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, 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 #ifndef EXPORT_SYMTAB
  46 # define EXPORT_SYMTAB
  47 #endif
  48 #define DEBUG_SUBSYSTEM S_MDS
  49
  50 #include "osp_internal.h"
  51
  52 /*
  53  * there are two specific states to take care about:
  54  *
  55  * = import is disconnected =
  56  *
  57  * = import is inactive =
  58  *   in this case osp_declare_object_create() returns an error
  59  *
  60  */
  61
  62 /*
  63  * statfs
  64  */
  65 static inline int osp_statfs_need_update(struct osp_device *d)
  66 {
  67         return !cfs_time_before(cfs_time_current(),
  68                                 d->opd_statfs_fresh_till);
  69 }
  70
  71 static void osp_statfs_timer_cb(unsigned long _d)
  72 {
  73         struct osp_device *d = (struct osp_device *) _d;
  74
  75         LASSERT(d);
  76         cfs_waitq_signal(&d->opd_pre_waitq);
  77 }
  78
  79 static int osp_statfs_interpret(const struct lu_env *env,
  80                                 struct ptlrpc_request *req,
  81                                 union ptlrpc_async_args *aa, int rc)
  82 {
  83         struct obd_import       *imp = req->rq_import;
  84         struct obd_statfs       *msfs;
  85         struct osp_device       *d;
  86
  87         ENTRY;
  88
  89         aa = ptlrpc_req_async_args(req);
  90         d = aa->pointer_arg[0];
  91         LASSERT(d);
  92
  93         if (rc != 0)
  94                 GOTO(out, rc);
  95
  96         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
  97         if (msfs == NULL)
  98                 GOTO(out, rc = -EPROTO);
  99
 100         d->opd_statfs = *msfs;
 101
 102         osp_pre_update_status(d, rc);
 103
 104         /* schedule next update */
 105         d->opd_statfs_fresh_till = cfs_time_shift(d->opd_statfs_maxage);
 106         cfs_timer_arm(&d->opd_statfs_timer, d->opd_statfs_fresh_till);
 107         d->opd_statfs_update_in_progress = 0;
 108
 109         CDEBUG(D_CACHE, "updated statfs %p\n", d);
 110
 111         RETURN(0);
 112 out:
 113         /* couldn't update statfs, try again as soon as possible */
 114         cfs_waitq_signal(&d->opd_pre_waitq);
 115         if (req->rq_import_generation == imp->imp_generation)
 116                 CDEBUG(D_CACHE, "%s: couldn't update statfs: rc = %d\n",
 117                        d->opd_obd->obd_name, rc);
 118         RETURN(rc);
 119 }
 120
 121 static int osp_statfs_update(struct osp_device *d)
 122 {
 123         struct ptlrpc_request   *req;
 124         struct obd_import       *imp;
 125         union ptlrpc_async_args *aa;
 126         int                      rc;
 127
 128         ENTRY;
 129
 130         CDEBUG(D_CACHE, "going to update statfs\n");
 131
 132         imp = d->opd_obd->u.cli.cl_import;
 133         LASSERT(imp);
 134
 135         req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
 136         if (req == NULL)
 137                 RETURN(-ENOMEM);
 138
 139         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
 140         if (rc) {
 141                 ptlrpc_request_free(req);
 142                 RETURN(rc);
 143         }
 144         ptlrpc_request_set_replen(req);
 145         req->rq_request_portal = OST_CREATE_PORTAL;
 146         ptlrpc_at_set_req_timeout(req);
 147
 148         req->rq_interpret_reply = (ptlrpc_interpterer_t)osp_statfs_interpret;
 149         aa = ptlrpc_req_async_args(req);
 150         aa->pointer_arg[0] = d;
 151
 152         /*
 153          * no updates till reply
 154          */
 155         cfs_timer_disarm(&d->opd_statfs_timer);
 156         d->opd_statfs_fresh_till = cfs_time_shift(obd_timeout * 1000);
 157         d->opd_statfs_update_in_progress = 1;
 158
 159         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
 160
 161         RETURN(0);
 162 }
 163
 164 /*
 165  * XXX: there might be a case where removed object(s) do not add free
 166  * space (empty object). if the number of such deletions is high, then
 167  * we can start to update statfs too often - a rpc storm
 168  * TODO: some throttling is needed
 169  */
 170 void osp_statfs_need_now(struct osp_device *d)
 171 {
 172         if (!d->opd_statfs_update_in_progress) {
 173                 /*
 174                  * if current status is -ENOSPC (lack of free space on OST)
 175                  * then we should poll OST immediately once object destroy
 176                  * is replied
 177                  */
 178                 d->opd_statfs_fresh_till = cfs_time_shift(-1);
 179                 cfs_timer_disarm(&d->opd_statfs_timer);
 180                 cfs_waitq_signal(&d->opd_pre_waitq);
 181         }
 182 }
 183
 184
 185 /*
 186  * OSP tries to maintain pool of available objects so that calls to create
 187  * objects don't block most of time
 188  *
 189  * each time OSP gets connected to OST, we should start from precreation cleanup
 190  */
 191 static inline int osp_precreate_running(struct osp_device *d)
 192 {
 193         return !!(d->opd_pre_thread.t_flags & SVC_RUNNING);
 194 }
 195
 196 static inline int osp_precreate_stopped(struct osp_device *d)
 197 {
 198         return !!(d->opd_pre_thread.t_flags & SVC_STOPPED);
 199 }
 200
 201 static inline int osp_objs_precreated(const struct lu_env *env,
 202                                       struct osp_device *osp)
 203 {
 204         struct lu_fid *fid1 = &osp->opd_pre_last_created_fid;
 205         struct lu_fid *fid2 = &osp->opd_pre_used_fid;
 206
 207         LASSERTF(fid_seq(fid1) == fid_seq(fid2),
 208                  "Created fid"DFID" Next fid "DFID"\n", PFID(fid1), PFID(fid2));
 209
 210         if (fid_is_idif(fid1)) {
 211                 struct ost_id *oi1 = &osp_env_info(env)->osi_oi;
 212                 struct ost_id *oi2 = &osp_env_info(env)->osi_oi2;
 213
 214                 LASSERT(fid_is_idif(fid1) && fid_is_idif(fid2));
 215                 ostid_idif_pack(fid1, oi1);
 216                 ostid_idif_pack(fid2, oi2);
 217                 LASSERT(oi1->oi_id >= oi2->oi_id);
 218
 219                 return oi1->oi_id - oi2->oi_id;
 220         }
 221
 222         return fid_oid(fid1) - fid_oid(fid2);
 223 }
 224
 225 static inline int osp_precreate_near_empty_nolock(const struct lu_env *env,
 226                                                   struct osp_device *d)
 227 {
 228         int window = osp_objs_precreated(env, d);
 229
 230         /* don't consider new precreation till OST is healty and
 231          * has free space */
 232         return ((window - d->opd_pre_reserved < d->opd_pre_grow_count / 2) &&
 233                 (d->opd_pre_status == 0));
 234 }
 235
 236 static inline int osp_precreate_near_empty(const struct lu_env *env,
 237                                            struct osp_device *d)
 238 {
 239         int rc;
 240
 241         /* XXX: do we really need locking here? */
 242         spin_lock(&d->opd_pre_lock);
 243         rc = osp_precreate_near_empty_nolock(env, d);
 244         spin_unlock(&d->opd_pre_lock);
 245         return rc;
 246 }
 247
 248 static inline int osp_create_end_seq(const struct lu_env *env,
 249                                      struct osp_device *osp)
 250 {
 251         struct lu_fid *fid = &osp->opd_pre_used_fid;
 252         int rc;
 253
 254         spin_lock(&osp->opd_pre_lock);
 255         rc = osp_fid_end_seq(env, fid);
 256         spin_unlock(&osp->opd_pre_lock);
 257         return rc;
 258 }
 259
 260 /**
 261  * Write fid into last_oid/last_seq file.
 262  **/
 263 int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp,
 264                                  struct lu_fid *fid, int sync)
 265 {
 266         struct osp_thread_info  *oti = osp_env_info(env);
 267         struct lu_buf      *lb_oid = &oti->osi_lb;
 268         struct lu_buf      *lb_oseq = &oti->osi_lb2;
 269         loff_t             oid_off;
 270         loff_t             oseq_off;
 271         struct thandle    *th;
 272         int                   rc;
 273         ENTRY;
 274
 275         /* Note: through f_oid is only 32bits, it will also write
 276          * 64 bits for oid to keep compatiblity with the previous
 277          * version. */
 278         lb_oid->lb_buf = &fid->f_oid;
 279         lb_oid->lb_len = sizeof(obd_id);
 280         oid_off = sizeof(obd_id) * osp->opd_index;
 281
 282         lb_oseq->lb_buf = &fid->f_seq;
 283         lb_oseq->lb_len = sizeof(obd_id);
 284         oseq_off = sizeof(obd_id) * osp->opd_index;
 285
 286         th = dt_trans_create(env, osp->opd_storage);
 287         if (IS_ERR(th))
 288                 RETURN(PTR_ERR(th));
 289
 290         th->th_sync |= sync;
 291         rc = dt_declare_record_write(env, osp->opd_last_used_oid_file,
 292                                      lb_oid->lb_len, oid_off, th);
 293         if (rc != 0)
 294                 GOTO(out, rc);
 295
 296         rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
 297                                      lb_oseq->lb_len, oseq_off, th);
 298         if (rc != 0)
 299                 GOTO(out, rc);
 300
 301         rc = dt_trans_start_local(env, osp->opd_storage, th);
 302         if (rc != 0)
 303                 GOTO(out, rc);
 304
 305         rc = dt_record_write(env, osp->opd_last_used_oid_file, lb_oid,
 306                              &oid_off, th);
 307         if (rc != 0) {
 308                 CERROR("%s: can not write to last seq file: rc = %d\n",
 309                         osp->opd_obd->obd_name, rc);
 310                 GOTO(out, rc);
 311         }
 312         rc = dt_record_write(env, osp->opd_last_used_seq_file, lb_oseq,
 313                              &oseq_off, th);
 314         if (rc) {
 315                 CERROR("%s: can not write to last seq file: rc = %d\n",
 316                         osp->opd_obd->obd_name, rc);
 317                 GOTO(out, rc);
 318         }
 319 out:
 320         dt_trans_stop(env, osp->opd_storage, th);
 321         RETURN(rc);
 322 }
 323
 324 int osp_precreate_rollover_new_seq(struct lu_env *env, struct osp_device *osp)
 325 {
 326         struct lu_fid   *fid = &osp_env_info(env)->osi_fid;
 327         struct lu_fid   *last_fid = &osp->opd_last_used_fid;
 328         int             rc;
 329         ENTRY;
 330
 331         rc = seq_client_get_seq(env, osp->opd_obd->u.cli.cl_seq, &fid->f_seq);
 332         if (rc != 0) {
 333                 CERROR("%s: alloc fid error: rc = %d\n",
 334                        osp->opd_obd->obd_name, rc);
 335                 RETURN(rc);
 336         }
 337
 338         fid->f_oid = 1;
 339         fid->f_ver = 0;
 340         LASSERTF(fid_seq(fid) != fid_seq(last_fid),
 341                  "fid "DFID", last_fid "DFID"\n", PFID(fid),
 342                  PFID(last_fid));
 343
 344         rc = osp_write_last_oid_seq_files(env, osp, fid, 1);
 345         if (rc != 0) {
 346                 CERROR("%s: Can not update oid/seq file: rc = %d\n",
 347                        osp->opd_obd->obd_name, rc);
 348                 RETURN(rc);
 349         }
 350
 351         LCONSOLE_INFO("%s: update sequence from "LPX64" to "LPX64"\n",
 352                       osp->opd_obd->obd_name, fid_seq(last_fid),
 353                       fid_seq(fid));
 354         /* Update last_xxx to the new seq */
 355         spin_lock(&osp->opd_pre_lock);
 356         osp->opd_last_used_fid = *fid;
 357         osp->opd_gap_start_fid = *fid;
 358         osp->opd_pre_used_fid = *fid;
 359         osp->opd_pre_last_created_fid = *fid;
 360         spin_unlock(&osp->opd_pre_lock);
 361
 362         RETURN(rc);
 363 }
 364
 365 /**
 366  * alloc fids for precreation.
 367  * rc = 0 Success, @grow is the count of real allocation.
 368  * rc = 1 Current seq is used up.
 369  * rc < 0 Other error.
 370  **/
 371 static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp,
 372                               struct lu_fid *fid, int *grow)
 373 {
 374         struct osp_thread_info  *osi = osp_env_info(env);
 375         __u64                   end;
 376         int                     i = 0;
 377
 378         if (fid_is_idif(fid)) {
 379                 struct lu_fid   *last_fid;
 380                 struct ost_id   *oi = &osi->osi_oi;
 381
 382                 spin_lock(&osp->opd_pre_lock);
 383                 last_fid = &osp->opd_pre_last_created_fid;
 384                 ostid_idif_pack(last_fid, oi);
 385                 end = min(oi->oi_id + *grow, IDIF_MAX_OID);
 386                 *grow = end - oi->oi_id;
 387                 oi->oi_id += *grow;
 388                 spin_unlock(&osp->opd_pre_lock);
 389
 390                 if (*grow == 0)
 391                         return 1;
 392
 393                 ostid_idif_unpack(oi, fid, osp->opd_index);
 394                 return 0;
 395         }
 396
 397         spin_lock(&osp->opd_pre_lock);
 398         *fid = osp->opd_pre_last_created_fid;
 399         end = fid->f_oid;
 400         end = min((end + *grow), (__u64)LUSTRE_DATA_SEQ_MAX_WIDTH);
 401         *grow = end - fid->f_oid;
 402         fid->f_oid += end - fid->f_oid;
 403         spin_unlock(&osp->opd_pre_lock);
 404
 405         CDEBUG(D_INFO, "Expect %d, actual %d ["DFID" -- "DFID"]\n",
 406                *grow, i, PFID(fid), PFID(&osp->opd_pre_last_created_fid));
 407
 408         return *grow > 0 ? 0 : 1;
 409 }
 410
 411 static int osp_precreate_send(const struct lu_env *env, struct osp_device *d)
 412 {
 413         struct osp_thread_info  *oti = osp_env_info(env);
 414         struct ptlrpc_request   *req;
 415         struct obd_import       *imp;
 416         struct ost_body         *body;
 417         int                      rc, grow, diff;
 418         struct lu_fid           *fid = &oti->osi_fid;
 419         ENTRY;
 420
 421         /* don't precreate new objects till OST healthy and has free space */
 422         if (unlikely(d->opd_pre_status)) {
 423                 CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
 424                        d->opd_obd->obd_name, d->opd_pre_status);
 425                 RETURN(0);
 426         }
 427
 428         /*
 429          * if not connection/initialization is compeleted, ignore
 430          */
 431         imp = d->opd_obd->u.cli.cl_import;
 432         LASSERT(imp);
 433
 434         req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
 435         if (req == NULL)
 436                 RETURN(-ENOMEM);
 437         req->rq_request_portal = OST_CREATE_PORTAL;
 438         /* we should not resend create request - anyway we will have delorphan
 439          * and kill these objects */
 440         req->rq_no_delay = req->rq_no_resend = 1;
 441
 442         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
 443         if (rc) {
 444                 ptlrpc_request_free(req);
 445                 RETURN(rc);
 446         }
 447
 448         spin_lock(&d->opd_pre_lock);
 449         if (d->opd_pre_grow_count > d->opd_pre_max_grow_count / 2)
 450                 d->opd_pre_grow_count = d->opd_pre_max_grow_count / 2;
 451         grow = d->opd_pre_grow_count;
 452         spin_unlock(&d->opd_pre_lock);
 453
 454         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 455         LASSERT(body);
 456
 457         rc = osp_precreate_fids(env, d, fid, &grow);
 458         if (rc == 1) {
 459                 /* Current seq has been used up*/
 460                 if (!osp_is_fid_client(d)) {
 461                         osp_pre_update_status(d, -ENOSPC);
 462                         rc = -ENOSPC;
 463                 }
 464                 cfs_waitq_signal(&d->opd_pre_waitq);
 465                 GOTO(out_req, rc);
 466         }
 467
 468         if (!osp_is_fid_client(d)) {
 469                 /* Non-FID client will always send seq 0 because of
 470                  * compatiblity */
 471                 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
 472                 fid->f_seq = 0;
 473         }
 474
 475         ostid_fid_pack(fid, &body->oa.o_oi);
 476         body->oa.o_valid = OBD_MD_FLGROUP;
 477
 478         ptlrpc_request_set_replen(req);
 479
 480         rc = ptlrpc_queue_wait(req);
 481         if (rc) {
 482                 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
 483                        rc);
 484                 GOTO(out_req, rc);
 485         }
 486         LASSERT(req->rq_transno == 0);
 487
 488         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 489         if (body == NULL)
 490                 GOTO(out_req, rc = -EPROTO);
 491
 492         fid_ostid_unpack(fid, &body->oa.o_oi, d->opd_index);
 493         LASSERTF(lu_fid_diff(fid, &d->opd_pre_used_fid) > 0,
 494                  "reply fid "DFID" pre used fid "DFID"\n", PFID(fid),
 495                  PFID(&d->opd_pre_used_fid));
 496
 497         CDEBUG(D_HA, "%s: new last_created "DFID"\n", d->opd_obd->obd_name,
 498                PFID(fid));
 499
 500         diff = lu_fid_diff(fid, &d->opd_pre_last_created_fid);
 501
 502         spin_lock(&d->opd_pre_lock);
 503         if (diff < grow) {
 504                 /* the OST has not managed to create all the
 505                  * objects we asked for */
 506                 d->opd_pre_grow_count = max(diff, OST_MIN_PRECREATE);
 507                 d->opd_pre_grow_slow = 1;
 508         } else {
 509                 /* the OST is able to keep up with the work,
 510                  * we could consider increasing grow_count
 511                  * next time if needed */
 512                 d->opd_pre_grow_slow = 0;
 513         }
 514
 515         d->opd_pre_last_created_fid = *fid;
 516         spin_unlock(&d->opd_pre_lock);
 517
 518         CDEBUG(D_OTHER, "current precreated pool: "DFID"-"DFID"\n",
 519                PFID(&d->opd_pre_used_fid), PFID(&d->opd_pre_last_created_fid));
 520 out_req:
 521         /* now we can wakeup all users awaiting for objects */
 522         osp_pre_update_status(d, rc);
 523         cfs_waitq_signal(&d->opd_pre_user_waitq);
 524
 525         ptlrpc_req_finished(req);
 526         RETURN(rc);
 527 }
 528
 529 static int osp_get_lastfid_from_ost(struct osp_device *d)
 530 {
 531         struct ptlrpc_request   *req = NULL;
 532         struct obd_import       *imp;
 533         struct lu_fid           *last_fid = &d->opd_last_used_fid;
 534         char                    *tmp;
 535         int                     rc;
 536         ENTRY;
 537
 538         imp = d->opd_obd->u.cli.cl_import;
 539         LASSERT(imp);
 540
 541         req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
 542         if (req == NULL)
 543                 RETURN(-ENOMEM);
 544
 545         req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
 546                              RCL_CLIENT, sizeof(KEY_LAST_FID));
 547
 548         req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
 549                              RCL_CLIENT, sizeof(*last_fid));
 550
 551         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
 552         if (rc) {
 553                 ptlrpc_request_free(req);
 554                 RETURN(rc);
 555         }
 556
 557         tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
 558         memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
 559
 560         req->rq_no_delay = req->rq_no_resend = 1;
 561         fid_cpu_to_le(last_fid, last_fid);
 562         tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
 563         memcpy(tmp, last_fid, sizeof(*last_fid));
 564         ptlrpc_request_set_replen(req);
 565
 566         rc = ptlrpc_queue_wait(req);
 567         if (rc) {
 568                 /* bad-bad OST.. let sysadm sort this out */
 569                 if (rc == -ENOTSUPP) {
 570                         CERROR("%s: server does not support FID: rc = %d\n",
 571                                d->opd_obd->obd_name, -ENOTSUPP);
 572                 }
 573                 ptlrpc_set_import_active(imp, 0);
 574                 GOTO(out, rc);
 575         }
 576
 577         last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
 578         if (last_fid == NULL || !fid_is_sane(last_fid)) {
 579                 CERROR("%s: Got insane last_fid "DFID"\n",
 580                        d->opd_obd->obd_name, PFID(last_fid));
 581                 GOTO(out, rc = -EPROTO);
 582         }
 583
 584         /* Only update the last used fid, if the OST has objects for
 585          * this sequence, i.e. fid_oid > 0 */
 586         if (fid_oid(last_fid) > 0)
 587                 d->opd_last_used_fid = *last_fid;
 588
 589         CDEBUG(D_HA, "%s: Got insane last_fid "DFID"\n", d->opd_obd->obd_name,
 590                PFID(last_fid));
 591
 592 out:
 593         ptlrpc_req_finished(req);
 594         RETURN(rc);
 595 }
 596
 597 /**
 598  * asks OST to clean precreate orphans
 599  * and gets next id for new objects
 600  */
 601 static int osp_precreate_cleanup_orphans(struct lu_env *env,
 602                                          struct osp_device *d)
 603 {
 604         struct osp_thread_info  *osi = osp_env_info(env);
 605         struct lu_fid           *last_fid = &osi->osi_fid;
 606         struct ptlrpc_request   *req = NULL;
 607         struct obd_import       *imp;
 608         struct ost_body         *body;
 609         struct l_wait_info       lwi = { 0 };
 610         int                      update_status = 0;
 611         int                      rc;
 612         int                      diff;
 613
 614         ENTRY;
 615
 616         /*
 617          * wait for local recovery to finish, so we can cleanup orphans
 618          * orphans are all objects since "last used" (assigned), but
 619          * there might be objects reserved and in some cases they won't
 620          * be used. we can't cleanup them till we're sure they won't be
 621          * used. also can't we allow new reservations because they may
 622          * end up getting orphans being cleaned up below. so we block
 623          * new reservations and wait till all reserved objects either
 624          * user or released.
 625          */
 626         spin_lock(&d->opd_pre_lock);
 627         d->opd_pre_recovering = 1;
 628         spin_unlock(&d->opd_pre_lock);
 629         /*
 630          * The locking above makes sure the opd_pre_reserved check below will
 631          * catch all osp_precreate_reserve() calls who find
 632          * "!opd_pre_recovering".
 633          */
 634         l_wait_event(d->opd_pre_waitq,
 635                      (!d->opd_pre_reserved && d->opd_recovery_completed) ||
 636                      !osp_precreate_running(d) || d->opd_got_disconnected,
 637                      &lwi);
 638         if (!osp_precreate_running(d) || d->opd_got_disconnected)
 639                 GOTO(out, rc = -EAGAIN);
 640
 641         CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
 642                d->opd_obd->obd_name, PFID(&d->opd_last_used_fid));
 643
 644         *last_fid = d->opd_last_used_fid;
 645         /* The OSP should already get the valid seq now */
 646         LASSERT(!fid_is_zero(last_fid));
 647         if (fid_oid(&d->opd_last_used_fid) < 2) {
 648                 /* lastfid looks strange... ask OST */
 649                 rc = osp_get_lastfid_from_ost(d);
 650                 if (rc)
 651                         GOTO(out, rc);
 652         }
 653
 654         imp = d->opd_obd->u.cli.cl_import;
 655         LASSERT(imp);
 656
 657         req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
 658         if (req == NULL)
 659                 GOTO(out, rc = -ENOMEM);
 660
 661         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
 662         if (rc) {
 663                 ptlrpc_request_free(req);
 664                 req = NULL;
 665                 GOTO(out, rc);
 666         }
 667
 668         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 669         if (body == NULL)
 670                 GOTO(out, rc = -EPROTO);
 671
 672         body->oa.o_flags = OBD_FL_DELORPHAN;
 673         body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
 674         if (osp_is_fid_client(d))
 675                 body->oa.o_seq = fid_seq(&d->opd_last_used_fid);
 676         else
 677                 body->oa.o_seq = 0;
 678         /* remove from NEXT after used one */
 679         body->oa.o_id = fid_oid(&d->opd_last_used_fid);
 680
 681         ptlrpc_request_set_replen(req);
 682
 683         /* Don't resend the delorphan req */
 684         req->rq_no_resend = req->rq_no_delay = 1;
 685
 686         rc = ptlrpc_queue_wait(req);
 687         if (rc) {
 688                 update_status = 1;
 689                 GOTO(out, rc);
 690         }
 691
 692         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 693         if (body == NULL)
 694                 GOTO(out, rc = -EPROTO);
 695
 696         /*
 697          * OST provides us with id new pool starts from in body->oa.o_id
 698          */
 699         fid_ostid_unpack(last_fid, &body->oa.o_oi, d->opd_index);
 700         CDEBUG(D_INFO, "%s: last_fid "DFID" server last fid "DFID"\n",
 701                d->opd_obd->obd_name, PFID(&d->opd_last_used_fid),
 702                PFID(last_fid));
 703
 704         spin_lock(&d->opd_pre_lock);
 705         diff = lu_fid_diff(&d->opd_last_used_fid, last_fid);
 706         if (diff > 0) {
 707                 d->opd_pre_grow_count = OST_MIN_PRECREATE + diff;
 708                 d->opd_pre_last_created_fid = d->opd_last_used_fid;
 709         } else {
 710                 d->opd_pre_grow_count = OST_MIN_PRECREATE;
 711                 d->opd_pre_last_created_fid = *last_fid;
 712         }
 713         /*
 714          * This empties the pre-creation pool and effectively blocks any new
 715          * reservations.
 716          */
 717         LASSERT(fid_oid(&d->opd_pre_last_created_fid) <=
 718                 LUSTRE_DATA_SEQ_MAX_WIDTH);
 719         d->opd_pre_used_fid = d->opd_pre_last_created_fid;
 720         d->opd_pre_grow_slow = 0;
 721         spin_unlock(&d->opd_pre_lock);
 722
 723         CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
 724                "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
 725                PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
 726 out:
 727         if (req)
 728                 ptlrpc_req_finished(req);
 729
 730         d->opd_pre_recovering = 0;
 731
 732         /*
 733          * If rc is zero, the pre-creation window should have been emptied.
 734          * Since waking up the herd would be useless without pre-created
 735          * objects, we defer the signal to osp_precreate_send() in that case.
 736          */
 737         if (rc != 0) {
 738                 if (update_status) {
 739                         CERROR("%s: cannot cleanup orphans: rc = %d\n",
 740                                d->opd_obd->obd_name, rc);
 741                         /* we can't proceed from here, OST seem to
 742                          * be in a bad shape, better to wait for
 743                          * a new instance of the server and repeat
 744                          * from the beginning. notify possible waiters
 745                          * this OSP isn't quite functional yet */
 746                         osp_pre_update_status(d, rc);
 747                 } else {
 748                         cfs_waitq_signal(&d->opd_pre_user_waitq);
 749                 }
 750         }
 751
 752         RETURN(rc);
 753 }
 754
 755 /*
 756  * the function updates current precreation status used: functional or not
 757  *
 758  * rc is a last code from the transport, rc == 0 meaning transport works
 759  * well and users of lod can use objects from this OSP
 760  *
 761  * the status depends on current usage of OST
 762  */
 763 void osp_pre_update_status(struct osp_device *d, int rc)
 764 {
 765         struct obd_statfs       *msfs = &d->opd_statfs;
 766         int                      old = d->opd_pre_status;
 767         __u64                    used;
 768
 769         d->opd_pre_status = rc;
 770         if (rc)
 771                 goto out;
 772
 773         /* Add a bit of hysteresis so this flag isn't continually flapping,
 774          * and ensure that new files don't get extremely fragmented due to
 775          * only a small amount of available space in the filesystem.
 776          * We want to set the NOSPC flag when there is less than ~0.1% free
 777          * and clear it when there is at least ~0.2% free space, so:
 778          *                   avail < ~0.1% max          max = avail + used
 779          *            1025 * avail < avail + used       used = blocks - free
 780          *            1024 * avail < used
 781          *            1024 * avail < blocks - free
 782          *                   avail < ((blocks - free) >> 10)
 783          *
 784          * On very large disk, say 16TB 0.1% will be 16 GB. We don't want to
 785          * lose that amount of space so in those cases we report no space left
 786          * if their is less than 1 GB left.                             */
 787         if (likely(msfs->os_type)) {
 788                 used = min_t(__u64, (msfs->os_blocks - msfs->os_bfree) >> 10,
 789                                     1 << 30);
 790                 if ((msfs->os_ffree < 32) || (msfs->os_bavail < used)) {
 791                         d->opd_pre_status = -ENOSPC;
 792                         if (old != -ENOSPC)
 793                                 CDEBUG(D_INFO, "%s: status: "LPU64" blocks, "
 794                                        LPU64" free, "LPU64" used, "LPU64" "
 795                                        "avail -> %d: rc = %d\n",
 796                                        d->opd_obd->obd_name, msfs->os_blocks,
 797                                        msfs->os_bfree, used, msfs->os_bavail,
 798                                        d->opd_pre_status, rc);
 799                         CDEBUG(D_INFO,
 800                                "non-commited changes: %lu, in progress: %u\n",
 801                                d->opd_syn_changes, d->opd_syn_rpc_in_progress);
 802                 } else if (old == -ENOSPC) {
 803                         d->opd_pre_status = 0;
 804                         d->opd_pre_grow_slow = 0;
 805                         d->opd_pre_grow_count = OST_MIN_PRECREATE;
 806                         cfs_waitq_signal(&d->opd_pre_waitq);
 807                         CDEBUG(D_INFO, "%s: no space: "LPU64" blocks, "LPU64
 808                                " free, "LPU64" used, "LPU64" avail -> %d: "
 809                                "rc = %d\n", d->opd_obd->obd_name,
 810                                msfs->os_blocks, msfs->os_bfree, used,
 811                                msfs->os_bavail, d->opd_pre_status, rc);
 812                 }
 813         }
 814
 815 out:
 816         cfs_waitq_signal(&d->opd_pre_user_waitq);
 817 }
 818
 819 static int osp_init_pre_fid(struct osp_device *osp)
 820 {
 821         struct lu_env           env;
 822         struct osp_thread_info  *osi;
 823         struct lu_client_seq    *cli_seq;
 824         struct lu_fid           *last_fid;
 825         int                     rc;
 826         ENTRY;
 827
 828         /* Return if last_used fid has been initialized */
 829         if (!fid_is_zero(&osp->opd_last_used_fid))
 830                 RETURN(0);
 831
 832         rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
 833         if (rc) {
 834                 CERROR("%s: init env error: rc = %d\n",
 835                        osp->opd_obd->obd_name, rc);
 836                 RETURN(rc);
 837         }
 838
 839         osi = osp_env_info(&env);
 840         last_fid = &osi->osi_fid;
 841         fid_zero(last_fid);
 842         /* For a freshed fs, it will allocate a new sequence first */
 843         if (osp_is_fid_client(osp)) {
 844                 cli_seq = osp->opd_obd->u.cli.cl_seq;
 845                 rc = seq_client_get_seq(&env, cli_seq, &last_fid->f_seq);
 846                 if (rc != 0) {
 847                         CERROR("%s: alloc fid error: rc = %d\n",
 848                                osp->opd_obd->obd_name, rc);
 849                         GOTO(out, rc);
 850                 }
 851         } else {
 852                 last_fid->f_seq = fid_idif_seq(1, osp->opd_index);
 853         }
 854         last_fid->f_oid = 1;
 855         last_fid->f_ver = 0;
 856
 857         spin_lock(&osp->opd_pre_lock);
 858         osp->opd_last_used_fid = *last_fid;
 859         osp->opd_pre_used_fid = *last_fid;
 860         osp->opd_pre_last_created_fid = *last_fid;
 861         spin_unlock(&osp->opd_pre_lock);
 862         rc = osp_write_last_oid_seq_files(&env, osp, last_fid, 1);
 863         if (rc != 0) {
 864                 CERROR("%s: write fid error: rc = %d\n",
 865                        osp->opd_obd->obd_name, rc);
 866                 GOTO(out, rc);
 867         }
 868 out:
 869         lu_env_fini(&env);
 870         RETURN(rc);
 871 }
 872
 873 static int osp_precreate_thread(void *_arg)
 874 {
 875         struct osp_device       *d = _arg;
 876         struct ptlrpc_thread    *thread = &d->opd_pre_thread;
 877         struct l_wait_info       lwi = { 0 };
 878         char                     pname[16];
 879         struct lu_env            env;
 880         int                      rc;
 881
 882         ENTRY;
 883
 884         sprintf(pname, "osp-pre-%u", d->opd_index);
 885         cfs_daemonize(pname);
 886
 887         rc = lu_env_init(&env, d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
 888         if (rc) {
 889                 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
 890                        rc);
 891                 RETURN(rc);
 892         }
 893
 894         spin_lock(&d->opd_pre_lock);
 895         thread->t_flags = SVC_RUNNING;
 896         spin_unlock(&d->opd_pre_lock);
 897         cfs_waitq_signal(&thread->t_ctl_waitq);
 898
 899         while (osp_precreate_running(d)) {
 900                 /*
 901                  * need to be connected to OST
 902                  */
 903                 while (osp_precreate_running(d)) {
 904                         l_wait_event(d->opd_pre_waitq,
 905                                      !osp_precreate_running(d) ||
 906                                      d->opd_new_connection,
 907                                      &lwi);
 908
 909                         if (!osp_precreate_running(d))
 910                                 break;
 911
 912                         if (!d->opd_new_connection)
 913                                 continue;
 914
 915                         d->opd_new_connection = 0;
 916                         d->opd_got_disconnected = 0;
 917                         break;
 918                 }
 919
 920                 if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL) {
 921                         /* Get new sequence for client first */
 922                         LASSERT(d->opd_exp != NULL);
 923                         d->opd_obd->u.cli.cl_seq->lcs_exp =
 924                         class_export_get(d->opd_exp);
 925                         rc = osp_init_pre_fid(d);
 926                         if (rc != 0) {
 927                                 class_export_put(d->opd_exp);
 928                                 d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
 929                                 CERROR("%s: init pre fid error: rc = %d\n",
 930                                        d->opd_obd->obd_name, rc);
 931                                 continue;
 932                         }
 933                 }
 934
 935                 osp_statfs_update(d);
 936
 937                 /*
 938                  * Clean up orphans or recreate missing objects.
 939                  */
 940                 rc = osp_precreate_cleanup_orphans(&env, d);
 941                 if (rc != 0)
 942                         continue;
 943                 /*
 944                  * connected, can handle precreates now
 945                  */
 946                 while (osp_precreate_running(d)) {
 947                         l_wait_event(d->opd_pre_waitq,
 948                                      !osp_precreate_running(d) ||
 949                                      osp_precreate_near_empty(&env, d) ||
 950                                      osp_statfs_need_update(d) ||
 951                                      d->opd_got_disconnected, &lwi);
 952
 953                         if (!osp_precreate_running(d))
 954                                 break;
 955
 956                         /* something happened to the connection
 957                          * have to start from the beginning */
 958                         if (d->opd_got_disconnected)
 959                                 break;
 960
 961                         if (osp_statfs_need_update(d))
 962                                 osp_statfs_update(d);
 963
 964                         /* To avoid handling different seq in precreate/orphan
 965                          * cleanup, it will hold precreate until current seq is
 966                          * used up. */
 967                         if (unlikely(osp_precreate_end_seq(&env, d) &&
 968                             !osp_create_end_seq(&env, d)))
 969                                 continue;
 970
 971                         if (unlikely(osp_precreate_end_seq(&env, d) &&
 972                                      osp_create_end_seq(&env, d))) {
 973                                 LCONSOLE_INFO("%s:"LPX64" is used up."
 974                                               " Update to new seq\n",
 975                                               d->opd_obd->obd_name,
 976                                          fid_seq(&d->opd_pre_last_created_fid));
 977                                 rc = osp_precreate_rollover_new_seq(&env, d);
 978                                 if (rc)
 979                                         continue;
 980                         }
 981
 982                         if (osp_precreate_near_empty(&env, d)) {
 983                                 rc = osp_precreate_send(&env, d);
 984                                 /* osp_precreate_send() sets opd_pre_status
 985                                  * in case of error, that prevent the using of
 986                                  * failed device. */
 987                                 if (rc < 0 && rc != -ENOSPC &&
 988                                     rc != -ETIMEDOUT && rc != -ENOTCONN)
 989                                         CERROR("%s: cannot precreate objects:"
 990                                                " rc = %d\n",
 991                                                d->opd_obd->obd_name, rc);
 992                         }
 993                 }
 994         }
 995
 996         thread->t_flags = SVC_STOPPED;
 997         lu_env_fini(&env);
 998         cfs_waitq_signal(&thread->t_ctl_waitq);
 999
1000         RETURN(0);
1001 }
1002
1003 static int osp_precreate_ready_condition(const struct lu_env *env,
1004                                          struct osp_device *d)
1005 {
1006         if (d->opd_pre_recovering)
1007                 return 0;
1008
1009         /* ready if got enough precreated objects */
1010         /* we need to wait for others (opd_pre_reserved) and our object (+1) */
1011         if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d))
1012                 return 1;
1013
1014         /* ready if OST reported no space and no destoys in progress */
1015         if (d->opd_syn_changes + d->opd_syn_rpc_in_progress == 0 &&
1016             d->opd_pre_status == -ENOSPC)
1017                 return 1;
1018
1019         return 0;
1020 }
1021
1022 static int osp_precreate_timeout_condition(void *data)
1023 {
1024         struct osp_device *d = data;
1025
1026         LCONSOLE_WARN("%s: slow creates, last="DFID", next="DFID", "
1027                       "reserved="LPU64", syn_changes=%lu, "
1028                       "syn_rpc_in_progress=%d, status=%d\n",
1029                       d->opd_obd->obd_name, PFID(&d->opd_pre_last_created_fid),
1030                       PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
1031                       d->opd_syn_changes, d->opd_syn_rpc_in_progress,
1032                       d->opd_pre_status);
1033
1034         return 1;
1035 }
1036
1037 /*
1038  * called to reserve object in the pool
1039  * return codes:
1040  *  ENOSPC - no space on corresponded OST
1041  *  EAGAIN - precreation is in progress, try later
1042  *  EIO    - no access to OST
1043  */
1044 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d)
1045 {
1046         struct l_wait_info       lwi;
1047         cfs_time_t               expire = cfs_time_shift(obd_timeout);
1048         int                      precreated, rc;
1049         int                      count = 0;
1050
1051         ENTRY;
1052
1053         LASSERTF(osp_objs_precreated(env, d) >= 0, "Last created FID "DFID
1054                  "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
1055                  PFID(&d->opd_pre_used_fid));
1056
1057         /*
1058          * wait till:
1059          *  - preallocation is done
1060          *  - no free space expected soon
1061          *  - can't connect to OST for too long (obd_timeout)
1062          *  - OST can allocate fid sequence.
1063          */
1064         while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
1065                 rc == -ENODEV || rc == -EAGAIN) {
1066
1067 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 3, 90, 0)
1068                 /*
1069                  * to address Andreas's concern on possible busy-loop
1070                  * between this thread and osp_precreate_send()
1071                  */
1072                 if (unlikely(count++ == 1000)) {
1073                         osp_precreate_timeout_condition(d);
1074                         LBUG();
1075                 }
1076 #endif
1077
1078                 /*
1079                  * increase number of precreations
1080                  */
1081                 precreated = osp_objs_precreated(env, d);
1082                 if (d->opd_pre_grow_count < d->opd_pre_max_grow_count &&
1083                     d->opd_pre_grow_slow == 0 &&
1084                     precreated <= (d->opd_pre_grow_count / 4 + 1)) {
1085                         spin_lock(&d->opd_pre_lock);
1086                         d->opd_pre_grow_slow = 1;
1087                         d->opd_pre_grow_count *= 2;
1088                         spin_unlock(&d->opd_pre_lock);
1089                 }
1090
1091                 spin_lock(&d->opd_pre_lock);
1092                 precreated = osp_objs_precreated(env, d);
1093                 if (precreated > d->opd_pre_reserved &&
1094                     !d->opd_pre_recovering) {
1095                         d->opd_pre_reserved++;
1096                         spin_unlock(&d->opd_pre_lock);
1097                         rc = 0;
1098
1099                         /* XXX: don't wake up if precreation is in progress */
1100                         if (osp_precreate_near_empty_nolock(env, d) &&
1101                            !osp_precreate_end_seq_nolock(env, d))
1102                                 cfs_waitq_signal(&d->opd_pre_waitq);
1103
1104                         break;
1105                 }
1106                 spin_unlock(&d->opd_pre_lock);
1107
1108                 /*
1109                  * all precreated objects have been used and no-space
1110                  * status leave us no chance to succeed very soon
1111                  * but if there is destroy in progress, then we should
1112                  * wait till that is done - some space might be released
1113                  */
1114                 if (unlikely(rc == -ENOSPC)) {
1115                         if (d->opd_syn_changes) {
1116                                 /* force local commit to release space */
1117                                 dt_commit_async(env, d->opd_storage);
1118                         }
1119                         if (d->opd_syn_rpc_in_progress) {
1120                                 /* just wait till destroys are done */
1121                                 /* see l_wait_even() few lines below */
1122                         }
1123                         if (d->opd_syn_changes +
1124                             d->opd_syn_rpc_in_progress == 0) {
1125                                 /* no hope for free space */
1126                                 break;
1127                         }
1128                 }
1129
1130                 /* XXX: don't wake up if precreation is in progress */
1131                 cfs_waitq_signal(&d->opd_pre_waitq);
1132
1133                 lwi = LWI_TIMEOUT(expire - cfs_time_current(),
1134                                 osp_precreate_timeout_condition, d);
1135                 if (cfs_time_aftereq(cfs_time_current(), expire))
1136                         break;
1137
1138                 l_wait_event(d->opd_pre_user_waitq,
1139                              osp_precreate_ready_condition(env, d), &lwi);
1140         }
1141
1142         RETURN(rc);
1143 }
1144
1145 /*
1146  * this function relies on reservation made before
1147  */
1148 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1149                           struct lu_fid *fid)
1150 {
1151         /* grab next id from the pool */
1152         spin_lock(&d->opd_pre_lock);
1153
1154         LASSERTF(lu_fid_diff(&d->opd_pre_used_fid,
1155                              &d->opd_pre_last_created_fid) < 0,
1156                  "next fid "DFID" last created fid "DFID"\n",
1157                  PFID(&d->opd_pre_used_fid),
1158                  PFID(&d->opd_pre_last_created_fid));
1159
1160         d->opd_pre_used_fid.f_oid++;
1161         memcpy(fid, &d->opd_pre_used_fid, sizeof(*fid));
1162         d->opd_pre_reserved--;
1163         /*
1164          * last_used_id must be changed along with getting new id otherwise
1165          * we might miscalculate gap causing object loss or leak
1166          */
1167         osp_update_last_fid(d, fid);
1168         spin_unlock(&d->opd_pre_lock);
1169
1170         /*
1171          * probably main thread suspended orphan cleanup till
1172          * all reservations are released, see comment in
1173          * osp_precreate_thread() just before orphan cleanup
1174          */
1175         if (unlikely(d->opd_pre_reserved == 0 && d->opd_pre_status))
1176                 cfs_waitq_signal(&d->opd_pre_waitq);
1177
1178         return 0;
1179 }
1180
1181 /*
1182  *
1183  */
1184 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1185                         __u64 size)
1186 {
1187         struct osp_device       *d = lu2osp_dev(dt->do_lu.lo_dev);
1188         struct ptlrpc_request   *req = NULL;
1189         struct obd_import       *imp;
1190         struct ost_body         *body;
1191         struct obdo             *oa = NULL;
1192         int                      rc;
1193
1194         ENTRY;
1195
1196         imp = d->opd_obd->u.cli.cl_import;
1197         LASSERT(imp);
1198
1199         req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1200         if (req == NULL)
1201                 RETURN(-ENOMEM);
1202
1203         /* XXX: capa support? */
1204         /* osc_set_capa_size(req, &RMF_CAPA1, capa); */
1205         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1206         if (rc) {
1207                 ptlrpc_request_free(req);
1208                 RETURN(rc);
1209         }
1210
1211         /*
1212          * XXX: decide how do we do here with resend
1213          * if we don't resend, then client may see wrong file size
1214          * if we do resend, then MDS thread can get stuck for quite long
1215          */
1216         req->rq_no_resend = req->rq_no_delay = 1;
1217
1218         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1219         ptlrpc_at_set_req_timeout(req);
1220
1221         OBD_ALLOC_PTR(oa);
1222         if (oa == NULL)
1223                 GOTO(out, rc = -ENOMEM);
1224
1225         rc = fid_ostid_pack(lu_object_fid(&dt->do_lu), &oa->o_oi);
1226         LASSERT(rc == 0);
1227         oa->o_size = size;
1228         oa->o_blocks = OBD_OBJECT_EOF;
1229         oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1230                       OBD_MD_FLID | OBD_MD_FLGROUP;
1231
1232         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1233         LASSERT(body);
1234         lustre_set_wire_obdo(&body->oa, oa);
1235
1236         /* XXX: capa support? */
1237         /* osc_pack_capa(req, body, capa); */
1238
1239         ptlrpc_request_set_replen(req);
1240
1241         rc = ptlrpc_queue_wait(req);
1242         if (rc)
1243                 CERROR("can't punch object: %d\n", rc);
1244 out:
1245         ptlrpc_req_finished(req);
1246         if (oa)
1247                 OBD_FREE_PTR(oa);
1248         RETURN(rc);
1249 }
1250
1251 int osp_init_precreate(struct osp_device *d)
1252 {
1253         struct l_wait_info       lwi = { 0 };
1254         int                      rc;
1255
1256         ENTRY;
1257
1258         /* initially precreation isn't ready */
1259         d->opd_pre_status = -EAGAIN;
1260         fid_zero(&d->opd_pre_used_fid);
1261         d->opd_pre_used_fid.f_oid = 1;
1262         fid_zero(&d->opd_pre_last_created_fid);
1263         d->opd_pre_last_created_fid.f_oid = 1;
1264         d->opd_pre_reserved = 0;
1265         d->opd_got_disconnected = 1;
1266         d->opd_pre_grow_slow = 0;
1267         d->opd_pre_grow_count = OST_MIN_PRECREATE;
1268         d->opd_pre_min_grow_count = OST_MIN_PRECREATE;
1269         d->opd_pre_max_grow_count = OST_MAX_PRECREATE;
1270
1271         spin_lock_init(&d->opd_pre_lock);
1272         cfs_waitq_init(&d->opd_pre_waitq);
1273         cfs_waitq_init(&d->opd_pre_user_waitq);
1274         cfs_waitq_init(&d->opd_pre_thread.t_ctl_waitq);
1275
1276         /*
1277          * Initialize statfs-related things
1278          */
1279         d->opd_statfs_maxage = 5; /* default update interval */
1280         d->opd_statfs_fresh_till = cfs_time_shift(-1000);
1281         CDEBUG(D_OTHER, "current %llu, fresh till %llu\n",
1282                (unsigned long long)cfs_time_current(),
1283                (unsigned long long)d->opd_statfs_fresh_till);
1284         cfs_timer_init(&d->opd_statfs_timer, osp_statfs_timer_cb, d);
1285
1286         /*
1287          * start thread handling precreation and statfs updates
1288          */
1289         rc = cfs_create_thread(osp_precreate_thread, d, 0);
1290         if (rc < 0) {
1291                 CERROR("can't start precreate thread %d\n", rc);
1292                 RETURN(rc);
1293         }
1294
1295         l_wait_event(d->opd_pre_thread.t_ctl_waitq,
1296                      osp_precreate_running(d) || osp_precreate_stopped(d),
1297                      &lwi);
1298
1299         RETURN(0);
1300 }
1301
1302 void osp_precreate_fini(struct osp_device *d)
1303 {
1304         struct ptlrpc_thread *thread = &d->opd_pre_thread;
1305
1306         ENTRY;
1307
1308         cfs_timer_disarm(&d->opd_statfs_timer);
1309
1310         thread->t_flags = SVC_STOPPING;
1311         cfs_waitq_signal(&d->opd_pre_waitq);
1312
1313         cfs_wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
1314
1315         EXIT;
1316 }
1317