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