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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2016, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/osp/osp_sync.c
  33  *
  34  * Lustre OST Proxy Device
  35  *
  36  * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
  37  * Author: Mikhail Pershin <mike.pershin@intel.com>
  38  * Author: Di Wang <di.wang@intel.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_MDS
  42
  43 #include <linux/kthread.h>
  44
  45 #include <lustre_obdo.h>
  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  **
  61  * Check whether statfs data is expired
  62  *
  63  * OSP device caches statfs data for the target, the function checks
  64  * whether the data is expired or not.
  65  *
  66  * \param[in] d         OSP device
  67  *
  68  * \retval              0 - not expired, 1 - expired
  69  */
  70 static inline int osp_statfs_need_update(struct osp_device *d)
  71 {
  72         return !cfs_time_before(cfs_time_current(),
  73                                 d->opd_statfs_fresh_till);
  74 }
  75
  76 /*
  77  * OSP tries to maintain pool of available objects so that calls to create
  78  * objects don't block most of time
  79  *
  80  * each time OSP gets connected to OST, we should start from precreation cleanup
  81  */
  82 static inline bool osp_precreate_running(struct osp_device *d)
  83 {
  84         return !!(d->opd_pre_thread.t_flags & SVC_RUNNING);
  85 }
  86
  87 static inline bool osp_precreate_stopped(struct osp_device *d)
  88 {
  89         return !!(d->opd_pre_thread.t_flags & SVC_STOPPED);
  90 }
  91
  92 static void osp_statfs_timer_cb(unsigned long _d)
  93 {
  94         struct osp_device *d = (struct osp_device *) _d;
  95
  96         LASSERT(d);
  97         if (d->opd_pre != NULL && osp_precreate_running(d))
  98                 wake_up(&d->opd_pre_waitq);
  99 }
 100
 101 /**
 102  * RPC interpret callback for OST_STATFS RPC
 103  *
 104  * An interpretation callback called by ptlrpc for OST_STATFS RPC when it is
 105  * replied by the target. It's used to maintain statfs cache for the target.
 106  * The function fills data from the reply if successful and schedules another
 107  * update.
 108  *
 109  * \param[in] env       LU environment provided by the caller
 110  * \param[in] req       RPC replied
 111  * \param[in] aa        callback data
 112  * \param[in] rc        RPC result
 113  *
 114  * \retval 0            on success
 115  * \retval negative     negated errno on error
 116  */
 117 static int osp_statfs_interpret(const struct lu_env *env,
 118                                 struct ptlrpc_request *req,
 119                                 union ptlrpc_async_args *aa, int rc)
 120 {
 121         struct obd_import       *imp = req->rq_import;
 122         struct obd_statfs       *msfs;
 123         struct osp_device       *d;
 124
 125         ENTRY;
 126
 127         aa = ptlrpc_req_async_args(req);
 128         d = aa->pointer_arg[0];
 129         LASSERT(d);
 130
 131         if (rc != 0)
 132                 GOTO(out, rc);
 133
 134         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
 135         if (msfs == NULL)
 136                 GOTO(out, rc = -EPROTO);
 137
 138         d->opd_statfs = *msfs;
 139
 140         osp_pre_update_status(d, rc);
 141
 142         /* schedule next update */
 143         d->opd_statfs_fresh_till = cfs_time_shift(d->opd_statfs_maxage);
 144         mod_timer(&d->opd_statfs_timer, d->opd_statfs_fresh_till);
 145         d->opd_statfs_update_in_progress = 0;
 146
 147         CDEBUG(D_CACHE, "updated statfs %p\n", d);
 148
 149         RETURN(0);
 150 out:
 151         /* couldn't update statfs, try again as soon as possible */
 152         if (d->opd_pre != NULL && osp_precreate_running(d))
 153                 wake_up(&d->opd_pre_waitq);
 154
 155         if (req->rq_import_generation == imp->imp_generation)
 156                 CDEBUG(D_CACHE, "%s: couldn't update statfs: rc = %d\n",
 157                        d->opd_obd->obd_name, rc);
 158         RETURN(rc);
 159 }
 160
 161 /**
 162  * Send OST_STATFS RPC
 163  *
 164  * Sends OST_STATFS RPC to refresh cached statfs data for the target.
 165  * Also disables scheduled updates as times OSP may need to refresh
 166  * statfs data before expiration. The function doesn't block, instead
 167  * an interpretation callback osp_statfs_interpret() is used.
 168  *
 169  * \param[in] d         OSP device
 170  */
 171 static int osp_statfs_update(struct osp_device *d)
 172 {
 173         struct ptlrpc_request   *req;
 174         struct obd_import       *imp;
 175         union ptlrpc_async_args *aa;
 176         int                      rc;
 177
 178         ENTRY;
 179
 180         CDEBUG(D_CACHE, "going to update statfs\n");
 181
 182         imp = d->opd_obd->u.cli.cl_import;
 183         LASSERT(imp);
 184
 185         req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
 186         if (req == NULL)
 187                 RETURN(-ENOMEM);
 188
 189         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
 190         if (rc) {
 191                 ptlrpc_request_free(req);
 192                 RETURN(rc);
 193         }
 194         ptlrpc_request_set_replen(req);
 195         req->rq_request_portal = OST_CREATE_PORTAL;
 196         ptlrpc_at_set_req_timeout(req);
 197
 198         req->rq_interpret_reply = (ptlrpc_interpterer_t)osp_statfs_interpret;
 199         aa = ptlrpc_req_async_args(req);
 200         aa->pointer_arg[0] = d;
 201
 202         /*
 203          * no updates till reply
 204          */
 205         del_timer(&d->opd_statfs_timer);
 206         d->opd_statfs_fresh_till = cfs_time_shift(obd_timeout * 1000);
 207         d->opd_statfs_update_in_progress = 1;
 208
 209         ptlrpcd_add_req(req);
 210
 211         RETURN(0);
 212 }
 213
 214 /**
 215  * Schedule an immediate update for statfs data
 216  *
 217  * If cached statfs data claim no free space, but OSP has got a request to
 218  * destroy an object (so release some space probably), then we may need to
 219  * refresh cached statfs data sooner than planned. The function checks there
 220  * is no statfs update going and schedules immediate update if so.
 221  * XXX: there might be a case where removed object(s) do not add free space (empty
 222  * object). If the number of such deletions is high, then we can start to update
 223  * statfs too often causing a RPC storm. some throttling is needed...
 224  *
 225  * \param[in] d         OSP device where statfs data needs to be refreshed
 226  */
 227 void osp_statfs_need_now(struct osp_device *d)
 228 {
 229         if (!d->opd_statfs_update_in_progress) {
 230                 /*
 231                  * if current status is -ENOSPC (lack of free space on OST)
 232                  * then we should poll OST immediately once object destroy
 233                  * is replied
 234                  */
 235                 d->opd_statfs_fresh_till = cfs_time_shift(-1);
 236                 del_timer(&d->opd_statfs_timer);
 237                 wake_up(&d->opd_pre_waitq);
 238         }
 239 }
 240
 241 /**
 242  * Return number of precreated objects
 243  *
 244  * A simple helper to calculate the number of precreated objects on the device.
 245  *
 246  * \param[in] env       LU environment provided by the caller
 247  * \param[in] osp       OSP device
 248  *
 249  * \retval              the number of the precreated objects
 250  */
 251 static inline int osp_objs_precreated(const struct lu_env *env,
 252                                       struct osp_device *osp)
 253 {
 254         return osp_fid_diff(&osp->opd_pre_last_created_fid,
 255                             &osp->opd_pre_used_fid);
 256 }
 257
 258 /**
 259  * Check pool of precreated objects is nearly empty
 260  *
 261  * We should not wait till the pool of the precreated objects is exhausted,
 262  * because then there will be a long period of OSP being unavailable for the
 263  * new creations due to lenghty precreate RPC. Instead we ask for another
 264  * precreation ahead and hopefully have it ready before the current pool is
 265  * empty. Notice this function relies on an external locking.
 266  *
 267  * \param[in] env       LU environment provided by the caller
 268  * \param[in] d         OSP device
 269  *
 270  * \retval              0 - current pool is good enough, 1 - time to precreate
 271  */
 272 static inline int osp_precreate_near_empty_nolock(const struct lu_env *env,
 273                                                   struct osp_device *d)
 274 {
 275         int window = osp_objs_precreated(env, d);
 276
 277         /* don't consider new precreation till OST is healty and
 278          * has free space */
 279         return ((window - d->opd_pre_reserved < d->opd_pre_create_count / 2) &&
 280                 (d->opd_pre_status == 0));
 281 }
 282
 283 /**
 284  * Check pool of precreated objects
 285  *
 286  * This is protected version of osp_precreate_near_empty_nolock(), check that
 287  * for the details.
 288  *
 289  * \param[in] env       LU environment provided by the caller
 290  * \param[in] d         OSP device
 291  *
 292  * \retval              0 - current pool is good enough, 1 - time to precreate
 293  */
 294 static inline int osp_precreate_near_empty(const struct lu_env *env,
 295                                            struct osp_device *d)
 296 {
 297         int rc;
 298
 299         /* XXX: do we really need locking here? */
 300         spin_lock(&d->opd_pre_lock);
 301         rc = osp_precreate_near_empty_nolock(env, d);
 302         spin_unlock(&d->opd_pre_lock);
 303         return rc;
 304 }
 305
 306 /**
 307  * Check given sequence is empty
 308  *
 309  * Returns a binary result whether the given sequence has some IDs left
 310  * or not. Find the details in osp_fid_end_seq(). This is a lock protected
 311  * version of that function.
 312  *
 313  * \param[in] env       LU environment provided by the caller
 314  * \param[in] osp       OSP device
 315  *
 316  * \retval              0 - current sequence has no IDs, 1 - otherwise
 317  */
 318 static inline int osp_create_end_seq(const struct lu_env *env,
 319                                      struct osp_device *osp)
 320 {
 321         struct lu_fid *fid = &osp->opd_pre_used_fid;
 322         int rc;
 323
 324         spin_lock(&osp->opd_pre_lock);
 325         rc = osp_fid_end_seq(env, fid);
 326         spin_unlock(&osp->opd_pre_lock);
 327         return rc;
 328 }
 329
 330 /**
 331  * Write FID into into last_oid/last_seq file
 332  *
 333  * The function stores the sequence and the in-sequence id into two dedicated
 334  * files. The sync argument can be used to request synchronous commit, so the
 335  * function won't return until the updates are committed.
 336  *
 337  * \param[in] env       LU environment provided by the caller
 338  * \param[in] osp       OSP device
 339  * \param[in] fid       fid where sequence/id is taken
 340  * \param[in] sync      update mode: 0 - asynchronously, 1 - synchronously
 341  *
 342  * \retval 0            on success
 343  * \retval negative     negated errno on error
 344  **/
 345 int osp_write_last_oid_seq_files(struct lu_env *env, struct osp_device *osp,
 346                                  struct lu_fid *fid, int sync)
 347 {
 348         struct osp_thread_info  *oti = osp_env_info(env);
 349         struct lu_buf      *lb_oid = &oti->osi_lb;
 350         struct lu_buf      *lb_oseq = &oti->osi_lb2;
 351         loff_t             oid_off;
 352         loff_t             oseq_off;
 353         struct thandle    *th;
 354         int                   rc;
 355         ENTRY;
 356
 357         /* Note: through f_oid is only 32 bits, it will also write 64 bits
 358          * for oid to keep compatibility with the previous version. */
 359         lb_oid->lb_buf = &fid->f_oid;
 360         lb_oid->lb_len = sizeof(u64);
 361         oid_off = sizeof(u64) * osp->opd_index;
 362
 363         lb_oseq->lb_buf = &fid->f_seq;
 364         lb_oseq->lb_len = sizeof(u64);
 365         oseq_off = sizeof(u64) * osp->opd_index;
 366
 367         th = dt_trans_create(env, osp->opd_storage);
 368         if (IS_ERR(th))
 369                 RETURN(PTR_ERR(th));
 370
 371         th->th_sync |= sync;
 372         rc = dt_declare_record_write(env, osp->opd_last_used_oid_file,
 373                                      lb_oid, oid_off, th);
 374         if (rc != 0)
 375                 GOTO(out, rc);
 376
 377         rc = dt_declare_record_write(env, osp->opd_last_used_seq_file,
 378                                      lb_oseq, oseq_off, th);
 379         if (rc != 0)
 380                 GOTO(out, rc);
 381
 382         rc = dt_trans_start_local(env, osp->opd_storage, th);
 383         if (rc != 0)
 384                 GOTO(out, rc);
 385
 386         rc = dt_record_write(env, osp->opd_last_used_oid_file, lb_oid,
 387                              &oid_off, th);
 388         if (rc != 0) {
 389                 CERROR("%s: can not write to last seq file: rc = %d\n",
 390                         osp->opd_obd->obd_name, rc);
 391                 GOTO(out, rc);
 392         }
 393         rc = dt_record_write(env, osp->opd_last_used_seq_file, lb_oseq,
 394                              &oseq_off, th);
 395         if (rc) {
 396                 CERROR("%s: can not write to last seq file: rc = %d\n",
 397                         osp->opd_obd->obd_name, rc);
 398                 GOTO(out, rc);
 399         }
 400 out:
 401         dt_trans_stop(env, osp->opd_storage, th);
 402         RETURN(rc);
 403 }
 404
 405 /**
 406  * Switch to another sequence
 407  *
 408  * When a current sequence has no available IDs left, OSP has to switch to
 409  * another new sequence. OSP requests it using the regular FLDB protocol
 410  * and stores synchronously before that is used in precreated. This is needed
 411  * to basically have the sequences referenced (not orphaned), otherwise it's
 412  * possible that OST has some objects precreated and the clients have data
 413  * written to it, but after MDT failover nobody refers those objects and OSP
 414  * has no idea that the sequence need cleanup to be done.
 415  * While this is very expensive operation, it's supposed to happen very very
 416  * infrequently because sequence has 2^32 or 2^48 objects (depending on type)
 417  *
 418  * \param[in] env       LU environment provided by the caller
 419  * \param[in] osp       OSP device
 420  *
 421  * \retval 0            on success
 422  * \retval negative     negated errno on error
 423  */
 424 static int osp_precreate_rollover_new_seq(struct lu_env *env,
 425                                           struct osp_device *osp)
 426 {
 427         struct lu_fid   *fid = &osp_env_info(env)->osi_fid;
 428         struct lu_fid   *last_fid = &osp->opd_last_used_fid;
 429         int             rc;
 430         ENTRY;
 431
 432         rc = seq_client_get_seq(env, osp->opd_obd->u.cli.cl_seq, &fid->f_seq);
 433         if (rc != 0) {
 434                 CERROR("%s: alloc fid error: rc = %d\n",
 435                        osp->opd_obd->obd_name, rc);
 436                 RETURN(rc);
 437         }
 438
 439         fid->f_oid = 1;
 440         fid->f_ver = 0;
 441         LASSERTF(fid_seq(fid) != fid_seq(last_fid),
 442                  "fid "DFID", last_fid "DFID"\n", PFID(fid),
 443                  PFID(last_fid));
 444
 445         rc = osp_write_last_oid_seq_files(env, osp, fid, 1);
 446         if (rc != 0) {
 447                 CERROR("%s: Can not update oid/seq file: rc = %d\n",
 448                        osp->opd_obd->obd_name, rc);
 449                 RETURN(rc);
 450         }
 451
 452         LCONSOLE_INFO("%s: update sequence from %#llx to %#llx\n",
 453                       osp->opd_obd->obd_name, fid_seq(last_fid),
 454                       fid_seq(fid));
 455         /* Update last_xxx to the new seq */
 456         spin_lock(&osp->opd_pre_lock);
 457         osp->opd_last_used_fid = *fid;
 458         osp->opd_gap_start_fid = *fid;
 459         osp->opd_pre_used_fid = *fid;
 460         osp->opd_pre_last_created_fid = *fid;
 461         spin_unlock(&osp->opd_pre_lock);
 462
 463         RETURN(rc);
 464 }
 465
 466 /**
 467  * Find IDs available in current sequence
 468  *
 469  * The function calculates the highest possible ID and the number of IDs
 470  * available in the current sequence OSP is using. The number is limited
 471  * artifically by the caller (grow param) and the number of IDs available
 472  * in the sequence by nature. The function doesn't require an external
 473  * locking.
 474  *
 475  * \param[in] env       LU environment provided by the caller
 476  * \param[in] osp       OSP device
 477  * \param[in] fid       FID the caller wants to start with
 478  * \param[in] grow      how many the caller wants
 479  * \param[out] fid      the highest calculated FID
 480  * \param[out] grow     the number of available IDs calculated
 481  *
 482  * \retval              0 on success, 1 - the sequence is empty
 483  */
 484 static int osp_precreate_fids(const struct lu_env *env, struct osp_device *osp,
 485                               struct lu_fid *fid, int *grow)
 486 {
 487         struct osp_thread_info  *osi = osp_env_info(env);
 488         __u64                   end;
 489         int                     i = 0;
 490
 491         if (fid_is_idif(fid)) {
 492                 struct lu_fid   *last_fid;
 493                 struct ost_id   *oi = &osi->osi_oi;
 494
 495                 spin_lock(&osp->opd_pre_lock);
 496                 last_fid = &osp->opd_pre_last_created_fid;
 497                 fid_to_ostid(last_fid, oi);
 498                 end = min(ostid_id(oi) + *grow, IDIF_MAX_OID);
 499                 *grow = end - ostid_id(oi);
 500                 ostid_set_id(oi, ostid_id(oi) + *grow);
 501                 spin_unlock(&osp->opd_pre_lock);
 502
 503                 if (*grow == 0)
 504                         return 1;
 505
 506                 ostid_to_fid(fid, oi, osp->opd_index);
 507                 return 0;
 508         }
 509
 510         spin_lock(&osp->opd_pre_lock);
 511         *fid = osp->opd_pre_last_created_fid;
 512         end = fid->f_oid;
 513         end = min((end + *grow), (__u64)LUSTRE_DATA_SEQ_MAX_WIDTH);
 514         *grow = end - fid->f_oid;
 515         fid->f_oid += end - fid->f_oid;
 516         spin_unlock(&osp->opd_pre_lock);
 517
 518         CDEBUG(D_INFO, "Expect %d, actual %d ["DFID" -- "DFID"]\n",
 519                *grow, i, PFID(fid), PFID(&osp->opd_pre_last_created_fid));
 520
 521         return *grow > 0 ? 0 : 1;
 522 }
 523
 524 /**
 525  * Prepare and send precreate RPC
 526  *
 527  * The function finds how many objects should be precreated.  Then allocates,
 528  * prepares and schedules precreate RPC synchronously. Upon reply the function
 529  * wake ups the threads waiting for the new objects on this target. If the
 530  * target wasn't able to create all the objects requested, then the next
 531  * precreate will be asking less objects (i.e. slow precreate down).
 532  *
 533  * \param[in] env       LU environment provided by the caller
 534  * \param[in] d         OSP device
 535  *
 536  * \retval 0            on success
 537  * \retval negative     negated errno on error
 538  **/
 539 static int osp_precreate_send(const struct lu_env *env, struct osp_device *d)
 540 {
 541         struct osp_thread_info  *oti = osp_env_info(env);
 542         struct ptlrpc_request   *req;
 543         struct obd_import       *imp;
 544         struct ost_body         *body;
 545         int                      rc, grow, diff;
 546         struct lu_fid           *fid = &oti->osi_fid;
 547         ENTRY;
 548
 549         /* don't precreate new objects till OST healthy and has free space */
 550         if (unlikely(d->opd_pre_status)) {
 551                 CDEBUG(D_INFO, "%s: don't send new precreate: rc = %d\n",
 552                        d->opd_obd->obd_name, d->opd_pre_status);
 553                 RETURN(0);
 554         }
 555
 556         /*
 557          * if not connection/initialization is compeleted, ignore
 558          */
 559         imp = d->opd_obd->u.cli.cl_import;
 560         LASSERT(imp);
 561
 562         req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
 563         if (req == NULL)
 564                 RETURN(-ENOMEM);
 565         req->rq_request_portal = OST_CREATE_PORTAL;
 566         /* we should not resend create request - anyway we will have delorphan
 567          * and kill these objects */
 568         req->rq_no_delay = req->rq_no_resend = 1;
 569
 570         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
 571         if (rc) {
 572                 ptlrpc_request_free(req);
 573                 RETURN(rc);
 574         }
 575
 576         LASSERT(d->opd_pre->osp_pre_delorphan_sent != 0);
 577         spin_lock(&d->opd_pre_lock);
 578         if (d->opd_pre_create_count > d->opd_pre_max_create_count / 2)
 579                 d->opd_pre_create_count = d->opd_pre_max_create_count / 2;
 580         grow = d->opd_pre_create_count;
 581         spin_unlock(&d->opd_pre_lock);
 582
 583         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 584         LASSERT(body);
 585
 586         *fid = d->opd_pre_last_created_fid;
 587         rc = osp_precreate_fids(env, d, fid, &grow);
 588         if (rc == 1) {
 589                 /* Current seq has been used up*/
 590                 if (!osp_is_fid_client(d)) {
 591                         osp_pre_update_status(d, -ENOSPC);
 592                         rc = -ENOSPC;
 593                 }
 594                 wake_up(&d->opd_pre_waitq);
 595                 GOTO(out_req, rc);
 596         }
 597
 598         if (!osp_is_fid_client(d)) {
 599                 /* Non-FID client will always send seq 0 because of
 600                  * compatiblity */
 601                 LASSERTF(fid_is_idif(fid), "Invalid fid "DFID"\n", PFID(fid));
 602                 fid->f_seq = 0;
 603         }
 604
 605         fid_to_ostid(fid, &body->oa.o_oi);
 606         body->oa.o_valid = OBD_MD_FLGROUP;
 607
 608         ptlrpc_request_set_replen(req);
 609
 610         if (OBD_FAIL_CHECK(OBD_FAIL_OSP_FAKE_PRECREATE))
 611                 GOTO(ready, rc = 0);
 612
 613         rc = ptlrpc_queue_wait(req);
 614         if (rc) {
 615                 CERROR("%s: can't precreate: rc = %d\n", d->opd_obd->obd_name,
 616                        rc);
 617                 GOTO(out_req, rc);
 618         }
 619         LASSERT(req->rq_transno == 0);
 620
 621         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 622         if (body == NULL)
 623                 GOTO(out_req, rc = -EPROTO);
 624
 625         ostid_to_fid(fid, &body->oa.o_oi, d->opd_index);
 626
 627 ready:
 628         if (osp_fid_diff(fid, &d->opd_pre_used_fid) <= 0) {
 629                 CERROR("%s: precreate fid "DFID" < local used fid "DFID
 630                        ": rc = %d\n", d->opd_obd->obd_name,
 631                        PFID(fid), PFID(&d->opd_pre_used_fid), -ESTALE);
 632                 GOTO(out_req, rc = -ESTALE);
 633         }
 634
 635         diff = osp_fid_diff(fid, &d->opd_pre_last_created_fid);
 636
 637         spin_lock(&d->opd_pre_lock);
 638         if (diff < grow) {
 639                 /* the OST has not managed to create all the
 640                  * objects we asked for */
 641                 d->opd_pre_create_count = max(diff, OST_MIN_PRECREATE);
 642                 d->opd_pre_create_slow = 1;
 643         } else {
 644                 /* the OST is able to keep up with the work,
 645                  * we could consider increasing create_count
 646                  * next time if needed */
 647                 d->opd_pre_create_slow = 0;
 648         }
 649
 650         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 651         fid_to_ostid(fid, &body->oa.o_oi);
 652
 653         d->opd_pre_last_created_fid = *fid;
 654         spin_unlock(&d->opd_pre_lock);
 655
 656         CDEBUG(D_HA, "%s: current precreated pool: "DFID"-"DFID"\n",
 657                d->opd_obd->obd_name, PFID(&d->opd_pre_used_fid),
 658                PFID(&d->opd_pre_last_created_fid));
 659 out_req:
 660         /* now we can wakeup all users awaiting for objects */
 661         osp_pre_update_status(d, rc);
 662         wake_up(&d->opd_pre_user_waitq);
 663
 664         ptlrpc_req_finished(req);
 665         RETURN(rc);
 666 }
 667
 668 /**
 669  * Get last precreated object from target (OST)
 670  *
 671  * Sends synchronous RPC to the target (OST) to learn the last precreated
 672  * object. This later is used to remove all unused objects (cleanup orphan
 673  * procedure). Also, the next object after one we got will be used as a
 674  * starting point for the new precreates.
 675  *
 676  * \param[in] env       LU environment provided by the caller
 677  * \param[in] d         OSP device
 678  *
 679  * \retval 0            on success
 680  * \retval negative     negated errno on error
 681  **/
 682 static int osp_get_lastfid_from_ost(const struct lu_env *env,
 683                                     struct osp_device *d)
 684 {
 685         struct ptlrpc_request   *req = NULL;
 686         struct obd_import       *imp;
 687         struct lu_fid           *last_fid;
 688         char                    *tmp;
 689         int                     rc;
 690         ENTRY;
 691
 692         imp = d->opd_obd->u.cli.cl_import;
 693         LASSERT(imp);
 694
 695         req = ptlrpc_request_alloc(imp, &RQF_OST_GET_INFO_LAST_FID);
 696         if (req == NULL)
 697                 RETURN(-ENOMEM);
 698
 699         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY, RCL_CLIENT,
 700                              sizeof(KEY_LAST_FID));
 701
 702         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
 703         if (rc) {
 704                 ptlrpc_request_free(req);
 705                 RETURN(rc);
 706         }
 707
 708         tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
 709         memcpy(tmp, KEY_LAST_FID, sizeof(KEY_LAST_FID));
 710
 711         req->rq_no_delay = req->rq_no_resend = 1;
 712         last_fid = req_capsule_client_get(&req->rq_pill, &RMF_FID);
 713         fid_cpu_to_le(last_fid, &d->opd_last_used_fid);
 714
 715         ptlrpc_request_set_replen(req);
 716
 717         rc = ptlrpc_queue_wait(req);
 718         if (rc) {
 719                 /* bad-bad OST.. let sysadm sort this out */
 720                 if (rc == -ENOTSUPP) {
 721                         CERROR("%s: server does not support FID: rc = %d\n",
 722                                d->opd_obd->obd_name, -ENOTSUPP);
 723                 }
 724                 ptlrpc_set_import_active(imp, 0);
 725                 GOTO(out, rc);
 726         }
 727
 728         last_fid = req_capsule_server_get(&req->rq_pill, &RMF_FID);
 729         if (last_fid == NULL) {
 730                 CERROR("%s: Got last_fid failed.\n", d->opd_obd->obd_name);
 731                 GOTO(out, rc = -EPROTO);
 732         }
 733
 734         if (!fid_is_sane(last_fid)) {
 735                 CERROR("%s: Got insane last_fid "DFID"\n",
 736                        d->opd_obd->obd_name, PFID(last_fid));
 737                 GOTO(out, rc = -EPROTO);
 738         }
 739
 740         /* Only update the last used fid, if the OST has objects for
 741          * this sequence, i.e. fid_oid > 0 */
 742         if (fid_oid(last_fid) > 0)
 743                 d->opd_last_used_fid = *last_fid;
 744
 745         CDEBUG(D_HA, "%s: Got last_fid "DFID"\n", d->opd_obd->obd_name,
 746                PFID(last_fid));
 747
 748 out:
 749         ptlrpc_req_finished(req);
 750         RETURN(rc);
 751 }
 752
 753 /**
 754  * Cleanup orphans on OST
 755  *
 756  * This function is called in a contex of a dedicated thread handling
 757  * all the precreation suff. The function waits till local recovery
 758  * is complete, then identify all the unreferenced objects (orphans)
 759  * using the highest ID referenced by a local and the highest object
 760  * precreated by the target. The found range is a subject to removal
 761  * using specially flagged RPC. During this process OSP is marked
 762  * unavailable for new objects.
 763  *
 764  * \param[in] env       LU environment provided by the caller
 765  * \param[in] d         OSP device
 766  *
 767  * \retval 0            on success
 768  * \retval negative     negated errno on error
 769  */
 770 static int osp_precreate_cleanup_orphans(struct lu_env *env,
 771                                          struct osp_device *d)
 772 {
 773         struct osp_thread_info  *osi = osp_env_info(env);
 774         struct lu_fid           *last_fid = &osi->osi_fid;
 775         struct ptlrpc_request   *req = NULL;
 776         struct obd_import       *imp;
 777         struct ost_body         *body;
 778         struct l_wait_info       lwi = { 0 };
 779         int                      update_status = 0;
 780         int                      rc;
 781         int                      diff;
 782         struct lu_fid            fid;
 783
 784         ENTRY;
 785
 786         /*
 787          * wait for local recovery to finish, so we can cleanup orphans.
 788          * orphans are all objects since "last used" (assigned).
 789          * consider reserved objects as created otherwise we can get into
 790          * a livelock when one blocked thread holding a reservation can
 791          * block recovery. see LU-8367 for the details. in some cases this
 792          * can result in gaps (i.e. leaked objects), but we've got LFSCK...
 793          *
 794          * do not allow new reservations because they may end up getting
 795          * orphans being cleaned up below. so we block new reservations.
 796          */
 797         spin_lock(&d->opd_pre_lock);
 798         d->opd_pre_recovering = 1;
 799         spin_unlock(&d->opd_pre_lock);
 800         /*
 801          * The locking above makes sure the opd_pre_reserved check below will
 802          * catch all osp_precreate_reserve() calls who find
 803          * "!opd_pre_recovering".
 804          */
 805         l_wait_event(d->opd_pre_waitq, d->opd_recovery_completed ||
 806                      !osp_precreate_running(d) || d->opd_got_disconnected,
 807                      &lwi);
 808         if (!osp_precreate_running(d) || d->opd_got_disconnected)
 809                 GOTO(out, rc = -EAGAIN);
 810
 811         *last_fid = d->opd_last_used_fid;
 812         /* The OSP should already get the valid seq now */
 813         LASSERT(!fid_is_zero(last_fid));
 814         if (fid_oid(&d->opd_last_used_fid) < 2) {
 815                 /* lastfid looks strange... ask OST */
 816                 rc = osp_get_lastfid_from_ost(env, d);
 817                 if (rc)
 818                         GOTO(out, rc);
 819         }
 820
 821         imp = d->opd_obd->u.cli.cl_import;
 822         LASSERT(imp);
 823
 824         req = ptlrpc_request_alloc(imp, &RQF_OST_CREATE);
 825         if (req == NULL)
 826                 GOTO(out, rc = -ENOMEM);
 827
 828         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
 829         if (rc) {
 830                 ptlrpc_request_free(req);
 831                 req = NULL;
 832                 GOTO(out, rc);
 833         }
 834
 835         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 836         if (body == NULL)
 837                 GOTO(out, rc = -EPROTO);
 838
 839         body->oa.o_flags = OBD_FL_DELORPHAN;
 840         body->oa.o_valid = OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
 841
 842         /* unless this is the very first DELORPHAN (when we really
 843          * can destroy some orphans), just tell OST to recreate
 844          * missing objects in our precreate pool */
 845         spin_lock(&d->opd_pre_lock);
 846         if (d->opd_pre->osp_pre_delorphan_sent)
 847                 fid = d->opd_pre_last_created_fid;
 848         else
 849                 fid = d->opd_last_used_fid;
 850         spin_unlock(&d->opd_pre_lock);
 851         fid_to_ostid(&fid, &body->oa.o_oi);
 852
 853         CDEBUG(D_HA, "%s: going to cleanup orphans since "DFID"\n",
 854                d->opd_obd->obd_name, PFID(&fid));
 855
 856         ptlrpc_request_set_replen(req);
 857
 858         /* Don't resend the delorphan req */
 859         req->rq_no_resend = req->rq_no_delay = 1;
 860
 861         rc = ptlrpc_queue_wait(req);
 862         if (rc) {
 863                 update_status = 1;
 864                 GOTO(out, rc);
 865         }
 866
 867         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 868         if (body == NULL)
 869                 GOTO(out, rc = -EPROTO);
 870
 871         /*
 872          * OST provides us with id new pool starts from in body->oa.o_id
 873          */
 874         ostid_to_fid(last_fid, &body->oa.o_oi, d->opd_index);
 875
 876         spin_lock(&d->opd_pre_lock);
 877         diff = osp_fid_diff(&fid, last_fid);
 878         if (diff > 0) {
 879                 d->opd_pre_create_count = OST_MIN_PRECREATE + diff;
 880                 d->opd_pre_last_created_fid = *last_fid;
 881         } else {
 882                 d->opd_pre_create_count = OST_MIN_PRECREATE;
 883                 d->opd_pre_last_created_fid = *last_fid;
 884         }
 885         /*
 886          * This empties the pre-creation pool and effectively blocks any new
 887          * reservations.
 888          */
 889         LASSERT(fid_oid(&d->opd_pre_last_created_fid) <=
 890                 LUSTRE_DATA_SEQ_MAX_WIDTH);
 891         if (d->opd_pre->osp_pre_delorphan_sent == 0)
 892                 d->opd_pre_used_fid = d->opd_pre_last_created_fid;
 893         d->opd_pre_create_slow = 0;
 894         spin_unlock(&d->opd_pre_lock);
 895         d->opd_pre->osp_pre_delorphan_sent = 1;
 896
 897         CDEBUG(D_HA, "%s: Got last_id "DFID" from OST, last_created "DFID
 898                "last_used is "DFID"\n", d->opd_obd->obd_name, PFID(last_fid),
 899                PFID(&d->opd_pre_last_created_fid), PFID(&d->opd_last_used_fid));
 900 out:
 901         if (req)
 902                 ptlrpc_req_finished(req);
 903
 904         /*
 905          * If rc is zero, the pre-creation window should have been emptied.
 906          * Since waking up the herd would be useless without pre-created
 907          * objects, we defer the signal to osp_precreate_send() in that case.
 908          */
 909         if (rc != 0) {
 910                 if (update_status) {
 911                         CERROR("%s: cannot cleanup orphans: rc = %d\n",
 912                                d->opd_obd->obd_name, rc);
 913                         /* we can't proceed from here, OST seem to
 914                          * be in a bad shape, better to wait for
 915                          * a new instance of the server and repeat
 916                          * from the beginning. notify possible waiters
 917                          * this OSP isn't quite functional yet */
 918                         osp_pre_update_status(d, rc);
 919                 } else {
 920                         wake_up(&d->opd_pre_user_waitq);
 921                 }
 922         } else {
 923                 spin_lock(&d->opd_pre_lock);
 924                 d->opd_pre_recovering = 0;
 925                 spin_unlock(&d->opd_pre_lock);
 926         }
 927
 928         RETURN(rc);
 929 }
 930
 931 /**
 932  * Update precreate status using statfs data
 933  *
 934  * The function decides whether this OSP should be used for new objects.
 935  * IOW, whether this OST is used up or has some free space. Cached statfs
 936  * data is used to make this decision. If the latest result of statfs
 937  * request (rc argument) is not success, then just mark OSP unavailable
 938  * right away.
 939
 940  * Add a bit of hysteresis so this flag isn't continually flapping,
 941  * and ensure that new files don't get extremely fragmented due to
 942  * only a small amount of available space in the filesystem.
 943  * We want to set the ENOSPC when there is less than reserved size
 944  * free and clear it when there is at least 2*reserved size free space.
 945  * the function updates current precreation status used: functional or not
 946  *
 947  * \param[in] d         OSP device
 948  * \param[in] rc        new precreate status for device \a d
 949  *
 950  * \retval 0            on success
 951  * \retval negative     negated errno on error
 952  */
 953 void osp_pre_update_status(struct osp_device *d, int rc)
 954 {
 955         struct obd_statfs       *msfs = &d->opd_statfs;
 956         int                      old = d->opd_pre_status;
 957         __u64                    available;
 958
 959         d->opd_pre_status = rc;
 960         if (rc)
 961                 goto out;
 962
 963         if (likely(msfs->os_type)) {
 964                 if (unlikely(d->opd_reserved_mb_high == 0 &&
 965                              d->opd_reserved_mb_low == 0)) {
 966                         /* Use ~0.1% by default to disable object allocation,
 967                          * and ~0.2% to enable, size in MB, set both watermark
 968                          */
 969                         spin_lock(&d->opd_pre_lock);
 970                         if (d->opd_reserved_mb_high == 0 &&
 971                             d->opd_reserved_mb_low == 0) {
 972                                 d->opd_reserved_mb_low =
 973                                         ((msfs->os_bsize >> 10) *
 974                                         msfs->os_blocks) >> 20;
 975                                 if (d->opd_reserved_mb_low == 0)
 976                                         d->opd_reserved_mb_low = 1;
 977                                 d->opd_reserved_mb_high =
 978                                         (d->opd_reserved_mb_low << 1) + 1;
 979                         }
 980                         spin_unlock(&d->opd_pre_lock);
 981                 }
 982                 /* in MB */
 983                 available = (msfs->os_bavail * (msfs->os_bsize >> 10)) >> 10;
 984                 if (msfs->os_ffree < 32)
 985                         msfs->os_state |= OS_STATE_ENOINO;
 986                 else if (msfs->os_ffree > 64)
 987                         msfs->os_state &= ~OS_STATE_ENOINO;
 988
 989                 if (available < d->opd_reserved_mb_low)
 990                         msfs->os_state |= OS_STATE_ENOSPC;
 991                 else if (available > d->opd_reserved_mb_high)
 992                         msfs->os_state &= ~OS_STATE_ENOSPC;
 993                 if (msfs->os_state & (OS_STATE_ENOINO | OS_STATE_ENOSPC)) {
 994                         d->opd_pre_status = -ENOSPC;
 995                         if (old != -ENOSPC)
 996                                 CDEBUG(D_INFO, "%s: status: %llu blocks, %llu "
 997                                        "free, %llu avail, %llu MB avail, %u "
 998                                        "hwm -> %d: rc = %d\n",
 999                                        d->opd_obd->obd_name, msfs->os_blocks,
1000                                        msfs->os_bfree, msfs->os_bavail,
1001                                        available, d->opd_reserved_mb_high,
1002                                        d->opd_pre_status, rc);
1003                         CDEBUG(D_INFO,
1004                                "non-committed changes: %u, in progress: %u\n",
1005                                atomic_read(&d->opd_syn_changes),
1006                                atomic_read(&d->opd_syn_rpc_in_progress));
1007                 } else if (unlikely(old == -ENOSPC)) {
1008                         d->opd_pre_status = 0;
1009                         spin_lock(&d->opd_pre_lock);
1010                         d->opd_pre_create_slow = 0;
1011                         d->opd_pre_create_count = OST_MIN_PRECREATE;
1012                         spin_unlock(&d->opd_pre_lock);
1013                         wake_up(&d->opd_pre_waitq);
1014
1015                         CDEBUG(D_INFO, "%s: space available: %llu blocks, %llu"
1016                                " free, %llu avail, %lluMB avail, %u lwm"
1017                                " -> %d: rc = %d\n", d->opd_obd->obd_name,
1018                                msfs->os_blocks, msfs->os_bfree, msfs->os_bavail,
1019                                available, d->opd_reserved_mb_low,
1020                                d->opd_pre_status, rc);
1021                 }
1022         }
1023 out:
1024         wake_up(&d->opd_pre_user_waitq);
1025 }
1026
1027 /**
1028  * Initialize FID for precreation
1029  *
1030  * For a just created new target, a new sequence should be taken.
1031  * The function checks there is no IDIF in use (if the target was
1032  * added with the older version of Lustre), then requests a new
1033  * sequence from FLDB using the regular protocol. Then this new
1034  * sequence is stored on a persisten storage synchronously to prevent
1035  * possible object leakage (for the detail see the description for
1036  * osp_precreate_rollover_new_seq()).
1037  *
1038  * \param[in] osp       OSP device
1039  *
1040  * \retval 0            on success
1041  * \retval negative     negated errno on error
1042  */
1043 int osp_init_pre_fid(struct osp_device *osp)
1044 {
1045         struct lu_env           env;
1046         struct osp_thread_info  *osi;
1047         struct lu_client_seq    *cli_seq;
1048         struct lu_fid           *last_fid;
1049         int                     rc;
1050         ENTRY;
1051
1052         LASSERT(osp->opd_pre != NULL);
1053
1054         /* Let's check if the current last_seq/fid is valid,
1055          * otherwise request new sequence from the controller */
1056         if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1057                 /* Non-MDT0 can only use normal sequence for
1058                  * OST objects */
1059                 if (fid_is_norm(&osp->opd_last_used_fid))
1060                         RETURN(0);
1061         } else {
1062                 /* Initially MDT0 will start with IDIF, after
1063                  * that it will request new sequence from the
1064                  * controller */
1065                 if (fid_is_idif(&osp->opd_last_used_fid) ||
1066                     fid_is_norm(&osp->opd_last_used_fid))
1067                         RETURN(0);
1068         }
1069
1070         if (!fid_is_zero(&osp->opd_last_used_fid))
1071                 CWARN("%s: invalid last used fid "DFID
1072                       ", try to get new sequence.\n",
1073                       osp->opd_obd->obd_name,
1074                       PFID(&osp->opd_last_used_fid));
1075
1076         rc = lu_env_init(&env, osp->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1077         if (rc) {
1078                 CERROR("%s: init env error: rc = %d\n",
1079                        osp->opd_obd->obd_name, rc);
1080                 RETURN(rc);
1081         }
1082
1083         osi = osp_env_info(&env);
1084         last_fid = &osi->osi_fid;
1085         fid_zero(last_fid);
1086         /* For a freshed fs, it will allocate a new sequence first */
1087         if (osp_is_fid_client(osp) && osp->opd_group != 0) {
1088                 cli_seq = osp->opd_obd->u.cli.cl_seq;
1089                 rc = seq_client_get_seq(&env, cli_seq, &last_fid->f_seq);
1090                 if (rc != 0) {
1091                         CERROR("%s: alloc fid error: rc = %d\n",
1092                                osp->opd_obd->obd_name, rc);
1093                         GOTO(out, rc);
1094                 }
1095         } else {
1096                 last_fid->f_seq = fid_idif_seq(0, osp->opd_index);
1097         }
1098         last_fid->f_oid = 1;
1099         last_fid->f_ver = 0;
1100
1101         spin_lock(&osp->opd_pre_lock);
1102         osp->opd_last_used_fid = *last_fid;
1103         osp->opd_pre_used_fid = *last_fid;
1104         osp->opd_pre_last_created_fid = *last_fid;
1105         spin_unlock(&osp->opd_pre_lock);
1106         rc = osp_write_last_oid_seq_files(&env, osp, last_fid, 1);
1107         if (rc != 0) {
1108                 CERROR("%s: write fid error: rc = %d\n",
1109                        osp->opd_obd->obd_name, rc);
1110                 GOTO(out, rc);
1111         }
1112 out:
1113         lu_env_fini(&env);
1114         RETURN(rc);
1115 }
1116
1117 /**
1118  * The core of precreate functionality
1119  *
1120  * The function implements the main precreation loop. Basically it
1121  * involves connecting to the target, precerate FID initialization,
1122  * identifying and removing orphans, then serving precreation. As
1123  * part of the latter, the thread is responsible for statfs data
1124  * updates. The precreation is mostly driven by another threads
1125  * asking for new OST objects - those askers wake the thread when
1126  * the number of precreated objects reach low watermark.
1127  * After a disconnect, the sequence above repeats. This is keep going
1128  * until the thread is requested to stop.
1129  *
1130  * \param[in] _arg      private data the thread (OSP device to handle)
1131  *
1132  * \retval 0            on success
1133  * \retval negative     negated errno on error
1134  */
1135 static int osp_precreate_thread(void *_arg)
1136 {
1137         struct osp_device       *d = _arg;
1138         struct ptlrpc_thread    *thread = &d->opd_pre_thread;
1139         struct l_wait_info       lwi = { 0 };
1140         struct l_wait_info       lwi2 = LWI_TIMEOUT(cfs_time_seconds(5),
1141                                                     back_to_sleep, NULL);
1142         struct lu_env            env;
1143         int                      rc;
1144
1145         ENTRY;
1146
1147         rc = lu_env_init(&env, d->opd_dt_dev.dd_lu_dev.ld_type->ldt_ctx_tags);
1148         if (rc) {
1149                 CERROR("%s: init env error: rc = %d\n", d->opd_obd->obd_name,
1150                        rc);
1151                 RETURN(rc);
1152         }
1153
1154         spin_lock(&d->opd_pre_lock);
1155         thread->t_flags = SVC_RUNNING;
1156         spin_unlock(&d->opd_pre_lock);
1157         wake_up(&thread->t_ctl_waitq);
1158
1159         while (osp_precreate_running(d)) {
1160                 /*
1161                  * need to be connected to OST
1162                  */
1163                 while (osp_precreate_running(d)) {
1164                         if (d->opd_pre_recovering &&
1165                             d->opd_imp_connected &&
1166                             !d->opd_got_disconnected)
1167                                 break;
1168                         l_wait_event(d->opd_pre_waitq,
1169                                      !osp_precreate_running(d) ||
1170                                      d->opd_new_connection,
1171                                      &lwi);
1172
1173                         if (!d->opd_new_connection)
1174                                 continue;
1175
1176                         d->opd_new_connection = 0;
1177                         d->opd_got_disconnected = 0;
1178                         break;
1179                 }
1180
1181                 if (!osp_precreate_running(d))
1182                         break;
1183
1184                 LASSERT(d->opd_obd->u.cli.cl_seq != NULL);
1185                 /* Sigh, fid client is not ready yet */
1186                 if (d->opd_obd->u.cli.cl_seq->lcs_exp == NULL)
1187                         continue;
1188
1189                 /* Init fid for osp_precreate if necessary */
1190                 rc = osp_init_pre_fid(d);
1191                 if (rc != 0) {
1192                         class_export_put(d->opd_exp);
1193                         d->opd_obd->u.cli.cl_seq->lcs_exp = NULL;
1194                         CERROR("%s: init pre fid error: rc = %d\n",
1195                                d->opd_obd->obd_name, rc);
1196                         continue;
1197                 }
1198
1199                 if (osp_statfs_update(d)) {
1200                         l_wait_event(d->opd_pre_waitq,
1201                                      !osp_precreate_running(d), &lwi2);
1202                         continue;
1203                 }
1204
1205                 /*
1206                  * Clean up orphans or recreate missing objects.
1207                  */
1208                 rc = osp_precreate_cleanup_orphans(&env, d);
1209                 if (rc != 0) {
1210                         schedule_timeout_interruptible(cfs_time_seconds(1));
1211                         continue;
1212                 }
1213                 /*
1214                  * connected, can handle precreates now
1215                  */
1216                 while (osp_precreate_running(d)) {
1217                         l_wait_event(d->opd_pre_waitq,
1218                                      !osp_precreate_running(d) ||
1219                                      osp_precreate_near_empty(&env, d) ||
1220                                      osp_statfs_need_update(d) ||
1221                                      d->opd_got_disconnected, &lwi);
1222
1223                         if (!osp_precreate_running(d))
1224                                 break;
1225
1226                         /* something happened to the connection
1227                          * have to start from the beginning */
1228                         if (d->opd_got_disconnected)
1229                                 break;
1230
1231                         if (osp_statfs_need_update(d))
1232                                 if (osp_statfs_update(d))
1233                                         break;
1234
1235                         /* To avoid handling different seq in precreate/orphan
1236                          * cleanup, it will hold precreate until current seq is
1237                          * used up. */
1238                         if (unlikely(osp_precreate_end_seq(&env, d) &&
1239                             !osp_create_end_seq(&env, d)))
1240                                 continue;
1241
1242                         if (unlikely(osp_precreate_end_seq(&env, d) &&
1243                                      osp_create_end_seq(&env, d))) {
1244                                 LCONSOLE_INFO("%s:%#llx is used up."
1245                                               " Update to new seq\n",
1246                                               d->opd_obd->obd_name,
1247                                          fid_seq(&d->opd_pre_last_created_fid));
1248                                 rc = osp_precreate_rollover_new_seq(&env, d);
1249                                 if (rc)
1250                                         continue;
1251                         }
1252
1253                         if (osp_precreate_near_empty(&env, d)) {
1254                                 rc = osp_precreate_send(&env, d);
1255                                 /* osp_precreate_send() sets opd_pre_status
1256                                  * in case of error, that prevent the using of
1257                                  * failed device. */
1258                                 if (rc < 0 && rc != -ENOSPC &&
1259                                     rc != -ETIMEDOUT && rc != -ENOTCONN)
1260                                         CERROR("%s: cannot precreate objects:"
1261                                                " rc = %d\n",
1262                                                d->opd_obd->obd_name, rc);
1263                         }
1264                 }
1265         }
1266
1267         thread->t_flags = SVC_STOPPED;
1268         lu_env_fini(&env);
1269         wake_up(&thread->t_ctl_waitq);
1270
1271         RETURN(0);
1272 }
1273
1274 /**
1275  * Check when to stop to wait for precreate objects.
1276  *
1277  * The caller wanting a new OST object can't wait undefinitely. The
1278  * function checks for few conditions including available new OST
1279  * objects, disconnected OST, lack of space with no pending destroys,
1280  * etc. IOW, it checks whether the current OSP state is good to keep
1281  * waiting or it's better to give up.
1282  *
1283  * \param[in] env       LU environment provided by the caller
1284  * \param[in] d         OSP device
1285  *
1286  * \retval              0 - keep waiting, 1 - no luck
1287  */
1288 static int osp_precreate_ready_condition(const struct lu_env *env,
1289                                          struct osp_device *d)
1290 {
1291         if (d->opd_pre_recovering)
1292                 return 0;
1293
1294         /* ready if got enough precreated objects */
1295         /* we need to wait for others (opd_pre_reserved) and our object (+1) */
1296         if (d->opd_pre_reserved + 1 < osp_objs_precreated(env, d))
1297                 return 1;
1298
1299         /* ready if OST reported no space and no destroys in progress */
1300         if (atomic_read(&d->opd_syn_changes) +
1301             atomic_read(&d->opd_syn_rpc_in_progress) == 0 &&
1302             d->opd_pre_status == -ENOSPC)
1303                 return 1;
1304
1305         /* Bail out I/O fails to OST */
1306         if (d->opd_pre_status != 0 &&
1307             d->opd_pre_status != -EAGAIN &&
1308             d->opd_pre_status != -ENODEV &&
1309             d->opd_pre_status != -ENOTCONN &&
1310             d->opd_pre_status != -ENOSPC) {
1311                 /* DEBUG LU-3230 */
1312                 if (d->opd_pre_status != -EIO)
1313                         CERROR("%s: precreate failed opd_pre_status %d\n",
1314                                d->opd_obd->obd_name, d->opd_pre_status);
1315                 return 1;
1316         }
1317
1318         return 0;
1319 }
1320
1321 static int osp_precreate_timeout_condition(void *data)
1322 {
1323         struct osp_device *d = data;
1324
1325         CDEBUG(D_HA, "%s: slow creates, last="DFID", next="DFID", "
1326               "reserved=%llu, syn_changes=%u, "
1327               "syn_rpc_in_progress=%d, status=%d\n",
1328               d->opd_obd->obd_name, PFID(&d->opd_pre_last_created_fid),
1329               PFID(&d->opd_pre_used_fid), d->opd_pre_reserved,
1330               atomic_read(&d->opd_syn_changes),
1331               atomic_read(&d->opd_syn_rpc_in_progress),
1332               d->opd_pre_status);
1333
1334         return 1;
1335 }
1336
1337 /**
1338  * Reserve object in precreate pool
1339  *
1340  * When the caller wants to create a new object on this target (target
1341  * represented by the given OSP), it should declare this intention using
1342  * a regular ->dt_declare_create() OSD API method. Then OSP will be trying
1343  * to reserve an object in the existing precreated pool or wait up to
1344  * obd_timeout for the available object to appear in the pool (a dedicated
1345  * thread will be doing real precreation in background). The object can be
1346  * consumed later with osp_precreate_get_fid() or be released with call to
1347  * lu_object_put(). Notice the function doesn't reserve a specific ID, just
1348  * some ID. The actual ID assignment happen in osp_precreate_get_fid().
1349  * If the space on the target is short and there is a pending object destroy,
1350  * then the function forces local commit to speedup space release (see
1351  * osp_sync.c for the details).
1352  *
1353  * \param[in] env       LU environment provided by the caller
1354  * \param[in] d         OSP device
1355  *
1356  * \retval              0 on success
1357  * \retval              -ENOSPC when no space on OST
1358  * \retval              -EAGAIN try later, slow precreation in progress
1359  * \retval              -EIO when no access to OST
1360  */
1361 int osp_precreate_reserve(const struct lu_env *env, struct osp_device *d)
1362 {
1363         struct l_wait_info       lwi;
1364         cfs_time_t               expire = cfs_time_shift(obd_timeout);
1365         int                      precreated, rc;
1366
1367         ENTRY;
1368
1369         LASSERTF(osp_objs_precreated(env, d) >= 0, "Last created FID "DFID
1370                  "Next FID "DFID"\n", PFID(&d->opd_pre_last_created_fid),
1371                  PFID(&d->opd_pre_used_fid));
1372
1373         /* opd_pre_max_create_count 0 to not use specified OST. */
1374         if (d->opd_pre_max_create_count == 0)
1375                 RETURN(-ENOBUFS);
1376
1377         if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_OSP_PRECREATE_WAIT)) {
1378                 if (d->opd_index == cfs_fail_val)
1379                         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_OSP_PRECREATE_WAIT,
1380                                          obd_timeout);
1381         }
1382
1383         /*
1384          * wait till:
1385          *  - preallocation is done
1386          *  - no free space expected soon
1387          *  - can't connect to OST for too long (obd_timeout)
1388          *  - OST can allocate fid sequence.
1389          */
1390         while ((rc = d->opd_pre_status) == 0 || rc == -ENOSPC ||
1391                 rc == -ENODEV || rc == -EAGAIN || rc == -ENOTCONN) {
1392
1393                 /*
1394                  * increase number of precreations
1395                  */
1396                 precreated = osp_objs_precreated(env, d);
1397                 if (d->opd_pre_create_count < d->opd_pre_max_create_count &&
1398                     d->opd_pre_create_slow == 0 &&
1399                     precreated <= (d->opd_pre_create_count / 4 + 1)) {
1400                         spin_lock(&d->opd_pre_lock);
1401                         d->opd_pre_create_slow = 1;
1402                         d->opd_pre_create_count *= 2;
1403                         spin_unlock(&d->opd_pre_lock);
1404                 }
1405
1406                 spin_lock(&d->opd_pre_lock);
1407                 precreated = osp_objs_precreated(env, d);
1408                 if (precreated > d->opd_pre_reserved &&
1409                     !d->opd_pre_recovering) {
1410                         d->opd_pre_reserved++;
1411                         spin_unlock(&d->opd_pre_lock);
1412                         rc = 0;
1413
1414                         /* XXX: don't wake up if precreation is in progress */
1415                         if (osp_precreate_near_empty_nolock(env, d) &&
1416                            !osp_precreate_end_seq_nolock(env, d))
1417                                 wake_up(&d->opd_pre_waitq);
1418
1419                         break;
1420                 }
1421                 spin_unlock(&d->opd_pre_lock);
1422
1423                 /*
1424                  * all precreated objects have been used and no-space
1425                  * status leave us no chance to succeed very soon
1426                  * but if there is destroy in progress, then we should
1427                  * wait till that is done - some space might be released
1428                  */
1429                 if (unlikely(rc == -ENOSPC)) {
1430                         if (atomic_read(&d->opd_syn_changes)) {
1431                                 /* force local commit to release space */
1432                                 dt_commit_async(env, d->opd_storage);
1433                         }
1434                         if (atomic_read(&d->opd_syn_rpc_in_progress)) {
1435                                 /* just wait till destroys are done */
1436                                 /* see l_wait_even() few lines below */
1437                         }
1438                         if (atomic_read(&d->opd_syn_changes) +
1439                             atomic_read(&d->opd_syn_rpc_in_progress) == 0) {
1440                                 /* no hope for free space */
1441                                 break;
1442                         }
1443                 }
1444
1445                 /* XXX: don't wake up if precreation is in progress */
1446                 wake_up(&d->opd_pre_waitq);
1447
1448                 lwi = LWI_TIMEOUT(expire - cfs_time_current(),
1449                                 osp_precreate_timeout_condition, d);
1450                 if (cfs_time_aftereq(cfs_time_current(), expire)) {
1451                         rc = -ETIMEDOUT;
1452                         break;
1453                 }
1454
1455                 l_wait_event(d->opd_pre_user_waitq,
1456                              osp_precreate_ready_condition(env, d), &lwi);
1457         }
1458
1459         RETURN(rc);
1460 }
1461
1462 /**
1463  * Get a FID from precreation pool
1464  *
1465  * The function is a companion for osp_precreate_reserve() - it assigns
1466  * a specific FID from the precreate. The function should be called only
1467  * if the call to osp_precreate_reserve() was successful. The function
1468  * updates a local storage to remember the highest object ID referenced
1469  * by the node in the given sequence.
1470  *
1471  * A very importan details: this is supposed to be called once the
1472  * transaction is started, so on-disk update will be atomic with the
1473  * data (like LOVEA) refering this object. Then the object won't be leaked:
1474  * either it's referenced by the committed transaction or it's a subject
1475  * to the orphan cleanup procedure.
1476  *
1477  * \param[in] env       LU environment provided by the caller
1478  * \param[in] d         OSP device
1479  * \param[out] fid      generated FID
1480  *
1481  * \retval 0            on success
1482  * \retval negative     negated errno on error
1483  */
1484 int osp_precreate_get_fid(const struct lu_env *env, struct osp_device *d,
1485                           struct lu_fid *fid)
1486 {
1487         /* grab next id from the pool */
1488         spin_lock(&d->opd_pre_lock);
1489
1490         LASSERTF(osp_fid_diff(&d->opd_pre_used_fid,
1491                              &d->opd_pre_last_created_fid) < 0,
1492                  "next fid "DFID" last created fid "DFID"\n",
1493                  PFID(&d->opd_pre_used_fid),
1494                  PFID(&d->opd_pre_last_created_fid));
1495
1496         d->opd_pre_used_fid.f_oid++;
1497         memcpy(fid, &d->opd_pre_used_fid, sizeof(*fid));
1498         d->opd_pre_reserved--;
1499         /*
1500          * last_used_id must be changed along with getting new id otherwise
1501          * we might miscalculate gap causing object loss or leak
1502          */
1503         osp_update_last_fid(d, fid);
1504         spin_unlock(&d->opd_pre_lock);
1505
1506         /*
1507          * probably main thread suspended orphan cleanup till
1508          * all reservations are released, see comment in
1509          * osp_precreate_thread() just before orphan cleanup
1510          */
1511         if (unlikely(d->opd_pre_reserved == 0 && d->opd_pre_status))
1512                 wake_up(&d->opd_pre_waitq);
1513
1514         return 0;
1515 }
1516
1517 /*
1518  * Set size regular attribute on an object
1519  *
1520  * When a striping is created late, it's possible that size is already
1521  * initialized on the file. Then the new striping should inherit size
1522  * from the file. The function sets size on the object using the regular
1523  * protocol (OST_PUNCH).
1524  * XXX: should be re-implemented using OUT ?
1525  *
1526  * \param[in] env       LU environment provided by the caller
1527  * \param[in] dt        object
1528  * \param[in] size      size to set.
1529  *
1530  * \retval 0            on success
1531  * \retval negative     negated errno on error
1532  */
1533 int osp_object_truncate(const struct lu_env *env, struct dt_object *dt,
1534                         __u64 size)
1535 {
1536         struct osp_device       *d = lu2osp_dev(dt->do_lu.lo_dev);
1537         struct ptlrpc_request   *req = NULL;
1538         struct obd_import       *imp;
1539         struct ost_body         *body;
1540         struct obdo             *oa = NULL;
1541         int                      rc;
1542
1543         ENTRY;
1544
1545         imp = d->opd_obd->u.cli.cl_import;
1546         LASSERT(imp);
1547
1548         req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
1549         if (req == NULL)
1550                 RETURN(-ENOMEM);
1551
1552         /* XXX: capa support? */
1553         /* osc_set_capa_size(req, &RMF_CAPA1, capa); */
1554         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
1555         if (rc) {
1556                 ptlrpc_request_free(req);
1557                 RETURN(rc);
1558         }
1559
1560         /*
1561          * XXX: decide how do we do here with resend
1562          * if we don't resend, then client may see wrong file size
1563          * if we do resend, then MDS thread can get stuck for quite long
1564          * and if we don't resend, then client will also get -EWOULDBLOCK !!
1565          * (see LU-7975 and sanity/test_27F use cases)
1566          * but let's decide not to resend/delay this truncate request to OST
1567          * and allow Client to decide to resend, in a less agressive way from
1568          * after_reply(), by returning -EINPROGRESS instead of
1569          * -EAGAIN/-EWOULDBLOCK upon return from ptlrpc_queue_wait() at the
1570          * end of this routine
1571          */
1572         req->rq_no_resend = req->rq_no_delay = 1;
1573
1574         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1575         ptlrpc_at_set_req_timeout(req);
1576
1577         OBD_ALLOC_PTR(oa);
1578         if (oa == NULL)
1579                 GOTO(out, rc = -ENOMEM);
1580
1581         rc = fid_to_ostid(lu_object_fid(&dt->do_lu), &oa->o_oi);
1582         LASSERT(rc == 0);
1583         oa->o_size = size;
1584         oa->o_blocks = OBD_OBJECT_EOF;
1585         oa->o_valid = OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1586                       OBD_MD_FLID | OBD_MD_FLGROUP;
1587
1588         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1589         LASSERT(body);
1590         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1591
1592         /* XXX: capa support? */
1593         /* osc_pack_capa(req, body, capa); */
1594
1595         ptlrpc_request_set_replen(req);
1596
1597         rc = ptlrpc_queue_wait(req);
1598         if (rc) {
1599                 /* -EWOULDBLOCK/-EAGAIN means OST is unreachable at the moment
1600                  * since we have decided not to resend/delay, but this could
1601                  * lead to wrong size to be seen at Client side and even process
1602                  * trying to open to exit/fail if not itself handling -EAGAIN.
1603                  * So it should be better to return -EINPROGRESS instead and
1604                  * leave the decision to resend at Client side in after_reply()
1605                  */
1606                 if (rc == -EWOULDBLOCK) {
1607                         rc = -EINPROGRESS;
1608                         CDEBUG(D_HA, "returning -EINPROGRESS instead of "
1609                                "-EWOULDBLOCK/-EAGAIN to allow Client to "
1610                                "resend\n");
1611                 } else {
1612                         CERROR("can't punch object: %d\n", rc);
1613                 }
1614         }
1615 out:
1616         ptlrpc_req_finished(req);
1617         if (oa)
1618                 OBD_FREE_PTR(oa);
1619         RETURN(rc);
1620 }
1621
1622 /**
1623  * Initialize precreation functionality of OSP
1624  *
1625  * Prepares all the internal structures and starts the precreate thread
1626  *
1627  * \param[in] d         OSP device
1628  *
1629  * \retval 0            on success
1630  * \retval negative     negated errno on error
1631  */
1632 int osp_init_precreate(struct osp_device *d)
1633 {
1634         struct l_wait_info       lwi = { 0 };
1635         struct task_struct              *task;
1636
1637         ENTRY;
1638
1639         OBD_ALLOC_PTR(d->opd_pre);
1640         if (d->opd_pre == NULL)
1641                 RETURN(-ENOMEM);
1642
1643         /* initially precreation isn't ready */
1644         d->opd_pre_status = -EAGAIN;
1645         fid_zero(&d->opd_pre_used_fid);
1646         d->opd_pre_used_fid.f_oid = 1;
1647         fid_zero(&d->opd_pre_last_created_fid);
1648         d->opd_pre_last_created_fid.f_oid = 1;
1649         d->opd_pre_reserved = 0;
1650         d->opd_got_disconnected = 1;
1651         d->opd_pre_create_slow = 0;
1652         d->opd_pre_create_count = OST_MIN_PRECREATE;
1653         d->opd_pre_min_create_count = OST_MIN_PRECREATE;
1654         d->opd_pre_max_create_count = OST_MAX_PRECREATE;
1655         d->opd_reserved_mb_high = 0;
1656         d->opd_reserved_mb_low = 0;
1657
1658         spin_lock_init(&d->opd_pre_lock);
1659         init_waitqueue_head(&d->opd_pre_waitq);
1660         init_waitqueue_head(&d->opd_pre_user_waitq);
1661         init_waitqueue_head(&d->opd_pre_thread.t_ctl_waitq);
1662
1663         /*
1664          * Initialize statfs-related things
1665          */
1666         d->opd_statfs_maxage = 5; /* default update interval */
1667         d->opd_statfs_fresh_till = cfs_time_shift(-1000);
1668         CDEBUG(D_OTHER, "current %llu, fresh till %llu\n",
1669                (unsigned long long)cfs_time_current(),
1670                (unsigned long long)d->opd_statfs_fresh_till);
1671         setup_timer(&d->opd_statfs_timer, osp_statfs_timer_cb,
1672                     (unsigned long)d);
1673
1674         /*
1675          * start thread handling precreation and statfs updates
1676          */
1677         task = kthread_run(osp_precreate_thread, d,
1678                            "osp-pre-%u-%u", d->opd_index, d->opd_group);
1679         if (IS_ERR(task)) {
1680                 CERROR("can't start precreate thread %ld\n", PTR_ERR(task));
1681                 RETURN(PTR_ERR(task));
1682         }
1683
1684         l_wait_event(d->opd_pre_thread.t_ctl_waitq,
1685                      osp_precreate_running(d) || osp_precreate_stopped(d),
1686                      &lwi);
1687
1688         RETURN(0);
1689 }
1690
1691 /**
1692  * Finish precreate functionality of OSP
1693  *
1694  *
1695  * Asks all the activity (the thread, update timer) to stop, then
1696  * wait till that is done.
1697  *
1698  * \param[in] d         OSP device
1699  */
1700 void osp_precreate_fini(struct osp_device *d)
1701 {
1702         struct ptlrpc_thread *thread;
1703
1704         ENTRY;
1705
1706         del_timer(&d->opd_statfs_timer);
1707
1708         if (d->opd_pre == NULL)
1709                 RETURN_EXIT;
1710
1711         thread = &d->opd_pre_thread;
1712
1713         thread->t_flags = SVC_STOPPING;
1714         wake_up(&d->opd_pre_waitq);
1715
1716         wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
1717
1718         OBD_FREE_PTR(d->opd_pre);
1719         d->opd_pre = NULL;
1720
1721         EXIT;
1722 }
1723