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
  41 /* #define QOS_DEBUG 1 */
  42 #define D_QOS D_OTHER
  43
  44 #define TGT_BAVAIL(i) (lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bavail * \
  45                        lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bsize)
  46 #define TGT_FFREE(i)  (lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_ffree)
  47
  48
  49 int qos_add_tgt(struct obd_device *obd, __u32 index)
  50 {
  51         struct lov_obd *lov = &obd->u.lov;
  52         struct lov_qos_oss *oss, *temposs;
  53         struct obd_export *exp = lov->lov_tgts[index]->ltd_exp;
  54         int rc = 0, found = 0;
  55         ENTRY;
  56
  57         /* We only need this QOS struct on MDT, not clients - but we may not
  58            have registered the LOV's observer yet, so there's no way to know */
  59         if (!exp || !exp->exp_connection) {
  60                 CERROR("Missing connection\n");
  61                 RETURN(-ENOTCONN);
  62         }
  63
  64         down_write(&lov->lov_qos.lq_rw_sem);
  65         mutex_down(&lov->lov_lock);
  66         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
  67                 if (obd_uuid_equals(&oss->lqo_uuid,
  68                                     &exp->exp_connection->c_remote_uuid)) {
  69                         found++;
  70                         break;
  71                 }
  72         }
  73
  74         if (!found) {
  75                 OBD_ALLOC_PTR(oss);
  76                 if (!oss)
  77                         GOTO(out, rc = -ENOMEM);
  78                 memcpy(&oss->lqo_uuid,
  79                        &exp->exp_connection->c_remote_uuid,
  80                        sizeof(oss->lqo_uuid));
  81         } else {
  82                 /* Assume we have to move this one */
  83                 list_del(&oss->lqo_oss_list);
  84         }
  85
  86         oss->lqo_ost_count++;
  87         lov->lov_tgts[index]->ltd_qos.ltq_oss = oss;
  88
  89         /* Add sorted by # of OSTs.  Find the first entry that we're
  90            bigger than... */
  91         list_for_each_entry(temposs, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
  92                 if (oss->lqo_ost_count > temposs->lqo_ost_count)
  93                         break;
  94         }
  95         /* ...and add before it.  If we're the first or smallest, temposs
  96            points to the list head, and we add to the end. */
  97         list_add_tail(&oss->lqo_oss_list, &temposs->lqo_oss_list);
  98
  99         lov->lov_qos.lq_dirty = 1;
 100         lov->lov_qos.lq_dirty_rr = 1;
 101
 102         CDEBUG(D_QOS, "add tgt %s to OSS %s (%d OSTs)\n",
 103                obd_uuid2str(&lov->lov_tgts[index]->ltd_uuid),
 104                obd_uuid2str(&oss->lqo_uuid),
 105                oss->lqo_ost_count);
 106
 107 out:
 108         mutex_up(&lov->lov_lock);
 109         up_write(&lov->lov_qos.lq_rw_sem);
 110         RETURN(rc);
 111 }
 112
 113 int qos_del_tgt(struct obd_device *obd, __u32 index)
 114 {
 115         struct lov_obd *lov = &obd->u.lov;
 116         struct lov_qos_oss *oss;
 117         int rc = 0;
 118         ENTRY;
 119
 120         if (!lov->lov_tgts[index])
 121                 RETURN(0);
 122
 123         down_write(&lov->lov_qos.lq_rw_sem);
 124
 125         oss = lov->lov_tgts[index]->ltd_qos.ltq_oss;
 126         if (!oss)
 127                 GOTO(out, rc = -ENOENT);
 128
 129         oss->lqo_ost_count--;
 130         if (oss->lqo_ost_count == 0) {
 131                 CDEBUG(D_QOS, "removing OSS %s\n",
 132                        obd_uuid2str(&oss->lqo_uuid));
 133                 list_del(&oss->lqo_oss_list);
 134                 OBD_FREE_PTR(oss);
 135         }
 136
 137         lov->lov_qos.lq_dirty = 1;
 138         lov->lov_qos.lq_dirty_rr = 1;
 139 out:
 140         up_write(&lov->lov_qos.lq_rw_sem);
 141         RETURN(rc);
 142 }
 143
 144 /* Recalculate per-object penalties for OSSs and OSTs,
 145    depends on size of each ost in an oss */
 146 static int qos_calc_ppo(struct obd_device *obd)
 147 {
 148         struct lov_obd *lov = &obd->u.lov;
 149         struct lov_qos_oss *oss;
 150         __u64 ba_max, ba_min, temp;
 151         __u32 num_active;
 152         int rc, i, prio_wide;
 153         ENTRY;
 154
 155         if (!lov->lov_qos.lq_dirty)
 156                 GOTO(out, rc = 0);
 157
 158         num_active = lov->desc.ld_active_tgt_count - 1;
 159         if (num_active < 1)
 160                 GOTO(out, rc = -EAGAIN);
 161
 162         /* find bavail on each OSS */
 163         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 164                 oss->lqo_bavail = 0;
 165         }
 166         lov->lov_qos.lq_active_oss_count = 0;
 167
 168         /* How badly user wants to select osts "widely" (not recently chosen
 169            and not on recent oss's).  As opposed to "freely" (free space
 170            avail.) 0-256. */
 171         prio_wide = 256 - lov->lov_qos.lq_prio_free;
 172
 173         ba_min = (__u64)(-1);
 174         ba_max = 0;
 175         /* Calculate OST penalty per object */
 176         /* (lov ref taken in alloc_qos) */
 177         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 178                 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
 179                         continue;
 180                 temp = TGT_BAVAIL(i);
 181                 if (!temp)
 182                         continue;
 183                 ba_min = min(temp, ba_min);
 184                 ba_max = max(temp, ba_max);
 185
 186                 /* Count the number of usable OSS's */
 187                 if (lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_bavail == 0)
 188                         lov->lov_qos.lq_active_oss_count++;
 189                 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_bavail += temp;
 190
 191                 /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
 192                 temp >>= 1;
 193                 do_div(temp, num_active);
 194                 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj =
 195                         (temp * prio_wide) >> 8;
 196
 197                 if (lov->lov_qos.lq_reset == 0)
 198                         lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
 199         }
 200
 201         num_active = lov->lov_qos.lq_active_oss_count - 1;
 202         if (num_active < 1) {
 203                 /* If there's only 1 OSS, we can't penalize it, so instead
 204                    we have to double the OST penalty */
 205                 num_active = 1;
 206                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 207                         if (lov->lov_tgts[i] == NULL)
 208                                 continue;
 209                         lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj <<= 1;
 210                 }
 211         }
 212
 213         /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
 214         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 215                 temp = oss->lqo_bavail >> 1;
 216                 do_div(temp, oss->lqo_ost_count * num_active);
 217                 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
 218                 if (lov->lov_qos.lq_reset == 0)
 219                         oss->lqo_penalty = 0;
 220         }
 221
 222         lov->lov_qos.lq_dirty = 0;
 223         lov->lov_qos.lq_reset = 0;
 224
 225         /* If each ost has almost same free space,
 226          * do rr allocation for better creation performance */
 227         lov->lov_qos.lq_same_space = 0;
 228         temp = ba_max - ba_min;
 229         ba_min = (ba_min * 51) >> 8;     /* 51/256 = .20 */
 230         if (temp < ba_min) {
 231                 /* Difference is less than 20% */
 232                 lov->lov_qos.lq_same_space = 1;
 233                 /* Reset weights for the next time we enter qos mode */
 234                 lov->lov_qos.lq_reset = 0;
 235         }
 236         rc = 0;
 237
 238 out:
 239         if (!rc && lov->lov_qos.lq_same_space)
 240                 RETURN(-EAGAIN);
 241         RETURN(rc);
 242 }
 243
 244 static int qos_calc_weight(struct lov_obd *lov, int i)
 245 {
 246         __u64 temp, temp2;
 247
 248         /* Final ost weight = TGT_BAVAIL - ost_penalty - oss_penalty */
 249         temp = TGT_BAVAIL(i);
 250         temp2 = lov->lov_tgts[i]->ltd_qos.ltq_penalty +
 251                 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty;
 252         if (temp < temp2)
 253                 lov->lov_tgts[i]->ltd_qos.ltq_weight = 0;
 254         else
 255                 lov->lov_tgts[i]->ltd_qos.ltq_weight = temp - temp2;
 256         return 0;
 257 }
 258
 259 /* We just used this index for a stripe; adjust everyone's weights */
 260 static int qos_used(struct lov_obd *lov, __u32 index, __u64 *total_wt)
 261 {
 262         struct lov_qos_oss *oss;
 263         int i;
 264         ENTRY;
 265
 266         /* Don't allocate from this stripe anymore, until the next alloc_qos */
 267         lov->lov_tgts[index]->ltd_qos.ltq_usable = 0;
 268
 269         oss = lov->lov_tgts[index]->ltd_qos.ltq_oss;
 270
 271         /* Decay old penalty by half (we're adding max penalty, and don't
 272            want it to run away.) */
 273         lov->lov_tgts[index]->ltd_qos.ltq_penalty >>= 1;
 274         oss->lqo_penalty >>= 1;
 275
 276         /* Set max penalties for this OST and OSS */
 277         lov->lov_tgts[index]->ltd_qos.ltq_penalty +=
 278                 lov->lov_tgts[index]->ltd_qos.ltq_penalty_per_obj *
 279                 lov->desc.ld_active_tgt_count;
 280         oss->lqo_penalty += oss->lqo_penalty_per_obj *
 281                 lov->lov_qos.lq_active_oss_count;
 282
 283         /* Decrease all OSS penalties */
 284         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 285                 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
 286                         oss->lqo_penalty = 0;
 287                 else
 288                         oss->lqo_penalty -= oss->lqo_penalty_per_obj;
 289         }
 290
 291         *total_wt = 0;
 292         /* Decrease all OST penalties */
 293         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 294                 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
 295                         continue;
 296                 if (lov->lov_tgts[i]->ltd_qos.ltq_penalty <
 297                     lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj)
 298                         lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
 299                 else
 300                         lov->lov_tgts[i]->ltd_qos.ltq_penalty -=
 301                         lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj;
 302
 303                 qos_calc_weight(lov, i);
 304
 305                 /* Recalc the total weight of usable osts */
 306                 if (lov->lov_tgts[i]->ltd_qos.ltq_usable)
 307                         *total_wt += lov->lov_tgts[i]->ltd_qos.ltq_weight;
 308
 309 #ifdef QOS_DEBUG
 310                 CDEBUG(D_QOS, "recalc tgt %d avail="LPU64
 311                        " ostppo="LPU64" ostp="LPU64" ossppo="LPU64
 312                        " ossp="LPU64" wt="LPU64"\n",
 313                        i, TGT_BAVAIL(i) >> 10,
 314                        lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj >> 10,
 315                        lov->lov_tgts[i]->ltd_qos.ltq_penalty >> 10,
 316                        lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty_per_obj >> 10,
 317                        lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty >> 10,
 318                        lov->lov_tgts[i]->ltd_qos.ltq_weight >> 10);
 319 #endif
 320         }
 321
 322         RETURN(0);
 323 }
 324
 325 #define LOV_QOS_EMPTY ((__u32)-1)
 326 /* compute optimal round-robin order, based on OSTs per OSS */
 327 static int qos_calc_rr(struct lov_obd *lov)
 328 {
 329         struct lov_qos_oss *oss;
 330         unsigned ost_count, placed, real_count;
 331         int i;
 332         ENTRY;
 333
 334         if (!lov->lov_qos.lq_dirty_rr) {
 335                 LASSERT(lov->lov_qos.lq_rr_size);
 336                 RETURN(0);
 337         }
 338
 339         /* Do actuall allocation. */
 340         down_write(&lov->lov_qos.lq_rw_sem);
 341
 342         /*
 343          * Check again. While we were sleeping on @lq_rw_sem something could
 344          * change.
 345          */
 346         if (!lov->lov_qos.lq_dirty_rr) {
 347                 LASSERT(lov->lov_qos.lq_rr_size);
 348                 up_write(&lov->lov_qos.lq_rw_sem);
 349                 RETURN(0);
 350         }
 351
 352         ost_count = lov->desc.ld_tgt_count;
 353
 354         if (lov->lov_qos.lq_rr_size)
 355                 OBD_FREE(lov->lov_qos.lq_rr_array, lov->lov_qos.lq_rr_size);
 356         lov->lov_qos.lq_rr_size = ost_count *
 357                 sizeof(lov->lov_qos.lq_rr_array[0]);
 358         OBD_ALLOC(lov->lov_qos.lq_rr_array, lov->lov_qos.lq_rr_size);
 359         if (!lov->lov_qos.lq_rr_array) {
 360                 lov->lov_qos.lq_rr_size = 0;
 361                 up_write(&lov->lov_qos.lq_rw_sem);
 362                 RETURN(-ENOMEM);
 363         }
 364
 365         real_count = 0;
 366         for (i = 0; i < ost_count; i++) {
 367                 lov->lov_qos.lq_rr_array[i] = LOV_QOS_EMPTY;
 368                 if (lov->lov_tgts[i])
 369                         real_count++;
 370         }
 371
 372         /* Place all the OSTs from 1 OSS at the same time. */
 373         placed = 0;
 374         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 375                 int j = 0;
 376                 for (i = 0; i < ost_count; i++) {
 377                       LASSERT(lov->lov_tgts[i] != NULL);
 378                       if (lov->lov_tgts[i]->ltd_qos.ltq_oss == oss) {
 379                               /* Evenly space these OSTs across arrayspace */
 380                               int next = j * ost_count / oss->lqo_ost_count;
 381                               LASSERT(next < ost_count);
 382                               while (lov->lov_qos.lq_rr_array[next] !=
 383                                      LOV_QOS_EMPTY)
 384                                       next = (next + 1) % ost_count;
 385                               lov->lov_qos.lq_rr_array[next] = i;
 386                               j++;
 387                               placed++;
 388                       }
 389                 }
 390                 LASSERT(j == oss->lqo_ost_count);
 391         }
 392
 393         lov->lov_qos.lq_dirty_rr = 0;
 394         up_write(&lov->lov_qos.lq_rw_sem);
 395
 396         if (placed != real_count) {
 397                 /* This should never happen */
 398                 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
 399                                    "round-robin list (%d of %d).\n",
 400                                    placed, real_count);
 401                 for (i = 0; i < ost_count; i++) {
 402                         LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
 403                                  lov->lov_qos.lq_rr_array[i]);
 404                 }
 405                 lov->lov_qos.lq_dirty_rr = 1;
 406                 RETURN(-EAGAIN);
 407         }
 408
 409 #ifdef QOS_DEBUG
 410         for (i = 0; i < ost_count; i++) {
 411                 LCONSOLE(D_QOS, "rr #%d ost idx=%d\n", i,
 412                          lov->lov_qos.lq_rr_array[i]);
 413         }
 414 #endif
 415
 416         RETURN(0);
 417 }
 418
 419
 420 void qos_shrink_lsm(struct lov_request_set *set)
 421 {
 422         struct lov_stripe_md *lsm = set->set_oi->oi_md, *lsm_new;
 423         /* XXX LOV STACKING call into osc for sizes */
 424         unsigned oldsize, newsize;
 425
 426         if (set->set_oti && set->set_cookies && set->set_cookie_sent) {
 427                 struct llog_cookie *cookies;
 428                 oldsize = lsm->lsm_stripe_count * sizeof(*cookies);
 429                 newsize = set->set_count * sizeof(*cookies);
 430
 431                 cookies = set->set_cookies;
 432                 oti_alloc_cookies(set->set_oti, set->set_count);
 433                 if (set->set_oti->oti_logcookies) {
 434                         memcpy(set->set_oti->oti_logcookies, cookies, newsize);
 435                         OBD_FREE(cookies, oldsize);
 436                         set->set_cookies = set->set_oti->oti_logcookies;
 437                 } else {
 438                         CWARN("'leaking' %d bytes\n", oldsize - newsize);
 439                 }
 440         }
 441
 442         CWARN("using fewer stripes for object "LPX64": old %u new %u\n",
 443               lsm->lsm_object_id, lsm->lsm_stripe_count, set->set_count);
 444         LASSERT(lsm->lsm_stripe_count >= set->set_count);
 445
 446         newsize = lov_stripe_md_size(set->set_count);
 447         OBD_ALLOC(lsm_new, newsize);
 448         if (lsm_new != NULL) {
 449                 int i;
 450                 memcpy(lsm_new, lsm, sizeof(*lsm));
 451                 for (i = 0; i < lsm->lsm_stripe_count; i++) {
 452                         if (i < set->set_count) {
 453                                 lsm_new->lsm_oinfo[i] = lsm->lsm_oinfo[i];
 454                                 continue;
 455                         }
 456                         OBD_SLAB_FREE(lsm->lsm_oinfo[i], lov_oinfo_slab,
 457                                       sizeof(struct lov_oinfo));
 458                 }
 459                 lsm_new->lsm_stripe_count = set->set_count;
 460                 OBD_FREE(lsm, sizeof(struct lov_stripe_md) +
 461                          lsm->lsm_stripe_count * sizeof(struct lov_oinfo *));
 462                 set->set_oi->oi_md = lsm_new;
 463         } else {
 464                 CWARN("'leaking' few bytes\n");
 465         }
 466 }
 467
 468 int qos_remedy_create(struct lov_request_set *set, struct lov_request *req)
 469 {
 470         struct lov_stripe_md *lsm = set->set_oi->oi_md;
 471         struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
 472         unsigned ost_idx, ost_count = lov->desc.ld_tgt_count;
 473         int stripe, i, rc = -EIO;
 474         ENTRY;
 475
 476         ost_idx = (req->rq_idx + lsm->lsm_stripe_count) % ost_count;
 477         for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
 478                 if (!lov->lov_tgts[ost_idx] ||
 479                     !lov->lov_tgts[ost_idx]->ltd_active)
 480                         continue;
 481                 /* check if objects has been created on this ost */
 482                 for (stripe = 0; stripe < lsm->lsm_stripe_count; stripe++) {
 483                         if (stripe == req->rq_stripe)
 484                                 continue;
 485                         if (ost_idx == lsm->lsm_oinfo[stripe]->loi_ost_idx)
 486                                 break;
 487                 }
 488
 489                 if (stripe >= lsm->lsm_stripe_count) {
 490                         req->rq_idx = ost_idx;
 491                         rc = obd_create(lov->lov_tgts[ost_idx]->ltd_exp,
 492                                         req->rq_oi.oi_oa, &req->rq_oi.oi_md,
 493                                         set->set_oti);
 494                         if (!rc)
 495                                 break;
 496                 }
 497         }
 498         RETURN(rc);
 499 }
 500
 501 #define LOV_CREATE_RESEED_MULT 4
 502 #define LOV_CREATE_RESEED_MIN  1000
 503 /* Allocate objects on osts with round-robin algorithm */
 504 static int alloc_rr(struct lov_obd *lov, int *idx_arr, int *stripe_cnt)
 505 {
 506         unsigned array_idx, ost_count = lov->desc.ld_tgt_count;
 507         unsigned ost_active_count = lov->desc.ld_active_tgt_count;
 508         int i, *idx_pos = idx_arr;
 509         __u32 ost_idx;
 510         ENTRY;
 511
 512         i = qos_calc_rr(lov);
 513         if (i)
 514                 RETURN(i);
 515
 516         if (--lov->lov_start_count <= 0) {
 517                 lov->lov_start_idx = ll_rand() % ost_count;
 518                 lov->lov_start_count =
 519                         (LOV_CREATE_RESEED_MIN / max(ost_active_count, 1U) +
 520                          LOV_CREATE_RESEED_MULT) * max(ost_active_count, 1U);
 521         } else if (*stripe_cnt >= ost_active_count ||
 522                    lov->lov_start_idx > ost_count) {
 523                 /* If we have allocated from all of the OSTs, slowly
 524                    precess the next start */
 525                 lov->lov_start_idx %= ost_count;
 526                 ++lov->lov_offset_idx;
 527         }
 528         array_idx = (lov->lov_start_idx + lov->lov_offset_idx) % ost_count;
 529 #ifdef QOS_DEBUG
 530         CDEBUG(D_QOS, "want %d startidx %d startcnt %d offset %d arrayidx %d\n",
 531                *stripe_cnt, lov->lov_start_idx, lov->lov_start_count,
 532                lov->lov_offset_idx, array_idx);
 533 #endif
 534
 535         down_read(&lov->lov_qos.lq_rw_sem);
 536         for (i = 0; i < ost_count; i++, array_idx=(array_idx + 1) % ost_count) {
 537                 ++lov->lov_start_idx;
 538                 ost_idx = lov->lov_qos.lq_rr_array[array_idx];
 539 #ifdef QOS_DEBUG
 540                 CDEBUG(D_QOS, "#%d strt %d act %d strp %d ary %d idx %d\n",
 541                        i, lov->lov_start_idx,
 542                        ((ost_idx != LOV_QOS_EMPTY) && lov->lov_tgts[ost_idx]) ?
 543                        lov->lov_tgts[ost_idx]->ltd_active : 0,
 544                        idx_pos - idx_arr, array_idx, ost_idx);
 545 #endif
 546                 if ((ost_idx == LOV_QOS_EMPTY) || !lov->lov_tgts[ost_idx] ||
 547                     !lov->lov_tgts[ost_idx]->ltd_active)
 548                         continue;
 549                 *idx_pos = ost_idx;
 550                 idx_pos++;
 551                 /* We have enough stripes */
 552                 if (idx_pos - idx_arr == *stripe_cnt)
 553                         break;
 554         }
 555         up_read(&lov->lov_qos.lq_rw_sem);
 556
 557         *stripe_cnt = idx_pos - idx_arr;
 558         RETURN(0);
 559 }
 560
 561 /* alloc objects on osts with specific stripe offset */
 562 static int alloc_specific(struct lov_obd *lov, struct lov_stripe_md *lsm,
 563                           int *idx_arr)
 564 {
 565         unsigned ost_idx, ost_count = lov->desc.ld_tgt_count;
 566         int i, *idx_pos = idx_arr;
 567         ENTRY;
 568
 569         ost_idx = lsm->lsm_oinfo[0]->loi_ost_idx;
 570         for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
 571                 if (!lov->lov_tgts[ost_idx] ||
 572                     !lov->lov_tgts[ost_idx]->ltd_active) {
 573                         continue;
 574                 }
 575                 *idx_pos = ost_idx;
 576                 idx_pos++;
 577                 /* got enough ost */
 578                 if (idx_pos - idx_arr == lsm->lsm_stripe_count)
 579                         RETURN(0);
 580         }
 581         /* If we were passed specific striping params, then a failure to
 582          * meet those requirements is an error, since we can't reallocate
 583          * that memory (it might be part of a larger array or something).
 584          *
 585          * We can only get here if lsm_stripe_count was originally > 1.
 586          */
 587         CERROR("can't lstripe objid "LPX64": have "LPSZ" want %u\n",
 588                lsm->lsm_object_id, idx_pos - idx_arr, lsm->lsm_stripe_count);
 589         RETURN(-EFBIG);
 590 }
 591
 592 /* Alloc objects on osts with optimization based on:
 593    - free space
 594    - network resources (shared OSS's)
 595 */
 596 static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt)
 597 {
 598         struct lov_obd *lov = &exp->exp_obd->u.lov;
 599         static time_t last_warn = 0;
 600         time_t now = cfs_time_current_sec();
 601         __u64 total_bavail, total_weight = 0;
 602         __u32 ost_count;
 603         int nfound, good_osts, i, warn = 0, rc = 0;
 604         ENTRY;
 605
 606         lov_getref(exp->exp_obd);
 607
 608         /* Detect -EAGAIN early, before expensive lock is taken. */
 609         if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space)
 610                 GOTO(out, rc = -EAGAIN);
 611
 612         /* Do actuall allocation, use write lock here. */
 613         down_write(&lov->lov_qos.lq_rw_sem);
 614
 615         /*
 616          * Check again, while we were sleeping on @lq_rw_sem things could
 617          * change.
 618          */
 619         if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space) {
 620                 up_write(&lov->lov_qos.lq_rw_sem);
 621                 GOTO(out, rc = -EAGAIN);
 622         }
 623         ost_count = lov->desc.ld_tgt_count;
 624
 625         if (lov->desc.ld_active_tgt_count < 2)
 626                 GOTO(out_up_write, rc = -EAGAIN);
 627
 628         rc = qos_calc_ppo(exp->exp_obd);
 629         if (rc)
 630                 GOTO(out_up_write, rc);
 631
 632         total_bavail = 0;
 633         good_osts = 0;
 634         /* Warn users about zero available space/inode every 30 min */
 635         if (cfs_time_sub(now, last_warn) > 60 * 30)
 636                 warn = 1;
 637         /* Find all the OSTs that are valid stripe candidates */
 638         for (i = 0; i < ost_count; i++) {
 639                 __u64 bavail;
 640
 641                 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
 642                         continue;
 643                 bavail = TGT_BAVAIL(i);
 644                 if (!bavail) {
 645                         if (warn) {
 646                                 CDEBUG(D_QOS, "no free space on %s\n",
 647                                        obd_uuid2str(&lov->lov_tgts[i]->ltd_uuid));
 648                                 last_warn = now;
 649                         }
 650                         continue;
 651                 }
 652                 if (!TGT_FFREE(i)) {
 653                         if (warn) {
 654                                 CDEBUG(D_QOS, "no free inodes on %s\n",
 655                                        obd_uuid2str(&lov->lov_tgts[i]->ltd_uuid));
 656                                 last_warn = now;
 657                         }
 658                         continue;
 659                 }
 660
 661                 lov->lov_tgts[i]->ltd_qos.ltq_usable = 1;
 662                 qos_calc_weight(lov, i);
 663                 total_bavail += bavail;
 664                 total_weight += lov->lov_tgts[i]->ltd_qos.ltq_weight;
 665
 666                 good_osts++;
 667         }
 668
 669         if (!total_bavail)
 670                 GOTO(out_up_write, rc = -ENOSPC);
 671
 672         /* if we don't have enough good OSTs, we reduce the stripe count. */
 673         if (good_osts < *stripe_cnt)
 674                 *stripe_cnt = good_osts;
 675
 676         if (!*stripe_cnt)
 677                 GOTO(out_up_write, rc = -EAGAIN);
 678
 679         /* Find enough OSTs with weighted random allocation. */
 680         nfound = 0;
 681         while (nfound < *stripe_cnt) {
 682                 __u64 rand, cur_weight = 0;
 683
 684                 rc = -ENODEV;
 685
 686                 if (total_weight) {
 687 #if BITS_PER_LONG == 32
 688                         rand = ll_rand() % (unsigned)total_weight;
 689                         /* If total_weight > 32-bit, first generate the high
 690                          * 32 bits of the random number, then add in the low
 691                          * 32 bits (truncated to the upper limit, if needed) */
 692                         if (total_weight > 0xffffffffULL)
 693                                 rand = (__u64)(ll_rand() %
 694                                           (unsigned)(total_weight >> 32)) << 32;
 695                         else
 696                                 rand = 0;
 697
 698                         if (rand == (total_weight & 0xffffffff00000000ULL))
 699                                 rand |= ll_rand() % (unsigned)total_weight;
 700                         else
 701                                 rand |= ll_rand();
 702 #else
 703                         rand = ((__u64)ll_rand() << 32 | ll_rand()) %
 704                                 total_weight;
 705 #endif
 706                 } else {
 707                         rand = 0;
 708                 }
 709
 710                 /* On average, this will hit larger-weighted osts more often.
 711                    0-weight osts will always get used last (only when rand=0).*/
 712                 for (i = 0; i < ost_count; i++) {
 713                         if (!lov->lov_tgts[i] ||
 714                             !lov->lov_tgts[i]->ltd_qos.ltq_usable)
 715                                 continue;
 716                         cur_weight += lov->lov_tgts[i]->ltd_qos.ltq_weight;
 717                         if (cur_weight >= rand) {
 718 #ifdef QOS_DEBUG
 719                                 CDEBUG(D_QOS, "assigned stripe=%d to idx=%d\n",
 720                                        nfound, i);
 721 #endif
 722                                 idx_arr[nfound++] = i;
 723                                 qos_used(lov, i, &total_weight);
 724                                 rc = 0;
 725                                 break;
 726                         }
 727                 }
 728                 /* should never satisfy below condition */
 729                 if (rc) {
 730                         CERROR("Didn't find any OSTs?\n");
 731                         break;
 732                 }
 733         }
 734         LASSERT(nfound == *stripe_cnt);
 735
 736 out_up_write:
 737         up_write(&lov->lov_qos.lq_rw_sem);
 738
 739 out:
 740         if (rc == -EAGAIN)
 741                 rc = alloc_rr(lov, idx_arr, stripe_cnt);
 742
 743         lov_putref(exp->exp_obd);
 744         RETURN(rc);
 745 }
 746
 747 /* return new alloced stripe count on success */
 748 static int alloc_idx_array(struct obd_export *exp, struct lov_stripe_md *lsm,
 749                            int newea, int **idx_arr, int *arr_cnt)
 750 {
 751         struct lov_obd *lov = &exp->exp_obd->u.lov;
 752         int stripe_cnt = lsm->lsm_stripe_count;
 753         int i, rc = 0;
 754         int *tmp_arr = NULL;
 755         ENTRY;
 756
 757         *arr_cnt = stripe_cnt;
 758         OBD_ALLOC(tmp_arr, *arr_cnt * sizeof(int));
 759         if (tmp_arr == NULL)
 760                 RETURN(-ENOMEM);
 761         for (i = 0; i < *arr_cnt; i++)
 762                 tmp_arr[i] = -1;
 763
 764         if (newea ||
 765             lsm->lsm_oinfo[0]->loi_ost_idx >= lov->desc.ld_tgt_count)
 766                 rc = alloc_qos(exp, tmp_arr, &stripe_cnt);
 767         else
 768                 rc = alloc_specific(lov, lsm, tmp_arr);
 769
 770         if (rc)
 771                 GOTO(out_arr, rc);
 772
 773         *idx_arr = tmp_arr;
 774         RETURN(stripe_cnt);
 775 out_arr:
 776         OBD_FREE(tmp_arr, *arr_cnt * sizeof(int));
 777         *arr_cnt = 0;
 778         RETURN(rc);
 779 }
 780
 781 static void free_idx_array(int *idx_arr, int arr_cnt)
 782 {
 783         if (arr_cnt)
 784                 OBD_FREE(idx_arr, arr_cnt * sizeof(int));
 785 }
 786
 787 int qos_prep_create(struct obd_export *exp, struct lov_request_set *set)
 788 {
 789         struct lov_obd *lov = &exp->exp_obd->u.lov;
 790         struct lov_stripe_md *lsm;
 791         struct obdo *src_oa = set->set_oi->oi_oa;
 792         struct obd_trans_info *oti = set->set_oti;
 793         int i, stripes, rc = 0, newea = 0;
 794         int *idx_arr, idx_cnt = 0;
 795         ENTRY;
 796
 797         LASSERT(src_oa->o_valid & OBD_MD_FLID);
 798         LASSERT(src_oa->o_valid & OBD_MD_FLGROUP);
 799
 800         if (set->set_oi->oi_md == NULL) {
 801                 int stripe_cnt = lov_get_stripecnt(lov, 0);
 802
 803                 /* If the MDS file was truncated up to some size, stripe over
 804                  * enough OSTs to allow the file to be created at that size.
 805                  * This may mean we use more than the default # of stripes. */
 806                 if (src_oa->o_valid & OBD_MD_FLSIZE) {
 807                         obd_size min_bavail = LUSTRE_STRIPE_MAXBYTES;
 808
 809                         /* Find a small number of stripes we can use
 810                            (up to # of active osts). */
 811                         stripes = 1;
 812                         lov_getref(exp->exp_obd);
 813                         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 814                                 if (!lov->lov_tgts[i] ||
 815                                     !lov->lov_tgts[i]->ltd_active)
 816                                         continue;
 817                                 min_bavail = min(min_bavail, TGT_BAVAIL(i));
 818                                 if (min_bavail * stripes > src_oa->o_size)
 819                                         break;
 820                                 stripes++;
 821                         }
 822                         lov_putref(exp->exp_obd);
 823
 824                         if (stripes < stripe_cnt)
 825                                 stripes = stripe_cnt;
 826                 } else {
 827                         stripes = stripe_cnt;
 828                 }
 829
 830                 rc = lov_alloc_memmd(&set->set_oi->oi_md, stripes,
 831                                      lov->desc.ld_pattern ?
 832                                      lov->desc.ld_pattern : LOV_PATTERN_RAID0,
 833                                      LOV_MAGIC);
 834                 if (rc < 0)
 835                         GOTO(out_err, rc);
 836                 rc = 0;
 837                 newea = 1;
 838         }
 839
 840         lsm = set->set_oi->oi_md;
 841         lsm->lsm_object_id = src_oa->o_id;
 842         lsm->lsm_object_gr = src_oa->o_gr;
 843
 844         if (!lsm->lsm_stripe_size)
 845                 lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size;
 846         if (!lsm->lsm_pattern) {
 847                 LASSERT(lov->desc.ld_pattern);
 848                 lsm->lsm_pattern = lov->desc.ld_pattern;
 849         }
 850
 851         stripes = alloc_idx_array(exp, lsm, newea, &idx_arr, &idx_cnt);
 852         if (stripes <= 0)
 853                 GOTO(out_err, rc = stripes ? stripes : -EIO);
 854         LASSERTF(stripes <= lsm->lsm_stripe_count,"requested %d allocated %d\n",
 855                  lsm->lsm_stripe_count, stripes);
 856
 857         for (i = 0; i < stripes; i++) {
 858                 struct lov_request *req;
 859                 int ost_idx = idx_arr[i];
 860                 LASSERT(ost_idx >= 0);
 861
 862                 OBD_ALLOC(req, sizeof(*req));
 863                 if (req == NULL)
 864                         GOTO(out_err, rc = -ENOMEM);
 865                 lov_set_add_req(req, set);
 866
 867                 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
 868                 OBD_ALLOC(req->rq_oi.oi_md, req->rq_buflen);
 869                 if (req->rq_oi.oi_md == NULL)
 870                         GOTO(out_err, rc = -ENOMEM);
 871
 872                 OBDO_ALLOC(req->rq_oi.oi_oa);
 873                 if (req->rq_oi.oi_oa == NULL)
 874                         GOTO(out_err, rc = -ENOMEM);
 875
 876                 req->rq_idx = ost_idx;
 877                 req->rq_stripe = i;
 878                 /* create data objects with "parent" OA */
 879                 memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
 880
 881                 /* XXX When we start creating objects on demand, we need to
 882                  *     make sure that we always create the object on the
 883                  *     stripe which holds the existing file size.
 884                  */
 885                 if (src_oa->o_valid & OBD_MD_FLSIZE) {
 886                         req->rq_oi.oi_oa->o_size =
 887                                 lov_size_to_stripe(lsm, src_oa->o_size, i);
 888
 889                         CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
 890                                i, req->rq_oi.oi_oa->o_size, src_oa->o_size);
 891                 }
 892         }
 893         LASSERT(set->set_count == stripes);
 894
 895         if (stripes < lsm->lsm_stripe_count)
 896                 qos_shrink_lsm(set);
 897
 898         if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) {
 899                 oti_alloc_cookies(oti, set->set_count);
 900                 if (!oti->oti_logcookies)
 901                         GOTO(out_err, rc = -ENOMEM);
 902                 set->set_cookies = oti->oti_logcookies;
 903         }
 904 out_err:
 905         if (newea && rc)
 906                 obd_free_memmd(exp, &set->set_oi->oi_md);
 907         free_idx_array(idx_arr, idx_cnt);
 908         EXIT;
 909         return rc;
 910 }
 911
 912 void qos_update(struct lov_obd *lov)
 913 {
 914         ENTRY;
 915         lov->lov_qos.lq_dirty = 1;
 916 }
 917