lustre/lov/lov_qos.c

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