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