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) 2012, 2013, Intel Corporation.
  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  * Author: Di Wang <di.wang@intel.com>
  43  */
  44
  45 #define DEBUG_SUBSYSTEM S_MDS
  46
  47 #include "osp_internal.h"
  48
  49 /*
  50  * there are two specific states to take care about:
  51  *
  52  * = import is disconnected =
  53  *
  54  * = import is inactive =
  55  *   in this case osp_declare_object_create() returns an error
  56  *
  57  */
  58
  59 /*
  60  * statfs
  61  */
  62 static inline int osp_statfs_need_update(struct osp_device *d)
  63 {
  64         return !cfs_time_before(cfs_time_current(),
  65                                 d->opd_statfs_fresh_till);
  66 }
  67
  68 /*
  69  * OSP tries to maintain pool of available objects so that calls to create
  70  * objects don't block most of time
  71  *
  72  * each time OSP gets connected to OST, we should start from precreation cleanup
  73  */
  74 static inline bool osp_precreate_running(struct osp_device *d)
  75 {
  76         return !!(d->opd_pre_thread.t_flags & SVC_RUNNING);
  77 }
  78
  79 static inline bool osp_precreate_stopped(struct osp_device *d)
  80 {
  81         return !!(d->opd_pre_thread.t_flags & SVC_STOPPED);
  82 }
  83
  84 static void osp_statfs_timer_cb(unsigned long _d)
  85 {
  86         struct osp_device *d = (struct osp_device *) _d;
  87
  88         LASSERT(d);
  89         if (d->opd_pre != NULL && osp_precreate_running(d))
  90                 wake_up(&d->opd_pre_waitq);
  91 }
  92
  93 static int osp_statfs_interpret(const struct lu_env *env,
  94                                 struct ptlrpc_request *req,
  95                                 union ptlrpc_async_args *aa, int rc)
  96 {
  97         struct obd_import       *imp = req->rq_import;
  98         struct obd_statfs       *msfs;
  99         struct osp_device       *d;
 100
 101         ENTRY;
 102
 103         aa = ptlrpc_req_async_args(req);
 104         d = aa->pointer_arg[0];
 105         LASSERT(d);
 106
 107         if (rc != 0)
 108                 GOTO(out, rc);
 109
 110         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
 111         if (msfs == NULL)
 112                 GOTO(out, rc = -EPROTO);
 113
 114         d->opd_statfs = *msfs;
 115
 116         osp_pre_update_status(d, rc);
 117
 118         /* schedule next update */
 119         d->opd_statfs_fresh_till = cfs_time_shift(d->opd_statfs_maxage);
 120         cfs_timer_arm(&d->opd_statfs_timer, d->opd_statfs_fresh_till);
 121         d->opd_statfs_update_in_progress = 0;
 122
 123         CDEBUG(D_CACHE, "updated statfs %p\n", d);
 124
 125         RETURN(0);
 126 out:
 127         /* couldn't update statfs, try again as soon as possible */
 128         if (d->opd_pre != NULL && osp_precreate_running(d))
 129                 wake_up(&d->opd_pre_waitq);
 130
 131         if (req->rq_import_generation == imp->imp_generation)
 132                 CDEBUG(D_CACHE, "%s: couldn't update statfs: rc = %d\n",
 133                        d->opd_obd->obd_name, rc);
 134         RETURN(rc);
 135 }
 136
 137 static int osp_statfs_update(struct osp_device *d)
 138 {
 139         struct ptlrpc_request   *req;
 140         struct obd_import       *imp;
 141         union ptlrpc_async_args *aa;
 142         int                      rc;
 143
 144         ENTRY;
 145
 146         CDEBUG(D_CACHE, "going to update statfs\n");
 147
 148         imp = d->opd_obd->u.cli.cl_import;
 149         LASSERT(imp);
 150
 151         req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
 152         if (req == NULL)
 153                 RETURN(-ENOMEM);
 154
 155         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
 156         if (rc) {
 157                 ptlrpc_request_free(req);
 158                 RETURN(rc);
 159         }
 160         ptlrpc_request_set_replen(req);
 161         req->rq_request_portal = OST_CREATE_PORTAL;
 162         ptlrpc_at_set_req_timeout(req);
 163
 164         req->rq_interpret_reply = (ptlrpc_interpterer_t)osp_statfs_interpret;
 165         aa = ptlrpc_req_async_args(req);
 166         aa->pointer_arg[0] = d;
 167
 168         /*
 169          * no updates till reply
 170          */
 171         cfs_timer_disarm(&d->opd_statfs_timer);
 172         d->opd_statfs_fresh_till = cfs_time_shift(obd_timeout * 1000);
 173         d->opd_statfs_update_in_progress = 1;
 174
 175         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
 176
 177         RETURN(0);
 178 }
 179
 180 /*
 181  * XXX: there might be a case where removed object(s) do not add free
 182  * space (empty object). if the number of such deletions is high, then
 183  * we can start to update statfs too often - a rpc storm
 184  * TODO: some throttling is needed
 185  */
 186 void osp_statfs_need_now(struct osp_device *d)
 187 {
 188         if (!d->opd_statfs_update_in_progress) {
 189                 /*
 190                  * if current status is -ENOSPC (lack of free space on OST)
 191                  * then we should poll OST immediately once object destroy
 192                  * is replied
 193                  */
 194                 d->opd_statfs_fresh_till = cfs_time_shift(-1);
 195                 cfs_timer_disarm(&d->opd_statfs_timer);
 196                 wake_up(&d->opd_pre_waitq);
 197         }
 198 }
 199
 200 static inline int osp_objs_precreated(const struct lu_env *env,
 201                                       struct osp_device *osp)
 202 {
 203         return osp_fid_diff(&osp->opd_pre_last_created_fid,
 204                             &osp->opd_pre_used_fid);
 205 }
 206
 207 static inline int osp_precreate_near_empty_nolock(const struct lu_env *env,
 208                                                   struct osp_device *d)
 209 {
 210         int window = osp_objs_precreated(env, d);
 211
 212         /* don't consider new precreation till OST is healty and
 213          * has free space */
 214         return ((window - d->opd_pre_reserved < d->opd_pre_grow_count / 2) &&
 215                 (d->opd_pre_status == 0));
 216 }
 217
 218 static inline int osp_precreate_near_empty(const struct lu_env *env,
 219                                            struct osp_device *d)
 220 {
 221         int rc;
 222
 223         /* XXX: do we really need locking here? */
 224         spin_lock(&d->opd_pre_lock);
 225         rc = osp_precreate_near_empty_nolock(env, d);
 226         spin_unlock(&d->opd_pre_lock);
 227         return rc;
 228 }
 229
 230 static inline int osp_create_end_seq(const struct lu_env *env,
 231                                      struct osp_device *osp)
 232 {
 233         struct lu_fid *fid = &osp->opd_pre_used_fid;
 234         int rc;
 235
 236         spin_lock(&osp->opd_pre_lock);
 237         rc = osp_fid_end_seq(env, fid);
 238         spin_unlock(&osp->opd_pre_lock);
 239         return rc;
 240 }
 241
 242 /**
 243  * Write fid into last_oid/last_seq file.
 244  **/
 245 int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp,
 246                                  struct lu_fid *fid, int sync)
 247 {
 248         struct osp_thread_info  *oti = osp_env_info(env);
 249         struct lu_buf      *lb_oid = &oti->osi_lb;
 250         struct lu_buf      *lb_oseq = &oti->osi_lb2;
 251         loff_t             oid_off;
 252         loff_t             oseq_off;
 253         struct thandle    *th;
 254         int                   rc;
 255         ENTRY;
 256
 257         /* Note: through f_oid is only 32bits, it will also write
 258          * 64 bits for oid to keep compatiblity with the previous
 259          * version. */
 260         lb_oid->lb_buf = &fid->f_oid;
 261         lb_oid->lb_len = sizeof(obd_id);
 262         oid_off = sizeof(obd_id) * osp->opd_index;
 263
 264         lb_oseq->lb_buf = &fid->f_seq;
 265         lb_oseq->lb_len = sizeof(obd_id);
 266         oseq_off = sizeof(obd_id) * osp->opd_index;
 267
 268         th = dt_trans_create(env, osp->opd_storage);
 269         if (IS_ERR(th))
 270                 RETURN(PTR_ERR(th));
 271
 272         th->th_sync |= sync;
 273         rc = dt_declare_record_write(env, osp->opd_last_used_oid_file,
 274                                      lb_oid, oid_off, th);
 275         if (rc != 0)
 276                 GOTO(out, rc);
 277
 278         rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
 279                                      lb_oseq, oseq_off, th);
 280         if (rc != 0)
 281                 GOTO(out, rc);
 282
 283         rc = dt_trans_start_local(env, osp->opd_storage, th);
 284         if (rc != 0)
 285                 GOTO(out, rc);
 286
 287         rc = dt_record_write(env, osp->opd_last_used_oid_file, lb_oid,
 288                              &oid_off, th);
 289         if (rc != 0) {
 290                 CERROR("%s: can not write to last seq file: rc = %d\n",
 291                         osp->opd_obd->obd_name, rc);
 292                 GOTO(out, rc);
 293         }
 294         rc = dt_record_write(env, osp->opd_last_used_seq_file, lb_oseq,
 295                              &oseq_off, th);
 296         if (rc) {
 297                 CERROR("%s: can not write to last seq file: rc = %d\n",
 298                         osp->opd_obd->obd_name, rc);
 299                 GOTO(out, rc);
 300         }
 301 out:
 302         dt_trans_stop(env, osp->opd_storage, th);
 303         RETURN(rc);
 304 }
 305
 306 int osp_precreate_rollover_new_seq(struct lu_env *env, struct osp_device *osp)
 307 {
 308         struct lu_fid   *fid = &osp_env_info(env)->osi_fid;
 309         struct lu_fid   *last_fid = &osp->opd_last_used_fid;
 310         int             rc;
 311         ENTRY;
 312
 313         rc = seq_client_get_seq(env, osp->opd_obd->u.cli.cl_seq, &fid->f_seq);
 314         if (rc != 0) {
 315                 CERROR("%s: alloc fid error: rc = %d\n",
 316                        osp->opd_obd->obd_name, rc);
 317                 RETURN(rc);
 318         }
 319
 320         fid->f_oid = 1;
 321         fid->f_ver = 0;
 322         LASSERTF(fid_seq(fid) != fid_seq(last_fid),
 323                  "fid "DFID", last_fid "DFID"\n", PFID(fid),
 324                  PFID(last_fid));
 325
 326         rc = osp_write_last_oid_seq_files(env, osp, fid, 1);
 327         if (rc != 0) {
 328                 CERROR("%s: Can not update oid/seq file: rc = %d\n",
 329                        osp->opd_obd->obd_name, rc);
 330                 RETURN(rc);
 331         }
 332
 333         LCONSOLE_INFO("%s: update sequence from "LPX64" to "LPX64"\n",
 334                       osp->opd_obd->obd_name, fid_seq(last_fid),
 335                       fid_seq(fid));
 336         /* Update last_xxx to the new seq */
 337         spin_lock(&osp->opd_pre_lock);
 338         osp->opd_last_used_fid = *fid;
 339         osp->opd_gap_start_fid = *fid;
 340         osp->opd_pre_used_fid = *fid;
 341         osp->opd_pre_last_created_fid = *fid;
 342         spin_unlock(&osp->opd_pre_lock);
 343
 344         RETURN(rc);
 345 }
 346
 347 /**
 348  * alloc fids for precreation.
 349  * rc = 0 Success, @grow is the count of real allocation.
 350  * rc = 1 Current seq is used up.
 351  * rc < 0 Other error.
 352  **/
 353 static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp,
 354                               struct lu_fid *fid, int *grow)
 355 {
 356         struct osp_thread_info  *osi = osp_env_info(env);
 357         __u64                   end;
 358         int                     i = 0;
 359
 360         if (fid_is_idif(fid)) {
 361                 struct lu_fid   *last_fid;
 362                 struct ost_id   *oi = &osi->osi_oi;
 363
 364                 spin_lock(&osp->opd_pre_lock);
 365                 last_fid = &osp->opd_pre_last_created_fid;
 366                 fid_to_ostid(last_fid, oi);
 367                 end = min(ostid_id(oi) + *grow, IDIF_MAX_OID);
 368                 *grow = end - ostid_id(oi);
 369                 ostid_set_id(oi, ostid_id(oi) + *grow);
 370                 spin_unlock(&osp->opd_pre_lock);
 371
 372                 if (*grow == 0)
 373                         return 1;
 374
 375                 ostid_to_fid(fid, oi, osp->opd_index);
 376                 return 0;
 377         }
 378
 379         spin_lock(&osp->opd_pre_lock);
 380         *fid = osp->opd_pre_last_created_fid;
 381         end = fid->f_oid;
 382         end = min((end + *grow), (__u64)LUSTRE_DATA_SEQ_MAX_WIDTH);
 383         *grow = end - fid->f_oid;
 384         fid->f_oid += end - fid->f_oid;
 385         spin_unlock(&osp->opd_pre_lock);
 386
 387         CDEBUG(D_INFO, "Expect %d, actual %d ["DFID" -- "DFID"]\n",
 388                *grow, i, PFID(fid), PFID(&osp->opd_pre_last_created_fid));
 389
 390         return *grow > 0 ? 0 : 1;
 391 }
 392
 393 static int osp_precreate_send(const struct lu_env *env, struct osp_device *d)
 394 {
 395         struct osp_thread_info  *oti = osp_env_info(env);
 396         struct ptlrpc_request   *req;
 397         struct obd_import       *imp;
 398         struct ost_body         *body;
 399         int                      rc, grow, diff;
 400         struct lu_fid           *fid = &oti->osi_fid;
 401         ENTRY;
 402
 403         /* don't precreate new objects till OST healthy and has free space */
 404         if (unlikely(d->opd_pre_status)) {
 405                 CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
 406                        d->opd_obd->obd_name, d->opd_pre_status);
 407                 RETURN(0);
 408         }
 409
 410         /*
 411          * if not connection/initialization is compeleted, ignore
 412          */
 413         imp = d->opd_obd->u.cli.cl_import;
 414         LASSERT(imp);
 415
 416         req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
 417         if (req == NULL)
 418                 RETURN(-ENOMEM);
 419         req->rq_request_portal = OST_CREATE_PORTAL;
 420         /* we should not resend create request - anyway we will have delorphan
 421          * and kill these objects */
 422         req->rq_no_delay = req->rq_no_resend = 1;
 423
 424         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
 425         if (rc) {
 426                 ptlrpc_request_free(req);
 427                 RETURN(rc);
 428         }
 429
 430         spin_lock(&d->opd_pre_lock);
 431         if (d->opd_pre_grow_count > d->opd_pre_max_grow_count / 2)
 432                 d->opd_pre_grow_count = d->opd_pre_max_grow_count / 2;
 433         grow = d->opd_pre_grow_count;
 434         spin_unlock(&d->opd_pre_lock);
 435
 436         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 437         LASSERT(body);
 438
 439         *fid = d->opd_pre_last_created_fid;
 440         rc = osp_precreate_fids(env, d, fid, &grow);
 441         if (rc == 1) {
 442                 /* Current seq has been used up*/
 443                 if (!osp_is_fid_client(d)) {
 444                         osp_pre_update_status(d, -ENOSPC);
 445                         rc = -ENOSPC;
 446                 }
 447                 wake_up(&d->opd_pre_waitq);
 448                 GOTO(out_req, rc);
 449         }
 450
 451         if (!osp_is_fid_client(d)) {
 452                 /* Non-FID client will always send seq 0 because of
 453                  * compatiblity */
 454                 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
 455                 fid->f_seq = 0;
 456         }
 457
 458         fid_to_ostid(fid, &body->oa.o_oi);
 459         body->oa.o_valid = OBD_MD_FLGROUP;
 460
 461         ptlrpc_request_set_replen(req);
 462
 463         rc = ptlrpc_queue_wait(req);
 464         if (rc) {
 465                 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
 466                        rc);
 467                 GOTO(out_req, rc);
 468         }
 469         LASSERT(req->rq_transno == 0);
 470
 471         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 472         if (body == NULL)
 473                 GOTO(out_req, rc = -EPROTO);
 474
 475         ostid_to_fid(fid, &body->oa.o_oi, d->opd_index);
 476         if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) {
 477                 CERROR("%s: precreate fid "DFID" < local used fid "DFID
 478                        ": rc = %d\n", d->opd_obd->obd_name,
 479                        PFID(fid), PFID(&d->opd_pre_used_fid), -ESTALE);
 480                 GOTO(out_req, rc = -ESTALE);
 481         }
 482
 483         diff = osp_fid_diff(fid, &d->opd_pre_last_created_fid);
 484
 485         spin_lock(&d->opd_pre_lock);
 486         if (diff < grow) {
 487                 /* the OST has not managed to create all the
 488                  * objects we asked for */
 489                 d->opd_pre_grow_count = max(diff, OST_MIN_PRECREATE);
 490                 d->opd_pre_grow_slow = 1;
 491         } else {
 492                 /* the OST is able to keep up with the work,
 493                  * we could consider increasing grow_count
 494                  * next time if needed */
 495                 d->opd_pre_grow_slow = 0;
 496         }
 497
 498         d->opd_pre_last_created_fid = *fid;
 499         spin_unlock(&d->opd_pre_lock);
 500
 501         CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n",
 502                d->opd_obd->obd_name, PFID(&d->opd_pre_used_fid),
 503                PFID(&d->opd_pre_last_created_fid));
 504 out_req:
 505         /* now we can wakeup all users awaiting for objects */
 506         osp_pre_update_status(d, rc);
 507         wake_up(&d->opd_pre_user_waitq);
 508
 509         ptlrpc_req_finished(req);
 510         RETURN(rc);
 511 }
 512
 513 static int osp_get_lastfid_from_ost(const struct lu_env *env,
 514                                     struct osp_device *d)
 515 {
 516         struct ptlrpc_request   *req = NULL;
 517         struct obd_import       *imp;
 518         struct lu_fid           *last_fid;
 519         char                    *tmp;
 520         int                     rc;
 521         ENTRY;
 522
 523         imp = d->opd_obd->u.cli.cl_import;
 524         LASSERT(imp);
 525
 526         req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
 527         if (req == NULL)
 528                 RETURN(-ENOMEM);
 529
 530         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY, RCL_CLIENT,
 531                              sizeof(KEY_LAST_FID));
 532
 533         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
 534         if (rc) {
 535                 ptlrpc_request_free(req);
 536                 RETURN(rc);
 537         }
 538
 539         tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
 540         memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
 541
 542         req->rq_no_delay = req->rq_no_resend = 1;
 543         last_fid = req_capsule_client_get(&req->rq_pill, &RMF_FID);
 544         fid_cpu_to_le(last_fid, &d->opd_last_used_fid);
 545
 546         ptlrpc_request_set_replen(req);
 547
 548         rc = ptlrpc_queue_wait(req);
 549         if (rc) {
 550                 /* bad-bad OST.. let sysadm sort this out */
 551                 if (rc == -ENOTSUPP) {
 552                         CERROR("%s: server does not support FID: rc = %d\n",
 553                                d->opd_obd->obd_name, -ENOTSUPP);
 554                 }
 555                 ptlrpc_set_import_active(imp, 0);
 556                 GOTO(out, rc);
 557         }
 558
 559         last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
 560         if (last_fid == NULL) {
 561                 CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name);
 562                 GOTO(out, rc = -EPROTO);
 563         }
 564
 565         if (!fid_is_sane(last_fid)) {
 566                 CERROR("%s: Got insane last_fid "DFID"\n",
 567                        d->opd_obd->obd_name, PFID(last_fid));
 568                 GOTO(out, rc = -EPROTO);
 569         }
 570
 571         /* Only update the last used fid, if the OST has objects for
 572          * this sequence, i.e. fid_oid > 0 */
 573         if (fid_oid(last_fid) > 0)
 574                 d->opd_last_used_fid = *last_fid;
 575
 576         CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
 577                PFID(last_fid));
 578
 579 out:
 580         ptlrpc_req_finished(req);
 581         RETURN(rc);
 582 }
 583
 584 /**
 585  * asks OST to clean precreate orphans
 586  * and gets next id for new objects
 587  */
 588 static int osp_precreate_cleanup_orphans(struct lu_env *env,
 589                                          struct osp_device *d)
 590 {
 591         struct osp_thread_info  *osi = osp_env_info(env);
 592         struct lu_fid           *last_fid = &osi->osi_fid;
 593         struct ptlrpc_request   *req = NULL;
 594         struct obd_import       *imp;
 595         struct ost_body         *body;
 596         struct l_wait_info       lwi = { 0 };
 597         int                      update_status = 0;
 598         int                      rc;
 599         int                      diff;
 600
 601         ENTRY;
 602
 603         /*
 604          * wait for local recovery to finish, so we can cleanup orphans
 605          * orphans are all objects since "last used" (assigned), but
 606          * there might be objects reserved and in some cases they won't
 607          * be used. we can't cleanup them till we're sure they won't be
 608          * used. also can't we allow new reservations because they may
 609          * end up getting orphans being cleaned up below. so we block
 610          * new reservations and wait till all reserved objects either
 611          * user or released.
 612          */
 613         spin_lock(&d->opd_pre_lock);
 614         d->opd_pre_recovering = 1;
 615         spin_unlock(&d->opd_pre_lock);
 616         /*
 617          * The locking above makes sure the opd_pre_reserved check below will
 618          * catch all osp_precreate_reserve() calls who find
 619          * "!opd_pre_recovering".
 620          */
 621         l_wait_event(d->opd_pre_waitq,
 622                      (!d->opd_pre_reserved && d->opd_recovery_completed) ||
 623                      !osp_precreate_running(d) || d->opd_got_disconnected,
 624                      &lwi);
 625         if (!osp_precreate_running(d) || d->opd_got_disconnected)
 626                 GOTO(out, rc = -EAGAIN);
 627
 628         CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
 629                d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
 630
 631         *last_fid = d->opd_last_used_fid;
 632         /* The OSP should already get the valid seq now */
 633         LASSERT(!fid_is_zero(last_fid));
 634         if (fid_oid(&d->opd_last_used_fid) < 2) {
 635                 /* lastfid looks strange... ask OST */
 636                 rc = osp_get_lastfid_from_ost(env, d);
 637                 if (rc)
 638                         GOTO(out, rc);
 639         }
 640
 641         imp = d->opd_obd->u.cli.cl_import;
 642         LASSERT(imp);
 643
 644         req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
 645         if (req == NULL)
 646                 GOTO(out, rc = -ENOMEM);
 647
 648         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
 649         if (rc) {
 650                 ptlrpc_request_free(req);
 651                 req = NULL;
 652                 GOTO(out, rc);
 653         }
 654
 655         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 656         if (body == NULL)
 657                 GOTO(out, rc = -EPROTO);
 658
 659         body->oa.o_flags = OBD_FL_DELORPHAN;
 660         body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
 661
 662         fid_to_ostid(&d->opd_last_used_fid, &body->oa.o_oi);
 663
 664         ptlrpc_request_set_replen(req);
 665
 666         /* Don't resend the delorphan req */
 667         req->rq_no_resend = req->rq_no_delay = 1;
 668
 669         rc = ptlrpc_queue_wait(req);
 670         if (rc) {
 671                 update_status = 1;
 672                 GOTO(out, rc);
 673         }
 674
 675         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 676         if (body == NULL)
 677                 GOTO(out, rc = -EPROTO);
 678
 679         /*
 680          * OST provides us with id new pool starts from in body->oa.o_id
 681          */
 682         ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
 683
 684         spin_lock(&d->opd_pre_lock);
 685         diff = osp_fid_diff(&d->opd_last_used_fid, last_fid);
 686         if (diff > 0) {
 687                 d->opd_pre_grow_count = OST_MIN_PRECREATE + diff;
 688                 d->opd_pre_last_created_fid = d->opd_last_used_fid;
 689         } else {
 690                 d->opd_pre_grow_count = OST_MIN_PRECREATE;
 691                 d->opd_pre_last_created_fid = *last_fid;
 692         }
 693         /*
 694          * This empties the pre-creation pool and effectively blocks any new
 695          * reservations.
 696          */
 697         LASSERT(fid_oid(&d->opd_pre_last_created_fid) <=
 698                 LUSTRE_DATA_SEQ_MAX_WIDTH);
 699         d->opd_pre_used_fid = d->opd_pre_last_created_fid;
 700         d->opd_pre_grow_slow = 0;
 701         spin_unlock(&d->opd_pre_lock);
 702
 703         CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
 704                "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
 705                PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
 706 out:
 707         if (req)
 708                 ptlrpc_req_finished(req);
 709
 710         spin_lock(&d->opd_pre_lock);
 711         d->opd_pre_recovering = 0;
 712         spin_unlock(&d->opd_pre_lock);
 713
 714         /*
 715          * If rc is zero, the pre-creation window should have been emptied.
 716          * Since waking up the herd would be useless without pre-created
 717          * objects, we defer the signal to osp_precreate_send() in that case.
 718          */
 719         if (rc != 0) {
 720                 if (update_status) {
 721                         CERROR("%s: cannot cleanup orphans: rc = %d\n",
 722                                d->opd_obd->obd_name, rc);
 723                         /* we can't proceed from here, OST seem to
 724                          * be in a bad shape, better to wait for
 725                          * a new instance of the server and repeat
 726                          * from the beginning. notify possible waiters
 727                          * this OSP isn't quite functional yet */
 728                         osp_pre_update_status(d, rc);
 729                 } else {
 730                         wake_up(&d->opd_pre_user_waitq);
 731                 }
 732         }
 733
 734         RETURN(rc);
 735 }
 736
 737 /*
 738  * the function updates current precreation status used: functional or not
 739  *
 740  * rc is a last code from the transport, rc == 0 meaning transport works
 741  * well and users of lod can use objects from this OSP
 742  *
 743  * the status depends on current usage of OST
 744  */
 745 void osp_pre_update_status(struct osp_device *d, int rc)
 746 {
 747         struct obd_statfs       *msfs = &d->opd_statfs;
 748         int                      old = d->opd_pre_status;
 749         __u64                    used;
 750
 751         d->opd_pre_status = rc;
 752         if (rc)
 753                 goto out;
 754
 755         /* Add a bit of hysteresis so this flag isn't continually flapping,
 756          * and ensure that new files don't get extremely fragmented due to
 757          * only a small amount of available space in the filesystem.
 758          * We want to set the NOSPC flag when there is less than ~0.1% free
 759          * and clear it when there is at least ~0.2% free space, so:
 760          *                   avail < ~0.1% max          max = avail + used
 761          *            1025 * avail < avail + used       used = blocks - free
 762          *            1024 * avail < used
 763          *            1024 * avail < blocks - free
 764          *                   avail < ((blocks - free) >> 10)
 765          *
 766          * On very large disk, say 16TB 0.1% will be 16 GB. We don't want to
 767          * lose that amount of space so in those cases we report no space left
 768          * if their is less than 1 GB left.                             */
 769         if (likely(msfs->os_type)) {
 770                 used = min_t(__u64, (msfs->os_blocks - msfs->os_bfree) >> 10,
 771                                     1 << 30);
 772                 if ((msfs->os_ffree < 32) || (msfs->os_bavail < used)) {
 773                         d->opd_pre_status = -ENOSPC;
 774                         if (old != -ENOSPC)
 775                                 CDEBUG(D_INFO, "%s: status: "LPU64" blocks, "
 776                                        LPU64" free, "LPU64" used, "LPU64" "
 777                                        "avail -> %d: rc = %d\n",
 778                                        d->opd_obd->obd_name, msfs->os_blocks,
 779                                        msfs->os_bfree, used, msfs->os_bavail,
 780                                        d->opd_pre_status, rc);
 781                         CDEBUG(D_INFO,
 782                                "non-commited changes: %lu, in progress: %u\n",
 783                                d->opd_syn_changes, d->opd_syn_rpc_in_progress);
 784                 } else if (old == -ENOSPC) {
 785                         d->opd_pre_status = 0;
 786                         spin_lock(&d->opd_pre_lock);
 787                         d->opd_pre_grow_slow = 0;
 788                         d->opd_pre_grow_count = OST_MIN_PRECREATE;
 789                         spin_unlock(&d->opd_pre_lock);
 790                         wake_up(&d->opd_pre_waitq);
 791                         CDEBUG(D_INFO, "%s: no space: "LPU64" blocks, "LPU64
 792                                " free, "LPU64" used, "LPU64" avail -> %d: "
 793                                "rc = %d\n", d->opd_obd->obd_name,
 794                                msfs->os_blocks, msfs->os_bfree, used,
 795                                msfs->os_bavail, d->opd_pre_status, rc);
 796                 }
 797         }
 798
 799 out:
 800         wake_up(&d->opd_pre_user_waitq);
 801 }
 802
 803 int osp_init_pre_fid(struct osp_device *osp)
 804 {
 805         struct lu_env           env;
 806         struct osp_thread_info  *osi;
 807         struct lu_client_seq    *cli_seq;
 808         struct lu_fid           *last_fid;
 809         int                     rc;
 810         ENTRY;
 811
 812         LASSERT(osp->opd_pre != NULL);
 813
 814         /* Return if last_used fid has been initialized */
 815         if (!fid_is_zero(&osp->opd_last_used_fid))
 816                 RETURN(0);
 817
 818         rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
 819         if (rc) {
 820                 CERROR("%s: init env error: rc = %d\n",
 821                        osp->opd_obd->obd_name, rc);
 822                 RETURN(rc);
 823         }
 824
 825         osi = osp_env_info(&env);
 826         last_fid = &osi->osi_fid;
 827         fid_zero(last_fid);
 828         /* For a freshed fs, it will allocate a new sequence first */
 829         if (osp_is_fid_client(osp) && osp->opd_group != 0) {
 830                 cli_seq = osp->opd_obd->u.cli.cl_seq;
 831                 rc = seq_client_get_seq(&env, cli_seq, &last_fid->f_seq);
 832                 if (rc != 0) {
 833                         CERROR("%s: alloc fid error: rc = %d\n",
 834                                osp->opd_obd->obd_name, rc);
 835                         GOTO(out, rc);
 836                 }
 837         } else {
 838                 last_fid->f_seq = fid_idif_seq(0, osp->opd_index);
 839         }
 840         last_fid->f_oid = 1;
 841         last_fid->f_ver = 0;
 842
 843         spin_lock(&osp->opd_pre_lock);
 844         osp->opd_last_used_fid = *last_fid;
 845         osp->opd_pre_used_fid = *last_fid;
 846         osp->opd_pre_last_created_fid = *last_fid;
 847         spin_unlock(&osp->opd_pre_lock);
 848         rc = osp_write_last_oid_seq_files(&env, osp, last_fid, 1);
 849         if (rc != 0) {
 850                 CERROR("%s: write fid error: rc = %d\n",
 851                        osp->opd_obd->obd_name, rc);
 852                 GOTO(out, rc);
 853         }
 854 out:
 855         lu_env_fini(&env);
 856         RETURN(rc);
 857 }
 858
 859 static int osp_precreate_thread(void *_arg)
 860 {
 861         struct osp_device       *d = _arg;
 862         struct ptlrpc_thread    *thread = &d->opd_pre_thread;
 863         struct l_wait_info       lwi = { 0 };
 864         struct lu_env            env;
 865         int                      rc;
 866
 867         ENTRY;
 868
 869         rc = lu_env_init(&env, d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
 870         if (rc) {
 871                 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
 872                        rc);
 873                 RETURN(rc);
 874         }
 875
 876         spin_lock(&d->opd_pre_lock);
 877         thread->t_flags = SVC_RUNNING;
 878         spin_unlock(&d->opd_pre_lock);
 879         wake_up(&thread->t_ctl_waitq);
 880
 881         while (osp_precreate_running(d)) {
 882                 /*
 883                  * need to be connected to OST
 884                  */
 885                 while (osp_precreate_running(d)) {
 886                         l_wait_event(d->opd_pre_waitq,
 887                                      !osp_precreate_running(d) ||
 888                                      d->opd_new_connection,
 889                                      &lwi);
 890
 891                         if (!d->opd_new_connection)
 892                                 continue;
 893
 894                         d->opd_new_connection = 0;
 895                         d->opd_got_disconnected = 0;
 896                         break;
 897                 }
 898
 899                 if (!osp_precreate_running(d))
 900                         break;
 901
 902                 LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
 903                 /* Sigh, fid client is not ready yet */
 904                 if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL)
 905                         continue;
 906
 907                 /* Init fid for osp_precreate if necessary */
 908                 rc = osp_init_pre_fid(d);
 909                 if (rc != 0) {
 910                         class_export_put(d->opd_exp);
 911                         d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
 912                         CERROR("%s: init pre fid error: rc = %d\n",
 913                                d->opd_obd->obd_name, rc);
 914                         continue;
 915                 }
 916
 917                 osp_statfs_update(d);
 918
 919                 /*
 920                  * Clean up orphans or recreate missing objects.
 921                  */
 922                 rc = osp_precreate_cleanup_orphans(&env, d);
 923                 if (rc != 0)
 924                         continue;
 925                 /*
 926                  * connected, can handle precreates now
 927                  */
 928                 while (osp_precreate_running(d)) {
 929                         l_wait_event(d->opd_pre_waitq,
 930                                      !osp_precreate_running(d) ||
 931                                      osp_precreate_near_empty(&env, d) ||
 932                                      osp_statfs_need_update(d) ||
 933                                      d->opd_got_disconnected, &lwi);
 934
 935                         if (!osp_precreate_running(d))
 936                                 break;
 937
 938                         /* something happened to the connection
 939                          * have to start from the beginning */
 940                         if (d->opd_got_disconnected)
 941                                 break;
 942
 943                         if (osp_statfs_need_update(d))
 944                                 osp_statfs_update(d);
 945
 946                         /* To avoid handling different seq in precreate/orphan
 947                          * cleanup, it will hold precreate until current seq is
 948                          * used up. */
 949                         if (unlikely(osp_precreate_end_seq(&env, d) &&
 950                             !osp_create_end_seq(&env, d)))
 951                                 continue;
 952
 953                         if (unlikely(osp_precreate_end_seq(&env, d) &&
 954                                      osp_create_end_seq(&env, d))) {
 955                                 LCONSOLE_INFO("%s:"LPX64" is used up."
 956                                               " Update to new seq\n",
 957                                               d->opd_obd->obd_name,
 958                                          fid_seq(&d->opd_pre_last_created_fid));
 959                                 rc = osp_precreate_rollover_new_seq(&env, d);
 960                                 if (rc)
 961                                         continue;
 962                         }
 963
 964                         if (osp_precreate_near_empty(&env, d)) {
 965                                 rc = osp_precreate_send(&env, d);
 966                                 /* osp_precreate_send() sets opd_pre_status
 967                                  * in case of error, that prevent the using of
 968                                  * failed device. */
 969                                 if (rc < 0 && rc != -ENOSPC &&
 970                                     rc != -ETIMEDOUT && rc != -ENOTCONN)
 971                                         CERROR("%s: cannot precreate objects:"
 972                                                " rc = %d\n",
 973                                                d->opd_obd->obd_name, rc);
 974                         }
 975                 }
 976         }
 977
 978         thread->t_flags = SVC_STOPPED;
 979         lu_env_fini(&env);
 980         wake_up(&thread->t_ctl_waitq);
 981
 982         RETURN(0);
 983 }
 984
 985 static int osp_precreate_ready_condition(const struct lu_env *env,
 986                                          struct osp_device *d)
 987 {
 988         if (d->opd_pre_recovering)
 989                 return 0;
 990
 991         /* ready if got enough precreated objects */
 992         /* we need to wait for others (opd_pre_reserved) and our object (+1) */
 993         if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d))
 994                 return 1;
 995
 996         /* ready if OST reported no space and no destroys in progress */
 997         if (d->opd_syn_changes + d->opd_syn_rpc_in_progress == 0 &&
 998             d->opd_pre_status == -ENOSPC)
 999                 return 1;
1000
1001         /* Bail out I/O fails to OST */
1002         if (d->opd_pre_status != 0 &&
1003             d->opd_pre_status != -EAGAIN &&
1004             d->opd_pre_status != -ENODEV &&
1005             d->opd_pre_status != -ENOSPC) {
1006                 /* DEBUG LU-3230 */
1007                 if (d->opd_pre_status != -EIO)
1008                         CERROR("%s: precreate failed opd_pre_status %d\n",
1009                                d->opd_obd->obd_name, d->opd_pre_status);
1010                 return 1;
1011         }
1012
1013         return 0;
1014 }
1015
1016 static int osp_precreate_timeout_condition(void *data)
1017 {
1018         struct osp_device *d = data;
1019
1020         CDEBUG(D_HA, "%s: slow creates, last="DFID", next="DFID", "
1021               "reserved="LPU64", syn_changes=%lu, "
1022               "syn_rpc_in_progress=%d, status=%d\n",
1023               d->opd_obd->obd_name, PFID(&d->opd_pre_last_created_fid),
1024               PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
1025               d->opd_syn_changes, d->opd_syn_rpc_in_progress,
1026               d->opd_pre_status);
1027
1028         return 1;
1029 }
1030
1031 /*
1032  * called to reserve object in the pool
1033  * return codes:
1034  *  ENOSPC - no space on corresponded OST
1035  *  EAGAIN - precreation is in progress, try later
1036  *  EIO    - no access to OST
1037  */
1038 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d)
1039 {
1040         struct l_wait_info       lwi;
1041         cfs_time_t               expire = cfs_time_shift(obd_timeout);
1042         int                      precreated, rc;
1043
1044         ENTRY;
1045
1046         LASSERTF(osp_objs_precreated(env, d) >= 0, "Last created FID "DFID
1047                  "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
1048                  PFID(&d->opd_pre_used_fid));
1049
1050         /*
1051          * wait till:
1052          *  - preallocation is done
1053          *  - no free space expected soon
1054          *  - can't connect to OST for too long (obd_timeout)
1055          *  - OST can allocate fid sequence.
1056          */
1057         while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
1058                 rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) {
1059
1060                 /*
1061                  * increase number of precreations
1062                  */
1063                 precreated = osp_objs_precreated(env, d);
1064                 if (d->opd_pre_grow_count < d->opd_pre_max_grow_count &&
1065                     d->opd_pre_grow_slow == 0 &&
1066                     precreated <= (d->opd_pre_grow_count / 4 + 1)) {
1067                         spin_lock(&d->opd_pre_lock);
1068                         d->opd_pre_grow_slow = 1;
1069                         d->opd_pre_grow_count *= 2;
1070                         spin_unlock(&d->opd_pre_lock);
1071                 }
1072
1073                 spin_lock(&d->opd_pre_lock);
1074                 precreated = osp_objs_precreated(env, d);
1075                 if (precreated > d->opd_pre_reserved &&
1076                     !d->opd_pre_recovering) {
1077                         d->opd_pre_reserved++;
1078                         spin_unlock(&d->opd_pre_lock);
1079                         rc = 0;
1080
1081                         /* XXX: don't wake up if precreation is in progress */
1082                         if (osp_precreate_near_empty_nolock(env, d) &&
1083                            !osp_precreate_end_seq_nolock(env, d))
1084                                 wake_up(&d->opd_pre_waitq);
1085
1086                         break;
1087                 }
1088                 spin_unlock(&d->opd_pre_lock);
1089
1090                 /*
1091                  * all precreated objects have been used and no-space
1092                  * status leave us no chance to succeed very soon
1093                  * but if there is destroy in progress, then we should
1094                  * wait till that is done - some space might be released
1095                  */
1096                 if (unlikely(rc == -ENOSPC)) {
1097                         if (d->opd_syn_changes) {
1098                                 /* force local commit to release space */
1099                                 dt_commit_async(env, d->opd_storage);
1100                         }
1101                         if (d->opd_syn_rpc_in_progress) {
1102                                 /* just wait till destroys are done */
1103                                 /* see l_wait_even() few lines below */
1104                         }
1105                         if (d->opd_syn_changes +
1106                             d->opd_syn_rpc_in_progress == 0) {
1107                                 /* no hope for free space */
1108                                 break;
1109                         }
1110                 }
1111
1112                 /* XXX: don't wake up if precreation is in progress */
1113                 wake_up(&d->opd_pre_waitq);
1114
1115                 lwi = LWI_TIMEOUT(expire - cfs_time_current(),
1116                                 osp_precreate_timeout_condition, d);
1117                 if (cfs_time_aftereq(cfs_time_current(), expire)) {
1118                         rc = -ETIMEDOUT;
1119                         break;
1120                 }
1121
1122                 l_wait_event(d->opd_pre_user_waitq,
1123                              osp_precreate_ready_condition(env, d), &lwi);
1124         }
1125
1126         RETURN(rc);
1127 }
1128
1129 /*
1130  * this function relies on reservation made before
1131  */
1132 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1133                           struct lu_fid *fid)
1134 {
1135         /* grab next id from the pool */
1136         spin_lock(&d->opd_pre_lock);
1137
1138         LASSERTF(osp_fid_diff(&d->opd_pre_used_fid,
1139                              &d->opd_pre_last_created_fid) < 0,
1140                  "next fid "DFID" last created fid "DFID"\n",
1141                  PFID(&d->opd_pre_used_fid),
1142                  PFID(&d->opd_pre_last_created_fid));
1143
1144         d->opd_pre_used_fid.f_oid++;
1145         memcpy(fid, &d->opd_pre_used_fid, sizeof(*fid));
1146         d->opd_pre_reserved--;
1147         /*
1148          * last_used_id must be changed along with getting new id otherwise
1149          * we might miscalculate gap causing object loss or leak
1150          */
1151         osp_update_last_fid(d, fid);
1152         spin_unlock(&d->opd_pre_lock);
1153
1154         /*
1155          * probably main thread suspended orphan cleanup till
1156          * all reservations are released, see comment in
1157          * osp_precreate_thread() just before orphan cleanup
1158          */
1159         if (unlikely(d->opd_pre_reserved == 0 && d->opd_pre_status))
1160                 wake_up(&d->opd_pre_waitq);
1161
1162         return 0;
1163 }
1164
1165 /*
1166  *
1167  */
1168 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1169                         __u64 size)
1170 {
1171         struct osp_device       *d = lu2osp_dev(dt->do_lu.lo_dev);
1172         struct ptlrpc_request   *req = NULL;
1173         struct obd_import       *imp;
1174         struct ost_body         *body;
1175         struct obdo             *oa = NULL;
1176         int                      rc;
1177
1178         ENTRY;
1179
1180         imp = d->opd_obd->u.cli.cl_import;
1181         LASSERT(imp);
1182
1183         req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1184         if (req == NULL)
1185                 RETURN(-ENOMEM);
1186
1187         /* XXX: capa support? */
1188         /* osc_set_capa_size(req, &RMF_CAPA1, capa); */
1189         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1190         if (rc) {
1191                 ptlrpc_request_free(req);
1192                 RETURN(rc);
1193         }
1194
1195         /*
1196          * XXX: decide how do we do here with resend
1197          * if we don't resend, then client may see wrong file size
1198          * if we do resend, then MDS thread can get stuck for quite long
1199          */
1200         req->rq_no_resend = req->rq_no_delay = 1;
1201
1202         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1203         ptlrpc_at_set_req_timeout(req);
1204
1205         OBD_ALLOC_PTR(oa);
1206         if (oa == NULL)
1207                 GOTO(out, rc = -ENOMEM);
1208
1209         rc = fid_to_ostid(lu_object_fid(&dt->do_lu), &oa->o_oi);
1210         LASSERT(rc == 0);
1211         oa->o_size = size;
1212         oa->o_blocks = OBD_OBJECT_EOF;
1213         oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1214                       OBD_MD_FLID | OBD_MD_FLGROUP;
1215
1216         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1217         LASSERT(body);
1218         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1219
1220         /* XXX: capa support? */
1221         /* osc_pack_capa(req, body, capa); */
1222
1223         ptlrpc_request_set_replen(req);
1224
1225         rc = ptlrpc_queue_wait(req);
1226         if (rc)
1227                 CERROR("can't punch object: %d\n", rc);
1228 out:
1229         ptlrpc_req_finished(req);
1230         if (oa)
1231                 OBD_FREE_PTR(oa);
1232         RETURN(rc);
1233 }
1234
1235 int osp_init_precreate(struct osp_device *d)
1236 {
1237         struct l_wait_info       lwi = { 0 };
1238         struct task_struct              *task;
1239
1240         ENTRY;
1241
1242         OBD_ALLOC_PTR(d->opd_pre);
1243         if (d->opd_pre == NULL)
1244                 RETURN(-ENOMEM);
1245
1246         /* initially precreation isn't ready */
1247         d->opd_pre_status = -EAGAIN;
1248         fid_zero(&d->opd_pre_used_fid);
1249         d->opd_pre_used_fid.f_oid = 1;
1250         fid_zero(&d->opd_pre_last_created_fid);
1251         d->opd_pre_last_created_fid.f_oid = 1;
1252         d->opd_pre_reserved = 0;
1253         d->opd_got_disconnected = 1;
1254         d->opd_pre_grow_slow = 0;
1255         d->opd_pre_grow_count = OST_MIN_PRECREATE;
1256         d->opd_pre_min_grow_count = OST_MIN_PRECREATE;
1257         d->opd_pre_max_grow_count = OST_MAX_PRECREATE;
1258
1259         spin_lock_init(&d->opd_pre_lock);
1260         init_waitqueue_head(&d->opd_pre_waitq);
1261         init_waitqueue_head(&d->opd_pre_user_waitq);
1262         init_waitqueue_head(&d->opd_pre_thread.t_ctl_waitq);
1263
1264         /*
1265          * Initialize statfs-related things
1266          */
1267         d->opd_statfs_maxage = 5; /* default update interval */
1268         d->opd_statfs_fresh_till = cfs_time_shift(-1000);
1269         CDEBUG(D_OTHER, "current %llu, fresh till %llu\n",
1270                (unsigned long long)cfs_time_current(),
1271                (unsigned long long)d->opd_statfs_fresh_till);
1272         cfs_timer_init(&d->opd_statfs_timer, osp_statfs_timer_cb, d);
1273
1274         /*
1275          * start thread handling precreation and statfs updates
1276          */
1277         task = kthread_run(osp_precreate_thread, d,
1278                            "osp-pre-%u-%u", d->opd_index, d->opd_group);
1279         if (IS_ERR(task)) {
1280                 CERROR("can't start precreate thread %ld\n", PTR_ERR(task));
1281                 RETURN(PTR_ERR(task));
1282         }
1283
1284         l_wait_event(d->opd_pre_thread.t_ctl_waitq,
1285                      osp_precreate_running(d) || osp_precreate_stopped(d),
1286                      &lwi);
1287
1288         RETURN(0);
1289 }
1290
1291 void osp_precreate_fini(struct osp_device *d)
1292 {
1293         struct ptlrpc_thread *thread;
1294
1295         ENTRY;
1296
1297         cfs_timer_disarm(&d->opd_statfs_timer);
1298
1299         if (d->opd_pre == NULL)
1300                 RETURN_EXIT;
1301
1302         thread = &d->opd_pre_thread;
1303
1304         thread->t_flags = SVC_STOPPING;
1305         wake_up(&d->opd_pre_waitq);
1306
1307         wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
1308
1309         OBD_FREE_PTR(d->opd_pre);
1310         d->opd_pre = NULL;
1311
1312         EXIT;
1313 }
1314