lustre/osp/osp_precreate.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Intel, Inc.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/osp/osp_sync.c
  37  *
  38  * Lustre OST Proxy Device
  39  *
  40  * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
  41  * Author: Mikhail Pershin <mike.pershin@intel.com>
  42  */
  43
  44 #ifndef EXPORT_SYMTAB
  45 # define EXPORT_SYMTAB
  46 #endif
  47 #define DEBUG_SUBSYSTEM S_MDS
  48
  49 #include "osp_internal.h"
  50
  51 /*
  52  * there are two specific states to take care about:
  53  *
  54  * = import is disconnected =
  55  *
  56  * = import is inactive =
  57  *   in this case osp_declare_object_create() returns an error
  58  *
  59  */
  60
  61 /*
  62  * statfs
  63  */
  64 static inline int osp_statfs_need_update(struct osp_device *d)
  65 {
  66         return !cfs_time_before(cfs_time_current(),
  67                                 d->opd_statfs_fresh_till);
  68 }
  69
  70 static void osp_statfs_timer_cb(unsigned long _d)
  71 {
  72         struct osp_device *d = (struct osp_device *) _d;
  73
  74         LASSERT(d);
  75         cfs_waitq_signal(&d->opd_pre_waitq);
  76 }
  77
  78 static int osp_statfs_interpret(const struct lu_env *env,
  79                                 struct ptlrpc_request *req,
  80                                 union ptlrpc_async_args *aa, int rc)
  81 {
  82         struct obd_import       *imp = req->rq_import;
  83         struct obd_statfs       *msfs;
  84         struct osp_device       *d;
  85
  86         ENTRY;
  87
  88         aa = ptlrpc_req_async_args(req);
  89         d = aa->pointer_arg[0];
  90         LASSERT(d);
  91
  92         if (rc != 0)
  93                 GOTO(out, rc);
  94
  95         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
  96         if (msfs == NULL)
  97                 GOTO(out, rc = -EPROTO);
  98
  99         d->opd_statfs = *msfs;
 100
 101         osp_pre_update_status(d, rc);
 102
 103         /* schedule next update */
 104         d->opd_statfs_fresh_till = cfs_time_shift(d->opd_statfs_maxage);
 105         cfs_timer_arm(&d->opd_statfs_timer, d->opd_statfs_fresh_till);
 106         d->opd_statfs_update_in_progress = 0;
 107
 108         CDEBUG(D_CACHE, "updated statfs %p\n", d);
 109
 110         RETURN(0);
 111 out:
 112         /* couldn't update statfs, try again as soon as possible */
 113         cfs_waitq_signal(&d->opd_pre_waitq);
 114         if (req->rq_import_generation == imp->imp_generation)
 115                 CERROR("%s: couldn't update statfs: rc = %d\n",
 116                        d->opd_obd->obd_name, rc);
 117         RETURN(rc);
 118 }
 119
 120 static int osp_statfs_update(struct osp_device *d)
 121 {
 122         struct ptlrpc_request   *req;
 123         struct obd_import       *imp;
 124         union ptlrpc_async_args *aa;
 125         int                      rc;
 126
 127         ENTRY;
 128
 129         CDEBUG(D_CACHE, "going to update statfs\n");
 130
 131         imp = d->opd_obd->u.cli.cl_import;
 132         LASSERT(imp);
 133
 134         req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
 135         if (req == NULL)
 136                 RETURN(-ENOMEM);
 137
 138         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
 139         if (rc) {
 140                 ptlrpc_request_free(req);
 141                 RETURN(rc);
 142         }
 143         ptlrpc_request_set_replen(req);
 144         req->rq_request_portal = OST_CREATE_PORTAL;
 145         ptlrpc_at_set_req_timeout(req);
 146
 147         req->rq_interpret_reply = (ptlrpc_interpterer_t)osp_statfs_interpret;
 148         aa = ptlrpc_req_async_args(req);
 149         aa->pointer_arg[0] = d;
 150
 151         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
 152
 153         cfs_timer_disarm(&d->opd_statfs_timer);
 154
 155         /*
 156          * no updates till reply
 157          */
 158         d->opd_statfs_fresh_till = cfs_time_shift(obd_timeout * 1000);
 159         d->opd_statfs_update_in_progress = 1;
 160
 161         RETURN(0);
 162 }
 163
 164 /*
 165  * XXX: there might be a case where removed object(s) do not add free
 166  * space (empty object). if the number of such deletions is high, then
 167  * we can start to update statfs too often - a rpc storm
 168  * TODO: some throttling is needed
 169  */
 170 void osp_statfs_need_now(struct osp_device *d)
 171 {
 172         if (!d->opd_statfs_update_in_progress) {
 173                 /*
 174                  * if current status is -ENOSPC (lack of free space on OST)
 175                  * then we should poll OST immediately once object destroy
 176                  * is replied
 177                  */
 178                 d->opd_statfs_fresh_till = cfs_time_shift(-1);
 179                 cfs_timer_disarm(&d->opd_statfs_timer);
 180                 cfs_waitq_signal(&d->opd_pre_waitq);
 181         }
 182 }
 183
 184
 185 /*
 186  * OSP tries to maintain pool of available objects so that calls to create
 187  * objects don't block most of time
 188  *
 189  * each time OSP gets connected to OST, we should start from precreation cleanup
 190  */
 191 static inline int osp_precreate_running(struct osp_device *d)
 192 {
 193         return !!(d->opd_pre_thread.t_flags & SVC_RUNNING);
 194 }
 195
 196 static inline int osp_precreate_stopped(struct osp_device *d)
 197 {
 198         return !!(d->opd_pre_thread.t_flags & SVC_STOPPED);
 199 }
 200
 201 static inline int osp_precreate_near_empty_nolock(struct osp_device *d)
 202 {
 203         int window = d->opd_pre_last_created - d->opd_pre_next;
 204
 205         /* don't consider new precreation till OST is healty and
 206          * has free space */
 207         return ((window - d->opd_pre_reserved < d->opd_pre_grow_count / 2) &&
 208                 (d->opd_pre_status == 0));
 209 }
 210
 211 static inline int osp_precreate_near_empty(struct osp_device *d)
 212 {
 213         int rc;
 214
 215         /* XXX: do we really need locking here? */
 216         cfs_spin_lock(&d->opd_pre_lock);
 217         rc = osp_precreate_near_empty_nolock(d);
 218         cfs_spin_unlock(&d->opd_pre_lock);
 219         return rc;
 220 }
 221
 222 static int osp_precreate_send(struct osp_device *d)
 223 {
 224         struct ptlrpc_request   *req;
 225         struct obd_import       *imp;
 226         struct ost_body         *body;
 227         int                      rc, grow, diff;
 228
 229         ENTRY;
 230
 231         /* don't precreate new objects till OST healthy and has free space */
 232         if (unlikely(d->opd_pre_status)) {
 233                 CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
 234                        d->opd_obd->obd_name, d->opd_pre_status);
 235                 RETURN(0);
 236         }
 237
 238         /*
 239          * if not connection/initialization is compeleted, ignore
 240          */
 241         imp = d->opd_obd->u.cli.cl_import;
 242         LASSERT(imp);
 243
 244         req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
 245         if (req == NULL)
 246                 RETURN(-ENOMEM);
 247         req->rq_request_portal = OST_CREATE_PORTAL;
 248         /* we should not resend create request - anyway we will have delorphan
 249          * and kill these objects */
 250         req->rq_no_delay = req->rq_no_resend = 1;
 251
 252         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
 253         if (rc) {
 254                 ptlrpc_request_free(req);
 255                 RETURN(rc);
 256         }
 257
 258         cfs_spin_lock(&d->opd_pre_lock);
 259         if (d->opd_pre_grow_count > d->opd_pre_max_grow_count / 2)
 260                 d->opd_pre_grow_count = d->opd_pre_max_grow_count / 2;
 261         grow = d->opd_pre_grow_count;
 262         cfs_spin_unlock(&d->opd_pre_lock);
 263
 264         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 265         LASSERT(body);
 266         body->oa.o_id = d->opd_pre_last_created + grow;
 267         body->oa.o_seq = FID_SEQ_OST_MDT0; /* XXX: support for CMD? */
 268         body->oa.o_valid = OBD_MD_FLGROUP;
 269
 270         ptlrpc_request_set_replen(req);
 271
 272         rc = ptlrpc_queue_wait(req);
 273         if (rc) {
 274                 CERROR("%s: can't precreate: rc = %d\n",
 275                        d->opd_obd->obd_name, rc);
 276                 GOTO(out_req, rc);
 277         }
 278         LASSERT(req->rq_transno == 0);
 279
 280         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 281         if (body == NULL)
 282                 GOTO(out_req, rc = -EPROTO);
 283
 284         CDEBUG(D_HA, "new last_created %lu\n", (unsigned long) body->oa.o_id);
 285         LASSERT(body->oa.o_id > d->opd_pre_next);
 286
 287         diff = body->oa.o_id - d->opd_pre_last_created;
 288
 289         cfs_spin_lock(&d->opd_pre_lock);
 290         if (diff < grow) {
 291                 /* the OST has not managed to create all the
 292                  * objects we asked for */
 293                 d->opd_pre_grow_count = max(diff, OST_MIN_PRECREATE);
 294                 d->opd_pre_grow_slow = 1;
 295         } else {
 296                 /* the OST is able to keep up with the work,
 297                  * we could consider increasing grow_count
 298                  * next time if needed */
 299                 d->opd_pre_grow_slow = 0;
 300         }
 301         d->opd_pre_last_created = body->oa.o_id;
 302         cfs_spin_unlock(&d->opd_pre_lock);
 303         CDEBUG(D_OTHER, "current precreated pool: %llu-%llu\n",
 304                d->opd_pre_next, d->opd_pre_last_created);
 305
 306 out_req:
 307         /* now we can wakeup all users awaiting for objects */
 308         osp_pre_update_status(d, rc);
 309         cfs_waitq_signal(&d->opd_pre_user_waitq);
 310
 311         ptlrpc_req_finished(req);
 312         RETURN(rc);
 313 }
 314
 315 /**
 316  * asks OST to clean precreate orphans
 317  * and gets next id for new objects
 318  */
 319 static int osp_precreate_cleanup_orphans(struct osp_device *d)
 320 {
 321         struct ptlrpc_request   *req = NULL;
 322         struct obd_import       *imp;
 323         struct ost_body         *body;
 324         int                      rc;
 325
 326         ENTRY;
 327
 328         LASSERT(d->opd_recovery_completed);
 329         LASSERT(d->opd_pre_reserved == 0);
 330
 331         imp = d->opd_obd->u.cli.cl_import;
 332         LASSERT(imp);
 333
 334         req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
 335         if (req == NULL)
 336                 RETURN(-ENOMEM);
 337
 338         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
 339         if (rc) {
 340                 ptlrpc_request_free(req);
 341                 RETURN(rc);
 342         }
 343
 344         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 345         if (body == NULL)
 346                 GOTO(out_req, rc = -EPROTO);
 347
 348         body->oa.o_flags = OBD_FL_DELORPHAN;
 349         body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
 350         body->oa.o_seq = FID_SEQ_OST_MDT0;
 351
 352         /* remove from NEXT after used one */
 353         body->oa.o_id = d->opd_last_used_id + 1;
 354
 355         ptlrpc_request_set_replen(req);
 356
 357         /* Don't resend the delorphan req */
 358         req->rq_no_resend = req->rq_no_delay = 1;
 359
 360         rc = ptlrpc_queue_wait(req);
 361         if (rc)
 362                 GOTO(out_req, rc);
 363
 364         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 365         if (body == NULL)
 366                 GOTO(out_req, rc = -EPROTO);
 367
 368         /*
 369          * OST provides us with id new pool starts from in body->oa.o_id
 370          */
 371         cfs_spin_lock(&d->opd_pre_lock);
 372         if (le64_to_cpu(d->opd_last_used_id) > body->oa.o_id) {
 373                 d->opd_pre_grow_count = OST_MIN_PRECREATE +
 374                                         le64_to_cpu(d->opd_last_used_id) -
 375                                         body->oa.o_id;
 376                 d->opd_pre_last_created = le64_to_cpu(d->opd_last_used_id) + 1;
 377         } else {
 378                 d->opd_pre_grow_count = OST_MIN_PRECREATE;
 379                 d->opd_pre_last_created = body->oa.o_id + 1;
 380         }
 381         d->opd_pre_next = d->opd_pre_last_created;
 382         d->opd_pre_grow_slow = 0;
 383         cfs_spin_unlock(&d->opd_pre_lock);
 384
 385         /* now we can wakeup all users awaiting for objects */
 386         osp_pre_update_status(d, rc);
 387         cfs_waitq_signal(&d->opd_pre_user_waitq);
 388
 389         CDEBUG(D_HA, "Got last_id "LPU64" from OST, last_used is "LPU64
 390                ", next "LPU64"\n", body->oa.o_id,
 391                le64_to_cpu(d->opd_last_used_id), d->opd_pre_next);
 392
 393 out_req:
 394         ptlrpc_req_finished(req);
 395         RETURN(rc);
 396 }
 397
 398 /*
 399  * the function updates current precreation status used: functional or not
 400  *
 401  * rc is a last code from the transport, rc == 0 meaning transport works
 402  * well and users of lod can use objects from this OSP
 403  *
 404  * the status depends on current usage of OST
 405  */
 406 void osp_pre_update_status(struct osp_device *d, int rc)
 407 {
 408         struct obd_statfs       *msfs = &d->opd_statfs;
 409         int                      old = d->opd_pre_status;
 410         __u64                    used;
 411
 412         d->opd_pre_status = rc;
 413         if (rc)
 414                 goto out;
 415
 416         if (likely(msfs->os_type)) {
 417                 used = min_t(__u64, (msfs->os_blocks - msfs->os_bfree) >> 10,
 418                                     1 << 30);
 419                 if ((msfs->os_ffree < 32) || (msfs->os_bavail < used)) {
 420                         d->opd_pre_status = -ENOSPC;
 421                         if (old != -ENOSPC)
 422                                 CDEBUG(D_INFO, "%s: status: "LPU64" blocks, "
 423                                        LPU64" free, "LPU64" used, "LPU64" "
 424                                        "avail -> %d: rc = %d\n",
 425                                        d->opd_obd->obd_name, msfs->os_blocks,
 426                                        msfs->os_bfree, used, msfs->os_bavail,
 427                                        d->opd_pre_status, rc);
 428                 } else if (old == -ENOSPC) {
 429                         d->opd_pre_status = 0;
 430                         d->opd_pre_grow_slow = 0;
 431                         d->opd_pre_grow_count = OST_MIN_PRECREATE;
 432                         cfs_waitq_signal(&d->opd_pre_waitq);
 433                         CDEBUG(D_INFO, "%s: no space: "LPU64" blocks, "LPU64
 434                                " free, "LPU64" used, "LPU64" avail -> %d: "
 435                                "rc = %d\n", d->opd_obd->obd_name,
 436                                msfs->os_blocks, msfs->os_bfree, used,
 437                                msfs->os_bavail, d->opd_pre_status, rc);
 438                 }
 439         }
 440
 441 out:
 442         cfs_waitq_signal(&d->opd_pre_user_waitq);
 443 }
 444
 445 static int osp_precreate_thread(void *_arg)
 446 {
 447         struct osp_device       *d = _arg;
 448         struct ptlrpc_thread    *thread = &d->opd_pre_thread;
 449         struct l_wait_info       lwi = { 0 };
 450         char                     pname[16];
 451         int                      rc;
 452
 453         ENTRY;
 454
 455         sprintf(pname, "osp-pre-%u\n", d->opd_index);
 456         cfs_daemonize(pname);
 457
 458         cfs_spin_lock(&d->opd_pre_lock);
 459         thread->t_flags = SVC_RUNNING;
 460         cfs_spin_unlock(&d->opd_pre_lock);
 461         cfs_waitq_signal(&thread->t_ctl_waitq);
 462
 463         while (osp_precreate_running(d)) {
 464                 /*
 465                  * need to be connected to OST
 466                  */
 467                 while (osp_precreate_running(d)) {
 468                         l_wait_event(d->opd_pre_waitq,
 469                                      !osp_precreate_running(d) ||
 470                                      d->opd_new_connection, &lwi);
 471
 472                         if (!osp_precreate_running(d))
 473                                 break;
 474
 475                         if (!d->opd_new_connection)
 476                                 continue;
 477
 478                         /* got connected */
 479                         d->opd_new_connection = 0;
 480                         d->opd_got_disconnected = 0;
 481                         break;
 482                 }
 483
 484                 osp_statfs_update(d);
 485
 486                 /*
 487                  * wait for local recovery to finish, so we can cleanup orphans
 488                  * orphans are all objects since "last used" (assigned), but
 489                  * there might be objects reserved and in some cases they won't
 490                  * be used. we can't cleanup them till we're sure they won't be
 491                  * used. so we block new reservations and wait till all reserved
 492                  * objects either user or released.
 493                  */
 494                 l_wait_event(d->opd_pre_waitq, (!d->opd_pre_reserved &&
 495                                                 d->opd_recovery_completed) ||
 496                              !osp_precreate_running(d) ||
 497                              d->opd_got_disconnected, &lwi);
 498
 499                 if (osp_precreate_running(d) && !d->opd_got_disconnected) {
 500                         rc = osp_precreate_cleanup_orphans(d);
 501                         if (rc) {
 502                                 CERROR("%s: cannot cleanup orphans: rc = %d\n",
 503                                        d->opd_obd->obd_name,  rc);
 504                         }
 505                 }
 506
 507                 /*
 508                  * connected, can handle precreates now
 509                  */
 510                 while (osp_precreate_running(d)) {
 511                         l_wait_event(d->opd_pre_waitq,
 512                                      !osp_precreate_running(d) ||
 513                                      osp_precreate_near_empty(d) ||
 514                                      osp_statfs_need_update(d) ||
 515                                      d->opd_got_disconnected, &lwi);
 516
 517                         if (!osp_precreate_running(d))
 518                                 break;
 519
 520                         /* something happened to the connection
 521                          * have to start from the beginning */
 522                         if (d->opd_got_disconnected)
 523                                 break;
 524
 525                         if (osp_statfs_need_update(d))
 526                                 osp_statfs_update(d);
 527
 528                         if (osp_precreate_near_empty(d)) {
 529                                 rc = osp_precreate_send(d);
 530                                 /* osp_precreate_send() sets opd_pre_status
 531                                  * in case of error, that prevent the using of
 532                                  * failed device. */
 533                                 if (rc != 0 && rc != -ENOSPC &&
 534                                     rc != -ETIMEDOUT && rc != -ENOTCONN)
 535                                         CERROR("%s: cannot precreate objects:"
 536                                                " rc = %d\n",
 537                                                d->opd_obd->obd_name, rc);
 538                         }
 539                 }
 540         }
 541
 542         thread->t_flags = SVC_STOPPED;
 543         cfs_waitq_signal(&thread->t_ctl_waitq);
 544
 545         RETURN(0);
 546 }
 547
 548 static int osp_precreate_ready_condition(struct osp_device *d)
 549 {
 550         /* ready if got enough precreated objects */
 551         if (d->opd_pre_next + d->opd_pre_reserved < d->opd_pre_last_created)
 552                 return 1;
 553
 554         /* ready if OST reported no space */
 555         if (d->opd_pre_status != 0)
 556                 return 1;
 557
 558         return 0;
 559 }
 560
 561 static int osp_precreate_timeout_condition(void *data)
 562 {
 563         struct osp_device *d = data;
 564
 565         LCONSOLE_WARN("%s: slow creates, last="LPU64", next="LPU64", "
 566                       "reserved="LPU64", status=%d\n",
 567                       d->opd_obd->obd_name, d->opd_pre_last_created,
 568                       d->opd_pre_next, d->opd_pre_reserved,
 569                       d->opd_pre_status);
 570
 571         return 0;
 572 }
 573
 574 /*
 575  * called to reserve object in the pool
 576  * return codes:
 577  *  ENOSPC - no space on corresponded OST
 578  *  EAGAIN - precreation is in progress, try later
 579  *  EIO    - no access to OST
 580  */
 581 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d)
 582 {
 583         struct l_wait_info       lwi;
 584         cfs_time_t               expire = cfs_time_shift(obd_timeout);
 585         int                      precreated, rc;
 586
 587         ENTRY;
 588
 589         LASSERT(d->opd_pre_last_created >= d->opd_pre_next);
 590
 591         lwi = LWI_TIMEOUT(cfs_time_seconds(obd_timeout),
 592                           osp_precreate_timeout_condition, d);
 593
 594         /*
 595          * wait till:
 596          *  - preallocation is done
 597          *  - no free space expected soon
 598          *  - can't connect to OST for too long (obd_timeout)
 599          */
 600         while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
 601                 rc == -ENODEV) {
 602                 if (unlikely(rc == -ENODEV)) {
 603                         if (cfs_time_aftereq(cfs_time_current(), expire))
 604                                 break;
 605                 }
 606
 607                 /*
 608                  * increase number of precreations
 609                  */
 610                 if (d->opd_pre_grow_count < d->opd_pre_max_grow_count &&
 611                     d->opd_pre_grow_slow == 0 &&
 612                     (d->opd_pre_last_created - d->opd_pre_next <=
 613                      d->opd_pre_grow_count / 4 + 1)) {
 614                         cfs_spin_lock(&d->opd_pre_lock);
 615                         d->opd_pre_grow_slow = 1;
 616                         d->opd_pre_grow_count *= 2;
 617                         cfs_spin_unlock(&d->opd_pre_lock);
 618                 }
 619
 620                 /*
 621                  * we never use the last object in the window
 622                  */
 623                 cfs_spin_lock(&d->opd_pre_lock);
 624                 precreated = d->opd_pre_last_created - d->opd_pre_next;
 625                 if (precreated > d->opd_pre_reserved) {
 626                         d->opd_pre_reserved++;
 627                         cfs_spin_unlock(&d->opd_pre_lock);
 628                         rc = 0;
 629
 630                         /* XXX: don't wake up if precreation is in progress */
 631                         if (osp_precreate_near_empty_nolock(d))
 632                                 cfs_waitq_signal(&d->opd_pre_waitq);
 633
 634                         break;
 635                 }
 636                 cfs_spin_unlock(&d->opd_pre_lock);
 637
 638                 /* XXX: don't wake up if precreation is in progress */
 639                 cfs_waitq_signal(&d->opd_pre_waitq);
 640
 641                 l_wait_event(d->opd_pre_user_waitq,
 642                              osp_precreate_ready_condition(d), &lwi);
 643         }
 644
 645         RETURN(rc);
 646 }
 647
 648 /*
 649  * this function relies on reservation made before
 650  */
 651 __u64 osp_precreate_get_id(struct osp_device *d)
 652 {
 653         obd_id objid;
 654
 655         /* grab next id from the pool */
 656         cfs_spin_lock(&d->opd_pre_lock);
 657         LASSERT(d->opd_pre_next <= d->opd_pre_last_created);
 658         objid = d->opd_pre_next++;
 659         d->opd_pre_reserved--;
 660         /*
 661          * last_used_id must be changed along with getting new id otherwise
 662          * we might miscalculate gap causing object loss or leak
 663          */
 664         osp_update_last_id(d, objid);
 665         cfs_spin_unlock(&d->opd_pre_lock);
 666
 667         /*
 668          * probably main thread suspended orphan cleanup till
 669          * all reservations are released, see comment in
 670          * osp_precreate_thread() just before orphan cleanup
 671          */
 672         if (unlikely(d->opd_pre_reserved == 0 && d->opd_pre_status))
 673                 cfs_waitq_signal(&d->opd_pre_waitq);
 674
 675         return objid;
 676 }
 677
 678 /*
 679  *
 680  */
 681 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
 682                         __u64 size)
 683 {
 684         struct osp_device       *d = lu2osp_dev(dt->do_lu.lo_dev);
 685         struct ptlrpc_request   *req = NULL;
 686         struct obd_import       *imp;
 687         struct ost_body         *body;
 688         struct obdo             *oa = NULL;
 689         int                      rc;
 690
 691         ENTRY;
 692
 693         imp = d->opd_obd->u.cli.cl_import;
 694         LASSERT(imp);
 695
 696         req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
 697         if (req == NULL)
 698                 RETURN(-ENOMEM);
 699
 700         /* XXX: capa support? */
 701         /* osc_set_capa_size(req, &RMF_CAPA1, capa); */
 702         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
 703         if (rc) {
 704                 ptlrpc_request_free(req);
 705                 RETURN(rc);
 706         }
 707
 708         /*
 709          * XXX: decide how do we do here with resend
 710          * if we don't resend, then client may see wrong file size
 711          * if we do resend, then MDS thread can get stuck for quite long
 712          */
 713         req->rq_no_resend = req->rq_no_delay = 1;
 714
 715         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
 716         ptlrpc_at_set_req_timeout(req);
 717
 718         OBD_ALLOC_PTR(oa);
 719         if (oa == NULL)
 720                 GOTO(out, rc = -ENOMEM);
 721
 722         rc = fid_ostid_pack(lu_object_fid(&dt->do_lu), &oa->o_oi);
 723         LASSERT(rc == 0);
 724         oa->o_size = size;
 725         oa->o_blocks = OBD_OBJECT_EOF;
 726         oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 727                       OBD_MD_FLID | OBD_MD_FLGROUP;
 728
 729         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 730         LASSERT(body);
 731         lustre_set_wire_obdo(&body->oa, oa);
 732
 733         /* XXX: capa support? */
 734         /* osc_pack_capa(req, body, capa); */
 735
 736         ptlrpc_request_set_replen(req);
 737
 738         rc = ptlrpc_queue_wait(req);
 739         if (rc)
 740                 CERROR("can't punch object: %d\n", rc);
 741 out:
 742         ptlrpc_req_finished(req);
 743         if (oa)
 744                 OBD_FREE_PTR(oa);
 745         RETURN(rc);
 746 }
 747
 748 int osp_init_precreate(struct osp_device *d)
 749 {
 750         struct l_wait_info       lwi = { 0 };
 751         int                      rc;
 752
 753         ENTRY;
 754
 755         /* initially precreation isn't ready */
 756         d->opd_pre_status = -EAGAIN;
 757         d->opd_pre_next = 1;
 758         d->opd_pre_last_created = 1;
 759         d->opd_pre_reserved = 0;
 760         d->opd_got_disconnected = 1;
 761         d->opd_pre_grow_slow = 0;
 762         d->opd_pre_grow_count = OST_MIN_PRECREATE;
 763         d->opd_pre_min_grow_count = OST_MIN_PRECREATE;
 764         d->opd_pre_max_grow_count = OST_MAX_PRECREATE;
 765
 766         cfs_spin_lock_init(&d->opd_pre_lock);
 767         cfs_waitq_init(&d->opd_pre_waitq);
 768         cfs_waitq_init(&d->opd_pre_user_waitq);
 769         cfs_waitq_init(&d->opd_pre_thread.t_ctl_waitq);
 770
 771         /*
 772          * Initialize statfs-related things
 773          */
 774         d->opd_statfs_maxage = 5; /* default update interval */
 775         d->opd_statfs_fresh_till = cfs_time_shift(-1000);
 776         CDEBUG(D_OTHER, "current %llu, fresh till %llu\n",
 777                (unsigned long long)cfs_time_current(),
 778                (unsigned long long)d->opd_statfs_fresh_till);
 779         cfs_timer_init(&d->opd_statfs_timer, osp_statfs_timer_cb, d);
 780
 781         /*
 782          * start thread handling precreation and statfs updates
 783          */
 784         rc = cfs_create_thread(osp_precreate_thread, d, 0);
 785         if (rc < 0) {
 786                 CERROR("can't start precreate thread %d\n", rc);
 787                 RETURN(rc);
 788         }
 789
 790         l_wait_event(d->opd_pre_thread.t_ctl_waitq,
 791                      osp_precreate_running(d) || osp_precreate_stopped(d),
 792                      &lwi);
 793
 794         RETURN(0);
 795 }
 796
 797 void osp_precreate_fini(struct osp_device *d)
 798 {
 799         struct ptlrpc_thread *thread = &d->opd_pre_thread;
 800
 801         ENTRY;
 802
 803         cfs_timer_disarm(&d->opd_statfs_timer);
 804
 805         thread->t_flags = SVC_STOPPING;
 806         cfs_waitq_signal(&d->opd_pre_waitq);
 807
 808         cfs_wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
 809
 810         EXIT;
 811 }
 812