lustre/lov/lov_qos.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (C) 2002, 2003 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of the Lustre file system, http://www.lustre.org
   7  *   Lustre is a trademark of Cluster File Systems, Inc.
   8  *
   9  *   You may have signed or agreed to another license before downloading
  10  *   this software.  If so, you are bound by the terms and conditions
  11  *   of that agreement, and the following does not apply to you.  See the
  12  *   LICENSE file included with this distribution for more information.
  13  *
  14  *   If you did not agree to a different license, then this copy of Lustre
  15  *   is open source software; you can redistribute it and/or modify it
  16  *   under the terms of version 2 of the GNU General Public License as
  17  *   published by the Free Software Foundation.
  18  *
  19  *   In either case, Lustre is distributed in the hope that it will be
  20  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  21  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  22  *   license text for more details.
  23  */
  24
  25 #ifndef EXPORT_SYMTAB
  26 # define EXPORT_SYMTAB
  27 #endif
  28 #define DEBUG_SUBSYSTEM S_LOV
  29
  30 #ifdef __KERNEL__
  31 #include <libcfs/libcfs.h>
  32 #else
  33 #include <liblustre.h>
  34 #endif
  35
  36 #include <obd_class.h>
  37 #include <obd_lov.h>
  38 #include "lov_internal.h"
  39
  40 /* #define QOS_DEBUG 1  */
  41 #define D_QOS D_OTHER
  42
  43 #define TGT_BAVAIL(i) (lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bavail * \
  44                        lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_bsize)
  45 #define TGT_FFREE(i)  (lov->lov_tgts[i]->ltd_exp->exp_obd->obd_osfs.os_ffree)
  46
  47
  48 int qos_add_tgt(struct obd_device *obd, __u32 index)
  49 {
  50         struct lov_obd *lov = &obd->u.lov;
  51         struct lov_qos_oss *oss, *temposs;
  52         struct obd_export *exp = lov->lov_tgts[index]->ltd_exp;
  53         int rc = 0, found = 0;
  54         ENTRY;
  55
  56         /* We only need this QOS struct on MDT, not clients - but we may not
  57            have registered the LOV's observer yet, so there's no way to know */
  58         if (!exp || !exp->exp_connection) {
  59                 CERROR("Missing connection\n");
  60                 RETURN(-ENOTCONN);
  61         }
  62
  63         down_write(&lov->lov_qos.lq_rw_sem);
  64         mutex_down(&lov->lov_lock);
  65         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
  66                 if (obd_uuid_equals(&oss->lqo_uuid,
  67                                     &exp->exp_connection->c_remote_uuid)) {
  68                         found++;
  69                         break;
  70                 }
  71         }
  72
  73         if (!found) {
  74                 OBD_ALLOC_PTR(oss);
  75                 if (!oss)
  76                         GOTO(out, rc = -ENOMEM);
  77                 memcpy(&oss->lqo_uuid,
  78                        &exp->exp_connection->c_remote_uuid,
  79                        sizeof(oss->lqo_uuid));
  80         } else {
  81                 /* Assume we have to move this one */
  82                 list_del(&oss->lqo_oss_list);
  83         }
  84
  85         oss->lqo_ost_count++;
  86         lov->lov_tgts[index]->ltd_qos.ltq_oss = oss;
  87
  88         /* Add sorted by # of OSTs.  Find the first entry that we're
  89            bigger than... */
  90         list_for_each_entry(temposs, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
  91                 if (oss->lqo_ost_count > temposs->lqo_ost_count)
  92                         break;
  93         }
  94         /* ...and add before it.  If we're the first or smallest, temposs
  95            points to the list head, and we add to the end. */
  96         list_add_tail(&oss->lqo_oss_list, &temposs->lqo_oss_list);
  97
  98         lov->lov_qos.lq_dirty = 1;
  99         lov->lov_qos.lq_dirty_rr = 1;
 100
 101         CDEBUG(D_QOS, "add tgt %s to OSS %s (%d OSTs)\n",
 102                obd_uuid2str(&lov->lov_tgts[index]->ltd_uuid),
 103                obd_uuid2str(&oss->lqo_uuid),
 104                oss->lqo_ost_count);
 105
 106 out:
 107         mutex_up(&lov->lov_lock);
 108         up_write(&lov->lov_qos.lq_rw_sem);
 109         RETURN(rc);
 110 }
 111
 112 int qos_del_tgt(struct obd_device *obd, __u32 index)
 113 {
 114         struct lov_obd *lov = &obd->u.lov;
 115         struct lov_qos_oss *oss;
 116         int rc = 0;
 117         ENTRY;
 118
 119         if (!lov->lov_tgts[index])
 120                 RETURN(0);
 121
 122         down_write(&lov->lov_qos.lq_rw_sem);
 123
 124         oss = lov->lov_tgts[index]->ltd_qos.ltq_oss;
 125         if (!oss)
 126                 GOTO(out, rc = -ENOENT);
 127
 128         oss->lqo_ost_count--;
 129         if (oss->lqo_ost_count == 0) {
 130                 CDEBUG(D_QOS, "removing OSS %s\n",
 131                        obd_uuid2str(&oss->lqo_uuid));
 132                 list_del(&oss->lqo_oss_list);
 133                 OBD_FREE_PTR(oss);
 134         }
 135
 136         lov->lov_qos.lq_dirty = 1;
 137         lov->lov_qos.lq_dirty_rr = 1;
 138 out:
 139         up_write(&lov->lov_qos.lq_rw_sem);
 140         RETURN(rc);
 141 }
 142
 143 /* Recalculate per-object penalties for OSSs and OSTs,
 144    depends on size of each ost in an oss */
 145 static int qos_calc_ppo(struct obd_device *obd)
 146 {
 147         struct lov_obd *lov = &obd->u.lov;
 148         struct lov_qos_oss *oss;
 149         __u64 ba_max, ba_min, temp;
 150         __u32 num_active;
 151         int rc, i, prio_wide;
 152         ENTRY;
 153
 154         if (!lov->lov_qos.lq_dirty)
 155                 GOTO(out, rc = 0);
 156
 157         num_active = lov->desc.ld_active_tgt_count - 1;
 158         if (num_active < 1)
 159                 GOTO(out, rc = -EAGAIN);
 160
 161         /* find bavail on each OSS */
 162         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 163                 oss->lqo_bavail = 0;
 164         }
 165         lov->lov_qos.lq_active_oss_count = 0;
 166
 167         /* How badly user wants to select osts "widely" (not recently chosen
 168            and not on recent oss's).  As opposed to "freely" (free space
 169            avail.) 0-256. */
 170         prio_wide = 256 - lov->lov_qos.lq_prio_free;
 171
 172         ba_min = (__u64)(-1);
 173         ba_max = 0;
 174         /* Calculate OST penalty per object */
 175         /* (lov ref taken in alloc_qos) */
 176         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 177                 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
 178                         continue;
 179                 temp = TGT_BAVAIL(i);
 180                 if (!temp)
 181                         continue;
 182                 ba_min = min(temp, ba_min);
 183                 ba_max = max(temp, ba_max);
 184
 185                 /* Count the number of usable OSS's */
 186                 if (lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_bavail == 0)
 187                         lov->lov_qos.lq_active_oss_count++;
 188                 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_bavail += temp;
 189
 190                 /* per-OST penalty is prio * TGT_bavail / (num_ost - 1) / 2 */
 191                 temp >>= 1;
 192                 do_div(temp, num_active);
 193                 lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj =
 194                         (temp * prio_wide) >> 8;
 195
 196                 if (lov->lov_qos.lq_reset == 0)
 197                         lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
 198         }
 199
 200         num_active = lov->lov_qos.lq_active_oss_count - 1;
 201         if (num_active < 1) {
 202                 /* If there's only 1 OSS, we can't penalize it, so instead
 203                    we have to double the OST penalty */
 204                 num_active = 1;
 205                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 206                         if (lov->lov_tgts[i] == NULL)
 207                                 continue;
 208                         lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj <<= 1;
 209                 }
 210         }
 211
 212         /* Per-OSS penalty is prio * oss_avail / oss_osts / (num_oss - 1) / 2 */
 213         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 214                 temp = oss->lqo_bavail >> 1;
 215                 do_div(temp, oss->lqo_ost_count * num_active);
 216                 oss->lqo_penalty_per_obj = (temp * prio_wide) >> 8;
 217                 if (lov->lov_qos.lq_reset == 0)
 218                         oss->lqo_penalty = 0;
 219         }
 220
 221         lov->lov_qos.lq_dirty = 0;
 222         lov->lov_qos.lq_reset = 0;
 223
 224         /* If each ost has almost same free space,
 225          * do rr allocation for better creation performance */
 226         lov->lov_qos.lq_same_space = 0;
 227         temp = ba_max - ba_min;
 228         ba_min = (ba_min * 51) >> 8;     /* 51/256 = .20 */
 229         if (temp < ba_min) {
 230                 /* Difference is less than 20% */
 231                 lov->lov_qos.lq_same_space = 1;
 232                 /* Reset weights for the next time we enter qos mode */
 233                 lov->lov_qos.lq_reset = 0;
 234         }
 235         rc = 0;
 236
 237 out:
 238         if (!rc && lov->lov_qos.lq_same_space)
 239                 RETURN(-EAGAIN);
 240         RETURN(rc);
 241 }
 242
 243 static int qos_calc_weight(struct lov_obd *lov, int i)
 244 {
 245         __u64 temp, temp2;
 246
 247         /* Final ost weight = TGT_BAVAIL - ost_penalty - oss_penalty */
 248         temp = TGT_BAVAIL(i);
 249         temp2 = lov->lov_tgts[i]->ltd_qos.ltq_penalty +
 250                 lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty;
 251         if (temp < temp2)
 252                 lov->lov_tgts[i]->ltd_qos.ltq_weight = 0;
 253         else
 254                 lov->lov_tgts[i]->ltd_qos.ltq_weight = temp - temp2;
 255         return 0;
 256 }
 257
 258 /* We just used this index for a stripe; adjust everyone's weights */
 259 static int qos_used(struct lov_obd *lov, __u32 index, __u64 *total_wt)
 260 {
 261         struct lov_qos_oss *oss;
 262         int i;
 263         ENTRY;
 264
 265         /* Don't allocate from this stripe anymore, until the next alloc_qos */
 266         lov->lov_tgts[index]->ltd_qos.ltq_usable = 0;
 267
 268         oss = lov->lov_tgts[index]->ltd_qos.ltq_oss;
 269
 270         /* Decay old penalty by half (we're adding max penalty, and don't
 271            want it to run away.) */
 272         lov->lov_tgts[index]->ltd_qos.ltq_penalty >>= 1;
 273         oss->lqo_penalty >>= 1;
 274
 275         /* Set max penalties for this OST and OSS */
 276         lov->lov_tgts[index]->ltd_qos.ltq_penalty +=
 277                 lov->lov_tgts[index]->ltd_qos.ltq_penalty_per_obj *
 278                 lov->desc.ld_active_tgt_count;
 279         oss->lqo_penalty += oss->lqo_penalty_per_obj *
 280                 lov->lov_qos.lq_active_oss_count;
 281
 282         /* Decrease all OSS penalties */
 283         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 284                 if (oss->lqo_penalty < oss->lqo_penalty_per_obj)
 285                         oss->lqo_penalty = 0;
 286                 else
 287                         oss->lqo_penalty -= oss->lqo_penalty_per_obj;
 288         }
 289
 290         *total_wt = 0;
 291         /* Decrease all OST penalties */
 292         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 293                 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
 294                         continue;
 295                 if (lov->lov_tgts[i]->ltd_qos.ltq_penalty <
 296                     lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj)
 297                         lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
 298                 else
 299                         lov->lov_tgts[i]->ltd_qos.ltq_penalty -=
 300                         lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj;
 301
 302                 qos_calc_weight(lov, i);
 303
 304                 /* Recalc the total weight of usable osts */
 305                 if (lov->lov_tgts[i]->ltd_qos.ltq_usable)
 306                         *total_wt += lov->lov_tgts[i]->ltd_qos.ltq_weight;
 307
 308 #ifdef QOS_DEBUG
 309                 CDEBUG(D_QOS, "recalc tgt %d avail="LPU64
 310                        " ostppo="LPU64" ostp="LPU64" ossppo="LPU64
 311                        " ossp="LPU64" wt="LPU64"\n",
 312                        i, TGT_BAVAIL(i) >> 10,
 313                        lov->lov_tgts[i]->ltd_qos.ltq_penalty_per_obj >> 10,
 314                        lov->lov_tgts[i]->ltd_qos.ltq_penalty >> 10,
 315                        lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty_per_obj >> 10,
 316                        lov->lov_tgts[i]->ltd_qos.ltq_oss->lqo_penalty >> 10,
 317                        lov->lov_tgts[i]->ltd_qos.ltq_weight >> 10);
 318 #endif
 319         }
 320
 321         RETURN(0);
 322 }
 323
 324 #define LOV_QOS_EMPTY ((__u32)-1)
 325 /* compute optimal round-robin order, based on OSTs per OSS */
 326 static int qos_calc_rr(struct lov_obd *lov)
 327 {
 328         struct lov_qos_oss *oss;
 329         unsigned ost_count, placed, real_count;
 330         int i;
 331         ENTRY;
 332
 333         if (!lov->lov_qos.lq_dirty_rr) {
 334                 LASSERT(lov->lov_qos.lq_rr_size);
 335                 RETURN(0);
 336         }
 337
 338         /* Do actuall allocation. */
 339         down_write(&lov->lov_qos.lq_rw_sem);
 340
 341         /*
 342          * Check again. While we were sleeping on @lq_rw_sem something could
 343          * change.
 344          */
 345         if (!lov->lov_qos.lq_dirty_rr) {
 346                 LASSERT(lov->lov_qos.lq_rr_size);
 347                 up_write(&lov->lov_qos.lq_rw_sem);
 348                 RETURN(0);
 349         }
 350
 351         ost_count = lov->desc.ld_tgt_count;
 352
 353         if (lov->lov_qos.lq_rr_size)
 354                 OBD_FREE(lov->lov_qos.lq_rr_array, lov->lov_qos.lq_rr_size);
 355         lov->lov_qos.lq_rr_size = ost_count *
 356                 sizeof(lov->lov_qos.lq_rr_array[0]);
 357         OBD_ALLOC(lov->lov_qos.lq_rr_array, lov->lov_qos.lq_rr_size);
 358         if (!lov->lov_qos.lq_rr_array) {
 359                 lov->lov_qos.lq_rr_size = 0;
 360                 up_write(&lov->lov_qos.lq_rw_sem);
 361                 RETURN(-ENOMEM);
 362         }
 363
 364         real_count = 0;
 365         for (i = 0; i < ost_count; i++) {
 366                 lov->lov_qos.lq_rr_array[i] = LOV_QOS_EMPTY;
 367                 if (lov->lov_tgts[i])
 368                         real_count++;
 369         }
 370
 371         /* Place all the OSTs from 1 OSS at the same time. */
 372         placed = 0;
 373         list_for_each_entry(oss, &lov->lov_qos.lq_oss_list, lqo_oss_list) {
 374                 int j = 0;
 375                 for (i = 0; i < ost_count; i++) {
 376                       LASSERT(lov->lov_tgts[i] != NULL);
 377                       if (lov->lov_tgts[i]->ltd_qos.ltq_oss == oss) {
 378                               /* Evenly space these OSTs across arrayspace */
 379                               int next = j * ost_count / oss->lqo_ost_count;
 380                               LASSERT(next < ost_count);
 381                               while (lov->lov_qos.lq_rr_array[next] !=
 382                                      LOV_QOS_EMPTY)
 383                                       next = (next + 1) % ost_count;
 384                               lov->lov_qos.lq_rr_array[next] = i;
 385                               j++;
 386                               placed++;
 387                       }
 388                 }
 389                 LASSERT(j == oss->lqo_ost_count);
 390         }
 391
 392         lov->lov_qos.lq_dirty_rr = 0;
 393         up_write(&lov->lov_qos.lq_rw_sem);
 394
 395         if (placed != real_count) {
 396                 /* This should never happen */
 397                 LCONSOLE_ERROR_MSG(0x14e, "Failed to place all OSTs in the "
 398                                    "round-robin list (%d of %d).\n",
 399                                    placed, real_count);
 400                 for (i = 0; i < ost_count; i++) {
 401                         LCONSOLE(D_WARNING, "rr #%d ost idx=%d\n", i,
 402                                  lov->lov_qos.lq_rr_array[i]);
 403                 }
 404                 lov->lov_qos.lq_dirty_rr = 1;
 405                 RETURN(-EAGAIN);
 406         }
 407
 408 #ifdef QOS_DEBUG
 409         for (i = 0; i < ost_count; i++) {
 410                 LCONSOLE(D_QOS, "rr #%d ost idx=%d\n", i,
 411                          lov->lov_qos.lq_rr_array[i]);
 412         }
 413 #endif
 414
 415         RETURN(0);
 416 }
 417
 418
 419 void qos_shrink_lsm(struct lov_request_set *set)
 420 {
 421         struct lov_stripe_md *lsm = set->set_oi->oi_md, *lsm_new;
 422         /* XXX LOV STACKING call into osc for sizes */
 423         unsigned oldsize, newsize;
 424
 425         if (set->set_oti && set->set_cookies && set->set_cookie_sent) {
 426                 struct llog_cookie *cookies;
 427                 oldsize = lsm->lsm_stripe_count * sizeof(*cookies);
 428                 newsize = set->set_count * sizeof(*cookies);
 429
 430                 cookies = set->set_cookies;
 431                 oti_alloc_cookies(set->set_oti, set->set_count);
 432                 if (set->set_oti->oti_logcookies) {
 433                         memcpy(set->set_oti->oti_logcookies, cookies, newsize);
 434                         OBD_FREE(cookies, oldsize);
 435                         set->set_cookies = set->set_oti->oti_logcookies;
 436                 } else {
 437                         CWARN("'leaking' %d bytes\n", oldsize - newsize);
 438                 }
 439         }
 440
 441         CWARN("using fewer stripes for object "LPX64": old %u new %u\n",
 442               lsm->lsm_object_id, lsm->lsm_stripe_count, set->set_count);
 443         LASSERT(lsm->lsm_stripe_count >= set->set_count);
 444
 445         newsize = lov_stripe_md_size(set->set_count);
 446         OBD_ALLOC(lsm_new, newsize);
 447         if (lsm_new != NULL) {
 448                 int i;
 449                 memcpy(lsm_new, lsm, sizeof(*lsm));
 450                 for (i = 0; i < lsm->lsm_stripe_count; i++) {
 451                         if (i < set->set_count) {
 452                                 lsm_new->lsm_oinfo[i] = lsm->lsm_oinfo[i];
 453                                 continue;
 454                         }
 455                         OBD_SLAB_FREE(lsm->lsm_oinfo[i], lov_oinfo_slab,
 456                                       sizeof(struct lov_oinfo));
 457                 }
 458                 lsm_new->lsm_stripe_count = set->set_count;
 459                 OBD_FREE(lsm, sizeof(struct lov_stripe_md) +
 460                          lsm->lsm_stripe_count * sizeof(struct lov_oinfo *));
 461                 set->set_oi->oi_md = lsm_new;
 462         } else {
 463                 CWARN("'leaking' few bytes\n");
 464         }
 465 }
 466
 467 int qos_remedy_create(struct lov_request_set *set, struct lov_request *req)
 468 {
 469         struct lov_stripe_md *lsm = set->set_oi->oi_md;
 470         struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
 471         unsigned ost_idx, ost_count = lov->desc.ld_tgt_count;
 472         int stripe, i, rc = -EIO;
 473         ENTRY;
 474
 475         ost_idx = (req->rq_idx + lsm->lsm_stripe_count) % ost_count;
 476         for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
 477                 if (!lov->lov_tgts[ost_idx] ||
 478                     !lov->lov_tgts[ost_idx]->ltd_active)
 479                         continue;
 480                 /* check if objects has been created on this ost */
 481                 for (stripe = 0; stripe < lsm->lsm_stripe_count; stripe++) {
 482                         if (stripe == req->rq_stripe)
 483                                 continue;
 484                         if (ost_idx == lsm->lsm_oinfo[stripe]->loi_ost_idx)
 485                                 break;
 486                 }
 487
 488                 if (stripe >= lsm->lsm_stripe_count) {
 489                         req->rq_idx = ost_idx;
 490                         rc = obd_create(lov->lov_tgts[ost_idx]->ltd_exp,
 491                                         req->rq_oi.oi_oa, &req->rq_oi.oi_md,
 492                                         set->set_oti);
 493                         if (!rc)
 494                                 break;
 495                 }
 496         }
 497         RETURN(rc);
 498 }
 499
 500 static int min_stripe_count(int stripe_cnt, int flags)
 501 {
 502         return (flags & LOV_USES_DEFAULT_STRIPE ?
 503                 stripe_cnt - (stripe_cnt / 4) : stripe_cnt);
 504 }
 505
 506 #define LOV_CREATE_RESEED_MULT 4
 507 #define LOV_CREATE_RESEED_MIN  1000
 508 /* Allocate objects on osts with round-robin algorithm */
 509 static int alloc_rr(struct lov_obd *lov, int *idx_arr, int *stripe_cnt,
 510                     int flags)
 511 {
 512         unsigned array_idx, ost_count = lov->desc.ld_tgt_count;
 513         unsigned ost_active_count = lov->desc.ld_active_tgt_count;
 514         int i, *idx_pos;
 515         __u32 ost_idx;
 516         int ost_start_idx_temp;
 517         int speed = 0;
 518         int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
 519         ENTRY;
 520
 521         i = qos_calc_rr(lov);
 522         if (i)
 523                 RETURN(i);
 524
 525         if (--lov->lov_start_count <= 0) {
 526                 lov->lov_start_idx = ll_rand() % ost_count;
 527                 lov->lov_start_count =
 528                         (LOV_CREATE_RESEED_MIN / max(ost_active_count, 1U) +
 529                          LOV_CREATE_RESEED_MULT) * max(ost_active_count, 1U);
 530         } else if (stripe_cnt_min >= ost_active_count ||
 531                    lov->lov_start_idx > ost_count) {
 532                 /* If we have allocated from all of the OSTs, slowly
 533                    precess the next start */
 534                 lov->lov_start_idx %= ost_count;
 535                 if (stripe_cnt_min > 1 &&
 536                     (ost_active_count % stripe_cnt_min) != 1)
 537                         ++lov->lov_offset_idx;
 538         }
 539         down_read(&lov->lov_qos.lq_rw_sem);
 540         ost_start_idx_temp = lov->lov_start_idx;
 541
 542 repeat_find:
 543         array_idx = (lov->lov_start_idx + lov->lov_offset_idx) % ost_count;
 544         idx_pos = idx_arr;
 545 #ifdef QOS_DEBUG
 546         CDEBUG(D_QOS, "want %d startidx %d startcnt %d offset %d arrayidx %d\n",
 547                stripe_cnt_min, lov->lov_start_idx, lov->lov_start_count,
 548                lov->lov_offset_idx, array_idx);
 549 #endif
 550
 551         for (i = 0; i < ost_count; i++, array_idx=(array_idx + 1) % ost_count) {
 552                 ++lov->lov_start_idx;
 553                 ost_idx = lov->lov_qos.lq_rr_array[array_idx];
 554 #ifdef QOS_DEBUG
 555                 CDEBUG(D_QOS, "#%d strt %d act %d strp %d ary %d idx %d\n",
 556                        i, lov->lov_start_idx,
 557                        ((ost_idx != LOV_QOS_EMPTY) && lov->lov_tgts[ost_idx]) ?
 558                        lov->lov_tgts[ost_idx]->ltd_active : 0,
 559                        idx_pos - idx_arr, array_idx, ost_idx);
 560 #endif
 561                 if ((ost_idx == LOV_QOS_EMPTY) || !lov->lov_tgts[ost_idx] ||
 562                     !lov->lov_tgts[ost_idx]->ltd_active)
 563                         continue;
 564                 /* Fail Check before osc_precreate() is called
 565                    so we can only 'fail' single OSC. */
 566                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && ost_idx == 0)
 567                         continue;
 568
 569                 /* Drop slow OSCs if we can */
 570                 if (obd_precreate(lov->lov_tgts[ost_idx]->ltd_exp, speed == 0) >
 571                                   speed)
 572                         continue;
 573
 574                 *idx_pos = ost_idx;
 575                 idx_pos++;
 576                 /* We have enough stripes */
 577                 if (idx_pos - idx_arr == *stripe_cnt)
 578                         break;
 579         }
 580         if ((speed < 2) && (idx_pos - idx_arr < stripe_cnt_min)) {
 581                 /* Try again, allowing slower OSCs */
 582                 speed++;
 583                 lov->lov_start_idx = ost_start_idx_temp;
 584                 goto repeat_find;
 585         }
 586
 587         up_read(&lov->lov_qos.lq_rw_sem);
 588
 589         *stripe_cnt = idx_pos - idx_arr;
 590         RETURN(0);
 591 }
 592
 593 /* alloc objects on osts with specific stripe offset */
 594 static int alloc_specific(struct lov_obd *lov, struct lov_stripe_md *lsm,
 595                           int *idx_arr)
 596 {
 597         unsigned ost_idx, ost_count = lov->desc.ld_tgt_count;
 598         int i, *idx_pos;
 599         int speed = 0;
 600         ENTRY;
 601
 602 repeat_find:
 603         idx_pos = idx_arr;
 604         ost_idx = lsm->lsm_oinfo[0]->loi_ost_idx;
 605         for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) {
 606                 if (!lov->lov_tgts[ost_idx] ||
 607                     !lov->lov_tgts[ost_idx]->ltd_active) {
 608                         continue;
 609                 }
 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 == 0) >
 618                     speed)
 619                         continue;
 620
 621                 *idx_pos = ost_idx;
 622                 idx_pos++;
 623                 /* We have enough stripes */
 624                 if (idx_pos - idx_arr == lsm->lsm_stripe_count)
 625                         RETURN(0);
 626         }
 627         if (speed < 2) {
 628                 /* Try again, allowing slower OSCs */
 629                 speed++;
 630                 goto repeat_find;
 631         }
 632
 633         /* If we were passed specific striping params, then a failure to
 634          * meet those requirements is an error, since we can't reallocate
 635          * that memory (it might be part of a larger array or something).
 636          *
 637          * We can only get here if lsm_stripe_count was originally > 1.
 638          */
 639         CERROR("can't lstripe objid "LPX64": have "LPSZ" want %u\n",
 640                lsm->lsm_object_id, idx_pos - idx_arr, lsm->lsm_stripe_count);
 641         RETURN(-EFBIG);
 642 }
 643
 644 /* Alloc objects on osts with optimization based on:
 645    - free space
 646    - network resources (shared OSS's)
 647 */
 648 static int alloc_qos(struct obd_export *exp, int *idx_arr, int *stripe_cnt,
 649                      int flags)
 650 {
 651         struct lov_obd *lov = &exp->exp_obd->u.lov;
 652         static time_t last_warn = 0;
 653         time_t now = cfs_time_current_sec();
 654         __u64 total_bavail, total_weight = 0;
 655         __u32 ost_count;
 656         int nfound, good_osts, i, warn = 0, rc = 0;
 657         int stripe_cnt_min = min_stripe_count(*stripe_cnt, flags);
 658         ENTRY;
 659
 660         if (stripe_cnt_min < 1)
 661                 GOTO(out, rc = -EINVAL);
 662
 663         lov_getref(exp->exp_obd);
 664
 665         /* Detect -EAGAIN early, before expensive lock is taken. */
 666         if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space)
 667                 GOTO(out, rc = -EAGAIN);
 668
 669         /* Do actuall allocation, use write lock here. */
 670         down_write(&lov->lov_qos.lq_rw_sem);
 671
 672         /*
 673          * Check again, while we were sleeping on @lq_rw_sem things could
 674          * change.
 675          */
 676         if (!lov->lov_qos.lq_dirty && lov->lov_qos.lq_same_space) {
 677                 up_write(&lov->lov_qos.lq_rw_sem);
 678                 GOTO(out, rc = -EAGAIN);
 679         }
 680         ost_count = lov->desc.ld_tgt_count;
 681
 682         if (lov->desc.ld_active_tgt_count < 2)
 683                 GOTO(out_up_write, rc = -EAGAIN);
 684
 685         rc = qos_calc_ppo(exp->exp_obd);
 686         if (rc)
 687                 GOTO(out_up_write, rc);
 688
 689         total_bavail = 0;
 690         good_osts = 0;
 691         /* Warn users about zero available space/inode every 30 min */
 692         if (cfs_time_sub(now, last_warn) > 60 * 30)
 693                 warn = 1;
 694         /* Find all the OSTs that are valid stripe candidates */
 695         for (i = 0; i < ost_count; i++) {
 696                 __u64 bavail;
 697
 698                 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
 699                         continue;
 700                 bavail = TGT_BAVAIL(i);
 701                 if (!bavail) {
 702                         if (warn) {
 703                                 CDEBUG(D_QOS, "no free space on %s\n",
 704                                        obd_uuid2str(&lov->lov_tgts[i]->ltd_uuid));
 705                                 last_warn = now;
 706                         }
 707                         continue;
 708                 }
 709                 if (!TGT_FFREE(i)) {
 710                         if (warn) {
 711                                 CDEBUG(D_QOS, "no free inodes on %s\n",
 712                                        obd_uuid2str(&lov->lov_tgts[i]->ltd_uuid));
 713                                 last_warn = now;
 714                         }
 715                         continue;
 716                 }
 717
 718                 /* Fail Check before osc_precreate() is called
 719                    so we can only 'fail' single OSC. */
 720                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OSC_PRECREATE) && i == 0)
 721                         continue;
 722
 723                 if (obd_precreate(lov->lov_tgts[i]->ltd_exp, 1) >= 2)
 724                         continue;
 725
 726                 lov->lov_tgts[i]->ltd_qos.ltq_usable = 1;
 727                 qos_calc_weight(lov, i);
 728                 total_bavail += bavail;
 729                 total_weight += lov->lov_tgts[i]->ltd_qos.ltq_weight;
 730
 731                 good_osts++;
 732         }
 733
 734         if (!total_bavail)
 735                 GOTO(out_up_write, rc = -ENOSPC);
 736
 737          if (good_osts < stripe_cnt_min)
 738                 GOTO(out_up_write, rc = -EAGAIN);
 739
 740         /* We have enough osts */
 741         if (good_osts < *stripe_cnt)
 742                 *stripe_cnt = good_osts;
 743
 744         if (!*stripe_cnt)
 745                 GOTO(out_up_write, rc = -EAGAIN);
 746
 747         /* Find enough OSTs with weighted random allocation. */
 748         nfound = 0;
 749         while (nfound < *stripe_cnt) {
 750                 __u64 rand, cur_weight = 0;
 751
 752                 rc = -ENODEV;
 753
 754                 if (total_weight) {
 755 #if BITS_PER_LONG == 32
 756                         rand = ll_rand() % (unsigned)total_weight;
 757                         /* If total_weight > 32-bit, first generate the high
 758                          * 32 bits of the random number, then add in the low
 759                          * 32 bits (truncated to the upper limit, if needed) */
 760                         if (total_weight > 0xffffffffULL)
 761                                 rand = (__u64)(ll_rand() %
 762                                           (unsigned)(total_weight >> 32)) << 32;
 763                         else
 764                                 rand = 0;
 765
 766                         if (rand == (total_weight & 0xffffffff00000000ULL))
 767                                 rand |= ll_rand() % (unsigned)total_weight;
 768                         else
 769                                 rand |= ll_rand();
 770 #else
 771                         rand = ((__u64)ll_rand() << 32 | ll_rand()) %
 772                                 total_weight;
 773 #endif
 774                 } else {
 775                         rand = 0;
 776                 }
 777
 778                 /* On average, this will hit larger-weighted osts more often.
 779                    0-weight osts will always get used last (only when rand=0).*/
 780                 for (i = 0; i < ost_count; i++) {
 781                         if (!lov->lov_tgts[i] ||
 782                             !lov->lov_tgts[i]->ltd_qos.ltq_usable)
 783                                 continue;
 784
 785                         cur_weight += lov->lov_tgts[i]->ltd_qos.ltq_weight;
 786                         if (cur_weight >= rand) {
 787 #ifdef QOS_DEBUG
 788                                 CDEBUG(D_QOS, "assigned stripe=%d to idx=%d\n",
 789                                        nfound, i);
 790 #endif
 791                                 idx_arr[nfound++] = i;
 792                                 qos_used(lov, i, &total_weight);
 793                                 rc = 0;
 794                                 break;
 795                         }
 796                 }
 797                 /* should never satisfy below condition */
 798                 if (rc) {
 799                         CERROR("Didn't find any OSTs?\n");
 800                         break;
 801                 }
 802         }
 803         LASSERT(nfound == *stripe_cnt);
 804
 805 out_up_write:
 806         up_write(&lov->lov_qos.lq_rw_sem);
 807
 808 out:
 809         if (rc == -EAGAIN)
 810                 rc = alloc_rr(lov, idx_arr, stripe_cnt, flags);
 811
 812         lov_putref(exp->exp_obd);
 813         RETURN(rc);
 814 }
 815
 816 /* return new alloced stripe count on success */
 817 static int alloc_idx_array(struct obd_export *exp, struct lov_stripe_md *lsm,
 818                            int newea, int **idx_arr, int *arr_cnt, int flags)
 819 {
 820         struct lov_obd *lov = &exp->exp_obd->u.lov;
 821         int stripe_cnt = lsm->lsm_stripe_count;
 822         int i, rc = 0;
 823         int *tmp_arr = NULL;
 824         ENTRY;
 825
 826         *arr_cnt = stripe_cnt;
 827         OBD_ALLOC(tmp_arr, *arr_cnt * sizeof(int));
 828         if (tmp_arr == NULL)
 829                 RETURN(-ENOMEM);
 830         for (i = 0; i < *arr_cnt; i++)
 831                 tmp_arr[i] = -1;
 832
 833         if (newea ||
 834             lsm->lsm_oinfo[0]->loi_ost_idx >= lov->desc.ld_tgt_count)
 835                 rc = alloc_qos(exp, tmp_arr, &stripe_cnt, flags);
 836         else
 837                 rc = alloc_specific(lov, lsm, tmp_arr);
 838
 839         if (rc)
 840                 GOTO(out_arr, rc);
 841
 842         *idx_arr = tmp_arr;
 843         RETURN(stripe_cnt);
 844 out_arr:
 845         OBD_FREE(tmp_arr, *arr_cnt * sizeof(int));
 846         *arr_cnt = 0;
 847         RETURN(rc);
 848 }
 849
 850 static void free_idx_array(int *idx_arr, int arr_cnt)
 851 {
 852         if (arr_cnt)
 853                 OBD_FREE(idx_arr, arr_cnt * sizeof(int));
 854 }
 855
 856 int qos_prep_create(struct obd_export *exp, struct lov_request_set *set)
 857 {
 858         struct lov_obd *lov = &exp->exp_obd->u.lov;
 859         struct lov_stripe_md *lsm;
 860         struct obdo *src_oa = set->set_oi->oi_oa;
 861         struct obd_trans_info *oti = set->set_oti;
 862         int i, stripes, rc = 0, newea = 0;
 863         int *idx_arr, idx_cnt = 0;
 864         int flag = LOV_USES_ASSIGNED_STRIPE;
 865         ENTRY;
 866
 867         LASSERT(src_oa->o_valid & OBD_MD_FLID);
 868         LASSERT(src_oa->o_valid & OBD_MD_FLGROUP);
 869
 870         if (set->set_oi->oi_md == NULL) {
 871                 int stripes_def = lov_get_stripecnt(lov, 0);
 872
 873                 /* If the MDS file was truncated up to some size, stripe over
 874                  * enough OSTs to allow the file to be created at that size.
 875                  * This may mean we use more than the default # of stripes. */
 876                 if (src_oa->o_valid & OBD_MD_FLSIZE) {
 877                         obd_size min_bavail = LUSTRE_STRIPE_MAXBYTES;
 878
 879                         /* Find a small number of stripes we can use
 880                            (up to # of active osts). */
 881                         stripes = 1;
 882                         lov_getref(exp->exp_obd);
 883                         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 884                                 if (!lov->lov_tgts[i] ||
 885                                     !lov->lov_tgts[i]->ltd_active)
 886                                         continue;
 887                                 min_bavail = min(min_bavail, TGT_BAVAIL(i));
 888                                 if (min_bavail * stripes > src_oa->o_size)
 889                                         break;
 890                                 stripes++;
 891                         }
 892                         lov_putref(exp->exp_obd);
 893
 894                         if (stripes < stripes_def)
 895                                 stripes = stripes_def;
 896                 } else {
 897                          flag = LOV_USES_DEFAULT_STRIPE;
 898                          stripes = stripes_def;
 899                 }
 900
 901                 rc = lov_alloc_memmd(&set->set_oi->oi_md, stripes,
 902                                      lov->desc.ld_pattern ?
 903                                      lov->desc.ld_pattern : LOV_PATTERN_RAID0,
 904                                      LOV_MAGIC);
 905                 if (rc < 0)
 906                         GOTO(out_err, rc);
 907                 newea = 1;
 908                 rc = 0;
 909         }
 910
 911         lsm = set->set_oi->oi_md;
 912         lsm->lsm_object_id = src_oa->o_id;
 913         lsm->lsm_object_gr = src_oa->o_gr;
 914
 915         if (!lsm->lsm_stripe_size)
 916                 lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size;
 917         if (!lsm->lsm_pattern) {
 918                 LASSERT(lov->desc.ld_pattern);
 919                 lsm->lsm_pattern = lov->desc.ld_pattern;
 920         }
 921
 922         stripes = alloc_idx_array(exp, lsm, newea, &idx_arr, &idx_cnt, flag);
 923         if (stripes <= 0)
 924                 GOTO(out_err, rc = stripes ? stripes : -EIO);
 925         LASSERTF(stripes <= lsm->lsm_stripe_count,"requested %d allocated %d\n",
 926                  lsm->lsm_stripe_count, stripes);
 927
 928         for (i = 0; i < stripes; i++) {
 929                 struct lov_request *req;
 930                 int ost_idx = idx_arr[i];
 931                 LASSERT(ost_idx >= 0);
 932
 933                 OBD_ALLOC(req, sizeof(*req));
 934                 if (req == NULL)
 935                         GOTO(out_err, rc = -ENOMEM);
 936                 lov_set_add_req(req, set);
 937
 938                 req->rq_buflen = sizeof(*req->rq_oi.oi_md);
 939                 OBD_ALLOC(req->rq_oi.oi_md, req->rq_buflen);
 940                 if (req->rq_oi.oi_md == NULL)
 941                         GOTO(out_err, rc = -ENOMEM);
 942
 943                 OBDO_ALLOC(req->rq_oi.oi_oa);
 944                 if (req->rq_oi.oi_oa == NULL)
 945                         GOTO(out_err, rc = -ENOMEM);
 946
 947                 req->rq_idx = ost_idx;
 948                 req->rq_stripe = i;
 949                 /* create data objects with "parent" OA */
 950                 memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
 951
 952                 /* XXX When we start creating objects on demand, we need to
 953                  *     make sure that we always create the object on the
 954                  *     stripe which holds the existing file size.
 955                  */
 956                 if (src_oa->o_valid & OBD_MD_FLSIZE) {
 957                         req->rq_oi.oi_oa->o_size =
 958                                 lov_size_to_stripe(lsm, src_oa->o_size, i);
 959
 960                         CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
 961                                i, req->rq_oi.oi_oa->o_size, src_oa->o_size);
 962                 }
 963         }
 964         LASSERT(set->set_count == stripes);
 965
 966         if (stripes < lsm->lsm_stripe_count)
 967                 qos_shrink_lsm(set);
 968
 969         if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) {
 970                 oti_alloc_cookies(oti, set->set_count);
 971                 if (!oti->oti_logcookies)
 972                         GOTO(out_err, rc = -ENOMEM);
 973                 set->set_cookies = oti->oti_logcookies;
 974         }
 975 out_err:
 976         if (newea && rc)
 977                 obd_free_memmd(exp, &set->set_oi->oi_md);
 978         free_idx_array(idx_arr, idx_cnt);
 979         EXIT;
 980         return rc;
 981 }
 982
 983 void qos_update(struct lov_obd *lov)
 984 {
 985         ENTRY;
 986         lov->lov_qos.lq_dirty = 1;
 987 }
 988