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, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  */
  31
  32 #define DEBUG_SUBSYSTEM S_OSC
  33
  34 #include <linux/workqueue.h>
  35 #include <libcfs/libcfs.h>
  36 #include <linux/falloc.h>
  37 #include <lprocfs_status.h>
  38 #include <lustre_dlm.h>
  39 #include <lustre_fid.h>
  40 #include <lustre_ha.h>
  41 #include <uapi/linux/lustre/lustre_ioctl.h>
  42 #include <lustre_net.h>
  43 #include <lustre_obdo.h>
  44 #include <obd.h>
  45 #include <obd_cksum.h>
  46 #include <obd_class.h>
  47 #include <lustre_osc.h>
  48 #include <linux/falloc.h>
  49
  50 #include "osc_internal.h"
  51 #include <lnet/lnet_rdma.h>
  52
  53 atomic_t osc_pool_req_count;
  54 unsigned int osc_reqpool_maxreqcount;
  55 struct ptlrpc_request_pool *osc_rq_pool;
  56
  57 /* max memory used for request pool, unit is MB */
  58 static unsigned int osc_reqpool_mem_max = 5;
  59 module_param(osc_reqpool_mem_max, uint, 0444);
  60
  61 static int osc_idle_timeout = 20;
  62 module_param(osc_idle_timeout, uint, 0644);
  63
  64 #define osc_grant_args osc_brw_async_args
  65
  66 struct osc_setattr_args {
  67         struct obdo             *sa_oa;
  68         obd_enqueue_update_f     sa_upcall;
  69         void                    *sa_cookie;
  70 };
  71
  72 struct osc_fsync_args {
  73         struct osc_object       *fa_obj;
  74         struct obdo             *fa_oa;
  75         obd_enqueue_update_f    fa_upcall;
  76         void                    *fa_cookie;
  77 };
  78
  79 struct osc_ladvise_args {
  80         struct obdo             *la_oa;
  81         obd_enqueue_update_f     la_upcall;
  82         void                    *la_cookie;
  83 };
  84
  85 static void osc_release_ppga(struct brw_page **ppga, size_t count);
  86 static int brw_interpret(const struct lu_env *env, struct ptlrpc_request *req,
  87                          void *data, int rc);
  88
  89 void osc_pack_req_body(struct ptlrpc_request *req, struct obdo *oa)
  90 {
  91         struct ost_body *body;
  92
  93         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
  94         LASSERT(body);
  95
  96         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
  97 }
  98
  99 static int osc_getattr(const struct lu_env *env, struct obd_export *exp,
 100                        struct obdo *oa)
 101 {
 102         struct ptlrpc_request   *req;
 103         struct ost_body         *body;
 104         int                      rc;
 105
 106         ENTRY;
 107         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
 108         if (req == NULL)
 109                 RETURN(-ENOMEM);
 110
 111         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
 112         if (rc) {
 113                 ptlrpc_request_free(req);
 114                 RETURN(rc);
 115         }
 116
 117         osc_pack_req_body(req, oa);
 118
 119         ptlrpc_request_set_replen(req);
 120
 121         rc = ptlrpc_queue_wait(req);
 122         if (rc)
 123                 GOTO(out, rc);
 124
 125         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 126         if (body == NULL)
 127                 GOTO(out, rc = -EPROTO);
 128
 129         CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
 130         lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
 131
 132         oa->o_blksize = cli_brw_size(exp->exp_obd);
 133         oa->o_valid |= OBD_MD_FLBLKSZ;
 134
 135         EXIT;
 136 out:
 137         ptlrpc_req_finished(req);
 138
 139         return rc;
 140 }
 141
 142 static int osc_setattr(const struct lu_env *env, struct obd_export *exp,
 143                        struct obdo *oa)
 144 {
 145         struct ptlrpc_request   *req;
 146         struct ost_body         *body;
 147         int                      rc;
 148
 149         ENTRY;
 150         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
 151
 152         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
 153         if (req == NULL)
 154                 RETURN(-ENOMEM);
 155
 156         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
 157         if (rc) {
 158                 ptlrpc_request_free(req);
 159                 RETURN(rc);
 160         }
 161
 162         osc_pack_req_body(req, oa);
 163
 164         ptlrpc_request_set_replen(req);
 165
 166         rc = ptlrpc_queue_wait(req);
 167         if (rc)
 168                 GOTO(out, rc);
 169
 170         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 171         if (body == NULL)
 172                 GOTO(out, rc = -EPROTO);
 173
 174         lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
 175
 176         EXIT;
 177 out:
 178         ptlrpc_req_finished(req);
 179
 180         RETURN(rc);
 181 }
 182
 183 static int osc_setattr_interpret(const struct lu_env *env,
 184                                  struct ptlrpc_request *req, void *args, int rc)
 185 {
 186         struct osc_setattr_args *sa = args;
 187         struct ost_body *body;
 188
 189         ENTRY;
 190
 191         if (rc != 0)
 192                 GOTO(out, rc);
 193
 194         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 195         if (body == NULL)
 196                 GOTO(out, rc = -EPROTO);
 197
 198         lustre_get_wire_obdo(&req->rq_import->imp_connect_data, sa->sa_oa,
 199                              &body->oa);
 200 out:
 201         rc = sa->sa_upcall(sa->sa_cookie, rc);
 202         RETURN(rc);
 203 }
 204
 205 int osc_setattr_async(struct obd_export *exp, struct obdo *oa,
 206                       obd_enqueue_update_f upcall, void *cookie,
 207                       struct ptlrpc_request_set *rqset)
 208 {
 209         struct ptlrpc_request   *req;
 210         struct osc_setattr_args *sa;
 211         int                      rc;
 212
 213         ENTRY;
 214
 215         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
 216         if (req == NULL)
 217                 RETURN(-ENOMEM);
 218
 219         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
 220         if (rc) {
 221                 ptlrpc_request_free(req);
 222                 RETURN(rc);
 223         }
 224
 225         osc_pack_req_body(req, oa);
 226
 227         ptlrpc_request_set_replen(req);
 228
 229         /* do mds to ost setattr asynchronously */
 230         if (!rqset) {
 231                 /* Do not wait for response. */
 232                 ptlrpcd_add_req(req);
 233         } else {
 234                 req->rq_interpret_reply = osc_setattr_interpret;
 235
 236                 sa = ptlrpc_req_async_args(sa, req);
 237                 sa->sa_oa = oa;
 238                 sa->sa_upcall = upcall;
 239                 sa->sa_cookie = cookie;
 240
 241                 ptlrpc_set_add_req(rqset, req);
 242         }
 243
 244         RETURN(0);
 245 }
 246
 247 static int osc_ladvise_interpret(const struct lu_env *env,
 248                                  struct ptlrpc_request *req,
 249                                  void *arg, int rc)
 250 {
 251         struct osc_ladvise_args *la = arg;
 252         struct ost_body *body;
 253         ENTRY;
 254
 255         if (rc != 0)
 256                 GOTO(out, rc);
 257
 258         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 259         if (body == NULL)
 260                 GOTO(out, rc = -EPROTO);
 261
 262         *la->la_oa = body->oa;
 263 out:
 264         rc = la->la_upcall(la->la_cookie, rc);
 265         RETURN(rc);
 266 }
 267
 268 /**
 269  * If rqset is NULL, do not wait for response. Upcall and cookie could also
 270  * be NULL in this case
 271  */
 272 int osc_ladvise_base(struct obd_export *exp, struct obdo *oa,
 273                      struct ladvise_hdr *ladvise_hdr,
 274                      obd_enqueue_update_f upcall, void *cookie,
 275                      struct ptlrpc_request_set *rqset)
 276 {
 277         struct ptlrpc_request   *req;
 278         struct ost_body         *body;
 279         struct osc_ladvise_args *la;
 280         int                      rc;
 281         struct lu_ladvise       *req_ladvise;
 282         struct lu_ladvise       *ladvise = ladvise_hdr->lah_advise;
 283         int                      num_advise = ladvise_hdr->lah_count;
 284         struct ladvise_hdr      *req_ladvise_hdr;
 285         ENTRY;
 286
 287         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_LADVISE);
 288         if (req == NULL)
 289                 RETURN(-ENOMEM);
 290
 291         req_capsule_set_size(&req->rq_pill, &RMF_OST_LADVISE, RCL_CLIENT,
 292                              num_advise * sizeof(*ladvise));
 293         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_LADVISE);
 294         if (rc != 0) {
 295                 ptlrpc_request_free(req);
 296                 RETURN(rc);
 297         }
 298         req->rq_request_portal = OST_IO_PORTAL;
 299         ptlrpc_at_set_req_timeout(req);
 300
 301         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 302         LASSERT(body);
 303         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
 304                              oa);
 305
 306         req_ladvise_hdr = req_capsule_client_get(&req->rq_pill,
 307                                                  &RMF_OST_LADVISE_HDR);
 308         memcpy(req_ladvise_hdr, ladvise_hdr, sizeof(*ladvise_hdr));
 309
 310         req_ladvise = req_capsule_client_get(&req->rq_pill, &RMF_OST_LADVISE);
 311         memcpy(req_ladvise, ladvise, sizeof(*ladvise) * num_advise);
 312         ptlrpc_request_set_replen(req);
 313
 314         if (rqset == NULL) {
 315                 /* Do not wait for response. */
 316                 ptlrpcd_add_req(req);
 317                 RETURN(0);
 318         }
 319
 320         req->rq_interpret_reply = osc_ladvise_interpret;
 321         la = ptlrpc_req_async_args(la, req);
 322         la->la_oa = oa;
 323         la->la_upcall = upcall;
 324         la->la_cookie = cookie;
 325
 326         ptlrpc_set_add_req(rqset, req);
 327
 328         RETURN(0);
 329 }
 330
 331 static int osc_create(const struct lu_env *env, struct obd_export *exp,
 332                       struct obdo *oa)
 333 {
 334         struct ptlrpc_request *req;
 335         struct ost_body       *body;
 336         int                    rc;
 337         ENTRY;
 338
 339         LASSERT(oa != NULL);
 340         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
 341         LASSERT(fid_seq_is_echo(ostid_seq(&oa->o_oi)));
 342
 343         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
 344         if (req == NULL)
 345                 GOTO(out, rc = -ENOMEM);
 346
 347         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
 348         if (rc) {
 349                 ptlrpc_request_free(req);
 350                 GOTO(out, rc);
 351         }
 352
 353         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 354         LASSERT(body);
 355
 356         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
 357
 358         ptlrpc_request_set_replen(req);
 359
 360         rc = ptlrpc_queue_wait(req);
 361         if (rc)
 362                 GOTO(out_req, rc);
 363
 364         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 365         if (body == NULL)
 366                 GOTO(out_req, rc = -EPROTO);
 367
 368         CDEBUG(D_INFO, "oa flags %x\n", oa->o_flags);
 369         lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
 370
 371         oa->o_blksize = cli_brw_size(exp->exp_obd);
 372         oa->o_valid |= OBD_MD_FLBLKSZ;
 373
 374         CDEBUG(D_HA, "transno: %lld\n",
 375                lustre_msg_get_transno(req->rq_repmsg));
 376 out_req:
 377         ptlrpc_req_finished(req);
 378 out:
 379         RETURN(rc);
 380 }
 381
 382 int osc_punch_send(struct obd_export *exp, struct obdo *oa,
 383                    obd_enqueue_update_f upcall, void *cookie)
 384 {
 385         struct ptlrpc_request *req;
 386         struct osc_setattr_args *sa;
 387         struct obd_import *imp = class_exp2cliimp(exp);
 388         struct ost_body *body;
 389         int rc;
 390
 391         ENTRY;
 392
 393         req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
 394         if (req == NULL)
 395                 RETURN(-ENOMEM);
 396
 397         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
 398         if (rc < 0) {
 399                 ptlrpc_request_free(req);
 400                 RETURN(rc);
 401         }
 402
 403         osc_set_io_portal(req);
 404
 405         ptlrpc_at_set_req_timeout(req);
 406
 407         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 408
 409         lustre_set_wire_obdo(&imp->imp_connect_data, &body->oa, oa);
 410
 411         ptlrpc_request_set_replen(req);
 412
 413         req->rq_interpret_reply = osc_setattr_interpret;
 414         sa = ptlrpc_req_async_args(sa, req);
 415         sa->sa_oa = oa;
 416         sa->sa_upcall = upcall;
 417         sa->sa_cookie = cookie;
 418
 419         ptlrpcd_add_req(req);
 420
 421         RETURN(0);
 422 }
 423 EXPORT_SYMBOL(osc_punch_send);
 424
 425 /**
 426  * osc_fallocate_base() - Handles fallocate request.
 427  *
 428  * @exp:        Export structure
 429  * @oa:         Attributes passed to OSS from client (obdo structure)
 430  * @upcall:     Primary & supplementary group information
 431  * @cookie:     Exclusive identifier
 432  * @rqset:      Request list.
 433  * @mode:       Operation done on given range.
 434  *
 435  * osc_fallocate_base() - Handles fallocate requests only. Only block
 436  * allocation or standard preallocate operation is supported currently.
 437  * Other mode flags is not supported yet. ftruncate(2) or truncate(2)
 438  * is supported via SETATTR request.
 439  *
 440  * Return: Non-zero on failure and O on success.
 441  */
 442 int osc_fallocate_base(struct obd_export *exp, struct obdo *oa,
 443                        obd_enqueue_update_f upcall, void *cookie, int mode)
 444 {
 445         struct ptlrpc_request *req;
 446         struct osc_setattr_args *sa;
 447         struct ost_body *body;
 448         struct obd_import *imp = class_exp2cliimp(exp);
 449         int rc;
 450         ENTRY;
 451
 452         oa->o_falloc_mode = mode;
 453         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 454                                    &RQF_OST_FALLOCATE);
 455         if (req == NULL)
 456                 RETURN(-ENOMEM);
 457
 458         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_FALLOCATE);
 459         if (rc != 0) {
 460                 ptlrpc_request_free(req);
 461                 RETURN(rc);
 462         }
 463
 464         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 465         LASSERT(body);
 466
 467         lustre_set_wire_obdo(&imp->imp_connect_data, &body->oa, oa);
 468
 469         ptlrpc_request_set_replen(req);
 470
 471         req->rq_interpret_reply = osc_setattr_interpret;
 472         BUILD_BUG_ON(sizeof(*sa) > sizeof(req->rq_async_args));
 473         sa = ptlrpc_req_async_args(sa, req);
 474         sa->sa_oa = oa;
 475         sa->sa_upcall = upcall;
 476         sa->sa_cookie = cookie;
 477
 478         ptlrpcd_add_req(req);
 479
 480         RETURN(0);
 481 }
 482 EXPORT_SYMBOL(osc_fallocate_base);
 483
 484 static int osc_sync_interpret(const struct lu_env *env,
 485                               struct ptlrpc_request *req, void *args, int rc)
 486 {
 487         struct osc_fsync_args *fa = args;
 488         struct ost_body *body;
 489         struct cl_attr *attr = &osc_env_info(env)->oti_attr;
 490         unsigned long valid = 0;
 491         struct cl_object *obj;
 492         ENTRY;
 493
 494         if (rc != 0)
 495                 GOTO(out, rc);
 496
 497         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 498         if (body == NULL) {
 499                 CERROR("can't unpack ost_body\n");
 500                 GOTO(out, rc = -EPROTO);
 501         }
 502
 503         *fa->fa_oa = body->oa;
 504         obj = osc2cl(fa->fa_obj);
 505
 506         /* Update osc object's blocks attribute */
 507         cl_object_attr_lock(obj);
 508         if (body->oa.o_valid & OBD_MD_FLBLOCKS) {
 509                 attr->cat_blocks = body->oa.o_blocks;
 510                 valid |= CAT_BLOCKS;
 511         }
 512
 513         if (valid != 0)
 514                 cl_object_attr_update(env, obj, attr, valid);
 515         cl_object_attr_unlock(obj);
 516
 517 out:
 518         rc = fa->fa_upcall(fa->fa_cookie, rc);
 519         RETURN(rc);
 520 }
 521
 522 int osc_sync_base(struct osc_object *obj, struct obdo *oa,
 523                   obd_enqueue_update_f upcall, void *cookie,
 524                   struct ptlrpc_request_set *rqset)
 525 {
 526         struct obd_export     *exp = osc_export(obj);
 527         struct ptlrpc_request *req;
 528         struct ost_body       *body;
 529         struct osc_fsync_args *fa;
 530         int                    rc;
 531         ENTRY;
 532
 533         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
 534         if (req == NULL)
 535                 RETURN(-ENOMEM);
 536
 537         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
 538         if (rc) {
 539                 ptlrpc_request_free(req);
 540                 RETURN(rc);
 541         }
 542
 543         /* overload the size and blocks fields in the oa with start/end */
 544         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 545         LASSERT(body);
 546         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
 547
 548         ptlrpc_request_set_replen(req);
 549         req->rq_interpret_reply = osc_sync_interpret;
 550
 551         fa = ptlrpc_req_async_args(fa, req);
 552         fa->fa_obj = obj;
 553         fa->fa_oa = oa;
 554         fa->fa_upcall = upcall;
 555         fa->fa_cookie = cookie;
 556
 557         ptlrpc_set_add_req(rqset, req);
 558
 559         RETURN (0);
 560 }
 561
 562 /* Find and cancel locally locks matched by @mode in the resource found by
 563  * @objid. Found locks are added into @cancel list. Returns the amount of
 564  * locks added to @cancels list. */
 565 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
 566                                    struct list_head *cancels,
 567                                    enum ldlm_mode mode, __u64 lock_flags)
 568 {
 569         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 570         struct ldlm_res_id res_id;
 571         struct ldlm_resource *res;
 572         int count;
 573         ENTRY;
 574
 575         /* Return, i.e. cancel nothing, only if ELC is supported (flag in
 576          * export) but disabled through procfs (flag in NS).
 577          *
 578          * This distinguishes from a case when ELC is not supported originally,
 579          * when we still want to cancel locks in advance and just cancel them
 580          * locally, without sending any RPC. */
 581         if (exp_connect_cancelset(exp) && !ns_connect_cancelset(ns))
 582                 RETURN(0);
 583
 584         ostid_build_res_name(&oa->o_oi, &res_id);
 585         res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
 586         if (IS_ERR(res))
 587                 RETURN(0);
 588
 589         LDLM_RESOURCE_ADDREF(res);
 590         count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
 591                                            lock_flags, 0, NULL);
 592         LDLM_RESOURCE_DELREF(res);
 593         ldlm_resource_putref(res);
 594         RETURN(count);
 595 }
 596
 597 static int osc_destroy_interpret(const struct lu_env *env,
 598                                  struct ptlrpc_request *req, void *args, int rc)
 599 {
 600         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
 601
 602         atomic_dec(&cli->cl_destroy_in_flight);
 603         wake_up(&cli->cl_destroy_waitq);
 604
 605         return 0;
 606 }
 607
 608 static int osc_can_send_destroy(struct client_obd *cli)
 609 {
 610         if (atomic_inc_return(&cli->cl_destroy_in_flight) <=
 611             cli->cl_max_rpcs_in_flight) {
 612                 /* The destroy request can be sent */
 613                 return 1;
 614         }
 615         if (atomic_dec_return(&cli->cl_destroy_in_flight) <
 616             cli->cl_max_rpcs_in_flight) {
 617                 /*
 618                  * The counter has been modified between the two atomic
 619                  * operations.
 620                  */
 621                 wake_up(&cli->cl_destroy_waitq);
 622         }
 623         return 0;
 624 }
 625
 626 static int osc_destroy(const struct lu_env *env, struct obd_export *exp,
 627                        struct obdo *oa)
 628 {
 629         struct client_obd     *cli = &exp->exp_obd->u.cli;
 630         struct ptlrpc_request *req;
 631         struct ost_body       *body;
 632         LIST_HEAD(cancels);
 633         int rc, count;
 634         ENTRY;
 635
 636         if (!oa) {
 637                 CDEBUG(D_INFO, "oa NULL\n");
 638                 RETURN(-EINVAL);
 639         }
 640
 641         count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
 642                                         LDLM_FL_DISCARD_DATA);
 643
 644         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
 645         if (req == NULL) {
 646                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
 647                 RETURN(-ENOMEM);
 648         }
 649
 650         rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
 651                                0, &cancels, count);
 652         if (rc) {
 653                 ptlrpc_request_free(req);
 654                 RETURN(rc);
 655         }
 656
 657         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
 658         ptlrpc_at_set_req_timeout(req);
 659
 660         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 661         LASSERT(body);
 662         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
 663
 664         ptlrpc_request_set_replen(req);
 665
 666         req->rq_interpret_reply = osc_destroy_interpret;
 667         if (!osc_can_send_destroy(cli)) {
 668                 /*
 669                  * Wait until the number of on-going destroy RPCs drops
 670                  * under max_rpc_in_flight
 671                  */
 672                 rc = l_wait_event_abortable_exclusive(
 673                         cli->cl_destroy_waitq,
 674                         osc_can_send_destroy(cli));
 675                 if (rc) {
 676                         ptlrpc_req_finished(req);
 677                         RETURN(-EINTR);
 678                 }
 679         }
 680
 681         /* Do not wait for response */
 682         ptlrpcd_add_req(req);
 683         RETURN(0);
 684 }
 685
 686 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
 687                                 long writing_bytes)
 688 {
 689         u64 bits = OBD_MD_FLBLOCKS | OBD_MD_FLGRANT;
 690
 691         LASSERT(!(oa->o_valid & bits));
 692
 693         oa->o_valid |= bits;
 694         spin_lock(&cli->cl_loi_list_lock);
 695         if (cli->cl_ocd_grant_param)
 696                 oa->o_dirty = cli->cl_dirty_grant;
 697         else
 698                 oa->o_dirty = cli->cl_dirty_pages << PAGE_SHIFT;
 699         if (unlikely(cli->cl_dirty_pages > cli->cl_dirty_max_pages)) {
 700                 CERROR("dirty %lu > dirty_max %lu\n",
 701                        cli->cl_dirty_pages,
 702                        cli->cl_dirty_max_pages);
 703                 oa->o_undirty = 0;
 704         } else if (unlikely(atomic_long_read(&obd_dirty_pages) >
 705                             (long)(obd_max_dirty_pages + 1))) {
 706                 /* The atomic_read() allowing the atomic_inc() are
 707                  * not covered by a lock thus they may safely race and trip
 708                  * this CERROR() unless we add in a small fudge factor (+1). */
 709                 CERROR("%s: dirty %ld > system dirty_max %ld\n",
 710                        cli_name(cli), atomic_long_read(&obd_dirty_pages),
 711                        obd_max_dirty_pages);
 712                 oa->o_undirty = 0;
 713         } else if (unlikely(cli->cl_dirty_max_pages - cli->cl_dirty_pages >
 714                             0x7fffffff)) {
 715                 CERROR("dirty %lu - dirty_max %lu too big???\n",
 716                        cli->cl_dirty_pages, cli->cl_dirty_max_pages);
 717                 oa->o_undirty = 0;
 718         } else {
 719                 unsigned long nrpages;
 720                 unsigned long undirty;
 721
 722                 nrpages = cli->cl_max_pages_per_rpc;
 723                 nrpages *= cli->cl_max_rpcs_in_flight + 1;
 724                 nrpages = max(nrpages, cli->cl_dirty_max_pages);
 725                 undirty = nrpages << PAGE_SHIFT;
 726                 if (cli->cl_ocd_grant_param) {
 727                         int nrextents;
 728
 729                         /* take extent tax into account when asking for more
 730                          * grant space */
 731                         nrextents = (nrpages + cli->cl_max_extent_pages - 1) /
 732                                      cli->cl_max_extent_pages;
 733                         undirty += nrextents * cli->cl_grant_extent_tax;
 734                 }
 735                 /* Do not ask for more than OBD_MAX_GRANT - a margin for server
 736                  * to add extent tax, etc.
 737                  */
 738                 oa->o_undirty = min(undirty, OBD_MAX_GRANT &
 739                                     ~(PTLRPC_MAX_BRW_SIZE * 4UL));
 740         }
 741         oa->o_grant = cli->cl_avail_grant + cli->cl_reserved_grant;
 742         /* o_dropped AKA o_misc is 32 bits, but cl_lost_grant is 64 bits */
 743         if (cli->cl_lost_grant > INT_MAX) {
 744                 CDEBUG(D_CACHE,
 745                       "%s: avoided o_dropped overflow: cl_lost_grant %lu\n",
 746                       cli_name(cli), cli->cl_lost_grant);
 747                 oa->o_dropped = INT_MAX;
 748         } else {
 749                 oa->o_dropped = cli->cl_lost_grant;
 750         }
 751         cli->cl_lost_grant -= oa->o_dropped;
 752         spin_unlock(&cli->cl_loi_list_lock);
 753         CDEBUG(D_CACHE, "%s: dirty: %llu undirty: %u dropped %u grant: %llu"
 754                " cl_lost_grant %lu\n", cli_name(cli), oa->o_dirty,
 755                oa->o_undirty, oa->o_dropped, oa->o_grant, cli->cl_lost_grant);
 756 }
 757
 758 void osc_update_next_shrink(struct client_obd *cli)
 759 {
 760         cli->cl_next_shrink_grant = ktime_get_seconds() +
 761                                     cli->cl_grant_shrink_interval;
 762
 763         CDEBUG(D_CACHE, "next time %lld to shrink grant\n",
 764                cli->cl_next_shrink_grant);
 765 }
 766 EXPORT_SYMBOL(osc_update_next_shrink);
 767
 768 static void __osc_update_grant(struct client_obd *cli, u64 grant)
 769 {
 770         spin_lock(&cli->cl_loi_list_lock);
 771         cli->cl_avail_grant += grant;
 772         spin_unlock(&cli->cl_loi_list_lock);
 773 }
 774
 775 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
 776 {
 777         if (body->oa.o_valid & OBD_MD_FLGRANT) {
 778                 CDEBUG(D_CACHE, "got %llu extra grant\n", body->oa.o_grant);
 779                 __osc_update_grant(cli, body->oa.o_grant);
 780         }
 781 }
 782
 783 /**
 784  * grant thread data for shrinking space.
 785  */
 786 struct grant_thread_data {
 787         struct list_head        gtd_clients;
 788         struct mutex            gtd_mutex;
 789         unsigned long           gtd_stopped:1;
 790 };
 791 static struct grant_thread_data client_gtd;
 792
 793 static int osc_shrink_grant_interpret(const struct lu_env *env,
 794                                       struct ptlrpc_request *req,
 795                                       void *args, int rc)
 796 {
 797         struct osc_grant_args *aa = args;
 798         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
 799         struct ost_body *body;
 800
 801         if (rc != 0) {
 802                 __osc_update_grant(cli, aa->aa_oa->o_grant);
 803                 GOTO(out, rc);
 804         }
 805
 806         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 807         LASSERT(body);
 808         osc_update_grant(cli, body);
 809 out:
 810         OBD_SLAB_FREE_PTR(aa->aa_oa, osc_obdo_kmem);
 811         aa->aa_oa = NULL;
 812
 813         return rc;
 814 }
 815
 816 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
 817 {
 818         spin_lock(&cli->cl_loi_list_lock);
 819         oa->o_grant = cli->cl_avail_grant / 4;
 820         cli->cl_avail_grant -= oa->o_grant;
 821         spin_unlock(&cli->cl_loi_list_lock);
 822         if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
 823                 oa->o_valid |= OBD_MD_FLFLAGS;
 824                 oa->o_flags = 0;
 825         }
 826         oa->o_flags |= OBD_FL_SHRINK_GRANT;
 827         osc_update_next_shrink(cli);
 828 }
 829
 830 /* Shrink the current grant, either from some large amount to enough for a
 831  * full set of in-flight RPCs, or if we have already shrunk to that limit
 832  * then to enough for a single RPC.  This avoids keeping more grant than
 833  * needed, and avoids shrinking the grant piecemeal. */
 834 static int osc_shrink_grant(struct client_obd *cli)
 835 {
 836         __u64 target_bytes = (cli->cl_max_rpcs_in_flight + 1) *
 837                              (cli->cl_max_pages_per_rpc << PAGE_SHIFT);
 838
 839         spin_lock(&cli->cl_loi_list_lock);
 840         if (cli->cl_avail_grant <= target_bytes)
 841                 target_bytes = cli->cl_max_pages_per_rpc << PAGE_SHIFT;
 842         spin_unlock(&cli->cl_loi_list_lock);
 843
 844         return osc_shrink_grant_to_target(cli, target_bytes);
 845 }
 846
 847 int osc_shrink_grant_to_target(struct client_obd *cli, __u64 target_bytes)
 848 {
 849         int                     rc = 0;
 850         struct ost_body        *body;
 851         ENTRY;
 852
 853         spin_lock(&cli->cl_loi_list_lock);
 854         /* Don't shrink if we are already above or below the desired limit
 855          * We don't want to shrink below a single RPC, as that will negatively
 856          * impact block allocation and long-term performance. */
 857         if (target_bytes < cli->cl_max_pages_per_rpc << PAGE_SHIFT)
 858                 target_bytes = cli->cl_max_pages_per_rpc << PAGE_SHIFT;
 859
 860         if (target_bytes >= cli->cl_avail_grant) {
 861                 spin_unlock(&cli->cl_loi_list_lock);
 862                 RETURN(0);
 863         }
 864         spin_unlock(&cli->cl_loi_list_lock);
 865
 866         OBD_ALLOC_PTR(body);
 867         if (!body)
 868                 RETURN(-ENOMEM);
 869
 870         osc_announce_cached(cli, &body->oa, 0);
 871
 872         spin_lock(&cli->cl_loi_list_lock);
 873         if (target_bytes >= cli->cl_avail_grant) {
 874                 /* available grant has changed since target calculation */
 875                 spin_unlock(&cli->cl_loi_list_lock);
 876                 GOTO(out_free, rc = 0);
 877         }
 878         body->oa.o_grant = cli->cl_avail_grant - target_bytes;
 879         cli->cl_avail_grant = target_bytes;
 880         spin_unlock(&cli->cl_loi_list_lock);
 881         if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
 882                 body->oa.o_valid |= OBD_MD_FLFLAGS;
 883                 body->oa.o_flags = 0;
 884         }
 885         body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
 886         osc_update_next_shrink(cli);
 887
 888         rc = osc_set_info_async(NULL, cli->cl_import->imp_obd->obd_self_export,
 889                                 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
 890                                 sizeof(*body), body, NULL);
 891         if (rc != 0)
 892                 __osc_update_grant(cli, body->oa.o_grant);
 893 out_free:
 894         OBD_FREE_PTR(body);
 895         RETURN(rc);
 896 }
 897
 898 static int osc_should_shrink_grant(struct client_obd *client)
 899 {
 900         time64_t next_shrink = client->cl_next_shrink_grant;
 901
 902         if (client->cl_import == NULL)
 903                 return 0;
 904
 905         if (!OCD_HAS_FLAG(&client->cl_import->imp_connect_data, GRANT_SHRINK) ||
 906             client->cl_import->imp_grant_shrink_disabled) {
 907                 osc_update_next_shrink(client);
 908                 return 0;
 909         }
 910
 911         if (ktime_get_seconds() >= next_shrink - 5) {
 912                 /* Get the current RPC size directly, instead of going via:
 913                  * cli_brw_size(obd->u.cli.cl_import->imp_obd->obd_self_export)
 914                  * Keep comment here so that it can be found by searching. */
 915                 int brw_size = client->cl_max_pages_per_rpc << PAGE_SHIFT;
 916
 917                 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
 918                     client->cl_avail_grant > brw_size)
 919                         return 1;
 920                 else
 921                         osc_update_next_shrink(client);
 922         }
 923         return 0;
 924 }
 925
 926 #define GRANT_SHRINK_RPC_BATCH  100
 927
 928 static struct delayed_work work;
 929
 930 static void osc_grant_work_handler(struct work_struct *data)
 931 {
 932         struct client_obd *cli;
 933         int rpc_sent;
 934         bool init_next_shrink = true;
 935         time64_t next_shrink = ktime_get_seconds() + GRANT_SHRINK_INTERVAL;
 936
 937         rpc_sent = 0;
 938         mutex_lock(&client_gtd.gtd_mutex);
 939         list_for_each_entry(cli, &client_gtd.gtd_clients,
 940                             cl_grant_chain) {
 941                 if (rpc_sent < GRANT_SHRINK_RPC_BATCH &&
 942                     osc_should_shrink_grant(cli)) {
 943                         osc_shrink_grant(cli);
 944                         rpc_sent++;
 945                 }
 946
 947                 if (!init_next_shrink) {
 948                         if (cli->cl_next_shrink_grant < next_shrink &&
 949                             cli->cl_next_shrink_grant > ktime_get_seconds())
 950                                 next_shrink = cli->cl_next_shrink_grant;
 951                 } else {
 952                         init_next_shrink = false;
 953                         next_shrink = cli->cl_next_shrink_grant;
 954                 }
 955         }
 956         mutex_unlock(&client_gtd.gtd_mutex);
 957
 958         if (client_gtd.gtd_stopped == 1)
 959                 return;
 960
 961         if (next_shrink > ktime_get_seconds()) {
 962                 time64_t delay = next_shrink - ktime_get_seconds();
 963
 964                 schedule_delayed_work(&work, cfs_time_seconds(delay));
 965         } else {
 966                 schedule_work(&work.work);
 967         }
 968 }
 969
 970 void osc_schedule_grant_work(void)
 971 {
 972         cancel_delayed_work_sync(&work);
 973         schedule_work(&work.work);
 974 }
 975 EXPORT_SYMBOL(osc_schedule_grant_work);
 976
 977 /**
 978  * Start grant thread for returing grant to server for idle clients.
 979  */
 980 static int osc_start_grant_work(void)
 981 {
 982         client_gtd.gtd_stopped = 0;
 983         mutex_init(&client_gtd.gtd_mutex);
 984         INIT_LIST_HEAD(&client_gtd.gtd_clients);
 985
 986         INIT_DELAYED_WORK(&work, osc_grant_work_handler);
 987         schedule_work(&work.work);
 988
 989         return 0;
 990 }
 991
 992 static void osc_stop_grant_work(void)
 993 {
 994         client_gtd.gtd_stopped = 1;
 995         cancel_delayed_work_sync(&work);
 996 }
 997
 998 static void osc_add_grant_list(struct client_obd *client)
 999 {
1000         mutex_lock(&client_gtd.gtd_mutex);
1001         list_add(&client->cl_grant_chain, &client_gtd.gtd_clients);
1002         mutex_unlock(&client_gtd.gtd_mutex);
1003 }
1004
1005 static void osc_del_grant_list(struct client_obd *client)
1006 {
1007         if (list_empty(&client->cl_grant_chain))
1008                 return;
1009
1010         mutex_lock(&client_gtd.gtd_mutex);
1011         list_del_init(&client->cl_grant_chain);
1012         mutex_unlock(&client_gtd.gtd_mutex);
1013 }
1014
1015 void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1016 {
1017         /*
1018          * ocd_grant is the total grant amount we're expect to hold: if we've
1019          * been evicted, it's the new avail_grant amount, cl_dirty_pages will
1020          * drop to 0 as inflight RPCs fail out; otherwise, it's avail_grant +
1021          * dirty.
1022          *
1023          * race is tolerable here: if we're evicted, but imp_state already
1024          * left EVICTED state, then cl_dirty_pages must be 0 already.
1025          */
1026         spin_lock(&cli->cl_loi_list_lock);
1027         cli->cl_avail_grant = ocd->ocd_grant;
1028         if (cli->cl_import->imp_state != LUSTRE_IMP_EVICTED) {
1029                 unsigned long consumed = cli->cl_reserved_grant;
1030
1031                 if (OCD_HAS_FLAG(ocd, GRANT_PARAM))
1032                         consumed += cli->cl_dirty_grant;
1033                 else
1034                         consumed += cli->cl_dirty_pages << PAGE_SHIFT;
1035                 if (cli->cl_avail_grant < consumed) {
1036                         CERROR("%s: granted %ld but already consumed %ld\n",
1037                                cli_name(cli), cli->cl_avail_grant, consumed);
1038                         cli->cl_avail_grant = 0;
1039                 } else {
1040                         cli->cl_avail_grant -= consumed;
1041                 }
1042         }
1043
1044         if (OCD_HAS_FLAG(ocd, GRANT_PARAM)) {
1045                 u64 size;
1046                 int chunk_mask;
1047
1048                 /* overhead for each extent insertion */
1049                 cli->cl_grant_extent_tax = ocd->ocd_grant_tax_kb << 10;
1050                 /* determine the appropriate chunk size used by osc_extent. */
1051                 cli->cl_chunkbits = max_t(int, PAGE_SHIFT,
1052                                           ocd->ocd_grant_blkbits);
1053                 /* max_pages_per_rpc must be chunk aligned */
1054                 chunk_mask = ~((1 << (cli->cl_chunkbits - PAGE_SHIFT)) - 1);
1055                 cli->cl_max_pages_per_rpc = (cli->cl_max_pages_per_rpc +
1056                                              ~chunk_mask) & chunk_mask;
1057                 /* determine maximum extent size, in #pages */
1058                 size = (u64)ocd->ocd_grant_max_blks << ocd->ocd_grant_blkbits;
1059                 cli->cl_max_extent_pages = (size >> PAGE_SHIFT) ?: 1;
1060                 cli->cl_ocd_grant_param = 1;
1061         } else {
1062                 cli->cl_ocd_grant_param = 0;
1063                 cli->cl_grant_extent_tax = 0;
1064                 cli->cl_chunkbits = PAGE_SHIFT;
1065                 cli->cl_max_extent_pages = DT_MAX_BRW_PAGES;
1066         }
1067         spin_unlock(&cli->cl_loi_list_lock);
1068
1069         CDEBUG(D_CACHE,
1070                "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld. chunk bits: %d cl_max_extent_pages: %d\n",
1071                cli_name(cli),
1072                cli->cl_avail_grant, cli->cl_lost_grant, cli->cl_chunkbits,
1073                cli->cl_max_extent_pages);
1074
1075         if (OCD_HAS_FLAG(ocd, GRANT_SHRINK) && list_empty(&cli->cl_grant_chain))
1076                 osc_add_grant_list(cli);
1077 }
1078 EXPORT_SYMBOL(osc_init_grant);
1079
1080 /* We assume that the reason this OSC got a short read is because it read
1081  * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1082  * via the LOV, and it _knows_ it's reading inside the file, it's just that
1083  * this stripe never got written at or beyond this stripe offset yet. */
1084 static void handle_short_read(int nob_read, size_t page_count,
1085                               struct brw_page **pga)
1086 {
1087         char *ptr;
1088         int i = 0;
1089
1090         /* skip bytes read OK */
1091         while (nob_read > 0) {
1092                 LASSERT (page_count > 0);
1093
1094                 if (pga[i]->count > nob_read) {
1095                         /* EOF inside this page */
1096                         ptr = kmap(pga[i]->pg) +
1097                                 (pga[i]->off & ~PAGE_MASK);
1098                         memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1099                         kunmap(pga[i]->pg);
1100                         page_count--;
1101                         i++;
1102                         break;
1103                 }
1104
1105                 nob_read -= pga[i]->count;
1106                 page_count--;
1107                 i++;
1108         }
1109
1110         /* zero remaining pages */
1111         while (page_count-- > 0) {
1112                 ptr = kmap(pga[i]->pg) + (pga[i]->off & ~PAGE_MASK);
1113                 memset(ptr, 0, pga[i]->count);
1114                 kunmap(pga[i]->pg);
1115                 i++;
1116         }
1117 }
1118
1119 static int check_write_rcs(struct ptlrpc_request *req,
1120                            int requested_nob, int niocount,
1121                            size_t page_count, struct brw_page **pga)
1122 {
1123         int     i;
1124         __u32   *remote_rcs;
1125
1126         remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1127                                                   sizeof(*remote_rcs) *
1128                                                   niocount);
1129         if (remote_rcs == NULL) {
1130                 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1131                 return(-EPROTO);
1132         }
1133
1134         /* return error if any niobuf was in error */
1135         for (i = 0; i < niocount; i++) {
1136                 if ((int)remote_rcs[i] < 0) {
1137                         CDEBUG(D_INFO, "rc[%d]: %d req %p\n",
1138                                i, remote_rcs[i], req);
1139                         return remote_rcs[i];
1140                 }
1141
1142                 if (remote_rcs[i] != 0) {
1143                         CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1144                                 i, remote_rcs[i], req);
1145                         return(-EPROTO);
1146                 }
1147         }
1148         if (req->rq_bulk != NULL &&
1149             req->rq_bulk->bd_nob_transferred != requested_nob) {
1150                 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1151                        req->rq_bulk->bd_nob_transferred, requested_nob);
1152                 return(-EPROTO);
1153         }
1154
1155         return (0);
1156 }
1157
1158 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1159 {
1160         if (p1->flag != p2->flag) {
1161                 unsigned mask = ~(OBD_BRW_FROM_GRANT | OBD_BRW_NOCACHE |
1162                                   OBD_BRW_SYNC       | OBD_BRW_ASYNC   |
1163                                   OBD_BRW_NOQUOTA    | OBD_BRW_SOFT_SYNC |
1164                                   OBD_BRW_SYS_RESOURCE);
1165
1166                 /* warn if we try to combine flags that we don't know to be
1167                  * safe to combine */
1168                 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
1169                         CWARN("Saw flags 0x%x and 0x%x in the same brw, please "
1170                               "report this at https://jira.whamcloud.com/\n",
1171                               p1->flag, p2->flag);
1172                 }
1173                 return 0;
1174         }
1175
1176         return (p1->off + p1->count == p2->off);
1177 }
1178
1179 #if IS_ENABLED(CONFIG_CRC_T10DIF)
1180 static int osc_checksum_bulk_t10pi(const char *obd_name, int nob,
1181                                    size_t pg_count, struct brw_page **pga,
1182                                    int opc, obd_dif_csum_fn *fn,
1183                                    int sector_size,
1184                                    u32 *check_sum, bool resend)
1185 {
1186         struct ahash_request *req;
1187         /* Used Adler as the default checksum type on top of DIF tags */
1188         unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
1189         struct page *__page;
1190         unsigned char *buffer;
1191         __u16 *guard_start;
1192         unsigned int bufsize;
1193         int guard_number;
1194         int used_number = 0;
1195         int used;
1196         u32 cksum;
1197         int rc = 0;
1198         int i = 0;
1199
1200         LASSERT(pg_count > 0);
1201
1202         __page = alloc_page(GFP_KERNEL);
1203         if (__page == NULL)
1204                 return -ENOMEM;
1205
1206         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1207         if (IS_ERR(req)) {
1208                 rc = PTR_ERR(req);
1209                 CERROR("%s: unable to initialize checksum hash %s: rc = %d\n",
1210                        obd_name, cfs_crypto_hash_name(cfs_alg), rc);
1211                 GOTO(out, rc);
1212         }
1213
1214         buffer = kmap(__page);
1215         guard_start = (__u16 *)buffer;
1216         guard_number = PAGE_SIZE / sizeof(*guard_start);
1217         CDEBUG(D_PAGE | (resend ? D_HA : 0),
1218                "GRD tags per page=%u, resend=%u, bytes=%u, pages=%zu\n",
1219                guard_number, resend, nob, pg_count);
1220
1221         while (nob > 0 && pg_count > 0) {
1222                 unsigned int count = pga[i]->count > nob ? nob : pga[i]->count;
1223
1224                 /* corrupt the data before we compute the checksum, to
1225                  * simulate an OST->client data error */
1226                 if (unlikely(i == 0 && opc == OST_READ &&
1227                              OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))) {
1228                         unsigned char *ptr = kmap(pga[i]->pg);
1229                         int off = pga[i]->off & ~PAGE_MASK;
1230
1231                         memcpy(ptr + off, "bad1", min_t(typeof(nob), 4, nob));
1232                         kunmap(pga[i]->pg);
1233                 }
1234
1235                 /*
1236                  * The left guard number should be able to hold checksums of a
1237                  * whole page
1238                  */
1239                 rc = obd_page_dif_generate_buffer(obd_name, pga[i]->pg,
1240                                                   pga[i]->off & ~PAGE_MASK,
1241                                                   count,
1242                                                   guard_start + used_number,
1243                                                   guard_number - used_number,
1244                                                   &used, sector_size,
1245                                                   fn);
1246                 if (unlikely(resend))
1247                         CDEBUG(D_PAGE | D_HA,
1248                                "pga[%u]: used %u off %llu+%u gen checksum: %*phN\n",
1249                                i, used, pga[i]->off & ~PAGE_MASK, count,
1250                                (int)(used * sizeof(*guard_start)),
1251                                guard_start + used_number);
1252                 if (rc)
1253                         break;
1254
1255                 used_number += used;
1256                 if (used_number == guard_number) {
1257                         cfs_crypto_hash_update_page(req, __page, 0,
1258                                 used_number * sizeof(*guard_start));
1259                         used_number = 0;
1260                 }
1261
1262                 nob -= pga[i]->count;
1263                 pg_count--;
1264                 i++;
1265         }
1266         kunmap(__page);
1267         if (rc)
1268                 GOTO(out, rc);
1269
1270         if (used_number != 0)
1271                 cfs_crypto_hash_update_page(req, __page, 0,
1272                         used_number * sizeof(*guard_start));
1273
1274         bufsize = sizeof(cksum);
1275         cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
1276
1277         /* For sending we only compute the wrong checksum instead
1278          * of corrupting the data so it is still correct on a redo */
1279         if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1280                 cksum++;
1281
1282         *check_sum = cksum;
1283 out:
1284         __free_page(__page);
1285         return rc;
1286 }
1287 #else /* !CONFIG_CRC_T10DIF */
1288 #define obd_dif_ip_fn NULL
1289 #define obd_dif_crc_fn NULL
1290 #define osc_checksum_bulk_t10pi(name, nob, pgc, pga, opc, fn, ssize, csum, re) \
1291         -EOPNOTSUPP
1292 #endif /* CONFIG_CRC_T10DIF */
1293
1294 static int osc_checksum_bulk(int nob, size_t pg_count,
1295                              struct brw_page **pga, int opc,
1296                              enum cksum_types cksum_type,
1297                              u32 *cksum)
1298 {
1299         int                             i = 0;
1300         struct ahash_request           *req;
1301         unsigned int                    bufsize;
1302         unsigned char                   cfs_alg = cksum_obd2cfs(cksum_type);
1303
1304         LASSERT(pg_count > 0);
1305
1306         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1307         if (IS_ERR(req)) {
1308                 CERROR("Unable to initialize checksum hash %s\n",
1309                        cfs_crypto_hash_name(cfs_alg));
1310                 return PTR_ERR(req);
1311         }
1312
1313         while (nob > 0 && pg_count > 0) {
1314                 unsigned int count = pga[i]->count > nob ? nob : pga[i]->count;
1315
1316                 /* corrupt the data before we compute the checksum, to
1317                  * simulate an OST->client data error */
1318                 if (i == 0 && opc == OST_READ &&
1319                     OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE)) {
1320                         unsigned char *ptr = kmap(pga[i]->pg);
1321                         int off = pga[i]->off & ~PAGE_MASK;
1322
1323                         memcpy(ptr + off, "bad1", min_t(typeof(nob), 4, nob));
1324                         kunmap(pga[i]->pg);
1325                 }
1326                 cfs_crypto_hash_update_page(req, pga[i]->pg,
1327                                             pga[i]->off & ~PAGE_MASK,
1328                                             count);
1329                 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d\n",
1330                                (int)(pga[i]->off & ~PAGE_MASK));
1331
1332                 nob -= pga[i]->count;
1333                 pg_count--;
1334                 i++;
1335         }
1336
1337         bufsize = sizeof(*cksum);
1338         cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1339
1340         /* For sending we only compute the wrong checksum instead
1341          * of corrupting the data so it is still correct on a redo */
1342         if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1343                 (*cksum)++;
1344
1345         return 0;
1346 }
1347
1348 static int osc_checksum_bulk_rw(const char *obd_name,
1349                                 enum cksum_types cksum_type,
1350                                 int nob, size_t pg_count,
1351                                 struct brw_page **pga, int opc,
1352                                 u32 *check_sum, bool resend)
1353 {
1354         obd_dif_csum_fn *fn = NULL;
1355         int sector_size = 0;
1356         int rc;
1357
1358         ENTRY;
1359         obd_t10_cksum2dif(cksum_type, &fn, &sector_size);
1360
1361         if (fn)
1362                 rc = osc_checksum_bulk_t10pi(obd_name, nob, pg_count, pga,
1363                                              opc, fn, sector_size, check_sum,
1364                                              resend);
1365         else
1366                 rc = osc_checksum_bulk(nob, pg_count, pga, opc, cksum_type,
1367                                        check_sum);
1368
1369         RETURN(rc);
1370 }
1371
1372 static inline void osc_release_bounce_pages(struct brw_page **pga,
1373                                             u32 page_count)
1374 {
1375 #ifdef HAVE_LUSTRE_CRYPTO
1376         int i;
1377
1378         for (i = 0; i < page_count; i++) {
1379                 /* Bounce pages allocated by a call to
1380                  * llcrypt_encrypt_pagecache_blocks() in osc_brw_prep_request()
1381                  * are identified thanks to the PageChecked flag.
1382                  */
1383                 if (PageChecked(pga[i]->pg))
1384                         llcrypt_finalize_bounce_page(&pga[i]->pg);
1385                 pga[i]->count -= pga[i]->bp_count_diff;
1386                 pga[i]->off += pga[i]->bp_off_diff;
1387         }
1388 #endif
1389 }
1390
1391 static int
1392 osc_brw_prep_request(int cmd, struct client_obd *cli, struct obdo *oa,
1393                      u32 page_count, struct brw_page **pga,
1394                      struct ptlrpc_request **reqp, int resend)
1395 {
1396         struct ptlrpc_request *req;
1397         struct ptlrpc_bulk_desc *desc;
1398         struct ost_body *body;
1399         struct obd_ioobj *ioobj;
1400         struct niobuf_remote *niobuf;
1401         int niocount, i, requested_nob, opc, rc, short_io_size = 0;
1402         struct osc_brw_async_args *aa;
1403         struct req_capsule *pill;
1404         struct brw_page *pg_prev;
1405         void *short_io_buf;
1406         const char *obd_name = cli->cl_import->imp_obd->obd_name;
1407         struct inode *inode = NULL;
1408         bool directio = false;
1409         bool gpu = 0;
1410         bool enable_checksum = true;
1411         struct cl_page *clpage;
1412
1413         ENTRY;
1414         if (pga[0]->pg) {
1415                 clpage = oap2cl_page(brw_page2oap(pga[0]));
1416                 inode = clpage->cp_inode;
1417                 if (clpage->cp_type == CPT_TRANSIENT)
1418                         directio = true;
1419         }
1420         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1421                 RETURN(-ENOMEM); /* Recoverable */
1422         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1423                 RETURN(-EINVAL); /* Fatal */
1424
1425         if ((cmd & OBD_BRW_WRITE) != 0) {
1426                 opc = OST_WRITE;
1427                 req = ptlrpc_request_alloc_pool(cli->cl_import,
1428                                                 osc_rq_pool,
1429                                                 &RQF_OST_BRW_WRITE);
1430         } else {
1431                 opc = OST_READ;
1432                 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1433         }
1434         if (req == NULL)
1435                 RETURN(-ENOMEM);
1436
1437         if (opc == OST_WRITE && inode && IS_ENCRYPTED(inode) &&
1438             llcrypt_has_encryption_key(inode)) {
1439                 for (i = 0; i < page_count; i++) {
1440                         struct brw_page *brwpg = pga[i];
1441                         struct page *data_page = NULL;
1442                         bool retried = false;
1443                         bool lockedbymyself;
1444                         u32 nunits = (brwpg->off & ~PAGE_MASK) + brwpg->count;
1445                         struct address_space *map_orig = NULL;
1446                         pgoff_t index_orig;
1447
1448 retry_encrypt:
1449                         nunits = round_up(nunits, LUSTRE_ENCRYPTION_UNIT_SIZE);
1450                         /* The page can already be locked when we arrive here.
1451                          * This is possible when cl_page_assume/vvp_page_assume
1452                          * is stuck on wait_on_page_writeback with page lock
1453                          * held. In this case there is no risk for the lock to
1454                          * be released while we are doing our encryption
1455                          * processing, because writeback against that page will
1456                          * end in vvp_page_completion_write/cl_page_completion,
1457                          * which means only once the page is fully processed.
1458                          */
1459                         lockedbymyself = trylock_page(brwpg->pg);
1460                         if (directio) {
1461                                 map_orig = brwpg->pg->mapping;
1462                                 brwpg->pg->mapping = inode->i_mapping;
1463                                 index_orig = brwpg->pg->index;
1464                                 clpage = oap2cl_page(brw_page2oap(brwpg));
1465                                 brwpg->pg->index = clpage->cp_page_index;
1466                         }
1467                         data_page =
1468                                 llcrypt_encrypt_pagecache_blocks(brwpg->pg,
1469                                                                  nunits, 0,
1470                                                                  GFP_NOFS);
1471                         if (directio) {
1472                                 brwpg->pg->mapping = map_orig;
1473                                 brwpg->pg->index = index_orig;
1474                         }
1475                         if (lockedbymyself)
1476                                 unlock_page(brwpg->pg);
1477                         if (IS_ERR(data_page)) {
1478                                 rc = PTR_ERR(data_page);
1479                                 if (rc == -ENOMEM && !retried) {
1480                                         retried = true;
1481                                         rc = 0;
1482                                         goto retry_encrypt;
1483                                 }
1484                                 ptlrpc_request_free(req);
1485                                 RETURN(rc);
1486                         }
1487                         /* Set PageChecked flag on bounce page for
1488                          * disambiguation in osc_release_bounce_pages().
1489                          */
1490                         SetPageChecked(data_page);
1491                         brwpg->pg = data_page;
1492                         /* there should be no gap in the middle of page array */
1493                         if (i == page_count - 1) {
1494                                 struct osc_async_page *oap =
1495                                         brw_page2oap(brwpg);
1496
1497                                 oa->o_size = oap->oap_count +
1498                                         oap->oap_obj_off + oap->oap_page_off;
1499                         }
1500                         /* len is forced to nunits, and relative offset to 0
1501                          * so store the old, clear text info
1502                          */
1503                         brwpg->bp_count_diff = nunits - brwpg->count;
1504                         brwpg->count = nunits;
1505                         brwpg->bp_off_diff = brwpg->off & ~PAGE_MASK;
1506                         brwpg->off = brwpg->off & PAGE_MASK;
1507                 }
1508         } else if (opc == OST_WRITE && inode && IS_ENCRYPTED(inode)) {
1509                 struct osc_async_page *oap = brw_page2oap(pga[0]);
1510                 struct cl_page *clpage = oap2cl_page(oap);
1511                 struct cl_object *clobj = clpage->cp_obj;
1512                 struct cl_attr attr = { 0 };
1513                 struct lu_env *env;
1514                 __u16 refcheck;
1515
1516                 env = cl_env_get(&refcheck);
1517                 if (IS_ERR(env)) {
1518                         rc = PTR_ERR(env);
1519                         ptlrpc_request_free(req);
1520                         RETURN(rc);
1521                 }
1522
1523                 cl_object_attr_lock(clobj);
1524                 rc = cl_object_attr_get(env, clobj, &attr);
1525                 cl_object_attr_unlock(clobj);
1526                 cl_env_put(env, &refcheck);
1527                 if (rc != 0) {
1528                         ptlrpc_request_free(req);
1529                         RETURN(rc);
1530                 }
1531                 if (attr.cat_size)
1532                         oa->o_size = attr.cat_size;
1533         } else if (opc == OST_READ && inode && IS_ENCRYPTED(inode) &&
1534                    llcrypt_has_encryption_key(inode)) {
1535                 for (i = 0; i < page_count; i++) {
1536                         struct brw_page *pg = pga[i];
1537                         u32 nunits = (pg->off & ~PAGE_MASK) + pg->count;
1538
1539                         nunits = round_up(nunits, LUSTRE_ENCRYPTION_UNIT_SIZE);
1540                         /* count/off are forced to cover the whole encryption
1541                          * unit size so that all encrypted data is stored on the
1542                          * OST, so adjust bp_{count,off}_diff for the size of
1543                          * the clear text.
1544                          */
1545                         pg->bp_count_diff = nunits - pg->count;
1546                         pg->count = nunits;
1547                         pg->bp_off_diff = pg->off & ~PAGE_MASK;
1548                         pg->off = pg->off & PAGE_MASK;
1549                 }
1550         }
1551
1552         for (niocount = i = 1; i < page_count; i++) {
1553                 if (!can_merge_pages(pga[i - 1], pga[i]))
1554                         niocount++;
1555         }
1556
1557         pill = &req->rq_pill;
1558         req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1559                              sizeof(*ioobj));
1560         req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1561                              niocount * sizeof(*niobuf));
1562
1563         for (i = 0; i < page_count; i++) {
1564                 short_io_size += pga[i]->count;
1565                 if (!inode || !IS_ENCRYPTED(inode) ||
1566                     !llcrypt_has_encryption_key(inode)) {
1567                         pga[i]->bp_count_diff = 0;
1568                         pga[i]->bp_off_diff = 0;
1569                 }
1570         }
1571
1572         if (brw_page2oap(pga[0])->oap_brw_flags & OBD_BRW_RDMA_ONLY) {
1573                 enable_checksum = false;
1574                 short_io_size = 0;
1575                 gpu = 1;
1576         }
1577
1578         /* Check if read/write is small enough to be a short io. */
1579         if (short_io_size > cli->cl_max_short_io_bytes || niocount > 1 ||
1580             !imp_connect_shortio(cli->cl_import))
1581                 short_io_size = 0;
1582
1583         /* If this is an empty RPC to old server, just ignore it */
1584         if (!short_io_size && !pga[0]->pg) {
1585                 ptlrpc_request_free(req);
1586                 RETURN(-ENODATA);
1587         }
1588
1589         req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_CLIENT,
1590                              opc == OST_READ ? 0 : short_io_size);
1591         if (opc == OST_READ)
1592                 req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_SERVER,
1593                                      short_io_size);
1594
1595         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1596         if (rc) {
1597                 ptlrpc_request_free(req);
1598                 RETURN(rc);
1599         }
1600         osc_set_io_portal(req);
1601
1602         ptlrpc_at_set_req_timeout(req);
1603         /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1604          * retry logic */
1605         req->rq_no_retry_einprogress = 1;
1606
1607         if (short_io_size != 0) {
1608                 desc = NULL;
1609                 short_io_buf = NULL;
1610                 goto no_bulk;
1611         }
1612
1613         desc = ptlrpc_prep_bulk_imp(req, page_count,
1614                 cli->cl_import->imp_connect_data.ocd_brw_size >> LNET_MTU_BITS,
1615                 (opc == OST_WRITE ? PTLRPC_BULK_GET_SOURCE :
1616                         PTLRPC_BULK_PUT_SINK),
1617                 OST_BULK_PORTAL,
1618                 &ptlrpc_bulk_kiov_pin_ops);
1619
1620         if (desc == NULL)
1621                 GOTO(out, rc = -ENOMEM);
1622         /* NB request now owns desc and will free it when it gets freed */
1623         desc->bd_is_rdma = gpu;
1624 no_bulk:
1625         body = req_capsule_client_get(pill, &RMF_OST_BODY);
1626         ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1627         niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1628         LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1629
1630         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1631
1632         /* For READ and WRITE, we can't fill o_uid and o_gid using from_kuid()
1633          * and from_kgid(), because they are asynchronous. Fortunately, variable
1634          * oa contains valid o_uid and o_gid in these two operations.
1635          * Besides, filling o_uid and o_gid is enough for nrs-tbf, see LU-9658.
1636          * OBD_MD_FLUID and OBD_MD_FLUID is not set in order to avoid breaking
1637          * other process logic */
1638         body->oa.o_uid = oa->o_uid;
1639         body->oa.o_gid = oa->o_gid;
1640
1641         obdo_to_ioobj(oa, ioobj);
1642         ioobj->ioo_bufcnt = niocount;
1643         /* The high bits of ioo_max_brw tells server _maximum_ number of bulks
1644          * that might be send for this request.  The actual number is decided
1645          * when the RPC is finally sent in ptlrpc_register_bulk(). It sends
1646          * "max - 1" for old client compatibility sending "0", and also so the
1647          * the actual maximum is a power-of-two number, not one less. LU-1431 */
1648         if (desc != NULL)
1649                 ioobj_max_brw_set(ioobj, desc->bd_md_max_brw);
1650         else /* short io */
1651                 ioobj_max_brw_set(ioobj, 0);
1652
1653         if (short_io_size != 0) {
1654                 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1655                         body->oa.o_valid |= OBD_MD_FLFLAGS;
1656                         body->oa.o_flags = 0;
1657                 }
1658                 body->oa.o_flags |= OBD_FL_SHORT_IO;
1659                 CDEBUG(D_CACHE, "Using short io for data transfer, size = %d\n",
1660                        short_io_size);
1661                 if (opc == OST_WRITE) {
1662                         short_io_buf = req_capsule_client_get(pill,
1663                                                               &RMF_SHORT_IO);
1664                         LASSERT(short_io_buf != NULL);
1665                 }
1666         }
1667
1668         LASSERT(page_count > 0);
1669         pg_prev = pga[0];
1670         for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1671                 struct brw_page *pg = pga[i];
1672                 int poff = pg->off & ~PAGE_MASK;
1673
1674                 LASSERT(pg->count > 0);
1675                 /* make sure there is no gap in the middle of page array */
1676                 LASSERTF(page_count == 1 ||
1677                          (ergo(i == 0, poff + pg->count == PAGE_SIZE) &&
1678                           ergo(i > 0 && i < page_count - 1,
1679                                poff == 0 && pg->count == PAGE_SIZE)   &&
1680                           ergo(i == page_count - 1, poff == 0)),
1681                          "i: %d/%d pg: %p off: %llu, count: %u\n",
1682                          i, page_count, pg, pg->off, pg->count);
1683                 LASSERTF(i == 0 || pg->off > pg_prev->off,
1684                          "i %d p_c %u pg %p [pri %lu ind %lu] off %llu"
1685                          " prev_pg %p [pri %lu ind %lu] off %llu\n",
1686                          i, page_count,
1687                          pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1688                          pg_prev->pg, page_private(pg_prev->pg),
1689                          pg_prev->pg->index, pg_prev->off);
1690                 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1691                         (pg->flag & OBD_BRW_SRVLOCK));
1692                 if (short_io_size != 0 && opc == OST_WRITE) {
1693                         unsigned char *ptr = kmap_atomic(pg->pg);
1694
1695                         LASSERT(short_io_size >= requested_nob + pg->count);
1696                         memcpy(short_io_buf + requested_nob,
1697                                ptr + poff,
1698                                pg->count);
1699                         kunmap_atomic(ptr);
1700                 } else if (short_io_size == 0) {
1701                         desc->bd_frag_ops->add_kiov_frag(desc, pg->pg, poff,
1702                                                          pg->count);
1703                 }
1704                 requested_nob += pg->count;
1705
1706                 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1707                         niobuf--;
1708                         niobuf->rnb_len += pg->count;
1709                 } else {
1710                         niobuf->rnb_offset = pg->off;
1711                         niobuf->rnb_len    = pg->count;
1712                         niobuf->rnb_flags  = pg->flag;
1713                 }
1714                 pg_prev = pg;
1715         }
1716
1717         LASSERTF((void *)(niobuf - niocount) ==
1718                 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1719                 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1720                 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1721
1722         osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1723         if (resend) {
1724                 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1725                         body->oa.o_valid |= OBD_MD_FLFLAGS;
1726                         body->oa.o_flags = 0;
1727                 }
1728                 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1729         }
1730
1731         if (osc_should_shrink_grant(cli))
1732                 osc_shrink_grant_local(cli, &body->oa);
1733
1734         if (!cli->cl_checksum || sptlrpc_flavor_has_bulk(&req->rq_flvr))
1735                 enable_checksum = false;
1736
1737         /* size[REQ_REC_OFF] still sizeof (*body) */
1738         if (opc == OST_WRITE) {
1739                 if (enable_checksum) {
1740                         /* store cl_cksum_type in a local variable since
1741                          * it can be changed via lprocfs */
1742                         enum cksum_types cksum_type = cli->cl_cksum_type;
1743
1744                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1745                                 body->oa.o_flags = 0;
1746
1747                         body->oa.o_flags |= obd_cksum_type_pack(obd_name,
1748                                                                 cksum_type);
1749                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1750
1751                         rc = osc_checksum_bulk_rw(obd_name, cksum_type,
1752                                                   requested_nob, page_count,
1753                                                   pga, OST_WRITE,
1754                                                   &body->oa.o_cksum, resend);
1755                         if (rc < 0) {
1756                                 CDEBUG(D_PAGE, "failed to checksum: rc = %d\n",
1757                                        rc);
1758                                 GOTO(out, rc);
1759                         }
1760                         CDEBUG(D_PAGE | (resend ? D_HA : 0),
1761                                "checksum at write origin: %x (%x)\n",
1762                                body->oa.o_cksum, cksum_type);
1763
1764                         /* save this in 'oa', too, for later checking */
1765                         oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1766                         oa->o_flags |= obd_cksum_type_pack(obd_name,
1767                                                            cksum_type);
1768                 } else {
1769                         /* clear out the checksum flag, in case this is a
1770                          * resend but cl_checksum is no longer set. b=11238 */
1771                         oa->o_valid &= ~OBD_MD_FLCKSUM;
1772                 }
1773                 oa->o_cksum = body->oa.o_cksum;
1774                 /* 1 RC per niobuf */
1775                 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1776                                      sizeof(__u32) * niocount);
1777         } else {
1778                 if (enable_checksum) {
1779                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1780                                 body->oa.o_flags = 0;
1781                         body->oa.o_flags |= obd_cksum_type_pack(obd_name,
1782                                 cli->cl_cksum_type);
1783                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1784                 }
1785
1786                 /* Client cksum has been already copied to wire obdo in previous
1787                  * lustre_set_wire_obdo(), and in the case a bulk-read is being
1788                  * resent due to cksum error, this will allow Server to
1789                  * check+dump pages on its side */
1790         }
1791         ptlrpc_request_set_replen(req);
1792
1793         aa = ptlrpc_req_async_args(aa, req);
1794         aa->aa_oa = oa;
1795         aa->aa_requested_nob = requested_nob;
1796         aa->aa_nio_count = niocount;
1797         aa->aa_page_count = page_count;
1798         aa->aa_resends = 0;
1799         aa->aa_ppga = pga;
1800         aa->aa_cli = cli;
1801         INIT_LIST_HEAD(&aa->aa_oaps);
1802
1803         *reqp = req;
1804         niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1805         CDEBUG(D_RPCTRACE, "brw rpc %p - object "DOSTID" offset %lld<>%lld\n",
1806                 req, POSTID(&oa->o_oi), niobuf[0].rnb_offset,
1807                 niobuf[niocount - 1].rnb_offset + niobuf[niocount - 1].rnb_len);
1808         RETURN(0);
1809
1810  out:
1811         ptlrpc_req_finished(req);
1812         RETURN(rc);
1813 }
1814
1815 char dbgcksum_file_name[PATH_MAX];
1816
1817 static void dump_all_bulk_pages(struct obdo *oa, __u32 page_count,
1818                                 struct brw_page **pga, __u32 server_cksum,
1819                                 __u32 client_cksum)
1820 {
1821         struct file *filp;
1822         int rc, i;
1823         unsigned int len;
1824         char *buf;
1825
1826         /* will only keep dump of pages on first error for the same range in
1827          * file/fid, not during the resends/retries. */
1828         snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1829                  "%s-checksum_dump-osc-"DFID":[%llu-%llu]-%x-%x",
1830                  (strncmp(libcfs_debug_file_path, "NONE", 4) != 0 ?
1831                   libcfs_debug_file_path : LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1832                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1833                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1834                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1835                  pga[0]->off,
1836                  pga[page_count-1]->off + pga[page_count-1]->count - 1,
1837                  client_cksum, server_cksum);
1838         CWARN("dumping checksum data to %s\n", dbgcksum_file_name);
1839         filp = filp_open(dbgcksum_file_name,
1840                          O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1841         if (IS_ERR(filp)) {
1842                 rc = PTR_ERR(filp);
1843                 if (rc == -EEXIST)
1844                         CDEBUG(D_INFO, "%s: can't open to dump pages with "
1845                                "checksum error: rc = %d\n", dbgcksum_file_name,
1846                                rc);
1847                 else
1848                         CERROR("%s: can't open to dump pages with checksum "
1849                                "error: rc = %d\n", dbgcksum_file_name, rc);
1850                 return;
1851         }
1852
1853         for (i = 0; i < page_count; i++) {
1854                 len = pga[i]->count;
1855                 buf = kmap(pga[i]->pg);
1856                 while (len != 0) {
1857                         rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1858                         if (rc < 0) {
1859                                 CERROR("%s: wanted to write %u but got %d "
1860                                        "error\n", dbgcksum_file_name, len, rc);
1861                                 break;
1862                         }
1863                         len -= rc;
1864                         buf += rc;
1865                 }
1866                 kunmap(pga[i]->pg);
1867         }
1868
1869         rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1870         if (rc)
1871                 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1872         filp_close(filp, NULL);
1873
1874         libcfs_debug_dumplog();
1875 }
1876
1877 static int
1878 check_write_checksum(struct obdo *oa, const struct lnet_processid *peer,
1879                      __u32 client_cksum, __u32 server_cksum,
1880                      struct osc_brw_async_args *aa)
1881 {
1882         const char *obd_name = aa->aa_cli->cl_import->imp_obd->obd_name;
1883         enum cksum_types cksum_type;
1884         obd_dif_csum_fn *fn = NULL;
1885         int sector_size = 0;
1886         __u32 new_cksum;
1887         char *msg;
1888         int rc;
1889
1890         if (server_cksum == client_cksum) {
1891                 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1892                 return 0;
1893         }
1894
1895         if (aa->aa_cli->cl_checksum_dump)
1896                 dump_all_bulk_pages(oa, aa->aa_page_count, aa->aa_ppga,
1897                                     server_cksum, client_cksum);
1898
1899         cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1900                                            oa->o_flags : 0);
1901
1902         switch (cksum_type) {
1903         case OBD_CKSUM_T10IP512:
1904                 fn = obd_dif_ip_fn;
1905                 sector_size = 512;
1906                 break;
1907         case OBD_CKSUM_T10IP4K:
1908                 fn = obd_dif_ip_fn;
1909                 sector_size = 4096;
1910                 break;
1911         case OBD_CKSUM_T10CRC512:
1912                 fn = obd_dif_crc_fn;
1913                 sector_size = 512;
1914                 break;
1915         case OBD_CKSUM_T10CRC4K:
1916                 fn = obd_dif_crc_fn;
1917                 sector_size = 4096;
1918                 break;
1919         default:
1920                 break;
1921         }
1922
1923         if (fn)
1924                 rc = osc_checksum_bulk_t10pi(obd_name, aa->aa_requested_nob,
1925                                              aa->aa_page_count, aa->aa_ppga,
1926                                              OST_WRITE, fn, sector_size,
1927                                              &new_cksum, true);
1928         else
1929                 rc = osc_checksum_bulk(aa->aa_requested_nob, aa->aa_page_count,
1930                                        aa->aa_ppga, OST_WRITE, cksum_type,
1931                                        &new_cksum);
1932
1933         if (rc < 0)
1934                 msg = "failed to calculate the client write checksum";
1935         else if (cksum_type != obd_cksum_type_unpack(aa->aa_oa->o_flags))
1936                 msg = "the server did not use the checksum type specified in "
1937                       "the original request - likely a protocol problem";
1938         else if (new_cksum == server_cksum)
1939                 msg = "changed on the client after we checksummed it - "
1940                       "likely false positive due to mmap IO (bug 11742)";
1941         else if (new_cksum == client_cksum)
1942                 msg = "changed in transit before arrival at OST";
1943         else
1944                 msg = "changed in transit AND doesn't match the original - "
1945                       "likely false positive due to mmap IO (bug 11742)";
1946
1947         LCONSOLE_ERROR_MSG(0x132, "%s: BAD WRITE CHECKSUM: %s: from %s inode "
1948                            DFID " object "DOSTID" extent [%llu-%llu], original "
1949                            "client csum %x (type %x), server csum %x (type %x),"
1950                            " client csum now %x\n",
1951                            obd_name, msg, libcfs_nidstr(&peer->nid),
1952                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1953                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1954                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1955                            POSTID(&oa->o_oi), aa->aa_ppga[0]->off,
1956                            aa->aa_ppga[aa->aa_page_count - 1]->off +
1957                                 aa->aa_ppga[aa->aa_page_count-1]->count - 1,
1958                            client_cksum,
1959                            obd_cksum_type_unpack(aa->aa_oa->o_flags),
1960                            server_cksum, cksum_type, new_cksum);
1961         return 1;
1962 }
1963
1964 /* Note rc enters this function as number of bytes transferred */
1965 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1966 {
1967         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1968         struct client_obd *cli = aa->aa_cli;
1969         const char *obd_name = cli->cl_import->imp_obd->obd_name;
1970         const struct lnet_processid *peer =
1971                 &req->rq_import->imp_connection->c_peer;
1972         struct ost_body *body;
1973         u32 client_cksum = 0;
1974         struct inode *inode = NULL;
1975         unsigned int blockbits = 0, blocksize = 0;
1976         struct cl_page *clpage;
1977
1978         ENTRY;
1979
1980         if (rc < 0 && rc != -EDQUOT) {
1981                 DEBUG_REQ(D_INFO, req, "Failed request: rc = %d", rc);
1982                 RETURN(rc);
1983         }
1984
1985         LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1986         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1987         if (body == NULL) {
1988                 DEBUG_REQ(D_INFO, req, "cannot unpack body");
1989                 RETURN(-EPROTO);
1990         }
1991
1992         /* set/clear over quota flag for a uid/gid/projid */
1993         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1994             body->oa.o_valid & (OBD_MD_FLALLQUOTA)) {
1995                 unsigned qid[LL_MAXQUOTAS] = {
1996                                          body->oa.o_uid, body->oa.o_gid,
1997                                          body->oa.o_projid };
1998                 CDEBUG(D_QUOTA,
1999                        "setdq for [%u %u %u] with valid %#llx, flags %x\n",
2000                        body->oa.o_uid, body->oa.o_gid, body->oa.o_projid,
2001                        body->oa.o_valid, body->oa.o_flags);
2002                 osc_quota_setdq(cli, req->rq_xid, qid, body->oa.o_valid,
2003                                 body->oa.o_flags);
2004         }
2005
2006         osc_update_grant(cli, body);
2007
2008         if (rc < 0)
2009                 RETURN(rc);
2010
2011         if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
2012                 client_cksum = aa->aa_oa->o_cksum; /* save for later */
2013
2014         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
2015                 if (rc > 0) {
2016                         CERROR("%s: unexpected positive size %d\n",
2017                                obd_name, rc);
2018                         RETURN(-EPROTO);
2019                 }
2020
2021                 if (req->rq_bulk != NULL &&
2022                     sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
2023                         RETURN(-EAGAIN);
2024
2025                 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
2026                     check_write_checksum(&body->oa, peer, client_cksum,
2027                                          body->oa.o_cksum, aa))
2028                         RETURN(-EAGAIN);
2029
2030                 rc = check_write_rcs(req, aa->aa_requested_nob,
2031                                      aa->aa_nio_count, aa->aa_page_count,
2032                                      aa->aa_ppga);
2033                 GOTO(out, rc);
2034         }
2035
2036         /* The rest of this function executes only for OST_READs */
2037
2038         if (req->rq_bulk == NULL) {
2039                 rc = req_capsule_get_size(&req->rq_pill, &RMF_SHORT_IO,
2040                                           RCL_SERVER);
2041                 LASSERT(rc == req->rq_status);
2042         } else {
2043                 /* if unwrap_bulk failed, return -EAGAIN to retry */
2044                 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
2045         }
2046         if (rc < 0)
2047                 GOTO(out, rc = -EAGAIN);
2048
2049         if (rc > aa->aa_requested_nob) {
2050                 CERROR("%s: unexpected size %d, requested %d\n", obd_name,
2051                        rc, aa->aa_requested_nob);
2052                 RETURN(-EPROTO);
2053         }
2054
2055         if (req->rq_bulk != NULL && rc != req->rq_bulk->bd_nob_transferred) {
2056                 CERROR("%s: unexpected size %d, transferred %d\n", obd_name,
2057                        rc, req->rq_bulk->bd_nob_transferred);
2058                 RETURN(-EPROTO);
2059         }
2060
2061         if (req->rq_bulk == NULL) {
2062                 /* short io */
2063                 int nob, pg_count, i = 0;
2064                 unsigned char *buf;
2065
2066                 CDEBUG(D_CACHE, "Using short io read, size %d\n", rc);
2067                 pg_count = aa->aa_page_count;
2068                 buf = req_capsule_server_sized_get(&req->rq_pill, &RMF_SHORT_IO,
2069                                                    rc);
2070                 nob = rc;
2071                 while (nob > 0 && pg_count > 0) {
2072                         unsigned char *ptr;
2073                         int count = aa->aa_ppga[i]->count > nob ?
2074                                     nob : aa->aa_ppga[i]->count;
2075
2076                         CDEBUG(D_CACHE, "page %p count %d\n",
2077                                aa->aa_ppga[i]->pg, count);
2078                         ptr = kmap_atomic(aa->aa_ppga[i]->pg);
2079                         memcpy(ptr + (aa->aa_ppga[i]->off & ~PAGE_MASK), buf,
2080                                count);
2081                         kunmap_atomic((void *) ptr);
2082
2083                         buf += count;
2084                         nob -= count;
2085                         i++;
2086                         pg_count--;
2087                 }
2088         }
2089
2090         if (rc < aa->aa_requested_nob)
2091                 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
2092
2093         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
2094                 static int cksum_counter;
2095                 u32 server_cksum = body->oa.o_cksum;
2096                 int nob = rc;
2097                 char *via = "";
2098                 char *router = "";
2099                 enum cksum_types cksum_type;
2100                 u32 o_flags = body->oa.o_valid & OBD_MD_FLFLAGS ?
2101                         body->oa.o_flags : 0;
2102
2103                 cksum_type = obd_cksum_type_unpack(o_flags);
2104                 rc = osc_checksum_bulk_rw(obd_name, cksum_type, nob,
2105                                           aa->aa_page_count, aa->aa_ppga,
2106                                           OST_READ, &client_cksum, false);
2107                 if (rc < 0)
2108                         GOTO(out, rc);
2109
2110                 if (req->rq_bulk != NULL &&
2111                     lnet_nid_to_nid4(&peer->nid) != req->rq_bulk->bd_sender) {
2112                         via = " via ";
2113                         router = libcfs_nid2str(req->rq_bulk->bd_sender);
2114                 }
2115
2116                 if (server_cksum != client_cksum) {
2117                         struct ost_body *clbody;
2118                         __u32 client_cksum2;
2119                         u32 page_count = aa->aa_page_count;
2120
2121                         osc_checksum_bulk_rw(obd_name, cksum_type, nob,
2122                                              page_count, aa->aa_ppga,
2123                                              OST_READ, &client_cksum2, true);
2124                         clbody = req_capsule_client_get(&req->rq_pill,
2125                                                         &RMF_OST_BODY);
2126                         if (cli->cl_checksum_dump)
2127                                 dump_all_bulk_pages(&clbody->oa, page_count,
2128                                                     aa->aa_ppga, server_cksum,
2129                                                     client_cksum);
2130
2131                         LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
2132                                            "%s%s%s inode "DFID" object "DOSTID
2133                                            " extent [%llu-%llu], client %x/%x, "
2134                                            "server %x, cksum_type %x\n",
2135                                            obd_name,
2136                                            libcfs_nidstr(&peer->nid),
2137                                            via, router,
2138                                            clbody->oa.o_valid & OBD_MD_FLFID ?
2139                                                 clbody->oa.o_parent_seq : 0ULL,
2140                                            clbody->oa.o_valid & OBD_MD_FLFID ?
2141                                                 clbody->oa.o_parent_oid : 0,
2142                                            clbody->oa.o_valid & OBD_MD_FLFID ?
2143                                                 clbody->oa.o_parent_ver : 0,
2144                                            POSTID(&body->oa.o_oi),
2145                                            aa->aa_ppga[0]->off,
2146                                            aa->aa_ppga[page_count-1]->off +
2147                                            aa->aa_ppga[page_count-1]->count - 1,
2148                                            client_cksum, client_cksum2,
2149                                            server_cksum, cksum_type);
2150                         cksum_counter = 0;
2151                         aa->aa_oa->o_cksum = client_cksum;
2152                         rc = -EAGAIN;
2153                 } else {
2154                         cksum_counter++;
2155                         CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
2156                         rc = 0;
2157                 }
2158         } else if (unlikely(client_cksum)) {
2159                 static int cksum_missed;
2160
2161                 cksum_missed++;
2162                 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
2163                         CERROR("%s: checksum %u requested from %s but not sent\n",
2164                                obd_name, cksum_missed,
2165                                libcfs_nidstr(&peer->nid));
2166         } else {
2167                 rc = 0;
2168         }
2169
2170         /* get the inode from the first cl_page */
2171         clpage = oap2cl_page(brw_page2oap(aa->aa_ppga[0]));
2172         inode = clpage->cp_inode;
2173         if (clpage->cp_type == CPT_TRANSIENT && inode) {
2174                 blockbits = inode->i_blkbits;
2175                 blocksize = 1 << blockbits;
2176         }
2177         if (inode && IS_ENCRYPTED(inode)) {
2178                 int idx;
2179
2180                 if (!llcrypt_has_encryption_key(inode)) {
2181                         CDEBUG(D_SEC, "no enc key for ino %lu\n", inode->i_ino);
2182                         GOTO(out, rc);
2183                 }
2184                 for (idx = 0; idx < aa->aa_page_count; idx++) {
2185                         struct brw_page *brwpg = aa->aa_ppga[idx];
2186                         unsigned int offs = 0;
2187
2188                         while (offs < PAGE_SIZE) {
2189                                 /* do not decrypt if page is all 0s */
2190                                 if (memchr_inv(page_address(brwpg->pg) + offs,
2191                                       0, LUSTRE_ENCRYPTION_UNIT_SIZE) == NULL) {
2192                                         /* if page is empty forward info to
2193                                          * upper layers (ll_io_zero_page) by
2194                                          * clearing PagePrivate2
2195                                          */
2196                                         if (!offs)
2197                                                 ClearPagePrivate2(brwpg->pg);
2198                                         break;
2199                                 }
2200
2201                                 if (blockbits) {
2202                                         /* This is direct IO case. Directly call
2203                                          * decrypt function that takes inode as
2204                                          * input parameter. Page does not need
2205                                          * to be locked.
2206                                          */
2207                                         u64 lblk_num;
2208                                         unsigned int i;
2209
2210                                         clpage =
2211                                                oap2cl_page(brw_page2oap(brwpg));
2212                                         lblk_num =
2213                                                 ((u64)(clpage->cp_page_index) <<
2214                                                 (PAGE_SHIFT - blockbits)) +
2215                                                 (offs >> blockbits);
2216                                         for (i = offs;
2217                                              i < offs +
2218                                                     LUSTRE_ENCRYPTION_UNIT_SIZE;
2219                                              i += blocksize, lblk_num++) {
2220                                                 rc =
2221                                                   llcrypt_decrypt_block_inplace(
2222                                                           inode, brwpg->pg,
2223                                                           blocksize, i,
2224                                                           lblk_num);
2225                                                 if (rc)
2226                                                         break;
2227                                         }
2228                                 } else {
2229                                         rc = llcrypt_decrypt_pagecache_blocks(
2230                                                 brwpg->pg,
2231                                                 LUSTRE_ENCRYPTION_UNIT_SIZE,
2232                                                 offs);
2233                                 }
2234                                 if (rc)
2235                                         GOTO(out, rc);
2236
2237                                 offs += LUSTRE_ENCRYPTION_UNIT_SIZE;
2238                         }
2239                 }
2240         }
2241
2242 out:
2243         if (rc >= 0)
2244                 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
2245                                      aa->aa_oa, &body->oa);
2246
2247         RETURN(rc);
2248 }
2249
2250 static int osc_brw_redo_request(struct ptlrpc_request *request,
2251                                 struct osc_brw_async_args *aa, int rc)
2252 {
2253         struct ptlrpc_request *new_req;
2254         struct osc_brw_async_args *new_aa;
2255         struct osc_async_page *oap;
2256         ENTRY;
2257
2258         /* The below message is checked in replay-ost-single.sh test_8ae*/
2259         DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
2260                   "redo for recoverable error %d", rc);
2261
2262         rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
2263                                 OST_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
2264                                   aa->aa_cli, aa->aa_oa, aa->aa_page_count,
2265                                   aa->aa_ppga, &new_req, 1);
2266         if (rc)
2267                 RETURN(rc);
2268
2269         list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2270                 if (oap->oap_request != NULL) {
2271                         LASSERTF(request == oap->oap_request,
2272                                  "request %p != oap_request %p\n",
2273                                  request, oap->oap_request);
2274                 }
2275         }
2276         /*
2277          * New request takes over pga and oaps from old request.
2278          * Note that copying a list_head doesn't work, need to move it...
2279          */
2280         aa->aa_resends++;
2281         new_req->rq_interpret_reply = request->rq_interpret_reply;
2282         new_req->rq_async_args = request->rq_async_args;
2283         new_req->rq_commit_cb = request->rq_commit_cb;
2284         /* cap resend delay to the current request timeout, this is similar to
2285          * what ptlrpc does (see after_reply()) */
2286         if (aa->aa_resends > new_req->rq_timeout)
2287                 new_req->rq_sent = ktime_get_real_seconds() + new_req->rq_timeout;
2288         else
2289                 new_req->rq_sent = ktime_get_real_seconds() + aa->aa_resends;
2290         new_req->rq_generation_set = 1;
2291         new_req->rq_import_generation = request->rq_import_generation;
2292
2293         new_aa = ptlrpc_req_async_args(new_aa, new_req);
2294
2295         INIT_LIST_HEAD(&new_aa->aa_oaps);
2296         list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
2297         INIT_LIST_HEAD(&new_aa->aa_exts);
2298         list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
2299         new_aa->aa_resends = aa->aa_resends;
2300
2301         list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
2302                 if (oap->oap_request) {
2303                         ptlrpc_req_finished(oap->oap_request);
2304                         oap->oap_request = ptlrpc_request_addref(new_req);
2305                 }
2306         }
2307
2308         /* XXX: This code will run into problem if we're going to support
2309          * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
2310          * and wait for all of them to be finished. We should inherit request
2311          * set from old request. */
2312         ptlrpcd_add_req(new_req);
2313
2314         DEBUG_REQ(D_INFO, new_req, "new request");
2315         RETURN(0);
2316 }
2317
2318 /*
2319  * ugh, we want disk allocation on the target to happen in offset order.  we'll
2320  * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
2321  * fine for our small page arrays and doesn't require allocation.  its an
2322  * insertion sort that swaps elements that are strides apart, shrinking the
2323  * stride down until its '1' and the array is sorted.
2324  */
2325 static void sort_brw_pages(struct brw_page **array, int num)
2326 {
2327         int stride, i, j;
2328         struct brw_page *tmp;
2329
2330         if (num == 1)
2331                 return;
2332         for (stride = 1; stride < num ; stride = (stride * 3) + 1)
2333                 ;
2334
2335         do {
2336                 stride /= 3;
2337                 for (i = stride ; i < num ; i++) {
2338                         tmp = array[i];
2339                         j = i;
2340                         while (j >= stride && array[j - stride]->off > tmp->off) {
2341                                 array[j] = array[j - stride];
2342                                 j -= stride;
2343                         }
2344                         array[j] = tmp;
2345                 }
2346         } while (stride > 1);
2347 }
2348
2349 static void osc_release_ppga(struct brw_page **ppga, size_t count)
2350 {
2351         LASSERT(ppga != NULL);
2352         OBD_FREE_PTR_ARRAY_LARGE(ppga, count);
2353 }
2354
2355 static int brw_interpret(const struct lu_env *env,
2356                          struct ptlrpc_request *req, void *args, int rc)
2357 {
2358         struct osc_brw_async_args *aa = args;
2359         struct osc_extent *ext;
2360         struct osc_extent *tmp;
2361         struct client_obd *cli = aa->aa_cli;
2362         unsigned long transferred = 0;
2363
2364         ENTRY;
2365
2366         rc = osc_brw_fini_request(req, rc);
2367         CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2368
2369         /* restore clear text pages */
2370         osc_release_bounce_pages(aa->aa_ppga, aa->aa_page_count);
2371
2372         /*
2373          * When server returns -EINPROGRESS, client should always retry
2374          * regardless of the number of times the bulk was resent already.
2375          */
2376         if (osc_recoverable_error(rc) && !req->rq_no_delay) {
2377                 if (req->rq_import_generation !=
2378                     req->rq_import->imp_generation) {
2379                         CDEBUG(D_HA, "%s: resend cross eviction for object: "
2380                                ""DOSTID", rc = %d.\n",
2381                                req->rq_import->imp_obd->obd_name,
2382                                POSTID(&aa->aa_oa->o_oi), rc);
2383                 } else if (rc == -EINPROGRESS ||
2384                            client_should_resend(aa->aa_resends, aa->aa_cli)) {
2385                         rc = osc_brw_redo_request(req, aa, rc);
2386                 } else {
2387                         CERROR("%s: too many resent retries for object: "
2388                                "%llu:%llu, rc = %d.\n",
2389                                req->rq_import->imp_obd->obd_name,
2390                                POSTID(&aa->aa_oa->o_oi), rc);
2391                 }
2392
2393                 if (rc == 0)
2394                         RETURN(0);
2395                 else if (rc == -EAGAIN || rc == -EINPROGRESS)
2396                         rc = -EIO;
2397         }
2398
2399         if (rc == 0) {
2400                 struct obdo *oa = aa->aa_oa;
2401                 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
2402                 unsigned long valid = 0;
2403                 struct cl_object *obj;
2404                 struct osc_async_page *last;
2405
2406                 last = brw_page2oap(aa->aa_ppga[aa->aa_page_count - 1]);
2407                 obj = osc2cl(last->oap_obj);
2408
2409                 cl_object_attr_lock(obj);
2410                 if (oa->o_valid & OBD_MD_FLBLOCKS) {
2411                         attr->cat_blocks = oa->o_blocks;
2412                         valid |= CAT_BLOCKS;
2413                 }
2414                 if (oa->o_valid & OBD_MD_FLMTIME) {
2415                         attr->cat_mtime = oa->o_mtime;
2416                         valid |= CAT_MTIME;
2417                 }
2418                 if (oa->o_valid & OBD_MD_FLATIME) {
2419                         attr->cat_atime = oa->o_atime;
2420                         valid |= CAT_ATIME;
2421                 }
2422                 if (oa->o_valid & OBD_MD_FLCTIME) {
2423                         attr->cat_ctime = oa->o_ctime;
2424                         valid |= CAT_CTIME;
2425                 }
2426
2427                 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
2428                         struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
2429                         loff_t last_off = last->oap_count + last->oap_obj_off +
2430                                 last->oap_page_off;
2431
2432                         /* Change file size if this is an out of quota or
2433                          * direct IO write and it extends the file size */
2434                         if (loi->loi_lvb.lvb_size < last_off) {
2435                                 attr->cat_size = last_off;
2436                                 valid |= CAT_SIZE;
2437                         }
2438                         /* Extend KMS if it's not a lockless write */
2439                         if (loi->loi_kms < last_off &&
2440                             oap2osc_page(last)->ops_srvlock == 0) {
2441                                 attr->cat_kms = last_off;
2442                                 valid |= CAT_KMS;
2443                         }
2444                 }
2445
2446                 if (valid != 0)
2447                         cl_object_attr_update(env, obj, attr, valid);
2448                 cl_object_attr_unlock(obj);
2449         }
2450         OBD_SLAB_FREE_PTR(aa->aa_oa, osc_obdo_kmem);
2451         aa->aa_oa = NULL;
2452
2453         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE && rc == 0)
2454                 osc_inc_unstable_pages(req);
2455
2456         list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
2457                 list_del_init(&ext->oe_link);
2458                 osc_extent_finish(env, ext, 1,
2459                                   rc && req->rq_no_delay ? -EAGAIN : rc);
2460         }
2461         LASSERT(list_empty(&aa->aa_exts));
2462         LASSERT(list_empty(&aa->aa_oaps));
2463
2464         transferred = (req->rq_bulk == NULL ? /* short io */
2465                        aa->aa_requested_nob :
2466                        req->rq_bulk->bd_nob_transferred);
2467
2468         osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2469         ptlrpc_lprocfs_brw(req, transferred);
2470
2471         spin_lock(&cli->cl_loi_list_lock);
2472         /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2473          * is called so we know whether to go to sync BRWs or wait for more
2474          * RPCs to complete */
2475         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2476                 cli->cl_w_in_flight--;
2477         else
2478                 cli->cl_r_in_flight--;
2479         osc_wake_cache_waiters(cli);
2480         spin_unlock(&cli->cl_loi_list_lock);
2481
2482         osc_io_unplug(env, cli, NULL);
2483         RETURN(rc);
2484 }
2485
2486 static void brw_commit(struct ptlrpc_request *req)
2487 {
2488         /* If osc_inc_unstable_pages (via osc_extent_finish) races with
2489          * this called via the rq_commit_cb, I need to ensure
2490          * osc_dec_unstable_pages is still called. Otherwise unstable
2491          * pages may be leaked. */
2492         spin_lock(&req->rq_lock);
2493         if (likely(req->rq_unstable)) {
2494                 req->rq_unstable = 0;
2495                 spin_unlock(&req->rq_lock);
2496
2497                 osc_dec_unstable_pages(req);
2498         } else {
2499                 req->rq_committed = 1;
2500                 spin_unlock(&req->rq_lock);
2501         }
2502 }
2503
2504 /**
2505  * Build an RPC by the list of extent @ext_list. The caller must ensure
2506  * that the total pages in this list are NOT over max pages per RPC.
2507  * Extents in the list must be in OES_RPC state.
2508  */
2509 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
2510                   struct list_head *ext_list, int cmd)
2511 {
2512         struct ptlrpc_request           *req = NULL;
2513         struct osc_extent               *ext;
2514         struct brw_page                 **pga = NULL;
2515         struct osc_brw_async_args       *aa = NULL;
2516         struct obdo                     *oa = NULL;
2517         struct osc_async_page           *oap;
2518         struct osc_object               *obj = NULL;
2519         struct cl_req_attr              *crattr = NULL;
2520         loff_t                          starting_offset = OBD_OBJECT_EOF;
2521         loff_t                          ending_offset = 0;
2522         /* '1' for consistency with code that checks !mpflag to restore */
2523         int mpflag = 1;
2524         int                             mem_tight = 0;
2525         int                             page_count = 0;
2526         bool                            soft_sync = false;
2527         bool                            ndelay = false;
2528         int                             i;
2529         int                             grant = 0;
2530         int                             rc;
2531         __u32                           layout_version = 0;
2532         LIST_HEAD(rpc_list);
2533         struct ost_body                 *body;
2534         ENTRY;
2535         LASSERT(!list_empty(ext_list));
2536
2537         /* add pages into rpc_list to build BRW rpc */
2538         list_for_each_entry(ext, ext_list, oe_link) {
2539                 LASSERT(ext->oe_state == OES_RPC);
2540                 mem_tight |= ext->oe_memalloc;
2541                 grant += ext->oe_grants;
2542                 page_count += ext->oe_nr_pages;
2543                 layout_version = max(layout_version, ext->oe_layout_version);
2544                 if (obj == NULL)
2545                         obj = ext->oe_obj;
2546         }
2547
2548         soft_sync = osc_over_unstable_soft_limit(cli);
2549         if (mem_tight)
2550                 mpflag = memalloc_noreclaim_save();
2551
2552         OBD_ALLOC_PTR_ARRAY_LARGE(pga, page_count);
2553         if (pga == NULL)
2554                 GOTO(out, rc = -ENOMEM);
2555
2556         OBD_SLAB_ALLOC_PTR_GFP(oa, osc_obdo_kmem, GFP_NOFS);
2557         if (oa == NULL)
2558                 GOTO(out, rc = -ENOMEM);
2559
2560         i = 0;
2561         list_for_each_entry(ext, ext_list, oe_link) {
2562                 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
2563                         if (mem_tight)
2564                                 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2565                         if (soft_sync)
2566                                 oap->oap_brw_flags |= OBD_BRW_SOFT_SYNC;
2567                         pga[i] = &oap->oap_brw_page;
2568                         pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2569                         i++;
2570
2571                         list_add_tail(&oap->oap_rpc_item, &rpc_list);
2572                         if (starting_offset == OBD_OBJECT_EOF ||
2573                             starting_offset > oap->oap_obj_off)
2574                                 starting_offset = oap->oap_obj_off;
2575                         else
2576                                 LASSERT(oap->oap_page_off == 0);
2577                         if (ending_offset < oap->oap_obj_off + oap->oap_count)
2578                                 ending_offset = oap->oap_obj_off +
2579                                                 oap->oap_count;
2580                         else
2581                                 LASSERT(oap->oap_page_off + oap->oap_count ==
2582                                         PAGE_SIZE);
2583                 }
2584                 if (ext->oe_ndelay)
2585                         ndelay = true;
2586         }
2587
2588         /* first page in the list */
2589         oap = list_first_entry(&rpc_list, typeof(*oap), oap_rpc_item);
2590
2591         crattr = &osc_env_info(env)->oti_req_attr;
2592         memset(crattr, 0, sizeof(*crattr));
2593         crattr->cra_type = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2594         crattr->cra_flags = ~0ULL;
2595         crattr->cra_page = oap2cl_page(oap);
2596         crattr->cra_oa = oa;
2597         cl_req_attr_set(env, osc2cl(obj), crattr);
2598
2599         if (cmd == OBD_BRW_WRITE) {
2600                 oa->o_grant_used = grant;
2601                 if (layout_version > 0) {
2602                         CDEBUG(D_LAYOUT, DFID": write with layout version %u\n",
2603                                PFID(&oa->o_oi.oi_fid), layout_version);
2604
2605                         oa->o_layout_version = layout_version;
2606                         oa->o_valid |= OBD_MD_LAYOUT_VERSION;
2607                 }
2608         }
2609
2610         sort_brw_pages(pga, page_count);
2611         rc = osc_brw_prep_request(cmd, cli, oa, page_count, pga, &req, 0);
2612         if (rc != 0) {
2613                 CERROR("prep_req failed: %d\n", rc);
2614                 GOTO(out, rc);
2615         }
2616
2617         req->rq_commit_cb = brw_commit;
2618         req->rq_interpret_reply = brw_interpret;
2619         req->rq_memalloc = mem_tight != 0;
2620         oap->oap_request = ptlrpc_request_addref(req);
2621         if (ndelay) {
2622                 req->rq_no_resend = req->rq_no_delay = 1;
2623                 /* probably set a shorter timeout value.
2624                  * to handle ETIMEDOUT in brw_interpret() correctly. */
2625                 /* lustre_msg_set_timeout(req, req->rq_timeout / 2); */
2626         }
2627
2628         /* Need to update the timestamps after the request is built in case
2629          * we race with setattr (locally or in queue at OST).  If OST gets
2630          * later setattr before earlier BRW (as determined by the request xid),
2631          * the OST will not use BRW timestamps.  Sadly, there is no obvious
2632          * way to do this in a single call.  bug 10150 */
2633         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2634         crattr->cra_oa = &body->oa;
2635         crattr->cra_flags = OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLATIME;
2636         cl_req_attr_set(env, osc2cl(obj), crattr);
2637         lustre_msg_set_jobid(req->rq_reqmsg, crattr->cra_jobid);
2638
2639         aa = ptlrpc_req_async_args(aa, req);
2640         INIT_LIST_HEAD(&aa->aa_oaps);
2641         list_splice_init(&rpc_list, &aa->aa_oaps);
2642         INIT_LIST_HEAD(&aa->aa_exts);
2643         list_splice_init(ext_list, &aa->aa_exts);
2644
2645         spin_lock(&cli->cl_loi_list_lock);
2646         starting_offset >>= PAGE_SHIFT;
2647         if (cmd == OBD_BRW_READ) {
2648                 cli->cl_r_in_flight++;
2649                 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2650                 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2651                 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2652                                       starting_offset + 1);
2653         } else {
2654                 cli->cl_w_in_flight++;
2655                 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2656                 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
2657                 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2658                                       starting_offset + 1);
2659         }
2660         spin_unlock(&cli->cl_loi_list_lock);
2661
2662         DEBUG_REQ(D_INODE, req, "%d pages, aa %p, now %ur/%uw in flight",
2663                   page_count, aa, cli->cl_r_in_flight,
2664                   cli->cl_w_in_flight);
2665         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_IO, cfs_fail_val);
2666
2667         ptlrpcd_add_req(req);
2668         rc = 0;
2669         EXIT;
2670
2671 out:
2672         if (mem_tight)
2673                 memalloc_noreclaim_restore(mpflag);
2674
2675         if (rc != 0) {
2676                 LASSERT(req == NULL);
2677
2678                 if (oa)
2679                         OBD_SLAB_FREE_PTR(oa, osc_obdo_kmem);
2680                 if (pga) {
2681                         osc_release_bounce_pages(pga, page_count);
2682                         osc_release_ppga(pga, page_count);
2683                 }
2684                 /* this should happen rarely and is pretty bad, it makes the
2685                  * pending list not follow the dirty order
2686                  */
2687                 while ((ext = list_first_entry_or_null(ext_list,
2688                                                        struct osc_extent,
2689                                                        oe_link)) != NULL) {
2690                         list_del_init(&ext->oe_link);
2691                         osc_extent_finish(env, ext, 0, rc);
2692                 }
2693         }
2694         RETURN(rc);
2695 }
2696
2697 /* This is to refresh our lock in face of no RPCs. */
2698 void osc_send_empty_rpc(struct osc_object *osc, pgoff_t start)
2699 {
2700         struct ptlrpc_request *req;
2701         struct obdo oa;
2702         struct brw_page bpg = { .off = start, .count = 1};
2703         struct brw_page *pga = &bpg;
2704         int rc;
2705
2706         memset(&oa, 0, sizeof(oa));
2707         oa.o_oi = osc->oo_oinfo->loi_oi;
2708         oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP | OBD_MD_FLFLAGS;
2709         /* For updated servers - don't do a read */
2710         oa.o_flags = OBD_FL_NORPC;
2711
2712         rc = osc_brw_prep_request(OBD_BRW_READ, osc_cli(osc), &oa, 1, &pga,
2713                                   &req, 0);
2714
2715         /* If we succeeded we ship it off, if not there's no point in doing
2716          * anything. Also no resends.
2717          * No interpret callback, no commit callback.
2718          */
2719         if (!rc) {
2720                 req->rq_no_resend = 1;
2721                 ptlrpcd_add_req(req);
2722         }
2723 }
2724
2725 static int osc_set_lock_data(struct ldlm_lock *lock, void *data)
2726 {
2727         int set = 0;
2728
2729         LASSERT(lock != NULL);
2730
2731         lock_res_and_lock(lock);
2732
2733         if (lock->l_ast_data == NULL)
2734                 lock->l_ast_data = data;
2735         if (lock->l_ast_data == data)
2736                 set = 1;
2737
2738         unlock_res_and_lock(lock);
2739
2740         return set;
2741 }
2742
2743 int osc_enqueue_fini(struct ptlrpc_request *req, osc_enqueue_upcall_f upcall,
2744                      void *cookie, struct lustre_handle *lockh,
2745                      enum ldlm_mode mode, __u64 *flags, bool speculative,
2746                      int errcode)
2747 {
2748         bool intent = *flags & LDLM_FL_HAS_INTENT;
2749         int rc;
2750         ENTRY;
2751
2752         /* The request was created before ldlm_cli_enqueue call. */
2753         if (intent && errcode == ELDLM_LOCK_ABORTED) {
2754                 struct ldlm_reply *rep;
2755
2756                 rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2757                 LASSERT(rep != NULL);
2758
2759                 rep->lock_policy_res1 =
2760                         ptlrpc_status_ntoh(rep->lock_policy_res1);
2761                 if (rep->lock_policy_res1)
2762                         errcode = rep->lock_policy_res1;
2763                 if (!speculative)
2764                         *flags |= LDLM_FL_LVB_READY;
2765         } else if (errcode == ELDLM_OK) {
2766                 *flags |= LDLM_FL_LVB_READY;
2767         }
2768
2769         /* Call the update callback. */
2770         rc = (*upcall)(cookie, lockh, errcode);
2771
2772         /* release the reference taken in ldlm_cli_enqueue() */
2773         if (errcode == ELDLM_LOCK_MATCHED)
2774                 errcode = ELDLM_OK;
2775         if (errcode == ELDLM_OK && lustre_handle_is_used(lockh))
2776                 ldlm_lock_decref(lockh, mode);
2777
2778         RETURN(rc);
2779 }
2780
2781 int osc_enqueue_interpret(const struct lu_env *env, struct ptlrpc_request *req,
2782                           void *args, int rc)
2783 {
2784         struct osc_enqueue_args *aa = args;
2785         struct ldlm_lock *lock;
2786         struct lustre_handle *lockh = &aa->oa_lockh;
2787         enum ldlm_mode mode = aa->oa_mode;
2788         struct ost_lvb *lvb = aa->oa_lvb;
2789         __u32 lvb_len = sizeof(*lvb);
2790         __u64 flags = 0;
2791         struct ldlm_enqueue_info einfo = {
2792                 .ei_type = aa->oa_type,
2793                 .ei_mode = mode,
2794         };
2795
2796         ENTRY;
2797
2798         /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2799          * be valid. */
2800         lock = ldlm_handle2lock(lockh);
2801         LASSERTF(lock != NULL,
2802                  "lockh %#llx, req %p, aa %p - client evicted?\n",
2803                  lockh->cookie, req, aa);
2804
2805         /* Take an additional reference so that a blocking AST that
2806          * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2807          * to arrive after an upcall has been executed by
2808          * osc_enqueue_fini(). */
2809         ldlm_lock_addref(lockh, mode);
2810
2811         /* Let cl_lock_state_wait fail with -ERESTARTSYS to unuse sublocks. */
2812         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_HANG, 2);
2813
2814         /* Let CP AST to grant the lock first. */
2815         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
2816
2817         if (aa->oa_speculative) {
2818                 LASSERT(aa->oa_lvb == NULL);
2819                 LASSERT(aa->oa_flags == NULL);
2820                 aa->oa_flags = &flags;
2821         }
2822
2823         /* Complete obtaining the lock procedure. */
2824         rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, &einfo, 1, aa->oa_flags,
2825                                    lvb, lvb_len, lockh, rc, false);
2826         /* Complete osc stuff. */
2827         rc = osc_enqueue_fini(req, aa->oa_upcall, aa->oa_cookie, lockh, mode,
2828                               aa->oa_flags, aa->oa_speculative, rc);
2829
2830         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
2831
2832         ldlm_lock_decref(lockh, mode);
2833         LDLM_LOCK_PUT(lock);
2834         RETURN(rc);
2835 }
2836
2837 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2838  * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2839  * other synchronous requests, however keeping some locks and trying to obtain
2840  * others may take a considerable amount of time in a case of ost failure; and
2841  * when other sync requests do not get released lock from a client, the client
2842  * is evicted from the cluster -- such scenarious make the life difficult, so
2843  * release locks just after they are obtained. */
2844 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2845                      __u64 *flags, union ldlm_policy_data *policy,
2846                      struct ost_lvb *lvb, osc_enqueue_upcall_f upcall,
2847                      void *cookie, struct ldlm_enqueue_info *einfo,
2848                      struct ptlrpc_request_set *rqset, int async,
2849                      bool speculative)
2850 {
2851         struct obd_device *obd = exp->exp_obd;
2852         struct lustre_handle lockh = { 0 };
2853         struct ptlrpc_request *req = NULL;
2854         int intent = *flags & LDLM_FL_HAS_INTENT;
2855         __u64 match_flags = *flags;
2856         enum ldlm_mode mode;
2857         int rc;
2858         ENTRY;
2859
2860         /* Filesystem lock extents are extended to page boundaries so that
2861          * dealing with the page cache is a little smoother.  */
2862         policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2863         policy->l_extent.end |= ~PAGE_MASK;
2864
2865         /* Next, search for already existing extent locks that will cover us */
2866         /* If we're trying to read, we also search for an existing PW lock.  The
2867          * VFS and page cache already protect us locally, so lots of readers/
2868          * writers can share a single PW lock.
2869          *
2870          * There are problems with conversion deadlocks, so instead of
2871          * converting a read lock to a write lock, we'll just enqueue a new
2872          * one.
2873          *
2874          * At some point we should cancel the read lock instead of making them
2875          * send us a blocking callback, but there are problems with canceling
2876          * locks out from other users right now, too. */
2877         mode = einfo->ei_mode;
2878         if (einfo->ei_mode == LCK_PR)
2879                 mode |= LCK_PW;
2880         /* Normal lock requests must wait for the LVB to be ready before
2881          * matching a lock; speculative lock requests do not need to,
2882          * because they will not actually use the lock. */
2883         if (!speculative)
2884                 match_flags |= LDLM_FL_LVB_READY;
2885         if (intent != 0)
2886                 match_flags |= LDLM_FL_BLOCK_GRANTED;
2887         mode = ldlm_lock_match(obd->obd_namespace, match_flags, res_id,
2888                                einfo->ei_type, policy, mode, &lockh);
2889         if (mode) {
2890                 struct ldlm_lock *matched;
2891
2892                 if (*flags & LDLM_FL_TEST_LOCK)
2893                         RETURN(ELDLM_OK);
2894
2895                 matched = ldlm_handle2lock(&lockh);
2896                 if (speculative) {
2897                         /* This DLM lock request is speculative, and does not
2898                          * have an associated IO request. Therefore if there
2899                          * is already a DLM lock, it wll just inform the
2900                          * caller to cancel the request for this stripe.*/
2901                         lock_res_and_lock(matched);
2902                         if (ldlm_extent_equal(&policy->l_extent,
2903                             &matched->l_policy_data.l_extent))
2904                                 rc = -EEXIST;
2905                         else
2906                                 rc = -ECANCELED;
2907                         unlock_res_and_lock(matched);
2908
2909                         ldlm_lock_decref(&lockh, mode);
2910                         LDLM_LOCK_PUT(matched);
2911                         RETURN(rc);
2912                 } else if (osc_set_lock_data(matched, einfo->ei_cbdata)) {
2913                         *flags |= LDLM_FL_LVB_READY;
2914
2915                         /* We already have a lock, and it's referenced. */
2916                         (*upcall)(cookie, &lockh, ELDLM_LOCK_MATCHED);
2917
2918                         ldlm_lock_decref(&lockh, mode);
2919                         LDLM_LOCK_PUT(matched);
2920                         RETURN(ELDLM_OK);
2921                 } else {
2922                         ldlm_lock_decref(&lockh, mode);
2923                         LDLM_LOCK_PUT(matched);
2924                 }
2925         }
2926
2927         if (*flags & (LDLM_FL_TEST_LOCK | LDLM_FL_MATCH_LOCK))
2928                 RETURN(-ENOLCK);
2929
2930         /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2931         *flags &= ~LDLM_FL_BLOCK_GRANTED;
2932
2933         rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2934                               sizeof(*lvb), LVB_T_OST, &lockh, async);
2935         if (async) {
2936                 if (!rc) {
2937                         struct osc_enqueue_args *aa;
2938                         aa = ptlrpc_req_async_args(aa, req);
2939                         aa->oa_exp         = exp;
2940                         aa->oa_mode        = einfo->ei_mode;
2941                         aa->oa_type        = einfo->ei_type;
2942                         lustre_handle_copy(&aa->oa_lockh, &lockh);
2943                         aa->oa_upcall      = upcall;
2944                         aa->oa_cookie      = cookie;
2945                         aa->oa_speculative = speculative;
2946                         if (!speculative) {
2947                                 aa->oa_flags  = flags;
2948                                 aa->oa_lvb    = lvb;
2949                         } else {
2950                                 /* speculative locks are essentially to enqueue
2951                                  * a DLM lock  in advance, so we don't care
2952                                  * about the result of the enqueue. */
2953                                 aa->oa_lvb    = NULL;
2954                                 aa->oa_flags  = NULL;
2955                         }
2956
2957                         req->rq_interpret_reply = osc_enqueue_interpret;
2958                         ptlrpc_set_add_req(rqset, req);
2959                 }
2960                 RETURN(rc);
2961         }
2962
2963         rc = osc_enqueue_fini(req, upcall, cookie, &lockh, einfo->ei_mode,
2964                               flags, speculative, rc);
2965
2966         RETURN(rc);
2967 }
2968
2969 int osc_match_base(const struct lu_env *env, struct obd_export *exp,
2970                    struct ldlm_res_id *res_id, enum ldlm_type type,
2971                    union ldlm_policy_data *policy, enum ldlm_mode mode,
2972                    __u64 *flags, struct osc_object *obj,
2973                    struct lustre_handle *lockh, enum ldlm_match_flags match_flags)
2974 {
2975         struct obd_device *obd = exp->exp_obd;
2976         __u64 lflags = *flags;
2977         enum ldlm_mode rc;
2978         ENTRY;
2979
2980         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2981                 RETURN(-EIO);
2982
2983         /* Filesystem lock extents are extended to page boundaries so that
2984          * dealing with the page cache is a little smoother */
2985         policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2986         policy->l_extent.end |= ~PAGE_MASK;
2987
2988         /* Next, search for already existing extent locks that will cover us */
2989         rc = ldlm_lock_match_with_skip(obd->obd_namespace, lflags, 0,
2990                                         res_id, type, policy, mode, lockh,
2991                                         match_flags);
2992         if (rc == 0 || lflags & LDLM_FL_TEST_LOCK)
2993                 RETURN(rc);
2994
2995         if (obj != NULL) {
2996                 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2997
2998                 LASSERT(lock != NULL);
2999                 if (osc_set_lock_data(lock, obj)) {
3000                         lock_res_and_lock(lock);
3001                         if (!ldlm_is_lvb_cached(lock)) {
3002                                 LASSERT(lock->l_ast_data == obj);
3003                                 osc_lock_lvb_update(env, obj, lock, NULL);
3004                                 ldlm_set_lvb_cached(lock);
3005                         }
3006                         unlock_res_and_lock(lock);
3007                 } else {
3008                         ldlm_lock_decref(lockh, rc);
3009                         rc = 0;
3010                 }
3011                 LDLM_LOCK_PUT(lock);
3012         }
3013         RETURN(rc);
3014 }
3015
3016 static int osc_statfs_interpret(const struct lu_env *env,
3017                                 struct ptlrpc_request *req, void *args, int rc)
3018 {
3019         struct osc_async_args *aa = args;
3020         struct obd_statfs *msfs;
3021
3022         ENTRY;
3023         if (rc == -EBADR)
3024                 /*
3025                  * The request has in fact never been sent due to issues at
3026                  * a higher level (LOV).  Exit immediately since the caller
3027                  * is aware of the problem and takes care of the clean up.
3028                  */
3029                 RETURN(rc);
3030
3031         if ((rc == -ENOTCONN || rc == -EAGAIN) &&
3032             (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
3033                 GOTO(out, rc = 0);
3034
3035         if (rc != 0)
3036                 GOTO(out, rc);
3037
3038         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3039         if (msfs == NULL)
3040                 GOTO(out, rc = -EPROTO);
3041
3042         *aa->aa_oi->oi_osfs = *msfs;
3043 out:
3044         rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
3045
3046         RETURN(rc);
3047 }
3048
3049 static int osc_statfs_async(struct obd_export *exp,
3050                             struct obd_info *oinfo, time64_t max_age,
3051                             struct ptlrpc_request_set *rqset)
3052 {
3053         struct obd_device     *obd = class_exp2obd(exp);
3054         struct ptlrpc_request *req;
3055         struct osc_async_args *aa;
3056         int rc;
3057         ENTRY;
3058
3059         if (obd->obd_osfs_age >= max_age) {
3060                 CDEBUG(D_SUPER,
3061                        "%s: use %p cache blocks %llu/%llu objects %llu/%llu\n",
3062                        obd->obd_name, &obd->obd_osfs,
3063                        obd->obd_osfs.os_bavail, obd->obd_osfs.os_blocks,
3064                        obd->obd_osfs.os_ffree, obd->obd_osfs.os_files);
3065                 spin_lock(&obd->obd_osfs_lock);
3066                 memcpy(oinfo->oi_osfs, &obd->obd_osfs, sizeof(*oinfo->oi_osfs));
3067                 spin_unlock(&obd->obd_osfs_lock);
3068                 oinfo->oi_flags |= OBD_STATFS_FROM_CACHE;
3069                 if (oinfo->oi_cb_up)
3070                         oinfo->oi_cb_up(oinfo, 0);
3071
3072                 RETURN(0);
3073         }
3074
3075         /* We could possibly pass max_age in the request (as an absolute
3076          * timestamp or a "seconds.usec ago") so the target can avoid doing
3077          * extra calls into the filesystem if that isn't necessary (e.g.
3078          * during mount that would help a bit).  Having relative timestamps
3079          * is not so great if request processing is slow, while absolute
3080          * timestamps are not ideal because they need time synchronization. */
3081         req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
3082         if (req == NULL)
3083                 RETURN(-ENOMEM);
3084
3085         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3086         if (rc) {
3087                 ptlrpc_request_free(req);
3088                 RETURN(rc);
3089         }
3090         ptlrpc_request_set_replen(req);
3091         req->rq_request_portal = OST_CREATE_PORTAL;
3092         ptlrpc_at_set_req_timeout(req);
3093
3094         if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
3095                 /* procfs requests not want stat in wait for avoid deadlock */
3096                 req->rq_no_resend = 1;
3097                 req->rq_no_delay = 1;
3098         }
3099
3100         req->rq_interpret_reply = osc_statfs_interpret;
3101         aa = ptlrpc_req_async_args(aa, req);
3102         aa->aa_oi = oinfo;
3103
3104         ptlrpc_set_add_req(rqset, req);
3105         RETURN(0);
3106 }
3107
3108 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
3109                       struct obd_statfs *osfs, time64_t max_age, __u32 flags)
3110 {
3111         struct obd_device     *obd = class_exp2obd(exp);
3112         struct obd_statfs     *msfs;
3113         struct ptlrpc_request *req;
3114         struct obd_import     *imp, *imp0;
3115         int rc;
3116         ENTRY;
3117
3118         /*Since the request might also come from lprocfs, so we need
3119          *sync this with client_disconnect_export Bug15684
3120          */
3121         with_imp_locked(obd, imp0, rc)
3122                 imp = class_import_get(imp0);
3123         if (rc)
3124                 RETURN(rc);
3125
3126         /* We could possibly pass max_age in the request (as an absolute
3127          * timestamp or a "seconds.usec ago") so the target can avoid doing
3128          * extra calls into the filesystem if that isn't necessary (e.g.
3129          * during mount that would help a bit).  Having relative timestamps
3130          * is not so great if request processing is slow, while absolute
3131          * timestamps are not ideal because they need time synchronization. */
3132         req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
3133
3134         class_import_put(imp);
3135
3136         if (req == NULL)
3137                 RETURN(-ENOMEM);
3138
3139         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3140         if (rc) {
3141                 ptlrpc_request_free(req);
3142                 RETURN(rc);
3143         }
3144         ptlrpc_request_set_replen(req);
3145         req->rq_request_portal = OST_CREATE_PORTAL;
3146         ptlrpc_at_set_req_timeout(req);
3147
3148         if (flags & OBD_STATFS_NODELAY) {
3149                 /* procfs requests not want stat in wait for avoid deadlock */
3150                 req->rq_no_resend = 1;
3151                 req->rq_no_delay = 1;
3152         }
3153
3154         rc = ptlrpc_queue_wait(req);
3155         if (rc)
3156                 GOTO(out, rc);
3157
3158         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3159         if (msfs == NULL)
3160                 GOTO(out, rc = -EPROTO);
3161
3162         *osfs = *msfs;
3163
3164         EXIT;
3165 out:
3166         ptlrpc_req_finished(req);
3167         return rc;
3168 }
3169
3170 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
3171                          void *karg, void __user *uarg)
3172 {
3173         struct obd_device *obd = exp->exp_obd;
3174         struct obd_ioctl_data *data = karg;
3175         int rc = 0;
3176
3177         ENTRY;
3178         if (!try_module_get(THIS_MODULE)) {
3179                 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
3180                        module_name(THIS_MODULE));
3181                 return -EINVAL;
3182         }
3183         switch (cmd) {
3184         case OBD_IOC_CLIENT_RECOVER:
3185                 rc = ptlrpc_recover_import(obd->u.cli.cl_import,
3186                                            data->ioc_inlbuf1, 0);
3187                 if (rc > 0)
3188                         rc = 0;
3189                 break;
3190         case OBD_IOC_GETATTR:
3191                 rc = obd_getattr(NULL, exp, &data->ioc_obdo1);
3192                 break;
3193         case IOC_OSC_SET_ACTIVE:
3194                 rc = ptlrpc_set_import_active(obd->u.cli.cl_import,
3195                                               data->ioc_offset);
3196                 break;
3197         default:
3198                 rc = -ENOTTY;
3199                 CDEBUG(D_INODE, "%s: unrecognised ioctl %#x by %s: rc = %d\n",
3200                        obd->obd_name, cmd, current->comm, rc);
3201                 break;
3202         }
3203
3204         module_put(THIS_MODULE);
3205         return rc;
3206 }
3207
3208 int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
3209                        u32 keylen, void *key, u32 vallen, void *val,
3210                        struct ptlrpc_request_set *set)
3211 {
3212         struct ptlrpc_request *req;
3213         struct obd_device     *obd = exp->exp_obd;
3214         struct obd_import     *imp = class_exp2cliimp(exp);
3215         char                  *tmp;
3216         int                    rc;
3217         ENTRY;
3218
3219         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
3220
3221         if (KEY_IS(KEY_CHECKSUM)) {
3222                 if (vallen != sizeof(int))
3223                         RETURN(-EINVAL);
3224                 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
3225                 RETURN(0);
3226         }
3227
3228         if (KEY_IS(KEY_SPTLRPC_CONF)) {
3229                 sptlrpc_conf_client_adapt(obd);
3230                 RETURN(0);
3231         }
3232
3233         if (KEY_IS(KEY_FLUSH_CTX)) {
3234                 sptlrpc_import_flush_my_ctx(imp);
3235                 RETURN(0);
3236         }
3237
3238         if (KEY_IS(KEY_CACHE_LRU_SHRINK)) {
3239                 struct client_obd *cli = &obd->u.cli;
3240                 long nr = atomic_long_read(&cli->cl_lru_in_list) >> 1;
3241                 long target = *(long *)val;
3242
3243                 nr = osc_lru_shrink(env, cli, min(nr, target), true);
3244                 *(long *)val -= nr;
3245                 RETURN(0);
3246         }
3247
3248         if (!set && !KEY_IS(KEY_GRANT_SHRINK))
3249                 RETURN(-EINVAL);
3250
3251         /* We pass all other commands directly to OST. Since nobody calls osc
3252            methods directly and everybody is supposed to go through LOV, we
3253            assume lov checked invalid values for us.
3254            The only recognised values so far are evict_by_nid and mds_conn.
3255            Even if something bad goes through, we'd get a -EINVAL from OST
3256            anyway. */
3257
3258         req = ptlrpc_request_alloc(imp, KEY_IS(KEY_GRANT_SHRINK) ?
3259                                                 &RQF_OST_SET_GRANT_INFO :
3260                                                 &RQF_OBD_SET_INFO);
3261         if (req == NULL)
3262                 RETURN(-ENOMEM);
3263
3264         req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
3265                              RCL_CLIENT, keylen);
3266         if (!KEY_IS(KEY_GRANT_SHRINK))
3267                 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
3268                                      RCL_CLIENT, vallen);
3269         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
3270         if (rc) {
3271                 ptlrpc_request_free(req);
3272                 RETURN(rc);
3273         }
3274
3275         tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
3276         memcpy(tmp, key, keylen);
3277         tmp = req_capsule_client_get(&req->rq_pill, KEY_IS(KEY_GRANT_SHRINK) ?
3278                                                         &RMF_OST_BODY :
3279                                                         &RMF_SETINFO_VAL);
3280         memcpy(tmp, val, vallen);
3281
3282         if (KEY_IS(KEY_GRANT_SHRINK)) {
3283                 struct osc_grant_args *aa;
3284                 struct obdo *oa;
3285
3286                 aa = ptlrpc_req_async_args(aa, req);
3287                 OBD_SLAB_ALLOC_PTR_GFP(oa, osc_obdo_kmem, GFP_NOFS);
3288                 if (!oa) {
3289                         ptlrpc_req_finished(req);
3290                         RETURN(-ENOMEM);
3291                 }
3292                 *oa = ((struct ost_body *)val)->oa;
3293                 aa->aa_oa = oa;
3294                 req->rq_interpret_reply = osc_shrink_grant_interpret;
3295         }
3296
3297         ptlrpc_request_set_replen(req);
3298         if (!KEY_IS(KEY_GRANT_SHRINK)) {
3299                 LASSERT(set != NULL);
3300                 ptlrpc_set_add_req(set, req);
3301                 ptlrpc_check_set(NULL, set);
3302         } else {
3303                 ptlrpcd_add_req(req);
3304         }
3305
3306         RETURN(0);
3307 }
3308 EXPORT_SYMBOL(osc_set_info_async);
3309
3310 int osc_reconnect(const struct lu_env *env, struct obd_export *exp,
3311                   struct obd_device *obd, struct obd_uuid *cluuid,
3312                   struct obd_connect_data *data, void *localdata)
3313 {
3314         struct client_obd *cli = &obd->u.cli;
3315
3316         if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
3317                 long lost_grant;
3318                 long grant;
3319
3320                 spin_lock(&cli->cl_loi_list_lock);
3321                 grant = cli->cl_avail_grant + cli->cl_reserved_grant;
3322                 if (data->ocd_connect_flags & OBD_CONNECT_GRANT_PARAM) {
3323                         /* restore ocd_grant_blkbits as client page bits */
3324                         data->ocd_grant_blkbits = PAGE_SHIFT;
3325                         grant += cli->cl_dirty_grant;
3326                 } else {
3327                         grant += cli->cl_dirty_pages << PAGE_SHIFT;
3328                 }
3329                 data->ocd_grant = grant ? : 2 * cli_brw_size(obd);
3330                 lost_grant = cli->cl_lost_grant;
3331                 cli->cl_lost_grant = 0;
3332                 spin_unlock(&cli->cl_loi_list_lock);
3333
3334                 CDEBUG(D_RPCTRACE, "ocd_connect_flags: %#llx ocd_version: %d"
3335                        " ocd_grant: %d, lost: %ld.\n", data->ocd_connect_flags,
3336                        data->ocd_version, data->ocd_grant, lost_grant);
3337         }
3338
3339         RETURN(0);
3340 }
3341 EXPORT_SYMBOL(osc_reconnect);
3342
3343 int osc_disconnect(struct obd_export *exp)
3344 {
3345         struct obd_device *obd = class_exp2obd(exp);
3346         int rc;
3347
3348         rc = client_disconnect_export(exp);
3349         /**
3350          * Initially we put del_shrink_grant before disconnect_export, but it
3351          * causes the following problem if setup (connect) and cleanup
3352          * (disconnect) are tangled together.
3353          *      connect p1                     disconnect p2
3354          *   ptlrpc_connect_import
3355          *     ...............               class_manual_cleanup
3356          *                                     osc_disconnect
3357          *                                     del_shrink_grant
3358          *   ptlrpc_connect_interrupt
3359          *     osc_init_grant
3360          *   add this client to shrink list
3361          *                                      cleanup_osc
3362          * Bang! grant shrink thread trigger the shrink. BUG18662
3363          */
3364         osc_del_grant_list(&obd->u.cli);
3365         return rc;
3366 }
3367 EXPORT_SYMBOL(osc_disconnect);
3368
3369 int osc_ldlm_resource_invalidate(struct cfs_hash *hs, struct cfs_hash_bd *bd,
3370                                  struct hlist_node *hnode, void *arg)
3371 {
3372         struct lu_env *env = arg;
3373         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
3374         struct ldlm_lock *lock;
3375         struct osc_object *osc = NULL;
3376         ENTRY;
3377
3378         lock_res(res);
3379         list_for_each_entry(lock, &res->lr_granted, l_res_link) {
3380                 if (lock->l_ast_data != NULL && osc == NULL) {
3381                         osc = lock->l_ast_data;
3382                         cl_object_get(osc2cl(osc));
3383                 }
3384
3385                 /* clear LDLM_FL_CLEANED flag to make sure it will be canceled
3386                  * by the 2nd round of ldlm_namespace_clean() call in
3387                  * osc_import_event(). */
3388                 ldlm_clear_cleaned(lock);
3389         }
3390         unlock_res(res);
3391
3392         if (osc != NULL) {
3393                 osc_object_invalidate(env, osc);
3394                 cl_object_put(env, osc2cl(osc));
3395         }
3396
3397         RETURN(0);
3398 }
3399 EXPORT_SYMBOL(osc_ldlm_resource_invalidate);
3400
3401 static int osc_import_event(struct obd_device *obd,
3402                             struct obd_import *imp,
3403                             enum obd_import_event event)
3404 {
3405         struct client_obd *cli;
3406         int rc = 0;
3407
3408         ENTRY;
3409         LASSERT(imp->imp_obd == obd);
3410
3411         switch (event) {
3412         case IMP_EVENT_DISCON: {
3413                 cli = &obd->u.cli;
3414                 spin_lock(&cli->cl_loi_list_lock);
3415                 cli->cl_avail_grant = 0;
3416                 cli->cl_lost_grant = 0;
3417                 spin_unlock(&cli->cl_loi_list_lock);
3418                 break;
3419         }
3420         case IMP_EVENT_INACTIVE: {
3421                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
3422                 break;
3423         }
3424         case IMP_EVENT_INVALIDATE: {
3425                 struct ldlm_namespace *ns = obd->obd_namespace;
3426                 struct lu_env         *env;
3427                 __u16                  refcheck;
3428
3429                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3430
3431                 env = cl_env_get(&refcheck);
3432                 if (!IS_ERR(env)) {
3433                         osc_io_unplug(env, &obd->u.cli, NULL);
3434
3435                         cfs_hash_for_each_nolock(ns->ns_rs_hash,
3436                                                  osc_ldlm_resource_invalidate,
3437                                                  env, 0);
3438                         cl_env_put(env, &refcheck);
3439
3440                         ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3441                 } else
3442                         rc = PTR_ERR(env);
3443                 break;
3444         }
3445         case IMP_EVENT_ACTIVE: {
3446                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
3447                 break;
3448         }
3449         case IMP_EVENT_OCD: {
3450                 struct obd_connect_data *ocd = &imp->imp_connect_data;
3451
3452                 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3453                         osc_init_grant(&obd->u.cli, ocd);
3454
3455                 /* See bug 7198 */
3456                 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3457                         imp->imp_client->cli_request_portal =OST_REQUEST_PORTAL;
3458
3459                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
3460                 break;
3461         }
3462         case IMP_EVENT_DEACTIVATE: {
3463                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE);
3464                 break;
3465         }
3466         case IMP_EVENT_ACTIVATE: {
3467                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE);
3468                 break;
3469         }
3470         default:
3471                 CERROR("Unknown import event %d\n", event);
3472                 LBUG();
3473         }
3474         RETURN(rc);
3475 }
3476
3477 /**
3478  * Determine whether the lock can be canceled before replaying the lock
3479  * during recovery, see bug16774 for detailed information.
3480  *
3481  * \retval zero the lock can't be canceled
3482  * \retval other ok to cancel
3483  */
3484 static int osc_cancel_weight(struct ldlm_lock *lock)
3485 {
3486         /*
3487          * Cancel all unused and granted extent lock.
3488          */
3489         if (lock->l_resource->lr_type == LDLM_EXTENT &&
3490             ldlm_is_granted(lock) &&
3491             osc_ldlm_weigh_ast(lock) == 0)
3492                 RETURN(1);
3493
3494         RETURN(0);
3495 }
3496
3497 static int brw_queue_work(const struct lu_env *env, void *data)
3498 {
3499         struct client_obd *cli = data;
3500
3501         CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
3502
3503         osc_io_unplug(env, cli, NULL);
3504         RETURN(0);
3505 }
3506
3507 int osc_setup_common(struct obd_device *obd, struct lustre_cfg *lcfg)
3508 {
3509         struct client_obd *cli = &obd->u.cli;
3510         void *handler;
3511         int rc;
3512
3513         ENTRY;
3514
3515         rc = ptlrpcd_addref();
3516         if (rc)
3517                 RETURN(rc);
3518
3519         rc = client_obd_setup(obd, lcfg);
3520         if (rc)
3521                 GOTO(out_ptlrpcd, rc);
3522
3523
3524         handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
3525         if (IS_ERR(handler))
3526                 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3527         cli->cl_writeback_work = handler;
3528
3529         handler = ptlrpcd_alloc_work(cli->cl_import, lru_queue_work, cli);
3530         if (IS_ERR(handler))
3531                 GOTO(out_ptlrpcd_work, rc = PTR_ERR(handler));
3532         cli->cl_lru_work = handler;
3533
3534         rc = osc_quota_setup(obd);
3535         if (rc)
3536                 GOTO(out_ptlrpcd_work, rc);
3537
3538         cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
3539         cli->cl_root_squash = 0;
3540         osc_update_next_shrink(cli);
3541
3542         RETURN(rc);
3543
3544 out_ptlrpcd_work:
3545         if (cli->cl_writeback_work != NULL) {
3546                 ptlrpcd_destroy_work(cli->cl_writeback_work);
3547                 cli->cl_writeback_work = NULL;
3548         }
3549         if (cli->cl_lru_work != NULL) {
3550                 ptlrpcd_destroy_work(cli->cl_lru_work);
3551                 cli->cl_lru_work = NULL;
3552         }
3553         client_obd_cleanup(obd);
3554 out_ptlrpcd:
3555         ptlrpcd_decref();
3556         RETURN(rc);
3557 }
3558 EXPORT_SYMBOL(osc_setup_common);
3559
3560 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3561 {
3562         struct client_obd *cli = &obd->u.cli;
3563         int                adding;
3564         int                added;
3565         int                req_count;
3566         int                rc;
3567
3568         ENTRY;
3569
3570         rc = osc_setup_common(obd, lcfg);
3571         if (rc < 0)
3572                 RETURN(rc);
3573
3574         rc = osc_tunables_init(obd);
3575         if (rc)
3576                 RETURN(rc);
3577
3578         /*
3579          * We try to control the total number of requests with a upper limit
3580          * osc_reqpool_maxreqcount. There might be some race which will cause
3581          * over-limit allocation, but it is fine.
3582          */
3583         req_count = atomic_read(&osc_pool_req_count);
3584         if (req_count < osc_reqpool_maxreqcount) {
3585                 adding = cli->cl_max_rpcs_in_flight + 2;
3586                 if (req_count + adding > osc_reqpool_maxreqcount)
3587                         adding = osc_reqpool_maxreqcount - req_count;
3588
3589                 added = ptlrpc_add_rqs_to_pool(osc_rq_pool, adding);
3590                 atomic_add(added, &osc_pool_req_count);
3591         }
3592
3593         ns_register_cancel(obd->obd_namespace, osc_cancel_weight);
3594
3595         spin_lock(&osc_shrink_lock);
3596         list_add_tail(&cli->cl_shrink_list, &osc_shrink_list);
3597         spin_unlock(&osc_shrink_lock);
3598         cli->cl_import->imp_idle_timeout = osc_idle_timeout;
3599         cli->cl_import->imp_idle_debug = D_HA;
3600
3601         RETURN(0);
3602 }
3603
3604 int osc_precleanup_common(struct obd_device *obd)
3605 {
3606         struct client_obd *cli = &obd->u.cli;
3607         ENTRY;
3608
3609         /* LU-464
3610          * for echo client, export may be on zombie list, wait for
3611          * zombie thread to cull it, because cli.cl_import will be
3612          * cleared in client_disconnect_export():
3613          *   class_export_destroy() -> obd_cleanup() ->
3614          *   echo_device_free() -> echo_client_cleanup() ->
3615          *   obd_disconnect() -> osc_disconnect() ->
3616          *   client_disconnect_export()
3617          */
3618         obd_zombie_barrier();
3619         if (cli->cl_writeback_work) {
3620                 ptlrpcd_destroy_work(cli->cl_writeback_work);
3621                 cli->cl_writeback_work = NULL;
3622         }
3623
3624         if (cli->cl_lru_work) {
3625                 ptlrpcd_destroy_work(cli->cl_lru_work);
3626                 cli->cl_lru_work = NULL;
3627         }
3628
3629         obd_cleanup_client_import(obd);
3630         RETURN(0);
3631 }
3632 EXPORT_SYMBOL(osc_precleanup_common);
3633
3634 static int osc_precleanup(struct obd_device *obd)
3635 {
3636         ENTRY;
3637
3638         osc_precleanup_common(obd);
3639
3640         ptlrpc_lprocfs_unregister_obd(obd);
3641         RETURN(0);
3642 }
3643
3644 int osc_cleanup_common(struct obd_device *obd)
3645 {
3646         struct client_obd *cli = &obd->u.cli;
3647         int rc;
3648
3649         ENTRY;
3650
3651         spin_lock(&osc_shrink_lock);
3652         list_del(&cli->cl_shrink_list);
3653         spin_unlock(&osc_shrink_lock);
3654
3655         /* lru cleanup */
3656         if (cli->cl_cache != NULL) {
3657                 LASSERT(atomic_read(&cli->cl_cache->ccc_users) > 0);
3658                 spin_lock(&cli->cl_cache->ccc_lru_lock);
3659                 list_del_init(&cli->cl_lru_osc);
3660                 spin_unlock(&cli->cl_cache->ccc_lru_lock);
3661                 cli->cl_lru_left = NULL;
3662                 cl_cache_decref(cli->cl_cache);
3663                 cli->cl_cache = NULL;
3664         }
3665
3666         /* free memory of osc quota cache */
3667         osc_quota_cleanup(obd);
3668
3669         rc = client_obd_cleanup(obd);
3670
3671         ptlrpcd_decref();
3672         RETURN(rc);
3673 }
3674 EXPORT_SYMBOL(osc_cleanup_common);
3675
3676 static const struct obd_ops osc_obd_ops = {
3677         .o_owner                = THIS_MODULE,
3678         .o_setup                = osc_setup,
3679         .o_precleanup           = osc_precleanup,
3680         .o_cleanup              = osc_cleanup_common,
3681         .o_add_conn             = client_import_add_conn,
3682         .o_del_conn             = client_import_del_conn,
3683         .o_connect              = client_connect_import,
3684         .o_reconnect            = osc_reconnect,
3685         .o_disconnect           = osc_disconnect,
3686         .o_statfs               = osc_statfs,
3687         .o_statfs_async         = osc_statfs_async,
3688         .o_create               = osc_create,
3689         .o_destroy              = osc_destroy,
3690         .o_getattr              = osc_getattr,
3691         .o_setattr              = osc_setattr,
3692         .o_iocontrol            = osc_iocontrol,
3693         .o_set_info_async       = osc_set_info_async,
3694         .o_import_event         = osc_import_event,
3695         .o_quotactl             = osc_quotactl,
3696 };
3697
3698 LIST_HEAD(osc_shrink_list);
3699 DEFINE_SPINLOCK(osc_shrink_lock);
3700
3701 #ifdef HAVE_SHRINKER_COUNT
3702 static struct shrinker osc_cache_shrinker = {
3703         .count_objects  = osc_cache_shrink_count,
3704         .scan_objects   = osc_cache_shrink_scan,
3705         .seeks          = DEFAULT_SEEKS,
3706 };
3707 #else
3708 static int osc_cache_shrink(struct shrinker *shrinker,
3709                             struct shrink_control *sc)
3710 {
3711         (void)osc_cache_shrink_scan(shrinker, sc);
3712
3713         return osc_cache_shrink_count(shrinker, sc);
3714 }
3715
3716 static struct shrinker osc_cache_shrinker = {
3717         .shrink   = osc_cache_shrink,
3718         .seeks    = DEFAULT_SEEKS,
3719 };
3720 #endif
3721
3722 static int __init osc_init(void)
3723 {
3724         unsigned int reqpool_size;
3725         unsigned int reqsize;
3726         int rc;
3727         ENTRY;
3728
3729         /* print an address of _any_ initialized kernel symbol from this
3730          * module, to allow debugging with gdb that doesn't support data
3731          * symbols from modules.*/
3732         CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
3733
3734         rc = lu_kmem_init(osc_caches);
3735         if (rc)
3736                 RETURN(rc);
3737
3738         rc = class_register_type(&osc_obd_ops, NULL, true,
3739                                  LUSTRE_OSC_NAME, &osc_device_type);
3740         if (rc)
3741                 GOTO(out_kmem, rc);
3742
3743         rc = register_shrinker(&osc_cache_shrinker);
3744         if (rc)
3745                 GOTO(out_type, rc);
3746
3747         /* This is obviously too much memory, only prevent overflow here */
3748         if (osc_reqpool_mem_max >= 1 << 12 || osc_reqpool_mem_max == 0)
3749                 GOTO(out_shrinker, rc = -EINVAL);
3750
3751         reqpool_size = osc_reqpool_mem_max << 20;
3752
3753         reqsize = 1;
3754         while (reqsize < OST_IO_MAXREQSIZE)
3755                 reqsize = reqsize << 1;
3756
3757         /*
3758          * We don't enlarge the request count in OSC pool according to
3759          * cl_max_rpcs_in_flight. The allocation from the pool will only be
3760          * tried after normal allocation failed. So a small OSC pool won't
3761          * cause much performance degression in most of cases.
3762          */
3763         osc_reqpool_maxreqcount = reqpool_size / reqsize;
3764
3765         atomic_set(&osc_pool_req_count, 0);
3766         osc_rq_pool = ptlrpc_init_rq_pool(0, OST_IO_MAXREQSIZE,
3767                                           ptlrpc_add_rqs_to_pool);
3768
3769         if (osc_rq_pool == NULL)
3770                 GOTO(out_shrinker, rc = -ENOMEM);
3771
3772         rc = osc_start_grant_work();
3773         if (rc != 0)
3774                 GOTO(out_req_pool, rc);
3775
3776         RETURN(rc);
3777
3778 out_req_pool:
3779         ptlrpc_free_rq_pool(osc_rq_pool);
3780 out_shrinker:
3781         unregister_shrinker(&osc_cache_shrinker);
3782 out_type:
3783         class_unregister_type(LUSTRE_OSC_NAME);
3784 out_kmem:
3785         lu_kmem_fini(osc_caches);
3786
3787         RETURN(rc);
3788 }
3789
3790 static void __exit osc_exit(void)
3791 {
3792         osc_stop_grant_work();
3793         unregister_shrinker(&osc_cache_shrinker);
3794         class_unregister_type(LUSTRE_OSC_NAME);
3795         lu_kmem_fini(osc_caches);
3796         ptlrpc_free_rq_pool(osc_rq_pool);
3797 }
3798
3799 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3800 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3801 MODULE_VERSION(LUSTRE_VERSION_STRING);
3802 MODULE_LICENSE("GPL");
3803
3804 module_init(osc_init);
3805 module_exit(osc_exit);