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                 rc = l_wait_event_exclusive(cli->cl_destroy_waitq,
 651                                             osc_can_send_destroy(cli), &lwi);
 652                 if (rc) {
 653                         ptlrpc_request_free(req);
 654                         RETURN(rc);
 655                 }
 656         }
 657
 658         /* Do not wait for response */
 659         ptlrpcd_add_req(req);
 660         RETURN(0);
 661 }
 662
 663 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
 664                                 long writing_bytes)
 665 {
 666         u64 bits = OBD_MD_FLBLOCKS | OBD_MD_FLGRANT;
 667
 668         LASSERT(!(oa->o_valid & bits));
 669
 670         oa->o_valid |= bits;
 671         spin_lock(&cli->cl_loi_list_lock);
 672         if (OCD_HAS_FLAG(&cli->cl_import->imp_connect_data, GRANT_PARAM))
 673                 oa->o_dirty = cli->cl_dirty_grant;
 674         else
 675                 oa->o_dirty = cli->cl_dirty_pages << PAGE_SHIFT;
 676         if (unlikely(cli->cl_dirty_pages - cli->cl_dirty_transit >
 677                      cli->cl_dirty_max_pages)) {
 678                 CERROR("dirty %lu - %lu > dirty_max %lu\n",
 679                        cli->cl_dirty_pages, cli->cl_dirty_transit,
 680                        cli->cl_dirty_max_pages);
 681                 oa->o_undirty = 0;
 682         } else if (unlikely(atomic_long_read(&obd_dirty_pages) -
 683                             atomic_long_read(&obd_dirty_transit_pages) >
 684                             (long)(obd_max_dirty_pages + 1))) {
 685                 /* The atomic_read() allowing the atomic_inc() are
 686                  * not covered by a lock thus they may safely race and trip
 687                  * this CERROR() unless we add in a small fudge factor (+1). */
 688                 CERROR("%s: dirty %ld - %ld > system dirty_max %ld\n",
 689                        cli_name(cli), atomic_long_read(&obd_dirty_pages),
 690                        atomic_long_read(&obd_dirty_transit_pages),
 691                        obd_max_dirty_pages);
 692                 oa->o_undirty = 0;
 693         } else if (unlikely(cli->cl_dirty_max_pages - cli->cl_dirty_pages >
 694                             0x7fffffff)) {
 695                 CERROR("dirty %lu - dirty_max %lu too big???\n",
 696                        cli->cl_dirty_pages, cli->cl_dirty_max_pages);
 697                 oa->o_undirty = 0;
 698         } else {
 699                 unsigned long nrpages;
 700
 701                 nrpages = cli->cl_max_pages_per_rpc;
 702                 nrpages *= cli->cl_max_rpcs_in_flight + 1;
 703                 nrpages = max(nrpages, cli->cl_dirty_max_pages);
 704                 oa->o_undirty = nrpages << PAGE_SHIFT;
 705                 if (OCD_HAS_FLAG(&cli->cl_import->imp_connect_data,
 706                                  GRANT_PARAM)) {
 707                         int nrextents;
 708
 709                         /* take extent tax into account when asking for more
 710                          * grant space */
 711                         nrextents = (nrpages + cli->cl_max_extent_pages - 1)  /
 712                                      cli->cl_max_extent_pages;
 713                         oa->o_undirty += nrextents * cli->cl_grant_extent_tax;
 714                 }
 715         }
 716         oa->o_grant = cli->cl_avail_grant + cli->cl_reserved_grant;
 717         oa->o_dropped = cli->cl_lost_grant;
 718         cli->cl_lost_grant = 0;
 719         spin_unlock(&cli->cl_loi_list_lock);
 720         CDEBUG(D_CACHE, "dirty: %llu undirty: %u dropped %u grant: %llu\n",
 721                oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
 722 }
 723
 724 void osc_update_next_shrink(struct client_obd *cli)
 725 {
 726         cli->cl_next_shrink_grant =
 727                 cfs_time_shift(cli->cl_grant_shrink_interval);
 728         CDEBUG(D_CACHE, "next time %ld to shrink grant \n",
 729                cli->cl_next_shrink_grant);
 730 }
 731
 732 static void __osc_update_grant(struct client_obd *cli, u64 grant)
 733 {
 734         spin_lock(&cli->cl_loi_list_lock);
 735         cli->cl_avail_grant += grant;
 736         spin_unlock(&cli->cl_loi_list_lock);
 737 }
 738
 739 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
 740 {
 741         if (body->oa.o_valid & OBD_MD_FLGRANT) {
 742                 CDEBUG(D_CACHE, "got %llu extra grant\n", body->oa.o_grant);
 743                 __osc_update_grant(cli, body->oa.o_grant);
 744         }
 745 }
 746
 747 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
 748                               u32 keylen, void *key,
 749                               u32 vallen, void *val,
 750                               struct ptlrpc_request_set *set);
 751
 752 static int osc_shrink_grant_interpret(const struct lu_env *env,
 753                                       struct ptlrpc_request *req,
 754                                       void *aa, int rc)
 755 {
 756         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
 757         struct obdo *oa = ((struct osc_grant_args *)aa)->aa_oa;
 758         struct ost_body *body;
 759
 760         if (rc != 0) {
 761                 __osc_update_grant(cli, oa->o_grant);
 762                 GOTO(out, rc);
 763         }
 764
 765         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 766         LASSERT(body);
 767         osc_update_grant(cli, body);
 768 out:
 769         OBDO_FREE(oa);
 770         return rc;
 771 }
 772
 773 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
 774 {
 775         spin_lock(&cli->cl_loi_list_lock);
 776         oa->o_grant = cli->cl_avail_grant / 4;
 777         cli->cl_avail_grant -= oa->o_grant;
 778         spin_unlock(&cli->cl_loi_list_lock);
 779         if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
 780                 oa->o_valid |= OBD_MD_FLFLAGS;
 781                 oa->o_flags = 0;
 782         }
 783         oa->o_flags |= OBD_FL_SHRINK_GRANT;
 784         osc_update_next_shrink(cli);
 785 }
 786
 787 /* Shrink the current grant, either from some large amount to enough for a
 788  * full set of in-flight RPCs, or if we have already shrunk to that limit
 789  * then to enough for a single RPC.  This avoids keeping more grant than
 790  * needed, and avoids shrinking the grant piecemeal. */
 791 static int osc_shrink_grant(struct client_obd *cli)
 792 {
 793         __u64 target_bytes = (cli->cl_max_rpcs_in_flight + 1) *
 794                              (cli->cl_max_pages_per_rpc << PAGE_SHIFT);
 795
 796         spin_lock(&cli->cl_loi_list_lock);
 797         if (cli->cl_avail_grant <= target_bytes)
 798                 target_bytes = cli->cl_max_pages_per_rpc << PAGE_SHIFT;
 799         spin_unlock(&cli->cl_loi_list_lock);
 800
 801         return osc_shrink_grant_to_target(cli, target_bytes);
 802 }
 803
 804 int osc_shrink_grant_to_target(struct client_obd *cli, __u64 target_bytes)
 805 {
 806         int                     rc = 0;
 807         struct ost_body        *body;
 808         ENTRY;
 809
 810         spin_lock(&cli->cl_loi_list_lock);
 811         /* Don't shrink if we are already above or below the desired limit
 812          * We don't want to shrink below a single RPC, as that will negatively
 813          * impact block allocation and long-term performance. */
 814         if (target_bytes < cli->cl_max_pages_per_rpc << PAGE_SHIFT)
 815                 target_bytes = cli->cl_max_pages_per_rpc << PAGE_SHIFT;
 816
 817         if (target_bytes >= cli->cl_avail_grant) {
 818                 spin_unlock(&cli->cl_loi_list_lock);
 819                 RETURN(0);
 820         }
 821         spin_unlock(&cli->cl_loi_list_lock);
 822
 823         OBD_ALLOC_PTR(body);
 824         if (!body)
 825                 RETURN(-ENOMEM);
 826
 827         osc_announce_cached(cli, &body->oa, 0);
 828
 829         spin_lock(&cli->cl_loi_list_lock);
 830         body->oa.o_grant = cli->cl_avail_grant - target_bytes;
 831         cli->cl_avail_grant = target_bytes;
 832         spin_unlock(&cli->cl_loi_list_lock);
 833         if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
 834                 body->oa.o_valid |= OBD_MD_FLFLAGS;
 835                 body->oa.o_flags = 0;
 836         }
 837         body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
 838         osc_update_next_shrink(cli);
 839
 840         rc = osc_set_info_async(NULL, cli->cl_import->imp_obd->obd_self_export,
 841                                 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
 842                                 sizeof(*body), body, NULL);
 843         if (rc != 0)
 844                 __osc_update_grant(cli, body->oa.o_grant);
 845         OBD_FREE_PTR(body);
 846         RETURN(rc);
 847 }
 848
 849 static int osc_should_shrink_grant(struct client_obd *client)
 850 {
 851         cfs_time_t time = cfs_time_current();
 852         cfs_time_t next_shrink = client->cl_next_shrink_grant;
 853
 854         if ((client->cl_import->imp_connect_data.ocd_connect_flags &
 855              OBD_CONNECT_GRANT_SHRINK) == 0)
 856                 return 0;
 857
 858         if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
 859                 /* Get the current RPC size directly, instead of going via:
 860                  * cli_brw_size(obd->u.cli.cl_import->imp_obd->obd_self_export)
 861                  * Keep comment here so that it can be found by searching. */
 862                 int brw_size = client->cl_max_pages_per_rpc << PAGE_SHIFT;
 863
 864                 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
 865                     client->cl_avail_grant > brw_size)
 866                         return 1;
 867                 else
 868                         osc_update_next_shrink(client);
 869         }
 870         return 0;
 871 }
 872
 873 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
 874 {
 875         struct client_obd *client;
 876
 877         list_for_each_entry(client, &item->ti_obd_list, cl_grant_shrink_list) {
 878                 if (osc_should_shrink_grant(client))
 879                         osc_shrink_grant(client);
 880         }
 881         return 0;
 882 }
 883
 884 static int osc_add_shrink_grant(struct client_obd *client)
 885 {
 886         int rc;
 887
 888         rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
 889                                        TIMEOUT_GRANT,
 890                                        osc_grant_shrink_grant_cb, NULL,
 891                                        &client->cl_grant_shrink_list);
 892         if (rc) {
 893                 CERROR("add grant client %s error %d\n", cli_name(client), rc);
 894                 return rc;
 895         }
 896         CDEBUG(D_CACHE, "add grant client %s\n", cli_name(client));
 897         osc_update_next_shrink(client);
 898         return 0;
 899 }
 900
 901 static int osc_del_shrink_grant(struct client_obd *client)
 902 {
 903         return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
 904                                          TIMEOUT_GRANT);
 905 }
 906
 907 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
 908 {
 909         /*
 910          * ocd_grant is the total grant amount we're expect to hold: if we've
 911          * been evicted, it's the new avail_grant amount, cl_dirty_pages will
 912          * drop to 0 as inflight RPCs fail out; otherwise, it's avail_grant +
 913          * dirty.
 914          *
 915          * race is tolerable here: if we're evicted, but imp_state already
 916          * left EVICTED state, then cl_dirty_pages must be 0 already.
 917          */
 918         spin_lock(&cli->cl_loi_list_lock);
 919         cli->cl_avail_grant = ocd->ocd_grant;
 920         if (cli->cl_import->imp_state != LUSTRE_IMP_EVICTED) {
 921                 cli->cl_avail_grant -= cli->cl_reserved_grant;
 922                 if (OCD_HAS_FLAG(ocd, GRANT_PARAM))
 923                         cli->cl_avail_grant -= cli->cl_dirty_grant;
 924                 else
 925                         cli->cl_avail_grant -=
 926                                         cli->cl_dirty_pages << PAGE_SHIFT;
 927         }
 928
 929         if (OCD_HAS_FLAG(ocd, GRANT_PARAM)) {
 930                 u64 size;
 931                 int chunk_mask;
 932
 933                 /* overhead for each extent insertion */
 934                 cli->cl_grant_extent_tax = ocd->ocd_grant_tax_kb << 10;
 935                 /* determine the appropriate chunk size used by osc_extent. */
 936                 cli->cl_chunkbits = max_t(int, PAGE_SHIFT,
 937                                           ocd->ocd_grant_blkbits);
 938                 /* max_pages_per_rpc must be chunk aligned */
 939                 chunk_mask = ~((1 << (cli->cl_chunkbits - PAGE_SHIFT)) - 1);
 940                 cli->cl_max_pages_per_rpc = (cli->cl_max_pages_per_rpc +
 941                                              ~chunk_mask) & chunk_mask;
 942                 /* determine maximum extent size, in #pages */
 943                 size = (u64)ocd->ocd_grant_max_blks << ocd->ocd_grant_blkbits;
 944                 cli->cl_max_extent_pages = size >> PAGE_SHIFT;
 945                 if (cli->cl_max_extent_pages == 0)
 946                         cli->cl_max_extent_pages = 1;
 947         } else {
 948                 cli->cl_grant_extent_tax = 0;
 949                 cli->cl_chunkbits = PAGE_SHIFT;
 950                 cli->cl_max_extent_pages = DT_MAX_BRW_PAGES;
 951         }
 952         spin_unlock(&cli->cl_loi_list_lock);
 953
 954         CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld."
 955                 "chunk bits: %d cl_max_extent_pages: %d\n",
 956                 cli_name(cli),
 957                 cli->cl_avail_grant, cli->cl_lost_grant, cli->cl_chunkbits,
 958                 cli->cl_max_extent_pages);
 959
 960         if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
 961             list_empty(&cli->cl_grant_shrink_list))
 962                 osc_add_shrink_grant(cli);
 963 }
 964
 965 /* We assume that the reason this OSC got a short read is because it read
 966  * beyond the end of a stripe file; i.e. lustre is reading a sparse file
 967  * via the LOV, and it _knows_ it's reading inside the file, it's just that
 968  * this stripe never got written at or beyond this stripe offset yet. */
 969 static void handle_short_read(int nob_read, size_t page_count,
 970                               struct brw_page **pga)
 971 {
 972         char *ptr;
 973         int i = 0;
 974
 975         /* skip bytes read OK */
 976         while (nob_read > 0) {
 977                 LASSERT (page_count > 0);
 978
 979                 if (pga[i]->count > nob_read) {
 980                         /* EOF inside this page */
 981                         ptr = kmap(pga[i]->pg) +
 982                                 (pga[i]->off & ~PAGE_MASK);
 983                         memset(ptr + nob_read, 0, pga[i]->count - nob_read);
 984                         kunmap(pga[i]->pg);
 985                         page_count--;
 986                         i++;
 987                         break;
 988                 }
 989
 990                 nob_read -= pga[i]->count;
 991                 page_count--;
 992                 i++;
 993         }
 994
 995         /* zero remaining pages */
 996         while (page_count-- > 0) {
 997                 ptr = kmap(pga[i]->pg) + (pga[i]->off & ~PAGE_MASK);
 998                 memset(ptr, 0, pga[i]->count);
 999                 kunmap(pga[i]->pg);
1000                 i++;
1001         }
1002 }
1003
1004 static int check_write_rcs(struct ptlrpc_request *req,
1005                            int requested_nob, int niocount,
1006                            size_t page_count, struct brw_page **pga)
1007 {
1008         int     i;
1009         __u32   *remote_rcs;
1010
1011         remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1012                                                   sizeof(*remote_rcs) *
1013                                                   niocount);
1014         if (remote_rcs == NULL) {
1015                 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1016                 return(-EPROTO);
1017         }
1018
1019         /* return error if any niobuf was in error */
1020         for (i = 0; i < niocount; i++) {
1021                 if ((int)remote_rcs[i] < 0)
1022                         return(remote_rcs[i]);
1023
1024                 if (remote_rcs[i] != 0) {
1025                         CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1026                                 i, remote_rcs[i], req);
1027                         return(-EPROTO);
1028                 }
1029         }
1030
1031         if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1032                 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1033                        req->rq_bulk->bd_nob_transferred, requested_nob);
1034                 return(-EPROTO);
1035         }
1036
1037         return (0);
1038 }
1039
1040 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1041 {
1042         if (p1->flag != p2->flag) {
1043                 unsigned mask = ~(OBD_BRW_FROM_GRANT | OBD_BRW_NOCACHE |
1044                                   OBD_BRW_SYNC       | OBD_BRW_ASYNC   |
1045                                   OBD_BRW_NOQUOTA    | OBD_BRW_SOFT_SYNC);
1046
1047                 /* warn if we try to combine flags that we don't know to be
1048                  * safe to combine */
1049                 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
1050                         CWARN("Saw flags 0x%x and 0x%x in the same brw, please "
1051                               "report this at https://jira.hpdd.intel.com/\n",
1052                               p1->flag, p2->flag);
1053                 }
1054                 return 0;
1055         }
1056
1057         return (p1->off + p1->count == p2->off);
1058 }
1059
1060 static u32 osc_checksum_bulk(int nob, size_t pg_count,
1061                              struct brw_page **pga, int opc,
1062                              cksum_type_t cksum_type)
1063 {
1064         u32                             cksum;
1065         int                             i = 0;
1066         struct cfs_crypto_hash_desc     *hdesc;
1067         unsigned int                    bufsize;
1068         unsigned char                   cfs_alg = cksum_obd2cfs(cksum_type);
1069
1070         LASSERT(pg_count > 0);
1071
1072         hdesc = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1073         if (IS_ERR(hdesc)) {
1074                 CERROR("Unable to initialize checksum hash %s\n",
1075                        cfs_crypto_hash_name(cfs_alg));
1076                 return PTR_ERR(hdesc);
1077         }
1078
1079         while (nob > 0 && pg_count > 0) {
1080                 unsigned int count = pga[i]->count > nob ? nob : pga[i]->count;
1081
1082                 /* corrupt the data before we compute the checksum, to
1083                  * simulate an OST->client data error */
1084                 if (i == 0 && opc == OST_READ &&
1085                     OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE)) {
1086                         unsigned char *ptr = kmap(pga[i]->pg);
1087                         int off = pga[i]->off & ~PAGE_MASK;
1088
1089                         memcpy(ptr + off, "bad1", min_t(typeof(nob), 4, nob));
1090                         kunmap(pga[i]->pg);
1091                 }
1092                 cfs_crypto_hash_update_page(hdesc, pga[i]->pg,
1093                                             pga[i]->off & ~PAGE_MASK,
1094                                             count);
1095                 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d\n",
1096                                (int)(pga[i]->off & ~PAGE_MASK));
1097
1098                 nob -= pga[i]->count;
1099                 pg_count--;
1100                 i++;
1101         }
1102
1103         bufsize = sizeof(cksum);
1104         cfs_crypto_hash_final(hdesc, (unsigned char *)&cksum, &bufsize);
1105
1106         /* For sending we only compute the wrong checksum instead
1107          * of corrupting the data so it is still correct on a redo */
1108         if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1109                 cksum++;
1110
1111         return cksum;
1112 }
1113
1114 static int
1115 osc_brw_prep_request(int cmd, struct client_obd *cli, struct obdo *oa,
1116                      u32 page_count, struct brw_page **pga,
1117                      struct ptlrpc_request **reqp, int resend)
1118 {
1119         struct ptlrpc_request   *req;
1120         struct ptlrpc_bulk_desc *desc;
1121         struct ost_body         *body;
1122         struct obd_ioobj        *ioobj;
1123         struct niobuf_remote    *niobuf;
1124         int niocount, i, requested_nob, opc, rc;
1125         struct osc_brw_async_args *aa;
1126         struct req_capsule      *pill;
1127         struct brw_page *pg_prev;
1128
1129         ENTRY;
1130         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1131                 RETURN(-ENOMEM); /* Recoverable */
1132         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1133                 RETURN(-EINVAL); /* Fatal */
1134
1135         if ((cmd & OBD_BRW_WRITE) != 0) {
1136                 opc = OST_WRITE;
1137                 req = ptlrpc_request_alloc_pool(cli->cl_import,
1138                                                 osc_rq_pool,
1139                                                 &RQF_OST_BRW_WRITE);
1140         } else {
1141                 opc = OST_READ;
1142                 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1143         }
1144         if (req == NULL)
1145                 RETURN(-ENOMEM);
1146
1147         for (niocount = i = 1; i < page_count; i++) {
1148                 if (!can_merge_pages(pga[i - 1], pga[i]))
1149                         niocount++;
1150         }
1151
1152         pill = &req->rq_pill;
1153         req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1154                              sizeof(*ioobj));
1155         req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1156                              niocount * sizeof(*niobuf));
1157
1158         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1159         if (rc) {
1160                 ptlrpc_request_free(req);
1161                 RETURN(rc);
1162         }
1163         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1164         ptlrpc_at_set_req_timeout(req);
1165         /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1166          * retry logic */
1167         req->rq_no_retry_einprogress = 1;
1168
1169         desc = ptlrpc_prep_bulk_imp(req, page_count,
1170                 cli->cl_import->imp_connect_data.ocd_brw_size >> LNET_MTU_BITS,
1171                 (opc == OST_WRITE ? PTLRPC_BULK_GET_SOURCE :
1172                         PTLRPC_BULK_PUT_SINK) |
1173                         PTLRPC_BULK_BUF_KIOV,
1174                 OST_BULK_PORTAL,
1175                 &ptlrpc_bulk_kiov_pin_ops);
1176
1177         if (desc == NULL)
1178                 GOTO(out, rc = -ENOMEM);
1179         /* NB request now owns desc and will free it when it gets freed */
1180
1181         body = req_capsule_client_get(pill, &RMF_OST_BODY);
1182         ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1183         niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1184         LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1185
1186         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1187
1188         obdo_to_ioobj(oa, ioobj);
1189         ioobj->ioo_bufcnt = niocount;
1190         /* The high bits of ioo_max_brw tells server _maximum_ number of bulks
1191          * that might be send for this request.  The actual number is decided
1192          * when the RPC is finally sent in ptlrpc_register_bulk(). It sends
1193          * "max - 1" for old client compatibility sending "0", and also so the
1194          * the actual maximum is a power-of-two number, not one less. LU-1431 */
1195         ioobj_max_brw_set(ioobj, desc->bd_md_max_brw);
1196         LASSERT(page_count > 0);
1197         pg_prev = pga[0];
1198         for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1199                 struct brw_page *pg = pga[i];
1200                 int poff = pg->off & ~PAGE_MASK;
1201
1202                 LASSERT(pg->count > 0);
1203                 /* make sure there is no gap in the middle of page array */
1204                 LASSERTF(page_count == 1 ||
1205                          (ergo(i == 0, poff + pg->count == PAGE_SIZE) &&
1206                           ergo(i > 0 && i < page_count - 1,
1207                                poff == 0 && pg->count == PAGE_SIZE)   &&
1208                           ergo(i == page_count - 1, poff == 0)),
1209                          "i: %d/%d pg: %p off: %llu, count: %u\n",
1210                          i, page_count, pg, pg->off, pg->count);
1211                 LASSERTF(i == 0 || pg->off > pg_prev->off,
1212                          "i %d p_c %u pg %p [pri %lu ind %lu] off %llu"
1213                          " prev_pg %p [pri %lu ind %lu] off %llu\n",
1214                          i, page_count,
1215                          pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1216                          pg_prev->pg, page_private(pg_prev->pg),
1217                          pg_prev->pg->index, pg_prev->off);
1218                 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1219                         (pg->flag & OBD_BRW_SRVLOCK));
1220
1221                 desc->bd_frag_ops->add_kiov_frag(desc, pg->pg, poff, pg->count);
1222                 requested_nob += pg->count;
1223
1224                 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1225                         niobuf--;
1226                         niobuf->rnb_len += pg->count;
1227                 } else {
1228                         niobuf->rnb_offset = pg->off;
1229                         niobuf->rnb_len    = pg->count;
1230                         niobuf->rnb_flags  = pg->flag;
1231                 }
1232                 pg_prev = pg;
1233         }
1234
1235         LASSERTF((void *)(niobuf - niocount) ==
1236                 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1237                 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1238                 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1239
1240         osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1241         if (resend) {
1242                 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1243                         body->oa.o_valid |= OBD_MD_FLFLAGS;
1244                         body->oa.o_flags = 0;
1245                 }
1246                 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1247         }
1248
1249         if (osc_should_shrink_grant(cli))
1250                 osc_shrink_grant_local(cli, &body->oa);
1251
1252         /* size[REQ_REC_OFF] still sizeof (*body) */
1253         if (opc == OST_WRITE) {
1254                 if (cli->cl_checksum &&
1255                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1256                         /* store cl_cksum_type in a local variable since
1257                          * it can be changed via lprocfs */
1258                         cksum_type_t cksum_type = cli->cl_cksum_type;
1259
1260                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1261                                 oa->o_flags &= OBD_FL_LOCAL_MASK;
1262                                 body->oa.o_flags = 0;
1263                         }
1264                         body->oa.o_flags |= cksum_type_pack(cksum_type);
1265                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1266                         body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1267                                                              page_count, pga,
1268                                                              OST_WRITE,
1269                                                              cksum_type);
1270                         CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1271                                body->oa.o_cksum);
1272                         /* save this in 'oa', too, for later checking */
1273                         oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1274                         oa->o_flags |= cksum_type_pack(cksum_type);
1275                 } else {
1276                         /* clear out the checksum flag, in case this is a
1277                          * resend but cl_checksum is no longer set. b=11238 */
1278                         oa->o_valid &= ~OBD_MD_FLCKSUM;
1279                 }
1280                 oa->o_cksum = body->oa.o_cksum;
1281                 /* 1 RC per niobuf */
1282                 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1283                                      sizeof(__u32) * niocount);
1284         } else {
1285                 if (cli->cl_checksum &&
1286                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1287                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1288                                 body->oa.o_flags = 0;
1289                         body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1290                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1291                 }
1292         }
1293         ptlrpc_request_set_replen(req);
1294
1295         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1296         aa = ptlrpc_req_async_args(req);
1297         aa->aa_oa = oa;
1298         aa->aa_requested_nob = requested_nob;
1299         aa->aa_nio_count = niocount;
1300         aa->aa_page_count = page_count;
1301         aa->aa_resends = 0;
1302         aa->aa_ppga = pga;
1303         aa->aa_cli = cli;
1304         INIT_LIST_HEAD(&aa->aa_oaps);
1305
1306         *reqp = req;
1307         niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1308         CDEBUG(D_RPCTRACE, "brw rpc %p - object "DOSTID" offset %lld<>%lld\n",
1309                 req, POSTID(&oa->o_oi), niobuf[0].rnb_offset,
1310                 niobuf[niocount - 1].rnb_offset + niobuf[niocount - 1].rnb_len);
1311         RETURN(0);
1312
1313  out:
1314         ptlrpc_req_finished(req);
1315         RETURN(rc);
1316 }
1317
1318 static int
1319 check_write_checksum(struct obdo *oa, const struct lnet_process_id *peer,
1320                      __u32 client_cksum, __u32 server_cksum, int nob,
1321                      size_t page_count, struct brw_page **pga,
1322                      cksum_type_t client_cksum_type)
1323 {
1324         __u32 new_cksum;
1325         char *msg;
1326         cksum_type_t cksum_type;
1327
1328         if (server_cksum == client_cksum) {
1329                 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1330                 return 0;
1331         }
1332
1333         cksum_type = cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1334                                        oa->o_flags : 0);
1335         new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1336                                       cksum_type);
1337
1338         if (cksum_type != client_cksum_type)
1339                 msg = "the server did not use the checksum type specified in "
1340                       "the original request - likely a protocol problem";
1341         else if (new_cksum == server_cksum)
1342                 msg = "changed on the client after we checksummed it - "
1343                       "likely false positive due to mmap IO (bug 11742)";
1344         else if (new_cksum == client_cksum)
1345                 msg = "changed in transit before arrival at OST";
1346         else
1347                 msg = "changed in transit AND doesn't match the original - "
1348                       "likely false positive due to mmap IO (bug 11742)";
1349
1350         LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode "DFID
1351                            " object "DOSTID" extent [%llu-%llu]\n",
1352                            msg, libcfs_nid2str(peer->nid),
1353                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1354                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1355                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1356                            POSTID(&oa->o_oi), pga[0]->off,
1357                            pga[page_count-1]->off + pga[page_count-1]->count - 1);
1358         CERROR("original client csum %x (type %x), server csum %x (type %x), "
1359                "client csum now %x\n", client_cksum, client_cksum_type,
1360                server_cksum, cksum_type, new_cksum);
1361         return 1;
1362 }
1363
1364 /* Note rc enters this function as number of bytes transferred */
1365 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1366 {
1367         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1368         const struct lnet_process_id *peer =
1369                         &req->rq_import->imp_connection->c_peer;
1370         struct client_obd *cli = aa->aa_cli;
1371         struct ost_body *body;
1372         u32 client_cksum = 0;
1373         ENTRY;
1374
1375         if (rc < 0 && rc != -EDQUOT) {
1376                 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1377                 RETURN(rc);
1378         }
1379
1380         LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1381         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1382         if (body == NULL) {
1383                 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1384                 RETURN(-EPROTO);
1385         }
1386
1387         /* set/clear over quota flag for a uid/gid */
1388         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1389             body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1390                 unsigned int qid[LL_MAXQUOTAS] =
1391                                         {body->oa.o_uid, body->oa.o_gid};
1392
1393                 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid %#llx, flags %x\n",
1394                        body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1395                        body->oa.o_flags);
1396                 osc_quota_setdq(cli, qid, body->oa.o_valid, body->oa.o_flags);
1397         }
1398
1399         osc_update_grant(cli, body);
1400
1401         if (rc < 0)
1402                 RETURN(rc);
1403
1404         if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1405                 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1406
1407         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1408                 if (rc > 0) {
1409                         CERROR("Unexpected +ve rc %d\n", rc);
1410                         RETURN(-EPROTO);
1411                 }
1412                 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1413
1414                 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1415                         RETURN(-EAGAIN);
1416
1417                 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1418                     check_write_checksum(&body->oa, peer, client_cksum,
1419                                          body->oa.o_cksum, aa->aa_requested_nob,
1420                                          aa->aa_page_count, aa->aa_ppga,
1421                                          cksum_type_unpack(aa->aa_oa->o_flags)))
1422                         RETURN(-EAGAIN);
1423
1424                 rc = check_write_rcs(req, aa->aa_requested_nob,aa->aa_nio_count,
1425                                      aa->aa_page_count, aa->aa_ppga);
1426                 GOTO(out, rc);
1427         }
1428
1429         /* The rest of this function executes only for OST_READs */
1430
1431         /* if unwrap_bulk failed, return -EAGAIN to retry */
1432         rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1433         if (rc < 0)
1434                 GOTO(out, rc = -EAGAIN);
1435
1436         if (rc > aa->aa_requested_nob) {
1437                 CERROR("Unexpected rc %d (%d requested)\n", rc,
1438                        aa->aa_requested_nob);
1439                 RETURN(-EPROTO);
1440         }
1441
1442         if (rc != req->rq_bulk->bd_nob_transferred) {
1443                 CERROR ("Unexpected rc %d (%d transferred)\n",
1444                         rc, req->rq_bulk->bd_nob_transferred);
1445                 return (-EPROTO);
1446         }
1447
1448         if (rc < aa->aa_requested_nob)
1449                 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1450
1451         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1452                 static int cksum_counter;
1453                 u32        server_cksum = body->oa.o_cksum;
1454                 char      *via = "";
1455                 char      *router = "";
1456                 cksum_type_t cksum_type;
1457
1458                 cksum_type = cksum_type_unpack(body->oa.o_valid &OBD_MD_FLFLAGS?
1459                                                body->oa.o_flags : 0);
1460                 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1461                                                  aa->aa_ppga, OST_READ,
1462                                                  cksum_type);
1463
1464                 if (peer->nid != req->rq_bulk->bd_sender) {
1465                         via = " via ";
1466                         router = libcfs_nid2str(req->rq_bulk->bd_sender);
1467                 }
1468
1469                 if (server_cksum != client_cksum) {
1470                         LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
1471                                            "%s%s%s inode "DFID" object "DOSTID
1472                                            " extent [%llu-%llu]\n",
1473                                            req->rq_import->imp_obd->obd_name,
1474                                            libcfs_nid2str(peer->nid),
1475                                            via, router,
1476                                            body->oa.o_valid & OBD_MD_FLFID ?
1477                                                 body->oa.o_parent_seq : (__u64)0,
1478                                            body->oa.o_valid & OBD_MD_FLFID ?
1479                                                 body->oa.o_parent_oid : 0,
1480                                            body->oa.o_valid & OBD_MD_FLFID ?
1481                                                 body->oa.o_parent_ver : 0,
1482                                            POSTID(&body->oa.o_oi),
1483                                            aa->aa_ppga[0]->off,
1484                                            aa->aa_ppga[aa->aa_page_count-1]->off +
1485                                            aa->aa_ppga[aa->aa_page_count-1]->count -
1486                                                                         1);
1487                         CERROR("client %x, server %x, cksum_type %x\n",
1488                                client_cksum, server_cksum, cksum_type);
1489                         cksum_counter = 0;
1490                         aa->aa_oa->o_cksum = client_cksum;
1491                         rc = -EAGAIN;
1492                 } else {
1493                         cksum_counter++;
1494                         CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1495                         rc = 0;
1496                 }
1497         } else if (unlikely(client_cksum)) {
1498                 static int cksum_missed;
1499
1500                 cksum_missed++;
1501                 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1502                         CERROR("Checksum %u requested from %s but not sent\n",
1503                                cksum_missed, libcfs_nid2str(peer->nid));
1504         } else {
1505                 rc = 0;
1506         }
1507 out:
1508         if (rc >= 0)
1509                 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
1510                                      aa->aa_oa, &body->oa);
1511
1512         RETURN(rc);
1513 }
1514
1515 static int osc_brw_redo_request(struct ptlrpc_request *request,
1516                                 struct osc_brw_async_args *aa, int rc)
1517 {
1518         struct ptlrpc_request *new_req;
1519         struct osc_brw_async_args *new_aa;
1520         struct osc_async_page *oap;
1521         ENTRY;
1522
1523         DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
1524                   "redo for recoverable error %d", rc);
1525
1526         rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1527                                 OST_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
1528                                   aa->aa_cli, aa->aa_oa, aa->aa_page_count,
1529                                   aa->aa_ppga, &new_req, 1);
1530         if (rc)
1531                 RETURN(rc);
1532
1533         list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1534                 if (oap->oap_request != NULL) {
1535                         LASSERTF(request == oap->oap_request,
1536                                  "request %p != oap_request %p\n",
1537                                  request, oap->oap_request);
1538                         if (oap->oap_interrupted) {
1539                                 ptlrpc_req_finished(new_req);
1540                                 RETURN(-EINTR);
1541                         }
1542                 }
1543         }
1544         /* New request takes over pga and oaps from old request.
1545          * Note that copying a list_head doesn't work, need to move it... */
1546         aa->aa_resends++;
1547         new_req->rq_interpret_reply = request->rq_interpret_reply;
1548         new_req->rq_async_args = request->rq_async_args;
1549         new_req->rq_commit_cb = request->rq_commit_cb;
1550         /* cap resend delay to the current request timeout, this is similar to
1551          * what ptlrpc does (see after_reply()) */
1552         if (aa->aa_resends > new_req->rq_timeout)
1553                 new_req->rq_sent = cfs_time_current_sec() + new_req->rq_timeout;
1554         else
1555                 new_req->rq_sent = cfs_time_current_sec() + aa->aa_resends;
1556         new_req->rq_generation_set = 1;
1557         new_req->rq_import_generation = request->rq_import_generation;
1558
1559         new_aa = ptlrpc_req_async_args(new_req);
1560
1561         INIT_LIST_HEAD(&new_aa->aa_oaps);
1562         list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
1563         INIT_LIST_HEAD(&new_aa->aa_exts);
1564         list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
1565         new_aa->aa_resends = aa->aa_resends;
1566
1567         list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1568                 if (oap->oap_request) {
1569                         ptlrpc_req_finished(oap->oap_request);
1570                         oap->oap_request = ptlrpc_request_addref(new_req);
1571                 }
1572         }
1573
1574         /* XXX: This code will run into problem if we're going to support
1575          * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
1576          * and wait for all of them to be finished. We should inherit request
1577          * set from old request. */
1578         ptlrpcd_add_req(new_req);
1579
1580         DEBUG_REQ(D_INFO, new_req, "new request");
1581         RETURN(0);
1582 }
1583
1584 /*
1585  * ugh, we want disk allocation on the target to happen in offset order.  we'll
1586  * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1587  * fine for our small page arrays and doesn't require allocation.  its an
1588  * insertion sort that swaps elements that are strides apart, shrinking the
1589  * stride down until its '1' and the array is sorted.
1590  */
1591 static void sort_brw_pages(struct brw_page **array, int num)
1592 {
1593         int stride, i, j;
1594         struct brw_page *tmp;
1595
1596         if (num == 1)
1597                 return;
1598         for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1599                 ;
1600
1601         do {
1602                 stride /= 3;
1603                 for (i = stride ; i < num ; i++) {
1604                         tmp = array[i];
1605                         j = i;
1606                         while (j >= stride && array[j - stride]->off > tmp->off) {
1607                                 array[j] = array[j - stride];
1608                                 j -= stride;
1609                         }
1610                         array[j] = tmp;
1611                 }
1612         } while (stride > 1);
1613 }
1614
1615 static void osc_release_ppga(struct brw_page **ppga, size_t count)
1616 {
1617         LASSERT(ppga != NULL);
1618         OBD_FREE(ppga, sizeof(*ppga) * count);
1619 }
1620
1621 static int brw_interpret(const struct lu_env *env,
1622                          struct ptlrpc_request *req, void *data, int rc)
1623 {
1624         struct osc_brw_async_args *aa = data;
1625         struct osc_extent *ext;
1626         struct osc_extent *tmp;
1627         struct client_obd *cli = aa->aa_cli;
1628         ENTRY;
1629
1630         rc = osc_brw_fini_request(req, rc);
1631         CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
1632         /* When server return -EINPROGRESS, client should always retry
1633          * regardless of the number of times the bulk was resent already. */
1634         if (osc_recoverable_error(rc)) {
1635                 if (req->rq_import_generation !=
1636                     req->rq_import->imp_generation) {
1637                         CDEBUG(D_HA, "%s: resend cross eviction for object: "
1638                                ""DOSTID", rc = %d.\n",
1639                                req->rq_import->imp_obd->obd_name,
1640                                POSTID(&aa->aa_oa->o_oi), rc);
1641                 } else if (rc == -EINPROGRESS ||
1642                     client_should_resend(aa->aa_resends, aa->aa_cli)) {
1643                         rc = osc_brw_redo_request(req, aa, rc);
1644                 } else {
1645                         CERROR("%s: too many resent retries for object: "
1646                                "%llu:%llu, rc = %d.\n",
1647                                req->rq_import->imp_obd->obd_name,
1648                                POSTID(&aa->aa_oa->o_oi), rc);
1649                 }
1650
1651                 if (rc == 0)
1652                         RETURN(0);
1653                 else if (rc == -EAGAIN || rc == -EINPROGRESS)
1654                         rc = -EIO;
1655         }
1656
1657         if (rc == 0) {
1658                 struct obdo *oa = aa->aa_oa;
1659                 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
1660                 unsigned long valid = 0;
1661                 struct cl_object *obj;
1662                 struct osc_async_page *last;
1663
1664                 last = brw_page2oap(aa->aa_ppga[aa->aa_page_count - 1]);
1665                 obj = osc2cl(last->oap_obj);
1666
1667                 cl_object_attr_lock(obj);
1668                 if (oa->o_valid & OBD_MD_FLBLOCKS) {
1669                         attr->cat_blocks = oa->o_blocks;
1670                         valid |= CAT_BLOCKS;
1671                 }
1672                 if (oa->o_valid & OBD_MD_FLMTIME) {
1673                         attr->cat_mtime = oa->o_mtime;
1674                         valid |= CAT_MTIME;
1675                 }
1676                 if (oa->o_valid & OBD_MD_FLATIME) {
1677                         attr->cat_atime = oa->o_atime;
1678                         valid |= CAT_ATIME;
1679                 }
1680                 if (oa->o_valid & OBD_MD_FLCTIME) {
1681                         attr->cat_ctime = oa->o_ctime;
1682                         valid |= CAT_CTIME;
1683                 }
1684
1685                 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1686                         struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
1687                         loff_t last_off = last->oap_count + last->oap_obj_off +
1688                                 last->oap_page_off;
1689
1690                         /* Change file size if this is an out of quota or
1691                          * direct IO write and it extends the file size */
1692                         if (loi->loi_lvb.lvb_size < last_off) {
1693                                 attr->cat_size = last_off;
1694                                 valid |= CAT_SIZE;
1695                         }
1696                         /* Extend KMS if it's not a lockless write */
1697                         if (loi->loi_kms < last_off &&
1698                             oap2osc_page(last)->ops_srvlock == 0) {
1699                                 attr->cat_kms = last_off;
1700                                 valid |= CAT_KMS;
1701                         }
1702                 }
1703
1704                 if (valid != 0)
1705                         cl_object_attr_update(env, obj, attr, valid);
1706                 cl_object_attr_unlock(obj);
1707         }
1708         OBDO_FREE(aa->aa_oa);
1709
1710         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE && rc == 0)
1711                 osc_inc_unstable_pages(req);
1712
1713         list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
1714                 list_del_init(&ext->oe_link);
1715                 osc_extent_finish(env, ext, 1, rc);
1716         }
1717         LASSERT(list_empty(&aa->aa_exts));
1718         LASSERT(list_empty(&aa->aa_oaps));
1719
1720         osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
1721         ptlrpc_lprocfs_brw(req, req->rq_bulk->bd_nob_transferred);
1722
1723         spin_lock(&cli->cl_loi_list_lock);
1724         /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
1725          * is called so we know whether to go to sync BRWs or wait for more
1726          * RPCs to complete */
1727         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
1728                 cli->cl_w_in_flight--;
1729         else
1730                 cli->cl_r_in_flight--;
1731         osc_wake_cache_waiters(cli);
1732         spin_unlock(&cli->cl_loi_list_lock);
1733
1734         osc_io_unplug(env, cli, NULL);
1735         RETURN(rc);
1736 }
1737
1738 static void brw_commit(struct ptlrpc_request *req)
1739 {
1740         /* If osc_inc_unstable_pages (via osc_extent_finish) races with
1741          * this called via the rq_commit_cb, I need to ensure
1742          * osc_dec_unstable_pages is still called. Otherwise unstable
1743          * pages may be leaked. */
1744         spin_lock(&req->rq_lock);
1745         if (likely(req->rq_unstable)) {
1746                 req->rq_unstable = 0;
1747                 spin_unlock(&req->rq_lock);
1748
1749                 osc_dec_unstable_pages(req);
1750         } else {
1751                 req->rq_committed = 1;
1752                 spin_unlock(&req->rq_lock);
1753         }
1754 }
1755
1756 /**
1757  * Build an RPC by the list of extent @ext_list. The caller must ensure
1758  * that the total pages in this list are NOT over max pages per RPC.
1759  * Extents in the list must be in OES_RPC state.
1760  */
1761 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
1762                   struct list_head *ext_list, int cmd)
1763 {
1764         struct ptlrpc_request           *req = NULL;
1765         struct osc_extent               *ext;
1766         struct brw_page                 **pga = NULL;
1767         struct osc_brw_async_args       *aa = NULL;
1768         struct obdo                     *oa = NULL;
1769         struct osc_async_page           *oap;
1770         struct osc_object               *obj = NULL;
1771         struct cl_req_attr              *crattr = NULL;
1772         loff_t                          starting_offset = OBD_OBJECT_EOF;
1773         loff_t                          ending_offset = 0;
1774         int                             mpflag = 0;
1775         int                             mem_tight = 0;
1776         int                             page_count = 0;
1777         bool                            soft_sync = false;
1778         bool                            interrupted = false;
1779         int                             i;
1780         int                             grant = 0;
1781         int                             rc;
1782         struct list_head                rpc_list = LIST_HEAD_INIT(rpc_list);
1783         struct ost_body                 *body;
1784         ENTRY;
1785         LASSERT(!list_empty(ext_list));
1786
1787         /* add pages into rpc_list to build BRW rpc */
1788         list_for_each_entry(ext, ext_list, oe_link) {
1789                 LASSERT(ext->oe_state == OES_RPC);
1790                 mem_tight |= ext->oe_memalloc;
1791                 grant += ext->oe_grants;
1792                 page_count += ext->oe_nr_pages;
1793                 if (obj == NULL)
1794                         obj = ext->oe_obj;
1795         }
1796
1797         soft_sync = osc_over_unstable_soft_limit(cli);
1798         if (mem_tight)
1799                 mpflag = cfs_memory_pressure_get_and_set();
1800
1801         OBD_ALLOC(pga, sizeof(*pga) * page_count);
1802         if (pga == NULL)
1803                 GOTO(out, rc = -ENOMEM);
1804
1805         OBDO_ALLOC(oa);
1806         if (oa == NULL)
1807                 GOTO(out, rc = -ENOMEM);
1808
1809         i = 0;
1810         list_for_each_entry(ext, ext_list, oe_link) {
1811                 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1812                         if (mem_tight)
1813                                 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
1814                         if (soft_sync)
1815                                 oap->oap_brw_flags |= OBD_BRW_SOFT_SYNC;
1816                         pga[i] = &oap->oap_brw_page;
1817                         pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
1818                         i++;
1819
1820                         list_add_tail(&oap->oap_rpc_item, &rpc_list);
1821                         if (starting_offset == OBD_OBJECT_EOF ||
1822                             starting_offset > oap->oap_obj_off)
1823                                 starting_offset = oap->oap_obj_off;
1824                         else
1825                                 LASSERT(oap->oap_page_off == 0);
1826                         if (ending_offset < oap->oap_obj_off + oap->oap_count)
1827                                 ending_offset = oap->oap_obj_off +
1828                                                 oap->oap_count;
1829                         else
1830                                 LASSERT(oap->oap_page_off + oap->oap_count ==
1831                                         PAGE_SIZE);
1832                         if (oap->oap_interrupted)
1833                                 interrupted = true;
1834                 }
1835         }
1836
1837         /* first page in the list */
1838         oap = list_entry(rpc_list.next, typeof(*oap), oap_rpc_item);
1839
1840         crattr = &osc_env_info(env)->oti_req_attr;
1841         memset(crattr, 0, sizeof(*crattr));
1842         crattr->cra_type = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
1843         crattr->cra_flags = ~0ULL;
1844         crattr->cra_page = oap2cl_page(oap);
1845         crattr->cra_oa = oa;
1846         cl_req_attr_set(env, osc2cl(obj), crattr);
1847
1848         if (cmd == OBD_BRW_WRITE)
1849                 oa->o_grant_used = grant;
1850
1851         sort_brw_pages(pga, page_count);
1852         rc = osc_brw_prep_request(cmd, cli, oa, page_count, pga, &req, 0);
1853         if (rc != 0) {
1854                 CERROR("prep_req failed: %d\n", rc);
1855                 GOTO(out, rc);
1856         }
1857
1858         req->rq_commit_cb = brw_commit;
1859         req->rq_interpret_reply = brw_interpret;
1860         req->rq_memalloc = mem_tight != 0;
1861         oap->oap_request = ptlrpc_request_addref(req);
1862         if (interrupted && !req->rq_intr)
1863                 ptlrpc_mark_interrupted(req);
1864
1865         /* Need to update the timestamps after the request is built in case
1866          * we race with setattr (locally or in queue at OST).  If OST gets
1867          * later setattr before earlier BRW (as determined by the request xid),
1868          * the OST will not use BRW timestamps.  Sadly, there is no obvious
1869          * way to do this in a single call.  bug 10150 */
1870         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1871         crattr->cra_oa = &body->oa;
1872         crattr->cra_flags = OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME;
1873         cl_req_attr_set(env, osc2cl(obj), crattr);
1874         lustre_msg_set_jobid(req->rq_reqmsg, crattr->cra_jobid);
1875
1876         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1877         aa = ptlrpc_req_async_args(req);
1878         INIT_LIST_HEAD(&aa->aa_oaps);
1879         list_splice_init(&rpc_list, &aa->aa_oaps);
1880         INIT_LIST_HEAD(&aa->aa_exts);
1881         list_splice_init(ext_list, &aa->aa_exts);
1882
1883         spin_lock(&cli->cl_loi_list_lock);
1884         starting_offset >>= PAGE_SHIFT;
1885         if (cmd == OBD_BRW_READ) {
1886                 cli->cl_r_in_flight++;
1887                 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
1888                 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
1889                 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
1890                                       starting_offset + 1);
1891         } else {
1892                 cli->cl_w_in_flight++;
1893                 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
1894                 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
1895                 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
1896                                       starting_offset + 1);
1897         }
1898         spin_unlock(&cli->cl_loi_list_lock);
1899
1900         DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %ur/%uw in flight",
1901                   page_count, aa, cli->cl_r_in_flight,
1902                   cli->cl_w_in_flight);
1903         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_IO, cfs_fail_val);
1904
1905         ptlrpcd_add_req(req);
1906         rc = 0;
1907         EXIT;
1908
1909 out:
1910         if (mem_tight != 0)
1911                 cfs_memory_pressure_restore(mpflag);
1912
1913         if (rc != 0) {
1914                 LASSERT(req == NULL);
1915
1916                 if (oa)
1917                         OBDO_FREE(oa);
1918                 if (pga)
1919                         OBD_FREE(pga, sizeof(*pga) * page_count);
1920                 /* this should happen rarely and is pretty bad, it makes the
1921                  * pending list not follow the dirty order */
1922                 while (!list_empty(ext_list)) {
1923                         ext = list_entry(ext_list->next, struct osc_extent,
1924                                          oe_link);
1925                         list_del_init(&ext->oe_link);
1926                         osc_extent_finish(env, ext, 0, rc);
1927                 }
1928         }
1929         RETURN(rc);
1930 }
1931
1932 static int osc_set_lock_data(struct ldlm_lock *lock, void *data)
1933 {
1934         int set = 0;
1935
1936         LASSERT(lock != NULL);
1937
1938         lock_res_and_lock(lock);
1939
1940         if (lock->l_ast_data == NULL)
1941                 lock->l_ast_data = data;
1942         if (lock->l_ast_data == data)
1943                 set = 1;
1944
1945         unlock_res_and_lock(lock);
1946
1947         return set;
1948 }
1949
1950 static int osc_enqueue_fini(struct ptlrpc_request *req,
1951                             osc_enqueue_upcall_f upcall, void *cookie,
1952                             struct lustre_handle *lockh, enum ldlm_mode mode,
1953                             __u64 *flags, int agl, int errcode)
1954 {
1955         bool intent = *flags & LDLM_FL_HAS_INTENT;
1956         int rc;
1957         ENTRY;
1958
1959         /* The request was created before ldlm_cli_enqueue call. */
1960         if (intent && errcode == ELDLM_LOCK_ABORTED) {
1961                 struct ldlm_reply *rep;
1962
1963                 rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1964                 LASSERT(rep != NULL);
1965
1966                 rep->lock_policy_res1 =
1967                         ptlrpc_status_ntoh(rep->lock_policy_res1);
1968                 if (rep->lock_policy_res1)
1969                         errcode = rep->lock_policy_res1;
1970                 if (!agl)
1971                         *flags |= LDLM_FL_LVB_READY;
1972         } else if (errcode == ELDLM_OK) {
1973                 *flags |= LDLM_FL_LVB_READY;
1974         }
1975
1976         /* Call the update callback. */
1977         rc = (*upcall)(cookie, lockh, errcode);
1978
1979         /* release the reference taken in ldlm_cli_enqueue() */
1980         if (errcode == ELDLM_LOCK_MATCHED)
1981                 errcode = ELDLM_OK;
1982         if (errcode == ELDLM_OK && lustre_handle_is_used(lockh))
1983                 ldlm_lock_decref(lockh, mode);
1984
1985         RETURN(rc);
1986 }
1987
1988 static int osc_enqueue_interpret(const struct lu_env *env,
1989                                  struct ptlrpc_request *req,
1990                                  struct osc_enqueue_args *aa, int rc)
1991 {
1992         struct ldlm_lock *lock;
1993         struct lustre_handle *lockh = &aa->oa_lockh;
1994         enum ldlm_mode mode = aa->oa_mode;
1995         struct ost_lvb *lvb = aa->oa_lvb;
1996         __u32 lvb_len = sizeof(*lvb);
1997         __u64 flags = 0;
1998
1999         ENTRY;
2000
2001         /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2002          * be valid. */
2003         lock = ldlm_handle2lock(lockh);
2004         LASSERTF(lock != NULL,
2005                  "lockh %#llx, req %p, aa %p - client evicted?\n",
2006                  lockh->cookie, req, aa);
2007
2008         /* Take an additional reference so that a blocking AST that
2009          * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2010          * to arrive after an upcall has been executed by
2011          * osc_enqueue_fini(). */
2012         ldlm_lock_addref(lockh, mode);
2013
2014         /* Let cl_lock_state_wait fail with -ERESTARTSYS to unuse sublocks. */
2015         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_HANG, 2);
2016
2017         /* Let CP AST to grant the lock first. */
2018         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
2019
2020         if (aa->oa_agl) {
2021                 LASSERT(aa->oa_lvb == NULL);
2022                 LASSERT(aa->oa_flags == NULL);
2023                 aa->oa_flags = &flags;
2024         }
2025
2026         /* Complete obtaining the lock procedure. */
2027         rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_type, 1,
2028                                    aa->oa_mode, aa->oa_flags, lvb, lvb_len,
2029                                    lockh, rc);
2030         /* Complete osc stuff. */
2031         rc = osc_enqueue_fini(req, aa->oa_upcall, aa->oa_cookie, lockh, mode,
2032                               aa->oa_flags, aa->oa_agl, rc);
2033
2034         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
2035
2036         ldlm_lock_decref(lockh, mode);
2037         LDLM_LOCK_PUT(lock);
2038         RETURN(rc);
2039 }
2040
2041 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
2042
2043 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2044  * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2045  * other synchronous requests, however keeping some locks and trying to obtain
2046  * others may take a considerable amount of time in a case of ost failure; and
2047  * when other sync requests do not get released lock from a client, the client
2048  * is evicted from the cluster -- such scenarious make the life difficult, so
2049  * release locks just after they are obtained. */
2050 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2051                      __u64 *flags, union ldlm_policy_data *policy,
2052                      struct ost_lvb *lvb, int kms_valid,
2053                      osc_enqueue_upcall_f upcall, void *cookie,
2054                      struct ldlm_enqueue_info *einfo,
2055                      struct ptlrpc_request_set *rqset, int async, int agl)
2056 {
2057         struct obd_device *obd = exp->exp_obd;
2058         struct lustre_handle lockh = { 0 };
2059         struct ptlrpc_request *req = NULL;
2060         int intent = *flags & LDLM_FL_HAS_INTENT;
2061         __u64 match_flags = *flags;
2062         enum ldlm_mode mode;
2063         int rc;
2064         ENTRY;
2065
2066         /* Filesystem lock extents are extended to page boundaries so that
2067          * dealing with the page cache is a little smoother.  */
2068         policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2069         policy->l_extent.end |= ~PAGE_MASK;
2070
2071         /*
2072          * kms is not valid when either object is completely fresh (so that no
2073          * locks are cached), or object was evicted. In the latter case cached
2074          * lock cannot be used, because it would prime inode state with
2075          * potentially stale LVB.
2076          */
2077         if (!kms_valid)
2078                 goto no_match;
2079
2080         /* Next, search for already existing extent locks that will cover us */
2081         /* If we're trying to read, we also search for an existing PW lock.  The
2082          * VFS and page cache already protect us locally, so lots of readers/
2083          * writers can share a single PW lock.
2084          *
2085          * There are problems with conversion deadlocks, so instead of
2086          * converting a read lock to a write lock, we'll just enqueue a new
2087          * one.
2088          *
2089          * At some point we should cancel the read lock instead of making them
2090          * send us a blocking callback, but there are problems with canceling
2091          * locks out from other users right now, too. */
2092         mode = einfo->ei_mode;
2093         if (einfo->ei_mode == LCK_PR)
2094                 mode |= LCK_PW;
2095         if (agl == 0)
2096                 match_flags |= LDLM_FL_LVB_READY;
2097         if (intent != 0)
2098                 match_flags |= LDLM_FL_BLOCK_GRANTED;
2099         mode = ldlm_lock_match(obd->obd_namespace, match_flags, res_id,
2100                                einfo->ei_type, policy, mode, &lockh, 0);
2101         if (mode) {
2102                 struct ldlm_lock *matched;
2103
2104                 if (*flags & LDLM_FL_TEST_LOCK)
2105                         RETURN(ELDLM_OK);
2106
2107                 matched = ldlm_handle2lock(&lockh);
2108                 if (agl) {
2109                         /* AGL enqueues DLM locks speculatively. Therefore if
2110                          * it already exists a DLM lock, it wll just inform the
2111                          * caller to cancel the AGL process for this stripe. */
2112                         ldlm_lock_decref(&lockh, mode);
2113                         LDLM_LOCK_PUT(matched);
2114                         RETURN(-ECANCELED);
2115                 } else if (osc_set_lock_data(matched, einfo->ei_cbdata)) {
2116                         *flags |= LDLM_FL_LVB_READY;
2117
2118                         /* We already have a lock, and it's referenced. */
2119                         (*upcall)(cookie, &lockh, ELDLM_LOCK_MATCHED);
2120
2121                         ldlm_lock_decref(&lockh, mode);
2122                         LDLM_LOCK_PUT(matched);
2123                         RETURN(ELDLM_OK);
2124                 } else {
2125                         ldlm_lock_decref(&lockh, mode);
2126                         LDLM_LOCK_PUT(matched);
2127                 }
2128         }
2129
2130 no_match:
2131         if (*flags & (LDLM_FL_TEST_LOCK | LDLM_FL_MATCH_LOCK))
2132                 RETURN(-ENOLCK);
2133
2134         if (intent) {
2135                 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2136                                            &RQF_LDLM_ENQUEUE_LVB);
2137                 if (req == NULL)
2138                         RETURN(-ENOMEM);
2139
2140                 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
2141                 if (rc) {
2142                         ptlrpc_request_free(req);
2143                         RETURN(rc);
2144                 }
2145
2146                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2147                                      sizeof *lvb);
2148                 ptlrpc_request_set_replen(req);
2149         }
2150
2151         /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2152         *flags &= ~LDLM_FL_BLOCK_GRANTED;
2153
2154         rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2155                               sizeof(*lvb), LVB_T_OST, &lockh, async);
2156         if (async) {
2157                 if (!rc) {
2158                         struct osc_enqueue_args *aa;
2159                         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2160                         aa = ptlrpc_req_async_args(req);
2161                         aa->oa_exp    = exp;
2162                         aa->oa_mode   = einfo->ei_mode;
2163                         aa->oa_type   = einfo->ei_type;
2164                         lustre_handle_copy(&aa->oa_lockh, &lockh);
2165                         aa->oa_upcall = upcall;
2166                         aa->oa_cookie = cookie;
2167                         aa->oa_agl    = !!agl;
2168                         if (!agl) {
2169                                 aa->oa_flags  = flags;
2170                                 aa->oa_lvb    = lvb;
2171                         } else {
2172                                 /* AGL is essentially to enqueue an DLM lock
2173                                  * in advance, so we don't care about the
2174                                  * result of AGL enqueue. */
2175                                 aa->oa_lvb    = NULL;
2176                                 aa->oa_flags  = NULL;
2177                         }
2178
2179                         req->rq_interpret_reply =
2180                                 (ptlrpc_interpterer_t)osc_enqueue_interpret;
2181                         if (rqset == PTLRPCD_SET)
2182                                 ptlrpcd_add_req(req);
2183                         else
2184                                 ptlrpc_set_add_req(rqset, req);
2185                 } else if (intent) {
2186                         ptlrpc_req_finished(req);
2187                 }
2188                 RETURN(rc);
2189         }
2190
2191         rc = osc_enqueue_fini(req, upcall, cookie, &lockh, einfo->ei_mode,
2192                               flags, agl, rc);
2193         if (intent)
2194                 ptlrpc_req_finished(req);
2195
2196         RETURN(rc);
2197 }
2198
2199 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2200                    enum ldlm_type type, union ldlm_policy_data *policy,
2201                    enum ldlm_mode mode, __u64 *flags, void *data,
2202                    struct lustre_handle *lockh, int unref)
2203 {
2204         struct obd_device *obd = exp->exp_obd;
2205         __u64 lflags = *flags;
2206         enum ldlm_mode rc;
2207         ENTRY;
2208
2209         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2210                 RETURN(-EIO);
2211
2212         /* Filesystem lock extents are extended to page boundaries so that
2213          * dealing with the page cache is a little smoother */
2214         policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2215         policy->l_extent.end |= ~PAGE_MASK;
2216
2217         /* Next, search for already existing extent locks that will cover us */
2218         /* If we're trying to read, we also search for an existing PW lock.  The
2219          * VFS and page cache already protect us locally, so lots of readers/
2220          * writers can share a single PW lock. */
2221         rc = mode;
2222         if (mode == LCK_PR)
2223                 rc |= LCK_PW;
2224         rc = ldlm_lock_match(obd->obd_namespace, lflags,
2225                              res_id, type, policy, rc, lockh, unref);
2226         if (rc == 0 || lflags & LDLM_FL_TEST_LOCK)
2227                 RETURN(rc);
2228
2229         if (data != NULL) {
2230                 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2231
2232                 LASSERT(lock != NULL);
2233                 if (!osc_set_lock_data(lock, data)) {
2234                         ldlm_lock_decref(lockh, rc);
2235                         rc = 0;
2236                 }
2237                 LDLM_LOCK_PUT(lock);
2238         }
2239         RETURN(rc);
2240 }
2241
2242 static int osc_statfs_interpret(const struct lu_env *env,
2243                                 struct ptlrpc_request *req,
2244                                 struct osc_async_args *aa, int rc)
2245 {
2246         struct obd_statfs *msfs;
2247         ENTRY;
2248
2249         if (rc == -EBADR)
2250                 /* The request has in fact never been sent
2251                  * due to issues at a higher level (LOV).
2252                  * Exit immediately since the caller is
2253                  * aware of the problem and takes care
2254                  * of the clean up */
2255                  RETURN(rc);
2256
2257         if ((rc == -ENOTCONN || rc == -EAGAIN) &&
2258             (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
2259                 GOTO(out, rc = 0);
2260
2261         if (rc != 0)
2262                 GOTO(out, rc);
2263
2264         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2265         if (msfs == NULL) {
2266                 GOTO(out, rc = -EPROTO);
2267         }
2268
2269         *aa->aa_oi->oi_osfs = *msfs;
2270 out:
2271         rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
2272         RETURN(rc);
2273 }
2274
2275 static int osc_statfs_async(struct obd_export *exp,
2276                             struct obd_info *oinfo, __u64 max_age,
2277                             struct ptlrpc_request_set *rqset)
2278 {
2279         struct obd_device     *obd = class_exp2obd(exp);
2280         struct ptlrpc_request *req;
2281         struct osc_async_args *aa;
2282         int                    rc;
2283         ENTRY;
2284
2285         /* We could possibly pass max_age in the request (as an absolute
2286          * timestamp or a "seconds.usec ago") so the target can avoid doing
2287          * extra calls into the filesystem if that isn't necessary (e.g.
2288          * during mount that would help a bit).  Having relative timestamps
2289          * is not so great if request processing is slow, while absolute
2290          * timestamps are not ideal because they need time synchronization. */
2291         req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
2292         if (req == NULL)
2293                 RETURN(-ENOMEM);
2294
2295         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2296         if (rc) {
2297                 ptlrpc_request_free(req);
2298                 RETURN(rc);
2299         }
2300         ptlrpc_request_set_replen(req);
2301         req->rq_request_portal = OST_CREATE_PORTAL;
2302         ptlrpc_at_set_req_timeout(req);
2303
2304         if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
2305                 /* procfs requests not want stat in wait for avoid deadlock */
2306                 req->rq_no_resend = 1;
2307                 req->rq_no_delay = 1;
2308         }
2309
2310         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
2311         CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
2312         aa = ptlrpc_req_async_args(req);
2313         aa->aa_oi = oinfo;
2314
2315         ptlrpc_set_add_req(rqset, req);
2316         RETURN(0);
2317 }
2318
2319 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
2320                       struct obd_statfs *osfs, __u64 max_age, __u32 flags)
2321 {
2322         struct obd_device     *obd = class_exp2obd(exp);
2323         struct obd_statfs     *msfs;
2324         struct ptlrpc_request *req;
2325         struct obd_import     *imp = NULL;
2326         int rc;
2327         ENTRY;
2328
2329         /*Since the request might also come from lprocfs, so we need
2330          *sync this with client_disconnect_export Bug15684*/
2331         down_read(&obd->u.cli.cl_sem);
2332         if (obd->u.cli.cl_import)
2333                 imp = class_import_get(obd->u.cli.cl_import);
2334         up_read(&obd->u.cli.cl_sem);
2335         if (!imp)
2336                 RETURN(-ENODEV);
2337
2338         /* We could possibly pass max_age in the request (as an absolute
2339          * timestamp or a "seconds.usec ago") so the target can avoid doing
2340          * extra calls into the filesystem if that isn't necessary (e.g.
2341          * during mount that would help a bit).  Having relative timestamps
2342          * is not so great if request processing is slow, while absolute
2343          * timestamps are not ideal because they need time synchronization. */
2344         req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
2345
2346         class_import_put(imp);
2347
2348         if (req == NULL)
2349                 RETURN(-ENOMEM);
2350
2351         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2352         if (rc) {
2353                 ptlrpc_request_free(req);
2354                 RETURN(rc);
2355         }
2356         ptlrpc_request_set_replen(req);
2357         req->rq_request_portal = OST_CREATE_PORTAL;
2358         ptlrpc_at_set_req_timeout(req);
2359
2360         if (flags & OBD_STATFS_NODELAY) {
2361                 /* procfs requests not want stat in wait for avoid deadlock */
2362                 req->rq_no_resend = 1;
2363                 req->rq_no_delay = 1;
2364         }
2365
2366         rc = ptlrpc_queue_wait(req);
2367         if (rc)
2368                 GOTO(out, rc);
2369
2370         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2371         if (msfs == NULL) {
2372                 GOTO(out, rc = -EPROTO);
2373         }
2374
2375         *osfs = *msfs;
2376
2377         EXIT;
2378  out:
2379         ptlrpc_req_finished(req);
2380         return rc;
2381 }
2382
2383 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2384                          void *karg, void __user *uarg)
2385 {
2386         struct obd_device *obd = exp->exp_obd;
2387         struct obd_ioctl_data *data = karg;
2388         int err = 0;
2389         ENTRY;
2390
2391         if (!try_module_get(THIS_MODULE)) {
2392                 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2393                        module_name(THIS_MODULE));
2394                 return -EINVAL;
2395         }
2396         switch (cmd) {
2397         case OBD_IOC_CLIENT_RECOVER:
2398                 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2399                                             data->ioc_inlbuf1, 0);
2400                 if (err > 0)
2401                         err = 0;
2402                 GOTO(out, err);
2403         case IOC_OSC_SET_ACTIVE:
2404                 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2405                                                data->ioc_offset);
2406                 GOTO(out, err);
2407         case OBD_IOC_PING_TARGET:
2408                 err = ptlrpc_obd_ping(obd);
2409                 GOTO(out, err);
2410         default:
2411                 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
2412                        cmd, current_comm());
2413                 GOTO(out, err = -ENOTTY);
2414         }
2415 out:
2416         module_put(THIS_MODULE);
2417         return err;
2418 }
2419
2420 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
2421                               u32 keylen, void *key,
2422                               u32 vallen, void *val,
2423                               struct ptlrpc_request_set *set)
2424 {
2425         struct ptlrpc_request *req;
2426         struct obd_device     *obd = exp->exp_obd;
2427         struct obd_import     *imp = class_exp2cliimp(exp);
2428         char                  *tmp;
2429         int                    rc;
2430         ENTRY;
2431
2432         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
2433
2434         if (KEY_IS(KEY_CHECKSUM)) {
2435                 if (vallen != sizeof(int))
2436                         RETURN(-EINVAL);
2437                 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
2438                 RETURN(0);
2439         }
2440
2441         if (KEY_IS(KEY_SPTLRPC_CONF)) {
2442                 sptlrpc_conf_client_adapt(obd);
2443                 RETURN(0);
2444         }
2445
2446         if (KEY_IS(KEY_FLUSH_CTX)) {
2447                 sptlrpc_import_flush_my_ctx(imp);
2448                 RETURN(0);
2449         }
2450
2451         if (KEY_IS(KEY_CACHE_SET)) {
2452                 struct client_obd *cli = &obd->u.cli;
2453
2454                 LASSERT(cli->cl_cache == NULL); /* only once */
2455                 cli->cl_cache = (struct cl_client_cache *)val;
2456                 cl_cache_incref(cli->cl_cache);
2457                 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
2458
2459                 /* add this osc into entity list */
2460                 LASSERT(list_empty(&cli->cl_lru_osc));
2461                 spin_lock(&cli->cl_cache->ccc_lru_lock);
2462                 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
2463                 spin_unlock(&cli->cl_cache->ccc_lru_lock);
2464
2465                 RETURN(0);
2466         }
2467
2468         if (KEY_IS(KEY_CACHE_LRU_SHRINK)) {
2469                 struct client_obd *cli = &obd->u.cli;
2470                 long nr = atomic_long_read(&cli->cl_lru_in_list) >> 1;
2471                 long target = *(long *)val;
2472
2473                 nr = osc_lru_shrink(env, cli, min(nr, target), true);
2474                 *(long *)val -= nr;
2475                 RETURN(0);
2476         }
2477
2478         if (!set && !KEY_IS(KEY_GRANT_SHRINK))
2479                 RETURN(-EINVAL);
2480
2481         /* We pass all other commands directly to OST. Since nobody calls osc
2482            methods directly and everybody is supposed to go through LOV, we
2483            assume lov checked invalid values for us.
2484            The only recognised values so far are evict_by_nid and mds_conn.
2485            Even if something bad goes through, we'd get a -EINVAL from OST
2486            anyway. */
2487
2488         req = ptlrpc_request_alloc(imp, KEY_IS(KEY_GRANT_SHRINK) ?
2489                                                 &RQF_OST_SET_GRANT_INFO :
2490                                                 &RQF_OBD_SET_INFO);
2491         if (req == NULL)
2492                 RETURN(-ENOMEM);
2493
2494         req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2495                              RCL_CLIENT, keylen);
2496         if (!KEY_IS(KEY_GRANT_SHRINK))
2497                 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
2498                                      RCL_CLIENT, vallen);
2499         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
2500         if (rc) {
2501                 ptlrpc_request_free(req);
2502                 RETURN(rc);
2503         }
2504
2505         tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2506         memcpy(tmp, key, keylen);
2507         tmp = req_capsule_client_get(&req->rq_pill, KEY_IS(KEY_GRANT_SHRINK) ?
2508                                                         &RMF_OST_BODY :
2509                                                         &RMF_SETINFO_VAL);
2510         memcpy(tmp, val, vallen);
2511
2512         if (KEY_IS(KEY_GRANT_SHRINK)) {
2513                 struct osc_grant_args *aa;
2514                 struct obdo *oa;
2515
2516                 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2517                 aa = ptlrpc_req_async_args(req);
2518                 OBDO_ALLOC(oa);
2519                 if (!oa) {
2520                         ptlrpc_req_finished(req);
2521                         RETURN(-ENOMEM);
2522                 }
2523                 *oa = ((struct ost_body *)val)->oa;
2524                 aa->aa_oa = oa;
2525                 req->rq_interpret_reply = osc_shrink_grant_interpret;
2526         }
2527
2528         ptlrpc_request_set_replen(req);
2529         if (!KEY_IS(KEY_GRANT_SHRINK)) {
2530                 LASSERT(set != NULL);
2531                 ptlrpc_set_add_req(set, req);
2532                 ptlrpc_check_set(NULL, set);
2533         } else {
2534                 ptlrpcd_add_req(req);
2535         }
2536
2537         RETURN(0);
2538 }
2539
2540 static int osc_reconnect(const struct lu_env *env,
2541                          struct obd_export *exp, struct obd_device *obd,
2542                          struct obd_uuid *cluuid,
2543                          struct obd_connect_data *data,
2544                          void *localdata)
2545 {
2546         struct client_obd *cli = &obd->u.cli;
2547
2548         if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
2549                 long lost_grant;
2550                 long grant;
2551
2552                 spin_lock(&cli->cl_loi_list_lock);
2553                 grant = cli->cl_avail_grant + cli->cl_reserved_grant;
2554                 if (data->ocd_connect_flags & OBD_CONNECT_GRANT_PARAM)
2555                         grant += cli->cl_dirty_grant;
2556                 else
2557                         grant += cli->cl_dirty_pages << PAGE_SHIFT;
2558                 data->ocd_grant = grant ? : 2 * cli_brw_size(obd);
2559                 lost_grant = cli->cl_lost_grant;
2560                 cli->cl_lost_grant = 0;
2561                 spin_unlock(&cli->cl_loi_list_lock);
2562
2563                 CDEBUG(D_RPCTRACE, "ocd_connect_flags: %#llx ocd_version: %d"
2564                        " ocd_grant: %d, lost: %ld.\n", data->ocd_connect_flags,
2565                        data->ocd_version, data->ocd_grant, lost_grant);
2566         }
2567
2568         RETURN(0);
2569 }
2570
2571 static int osc_disconnect(struct obd_export *exp)
2572 {
2573         struct obd_device *obd = class_exp2obd(exp);
2574         int rc;
2575
2576         rc = client_disconnect_export(exp);
2577         /**
2578          * Initially we put del_shrink_grant before disconnect_export, but it
2579          * causes the following problem if setup (connect) and cleanup
2580          * (disconnect) are tangled together.
2581          *      connect p1                     disconnect p2
2582          *   ptlrpc_connect_import
2583          *     ...............               class_manual_cleanup
2584          *                                     osc_disconnect
2585          *                                     del_shrink_grant
2586          *   ptlrpc_connect_interrupt
2587          *     init_grant_shrink
2588          *   add this client to shrink list
2589          *                                      cleanup_osc
2590          * Bang! pinger trigger the shrink.
2591          * So the osc should be disconnected from the shrink list, after we
2592          * are sure the import has been destroyed. BUG18662
2593          */
2594         if (obd->u.cli.cl_import == NULL)
2595                 osc_del_shrink_grant(&obd->u.cli);
2596         return rc;
2597 }
2598
2599 static int osc_ldlm_resource_invalidate(struct cfs_hash *hs,
2600         struct cfs_hash_bd *bd, struct hlist_node *hnode, void *arg)
2601 {
2602         struct lu_env *env = arg;
2603         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2604         struct ldlm_lock *lock;
2605         struct osc_object *osc = NULL;
2606         ENTRY;
2607
2608         lock_res(res);
2609         list_for_each_entry(lock, &res->lr_granted, l_res_link) {
2610                 if (lock->l_ast_data != NULL && osc == NULL) {
2611                         osc = lock->l_ast_data;
2612                         cl_object_get(osc2cl(osc));
2613                 }
2614
2615                 /* clear LDLM_FL_CLEANED flag to make sure it will be canceled
2616                  * by the 2nd round of ldlm_namespace_clean() call in
2617                  * osc_import_event(). */
2618                 ldlm_clear_cleaned(lock);
2619         }
2620         unlock_res(res);
2621
2622         if (osc != NULL) {
2623                 osc_object_invalidate(env, osc);
2624                 cl_object_put(env, osc2cl(osc));
2625         }
2626
2627         RETURN(0);
2628 }
2629
2630 static int osc_import_event(struct obd_device *obd,
2631                             struct obd_import *imp,
2632                             enum obd_import_event event)
2633 {
2634         struct client_obd *cli;
2635         int rc = 0;
2636
2637         ENTRY;
2638         LASSERT(imp->imp_obd == obd);
2639
2640         switch (event) {
2641         case IMP_EVENT_DISCON: {
2642                 cli = &obd->u.cli;
2643                 spin_lock(&cli->cl_loi_list_lock);
2644                 cli->cl_avail_grant = 0;
2645                 cli->cl_lost_grant = 0;
2646                 spin_unlock(&cli->cl_loi_list_lock);
2647                 break;
2648         }
2649         case IMP_EVENT_INACTIVE: {
2650                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
2651                 break;
2652         }
2653         case IMP_EVENT_INVALIDATE: {
2654                 struct ldlm_namespace *ns = obd->obd_namespace;
2655                 struct lu_env         *env;
2656                 __u16                  refcheck;
2657
2658                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2659
2660                 env = cl_env_get(&refcheck);
2661                 if (!IS_ERR(env)) {
2662                         osc_io_unplug(env, &obd->u.cli, NULL);
2663
2664                         cfs_hash_for_each_nolock(ns->ns_rs_hash,
2665                                                  osc_ldlm_resource_invalidate,
2666                                                  env, 0);
2667                         cl_env_put(env, &refcheck);
2668
2669                         ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2670                 } else
2671                         rc = PTR_ERR(env);
2672                 break;
2673         }
2674         case IMP_EVENT_ACTIVE: {
2675                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
2676                 break;
2677         }
2678         case IMP_EVENT_OCD: {
2679                 struct obd_connect_data *ocd = &imp->imp_connect_data;
2680
2681                 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
2682                         osc_init_grant(&obd->u.cli, ocd);
2683
2684                 /* See bug 7198 */
2685                 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
2686                         imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
2687
2688                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
2689                 break;
2690         }
2691         case IMP_EVENT_DEACTIVATE: {
2692                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE, NULL);
2693                 break;
2694         }
2695         case IMP_EVENT_ACTIVATE: {
2696                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE, NULL);
2697                 break;
2698         }
2699         default:
2700                 CERROR("Unknown import event %d\n", event);
2701                 LBUG();
2702         }
2703         RETURN(rc);
2704 }
2705
2706 /**
2707  * Determine whether the lock can be canceled before replaying the lock
2708  * during recovery, see bug16774 for detailed information.
2709  *
2710  * \retval zero the lock can't be canceled
2711  * \retval other ok to cancel
2712  */
2713 static int osc_cancel_weight(struct ldlm_lock *lock)
2714 {
2715         /*
2716          * Cancel all unused and granted extent lock.
2717          */
2718         if (lock->l_resource->lr_type == LDLM_EXTENT &&
2719             lock->l_granted_mode == lock->l_req_mode &&
2720             osc_ldlm_weigh_ast(lock) == 0)
2721                 RETURN(1);
2722
2723         RETURN(0);
2724 }
2725
2726 static int brw_queue_work(const struct lu_env *env, void *data)
2727 {
2728         struct client_obd *cli = data;
2729
2730         CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
2731
2732         osc_io_unplug(env, cli, NULL);
2733         RETURN(0);
2734 }
2735
2736 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
2737 {
2738         struct client_obd *cli = &obd->u.cli;
2739         struct obd_type   *type;
2740         void              *handler;
2741         int                rc;
2742         int                adding;
2743         int                added;
2744         int                req_count;
2745         ENTRY;
2746
2747         rc = ptlrpcd_addref();
2748         if (rc)
2749                 RETURN(rc);
2750
2751         rc = client_obd_setup(obd, lcfg);
2752         if (rc)
2753                 GOTO(out_ptlrpcd, rc);
2754
2755         handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
2756         if (IS_ERR(handler))
2757                 GOTO(out_client_setup, rc = PTR_ERR(handler));
2758         cli->cl_writeback_work = handler;
2759
2760         handler = ptlrpcd_alloc_work(cli->cl_import, lru_queue_work, cli);
2761         if (IS_ERR(handler))
2762                 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
2763         cli->cl_lru_work = handler;
2764
2765         rc = osc_quota_setup(obd);
2766         if (rc)
2767                 GOTO(out_ptlrpcd_work, rc);
2768
2769         cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
2770
2771 #ifdef CONFIG_PROC_FS
2772         obd->obd_vars = lprocfs_osc_obd_vars;
2773 #endif
2774         /* If this is true then both client (osc) and server (osp) are on the
2775          * same node. The osp layer if loaded first will register the osc proc
2776          * directory. In that case this obd_device will be attached its proc
2777          * tree to type->typ_procsym instead of obd->obd_type->typ_procroot. */
2778         type = class_search_type(LUSTRE_OSP_NAME);
2779         if (type && type->typ_procsym) {
2780                 obd->obd_proc_entry = lprocfs_register(obd->obd_name,
2781                                                        type->typ_procsym,
2782                                                        obd->obd_vars, obd);
2783                 if (IS_ERR(obd->obd_proc_entry)) {
2784                         rc = PTR_ERR(obd->obd_proc_entry);
2785                         CERROR("error %d setting up lprocfs for %s\n", rc,
2786                                obd->obd_name);
2787                         obd->obd_proc_entry = NULL;
2788                 }
2789         } else {
2790                 rc = lprocfs_obd_setup(obd);
2791         }
2792
2793         /* If the basic OSC proc tree construction succeeded then
2794          * lets do the rest. */
2795         if (rc == 0) {
2796                 lproc_osc_attach_seqstat(obd);
2797                 sptlrpc_lprocfs_cliobd_attach(obd);
2798                 ptlrpc_lprocfs_register_obd(obd);
2799         }
2800
2801         /*
2802          * We try to control the total number of requests with a upper limit
2803          * osc_reqpool_maxreqcount. There might be some race which will cause
2804          * over-limit allocation, but it is fine.
2805          */
2806         req_count = atomic_read(&osc_pool_req_count);
2807         if (req_count < osc_reqpool_maxreqcount) {
2808                 adding = cli->cl_max_rpcs_in_flight + 2;
2809                 if (req_count + adding > osc_reqpool_maxreqcount)
2810                         adding = osc_reqpool_maxreqcount - req_count;
2811
2812                 added = ptlrpc_add_rqs_to_pool(osc_rq_pool, adding);
2813                 atomic_add(added, &osc_pool_req_count);
2814         }
2815
2816         INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
2817         ns_register_cancel(obd->obd_namespace, osc_cancel_weight);
2818
2819         spin_lock(&osc_shrink_lock);
2820         list_add_tail(&cli->cl_shrink_list, &osc_shrink_list);
2821         spin_unlock(&osc_shrink_lock);
2822
2823         RETURN(0);
2824
2825 out_ptlrpcd_work:
2826         if (cli->cl_writeback_work != NULL) {
2827                 ptlrpcd_destroy_work(cli->cl_writeback_work);
2828                 cli->cl_writeback_work = NULL;
2829         }
2830         if (cli->cl_lru_work != NULL) {
2831                 ptlrpcd_destroy_work(cli->cl_lru_work);
2832                 cli->cl_lru_work = NULL;
2833         }
2834 out_client_setup:
2835         client_obd_cleanup(obd);
2836 out_ptlrpcd:
2837         ptlrpcd_decref();
2838         RETURN(rc);
2839 }
2840
2841 static int osc_precleanup(struct obd_device *obd)
2842 {
2843         struct client_obd *cli = &obd->u.cli;
2844         ENTRY;
2845
2846         /* LU-464
2847          * for echo client, export may be on zombie list, wait for
2848          * zombie thread to cull it, because cli.cl_import will be
2849          * cleared in client_disconnect_export():
2850          *   class_export_destroy() -> obd_cleanup() ->
2851          *   echo_device_free() -> echo_client_cleanup() ->
2852          *   obd_disconnect() -> osc_disconnect() ->
2853          *   client_disconnect_export()
2854          */
2855         obd_zombie_barrier();
2856         if (cli->cl_writeback_work) {
2857                 ptlrpcd_destroy_work(cli->cl_writeback_work);
2858                 cli->cl_writeback_work = NULL;
2859         }
2860
2861         if (cli->cl_lru_work) {
2862                 ptlrpcd_destroy_work(cli->cl_lru_work);
2863                 cli->cl_lru_work = NULL;
2864         }
2865
2866         obd_cleanup_client_import(obd);
2867         ptlrpc_lprocfs_unregister_obd(obd);
2868         lprocfs_obd_cleanup(obd);
2869         RETURN(0);
2870 }
2871
2872 int osc_cleanup(struct obd_device *obd)
2873 {
2874         struct client_obd *cli = &obd->u.cli;
2875         int rc;
2876
2877         ENTRY;
2878
2879         spin_lock(&osc_shrink_lock);
2880         list_del(&cli->cl_shrink_list);
2881         spin_unlock(&osc_shrink_lock);
2882
2883         /* lru cleanup */
2884         if (cli->cl_cache != NULL) {
2885                 LASSERT(atomic_read(&cli->cl_cache->ccc_users) > 0);
2886                 spin_lock(&cli->cl_cache->ccc_lru_lock);
2887                 list_del_init(&cli->cl_lru_osc);
2888                 spin_unlock(&cli->cl_cache->ccc_lru_lock);
2889                 cli->cl_lru_left = NULL;
2890                 cl_cache_decref(cli->cl_cache);
2891                 cli->cl_cache = NULL;
2892         }
2893
2894         /* free memory of osc quota cache */
2895         osc_quota_cleanup(obd);
2896
2897         rc = client_obd_cleanup(obd);
2898
2899         ptlrpcd_decref();
2900         RETURN(rc);
2901 }
2902
2903 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
2904 {
2905         int rc = class_process_proc_param(PARAM_OSC, obd->obd_vars, lcfg, obd);
2906         return rc > 0 ? 0: rc;
2907 }
2908
2909 static int osc_process_config(struct obd_device *obd, size_t len, void *buf)
2910 {
2911         return osc_process_config_base(obd, buf);
2912 }
2913
2914 static struct obd_ops osc_obd_ops = {
2915         .o_owner                = THIS_MODULE,
2916         .o_setup                = osc_setup,
2917         .o_precleanup           = osc_precleanup,
2918         .o_cleanup              = osc_cleanup,
2919         .o_add_conn             = client_import_add_conn,
2920         .o_del_conn             = client_import_del_conn,
2921         .o_connect              = client_connect_import,
2922         .o_reconnect            = osc_reconnect,
2923         .o_disconnect           = osc_disconnect,
2924         .o_statfs               = osc_statfs,
2925         .o_statfs_async         = osc_statfs_async,
2926         .o_create               = osc_create,
2927         .o_destroy              = osc_destroy,
2928         .o_getattr              = osc_getattr,
2929         .o_setattr              = osc_setattr,
2930         .o_iocontrol            = osc_iocontrol,
2931         .o_set_info_async       = osc_set_info_async,
2932         .o_import_event         = osc_import_event,
2933         .o_process_config       = osc_process_config,
2934         .o_quotactl             = osc_quotactl,
2935 };
2936
2937 static struct shrinker *osc_cache_shrinker;
2938 struct list_head osc_shrink_list = LIST_HEAD_INIT(osc_shrink_list);
2939 DEFINE_SPINLOCK(osc_shrink_lock);
2940
2941 #ifndef HAVE_SHRINKER_COUNT
2942 static int osc_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
2943 {
2944         struct shrink_control scv = {
2945                 .nr_to_scan = shrink_param(sc, nr_to_scan),
2946                 .gfp_mask   = shrink_param(sc, gfp_mask)
2947         };
2948 #if !defined(HAVE_SHRINKER_WANT_SHRINK_PTR) && !defined(HAVE_SHRINK_CONTROL)
2949         struct shrinker *shrinker = NULL;
2950 #endif
2951
2952         (void)osc_cache_shrink_scan(shrinker, &scv);
2953
2954         return osc_cache_shrink_count(shrinker, &scv);
2955 }
2956 #endif
2957
2958 static int __init osc_init(void)
2959 {
2960         bool enable_proc = true;
2961         struct obd_type *type;
2962         unsigned int reqpool_size;
2963         unsigned int reqsize;
2964         int rc;
2965         DEF_SHRINKER_VAR(osc_shvar, osc_cache_shrink,
2966                          osc_cache_shrink_count, osc_cache_shrink_scan);
2967         ENTRY;
2968
2969         /* print an address of _any_ initialized kernel symbol from this
2970          * module, to allow debugging with gdb that doesn't support data
2971          * symbols from modules.*/
2972         CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
2973
2974         rc = lu_kmem_init(osc_caches);
2975         if (rc)
2976                 RETURN(rc);
2977
2978         type = class_search_type(LUSTRE_OSP_NAME);
2979         if (type != NULL && type->typ_procsym != NULL)
2980                 enable_proc = false;
2981
2982         rc = class_register_type(&osc_obd_ops, NULL, enable_proc, NULL,
2983                                  LUSTRE_OSC_NAME, &osc_device_type);
2984         if (rc)
2985                 GOTO(out_kmem, rc);
2986
2987         osc_cache_shrinker = set_shrinker(DEFAULT_SEEKS, &osc_shvar);
2988
2989         /* This is obviously too much memory, only prevent overflow here */
2990         if (osc_reqpool_mem_max >= 1 << 12 || osc_reqpool_mem_max == 0)
2991                 GOTO(out_type, rc = -EINVAL);
2992
2993         reqpool_size = osc_reqpool_mem_max << 20;
2994
2995         reqsize = 1;
2996         while (reqsize < OST_IO_MAXREQSIZE)
2997                 reqsize = reqsize << 1;
2998
2999         /*
3000          * We don't enlarge the request count in OSC pool according to
3001          * cl_max_rpcs_in_flight. The allocation from the pool will only be
3002          * tried after normal allocation failed. So a small OSC pool won't
3003          * cause much performance degression in most of cases.
3004          */
3005         osc_reqpool_maxreqcount = reqpool_size / reqsize;
3006
3007         atomic_set(&osc_pool_req_count, 0);
3008         osc_rq_pool = ptlrpc_init_rq_pool(0, OST_IO_MAXREQSIZE,
3009                                           ptlrpc_add_rqs_to_pool);
3010
3011         if (osc_rq_pool != NULL)
3012                 GOTO(out, rc);
3013         rc = -ENOMEM;
3014 out_type:
3015         class_unregister_type(LUSTRE_OSC_NAME);
3016 out_kmem:
3017         lu_kmem_fini(osc_caches);
3018 out:
3019         RETURN(rc);
3020 }
3021
3022 static void __exit osc_exit(void)
3023 {
3024         remove_shrinker(osc_cache_shrinker);
3025         class_unregister_type(LUSTRE_OSC_NAME);
3026         lu_kmem_fini(osc_caches);
3027         ptlrpc_free_rq_pool(osc_rq_pool);
3028 }
3029
3030 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3031 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3032 MODULE_VERSION(LUSTRE_VERSION_STRING);
3033 MODULE_LICENSE("GPL");
3034
3035 module_init(osc_init);
3036 module_exit(osc_exit);