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         /* Do actual allocation. */
 337         down_write(&lov->lov_qos.lq_rw_sem);
 338         ost_count = lov->desc.ld_tgt_count;
 339
 340         if (lov->lov_qos.lq_rr_size)
 341                 OBD_FREE(lov->lov_qos.lq_rr_array, lov->lov_qos.lq_rr_size);
 342         lov->lov_qos.lq_rr_size = ost_count *
 343                 sizeof(lov->lov_qos.lq_rr_array[0]);
 344         OBD_ALLOC(lov->lov_qos.lq_rr_array, lov->lov_qos.lq_rr_size);
 345         if (!lov->lov_qos.lq_rr_array) {
 346                 lov->lov_qos.lq_rr_size = 0;
 347                 up_write(&lov->lov_qos.lq_rw_sem);
 348                 RETURN(-ENOMEM);
 349         }
 350
 351         real_count = 0;
 352         for (i = 0; i < ost_count; i++) {
 353                 lov->lov_qos.lq_rr_array[i] = LOV_QOS_EMPTY;
 354                 if (lov->lov_tgts[i])
 355                         real_count++;
 356         }
 357
 358         /* Place all the OSTs from 1 OSS at the same time. */
 359         placed = 0;
 360         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 361                 int j = 0;
 362                 for (i = 0; i < ost_count; i++) {
 363                         if (lov->lov_tgts[i] &&
 364                             lov->lov_tgts[i]->ltd_qos.ltq_oss == oss) {
 365                               /* Evenly space these OSTs across arrayspace */
 366                               int next = j * ost_count / oss->lqo_ost_count;
 367                               while (lov->lov_qos.lq_rr_array[next] !=
 368                                      LOV_QOS_EMPTY)
 369                                       next = (next + 1) % ost_count;
 370                               lov->lov_qos.lq_rr_array[next] = i;
 371                               j++;
 372                               placed++;
 373                         }
 374                 }
 375                 LASSERT(j == oss->lqo_ost_count);
 376         }
 377
 378         lov->lov_qos.lq_dirty_rr = 0;
 379         up_write(&lov->lov_qos.lq_rw_sem);
 380
 381         if (placed != real_count) {
 382                 /* This should never happen */
 383                 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
 384                                    "round-robin list (%d of %d).\n",
 385                                    placed, real_count);
 386                 for (i = 0; i < ost_count; i++) {
 387                         LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
 388                                  lov->lov_qos.lq_rr_array[i]);
 389                 }
 390                 lov->lov_qos.lq_dirty_rr = 1;
 391                 RETURN(-EAGAIN);
 392         }
 393
 394 #ifdef QOS_DEBUG
 395         for (i = 0; i < ost_count; i++) {
 396                 LCONSOLE(D_QOS, "rr #%d ost idx=%d\n", i,
 397                          lov->lov_qos.lq_rr_array[i]);
 398         }
 399 #endif
 400
 401         RETURN(0);
 402 }
 403
 404
 405 void qos_shrink_lsm(struct lov_request_set *set)
 406 {
 407         struct lov_stripe_md *lsm = set->set_oi->oi_md, *lsm_new;
 408         /* XXX LOV STACKING call into osc for sizes */
 409         unsigned oldsize, newsize;
 410
 411         if (set->set_oti && set->set_cookies && set->set_cookie_sent) {
 412                 struct llog_cookie *cookies;
 413                 oldsize = lsm->lsm_stripe_count * sizeof(*cookies);
 414                 newsize = set->set_count * sizeof(*cookies);
 415
 416                 cookies = set->set_cookies;
 417                 oti_alloc_cookies(set->set_oti, set->set_count);
 418                 if (set->set_oti->oti_logcookies) {
 419                         memcpy(set->set_oti->oti_logcookies, cookies, newsize);
 420                         OBD_FREE(cookies, oldsize);
 421                         set->set_cookies = set->set_oti->oti_logcookies;
 422                 } else {
 423                         CWARN("'leaking' %d bytes\n", oldsize - newsize);
 424                 }
 425         }
 426
 427         CWARN("using fewer stripes for object "LPU64": old %u new %u\n",
 428               lsm->lsm_object_id, lsm->lsm_stripe_count, set->set_count);
 429         LASSERT(lsm->lsm_stripe_count >= set->set_count);
 430
 431         newsize = lov_stripe_md_size(set->set_count);
 432         OBD_ALLOC(lsm_new, newsize);
 433         if (lsm_new != NULL) {
 434                 int i;
 435                 memcpy(lsm_new, lsm, sizeof(*lsm));
 436                 for (i = 0; i < lsm->lsm_stripe_count; i++) {
 437                         if (i < set->set_count) {
 438                                 lsm_new->lsm_oinfo[i] = lsm->lsm_oinfo[i];
 439                                 continue;
 440                         }
 441                         OBD_SLAB_FREE(lsm->lsm_oinfo[i], lov_oinfo_slab,
 442                                       sizeof(struct lov_oinfo));
 443                 }
 444                 lsm_new->lsm_stripe_count = set->set_count;
 445                 OBD_FREE(lsm, sizeof(struct lov_stripe_md) +
 446                          lsm->lsm_stripe_count * sizeof(struct lov_oinfo *));
 447                 set->set_oi->oi_md = lsm_new;
 448         } else {
 449                 CWARN("'leaking' few bytes\n");
 450         }
 451 }
 452
 453 int qos_remedy_create(struct lov_request_set *set, struct lov_request *req)
 454 {
 455         struct lov_stripe_md *lsm = set->set_oi->oi_md;
 456         struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
 457         unsigned ost_idx, ost_count = lov->desc.ld_tgt_count;
 458         int stripe, i, rc = -EIO;
 459         ENTRY;
 460
 461         ost_idx = (req->rq_idx + lsm->lsm_stripe_count) % ost_count;
 462         for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
 463                 if (!lov->lov_tgts[ost_idx] ||
 464                     !lov->lov_tgts[ost_idx]->ltd_active)
 465                         continue;
 466                 /* check if objects has been created on this ost */
 467                 for (stripe = 0; stripe < lsm->lsm_stripe_count; stripe++) {
 468                         if (stripe == req->rq_stripe)
 469                                 continue;
 470                         if (ost_idx == lsm->lsm_oinfo[stripe]->loi_ost_idx)
 471                                 break;
 472                 }
 473
 474                 if (stripe >= lsm->lsm_stripe_count) {
 475                         req->rq_idx = ost_idx;
 476                         rc = obd_create(lov->lov_tgts[ost_idx]->ltd_exp,
 477                                         req->rq_oi.oi_oa, &req->rq_oi.oi_md,
 478                                         set->set_oti);
 479                         if (!rc)
 480                                 break;
 481                 }
 482         }
 483         RETURN(rc);
 484 }
 485
 486 static int min_stripe_count(int stripe_cnt, int flags)
 487 {
 488         return (flags & LOV_USES_DEFAULT_STRIPE ?
 489                 stripe_cnt - (stripe_cnt / 4) : stripe_cnt);
 490 }
 491
 492 #define LOV_CREATE_RESEED_MULT 4
 493 #define LOV_CREATE_RESEED_MIN  1000
 494 /* Allocate objects on osts with round-robin algorithm */
 495 static int alloc_rr(struct lov_obd *lov, int *idx_arr, int *stripe_cnt,
 496                     int flags)
 497 {
 498         unsigned array_idx, ost_count = lov->desc.ld_tgt_count;
 499         unsigned ost_active_count = lov->desc.ld_active_tgt_count;
 500         int i, *idx_pos;
 501         __u32 ost_idx;
 502         int ost_start_idx_temp;
 503         int speed = 0;
 504         int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
 505         ENTRY;
 506
 507         i = qos_calc_rr(lov);
 508         if (i)
 509                 RETURN(i);
 510
 511         if (--lov->lov_start_count <= 0) {
 512                 lov->lov_start_idx = ll_rand() % ost_count;
 513                 lov->lov_start_count =
 514                         (LOV_CREATE_RESEED_MIN / max(ost_active_count, 1U) +
 515                          LOV_CREATE_RESEED_MULT) * max(ost_active_count, 1U);
 516         } else if (stripe_cnt_min >= ost_active_count ||
 517                    lov->lov_start_idx > ost_count) {
 518                 /* If we have allocated from all of the OSTs, slowly
 519                  * precess the next start if the OST/stripe count isn't
 520                  * already doing this for us. */
 521                 lov->lov_start_idx %= ost_count;
 522                 if (*stripe_cnt > 1 && (ost_active_count % (*stripe_cnt)) != 1)
 523                         ++lov->lov_offset_idx;
 524         }
 525         down_read(&lov->lov_qos.lq_rw_sem);
 526         ost_start_idx_temp = lov->lov_start_idx;
 527
 528 repeat_find:
 529         array_idx = (lov->lov_start_idx + lov->lov_offset_idx) % ost_count;
 530         idx_pos = idx_arr;
 531 #ifdef QOS_DEBUG
 532         CDEBUG(D_QOS, "want %d startidx %d startcnt %d offset %d active %d "
 533                "count %d arrayidx %d\n",
 534                stripe_cnt, lov->lov_start_idx, lov->lov_start_count,
 535                lov->lov_offset_idx, ost_active_count, ost_count, array_idx);
 536 #endif
 537
 538         for (i = 0; i < ost_count; i++, array_idx=(array_idx + 1) % ost_count) {
 539                 ++lov->lov_start_idx;
 540                 ost_idx = lov->lov_qos.lq_rr_array[array_idx];
 541 #ifdef QOS_DEBUG
 542                 CDEBUG(D_QOS, "#%d strt %d act %d strp %d ary %d idx %d\n",
 543                        i, lov->lov_start_idx,
 544                        ((ost_idx != LOV_QOS_EMPTY) && lov->lov_tgts[ost_idx]) ?
 545                        lov->lov_tgts[ost_idx]->ltd_active : 0,
 546                        idx_pos - idx_arr, array_idx, ost_idx);
 547 #endif
 548                 if ((ost_idx == LOV_QOS_EMPTY) || !lov->lov_tgts[ost_idx] ||
 549                     !lov->lov_tgts[ost_idx]->ltd_active)
 550                         continue;
 551
 552                 /* Fail Check before osc_precreate() is called
 553                    so we can only 'fail' single OSC. */
 554                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
 555                         continue;
 556
 557                 /* Drop slow OSCs if we can */
 558                 if (obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed)
 559                         continue;
 560
 561                 *idx_pos = ost_idx;
 562                 idx_pos++;
 563                 /* We have enough stripes */
 564                 if (idx_pos - idx_arr == *stripe_cnt)
 565                         break;
 566         }
 567         if ((speed < 2) && (idx_pos - idx_arr < stripe_cnt_min)) {
 568                 /* Try again, allowing slower OSCs */
 569                 speed++;
 570                 lov->lov_start_idx = ost_start_idx_temp;
 571                 goto repeat_find;
 572         }
 573
 574         up_read(&lov->lov_qos.lq_rw_sem);
 575
 576         *stripe_cnt = idx_pos - idx_arr;
 577         RETURN(0);
 578 }
 579
 580 /* alloc objects on osts with specific stripe offset */
 581 static int alloc_specific(struct lov_obd *lov, struct lov_stripe_md *lsm,
 582                           int *idx_arr)
 583 {
 584         unsigned ost_idx, ost_count = lov->desc.ld_tgt_count;
 585         int i, *idx_pos;
 586         int speed = 0;
 587         ENTRY;
 588
 589 repeat_find:
 590         ost_idx = lsm->lsm_oinfo[0]->loi_ost_idx;
 591         idx_pos = idx_arr;
 592         for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
 593                 if (!lov->lov_tgts[ost_idx] ||
 594                     !lov->lov_tgts[ost_idx]->ltd_active) {
 595                         continue;
 596                 }
 597
 598                 /* Fail Check before osc_precreate() is called
 599                    so we can only 'fail' single OSC. */
 600                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
 601                         continue;
 602
 603                 /* Drop slow OSCs if we can, but not for requested start idx */
 604                 if ((obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed) &&
 605                     (i != 0 || speed < 2))
 606                         continue;
 607
 608                 *idx_pos = ost_idx;
 609                 idx_pos++;
 610                 /* We have enough stripes */
 611                 if (idx_pos - idx_arr == lsm->lsm_stripe_count)
 612                         RETURN(0);
 613         }
 614         if (speed < 2) {
 615                 /* Try again, allowing slower OSCs */
 616                 speed++;
 617                 goto repeat_find;
 618         }
 619
 620         /* If we were passed specific striping params, then a failure to
 621          * meet those requirements is an error, since we can't reallocate
 622          * that memory (it might be part of a larger array or something).
 623          *
 624          * We can only get here if lsm_stripe_count was originally > 1.
 625          */
 626         CERROR("can't lstripe objid "LPX64": have %d want %u\n",
 627                lsm->lsm_object_id, (int)(idx_pos - idx_arr),
 628                lsm->lsm_stripe_count);
 629         RETURN(-EFBIG);
 630 }
 631
 632 /* Alloc objects on osts with optimization based on:
 633    - free space
 634    - network resources (shared OSS's)
 635 */
 636 static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt,
 637                      int flags)
 638 {
 639         struct lov_obd *lov = &exp->exp_obd->u.lov;
 640         static time_t last_warn = 0;
 641         time_t now = cfs_time_current_sec();
 642         __u64 total_bavail, total_weight = 0;
 643         __u32 ost_count;
 644         int nfound, good_osts, i, warn = 0, rc = 0;
 645         int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
 646         ENTRY;
 647
 648         if (stripe_cnt_min < 1)
 649                 GOTO(out, rc = -EINVAL);
 650
 651         lov_getref(exp->exp_obd);
 652         down_write(&lov->lov_qos.lq_rw_sem);
 653
 654         ost_count = lov->desc.ld_tgt_count;
 655
 656         if (lov->desc.ld_active_tgt_count < 2)
 657                 GOTO(out, rc = -EAGAIN);
 658
 659         rc = qos_calc_ppo(exp->exp_obd);
 660         if (rc)
 661                 GOTO(out, rc);
 662
 663         total_bavail = 0;
 664         good_osts = 0;
 665         /* Warn users about zero available space/inode every 30 min */
 666         if (cfs_time_sub(now, last_warn) > 60 * 30)
 667                 warn = 1;
 668         /* Find all the OSTs that are valid stripe candidates */
 669         for (i = 0; i < ost_count; i++) {
 670                 __u64 bavail;
 671
 672                 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
 673                         continue;
 674                 bavail = TGT_BAVAIL(i);
 675                 if (!bavail) {
 676                         if (warn) {
 677                                 CDEBUG(D_QOS, "no free space on %s\n",
 678                                      obd_uuid2str(&lov->lov_tgts[i]->ltd_uuid));
 679                                 last_warn = now;
 680                         }
 681                         continue;
 682                 }
 683                 if (!TGT_FFREE(i)) {
 684                         if (warn) {
 685                                 CDEBUG(D_QOS, "no free inodes on %s\n",
 686                                      obd_uuid2str(&lov->lov_tgts[i]->ltd_uuid));
 687                                 last_warn = now;
 688                         }
 689                         continue;
 690                 }
 691
 692                 /* Fail Check before osc_precreate() is called
 693                    so we can only 'fail' single OSC. */
 694                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && i == 0)
 695                         continue;
 696
 697                 if (obd_precreate(lov->lov_tgts[i]->ltd_exp) > 2)
 698                         continue;
 699
 700                 lov->lov_tgts[i]->ltd_qos.ltq_usable = 1;
 701                 qos_calc_weight(lov, i);
 702                 total_bavail += bavail;
 703                 total_weight += lov->lov_tgts[i]->ltd_qos.ltq_weight;
 704
 705                 good_osts++;
 706         }
 707
 708         if (good_osts < stripe_cnt_min)
 709                 GOTO(out, rc = -EAGAIN);
 710
 711         if (!total_bavail)
 712                 GOTO(out, rc = -ENOSPC);
 713
 714         /* We have enough osts */
 715         if (good_osts < *stripe_cnt)
 716                 *stripe_cnt = good_osts;
 717
 718         /* Find enough OSTs with weighted random allocation. */
 719         nfound = 0;
 720         while (nfound < *stripe_cnt) {
 721                 __u64 rand, cur_weight;
 722
 723                 cur_weight = 0;
 724                 rc = -ENODEV;
 725
 726                 if (total_weight) {
 727 #if BITS_PER_LONG == 32
 728                         rand = ll_rand() % (unsigned)total_weight;
 729                         /* If total_weight > 32-bit, first generate the high
 730                          * 32 bits of the random number, then add in the low
 731                          * 32 bits (truncated to the upper limit, if needed) */
 732                         if (total_weight > 0xffffffffULL)
 733                                 rand = (__u64)(ll_rand() %
 734                                           (unsigned)(total_weight >> 32)) << 32;
 735                         else
 736                                 rand = 0;
 737
 738                         if (rand == (total_weight & 0xffffffff00000000ULL))
 739                                 rand |= ll_rand() % (unsigned)total_weight;
 740                         else
 741                                 rand |= ll_rand();
 742
 743 #else
 744                         rand = ((__u64)ll_rand() << 32 | ll_rand()) %
 745                                 total_weight;
 746 #endif
 747                 } else {
 748                         rand = 0;
 749                 }
 750
 751                 /* On average, this will hit larger-weighted osts more often.
 752                    0-weight osts will always get used last (only when rand=0).*/
 753                 for (i = 0; i < ost_count; i++) {
 754                         if (!lov->lov_tgts[i] ||
 755                             !lov->lov_tgts[i]->ltd_qos.ltq_usable)
 756                                 continue;
 757
 758                         cur_weight += lov->lov_tgts[i]->ltd_qos.ltq_weight;
 759                         if (cur_weight >= rand) {
 760 #ifdef QOS_DEBUG
 761                                 CDEBUG(D_QOS, "assigned stripe=%d to idx=%d\n",
 762                                        nfound, i);
 763 #endif
 764                                 idx_arr[nfound++] = i;
 765                                 qos_used(lov, i, &total_weight);
 766                                 rc = 0;
 767                                 break;
 768                         }
 769                 }
 770                 /* should never satisfy below condition */
 771                 if (rc) {
 772                         CERROR("Didn't find any OSTs?\n");
 773                         break;
 774                 }
 775         }
 776         LASSERT(nfound == *stripe_cnt);
 777
 778 out:
 779         up_write(&lov->lov_qos.lq_rw_sem);
 780
 781         if (rc == -EAGAIN)
 782                 rc = alloc_rr(lov, idx_arr, stripe_cnt, flags);
 783
 784         lov_putref(exp->exp_obd);
 785         RETURN(rc);
 786 }
 787
 788 /* return new alloced stripe count on success */
 789 static int alloc_idx_array(struct obd_export *exp, struct lov_stripe_md *lsm,
 790                            int newea, int **idx_arr, int *arr_cnt, int flags)
 791 {
 792         struct lov_obd *lov = &exp->exp_obd->u.lov;
 793         int stripe_cnt = lsm->lsm_stripe_count;
 794         int i, rc = 0;
 795         int *tmp_arr = NULL;
 796         ENTRY;
 797
 798         *arr_cnt = stripe_cnt;
 799         OBD_ALLOC(tmp_arr, *arr_cnt * sizeof(int));
 800         if (tmp_arr == NULL)
 801                 RETURN(-ENOMEM);
 802         for (i = 0; i < *arr_cnt; i++)
 803                 tmp_arr[i] = -1;
 804
 805         if (newea ||
 806             lsm->lsm_oinfo[0]->loi_ost_idx >= lov->desc.ld_tgt_count)
 807                 rc = alloc_qos(exp, tmp_arr, &stripe_cnt, flags);
 808         else
 809                 rc = alloc_specific(lov, lsm, tmp_arr);
 810
 811         if (rc)
 812                 GOTO(out_arr, rc);
 813
 814         *idx_arr = tmp_arr;
 815         RETURN(stripe_cnt);
 816 out_arr:
 817         OBD_FREE(tmp_arr, *arr_cnt * sizeof(int));
 818         *arr_cnt = 0;
 819         RETURN(rc);
 820 }
 821
 822 static void free_idx_array(int *idx_arr, int arr_cnt)
 823 {
 824         if (arr_cnt)
 825                 OBD_FREE(idx_arr, arr_cnt * sizeof(int));
 826 }
 827
 828 int qos_prep_create(struct obd_export *exp, struct lov_request_set *set)
 829 {
 830         struct lov_obd *lov = &exp->exp_obd->u.lov;
 831         struct lov_stripe_md *lsm;
 832         struct obdo *src_oa = set->set_oi->oi_oa;
 833         struct obd_trans_info *oti = set->set_oti;
 834         int i, stripes, rc = 0, newea = 0;
 835         int flag = LOV_USES_ASSIGNED_STRIPE;
 836         int *idx_arr = NULL, idx_cnt = 0;
 837         ENTRY;
 838
 839         LASSERT(src_oa->o_valid & OBD_MD_FLID);
 840
 841         if (set->set_oi->oi_md == NULL) {
 842                 int stripes_def = lov_get_stripecnt(lov, 0);
 843
 844                 /* If the MDS file was truncated up to some size, stripe over
 845                  * enough OSTs to allow the file to be created at that size.
 846                  * This may mean we use more than the default # of stripes. */
 847                 if (src_oa->o_valid & OBD_MD_FLSIZE) {
 848                         obd_size min_bavail = LUSTRE_STRIPE_MAXBYTES;
 849
 850                         /* Find a small number of stripes we can use
 851                            (up to # of active osts). */
 852                         stripes = 1;
 853                         lov_getref(exp->exp_obd);
 854                         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 855                                 if (!lov->lov_tgts[i] ||
 856                                     !lov->lov_tgts[i]->ltd_active)
 857                                         continue;
 858                                 min_bavail = min(min_bavail, TGT_BAVAIL(i));
 859                                 if (min_bavail * stripes > src_oa->o_size)
 860                                         break;
 861                                 stripes++;
 862                         }
 863                         lov_putref(exp->exp_obd);
 864
 865                         if (stripes < stripes_def)
 866                                 stripes = stripes_def;
 867                 } else {
 868                          flag = LOV_USES_DEFAULT_STRIPE;
 869                          stripes = stripes_def;
 870                 }
 871
 872                 rc = lov_alloc_memmd(&set->set_oi->oi_md, stripes,
 873                                      lov->desc.ld_pattern ?
 874                                      lov->desc.ld_pattern : LOV_PATTERN_RAID0,
 875                                      LOV_MAGIC);
 876                 if (rc < 0)
 877                         GOTO(out_err, rc);
 878                 newea = 1;
 879                 rc = 0;
 880         }
 881
 882         lsm = set->set_oi->oi_md;
 883         lsm->lsm_object_id = src_oa->o_id;
 884         if (!lsm->lsm_stripe_size)
 885                 lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size;
 886         if (!lsm->lsm_pattern) {
 887                 LASSERT(lov->desc.ld_pattern);
 888                 lsm->lsm_pattern = lov->desc.ld_pattern;
 889         }
 890
 891         stripes = alloc_idx_array(exp, lsm, newea, &idx_arr, &idx_cnt, flag);
 892         if (stripes <= 0)
 893                 GOTO(out_err, rc = stripes ? stripes : -EIO);
 894         LASSERTF(stripes <= lsm->lsm_stripe_count,"requested %d allocated %d\n",
 895                  lsm->lsm_stripe_count, stripes);
 896
 897         for (i = 0; i < stripes; i++) {
 898                 struct lov_request *req;
 899                 int ost_idx = idx_arr[i];
 900                 LASSERT(ost_idx >= 0);
 901
 902                 OBD_ALLOC(req, sizeof(*req));
 903                 if (req == NULL)
 904                         GOTO(out_err, rc = -ENOMEM);
 905                 lov_set_add_req(req, set);
 906
 907                 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
 908                 OBD_ALLOC(req->rq_oi.oi_md, req->rq_buflen);
 909                 if (req->rq_oi.oi_md == NULL)
 910                         GOTO(out_err, rc = -ENOMEM);
 911
 912                 OBDO_ALLOC(req->rq_oi.oi_oa);
 913                 if (req->rq_oi.oi_oa == NULL)
 914                         GOTO(out_err, rc = -ENOMEM);
 915
 916                 req->rq_idx = ost_idx;
 917                 req->rq_stripe = i;
 918                 /* create data objects with "parent" OA */
 919                 memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
 920
 921                 /* XXX When we start creating objects on demand, we need to
 922                  *     make sure that we always create the object on the
 923                  *     stripe which holds the existing file size.
 924                  */
 925                 if (src_oa->o_valid & OBD_MD_FLSIZE) {
 926                         req->rq_oi.oi_oa->o_size =
 927                                 lov_size_to_stripe(lsm, src_oa->o_size, i);
 928
 929                         CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
 930                                i, req->rq_oi.oi_oa->o_size, src_oa->o_size);
 931                 }
 932         }
 933         LASSERT(set->set_count == stripes);
 934
 935         if (stripes < lsm->lsm_stripe_count)
 936                 qos_shrink_lsm(set);
 937
 938         if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) {
 939                 oti_alloc_cookies(oti, set->set_count);
 940                 if (!oti->oti_logcookies)
 941                         GOTO(out_err, rc = -ENOMEM);
 942                 set->set_cookies = oti->oti_logcookies;
 943         }
 944 out_err:
 945         if (newea && rc)
 946                 obd_free_memmd(exp, &set->set_oi->oi_md);
 947         if (idx_arr)
 948                 free_idx_array(idx_arr, idx_cnt);
 949         EXIT;
 950         return rc;
 951 }
 952
 953 void qos_update(struct lov_obd *lov)
 954 {
 955         ENTRY;
 956         lov->lov_qos.lq_dirty = 1;
 957 }
 958