lustre/osc/osc_request.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) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 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
  33 #define DEBUG_SUBSYSTEM S_OSC
  34
  35 #include <libcfs/libcfs.h>
  36
  37 #include <lustre/lustre_user.h>
  38
  39 #include <lprocfs_status.h>
  40 #include <lustre_debug.h>
  41 #include <lustre_dlm.h>
  42 #include <lustre_fid.h>
  43 #include <lustre_ha.h>
  44 #include <lustre_ioctl.h>
  45 #include <lustre_net.h>
  46 #include <lustre_obdo.h>
  47 #include <lustre_param.h>
  48 #include <obd.h>
  49 #include <obd_cksum.h>
  50 #include <obd_class.h>
  51
  52 #include "osc_cl_internal.h"
  53 #include "osc_internal.h"
  54
  55 atomic_t osc_pool_req_count;
  56 unsigned int osc_reqpool_maxreqcount;
  57 struct ptlrpc_request_pool *osc_rq_pool;
  58
  59 /* max memory used for request pool, unit is MB */
  60 static unsigned int osc_reqpool_mem_max = 5;
  61 module_param(osc_reqpool_mem_max, uint, 0444);
  62
  63 struct osc_brw_async_args {
  64         struct obdo              *aa_oa;
  65         int                       aa_requested_nob;
  66         int                       aa_nio_count;
  67         u32                       aa_page_count;
  68         int                       aa_resends;
  69         struct brw_page **aa_ppga;
  70         struct client_obd        *aa_cli;
  71         struct list_head          aa_oaps;
  72         struct list_head          aa_exts;
  73 };
  74
  75 #define osc_grant_args osc_brw_async_args
  76
  77 struct osc_setattr_args {
  78         struct obdo             *sa_oa;
  79         obd_enqueue_update_f     sa_upcall;
  80         void                    *sa_cookie;
  81 };
  82
  83 struct osc_fsync_args {
  84         struct osc_object       *fa_obj;
  85         struct obdo             *fa_oa;
  86         obd_enqueue_update_f    fa_upcall;
  87         void                    *fa_cookie;
  88 };
  89
  90 struct osc_ladvise_args {
  91         struct obdo             *la_oa;
  92         obd_enqueue_update_f     la_upcall;
  93         void                    *la_cookie;
  94 };
  95
  96 struct osc_enqueue_args {
  97         struct obd_export       *oa_exp;
  98         enum ldlm_type          oa_type;
  99         enum ldlm_mode          oa_mode;
 100         __u64                   *oa_flags;
 101         osc_enqueue_upcall_f    oa_upcall;
 102         void                    *oa_cookie;
 103         struct ost_lvb          *oa_lvb;
 104         struct lustre_handle    oa_lockh;
 105         unsigned int            oa_agl:1;
 106 };
 107
 108 static void osc_release_ppga(struct brw_page **ppga, size_t count);
 109 static int brw_interpret(const struct lu_env *env, struct ptlrpc_request *req,
 110                          void *data, int rc);
 111
 112 void osc_pack_req_body(struct ptlrpc_request *req, struct obdo *oa)
 113 {
 114         struct ost_body *body;
 115
 116         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 117         LASSERT(body);
 118
 119         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
 120 }
 121
 122 static int osc_getattr(const struct lu_env *env, struct obd_export *exp,
 123                        struct obdo *oa)
 124 {
 125         struct ptlrpc_request   *req;
 126         struct ost_body         *body;
 127         int                      rc;
 128
 129         ENTRY;
 130         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
 131         if (req == NULL)
 132                 RETURN(-ENOMEM);
 133
 134         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
 135         if (rc) {
 136                 ptlrpc_request_free(req);
 137                 RETURN(rc);
 138         }
 139
 140         osc_pack_req_body(req, oa);
 141
 142         ptlrpc_request_set_replen(req);
 143
 144         rc = ptlrpc_queue_wait(req);
 145         if (rc)
 146                 GOTO(out, rc);
 147
 148         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 149         if (body == NULL)
 150                 GOTO(out, rc = -EPROTO);
 151
 152         CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
 153         lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
 154
 155         oa->o_blksize = cli_brw_size(exp->exp_obd);
 156         oa->o_valid |= OBD_MD_FLBLKSZ;
 157
 158         EXIT;
 159 out:
 160         ptlrpc_req_finished(req);
 161
 162         return rc;
 163 }
 164
 165 static int osc_setattr(const struct lu_env *env, struct obd_export *exp,
 166                        struct obdo *oa)
 167 {
 168         struct ptlrpc_request   *req;
 169         struct ost_body         *body;
 170         int                      rc;
 171
 172         ENTRY;
 173         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
 174
 175         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
 176         if (req == NULL)
 177                 RETURN(-ENOMEM);
 178
 179         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
 180         if (rc) {
 181                 ptlrpc_request_free(req);
 182                 RETURN(rc);
 183         }
 184
 185         osc_pack_req_body(req, oa);
 186
 187         ptlrpc_request_set_replen(req);
 188
 189         rc = ptlrpc_queue_wait(req);
 190         if (rc)
 191                 GOTO(out, rc);
 192
 193         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 194         if (body == NULL)
 195                 GOTO(out, rc = -EPROTO);
 196
 197         lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
 198
 199         EXIT;
 200 out:
 201         ptlrpc_req_finished(req);
 202
 203         RETURN(rc);
 204 }
 205
 206 static int osc_setattr_interpret(const struct lu_env *env,
 207                                  struct ptlrpc_request *req,
 208                                  struct osc_setattr_args *sa, int rc)
 209 {
 210         struct ost_body *body;
 211         ENTRY;
 212
 213         if (rc != 0)
 214                 GOTO(out, rc);
 215
 216         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 217         if (body == NULL)
 218                 GOTO(out, rc = -EPROTO);
 219
 220         lustre_get_wire_obdo(&req->rq_import->imp_connect_data, sa->sa_oa,
 221                              &body->oa);
 222 out:
 223         rc = sa->sa_upcall(sa->sa_cookie, rc);
 224         RETURN(rc);
 225 }
 226
 227 int osc_setattr_async(struct obd_export *exp, struct obdo *oa,
 228                       obd_enqueue_update_f upcall, void *cookie,
 229                       struct ptlrpc_request_set *rqset)
 230 {
 231         struct ptlrpc_request   *req;
 232         struct osc_setattr_args *sa;
 233         int                      rc;
 234
 235         ENTRY;
 236
 237         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
 238         if (req == NULL)
 239                 RETURN(-ENOMEM);
 240
 241         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
 242         if (rc) {
 243                 ptlrpc_request_free(req);
 244                 RETURN(rc);
 245         }
 246
 247         osc_pack_req_body(req, oa);
 248
 249         ptlrpc_request_set_replen(req);
 250
 251         /* do mds to ost setattr asynchronously */
 252         if (!rqset) {
 253                 /* Do not wait for response. */
 254                 ptlrpcd_add_req(req);
 255         } else {
 256                 req->rq_interpret_reply =
 257                         (ptlrpc_interpterer_t)osc_setattr_interpret;
 258
 259                 CLASSERT(sizeof(*sa) <= sizeof(req->rq_async_args));
 260                 sa = ptlrpc_req_async_args(req);
 261                 sa->sa_oa = oa;
 262                 sa->sa_upcall = upcall;
 263                 sa->sa_cookie = cookie;
 264
 265                 if (rqset == PTLRPCD_SET)
 266                         ptlrpcd_add_req(req);
 267                 else
 268                         ptlrpc_set_add_req(rqset, req);
 269         }
 270
 271         RETURN(0);
 272 }
 273
 274 static int osc_ladvise_interpret(const struct lu_env *env,
 275                                  struct ptlrpc_request *req,
 276                                  void *arg, int rc)
 277 {
 278         struct osc_ladvise_args *la = arg;
 279         struct ost_body *body;
 280         ENTRY;
 281
 282         if (rc != 0)
 283                 GOTO(out, rc);
 284
 285         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 286         if (body == NULL)
 287                 GOTO(out, rc = -EPROTO);
 288
 289         *la->la_oa = body->oa;
 290 out:
 291         rc = la->la_upcall(la->la_cookie, rc);
 292         RETURN(rc);
 293 }
 294
 295 /**
 296  * If rqset is NULL, do not wait for response. Upcall and cookie could also
 297  * be NULL in this case
 298  */
 299 int osc_ladvise_base(struct obd_export *exp, struct obdo *oa,
 300                      struct ladvise_hdr *ladvise_hdr,
 301                      obd_enqueue_update_f upcall, void *cookie,
 302                      struct ptlrpc_request_set *rqset)
 303 {
 304         struct ptlrpc_request   *req;
 305         struct ost_body         *body;
 306         struct osc_ladvise_args *la;
 307         int                      rc;
 308         struct lu_ladvise       *req_ladvise;
 309         struct lu_ladvise       *ladvise = ladvise_hdr->lah_advise;
 310         int                      num_advise = ladvise_hdr->lah_count;
 311         struct ladvise_hdr      *req_ladvise_hdr;
 312         ENTRY;
 313
 314         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_LADVISE);
 315         if (req == NULL)
 316                 RETURN(-ENOMEM);
 317
 318         req_capsule_set_size(&req->rq_pill, &RMF_OST_LADVISE, RCL_CLIENT,
 319                              num_advise * sizeof(*ladvise));
 320         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_LADVISE);
 321         if (rc != 0) {
 322                 ptlrpc_request_free(req);
 323                 RETURN(rc);
 324         }
 325         req->rq_request_portal = OST_IO_PORTAL;
 326         ptlrpc_at_set_req_timeout(req);
 327
 328         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 329         LASSERT(body);
 330         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
 331                              oa);
 332
 333         req_ladvise_hdr = req_capsule_client_get(&req->rq_pill,
 334                                                  &RMF_OST_LADVISE_HDR);
 335         memcpy(req_ladvise_hdr, ladvise_hdr, sizeof(*ladvise_hdr));
 336
 337         req_ladvise = req_capsule_client_get(&req->rq_pill, &RMF_OST_LADVISE);
 338         memcpy(req_ladvise, ladvise, sizeof(*ladvise) * num_advise);
 339         ptlrpc_request_set_replen(req);
 340
 341         if (rqset == NULL) {
 342                 /* Do not wait for response. */
 343                 ptlrpcd_add_req(req);
 344                 RETURN(0);
 345         }
 346
 347         req->rq_interpret_reply = osc_ladvise_interpret;
 348         CLASSERT(sizeof(*la) <= sizeof(req->rq_async_args));
 349         la = ptlrpc_req_async_args(req);
 350         la->la_oa = oa;
 351         la->la_upcall = upcall;
 352         la->la_cookie = cookie;
 353
 354         if (rqset == PTLRPCD_SET)
 355                 ptlrpcd_add_req(req);
 356         else
 357                 ptlrpc_set_add_req(rqset, req);
 358
 359         RETURN(0);
 360 }
 361
 362 static int osc_create(const struct lu_env *env, struct obd_export *exp,
 363                       struct obdo *oa)
 364 {
 365         struct ptlrpc_request *req;
 366         struct ost_body       *body;
 367         int                    rc;
 368         ENTRY;
 369
 370         LASSERT(oa != NULL);
 371         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
 372         LASSERT(fid_seq_is_echo(ostid_seq(&oa->o_oi)));
 373
 374         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
 375         if (req == NULL)
 376                 GOTO(out, rc = -ENOMEM);
 377
 378         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
 379         if (rc) {
 380                 ptlrpc_request_free(req);
 381                 GOTO(out, rc);
 382         }
 383
 384         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 385         LASSERT(body);
 386
 387         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
 388
 389         ptlrpc_request_set_replen(req);
 390
 391         rc = ptlrpc_queue_wait(req);
 392         if (rc)
 393                 GOTO(out_req, rc);
 394
 395         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 396         if (body == NULL)
 397                 GOTO(out_req, rc = -EPROTO);
 398
 399         CDEBUG(D_INFO, "oa flags %x\n", oa->o_flags);
 400         lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
 401
 402         oa->o_blksize = cli_brw_size(exp->exp_obd);
 403         oa->o_valid |= OBD_MD_FLBLKSZ;
 404
 405         CDEBUG(D_HA, "transno: %lld\n",
 406                lustre_msg_get_transno(req->rq_repmsg));
 407 out_req:
 408         ptlrpc_req_finished(req);
 409 out:
 410         RETURN(rc);
 411 }
 412
 413 int osc_punch_base(struct obd_export *exp, struct obdo *oa,
 414                    obd_enqueue_update_f upcall, void *cookie,
 415                    struct ptlrpc_request_set *rqset)
 416 {
 417         struct ptlrpc_request   *req;
 418         struct osc_setattr_args *sa;
 419         struct ost_body         *body;
 420         int                      rc;
 421         ENTRY;
 422
 423         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
 424         if (req == NULL)
 425                 RETURN(-ENOMEM);
 426
 427         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
 428         if (rc) {
 429                 ptlrpc_request_free(req);
 430                 RETURN(rc);
 431         }
 432         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
 433         ptlrpc_at_set_req_timeout(req);
 434
 435         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 436         LASSERT(body);
 437         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
 438
 439         ptlrpc_request_set_replen(req);
 440
 441         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_setattr_interpret;
 442         CLASSERT(sizeof(*sa) <= sizeof(req->rq_async_args));
 443         sa = ptlrpc_req_async_args(req);
 444         sa->sa_oa = oa;
 445         sa->sa_upcall = upcall;
 446         sa->sa_cookie = cookie;
 447         if (rqset == PTLRPCD_SET)
 448                 ptlrpcd_add_req(req);
 449         else
 450                 ptlrpc_set_add_req(rqset, req);
 451
 452         RETURN(0);
 453 }
 454
 455 static int osc_sync_interpret(const struct lu_env *env,
 456                               struct ptlrpc_request *req,
 457                               void *arg, int rc)
 458 {
 459         struct osc_fsync_args   *fa = arg;
 460         struct ost_body         *body;
 461         struct cl_attr          *attr = &osc_env_info(env)->oti_attr;
 462         unsigned long           valid = 0;
 463         struct cl_object        *obj;
 464         ENTRY;
 465
 466         if (rc != 0)
 467                 GOTO(out, rc);
 468
 469         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 470         if (body == NULL) {
 471                 CERROR("can't unpack ost_body\n");
 472                 GOTO(out, rc = -EPROTO);
 473         }
 474
 475         *fa->fa_oa = body->oa;
 476         obj = osc2cl(fa->fa_obj);
 477
 478         /* Update osc object's blocks attribute */
 479         cl_object_attr_lock(obj);
 480         if (body->oa.o_valid & OBD_MD_FLBLOCKS) {
 481                 attr->cat_blocks = body->oa.o_blocks;
 482                 valid |= CAT_BLOCKS;
 483         }
 484
 485         if (valid != 0)
 486                 cl_object_attr_update(env, obj, attr, valid);
 487         cl_object_attr_unlock(obj);
 488
 489 out:
 490         rc = fa->fa_upcall(fa->fa_cookie, rc);
 491         RETURN(rc);
 492 }
 493
 494 int osc_sync_base(struct osc_object *obj, struct obdo *oa,
 495                   obd_enqueue_update_f upcall, void *cookie,
 496                   struct ptlrpc_request_set *rqset)
 497 {
 498         struct obd_export     *exp = osc_export(obj);
 499         struct ptlrpc_request *req;
 500         struct ost_body       *body;
 501         struct osc_fsync_args *fa;
 502         int                    rc;
 503         ENTRY;
 504
 505         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
 506         if (req == NULL)
 507                 RETURN(-ENOMEM);
 508
 509         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
 510         if (rc) {
 511                 ptlrpc_request_free(req);
 512                 RETURN(rc);
 513         }
 514
 515         /* overload the size and blocks fields in the oa with start/end */
 516         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 517         LASSERT(body);
 518         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
 519
 520         ptlrpc_request_set_replen(req);
 521         req->rq_interpret_reply = osc_sync_interpret;
 522
 523         CLASSERT(sizeof(*fa) <= sizeof(req->rq_async_args));
 524         fa = ptlrpc_req_async_args(req);
 525         fa->fa_obj = obj;
 526         fa->fa_oa = oa;
 527         fa->fa_upcall = upcall;
 528         fa->fa_cookie = cookie;
 529
 530         if (rqset == PTLRPCD_SET)
 531                 ptlrpcd_add_req(req);
 532         else
 533                 ptlrpc_set_add_req(rqset, req);
 534
 535         RETURN (0);
 536 }
 537
 538 /* Find and cancel locally locks matched by @mode in the resource found by
 539  * @objid. Found locks are added into @cancel list. Returns the amount of
 540  * locks added to @cancels list. */
 541 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
 542                                    struct list_head *cancels,
 543                                    enum ldlm_mode mode, __u64 lock_flags)
 544 {
 545         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 546         struct ldlm_res_id res_id;
 547         struct ldlm_resource *res;
 548         int count;
 549         ENTRY;
 550
 551         /* Return, i.e. cancel nothing, only if ELC is supported (flag in
 552          * export) but disabled through procfs (flag in NS).
 553          *
 554          * This distinguishes from a case when ELC is not supported originally,
 555          * when we still want to cancel locks in advance and just cancel them
 556          * locally, without sending any RPC. */
 557         if (exp_connect_cancelset(exp) && !ns_connect_cancelset(ns))
 558                 RETURN(0);
 559
 560         ostid_build_res_name(&oa->o_oi, &res_id);
 561         res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
 562         if (IS_ERR(res))
 563                 RETURN(0);
 564
 565         LDLM_RESOURCE_ADDREF(res);
 566         count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
 567                                            lock_flags, 0, NULL);
 568         LDLM_RESOURCE_DELREF(res);
 569         ldlm_resource_putref(res);
 570         RETURN(count);
 571 }
 572
 573 static int osc_destroy_interpret(const struct lu_env *env,
 574                                  struct ptlrpc_request *req, void *data,
 575                                  int rc)
 576 {
 577         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
 578
 579         atomic_dec(&cli->cl_destroy_in_flight);
 580         wake_up(&cli->cl_destroy_waitq);
 581         return 0;
 582 }
 583
 584 static int osc_can_send_destroy(struct client_obd *cli)
 585 {
 586         if (atomic_inc_return(&cli->cl_destroy_in_flight) <=
 587             cli->cl_max_rpcs_in_flight) {
 588                 /* The destroy request can be sent */
 589                 return 1;
 590         }
 591         if (atomic_dec_return(&cli->cl_destroy_in_flight) <
 592             cli->cl_max_rpcs_in_flight) {
 593                 /*
 594                  * The counter has been modified between the two atomic
 595                  * operations.
 596                  */
 597                 wake_up(&cli->cl_destroy_waitq);
 598         }
 599         return 0;
 600 }
 601
 602 static int osc_destroy(const struct lu_env *env, struct obd_export *exp,
 603                        struct obdo *oa)
 604 {
 605         struct client_obd     *cli = &exp->exp_obd->u.cli;
 606         struct ptlrpc_request *req;
 607         struct ost_body       *body;
 608         struct list_head       cancels = LIST_HEAD_INIT(cancels);
 609         int rc, count;
 610         ENTRY;
 611
 612         if (!oa) {
 613                 CDEBUG(D_INFO, "oa NULL\n");
 614                 RETURN(-EINVAL);
 615         }
 616
 617         count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
 618                                         LDLM_FL_DISCARD_DATA);
 619
 620         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
 621         if (req == NULL) {
 622                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
 623                 RETURN(-ENOMEM);
 624         }
 625
 626         rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
 627                                0, &cancels, count);
 628         if (rc) {
 629                 ptlrpc_request_free(req);
 630                 RETURN(rc);
 631         }
 632
 633         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
 634         ptlrpc_at_set_req_timeout(req);
 635
 636         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 637         LASSERT(body);
 638         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
 639
 640         ptlrpc_request_set_replen(req);
 641
 642         req->rq_interpret_reply = osc_destroy_interpret;
 643         if (!osc_can_send_destroy(cli)) {
 644                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 645
 646                 /*
 647                  * Wait until the number of on-going destroy RPCs drops
 648                  * under max_rpc_in_flight
 649                  */
 650                 l_wait_event_exclusive(cli->cl_destroy_waitq,
 651                                        osc_can_send_destroy(cli), &lwi);
 652         }
 653
 654         /* Do not wait for response */
 655         ptlrpcd_add_req(req);
 656         RETURN(0);
 657 }
 658
 659 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
 660                                 long writing_bytes)
 661 {
 662         u64 bits = OBD_MD_FLBLOCKS | OBD_MD_FLGRANT;
 663
 664         LASSERT(!(oa->o_valid & bits));
 665
 666         oa->o_valid |= bits;
 667         spin_lock(&cli->cl_loi_list_lock);
 668         if (OCD_HAS_FLAG(&cli->cl_import->imp_connect_data, GRANT_PARAM))
 669                 oa->o_dirty = cli->cl_dirty_grant;
 670         else
 671                 oa->o_dirty = cli->cl_dirty_pages << PAGE_SHIFT;
 672         if (unlikely(cli->cl_dirty_pages - cli->cl_dirty_transit >
 673                      cli->cl_dirty_max_pages)) {
 674                 CERROR("dirty %lu - %lu > dirty_max %lu\n",
 675                        cli->cl_dirty_pages, cli->cl_dirty_transit,
 676                        cli->cl_dirty_max_pages);
 677                 oa->o_undirty = 0;
 678         } else if (unlikely(atomic_long_read(&obd_dirty_pages) -
 679                             atomic_long_read(&obd_dirty_transit_pages) >
 680                             (long)(obd_max_dirty_pages + 1))) {
 681                 /* The atomic_read() allowing the atomic_inc() are
 682                  * not covered by a lock thus they may safely race and trip
 683                  * this CERROR() unless we add in a small fudge factor (+1). */
 684                 CERROR("%s: dirty %ld - %ld > system dirty_max %ld\n",
 685                        cli_name(cli), atomic_long_read(&obd_dirty_pages),
 686                        atomic_long_read(&obd_dirty_transit_pages),
 687                        obd_max_dirty_pages);
 688                 oa->o_undirty = 0;
 689         } else if (unlikely(cli->cl_dirty_max_pages - cli->cl_dirty_pages >
 690                             0x7fffffff)) {
 691                 CERROR("dirty %lu - dirty_max %lu too big???\n",
 692                        cli->cl_dirty_pages, cli->cl_dirty_max_pages);
 693                 oa->o_undirty = 0;
 694         } else {
 695                 unsigned long nrpages;
 696
 697                 nrpages = cli->cl_max_pages_per_rpc;
 698                 nrpages *= cli->cl_max_rpcs_in_flight + 1;
 699                 nrpages = max(nrpages, cli->cl_dirty_max_pages);
 700                 oa->o_undirty = nrpages << PAGE_SHIFT;
 701                 if (OCD_HAS_FLAG(&cli->cl_import->imp_connect_data,
 702                                  GRANT_PARAM)) {
 703                         int nrextents;
 704
 705                         /* take extent tax into account when asking for more
 706                          * grant space */
 707                         nrextents = (nrpages + cli->cl_max_extent_pages - 1)  /
 708                                      cli->cl_max_extent_pages;
 709                         oa->o_undirty += nrextents * cli->cl_grant_extent_tax;
 710                 }
 711         }
 712         oa->o_grant = cli->cl_avail_grant + cli->cl_reserved_grant;
 713         oa->o_dropped = cli->cl_lost_grant;
 714         cli->cl_lost_grant = 0;
 715         spin_unlock(&cli->cl_loi_list_lock);
 716         CDEBUG(D_CACHE, "dirty: %llu undirty: %u dropped %u grant: %llu\n",
 717                oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
 718 }
 719
 720 void osc_update_next_shrink(struct client_obd *cli)
 721 {
 722         cli->cl_next_shrink_grant =
 723                 cfs_time_shift(cli->cl_grant_shrink_interval);
 724         CDEBUG(D_CACHE, "next time %ld to shrink grant \n",
 725                cli->cl_next_shrink_grant);
 726 }
 727
 728 static void __osc_update_grant(struct client_obd *cli, u64 grant)
 729 {
 730         spin_lock(&cli->cl_loi_list_lock);
 731         cli->cl_avail_grant += grant;
 732         spin_unlock(&cli->cl_loi_list_lock);
 733 }
 734
 735 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
 736 {
 737         if (body->oa.o_valid & OBD_MD_FLGRANT) {
 738                 CDEBUG(D_CACHE, "got %llu extra grant\n", body->oa.o_grant);
 739                 __osc_update_grant(cli, body->oa.o_grant);
 740         }
 741 }
 742
 743 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
 744                               u32 keylen, void *key,
 745                               u32 vallen, void *val,
 746                               struct ptlrpc_request_set *set);
 747
 748 static int osc_shrink_grant_interpret(const struct lu_env *env,
 749                                       struct ptlrpc_request *req,
 750                                       void *aa, int rc)
 751 {
 752         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
 753         struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
 754         struct ost_body *body;
 755
 756         if (rc != 0) {
 757                 __osc_update_grant(cli, oa->o_grant);
 758                 GOTO(out, rc);
 759         }
 760
 761         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 762         LASSERT(body);
 763         osc_update_grant(cli, body);
 764 out:
 765         OBDO_FREE(oa);
 766         return rc;
 767 }
 768
 769 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
 770 {
 771         spin_lock(&cli->cl_loi_list_lock);
 772         oa->o_grant = cli->cl_avail_grant / 4;
 773         cli->cl_avail_grant -= oa->o_grant;
 774         spin_unlock(&cli->cl_loi_list_lock);
 775         if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
 776                 oa->o_valid |= OBD_MD_FLFLAGS;
 777                 oa->o_flags = 0;
 778         }
 779         oa->o_flags |= OBD_FL_SHRINK_GRANT;
 780         osc_update_next_shrink(cli);
 781 }
 782
 783 /* Shrink the current grant, either from some large amount to enough for a
 784  * full set of in-flight RPCs, or if we have already shrunk to that limit
 785  * then to enough for a single RPC.  This avoids keeping more grant than
 786  * needed, and avoids shrinking the grant piecemeal. */
 787 static int osc_shrink_grant(struct client_obd *cli)
 788 {
 789         __u64 target_bytes = (cli->cl_max_rpcs_in_flight + 1) *
 790                              (cli->cl_max_pages_per_rpc << PAGE_SHIFT);
 791
 792         spin_lock(&cli->cl_loi_list_lock);
 793         if (cli->cl_avail_grant <= target_bytes)
 794                 target_bytes = cli->cl_max_pages_per_rpc << PAGE_SHIFT;
 795         spin_unlock(&cli->cl_loi_list_lock);
 796
 797         return osc_shrink_grant_to_target(cli, target_bytes);
 798 }
 799
 800 int osc_shrink_grant_to_target(struct client_obd *cli, __u64 target_bytes)
 801 {
 802         int                     rc = 0;
 803         struct ost_body        *body;
 804         ENTRY;
 805
 806         spin_lock(&cli->cl_loi_list_lock);
 807         /* Don't shrink if we are already above or below the desired limit
 808          * We don't want to shrink below a single RPC, as that will negatively
 809          * impact block allocation and long-term performance. */
 810         if (target_bytes < cli->cl_max_pages_per_rpc << PAGE_SHIFT)
 811                 target_bytes = cli->cl_max_pages_per_rpc << PAGE_SHIFT;
 812
 813         if (target_bytes >= cli->cl_avail_grant) {
 814                 spin_unlock(&cli->cl_loi_list_lock);
 815                 RETURN(0);
 816         }
 817         spin_unlock(&cli->cl_loi_list_lock);
 818
 819         OBD_ALLOC_PTR(body);
 820         if (!body)
 821                 RETURN(-ENOMEM);
 822
 823         osc_announce_cached(cli, &body->oa, 0);
 824
 825         spin_lock(&cli->cl_loi_list_lock);
 826         body->oa.o_grant = cli->cl_avail_grant - target_bytes;
 827         cli->cl_avail_grant = target_bytes;
 828         spin_unlock(&cli->cl_loi_list_lock);
 829         if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
 830                 body->oa.o_valid |= OBD_MD_FLFLAGS;
 831                 body->oa.o_flags = 0;
 832         }
 833         body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
 834         osc_update_next_shrink(cli);
 835
 836         rc = osc_set_info_async(NULL, cli->cl_import->imp_obd->obd_self_export,
 837                                 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
 838                                 sizeof(*body), body, NULL);
 839         if (rc != 0)
 840                 __osc_update_grant(cli, body->oa.o_grant);
 841         OBD_FREE_PTR(body);
 842         RETURN(rc);
 843 }
 844
 845 static int osc_should_shrink_grant(struct client_obd *client)
 846 {
 847         cfs_time_t time = cfs_time_current();
 848         cfs_time_t next_shrink = client->cl_next_shrink_grant;
 849
 850         if ((client->cl_import->imp_connect_data.ocd_connect_flags &
 851              OBD_CONNECT_GRANT_SHRINK) == 0)
 852                 return 0;
 853
 854         if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
 855                 /* Get the current RPC size directly, instead of going via:
 856                  * cli_brw_size(obd->u.cli.cl_import->imp_obd->obd_self_export)
 857                  * Keep comment here so that it can be found by searching. */
 858                 int brw_size = client->cl_max_pages_per_rpc << PAGE_SHIFT;
 859
 860                 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
 861                     client->cl_avail_grant > brw_size)
 862                         return 1;
 863                 else
 864                         osc_update_next_shrink(client);
 865         }
 866         return 0;
 867 }
 868
 869 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
 870 {
 871         struct client_obd *client;
 872
 873         list_for_each_entry(client, &item->ti_obd_list, cl_grant_shrink_list) {
 874                 if (osc_should_shrink_grant(client))
 875                         osc_shrink_grant(client);
 876         }
 877         return 0;
 878 }
 879
 880 static int osc_add_shrink_grant(struct client_obd *client)
 881 {
 882         int rc;
 883
 884         rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
 885                                        TIMEOUT_GRANT,
 886                                        osc_grant_shrink_grant_cb, NULL,
 887                                        &client->cl_grant_shrink_list);
 888         if (rc) {
 889                 CERROR("add grant client %s error %d\n", cli_name(client), rc);
 890                 return rc;
 891         }
 892         CDEBUG(D_CACHE, "add grant client %s\n", cli_name(client));
 893         osc_update_next_shrink(client);
 894         return 0;
 895 }
 896
 897 static int osc_del_shrink_grant(struct client_obd *client)
 898 {
 899         return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
 900                                          TIMEOUT_GRANT);
 901 }
 902
 903 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
 904 {
 905         /*
 906          * ocd_grant is the total grant amount we're expect to hold: if we've
 907          * been evicted, it's the new avail_grant amount, cl_dirty_pages will
 908          * drop to 0 as inflight RPCs fail out; otherwise, it's avail_grant +
 909          * dirty.
 910          *
 911          * race is tolerable here: if we're evicted, but imp_state already
 912          * left EVICTED state, then cl_dirty_pages must be 0 already.
 913          */
 914         spin_lock(&cli->cl_loi_list_lock);
 915         cli->cl_avail_grant = ocd->ocd_grant;
 916         if (cli->cl_import->imp_state != LUSTRE_IMP_EVICTED) {
 917                 cli->cl_avail_grant -= cli->cl_reserved_grant;
 918                 if (OCD_HAS_FLAG(ocd, GRANT_PARAM))
 919                         cli->cl_avail_grant -= cli->cl_dirty_grant;
 920                 else
 921                         cli->cl_avail_grant -=
 922                                         cli->cl_dirty_pages << PAGE_SHIFT;
 923         }
 924
 925         if (OCD_HAS_FLAG(ocd, GRANT_PARAM)) {
 926                 u64 size;
 927                 int chunk_mask;
 928
 929                 /* overhead for each extent insertion */
 930                 cli->cl_grant_extent_tax = ocd->ocd_grant_tax_kb << 10;
 931                 /* determine the appropriate chunk size used by osc_extent. */
 932                 cli->cl_chunkbits = max_t(int, PAGE_SHIFT,
 933                                           ocd->ocd_grant_blkbits);
 934                 /* max_pages_per_rpc must be chunk aligned */
 935                 chunk_mask = ~((1 << (cli->cl_chunkbits - PAGE_SHIFT)) - 1);
 936                 cli->cl_max_pages_per_rpc = (cli->cl_max_pages_per_rpc +
 937                                              ~chunk_mask) & chunk_mask;
 938                 /* determine maximum extent size, in #pages */
 939                 size = (u64)ocd->ocd_grant_max_blks << ocd->ocd_grant_blkbits;
 940                 cli->cl_max_extent_pages = size >> PAGE_SHIFT;
 941                 if (cli->cl_max_extent_pages == 0)
 942                         cli->cl_max_extent_pages = 1;
 943         } else {
 944                 cli->cl_grant_extent_tax = 0;
 945                 cli->cl_chunkbits = PAGE_SHIFT;
 946                 cli->cl_max_extent_pages = DT_MAX_BRW_PAGES;
 947         }
 948         spin_unlock(&cli->cl_loi_list_lock);
 949
 950         CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld."
 951                 "chunk bits: %d cl_max_extent_pages: %d\n",
 952                 cli_name(cli),
 953                 cli->cl_avail_grant, cli->cl_lost_grant, cli->cl_chunkbits,
 954                 cli->cl_max_extent_pages);
 955
 956         if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
 957             list_empty(&cli->cl_grant_shrink_list))
 958                 osc_add_shrink_grant(cli);
 959 }
 960
 961 /* We assume that the reason this OSC got a short read is because it read
 962  * beyond the end of a stripe file; i.e. lustre is reading a sparse file
 963  * via the LOV, and it _knows_ it's reading inside the file, it's just that
 964  * this stripe never got written at or beyond this stripe offset yet. */
 965 static void handle_short_read(int nob_read, size_t page_count,
 966                               struct brw_page **pga)
 967 {
 968         char *ptr;
 969         int i = 0;
 970
 971         /* skip bytes read OK */
 972         while (nob_read > 0) {
 973                 LASSERT (page_count > 0);
 974
 975                 if (pga[i]->count > nob_read) {
 976                         /* EOF inside this page */
 977                         ptr = kmap(pga[i]->pg) +
 978                                 (pga[i]->off & ~PAGE_MASK);
 979                         memset(ptr + nob_read, 0, pga[i]->count - nob_read);
 980                         kunmap(pga[i]->pg);
 981                         page_count--;
 982                         i++;
 983                         break;
 984                 }
 985
 986                 nob_read -= pga[i]->count;
 987                 page_count--;
 988                 i++;
 989         }
 990
 991         /* zero remaining pages */
 992         while (page_count-- > 0) {
 993                 ptr = kmap(pga[i]->pg) + (pga[i]->off & ~PAGE_MASK);
 994                 memset(ptr, 0, pga[i]->count);
 995                 kunmap(pga[i]->pg);
 996                 i++;
 997         }
 998 }
 999
1000 static int check_write_rcs(struct ptlrpc_request *req,
1001                            int requested_nob, int niocount,
1002                            size_t page_count, struct brw_page **pga)
1003 {
1004         int     i;
1005         __u32   *remote_rcs;
1006
1007         remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1008                                                   sizeof(*remote_rcs) *
1009                                                   niocount);
1010         if (remote_rcs == NULL) {
1011                 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1012                 return(-EPROTO);
1013         }
1014
1015         /* return error if any niobuf was in error */
1016         for (i = 0; i < niocount; i++) {
1017                 if ((int)remote_rcs[i] < 0)
1018                         return(remote_rcs[i]);
1019
1020                 if (remote_rcs[i] != 0) {
1021                         CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1022                                 i, remote_rcs[i], req);
1023                         return(-EPROTO);
1024                 }
1025         }
1026
1027         if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1028                 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1029                        req->rq_bulk->bd_nob_transferred, requested_nob);
1030                 return(-EPROTO);
1031         }
1032
1033         return (0);
1034 }
1035
1036 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1037 {
1038         if (p1->flag != p2->flag) {
1039                 unsigned mask = ~(OBD_BRW_FROM_GRANT | OBD_BRW_NOCACHE |
1040                                   OBD_BRW_SYNC       | OBD_BRW_ASYNC   |
1041                                   OBD_BRW_NOQUOTA    | OBD_BRW_SOFT_SYNC);
1042
1043                 /* warn if we try to combine flags that we don't know to be
1044                  * safe to combine */
1045                 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
1046                         CWARN("Saw flags 0x%x and 0x%x in the same brw, please "
1047                               "report this at https://jira.hpdd.intel.com/\n",
1048                               p1->flag, p2->flag);
1049                 }
1050                 return 0;
1051         }
1052
1053         return (p1->off + p1->count == p2->off);
1054 }
1055
1056 static u32 osc_checksum_bulk(int nob, size_t pg_count,
1057                              struct brw_page **pga, int opc,
1058                              cksum_type_t cksum_type)
1059 {
1060         u32                             cksum;
1061         int                             i = 0;
1062         struct cfs_crypto_hash_desc     *hdesc;
1063         unsigned int                    bufsize;
1064         unsigned char                   cfs_alg = cksum_obd2cfs(cksum_type);
1065
1066         LASSERT(pg_count > 0);
1067
1068         hdesc = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1069         if (IS_ERR(hdesc)) {
1070                 CERROR("Unable to initialize checksum hash %s\n",
1071                        cfs_crypto_hash_name(cfs_alg));
1072                 return PTR_ERR(hdesc);
1073         }
1074
1075         while (nob > 0 && pg_count > 0) {
1076                 unsigned int count = pga[i]->count > nob ? nob : pga[i]->count;
1077
1078                 /* corrupt the data before we compute the checksum, to
1079                  * simulate an OST->client data error */
1080                 if (i == 0 && opc == OST_READ &&
1081                     OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE)) {
1082                         unsigned char *ptr = kmap(pga[i]->pg);
1083                         int off = pga[i]->off & ~PAGE_MASK;
1084
1085                         memcpy(ptr + off, "bad1", min_t(typeof(nob), 4, nob));
1086                         kunmap(pga[i]->pg);
1087                 }
1088                 cfs_crypto_hash_update_page(hdesc, pga[i]->pg,
1089                                             pga[i]->off & ~PAGE_MASK,
1090                                             count);
1091                 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d\n",
1092                                (int)(pga[i]->off & ~PAGE_MASK));
1093
1094                 nob -= pga[i]->count;
1095                 pg_count--;
1096                 i++;
1097         }
1098
1099         bufsize = sizeof(cksum);
1100         cfs_crypto_hash_final(hdesc, (unsigned char *)&cksum, &bufsize);
1101
1102         /* For sending we only compute the wrong checksum instead
1103          * of corrupting the data so it is still correct on a redo */
1104         if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1105                 cksum++;
1106
1107         return cksum;
1108 }
1109
1110 static int
1111 osc_brw_prep_request(int cmd, struct client_obd *cli, struct obdo *oa,
1112                      u32 page_count, struct brw_page **pga,
1113                      struct ptlrpc_request **reqp, int resend)
1114 {
1115         struct ptlrpc_request   *req;
1116         struct ptlrpc_bulk_desc *desc;
1117         struct ost_body         *body;
1118         struct obd_ioobj        *ioobj;
1119         struct niobuf_remote    *niobuf;
1120         int niocount, i, requested_nob, opc, rc;
1121         struct osc_brw_async_args *aa;
1122         struct req_capsule      *pill;
1123         struct brw_page *pg_prev;
1124
1125         ENTRY;
1126         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1127                 RETURN(-ENOMEM); /* Recoverable */
1128         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1129                 RETURN(-EINVAL); /* Fatal */
1130
1131         if ((cmd & OBD_BRW_WRITE) != 0) {
1132                 opc = OST_WRITE;
1133                 req = ptlrpc_request_alloc_pool(cli->cl_import,
1134                                                 osc_rq_pool,
1135                                                 &RQF_OST_BRW_WRITE);
1136         } else {
1137                 opc = OST_READ;
1138                 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1139         }
1140         if (req == NULL)
1141                 RETURN(-ENOMEM);
1142
1143         for (niocount = i = 1; i < page_count; i++) {
1144                 if (!can_merge_pages(pga[i - 1], pga[i]))
1145                         niocount++;
1146         }
1147
1148         pill = &req->rq_pill;
1149         req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1150                              sizeof(*ioobj));
1151         req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1152                              niocount * sizeof(*niobuf));
1153
1154         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1155         if (rc) {
1156                 ptlrpc_request_free(req);
1157                 RETURN(rc);
1158         }
1159         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1160         ptlrpc_at_set_req_timeout(req);
1161         /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1162          * retry logic */
1163         req->rq_no_retry_einprogress = 1;
1164
1165         desc = ptlrpc_prep_bulk_imp(req, page_count,
1166                 cli->cl_import->imp_connect_data.ocd_brw_size >> LNET_MTU_BITS,
1167                 (opc == OST_WRITE ? PTLRPC_BULK_GET_SOURCE :
1168                         PTLRPC_BULK_PUT_SINK) |
1169                         PTLRPC_BULK_BUF_KIOV,
1170                 OST_BULK_PORTAL,
1171                 &ptlrpc_bulk_kiov_pin_ops);
1172
1173         if (desc == NULL)
1174                 GOTO(out, rc = -ENOMEM);
1175         /* NB request now owns desc and will free it when it gets freed */
1176
1177         body = req_capsule_client_get(pill, &RMF_OST_BODY);
1178         ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1179         niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1180         LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1181
1182         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1183
1184         obdo_to_ioobj(oa, ioobj);
1185         ioobj->ioo_bufcnt = niocount;
1186         /* The high bits of ioo_max_brw tells server _maximum_ number of bulks
1187          * that might be send for this request.  The actual number is decided
1188          * when the RPC is finally sent in ptlrpc_register_bulk(). It sends
1189          * "max - 1" for old client compatibility sending "0", and also so the
1190          * the actual maximum is a power-of-two number, not one less. LU-1431 */
1191         ioobj_max_brw_set(ioobj, desc->bd_md_max_brw);
1192         LASSERT(page_count > 0);
1193         pg_prev = pga[0];
1194         for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1195                 struct brw_page *pg = pga[i];
1196                 int poff = pg->off & ~PAGE_MASK;
1197
1198                 LASSERT(pg->count > 0);
1199                 /* make sure there is no gap in the middle of page array */
1200                 LASSERTF(page_count == 1 ||
1201                          (ergo(i == 0, poff + pg->count == PAGE_SIZE) &&
1202                           ergo(i > 0 && i < page_count - 1,
1203                                poff == 0 && pg->count == PAGE_SIZE)   &&
1204                           ergo(i == page_count - 1, poff == 0)),
1205                          "i: %d/%d pg: %p off: %llu, count: %u\n",
1206                          i, page_count, pg, pg->off, pg->count);
1207                 LASSERTF(i == 0 || pg->off > pg_prev->off,
1208                          "i %d p_c %u pg %p [pri %lu ind %lu] off %llu"
1209                          " prev_pg %p [pri %lu ind %lu] off %llu\n",
1210                          i, page_count,
1211                          pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1212                          pg_prev->pg, page_private(pg_prev->pg),
1213                          pg_prev->pg->index, pg_prev->off);
1214                 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1215                         (pg->flag & OBD_BRW_SRVLOCK));
1216
1217                 desc->bd_frag_ops->add_kiov_frag(desc, pg->pg, poff, pg->count);
1218                 requested_nob += pg->count;
1219
1220                 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1221                         niobuf--;
1222                         niobuf->rnb_len += pg->count;
1223                 } else {
1224                         niobuf->rnb_offset = pg->off;
1225                         niobuf->rnb_len    = pg->count;
1226                         niobuf->rnb_flags  = pg->flag;
1227                 }
1228                 pg_prev = pg;
1229         }
1230
1231         LASSERTF((void *)(niobuf - niocount) ==
1232                 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1233                 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1234                 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1235
1236         osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1237         if (resend) {
1238                 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1239                         body->oa.o_valid |= OBD_MD_FLFLAGS;
1240                         body->oa.o_flags = 0;
1241                 }
1242                 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1243         }
1244
1245         if (osc_should_shrink_grant(cli))
1246                 osc_shrink_grant_local(cli, &body->oa);
1247
1248         /* size[REQ_REC_OFF] still sizeof (*body) */
1249         if (opc == OST_WRITE) {
1250                 if (cli->cl_checksum &&
1251                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1252                         /* store cl_cksum_type in a local variable since
1253                          * it can be changed via lprocfs */
1254                         cksum_type_t cksum_type = cli->cl_cksum_type;
1255
1256                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1257                                 oa->o_flags &= OBD_FL_LOCAL_MASK;
1258                                 body->oa.o_flags = 0;
1259                         }
1260                         body->oa.o_flags |= cksum_type_pack(cksum_type);
1261                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1262                         body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1263                                                              page_count, pga,
1264                                                              OST_WRITE,
1265                                                              cksum_type);
1266                         CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1267                                body->oa.o_cksum);
1268                         /* save this in 'oa', too, for later checking */
1269                         oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1270                         oa->o_flags |= cksum_type_pack(cksum_type);
1271                 } else {
1272                         /* clear out the checksum flag, in case this is a
1273                          * resend but cl_checksum is no longer set. b=11238 */
1274                         oa->o_valid &= ~OBD_MD_FLCKSUM;
1275                 }
1276                 oa->o_cksum = body->oa.o_cksum;
1277                 /* 1 RC per niobuf */
1278                 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1279                                      sizeof(__u32) * niocount);
1280         } else {
1281                 if (cli->cl_checksum &&
1282                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1283                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1284                                 body->oa.o_flags = 0;
1285                         body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1286                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1287                 }
1288         }
1289         ptlrpc_request_set_replen(req);
1290
1291         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1292         aa = ptlrpc_req_async_args(req);
1293         aa->aa_oa = oa;
1294         aa->aa_requested_nob = requested_nob;
1295         aa->aa_nio_count = niocount;
1296         aa->aa_page_count = page_count;
1297         aa->aa_resends = 0;
1298         aa->aa_ppga = pga;
1299         aa->aa_cli = cli;
1300         INIT_LIST_HEAD(&aa->aa_oaps);
1301
1302         *reqp = req;
1303         niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1304         CDEBUG(D_RPCTRACE, "brw rpc %p - object "DOSTID" offset %lld<>%lld\n",
1305                 req, POSTID(&oa->o_oi), niobuf[0].rnb_offset,
1306                 niobuf[niocount - 1].rnb_offset + niobuf[niocount - 1].rnb_len);
1307         RETURN(0);
1308
1309  out:
1310         ptlrpc_req_finished(req);
1311         RETURN(rc);
1312 }
1313
1314 static int
1315 check_write_checksum(struct obdo *oa, const struct lnet_process_id *peer,
1316                      __u32 client_cksum, __u32 server_cksum, int nob,
1317                      size_t page_count, struct brw_page **pga,
1318                      cksum_type_t client_cksum_type)
1319 {
1320         __u32 new_cksum;
1321         char *msg;
1322         cksum_type_t cksum_type;
1323
1324         if (server_cksum == client_cksum) {
1325                 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1326                 return 0;
1327         }
1328
1329         cksum_type = cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1330                                        oa->o_flags : 0);
1331         new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1332                                       cksum_type);
1333
1334         if (cksum_type != client_cksum_type)
1335                 msg = "the server did not use the checksum type specified in "
1336                       "the original request - likely a protocol problem";
1337         else if (new_cksum == server_cksum)
1338                 msg = "changed on the client after we checksummed it - "
1339                       "likely false positive due to mmap IO (bug 11742)";
1340         else if (new_cksum == client_cksum)
1341                 msg = "changed in transit before arrival at OST";
1342         else
1343                 msg = "changed in transit AND doesn't match the original - "
1344                       "likely false positive due to mmap IO (bug 11742)";
1345
1346         LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode "DFID
1347                            " object "DOSTID" extent [%llu-%llu]\n",
1348                            msg, libcfs_nid2str(peer->nid),
1349                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1350                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1351                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1352                            POSTID(&oa->o_oi), pga[0]->off,
1353                            pga[page_count-1]->off + pga[page_count-1]->count - 1);
1354         CERROR("original client csum %x (type %x), server csum %x (type %x), "
1355                "client csum now %x\n", client_cksum, client_cksum_type,
1356                server_cksum, cksum_type, new_cksum);
1357         return 1;
1358 }
1359
1360 /* Note rc enters this function as number of bytes transferred */
1361 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1362 {
1363         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1364         const struct lnet_process_id *peer =
1365                         &req->rq_import->imp_connection->c_peer;
1366         struct client_obd *cli = aa->aa_cli;
1367         struct ost_body *body;
1368         u32 client_cksum = 0;
1369         ENTRY;
1370
1371         if (rc < 0 && rc != -EDQUOT) {
1372                 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1373                 RETURN(rc);
1374         }
1375
1376         LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1377         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1378         if (body == NULL) {
1379                 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1380                 RETURN(-EPROTO);
1381         }
1382
1383         /* set/clear over quota flag for a uid/gid */
1384         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1385             body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1386                 unsigned int qid[LL_MAXQUOTAS] =
1387                                         {body->oa.o_uid, body->oa.o_gid};
1388
1389                 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid %#llx, flags %x\n",
1390                        body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1391                        body->oa.o_flags);
1392                 osc_quota_setdq(cli, qid, body->oa.o_valid, body->oa.o_flags);
1393         }
1394
1395         osc_update_grant(cli, body);
1396
1397         if (rc < 0)
1398                 RETURN(rc);
1399
1400         if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1401                 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1402
1403         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1404                 if (rc > 0) {
1405                         CERROR("Unexpected +ve rc %d\n", rc);
1406                         RETURN(-EPROTO);
1407                 }
1408                 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1409
1410                 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1411                         RETURN(-EAGAIN);
1412
1413                 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1414                     check_write_checksum(&body->oa, peer, client_cksum,
1415                                          body->oa.o_cksum, aa->aa_requested_nob,
1416                                          aa->aa_page_count, aa->aa_ppga,
1417                                          cksum_type_unpack(aa->aa_oa->o_flags)))
1418                         RETURN(-EAGAIN);
1419
1420                 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1421                                      aa->aa_page_count, aa->aa_ppga);
1422                 GOTO(out, rc);
1423         }
1424
1425         /* The rest of this function executes only for OST_READs */
1426
1427         /* if unwrap_bulk failed, return -EAGAIN to retry */
1428         rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1429         if (rc < 0)
1430                 GOTO(out, rc = -EAGAIN);
1431
1432         if (rc > aa->aa_requested_nob) {
1433                 CERROR("Unexpected rc %d (%d requested)\n", rc,
1434                        aa->aa_requested_nob);
1435                 RETURN(-EPROTO);
1436         }
1437
1438         if (rc != req->rq_bulk->bd_nob_transferred) {
1439                 CERROR ("Unexpected rc %d (%d transferred)\n",
1440                         rc, req->rq_bulk->bd_nob_transferred);
1441                 return (-EPROTO);
1442         }
1443
1444         if (rc < aa->aa_requested_nob)
1445                 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1446
1447         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1448                 static int cksum_counter;
1449                 u32        server_cksum = body->oa.o_cksum;
1450                 char      *via = "";
1451                 char      *router = "";
1452                 cksum_type_t cksum_type;
1453
1454                 cksum_type = cksum_type_unpack(body->oa.o_valid &OBD_MD_FLFLAGS?
1455                                                body->oa.o_flags : 0);
1456                 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1457                                                  aa->aa_ppga, OST_READ,
1458                                                  cksum_type);
1459
1460                 if (peer->nid != req->rq_bulk->bd_sender) {
1461                         via = " via ";
1462                         router = libcfs_nid2str(req->rq_bulk->bd_sender);
1463                 }
1464
1465                 if (server_cksum != client_cksum) {
1466                         LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1467                                            "%s%s%s inode "DFID" object "DOSTID
1468                                            " extent [%llu-%llu]\n",
1469                                            req->rq_import->imp_obd->obd_name,
1470                                            libcfs_nid2str(peer->nid),
1471                                            via, router,
1472                                            body->oa.o_valid & OBD_MD_FLFID ?
1473                                                 body->oa.o_parent_seq : (__u64)0,
1474                                            body->oa.o_valid & OBD_MD_FLFID ?
1475                                                 body->oa.o_parent_oid : 0,
1476                                            body->oa.o_valid & OBD_MD_FLFID ?
1477                                                 body->oa.o_parent_ver : 0,
1478                                            POSTID(&body->oa.o_oi),
1479                                            aa->aa_ppga[0]->off,
1480                                            aa->aa_ppga[aa->aa_page_count-1]->off +
1481                                            aa->aa_ppga[aa->aa_page_count-1]->count -
1482                                                                         1);
1483                         CERROR("client %x, server %x, cksum_type %x\n",
1484                                client_cksum, server_cksum, cksum_type);
1485                         cksum_counter = 0;
1486                         aa->aa_oa->o_cksum = client_cksum;
1487                         rc = -EAGAIN;
1488                 } else {
1489                         cksum_counter++;
1490                         CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1491                         rc = 0;
1492                 }
1493         } else if (unlikely(client_cksum)) {
1494                 static int cksum_missed;
1495
1496                 cksum_missed++;
1497                 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1498                         CERROR("Checksum %u requested from %s but not sent\n",
1499                                cksum_missed, libcfs_nid2str(peer->nid));
1500         } else {
1501                 rc = 0;
1502         }
1503 out:
1504         if (rc >= 0)
1505                 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
1506                                      aa->aa_oa, &body->oa);
1507
1508         RETURN(rc);
1509 }
1510
1511 static int osc_brw_redo_request(struct ptlrpc_request *request,
1512                                 struct osc_brw_async_args *aa, int rc)
1513 {
1514         struct ptlrpc_request *new_req;
1515         struct osc_brw_async_args *new_aa;
1516         struct osc_async_page *oap;
1517         ENTRY;
1518
1519         DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
1520                   "redo for recoverable error %d", rc);
1521
1522         rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1523                                 OST_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
1524                                   aa->aa_cli, aa->aa_oa, aa->aa_page_count,
1525                                   aa->aa_ppga, &new_req, 1);
1526         if (rc)
1527                 RETURN(rc);
1528
1529         list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1530                 if (oap->oap_request != NULL) {
1531                         LASSERTF(request == oap->oap_request,
1532                                  "request %p != oap_request %p\n",
1533                                  request, oap->oap_request);
1534                         if (oap->oap_interrupted) {
1535                                 ptlrpc_req_finished(new_req);
1536                                 RETURN(-EINTR);
1537                         }
1538                 }
1539         }
1540         /* New request takes over pga and oaps from old request.
1541          * Note that copying a list_head doesn't work, need to move it... */
1542         aa->aa_resends++;
1543         new_req->rq_interpret_reply = request->rq_interpret_reply;
1544         new_req->rq_async_args = request->rq_async_args;
1545         new_req->rq_commit_cb = request->rq_commit_cb;
1546         /* cap resend delay to the current request timeout, this is similar to
1547          * what ptlrpc does (see after_reply()) */
1548         if (aa->aa_resends > new_req->rq_timeout)
1549                 new_req->rq_sent = cfs_time_current_sec() + new_req->rq_timeout;
1550         else
1551                 new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1552         new_req->rq_generation_set = 1;
1553         new_req->rq_import_generation = request->rq_import_generation;
1554
1555         new_aa = ptlrpc_req_async_args(new_req);
1556
1557         INIT_LIST_HEAD(&new_aa->aa_oaps);
1558         list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
1559         INIT_LIST_HEAD(&new_aa->aa_exts);
1560         list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
1561         new_aa->aa_resends = aa->aa_resends;
1562
1563         list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1564                 if (oap->oap_request) {
1565                         ptlrpc_req_finished(oap->oap_request);
1566                         oap->oap_request = ptlrpc_request_addref(new_req);
1567                 }
1568         }
1569
1570         /* XXX: This code will run into problem if we're going to support
1571          * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
1572          * and wait for all of them to be finished. We should inherit request
1573          * set from old request. */
1574         ptlrpcd_add_req(new_req);
1575
1576         DEBUG_REQ(D_INFO, new_req, "new request");
1577         RETURN(0);
1578 }
1579
1580 /*
1581  * ugh, we want disk allocation on the target to happen in offset order.  we'll
1582  * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1583  * fine for our small page arrays and doesn't require allocation.  its an
1584  * insertion sort that swaps elements that are strides apart, shrinking the
1585  * stride down until its '1' and the array is sorted.
1586  */
1587 static void sort_brw_pages(struct brw_page **array, int num)
1588 {
1589         int stride, i, j;
1590         struct brw_page *tmp;
1591
1592         if (num == 1)
1593                 return;
1594         for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1595                 ;
1596
1597         do {
1598                 stride /= 3;
1599                 for (i = stride ; i < num ; i++) {
1600                         tmp = array[i];
1601                         j = i;
1602                         while (j >= stride && array[j - stride]->off > tmp->off) {
1603                                 array[j] = array[j - stride];
1604                                 j -= stride;
1605                         }
1606                         array[j] = tmp;
1607                 }
1608         } while (stride > 1);
1609 }
1610
1611 static void osc_release_ppga(struct brw_page **ppga, size_t count)
1612 {
1613         LASSERT(ppga != NULL);
1614         OBD_FREE(ppga, sizeof(*ppga) * count);
1615 }
1616
1617 static int brw_interpret(const struct lu_env *env,
1618                          struct ptlrpc_request *req, void *data, int rc)
1619 {
1620         struct osc_brw_async_args *aa = data;
1621         struct osc_extent *ext;
1622         struct osc_extent *tmp;
1623         struct client_obd *cli = aa->aa_cli;
1624         ENTRY;
1625
1626         rc = osc_brw_fini_request(req, rc);
1627         CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
1628         /* When server return -EINPROGRESS, client should always retry
1629          * regardless of the number of times the bulk was resent already. */
1630         if (osc_recoverable_error(rc)) {
1631                 if (req->rq_import_generation !=
1632                     req->rq_import->imp_generation) {
1633                         CDEBUG(D_HA, "%s: resend cross eviction for object: "
1634                                ""DOSTID", rc = %d.\n",
1635                                req->rq_import->imp_obd->obd_name,
1636                                POSTID(&aa->aa_oa->o_oi), rc);
1637                 } else if (rc == -EINPROGRESS ||
1638                     client_should_resend(aa->aa_resends, aa->aa_cli)) {
1639                         rc = osc_brw_redo_request(req, aa, rc);
1640                 } else {
1641                         CERROR("%s: too many resent retries for object: "
1642                                "%llu:%llu, rc = %d.\n",
1643                                req->rq_import->imp_obd->obd_name,
1644                                POSTID(&aa->aa_oa->o_oi), rc);
1645                 }
1646
1647                 if (rc == 0)
1648                         RETURN(0);
1649                 else if (rc == -EAGAIN || rc == -EINPROGRESS)
1650                         rc = -EIO;
1651         }
1652
1653         if (rc == 0) {
1654                 struct obdo *oa = aa->aa_oa;
1655                 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
1656                 unsigned long valid = 0;
1657                 struct cl_object *obj;
1658                 struct osc_async_page *last;
1659
1660                 last = brw_page2oap(aa->aa_ppga[aa->aa_page_count - 1]);
1661                 obj = osc2cl(last->oap_obj);
1662
1663                 cl_object_attr_lock(obj);
1664                 if (oa->o_valid & OBD_MD_FLBLOCKS) {
1665                         attr->cat_blocks = oa->o_blocks;
1666                         valid |= CAT_BLOCKS;
1667                 }
1668                 if (oa->o_valid & OBD_MD_FLMTIME) {
1669                         attr->cat_mtime = oa->o_mtime;
1670                         valid |= CAT_MTIME;
1671                 }
1672                 if (oa->o_valid & OBD_MD_FLATIME) {
1673                         attr->cat_atime = oa->o_atime;
1674                         valid |= CAT_ATIME;
1675                 }
1676                 if (oa->o_valid & OBD_MD_FLCTIME) {
1677                         attr->cat_ctime = oa->o_ctime;
1678                         valid |= CAT_CTIME;
1679                 }
1680
1681                 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1682                         struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
1683                         loff_t last_off = last->oap_count + last->oap_obj_off +
1684                                 last->oap_page_off;
1685
1686                         /* Change file size if this is an out of quota or
1687                          * direct IO write and it extends the file size */
1688                         if (loi->loi_lvb.lvb_size < last_off) {
1689                                 attr->cat_size = last_off;
1690                                 valid |= CAT_SIZE;
1691                         }
1692                         /* Extend KMS if it's not a lockless write */
1693                         if (loi->loi_kms < last_off &&
1694                             oap2osc_page(last)->ops_srvlock == 0) {
1695                                 attr->cat_kms = last_off;
1696                                 valid |= CAT_KMS;
1697                         }
1698                 }
1699
1700                 if (valid != 0)
1701                         cl_object_attr_update(env, obj, attr, valid);
1702                 cl_object_attr_unlock(obj);
1703         }
1704         OBDO_FREE(aa->aa_oa);
1705
1706         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE && rc == 0)
1707                 osc_inc_unstable_pages(req);
1708
1709         list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
1710                 list_del_init(&ext->oe_link);
1711                 osc_extent_finish(env, ext, 1, rc);
1712         }
1713         LASSERT(list_empty(&aa->aa_exts));
1714         LASSERT(list_empty(&aa->aa_oaps));
1715
1716         osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
1717         ptlrpc_lprocfs_brw(req, req->rq_bulk->bd_nob_transferred);
1718
1719         spin_lock(&cli->cl_loi_list_lock);
1720         /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
1721          * is called so we know whether to go to sync BRWs or wait for more
1722          * RPCs to complete */
1723         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
1724                 cli->cl_w_in_flight--;
1725         else
1726                 cli->cl_r_in_flight--;
1727         osc_wake_cache_waiters(cli);
1728         spin_unlock(&cli->cl_loi_list_lock);
1729
1730         osc_io_unplug(env, cli, NULL);
1731         RETURN(rc);
1732 }
1733
1734 static void brw_commit(struct ptlrpc_request *req)
1735 {
1736         /* If osc_inc_unstable_pages (via osc_extent_finish) races with
1737          * this called via the rq_commit_cb, I need to ensure
1738          * osc_dec_unstable_pages is still called. Otherwise unstable
1739          * pages may be leaked. */
1740         spin_lock(&req->rq_lock);
1741         if (likely(req->rq_unstable)) {
1742                 req->rq_unstable = 0;
1743                 spin_unlock(&req->rq_lock);
1744
1745                 osc_dec_unstable_pages(req);
1746         } else {
1747                 req->rq_committed = 1;
1748                 spin_unlock(&req->rq_lock);
1749         }
1750 }
1751
1752 /**
1753  * Build an RPC by the list of extent @ext_list. The caller must ensure
1754  * that the total pages in this list are NOT over max pages per RPC.
1755  * Extents in the list must be in OES_RPC state.
1756  */
1757 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
1758                   struct list_head *ext_list, int cmd)
1759 {
1760         struct ptlrpc_request           *req = NULL;
1761         struct osc_extent               *ext;
1762         struct brw_page                 **pga = NULL;
1763         struct osc_brw_async_args       *aa = NULL;
1764         struct obdo                     *oa = NULL;
1765         struct osc_async_page           *oap;
1766         struct osc_object               *obj = NULL;
1767         struct cl_req_attr              *crattr = NULL;
1768         loff_t                          starting_offset = OBD_OBJECT_EOF;
1769         loff_t                          ending_offset = 0;
1770         int                             mpflag = 0;
1771         int                             mem_tight = 0;
1772         int                             page_count = 0;
1773         bool                            soft_sync = false;
1774         bool                            interrupted = false;
1775         int                             i;
1776         int                             grant = 0;
1777         int                             rc;
1778         struct list_head                rpc_list = LIST_HEAD_INIT(rpc_list);
1779         struct ost_body                 *body;
1780         ENTRY;
1781         LASSERT(!list_empty(ext_list));
1782
1783         /* add pages into rpc_list to build BRW rpc */
1784         list_for_each_entry(ext, ext_list, oe_link) {
1785                 LASSERT(ext->oe_state == OES_RPC);
1786                 mem_tight |= ext->oe_memalloc;
1787                 grant += ext->oe_grants;
1788                 page_count += ext->oe_nr_pages;
1789                 if (obj == NULL)
1790                         obj = ext->oe_obj;
1791         }
1792
1793         soft_sync = osc_over_unstable_soft_limit(cli);
1794         if (mem_tight)
1795                 mpflag = cfs_memory_pressure_get_and_set();
1796
1797         OBD_ALLOC(pga, sizeof(*pga) * page_count);
1798         if (pga == NULL)
1799                 GOTO(out, rc = -ENOMEM);
1800
1801         OBDO_ALLOC(oa);
1802         if (oa == NULL)
1803                 GOTO(out, rc = -ENOMEM);
1804
1805         i = 0;
1806         list_for_each_entry(ext, ext_list, oe_link) {
1807                 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1808                         if (mem_tight)
1809                                 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
1810                         if (soft_sync)
1811                                 oap->oap_brw_flags |= OBD_BRW_SOFT_SYNC;
1812                         pga[i] = &oap->oap_brw_page;
1813                         pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
1814                         i++;
1815
1816                         list_add_tail(&oap->oap_rpc_item, &rpc_list);
1817                         if (starting_offset == OBD_OBJECT_EOF ||
1818                             starting_offset > oap->oap_obj_off)
1819                                 starting_offset = oap->oap_obj_off;
1820                         else
1821                                 LASSERT(oap->oap_page_off == 0);
1822                         if (ending_offset < oap->oap_obj_off + oap->oap_count)
1823                                 ending_offset = oap->oap_obj_off +
1824                                                 oap->oap_count;
1825                         else
1826                                 LASSERT(oap->oap_page_off + oap->oap_count ==
1827                                         PAGE_SIZE);
1828                         if (oap->oap_interrupted)
1829                                 interrupted = true;
1830                 }
1831         }
1832
1833         /* first page in the list */
1834         oap = list_entry(rpc_list.next, typeof(*oap), oap_rpc_item);
1835
1836         crattr = &osc_env_info(env)->oti_req_attr;
1837         memset(crattr, 0, sizeof(*crattr));
1838         crattr->cra_type = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
1839         crattr->cra_flags = ~0ULL;
1840         crattr->cra_page = oap2cl_page(oap);
1841         crattr->cra_oa = oa;
1842         cl_req_attr_set(env, osc2cl(obj), crattr);
1843
1844         if (cmd == OBD_BRW_WRITE)
1845                 oa->o_grant_used = grant;
1846
1847         sort_brw_pages(pga, page_count);
1848         rc = osc_brw_prep_request(cmd, cli, oa, page_count, pga, &req, 0);
1849         if (rc != 0) {
1850                 CERROR("prep_req failed: %d\n", rc);
1851                 GOTO(out, rc);
1852         }
1853
1854         req->rq_commit_cb = brw_commit;
1855         req->rq_interpret_reply = brw_interpret;
1856         req->rq_memalloc = mem_tight != 0;
1857         oap->oap_request = ptlrpc_request_addref(req);
1858         if (interrupted && !req->rq_intr)
1859                 ptlrpc_mark_interrupted(req);
1860
1861         /* Need to update the timestamps after the request is built in case
1862          * we race with setattr (locally or in queue at OST).  If OST gets
1863          * later setattr before earlier BRW (as determined by the request xid),
1864          * the OST will not use BRW timestamps.  Sadly, there is no obvious
1865          * way to do this in a single call.  bug 10150 */
1866         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1867         crattr->cra_oa = &body->oa;
1868         crattr->cra_flags = OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME;
1869         cl_req_attr_set(env, osc2cl(obj), crattr);
1870         lustre_msg_set_jobid(req->rq_reqmsg, crattr->cra_jobid);
1871
1872         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1873         aa = ptlrpc_req_async_args(req);
1874         INIT_LIST_HEAD(&aa->aa_oaps);
1875         list_splice_init(&rpc_list, &aa->aa_oaps);
1876         INIT_LIST_HEAD(&aa->aa_exts);
1877         list_splice_init(ext_list, &aa->aa_exts);
1878
1879         spin_lock(&cli->cl_loi_list_lock);
1880         starting_offset >>= PAGE_SHIFT;
1881         if (cmd == OBD_BRW_READ) {
1882                 cli->cl_r_in_flight++;
1883                 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
1884                 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
1885                 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
1886                                       starting_offset + 1);
1887         } else {
1888                 cli->cl_w_in_flight++;
1889                 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
1890                 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
1891                 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
1892                                       starting_offset + 1);
1893         }
1894         spin_unlock(&cli->cl_loi_list_lock);
1895
1896         DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %ur/%uw in flight",
1897                   page_count, aa, cli->cl_r_in_flight,
1898                   cli->cl_w_in_flight);
1899         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_IO, cfs_fail_val);
1900
1901         ptlrpcd_add_req(req);
1902         rc = 0;
1903         EXIT;
1904
1905 out:
1906         if (mem_tight != 0)
1907                 cfs_memory_pressure_restore(mpflag);
1908
1909         if (rc != 0) {
1910                 LASSERT(req == NULL);
1911
1912                 if (oa)
1913                         OBDO_FREE(oa);
1914                 if (pga)
1915                         OBD_FREE(pga, sizeof(*pga) * page_count);
1916                 /* this should happen rarely and is pretty bad, it makes the
1917                  * pending list not follow the dirty order */
1918                 while (!list_empty(ext_list)) {
1919                         ext = list_entry(ext_list->next, struct osc_extent,
1920                                          oe_link);
1921                         list_del_init(&ext->oe_link);
1922                         osc_extent_finish(env, ext, 0, rc);
1923                 }
1924         }
1925         RETURN(rc);
1926 }
1927
1928 static int osc_set_lock_data(struct ldlm_lock *lock, void *data)
1929 {
1930         int set = 0;
1931
1932         LASSERT(lock != NULL);
1933
1934         lock_res_and_lock(lock);
1935
1936         if (lock->l_ast_data == NULL)
1937                 lock->l_ast_data = data;
1938         if (lock->l_ast_data == data)
1939                 set = 1;
1940
1941         unlock_res_and_lock(lock);
1942
1943         return set;
1944 }
1945
1946 static int osc_enqueue_fini(struct ptlrpc_request *req,
1947                             osc_enqueue_upcall_f upcall, void *cookie,
1948                             struct lustre_handle *lockh, enum ldlm_mode mode,
1949                             __u64 *flags, int agl, int errcode)
1950 {
1951         bool intent = *flags & LDLM_FL_HAS_INTENT;
1952         int rc;
1953         ENTRY;
1954
1955         /* The request was created before ldlm_cli_enqueue call. */
1956         if (intent && errcode == ELDLM_LOCK_ABORTED) {
1957                 struct ldlm_reply *rep;
1958
1959                 rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1960                 LASSERT(rep != NULL);
1961
1962                 rep->lock_policy_res1 =
1963                         ptlrpc_status_ntoh(rep->lock_policy_res1);
1964                 if (rep->lock_policy_res1)
1965                         errcode = rep->lock_policy_res1;
1966                 if (!agl)
1967                         *flags |= LDLM_FL_LVB_READY;
1968         } else if (errcode == ELDLM_OK) {
1969                 *flags |= LDLM_FL_LVB_READY;
1970         }
1971
1972         /* Call the update callback. */
1973         rc = (*upcall)(cookie, lockh, errcode);
1974
1975         /* release the reference taken in ldlm_cli_enqueue() */
1976         if (errcode == ELDLM_LOCK_MATCHED)
1977                 errcode = ELDLM_OK;
1978         if (errcode == ELDLM_OK && lustre_handle_is_used(lockh))
1979                 ldlm_lock_decref(lockh, mode);
1980
1981         RETURN(rc);
1982 }
1983
1984 static int osc_enqueue_interpret(const struct lu_env *env,
1985                                  struct ptlrpc_request *req,
1986                                  struct osc_enqueue_args *aa, int rc)
1987 {
1988         struct ldlm_lock *lock;
1989         struct lustre_handle *lockh = &aa->oa_lockh;
1990         enum ldlm_mode mode = aa->oa_mode;
1991         struct ost_lvb *lvb = aa->oa_lvb;
1992         __u32 lvb_len = sizeof(*lvb);
1993         __u64 flags = 0;
1994
1995         ENTRY;
1996
1997         /* ldlm_cli_enqueue is holding a reference on the lock, so it must
1998          * be valid. */
1999         lock = ldlm_handle2lock(lockh);
2000         LASSERTF(lock != NULL,
2001                  "lockh %#llx, req %p, aa %p - client evicted?\n",
2002                  lockh->cookie, req, aa);
2003
2004         /* Take an additional reference so that a blocking AST that
2005          * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2006          * to arrive after an upcall has been executed by
2007          * osc_enqueue_fini(). */
2008         ldlm_lock_addref(lockh, mode);
2009
2010         /* Let cl_lock_state_wait fail with -ERESTARTSYS to unuse sublocks. */
2011         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_HANG, 2);
2012
2013         /* Let CP AST to grant the lock first. */
2014         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
2015
2016         if (aa->oa_agl) {
2017                 LASSERT(aa->oa_lvb == NULL);
2018                 LASSERT(aa->oa_flags == NULL);
2019                 aa->oa_flags = &flags;
2020         }
2021
2022         /* Complete obtaining the lock procedure. */
2023         rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_type, 1,
2024                                    aa->oa_mode, aa->oa_flags, lvb, lvb_len,
2025                                    lockh, rc);
2026         /* Complete osc stuff. */
2027         rc = osc_enqueue_fini(req, aa->oa_upcall, aa->oa_cookie, lockh, mode,
2028                               aa->oa_flags, aa->oa_agl, rc);
2029
2030         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
2031
2032         ldlm_lock_decref(lockh, mode);
2033         LDLM_LOCK_PUT(lock);
2034         RETURN(rc);
2035 }
2036
2037 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
2038
2039 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2040  * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2041  * other synchronous requests, however keeping some locks and trying to obtain
2042  * others may take a considerable amount of time in a case of ost failure; and
2043  * when other sync requests do not get released lock from a client, the client
2044  * is evicted from the cluster -- such scenarious make the life difficult, so
2045  * release locks just after they are obtained. */
2046 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2047                      __u64 *flags, union ldlm_policy_data *policy,
2048                      struct ost_lvb *lvb, int kms_valid,
2049                      osc_enqueue_upcall_f upcall, void *cookie,
2050                      struct ldlm_enqueue_info *einfo,
2051                      struct ptlrpc_request_set *rqset, int async, int agl)
2052 {
2053         struct obd_device *obd = exp->exp_obd;
2054         struct lustre_handle lockh = { 0 };
2055         struct ptlrpc_request *req = NULL;
2056         int intent = *flags & LDLM_FL_HAS_INTENT;
2057         __u64 match_flags = *flags;
2058         enum ldlm_mode mode;
2059         int rc;
2060         ENTRY;
2061
2062         /* Filesystem lock extents are extended to page boundaries so that
2063          * dealing with the page cache is a little smoother.  */
2064         policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2065         policy->l_extent.end |= ~PAGE_MASK;
2066
2067         /*
2068          * kms is not valid when either object is completely fresh (so that no
2069          * locks are cached), or object was evicted. In the latter case cached
2070          * lock cannot be used, because it would prime inode state with
2071          * potentially stale LVB.
2072          */
2073         if (!kms_valid)
2074                 goto no_match;
2075
2076         /* Next, search for already existing extent locks that will cover us */
2077         /* If we're trying to read, we also search for an existing PW lock.  The
2078          * VFS and page cache already protect us locally, so lots of readers/
2079          * writers can share a single PW lock.
2080          *
2081          * There are problems with conversion deadlocks, so instead of
2082          * converting a read lock to a write lock, we'll just enqueue a new
2083          * one.
2084          *
2085          * At some point we should cancel the read lock instead of making them
2086          * send us a blocking callback, but there are problems with canceling
2087          * locks out from other users right now, too. */
2088         mode = einfo->ei_mode;
2089         if (einfo->ei_mode == LCK_PR)
2090                 mode |= LCK_PW;
2091         if (agl == 0)
2092                 match_flags |= LDLM_FL_LVB_READY;
2093         if (intent != 0)
2094                 match_flags |= LDLM_FL_BLOCK_GRANTED;
2095         mode = ldlm_lock_match(obd->obd_namespace, match_flags, res_id,
2096                                einfo->ei_type, policy, mode, &lockh, 0);
2097         if (mode) {
2098                 struct ldlm_lock *matched;
2099
2100                 if (*flags & LDLM_FL_TEST_LOCK)
2101                         RETURN(ELDLM_OK);
2102
2103                 matched = ldlm_handle2lock(&lockh);
2104                 if (agl) {
2105                         /* AGL enqueues DLM locks speculatively. Therefore if
2106                          * it already exists a DLM lock, it wll just inform the
2107                          * caller to cancel the AGL process for this stripe. */
2108                         ldlm_lock_decref(&lockh, mode);
2109                         LDLM_LOCK_PUT(matched);
2110                         RETURN(-ECANCELED);
2111                 } else if (osc_set_lock_data(matched, einfo->ei_cbdata)) {
2112                         *flags |= LDLM_FL_LVB_READY;
2113
2114                         /* We already have a lock, and it's referenced. */
2115                         (*upcall)(cookie, &lockh, ELDLM_LOCK_MATCHED);
2116
2117                         ldlm_lock_decref(&lockh, mode);
2118                         LDLM_LOCK_PUT(matched);
2119                         RETURN(ELDLM_OK);
2120                 } else {
2121                         ldlm_lock_decref(&lockh, mode);
2122                         LDLM_LOCK_PUT(matched);
2123                 }
2124         }
2125
2126 no_match:
2127         if (*flags & (LDLM_FL_TEST_LOCK | LDLM_FL_MATCH_LOCK))
2128                 RETURN(-ENOLCK);
2129
2130         if (intent) {
2131                 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2132                                            &RQF_LDLM_ENQUEUE_LVB);
2133                 if (req == NULL)
2134                         RETURN(-ENOMEM);
2135
2136                 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
2137                 if (rc) {
2138                         ptlrpc_request_free(req);
2139                         RETURN(rc);
2140                 }
2141
2142                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2143                                      sizeof *lvb);
2144                 ptlrpc_request_set_replen(req);
2145         }
2146
2147         /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2148         *flags &= ~LDLM_FL_BLOCK_GRANTED;
2149
2150         rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2151                               sizeof(*lvb), LVB_T_OST, &lockh, async);
2152         if (async) {
2153                 if (!rc) {
2154                         struct osc_enqueue_args *aa;
2155                         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2156                         aa = ptlrpc_req_async_args(req);
2157                         aa->oa_exp    = exp;
2158                         aa->oa_mode   = einfo->ei_mode;
2159                         aa->oa_type   = einfo->ei_type;
2160                         lustre_handle_copy(&aa->oa_lockh, &lockh);
2161                         aa->oa_upcall = upcall;
2162                         aa->oa_cookie = cookie;
2163                         aa->oa_agl    = !!agl;
2164                         if (!agl) {
2165                                 aa->oa_flags  = flags;
2166                                 aa->oa_lvb    = lvb;
2167                         } else {
2168                                 /* AGL is essentially to enqueue an DLM lock
2169                                  * in advance, so we don't care about the
2170                                  * result of AGL enqueue. */
2171                                 aa->oa_lvb    = NULL;
2172                                 aa->oa_flags  = NULL;
2173                         }
2174
2175                         req->rq_interpret_reply =
2176                                 (ptlrpc_interpterer_t)osc_enqueue_interpret;
2177                         if (rqset == PTLRPCD_SET)
2178                                 ptlrpcd_add_req(req);
2179                         else
2180                                 ptlrpc_set_add_req(rqset, req);
2181                 } else if (intent) {
2182                         ptlrpc_req_finished(req);
2183                 }
2184                 RETURN(rc);
2185         }
2186
2187         rc = osc_enqueue_fini(req, upcall, cookie, &lockh, einfo->ei_mode,
2188                               flags, agl, rc);
2189         if (intent)
2190                 ptlrpc_req_finished(req);
2191
2192         RETURN(rc);
2193 }
2194
2195 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2196                    enum ldlm_type type, union ldlm_policy_data *policy,
2197                    enum ldlm_mode mode, __u64 *flags, void *data,
2198                    struct lustre_handle *lockh, int unref)
2199 {
2200         struct obd_device *obd = exp->exp_obd;
2201         __u64 lflags = *flags;
2202         enum ldlm_mode rc;
2203         ENTRY;
2204
2205         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2206                 RETURN(-EIO);
2207
2208         /* Filesystem lock extents are extended to page boundaries so that
2209          * dealing with the page cache is a little smoother */
2210         policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2211         policy->l_extent.end |= ~PAGE_MASK;
2212
2213         /* Next, search for already existing extent locks that will cover us */
2214         /* If we're trying to read, we also search for an existing PW lock.  The
2215          * VFS and page cache already protect us locally, so lots of readers/
2216          * writers can share a single PW lock. */
2217         rc = mode;
2218         if (mode == LCK_PR)
2219                 rc |= LCK_PW;
2220         rc = ldlm_lock_match(obd->obd_namespace, lflags,
2221                              res_id, type, policy, rc, lockh, unref);
2222         if (rc == 0 || lflags & LDLM_FL_TEST_LOCK)
2223                 RETURN(rc);
2224
2225         if (data != NULL) {
2226                 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2227
2228                 LASSERT(lock != NULL);
2229                 if (!osc_set_lock_data(lock, data)) {
2230                         ldlm_lock_decref(lockh, rc);
2231                         rc = 0;
2232                 }
2233                 LDLM_LOCK_PUT(lock);
2234         }
2235         RETURN(rc);
2236 }
2237
2238 static int osc_statfs_interpret(const struct lu_env *env,
2239                                 struct ptlrpc_request *req,
2240                                 struct osc_async_args *aa, int rc)
2241 {
2242         struct obd_statfs *msfs;
2243         ENTRY;
2244
2245         if (rc == -EBADR)
2246                 /* The request has in fact never been sent
2247                  * due to issues at a higher level (LOV).
2248                  * Exit immediately since the caller is
2249                  * aware of the problem and takes care
2250                  * of the clean up */
2251                  RETURN(rc);
2252
2253         if ((rc == -ENOTCONN || rc == -EAGAIN) &&
2254             (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
2255                 GOTO(out, rc = 0);
2256
2257         if (rc != 0)
2258                 GOTO(out, rc);
2259
2260         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2261         if (msfs == NULL) {
2262                 GOTO(out, rc = -EPROTO);
2263         }
2264
2265         *aa->aa_oi->oi_osfs = *msfs;
2266 out:
2267         rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
2268         RETURN(rc);
2269 }
2270
2271 static int osc_statfs_async(struct obd_export *exp,
2272                             struct obd_info *oinfo, __u64 max_age,
2273                             struct ptlrpc_request_set *rqset)
2274 {
2275         struct obd_device     *obd = class_exp2obd(exp);
2276         struct ptlrpc_request *req;
2277         struct osc_async_args *aa;
2278         int                    rc;
2279         ENTRY;
2280
2281         /* We could possibly pass max_age in the request (as an absolute
2282          * timestamp or a "seconds.usec ago") so the target can avoid doing
2283          * extra calls into the filesystem if that isn't necessary (e.g.
2284          * during mount that would help a bit).  Having relative timestamps
2285          * is not so great if request processing is slow, while absolute
2286          * timestamps are not ideal because they need time synchronization. */
2287         req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
2288         if (req == NULL)
2289                 RETURN(-ENOMEM);
2290
2291         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2292         if (rc) {
2293                 ptlrpc_request_free(req);
2294                 RETURN(rc);
2295         }
2296         ptlrpc_request_set_replen(req);
2297         req->rq_request_portal = OST_CREATE_PORTAL;
2298         ptlrpc_at_set_req_timeout(req);
2299
2300         if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
2301                 /* procfs requests not want stat in wait for avoid deadlock */
2302                 req->rq_no_resend = 1;
2303                 req->rq_no_delay = 1;
2304         }
2305
2306         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
2307         CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
2308         aa = ptlrpc_req_async_args(req);
2309         aa->aa_oi = oinfo;
2310
2311         ptlrpc_set_add_req(rqset, req);
2312         RETURN(0);
2313 }
2314
2315 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
2316                       struct obd_statfs *osfs, __u64 max_age, __u32 flags)
2317 {
2318         struct obd_device     *obd = class_exp2obd(exp);
2319         struct obd_statfs     *msfs;
2320         struct ptlrpc_request *req;
2321         struct obd_import     *imp = NULL;
2322         int rc;
2323         ENTRY;
2324
2325         /*Since the request might also come from lprocfs, so we need
2326          *sync this with client_disconnect_export Bug15684*/
2327         down_read(&obd->u.cli.cl_sem);
2328         if (obd->u.cli.cl_import)
2329                 imp = class_import_get(obd->u.cli.cl_import);
2330         up_read(&obd->u.cli.cl_sem);
2331         if (!imp)
2332                 RETURN(-ENODEV);
2333
2334         /* We could possibly pass max_age in the request (as an absolute
2335          * timestamp or a "seconds.usec ago") so the target can avoid doing
2336          * extra calls into the filesystem if that isn't necessary (e.g.
2337          * during mount that would help a bit).  Having relative timestamps
2338          * is not so great if request processing is slow, while absolute
2339          * timestamps are not ideal because they need time synchronization. */
2340         req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
2341
2342         class_import_put(imp);
2343
2344         if (req == NULL)
2345                 RETURN(-ENOMEM);
2346
2347         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2348         if (rc) {
2349                 ptlrpc_request_free(req);
2350                 RETURN(rc);
2351         }
2352         ptlrpc_request_set_replen(req);
2353         req->rq_request_portal = OST_CREATE_PORTAL;
2354         ptlrpc_at_set_req_timeout(req);
2355
2356         if (flags & OBD_STATFS_NODELAY) {
2357                 /* procfs requests not want stat in wait for avoid deadlock */
2358                 req->rq_no_resend = 1;
2359                 req->rq_no_delay = 1;
2360         }
2361
2362         rc = ptlrpc_queue_wait(req);
2363         if (rc)
2364                 GOTO(out, rc);
2365
2366         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2367         if (msfs == NULL) {
2368                 GOTO(out, rc = -EPROTO);
2369         }
2370
2371         *osfs = *msfs;
2372
2373         EXIT;
2374  out:
2375         ptlrpc_req_finished(req);
2376         return rc;
2377 }
2378
2379 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2380                          void *karg, void __user *uarg)
2381 {
2382         struct obd_device *obd = exp->exp_obd;
2383         struct obd_ioctl_data *data = karg;
2384         int err = 0;
2385         ENTRY;
2386
2387         if (!try_module_get(THIS_MODULE)) {
2388                 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2389                        module_name(THIS_MODULE));
2390                 return -EINVAL;
2391         }
2392         switch (cmd) {
2393         case OBD_IOC_CLIENT_RECOVER:
2394                 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2395                                             data->ioc_inlbuf1, 0);
2396                 if (err > 0)
2397                         err = 0;
2398                 GOTO(out, err);
2399         case IOC_OSC_SET_ACTIVE:
2400                 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2401                                                data->ioc_offset);
2402                 GOTO(out, err);
2403         case OBD_IOC_PING_TARGET:
2404                 err = ptlrpc_obd_ping(obd);
2405                 GOTO(out, err);
2406         default:
2407                 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
2408                        cmd, current_comm());
2409                 GOTO(out, err = -ENOTTY);
2410         }
2411 out:
2412         module_put(THIS_MODULE);
2413         return err;
2414 }
2415
2416 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
2417                               u32 keylen, void *key,
2418                               u32 vallen, void *val,
2419                               struct ptlrpc_request_set *set)
2420 {
2421         struct ptlrpc_request *req;
2422         struct obd_device     *obd = exp->exp_obd;
2423         struct obd_import     *imp = class_exp2cliimp(exp);
2424         char                  *tmp;
2425         int                    rc;
2426         ENTRY;
2427
2428         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
2429
2430         if (KEY_IS(KEY_CHECKSUM)) {
2431                 if (vallen != sizeof(int))
2432                         RETURN(-EINVAL);
2433                 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
2434                 RETURN(0);
2435         }
2436
2437         if (KEY_IS(KEY_SPTLRPC_CONF)) {
2438                 sptlrpc_conf_client_adapt(obd);
2439                 RETURN(0);
2440         }
2441
2442         if (KEY_IS(KEY_FLUSH_CTX)) {
2443                 sptlrpc_import_flush_my_ctx(imp);
2444                 RETURN(0);
2445         }
2446
2447         if (KEY_IS(KEY_CACHE_SET)) {
2448                 struct client_obd *cli = &obd->u.cli;
2449
2450                 LASSERT(cli->cl_cache == NULL); /* only once */
2451                 cli->cl_cache = (struct cl_client_cache *)val;
2452                 cl_cache_incref(cli->cl_cache);
2453                 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
2454
2455                 /* add this osc into entity list */
2456                 LASSERT(list_empty(&cli->cl_lru_osc));
2457                 spin_lock(&cli->cl_cache->ccc_lru_lock);
2458                 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
2459                 spin_unlock(&cli->cl_cache->ccc_lru_lock);
2460
2461                 RETURN(0);
2462         }
2463
2464         if (KEY_IS(KEY_CACHE_LRU_SHRINK)) {
2465                 struct client_obd *cli = &obd->u.cli;
2466                 long nr = atomic_long_read(&cli->cl_lru_in_list) >> 1;
2467                 long target = *(long *)val;
2468
2469                 nr = osc_lru_shrink(env, cli, min(nr, target), true);
2470                 *(long *)val -= nr;
2471                 RETURN(0);
2472         }
2473
2474         if (!set && !KEY_IS(KEY_GRANT_SHRINK))
2475                 RETURN(-EINVAL);
2476
2477         /* We pass all other commands directly to OST. Since nobody calls osc
2478            methods directly and everybody is supposed to go through LOV, we
2479            assume lov checked invalid values for us.
2480            The only recognised values so far are evict_by_nid and mds_conn.
2481            Even if something bad goes through, we'd get a -EINVAL from OST
2482            anyway. */
2483
2484         req = ptlrpc_request_alloc(imp, KEY_IS(KEY_GRANT_SHRINK) ?
2485                                                 &RQF_OST_SET_GRANT_INFO :
2486                                                 &RQF_OBD_SET_INFO);
2487         if (req == NULL)
2488                 RETURN(-ENOMEM);
2489
2490         req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2491                              RCL_CLIENT, keylen);
2492         if (!KEY_IS(KEY_GRANT_SHRINK))
2493                 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
2494                                      RCL_CLIENT, vallen);
2495         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
2496         if (rc) {
2497                 ptlrpc_request_free(req);
2498                 RETURN(rc);
2499         }
2500
2501         tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2502         memcpy(tmp, key, keylen);
2503         tmp = req_capsule_client_get(&req->rq_pill, KEY_IS(KEY_GRANT_SHRINK) ?
2504                                                         &RMF_OST_BODY :
2505                                                         &RMF_SETINFO_VAL);
2506         memcpy(tmp, val, vallen);
2507
2508         if (KEY_IS(KEY_GRANT_SHRINK)) {
2509                 struct osc_grant_args *aa;
2510                 struct obdo *oa;
2511
2512                 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2513                 aa = ptlrpc_req_async_args(req);
2514                 OBDO_ALLOC(oa);
2515                 if (!oa) {
2516                         ptlrpc_req_finished(req);
2517                         RETURN(-ENOMEM);
2518                 }
2519                 *oa = ((struct ost_body *)val)->oa;
2520                 aa->aa_oa = oa;
2521                 req->rq_interpret_reply = osc_shrink_grant_interpret;
2522         }
2523
2524         ptlrpc_request_set_replen(req);
2525         if (!KEY_IS(KEY_GRANT_SHRINK)) {
2526                 LASSERT(set != NULL);
2527                 ptlrpc_set_add_req(set, req);
2528                 ptlrpc_check_set(NULL, set);
2529         } else {
2530                 ptlrpcd_add_req(req);
2531         }
2532
2533         RETURN(0);
2534 }
2535
2536 static int osc_reconnect(const struct lu_env *env,
2537                          struct obd_export *exp, struct obd_device *obd,
2538                          struct obd_uuid *cluuid,
2539                          struct obd_connect_data *data,
2540                          void *localdata)
2541 {
2542         struct client_obd *cli = &obd->u.cli;
2543
2544         if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
2545                 long lost_grant;
2546                 long grant;
2547
2548                 spin_lock(&cli->cl_loi_list_lock);
2549                 grant = cli->cl_avail_grant + cli->cl_reserved_grant;
2550                 if (data->ocd_connect_flags & OBD_CONNECT_GRANT_PARAM)
2551                         grant += cli->cl_dirty_grant;
2552                 else
2553                         grant += cli->cl_dirty_pages << PAGE_SHIFT;
2554                 data->ocd_grant = grant ? : 2 * cli_brw_size(obd);
2555                 lost_grant = cli->cl_lost_grant;
2556                 cli->cl_lost_grant = 0;
2557                 spin_unlock(&cli->cl_loi_list_lock);
2558
2559                 CDEBUG(D_RPCTRACE, "ocd_connect_flags: %#llx ocd_version: %d"
2560                        " ocd_grant: %d, lost: %ld.\n", data->ocd_connect_flags,
2561                        data->ocd_version, data->ocd_grant, lost_grant);
2562         }
2563
2564         RETURN(0);
2565 }
2566
2567 static int osc_disconnect(struct obd_export *exp)
2568 {
2569         struct obd_device *obd = class_exp2obd(exp);
2570         int rc;
2571
2572         rc = client_disconnect_export(exp);
2573         /**
2574          * Initially we put del_shrink_grant before disconnect_export, but it
2575          * causes the following problem if setup (connect) and cleanup
2576          * (disconnect) are tangled together.
2577          *      connect p1                     disconnect p2
2578          *   ptlrpc_connect_import
2579          *     ...............               class_manual_cleanup
2580          *                                     osc_disconnect
2581          *                                     del_shrink_grant
2582          *   ptlrpc_connect_interrupt
2583          *     init_grant_shrink
2584          *   add this client to shrink list
2585          *                                      cleanup_osc
2586          * Bang! pinger trigger the shrink.
2587          * So the osc should be disconnected from the shrink list, after we
2588          * are sure the import has been destroyed. BUG18662
2589          */
2590         if (obd->u.cli.cl_import == NULL)
2591                 osc_del_shrink_grant(&obd->u.cli);
2592         return rc;
2593 }
2594
2595 static int osc_ldlm_resource_invalidate(struct cfs_hash *hs,
2596         struct cfs_hash_bd *bd, struct hlist_node *hnode, void *arg)
2597 {
2598         struct lu_env *env = arg;
2599         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2600         struct ldlm_lock *lock;
2601         struct osc_object *osc = NULL;
2602         ENTRY;
2603
2604         lock_res(res);
2605         list_for_each_entry(lock, &res->lr_granted, l_res_link) {
2606                 if (lock->l_ast_data != NULL && osc == NULL) {
2607                         osc = lock->l_ast_data;
2608                         cl_object_get(osc2cl(osc));
2609                 }
2610
2611                 /* clear LDLM_FL_CLEANED flag to make sure it will be canceled
2612                  * by the 2nd round of ldlm_namespace_clean() call in
2613                  * osc_import_event(). */
2614                 ldlm_clear_cleaned(lock);
2615         }
2616         unlock_res(res);
2617
2618         if (osc != NULL) {
2619                 osc_object_invalidate(env, osc);
2620                 cl_object_put(env, osc2cl(osc));
2621         }
2622
2623         RETURN(0);
2624 }
2625
2626 static int osc_import_event(struct obd_device *obd,
2627                             struct obd_import *imp,
2628                             enum obd_import_event event)
2629 {
2630         struct client_obd *cli;
2631         int rc = 0;
2632
2633         ENTRY;
2634         LASSERT(imp->imp_obd == obd);
2635
2636         switch (event) {
2637         case IMP_EVENT_DISCON: {
2638                 cli = &obd->u.cli;
2639                 spin_lock(&cli->cl_loi_list_lock);
2640                 cli->cl_avail_grant = 0;
2641                 cli->cl_lost_grant = 0;
2642                 spin_unlock(&cli->cl_loi_list_lock);
2643                 break;
2644         }
2645         case IMP_EVENT_INACTIVE: {
2646                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
2647                 break;
2648         }
2649         case IMP_EVENT_INVALIDATE: {
2650                 struct ldlm_namespace *ns = obd->obd_namespace;
2651                 struct lu_env         *env;
2652                 __u16                  refcheck;
2653
2654                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2655
2656                 env = cl_env_get(&refcheck);
2657                 if (!IS_ERR(env)) {
2658                         osc_io_unplug(env, &obd->u.cli, NULL);
2659
2660                         cfs_hash_for_each_nolock(ns->ns_rs_hash,
2661                                                  osc_ldlm_resource_invalidate,
2662                                                  env, 0);
2663                         cl_env_put(env, &refcheck);
2664
2665                         ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2666                 } else
2667                         rc = PTR_ERR(env);
2668                 break;
2669         }
2670         case IMP_EVENT_ACTIVE: {
2671                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
2672                 break;
2673         }
2674         case IMP_EVENT_OCD: {
2675                 struct obd_connect_data *ocd = &imp->imp_connect_data;
2676
2677                 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
2678                         osc_init_grant(&obd->u.cli, ocd);
2679
2680                 /* See bug 7198 */
2681                 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
2682                         imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
2683
2684                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
2685                 break;
2686         }
2687         case IMP_EVENT_DEACTIVATE: {
2688                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE, NULL);
2689                 break;
2690         }
2691         case IMP_EVENT_ACTIVATE: {
2692                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE, NULL);
2693                 break;
2694         }
2695         default:
2696                 CERROR("Unknown import event %d\n", event);
2697                 LBUG();
2698         }
2699         RETURN(rc);
2700 }
2701
2702 /**
2703  * Determine whether the lock can be canceled before replaying the lock
2704  * during recovery, see bug16774 for detailed information.
2705  *
2706  * \retval zero the lock can't be canceled
2707  * \retval other ok to cancel
2708  */
2709 static int osc_cancel_weight(struct ldlm_lock *lock)
2710 {
2711         /*
2712          * Cancel all unused and granted extent lock.
2713          */
2714         if (lock->l_resource->lr_type == LDLM_EXTENT &&
2715             lock->l_granted_mode == lock->l_req_mode &&
2716             osc_ldlm_weigh_ast(lock) == 0)
2717                 RETURN(1);
2718
2719         RETURN(0);
2720 }
2721
2722 static int brw_queue_work(const struct lu_env *env, void *data)
2723 {
2724         struct client_obd *cli = data;
2725
2726         CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
2727
2728         osc_io_unplug(env, cli, NULL);
2729         RETURN(0);
2730 }
2731
2732 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
2733 {
2734         struct client_obd *cli = &obd->u.cli;
2735         struct obd_type   *type;
2736         void              *handler;
2737         int                rc;
2738         int                adding;
2739         int                added;
2740         int                req_count;
2741         ENTRY;
2742
2743         rc = ptlrpcd_addref();
2744         if (rc)
2745                 RETURN(rc);
2746
2747         rc = client_obd_setup(obd, lcfg);
2748         if (rc)
2749                 GOTO(out_ptlrpcd, rc);
2750
2751         handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
2752         if (IS_ERR(handler))
2753                 GOTO(out_client_setup, rc = PTR_ERR(handler));
2754         cli->cl_writeback_work = handler;
2755
2756         handler = ptlrpcd_alloc_work(cli->cl_import, lru_queue_work, cli);
2757         if (IS_ERR(handler))
2758                 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
2759         cli->cl_lru_work = handler;
2760
2761         rc = osc_quota_setup(obd);
2762         if (rc)
2763                 GOTO(out_ptlrpcd_work, rc);
2764
2765         cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
2766
2767 #ifdef CONFIG_PROC_FS
2768         obd->obd_vars = lprocfs_osc_obd_vars;
2769 #endif
2770         /* If this is true then both client (osc) and server (osp) are on the
2771          * same node. The osp layer if loaded first will register the osc proc
2772          * directory. In that case this obd_device will be attached its proc
2773          * tree to type->typ_procsym instead of obd->obd_type->typ_procroot. */
2774         type = class_search_type(LUSTRE_OSP_NAME);
2775         if (type && type->typ_procsym) {
2776                 obd->obd_proc_entry = lprocfs_register(obd->obd_name,
2777                                                        type->typ_procsym,
2778                                                        obd->obd_vars, obd);
2779                 if (IS_ERR(obd->obd_proc_entry)) {
2780                         rc = PTR_ERR(obd->obd_proc_entry);
2781                         CERROR("error %d setting up lprocfs for %s\n", rc,
2782                                obd->obd_name);
2783                         obd->obd_proc_entry = NULL;
2784                 }
2785         } else {
2786                 rc = lprocfs_obd_setup(obd);
2787         }
2788
2789         /* If the basic OSC proc tree construction succeeded then
2790          * lets do the rest. */
2791         if (rc == 0) {
2792                 lproc_osc_attach_seqstat(obd);
2793                 sptlrpc_lprocfs_cliobd_attach(obd);
2794                 ptlrpc_lprocfs_register_obd(obd);
2795         }
2796
2797         /*
2798          * We try to control the total number of requests with a upper limit
2799          * osc_reqpool_maxreqcount. There might be some race which will cause
2800          * over-limit allocation, but it is fine.
2801          */
2802         req_count = atomic_read(&osc_pool_req_count);
2803         if (req_count < osc_reqpool_maxreqcount) {
2804                 adding = cli->cl_max_rpcs_in_flight + 2;
2805                 if (req_count + adding > osc_reqpool_maxreqcount)
2806                         adding = osc_reqpool_maxreqcount - req_count;
2807
2808                 added = ptlrpc_add_rqs_to_pool(osc_rq_pool, adding);
2809                 atomic_add(added, &osc_pool_req_count);
2810         }
2811
2812         INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
2813         ns_register_cancel(obd->obd_namespace, osc_cancel_weight);
2814
2815         spin_lock(&osc_shrink_lock);
2816         list_add_tail(&cli->cl_shrink_list, &osc_shrink_list);
2817         spin_unlock(&osc_shrink_lock);
2818
2819         RETURN(0);
2820
2821 out_ptlrpcd_work:
2822         if (cli->cl_writeback_work != NULL) {
2823                 ptlrpcd_destroy_work(cli->cl_writeback_work);
2824                 cli->cl_writeback_work = NULL;
2825         }
2826         if (cli->cl_lru_work != NULL) {
2827                 ptlrpcd_destroy_work(cli->cl_lru_work);
2828                 cli->cl_lru_work = NULL;
2829         }
2830 out_client_setup:
2831         client_obd_cleanup(obd);
2832 out_ptlrpcd:
2833         ptlrpcd_decref();
2834         RETURN(rc);
2835 }
2836
2837 static int osc_precleanup(struct obd_device *obd)
2838 {
2839         struct client_obd *cli = &obd->u.cli;
2840         ENTRY;
2841
2842         /* LU-464
2843          * for echo client, export may be on zombie list, wait for
2844          * zombie thread to cull it, because cli.cl_import will be
2845          * cleared in client_disconnect_export():
2846          *   class_export_destroy() -> obd_cleanup() ->
2847          *   echo_device_free() -> echo_client_cleanup() ->
2848          *   obd_disconnect() -> osc_disconnect() ->
2849          *   client_disconnect_export()
2850          */
2851         obd_zombie_barrier();
2852         if (cli->cl_writeback_work) {
2853                 ptlrpcd_destroy_work(cli->cl_writeback_work);
2854                 cli->cl_writeback_work = NULL;
2855         }
2856
2857         if (cli->cl_lru_work) {
2858                 ptlrpcd_destroy_work(cli->cl_lru_work);
2859                 cli->cl_lru_work = NULL;
2860         }
2861
2862         obd_cleanup_client_import(obd);
2863         ptlrpc_lprocfs_unregister_obd(obd);
2864         lprocfs_obd_cleanup(obd);
2865         RETURN(0);
2866 }
2867
2868 int osc_cleanup(struct obd_device *obd)
2869 {
2870         struct client_obd *cli = &obd->u.cli;
2871         int rc;
2872
2873         ENTRY;
2874
2875         spin_lock(&osc_shrink_lock);
2876         list_del(&cli->cl_shrink_list);
2877         spin_unlock(&osc_shrink_lock);
2878
2879         /* lru cleanup */
2880         if (cli->cl_cache != NULL) {
2881                 LASSERT(atomic_read(&cli->cl_cache->ccc_users) > 0);
2882                 spin_lock(&cli->cl_cache->ccc_lru_lock);
2883                 list_del_init(&cli->cl_lru_osc);
2884                 spin_unlock(&cli->cl_cache->ccc_lru_lock);
2885                 cli->cl_lru_left = NULL;
2886                 cl_cache_decref(cli->cl_cache);
2887                 cli->cl_cache = NULL;
2888         }
2889
2890         /* free memory of osc quota cache */
2891         osc_quota_cleanup(obd);
2892
2893         rc = client_obd_cleanup(obd);
2894
2895         ptlrpcd_decref();
2896         RETURN(rc);
2897 }
2898
2899 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
2900 {
2901         int rc = class_process_proc_param(PARAM_OSC, obd->obd_vars, lcfg, obd);
2902         return rc > 0 ? 0: rc;
2903 }
2904
2905 static int osc_process_config(struct obd_device *obd, size_t len, void *buf)
2906 {
2907         return osc_process_config_base(obd, buf);
2908 }
2909
2910 static struct obd_ops osc_obd_ops = {
2911         .o_owner                = THIS_MODULE,
2912         .o_setup                = osc_setup,
2913         .o_precleanup           = osc_precleanup,
2914         .o_cleanup              = osc_cleanup,
2915         .o_add_conn             = client_import_add_conn,
2916         .o_del_conn             = client_import_del_conn,
2917         .o_connect              = client_connect_import,
2918         .o_reconnect            = osc_reconnect,
2919         .o_disconnect           = osc_disconnect,
2920         .o_statfs               = osc_statfs,
2921         .o_statfs_async         = osc_statfs_async,
2922         .o_create               = osc_create,
2923         .o_destroy              = osc_destroy,
2924         .o_getattr              = osc_getattr,
2925         .o_setattr              = osc_setattr,
2926         .o_iocontrol            = osc_iocontrol,
2927         .o_set_info_async       = osc_set_info_async,
2928         .o_import_event         = osc_import_event,
2929         .o_process_config       = osc_process_config,
2930         .o_quotactl             = osc_quotactl,
2931 };
2932
2933 static struct shrinker *osc_cache_shrinker;
2934 struct list_head osc_shrink_list = LIST_HEAD_INIT(osc_shrink_list);
2935 DEFINE_SPINLOCK(osc_shrink_lock);
2936
2937 #ifndef HAVE_SHRINKER_COUNT
2938 static int osc_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
2939 {
2940         struct shrink_control scv = {
2941                 .nr_to_scan = shrink_param(sc, nr_to_scan),
2942                 .gfp_mask   = shrink_param(sc, gfp_mask)
2943         };
2944 #if !defined(HAVE_SHRINKER_WANT_SHRINK_PTR) && !defined(HAVE_SHRINK_CONTROL)
2945         struct shrinker *shrinker = NULL;
2946 #endif
2947
2948         (void)osc_cache_shrink_scan(shrinker, &scv);
2949
2950         return osc_cache_shrink_count(shrinker, &scv);
2951 }
2952 #endif
2953
2954 static int __init osc_init(void)
2955 {
2956         bool enable_proc = true;
2957         struct obd_type *type;
2958         unsigned int reqpool_size;
2959         unsigned int reqsize;
2960         int rc;
2961         DEF_SHRINKER_VAR(osc_shvar, osc_cache_shrink,
2962                          osc_cache_shrink_count, osc_cache_shrink_scan);
2963         ENTRY;
2964
2965         /* print an address of _any_ initialized kernel symbol from this
2966          * module, to allow debugging with gdb that doesn't support data
2967          * symbols from modules.*/
2968         CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
2969
2970         rc = lu_kmem_init(osc_caches);
2971         if (rc)
2972                 RETURN(rc);
2973
2974         type = class_search_type(LUSTRE_OSP_NAME);
2975         if (type != NULL && type->typ_procsym != NULL)
2976                 enable_proc = false;
2977
2978         rc = class_register_type(&osc_obd_ops, NULL, enable_proc, NULL,
2979                                  LUSTRE_OSC_NAME, &osc_device_type);
2980         if (rc)
2981                 GOTO(out_kmem, rc);
2982
2983         osc_cache_shrinker = set_shrinker(DEFAULT_SEEKS, &osc_shvar);
2984
2985         /* This is obviously too much memory, only prevent overflow here */
2986         if (osc_reqpool_mem_max >= 1 << 12 || osc_reqpool_mem_max == 0)
2987                 GOTO(out_type, rc = -EINVAL);
2988
2989         reqpool_size = osc_reqpool_mem_max << 20;
2990
2991         reqsize = 1;
2992         while (reqsize < OST_IO_MAXREQSIZE)
2993                 reqsize = reqsize << 1;
2994
2995         /*
2996          * We don't enlarge the request count in OSC pool according to
2997          * cl_max_rpcs_in_flight. The allocation from the pool will only be
2998          * tried after normal allocation failed. So a small OSC pool won't
2999          * cause much performance degression in most of cases.
3000          */
3001         osc_reqpool_maxreqcount = reqpool_size / reqsize;
3002
3003         atomic_set(&osc_pool_req_count, 0);
3004         osc_rq_pool = ptlrpc_init_rq_pool(0, OST_IO_MAXREQSIZE,
3005                                           ptlrpc_add_rqs_to_pool);
3006
3007         if (osc_rq_pool != NULL)
3008                 GOTO(out, rc);
3009         rc = -ENOMEM;
3010 out_type:
3011         class_unregister_type(LUSTRE_OSC_NAME);
3012 out_kmem:
3013         lu_kmem_fini(osc_caches);
3014 out:
3015         RETURN(rc);
3016 }
3017
3018 static void __exit osc_exit(void)
3019 {
3020         remove_shrinker(osc_cache_shrinker);
3021         class_unregister_type(LUSTRE_OSC_NAME);
3022         lu_kmem_fini(osc_caches);
3023         ptlrpc_free_rq_pool(osc_rq_pool);
3024 }
3025
3026 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3027 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3028 MODULE_VERSION(LUSTRE_VERSION_STRING);
3029 MODULE_LICENSE("GPL");
3030
3031 module_init(osc_init);
3032 module_exit(osc_exit);