drivers/staging/lustre/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, 2015, 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 <linux/libcfs/libcfs.h>
  36
  37 #include <lustre_dlm.h>
  38 #include <lustre_net.h>
  39 #include <uapi/linux/lustre/lustre_idl.h>
  40 #include <obd_cksum.h>
  41
  42 #include <lustre_ha.h>
  43 #include <lprocfs_status.h>
  44 #include <uapi/linux/lustre/lustre_ioctl.h>
  45 #include <lustre_debug.h>
  46 #include <lustre_obdo.h>
  47 #include <uapi/linux/lustre/lustre_param.h>
  48 #include <lustre_fid.h>
  49 #include <obd_class.h>
  50 #include <obd.h>
  51 #include "osc_internal.h"
  52 #include "osc_cl_internal.h"
  53
  54 atomic_t osc_pool_req_count;
  55 unsigned int osc_reqpool_maxreqcount;
  56 struct ptlrpc_request_pool *osc_rq_pool;
  57
  58 /* max memory used for request pool, unit is MB */
  59 static unsigned int osc_reqpool_mem_max = 5;
  60 module_param(osc_reqpool_mem_max, uint, 0444);
  61
  62 struct osc_brw_async_args {
  63         struct obdo       *aa_oa;
  64         int             aa_requested_nob;
  65         int             aa_nio_count;
  66         u32             aa_page_count;
  67         int             aa_resends;
  68         struct brw_page  **aa_ppga;
  69         struct client_obd *aa_cli;
  70         struct list_head         aa_oaps;
  71         struct list_head         aa_exts;
  72 };
  73
  74 struct osc_async_args {
  75         struct obd_info   *aa_oi;
  76 };
  77
  78 struct osc_setattr_args {
  79         struct obdo      *sa_oa;
  80         obd_enqueue_update_f sa_upcall;
  81         void            *sa_cookie;
  82 };
  83
  84 struct osc_fsync_args {
  85         struct osc_object       *fa_obj;
  86         struct obdo             *fa_oa;
  87         obd_enqueue_update_f fa_upcall;
  88         void            *fa_cookie;
  89 };
  90
  91 struct osc_enqueue_args {
  92         struct obd_export       *oa_exp;
  93         enum ldlm_type          oa_type;
  94         enum ldlm_mode          oa_mode;
  95         __u64               *oa_flags;
  96         osc_enqueue_upcall_f    oa_upcall;
  97         void                 *oa_cookie;
  98         struct ost_lvb     *oa_lvb;
  99         struct lustre_handle    oa_lockh;
 100         unsigned int          oa_agl:1;
 101 };
 102
 103 static void osc_release_ppga(struct brw_page **ppga, u32 count);
 104 static int brw_interpret(const struct lu_env *env,
 105                          struct ptlrpc_request *req, void *data, int rc);
 106
 107 static inline void osc_pack_req_body(struct ptlrpc_request *req,
 108                                      struct obdo *oa)
 109 {
 110         struct ost_body *body;
 111
 112         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 113         LASSERT(body);
 114
 115         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
 116 }
 117
 118 static int osc_getattr(const struct lu_env *env, struct obd_export *exp,
 119                        struct obdo *oa)
 120 {
 121         struct ptlrpc_request *req;
 122         struct ost_body *body;
 123         int rc;
 124
 125         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
 126         if (!req)
 127                 return -ENOMEM;
 128
 129         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
 130         if (rc) {
 131                 ptlrpc_request_free(req);
 132                 return rc;
 133         }
 134
 135         osc_pack_req_body(req, oa);
 136
 137         ptlrpc_request_set_replen(req);
 138
 139         rc = ptlrpc_queue_wait(req);
 140         if (rc)
 141                 goto out;
 142
 143         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 144         if (!body) {
 145                 rc = -EPROTO;
 146                 goto out;
 147         }
 148
 149         CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
 150         lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa,
 151                              &body->oa);
 152
 153         oa->o_blksize = cli_brw_size(exp->exp_obd);
 154         oa->o_valid |= OBD_MD_FLBLKSZ;
 155
 156  out:
 157         ptlrpc_req_finished(req);
 158         return rc;
 159 }
 160
 161 static int osc_setattr(const struct lu_env *env, struct obd_export *exp,
 162                        struct obdo *oa)
 163 {
 164         struct ptlrpc_request *req;
 165         struct ost_body *body;
 166         int rc;
 167
 168         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
 169
 170         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
 171         if (!req)
 172                 return -ENOMEM;
 173
 174         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
 175         if (rc) {
 176                 ptlrpc_request_free(req);
 177                 return rc;
 178         }
 179
 180         osc_pack_req_body(req, oa);
 181
 182         ptlrpc_request_set_replen(req);
 183
 184         rc = ptlrpc_queue_wait(req);
 185         if (rc)
 186                 goto out;
 187
 188         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 189         if (!body) {
 190                 rc = -EPROTO;
 191                 goto out;
 192         }
 193
 194         lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa,
 195                              &body->oa);
 196
 197 out:
 198         ptlrpc_req_finished(req);
 199         return rc;
 200 }
 201
 202 static int osc_setattr_interpret(const struct lu_env *env,
 203                                  struct ptlrpc_request *req,
 204                                  struct osc_setattr_args *sa, int rc)
 205 {
 206         struct ost_body *body;
 207
 208         if (rc != 0)
 209                 goto out;
 210
 211         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 212         if (!body) {
 213                 rc = -EPROTO;
 214                 goto out;
 215         }
 216
 217         lustre_get_wire_obdo(&req->rq_import->imp_connect_data, sa->sa_oa,
 218                              &body->oa);
 219 out:
 220         rc = sa->sa_upcall(sa->sa_cookie, rc);
 221         return rc;
 222 }
 223
 224 int osc_setattr_async(struct obd_export *exp, struct obdo *oa,
 225                       obd_enqueue_update_f upcall, void *cookie,
 226                       struct ptlrpc_request_set *rqset)
 227 {
 228         struct ptlrpc_request *req;
 229         struct osc_setattr_args *sa;
 230         int rc;
 231
 232         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
 233         if (!req)
 234                 return -ENOMEM;
 235
 236         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
 237         if (rc) {
 238                 ptlrpc_request_free(req);
 239                 return rc;
 240         }
 241
 242         osc_pack_req_body(req, oa);
 243
 244         ptlrpc_request_set_replen(req);
 245
 246         /* do mds to ost setattr asynchronously */
 247         if (!rqset) {
 248                 /* Do not wait for response. */
 249                 ptlrpcd_add_req(req);
 250         } else {
 251                 req->rq_interpret_reply =
 252                         (ptlrpc_interpterer_t)osc_setattr_interpret;
 253
 254                 BUILD_BUG_ON(sizeof(*sa) > sizeof(req->rq_async_args));
 255                 sa = ptlrpc_req_async_args(req);
 256                 sa->sa_oa = oa;
 257                 sa->sa_upcall = upcall;
 258                 sa->sa_cookie = cookie;
 259
 260                 if (rqset == PTLRPCD_SET)
 261                         ptlrpcd_add_req(req);
 262                 else
 263                         ptlrpc_set_add_req(rqset, req);
 264         }
 265
 266         return 0;
 267 }
 268
 269 static int osc_create(const struct lu_env *env, struct obd_export *exp,
 270                       struct obdo *oa)
 271 {
 272         struct ptlrpc_request *req;
 273         struct ost_body *body;
 274         int rc;
 275
 276         LASSERT(oa);
 277         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
 278         LASSERT(fid_seq_is_echo(ostid_seq(&oa->o_oi)));
 279
 280         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
 281         if (!req) {
 282                 rc = -ENOMEM;
 283                 goto out;
 284         }
 285
 286         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
 287         if (rc) {
 288                 ptlrpc_request_free(req);
 289                 goto out;
 290         }
 291
 292         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 293         LASSERT(body);
 294
 295         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
 296
 297         ptlrpc_request_set_replen(req);
 298
 299         rc = ptlrpc_queue_wait(req);
 300         if (rc)
 301                 goto out_req;
 302
 303         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 304         if (!body) {
 305                 rc = -EPROTO;
 306                 goto out_req;
 307         }
 308
 309         CDEBUG(D_INFO, "oa flags %x\n", oa->o_flags);
 310         lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
 311
 312         oa->o_blksize = cli_brw_size(exp->exp_obd);
 313         oa->o_valid |= OBD_MD_FLBLKSZ;
 314
 315         CDEBUG(D_HA, "transno: %lld\n",
 316                lustre_msg_get_transno(req->rq_repmsg));
 317 out_req:
 318         ptlrpc_req_finished(req);
 319 out:
 320         return rc;
 321 }
 322
 323 int osc_punch_base(struct obd_export *exp, struct obdo *oa,
 324                    obd_enqueue_update_f upcall, void *cookie,
 325                    struct ptlrpc_request_set *rqset)
 326 {
 327         struct ptlrpc_request *req;
 328         struct osc_setattr_args *sa;
 329         struct ost_body *body;
 330         int rc;
 331
 332         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
 333         if (!req)
 334                 return -ENOMEM;
 335
 336         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
 337         if (rc) {
 338                 ptlrpc_request_free(req);
 339                 return rc;
 340         }
 341         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
 342         ptlrpc_at_set_req_timeout(req);
 343
 344         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 345         LASSERT(body);
 346         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
 347                              oa);
 348
 349         ptlrpc_request_set_replen(req);
 350
 351         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_setattr_interpret;
 352         BUILD_BUG_ON(sizeof(*sa) > sizeof(req->rq_async_args));
 353         sa = ptlrpc_req_async_args(req);
 354         sa->sa_oa = oa;
 355         sa->sa_upcall = upcall;
 356         sa->sa_cookie = cookie;
 357         if (rqset == PTLRPCD_SET)
 358                 ptlrpcd_add_req(req);
 359         else
 360                 ptlrpc_set_add_req(rqset, req);
 361
 362         return 0;
 363 }
 364
 365 static int osc_sync_interpret(const struct lu_env *env,
 366                               struct ptlrpc_request *req,
 367                               void *arg, int rc)
 368 {
 369         struct cl_attr *attr = &osc_env_info(env)->oti_attr;
 370         struct osc_fsync_args *fa = arg;
 371         unsigned long valid = 0;
 372         struct ost_body *body;
 373         struct cl_object *obj;
 374
 375         if (rc)
 376                 goto out;
 377
 378         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 379         if (!body) {
 380                 CERROR("can't unpack ost_body\n");
 381                 rc = -EPROTO;
 382                 goto out;
 383         }
 384
 385         *fa->fa_oa = body->oa;
 386         obj = osc2cl(fa->fa_obj);
 387
 388         /* Update osc object's blocks attribute */
 389         cl_object_attr_lock(obj);
 390         if (body->oa.o_valid & OBD_MD_FLBLOCKS) {
 391                 attr->cat_blocks = body->oa.o_blocks;
 392                 valid |= CAT_BLOCKS;
 393         }
 394
 395         if (valid)
 396                 cl_object_attr_update(env, obj, attr, valid);
 397         cl_object_attr_unlock(obj);
 398
 399 out:
 400         rc = fa->fa_upcall(fa->fa_cookie, rc);
 401         return rc;
 402 }
 403
 404 int osc_sync_base(struct osc_object *obj, struct obdo *oa,
 405                   obd_enqueue_update_f upcall, void *cookie,
 406                   struct ptlrpc_request_set *rqset)
 407 {
 408         struct obd_export *exp = osc_export(obj);
 409         struct ptlrpc_request *req;
 410         struct ost_body *body;
 411         struct osc_fsync_args *fa;
 412         int rc;
 413
 414         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
 415         if (!req)
 416                 return -ENOMEM;
 417
 418         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
 419         if (rc) {
 420                 ptlrpc_request_free(req);
 421                 return rc;
 422         }
 423
 424         /* overload the size and blocks fields in the oa with start/end */
 425         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 426         LASSERT(body);
 427         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
 428                              oa);
 429
 430         ptlrpc_request_set_replen(req);
 431         req->rq_interpret_reply = osc_sync_interpret;
 432
 433         BUILD_BUG_ON(sizeof(*fa) > sizeof(req->rq_async_args));
 434         fa = ptlrpc_req_async_args(req);
 435         fa->fa_obj = obj;
 436         fa->fa_oa = oa;
 437         fa->fa_upcall = upcall;
 438         fa->fa_cookie = cookie;
 439
 440         if (rqset == PTLRPCD_SET)
 441                 ptlrpcd_add_req(req);
 442         else
 443                 ptlrpc_set_add_req(rqset, req);
 444
 445         return 0;
 446 }
 447
 448 /* Find and cancel locally locks matched by @mode in the resource found by
 449  * @objid. Found locks are added into @cancel list. Returns the amount of
 450  * locks added to @cancels list.
 451  */
 452 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
 453                                    struct list_head *cancels,
 454                                    enum ldlm_mode mode, __u64 lock_flags)
 455 {
 456         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 457         struct ldlm_res_id res_id;
 458         struct ldlm_resource *res;
 459         int count;
 460
 461         /* Return, i.e. cancel nothing, only if ELC is supported (flag in
 462          * export) but disabled through procfs (flag in NS).
 463          *
 464          * This distinguishes from a case when ELC is not supported originally,
 465          * when we still want to cancel locks in advance and just cancel them
 466          * locally, without sending any RPC.
 467          */
 468         if (exp_connect_cancelset(exp) && !ns_connect_cancelset(ns))
 469                 return 0;
 470
 471         ostid_build_res_name(&oa->o_oi, &res_id);
 472         res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
 473         if (IS_ERR(res))
 474                 return 0;
 475
 476         LDLM_RESOURCE_ADDREF(res);
 477         count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
 478                                            lock_flags, 0, NULL);
 479         LDLM_RESOURCE_DELREF(res);
 480         ldlm_resource_putref(res);
 481         return count;
 482 }
 483
 484 static int osc_destroy_interpret(const struct lu_env *env,
 485                                  struct ptlrpc_request *req, void *data,
 486                                  int rc)
 487 {
 488         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
 489
 490         atomic_dec(&cli->cl_destroy_in_flight);
 491         wake_up(&cli->cl_destroy_waitq);
 492         return 0;
 493 }
 494
 495 static int osc_can_send_destroy(struct client_obd *cli)
 496 {
 497         if (atomic_inc_return(&cli->cl_destroy_in_flight) <=
 498             cli->cl_max_rpcs_in_flight) {
 499                 /* The destroy request can be sent */
 500                 return 1;
 501         }
 502         if (atomic_dec_return(&cli->cl_destroy_in_flight) <
 503             cli->cl_max_rpcs_in_flight) {
 504                 /*
 505                  * The counter has been modified between the two atomic
 506                  * operations.
 507                  */
 508                 wake_up(&cli->cl_destroy_waitq);
 509         }
 510         return 0;
 511 }
 512
 513 static int osc_destroy(const struct lu_env *env, struct obd_export *exp,
 514                        struct obdo *oa)
 515 {
 516         struct client_obd *cli = &exp->exp_obd->u.cli;
 517         struct ptlrpc_request *req;
 518         struct ost_body *body;
 519         LIST_HEAD(cancels);
 520         int rc, count;
 521
 522         if (!oa) {
 523                 CDEBUG(D_INFO, "oa NULL\n");
 524                 return -EINVAL;
 525         }
 526
 527         count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
 528                                         LDLM_FL_DISCARD_DATA);
 529
 530         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
 531         if (!req) {
 532                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
 533                 return -ENOMEM;
 534         }
 535
 536         rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
 537                                0, &cancels, count);
 538         if (rc) {
 539                 ptlrpc_request_free(req);
 540                 return rc;
 541         }
 542
 543         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
 544         ptlrpc_at_set_req_timeout(req);
 545
 546         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 547         LASSERT(body);
 548         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
 549
 550         ptlrpc_request_set_replen(req);
 551
 552         req->rq_interpret_reply = osc_destroy_interpret;
 553         if (!osc_can_send_destroy(cli)) {
 554                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 555
 556                 /*
 557                  * Wait until the number of on-going destroy RPCs drops
 558                  * under max_rpc_in_flight
 559                  */
 560                 l_wait_event_exclusive(cli->cl_destroy_waitq,
 561                                        osc_can_send_destroy(cli), &lwi);
 562         }
 563
 564         /* Do not wait for response */
 565         ptlrpcd_add_req(req);
 566         return 0;
 567 }
 568
 569 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
 570                                 long writing_bytes)
 571 {
 572         u32 bits = OBD_MD_FLBLOCKS | OBD_MD_FLGRANT;
 573
 574         LASSERT(!(oa->o_valid & bits));
 575
 576         oa->o_valid |= bits;
 577         spin_lock(&cli->cl_loi_list_lock);
 578         oa->o_dirty = cli->cl_dirty_pages << PAGE_SHIFT;
 579         if (unlikely(cli->cl_dirty_pages - cli->cl_dirty_transit >
 580                      cli->cl_dirty_max_pages)) {
 581                 CERROR("dirty %lu - %lu > dirty_max %lu\n",
 582                        cli->cl_dirty_pages, cli->cl_dirty_transit,
 583                        cli->cl_dirty_max_pages);
 584                 oa->o_undirty = 0;
 585         } else if (unlikely(atomic_long_read(&obd_dirty_pages) -
 586                             atomic_long_read(&obd_dirty_transit_pages) >
 587                             (long)(obd_max_dirty_pages + 1))) {
 588                 /* The atomic_read() allowing the atomic_inc() are
 589                  * not covered by a lock thus they may safely race and trip
 590                  * this CERROR() unless we add in a small fudge factor (+1).
 591                  */
 592                 CERROR("%s: dirty %ld + %ld > system dirty_max %ld\n",
 593                        cli_name(cli), atomic_long_read(&obd_dirty_pages),
 594                        atomic_long_read(&obd_dirty_transit_pages),
 595                        obd_max_dirty_pages);
 596                 oa->o_undirty = 0;
 597         } else if (unlikely(cli->cl_dirty_max_pages - cli->cl_dirty_pages >
 598                    0x7fffffff)) {
 599                 CERROR("dirty %lu - dirty_max %lu too big???\n",
 600                        cli->cl_dirty_pages, cli->cl_dirty_max_pages);
 601                 oa->o_undirty = 0;
 602         } else {
 603                 unsigned long max_in_flight;
 604
 605                 max_in_flight = (cli->cl_max_pages_per_rpc << PAGE_SHIFT) *
 606                                 (cli->cl_max_rpcs_in_flight + 1);
 607                 oa->o_undirty = max(cli->cl_dirty_max_pages << PAGE_SHIFT,
 608                                     max_in_flight);
 609         }
 610         oa->o_grant = cli->cl_avail_grant + cli->cl_reserved_grant;
 611         oa->o_dropped = cli->cl_lost_grant;
 612         cli->cl_lost_grant = 0;
 613         spin_unlock(&cli->cl_loi_list_lock);
 614         CDEBUG(D_CACHE, "dirty: %llu undirty: %u dropped %u grant: %llu\n",
 615                oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
 616 }
 617
 618 void osc_update_next_shrink(struct client_obd *cli)
 619 {
 620         cli->cl_next_shrink_grant =
 621                 cfs_time_shift(cli->cl_grant_shrink_interval);
 622         CDEBUG(D_CACHE, "next time %ld to shrink grant\n",
 623                cli->cl_next_shrink_grant);
 624 }
 625
 626 static void __osc_update_grant(struct client_obd *cli, u64 grant)
 627 {
 628         spin_lock(&cli->cl_loi_list_lock);
 629         cli->cl_avail_grant += grant;
 630         spin_unlock(&cli->cl_loi_list_lock);
 631 }
 632
 633 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
 634 {
 635         if (body->oa.o_valid & OBD_MD_FLGRANT) {
 636                 CDEBUG(D_CACHE, "got %llu extra grant\n", body->oa.o_grant);
 637                 __osc_update_grant(cli, body->oa.o_grant);
 638         }
 639 }
 640
 641 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
 642                               u32 keylen, void *key, u32 vallen,
 643                               void *val, struct ptlrpc_request_set *set);
 644
 645 static int osc_shrink_grant_interpret(const struct lu_env *env,
 646                                       struct ptlrpc_request *req,
 647                                       void *aa, int rc)
 648 {
 649         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
 650         struct obdo *oa = ((struct osc_brw_async_args *)aa)->aa_oa;
 651         struct ost_body *body;
 652
 653         if (rc != 0) {
 654                 __osc_update_grant(cli, oa->o_grant);
 655                 goto out;
 656         }
 657
 658         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 659         LASSERT(body);
 660         osc_update_grant(cli, body);
 661 out:
 662         kmem_cache_free(obdo_cachep, oa);
 663         return rc;
 664 }
 665
 666 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
 667 {
 668         spin_lock(&cli->cl_loi_list_lock);
 669         oa->o_grant = cli->cl_avail_grant / 4;
 670         cli->cl_avail_grant -= oa->o_grant;
 671         spin_unlock(&cli->cl_loi_list_lock);
 672         if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
 673                 oa->o_valid |= OBD_MD_FLFLAGS;
 674                 oa->o_flags = 0;
 675         }
 676         oa->o_flags |= OBD_FL_SHRINK_GRANT;
 677         osc_update_next_shrink(cli);
 678 }
 679
 680 /* Shrink the current grant, either from some large amount to enough for a
 681  * full set of in-flight RPCs, or if we have already shrunk to that limit
 682  * then to enough for a single RPC.  This avoids keeping more grant than
 683  * needed, and avoids shrinking the grant piecemeal.
 684  */
 685 static int osc_shrink_grant(struct client_obd *cli)
 686 {
 687         __u64 target_bytes = (cli->cl_max_rpcs_in_flight + 1) *
 688                              (cli->cl_max_pages_per_rpc << PAGE_SHIFT);
 689
 690         spin_lock(&cli->cl_loi_list_lock);
 691         if (cli->cl_avail_grant <= target_bytes)
 692                 target_bytes = cli->cl_max_pages_per_rpc << PAGE_SHIFT;
 693         spin_unlock(&cli->cl_loi_list_lock);
 694
 695         return osc_shrink_grant_to_target(cli, target_bytes);
 696 }
 697
 698 int osc_shrink_grant_to_target(struct client_obd *cli, __u64 target_bytes)
 699 {
 700         int rc = 0;
 701         struct ost_body *body;
 702
 703         spin_lock(&cli->cl_loi_list_lock);
 704         /* Don't shrink if we are already above or below the desired limit
 705          * We don't want to shrink below a single RPC, as that will negatively
 706          * impact block allocation and long-term performance.
 707          */
 708         if (target_bytes < cli->cl_max_pages_per_rpc << PAGE_SHIFT)
 709                 target_bytes = cli->cl_max_pages_per_rpc << PAGE_SHIFT;
 710
 711         if (target_bytes >= cli->cl_avail_grant) {
 712                 spin_unlock(&cli->cl_loi_list_lock);
 713                 return 0;
 714         }
 715         spin_unlock(&cli->cl_loi_list_lock);
 716
 717         body = kzalloc(sizeof(*body), GFP_NOFS);
 718         if (!body)
 719                 return -ENOMEM;
 720
 721         osc_announce_cached(cli, &body->oa, 0);
 722
 723         spin_lock(&cli->cl_loi_list_lock);
 724         body->oa.o_grant = cli->cl_avail_grant - target_bytes;
 725         cli->cl_avail_grant = target_bytes;
 726         spin_unlock(&cli->cl_loi_list_lock);
 727         if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
 728                 body->oa.o_valid |= OBD_MD_FLFLAGS;
 729                 body->oa.o_flags = 0;
 730         }
 731         body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
 732         osc_update_next_shrink(cli);
 733
 734         rc = osc_set_info_async(NULL, cli->cl_import->imp_obd->obd_self_export,
 735                                 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
 736                                 sizeof(*body), body, NULL);
 737         if (rc != 0)
 738                 __osc_update_grant(cli, body->oa.o_grant);
 739         kfree(body);
 740         return rc;
 741 }
 742
 743 static int osc_should_shrink_grant(struct client_obd *client)
 744 {
 745         unsigned long time = cfs_time_current();
 746         unsigned long next_shrink = client->cl_next_shrink_grant;
 747
 748         if ((client->cl_import->imp_connect_data.ocd_connect_flags &
 749              OBD_CONNECT_GRANT_SHRINK) == 0)
 750                 return 0;
 751
 752         if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
 753                 /* Get the current RPC size directly, instead of going via:
 754                  * cli_brw_size(obd->u.cli.cl_import->imp_obd->obd_self_export)
 755                  * Keep comment here so that it can be found by searching.
 756                  */
 757                 int brw_size = client->cl_max_pages_per_rpc << PAGE_SHIFT;
 758
 759                 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
 760                     client->cl_avail_grant > brw_size)
 761                         return 1;
 762
 763                 osc_update_next_shrink(client);
 764         }
 765         return 0;
 766 }
 767
 768 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
 769 {
 770         struct client_obd *client;
 771
 772         list_for_each_entry(client, &item->ti_obd_list, cl_grant_shrink_list) {
 773                 if (osc_should_shrink_grant(client))
 774                         osc_shrink_grant(client);
 775         }
 776         return 0;
 777 }
 778
 779 static int osc_add_shrink_grant(struct client_obd *client)
 780 {
 781         int rc;
 782
 783         rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
 784                                        TIMEOUT_GRANT,
 785                                        osc_grant_shrink_grant_cb, NULL,
 786                                        &client->cl_grant_shrink_list);
 787         if (rc) {
 788                 CERROR("add grant client %s error %d\n", cli_name(client), rc);
 789                 return rc;
 790         }
 791         CDEBUG(D_CACHE, "add grant client %s\n", cli_name(client));
 792         osc_update_next_shrink(client);
 793         return 0;
 794 }
 795
 796 static int osc_del_shrink_grant(struct client_obd *client)
 797 {
 798         return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
 799                                          TIMEOUT_GRANT);
 800 }
 801
 802 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
 803 {
 804         /*
 805          * ocd_grant is the total grant amount we're expect to hold: if we've
 806          * been evicted, it's the new avail_grant amount, cl_dirty_pages will
 807          * drop to 0 as inflight RPCs fail out; otherwise, it's avail_grant +
 808          * dirty.
 809          *
 810          * race is tolerable here: if we're evicted, but imp_state already
 811          * left EVICTED state, then cl_dirty_pages must be 0 already.
 812          */
 813         spin_lock(&cli->cl_loi_list_lock);
 814         if (cli->cl_import->imp_state == LUSTRE_IMP_EVICTED)
 815                 cli->cl_avail_grant = ocd->ocd_grant;
 816         else
 817                 cli->cl_avail_grant = ocd->ocd_grant -
 818                                       (cli->cl_dirty_pages << PAGE_SHIFT);
 819
 820         /* determine the appropriate chunk size used by osc_extent. */
 821         cli->cl_chunkbits = max_t(int, PAGE_SHIFT, ocd->ocd_blocksize);
 822         spin_unlock(&cli->cl_loi_list_lock);
 823
 824         CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld chunk bits: %d\n",
 825                cli_name(cli), cli->cl_avail_grant, cli->cl_lost_grant,
 826                cli->cl_chunkbits);
 827
 828         if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
 829             list_empty(&cli->cl_grant_shrink_list))
 830                 osc_add_shrink_grant(cli);
 831 }
 832
 833 /* We assume that the reason this OSC got a short read is because it read
 834  * beyond the end of a stripe file; i.e. lustre is reading a sparse file
 835  * via the LOV, and it _knows_ it's reading inside the file, it's just that
 836  * this stripe never got written at or beyond this stripe offset yet.
 837  */
 838 static void handle_short_read(int nob_read, u32 page_count,
 839                               struct brw_page **pga)
 840 {
 841         char *ptr;
 842         int i = 0;
 843
 844         /* skip bytes read OK */
 845         while (nob_read > 0) {
 846                 LASSERT(page_count > 0);
 847
 848                 if (pga[i]->count > nob_read) {
 849                         /* EOF inside this page */
 850                         ptr = kmap(pga[i]->pg) +
 851                                 (pga[i]->off & ~PAGE_MASK);
 852                         memset(ptr + nob_read, 0, pga[i]->count - nob_read);
 853                         kunmap(pga[i]->pg);
 854                         page_count--;
 855                         i++;
 856                         break;
 857                 }
 858
 859                 nob_read -= pga[i]->count;
 860                 page_count--;
 861                 i++;
 862         }
 863
 864         /* zero remaining pages */
 865         while (page_count-- > 0) {
 866                 ptr = kmap(pga[i]->pg) + (pga[i]->off & ~PAGE_MASK);
 867                 memset(ptr, 0, pga[i]->count);
 868                 kunmap(pga[i]->pg);
 869                 i++;
 870         }
 871 }
 872
 873 static int check_write_rcs(struct ptlrpc_request *req,
 874                            int requested_nob, int niocount,
 875                            u32 page_count, struct brw_page **pga)
 876 {
 877         int i;
 878         __u32 *remote_rcs;
 879
 880         remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
 881                                                   sizeof(*remote_rcs) *
 882                                                   niocount);
 883         if (!remote_rcs) {
 884                 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
 885                 return -EPROTO;
 886         }
 887
 888         /* return error if any niobuf was in error */
 889         for (i = 0; i < niocount; i++) {
 890                 if ((int)remote_rcs[i] < 0)
 891                         return remote_rcs[i];
 892
 893                 if (remote_rcs[i] != 0) {
 894                         CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
 895                                i, remote_rcs[i], req);
 896                         return -EPROTO;
 897                 }
 898         }
 899
 900         if (req->rq_bulk->bd_nob_transferred != requested_nob) {
 901                 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
 902                        req->rq_bulk->bd_nob_transferred, requested_nob);
 903                 return -EPROTO;
 904         }
 905
 906         return 0;
 907 }
 908
 909 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
 910 {
 911         if (p1->flag != p2->flag) {
 912                 unsigned int mask = ~(OBD_BRW_FROM_GRANT | OBD_BRW_NOCACHE |
 913                                       OBD_BRW_SYNC | OBD_BRW_ASYNC |
 914                                       OBD_BRW_NOQUOTA | OBD_BRW_SOFT_SYNC);
 915
 916                 /* warn if we try to combine flags that we don't know to be
 917                  * safe to combine
 918                  */
 919                 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
 920                         CWARN("Saw flags 0x%x and 0x%x in the same brw, please report this at http://bugs.whamcloud.com/\n",
 921                               p1->flag, p2->flag);
 922                 }
 923                 return 0;
 924         }
 925
 926         return (p1->off + p1->count == p2->off);
 927 }
 928
 929 static u32 osc_checksum_bulk(int nob, u32 pg_count,
 930                              struct brw_page **pga, int opc,
 931                              enum cksum_type cksum_type)
 932 {
 933         __u32 cksum;
 934         int i = 0;
 935         struct cfs_crypto_hash_desc *hdesc;
 936         unsigned int bufsize;
 937         unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
 938
 939         LASSERT(pg_count > 0);
 940
 941         hdesc = cfs_crypto_hash_init(cfs_alg, NULL, 0);
 942         if (IS_ERR(hdesc)) {
 943                 CERROR("Unable to initialize checksum hash %s\n",
 944                        cfs_crypto_hash_name(cfs_alg));
 945                 return PTR_ERR(hdesc);
 946         }
 947
 948         while (nob > 0 && pg_count > 0) {
 949                 unsigned int count = pga[i]->count > nob ? nob : pga[i]->count;
 950
 951                 /* corrupt the data before we compute the checksum, to
 952                  * simulate an OST->client data error
 953                  */
 954                 if (i == 0 && opc == OST_READ &&
 955                     OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE)) {
 956                         unsigned char *ptr = kmap(pga[i]->pg);
 957                         int off = pga[i]->off & ~PAGE_MASK;
 958
 959                         memcpy(ptr + off, "bad1", min_t(typeof(nob), 4, nob));
 960                         kunmap(pga[i]->pg);
 961                 }
 962                 cfs_crypto_hash_update_page(hdesc, pga[i]->pg,
 963                                             pga[i]->off & ~PAGE_MASK,
 964                                   count);
 965                 CDEBUG(D_PAGE,
 966                        "page %p map %p index %lu flags %lx count %u priv %0lx: off %d\n",
 967                        pga[i]->pg, pga[i]->pg->mapping, pga[i]->pg->index,
 968                        (long)pga[i]->pg->flags, page_count(pga[i]->pg),
 969                        page_private(pga[i]->pg),
 970                        (int)(pga[i]->off & ~PAGE_MASK));
 971
 972                 nob -= pga[i]->count;
 973                 pg_count--;
 974                 i++;
 975         }
 976
 977         bufsize = sizeof(cksum);
 978         cfs_crypto_hash_final(hdesc, (unsigned char *)&cksum, &bufsize);
 979
 980         /* For sending we only compute the wrong checksum instead
 981          * of corrupting the data so it is still correct on a redo
 982          */
 983         if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
 984                 cksum++;
 985
 986         return cksum;
 987 }
 988
 989 static int osc_brw_prep_request(int cmd, struct client_obd *cli,
 990                                 struct obdo *oa, u32 page_count,
 991                                 struct brw_page **pga,
 992                                 struct ptlrpc_request **reqp,
 993                                 int reserve,
 994                                 int resend)
 995 {
 996         struct ptlrpc_request *req;
 997         struct ptlrpc_bulk_desc *desc;
 998         struct ost_body *body;
 999         struct obd_ioobj *ioobj;
1000         struct niobuf_remote *niobuf;
1001         int niocount, i, requested_nob, opc, rc;
1002         struct osc_brw_async_args *aa;
1003         struct req_capsule *pill;
1004         struct brw_page *pg_prev;
1005
1006         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1007                 return -ENOMEM; /* Recoverable */
1008         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1009                 return -EINVAL; /* Fatal */
1010
1011         if ((cmd & OBD_BRW_WRITE) != 0) {
1012                 opc = OST_WRITE;
1013                 req = ptlrpc_request_alloc_pool(cli->cl_import,
1014                                                 osc_rq_pool,
1015                                                 &RQF_OST_BRW_WRITE);
1016         } else {
1017                 opc = OST_READ;
1018                 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1019         }
1020         if (!req)
1021                 return -ENOMEM;
1022
1023         for (niocount = i = 1; i < page_count; i++) {
1024                 if (!can_merge_pages(pga[i - 1], pga[i]))
1025                         niocount++;
1026         }
1027
1028         pill = &req->rq_pill;
1029         req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1030                              sizeof(*ioobj));
1031         req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1032                              niocount * sizeof(*niobuf));
1033
1034         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1035         if (rc) {
1036                 ptlrpc_request_free(req);
1037                 return rc;
1038         }
1039         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1040         ptlrpc_at_set_req_timeout(req);
1041         /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1042          * retry logic
1043          */
1044         req->rq_no_retry_einprogress = 1;
1045
1046         desc = ptlrpc_prep_bulk_imp(req, page_count,
1047                 cli->cl_import->imp_connect_data.ocd_brw_size >> LNET_MTU_BITS,
1048                 (opc == OST_WRITE ? PTLRPC_BULK_GET_SOURCE :
1049                  PTLRPC_BULK_PUT_SINK) | PTLRPC_BULK_BUF_KIOV, OST_BULK_PORTAL,
1050                  &ptlrpc_bulk_kiov_pin_ops);
1051
1052         if (!desc) {
1053                 rc = -ENOMEM;
1054                 goto out;
1055         }
1056         /* NB request now owns desc and will free it when it gets freed */
1057
1058         body = req_capsule_client_get(pill, &RMF_OST_BODY);
1059         ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1060         niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1061         LASSERT(body && ioobj && niobuf);
1062
1063         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1064
1065         obdo_to_ioobj(oa, ioobj);
1066         ioobj->ioo_bufcnt = niocount;
1067         /* The high bits of ioo_max_brw tells server _maximum_ number of bulks
1068          * that might be send for this request.  The actual number is decided
1069          * when the RPC is finally sent in ptlrpc_register_bulk(). It sends
1070          * "max - 1" for old client compatibility sending "0", and also so the
1071          * the actual maximum is a power-of-two number, not one less. LU-1431
1072          */
1073         ioobj_max_brw_set(ioobj, desc->bd_md_max_brw);
1074         LASSERT(page_count > 0);
1075         pg_prev = pga[0];
1076         for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1077                 struct brw_page *pg = pga[i];
1078                 int poff = pg->off & ~PAGE_MASK;
1079
1080                 LASSERT(pg->count > 0);
1081                 /* make sure there is no gap in the middle of page array */
1082                 LASSERTF(page_count == 1 ||
1083                          (ergo(i == 0, poff + pg->count == PAGE_SIZE) &&
1084                           ergo(i > 0 && i < page_count - 1,
1085                                poff == 0 && pg->count == PAGE_SIZE)   &&
1086                           ergo(i == page_count - 1, poff == 0)),
1087                          "i: %d/%d pg: %p off: %llu, count: %u\n",
1088                          i, page_count, pg, pg->off, pg->count);
1089                 LASSERTF(i == 0 || pg->off > pg_prev->off,
1090                          "i %d p_c %u pg %p [pri %lu ind %lu] off %llu prev_pg %p [pri %lu ind %lu] off %llu\n",
1091                          i, page_count,
1092                          pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1093                          pg_prev->pg, page_private(pg_prev->pg),
1094                          pg_prev->pg->index, pg_prev->off);
1095                 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1096                         (pg->flag & OBD_BRW_SRVLOCK));
1097
1098                 desc->bd_frag_ops->add_kiov_frag(desc, pg->pg, poff, pg->count);
1099                 requested_nob += pg->count;
1100
1101                 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1102                         niobuf--;
1103                         niobuf->rnb_len += pg->count;
1104                 } else {
1105                         niobuf->rnb_offset = pg->off;
1106                         niobuf->rnb_len = pg->count;
1107                         niobuf->rnb_flags = pg->flag;
1108                 }
1109                 pg_prev = pg;
1110         }
1111
1112         LASSERTF((void *)(niobuf - niocount) ==
1113                 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1114                 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1115                 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1116
1117         osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1118         if (resend) {
1119                 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1120                         body->oa.o_valid |= OBD_MD_FLFLAGS;
1121                         body->oa.o_flags = 0;
1122                 }
1123                 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1124         }
1125
1126         if (osc_should_shrink_grant(cli))
1127                 osc_shrink_grant_local(cli, &body->oa);
1128
1129         /* size[REQ_REC_OFF] still sizeof (*body) */
1130         if (opc == OST_WRITE) {
1131                 if (cli->cl_checksum &&
1132                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1133                         /* store cl_cksum_type in a local variable since
1134                          * it can be changed via lprocfs
1135                          */
1136                         enum cksum_type cksum_type = cli->cl_cksum_type;
1137
1138                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1139                                 oa->o_flags &= OBD_FL_LOCAL_MASK;
1140                                 body->oa.o_flags = 0;
1141                         }
1142                         body->oa.o_flags |= cksum_type_pack(cksum_type);
1143                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1144                         body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1145                                                              page_count, pga,
1146                                                              OST_WRITE,
1147                                                              cksum_type);
1148                         CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1149                                body->oa.o_cksum);
1150                         /* save this in 'oa', too, for later checking */
1151                         oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1152                         oa->o_flags |= cksum_type_pack(cksum_type);
1153                 } else {
1154                         /* clear out the checksum flag, in case this is a
1155                          * resend but cl_checksum is no longer set. b=11238
1156                          */
1157                         oa->o_valid &= ~OBD_MD_FLCKSUM;
1158                 }
1159                 oa->o_cksum = body->oa.o_cksum;
1160                 /* 1 RC per niobuf */
1161                 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1162                                      sizeof(__u32) * niocount);
1163         } else {
1164                 if (cli->cl_checksum &&
1165                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1166                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1167                                 body->oa.o_flags = 0;
1168                         body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1169                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1170                 }
1171         }
1172         ptlrpc_request_set_replen(req);
1173
1174         BUILD_BUG_ON(sizeof(*aa) > sizeof(req->rq_async_args));
1175         aa = ptlrpc_req_async_args(req);
1176         aa->aa_oa = oa;
1177         aa->aa_requested_nob = requested_nob;
1178         aa->aa_nio_count = niocount;
1179         aa->aa_page_count = page_count;
1180         aa->aa_resends = 0;
1181         aa->aa_ppga = pga;
1182         aa->aa_cli = cli;
1183         INIT_LIST_HEAD(&aa->aa_oaps);
1184
1185         *reqp = req;
1186         niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1187         CDEBUG(D_RPCTRACE, "brw rpc %p - object " DOSTID " offset %lld<>%lld\n",
1188                req, POSTID(&oa->o_oi), niobuf[0].rnb_offset,
1189                niobuf[niocount - 1].rnb_offset + niobuf[niocount - 1].rnb_len);
1190
1191         return 0;
1192
1193  out:
1194         ptlrpc_req_finished(req);
1195         return rc;
1196 }
1197
1198 static int check_write_checksum(struct obdo *oa,
1199                                 const struct lnet_process_id *peer,
1200                                 __u32 client_cksum, __u32 server_cksum, int nob,
1201                                 u32 page_count, struct brw_page **pga,
1202                                 enum cksum_type client_cksum_type)
1203 {
1204         __u32 new_cksum;
1205         char *msg;
1206         enum cksum_type cksum_type;
1207
1208         if (server_cksum == client_cksum) {
1209                 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1210                 return 0;
1211         }
1212
1213         cksum_type = cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1214                                        oa->o_flags : 0);
1215         new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1216                                       cksum_type);
1217
1218         if (cksum_type != client_cksum_type)
1219                 msg = "the server did not use the checksum type specified in the original request - likely a protocol problem"
1220                         ;
1221         else if (new_cksum == server_cksum)
1222                 msg = "changed on the client after we checksummed it - likely false positive due to mmap IO (bug 11742)"
1223                         ;
1224         else if (new_cksum == client_cksum)
1225                 msg = "changed in transit before arrival at OST";
1226         else
1227                 msg = "changed in transit AND doesn't match the original - likely false positive due to mmap IO (bug 11742)"
1228                         ;
1229
1230         LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode " DFID " object " DOSTID " extent [%llu-%llu]\n",
1231                            msg, libcfs_nid2str(peer->nid),
1232                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1233                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1234                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1235                            POSTID(&oa->o_oi), pga[0]->off,
1236                            pga[page_count - 1]->off +
1237                            pga[page_count - 1]->count - 1);
1238         CERROR("original client csum %x (type %x), server csum %x (type %x), client csum now %x\n",
1239                client_cksum, client_cksum_type,
1240                server_cksum, cksum_type, new_cksum);
1241         return 1;
1242 }
1243
1244 /* Note rc enters this function as number of bytes transferred */
1245 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1246 {
1247         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1248         const struct lnet_process_id *peer =
1249                         &req->rq_import->imp_connection->c_peer;
1250         struct client_obd *cli = aa->aa_cli;
1251         struct ost_body *body;
1252         __u32 client_cksum = 0;
1253
1254         if (rc < 0 && rc != -EDQUOT) {
1255                 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1256                 return rc;
1257         }
1258
1259         LASSERTF(req->rq_repmsg, "rc = %d\n", rc);
1260         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1261         if (!body) {
1262                 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1263                 return -EPROTO;
1264         }
1265
1266         /* set/clear over quota flag for a uid/gid */
1267         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1268             body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1269                 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1270
1271                 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid %#llx, flags %x\n",
1272                        body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1273                        body->oa.o_flags);
1274                 osc_quota_setdq(cli, qid, body->oa.o_valid, body->oa.o_flags);
1275         }
1276
1277         osc_update_grant(cli, body);
1278
1279         if (rc < 0)
1280                 return rc;
1281
1282         if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1283                 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1284
1285         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1286                 if (rc > 0) {
1287                         CERROR("Unexpected +ve rc %d\n", rc);
1288                         return -EPROTO;
1289                 }
1290                 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1291
1292                 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1293                         return -EAGAIN;
1294
1295                 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1296                     check_write_checksum(&body->oa, peer, client_cksum,
1297                                          body->oa.o_cksum, aa->aa_requested_nob,
1298                                          aa->aa_page_count, aa->aa_ppga,
1299                                          cksum_type_unpack(aa->aa_oa->o_flags)))
1300                         return -EAGAIN;
1301
1302                 rc = check_write_rcs(req, aa->aa_requested_nob,
1303                                      aa->aa_nio_count,
1304                                      aa->aa_page_count, aa->aa_ppga);
1305                 goto out;
1306         }
1307
1308         /* The rest of this function executes only for OST_READs */
1309
1310         /* if unwrap_bulk failed, return -EAGAIN to retry */
1311         rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1312         if (rc < 0) {
1313                 rc = -EAGAIN;
1314                 goto out;
1315         }
1316
1317         if (rc > aa->aa_requested_nob) {
1318                 CERROR("Unexpected rc %d (%d requested)\n", rc,
1319                        aa->aa_requested_nob);
1320                 return -EPROTO;
1321         }
1322
1323         if (rc != req->rq_bulk->bd_nob_transferred) {
1324                 CERROR("Unexpected rc %d (%d transferred)\n",
1325                        rc, req->rq_bulk->bd_nob_transferred);
1326                 return -EPROTO;
1327         }
1328
1329         if (rc < aa->aa_requested_nob)
1330                 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1331
1332         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1333                 static int cksum_counter;
1334                 __u32 server_cksum = body->oa.o_cksum;
1335                 char *via = "";
1336                 char *router = "";
1337                 enum cksum_type cksum_type;
1338
1339                 cksum_type = cksum_type_unpack(body->oa.o_valid &
1340                                                OBD_MD_FLFLAGS ?
1341                                                body->oa.o_flags : 0);
1342                 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1343                                                  aa->aa_ppga, OST_READ,
1344                                                  cksum_type);
1345
1346                 if (peer->nid != req->rq_bulk->bd_sender) {
1347                         via = " via ";
1348                         router = libcfs_nid2str(req->rq_bulk->bd_sender);
1349                 }
1350
1351                 if (server_cksum != client_cksum) {
1352                         LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from %s%s%s inode " DFID " object " DOSTID " extent [%llu-%llu]\n",
1353                                            req->rq_import->imp_obd->obd_name,
1354                                            libcfs_nid2str(peer->nid),
1355                                            via, router,
1356                                            body->oa.o_valid & OBD_MD_FLFID ?
1357                                            body->oa.o_parent_seq : (__u64)0,
1358                                            body->oa.o_valid & OBD_MD_FLFID ?
1359                                            body->oa.o_parent_oid : 0,
1360                                            body->oa.o_valid & OBD_MD_FLFID ?
1361                                            body->oa.o_parent_ver : 0,
1362                                            POSTID(&body->oa.o_oi),
1363                                            aa->aa_ppga[0]->off,
1364                                            aa->aa_ppga[aa->aa_page_count-1]->off +
1365                                            aa->aa_ppga[aa->aa_page_count-1]->count -
1366                                            1);
1367                         CERROR("client %x, server %x, cksum_type %x\n",
1368                                client_cksum, server_cksum, cksum_type);
1369                         cksum_counter = 0;
1370                         aa->aa_oa->o_cksum = client_cksum;
1371                         rc = -EAGAIN;
1372                 } else {
1373                         cksum_counter++;
1374                         CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1375                         rc = 0;
1376                 }
1377         } else if (unlikely(client_cksum)) {
1378                 static int cksum_missed;
1379
1380                 cksum_missed++;
1381                 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1382                         CERROR("Checksum %u requested from %s but not sent\n",
1383                                cksum_missed, libcfs_nid2str(peer->nid));
1384         } else {
1385                 rc = 0;
1386         }
1387 out:
1388         if (rc >= 0)
1389                 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
1390                                      aa->aa_oa, &body->oa);
1391
1392         return rc;
1393 }
1394
1395 static int osc_brw_redo_request(struct ptlrpc_request *request,
1396                                 struct osc_brw_async_args *aa, int rc)
1397 {
1398         struct ptlrpc_request *new_req;
1399         struct osc_brw_async_args *new_aa;
1400         struct osc_async_page *oap;
1401
1402         DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
1403                   "redo for recoverable error %d", rc);
1404
1405         rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1406                                         OST_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
1407                                   aa->aa_cli, aa->aa_oa,
1408                                   aa->aa_page_count, aa->aa_ppga,
1409                                   &new_req, 0, 1);
1410         if (rc)
1411                 return rc;
1412
1413         list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1414                 if (oap->oap_request) {
1415                         LASSERTF(request == oap->oap_request,
1416                                  "request %p != oap_request %p\n",
1417                                  request, oap->oap_request);
1418                         if (oap->oap_interrupted) {
1419                                 ptlrpc_req_finished(new_req);
1420                                 return -EINTR;
1421                         }
1422                 }
1423         }
1424         /* New request takes over pga and oaps from old request.
1425          * Note that copying a list_head doesn't work, need to move it...
1426          */
1427         aa->aa_resends++;
1428         new_req->rq_interpret_reply = request->rq_interpret_reply;
1429         new_req->rq_async_args = request->rq_async_args;
1430         new_req->rq_commit_cb = request->rq_commit_cb;
1431         /* cap resend delay to the current request timeout, this is similar to
1432          * what ptlrpc does (see after_reply())
1433          */
1434         if (aa->aa_resends > new_req->rq_timeout)
1435                 new_req->rq_sent = ktime_get_real_seconds() + new_req->rq_timeout;
1436         else
1437                 new_req->rq_sent = ktime_get_real_seconds() + aa->aa_resends;
1438         new_req->rq_generation_set = 1;
1439         new_req->rq_import_generation = request->rq_import_generation;
1440
1441         new_aa = ptlrpc_req_async_args(new_req);
1442
1443         INIT_LIST_HEAD(&new_aa->aa_oaps);
1444         list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
1445         INIT_LIST_HEAD(&new_aa->aa_exts);
1446         list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
1447         new_aa->aa_resends = aa->aa_resends;
1448
1449         list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1450                 if (oap->oap_request) {
1451                         ptlrpc_req_finished(oap->oap_request);
1452                         oap->oap_request = ptlrpc_request_addref(new_req);
1453                 }
1454         }
1455
1456         /* XXX: This code will run into problem if we're going to support
1457          * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
1458          * and wait for all of them to be finished. We should inherit request
1459          * set from old request.
1460          */
1461         ptlrpcd_add_req(new_req);
1462
1463         DEBUG_REQ(D_INFO, new_req, "new request");
1464         return 0;
1465 }
1466
1467 /*
1468  * ugh, we want disk allocation on the target to happen in offset order.  we'll
1469  * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1470  * fine for our small page arrays and doesn't require allocation.  its an
1471  * insertion sort that swaps elements that are strides apart, shrinking the
1472  * stride down until its '1' and the array is sorted.
1473  */
1474 static void sort_brw_pages(struct brw_page **array, int num)
1475 {
1476         int stride, i, j;
1477         struct brw_page *tmp;
1478
1479         if (num == 1)
1480                 return;
1481         for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1482                 ;
1483
1484         do {
1485                 stride /= 3;
1486                 for (i = stride ; i < num ; i++) {
1487                         tmp = array[i];
1488                         j = i;
1489                         while (j >= stride && array[j - stride]->off > tmp->off) {
1490                                 array[j] = array[j - stride];
1491                                 j -= stride;
1492                         }
1493                         array[j] = tmp;
1494                 }
1495         } while (stride > 1);
1496 }
1497
1498 static void osc_release_ppga(struct brw_page **ppga, u32 count)
1499 {
1500         LASSERT(ppga);
1501         kfree(ppga);
1502 }
1503
1504 static int brw_interpret(const struct lu_env *env,
1505                          struct ptlrpc_request *req, void *data, int rc)
1506 {
1507         struct osc_brw_async_args *aa = data;
1508         struct osc_extent *ext;
1509         struct osc_extent *tmp;
1510         struct client_obd *cli = aa->aa_cli;
1511
1512         rc = osc_brw_fini_request(req, rc);
1513         CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
1514         /* When server return -EINPROGRESS, client should always retry
1515          * regardless of the number of times the bulk was resent already.
1516          */
1517         if (osc_recoverable_error(rc)) {
1518                 if (req->rq_import_generation !=
1519                     req->rq_import->imp_generation) {
1520                         CDEBUG(D_HA, "%s: resend cross eviction for object: " DOSTID ", rc = %d.\n",
1521                                req->rq_import->imp_obd->obd_name,
1522                                POSTID(&aa->aa_oa->o_oi), rc);
1523                 } else if (rc == -EINPROGRESS ||
1524                     client_should_resend(aa->aa_resends, aa->aa_cli)) {
1525                         rc = osc_brw_redo_request(req, aa, rc);
1526                 } else {
1527                         CERROR("%s: too many resent retries for object: %llu:%llu, rc = %d.\n",
1528                                req->rq_import->imp_obd->obd_name,
1529                                POSTID(&aa->aa_oa->o_oi), rc);
1530                 }
1531
1532                 if (rc == 0)
1533                         return 0;
1534                 else if (rc == -EAGAIN || rc == -EINPROGRESS)
1535                         rc = -EIO;
1536         }
1537
1538         if (rc == 0) {
1539                 struct obdo *oa = aa->aa_oa;
1540                 struct cl_attr *attr  = &osc_env_info(env)->oti_attr;
1541                 unsigned long valid = 0;
1542                 struct cl_object *obj;
1543                 struct osc_async_page *last;
1544
1545                 last = brw_page2oap(aa->aa_ppga[aa->aa_page_count - 1]);
1546                 obj = osc2cl(last->oap_obj);
1547
1548                 cl_object_attr_lock(obj);
1549                 if (oa->o_valid & OBD_MD_FLBLOCKS) {
1550                         attr->cat_blocks = oa->o_blocks;
1551                         valid |= CAT_BLOCKS;
1552                 }
1553                 if (oa->o_valid & OBD_MD_FLMTIME) {
1554                         attr->cat_mtime = oa->o_mtime;
1555                         valid |= CAT_MTIME;
1556                 }
1557                 if (oa->o_valid & OBD_MD_FLATIME) {
1558                         attr->cat_atime = oa->o_atime;
1559                         valid |= CAT_ATIME;
1560                 }
1561                 if (oa->o_valid & OBD_MD_FLCTIME) {
1562                         attr->cat_ctime = oa->o_ctime;
1563                         valid |= CAT_CTIME;
1564                 }
1565
1566                 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1567                         struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
1568                         loff_t last_off = last->oap_count + last->oap_obj_off +
1569                                           last->oap_page_off;
1570
1571                         /* Change file size if this is an out of quota or
1572                          * direct IO write and it extends the file size
1573                          */
1574                         if (loi->loi_lvb.lvb_size < last_off) {
1575                                 attr->cat_size = last_off;
1576                                 valid |= CAT_SIZE;
1577                         }
1578                         /* Extend KMS if it's not a lockless write */
1579                         if (loi->loi_kms < last_off &&
1580                             oap2osc_page(last)->ops_srvlock == 0) {
1581                                 attr->cat_kms = last_off;
1582                                 valid |= CAT_KMS;
1583                         }
1584                 }
1585
1586                 if (valid != 0)
1587                         cl_object_attr_update(env, obj, attr, valid);
1588                 cl_object_attr_unlock(obj);
1589         }
1590         kmem_cache_free(obdo_cachep, aa->aa_oa);
1591
1592         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE && rc == 0)
1593                 osc_inc_unstable_pages(req);
1594
1595         list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
1596                 list_del_init(&ext->oe_link);
1597                 osc_extent_finish(env, ext, 1, rc);
1598         }
1599         LASSERT(list_empty(&aa->aa_exts));
1600         LASSERT(list_empty(&aa->aa_oaps));
1601
1602         osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
1603         ptlrpc_lprocfs_brw(req, req->rq_bulk->bd_nob_transferred);
1604
1605         spin_lock(&cli->cl_loi_list_lock);
1606         /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
1607          * is called so we know whether to go to sync BRWs or wait for more
1608          * RPCs to complete
1609          */
1610         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
1611                 cli->cl_w_in_flight--;
1612         else
1613                 cli->cl_r_in_flight--;
1614         osc_wake_cache_waiters(cli);
1615         spin_unlock(&cli->cl_loi_list_lock);
1616
1617         osc_io_unplug(env, cli, NULL);
1618         return rc;
1619 }
1620
1621 static void brw_commit(struct ptlrpc_request *req)
1622 {
1623         /*
1624          * If osc_inc_unstable_pages (via osc_extent_finish) races with
1625          * this called via the rq_commit_cb, I need to ensure
1626          * osc_dec_unstable_pages is still called. Otherwise unstable
1627          * pages may be leaked.
1628          */
1629         spin_lock(&req->rq_lock);
1630         if (unlikely(req->rq_unstable)) {
1631                 req->rq_unstable = 0;
1632                 spin_unlock(&req->rq_lock);
1633                 osc_dec_unstable_pages(req);
1634         } else {
1635                 req->rq_committed = 1;
1636                 spin_unlock(&req->rq_lock);
1637         }
1638 }
1639
1640 /**
1641  * Build an RPC by the list of extent @ext_list. The caller must ensure
1642  * that the total pages in this list are NOT over max pages per RPC.
1643  * Extents in the list must be in OES_RPC state.
1644  */
1645 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
1646                   struct list_head *ext_list, int cmd)
1647 {
1648         struct ptlrpc_request *req = NULL;
1649         struct osc_extent *ext;
1650         struct brw_page **pga = NULL;
1651         struct osc_brw_async_args *aa = NULL;
1652         struct obdo *oa = NULL;
1653         struct osc_async_page *oap;
1654         struct osc_object *obj = NULL;
1655         struct cl_req_attr *crattr = NULL;
1656         u64 starting_offset = OBD_OBJECT_EOF;
1657         u64 ending_offset = 0;
1658         int mpflag = 0;
1659         int mem_tight = 0;
1660         int page_count = 0;
1661         bool soft_sync = false;
1662         bool interrupted = false;
1663         int i;
1664         int rc;
1665         struct ost_body *body;
1666         LIST_HEAD(rpc_list);
1667
1668         LASSERT(!list_empty(ext_list));
1669
1670         /* add pages into rpc_list to build BRW rpc */
1671         list_for_each_entry(ext, ext_list, oe_link) {
1672                 LASSERT(ext->oe_state == OES_RPC);
1673                 mem_tight |= ext->oe_memalloc;
1674                 page_count += ext->oe_nr_pages;
1675                 if (!obj)
1676                         obj = ext->oe_obj;
1677         }
1678
1679         soft_sync = osc_over_unstable_soft_limit(cli);
1680         if (mem_tight)
1681                 mpflag = cfs_memory_pressure_get_and_set();
1682
1683         pga = kcalloc(page_count, sizeof(*pga), GFP_NOFS);
1684         if (!pga) {
1685                 rc = -ENOMEM;
1686                 goto out;
1687         }
1688
1689         oa = kmem_cache_zalloc(obdo_cachep, GFP_NOFS);
1690         if (!oa) {
1691                 rc = -ENOMEM;
1692                 goto out;
1693         }
1694
1695         i = 0;
1696         list_for_each_entry(ext, ext_list, oe_link) {
1697                 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1698                         if (mem_tight)
1699                                 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
1700                         if (soft_sync)
1701                                 oap->oap_brw_flags |= OBD_BRW_SOFT_SYNC;
1702                         pga[i] = &oap->oap_brw_page;
1703                         pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
1704                         i++;
1705
1706                         list_add_tail(&oap->oap_rpc_item, &rpc_list);
1707                         if (starting_offset == OBD_OBJECT_EOF ||
1708                             starting_offset > oap->oap_obj_off)
1709                                 starting_offset = oap->oap_obj_off;
1710                         else
1711                                 LASSERT(!oap->oap_page_off);
1712                         if (ending_offset < oap->oap_obj_off + oap->oap_count)
1713                                 ending_offset = oap->oap_obj_off +
1714                                                 oap->oap_count;
1715                         else
1716                                 LASSERT(oap->oap_page_off + oap->oap_count ==
1717                                         PAGE_SIZE);
1718                         if (oap->oap_interrupted)
1719                                 interrupted = true;
1720                 }
1721         }
1722
1723         /* first page in the list */
1724         oap = list_entry(rpc_list.next, typeof(*oap), oap_rpc_item);
1725
1726         crattr = &osc_env_info(env)->oti_req_attr;
1727         memset(crattr, 0, sizeof(*crattr));
1728         crattr->cra_type = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
1729         crattr->cra_flags = ~0ULL;
1730         crattr->cra_page = oap2cl_page(oap);
1731         crattr->cra_oa = oa;
1732         cl_req_attr_set(env, osc2cl(obj), crattr);
1733
1734         sort_brw_pages(pga, page_count);
1735         rc = osc_brw_prep_request(cmd, cli, oa, page_count, pga, &req, 1, 0);
1736         if (rc != 0) {
1737                 CERROR("prep_req failed: %d\n", rc);
1738                 goto out;
1739         }
1740
1741         req->rq_commit_cb = brw_commit;
1742         req->rq_interpret_reply = brw_interpret;
1743
1744         req->rq_memalloc = mem_tight != 0;
1745         oap->oap_request = ptlrpc_request_addref(req);
1746         if (interrupted && !req->rq_intr)
1747                 ptlrpc_mark_interrupted(req);
1748
1749         /* Need to update the timestamps after the request is built in case
1750          * we race with setattr (locally or in queue at OST).  If OST gets
1751          * later setattr before earlier BRW (as determined by the request xid),
1752          * the OST will not use BRW timestamps.  Sadly, there is no obvious
1753          * way to do this in a single call.  bug 10150
1754          */
1755         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1756         crattr->cra_oa = &body->oa;
1757         crattr->cra_flags = OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLATIME;
1758         cl_req_attr_set(env, osc2cl(obj), crattr);
1759         lustre_msg_set_jobid(req->rq_reqmsg, crattr->cra_jobid);
1760
1761         BUILD_BUG_ON(sizeof(*aa) > sizeof(req->rq_async_args));
1762         aa = ptlrpc_req_async_args(req);
1763         INIT_LIST_HEAD(&aa->aa_oaps);
1764         list_splice_init(&rpc_list, &aa->aa_oaps);
1765         INIT_LIST_HEAD(&aa->aa_exts);
1766         list_splice_init(ext_list, &aa->aa_exts);
1767
1768         spin_lock(&cli->cl_loi_list_lock);
1769         starting_offset >>= PAGE_SHIFT;
1770         if (cmd == OBD_BRW_READ) {
1771                 cli->cl_r_in_flight++;
1772                 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
1773                 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
1774                 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
1775                                       starting_offset + 1);
1776         } else {
1777                 cli->cl_w_in_flight++;
1778                 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
1779                 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
1780                 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
1781                                       starting_offset + 1);
1782         }
1783         spin_unlock(&cli->cl_loi_list_lock);
1784
1785         DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %ur/%dw in flight",
1786                   page_count, aa, cli->cl_r_in_flight,
1787                   cli->cl_w_in_flight);
1788         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_IO, cfs_fail_val);
1789
1790         ptlrpcd_add_req(req);
1791         rc = 0;
1792
1793 out:
1794         if (mem_tight != 0)
1795                 cfs_memory_pressure_restore(mpflag);
1796
1797         if (rc != 0) {
1798                 LASSERT(!req);
1799
1800                 if (oa)
1801                         kmem_cache_free(obdo_cachep, oa);
1802                 kfree(pga);
1803                 /* this should happen rarely and is pretty bad, it makes the
1804                  * pending list not follow the dirty order
1805                  */
1806                 while (!list_empty(ext_list)) {
1807                         ext = list_entry(ext_list->next, struct osc_extent,
1808                                          oe_link);
1809                         list_del_init(&ext->oe_link);
1810                         osc_extent_finish(env, ext, 0, rc);
1811                 }
1812         }
1813         return rc;
1814 }
1815
1816 static int osc_set_lock_data(struct ldlm_lock *lock, void *data)
1817 {
1818         int set = 0;
1819
1820         LASSERT(lock);
1821
1822         lock_res_and_lock(lock);
1823
1824         if (!lock->l_ast_data)
1825                 lock->l_ast_data = data;
1826         if (lock->l_ast_data == data)
1827                 set = 1;
1828
1829         unlock_res_and_lock(lock);
1830
1831         return set;
1832 }
1833
1834 static int osc_enqueue_fini(struct ptlrpc_request *req,
1835                             osc_enqueue_upcall_f upcall, void *cookie,
1836                             struct lustre_handle *lockh, enum ldlm_mode mode,
1837                             __u64 *flags, int agl, int errcode)
1838 {
1839         bool intent = *flags & LDLM_FL_HAS_INTENT;
1840         int rc;
1841
1842         /* The request was created before ldlm_cli_enqueue call. */
1843         if (intent && errcode == ELDLM_LOCK_ABORTED) {
1844                 struct ldlm_reply *rep;
1845
1846                 rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1847
1848                 rep->lock_policy_res1 =
1849                         ptlrpc_status_ntoh(rep->lock_policy_res1);
1850                 if (rep->lock_policy_res1)
1851                         errcode = rep->lock_policy_res1;
1852                 if (!agl)
1853                         *flags |= LDLM_FL_LVB_READY;
1854         } else if (errcode == ELDLM_OK) {
1855                 *flags |= LDLM_FL_LVB_READY;
1856         }
1857
1858         /* Call the update callback. */
1859         rc = (*upcall)(cookie, lockh, errcode);
1860         /* release the reference taken in ldlm_cli_enqueue() */
1861         if (errcode == ELDLM_LOCK_MATCHED)
1862                 errcode = ELDLM_OK;
1863         if (errcode == ELDLM_OK && lustre_handle_is_used(lockh))
1864                 ldlm_lock_decref(lockh, mode);
1865
1866         return rc;
1867 }
1868
1869 static int osc_enqueue_interpret(const struct lu_env *env,
1870                                  struct ptlrpc_request *req,
1871                                  struct osc_enqueue_args *aa, int rc)
1872 {
1873         struct ldlm_lock *lock;
1874         struct lustre_handle *lockh = &aa->oa_lockh;
1875         enum ldlm_mode mode = aa->oa_mode;
1876         struct ost_lvb *lvb = aa->oa_lvb;
1877         __u32 lvb_len = sizeof(*lvb);
1878         __u64 flags = 0;
1879
1880
1881         /* ldlm_cli_enqueue is holding a reference on the lock, so it must
1882          * be valid.
1883          */
1884         lock = ldlm_handle2lock(lockh);
1885         LASSERTF(lock, "lockh %llx, req %p, aa %p - client evicted?\n",
1886                  lockh->cookie, req, aa);
1887
1888         /* Take an additional reference so that a blocking AST that
1889          * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
1890          * to arrive after an upcall has been executed by
1891          * osc_enqueue_fini().
1892          */
1893         ldlm_lock_addref(lockh, mode);
1894
1895         /* Let cl_lock_state_wait fail with -ERESTARTSYS to unuse sublocks. */
1896         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_HANG, 2);
1897
1898         /* Let CP AST to grant the lock first. */
1899         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
1900
1901         if (aa->oa_agl) {
1902                 LASSERT(!aa->oa_lvb);
1903                 LASSERT(!aa->oa_flags);
1904                 aa->oa_flags = &flags;
1905         }
1906
1907         /* Complete obtaining the lock procedure. */
1908         rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_type, 1,
1909                                    aa->oa_mode, aa->oa_flags, lvb, lvb_len,
1910                                    lockh, rc);
1911         /* Complete osc stuff. */
1912         rc = osc_enqueue_fini(req, aa->oa_upcall, aa->oa_cookie, lockh, mode,
1913                               aa->oa_flags, aa->oa_agl, rc);
1914
1915         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
1916
1917         ldlm_lock_decref(lockh, mode);
1918         LDLM_LOCK_PUT(lock);
1919         return rc;
1920 }
1921
1922 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
1923
1924 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
1925  * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
1926  * other synchronous requests, however keeping some locks and trying to obtain
1927  * others may take a considerable amount of time in a case of ost failure; and
1928  * when other sync requests do not get released lock from a client, the client
1929  * is evicted from the cluster -- such scenaries make the life difficult, so
1930  * release locks just after they are obtained.
1931  */
1932 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
1933                      __u64 *flags, union ldlm_policy_data *policy,
1934                      struct ost_lvb *lvb, int kms_valid,
1935                      osc_enqueue_upcall_f upcall, void *cookie,
1936                      struct ldlm_enqueue_info *einfo,
1937                      struct ptlrpc_request_set *rqset, int async, int agl)
1938 {
1939         struct obd_device *obd = exp->exp_obd;
1940         struct lustre_handle lockh = { 0 };
1941         struct ptlrpc_request *req = NULL;
1942         int intent = *flags & LDLM_FL_HAS_INTENT;
1943         __u64 match_flags = *flags;
1944         enum ldlm_mode mode;
1945         int rc;
1946
1947         /* Filesystem lock extents are extended to page boundaries so that
1948          * dealing with the page cache is a little smoother.
1949          */
1950         policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
1951         policy->l_extent.end |= ~PAGE_MASK;
1952
1953         /*
1954          * kms is not valid when either object is completely fresh (so that no
1955          * locks are cached), or object was evicted. In the latter case cached
1956          * lock cannot be used, because it would prime inode state with
1957          * potentially stale LVB.
1958          */
1959         if (!kms_valid)
1960                 goto no_match;
1961
1962         /* Next, search for already existing extent locks that will cover us */
1963         /* If we're trying to read, we also search for an existing PW lock.  The
1964          * VFS and page cache already protect us locally, so lots of readers/
1965          * writers can share a single PW lock.
1966          *
1967          * There are problems with conversion deadlocks, so instead of
1968          * converting a read lock to a write lock, we'll just enqueue a new
1969          * one.
1970          *
1971          * At some point we should cancel the read lock instead of making them
1972          * send us a blocking callback, but there are problems with canceling
1973          * locks out from other users right now, too.
1974          */
1975         mode = einfo->ei_mode;
1976         if (einfo->ei_mode == LCK_PR)
1977                 mode |= LCK_PW;
1978         if (agl == 0)
1979                 match_flags |= LDLM_FL_LVB_READY;
1980         if (intent != 0)
1981                 match_flags |= LDLM_FL_BLOCK_GRANTED;
1982         mode = ldlm_lock_match(obd->obd_namespace, match_flags, res_id,
1983                                einfo->ei_type, policy, mode, &lockh, 0);
1984         if (mode) {
1985                 struct ldlm_lock *matched;
1986
1987                 if (*flags & LDLM_FL_TEST_LOCK)
1988                         return ELDLM_OK;
1989
1990                 matched = ldlm_handle2lock(&lockh);
1991                 if (agl) {
1992                         /* AGL enqueues DLM locks speculatively. Therefore if
1993                          * it already exists a DLM lock, it wll just inform the
1994                          * caller to cancel the AGL process for this stripe.
1995                          */
1996                         ldlm_lock_decref(&lockh, mode);
1997                         LDLM_LOCK_PUT(matched);
1998                         return -ECANCELED;
1999                 } else if (osc_set_lock_data(matched, einfo->ei_cbdata)) {
2000                         *flags |= LDLM_FL_LVB_READY;
2001                         /* We already have a lock, and it's referenced. */
2002                         (*upcall)(cookie, &lockh, ELDLM_LOCK_MATCHED);
2003
2004                         ldlm_lock_decref(&lockh, mode);
2005                         LDLM_LOCK_PUT(matched);
2006                         return ELDLM_OK;
2007                 } else {
2008                         ldlm_lock_decref(&lockh, mode);
2009                         LDLM_LOCK_PUT(matched);
2010                 }
2011         }
2012
2013 no_match:
2014         if (*flags & (LDLM_FL_TEST_LOCK | LDLM_FL_MATCH_LOCK))
2015                 return -ENOLCK;
2016         if (intent) {
2017                 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2018                                            &RQF_LDLM_ENQUEUE_LVB);
2019                 if (!req)
2020                         return -ENOMEM;
2021
2022                 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
2023                 if (rc) {
2024                         ptlrpc_request_free(req);
2025                         return rc;
2026                 }
2027
2028                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2029                                      sizeof(*lvb));
2030                 ptlrpc_request_set_replen(req);
2031         }
2032
2033         /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2034         *flags &= ~LDLM_FL_BLOCK_GRANTED;
2035
2036         rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2037                               sizeof(*lvb), LVB_T_OST, &lockh, async);
2038         if (async) {
2039                 if (!rc) {
2040                         struct osc_enqueue_args *aa;
2041
2042                         BUILD_BUG_ON(sizeof(*aa) > sizeof(req->rq_async_args));
2043                         aa = ptlrpc_req_async_args(req);
2044                         aa->oa_exp = exp;
2045                         aa->oa_mode = einfo->ei_mode;
2046                         aa->oa_type = einfo->ei_type;
2047                         lustre_handle_copy(&aa->oa_lockh, &lockh);
2048                         aa->oa_upcall = upcall;
2049                         aa->oa_cookie = cookie;
2050                         aa->oa_agl    = !!agl;
2051                         if (!agl) {
2052                                 aa->oa_flags = flags;
2053                                 aa->oa_lvb = lvb;
2054                         } else {
2055                                 /* AGL is essentially to enqueue an DLM lock
2056                                 * in advance, so we don't care about the
2057                                 * result of AGL enqueue.
2058                                 */
2059                                 aa->oa_lvb = NULL;
2060                                 aa->oa_flags = NULL;
2061                         }
2062
2063                         req->rq_interpret_reply =
2064                                 (ptlrpc_interpterer_t)osc_enqueue_interpret;
2065                         if (rqset == PTLRPCD_SET)
2066                                 ptlrpcd_add_req(req);
2067                         else
2068                                 ptlrpc_set_add_req(rqset, req);
2069                 } else if (intent) {
2070                         ptlrpc_req_finished(req);
2071                 }
2072                 return rc;
2073         }
2074
2075         rc = osc_enqueue_fini(req, upcall, cookie, &lockh, einfo->ei_mode,
2076                               flags, agl, rc);
2077         if (intent)
2078                 ptlrpc_req_finished(req);
2079
2080         return rc;
2081 }
2082
2083 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2084                    enum ldlm_type type, union ldlm_policy_data *policy,
2085                    enum ldlm_mode mode, __u64 *flags, void *data,
2086                    struct lustre_handle *lockh, int unref)
2087 {
2088         struct obd_device *obd = exp->exp_obd;
2089         __u64 lflags = *flags;
2090         enum ldlm_mode rc;
2091
2092         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2093                 return -EIO;
2094
2095         /* Filesystem lock extents are extended to page boundaries so that
2096          * dealing with the page cache is a little smoother
2097          */
2098         policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2099         policy->l_extent.end |= ~PAGE_MASK;
2100
2101         /* Next, search for already existing extent locks that will cover us */
2102         /* If we're trying to read, we also search for an existing PW lock.  The
2103          * VFS and page cache already protect us locally, so lots of readers/
2104          * writers can share a single PW lock.
2105          */
2106         rc = mode;
2107         if (mode == LCK_PR)
2108                 rc |= LCK_PW;
2109         rc = ldlm_lock_match(obd->obd_namespace, lflags,
2110                              res_id, type, policy, rc, lockh, unref);
2111         if (!rc || lflags & LDLM_FL_TEST_LOCK)
2112                 return rc;
2113
2114         if (data) {
2115                 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2116
2117                 LASSERT(lock);
2118                 if (!osc_set_lock_data(lock, data)) {
2119                         ldlm_lock_decref(lockh, rc);
2120                         rc = 0;
2121                 }
2122                 LDLM_LOCK_PUT(lock);
2123         }
2124         return rc;
2125 }
2126
2127 static int osc_statfs_interpret(const struct lu_env *env,
2128                                 struct ptlrpc_request *req,
2129                                 struct osc_async_args *aa, int rc)
2130 {
2131         struct obd_statfs *msfs;
2132
2133         if (rc == -EBADR)
2134                 /* The request has in fact never been sent
2135                  * due to issues at a higher level (LOV).
2136                  * Exit immediately since the caller is
2137                  * aware of the problem and takes care
2138                  * of the clean up
2139                  */
2140                 return rc;
2141
2142         if ((rc == -ENOTCONN || rc == -EAGAIN) &&
2143             (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY)) {
2144                 rc = 0;
2145                 goto out;
2146         }
2147
2148         if (rc != 0)
2149                 goto out;
2150
2151         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2152         if (!msfs) {
2153                 rc = -EPROTO;
2154                 goto out;
2155         }
2156
2157         *aa->aa_oi->oi_osfs = *msfs;
2158 out:
2159         rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
2160         return rc;
2161 }
2162
2163 static int osc_statfs_async(struct obd_export *exp,
2164                             struct obd_info *oinfo, __u64 max_age,
2165                             struct ptlrpc_request_set *rqset)
2166 {
2167         struct obd_device *obd = class_exp2obd(exp);
2168         struct ptlrpc_request *req;
2169         struct osc_async_args *aa;
2170         int rc;
2171
2172         /* We could possibly pass max_age in the request (as an absolute
2173          * timestamp or a "seconds.usec ago") so the target can avoid doing
2174          * extra calls into the filesystem if that isn't necessary (e.g.
2175          * during mount that would help a bit).  Having relative timestamps
2176          * is not so great if request processing is slow, while absolute
2177          * timestamps are not ideal because they need time synchronization.
2178          */
2179         req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
2180         if (!req)
2181                 return -ENOMEM;
2182
2183         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2184         if (rc) {
2185                 ptlrpc_request_free(req);
2186                 return rc;
2187         }
2188         ptlrpc_request_set_replen(req);
2189         req->rq_request_portal = OST_CREATE_PORTAL;
2190         ptlrpc_at_set_req_timeout(req);
2191
2192         if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
2193                 /* procfs requests not want stat in wait for avoid deadlock */
2194                 req->rq_no_resend = 1;
2195                 req->rq_no_delay = 1;
2196         }
2197
2198         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
2199         BUILD_BUG_ON(sizeof(*aa) > sizeof(req->rq_async_args));
2200         aa = ptlrpc_req_async_args(req);
2201         aa->aa_oi = oinfo;
2202
2203         ptlrpc_set_add_req(rqset, req);
2204         return 0;
2205 }
2206
2207 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
2208                       struct obd_statfs *osfs, __u64 max_age, __u32 flags)
2209 {
2210         struct obd_device *obd = class_exp2obd(exp);
2211         struct obd_statfs *msfs;
2212         struct ptlrpc_request *req;
2213         struct obd_import *imp = NULL;
2214         int rc;
2215
2216         /* Since the request might also come from lprocfs, so we need
2217          * sync this with client_disconnect_export Bug15684
2218          */
2219         down_read(&obd->u.cli.cl_sem);
2220         if (obd->u.cli.cl_import)
2221                 imp = class_import_get(obd->u.cli.cl_import);
2222         up_read(&obd->u.cli.cl_sem);
2223         if (!imp)
2224                 return -ENODEV;
2225
2226         /* We could possibly pass max_age in the request (as an absolute
2227          * timestamp or a "seconds.usec ago") so the target can avoid doing
2228          * extra calls into the filesystem if that isn't necessary (e.g.
2229          * during mount that would help a bit).  Having relative timestamps
2230          * is not so great if request processing is slow, while absolute
2231          * timestamps are not ideal because they need time synchronization.
2232          */
2233         req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
2234
2235         class_import_put(imp);
2236
2237         if (!req)
2238                 return -ENOMEM;
2239
2240         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2241         if (rc) {
2242                 ptlrpc_request_free(req);
2243                 return rc;
2244         }
2245         ptlrpc_request_set_replen(req);
2246         req->rq_request_portal = OST_CREATE_PORTAL;
2247         ptlrpc_at_set_req_timeout(req);
2248
2249         if (flags & OBD_STATFS_NODELAY) {
2250                 /* procfs requests not want stat in wait for avoid deadlock */
2251                 req->rq_no_resend = 1;
2252                 req->rq_no_delay = 1;
2253         }
2254
2255         rc = ptlrpc_queue_wait(req);
2256         if (rc)
2257                 goto out;
2258
2259         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2260         if (!msfs) {
2261                 rc = -EPROTO;
2262                 goto out;
2263         }
2264
2265         *osfs = *msfs;
2266
2267  out:
2268         ptlrpc_req_finished(req);
2269         return rc;
2270 }
2271
2272 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2273                          void *karg, void __user *uarg)
2274 {
2275         struct obd_device *obd = exp->exp_obd;
2276         struct obd_ioctl_data *data = karg;
2277         int err = 0;
2278
2279         if (!try_module_get(THIS_MODULE)) {
2280                 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2281                        module_name(THIS_MODULE));
2282                 return -EINVAL;
2283         }
2284         switch (cmd) {
2285         case OBD_IOC_CLIENT_RECOVER:
2286                 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2287                                             data->ioc_inlbuf1, 0);
2288                 if (err > 0)
2289                         err = 0;
2290                 goto out;
2291         case IOC_OSC_SET_ACTIVE:
2292                 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2293                                                data->ioc_offset);
2294                 goto out;
2295         case OBD_IOC_PING_TARGET:
2296                 err = ptlrpc_obd_ping(obd);
2297                 goto out;
2298         default:
2299                 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
2300                        cmd, current_comm());
2301                 err = -ENOTTY;
2302                 goto out;
2303         }
2304 out:
2305         module_put(THIS_MODULE);
2306         return err;
2307 }
2308
2309 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
2310                               u32 keylen, void *key, u32 vallen,
2311                               void *val, struct ptlrpc_request_set *set)
2312 {
2313         struct ptlrpc_request *req;
2314         struct obd_device *obd = exp->exp_obd;
2315         struct obd_import *imp = class_exp2cliimp(exp);
2316         char *tmp;
2317         int rc;
2318
2319         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
2320
2321         if (KEY_IS(KEY_CHECKSUM)) {
2322                 if (vallen != sizeof(int))
2323                         return -EINVAL;
2324                 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
2325                 return 0;
2326         }
2327
2328         if (KEY_IS(KEY_SPTLRPC_CONF)) {
2329                 sptlrpc_conf_client_adapt(obd);
2330                 return 0;
2331         }
2332
2333         if (KEY_IS(KEY_FLUSH_CTX)) {
2334                 sptlrpc_import_flush_my_ctx(imp);
2335                 return 0;
2336         }
2337
2338         if (KEY_IS(KEY_CACHE_SET)) {
2339                 struct client_obd *cli = &obd->u.cli;
2340
2341                 LASSERT(!cli->cl_cache); /* only once */
2342                 cli->cl_cache = val;
2343                 cl_cache_incref(cli->cl_cache);
2344                 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
2345
2346                 /* add this osc into entity list */
2347                 LASSERT(list_empty(&cli->cl_lru_osc));
2348                 spin_lock(&cli->cl_cache->ccc_lru_lock);
2349                 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
2350                 spin_unlock(&cli->cl_cache->ccc_lru_lock);
2351
2352                 return 0;
2353         }
2354
2355         if (KEY_IS(KEY_CACHE_LRU_SHRINK)) {
2356                 struct client_obd *cli = &obd->u.cli;
2357                 long nr = atomic_long_read(&cli->cl_lru_in_list) >> 1;
2358                 long target = *(long *)val;
2359
2360                 nr = osc_lru_shrink(env, cli, min(nr, target), true);
2361                 *(long *)val -= nr;
2362                 return 0;
2363         }
2364
2365         if (!set && !KEY_IS(KEY_GRANT_SHRINK))
2366                 return -EINVAL;
2367
2368         /* We pass all other commands directly to OST. Since nobody calls osc
2369          * methods directly and everybody is supposed to go through LOV, we
2370          * assume lov checked invalid values for us.
2371          * The only recognised values so far are evict_by_nid and mds_conn.
2372          * Even if something bad goes through, we'd get a -EINVAL from OST
2373          * anyway.
2374          */
2375
2376         req = ptlrpc_request_alloc(imp, KEY_IS(KEY_GRANT_SHRINK) ?
2377                                                 &RQF_OST_SET_GRANT_INFO :
2378                                                 &RQF_OBD_SET_INFO);
2379         if (!req)
2380                 return -ENOMEM;
2381
2382         req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2383                              RCL_CLIENT, keylen);
2384         if (!KEY_IS(KEY_GRANT_SHRINK))
2385                 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
2386                                      RCL_CLIENT, vallen);
2387         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
2388         if (rc) {
2389                 ptlrpc_request_free(req);
2390                 return rc;
2391         }
2392
2393         tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2394         memcpy(tmp, key, keylen);
2395         tmp = req_capsule_client_get(&req->rq_pill, KEY_IS(KEY_GRANT_SHRINK) ?
2396                                                         &RMF_OST_BODY :
2397                                                         &RMF_SETINFO_VAL);
2398         memcpy(tmp, val, vallen);
2399
2400         if (KEY_IS(KEY_GRANT_SHRINK)) {
2401                 struct osc_brw_async_args *aa;
2402                 struct obdo *oa;
2403
2404                 BUILD_BUG_ON(sizeof(*aa) > sizeof(req->rq_async_args));
2405                 aa = ptlrpc_req_async_args(req);
2406                 oa = kmem_cache_zalloc(obdo_cachep, GFP_NOFS);
2407                 if (!oa) {
2408                         ptlrpc_req_finished(req);
2409                         return -ENOMEM;
2410                 }
2411                 *oa = ((struct ost_body *)val)->oa;
2412                 aa->aa_oa = oa;
2413                 req->rq_interpret_reply = osc_shrink_grant_interpret;
2414         }
2415
2416         ptlrpc_request_set_replen(req);
2417         if (!KEY_IS(KEY_GRANT_SHRINK)) {
2418                 LASSERT(set);
2419                 ptlrpc_set_add_req(set, req);
2420                 ptlrpc_check_set(NULL, set);
2421         } else {
2422                 ptlrpcd_add_req(req);
2423         }
2424
2425         return 0;
2426 }
2427
2428 static int osc_reconnect(const struct lu_env *env,
2429                          struct obd_export *exp, struct obd_device *obd,
2430                          struct obd_uuid *cluuid,
2431                          struct obd_connect_data *data,
2432                          void *localdata)
2433 {
2434         struct client_obd *cli = &obd->u.cli;
2435
2436         if (data && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
2437                 long lost_grant;
2438
2439                 spin_lock(&cli->cl_loi_list_lock);
2440                 data->ocd_grant = (cli->cl_avail_grant +
2441                                    (cli->cl_dirty_pages << PAGE_SHIFT)) ?:
2442                                    2 * cli_brw_size(obd);
2443                 lost_grant = cli->cl_lost_grant;
2444                 cli->cl_lost_grant = 0;
2445                 spin_unlock(&cli->cl_loi_list_lock);
2446
2447                 CDEBUG(D_RPCTRACE, "ocd_connect_flags: %#llx ocd_version: %d ocd_grant: %d, lost: %ld.\n",
2448                        data->ocd_connect_flags,
2449                        data->ocd_version, data->ocd_grant, lost_grant);
2450         }
2451
2452         return 0;
2453 }
2454
2455 static int osc_disconnect(struct obd_export *exp)
2456 {
2457         struct obd_device *obd = class_exp2obd(exp);
2458         int rc;
2459
2460         rc = client_disconnect_export(exp);
2461         /**
2462          * Initially we put del_shrink_grant before disconnect_export, but it
2463          * causes the following problem if setup (connect) and cleanup
2464          * (disconnect) are tangled together.
2465          *      connect p1                   disconnect p2
2466          *   ptlrpc_connect_import
2467          *     ...............         class_manual_cleanup
2468          *                                   osc_disconnect
2469          *                                   del_shrink_grant
2470          *   ptlrpc_connect_interrupt
2471          *     init_grant_shrink
2472          *   add this client to shrink list
2473          *                                    cleanup_osc
2474          * Bang! pinger trigger the shrink.
2475          * So the osc should be disconnected from the shrink list, after we
2476          * are sure the import has been destroyed. BUG18662
2477          */
2478         if (!obd->u.cli.cl_import)
2479                 osc_del_shrink_grant(&obd->u.cli);
2480         return rc;
2481 }
2482
2483 static int osc_ldlm_resource_invalidate(struct cfs_hash *hs,
2484                                         struct cfs_hash_bd *bd,
2485                                         struct hlist_node *hnode, void *arg)
2486 {
2487         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
2488         struct osc_object *osc = NULL;
2489         struct lu_env *env = arg;
2490         struct ldlm_lock *lock;
2491
2492         lock_res(res);
2493         list_for_each_entry(lock, &res->lr_granted, l_res_link) {
2494                 if (lock->l_ast_data && !osc) {
2495                         osc = lock->l_ast_data;
2496                         cl_object_get(osc2cl(osc));
2497                 }
2498
2499                 /*
2500                  * clear LDLM_FL_CLEANED flag to make sure it will be canceled
2501                  * by the 2nd round of ldlm_namespace_clean() call in
2502                  * osc_import_event().
2503                  */
2504                 ldlm_clear_cleaned(lock);
2505         }
2506         unlock_res(res);
2507
2508         if (osc) {
2509                 osc_object_invalidate(env, osc);
2510                 cl_object_put(env, osc2cl(osc));
2511         }
2512
2513         return 0;
2514 }
2515
2516 static int osc_import_event(struct obd_device *obd,
2517                             struct obd_import *imp,
2518                             enum obd_import_event event)
2519 {
2520         struct client_obd *cli;
2521         int rc = 0;
2522
2523         LASSERT(imp->imp_obd == obd);
2524
2525         switch (event) {
2526         case IMP_EVENT_DISCON: {
2527                 cli = &obd->u.cli;
2528                 spin_lock(&cli->cl_loi_list_lock);
2529                 cli->cl_avail_grant = 0;
2530                 cli->cl_lost_grant = 0;
2531                 spin_unlock(&cli->cl_loi_list_lock);
2532                 break;
2533         }
2534         case IMP_EVENT_INACTIVE: {
2535                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
2536                 break;
2537         }
2538         case IMP_EVENT_INVALIDATE: {
2539                 struct ldlm_namespace *ns = obd->obd_namespace;
2540                 struct lu_env *env;
2541                 u16 refcheck;
2542
2543                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2544
2545                 env = cl_env_get(&refcheck);
2546                 if (!IS_ERR(env)) {
2547                         osc_io_unplug(env, &obd->u.cli, NULL);
2548
2549                         cfs_hash_for_each_nolock(ns->ns_rs_hash,
2550                                                  osc_ldlm_resource_invalidate,
2551                                                  env, 0);
2552                         cl_env_put(env, &refcheck);
2553
2554                         ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2555                 } else {
2556                         rc = PTR_ERR(env);
2557                 }
2558                 break;
2559         }
2560         case IMP_EVENT_ACTIVE: {
2561                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
2562                 break;
2563         }
2564         case IMP_EVENT_OCD: {
2565                 struct obd_connect_data *ocd = &imp->imp_connect_data;
2566
2567                 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
2568                         osc_init_grant(&obd->u.cli, ocd);
2569
2570                 /* See bug 7198 */
2571                 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
2572                         imp->imp_client->cli_request_portal = OST_REQUEST_PORTAL;
2573
2574                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
2575                 break;
2576         }
2577         case IMP_EVENT_DEACTIVATE: {
2578                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE, NULL);
2579                 break;
2580         }
2581         case IMP_EVENT_ACTIVATE: {
2582                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE, NULL);
2583                 break;
2584         }
2585         default:
2586                 CERROR("Unknown import event %d\n", event);
2587                 LBUG();
2588         }
2589         return rc;
2590 }
2591
2592 /**
2593  * Determine whether the lock can be canceled before replaying the lock
2594  * during recovery, see bug16774 for detailed information.
2595  *
2596  * \retval zero the lock can't be canceled
2597  * \retval other ok to cancel
2598  */
2599 static int osc_cancel_weight(struct ldlm_lock *lock)
2600 {
2601         /*
2602          * Cancel all unused and granted extent lock.
2603          */
2604         if (lock->l_resource->lr_type == LDLM_EXTENT &&
2605             lock->l_granted_mode == lock->l_req_mode &&
2606             osc_ldlm_weigh_ast(lock) == 0)
2607                 return 1;
2608
2609         return 0;
2610 }
2611
2612 static int brw_queue_work(const struct lu_env *env, void *data)
2613 {
2614         struct client_obd *cli = data;
2615
2616         CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
2617
2618         osc_io_unplug(env, cli, NULL);
2619         return 0;
2620 }
2621
2622 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
2623 {
2624         struct lprocfs_static_vars lvars = { NULL };
2625         struct client_obd *cli = &obd->u.cli;
2626         void *handler;
2627         int rc;
2628         int adding;
2629         int added;
2630         int req_count;
2631
2632         rc = ptlrpcd_addref();
2633         if (rc)
2634                 return rc;
2635
2636         rc = client_obd_setup(obd, lcfg);
2637         if (rc)
2638                 goto out_ptlrpcd;
2639
2640         handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
2641         if (IS_ERR(handler)) {
2642                 rc = PTR_ERR(handler);
2643                 goto out_client_setup;
2644         }
2645         cli->cl_writeback_work = handler;
2646
2647         handler = ptlrpcd_alloc_work(cli->cl_import, lru_queue_work, cli);
2648         if (IS_ERR(handler)) {
2649                 rc = PTR_ERR(handler);
2650                 goto out_ptlrpcd_work;
2651         }
2652
2653         cli->cl_lru_work = handler;
2654
2655         rc = osc_quota_setup(obd);
2656         if (rc)
2657                 goto out_ptlrpcd_work;
2658
2659         cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
2660         lprocfs_osc_init_vars(&lvars);
2661         if (lprocfs_obd_setup(obd, lvars.obd_vars, lvars.sysfs_vars) == 0) {
2662                 lproc_osc_attach_seqstat(obd);
2663                 sptlrpc_lprocfs_cliobd_attach(obd);
2664                 ptlrpc_lprocfs_register_obd(obd);
2665         }
2666
2667         /*
2668          * We try to control the total number of requests with a upper limit
2669          * osc_reqpool_maxreqcount. There might be some race which will cause
2670          * over-limit allocation, but it is fine.
2671          */
2672         req_count = atomic_read(&osc_pool_req_count);
2673         if (req_count < osc_reqpool_maxreqcount) {
2674                 adding = cli->cl_max_rpcs_in_flight + 2;
2675                 if (req_count + adding > osc_reqpool_maxreqcount)
2676                         adding = osc_reqpool_maxreqcount - req_count;
2677
2678                 added = ptlrpc_add_rqs_to_pool(osc_rq_pool, adding);
2679                 atomic_add(added, &osc_pool_req_count);
2680         }
2681
2682         INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
2683         ns_register_cancel(obd->obd_namespace, osc_cancel_weight);
2684
2685         spin_lock(&osc_shrink_lock);
2686         list_add_tail(&cli->cl_shrink_list, &osc_shrink_list);
2687         spin_unlock(&osc_shrink_lock);
2688
2689         return rc;
2690
2691 out_ptlrpcd_work:
2692         if (cli->cl_writeback_work) {
2693                 ptlrpcd_destroy_work(cli->cl_writeback_work);
2694                 cli->cl_writeback_work = NULL;
2695         }
2696         if (cli->cl_lru_work) {
2697                 ptlrpcd_destroy_work(cli->cl_lru_work);
2698                 cli->cl_lru_work = NULL;
2699         }
2700 out_client_setup:
2701         client_obd_cleanup(obd);
2702 out_ptlrpcd:
2703         ptlrpcd_decref();
2704         return rc;
2705 }
2706
2707 static int osc_precleanup(struct obd_device *obd)
2708 {
2709         struct client_obd *cli = &obd->u.cli;
2710
2711         /* LU-464
2712          * for echo client, export may be on zombie list, wait for
2713          * zombie thread to cull it, because cli.cl_import will be
2714          * cleared in client_disconnect_export():
2715          *   class_export_destroy() -> obd_cleanup() ->
2716          *   echo_device_free() -> echo_client_cleanup() ->
2717          *   obd_disconnect() -> osc_disconnect() ->
2718          *   client_disconnect_export()
2719          */
2720         obd_zombie_barrier();
2721         if (cli->cl_writeback_work) {
2722                 ptlrpcd_destroy_work(cli->cl_writeback_work);
2723                 cli->cl_writeback_work = NULL;
2724         }
2725
2726         if (cli->cl_lru_work) {
2727                 ptlrpcd_destroy_work(cli->cl_lru_work);
2728                 cli->cl_lru_work = NULL;
2729         }
2730
2731         obd_cleanup_client_import(obd);
2732         ptlrpc_lprocfs_unregister_obd(obd);
2733         lprocfs_obd_cleanup(obd);
2734         return 0;
2735 }
2736
2737 static int osc_cleanup(struct obd_device *obd)
2738 {
2739         struct client_obd *cli = &obd->u.cli;
2740         int rc;
2741
2742         spin_lock(&osc_shrink_lock);
2743         list_del(&cli->cl_shrink_list);
2744         spin_unlock(&osc_shrink_lock);
2745
2746         /* lru cleanup */
2747         if (cli->cl_cache) {
2748                 LASSERT(atomic_read(&cli->cl_cache->ccc_users) > 0);
2749                 spin_lock(&cli->cl_cache->ccc_lru_lock);
2750                 list_del_init(&cli->cl_lru_osc);
2751                 spin_unlock(&cli->cl_cache->ccc_lru_lock);
2752                 cli->cl_lru_left = NULL;
2753                 cl_cache_decref(cli->cl_cache);
2754                 cli->cl_cache = NULL;
2755         }
2756
2757         /* free memory of osc quota cache */
2758         osc_quota_cleanup(obd);
2759
2760         rc = client_obd_cleanup(obd);
2761
2762         ptlrpcd_decref();
2763         return rc;
2764 }
2765
2766 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
2767 {
2768         struct lprocfs_static_vars lvars = { NULL };
2769         int rc = 0;
2770
2771         lprocfs_osc_init_vars(&lvars);
2772
2773         switch (lcfg->lcfg_command) {
2774         default:
2775                 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
2776                                               lcfg, obd);
2777                 if (rc > 0)
2778                         rc = 0;
2779                 break;
2780         }
2781
2782         return rc;
2783 }
2784
2785 static int osc_process_config(struct obd_device *obd, u32 len, void *buf)
2786 {
2787         return osc_process_config_base(obd, buf);
2788 }
2789
2790 static struct obd_ops osc_obd_ops = {
2791         .owner          = THIS_MODULE,
2792         .setup          = osc_setup,
2793         .precleanup     = osc_precleanup,
2794         .cleanup        = osc_cleanup,
2795         .add_conn       = client_import_add_conn,
2796         .del_conn       = client_import_del_conn,
2797         .connect        = client_connect_import,
2798         .reconnect      = osc_reconnect,
2799         .disconnect     = osc_disconnect,
2800         .statfs         = osc_statfs,
2801         .statfs_async   = osc_statfs_async,
2802         .create         = osc_create,
2803         .destroy        = osc_destroy,
2804         .getattr        = osc_getattr,
2805         .setattr        = osc_setattr,
2806         .iocontrol      = osc_iocontrol,
2807         .set_info_async = osc_set_info_async,
2808         .import_event   = osc_import_event,
2809         .process_config = osc_process_config,
2810         .quotactl       = osc_quotactl,
2811 };
2812
2813 struct list_head osc_shrink_list = LIST_HEAD_INIT(osc_shrink_list);
2814 DEFINE_SPINLOCK(osc_shrink_lock);
2815
2816 static struct shrinker osc_cache_shrinker = {
2817         .count_objects  = osc_cache_shrink_count,
2818         .scan_objects   = osc_cache_shrink_scan,
2819         .seeks          = DEFAULT_SEEKS,
2820 };
2821
2822 static int __init osc_init(void)
2823 {
2824         struct lprocfs_static_vars lvars = { NULL };
2825         unsigned int reqpool_size;
2826         unsigned int reqsize;
2827         int rc;
2828
2829         /* print an address of _any_ initialized kernel symbol from this
2830          * module, to allow debugging with gdb that doesn't support data
2831          * symbols from modules.
2832          */
2833         CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
2834
2835         rc = lu_kmem_init(osc_caches);
2836         if (rc)
2837                 return rc;
2838
2839         lprocfs_osc_init_vars(&lvars);
2840
2841         rc = class_register_type(&osc_obd_ops, NULL,
2842                                  LUSTRE_OSC_NAME, &osc_device_type);
2843         if (rc)
2844                 goto out_kmem;
2845
2846         register_shrinker(&osc_cache_shrinker);
2847
2848         /* This is obviously too much memory, only prevent overflow here */
2849         if (osc_reqpool_mem_max >= 1 << 12 || osc_reqpool_mem_max == 0) {
2850                 rc = -EINVAL;
2851                 goto out_type;
2852         }
2853
2854         reqpool_size = osc_reqpool_mem_max << 20;
2855
2856         reqsize = 1;
2857         while (reqsize < OST_MAXREQSIZE)
2858                 reqsize = reqsize << 1;
2859
2860         /*
2861          * We don't enlarge the request count in OSC pool according to
2862          * cl_max_rpcs_in_flight. The allocation from the pool will only be
2863          * tried after normal allocation failed. So a small OSC pool won't
2864          * cause much performance degression in most of cases.
2865          */
2866         osc_reqpool_maxreqcount = reqpool_size / reqsize;
2867
2868         atomic_set(&osc_pool_req_count, 0);
2869         osc_rq_pool = ptlrpc_init_rq_pool(0, OST_MAXREQSIZE,
2870                                           ptlrpc_add_rqs_to_pool);
2871
2872         if (osc_rq_pool)
2873                 return 0;
2874
2875         rc = -ENOMEM;
2876
2877 out_type:
2878         class_unregister_type(LUSTRE_OSC_NAME);
2879 out_kmem:
2880         lu_kmem_fini(osc_caches);
2881         return rc;
2882 }
2883
2884 static void /*__exit*/ osc_exit(void)
2885 {
2886         unregister_shrinker(&osc_cache_shrinker);
2887         class_unregister_type(LUSTRE_OSC_NAME);
2888         lu_kmem_fini(osc_caches);
2889         ptlrpc_free_rq_pool(osc_rq_pool);
2890 }
2891
2892 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2893 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
2894 MODULE_LICENSE("GPL");
2895 MODULE_VERSION(LUSTRE_VERSION_STRING);
2896
2897 module_init(osc_init);
2898 module_exit(osc_exit);