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 static int osp_get_lastid_from_ost(struct osp_device *d)
 317 {
 318         struct ptlrpc_request   *req;
 319         struct obd_import       *imp;
 320         obd_id                  *reply;
 321         char                    *tmp;
 322         int                      rc;
 323
 324         imp = d->opd_obd->u.cli.cl_import;
 325         LASSERT(imp);
 326
 327         req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_ID);
 328         if (req == NULL)
 329                 RETURN(-ENOMEM);
 330
 331         req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
 332                              RCL_CLIENT, sizeof(KEY_LAST_ID));
 333         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
 334         if (rc) {
 335                 ptlrpc_request_free(req);
 336                 RETURN(rc);
 337         }
 338
 339         tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
 340         memcpy(tmp, KEY_LAST_ID, sizeof(KEY_LAST_ID));
 341
 342         req->rq_no_delay = req->rq_no_resend = 1;
 343         ptlrpc_request_set_replen(req);
 344         rc = ptlrpc_queue_wait(req);
 345         if (rc) {
 346                 /* bad-bad OST.. let sysadm sort this out */
 347                 ptlrpc_set_import_active(imp, 0);
 348                 GOTO(out, rc);
 349         }
 350
 351         reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
 352         if (reply == NULL)
 353                 GOTO(out, rc = -EPROTO);
 354
 355         d->opd_last_used_id = *reply;
 356         CDEBUG(D_HA, "%s: got last_id "LPU64" from OST\n",
 357                d->opd_obd->obd_name, d->opd_last_used_id);
 358
 359 out:
 360         ptlrpc_req_finished(req);
 361         RETURN(rc);
 362
 363 }
 364
 365 /**
 366  * asks OST to clean precreate orphans
 367  * and gets next id for new objects
 368  */
 369 static int osp_precreate_cleanup_orphans(struct osp_device *d)
 370 {
 371         struct ptlrpc_request   *req = NULL;
 372         struct obd_import       *imp;
 373         struct ost_body         *body;
 374         int                      rc;
 375
 376         ENTRY;
 377
 378         LASSERT(d->opd_recovery_completed);
 379         LASSERT(d->opd_pre_reserved == 0);
 380
 381         CDEBUG(D_HA, "%s: going to cleanup orphans since "LPU64"\n",
 382                 d->opd_obd->obd_name, d->opd_last_used_id);
 383
 384         if (d->opd_last_used_id < 2) {
 385                 /* lastid looks strange... ask OST */
 386                 rc = osp_get_lastid_from_ost(d);
 387                 if (rc)
 388                         GOTO(out, rc);
 389         }
 390
 391         imp = d->opd_obd->u.cli.cl_import;
 392         LASSERT(imp);
 393
 394         req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
 395         if (req == NULL)
 396                 GOTO(out, rc = -ENOMEM);
 397
 398         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
 399         if (rc) {
 400                 ptlrpc_request_free(req);
 401                 GOTO(out, rc);
 402         }
 403
 404         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 405         if (body == NULL)
 406                 GOTO(out, rc = -EPROTO);
 407
 408         body->oa.o_flags = OBD_FL_DELORPHAN;
 409         body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
 410         body->oa.o_seq = FID_SEQ_OST_MDT0;
 411
 412         /* remove from NEXT after used one */
 413         body->oa.o_id = d->opd_last_used_id + 1;
 414
 415         ptlrpc_request_set_replen(req);
 416
 417         /* Don't resend the delorphan req */
 418         req->rq_no_resend = req->rq_no_delay = 1;
 419
 420         rc = ptlrpc_queue_wait(req);
 421         if (rc)
 422                 GOTO(out, rc);
 423
 424         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 425         if (body == NULL)
 426                 GOTO(out, rc = -EPROTO);
 427
 428         /*
 429          * OST provides us with id new pool starts from in body->oa.o_id
 430          */
 431         cfs_spin_lock(&d->opd_pre_lock);
 432         if (le64_to_cpu(d->opd_last_used_id) > body->oa.o_id) {
 433                 d->opd_pre_grow_count = OST_MIN_PRECREATE +
 434                                         le64_to_cpu(d->opd_last_used_id) -
 435                                         body->oa.o_id;
 436                 d->opd_pre_last_created = le64_to_cpu(d->opd_last_used_id) + 1;
 437         } else {
 438                 d->opd_pre_grow_count = OST_MIN_PRECREATE;
 439                 d->opd_pre_last_created = body->oa.o_id + 1;
 440         }
 441         d->opd_pre_next = d->opd_pre_last_created;
 442         d->opd_pre_grow_slow = 0;
 443         cfs_spin_unlock(&d->opd_pre_lock);
 444
 445         CDEBUG(D_HA, "Got last_id "LPU64" from OST, last_used is "LPU64
 446                ", next "LPU64"\n", body->oa.o_id,
 447                le64_to_cpu(d->opd_last_used_id), d->opd_pre_next);
 448
 449 out:
 450         if (req)
 451                 ptlrpc_req_finished(req);
 452
 453         RETURN(rc);
 454 }
 455
 456 /*
 457  * the function updates current precreation status used: functional or not
 458  *
 459  * rc is a last code from the transport, rc == 0 meaning transport works
 460  * well and users of lod can use objects from this OSP
 461  *
 462  * the status depends on current usage of OST
 463  */
 464 void osp_pre_update_status(struct osp_device *d, int rc)
 465 {
 466         struct obd_statfs       *msfs = &d->opd_statfs;
 467         int                      old = d->opd_pre_status;
 468         __u64                    used;
 469
 470         d->opd_pre_status = rc;
 471         if (rc)
 472                 goto out;
 473
 474         if (likely(msfs->os_type)) {
 475                 used = min_t(__u64, (msfs->os_blocks - msfs->os_bfree) >> 10,
 476                                     1 << 30);
 477                 if ((msfs->os_ffree < 32) || (msfs->os_bavail < used)) {
 478                         d->opd_pre_status = -ENOSPC;
 479                         if (old != -ENOSPC)
 480                                 CDEBUG(D_INFO, "%s: status: "LPU64" blocks, "
 481                                        LPU64" free, "LPU64" used, "LPU64" "
 482                                        "avail -> %d: rc = %d\n",
 483                                        d->opd_obd->obd_name, msfs->os_blocks,
 484                                        msfs->os_bfree, used, msfs->os_bavail,
 485                                        d->opd_pre_status, rc);
 486                         CDEBUG(D_INFO,
 487                                "non-commited changes: %lu, in progress: %u\n",
 488                                d->opd_syn_changes, d->opd_syn_rpc_in_progress);
 489                 } else if (old == -ENOSPC) {
 490                         d->opd_pre_status = 0;
 491                         d->opd_pre_grow_slow = 0;
 492                         d->opd_pre_grow_count = OST_MIN_PRECREATE;
 493                         cfs_waitq_signal(&d->opd_pre_waitq);
 494                         CDEBUG(D_INFO, "%s: no space: "LPU64" blocks, "LPU64
 495                                " free, "LPU64" used, "LPU64" avail -> %d: "
 496                                "rc = %d\n", d->opd_obd->obd_name,
 497                                msfs->os_blocks, msfs->os_bfree, used,
 498                                msfs->os_bavail, d->opd_pre_status, rc);
 499                 }
 500         }
 501
 502 out:
 503         cfs_waitq_signal(&d->opd_pre_user_waitq);
 504 }
 505
 506 static int osp_precreate_thread(void *_arg)
 507 {
 508         struct osp_device       *d = _arg;
 509         struct ptlrpc_thread    *thread = &d->opd_pre_thread;
 510         struct l_wait_info       lwi = { 0 };
 511         char                     pname[16];
 512         int                      rc;
 513
 514         ENTRY;
 515
 516         sprintf(pname, "osp-pre-%u\n", d->opd_index);
 517         cfs_daemonize(pname);
 518
 519         cfs_spin_lock(&d->opd_pre_lock);
 520         thread->t_flags = SVC_RUNNING;
 521         cfs_spin_unlock(&d->opd_pre_lock);
 522         cfs_waitq_signal(&thread->t_ctl_waitq);
 523
 524         while (osp_precreate_running(d)) {
 525                 /*
 526                  * need to be connected to OST
 527                  */
 528                 while (osp_precreate_running(d)) {
 529                         l_wait_event(d->opd_pre_waitq,
 530                                      !osp_precreate_running(d) ||
 531                                      d->opd_new_connection, &lwi);
 532
 533                         if (!osp_precreate_running(d))
 534                                 break;
 535
 536                         if (!d->opd_new_connection)
 537                                 continue;
 538
 539                         /* got connected */
 540                         d->opd_new_connection = 0;
 541                         d->opd_got_disconnected = 0;
 542                         break;
 543                 }
 544
 545                 osp_statfs_update(d);
 546
 547                 /*
 548                  * wait for local recovery to finish, so we can cleanup orphans
 549                  * orphans are all objects since "last used" (assigned), but
 550                  * there might be objects reserved and in some cases they won't
 551                  * be used. we can't cleanup them till we're sure they won't be
 552                  * used. so we block new reservations and wait till all reserved
 553                  * objects either user or released.
 554                  */
 555                 l_wait_event(d->opd_pre_waitq, (!d->opd_pre_reserved &&
 556                                                 d->opd_recovery_completed) ||
 557                              !osp_precreate_running(d) ||
 558                              d->opd_got_disconnected, &lwi);
 559
 560                 if (osp_precreate_running(d) && !d->opd_got_disconnected) {
 561                         rc = osp_precreate_cleanup_orphans(d);
 562                         if (rc) {
 563                                 CERROR("%s: cannot cleanup orphans: rc = %d\n",
 564                                        d->opd_obd->obd_name,  rc);
 565                                 /* we can't proceed from here, OST seem to
 566                                  * be in a bad shape, better to wait for
 567                                  * a new instance of the server and repeat
 568                                  * from the beginning. notify possible waiters
 569                                  * this OSP isn't quite functional yet */
 570                                 osp_pre_update_status(d, rc);
 571                                 cfs_waitq_signal(&d->opd_pre_user_waitq);
 572                                 l_wait_event(d->opd_pre_waitq,
 573                                              !osp_precreate_running(d) ||
 574                                              d->opd_new_connection, &lwi);
 575                                 continue;
 576
 577                         }
 578                 }
 579
 580                 /*
 581                  * connected, can handle precreates now
 582                  */
 583                 while (osp_precreate_running(d)) {
 584                         l_wait_event(d->opd_pre_waitq,
 585                                      !osp_precreate_running(d) ||
 586                                      osp_precreate_near_empty(d) ||
 587                                      osp_statfs_need_update(d) ||
 588                                      d->opd_got_disconnected, &lwi);
 589
 590                         if (!osp_precreate_running(d))
 591                                 break;
 592
 593                         /* something happened to the connection
 594                          * have to start from the beginning */
 595                         if (d->opd_got_disconnected)
 596                                 break;
 597
 598                         if (osp_statfs_need_update(d))
 599                                 osp_statfs_update(d);
 600
 601                         if (osp_precreate_near_empty(d)) {
 602                                 rc = osp_precreate_send(d);
 603                                 /* osp_precreate_send() sets opd_pre_status
 604                                  * in case of error, that prevent the using of
 605                                  * failed device. */
 606                                 if (rc != 0 && rc != -ENOSPC &&
 607                                     rc != -ETIMEDOUT && rc != -ENOTCONN)
 608                                         CERROR("%s: cannot precreate objects:"
 609                                                " rc = %d\n",
 610                                                d->opd_obd->obd_name, rc);
 611                         }
 612                 }
 613         }
 614
 615         thread->t_flags = SVC_STOPPED;
 616         cfs_waitq_signal(&thread->t_ctl_waitq);
 617
 618         RETURN(0);
 619 }
 620
 621 static int osp_precreate_ready_condition(struct osp_device *d)
 622 {
 623         /* ready if got enough precreated objects */
 624         if (d->opd_pre_next + d->opd_pre_reserved < d->opd_pre_last_created)
 625                 return 1;
 626
 627         /* ready if OST reported no space and no destoys in progress */
 628         if (d->opd_syn_changes + d->opd_syn_rpc_in_progress == 0 &&
 629             d->opd_pre_status != 0)
 630                 return 1;
 631
 632         return 0;
 633 }
 634
 635 static int osp_precreate_timeout_condition(void *data)
 636 {
 637         struct osp_device *d = data;
 638
 639         LCONSOLE_WARN("%s: slow creates, last="LPU64", next="LPU64", "
 640                       "reserved="LPU64", syn_changes=%lu, "
 641                       "syn_rpc_in_progress=%d, status=%d\n",
 642                       d->opd_obd->obd_name, d->opd_pre_last_created,
 643                       d->opd_pre_next, d->opd_pre_reserved,
 644                       d->opd_syn_changes, d->opd_syn_rpc_in_progress,
 645                       d->opd_pre_status);
 646
 647         return 0;
 648 }
 649
 650 /*
 651  * called to reserve object in the pool
 652  * return codes:
 653  *  ENOSPC - no space on corresponded OST
 654  *  EAGAIN - precreation is in progress, try later
 655  *  EIO    - no access to OST
 656  */
 657 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d)
 658 {
 659         struct l_wait_info       lwi;
 660         cfs_time_t               expire = cfs_time_shift(obd_timeout);
 661         int                      precreated, rc;
 662         int                      count = 0;
 663
 664         ENTRY;
 665
 666         LASSERT(d->opd_pre_last_created >= d->opd_pre_next);
 667
 668         lwi = LWI_TIMEOUT(cfs_time_seconds(obd_timeout),
 669                           osp_precreate_timeout_condition, d);
 670
 671         /*
 672          * wait till:
 673          *  - preallocation is done
 674          *  - no free space expected soon
 675          *  - can't connect to OST for too long (obd_timeout)
 676          */
 677         while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
 678                 rc == -ENODEV) {
 679                 if (unlikely(rc == -ENODEV)) {
 680                         if (cfs_time_aftereq(cfs_time_current(), expire))
 681                                 break;
 682                 }
 683
 684 #if LUSTRE_VERSION_CODE >= OBD_OCD_VERSION(2, 3, 90, 0)
 685 #error "remove this before the release"
 686 #endif
 687                 /*
 688                  * to address Andreas's concern on possible busy-loop
 689                  * between this thread and osp_precreate_send()
 690                  */
 691                 LASSERT(count++ < 1000);
 692
 693                 /*
 694                  * increase number of precreations
 695                  */
 696                 if (d->opd_pre_grow_count < d->opd_pre_max_grow_count &&
 697                     d->opd_pre_grow_slow == 0 &&
 698                     (d->opd_pre_last_created - d->opd_pre_next <=
 699                      d->opd_pre_grow_count / 4 + 1)) {
 700                         cfs_spin_lock(&d->opd_pre_lock);
 701                         d->opd_pre_grow_slow = 1;
 702                         d->opd_pre_grow_count *= 2;
 703                         cfs_spin_unlock(&d->opd_pre_lock);
 704                 }
 705
 706                 /*
 707                  * we never use the last object in the window
 708                  */
 709                 cfs_spin_lock(&d->opd_pre_lock);
 710                 precreated = d->opd_pre_last_created - d->opd_pre_next;
 711                 if (precreated > d->opd_pre_reserved) {
 712                         d->opd_pre_reserved++;
 713                         cfs_spin_unlock(&d->opd_pre_lock);
 714                         rc = 0;
 715
 716                         /* XXX: don't wake up if precreation is in progress */
 717                         if (osp_precreate_near_empty_nolock(d))
 718                                 cfs_waitq_signal(&d->opd_pre_waitq);
 719
 720                         break;
 721                 }
 722                 cfs_spin_unlock(&d->opd_pre_lock);
 723
 724                 /*
 725                  * all precreated objects have been used and no-space
 726                  * status leave us no chance to succeed very soon
 727                  * but if there is destroy in progress, then we should
 728                  * wait till that is done - some space might be released
 729                  */
 730                 if (unlikely(rc == -ENOSPC)) {
 731                         if (d->opd_syn_changes) {
 732                                 /* force local commit to release space */
 733                                 dt_commit_async(env, d->opd_storage);
 734                         }
 735                         if (d->opd_syn_rpc_in_progress) {
 736                                 /* just wait till destroys are done */
 737                                 /* see l_wait_even() few lines below */
 738                         }
 739                         if (d->opd_syn_changes +
 740                             d->opd_syn_rpc_in_progress == 0) {
 741                                 /* no hope for free space */
 742                                 break;
 743                         }
 744                 }
 745
 746                 /* XXX: don't wake up if precreation is in progress */
 747                 cfs_waitq_signal(&d->opd_pre_waitq);
 748
 749                 l_wait_event(d->opd_pre_user_waitq,
 750                              osp_precreate_ready_condition(d), &lwi);
 751         }
 752
 753         RETURN(rc);
 754 }
 755
 756 /*
 757  * this function relies on reservation made before
 758  */
 759 __u64 osp_precreate_get_id(struct osp_device *d)
 760 {
 761         obd_id objid;
 762
 763         /* grab next id from the pool */
 764         cfs_spin_lock(&d->opd_pre_lock);
 765         LASSERT(d->opd_pre_next <= d->opd_pre_last_created);
 766         objid = d->opd_pre_next++;
 767         d->opd_pre_reserved--;
 768         /*
 769          * last_used_id must be changed along with getting new id otherwise
 770          * we might miscalculate gap causing object loss or leak
 771          */
 772         osp_update_last_id(d, objid);
 773         cfs_spin_unlock(&d->opd_pre_lock);
 774
 775         /*
 776          * probably main thread suspended orphan cleanup till
 777          * all reservations are released, see comment in
 778          * osp_precreate_thread() just before orphan cleanup
 779          */
 780         if (unlikely(d->opd_pre_reserved == 0 && d->opd_pre_status))
 781                 cfs_waitq_signal(&d->opd_pre_waitq);
 782
 783         return objid;
 784 }
 785
 786 /*
 787  *
 788  */
 789 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
 790                         __u64 size)
 791 {
 792         struct osp_device       *d = lu2osp_dev(dt->do_lu.lo_dev);
 793         struct ptlrpc_request   *req = NULL;
 794         struct obd_import       *imp;
 795         struct ost_body         *body;
 796         struct obdo             *oa = NULL;
 797         int                      rc;
 798
 799         ENTRY;
 800
 801         imp = d->opd_obd->u.cli.cl_import;
 802         LASSERT(imp);
 803
 804         req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
 805         if (req == NULL)
 806                 RETURN(-ENOMEM);
 807
 808         /* XXX: capa support? */
 809         /* osc_set_capa_size(req, &RMF_CAPA1, capa); */
 810         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
 811         if (rc) {
 812                 ptlrpc_request_free(req);
 813                 RETURN(rc);
 814         }
 815
 816         /*
 817          * XXX: decide how do we do here with resend
 818          * if we don't resend, then client may see wrong file size
 819          * if we do resend, then MDS thread can get stuck for quite long
 820          */
 821         req->rq_no_resend = req->rq_no_delay = 1;
 822
 823         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
 824         ptlrpc_at_set_req_timeout(req);
 825
 826         OBD_ALLOC_PTR(oa);
 827         if (oa == NULL)
 828                 GOTO(out, rc = -ENOMEM);
 829
 830         rc = fid_ostid_pack(lu_object_fid(&dt->do_lu), &oa->o_oi);
 831         LASSERT(rc == 0);
 832         oa->o_size = size;
 833         oa->o_blocks = OBD_OBJECT_EOF;
 834         oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 835                       OBD_MD_FLID | OBD_MD_FLGROUP;
 836
 837         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 838         LASSERT(body);
 839         lustre_set_wire_obdo(&body->oa, oa);
 840
 841         /* XXX: capa support? */
 842         /* osc_pack_capa(req, body, capa); */
 843
 844         ptlrpc_request_set_replen(req);
 845
 846         rc = ptlrpc_queue_wait(req);
 847         if (rc)
 848                 CERROR("can't punch object: %d\n", rc);
 849 out:
 850         ptlrpc_req_finished(req);
 851         if (oa)
 852                 OBD_FREE_PTR(oa);
 853         RETURN(rc);
 854 }
 855
 856 int osp_init_precreate(struct osp_device *d)
 857 {
 858         struct l_wait_info       lwi = { 0 };
 859         int                      rc;
 860
 861         ENTRY;
 862
 863         /* initially precreation isn't ready */
 864         d->opd_pre_status = -EAGAIN;
 865         d->opd_pre_next = 1;
 866         d->opd_pre_last_created = 1;
 867         d->opd_pre_reserved = 0;
 868         d->opd_got_disconnected = 1;
 869         d->opd_pre_grow_slow = 0;
 870         d->opd_pre_grow_count = OST_MIN_PRECREATE;
 871         d->opd_pre_min_grow_count = OST_MIN_PRECREATE;
 872         d->opd_pre_max_grow_count = OST_MAX_PRECREATE;
 873
 874         cfs_spin_lock_init(&d->opd_pre_lock);
 875         cfs_waitq_init(&d->opd_pre_waitq);
 876         cfs_waitq_init(&d->opd_pre_user_waitq);
 877         cfs_waitq_init(&d->opd_pre_thread.t_ctl_waitq);
 878
 879         /*
 880          * Initialize statfs-related things
 881          */
 882         d->opd_statfs_maxage = 5; /* default update interval */
 883         d->opd_statfs_fresh_till = cfs_time_shift(-1000);
 884         CDEBUG(D_OTHER, "current %llu, fresh till %llu\n",
 885                (unsigned long long)cfs_time_current(),
 886                (unsigned long long)d->opd_statfs_fresh_till);
 887         cfs_timer_init(&d->opd_statfs_timer, osp_statfs_timer_cb, d);
 888
 889         /*
 890          * start thread handling precreation and statfs updates
 891          */
 892         rc = cfs_create_thread(osp_precreate_thread, d, 0);
 893         if (rc < 0) {
 894                 CERROR("can't start precreate thread %d\n", rc);
 895                 RETURN(rc);
 896         }
 897
 898         l_wait_event(d->opd_pre_thread.t_ctl_waitq,
 899                      osp_precreate_running(d) || osp_precreate_stopped(d),
 900                      &lwi);
 901
 902         RETURN(0);
 903 }
 904
 905 void osp_precreate_fini(struct osp_device *d)
 906 {
 907         struct ptlrpc_thread *thread = &d->opd_pre_thread;
 908
 909         ENTRY;
 910
 911         cfs_timer_disarm(&d->opd_statfs_timer);
 912
 913         thread->t_flags = SVC_STOPPING;
 914         cfs_waitq_signal(&d->opd_pre_waitq);
 915
 916         cfs_wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
 917
 918         EXIT;
 919 }
 920