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, struct lov_tgt_desc *tgt)
 125 {
 126         struct lov_obd *lov = &obd->u.lov;
 127         struct lov_qos_oss *oss;
 128         int rc = 0;
 129         ENTRY;
 130
 131         down_write(&lov->lov_qos.lq_rw_sem);
 132
 133         oss = tgt->ltd_qos.ltq_oss;
 134         if (!oss)
 135                 GOTO(out, rc = -ENOENT);
 136
 137         oss->lqo_ost_count--;
 138         if (oss->lqo_ost_count == 0) {
 139                 CDEBUG(D_QOS, "removing OSS %s\n",
 140                        obd_uuid2str(&oss->lqo_uuid));
 141                 list_del(&oss->lqo_oss_list);
 142                 OBD_FREE_PTR(oss);
 143         }
 144
 145         lov->lov_qos.lq_dirty = 1;
 146         lov->lov_qos.lq_rr.lqr_dirty = 1;
 147 out:
 148         up_write(&lov->lov_qos.lq_rw_sem);
 149         RETURN(rc);
 150 }
 151
 152 /* Recalculate per-object penalties for OSSs and OSTs,
 153    depends on size of each ost in an oss */
 154 static int qos_calc_ppo(struct obd_device *obd)
 155 {
 156         struct lov_obd *lov = &obd->u.lov;
 157         struct lov_qos_oss *oss;
 158         __u64 ba_max, ba_min, temp;
 159         __u32 num_active;
 160         int rc, i, prio_wide;
 161         time_t now, age;
 162         ENTRY;
 163
 164         if (!lov->lov_qos.lq_dirty)
 165                 GOTO(out, rc = 0);
 166
 167         num_active = lov->desc.ld_active_tgt_count - 1;
 168         if (num_active < 1)
 169                 GOTO(out, rc = -EAGAIN);
 170
 171         /* find bavail on each OSS */
 172         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 173                 oss->lqo_bavail = 0;
 174         }
 175         lov->lov_qos.lq_active_oss_count = 0;
 176
 177         /* How badly user wants to select osts "widely" (not recently chosen
 178            and not on recent oss's).  As opposed to "freely" (free space
 179            avail.) 0-256. */
 180         prio_wide = 256 - lov->lov_qos.lq_prio_free;
 181
 182         ba_min = (__u64)(-1);
 183         ba_max = 0;
 184         now = cfs_time_current_sec();
 185         /* Calculate OST penalty per object */
 186         /* (lov ref taken in alloc_qos) */
 187         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 188                 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
 189                         continue;
 190                 temp = TGT_BAVAIL(i);
 191                 if (!temp)
 192                         continue;
 193                 ba_min = min(temp, ba_min);
 194                 ba_max = max(temp, ba_max);
 195
 196                 /* Count the number of usable OSS's */
 197                 if (lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_bavail == 0)
 198                         lov->lov_qos.lq_active_oss_count++;
 199                 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_bavail += temp;
 200
 201                 /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
 202                 temp >>= 1;
 203                 do_div(temp, num_active);
 204                 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj =
 205                         (temp * prio_wide) >> 8;
 206
 207                 age = (now - lov->lov_tgts[i]->ltd_qos.ltq_used) >> 3;
 208                 if (lov->lov_qos.lq_reset || age > 32 * lov->desc.ld_qos_maxage)
 209                         lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
 210                 else if (age > lov->desc.ld_qos_maxage)
 211                         /* Decay the penalty by half for every 8x the update
 212                          * interval that the device has been idle.  That gives
 213                          * lots of time for the statfs information to be
 214                          * updated (which the penalty is only a proxy for),
 215                          * and avoids penalizing OSS/OSTs under light load. */
 216                         lov->lov_tgts[i]->ltd_qos.ltq_penalty >>=
 217                                 (age / lov->desc.ld_qos_maxage);
 218         }
 219
 220         num_active = lov->lov_qos.lq_active_oss_count - 1;
 221         if (num_active < 1) {
 222                 /* If there's only 1 OSS, we can't penalize it, so instead
 223                    we have to double the OST penalty */
 224                 num_active = 1;
 225                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 226                         if (lov->lov_tgts[i] == NULL)
 227                                 continue;
 228                         lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj <<= 1;
 229                 }
 230         }
 231
 232         /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
 233         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 234                 temp = oss->lqo_bavail >> 1;
 235                 do_div(temp, oss->lqo_ost_count * num_active);
 236                 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
 237
 238                 age = (now - oss->lqo_used) >> 3;
 239                 if (lov->lov_qos.lq_reset || age > 32 * lov->desc.ld_qos_maxage)
 240                         oss->lqo_penalty = 0;
 241                 else if (age > lov->desc.ld_qos_maxage)
 242                         /* Decay the penalty by half for every 8x the update
 243                          * interval that the device has been idle.  That gives
 244                          * lots of time for the statfs information to be
 245                          * updated (which the penalty is only a proxy for),
 246                          * and avoids penalizing OSS/OSTs under light load. */
 247                         oss->lqo_penalty >>= (age / lov->desc.ld_qos_maxage);
 248         }
 249
 250         lov->lov_qos.lq_dirty = 0;
 251         lov->lov_qos.lq_reset = 0;
 252
 253         /* If each ost has almost same free space,
 254          * do rr allocation for better creation performance */
 255         lov->lov_qos.lq_same_space = 0;
 256         temp = ba_max - ba_min;
 257         ba_min = (ba_min * 51) >> 8;     /* 51/256 = .20 */
 258         if (temp < ba_min) {
 259                 /* Difference is less than 20% */
 260                 lov->lov_qos.lq_same_space = 1;
 261                 /* Reset weights for the next time we enter qos mode */
 262                 lov->lov_qos.lq_reset = 1;
 263         }
 264         rc = 0;
 265
 266 out:
 267         if (!rc && lov->lov_qos.lq_same_space)
 268                 RETURN(-EAGAIN);
 269         RETURN(rc);
 270 }
 271
 272 static int qos_calc_weight(struct lov_obd *lov, int i)
 273 {
 274         __u64 temp, temp2;
 275
 276         /* Final ost weight = TGT_BAVAIL - ost_penalty - oss_penalty */
 277         temp = TGT_BAVAIL(i);
 278         temp2 = lov->lov_tgts[i]->ltd_qos.ltq_penalty +
 279                 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty;
 280         if (temp < temp2)
 281                 lov->lov_tgts[i]->ltd_qos.ltq_weight = 0;
 282         else
 283                 lov->lov_tgts[i]->ltd_qos.ltq_weight = temp - temp2;
 284         return 0;
 285 }
 286
 287 /* We just used this index for a stripe; adjust everyone's weights */
 288 static int qos_used(struct lov_obd *lov, struct ost_pool *osts,
 289                     __u32 index, __u64 *total_wt)
 290 {
 291         struct lov_qos_oss *oss;
 292         int j;
 293         ENTRY;
 294
 295         /* Don't allocate from this stripe anymore, until the next alloc_qos */
 296         lov->lov_tgts[index]->ltd_qos.ltq_usable = 0;
 297
 298         oss = lov->lov_tgts[index]->ltd_qos.ltq_oss;
 299
 300         /* Decay old penalty by half (we're adding max penalty, and don't
 301            want it to run away.) */
 302         lov->lov_tgts[index]->ltd_qos.ltq_penalty >>= 1;
 303         oss->lqo_penalty >>= 1;
 304
 305         /* mark the OSS and OST as recently used */
 306         lov->lov_tgts[index]->ltd_qos.ltq_used =
 307                 oss->lqo_used = cfs_time_current_sec();
 308
 309         /* Set max penalties for this OST and OSS */
 310         lov->lov_tgts[index]->ltd_qos.ltq_penalty +=
 311                 lov->lov_tgts[index]->ltd_qos.ltq_penalty_per_obj *
 312                 lov->desc.ld_active_tgt_count;
 313         oss->lqo_penalty += oss->lqo_penalty_per_obj *
 314                 lov->lov_qos.lq_active_oss_count;
 315
 316         /* Decrease all OSS penalties */
 317         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 318                 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
 319                         oss->lqo_penalty = 0;
 320                 else
 321                         oss->lqo_penalty -= oss->lqo_penalty_per_obj;
 322         }
 323
 324         *total_wt = 0;
 325         /* Decrease all OST penalties */
 326         for (j = 0; j < osts->op_count; j++) {
 327                 int i;
 328
 329                 i = osts->op_array[j];
 330                 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
 331                         continue;
 332                 if (lov->lov_tgts[i]->ltd_qos.ltq_penalty <
 333                     lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj)
 334                         lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
 335                 else
 336                         lov->lov_tgts[i]->ltd_qos.ltq_penalty -=
 337                         lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj;
 338
 339                 qos_calc_weight(lov, i);
 340
 341                 /* Recalc the total weight of usable osts */
 342                 if (lov->lov_tgts[i]->ltd_qos.ltq_usable)
 343                         *total_wt += lov->lov_tgts[i]->ltd_qos.ltq_weight;
 344
 345 #ifdef QOS_DEBUG
 346                 CDEBUG(D_QOS, "recalc tgt %d usable=%d avail="LPU64
 347                        " ostppo="LPU64" ostp="LPU64" ossppo="LPU64
 348                        " ossp="LPU64" wt="LPU64"\n",
 349                        i, lov->lov_tgts[i]->ltd_qos.ltq_usable,
 350                        TGT_BAVAIL(i) >> 10,
 351                        lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj >> 10,
 352                        lov->lov_tgts[i]->ltd_qos.ltq_penalty >> 10,
 353                        lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty_per_obj>>10,
 354                        lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty >> 10,
 355                        lov->lov_tgts[i]->ltd_qos.ltq_weight >> 10);
 356 #endif
 357         }
 358
 359         RETURN(0);
 360 }
 361
 362 #define LOV_QOS_EMPTY ((__u32)-1)
 363 /* compute optimal round-robin order, based on OSTs per OSS */
 364 static int qos_calc_rr(struct lov_obd *lov, struct ost_pool *src_pool,
 365                        struct lov_qos_rr *lqr)
 366 {
 367         struct lov_qos_oss *oss;
 368         unsigned placed, real_count;
 369         int i, rc;
 370         ENTRY;
 371
 372         if (!lqr->lqr_dirty) {
 373                 LASSERT(lqr->lqr_pool.op_size);
 374                 RETURN(0);
 375         }
 376
 377         /* Do actual allocation. */
 378         down_write(&lov->lov_qos.lq_rw_sem);
 379
 380         /*
 381          * Check again. While we were sleeping on @lq_rw_sem something could
 382          * change.
 383          */
 384         if (!lqr->lqr_dirty) {
 385                 LASSERT(lqr->lqr_pool.op_size);
 386                 up_write(&lov->lov_qos.lq_rw_sem);
 387                 RETURN(0);
 388         }
 389
 390         real_count = src_pool->op_count;
 391
 392         /* Zero the pool array */
 393         /* alloc_rr is holding a read lock on the pool, so nobody is adding/
 394            deleting from the pool. The lq_rw_sem insures that nobody else
 395            is reading. */
 396         lqr->lqr_pool.op_count = real_count;
 397         rc = lov_ost_pool_extend(&lqr->lqr_pool, real_count);
 398         if (rc) {
 399                 up_write(&lov->lov_qos.lq_rw_sem);
 400                 RETURN(rc);
 401         }
 402         for (i = 0; i < lqr->lqr_pool.op_count; i++)
 403                 lqr->lqr_pool.op_array[i] = LOV_QOS_EMPTY;
 404
 405         /* Place all the OSTs from 1 OSS at the same time. */
 406         placed = 0;
 407         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 408                 int j = 0;
 409                 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
 410                         if (lov->lov_tgts[src_pool->op_array[i]] &&
 411                             (lov->lov_tgts[src_pool->op_array[i]]->ltd_qos.ltq_oss == oss)) {
 412                               /* Evenly space these OSTs across arrayspace */
 413                               int next = j * lqr->lqr_pool.op_count / oss->lqo_ost_count;
 414                               while (lqr->lqr_pool.op_array[next] !=
 415                                      LOV_QOS_EMPTY)
 416                                         next = (next + 1) % lqr->lqr_pool.op_count;
 417                               lqr->lqr_pool.op_array[next] = src_pool->op_array[i];
 418                               j++;
 419                               placed++;
 420                         }
 421                 }
 422         }
 423
 424         lqr->lqr_dirty = 0;
 425         up_write(&lov->lov_qos.lq_rw_sem);
 426
 427         if (placed != real_count) {
 428                 /* This should never happen */
 429                 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
 430                                    "round-robin list (%d of %d).\n",
 431                                    placed, real_count);
 432                 for (i = 0; i < lqr->lqr_pool.op_count; i++) {
 433                         LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
 434                                  lqr->lqr_pool.op_array[i]);
 435                 }
 436                 lqr->lqr_dirty = 1;
 437                 RETURN(-EAGAIN);
 438         }
 439
 440 #ifdef QOS_DEBUG
 441         for (i = 0; i < lqr->lqr_pool.op_count; i++) {
 442                 LCONSOLE(D_QOS, "rr #%d ost idx=%d\n", i,
 443                          lqr->lqr_pool.op_array[i]);
 444         }
 445 #endif
 446
 447         RETURN(0);
 448 }
 449
 450
 451 void qos_shrink_lsm(struct lov_request_set *set)
 452 {
 453         struct lov_stripe_md *lsm = set->set_oi->oi_md, *lsm_new;
 454         /* XXX LOV STACKING call into osc for sizes */
 455         unsigned oldsize, newsize;
 456
 457         if (set->set_oti && set->set_cookies && set->set_cookie_sent) {
 458                 struct llog_cookie *cookies;
 459                 oldsize = lsm->lsm_stripe_count * sizeof(*cookies);
 460                 newsize = set->set_count * sizeof(*cookies);
 461
 462                 cookies = set->set_cookies;
 463                 oti_alloc_cookies(set->set_oti, set->set_count);
 464                 if (set->set_oti->oti_logcookies) {
 465                         memcpy(set->set_oti->oti_logcookies, cookies, newsize);
 466                         OBD_FREE(cookies, oldsize);
 467                         set->set_cookies = set->set_oti->oti_logcookies;
 468                 } else {
 469                         CWARN("'leaking' %d bytes\n", oldsize - newsize);
 470                 }
 471         }
 472
 473         CWARN("using fewer stripes for object "LPU64": old %u new %u\n",
 474               lsm->lsm_object_id, lsm->lsm_stripe_count, set->set_count);
 475         LASSERT(lsm->lsm_stripe_count >= set->set_count);
 476
 477         newsize = lov_stripe_md_size(set->set_count);
 478         OBD_ALLOC(lsm_new, newsize);
 479         if (lsm_new != NULL) {
 480                 int i;
 481                 memcpy(lsm_new, lsm, sizeof(*lsm));
 482                 for (i = 0; i < lsm->lsm_stripe_count; i++) {
 483                         if (i < set->set_count) {
 484                                 lsm_new->lsm_oinfo[i] = lsm->lsm_oinfo[i];
 485                                 continue;
 486                         }
 487                         OBD_SLAB_FREE(lsm->lsm_oinfo[i], lov_oinfo_slab,
 488                                       sizeof(struct lov_oinfo));
 489                 }
 490                 lsm_new->lsm_stripe_count = set->set_count;
 491                 OBD_FREE(lsm, sizeof(struct lov_stripe_md) +
 492                          lsm->lsm_stripe_count * sizeof(struct lov_oinfo *));
 493                 set->set_oi->oi_md = lsm_new;
 494         } else {
 495                 CWARN("'leaking' few bytes\n");
 496         }
 497 }
 498
 499 int qos_remedy_create(struct lov_request_set *set, struct lov_request *req)
 500 {
 501         struct lov_stripe_md *lsm = set->set_oi->oi_md;
 502         struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
 503         unsigned ost_idx, ost_count = lov->desc.ld_tgt_count;
 504         int stripe, i, rc = -EIO;
 505         ENTRY;
 506
 507         ost_idx = (req->rq_idx + lsm->lsm_stripe_count) % ost_count;
 508         for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
 509                 if (!lov->lov_tgts[ost_idx] ||
 510                     !lov->lov_tgts[ost_idx]->ltd_active)
 511                         continue;
 512                 /* check if objects has been created on this ost */
 513                 for (stripe = 0; stripe < lsm->lsm_stripe_count; stripe++) {
 514                         if (stripe == req->rq_stripe)
 515                                 continue;
 516                         if (ost_idx == lsm->lsm_oinfo[stripe]->loi_ost_idx)
 517                                 break;
 518                 }
 519
 520                 if (stripe >= lsm->lsm_stripe_count) {
 521                         req->rq_idx = ost_idx;
 522                         rc = obd_create(lov->lov_tgts[ost_idx]->ltd_exp,
 523                                         req->rq_oi.oi_oa, &req->rq_oi.oi_md,
 524                                         set->set_oti);
 525                         if (!rc)
 526                                 break;
 527                 }
 528         }
 529         RETURN(rc);
 530 }
 531
 532 static int min_stripe_count(int stripe_cnt, int flags)
 533 {
 534         return (flags & LOV_USES_DEFAULT_STRIPE ?
 535                 stripe_cnt - (stripe_cnt / 4) : stripe_cnt);
 536 }
 537
 538 #define LOV_CREATE_RESEED_MULT 4
 539 #define LOV_CREATE_RESEED_MIN  1000
 540 /* Allocate objects on osts with round-robin algorithm */
 541 static int alloc_rr(struct lov_obd *lov, int *idx_arr, int *stripe_cnt,
 542                     char *poolname, int flags)
 543 {
 544         unsigned array_idx;
 545         int i, rc, *idx_pos;
 546         __u32 ost_idx;
 547         int ost_start_idx_temp;
 548         int speed = 0;
 549         int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
 550         struct pool_desc *pool;
 551         struct ost_pool *osts;
 552         struct lov_qos_rr *lqr;
 553         ENTRY;
 554
 555         pool = lov_find_pool(lov, poolname);
 556         if (pool == NULL) {
 557                 osts = &(lov->lov_packed);
 558                 lqr = &(lov->lov_qos.lq_rr);
 559         } else {
 560                 down_read(&pool_tgt_rw_sem(pool));
 561                 osts = &(pool->pool_obds);
 562                 lqr = &(pool->pool_rr);
 563         }
 564
 565         rc = qos_calc_rr(lov, osts, lqr);
 566         if (rc)
 567                 GOTO(out, rc);
 568
 569         if (--lqr->lqr_start_count <= 0) {
 570                 lqr->lqr_start_idx = ll_rand() % osts->op_count;
 571                 lqr->lqr_start_count =
 572                         (LOV_CREATE_RESEED_MIN / max(osts->op_count, 1U) +
 573                          LOV_CREATE_RESEED_MULT) * max(osts->op_count, 1U);
 574         } else if (stripe_cnt_min >= osts->op_count ||
 575                    lqr->lqr_start_idx > osts->op_count) {
 576                 /* If we have allocated from all of the OSTs, slowly
 577                  * precess the next start if the OST/stripe count isn't
 578                  * already doing this for us. */
 579                 lqr->lqr_start_idx %= osts->op_count;
 580                 if (*stripe_cnt > 1 && (osts->op_count % (*stripe_cnt)) != 1)
 581                         ++lqr->lqr_offset_idx;
 582         }
 583         down_read(&lov->lov_qos.lq_rw_sem);
 584         ost_start_idx_temp = lqr->lqr_start_idx;
 585
 586 repeat_find:
 587         array_idx = (lqr->lqr_start_idx + lqr->lqr_offset_idx) % osts->op_count;
 588         idx_pos = idx_arr;
 589 #ifdef QOS_DEBUG
 590         CDEBUG(D_QOS, "pool '%s' want %d startidx %d startcnt %d offset %d "
 591                "active %d count %d arrayidx %d\n", poolname,
 592                *stripe_cnt, lqr->lqr_start_idx, lqr->lqr_start_count,
 593                lqr->lqr_offset_idx, osts->op_count, osts->op_count, array_idx);
 594 #endif
 595
 596         for (i = 0; i < osts->op_count;
 597                     i++, array_idx=(array_idx + 1) % osts->op_count) {
 598                 ++lqr->lqr_start_idx;
 599                 ost_idx = lqr->lqr_pool.op_array[array_idx];
 600 #ifdef QOS_DEBUG
 601                 CDEBUG(D_QOS, "#%d strt %d act %d strp %d ary %d idx %d\n",
 602                        i, lqr->lqr_start_idx,
 603                        ((ost_idx != LOV_QOS_EMPTY) && lov->lov_tgts[ost_idx]) ?
 604                        lov->lov_tgts[ost_idx]->ltd_active : 0,
 605                        idx_pos - idx_arr, array_idx, ost_idx);
 606 #endif
 607                 if ((ost_idx == LOV_QOS_EMPTY) || !lov->lov_tgts[ost_idx] ||
 608                     !lov->lov_tgts[ost_idx]->ltd_active)
 609                         continue;
 610
 611                 /* Fail Check before osc_precreate() is called
 612                    so we can only 'fail' single OSC. */
 613                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
 614                         continue;
 615
 616                 /* Drop slow OSCs if we can */
 617                 if (obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed)
 618                         continue;
 619
 620                 *idx_pos = ost_idx;
 621                 idx_pos++;
 622                 /* We have enough stripes */
 623                 if (idx_pos - idx_arr == *stripe_cnt)
 624                         break;
 625         }
 626         if ((speed < 2) && (idx_pos - idx_arr < stripe_cnt_min)) {
 627                 /* Try again, allowing slower OSCs */
 628                 speed++;
 629                 lqr->lqr_start_idx = ost_start_idx_temp;
 630                 goto repeat_find;
 631         }
 632
 633         up_read(&lov->lov_qos.lq_rw_sem);
 634
 635         *stripe_cnt = idx_pos - idx_arr;
 636 out:
 637         if (pool != NULL) {
 638                 up_read(&pool_tgt_rw_sem(pool));
 639                 /* put back ref got by lov_find_pool() */
 640                 lh_put(lov->lov_pools_hash_body, &pool->pool_hash);
 641         }
 642
 643         RETURN(rc);
 644 }
 645
 646 /* alloc objects on osts with specific stripe offset */
 647 static int alloc_specific(struct lov_obd *lov, struct lov_stripe_md *lsm,
 648                           int *idx_arr)
 649 {
 650         unsigned ost_idx, array_idx, ost_count;
 651         int i, rc, *idx_pos;
 652         int speed = 0;
 653         struct pool_desc *pool;
 654         struct ost_pool *osts;
 655         ENTRY;
 656
 657         pool = lov_find_pool(lov, lsm->lsm_pool_name);
 658         if (pool == NULL) {
 659                 osts = &(lov->lov_packed);
 660         } else {
 661                 down_read(&pool_tgt_rw_sem(pool));
 662                 osts = &(pool->pool_obds);
 663         }
 664
 665         ost_count = osts->op_count;
 666
 667 repeat_find:
 668         /* search loi_ost_idx in ost array */
 669         array_idx = 0;
 670         for (i = 0; i < ost_count; i++) {
 671                 if (osts->op_array[i] == lsm->lsm_oinfo[0]->loi_ost_idx) {
 672                         array_idx = i;
 673                         break;
 674                 }
 675         }
 676         if (i == ost_count) {
 677                 CERROR("Start index %d not found in pool '%s'\n",
 678                        lsm->lsm_oinfo[0]->loi_ost_idx, lsm->lsm_pool_name);
 679                 GOTO(out, rc = -EINVAL);
 680         }
 681
 682         idx_pos = idx_arr;
 683         for (i = 0; i < ost_count;
 684              i++, array_idx = (array_idx + 1) % ost_count) {
 685                 ost_idx = osts->op_array[array_idx];
 686
 687                 if (!lov->lov_tgts[ost_idx] ||
 688                     !lov->lov_tgts[ost_idx]->ltd_active) {
 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) && ost_idx == 0)
 695                         continue;
 696
 697                 /* Drop slow OSCs if we can, but not for requested start idx.
 698                  *
 699                  * This means "if OSC is slow and it is not the requested
 700                  * start OST, then it can be skipped, otherwise skip it only
 701                  * if it is inactive/recovering/out-of-space." */
 702                 if ((obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp) > speed) &&
 703                     (i != 0 || speed >= 2))
 704                         continue;
 705
 706                 *idx_pos = ost_idx;
 707                 idx_pos++;
 708                 /* We have enough stripes */
 709                 if (idx_pos - idx_arr == lsm->lsm_stripe_count)
 710                         GOTO(out, rc = 0);
 711         }
 712         if (speed < 2) {
 713                 /* Try again, allowing slower OSCs */
 714                 speed++;
 715                 goto repeat_find;
 716         }
 717
 718         /* If we were passed specific striping params, then a failure to
 719          * meet those requirements is an error, since we can't reallocate
 720          * that memory (it might be part of a larger array or something).
 721          *
 722          * We can only get here if lsm_stripe_count was originally > 1.
 723          */
 724         CERROR("can't lstripe objid "LPX64": have %d want %u\n",
 725                lsm->lsm_object_id, (int)(idx_pos - idx_arr),
 726                lsm->lsm_stripe_count);
 727         rc = -EFBIG;
 728 out:
 729         if (pool != NULL) {
 730                 up_read(&pool_tgt_rw_sem(pool));
 731                 /* put back ref got by lov_find_pool() */
 732                 lh_put(lov->lov_pools_hash_body, &pool->pool_hash);
 733         }
 734
 735         RETURN(rc);
 736 }
 737
 738 /* Alloc objects on osts with optimization based on:
 739    - free space
 740    - network resources (shared OSS's)
 741 */
 742 static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt,
 743                      char *poolname, int flags)
 744 {
 745         struct lov_obd *lov = &exp->exp_obd->u.lov;
 746         static time_t last_warn = 0;
 747         time_t now = cfs_time_current_sec();
 748         __u64 total_bavail, total_weight = 0;
 749         int nfound, good_osts, i, warn = 0, rc = 0;
 750         int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
 751         struct pool_desc *pool;
 752         struct ost_pool *osts;
 753         struct lov_qos_rr *lqr;
 754         ENTRY;
 755
 756         if (stripe_cnt_min < 1)
 757                 RETURN(-EINVAL);
 758
 759         pool = lov_find_pool(lov, poolname);
 760         if (pool == NULL) {
 761                 osts = &(lov->lov_packed);
 762                 lqr = &(lov->lov_qos.lq_rr);
 763         } else {
 764                 down_read(&pool_tgt_rw_sem(pool));
 765                 osts = &(pool->pool_obds);
 766                 lqr = &(pool->pool_rr);
 767         }
 768
 769         lov_getref(exp->exp_obd);
 770
 771         /* Detect -EAGAIN early, before expensive lock is taken. */
 772         if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space)
 773                 GOTO(out_nolock, rc = -EAGAIN);
 774
 775         /* Do actual allocation, use write lock here. */
 776         down_write(&lov->lov_qos.lq_rw_sem);
 777
 778         /*
 779          * Check again, while we were sleeping on @lq_rw_sem things could
 780          * change.
 781          */
 782         if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space)
 783                 GOTO(out, rc = -EAGAIN);
 784
 785         if (lov->desc.ld_active_tgt_count < 2)
 786                 GOTO(out, rc = -EAGAIN);
 787
 788         rc = qos_calc_ppo(exp->exp_obd);
 789         if (rc)
 790                 GOTO(out, rc);
 791
 792         total_bavail = 0;
 793         good_osts = 0;
 794         /* Warn users about zero available space/inode every 30 min */
 795         if (cfs_time_sub(now, last_warn) > 60 * 30)
 796                 warn = 1;
 797         /* Find all the OSTs that are valid stripe candidates */
 798         for (i = 0; i < osts->op_count; i++) {
 799                 __u64 bavail;
 800
 801                 if (!lov->lov_tgts[osts->op_array[i]] ||
 802                     !lov->lov_tgts[osts->op_array[i]]->ltd_active)
 803                         continue;
 804                 bavail = TGT_BAVAIL(osts->op_array[i]);
 805                 if (!bavail) {
 806                         if (warn) {
 807                                 CDEBUG(D_QOS, "no free space on %s\n",
 808                                      obd_uuid2str(&lov->lov_tgts[osts->op_array[i]]->ltd_uuid));
 809                                 last_warn = now;
 810                         }
 811                         continue;
 812                 }
 813                 if (!TGT_FFREE(osts->op_array[i])) {
 814                         if (warn) {
 815                                 CDEBUG(D_QOS, "no free inodes on %s\n",
 816                                      obd_uuid2str(&lov->lov_tgts[osts->op_array[i]]->ltd_uuid));
 817                                 last_warn = now;
 818                         }
 819                         continue;
 820                 }
 821
 822                 /* Fail Check before osc_precreate() is called
 823                    so we can only 'fail' single OSC. */
 824                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && osts->op_array[i] == 0)
 825                         continue;
 826
 827                 if (obd_precreate(lov->lov_tgts[osts->op_array[i]]->ltd_exp) > 2)
 828                         continue;
 829
 830                 lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable = 1;
 831                 qos_calc_weight(lov, osts->op_array[i]);
 832                 total_bavail += bavail;
 833                 total_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight;
 834
 835                 good_osts++;
 836         }
 837
 838 #ifdef QOS_DEBUG
 839         CDEBUG(D_QOS, "found %d good osts\n", good_osts);
 840 #endif
 841
 842         if (good_osts < stripe_cnt_min)
 843                 GOTO(out, rc = -EAGAIN);
 844
 845         if (!total_bavail)
 846                 GOTO(out, rc = -ENOSPC);
 847
 848         /* We have enough osts */
 849         if (good_osts < *stripe_cnt)
 850                 *stripe_cnt = good_osts;
 851
 852         if (!*stripe_cnt)
 853                 GOTO(out, rc = -EAGAIN);
 854
 855         /* Find enough OSTs with weighted random allocation. */
 856         nfound = 0;
 857         while (nfound < *stripe_cnt) {
 858                 __u64 rand, cur_weight;
 859
 860                 cur_weight = 0;
 861                 rc = -ENODEV;
 862
 863                 if (total_weight) {
 864 #if BITS_PER_LONG == 32
 865                         rand = ll_rand() % (unsigned)total_weight;
 866                         /* If total_weight > 32-bit, first generate the high
 867                          * 32 bits of the random number, then add in the low
 868                          * 32 bits (truncated to the upper limit, if needed) */
 869                         if (total_weight > 0xffffffffULL)
 870                                 rand = (__u64)(ll_rand() %
 871                                           (unsigned)(total_weight >> 32)) << 32;
 872                         else
 873                                 rand = 0;
 874
 875                         if (rand == (total_weight & 0xffffffff00000000ULL))
 876                                 rand |= ll_rand() % (unsigned)total_weight;
 877                         else
 878                                 rand |= ll_rand();
 879
 880 #else
 881                         rand = ((__u64)ll_rand() << 32 | ll_rand()) %
 882                                 total_weight;
 883 #endif
 884                 } else {
 885                         rand = 0;
 886                 }
 887
 888                 /* On average, this will hit larger-weighted osts more often.
 889                    0-weight osts will always get used last (only when rand=0).*/
 890                 for (i = 0; i < osts->op_count; i++) {
 891                         if (!lov->lov_tgts[osts->op_array[i]] ||
 892                             !lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_usable)
 893                                 continue;
 894
 895                         cur_weight += lov->lov_tgts[osts->op_array[i]]->ltd_qos.ltq_weight;
 896 #ifdef QOS_DEBUG
 897                         CDEBUG(D_QOS, "stripe_cnt=%d nfound=%d cur_weight="LPU64
 898                                       " rand="LPU64" total_weight="LPU64"\n",
 899                                *stripe_cnt, nfound, cur_weight, rand, total_weight);
 900 #endif
 901                         if (cur_weight >= rand) {
 902 #ifdef QOS_DEBUG
 903                                 CDEBUG(D_QOS, "assigned stripe=%d to idx=%d\n",
 904                                        nfound, osts->op_array[i]);
 905 #endif
 906                                 idx_arr[nfound++] = osts->op_array[i];
 907                                 qos_used(lov, osts, osts->op_array[i], &total_weight);
 908                                 rc = 0;
 909                                 break;
 910                         }
 911                 }
 912                 /* should never satisfy below condition */
 913                 if (rc) {
 914                         CERROR("Didn't find any OSTs?\n");
 915                         break;
 916                 }
 917         }
 918         LASSERT(nfound == *stripe_cnt);
 919
 920 out:
 921         up_write(&lov->lov_qos.lq_rw_sem);
 922
 923 out_nolock:
 924         if (pool != NULL) {
 925                 up_read(&pool_tgt_rw_sem(pool));
 926                 /* put back ref got by lov_find_pool() */
 927                 lh_put(lov->lov_pools_hash_body, &pool->pool_hash);
 928         }
 929
 930         if (rc == -EAGAIN)
 931                 rc = alloc_rr(lov, idx_arr, stripe_cnt, poolname, flags);
 932
 933         lov_putref(exp->exp_obd);
 934         RETURN(rc);
 935 }
 936
 937 /* return new alloced stripe count on success */
 938 static int alloc_idx_array(struct obd_export *exp, struct lov_stripe_md *lsm,
 939                            int newea, int **idx_arr, int *arr_cnt, int flags)
 940 {
 941         struct lov_obd *lov = &exp->exp_obd->u.lov;
 942         int stripe_cnt = lsm->lsm_stripe_count;
 943         int i, rc = 0;
 944         int *tmp_arr = NULL;
 945         ENTRY;
 946
 947         *arr_cnt = stripe_cnt;
 948         OBD_ALLOC(tmp_arr, *arr_cnt * sizeof(int));
 949         if (tmp_arr == NULL)
 950                 RETURN(-ENOMEM);
 951         for (i = 0; i < *arr_cnt; i++)
 952                 tmp_arr[i] = -1;
 953
 954         if (newea ||
 955             lsm->lsm_oinfo[0]->loi_ost_idx >= lov->desc.ld_tgt_count)
 956                 rc = alloc_qos(exp, tmp_arr, &stripe_cnt,
 957                                lsm->lsm_pool_name, flags);
 958         else
 959                 rc = alloc_specific(lov, lsm, tmp_arr);
 960
 961         if (rc)
 962                 GOTO(out_arr, rc);
 963
 964         *idx_arr = tmp_arr;
 965         RETURN(stripe_cnt);
 966 out_arr:
 967         OBD_FREE(tmp_arr, *arr_cnt * sizeof(int));
 968         *arr_cnt = 0;
 969         RETURN(rc);
 970 }
 971
 972 static void free_idx_array(int *idx_arr, int arr_cnt)
 973 {
 974         if (arr_cnt)
 975                 OBD_FREE(idx_arr, arr_cnt * sizeof(int));
 976 }
 977
 978 int qos_prep_create(struct obd_export *exp, struct lov_request_set *set)
 979 {
 980         struct lov_obd *lov = &exp->exp_obd->u.lov;
 981         struct lov_stripe_md *lsm;
 982         struct obdo *src_oa = set->set_oi->oi_oa;
 983         struct obd_trans_info *oti = set->set_oti;
 984         int i, stripes, rc = 0, newea = 0;
 985         int flag = LOV_USES_ASSIGNED_STRIPE;
 986         int *idx_arr = NULL, idx_cnt = 0;
 987         ENTRY;
 988
 989         LASSERT(src_oa->o_valid & OBD_MD_FLID);
 990         LASSERT(src_oa->o_valid & OBD_MD_FLGROUP);
 991
 992         if (set->set_oi->oi_md == NULL) {
 993                 int stripes_def = lov_get_stripecnt(lov, 0);
 994
 995                 /* If the MDS file was truncated up to some size, stripe over
 996                  * enough OSTs to allow the file to be created at that size.
 997                  * This may mean we use more than the default # of stripes. */
 998                 if (src_oa->o_valid & OBD_MD_FLSIZE) {
 999                         obd_size min_bavail = LUSTRE_STRIPE_MAXBYTES;
1000
1001                         /* Find a small number of stripes we can use
1002                            (up to # of active osts). */
1003                         stripes = 1;
1004                         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1005                                 if (!lov->lov_tgts[i] ||
1006                                     !lov->lov_tgts[i]->ltd_active)
1007                                         continue;
1008                                 min_bavail = min(min_bavail, TGT_BAVAIL(i));
1009                                 if (min_bavail * stripes > src_oa->o_size)
1010                                         break;
1011                                 stripes++;
1012                         }
1013
1014                         if (stripes < stripes_def)
1015                                 stripes = stripes_def;
1016                 } else {
1017                          flag = LOV_USES_DEFAULT_STRIPE;
1018                          stripes = stripes_def;
1019                 }
1020
1021                 rc = lov_alloc_memmd(&set->set_oi->oi_md, stripes,
1022                                      lov->desc.ld_pattern ?
1023                                      lov->desc.ld_pattern : LOV_PATTERN_RAID0,
1024                                      LOV_MAGIC);
1025                 if (rc < 0)
1026                         GOTO(out_err, rc);
1027                 newea = 1;
1028                 rc = 0;
1029         }
1030
1031         lsm = set->set_oi->oi_md;
1032         lsm->lsm_object_id = src_oa->o_id;
1033         lsm->lsm_object_gr = src_oa->o_gr;
1034
1035         if (!lsm->lsm_stripe_size)
1036                 lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size;
1037         if (!lsm->lsm_pattern) {
1038                 LASSERT(lov->desc.ld_pattern);
1039                 lsm->lsm_pattern = lov->desc.ld_pattern;
1040         }
1041
1042         stripes = alloc_idx_array(exp, lsm, newea, &idx_arr, &idx_cnt, flag);
1043         if (stripes <= 0)
1044                 GOTO(out_err, rc = stripes ? stripes : -EIO);
1045         LASSERTF(stripes <= lsm->lsm_stripe_count,"requested %d allocated %d\n",
1046                  lsm->lsm_stripe_count, stripes);
1047
1048         for (i = 0; i < stripes; i++) {
1049                 struct lov_request *req;
1050                 int ost_idx = idx_arr[i];
1051                 LASSERT(ost_idx >= 0);
1052
1053                 OBD_ALLOC(req, sizeof(*req));
1054                 if (req == NULL)
1055                         GOTO(out_err, rc = -ENOMEM);
1056                 lov_set_add_req(req, set);
1057
1058                 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
1059                 OBD_ALLOC(req->rq_oi.oi_md, req->rq_buflen);
1060                 if (req->rq_oi.oi_md == NULL)
1061                         GOTO(out_err, rc = -ENOMEM);
1062
1063                 OBDO_ALLOC(req->rq_oi.oi_oa);
1064                 if (req->rq_oi.oi_oa == NULL)
1065                         GOTO(out_err, rc = -ENOMEM);
1066
1067                 req->rq_idx = ost_idx;
1068                 req->rq_stripe = i;
1069                 /* create data objects with "parent" OA */
1070                 memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
1071
1072                 /* XXX When we start creating objects on demand, we need to
1073                  *     make sure that we always create the object on the
1074                  *     stripe which holds the existing file size.
1075                  */
1076                 if (src_oa->o_valid & OBD_MD_FLSIZE) {
1077                         req->rq_oi.oi_oa->o_size =
1078                                 lov_size_to_stripe(lsm, src_oa->o_size, i);
1079
1080                         CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
1081                                i, req->rq_oi.oi_oa->o_size, src_oa->o_size);
1082                 }
1083         }
1084         LASSERT(set->set_count == stripes);
1085
1086         if (stripes < lsm->lsm_stripe_count)
1087                 qos_shrink_lsm(set);
1088
1089         if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) {
1090                 oti_alloc_cookies(oti, set->set_count);
1091                 if (!oti->oti_logcookies)
1092                         GOTO(out_err, rc = -ENOMEM);
1093                 set->set_cookies = oti->oti_logcookies;
1094         }
1095 out_err:
1096         if (newea && rc)
1097                 obd_free_memmd(exp, &set->set_oi->oi_md);
1098         if (idx_arr)
1099                 free_idx_array(idx_arr, idx_cnt);
1100         EXIT;
1101         return rc;
1102 }
1103
1104 void qos_update(struct lov_obd *lov)
1105 {
1106         ENTRY;
1107         lov->lov_qos.lq_dirty = 1;
1108 }