 32-bit, first generate the high
 660                          * 32 bits of the random number, then add in the low
 661                          * 32 bits (truncated to the upper limit, if needed) */
 662                         if (total_weight > 0xffffffffULL)
 663                                 rand = (__u64)(ll_rand() %
 664                                           (unsigned)(total_weight >> 32)) << 32;
 665                         else
 666                                 rand = 0;
 667
 668                         if (rand == (total_weight & 0xffffffff00000000ULL))
 669                                 rand |= ll_rand() % (unsigned)total_weight;
 670                         else
 671                                 rand |= ll_rand();
 672
 673 #else
 674                         rand = ((__u64)ll_rand() << 32 | ll_rand()) %
 675                                 total_weight;
 676 #endif
 677                 } else {
 678                         rand = 0;
 679                 }
 680
 681                 /* On average, this will hit larger-weighted osts more often.
 682                    0-weight osts will always get used last (only when rand=0).*/
 683                 for (i = 0; i < ost_count; i++) {
 684                         if (!lov->lov_tgts[i] ||
 685                             !lov->lov_tgts[i]->ltd_qos.ltq_usable)
 686                                 continue;
 687                         cur_weight += lov->lov_tgts[i]->ltd_qos.ltq_weight;
 688                         if (cur_weight >= rand) {
 689 #ifdef QOS_DEBUG
 690                                 CDEBUG(D_QOS, "assigned stripe=%d to idx=%d\n",
 691                                        nfound, i);
 692 #endif
 693                                 idx_arr[nfound++] = i;
 694                                 qos_used(lov, i, &total_weight);
 695                                 rc = 0;
 696                                 break;
 697                         }
 698                 }
 699                 /* should never satisfy below condition */
 700                 if (rc) {
 701                         CERROR("Didn't find any OSTs?\n");
 702                         break;
 703                 }
 704         }
 705         LASSERT(nfound == *stripe_cnt);
 706
 707 out:
 708         up_write(&lov->lov_qos.lq_rw_sem);
 709
 710         if (rc == -EAGAIN)
 711                 rc = alloc_rr(lov, idx_arr, stripe_cnt);
 712
 713         lov_putref(exp->exp_obd);
 714         RETURN(rc);
 715 }
 716
 717 /* return new alloced stripe count on success */
 718 static int alloc_idx_array(struct obd_export *exp, struct lov_stripe_md *lsm,
 719                            int newea, int **idx_arr, int *arr_cnt)
 720 {
 721         struct lov_obd *lov = &exp->exp_obd->u.lov;
 722         int stripe_cnt = lsm->lsm_stripe_count;
 723         int i, rc = 0;
 724         int *tmp_arr = NULL;
 725         ENTRY;
 726
 727         *arr_cnt = stripe_cnt;
 728         OBD_ALLOC(tmp_arr, *arr_cnt * sizeof(int));
 729         if (tmp_arr == NULL)
 730                 RETURN(-ENOMEM);
 731         for (i = 0; i < *arr_cnt; i++)
 732                 tmp_arr[i] = -1;
 733
 734         if (newea ||
 735             lsm->lsm_oinfo[0]->loi_ost_idx >= lov->desc.ld_tgt_count)
 736                 rc = alloc_qos(exp, tmp_arr, &stripe_cnt);
 737         else
 738                 rc = alloc_specific(lov, lsm, tmp_arr);
 739
 740         if (rc)
 741                 GOTO(out_arr, rc);
 742
 743         *idx_arr = tmp_arr;
 744         RETURN(stripe_cnt);
 745 out_arr:
 746         OBD_FREE(tmp_arr, *arr_cnt * sizeof(int));
 747         *arr_cnt = 0;
 748         RETURN(rc);
 749 }
 750
 751 static void free_idx_array(int *idx_arr, int arr_cnt)
 752 {
 753         if (arr_cnt)
 754                 OBD_FREE(idx_arr, arr_cnt * sizeof(int));
 755 }
 756
 757 int qos_prep_create(struct obd_export *exp, struct lov_request_set *set)
 758 {
 759         struct lov_obd *lov = &exp->exp_obd->u.lov;
 760         struct lov_stripe_md *lsm;
 761         struct obdo *src_oa = set->set_oi->oi_oa;
 762         struct obd_trans_info *oti = set->set_oti;
 763         int i, stripes, rc = 0, newea = 0;
 764         int *idx_arr, idx_cnt = 0;
 765         ENTRY;
 766
 767         LASSERT(src_oa->o_valid & OBD_MD_FLID);
 768
 769         if (set->set_oi->oi_md == NULL) {
 770                 int stripe_cnt = lov_get_stripecnt(lov, 0);
 771
 772                 /* If the MDS file was truncated up to some size, stripe over
 773                  * enough OSTs to allow the file to be created at that size.
 774                  * This may mean we use more than the default # of stripes. */
 775                 if (src_oa->o_valid & OBD_MD_FLSIZE) {
 776                         obd_size min_bavail = LUSTRE_STRIPE_MAXBYTES;
 777
 778                         /* Find a small number of stripes we can use
 779                            (up to # of active osts). */
 780                         stripes = 1;
 781                         lov_getref(exp->exp_obd);
 782                         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 783                                 if (!lov->lov_tgts[i] ||
 784                                     !lov->lov_tgts[i]->ltd_active)
 785                                         continue;
 786                                 min_bavail = min(min_bavail, TGT_BAVAIL(i));
 787                                 if (min_bavail * stripes > src_oa->o_size)
 788                                         break;
 789                                 stripes++;
 790                         }
 791                         lov_putref(exp->exp_obd);
 792
 793                         if (stripes < stripe_cnt)
 794                                 stripes = stripe_cnt;
 795                 } else {
 796                         stripes = stripe_cnt;
 797                 }
 798
 799                 rc = lov_alloc_memmd(&set->set_oi->oi_md, stripes,
 800                                      lov->desc.ld_pattern ?
 801                                      lov->desc.ld_pattern : LOV_PATTERN_RAID0,
 802                                      LOV_MAGIC);
 803                 if (rc < 0)
 804                         GOTO(out_err, rc);
 805                 rc = 0;
 806                 newea = 1;
 807         }
 808
 809         lsm = set->set_oi->oi_md;
 810         lsm->lsm_object_id = src_oa->o_id;
 811         if (!lsm->lsm_stripe_size)
 812                 lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size;
 813         if (!lsm->lsm_pattern) {
 814                 LASSERT(lov->desc.ld_pattern);
 815                 lsm->lsm_pattern = lov->desc.ld_pattern;
 816         }
 817
 818         stripes = alloc_idx_array(exp, lsm, newea, &idx_arr, &idx_cnt);
 819         if (stripes <= 0)
 820                 GOTO(out_err, rc = stripes ? stripes : -EIO);
 821         LASSERTF(stripes <= lsm->lsm_stripe_count,"requested %d allocated %d\n",
 822                  lsm->lsm_stripe_count, stripes);
 823
 824         for (i = 0; i < stripes; i++) {
 825                 struct lov_request *req;
 826                 int ost_idx = idx_arr[i];
 827                 LASSERT(ost_idx >= 0);
 828
 829                 OBD_ALLOC(req, sizeof(*req));
 830                 if (req == NULL)
 831                         GOTO(out_err, rc = -ENOMEM);
 832                 lov_set_add_req(req, set);
 833
 834                 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
 835                 OBD_ALLOC(req->rq_oi.oi_md, req->rq_buflen);
 836                 if (req->rq_oi.oi_md == NULL)
 837                         GOTO(out_err, rc = -ENOMEM);
 838
 839                 req->rq_oi.oi_oa = obdo_alloc();
 840                 if (req->rq_oi.oi_oa == NULL)
 841                         GOTO(out_err, rc = -ENOMEM);
 842
 843                 req->rq_idx = ost_idx;
 844                 req->rq_stripe = i;
 845                 /* create data objects with "parent" OA */
 846                 memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
 847
 848                 /* XXX When we start creating objects on demand, we need to
 849                  *     make sure that we always create the object on the
 850                  *     stripe which holds the existing file size.
 851                  */
 852                 if (src_oa->o_valid & OBD_MD_FLSIZE) {
 853                         req->rq_oi.oi_oa->o_size =
 854                                 lov_size_to_stripe(lsm, src_oa->o_size, i);
 855
 856                         CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
 857                                i, req->rq_oi.oi_oa->o_size, src_oa->o_size);
 858                 }
 859
 860         }
 861         LASSERT(set->set_count == stripes);
 862
 863         if (stripes < lsm->lsm_stripe_count)
 864                 qos_shrink_lsm(set);
 865
 866         if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) {
 867                 oti_alloc_cookies(oti, set->set_count);
 868                 if (!oti->oti_logcookies)
 869                         GOTO(out_err, rc = -ENOMEM);
 870                 set->set_cookies = oti->oti_logcookies;
 871         }
 872 out_err:
 873         if (newea && rc)
 874                 obd_free_memmd(exp, &set->set_oi->oi_md);
 875         free_idx_array(idx_arr, idx_cnt);
 876         EXIT;
 877         return rc;
 878 }
 879
 880 void qos_update(struct lov_obd *lov)
 881 {
 882         ENTRY;
 883         lov->lov_qos.lq_dirty = 1;
 884 }
 885