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                 CDEBUG(D_CACHE, "%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         /*
 152          * no updates till reply
 153          */
 154         cfs_timer_disarm(&d->opd_statfs_timer);
 155         d->opd_statfs_fresh_till = cfs_time_shift(obd_timeout * 1000);
 156         d->opd_statfs_update_in_progress = 1;
 157
 158         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
 159
 160         RETURN(0);
 161 }
 162
 163 /*
 164  * XXX: there might be a case where removed object(s) do not add free
 165  * space (empty object). if the number of such deletions is high, then
 166  * we can start to update statfs too often - a rpc storm
 167  * TODO: some throttling is needed
 168  */
 169 void osp_statfs_need_now(struct osp_device *d)
 170 {
 171         if (!d->opd_statfs_update_in_progress) {
 172                 /*
 173                  * if current status is -ENOSPC (lack of free space on OST)
 174                  * then we should poll OST immediately once object destroy
 175                  * is replied
 176                  */
 177                 d->opd_statfs_fresh_till = cfs_time_shift(-1);
 178                 cfs_timer_disarm(&d->opd_statfs_timer);
 179                 cfs_waitq_signal(&d->opd_pre_waitq);
 180         }
 181 }
 182
 183
 184 /*
 185  * OSP tries to maintain pool of available objects so that calls to create
 186  * objects don't block most of time
 187  *
 188  * each time OSP gets connected to OST, we should start from precreation cleanup
 189  */
 190 static inline int osp_precreate_running(struct osp_device *d)
 191 {
 192         return !!(d->opd_pre_thread.t_flags & SVC_RUNNING);
 193 }
 194
 195 static inline int osp_precreate_stopped(struct osp_device *d)
 196 {
 197         return !!(d->opd_pre_thread.t_flags & SVC_STOPPED);
 198 }
 199
 200 static inline int osp_precreate_near_empty_nolock(struct osp_device *d)
 201 {
 202         int window = d->opd_pre_last_created - d->opd_pre_used_id;
 203
 204         /* don't consider new precreation till OST is healty and
 205          * has free space */
 206         return ((window - d->opd_pre_reserved < d->opd_pre_grow_count / 2) &&
 207                 (d->opd_pre_status == 0));
 208 }
 209
 210 static inline int osp_precreate_near_empty(struct osp_device *d)
 211 {
 212         int rc;
 213
 214         /* XXX: do we really need locking here? */
 215         spin_lock(&d->opd_pre_lock);
 216         rc = osp_precreate_near_empty_nolock(d);
 217         spin_unlock(&d->opd_pre_lock);
 218         return rc;
 219 }
 220
 221 static int osp_precreate_send(struct osp_device *d)
 222 {
 223         struct ptlrpc_request   *req;
 224         struct obd_import       *imp;
 225         struct ost_body         *body;
 226         int                      rc, grow, diff;
 227
 228         ENTRY;
 229
 230         /* don't precreate new objects till OST healthy and has free space */
 231         if (unlikely(d->opd_pre_status)) {
 232                 CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
 233                        d->opd_obd->obd_name, d->opd_pre_status);
 234                 RETURN(0);
 235         }
 236
 237         /*
 238          * if not connection/initialization is compeleted, ignore
 239          */
 240         imp = d->opd_obd->u.cli.cl_import;
 241         LASSERT(imp);
 242
 243         req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
 244         if (req == NULL)
 245                 RETURN(-ENOMEM);
 246         req->rq_request_portal = OST_CREATE_PORTAL;
 247         /* we should not resend create request - anyway we will have delorphan
 248          * and kill these objects */
 249         req->rq_no_delay = req->rq_no_resend = 1;
 250
 251         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
 252         if (rc) {
 253                 ptlrpc_request_free(req);
 254                 RETURN(rc);
 255         }
 256
 257         spin_lock(&d->opd_pre_lock);
 258         if (d->opd_pre_grow_count > d->opd_pre_max_grow_count / 2)
 259                 d->opd_pre_grow_count = d->opd_pre_max_grow_count / 2;
 260         grow = d->opd_pre_grow_count;
 261         spin_unlock(&d->opd_pre_lock);
 262
 263         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 264         LASSERT(body);
 265         body->oa.o_id = d->opd_pre_last_created + grow;
 266         body->oa.o_seq = FID_SEQ_OST_MDT0; /* XXX: support for CMD? */
 267         body->oa.o_valid = OBD_MD_FLGROUP;
 268
 269         ptlrpc_request_set_replen(req);
 270
 271         rc = ptlrpc_queue_wait(req);
 272         if (rc) {
 273                 CERROR("%s: can't precreate: rc = %d\n",
 274                        d->opd_obd->obd_name, rc);
 275                 GOTO(out_req, rc);
 276         }
 277         LASSERT(req->rq_transno == 0);
 278
 279         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 280         if (body == NULL)
 281                 GOTO(out_req, rc = -EPROTO);
 282
 283         CDEBUG(D_HA, "%s: new last_created "LPU64"\n", d->opd_obd->obd_name,
 284                body->oa.o_id);
 285         LASSERT(body->oa.o_id > d->opd_pre_used_id);
 286
 287         diff = body->oa.o_id - d->opd_pre_last_created;
 288
 289         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         spin_unlock(&d->opd_pre_lock);
 303         CDEBUG(D_OTHER, "current precreated pool: %llu-%llu\n",
 304                d->opd_pre_used_id, 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         struct l_wait_info       lwi = { 0 };
 375         int                      update_status = 0;
 376         int                      rc;
 377
 378         ENTRY;
 379
 380         /*
 381          * wait for local recovery to finish, so we can cleanup orphans
 382          * orphans are all objects since "last used" (assigned), but
 383          * there might be objects reserved and in some cases they won't
 384          * be used. we can't cleanup them till we're sure they won't be
 385          * used. also can't we allow new reservations because they may
 386          * end up getting orphans being cleaned up below. so we block
 387          * new reservations and wait till all reserved objects either
 388          * user or released.
 389          */
 390         spin_lock(&d->opd_pre_lock);
 391         d->opd_pre_recovering = 1;
 392         spin_unlock(&d->opd_pre_lock);
 393         /*
 394          * The locking above makes sure the opd_pre_reserved check below will
 395          * catch all osp_precreate_reserve() calls who find
 396          * "!opd_pre_recovering".
 397          */
 398         l_wait_event(d->opd_pre_waitq,
 399                      (!d->opd_pre_reserved && d->opd_recovery_completed) ||
 400                      !osp_precreate_running(d) || d->opd_got_disconnected,
 401                      &lwi);
 402         if (!osp_precreate_running(d) || d->opd_got_disconnected)
 403                 GOTO(out, rc = -EAGAIN);
 404
 405         CDEBUG(D_HA, "%s: going to cleanup orphans since "LPU64"\n",
 406                 d->opd_obd->obd_name, d->opd_last_used_id);
 407
 408         if (d->opd_last_used_id < 2) {
 409                 /* lastid looks strange... ask OST */
 410                 rc = osp_get_lastid_from_ost(d);
 411                 if (rc)
 412                         GOTO(out, rc);
 413         }
 414
 415         imp = d->opd_obd->u.cli.cl_import;
 416         LASSERT(imp);
 417
 418         req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
 419         if (req == NULL)
 420                 GOTO(out, rc = -ENOMEM);
 421
 422         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
 423         if (rc) {
 424                 ptlrpc_request_free(req);
 425                 GOTO(out, rc);
 426         }
 427
 428         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 429         if (body == NULL)
 430                 GOTO(out, rc = -EPROTO);
 431
 432         body->oa.o_flags = OBD_FL_DELORPHAN;
 433         body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
 434         body->oa.o_seq = FID_SEQ_OST_MDT0;
 435
 436         body->oa.o_id = d->opd_last_used_id;
 437
 438         ptlrpc_request_set_replen(req);
 439
 440         /* Don't resend the delorphan req */
 441         req->rq_no_resend = req->rq_no_delay = 1;
 442
 443         rc = ptlrpc_queue_wait(req);
 444         if (rc) {
 445                 update_status = 1;
 446                 GOTO(out, rc);
 447         }
 448
 449         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 450         if (body == NULL)
 451                 GOTO(out, rc = -EPROTO);
 452
 453         /*
 454          * OST provides us with id new pool starts from in body->oa.o_id
 455          */
 456         spin_lock(&d->opd_pre_lock);
 457         if (le64_to_cpu(d->opd_last_used_id) > body->oa.o_id) {
 458                 d->opd_pre_grow_count = OST_MIN_PRECREATE +
 459                                         le64_to_cpu(d->opd_last_used_id) -
 460                                         body->oa.o_id;
 461                 d->opd_pre_last_created = le64_to_cpu(d->opd_last_used_id);
 462         } else {
 463                 d->opd_pre_grow_count = OST_MIN_PRECREATE;
 464                 d->opd_pre_last_created = body->oa.o_id;
 465         }
 466         /*
 467          * This empties the pre-creation pool and effectively blocks any new
 468          * reservations.
 469          */
 470         d->opd_pre_used_id = d->opd_pre_last_created;
 471         d->opd_pre_grow_slow = 0;
 472         spin_unlock(&d->opd_pre_lock);
 473
 474         CDEBUG(D_HA, "%s: Got last_id "LPU64" from OST, last_used is "LPU64
 475                ", pre_used "LPU64"\n", d->opd_obd->obd_name, body->oa.o_id,
 476                le64_to_cpu(d->opd_last_used_id), d->opd_pre_used_id);
 477
 478 out:
 479         if (req)
 480                 ptlrpc_req_finished(req);
 481
 482         d->opd_pre_recovering = 0;
 483
 484         /*
 485          * If rc is zero, the pre-creation window should have been emptied.
 486          * Since waking up the herd would be useless without pre-created
 487          * objects, we defer the signal to osp_precreate_send() in that case.
 488          */
 489         if (rc != 0) {
 490                 if (update_status) {
 491                         CERROR("%s: cannot cleanup orphans: rc = %d\n",
 492                                d->opd_obd->obd_name, rc);
 493                         /* we can't proceed from here, OST seem to
 494                          * be in a bad shape, better to wait for
 495                          * a new instance of the server and repeat
 496                          * from the beginning. notify possible waiters
 497                          * this OSP isn't quite functional yet */
 498                         osp_pre_update_status(d, rc);
 499                 } else {
 500                         cfs_waitq_signal(&d->opd_pre_user_waitq);
 501                 }
 502         }
 503
 504         RETURN(rc);
 505 }
 506
 507 /*
 508  * the function updates current precreation status used: functional or not
 509  *
 510  * rc is a last code from the transport, rc == 0 meaning transport works
 511  * well and users of lod can use objects from this OSP
 512  *
 513  * the status depends on current usage of OST
 514  */
 515 void osp_pre_update_status(struct osp_device *d, int rc)
 516 {
 517         struct obd_statfs       *msfs = &d->opd_statfs;
 518         int                      old = d->opd_pre_status;
 519         __u64                    used;
 520
 521         d->opd_pre_status = rc;
 522         if (rc)
 523                 goto out;
 524
 525         /* Add a bit of hysteresis so this flag isn't continually flapping,
 526          * and ensure that new files don't get extremely fragmented due to
 527          * only a small amount of available space in the filesystem.
 528          * We want to set the NOSPC flag when there is less than ~0.1% free
 529          * and clear it when there is at least ~0.2% free space, so:
 530          *                   avail < ~0.1% max          max = avail + used
 531          *            1025 * avail < avail + used       used = blocks - free
 532          *            1024 * avail < used
 533          *            1024 * avail < blocks - free
 534          *                   avail < ((blocks - free) >> 10)
 535          *
 536          * On very large disk, say 16TB 0.1% will be 16 GB. We don't want to
 537          * lose that amount of space so in those cases we report no space left
 538          * if their is less than 1 GB left.                             */
 539         if (likely(msfs->os_type)) {
 540                 used = min_t(__u64, (msfs->os_blocks - msfs->os_bfree) >> 10,
 541                                     1 << 30);
 542                 if ((msfs->os_ffree < 32) || (msfs->os_bavail < used)) {
 543                         d->opd_pre_status = -ENOSPC;
 544                         if (old != -ENOSPC)
 545                                 CDEBUG(D_INFO, "%s: status: "LPU64" blocks, "
 546                                        LPU64" free, "LPU64" used, "LPU64" "
 547                                        "avail -> %d: rc = %d\n",
 548                                        d->opd_obd->obd_name, msfs->os_blocks,
 549                                        msfs->os_bfree, used, msfs->os_bavail,
 550                                        d->opd_pre_status, rc);
 551                         CDEBUG(D_INFO,
 552                                "non-commited changes: %lu, in progress: %u\n",
 553                                d->opd_syn_changes, d->opd_syn_rpc_in_progress);
 554                 } else if (old == -ENOSPC) {
 555                         d->opd_pre_status = 0;
 556                         d->opd_pre_grow_slow = 0;
 557                         d->opd_pre_grow_count = OST_MIN_PRECREATE;
 558                         cfs_waitq_signal(&d->opd_pre_waitq);
 559                         CDEBUG(D_INFO, "%s: no space: "LPU64" blocks, "LPU64
 560                                " free, "LPU64" used, "LPU64" avail -> %d: "
 561                                "rc = %d\n", d->opd_obd->obd_name,
 562                                msfs->os_blocks, msfs->os_bfree, used,
 563                                msfs->os_bavail, d->opd_pre_status, rc);
 564                 }
 565         }
 566
 567 out:
 568         cfs_waitq_signal(&d->opd_pre_user_waitq);
 569 }
 570
 571 static int osp_precreate_thread(void *_arg)
 572 {
 573         struct osp_device       *d = _arg;
 574         struct ptlrpc_thread    *thread = &d->opd_pre_thread;
 575         struct l_wait_info       lwi = { 0 };
 576         char                     pname[16];
 577         int                      rc;
 578
 579         ENTRY;
 580
 581         sprintf(pname, "osp-pre-%u\n", d->opd_index);
 582         cfs_daemonize(pname);
 583
 584         spin_lock(&d->opd_pre_lock);
 585         thread->t_flags = SVC_RUNNING;
 586         spin_unlock(&d->opd_pre_lock);
 587         cfs_waitq_signal(&thread->t_ctl_waitq);
 588
 589         while (osp_precreate_running(d)) {
 590                 /*
 591                  * need to be connected to OST
 592                  */
 593                 while (osp_precreate_running(d)) {
 594                         l_wait_event(d->opd_pre_waitq,
 595                                      !osp_precreate_running(d) ||
 596                                      d->opd_new_connection, &lwi);
 597
 598                         if (!osp_precreate_running(d))
 599                                 break;
 600
 601                         if (!d->opd_new_connection)
 602                                 continue;
 603
 604                         /* got connected */
 605                         d->opd_new_connection = 0;
 606                         d->opd_got_disconnected = 0;
 607                         break;
 608                 }
 609
 610                 osp_statfs_update(d);
 611
 612                 /*
 613                  * Clean up orphans or recreate missing objects.
 614                  */
 615                 rc = osp_precreate_cleanup_orphans(d);
 616                 if (rc != 0)
 617                         continue;
 618
 619                 /*
 620                  * connected, can handle precreates now
 621                  */
 622                 while (osp_precreate_running(d)) {
 623                         l_wait_event(d->opd_pre_waitq,
 624                                      !osp_precreate_running(d) ||
 625                                      osp_precreate_near_empty(d) ||
 626                                      osp_statfs_need_update(d) ||
 627                                      d->opd_got_disconnected, &lwi);
 628
 629                         if (!osp_precreate_running(d))
 630                                 break;
 631
 632                         /* something happened to the connection
 633                          * have to start from the beginning */
 634                         if (d->opd_got_disconnected)
 635                                 break;
 636
 637                         if (osp_statfs_need_update(d))
 638                                 osp_statfs_update(d);
 639
 640                         if (osp_precreate_near_empty(d)) {
 641                                 rc = osp_precreate_send(d);
 642                                 /* osp_precreate_send() sets opd_pre_status
 643                                  * in case of error, that prevent the using of
 644                                  * failed device. */
 645                                 if (rc != 0 && rc != -ENOSPC &&
 646                                     rc != -ETIMEDOUT && rc != -ENOTCONN)
 647                                         CERROR("%s: cannot precreate objects:"
 648                                                " rc = %d\n",
 649                                                d->opd_obd->obd_name, rc);
 650                         }
 651                 }
 652         }
 653
 654         thread->t_flags = SVC_STOPPED;
 655         cfs_waitq_signal(&thread->t_ctl_waitq);
 656
 657         RETURN(0);
 658 }
 659
 660 static int osp_precreate_ready_condition(struct osp_device *d)
 661 {
 662         __u64 next;
 663
 664         if (d->opd_pre_recovering)
 665                 return 0;
 666
 667         /* ready if got enough precreated objects */
 668         /* we need to wait for others (opd_pre_reserved) and our object (+1) */
 669         next = d->opd_pre_used_id + d->opd_pre_reserved + 1;
 670         if (next <= d->opd_pre_last_created)
 671                 return 1;
 672
 673         /* ready if OST reported no space and no destoys in progress */
 674         if (d->opd_syn_changes + d->opd_syn_rpc_in_progress == 0 &&
 675             d->opd_pre_status != 0)
 676                 return 1;
 677
 678         return 0;
 679 }
 680
 681 static int osp_precreate_timeout_condition(void *data)
 682 {
 683         struct osp_device *d = data;
 684
 685         LCONSOLE_WARN("%s: slow creates, last="LPU64", next="LPU64", "
 686                       "reserved="LPU64", syn_changes=%lu, "
 687                       "syn_rpc_in_progress=%d, status=%d\n",
 688                       d->opd_obd->obd_name, d->opd_pre_last_created,
 689                       d->opd_pre_used_id, d->opd_pre_reserved,
 690                       d->opd_syn_changes, d->opd_syn_rpc_in_progress,
 691                       d->opd_pre_status);
 692
 693         return 1;
 694 }
 695
 696 /*
 697  * called to reserve object in the pool
 698  * return codes:
 699  *  ENOSPC - no space on corresponded OST
 700  *  EAGAIN - precreation is in progress, try later
 701  *  EIO    - no access to OST
 702  */
 703 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d)
 704 {
 705         struct l_wait_info       lwi;
 706         cfs_time_t               expire = cfs_time_shift(obd_timeout);
 707         int                      precreated, rc;
 708         int                      count = 0;
 709
 710         ENTRY;
 711
 712         LASSERT(d->opd_pre_last_created >= d->opd_pre_used_id);
 713
 714         lwi = LWI_TIMEOUT(cfs_time_seconds(obd_timeout),
 715                           osp_precreate_timeout_condition, d);
 716
 717         /*
 718          * wait till:
 719          *  - preallocation is done
 720          *  - no free space expected soon
 721          *  - can't connect to OST for too long (obd_timeout)
 722          */
 723         while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
 724                 rc == -ENODEV) {
 725                 if (unlikely(rc == -ENODEV)) {
 726                         if (cfs_time_aftereq(cfs_time_current(), expire))
 727                                 break;
 728                 }
 729
 730 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 3, 90, 0)
 731                 /*
 732                  * to address Andreas's concern on possible busy-loop
 733                  * between this thread and osp_precreate_send()
 734                  */
 735                 if (unlikely(count++ == 1000)) {
 736                         osp_precreate_timeout_condition(d);
 737                         LBUG();
 738                 }
 739 #endif
 740
 741                 /*
 742                  * increase number of precreations
 743                  */
 744                 if (d->opd_pre_grow_count < d->opd_pre_max_grow_count &&
 745                     d->opd_pre_grow_slow == 0 &&
 746                     (d->opd_pre_last_created - d->opd_pre_used_id <=
 747                      d->opd_pre_grow_count / 4 + 1)) {
 748                         spin_lock(&d->opd_pre_lock);
 749                         d->opd_pre_grow_slow = 1;
 750                         d->opd_pre_grow_count *= 2;
 751                         spin_unlock(&d->opd_pre_lock);
 752                 }
 753
 754                 spin_lock(&d->opd_pre_lock);
 755                 precreated = d->opd_pre_last_created - d->opd_pre_used_id;
 756                 if (precreated > d->opd_pre_reserved &&
 757                     !d->opd_pre_recovering) {
 758                         d->opd_pre_reserved++;
 759                         spin_unlock(&d->opd_pre_lock);
 760                         rc = 0;
 761
 762                         /* XXX: don't wake up if precreation is in progress */
 763                         if (osp_precreate_near_empty_nolock(d))
 764                                 cfs_waitq_signal(&d->opd_pre_waitq);
 765
 766                         break;
 767                 }
 768                 spin_unlock(&d->opd_pre_lock);
 769
 770                 /*
 771                  * all precreated objects have been used and no-space
 772                  * status leave us no chance to succeed very soon
 773                  * but if there is destroy in progress, then we should
 774                  * wait till that is done - some space might be released
 775                  */
 776                 if (unlikely(rc == -ENOSPC)) {
 777                         if (d->opd_syn_changes) {
 778                                 /* force local commit to release space */
 779                                 dt_commit_async(env, d->opd_storage);
 780                         }
 781                         if (d->opd_syn_rpc_in_progress) {
 782                                 /* just wait till destroys are done */
 783                                 /* see l_wait_even() few lines below */
 784                         }
 785                         if (d->opd_syn_changes +
 786                             d->opd_syn_rpc_in_progress == 0) {
 787                                 /* no hope for free space */
 788                                 break;
 789                         }
 790                 }
 791
 792                 /* XXX: don't wake up if precreation is in progress */
 793                 cfs_waitq_signal(&d->opd_pre_waitq);
 794
 795                 l_wait_event(d->opd_pre_user_waitq,
 796                              osp_precreate_ready_condition(d), &lwi);
 797         }
 798
 799         RETURN(rc);
 800 }
 801
 802 /*
 803  * this function relies on reservation made before
 804  */
 805 __u64 osp_precreate_get_id(struct osp_device *d)
 806 {
 807         obd_id objid;
 808
 809         /* grab next id from the pool */
 810         spin_lock(&d->opd_pre_lock);
 811         LASSERT(d->opd_pre_used_id < d->opd_pre_last_created);
 812         objid = ++d->opd_pre_used_id;
 813         d->opd_pre_reserved--;
 814         /*
 815          * last_used_id must be changed along with getting new id otherwise
 816          * we might miscalculate gap causing object loss or leak
 817          */
 818         osp_update_last_id(d, objid);
 819         spin_unlock(&d->opd_pre_lock);
 820
 821         /*
 822          * probably main thread suspended orphan cleanup till
 823          * all reservations are released, see comment in
 824          * osp_precreate_thread() just before orphan cleanup
 825          */
 826         if (unlikely(d->opd_pre_reserved == 0 && d->opd_pre_status))
 827                 cfs_waitq_signal(&d->opd_pre_waitq);
 828
 829         return objid;
 830 }
 831
 832 /*
 833  *
 834  */
 835 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
 836                         __u64 size)
 837 {
 838         struct osp_device       *d = lu2osp_dev(dt->do_lu.lo_dev);
 839         struct ptlrpc_request   *req = NULL;
 840         struct obd_import       *imp;
 841         struct ost_body         *body;
 842         struct obdo             *oa = NULL;
 843         int                      rc;
 844
 845         ENTRY;
 846
 847         imp = d->opd_obd->u.cli.cl_import;
 848         LASSERT(imp);
 849
 850         req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
 851         if (req == NULL)
 852                 RETURN(-ENOMEM);
 853
 854         /* XXX: capa support? */
 855         /* osc_set_capa_size(req, &RMF_CAPA1, capa); */
 856         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
 857         if (rc) {
 858                 ptlrpc_request_free(req);
 859                 RETURN(rc);
 860         }
 861
 862         /*
 863          * XXX: decide how do we do here with resend
 864          * if we don't resend, then client may see wrong file size
 865          * if we do resend, then MDS thread can get stuck for quite long
 866          */
 867         req->rq_no_resend = req->rq_no_delay = 1;
 868
 869         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
 870         ptlrpc_at_set_req_timeout(req);
 871
 872         OBD_ALLOC_PTR(oa);
 873         if (oa == NULL)
 874                 GOTO(out, rc = -ENOMEM);
 875
 876         rc = fid_ostid_pack(lu_object_fid(&dt->do_lu), &oa->o_oi);
 877         LASSERT(rc == 0);
 878         oa->o_size = size;
 879         oa->o_blocks = OBD_OBJECT_EOF;
 880         oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 881                       OBD_MD_FLID | OBD_MD_FLGROUP;
 882
 883         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 884         LASSERT(body);
 885         lustre_set_wire_obdo(&body->oa, oa);
 886
 887         /* XXX: capa support? */
 888         /* osc_pack_capa(req, body, capa); */
 889
 890         ptlrpc_request_set_replen(req);
 891
 892         rc = ptlrpc_queue_wait(req);
 893         if (rc)
 894                 CERROR("can't punch object: %d\n", rc);
 895 out:
 896         ptlrpc_req_finished(req);
 897         if (oa)
 898                 OBD_FREE_PTR(oa);
 899         RETURN(rc);
 900 }
 901
 902 int osp_init_precreate(struct osp_device *d)
 903 {
 904         struct l_wait_info       lwi = { 0 };
 905         int                      rc;
 906
 907         ENTRY;
 908
 909         /* initially precreation isn't ready */
 910         d->opd_pre_status = -EAGAIN;
 911         d->opd_pre_used_id = 0;
 912         d->opd_pre_last_created = 0;
 913         d->opd_pre_reserved = 0;
 914         d->opd_got_disconnected = 1;
 915         d->opd_pre_grow_slow = 0;
 916         d->opd_pre_grow_count = OST_MIN_PRECREATE;
 917         d->opd_pre_min_grow_count = OST_MIN_PRECREATE;
 918         d->opd_pre_max_grow_count = OST_MAX_PRECREATE;
 919
 920         spin_lock_init(&d->opd_pre_lock);
 921         cfs_waitq_init(&d->opd_pre_waitq);
 922         cfs_waitq_init(&d->opd_pre_user_waitq);
 923         cfs_waitq_init(&d->opd_pre_thread.t_ctl_waitq);
 924
 925         /*
 926          * Initialize statfs-related things
 927          */
 928         d->opd_statfs_maxage = 5; /* default update interval */
 929         d->opd_statfs_fresh_till = cfs_time_shift(-1000);
 930         CDEBUG(D_OTHER, "current %llu, fresh till %llu\n",
 931                (unsigned long long)cfs_time_current(),
 932                (unsigned long long)d->opd_statfs_fresh_till);
 933         cfs_timer_init(&d->opd_statfs_timer, osp_statfs_timer_cb, d);
 934
 935         /*
 936          * start thread handling precreation and statfs updates
 937          */
 938         rc = cfs_create_thread(osp_precreate_thread, d, 0);
 939         if (rc < 0) {
 940                 CERROR("can't start precreate thread %d\n", rc);
 941                 RETURN(rc);
 942         }
 943
 944         l_wait_event(d->opd_pre_thread.t_ctl_waitq,
 945                      osp_precreate_running(d) || osp_precreate_stopped(d),
 946                      &lwi);
 947
 948         RETURN(0);
 949 }
 950
 951 void osp_precreate_fini(struct osp_device *d)
 952 {
 953         struct ptlrpc_thread *thread = &d->opd_pre_thread;
 954
 955         ENTRY;
 956
 957         cfs_timer_disarm(&d->opd_statfs_timer);
 958
 959         thread->t_flags = SVC_STOPPING;
 960         cfs_waitq_signal(&d->opd_pre_waitq);
 961
 962         cfs_wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
 963
 964         EXIT;
 965 }
 966