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 static void osp_statfs_timer_cb(unsigned long _d)
  69 {
  70         struct osp_device *d = (struct osp_device *) _d;
  71
  72         LASSERT(d);
  73         wake_up(&d->opd_pre_waitq);
  74 }
  75
  76 static int osp_statfs_interpret(const struct lu_env *env,
  77                                 struct ptlrpc_request *req,
  78                                 union ptlrpc_async_args *aa, int rc)
  79 {
  80         struct obd_import       *imp = req->rq_import;
  81         struct obd_statfs       *msfs;
  82         struct osp_device       *d;
  83
  84         ENTRY;
  85
  86         aa = ptlrpc_req_async_args(req);
  87         d = aa->pointer_arg[0];
  88         LASSERT(d);
  89
  90         if (rc != 0)
  91                 GOTO(out, rc);
  92
  93         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
  94         if (msfs == NULL)
  95                 GOTO(out, rc = -EPROTO);
  96
  97         d->opd_statfs = *msfs;
  98
  99         osp_pre_update_status(d, rc);
 100
 101         /* schedule next update */
 102         d->opd_statfs_fresh_till = cfs_time_shift(d->opd_statfs_maxage);
 103         cfs_timer_arm(&d->opd_statfs_timer, d->opd_statfs_fresh_till);
 104         d->opd_statfs_update_in_progress = 0;
 105
 106         CDEBUG(D_CACHE, "updated statfs %p\n", d);
 107
 108         RETURN(0);
 109 out:
 110         /* couldn't update statfs, try again as soon as possible */
 111         wake_up(&d->opd_pre_waitq);
 112         if (req->rq_import_generation == imp->imp_generation)
 113                 CDEBUG(D_CACHE, "%s: couldn't update statfs: rc = %d\n",
 114                        d->opd_obd->obd_name, rc);
 115         RETURN(rc);
 116 }
 117
 118 static int osp_statfs_update(struct osp_device *d)
 119 {
 120         struct ptlrpc_request   *req;
 121         struct obd_import       *imp;
 122         union ptlrpc_async_args *aa;
 123         int                      rc;
 124
 125         ENTRY;
 126
 127         CDEBUG(D_CACHE, "going to update statfs\n");
 128
 129         imp = d->opd_obd->u.cli.cl_import;
 130         LASSERT(imp);
 131
 132         req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
 133         if (req == NULL)
 134                 RETURN(-ENOMEM);
 135
 136         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
 137         if (rc) {
 138                 ptlrpc_request_free(req);
 139                 RETURN(rc);
 140         }
 141         ptlrpc_request_set_replen(req);
 142         req->rq_request_portal = OST_CREATE_PORTAL;
 143         ptlrpc_at_set_req_timeout(req);
 144
 145         req->rq_interpret_reply = (ptlrpc_interpterer_t)osp_statfs_interpret;
 146         aa = ptlrpc_req_async_args(req);
 147         aa->pointer_arg[0] = d;
 148
 149         /*
 150          * no updates till reply
 151          */
 152         cfs_timer_disarm(&d->opd_statfs_timer);
 153         d->opd_statfs_fresh_till = cfs_time_shift(obd_timeout * 1000);
 154         d->opd_statfs_update_in_progress = 1;
 155
 156         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
 157
 158         RETURN(0);
 159 }
 160
 161 /*
 162  * XXX: there might be a case where removed object(s) do not add free
 163  * space (empty object). if the number of such deletions is high, then
 164  * we can start to update statfs too often - a rpc storm
 165  * TODO: some throttling is needed
 166  */
 167 void osp_statfs_need_now(struct osp_device *d)
 168 {
 169         if (!d->opd_statfs_update_in_progress) {
 170                 /*
 171                  * if current status is -ENOSPC (lack of free space on OST)
 172                  * then we should poll OST immediately once object destroy
 173                  * is replied
 174                  */
 175                 d->opd_statfs_fresh_till = cfs_time_shift(-1);
 176                 cfs_timer_disarm(&d->opd_statfs_timer);
 177                 wake_up(&d->opd_pre_waitq);
 178         }
 179 }
 180
 181
 182 /*
 183  * OSP tries to maintain pool of available objects so that calls to create
 184  * objects don't block most of time
 185  *
 186  * each time OSP gets connected to OST, we should start from precreation cleanup
 187  */
 188 static inline int osp_precreate_running(struct osp_device *d)
 189 {
 190         return !!(d->opd_pre_thread.t_flags & SVC_RUNNING);
 191 }
 192
 193 static inline int osp_precreate_stopped(struct osp_device *d)
 194 {
 195         return !!(d->opd_pre_thread.t_flags & SVC_STOPPED);
 196 }
 197
 198 static inline int osp_objs_precreated(const struct lu_env *env,
 199                                       struct osp_device *osp)
 200 {
 201         struct lu_fid *fid1 = &osp->opd_pre_last_created_fid;
 202         struct lu_fid *fid2 = &osp->opd_pre_used_fid;
 203
 204         LASSERTF(fid_seq(fid1) == fid_seq(fid2),
 205                  "Created fid"DFID" Next fid "DFID"\n", PFID(fid1), PFID(fid2));
 206
 207         if (fid_is_idif(fid1)) {
 208                 struct ost_id *oi1 = &osp_env_info(env)->osi_oi;
 209                 struct ost_id *oi2 = &osp_env_info(env)->osi_oi2;
 210
 211                 LASSERT(fid_is_idif(fid1) && fid_is_idif(fid2));
 212                 fid_to_ostid(fid1, oi1);
 213                 fid_to_ostid(fid2, oi2);
 214                 LASSERT(ostid_id(oi1) >= ostid_id(oi2));
 215
 216                 return ostid_id(oi1) - ostid_id(oi2);
 217         }
 218
 219         return fid_oid(fid1) - fid_oid(fid2);
 220 }
 221
 222 static inline int osp_precreate_near_empty_nolock(const struct lu_env *env,
 223                                                   struct osp_device *d)
 224 {
 225         int window = osp_objs_precreated(env, d);
 226
 227         /* don't consider new precreation till OST is healty and
 228          * has free space */
 229         return ((window - d->opd_pre_reserved < d->opd_pre_grow_count / 2) &&
 230                 (d->opd_pre_status == 0));
 231 }
 232
 233 static inline int osp_precreate_near_empty(const struct lu_env *env,
 234                                            struct osp_device *d)
 235 {
 236         int rc;
 237
 238         /* XXX: do we really need locking here? */
 239         spin_lock(&d->opd_pre_lock);
 240         rc = osp_precreate_near_empty_nolock(env, d);
 241         spin_unlock(&d->opd_pre_lock);
 242         return rc;
 243 }
 244
 245 static inline int osp_create_end_seq(const struct lu_env *env,
 246                                      struct osp_device *osp)
 247 {
 248         struct lu_fid *fid = &osp->opd_pre_used_fid;
 249         int rc;
 250
 251         spin_lock(&osp->opd_pre_lock);
 252         rc = osp_fid_end_seq(env, fid);
 253         spin_unlock(&osp->opd_pre_lock);
 254         return rc;
 255 }
 256
 257 /**
 258  * Write fid into last_oid/last_seq file.
 259  **/
 260 int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp,
 261                                  struct lu_fid *fid, int sync)
 262 {
 263         struct osp_thread_info  *oti = osp_env_info(env);
 264         struct lu_buf      *lb_oid = &oti->osi_lb;
 265         struct lu_buf      *lb_oseq = &oti->osi_lb2;
 266         loff_t             oid_off;
 267         loff_t             oseq_off;
 268         struct thandle    *th;
 269         int                   rc;
 270         ENTRY;
 271
 272         /* Note: through f_oid is only 32bits, it will also write
 273          * 64 bits for oid to keep compatiblity with the previous
 274          * version. */
 275         lb_oid->lb_buf = &fid->f_oid;
 276         lb_oid->lb_len = sizeof(obd_id);
 277         oid_off = sizeof(obd_id) * osp->opd_index;
 278
 279         lb_oseq->lb_buf = &fid->f_seq;
 280         lb_oseq->lb_len = sizeof(obd_id);
 281         oseq_off = sizeof(obd_id) * osp->opd_index;
 282
 283         th = dt_trans_create(env, osp->opd_storage);
 284         if (IS_ERR(th))
 285                 RETURN(PTR_ERR(th));
 286
 287         th->th_sync |= sync;
 288         rc = dt_declare_record_write(env, osp->opd_last_used_oid_file,
 289                                      lb_oid->lb_len, oid_off, th);
 290         if (rc != 0)
 291                 GOTO(out, rc);
 292
 293         rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
 294                                      lb_oseq->lb_len, oseq_off, th);
 295         if (rc != 0)
 296                 GOTO(out, rc);
 297
 298         rc = dt_trans_start_local(env, osp->opd_storage, th);
 299         if (rc != 0)
 300                 GOTO(out, rc);
 301
 302         rc = dt_record_write(env, osp->opd_last_used_oid_file, lb_oid,
 303                              &oid_off, th);
 304         if (rc != 0) {
 305                 CERROR("%s: can not write to last seq file: rc = %d\n",
 306                         osp->opd_obd->obd_name, rc);
 307                 GOTO(out, rc);
 308         }
 309         rc = dt_record_write(env, osp->opd_last_used_seq_file, lb_oseq,
 310                              &oseq_off, th);
 311         if (rc) {
 312                 CERROR("%s: can not write to last seq file: rc = %d\n",
 313                         osp->opd_obd->obd_name, rc);
 314                 GOTO(out, rc);
 315         }
 316 out:
 317         dt_trans_stop(env, osp->opd_storage, th);
 318         RETURN(rc);
 319 }
 320
 321 int osp_precreate_rollover_new_seq(struct lu_env *env, struct osp_device *osp)
 322 {
 323         struct lu_fid   *fid = &osp_env_info(env)->osi_fid;
 324         struct lu_fid   *last_fid = &osp->opd_last_used_fid;
 325         int             rc;
 326         ENTRY;
 327
 328         rc = seq_client_get_seq(env, osp->opd_obd->u.cli.cl_seq, &fid->f_seq);
 329         if (rc != 0) {
 330                 CERROR("%s: alloc fid error: rc = %d\n",
 331                        osp->opd_obd->obd_name, rc);
 332                 RETURN(rc);
 333         }
 334
 335         fid->f_oid = 1;
 336         fid->f_ver = 0;
 337         LASSERTF(fid_seq(fid) != fid_seq(last_fid),
 338                  "fid "DFID", last_fid "DFID"\n", PFID(fid),
 339                  PFID(last_fid));
 340
 341         rc = osp_write_last_oid_seq_files(env, osp, fid, 1);
 342         if (rc != 0) {
 343                 CERROR("%s: Can not update oid/seq file: rc = %d\n",
 344                        osp->opd_obd->obd_name, rc);
 345                 RETURN(rc);
 346         }
 347
 348         LCONSOLE_INFO("%s: update sequence from "LPX64" to "LPX64"\n",
 349                       osp->opd_obd->obd_name, fid_seq(last_fid),
 350                       fid_seq(fid));
 351         /* Update last_xxx to the new seq */
 352         spin_lock(&osp->opd_pre_lock);
 353         osp->opd_last_used_fid = *fid;
 354         osp->opd_gap_start_fid = *fid;
 355         osp->opd_pre_used_fid = *fid;
 356         osp->opd_pre_last_created_fid = *fid;
 357         spin_unlock(&osp->opd_pre_lock);
 358
 359         RETURN(rc);
 360 }
 361
 362 /**
 363  * alloc fids for precreation.
 364  * rc = 0 Success, @grow is the count of real allocation.
 365  * rc = 1 Current seq is used up.
 366  * rc < 0 Other error.
 367  **/
 368 static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp,
 369                               struct lu_fid *fid, int *grow)
 370 {
 371         struct osp_thread_info  *osi = osp_env_info(env);
 372         __u64                   end;
 373         int                     i = 0;
 374
 375         if (fid_is_idif(fid)) {
 376                 struct lu_fid   *last_fid;
 377                 struct ost_id   *oi = &osi->osi_oi;
 378
 379                 spin_lock(&osp->opd_pre_lock);
 380                 last_fid = &osp->opd_pre_last_created_fid;
 381                 fid_to_ostid(last_fid, oi);
 382                 end = min(ostid_id(oi) + *grow, IDIF_MAX_OID);
 383                 *grow = end - ostid_id(oi);
 384                 ostid_set_id(oi, ostid_id(oi) + *grow);
 385                 spin_unlock(&osp->opd_pre_lock);
 386
 387                 if (*grow == 0)
 388                         return 1;
 389
 390                 ostid_to_fid(fid, oi, osp->opd_index);
 391                 return 0;
 392         }
 393
 394         spin_lock(&osp->opd_pre_lock);
 395         *fid = osp->opd_pre_last_created_fid;
 396         end = fid->f_oid;
 397         end = min((end + *grow), (__u64)LUSTRE_DATA_SEQ_MAX_WIDTH);
 398         *grow = end - fid->f_oid;
 399         fid->f_oid += end - fid->f_oid;
 400         spin_unlock(&osp->opd_pre_lock);
 401
 402         CDEBUG(D_INFO, "Expect %d, actual %d ["DFID" -- "DFID"]\n",
 403                *grow, i, PFID(fid), PFID(&osp->opd_pre_last_created_fid));
 404
 405         return *grow > 0 ? 0 : 1;
 406 }
 407
 408 static int osp_precreate_send(const struct lu_env *env, struct osp_device *d)
 409 {
 410         struct osp_thread_info  *oti = osp_env_info(env);
 411         struct ptlrpc_request   *req;
 412         struct obd_import       *imp;
 413         struct ost_body         *body;
 414         int                      rc, grow, diff;
 415         struct lu_fid           *fid = &oti->osi_fid;
 416         ENTRY;
 417
 418         /* don't precreate new objects till OST healthy and has free space */
 419         if (unlikely(d->opd_pre_status)) {
 420                 CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
 421                        d->opd_obd->obd_name, d->opd_pre_status);
 422                 RETURN(0);
 423         }
 424
 425         /*
 426          * if not connection/initialization is compeleted, ignore
 427          */
 428         imp = d->opd_obd->u.cli.cl_import;
 429         LASSERT(imp);
 430
 431         req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
 432         if (req == NULL)
 433                 RETURN(-ENOMEM);
 434         req->rq_request_portal = OST_CREATE_PORTAL;
 435         /* we should not resend create request - anyway we will have delorphan
 436          * and kill these objects */
 437         req->rq_no_delay = req->rq_no_resend = 1;
 438
 439         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
 440         if (rc) {
 441                 ptlrpc_request_free(req);
 442                 RETURN(rc);
 443         }
 444
 445         spin_lock(&d->opd_pre_lock);
 446         if (d->opd_pre_grow_count > d->opd_pre_max_grow_count / 2)
 447                 d->opd_pre_grow_count = d->opd_pre_max_grow_count / 2;
 448         grow = d->opd_pre_grow_count;
 449         spin_unlock(&d->opd_pre_lock);
 450
 451         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 452         LASSERT(body);
 453
 454         *fid = d->opd_pre_last_created_fid;
 455         rc = osp_precreate_fids(env, d, fid, &grow);
 456         if (rc == 1) {
 457                 /* Current seq has been used up*/
 458                 if (!osp_is_fid_client(d)) {
 459                         osp_pre_update_status(d, -ENOSPC);
 460                         rc = -ENOSPC;
 461                 }
 462                 wake_up(&d->opd_pre_waitq);
 463                 GOTO(out_req, rc);
 464         }
 465
 466         if (!osp_is_fid_client(d)) {
 467                 /* Non-FID client will always send seq 0 because of
 468                  * compatiblity */
 469                 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
 470                 fid->f_seq = 0;
 471         }
 472
 473         fid_to_ostid(fid, &body->oa.o_oi);
 474         body->oa.o_valid = OBD_MD_FLGROUP;
 475
 476         ptlrpc_request_set_replen(req);
 477
 478         rc = ptlrpc_queue_wait(req);
 479         if (rc) {
 480                 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
 481                        rc);
 482                 GOTO(out_req, rc);
 483         }
 484         LASSERT(req->rq_transno == 0);
 485
 486         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 487         if (body == NULL)
 488                 GOTO(out_req, rc = -EPROTO);
 489
 490         ostid_to_fid(fid, &body->oa.o_oi, d->opd_index);
 491         LASSERTF(lu_fid_diff(fid, &d->opd_pre_used_fid) > 0,
 492                  "reply fid "DFID" pre used fid "DFID"\n", PFID(fid),
 493                  PFID(&d->opd_pre_used_fid));
 494
 495         diff = lu_fid_diff(fid, &d->opd_pre_last_created_fid);
 496
 497         spin_lock(&d->opd_pre_lock);
 498         if (diff < grow) {
 499                 /* the OST has not managed to create all the
 500                  * objects we asked for */
 501                 d->opd_pre_grow_count = max(diff, OST_MIN_PRECREATE);
 502                 d->opd_pre_grow_slow = 1;
 503         } else {
 504                 /* the OST is able to keep up with the work,
 505                  * we could consider increasing grow_count
 506                  * next time if needed */
 507                 d->opd_pre_grow_slow = 0;
 508         }
 509
 510         d->opd_pre_last_created_fid = *fid;
 511         spin_unlock(&d->opd_pre_lock);
 512
 513         CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n",
 514                d->opd_obd->obd_name, PFID(&d->opd_pre_used_fid),
 515                PFID(&d->opd_pre_last_created_fid));
 516 out_req:
 517         /* now we can wakeup all users awaiting for objects */
 518         osp_pre_update_status(d, rc);
 519         wake_up(&d->opd_pre_user_waitq);
 520
 521         ptlrpc_req_finished(req);
 522         RETURN(rc);
 523 }
 524
 525 static int osp_get_lastfid_from_ost(const struct lu_env *env,
 526                                     struct osp_device *d)
 527 {
 528         struct ptlrpc_request   *req = NULL;
 529         struct obd_import       *imp;
 530         struct lu_fid           *last_fid;
 531         char                    *tmp;
 532         int                     rc;
 533         ENTRY;
 534
 535         imp = d->opd_obd->u.cli.cl_import;
 536         LASSERT(imp);
 537
 538         req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
 539         if (req == NULL)
 540                 RETURN(-ENOMEM);
 541
 542         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY, RCL_CLIENT,
 543                              sizeof(KEY_LAST_FID));
 544
 545         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
 546         if (rc) {
 547                 ptlrpc_request_free(req);
 548                 RETURN(rc);
 549         }
 550
 551         tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
 552         memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
 553
 554         req->rq_no_delay = req->rq_no_resend = 1;
 555         last_fid = req_capsule_client_get(&req->rq_pill, &RMF_FID);
 556         fid_cpu_to_le(last_fid, &d->opd_last_used_fid);
 557
 558         ptlrpc_request_set_replen(req);
 559
 560         rc = ptlrpc_queue_wait(req);
 561         if (rc) {
 562                 /* bad-bad OST.. let sysadm sort this out */
 563                 if (rc == -ENOTSUPP) {
 564                         CERROR("%s: server does not support FID: rc = %d\n",
 565                                d->opd_obd->obd_name, -ENOTSUPP);
 566                 }
 567                 ptlrpc_set_import_active(imp, 0);
 568                 GOTO(out, rc);
 569         }
 570
 571         last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
 572         if (last_fid == NULL) {
 573                 CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name);
 574                 GOTO(out, rc = -EPROTO);
 575         }
 576
 577         if (!fid_is_sane(last_fid)) {
 578                 CERROR("%s: Got insane last_fid "DFID"\n",
 579                        d->opd_obd->obd_name, PFID(last_fid));
 580                 GOTO(out, rc = -EPROTO);
 581         }
 582
 583         /* Only update the last used fid, if the OST has objects for
 584          * this sequence, i.e. fid_oid > 0 */
 585         if (fid_oid(last_fid) > 0)
 586                 d->opd_last_used_fid = *last_fid;
 587
 588         CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
 589                PFID(last_fid));
 590
 591 out:
 592         ptlrpc_req_finished(req);
 593         RETURN(rc);
 594 }
 595
 596 /**
 597  * asks OST to clean precreate orphans
 598  * and gets next id for new objects
 599  */
 600 static int osp_precreate_cleanup_orphans(struct lu_env *env,
 601                                          struct osp_device *d)
 602 {
 603         struct osp_thread_info  *osi = osp_env_info(env);
 604         struct lu_fid           *last_fid = &osi->osi_fid;
 605         struct ptlrpc_request   *req = NULL;
 606         struct obd_import       *imp;
 607         struct ost_body         *body;
 608         struct l_wait_info       lwi = { 0 };
 609         int                      update_status = 0;
 610         int                      rc;
 611         int                      diff;
 612
 613         ENTRY;
 614
 615         /*
 616          * wait for local recovery to finish, so we can cleanup orphans
 617          * orphans are all objects since "last used" (assigned), but
 618          * there might be objects reserved and in some cases they won't
 619          * be used. we can't cleanup them till we're sure they won't be
 620          * used. also can't we allow new reservations because they may
 621          * end up getting orphans being cleaned up below. so we block
 622          * new reservations and wait till all reserved objects either
 623          * user or released.
 624          */
 625         spin_lock(&d->opd_pre_lock);
 626         d->opd_pre_recovering = 1;
 627         spin_unlock(&d->opd_pre_lock);
 628         /*
 629          * The locking above makes sure the opd_pre_reserved check below will
 630          * catch all osp_precreate_reserve() calls who find
 631          * "!opd_pre_recovering".
 632          */
 633         l_wait_event(d->opd_pre_waitq,
 634                      (!d->opd_pre_reserved && d->opd_recovery_completed) ||
 635                      !osp_precreate_running(d) || d->opd_got_disconnected,
 636                      &lwi);
 637         if (!osp_precreate_running(d) || d->opd_got_disconnected)
 638                 GOTO(out, rc = -EAGAIN);
 639
 640         CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
 641                d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
 642
 643         *last_fid = d->opd_last_used_fid;
 644         /* The OSP should already get the valid seq now */
 645         LASSERT(!fid_is_zero(last_fid));
 646         if (fid_oid(&d->opd_last_used_fid) < 2) {
 647                 /* lastfid looks strange... ask OST */
 648                 rc = osp_get_lastfid_from_ost(env, d);
 649                 if (rc)
 650                         GOTO(out, rc);
 651         }
 652
 653         imp = d->opd_obd->u.cli.cl_import;
 654         LASSERT(imp);
 655
 656         req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
 657         if (req == NULL)
 658                 GOTO(out, rc = -ENOMEM);
 659
 660         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
 661         if (rc) {
 662                 ptlrpc_request_free(req);
 663                 req = NULL;
 664                 GOTO(out, rc);
 665         }
 666
 667         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 668         if (body == NULL)
 669                 GOTO(out, rc = -EPROTO);
 670
 671         body->oa.o_flags = OBD_FL_DELORPHAN;
 672         body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
 673
 674         fid_to_ostid(&d->opd_last_used_fid, &body->oa.o_oi);
 675
 676         ptlrpc_request_set_replen(req);
 677
 678         /* Don't resend the delorphan req */
 679         req->rq_no_resend = req->rq_no_delay = 1;
 680
 681         rc = ptlrpc_queue_wait(req);
 682         if (rc) {
 683                 update_status = 1;
 684                 GOTO(out, rc);
 685         }
 686
 687         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 688         if (body == NULL)
 689                 GOTO(out, rc = -EPROTO);
 690
 691         /*
 692          * OST provides us with id new pool starts from in body->oa.o_id
 693          */
 694         ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
 695
 696         spin_lock(&d->opd_pre_lock);
 697         diff = lu_fid_diff(&d->opd_last_used_fid, last_fid);
 698         if (diff > 0) {
 699                 d->opd_pre_grow_count = OST_MIN_PRECREATE + diff;
 700                 d->opd_pre_last_created_fid = d->opd_last_used_fid;
 701         } else {
 702                 d->opd_pre_grow_count = OST_MIN_PRECREATE;
 703                 d->opd_pre_last_created_fid = *last_fid;
 704         }
 705         /*
 706          * This empties the pre-creation pool and effectively blocks any new
 707          * reservations.
 708          */
 709         LASSERT(fid_oid(&d->opd_pre_last_created_fid) <=
 710                 LUSTRE_DATA_SEQ_MAX_WIDTH);
 711         d->opd_pre_used_fid = d->opd_pre_last_created_fid;
 712         d->opd_pre_grow_slow = 0;
 713         spin_unlock(&d->opd_pre_lock);
 714
 715         CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
 716                "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
 717                PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
 718 out:
 719         if (req)
 720                 ptlrpc_req_finished(req);
 721
 722         spin_lock(&d->opd_pre_lock);
 723         d->opd_pre_recovering = 0;
 724         spin_unlock(&d->opd_pre_lock);
 725
 726         /*
 727          * If rc is zero, the pre-creation window should have been emptied.
 728          * Since waking up the herd would be useless without pre-created
 729          * objects, we defer the signal to osp_precreate_send() in that case.
 730          */
 731         if (rc != 0) {
 732                 if (update_status) {
 733                         CERROR("%s: cannot cleanup orphans: rc = %d\n",
 734                                d->opd_obd->obd_name, rc);
 735                         /* we can't proceed from here, OST seem to
 736                          * be in a bad shape, better to wait for
 737                          * a new instance of the server and repeat
 738                          * from the beginning. notify possible waiters
 739                          * this OSP isn't quite functional yet */
 740                         osp_pre_update_status(d, rc);
 741                 } else {
 742                         wake_up(&d->opd_pre_user_waitq);
 743                 }
 744         }
 745
 746         RETURN(rc);
 747 }
 748
 749 /*
 750  * the function updates current precreation status used: functional or not
 751  *
 752  * rc is a last code from the transport, rc == 0 meaning transport works
 753  * well and users of lod can use objects from this OSP
 754  *
 755  * the status depends on current usage of OST
 756  */
 757 void osp_pre_update_status(struct osp_device *d, int rc)
 758 {
 759         struct obd_statfs       *msfs = &d->opd_statfs;
 760         int                      old = d->opd_pre_status;
 761         __u64                    used;
 762
 763         d->opd_pre_status = rc;
 764         if (rc)
 765                 goto out;
 766
 767         /* Add a bit of hysteresis so this flag isn't continually flapping,
 768          * and ensure that new files don't get extremely fragmented due to
 769          * only a small amount of available space in the filesystem.
 770          * We want to set the NOSPC flag when there is less than ~0.1% free
 771          * and clear it when there is at least ~0.2% free space, so:
 772          *                   avail < ~0.1% max          max = avail + used
 773          *            1025 * avail < avail + used       used = blocks - free
 774          *            1024 * avail < used
 775          *            1024 * avail < blocks - free
 776          *                   avail < ((blocks - free) >> 10)
 777          *
 778          * On very large disk, say 16TB 0.1% will be 16 GB. We don't want to
 779          * lose that amount of space so in those cases we report no space left
 780          * if their is less than 1 GB left.                             */
 781         if (likely(msfs->os_type)) {
 782                 used = min_t(__u64, (msfs->os_blocks - msfs->os_bfree) >> 10,
 783                                     1 << 30);
 784                 if ((msfs->os_ffree < 32) || (msfs->os_bavail < used)) {
 785                         d->opd_pre_status = -ENOSPC;
 786                         if (old != -ENOSPC)
 787                                 CDEBUG(D_INFO, "%s: status: "LPU64" blocks, "
 788                                        LPU64" free, "LPU64" used, "LPU64" "
 789                                        "avail -> %d: rc = %d\n",
 790                                        d->opd_obd->obd_name, msfs->os_blocks,
 791                                        msfs->os_bfree, used, msfs->os_bavail,
 792                                        d->opd_pre_status, rc);
 793                         CDEBUG(D_INFO,
 794                                "non-commited changes: %lu, in progress: %u\n",
 795                                d->opd_syn_changes, d->opd_syn_rpc_in_progress);
 796                 } else if (old == -ENOSPC) {
 797                         d->opd_pre_status = 0;
 798                         spin_lock(&d->opd_pre_lock);
 799                         d->opd_pre_grow_slow = 0;
 800                         d->opd_pre_grow_count = OST_MIN_PRECREATE;
 801                         spin_unlock(&d->opd_pre_lock);
 802                         wake_up(&d->opd_pre_waitq);
 803                         CDEBUG(D_INFO, "%s: no space: "LPU64" blocks, "LPU64
 804                                " free, "LPU64" used, "LPU64" avail -> %d: "
 805                                "rc = %d\n", d->opd_obd->obd_name,
 806                                msfs->os_blocks, msfs->os_bfree, used,
 807                                msfs->os_bavail, d->opd_pre_status, rc);
 808                 }
 809         }
 810
 811 out:
 812         wake_up(&d->opd_pre_user_waitq);
 813 }
 814
 815 static int osp_init_pre_fid(struct osp_device *osp)
 816 {
 817         struct lu_env           env;
 818         struct osp_thread_info  *osi;
 819         struct lu_client_seq    *cli_seq;
 820         struct lu_fid           *last_fid;
 821         int                     rc;
 822         ENTRY;
 823
 824         /* Return if last_used fid has been initialized */
 825         if (!fid_is_zero(&osp->opd_last_used_fid))
 826                 RETURN(0);
 827
 828         rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
 829         if (rc) {
 830                 CERROR("%s: init env error: rc = %d\n",
 831                        osp->opd_obd->obd_name, rc);
 832                 RETURN(rc);
 833         }
 834
 835         osi = osp_env_info(&env);
 836         last_fid = &osi->osi_fid;
 837         fid_zero(last_fid);
 838         /* For a freshed fs, it will allocate a new sequence first */
 839         if (osp_is_fid_client(osp) && osp->opd_group != 0) {
 840                 cli_seq = osp->opd_obd->u.cli.cl_seq;
 841                 rc = seq_client_get_seq(&env, cli_seq, &last_fid->f_seq);
 842                 if (rc != 0) {
 843                         CERROR("%s: alloc fid error: rc = %d\n",
 844                                osp->opd_obd->obd_name, rc);
 845                         GOTO(out, rc);
 846                 }
 847         } else {
 848                 last_fid->f_seq = fid_idif_seq(0, osp->opd_index);
 849         }
 850         last_fid->f_oid = 1;
 851         last_fid->f_ver = 0;
 852
 853         spin_lock(&osp->opd_pre_lock);
 854         osp->opd_last_used_fid = *last_fid;
 855         osp->opd_pre_used_fid = *last_fid;
 856         osp->opd_pre_last_created_fid = *last_fid;
 857         spin_unlock(&osp->opd_pre_lock);
 858         rc = osp_write_last_oid_seq_files(&env, osp, last_fid, 1);
 859         if (rc != 0) {
 860                 CERROR("%s: write fid error: rc = %d\n",
 861                        osp->opd_obd->obd_name, rc);
 862                 GOTO(out, rc);
 863         }
 864 out:
 865         lu_env_fini(&env);
 866         RETURN(rc);
 867 }
 868
 869 static int osp_precreate_thread(void *_arg)
 870 {
 871         struct osp_device       *d = _arg;
 872         struct ptlrpc_thread    *thread = &d->opd_pre_thread;
 873         struct l_wait_info       lwi = { 0 };
 874         struct lu_env            env;
 875         int                      rc;
 876
 877         ENTRY;
 878
 879         rc = lu_env_init(&env, d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
 880         if (rc) {
 881                 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
 882                        rc);
 883                 RETURN(rc);
 884         }
 885
 886         spin_lock(&d->opd_pre_lock);
 887         thread->t_flags = SVC_RUNNING;
 888         spin_unlock(&d->opd_pre_lock);
 889         wake_up(&thread->t_ctl_waitq);
 890
 891         while (osp_precreate_running(d)) {
 892                 /*
 893                  * need to be connected to OST
 894                  */
 895                 while (osp_precreate_running(d)) {
 896                         l_wait_event(d->opd_pre_waitq,
 897                                      !osp_precreate_running(d) ||
 898                                      d->opd_new_connection,
 899                                      &lwi);
 900
 901                         if (!d->opd_new_connection)
 902                                 continue;
 903
 904                         d->opd_new_connection = 0;
 905                         d->opd_got_disconnected = 0;
 906                         break;
 907                 }
 908
 909                 if (!osp_precreate_running(d))
 910                         break;
 911
 912                 LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
 913                 if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL) {
 914                         /* Get new sequence for client first */
 915                         LASSERT(d->opd_exp != NULL);
 916                         d->opd_obd->u.cli.cl_seq->lcs_exp =
 917                         class_export_get(d->opd_exp);
 918                         rc = osp_init_pre_fid(d);
 919                         if (rc != 0) {
 920                                 class_export_put(d->opd_exp);
 921                                 d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
 922                                 CERROR("%s: init pre fid error: rc = %d\n",
 923                                        d->opd_obd->obd_name, rc);
 924                                 continue;
 925                         }
 926                 }
 927
 928                 osp_statfs_update(d);
 929
 930                 /*
 931                  * Clean up orphans or recreate missing objects.
 932                  */
 933                 rc = osp_precreate_cleanup_orphans(&env, d);
 934                 if (rc != 0)
 935                         continue;
 936                 /*
 937                  * connected, can handle precreates now
 938                  */
 939                 while (osp_precreate_running(d)) {
 940                         l_wait_event(d->opd_pre_waitq,
 941                                      !osp_precreate_running(d) ||
 942                                      osp_precreate_near_empty(&env, d) ||
 943                                      osp_statfs_need_update(d) ||
 944                                      d->opd_got_disconnected, &lwi);
 945
 946                         if (!osp_precreate_running(d))
 947                                 break;
 948
 949                         /* something happened to the connection
 950                          * have to start from the beginning */
 951                         if (d->opd_got_disconnected)
 952                                 break;
 953
 954                         if (osp_statfs_need_update(d))
 955                                 osp_statfs_update(d);
 956
 957                         /* To avoid handling different seq in precreate/orphan
 958                          * cleanup, it will hold precreate until current seq is
 959                          * used up. */
 960                         if (unlikely(osp_precreate_end_seq(&env, d) &&
 961                             !osp_create_end_seq(&env, d)))
 962                                 continue;
 963
 964                         if (unlikely(osp_precreate_end_seq(&env, d) &&
 965                                      osp_create_end_seq(&env, d))) {
 966                                 LCONSOLE_INFO("%s:"LPX64" is used up."
 967                                               " Update to new seq\n",
 968                                               d->opd_obd->obd_name,
 969                                          fid_seq(&d->opd_pre_last_created_fid));
 970                                 rc = osp_precreate_rollover_new_seq(&env, d);
 971                                 if (rc)
 972                                         continue;
 973                         }
 974
 975                         if (osp_precreate_near_empty(&env, d)) {
 976                                 rc = osp_precreate_send(&env, d);
 977                                 /* osp_precreate_send() sets opd_pre_status
 978                                  * in case of error, that prevent the using of
 979                                  * failed device. */
 980                                 if (rc < 0 && rc != -ENOSPC &&
 981                                     rc != -ETIMEDOUT && rc != -ENOTCONN)
 982                                         CERROR("%s: cannot precreate objects:"
 983                                                " rc = %d\n",
 984                                                d->opd_obd->obd_name, rc);
 985                         }
 986                 }
 987         }
 988
 989         thread->t_flags = SVC_STOPPED;
 990         lu_env_fini(&env);
 991         wake_up(&thread->t_ctl_waitq);
 992
 993         RETURN(0);
 994 }
 995
 996 static int osp_precreate_ready_condition(const struct lu_env *env,
 997                                          struct osp_device *d)
 998 {
 999         if (d->opd_pre_recovering)
1000                 return 0;
1001
1002         /* ready if got enough precreated objects */
1003         /* we need to wait for others (opd_pre_reserved) and our object (+1) */
1004         if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d))
1005                 return 1;
1006
1007         /* ready if OST reported no space and no destroys in progress */
1008         if (d->opd_syn_changes + d->opd_syn_rpc_in_progress == 0 &&
1009             d->opd_pre_status == -ENOSPC)
1010                 return 1;
1011
1012         /* Bail out I/O fails to OST */
1013         if (d->opd_pre_status != 0 &&
1014             d->opd_pre_status != -EAGAIN &&
1015             d->opd_pre_status != -ENODEV &&
1016             d->opd_pre_status != -ENOSPC) {
1017                 /* DEBUG LU-3230 */
1018                 if (d->opd_pre_status != -EIO)
1019                         CERROR("%s: precreate failed opd_pre_status %d\n",
1020                                d->opd_obd->obd_name, d->opd_pre_status);
1021                 return 1;
1022         }
1023
1024         return 0;
1025 }
1026
1027 static int osp_precreate_timeout_condition(void *data)
1028 {
1029         struct osp_device *d = data;
1030
1031         LCONSOLE_WARN("%s: slow creates, last="DFID", next="DFID", "
1032                       "reserved="LPU64", syn_changes=%lu, "
1033                       "syn_rpc_in_progress=%d, status=%d\n",
1034                       d->opd_obd->obd_name, PFID(&d->opd_pre_last_created_fid),
1035                       PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
1036                       d->opd_syn_changes, d->opd_syn_rpc_in_progress,
1037                       d->opd_pre_status);
1038
1039         return 1;
1040 }
1041
1042 /*
1043  * called to reserve object in the pool
1044  * return codes:
1045  *  ENOSPC - no space on corresponded OST
1046  *  EAGAIN - precreation is in progress, try later
1047  *  EIO    - no access to OST
1048  */
1049 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d)
1050 {
1051         struct l_wait_info       lwi;
1052         cfs_time_t               expire = cfs_time_shift(obd_timeout);
1053         int                      precreated, rc;
1054
1055         ENTRY;
1056
1057         LASSERTF(osp_objs_precreated(env, d) >= 0, "Last created FID "DFID
1058                  "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
1059                  PFID(&d->opd_pre_used_fid));
1060
1061         /*
1062          * wait till:
1063          *  - preallocation is done
1064          *  - no free space expected soon
1065          *  - can't connect to OST for too long (obd_timeout)
1066          *  - OST can allocate fid sequence.
1067          */
1068         while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
1069                 rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) {
1070
1071                 /*
1072                  * increase number of precreations
1073                  */
1074                 precreated = osp_objs_precreated(env, d);
1075                 if (d->opd_pre_grow_count < d->opd_pre_max_grow_count &&
1076                     d->opd_pre_grow_slow == 0 &&
1077                     precreated <= (d->opd_pre_grow_count / 4 + 1)) {
1078                         spin_lock(&d->opd_pre_lock);
1079                         d->opd_pre_grow_slow = 1;
1080                         d->opd_pre_grow_count *= 2;
1081                         spin_unlock(&d->opd_pre_lock);
1082                 }
1083
1084                 spin_lock(&d->opd_pre_lock);
1085                 precreated = osp_objs_precreated(env, d);
1086                 if (precreated > d->opd_pre_reserved &&
1087                     !d->opd_pre_recovering) {
1088                         d->opd_pre_reserved++;
1089                         spin_unlock(&d->opd_pre_lock);
1090                         rc = 0;
1091
1092                         /* XXX: don't wake up if precreation is in progress */
1093                         if (osp_precreate_near_empty_nolock(env, d) &&
1094                            !osp_precreate_end_seq_nolock(env, d))
1095                                 wake_up(&d->opd_pre_waitq);
1096
1097                         break;
1098                 }
1099                 spin_unlock(&d->opd_pre_lock);
1100
1101                 /*
1102                  * all precreated objects have been used and no-space
1103                  * status leave us no chance to succeed very soon
1104                  * but if there is destroy in progress, then we should
1105                  * wait till that is done - some space might be released
1106                  */
1107                 if (unlikely(rc == -ENOSPC)) {
1108                         if (d->opd_syn_changes) {
1109                                 /* force local commit to release space */
1110                                 dt_commit_async(env, d->opd_storage);
1111                         }
1112                         if (d->opd_syn_rpc_in_progress) {
1113                                 /* just wait till destroys are done */
1114                                 /* see l_wait_even() few lines below */
1115                         }
1116                         if (d->opd_syn_changes +
1117                             d->opd_syn_rpc_in_progress == 0) {
1118                                 /* no hope for free space */
1119                                 break;
1120                         }
1121                 }
1122
1123                 /* XXX: don't wake up if precreation is in progress */
1124                 wake_up(&d->opd_pre_waitq);
1125
1126                 lwi = LWI_TIMEOUT(expire - cfs_time_current(),
1127                                 osp_precreate_timeout_condition, d);
1128                 if (cfs_time_aftereq(cfs_time_current(), expire)) {
1129                         rc = -ETIMEDOUT;
1130                         break;
1131                 }
1132
1133                 l_wait_event(d->opd_pre_user_waitq,
1134                              osp_precreate_ready_condition(env, d), &lwi);
1135         }
1136
1137         RETURN(rc);
1138 }
1139
1140 /*
1141  * this function relies on reservation made before
1142  */
1143 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1144                           struct lu_fid *fid)
1145 {
1146         /* grab next id from the pool */
1147         spin_lock(&d->opd_pre_lock);
1148
1149         LASSERTF(lu_fid_diff(&d->opd_pre_used_fid,
1150                              &d->opd_pre_last_created_fid) < 0,
1151                  "next fid "DFID" last created fid "DFID"\n",
1152                  PFID(&d->opd_pre_used_fid),
1153                  PFID(&d->opd_pre_last_created_fid));
1154
1155         d->opd_pre_used_fid.f_oid++;
1156         memcpy(fid, &d->opd_pre_used_fid, sizeof(*fid));
1157         d->opd_pre_reserved--;
1158         /*
1159          * last_used_id must be changed along with getting new id otherwise
1160          * we might miscalculate gap causing object loss or leak
1161          */
1162         osp_update_last_fid(d, fid);
1163         spin_unlock(&d->opd_pre_lock);
1164
1165         /*
1166          * probably main thread suspended orphan cleanup till
1167          * all reservations are released, see comment in
1168          * osp_precreate_thread() just before orphan cleanup
1169          */
1170         if (unlikely(d->opd_pre_reserved == 0 && d->opd_pre_status))
1171                 wake_up(&d->opd_pre_waitq);
1172
1173         return 0;
1174 }
1175
1176 /*
1177  *
1178  */
1179 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1180                         __u64 size)
1181 {
1182         struct osp_device       *d = lu2osp_dev(dt->do_lu.lo_dev);
1183         struct ptlrpc_request   *req = NULL;
1184         struct obd_import       *imp;
1185         struct ost_body         *body;
1186         struct obdo             *oa = NULL;
1187         int                      rc;
1188
1189         ENTRY;
1190
1191         imp = d->opd_obd->u.cli.cl_import;
1192         LASSERT(imp);
1193
1194         req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1195         if (req == NULL)
1196                 RETURN(-ENOMEM);
1197
1198         /* XXX: capa support? */
1199         /* osc_set_capa_size(req, &RMF_CAPA1, capa); */
1200         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1201         if (rc) {
1202                 ptlrpc_request_free(req);
1203                 RETURN(rc);
1204         }
1205
1206         /*
1207          * XXX: decide how do we do here with resend
1208          * if we don't resend, then client may see wrong file size
1209          * if we do resend, then MDS thread can get stuck for quite long
1210          */
1211         req->rq_no_resend = req->rq_no_delay = 1;
1212
1213         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1214         ptlrpc_at_set_req_timeout(req);
1215
1216         OBD_ALLOC_PTR(oa);
1217         if (oa == NULL)
1218                 GOTO(out, rc = -ENOMEM);
1219
1220         rc = fid_to_ostid(lu_object_fid(&dt->do_lu), &oa->o_oi);
1221         LASSERT(rc == 0);
1222         oa->o_size = size;
1223         oa->o_blocks = OBD_OBJECT_EOF;
1224         oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1225                       OBD_MD_FLID | OBD_MD_FLGROUP;
1226
1227         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1228         LASSERT(body);
1229         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1230
1231         /* XXX: capa support? */
1232         /* osc_pack_capa(req, body, capa); */
1233
1234         ptlrpc_request_set_replen(req);
1235
1236         rc = ptlrpc_queue_wait(req);
1237         if (rc)
1238                 CERROR("can't punch object: %d\n", rc);
1239 out:
1240         ptlrpc_req_finished(req);
1241         if (oa)
1242                 OBD_FREE_PTR(oa);
1243         RETURN(rc);
1244 }
1245
1246 int osp_init_precreate(struct osp_device *d)
1247 {
1248         struct l_wait_info       lwi = { 0 };
1249         struct task_struct              *task;
1250
1251         ENTRY;
1252
1253         /* initially precreation isn't ready */
1254         d->opd_pre_status = -EAGAIN;
1255         fid_zero(&d->opd_pre_used_fid);
1256         d->opd_pre_used_fid.f_oid = 1;
1257         fid_zero(&d->opd_pre_last_created_fid);
1258         d->opd_pre_last_created_fid.f_oid = 1;
1259         d->opd_pre_reserved = 0;
1260         d->opd_got_disconnected = 1;
1261         d->opd_pre_grow_slow = 0;
1262         d->opd_pre_grow_count = OST_MIN_PRECREATE;
1263         d->opd_pre_min_grow_count = OST_MIN_PRECREATE;
1264         d->opd_pre_max_grow_count = OST_MAX_PRECREATE;
1265
1266         spin_lock_init(&d->opd_pre_lock);
1267         init_waitqueue_head(&d->opd_pre_waitq);
1268         init_waitqueue_head(&d->opd_pre_user_waitq);
1269         init_waitqueue_head(&d->opd_pre_thread.t_ctl_waitq);
1270
1271         /*
1272          * Initialize statfs-related things
1273          */
1274         d->opd_statfs_maxage = 5; /* default update interval */
1275         d->opd_statfs_fresh_till = cfs_time_shift(-1000);
1276         CDEBUG(D_OTHER, "current %llu, fresh till %llu\n",
1277                (unsigned long long)cfs_time_current(),
1278                (unsigned long long)d->opd_statfs_fresh_till);
1279         cfs_timer_init(&d->opd_statfs_timer, osp_statfs_timer_cb, d);
1280
1281         /*
1282          * start thread handling precreation and statfs updates
1283          */
1284         task = kthread_run(osp_precreate_thread, d,
1285                                "osp-pre-%u", d->opd_index);
1286         if (IS_ERR(task)) {
1287                 CERROR("can't start precreate thread %ld\n", PTR_ERR(task));
1288                 RETURN(PTR_ERR(task));
1289         }
1290
1291         l_wait_event(d->opd_pre_thread.t_ctl_waitq,
1292                      osp_precreate_running(d) || osp_precreate_stopped(d),
1293                      &lwi);
1294
1295         RETURN(0);
1296 }
1297
1298 void osp_precreate_fini(struct osp_device *d)
1299 {
1300         struct ptlrpc_thread *thread = &d->opd_pre_thread;
1301
1302         ENTRY;
1303
1304         cfs_timer_disarm(&d->opd_statfs_timer);
1305
1306         thread->t_flags = SVC_STOPPING;
1307         wake_up(&d->opd_pre_waitq);
1308
1309         wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
1310
1311         EXIT;
1312 }
1313