drivers/staging/lustre/lustre/osc/osc_request.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2015, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_OSC
  38
  39 #include "../../include/linux/libcfs/libcfs.h"
  40
  41 #include "../include/lustre_dlm.h"
  42 #include "../include/lustre_net.h"
  43 #include "../include/lustre/lustre_user.h"
  44 #include "../include/obd_cksum.h"
  45
  46 #include "../include/lustre_ha.h"
  47 #include "../include/lprocfs_status.h"
  48 #include "../include/lustre_debug.h"
  49 #include "../include/lustre_param.h"
  50 #include "../include/lustre_fid.h"
  51 #include "../include/obd_class.h"
  52 #include "../include/obd.h"
  53 #include "osc_internal.h"
  54 #include "osc_cl_internal.h"
  55
  56 atomic_t osc_pool_req_count;
  57 unsigned int osc_reqpool_maxreqcount;
  58 struct ptlrpc_request_pool *osc_rq_pool;
  59
  60 /* max memory used for request pool, unit is MB */
  61 static unsigned int osc_reqpool_mem_max = 5;
  62 module_param(osc_reqpool_mem_max, uint, 0444);
  63
  64 struct osc_brw_async_args {
  65         struct obdo       *aa_oa;
  66         int             aa_requested_nob;
  67         int             aa_nio_count;
  68         u32             aa_page_count;
  69         int             aa_resends;
  70         struct brw_page  **aa_ppga;
  71         struct client_obd *aa_cli;
  72         struct list_head         aa_oaps;
  73         struct list_head         aa_exts;
  74         struct cl_req     *aa_clerq;
  75 };
  76
  77 struct osc_async_args {
  78         struct obd_info   *aa_oi;
  79 };
  80
  81 struct osc_setattr_args {
  82         struct obdo      *sa_oa;
  83         obd_enqueue_update_f sa_upcall;
  84         void            *sa_cookie;
  85 };
  86
  87 struct osc_fsync_args {
  88         struct obd_info     *fa_oi;
  89         obd_enqueue_update_f fa_upcall;
  90         void            *fa_cookie;
  91 };
  92
  93 struct osc_enqueue_args {
  94         struct obd_export       *oa_exp;
  95         __u64               *oa_flags;
  96         obd_enqueue_update_f      oa_upcall;
  97         void                 *oa_cookie;
  98         struct ost_lvb     *oa_lvb;
  99         struct lustre_handle     *oa_lockh;
 100         struct ldlm_enqueue_info *oa_ei;
 101         unsigned int          oa_agl:1;
 102 };
 103
 104 static void osc_release_ppga(struct brw_page **ppga, u32 count);
 105 static int brw_interpret(const struct lu_env *env,
 106                          struct ptlrpc_request *req, void *data, int rc);
 107
 108 /* Pack OSC object metadata for disk storage (LE byte order). */
 109 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
 110                       struct lov_stripe_md *lsm)
 111 {
 112         int lmm_size;
 113
 114         lmm_size = sizeof(**lmmp);
 115         if (!lmmp)
 116                 return lmm_size;
 117
 118         if (*lmmp && !lsm) {
 119                 kfree(*lmmp);
 120                 *lmmp = NULL;
 121                 return 0;
 122         } else if (unlikely(lsm && ostid_id(&lsm->lsm_oi) == 0)) {
 123                 return -EBADF;
 124         }
 125
 126         if (!*lmmp) {
 127                 *lmmp = kzalloc(lmm_size, GFP_NOFS);
 128                 if (!*lmmp)
 129                         return -ENOMEM;
 130         }
 131
 132         if (lsm)
 133                 ostid_cpu_to_le(&lsm->lsm_oi, &(*lmmp)->lmm_oi);
 134
 135         return lmm_size;
 136 }
 137
 138 /* Unpack OSC object metadata from disk storage (LE byte order). */
 139 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
 140                         struct lov_mds_md *lmm, int lmm_bytes)
 141 {
 142         int lsm_size;
 143         struct obd_import *imp = class_exp2cliimp(exp);
 144
 145         if (lmm) {
 146                 if (lmm_bytes < sizeof(*lmm)) {
 147                         CERROR("%s: lov_mds_md too small: %d, need %d\n",
 148                                exp->exp_obd->obd_name, lmm_bytes,
 149                                (int)sizeof(*lmm));
 150                         return -EINVAL;
 151                 }
 152                 /* XXX LOV_MAGIC etc check? */
 153
 154                 if (unlikely(ostid_id(&lmm->lmm_oi) == 0)) {
 155                         CERROR("%s: zero lmm_object_id: rc = %d\n",
 156                                exp->exp_obd->obd_name, -EINVAL);
 157                         return -EINVAL;
 158                 }
 159         }
 160
 161         lsm_size = lov_stripe_md_size(1);
 162         if (!lsmp)
 163                 return lsm_size;
 164
 165         if (*lsmp && !lmm) {
 166                 kfree((*lsmp)->lsm_oinfo[0]);
 167                 kfree(*lsmp);
 168                 *lsmp = NULL;
 169                 return 0;
 170         }
 171
 172         if (!*lsmp) {
 173                 *lsmp = kzalloc(lsm_size, GFP_NOFS);
 174                 if (unlikely(!*lsmp))
 175                         return -ENOMEM;
 176                 (*lsmp)->lsm_oinfo[0] = kzalloc(sizeof(struct lov_oinfo),
 177                                                 GFP_NOFS);
 178                 if (unlikely(!(*lsmp)->lsm_oinfo[0])) {
 179                         kfree(*lsmp);
 180                         return -ENOMEM;
 181                 }
 182                 loi_init((*lsmp)->lsm_oinfo[0]);
 183         } else if (unlikely(ostid_id(&(*lsmp)->lsm_oi) == 0)) {
 184                 return -EBADF;
 185         }
 186
 187         if (lmm)
 188                 /* XXX zero *lsmp? */
 189                 ostid_le_to_cpu(&lmm->lmm_oi, &(*lsmp)->lsm_oi);
 190
 191         if (imp &&
 192             (imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_MAXBYTES))
 193                 (*lsmp)->lsm_maxbytes = imp->imp_connect_data.ocd_maxbytes;
 194         else
 195                 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
 196
 197         return lsm_size;
 198 }
 199
 200 static inline void osc_pack_req_body(struct ptlrpc_request *req,
 201                                      struct obd_info *oinfo)
 202 {
 203         struct ost_body *body;
 204
 205         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 206         LASSERT(body);
 207
 208         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
 209                              oinfo->oi_oa);
 210 }
 211
 212 static int osc_getattr_interpret(const struct lu_env *env,
 213                                  struct ptlrpc_request *req,
 214                                  struct osc_async_args *aa, int rc)
 215 {
 216         struct ost_body *body;
 217
 218         if (rc != 0)
 219                 goto out;
 220
 221         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 222         if (body) {
 223                 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
 224                 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
 225                                      aa->aa_oi->oi_oa, &body->oa);
 226
 227                 /* This should really be sent by the OST */
 228                 aa->aa_oi->oi_oa->o_blksize = DT_MAX_BRW_SIZE;
 229                 aa->aa_oi->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
 230         } else {
 231                 CDEBUG(D_INFO, "can't unpack ost_body\n");
 232                 rc = -EPROTO;
 233                 aa->aa_oi->oi_oa->o_valid = 0;
 234         }
 235 out:
 236         rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
 237         return rc;
 238 }
 239
 240 static int osc_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
 241                              struct ptlrpc_request_set *set)
 242 {
 243         struct ptlrpc_request *req;
 244         struct osc_async_args *aa;
 245         int rc;
 246
 247         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
 248         if (!req)
 249                 return -ENOMEM;
 250
 251         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
 252         if (rc) {
 253                 ptlrpc_request_free(req);
 254                 return rc;
 255         }
 256
 257         osc_pack_req_body(req, oinfo);
 258
 259         ptlrpc_request_set_replen(req);
 260         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_getattr_interpret;
 261
 262         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
 263         aa = ptlrpc_req_async_args(req);
 264         aa->aa_oi = oinfo;
 265
 266         ptlrpc_set_add_req(set, req);
 267         return 0;
 268 }
 269
 270 static int osc_getattr(const struct lu_env *env, struct obd_export *exp,
 271                        struct obd_info *oinfo)
 272 {
 273         struct ptlrpc_request *req;
 274         struct ost_body *body;
 275         int rc;
 276
 277         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
 278         if (!req)
 279                 return -ENOMEM;
 280
 281         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
 282         if (rc) {
 283                 ptlrpc_request_free(req);
 284                 return rc;
 285         }
 286
 287         osc_pack_req_body(req, oinfo);
 288
 289         ptlrpc_request_set_replen(req);
 290
 291         rc = ptlrpc_queue_wait(req);
 292         if (rc)
 293                 goto out;
 294
 295         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 296         if (!body) {
 297                 rc = -EPROTO;
 298                 goto out;
 299         }
 300
 301         CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
 302         lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oinfo->oi_oa,
 303                              &body->oa);
 304
 305         oinfo->oi_oa->o_blksize = cli_brw_size(exp->exp_obd);
 306         oinfo->oi_oa->o_valid |= OBD_MD_FLBLKSZ;
 307
 308  out:
 309         ptlrpc_req_finished(req);
 310         return rc;
 311 }
 312
 313 static int osc_setattr(const struct lu_env *env, struct obd_export *exp,
 314                        struct obd_info *oinfo, struct obd_trans_info *oti)
 315 {
 316         struct ptlrpc_request *req;
 317         struct ost_body *body;
 318         int rc;
 319
 320         LASSERT(oinfo->oi_oa->o_valid & OBD_MD_FLGROUP);
 321
 322         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
 323         if (!req)
 324                 return -ENOMEM;
 325
 326         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
 327         if (rc) {
 328                 ptlrpc_request_free(req);
 329                 return rc;
 330         }
 331
 332         osc_pack_req_body(req, oinfo);
 333
 334         ptlrpc_request_set_replen(req);
 335
 336         rc = ptlrpc_queue_wait(req);
 337         if (rc)
 338                 goto out;
 339
 340         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 341         if (!body) {
 342                 rc = -EPROTO;
 343                 goto out;
 344         }
 345
 346         lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oinfo->oi_oa,
 347                              &body->oa);
 348
 349 out:
 350         ptlrpc_req_finished(req);
 351         return rc;
 352 }
 353
 354 static int osc_setattr_interpret(const struct lu_env *env,
 355                                  struct ptlrpc_request *req,
 356                                  struct osc_setattr_args *sa, int rc)
 357 {
 358         struct ost_body *body;
 359
 360         if (rc != 0)
 361                 goto out;
 362
 363         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 364         if (!body) {
 365                 rc = -EPROTO;
 366                 goto out;
 367         }
 368
 369         lustre_get_wire_obdo(&req->rq_import->imp_connect_data, sa->sa_oa,
 370                              &body->oa);
 371 out:
 372         rc = sa->sa_upcall(sa->sa_cookie, rc);
 373         return rc;
 374 }
 375
 376 int osc_setattr_async_base(struct obd_export *exp, struct obd_info *oinfo,
 377                            struct obd_trans_info *oti,
 378                            obd_enqueue_update_f upcall, void *cookie,
 379                            struct ptlrpc_request_set *rqset)
 380 {
 381         struct ptlrpc_request *req;
 382         struct osc_setattr_args *sa;
 383         int rc;
 384
 385         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
 386         if (!req)
 387                 return -ENOMEM;
 388
 389         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
 390         if (rc) {
 391                 ptlrpc_request_free(req);
 392                 return rc;
 393         }
 394
 395         if (oti && oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
 396                 oinfo->oi_oa->o_lcookie = *oti->oti_logcookies;
 397
 398         osc_pack_req_body(req, oinfo);
 399
 400         ptlrpc_request_set_replen(req);
 401
 402         /* do mds to ost setattr asynchronously */
 403         if (!rqset) {
 404                 /* Do not wait for response. */
 405                 ptlrpcd_add_req(req);
 406         } else {
 407                 req->rq_interpret_reply =
 408                         (ptlrpc_interpterer_t)osc_setattr_interpret;
 409
 410                 CLASSERT(sizeof(*sa) <= sizeof(req->rq_async_args));
 411                 sa = ptlrpc_req_async_args(req);
 412                 sa->sa_oa = oinfo->oi_oa;
 413                 sa->sa_upcall = upcall;
 414                 sa->sa_cookie = cookie;
 415
 416                 if (rqset == PTLRPCD_SET)
 417                         ptlrpcd_add_req(req);
 418                 else
 419                         ptlrpc_set_add_req(rqset, req);
 420         }
 421
 422         return 0;
 423 }
 424
 425 static int osc_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
 426                              struct obd_trans_info *oti,
 427                              struct ptlrpc_request_set *rqset)
 428 {
 429         return osc_setattr_async_base(exp, oinfo, oti,
 430                                       oinfo->oi_cb_up, oinfo, rqset);
 431 }
 432
 433 static int osc_real_create(struct obd_export *exp, struct obdo *oa,
 434                            struct lov_stripe_md **ea,
 435                            struct obd_trans_info *oti)
 436 {
 437         struct ptlrpc_request *req;
 438         struct ost_body *body;
 439         struct lov_stripe_md *lsm;
 440         int rc;
 441
 442         LASSERT(oa);
 443         LASSERT(ea);
 444
 445         lsm = *ea;
 446         if (!lsm) {
 447                 rc = obd_alloc_memmd(exp, &lsm);
 448                 if (rc < 0)
 449                         return rc;
 450         }
 451
 452         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
 453         if (!req) {
 454                 rc = -ENOMEM;
 455                 goto out;
 456         }
 457
 458         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
 459         if (rc) {
 460                 ptlrpc_request_free(req);
 461                 goto out;
 462         }
 463
 464         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 465         LASSERT(body);
 466
 467         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
 468
 469         ptlrpc_request_set_replen(req);
 470
 471         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 472             oa->o_flags == OBD_FL_DELORPHAN) {
 473                 DEBUG_REQ(D_HA, req,
 474                           "delorphan from OST integration");
 475                 /* Don't resend the delorphan req */
 476                 req->rq_no_resend = req->rq_no_delay = 1;
 477         }
 478
 479         rc = ptlrpc_queue_wait(req);
 480         if (rc)
 481                 goto out_req;
 482
 483         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 484         if (!body) {
 485                 rc = -EPROTO;
 486                 goto out_req;
 487         }
 488
 489         CDEBUG(D_INFO, "oa flags %x\n", oa->o_flags);
 490         lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
 491
 492         oa->o_blksize = cli_brw_size(exp->exp_obd);
 493         oa->o_valid |= OBD_MD_FLBLKSZ;
 494
 495         /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
 496          * have valid lsm_oinfo data structs, so don't go touching that.
 497          * This needs to be fixed in a big way.
 498          */
 499         lsm->lsm_oi = oa->o_oi;
 500         *ea = lsm;
 501
 502         if (oti) {
 503                 oti->oti_transno = lustre_msg_get_transno(req->rq_repmsg);
 504
 505                 if (oa->o_valid & OBD_MD_FLCOOKIE) {
 506                         if (!oti->oti_logcookies)
 507                                 oti_alloc_cookies(oti, 1);
 508                         *oti->oti_logcookies = oa->o_lcookie;
 509                 }
 510         }
 511
 512         CDEBUG(D_HA, "transno: %lld\n",
 513                lustre_msg_get_transno(req->rq_repmsg));
 514 out_req:
 515         ptlrpc_req_finished(req);
 516 out:
 517         if (rc && !*ea)
 518                 obd_free_memmd(exp, &lsm);
 519         return rc;
 520 }
 521
 522 int osc_punch_base(struct obd_export *exp, struct obd_info *oinfo,
 523                    obd_enqueue_update_f upcall, void *cookie,
 524                    struct ptlrpc_request_set *rqset)
 525 {
 526         struct ptlrpc_request *req;
 527         struct osc_setattr_args *sa;
 528         struct ost_body *body;
 529         int rc;
 530
 531         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_PUNCH);
 532         if (!req)
 533                 return -ENOMEM;
 534
 535         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
 536         if (rc) {
 537                 ptlrpc_request_free(req);
 538                 return rc;
 539         }
 540         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
 541         ptlrpc_at_set_req_timeout(req);
 542
 543         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 544         LASSERT(body);
 545         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
 546                              oinfo->oi_oa);
 547
 548         ptlrpc_request_set_replen(req);
 549
 550         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_setattr_interpret;
 551         CLASSERT(sizeof(*sa) <= sizeof(req->rq_async_args));
 552         sa = ptlrpc_req_async_args(req);
 553         sa->sa_oa = oinfo->oi_oa;
 554         sa->sa_upcall = upcall;
 555         sa->sa_cookie = cookie;
 556         if (rqset == PTLRPCD_SET)
 557                 ptlrpcd_add_req(req);
 558         else
 559                 ptlrpc_set_add_req(rqset, req);
 560
 561         return 0;
 562 }
 563
 564 static int osc_sync_interpret(const struct lu_env *env,
 565                               struct ptlrpc_request *req,
 566                               void *arg, int rc)
 567 {
 568         struct osc_fsync_args *fa = arg;
 569         struct ost_body *body;
 570
 571         if (rc)
 572                 goto out;
 573
 574         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 575         if (!body) {
 576                 CERROR("can't unpack ost_body\n");
 577                 rc = -EPROTO;
 578                 goto out;
 579         }
 580
 581         *fa->fa_oi->oi_oa = body->oa;
 582 out:
 583         rc = fa->fa_upcall(fa->fa_cookie, rc);
 584         return rc;
 585 }
 586
 587 int osc_sync_base(struct obd_export *exp, struct obd_info *oinfo,
 588                   obd_enqueue_update_f upcall, void *cookie,
 589                   struct ptlrpc_request_set *rqset)
 590 {
 591         struct ptlrpc_request *req;
 592         struct ost_body *body;
 593         struct osc_fsync_args *fa;
 594         int rc;
 595
 596         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
 597         if (!req)
 598                 return -ENOMEM;
 599
 600         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
 601         if (rc) {
 602                 ptlrpc_request_free(req);
 603                 return rc;
 604         }
 605
 606         /* overload the size and blocks fields in the oa with start/end */
 607         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 608         LASSERT(body);
 609         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
 610                              oinfo->oi_oa);
 611
 612         ptlrpc_request_set_replen(req);
 613         req->rq_interpret_reply = osc_sync_interpret;
 614
 615         CLASSERT(sizeof(*fa) <= sizeof(req->rq_async_args));
 616         fa = ptlrpc_req_async_args(req);
 617         fa->fa_oi = oinfo;
 618         fa->fa_upcall = upcall;
 619         fa->fa_cookie = cookie;
 620
 621         if (rqset == PTLRPCD_SET)
 622                 ptlrpcd_add_req(req);
 623         else
 624                 ptlrpc_set_add_req(rqset, req);
 625
 626         return 0;
 627 }
 628
 629 /* Find and cancel locally locks matched by @mode in the resource found by
 630  * @objid. Found locks are added into @cancel list. Returns the amount of
 631  * locks added to @cancels list.
 632  */
 633 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
 634                                    struct list_head *cancels,
 635                                    enum ldlm_mode mode, __u64 lock_flags)
 636 {
 637         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
 638         struct ldlm_res_id res_id;
 639         struct ldlm_resource *res;
 640         int count;
 641
 642         /* Return, i.e. cancel nothing, only if ELC is supported (flag in
 643          * export) but disabled through procfs (flag in NS).
 644          *
 645          * This distinguishes from a case when ELC is not supported originally,
 646          * when we still want to cancel locks in advance and just cancel them
 647          * locally, without sending any RPC.
 648          */
 649         if (exp_connect_cancelset(exp) && !ns_connect_cancelset(ns))
 650                 return 0;
 651
 652         ostid_build_res_name(&oa->o_oi, &res_id);
 653         res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
 654         if (!res)
 655                 return 0;
 656
 657         LDLM_RESOURCE_ADDREF(res);
 658         count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
 659                                            lock_flags, 0, NULL);
 660         LDLM_RESOURCE_DELREF(res);
 661         ldlm_resource_putref(res);
 662         return count;
 663 }
 664
 665 static int osc_destroy_interpret(const struct lu_env *env,
 666                                  struct ptlrpc_request *req, void *data,
 667                                  int rc)
 668 {
 669         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
 670
 671         atomic_dec(&cli->cl_destroy_in_flight);
 672         wake_up(&cli->cl_destroy_waitq);
 673         return 0;
 674 }
 675
 676 static int osc_can_send_destroy(struct client_obd *cli)
 677 {
 678         if (atomic_inc_return(&cli->cl_destroy_in_flight) <=
 679             cli->cl_max_rpcs_in_flight) {
 680                 /* The destroy request can be sent */
 681                 return 1;
 682         }
 683         if (atomic_dec_return(&cli->cl_destroy_in_flight) <
 684             cli->cl_max_rpcs_in_flight) {
 685                 /*
 686                  * The counter has been modified between the two atomic
 687                  * operations.
 688                  */
 689                 wake_up(&cli->cl_destroy_waitq);
 690         }
 691         return 0;
 692 }
 693
 694 static int osc_create(const struct lu_env *env, struct obd_export *exp,
 695                       struct obdo *oa, struct lov_stripe_md **ea,
 696                       struct obd_trans_info *oti)
 697 {
 698         int rc = 0;
 699
 700         LASSERT(oa);
 701         LASSERT(ea);
 702         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
 703
 704         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 705             oa->o_flags == OBD_FL_RECREATE_OBJS) {
 706                 return osc_real_create(exp, oa, ea, oti);
 707         }
 708
 709         if (!fid_seq_is_mdt(ostid_seq(&oa->o_oi)))
 710                 return osc_real_create(exp, oa, ea, oti);
 711
 712         /* we should not get here anymore */
 713         LBUG();
 714
 715         return rc;
 716 }
 717
 718 /* Destroy requests can be async always on the client, and we don't even really
 719  * care about the return code since the client cannot do anything at all about
 720  * a destroy failure.
 721  * When the MDS is unlinking a filename, it saves the file objects into a
 722  * recovery llog, and these object records are cancelled when the OST reports
 723  * they were destroyed and sync'd to disk (i.e. transaction committed).
 724  * If the client dies, or the OST is down when the object should be destroyed,
 725  * the records are not cancelled, and when the OST reconnects to the MDS next,
 726  * it will retrieve the llog unlink logs and then sends the log cancellation
 727  * cookies to the MDS after committing destroy transactions.
 728  */
 729 static int osc_destroy(const struct lu_env *env, struct obd_export *exp,
 730                        struct obdo *oa, struct lov_stripe_md *ea,
 731                        struct obd_trans_info *oti, struct obd_export *md_export)
 732 {
 733         struct client_obd *cli = &exp->exp_obd->u.cli;
 734         struct ptlrpc_request *req;
 735         struct ost_body *body;
 736         LIST_HEAD(cancels);
 737         int rc, count;
 738
 739         if (!oa) {
 740                 CDEBUG(D_INFO, "oa NULL\n");
 741                 return -EINVAL;
 742         }
 743
 744         count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
 745                                         LDLM_FL_DISCARD_DATA);
 746
 747         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
 748         if (!req) {
 749                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
 750                 return -ENOMEM;
 751         }
 752
 753         rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
 754                                0, &cancels, count);
 755         if (rc) {
 756                 ptlrpc_request_free(req);
 757                 return rc;
 758         }
 759
 760         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
 761         ptlrpc_at_set_req_timeout(req);
 762
 763         if (oti && oa->o_valid & OBD_MD_FLCOOKIE)
 764                 oa->o_lcookie = *oti->oti_logcookies;
 765         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 766         LASSERT(body);
 767         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
 768
 769         ptlrpc_request_set_replen(req);
 770
 771         /* If osc_destroy is for destroying the unlink orphan,
 772          * sent from MDT to OST, which should not be blocked here,
 773          * because the process might be triggered by ptlrpcd, and
 774          * it is not good to block ptlrpcd thread (b=16006
 775          **/
 776         if (!(oa->o_flags & OBD_FL_DELORPHAN)) {
 777                 req->rq_interpret_reply = osc_destroy_interpret;
 778                 if (!osc_can_send_destroy(cli)) {
 779                         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP,
 780                                                           NULL);
 781
 782                         /*
 783                          * Wait until the number of on-going destroy RPCs drops
 784                          * under max_rpc_in_flight
 785                          */
 786                         l_wait_event_exclusive(cli->cl_destroy_waitq,
 787                                                osc_can_send_destroy(cli), &lwi);
 788                 }
 789         }
 790
 791         /* Do not wait for response */
 792         ptlrpcd_add_req(req);
 793         return 0;
 794 }
 795
 796 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
 797                                 long writing_bytes)
 798 {
 799         u32 bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
 800
 801         LASSERT(!(oa->o_valid & bits));
 802
 803         oa->o_valid |= bits;
 804         client_obd_list_lock(&cli->cl_loi_list_lock);
 805         oa->o_dirty = cli->cl_dirty;
 806         if (unlikely(cli->cl_dirty - cli->cl_dirty_transit >
 807                      cli->cl_dirty_max)) {
 808                 CERROR("dirty %lu - %lu > dirty_max %lu\n",
 809                        cli->cl_dirty, cli->cl_dirty_transit, cli->cl_dirty_max);
 810                 oa->o_undirty = 0;
 811         } else if (unlikely(atomic_read(&obd_dirty_pages) -
 812                             atomic_read(&obd_dirty_transit_pages) >
 813                             (long)(obd_max_dirty_pages + 1))) {
 814                 /* The atomic_read() allowing the atomic_inc() are
 815                  * not covered by a lock thus they may safely race and trip
 816                  * this CERROR() unless we add in a small fudge factor (+1).
 817                  */
 818                 CERROR("dirty %d - %d > system dirty_max %d\n",
 819                        atomic_read(&obd_dirty_pages),
 820                        atomic_read(&obd_dirty_transit_pages),
 821                        obd_max_dirty_pages);
 822                 oa->o_undirty = 0;
 823         } else if (unlikely(cli->cl_dirty_max - cli->cl_dirty > 0x7fffffff)) {
 824                 CERROR("dirty %lu - dirty_max %lu too big???\n",
 825                        cli->cl_dirty, cli->cl_dirty_max);
 826                 oa->o_undirty = 0;
 827         } else {
 828                 long max_in_flight = (cli->cl_max_pages_per_rpc <<
 829                                       PAGE_CACHE_SHIFT)*
 830                                      (cli->cl_max_rpcs_in_flight + 1);
 831                 oa->o_undirty = max(cli->cl_dirty_max, max_in_flight);
 832         }
 833         oa->o_grant = cli->cl_avail_grant + cli->cl_reserved_grant;
 834         oa->o_dropped = cli->cl_lost_grant;
 835         cli->cl_lost_grant = 0;
 836         client_obd_list_unlock(&cli->cl_loi_list_lock);
 837         CDEBUG(D_CACHE, "dirty: %llu undirty: %u dropped %u grant: %llu\n",
 838                oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
 839
 840 }
 841
 842 void osc_update_next_shrink(struct client_obd *cli)
 843 {
 844         cli->cl_next_shrink_grant =
 845                 cfs_time_shift(cli->cl_grant_shrink_interval);
 846         CDEBUG(D_CACHE, "next time %ld to shrink grant\n",
 847                cli->cl_next_shrink_grant);
 848 }
 849
 850 static void __osc_update_grant(struct client_obd *cli, u64 grant)
 851 {
 852         client_obd_list_lock(&cli->cl_loi_list_lock);
 853         cli->cl_avail_grant += grant;
 854         client_obd_list_unlock(&cli->cl_loi_list_lock);
 855 }
 856
 857 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
 858 {
 859         if (body->oa.o_valid & OBD_MD_FLGRANT) {
 860                 CDEBUG(D_CACHE, "got %llu extra grant\n", body->oa.o_grant);
 861                 __osc_update_grant(cli, body->oa.o_grant);
 862         }
 863 }
 864
 865 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
 866                               u32 keylen, void *key, u32 vallen,
 867                               void *val, struct ptlrpc_request_set *set);
 868
 869 static int osc_shrink_grant_interpret(const struct lu_env *env,
 870                                       struct ptlrpc_request *req,
 871                                       void *aa, int rc)
 872 {
 873         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
 874         struct obdo *oa = ((struct osc_brw_async_args *)aa)->aa_oa;
 875         struct ost_body *body;
 876
 877         if (rc != 0) {
 878                 __osc_update_grant(cli, oa->o_grant);
 879                 goto out;
 880         }
 881
 882         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
 883         LASSERT(body);
 884         osc_update_grant(cli, body);
 885 out:
 886         kmem_cache_free(obdo_cachep, oa);
 887         return rc;
 888 }
 889
 890 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
 891 {
 892         client_obd_list_lock(&cli->cl_loi_list_lock);
 893         oa->o_grant = cli->cl_avail_grant / 4;
 894         cli->cl_avail_grant -= oa->o_grant;
 895         client_obd_list_unlock(&cli->cl_loi_list_lock);
 896         if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
 897                 oa->o_valid |= OBD_MD_FLFLAGS;
 898                 oa->o_flags = 0;
 899         }
 900         oa->o_flags |= OBD_FL_SHRINK_GRANT;
 901         osc_update_next_shrink(cli);
 902 }
 903
 904 /* Shrink the current grant, either from some large amount to enough for a
 905  * full set of in-flight RPCs, or if we have already shrunk to that limit
 906  * then to enough for a single RPC.  This avoids keeping more grant than
 907  * needed, and avoids shrinking the grant piecemeal.
 908  */
 909 static int osc_shrink_grant(struct client_obd *cli)
 910 {
 911         __u64 target_bytes = (cli->cl_max_rpcs_in_flight + 1) *
 912                              (cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT);
 913
 914         client_obd_list_lock(&cli->cl_loi_list_lock);
 915         if (cli->cl_avail_grant <= target_bytes)
 916                 target_bytes = cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT;
 917         client_obd_list_unlock(&cli->cl_loi_list_lock);
 918
 919         return osc_shrink_grant_to_target(cli, target_bytes);
 920 }
 921
 922 int osc_shrink_grant_to_target(struct client_obd *cli, __u64 target_bytes)
 923 {
 924         int rc = 0;
 925         struct ost_body *body;
 926
 927         client_obd_list_lock(&cli->cl_loi_list_lock);
 928         /* Don't shrink if we are already above or below the desired limit
 929          * We don't want to shrink below a single RPC, as that will negatively
 930          * impact block allocation and long-term performance.
 931          */
 932         if (target_bytes < cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT)
 933                 target_bytes = cli->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT;
 934
 935         if (target_bytes >= cli->cl_avail_grant) {
 936                 client_obd_list_unlock(&cli->cl_loi_list_lock);
 937                 return 0;
 938         }
 939         client_obd_list_unlock(&cli->cl_loi_list_lock);
 940
 941         body = kzalloc(sizeof(*body), GFP_NOFS);
 942         if (!body)
 943                 return -ENOMEM;
 944
 945         osc_announce_cached(cli, &body->oa, 0);
 946
 947         client_obd_list_lock(&cli->cl_loi_list_lock);
 948         body->oa.o_grant = cli->cl_avail_grant - target_bytes;
 949         cli->cl_avail_grant = target_bytes;
 950         client_obd_list_unlock(&cli->cl_loi_list_lock);
 951         if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
 952                 body->oa.o_valid |= OBD_MD_FLFLAGS;
 953                 body->oa.o_flags = 0;
 954         }
 955         body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
 956         osc_update_next_shrink(cli);
 957
 958         rc = osc_set_info_async(NULL, cli->cl_import->imp_obd->obd_self_export,
 959                                 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
 960                                 sizeof(*body), body, NULL);
 961         if (rc != 0)
 962                 __osc_update_grant(cli, body->oa.o_grant);
 963         kfree(body);
 964         return rc;
 965 }
 966
 967 static int osc_should_shrink_grant(struct client_obd *client)
 968 {
 969         unsigned long time = cfs_time_current();
 970         unsigned long next_shrink = client->cl_next_shrink_grant;
 971
 972         if ((client->cl_import->imp_connect_data.ocd_connect_flags &
 973              OBD_CONNECT_GRANT_SHRINK) == 0)
 974                 return 0;
 975
 976         if (cfs_time_aftereq(time, next_shrink - 5 * CFS_TICK)) {
 977                 /* Get the current RPC size directly, instead of going via:
 978                  * cli_brw_size(obd->u.cli.cl_import->imp_obd->obd_self_export)
 979                  * Keep comment here so that it can be found by searching.
 980                  */
 981                 int brw_size = client->cl_max_pages_per_rpc << PAGE_CACHE_SHIFT;
 982
 983                 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
 984                     client->cl_avail_grant > brw_size)
 985                         return 1;
 986
 987                 osc_update_next_shrink(client);
 988         }
 989         return 0;
 990 }
 991
 992 static int osc_grant_shrink_grant_cb(struct timeout_item *item, void *data)
 993 {
 994         struct client_obd *client;
 995
 996         list_for_each_entry(client, &item->ti_obd_list,
 997                                 cl_grant_shrink_list) {
 998                 if (osc_should_shrink_grant(client))
 999                         osc_shrink_grant(client);
1000         }
1001         return 0;
1002 }
1003
1004 static int osc_add_shrink_grant(struct client_obd *client)
1005 {
1006         int rc;
1007
1008         rc = ptlrpc_add_timeout_client(client->cl_grant_shrink_interval,
1009                                        TIMEOUT_GRANT,
1010                                        osc_grant_shrink_grant_cb, NULL,
1011                                        &client->cl_grant_shrink_list);
1012         if (rc) {
1013                 CERROR("add grant client %s error %d\n",
1014                         client->cl_import->imp_obd->obd_name, rc);
1015                 return rc;
1016         }
1017         CDEBUG(D_CACHE, "add grant client %s\n",
1018                client->cl_import->imp_obd->obd_name);
1019         osc_update_next_shrink(client);
1020         return 0;
1021 }
1022
1023 static int osc_del_shrink_grant(struct client_obd *client)
1024 {
1025         return ptlrpc_del_timeout_client(&client->cl_grant_shrink_list,
1026                                          TIMEOUT_GRANT);
1027 }
1028
1029 static void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1030 {
1031         /*
1032          * ocd_grant is the total grant amount we're expect to hold: if we've
1033          * been evicted, it's the new avail_grant amount, cl_dirty will drop
1034          * to 0 as inflight RPCs fail out; otherwise, it's avail_grant + dirty.
1035          *
1036          * race is tolerable here: if we're evicted, but imp_state already
1037          * left EVICTED state, then cl_dirty must be 0 already.
1038          */
1039         client_obd_list_lock(&cli->cl_loi_list_lock);
1040         if (cli->cl_import->imp_state == LUSTRE_IMP_EVICTED)
1041                 cli->cl_avail_grant = ocd->ocd_grant;
1042         else
1043                 cli->cl_avail_grant = ocd->ocd_grant - cli->cl_dirty;
1044
1045         if (cli->cl_avail_grant < 0) {
1046                 CWARN("%s: available grant < 0: avail/ocd/dirty %ld/%u/%ld\n",
1047                       cli->cl_import->imp_obd->obd_name, cli->cl_avail_grant,
1048                       ocd->ocd_grant, cli->cl_dirty);
1049                 /* workaround for servers which do not have the patch from
1050                  * LU-2679
1051                  */
1052                 cli->cl_avail_grant = ocd->ocd_grant;
1053         }
1054
1055         /* determine the appropriate chunk size used by osc_extent. */
1056         cli->cl_chunkbits = max_t(int, PAGE_CACHE_SHIFT, ocd->ocd_blocksize);
1057         client_obd_list_unlock(&cli->cl_loi_list_lock);
1058
1059         CDEBUG(D_CACHE, "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld chunk bits: %d\n",
1060                cli->cl_import->imp_obd->obd_name,
1061                cli->cl_avail_grant, cli->cl_lost_grant, cli->cl_chunkbits);
1062
1063         if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT_SHRINK &&
1064             list_empty(&cli->cl_grant_shrink_list))
1065                 osc_add_shrink_grant(cli);
1066 }
1067
1068 /* We assume that the reason this OSC got a short read is because it read
1069  * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1070  * via the LOV, and it _knows_ it's reading inside the file, it's just that
1071  * this stripe never got written at or beyond this stripe offset yet.
1072  */
1073 static void handle_short_read(int nob_read, u32 page_count,
1074                               struct brw_page **pga)
1075 {
1076         char *ptr;
1077         int i = 0;
1078
1079         /* skip bytes read OK */
1080         while (nob_read > 0) {
1081                 LASSERT(page_count > 0);
1082
1083                 if (pga[i]->count > nob_read) {
1084                         /* EOF inside this page */
1085                         ptr = kmap(pga[i]->pg) +
1086                                 (pga[i]->off & ~CFS_PAGE_MASK);
1087                         memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1088                         kunmap(pga[i]->pg);
1089                         page_count--;
1090                         i++;
1091                         break;
1092                 }
1093
1094                 nob_read -= pga[i]->count;
1095                 page_count--;
1096                 i++;
1097         }
1098
1099         /* zero remaining pages */
1100         while (page_count-- > 0) {
1101                 ptr = kmap(pga[i]->pg) + (pga[i]->off & ~CFS_PAGE_MASK);
1102                 memset(ptr, 0, pga[i]->count);
1103                 kunmap(pga[i]->pg);
1104                 i++;
1105         }
1106 }
1107
1108 static int check_write_rcs(struct ptlrpc_request *req,
1109                            int requested_nob, int niocount,
1110                            u32 page_count, struct brw_page **pga)
1111 {
1112         int i;
1113         __u32 *remote_rcs;
1114
1115         remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1116                                                   sizeof(*remote_rcs) *
1117                                                   niocount);
1118         if (!remote_rcs) {
1119                 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1120                 return -EPROTO;
1121         }
1122
1123         /* return error if any niobuf was in error */
1124         for (i = 0; i < niocount; i++) {
1125                 if ((int)remote_rcs[i] < 0)
1126                         return remote_rcs[i];
1127
1128                 if (remote_rcs[i] != 0) {
1129                         CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1130                                 i, remote_rcs[i], req);
1131                         return -EPROTO;
1132                 }
1133         }
1134
1135         if (req->rq_bulk->bd_nob_transferred != requested_nob) {
1136                 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1137                        req->rq_bulk->bd_nob_transferred, requested_nob);
1138                 return -EPROTO;
1139         }
1140
1141         return 0;
1142 }
1143
1144 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1145 {
1146         if (p1->flag != p2->flag) {
1147                 unsigned mask = ~(OBD_BRW_FROM_GRANT | OBD_BRW_NOCACHE |
1148                                   OBD_BRW_SYNC | OBD_BRW_ASYNC|OBD_BRW_NOQUOTA);
1149
1150                 /* warn if we try to combine flags that we don't know to be
1151                  * safe to combine
1152                  */
1153                 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
1154                         CWARN("Saw flags 0x%x and 0x%x in the same brw, please report this at http://bugs.whamcloud.com/\n",
1155                               p1->flag, p2->flag);
1156                 }
1157                 return 0;
1158         }
1159
1160         return (p1->off + p1->count == p2->off);
1161 }
1162
1163 static u32 osc_checksum_bulk(int nob, u32 pg_count,
1164                              struct brw_page **pga, int opc,
1165                              enum cksum_type cksum_type)
1166 {
1167         __u32 cksum;
1168         int i = 0;
1169         struct cfs_crypto_hash_desc *hdesc;
1170         unsigned int bufsize;
1171         int err;
1172         unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1173
1174         LASSERT(pg_count > 0);
1175
1176         hdesc = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1177         if (IS_ERR(hdesc)) {
1178                 CERROR("Unable to initialize checksum hash %s\n",
1179                        cfs_crypto_hash_name(cfs_alg));
1180                 return PTR_ERR(hdesc);
1181         }
1182
1183         while (nob > 0 && pg_count > 0) {
1184                 int count = pga[i]->count > nob ? nob : pga[i]->count;
1185
1186                 /* corrupt the data before we compute the checksum, to
1187                  * simulate an OST->client data error
1188                  */
1189                 if (i == 0 && opc == OST_READ &&
1190                     OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE)) {
1191                         unsigned char *ptr = kmap(pga[i]->pg);
1192                         int off = pga[i]->off & ~CFS_PAGE_MASK;
1193
1194                         memcpy(ptr + off, "bad1", min(4, nob));
1195                         kunmap(pga[i]->pg);
1196                 }
1197                 cfs_crypto_hash_update_page(hdesc, pga[i]->pg,
1198                                   pga[i]->off & ~CFS_PAGE_MASK,
1199                                   count);
1200                 CDEBUG(D_PAGE,
1201                        "page %p map %p index %lu flags %lx count %u priv %0lx: off %d\n",
1202                        pga[i]->pg, pga[i]->pg->mapping, pga[i]->pg->index,
1203                        (long)pga[i]->pg->flags, page_count(pga[i]->pg),
1204                        page_private(pga[i]->pg),
1205                        (int)(pga[i]->off & ~CFS_PAGE_MASK));
1206
1207                 nob -= pga[i]->count;
1208                 pg_count--;
1209                 i++;
1210         }
1211
1212         bufsize = 4;
1213         err = cfs_crypto_hash_final(hdesc, (unsigned char *)&cksum, &bufsize);
1214
1215         if (err)
1216                 cfs_crypto_hash_final(hdesc, NULL, NULL);
1217
1218         /* For sending we only compute the wrong checksum instead
1219          * of corrupting the data so it is still correct on a redo
1220          */
1221         if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1222                 cksum++;
1223
1224         return cksum;
1225 }
1226
1227 static int osc_brw_prep_request(int cmd, struct client_obd *cli,
1228                                 struct obdo *oa,
1229                                 struct lov_stripe_md *lsm, u32 page_count,
1230                                 struct brw_page **pga,
1231                                 struct ptlrpc_request **reqp,
1232                                 int reserve,
1233                                 int resend)
1234 {
1235         struct ptlrpc_request *req;
1236         struct ptlrpc_bulk_desc *desc;
1237         struct ost_body *body;
1238         struct obd_ioobj *ioobj;
1239         struct niobuf_remote *niobuf;
1240         int niocount, i, requested_nob, opc, rc;
1241         struct osc_brw_async_args *aa;
1242         struct req_capsule *pill;
1243         struct brw_page *pg_prev;
1244
1245         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1246                 return -ENOMEM; /* Recoverable */
1247         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1248                 return -EINVAL; /* Fatal */
1249
1250         if ((cmd & OBD_BRW_WRITE) != 0) {
1251                 opc = OST_WRITE;
1252                 req = ptlrpc_request_alloc_pool(cli->cl_import,
1253                                                 osc_rq_pool,
1254                                                 &RQF_OST_BRW_WRITE);
1255         } else {
1256                 opc = OST_READ;
1257                 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1258         }
1259         if (!req)
1260                 return -ENOMEM;
1261
1262         for (niocount = i = 1; i < page_count; i++) {
1263                 if (!can_merge_pages(pga[i - 1], pga[i]))
1264                         niocount++;
1265         }
1266
1267         pill = &req->rq_pill;
1268         req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1269                              sizeof(*ioobj));
1270         req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1271                              niocount * sizeof(*niobuf));
1272
1273         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1274         if (rc) {
1275                 ptlrpc_request_free(req);
1276                 return rc;
1277         }
1278         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
1279         ptlrpc_at_set_req_timeout(req);
1280         /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1281          * retry logic
1282          */
1283         req->rq_no_retry_einprogress = 1;
1284
1285         desc = ptlrpc_prep_bulk_imp(req, page_count,
1286                 cli->cl_import->imp_connect_data.ocd_brw_size >> LNET_MTU_BITS,
1287                 opc == OST_WRITE ? BULK_GET_SOURCE : BULK_PUT_SINK,
1288                 OST_BULK_PORTAL);
1289
1290         if (!desc) {
1291                 rc = -ENOMEM;
1292                 goto out;
1293         }
1294         /* NB request now owns desc and will free it when it gets freed */
1295
1296         body = req_capsule_client_get(pill, &RMF_OST_BODY);
1297         ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1298         niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1299         LASSERT(body && ioobj && niobuf);
1300
1301         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1302
1303         obdo_to_ioobj(oa, ioobj);
1304         ioobj->ioo_bufcnt = niocount;
1305         /* The high bits of ioo_max_brw tells server _maximum_ number of bulks
1306          * that might be send for this request.  The actual number is decided
1307          * when the RPC is finally sent in ptlrpc_register_bulk(). It sends
1308          * "max - 1" for old client compatibility sending "0", and also so the
1309          * the actual maximum is a power-of-two number, not one less. LU-1431
1310          */
1311         ioobj_max_brw_set(ioobj, desc->bd_md_max_brw);
1312         LASSERT(page_count > 0);
1313         pg_prev = pga[0];
1314         for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1315                 struct brw_page *pg = pga[i];
1316                 int poff = pg->off & ~CFS_PAGE_MASK;
1317
1318                 LASSERT(pg->count > 0);
1319                 /* make sure there is no gap in the middle of page array */
1320                 LASSERTF(page_count == 1 ||
1321                          (ergo(i == 0, poff + pg->count == PAGE_CACHE_SIZE) &&
1322                           ergo(i > 0 && i < page_count - 1,
1323                                poff == 0 && pg->count == PAGE_CACHE_SIZE)   &&
1324                           ergo(i == page_count - 1, poff == 0)),
1325                          "i: %d/%d pg: %p off: %llu, count: %u\n",
1326                          i, page_count, pg, pg->off, pg->count);
1327                 LASSERTF(i == 0 || pg->off > pg_prev->off,
1328                          "i %d p_c %u pg %p [pri %lu ind %lu] off %llu prev_pg %p [pri %lu ind %lu] off %llu\n",
1329                          i, page_count,
1330                          pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1331                          pg_prev->pg, page_private(pg_prev->pg),
1332                          pg_prev->pg->index, pg_prev->off);
1333                 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1334                         (pg->flag & OBD_BRW_SRVLOCK));
1335
1336                 ptlrpc_prep_bulk_page_pin(desc, pg->pg, poff, pg->count);
1337                 requested_nob += pg->count;
1338
1339                 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1340                         niobuf--;
1341                         niobuf->len += pg->count;
1342                 } else {
1343                         niobuf->offset = pg->off;
1344                         niobuf->len = pg->count;
1345                         niobuf->flags = pg->flag;
1346                 }
1347                 pg_prev = pg;
1348         }
1349
1350         LASSERTF((void *)(niobuf - niocount) ==
1351                 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1352                 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1353                 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1354
1355         osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1356         if (resend) {
1357                 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1358                         body->oa.o_valid |= OBD_MD_FLFLAGS;
1359                         body->oa.o_flags = 0;
1360                 }
1361                 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1362         }
1363
1364         if (osc_should_shrink_grant(cli))
1365                 osc_shrink_grant_local(cli, &body->oa);
1366
1367         /* size[REQ_REC_OFF] still sizeof (*body) */
1368         if (opc == OST_WRITE) {
1369                 if (cli->cl_checksum &&
1370                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1371                         /* store cl_cksum_type in a local variable since
1372                          * it can be changed via lprocfs
1373                          */
1374                         enum cksum_type cksum_type = cli->cl_cksum_type;
1375
1376                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1377                                 oa->o_flags &= OBD_FL_LOCAL_MASK;
1378                                 body->oa.o_flags = 0;
1379                         }
1380                         body->oa.o_flags |= cksum_type_pack(cksum_type);
1381                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1382                         body->oa.o_cksum = osc_checksum_bulk(requested_nob,
1383                                                              page_count, pga,
1384                                                              OST_WRITE,
1385                                                              cksum_type);
1386                         CDEBUG(D_PAGE, "checksum at write origin: %x\n",
1387                                body->oa.o_cksum);
1388                         /* save this in 'oa', too, for later checking */
1389                         oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1390                         oa->o_flags |= cksum_type_pack(cksum_type);
1391                 } else {
1392                         /* clear out the checksum flag, in case this is a
1393                          * resend but cl_checksum is no longer set. b=11238
1394                          */
1395                         oa->o_valid &= ~OBD_MD_FLCKSUM;
1396                 }
1397                 oa->o_cksum = body->oa.o_cksum;
1398                 /* 1 RC per niobuf */
1399                 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1400                                      sizeof(__u32) * niocount);
1401         } else {
1402                 if (cli->cl_checksum &&
1403                     !sptlrpc_flavor_has_bulk(&req->rq_flvr)) {
1404                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1405                                 body->oa.o_flags = 0;
1406                         body->oa.o_flags |= cksum_type_pack(cli->cl_cksum_type);
1407                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1408                 }
1409         }
1410         ptlrpc_request_set_replen(req);
1411
1412         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1413         aa = ptlrpc_req_async_args(req);
1414         aa->aa_oa = oa;
1415         aa->aa_requested_nob = requested_nob;
1416         aa->aa_nio_count = niocount;
1417         aa->aa_page_count = page_count;
1418         aa->aa_resends = 0;
1419         aa->aa_ppga = pga;
1420         aa->aa_cli = cli;
1421         INIT_LIST_HEAD(&aa->aa_oaps);
1422
1423         *reqp = req;
1424         return 0;
1425
1426  out:
1427         ptlrpc_req_finished(req);
1428         return rc;
1429 }
1430
1431 static int check_write_checksum(struct obdo *oa, const lnet_process_id_t *peer,
1432                                 __u32 client_cksum, __u32 server_cksum, int nob,
1433                                 u32 page_count, struct brw_page **pga,
1434                                 enum cksum_type client_cksum_type)
1435 {
1436         __u32 new_cksum;
1437         char *msg;
1438         enum cksum_type cksum_type;
1439
1440         if (server_cksum == client_cksum) {
1441                 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1442                 return 0;
1443         }
1444
1445         cksum_type = cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1446                                        oa->o_flags : 0);
1447         new_cksum = osc_checksum_bulk(nob, page_count, pga, OST_WRITE,
1448                                       cksum_type);
1449
1450         if (cksum_type != client_cksum_type)
1451                 msg = "the server did not use the checksum type specified in the original request - likely a protocol problem"
1452                         ;
1453         else if (new_cksum == server_cksum)
1454                 msg = "changed on the client after we checksummed it - likely false positive due to mmap IO (bug 11742)"
1455                         ;
1456         else if (new_cksum == client_cksum)
1457                 msg = "changed in transit before arrival at OST";
1458         else
1459                 msg = "changed in transit AND doesn't match the original - likely false positive due to mmap IO (bug 11742)"
1460                         ;
1461
1462         LCONSOLE_ERROR_MSG(0x132, "BAD WRITE CHECKSUM: %s: from %s inode "DFID
1463                            " object "DOSTID" extent [%llu-%llu]\n",
1464                            msg, libcfs_nid2str(peer->nid),
1465                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1466                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1467                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1468                            POSTID(&oa->o_oi), pga[0]->off,
1469                            pga[page_count-1]->off + pga[page_count-1]->count - 1);
1470         CERROR("original client csum %x (type %x), server csum %x (type %x), client csum now %x\n",
1471                client_cksum, client_cksum_type,
1472                server_cksum, cksum_type, new_cksum);
1473         return 1;
1474 }
1475
1476 /* Note rc enters this function as number of bytes transferred */
1477 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1478 {
1479         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1480         const lnet_process_id_t *peer =
1481                         &req->rq_import->imp_connection->c_peer;
1482         struct client_obd *cli = aa->aa_cli;
1483         struct ost_body *body;
1484         __u32 client_cksum = 0;
1485
1486         if (rc < 0 && rc != -EDQUOT) {
1487                 DEBUG_REQ(D_INFO, req, "Failed request with rc = %d\n", rc);
1488                 return rc;
1489         }
1490
1491         LASSERTF(req->rq_repmsg, "rc = %d\n", rc);
1492         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1493         if (!body) {
1494                 DEBUG_REQ(D_INFO, req, "Can't unpack body\n");
1495                 return -EPROTO;
1496         }
1497
1498         /* set/clear over quota flag for a uid/gid */
1499         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1500             body->oa.o_valid & (OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA)) {
1501                 unsigned int qid[MAXQUOTAS] = { body->oa.o_uid, body->oa.o_gid };
1502
1503                 CDEBUG(D_QUOTA, "setdq for [%u %u] with valid %#llx, flags %x\n",
1504                        body->oa.o_uid, body->oa.o_gid, body->oa.o_valid,
1505                        body->oa.o_flags);
1506                 osc_quota_setdq(cli, qid, body->oa.o_valid, body->oa.o_flags);
1507         }
1508
1509         osc_update_grant(cli, body);
1510
1511         if (rc < 0)
1512                 return rc;
1513
1514         if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1515                 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1516
1517         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1518                 if (rc > 0) {
1519                         CERROR("Unexpected +ve rc %d\n", rc);
1520                         return -EPROTO;
1521                 }
1522                 LASSERT(req->rq_bulk->bd_nob == aa->aa_requested_nob);
1523
1524                 if (sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1525                         return -EAGAIN;
1526
1527                 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
1528                     check_write_checksum(&body->oa, peer, client_cksum,
1529                                          body->oa.o_cksum, aa->aa_requested_nob,
1530                                          aa->aa_page_count, aa->aa_ppga,
1531                                          cksum_type_unpack(aa->aa_oa->o_flags)))
1532                         return -EAGAIN;
1533
1534                 rc = check_write_rcs(req, aa->aa_requested_nob,
1535                                      aa->aa_nio_count,
1536                                      aa->aa_page_count, aa->aa_ppga);
1537                 goto out;
1538         }
1539
1540         /* The rest of this function executes only for OST_READs */
1541
1542         /* if unwrap_bulk failed, return -EAGAIN to retry */
1543         rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
1544         if (rc < 0) {
1545                 rc = -EAGAIN;
1546                 goto out;
1547         }
1548
1549         if (rc > aa->aa_requested_nob) {
1550                 CERROR("Unexpected rc %d (%d requested)\n", rc,
1551                        aa->aa_requested_nob);
1552                 return -EPROTO;
1553         }
1554
1555         if (rc != req->rq_bulk->bd_nob_transferred) {
1556                 CERROR("Unexpected rc %d (%d transferred)\n",
1557                         rc, req->rq_bulk->bd_nob_transferred);
1558                 return -EPROTO;
1559         }
1560
1561         if (rc < aa->aa_requested_nob)
1562                 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
1563
1564         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1565                 static int cksum_counter;
1566                 __u32 server_cksum = body->oa.o_cksum;
1567                 char *via = "";
1568                 char *router = "";
1569                 enum cksum_type cksum_type;
1570
1571                 cksum_type = cksum_type_unpack(body->oa.o_valid&OBD_MD_FLFLAGS ?
1572                                                body->oa.o_flags : 0);
1573                 client_cksum = osc_checksum_bulk(rc, aa->aa_page_count,
1574                                                  aa->aa_ppga, OST_READ,
1575                                                  cksum_type);
1576
1577                 if (peer->nid != req->rq_bulk->bd_sender) {
1578                         via = " via ";
1579                         router = libcfs_nid2str(req->rq_bulk->bd_sender);
1580                 }
1581
1582                 if (server_cksum != client_cksum) {
1583                         LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from %s%s%s inode " DFID " object " DOSTID " extent [%llu-%llu]\n",
1584                                            req->rq_import->imp_obd->obd_name,
1585                                            libcfs_nid2str(peer->nid),
1586                                            via, router,
1587                                            body->oa.o_valid & OBD_MD_FLFID ?
1588                                            body->oa.o_parent_seq : (__u64)0,
1589                                            body->oa.o_valid & OBD_MD_FLFID ?
1590                                            body->oa.o_parent_oid : 0,
1591                                            body->oa.o_valid & OBD_MD_FLFID ?
1592                                            body->oa.o_parent_ver : 0,
1593                                            POSTID(&body->oa.o_oi),
1594                                            aa->aa_ppga[0]->off,
1595                                            aa->aa_ppga[aa->aa_page_count-1]->off +
1596                                            aa->aa_ppga[aa->aa_page_count-1]->count -
1597                                            1);
1598                         CERROR("client %x, server %x, cksum_type %x\n",
1599                                client_cksum, server_cksum, cksum_type);
1600                         cksum_counter = 0;
1601                         aa->aa_oa->o_cksum = client_cksum;
1602                         rc = -EAGAIN;
1603                 } else {
1604                         cksum_counter++;
1605                         CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1606                         rc = 0;
1607                 }
1608         } else if (unlikely(client_cksum)) {
1609                 static int cksum_missed;
1610
1611                 cksum_missed++;
1612                 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1613                         CERROR("Checksum %u requested from %s but not sent\n",
1614                                cksum_missed, libcfs_nid2str(peer->nid));
1615         } else {
1616                 rc = 0;
1617         }
1618 out:
1619         if (rc >= 0)
1620                 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
1621                                      aa->aa_oa, &body->oa);
1622
1623         return rc;
1624 }
1625
1626 static int osc_brw_redo_request(struct ptlrpc_request *request,
1627                                 struct osc_brw_async_args *aa, int rc)
1628 {
1629         struct ptlrpc_request *new_req;
1630         struct osc_brw_async_args *new_aa;
1631         struct osc_async_page *oap;
1632
1633         DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
1634                   "redo for recoverable error %d", rc);
1635
1636         rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
1637                                         OST_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
1638                                   aa->aa_cli, aa->aa_oa,
1639                                   NULL /* lsm unused by osc currently */,
1640                                   aa->aa_page_count, aa->aa_ppga,
1641                                   &new_req, 0, 1);
1642         if (rc)
1643                 return rc;
1644
1645         list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1646                 if (oap->oap_request) {
1647                         LASSERTF(request == oap->oap_request,
1648                                  "request %p != oap_request %p\n",
1649                                  request, oap->oap_request);
1650                         if (oap->oap_interrupted) {
1651                                 ptlrpc_req_finished(new_req);
1652                                 return -EINTR;
1653                         }
1654                 }
1655         }
1656         /* New request takes over pga and oaps from old request.
1657          * Note that copying a list_head doesn't work, need to move it...
1658          */
1659         aa->aa_resends++;
1660         new_req->rq_interpret_reply = request->rq_interpret_reply;
1661         new_req->rq_async_args = request->rq_async_args;
1662         /* cap resend delay to the current request timeout, this is similar to
1663          * what ptlrpc does (see after_reply())
1664          */
1665         if (aa->aa_resends > new_req->rq_timeout)
1666                 new_req->rq_sent = ktime_get_real_seconds() + new_req->rq_timeout;
1667         else
1668                 new_req->rq_sent = ktime_get_real_seconds() + aa->aa_resends;
1669         new_req->rq_generation_set = 1;
1670         new_req->rq_import_generation = request->rq_import_generation;
1671
1672         new_aa = ptlrpc_req_async_args(new_req);
1673
1674         INIT_LIST_HEAD(&new_aa->aa_oaps);
1675         list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
1676         INIT_LIST_HEAD(&new_aa->aa_exts);
1677         list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
1678         new_aa->aa_resends = aa->aa_resends;
1679
1680         list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
1681                 if (oap->oap_request) {
1682                         ptlrpc_req_finished(oap->oap_request);
1683                         oap->oap_request = ptlrpc_request_addref(new_req);
1684                 }
1685         }
1686
1687         /* XXX: This code will run into problem if we're going to support
1688          * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
1689          * and wait for all of them to be finished. We should inherit request
1690          * set from old request.
1691          */
1692         ptlrpcd_add_req(new_req);
1693
1694         DEBUG_REQ(D_INFO, new_req, "new request");
1695         return 0;
1696 }
1697
1698 /*
1699  * ugh, we want disk allocation on the target to happen in offset order.  we'll
1700  * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1701  * fine for our small page arrays and doesn't require allocation.  its an
1702  * insertion sort that swaps elements that are strides apart, shrinking the
1703  * stride down until its '1' and the array is sorted.
1704  */
1705 static void sort_brw_pages(struct brw_page **array, int num)
1706 {
1707         int stride, i, j;
1708         struct brw_page *tmp;
1709
1710         if (num == 1)
1711                 return;
1712         for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1713                 ;
1714
1715         do {
1716                 stride /= 3;
1717                 for (i = stride ; i < num ; i++) {
1718                         tmp = array[i];
1719                         j = i;
1720                         while (j >= stride && array[j - stride]->off > tmp->off) {
1721                                 array[j] = array[j - stride];
1722                                 j -= stride;
1723                         }
1724                         array[j] = tmp;
1725                 }
1726         } while (stride > 1);
1727 }
1728
1729 static void osc_release_ppga(struct brw_page **ppga, u32 count)
1730 {
1731         LASSERT(ppga);
1732         kfree(ppga);
1733 }
1734
1735 static int brw_interpret(const struct lu_env *env,
1736                          struct ptlrpc_request *req, void *data, int rc)
1737 {
1738         struct osc_brw_async_args *aa = data;
1739         struct osc_extent *ext;
1740         struct osc_extent *tmp;
1741         struct cl_object *obj = NULL;
1742         struct client_obd *cli = aa->aa_cli;
1743
1744         rc = osc_brw_fini_request(req, rc);
1745         CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
1746         /* When server return -EINPROGRESS, client should always retry
1747          * regardless of the number of times the bulk was resent already.
1748          */
1749         if (osc_recoverable_error(rc)) {
1750                 if (req->rq_import_generation !=
1751                     req->rq_import->imp_generation) {
1752                         CDEBUG(D_HA, "%s: resend cross eviction for object: " DOSTID ", rc = %d.\n",
1753                                req->rq_import->imp_obd->obd_name,
1754                                POSTID(&aa->aa_oa->o_oi), rc);
1755                 } else if (rc == -EINPROGRESS ||
1756                     client_should_resend(aa->aa_resends, aa->aa_cli)) {
1757                         rc = osc_brw_redo_request(req, aa, rc);
1758                 } else {
1759                         CERROR("%s: too many resent retries for object: %llu:%llu, rc = %d.\n",
1760                                req->rq_import->imp_obd->obd_name,
1761                                POSTID(&aa->aa_oa->o_oi), rc);
1762                 }
1763
1764                 if (rc == 0)
1765                         return 0;
1766                 else if (rc == -EAGAIN || rc == -EINPROGRESS)
1767                         rc = -EIO;
1768         }
1769
1770         list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
1771                 if (!obj && rc == 0) {
1772                         obj = osc2cl(ext->oe_obj);
1773                         cl_object_get(obj);
1774                 }
1775
1776                 list_del_init(&ext->oe_link);
1777                 osc_extent_finish(env, ext, 1, rc);
1778         }
1779         LASSERT(list_empty(&aa->aa_exts));
1780         LASSERT(list_empty(&aa->aa_oaps));
1781
1782         if (obj) {
1783                 struct obdo *oa = aa->aa_oa;
1784                 struct cl_attr *attr  = &osc_env_info(env)->oti_attr;
1785                 unsigned long valid = 0;
1786
1787                 LASSERT(rc == 0);
1788                 if (oa->o_valid & OBD_MD_FLBLOCKS) {
1789                         attr->cat_blocks = oa->o_blocks;
1790                         valid |= CAT_BLOCKS;
1791                 }
1792                 if (oa->o_valid & OBD_MD_FLMTIME) {
1793                         attr->cat_mtime = oa->o_mtime;
1794                         valid |= CAT_MTIME;
1795                 }
1796                 if (oa->o_valid & OBD_MD_FLATIME) {
1797                         attr->cat_atime = oa->o_atime;
1798                         valid |= CAT_ATIME;
1799                 }
1800                 if (oa->o_valid & OBD_MD_FLCTIME) {
1801                         attr->cat_ctime = oa->o_ctime;
1802                         valid |= CAT_CTIME;
1803                 }
1804                 if (valid != 0) {
1805                         cl_object_attr_lock(obj);
1806                         cl_object_attr_set(env, obj, attr, valid);
1807                         cl_object_attr_unlock(obj);
1808                 }
1809                 cl_object_put(env, obj);
1810         }
1811         kmem_cache_free(obdo_cachep, aa->aa_oa);
1812
1813         cl_req_completion(env, aa->aa_clerq, rc < 0 ? rc :
1814                           req->rq_bulk->bd_nob_transferred);
1815         osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
1816         ptlrpc_lprocfs_brw(req, req->rq_bulk->bd_nob_transferred);
1817
1818         client_obd_list_lock(&cli->cl_loi_list_lock);
1819         /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
1820          * is called so we know whether to go to sync BRWs or wait for more
1821          * RPCs to complete
1822          */
1823         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
1824                 cli->cl_w_in_flight--;
1825         else
1826                 cli->cl_r_in_flight--;
1827         osc_wake_cache_waiters(cli);
1828         client_obd_list_unlock(&cli->cl_loi_list_lock);
1829
1830         osc_io_unplug(env, cli, NULL);
1831         return rc;
1832 }
1833
1834 /**
1835  * Build an RPC by the list of extent @ext_list. The caller must ensure
1836  * that the total pages in this list are NOT over max pages per RPC.
1837  * Extents in the list must be in OES_RPC state.
1838  */
1839 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
1840                   struct list_head *ext_list, int cmd)
1841 {
1842         struct ptlrpc_request *req = NULL;
1843         struct osc_extent *ext;
1844         struct brw_page **pga = NULL;
1845         struct osc_brw_async_args *aa = NULL;
1846         struct obdo *oa = NULL;
1847         struct osc_async_page *oap;
1848         struct osc_async_page *tmp;
1849         struct cl_req *clerq = NULL;
1850         enum cl_req_type crt = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
1851         struct ldlm_lock *lock = NULL;
1852         struct cl_req_attr *crattr = NULL;
1853         u64 starting_offset = OBD_OBJECT_EOF;
1854         u64 ending_offset = 0;
1855         int mpflag = 0;
1856         int mem_tight = 0;
1857         int page_count = 0;
1858         int i;
1859         int rc;
1860         struct ost_body *body;
1861         LIST_HEAD(rpc_list);
1862
1863         LASSERT(!list_empty(ext_list));
1864
1865         /* add pages into rpc_list to build BRW rpc */
1866         list_for_each_entry(ext, ext_list, oe_link) {
1867                 LASSERT(ext->oe_state == OES_RPC);
1868                 mem_tight |= ext->oe_memalloc;
1869                 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
1870                         ++page_count;
1871                         list_add_tail(&oap->oap_rpc_item, &rpc_list);
1872                         if (starting_offset > oap->oap_obj_off)
1873                                 starting_offset = oap->oap_obj_off;
1874                         else
1875                                 LASSERT(oap->oap_page_off == 0);
1876                         if (ending_offset < oap->oap_obj_off + oap->oap_count)
1877                                 ending_offset = oap->oap_obj_off +
1878                                                 oap->oap_count;
1879                         else
1880                                 LASSERT(oap->oap_page_off + oap->oap_count ==
1881                                         PAGE_CACHE_SIZE);
1882                 }
1883         }
1884
1885         if (mem_tight)
1886                 mpflag = cfs_memory_pressure_get_and_set();
1887
1888         crattr = kzalloc(sizeof(*crattr), GFP_NOFS);
1889         if (!crattr) {
1890                 rc = -ENOMEM;
1891                 goto out;
1892         }
1893
1894         pga = kcalloc(page_count, sizeof(*pga), GFP_NOFS);
1895         if (!pga) {
1896                 rc = -ENOMEM;
1897                 goto out;
1898         }
1899
1900         oa = kmem_cache_alloc(obdo_cachep, GFP_NOFS | __GFP_ZERO);
1901         if (!oa) {
1902                 rc = -ENOMEM;
1903                 goto out;
1904         }
1905
1906         i = 0;
1907         list_for_each_entry(oap, &rpc_list, oap_rpc_item) {
1908                 struct cl_page *page = oap2cl_page(oap);
1909
1910                 if (!clerq) {
1911                         clerq = cl_req_alloc(env, page, crt,
1912                                              1 /* only 1-object rpcs for now */);
1913                         if (IS_ERR(clerq)) {
1914                                 rc = PTR_ERR(clerq);
1915                                 goto out;
1916                         }
1917                         lock = oap->oap_ldlm_lock;
1918                 }
1919                 if (mem_tight)
1920                         oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
1921                 pga[i] = &oap->oap_brw_page;
1922                 pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
1923                 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
1924                        pga[i]->pg, page_index(oap->oap_page), oap,
1925                        pga[i]->flag);
1926                 i++;
1927                 cl_req_page_add(env, clerq, page);
1928         }
1929
1930         /* always get the data for the obdo for the rpc */
1931         LASSERT(clerq);
1932         crattr->cra_oa = oa;
1933         cl_req_attr_set(env, clerq, crattr, ~0ULL);
1934         if (lock) {
1935                 oa->o_handle = lock->l_remote_handle;
1936                 oa->o_valid |= OBD_MD_FLHANDLE;
1937         }
1938
1939         rc = cl_req_prep(env, clerq);
1940         if (rc != 0) {
1941                 CERROR("cl_req_prep failed: %d\n", rc);
1942                 goto out;
1943         }
1944
1945         sort_brw_pages(pga, page_count);
1946         rc = osc_brw_prep_request(cmd, cli, oa, NULL, page_count,
1947                         pga, &req, 1, 0);
1948         if (rc != 0) {
1949                 CERROR("prep_req failed: %d\n", rc);
1950                 goto out;
1951         }
1952
1953         req->rq_interpret_reply = brw_interpret;
1954
1955         if (mem_tight != 0)
1956                 req->rq_memalloc = 1;
1957
1958         /* Need to update the timestamps after the request is built in case
1959          * we race with setattr (locally or in queue at OST).  If OST gets
1960          * later setattr before earlier BRW (as determined by the request xid),
1961          * the OST will not use BRW timestamps.  Sadly, there is no obvious
1962          * way to do this in a single call.  bug 10150
1963          */
1964         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1965         crattr->cra_oa = &body->oa;
1966         cl_req_attr_set(env, clerq, crattr,
1967                         OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME);
1968
1969         lustre_msg_set_jobid(req->rq_reqmsg, crattr->cra_jobid);
1970
1971         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1972         aa = ptlrpc_req_async_args(req);
1973         INIT_LIST_HEAD(&aa->aa_oaps);
1974         list_splice_init(&rpc_list, &aa->aa_oaps);
1975         INIT_LIST_HEAD(&aa->aa_exts);
1976         list_splice_init(ext_list, &aa->aa_exts);
1977         aa->aa_clerq = clerq;
1978
1979         /* queued sync pages can be torn down while the pages
1980          * were between the pending list and the rpc
1981          */
1982         tmp = NULL;
1983         list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
1984                 /* only one oap gets a request reference */
1985                 if (!tmp)
1986                         tmp = oap;
1987                 if (oap->oap_interrupted && !req->rq_intr) {
1988                         CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
1989                                         oap, req);
1990                         ptlrpc_mark_interrupted(req);
1991                 }
1992         }
1993         if (tmp)
1994                 tmp->oap_request = ptlrpc_request_addref(req);
1995
1996         client_obd_list_lock(&cli->cl_loi_list_lock);
1997         starting_offset >>= PAGE_CACHE_SHIFT;
1998         if (cmd == OBD_BRW_READ) {
1999                 cli->cl_r_in_flight++;
2000                 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2001                 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2002                 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2003                                       starting_offset + 1);
2004         } else {
2005                 cli->cl_w_in_flight++;
2006                 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2007                 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
2008                 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2009                                       starting_offset + 1);
2010         }
2011         client_obd_list_unlock(&cli->cl_loi_list_lock);
2012
2013         DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %dr/%dw in flight",
2014                   page_count, aa, cli->cl_r_in_flight,
2015                   cli->cl_w_in_flight);
2016
2017         ptlrpcd_add_req(req);
2018         rc = 0;
2019
2020 out:
2021         if (mem_tight != 0)
2022                 cfs_memory_pressure_restore(mpflag);
2023
2024         kfree(crattr);
2025
2026         if (rc != 0) {
2027                 LASSERT(!req);
2028
2029                 if (oa)
2030                         kmem_cache_free(obdo_cachep, oa);
2031                 kfree(pga);
2032                 /* this should happen rarely and is pretty bad, it makes the
2033                  * pending list not follow the dirty order
2034                  */
2035                 while (!list_empty(ext_list)) {
2036                         ext = list_entry(ext_list->next, struct osc_extent,
2037                                              oe_link);
2038                         list_del_init(&ext->oe_link);
2039                         osc_extent_finish(env, ext, 0, rc);
2040                 }
2041                 if (clerq && !IS_ERR(clerq))
2042                         cl_req_completion(env, clerq, rc);
2043         }
2044         return rc;
2045 }
2046
2047 static int osc_set_lock_data_with_check(struct ldlm_lock *lock,
2048                                         struct ldlm_enqueue_info *einfo)
2049 {
2050         void *data = einfo->ei_cbdata;
2051         int set = 0;
2052
2053         LASSERT(lock->l_blocking_ast == einfo->ei_cb_bl);
2054         LASSERT(lock->l_resource->lr_type == einfo->ei_type);
2055         LASSERT(lock->l_completion_ast == einfo->ei_cb_cp);
2056         LASSERT(lock->l_glimpse_ast == einfo->ei_cb_gl);
2057
2058         lock_res_and_lock(lock);
2059         spin_lock(&osc_ast_guard);
2060
2061         if (!lock->l_ast_data)
2062                 lock->l_ast_data = data;
2063         if (lock->l_ast_data == data)
2064                 set = 1;
2065
2066         spin_unlock(&osc_ast_guard);
2067         unlock_res_and_lock(lock);
2068
2069         return set;
2070 }
2071
2072 static int osc_set_data_with_check(struct lustre_handle *lockh,
2073                                    struct ldlm_enqueue_info *einfo)
2074 {
2075         struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2076         int set = 0;
2077
2078         if (lock) {
2079                 set = osc_set_lock_data_with_check(lock, einfo);
2080                 LDLM_LOCK_PUT(lock);
2081         } else
2082                 CERROR("lockh %p, data %p - client evicted?\n",
2083                        lockh, einfo->ei_cbdata);
2084         return set;
2085 }
2086
2087 /* find any ldlm lock of the inode in osc
2088  * return 0    not find
2089  *      1    find one
2090  *      < 0    error
2091  */
2092 static int osc_find_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2093                            ldlm_iterator_t replace, void *data)
2094 {
2095         struct ldlm_res_id res_id;
2096         struct obd_device *obd = class_exp2obd(exp);
2097         int rc = 0;
2098
2099         ostid_build_res_name(&lsm->lsm_oi, &res_id);
2100         rc = ldlm_resource_iterate(obd->obd_namespace, &res_id, replace, data);
2101         if (rc == LDLM_ITER_STOP)
2102                 return 1;
2103         if (rc == LDLM_ITER_CONTINUE)
2104                 return 0;
2105         return rc;
2106 }
2107
2108 static int osc_enqueue_fini(struct ptlrpc_request *req, struct ost_lvb *lvb,
2109                             obd_enqueue_update_f upcall, void *cookie,
2110                             __u64 *flags, int agl, int rc)
2111 {
2112         int intent = *flags & LDLM_FL_HAS_INTENT;
2113
2114         if (intent) {
2115                 /* The request was created before ldlm_cli_enqueue call. */
2116                 if (rc == ELDLM_LOCK_ABORTED) {
2117                         struct ldlm_reply *rep;
2118
2119                         rep = req_capsule_server_get(&req->rq_pill,
2120                                                      &RMF_DLM_REP);
2121
2122                         rep->lock_policy_res1 =
2123                                 ptlrpc_status_ntoh(rep->lock_policy_res1);
2124                         if (rep->lock_policy_res1)
2125                                 rc = rep->lock_policy_res1;
2126                 }
2127         }
2128
2129         if ((intent != 0 && rc == ELDLM_LOCK_ABORTED && agl == 0) ||
2130             (rc == 0)) {
2131                 *flags |= LDLM_FL_LVB_READY;
2132                 CDEBUG(D_INODE, "got kms %llu blocks %llu mtime %llu\n",
2133                        lvb->lvb_size, lvb->lvb_blocks, lvb->lvb_mtime);
2134         }
2135
2136         /* Call the update callback. */
2137         rc = (*upcall)(cookie, rc);
2138         return rc;
2139 }
2140
2141 static int osc_enqueue_interpret(const struct lu_env *env,
2142                                  struct ptlrpc_request *req,
2143                                  struct osc_enqueue_args *aa, int rc)
2144 {
2145         struct ldlm_lock *lock;
2146         struct lustre_handle handle;
2147         __u32 mode;
2148         struct ost_lvb *lvb;
2149         __u32 lvb_len;
2150         __u64 *flags = aa->oa_flags;
2151
2152         /* Make a local copy of a lock handle and a mode, because aa->oa_*
2153          * might be freed anytime after lock upcall has been called.
2154          */
2155         lustre_handle_copy(&handle, aa->oa_lockh);
2156         mode = aa->oa_ei->ei_mode;
2157
2158         /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2159          * be valid.
2160          */
2161         lock = ldlm_handle2lock(&handle);
2162
2163         /* Take an additional reference so that a blocking AST that
2164          * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2165          * to arrive after an upcall has been executed by
2166          * osc_enqueue_fini().
2167          */
2168         ldlm_lock_addref(&handle, mode);
2169
2170         /* Let CP AST to grant the lock first. */
2171         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
2172
2173         if (aa->oa_agl && rc == ELDLM_LOCK_ABORTED) {
2174                 lvb = NULL;
2175                 lvb_len = 0;
2176         } else {
2177                 lvb = aa->oa_lvb;
2178                 lvb_len = sizeof(*aa->oa_lvb);
2179         }
2180
2181         /* Complete obtaining the lock procedure. */
2182         rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_ei->ei_type, 1,
2183                                    mode, flags, lvb, lvb_len, &handle, rc);
2184         /* Complete osc stuff. */
2185         rc = osc_enqueue_fini(req, aa->oa_lvb, aa->oa_upcall, aa->oa_cookie,
2186                               flags, aa->oa_agl, rc);
2187
2188         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
2189
2190         /* Release the lock for async request. */
2191         if (lustre_handle_is_used(&handle) && rc == ELDLM_OK)
2192                 /*
2193                  * Releases a reference taken by ldlm_cli_enqueue(), if it is
2194                  * not already released by
2195                  * ldlm_cli_enqueue_fini()->failed_lock_cleanup()
2196                  */
2197                 ldlm_lock_decref(&handle, mode);
2198
2199         LASSERTF(lock, "lockh %p, req %p, aa %p - client evicted?\n",
2200                  aa->oa_lockh, req, aa);
2201         ldlm_lock_decref(&handle, mode);
2202         LDLM_LOCK_PUT(lock);
2203         return rc;
2204 }
2205
2206 struct ptlrpc_request_set *PTLRPCD_SET = (void *)1;
2207
2208 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2209  * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2210  * other synchronous requests, however keeping some locks and trying to obtain
2211  * others may take a considerable amount of time in a case of ost failure; and
2212  * when other sync requests do not get released lock from a client, the client
2213  * is excluded from the cluster -- such scenarious make the life difficult, so
2214  * release locks just after they are obtained.
2215  */
2216 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2217                      __u64 *flags, ldlm_policy_data_t *policy,
2218                      struct ost_lvb *lvb, int kms_valid,
2219                      obd_enqueue_update_f upcall, void *cookie,
2220                      struct ldlm_enqueue_info *einfo,
2221                      struct lustre_handle *lockh,
2222                      struct ptlrpc_request_set *rqset, int async, int agl)
2223 {
2224         struct obd_device *obd = exp->exp_obd;
2225         struct ptlrpc_request *req = NULL;
2226         int intent = *flags & LDLM_FL_HAS_INTENT;
2227         __u64 match_lvb = (agl != 0 ? 0 : LDLM_FL_LVB_READY);
2228         enum ldlm_mode mode;
2229         int rc;
2230
2231         /* Filesystem lock extents are extended to page boundaries so that
2232          * dealing with the page cache is a little smoother.
2233          */
2234         policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
2235         policy->l_extent.end |= ~CFS_PAGE_MASK;
2236
2237         /*
2238          * kms is not valid when either object is completely fresh (so that no
2239          * locks are cached), or object was evicted. In the latter case cached
2240          * lock cannot be used, because it would prime inode state with
2241          * potentially stale LVB.
2242          */
2243         if (!kms_valid)
2244                 goto no_match;
2245
2246         /* Next, search for already existing extent locks that will cover us */
2247         /* If we're trying to read, we also search for an existing PW lock.  The
2248          * VFS and page cache already protect us locally, so lots of readers/
2249          * writers can share a single PW lock.
2250          *
2251          * There are problems with conversion deadlocks, so instead of
2252          * converting a read lock to a write lock, we'll just enqueue a new
2253          * one.
2254          *
2255          * At some point we should cancel the read lock instead of making them
2256          * send us a blocking callback, but there are problems with canceling
2257          * locks out from other users right now, too.
2258          */
2259         mode = einfo->ei_mode;
2260         if (einfo->ei_mode == LCK_PR)
2261                 mode |= LCK_PW;
2262         mode = ldlm_lock_match(obd->obd_namespace, *flags | match_lvb, res_id,
2263                                einfo->ei_type, policy, mode, lockh, 0);
2264         if (mode) {
2265                 struct ldlm_lock *matched = ldlm_handle2lock(lockh);
2266
2267                 if ((agl != 0) && !(matched->l_flags & LDLM_FL_LVB_READY)) {
2268                         /* For AGL, if enqueue RPC is sent but the lock is not
2269                          * granted, then skip to process this strpe.
2270                          * Return -ECANCELED to tell the caller.
2271                          */
2272                         ldlm_lock_decref(lockh, mode);
2273                         LDLM_LOCK_PUT(matched);
2274                         return -ECANCELED;
2275                 }
2276
2277                 if (osc_set_lock_data_with_check(matched, einfo)) {
2278                         *flags |= LDLM_FL_LVB_READY;
2279                         /* addref the lock only if not async requests and PW
2280                          * lock is matched whereas we asked for PR.
2281                          */
2282                         if (!rqset && einfo->ei_mode != mode)
2283                                 ldlm_lock_addref(lockh, LCK_PR);
2284                         if (intent) {
2285                                 /* I would like to be able to ASSERT here that
2286                                  * rss <= kms, but I can't, for reasons which
2287                                  * are explained in lov_enqueue()
2288                                  */
2289                         }
2290
2291                         /* We already have a lock, and it's referenced.
2292                          *
2293                          * At this point, the cl_lock::cll_state is CLS_QUEUING,
2294                          * AGL upcall may change it to CLS_HELD directly.
2295                          */
2296                         (*upcall)(cookie, ELDLM_OK);
2297
2298                         if (einfo->ei_mode != mode)
2299                                 ldlm_lock_decref(lockh, LCK_PW);
2300                         else if (rqset)
2301                                 /* For async requests, decref the lock. */
2302                                 ldlm_lock_decref(lockh, einfo->ei_mode);
2303                         LDLM_LOCK_PUT(matched);
2304                         return ELDLM_OK;
2305                 }
2306
2307                 ldlm_lock_decref(lockh, mode);
2308                 LDLM_LOCK_PUT(matched);
2309         }
2310
2311  no_match:
2312         if (intent) {
2313                 LIST_HEAD(cancels);
2314
2315                 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2316                                            &RQF_LDLM_ENQUEUE_LVB);
2317                 if (!req)
2318                         return -ENOMEM;
2319
2320                 rc = ldlm_prep_enqueue_req(exp, req, &cancels, 0);
2321                 if (rc) {
2322                         ptlrpc_request_free(req);
2323                         return rc;
2324                 }
2325
2326                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
2327                                      sizeof(*lvb));
2328                 ptlrpc_request_set_replen(req);
2329         }
2330
2331         /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2332         *flags &= ~LDLM_FL_BLOCK_GRANTED;
2333
2334         rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2335                               sizeof(*lvb), LVB_T_OST, lockh, async);
2336         if (rqset) {
2337                 if (!rc) {
2338                         struct osc_enqueue_args *aa;
2339
2340                         CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
2341                         aa = ptlrpc_req_async_args(req);
2342                         aa->oa_ei = einfo;
2343                         aa->oa_exp = exp;
2344                         aa->oa_flags  = flags;
2345                         aa->oa_upcall = upcall;
2346                         aa->oa_cookie = cookie;
2347                         aa->oa_lvb    = lvb;
2348                         aa->oa_lockh  = lockh;
2349                         aa->oa_agl    = !!agl;
2350
2351                         req->rq_interpret_reply =
2352                                 (ptlrpc_interpterer_t)osc_enqueue_interpret;
2353                         if (rqset == PTLRPCD_SET)
2354                                 ptlrpcd_add_req(req);
2355                         else
2356                                 ptlrpc_set_add_req(rqset, req);
2357                 } else if (intent) {
2358                         ptlrpc_req_finished(req);
2359                 }
2360                 return rc;
2361         }
2362
2363         rc = osc_enqueue_fini(req, lvb, upcall, cookie, flags, agl, rc);
2364         if (intent)
2365                 ptlrpc_req_finished(req);
2366
2367         return rc;
2368 }
2369
2370 int osc_match_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2371                    __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2372                    __u64 *flags, void *data, struct lustre_handle *lockh,
2373                    int unref)
2374 {
2375         struct obd_device *obd = exp->exp_obd;
2376         __u64 lflags = *flags;
2377         enum ldlm_mode rc;
2378
2379         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2380                 return -EIO;
2381
2382         /* Filesystem lock extents are extended to page boundaries so that
2383          * dealing with the page cache is a little smoother
2384          */
2385         policy->l_extent.start -= policy->l_extent.start & ~CFS_PAGE_MASK;
2386         policy->l_extent.end |= ~CFS_PAGE_MASK;
2387
2388         /* Next, search for already existing extent locks that will cover us */
2389         /* If we're trying to read, we also search for an existing PW lock.  The
2390          * VFS and page cache already protect us locally, so lots of readers/
2391          * writers can share a single PW lock.
2392          */
2393         rc = mode;
2394         if (mode == LCK_PR)
2395                 rc |= LCK_PW;
2396         rc = ldlm_lock_match(obd->obd_namespace, lflags,
2397                              res_id, type, policy, rc, lockh, unref);
2398         if (rc) {
2399                 if (data) {
2400                         if (!osc_set_data_with_check(lockh, data)) {
2401                                 if (!(lflags & LDLM_FL_TEST_LOCK))
2402                                         ldlm_lock_decref(lockh, rc);
2403                                 return 0;
2404                         }
2405                 }
2406                 if (!(lflags & LDLM_FL_TEST_LOCK) && mode != rc) {
2407                         ldlm_lock_addref(lockh, LCK_PR);
2408                         ldlm_lock_decref(lockh, LCK_PW);
2409                 }
2410                 return rc;
2411         }
2412         return rc;
2413 }
2414
2415 int osc_cancel_base(struct lustre_handle *lockh, __u32 mode)
2416 {
2417         if (unlikely(mode == LCK_GROUP))
2418                 ldlm_lock_decref_and_cancel(lockh, mode);
2419         else
2420                 ldlm_lock_decref(lockh, mode);
2421
2422         return 0;
2423 }
2424
2425 static int osc_statfs_interpret(const struct lu_env *env,
2426                                 struct ptlrpc_request *req,
2427                                 struct osc_async_args *aa, int rc)
2428 {
2429         struct obd_statfs *msfs;
2430
2431         if (rc == -EBADR)
2432                 /* The request has in fact never been sent
2433                  * due to issues at a higher level (LOV).
2434                  * Exit immediately since the caller is
2435                  * aware of the problem and takes care
2436                  * of the clean up
2437                  */
2438                 return rc;
2439
2440         if ((rc == -ENOTCONN || rc == -EAGAIN) &&
2441             (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY)) {
2442                 rc = 0;
2443                 goto out;
2444         }
2445
2446         if (rc != 0)
2447                 goto out;
2448
2449         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2450         if (!msfs) {
2451                 rc = -EPROTO;
2452                 goto out;
2453         }
2454
2455         *aa->aa_oi->oi_osfs = *msfs;
2456 out:
2457         rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
2458         return rc;
2459 }
2460
2461 static int osc_statfs_async(struct obd_export *exp,
2462                             struct obd_info *oinfo, __u64 max_age,
2463                             struct ptlrpc_request_set *rqset)
2464 {
2465         struct obd_device *obd = class_exp2obd(exp);
2466         struct ptlrpc_request *req;
2467         struct osc_async_args *aa;
2468         int rc;
2469
2470         /* We could possibly pass max_age in the request (as an absolute
2471          * timestamp or a "seconds.usec ago") so the target can avoid doing
2472          * extra calls into the filesystem if that isn't necessary (e.g.
2473          * during mount that would help a bit).  Having relative timestamps
2474          * is not so great if request processing is slow, while absolute
2475          * timestamps are not ideal because they need time synchronization.
2476          */
2477         req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
2478         if (!req)
2479                 return -ENOMEM;
2480
2481         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2482         if (rc) {
2483                 ptlrpc_request_free(req);
2484                 return rc;
2485         }
2486         ptlrpc_request_set_replen(req);
2487         req->rq_request_portal = OST_CREATE_PORTAL;
2488         ptlrpc_at_set_req_timeout(req);
2489
2490         if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
2491                 /* procfs requests not want stat in wait for avoid deadlock */
2492                 req->rq_no_resend = 1;
2493                 req->rq_no_delay = 1;
2494         }
2495
2496         req->rq_interpret_reply = (ptlrpc_interpterer_t)osc_statfs_interpret;
2497         CLASSERT (sizeof(*aa) <= sizeof(req->rq_async_args));
2498         aa = ptlrpc_req_async_args(req);
2499         aa->aa_oi = oinfo;
2500
2501         ptlrpc_set_add_req(rqset, req);
2502         return 0;
2503 }
2504
2505 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
2506                       struct obd_statfs *osfs, __u64 max_age, __u32 flags)
2507 {
2508         struct obd_device *obd = class_exp2obd(exp);
2509         struct obd_statfs *msfs;
2510         struct ptlrpc_request *req;
2511         struct obd_import *imp = NULL;
2512         int rc;
2513
2514         /* Since the request might also come from lprocfs, so we need
2515          * sync this with client_disconnect_export Bug15684
2516          */
2517         down_read(&obd->u.cli.cl_sem);
2518         if (obd->u.cli.cl_import)
2519                 imp = class_import_get(obd->u.cli.cl_import);
2520         up_read(&obd->u.cli.cl_sem);
2521         if (!imp)
2522                 return -ENODEV;
2523
2524         /* We could possibly pass max_age in the request (as an absolute
2525          * timestamp or a "seconds.usec ago") so the target can avoid doing
2526          * extra calls into the filesystem if that isn't necessary (e.g.
2527          * during mount that would help a bit).  Having relative timestamps
2528          * is not so great if request processing is slow, while absolute
2529          * timestamps are not ideal because they need time synchronization.
2530          */
2531         req = ptlrpc_request_alloc(imp, &RQF_OST_STATFS);
2532
2533         class_import_put(imp);
2534
2535         if (!req)
2536                 return -ENOMEM;
2537
2538         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
2539         if (rc) {
2540                 ptlrpc_request_free(req);
2541                 return rc;
2542         }
2543         ptlrpc_request_set_replen(req);
2544         req->rq_request_portal = OST_CREATE_PORTAL;
2545         ptlrpc_at_set_req_timeout(req);
2546
2547         if (flags & OBD_STATFS_NODELAY) {
2548                 /* procfs requests not want stat in wait for avoid deadlock */
2549                 req->rq_no_resend = 1;
2550                 req->rq_no_delay = 1;
2551         }
2552
2553         rc = ptlrpc_queue_wait(req);
2554         if (rc)
2555                 goto out;
2556
2557         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
2558         if (!msfs) {
2559                 rc = -EPROTO;
2560                 goto out;
2561         }
2562
2563         *osfs = *msfs;
2564
2565  out:
2566         ptlrpc_req_finished(req);
2567         return rc;
2568 }
2569
2570 /* Retrieve object striping information.
2571  *
2572  * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
2573  * the maximum number of OST indices which will fit in the user buffer.
2574  * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
2575  */
2576 static int osc_getstripe(struct lov_stripe_md *lsm,
2577                          struct lov_user_md __user *lump)
2578 {
2579         /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
2580         struct lov_user_md_v3 lum, *lumk;
2581         struct lov_user_ost_data_v1 *lmm_objects;
2582         int rc = 0, lum_size;
2583
2584         if (!lsm)
2585                 return -ENODATA;
2586
2587         /* we only need the header part from user space to get lmm_magic and
2588          * lmm_stripe_count, (the header part is common to v1 and v3)
2589          */
2590         lum_size = sizeof(struct lov_user_md_v1);
2591         if (copy_from_user(&lum, lump, lum_size))
2592                 return -EFAULT;
2593
2594         if ((lum.lmm_magic != LOV_USER_MAGIC_V1) &&
2595             (lum.lmm_magic != LOV_USER_MAGIC_V3))
2596                 return -EINVAL;
2597
2598         /* lov_user_md_vX and lov_mds_md_vX must have the same size */
2599         LASSERT(sizeof(struct lov_user_md_v1) == sizeof(struct lov_mds_md_v1));
2600         LASSERT(sizeof(struct lov_user_md_v3) == sizeof(struct lov_mds_md_v3));
2601         LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lumk->lmm_objects[0]));
2602
2603         /* we can use lov_mds_md_size() to compute lum_size
2604          * because lov_user_md_vX and lov_mds_md_vX have the same size
2605          */
2606         if (lum.lmm_stripe_count > 0) {
2607                 lum_size = lov_mds_md_size(lum.lmm_stripe_count, lum.lmm_magic);
2608                 lumk = kzalloc(lum_size, GFP_NOFS);
2609                 if (!lumk)
2610                         return -ENOMEM;
2611
2612                 if (lum.lmm_magic == LOV_USER_MAGIC_V1)
2613                         lmm_objects =
2614                             &(((struct lov_user_md_v1 *)lumk)->lmm_objects[0]);
2615                 else
2616                         lmm_objects = &(lumk->lmm_objects[0]);
2617                 lmm_objects->l_ost_oi = lsm->lsm_oi;
2618         } else {
2619                 lum_size = lov_mds_md_size(0, lum.lmm_magic);
2620                 lumk = &lum;
2621         }
2622
2623         lumk->lmm_oi = lsm->lsm_oi;
2624         lumk->lmm_stripe_count = 1;
2625
2626         if (copy_to_user(lump, lumk, lum_size))
2627                 rc = -EFAULT;
2628
2629         if (lumk != &lum)
2630                 kfree(lumk);
2631
2632         return rc;
2633 }
2634
2635 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2636                          void *karg, void __user *uarg)
2637 {
2638         struct obd_device *obd = exp->exp_obd;
2639         struct obd_ioctl_data *data = karg;
2640         int err = 0;
2641
2642         if (!try_module_get(THIS_MODULE)) {
2643                 CERROR("Can't get module. Is it alive?");
2644                 return -EINVAL;
2645         }
2646         switch (cmd) {
2647         case OBD_IOC_LOV_GET_CONFIG: {
2648                 char *buf;
2649                 struct lov_desc *desc;
2650                 struct obd_uuid uuid;
2651
2652                 buf = NULL;
2653                 len = 0;
2654                 if (obd_ioctl_getdata(&buf, &len, uarg)) {
2655                         err = -EINVAL;
2656                         goto out;
2657                 }
2658
2659                 data = (struct obd_ioctl_data *)buf;
2660
2661                 if (sizeof(*desc) > data->ioc_inllen1) {
2662                         obd_ioctl_freedata(buf, len);
2663                         err = -EINVAL;
2664                         goto out;
2665                 }
2666
2667                 if (data->ioc_inllen2 < sizeof(uuid)) {
2668                         obd_ioctl_freedata(buf, len);
2669                         err = -EINVAL;
2670                         goto out;
2671                 }
2672
2673                 desc = (struct lov_desc *)data->ioc_inlbuf1;
2674                 desc->ld_tgt_count = 1;
2675                 desc->ld_active_tgt_count = 1;
2676                 desc->ld_default_stripe_count = 1;
2677                 desc->ld_default_stripe_size = 0;
2678                 desc->ld_default_stripe_offset = 0;
2679                 desc->ld_pattern = 0;
2680                 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
2681
2682                 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
2683
2684                 err = copy_to_user(uarg, buf, len);
2685                 if (err)
2686                         err = -EFAULT;
2687                 obd_ioctl_freedata(buf, len);
2688                 goto out;
2689         }
2690         case LL_IOC_LOV_SETSTRIPE:
2691                 err = obd_alloc_memmd(exp, karg);
2692                 if (err > 0)
2693                         err = 0;
2694                 goto out;
2695         case LL_IOC_LOV_GETSTRIPE:
2696                 err = osc_getstripe(karg, uarg);
2697                 goto out;
2698         case OBD_IOC_CLIENT_RECOVER:
2699                 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2700                                             data->ioc_inlbuf1, 0);
2701                 if (err > 0)
2702                         err = 0;
2703                 goto out;
2704         case IOC_OSC_SET_ACTIVE:
2705                 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2706                                                data->ioc_offset);
2707                 goto out;
2708         case OBD_IOC_POLL_QUOTACHECK:
2709                 err = osc_quota_poll_check(exp, karg);
2710                 goto out;
2711         case OBD_IOC_PING_TARGET:
2712                 err = ptlrpc_obd_ping(obd);
2713                 goto out;
2714         default:
2715                 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n",
2716                        cmd, current_comm());
2717                 err = -ENOTTY;
2718                 goto out;
2719         }
2720 out:
2721         module_put(THIS_MODULE);
2722         return err;
2723 }
2724
2725 static int osc_get_info(const struct lu_env *env, struct obd_export *exp,
2726                         u32 keylen, void *key, __u32 *vallen, void *val,
2727                         struct lov_stripe_md *lsm)
2728 {
2729         if (!vallen || !val)
2730                 return -EFAULT;
2731
2732         if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
2733                 __u32 *stripe = val;
2734                 *vallen = sizeof(*stripe);
2735                 *stripe = 0;
2736                 return 0;
2737         } else if (KEY_IS(KEY_LAST_ID)) {
2738                 struct ptlrpc_request *req;
2739                 u64 *reply;
2740                 char *tmp;
2741                 int rc;
2742
2743                 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2744                                            &RQF_OST_GET_INFO_LAST_ID);
2745                 if (!req)
2746                         return -ENOMEM;
2747
2748                 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2749                                      RCL_CLIENT, keylen);
2750                 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
2751                 if (rc) {
2752                         ptlrpc_request_free(req);
2753                         return rc;
2754                 }
2755
2756                 tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2757                 memcpy(tmp, key, keylen);
2758
2759                 req->rq_no_delay = req->rq_no_resend = 1;
2760                 ptlrpc_request_set_replen(req);
2761                 rc = ptlrpc_queue_wait(req);
2762                 if (rc)
2763                         goto out;
2764
2765                 reply = req_capsule_server_get(&req->rq_pill, &RMF_OBD_ID);
2766                 if (!reply) {
2767                         rc = -EPROTO;
2768                         goto out;
2769                 }
2770
2771                 *((u64 *)val) = *reply;
2772 out:
2773                 ptlrpc_req_finished(req);
2774                 return rc;
2775         } else if (KEY_IS(KEY_FIEMAP)) {
2776                 struct ll_fiemap_info_key *fm_key = key;
2777                 struct ldlm_res_id res_id;
2778                 ldlm_policy_data_t policy;
2779                 struct lustre_handle lockh;
2780                 enum ldlm_mode mode = 0;
2781                 struct ptlrpc_request *req;
2782                 struct ll_user_fiemap *reply;
2783                 char *tmp;
2784                 int rc;
2785
2786                 if (!(fm_key->fiemap.fm_flags & FIEMAP_FLAG_SYNC))
2787                         goto skip_locking;
2788
2789                 policy.l_extent.start = fm_key->fiemap.fm_start &
2790                                                 CFS_PAGE_MASK;
2791
2792                 if (OBD_OBJECT_EOF - fm_key->fiemap.fm_length <=
2793                     fm_key->fiemap.fm_start + PAGE_CACHE_SIZE - 1)
2794                         policy.l_extent.end = OBD_OBJECT_EOF;
2795                 else
2796                         policy.l_extent.end = (fm_key->fiemap.fm_start +
2797                                 fm_key->fiemap.fm_length +
2798                                 PAGE_CACHE_SIZE - 1) & CFS_PAGE_MASK;
2799
2800                 ostid_build_res_name(&fm_key->oa.o_oi, &res_id);
2801                 mode = ldlm_lock_match(exp->exp_obd->obd_namespace,
2802                                        LDLM_FL_BLOCK_GRANTED |
2803                                        LDLM_FL_LVB_READY,
2804                                        &res_id, LDLM_EXTENT, &policy,
2805                                        LCK_PR | LCK_PW, &lockh, 0);
2806                 if (mode) { /* lock is cached on client */
2807                         if (mode != LCK_PR) {
2808                                 ldlm_lock_addref(&lockh, LCK_PR);
2809                                 ldlm_lock_decref(&lockh, LCK_PW);
2810                         }
2811                 } else { /* no cached lock, needs acquire lock on server side */
2812                         fm_key->oa.o_valid |= OBD_MD_FLFLAGS;
2813                         fm_key->oa.o_flags |= OBD_FL_SRVLOCK;
2814                 }
2815
2816 skip_locking:
2817                 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2818                                            &RQF_OST_GET_INFO_FIEMAP);
2819                 if (!req) {
2820                         rc = -ENOMEM;
2821                         goto drop_lock;
2822                 }
2823
2824                 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY,
2825                                      RCL_CLIENT, keylen);
2826                 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
2827                                      RCL_CLIENT, *vallen);
2828                 req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL,
2829                                      RCL_SERVER, *vallen);
2830
2831                 rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
2832                 if (rc) {
2833                         ptlrpc_request_free(req);
2834                         goto drop_lock;
2835                 }
2836
2837                 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
2838                 memcpy(tmp, key, keylen);
2839                 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
2840                 memcpy(tmp, val, *vallen);
2841
2842                 ptlrpc_request_set_replen(req);
2843                 rc = ptlrpc_queue_wait(req);
2844                 if (rc)
2845                         goto fini_req;
2846
2847                 reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
2848                 if (!reply) {
2849                         rc = -EPROTO;
2850                         goto fini_req;
2851                 }
2852
2853                 memcpy(val, reply, *vallen);
2854 fini_req:
2855                 ptlrpc_req_finished(req);
2856 drop_lock:
2857                 if (mode)
2858                         ldlm_lock_decref(&lockh, LCK_PR);
2859                 return rc;
2860         }
2861
2862         return -EINVAL;
2863 }
2864
2865 static int osc_set_info_async(const struct lu_env *env, struct obd_export *exp,
2866                               u32 keylen, void *key, u32 vallen,
2867                               void *val, struct ptlrpc_request_set *set)
2868 {
2869         struct ptlrpc_request *req;
2870         struct obd_device *obd = exp->exp_obd;
2871         struct obd_import *imp = class_exp2cliimp(exp);
2872         char *tmp;
2873         int rc;
2874
2875         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_SHUTDOWN, 10);
2876
2877         if (KEY_IS(KEY_CHECKSUM)) {
2878                 if (vallen != sizeof(int))
2879                         return -EINVAL;
2880                 exp->exp_obd->u.cli.cl_checksum = (*(int *)val) ? 1 : 0;
2881                 return 0;
2882         }
2883
2884         if (KEY_IS(KEY_SPTLRPC_CONF)) {
2885                 sptlrpc_conf_client_adapt(obd);
2886                 return 0;
2887         }
2888
2889         if (KEY_IS(KEY_FLUSH_CTX)) {
2890                 sptlrpc_import_flush_my_ctx(imp);
2891                 return 0;
2892         }
2893
2894         if (KEY_IS(KEY_CACHE_SET)) {
2895                 struct client_obd *cli = &obd->u.cli;
2896
2897                 LASSERT(!cli->cl_cache); /* only once */
2898                 cli->cl_cache = val;
2899                 atomic_inc(&cli->cl_cache->ccc_users);
2900                 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
2901
2902                 /* add this osc into entity list */
2903                 LASSERT(list_empty(&cli->cl_lru_osc));
2904                 spin_lock(&cli->cl_cache->ccc_lru_lock);
2905                 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
2906                 spin_unlock(&cli->cl_cache->ccc_lru_lock);
2907
2908                 return 0;
2909         }
2910
2911         if (KEY_IS(KEY_CACHE_LRU_SHRINK)) {
2912                 struct client_obd *cli = &obd->u.cli;
2913                 int nr = atomic_read(&cli->cl_lru_in_list) >> 1;
2914                 int target = *(int *)val;
2915
2916                 nr = osc_lru_shrink(cli, min(nr, target));
2917                 *(int *)val -= nr;
2918                 return 0;
2919         }
2920
2921         if (!set && !KEY_IS(KEY_GRANT_SHRINK))
2922                 return -EINVAL;
2923
2924         /* We pass all other commands directly to OST. Since nobody calls osc
2925          * methods directly and everybody is supposed to go through LOV, we
2926          * assume lov checked invalid values for us.
2927          * The only recognised values so far are evict_by_nid and mds_conn.
2928          * Even if something bad goes through, we'd get a -EINVAL from OST
2929          * anyway.
2930          */
2931
2932         req = ptlrpc_request_alloc(imp, KEY_IS(KEY_GRANT_SHRINK) ?
2933                                                 &RQF_OST_SET_GRANT_INFO :
2934                                                 &RQF_OBD_SET_INFO);
2935         if (!req)
2936                 return -ENOMEM;
2937
2938         req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2939                              RCL_CLIENT, keylen);
2940         if (!KEY_IS(KEY_GRANT_SHRINK))
2941                 req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
2942                                      RCL_CLIENT, vallen);
2943         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SET_INFO);
2944         if (rc) {
2945                 ptlrpc_request_free(req);
2946                 return rc;
2947         }
2948
2949         tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2950         memcpy(tmp, key, keylen);
2951         tmp = req_capsule_client_get(&req->rq_pill, KEY_IS(KEY_GRANT_SHRINK) ?
2952                                                         &RMF_OST_BODY :
2953                                                         &RMF_SETINFO_VAL);
2954         memcpy(tmp, val, vallen);
2955
2956         if (KEY_IS(KEY_GRANT_SHRINK)) {
2957                 struct osc_brw_async_args *aa;
2958                 struct obdo *oa;
2959
2960                 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2961                 aa = ptlrpc_req_async_args(req);
2962                 oa = kmem_cache_alloc(obdo_cachep, GFP_NOFS | __GFP_ZERO);
2963                 if (!oa) {
2964                         ptlrpc_req_finished(req);
2965                         return -ENOMEM;
2966                 }
2967                 *oa = ((struct ost_body *)val)->oa;
2968                 aa->aa_oa = oa;
2969                 req->rq_interpret_reply = osc_shrink_grant_interpret;
2970         }
2971
2972         ptlrpc_request_set_replen(req);
2973         if (!KEY_IS(KEY_GRANT_SHRINK)) {
2974                 LASSERT(set);
2975                 ptlrpc_set_add_req(set, req);
2976                 ptlrpc_check_set(NULL, set);
2977         } else {
2978                 ptlrpcd_add_req(req);
2979         }
2980
2981         return 0;
2982 }
2983
2984 static int osc_reconnect(const struct lu_env *env,
2985                          struct obd_export *exp, struct obd_device *obd,
2986                          struct obd_uuid *cluuid,
2987                          struct obd_connect_data *data,
2988                          void *localdata)
2989 {
2990         struct client_obd *cli = &obd->u.cli;
2991
2992         if (data && (data->ocd_connect_flags & OBD_CONNECT_GRANT)) {
2993                 long lost_grant;
2994
2995                 client_obd_list_lock(&cli->cl_loi_list_lock);
2996                 data->ocd_grant = (cli->cl_avail_grant + cli->cl_dirty) ?:
2997                                 2 * cli_brw_size(obd);
2998                 lost_grant = cli->cl_lost_grant;
2999                 cli->cl_lost_grant = 0;
3000                 client_obd_list_unlock(&cli->cl_loi_list_lock);
3001
3002                 CDEBUG(D_RPCTRACE, "ocd_connect_flags: %#llx ocd_version: %d ocd_grant: %d, lost: %ld.\n",
3003                        data->ocd_connect_flags,
3004                        data->ocd_version, data->ocd_grant, lost_grant);
3005         }
3006
3007         return 0;
3008 }
3009
3010 static int osc_disconnect(struct obd_export *exp)
3011 {
3012         struct obd_device *obd = class_exp2obd(exp);
3013         int rc;
3014
3015         rc = client_disconnect_export(exp);
3016         /**
3017          * Initially we put del_shrink_grant before disconnect_export, but it
3018          * causes the following problem if setup (connect) and cleanup
3019          * (disconnect) are tangled together.
3020          *      connect p1                   disconnect p2
3021          *   ptlrpc_connect_import
3022          *     ...............         class_manual_cleanup
3023          *                                   osc_disconnect
3024          *                                   del_shrink_grant
3025          *   ptlrpc_connect_interrupt
3026          *     init_grant_shrink
3027          *   add this client to shrink list
3028          *                                    cleanup_osc
3029          * Bang! pinger trigger the shrink.
3030          * So the osc should be disconnected from the shrink list, after we
3031          * are sure the import has been destroyed. BUG18662
3032          */
3033         if (!obd->u.cli.cl_import)
3034                 osc_del_shrink_grant(&obd->u.cli);
3035         return rc;
3036 }
3037
3038 static int osc_import_event(struct obd_device *obd,
3039                             struct obd_import *imp,
3040                             enum obd_import_event event)
3041 {
3042         struct client_obd *cli;
3043         int rc = 0;
3044
3045         LASSERT(imp->imp_obd == obd);
3046
3047         switch (event) {
3048         case IMP_EVENT_DISCON: {
3049                 cli = &obd->u.cli;
3050                 client_obd_list_lock(&cli->cl_loi_list_lock);
3051                 cli->cl_avail_grant = 0;
3052                 cli->cl_lost_grant = 0;
3053                 client_obd_list_unlock(&cli->cl_loi_list_lock);
3054                 break;
3055         }
3056         case IMP_EVENT_INACTIVE: {
3057                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
3058                 break;
3059         }
3060         case IMP_EVENT_INVALIDATE: {
3061                 struct ldlm_namespace *ns = obd->obd_namespace;
3062                 struct lu_env *env;
3063                 int refcheck;
3064
3065                 env = cl_env_get(&refcheck);
3066                 if (!IS_ERR(env)) {
3067                         /* Reset grants */
3068                         cli = &obd->u.cli;
3069                         /* all pages go to failing rpcs due to the invalid
3070                          * import
3071                          */
3072                         osc_io_unplug(env, cli, NULL);
3073
3074                         ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3075                         cl_env_put(env, &refcheck);
3076                 } else
3077                         rc = PTR_ERR(env);
3078                 break;
3079         }
3080         case IMP_EVENT_ACTIVE: {
3081                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
3082                 break;
3083         }
3084         case IMP_EVENT_OCD: {
3085                 struct obd_connect_data *ocd = &imp->imp_connect_data;
3086
3087                 if (ocd->ocd_connect_flags & OBD_CONNECT_GRANT)
3088                         osc_init_grant(&obd->u.cli, ocd);
3089
3090                 /* See bug 7198 */
3091                 if (ocd->ocd_connect_flags & OBD_CONNECT_REQPORTAL)
3092                         imp->imp_client->cli_request_portal = OST_REQUEST_PORTAL;
3093
3094                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
3095                 break;
3096         }
3097         case IMP_EVENT_DEACTIVATE: {
3098                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DEACTIVATE, NULL);
3099                 break;
3100         }
3101         case IMP_EVENT_ACTIVATE: {
3102                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVATE, NULL);
3103                 break;
3104         }
3105         default:
3106                 CERROR("Unknown import event %d\n", event);
3107                 LBUG();
3108         }
3109         return rc;
3110 }
3111
3112 /**
3113  * Determine whether the lock can be canceled before replaying the lock
3114  * during recovery, see bug16774 for detailed information.
3115  *
3116  * \retval zero the lock can't be canceled
3117  * \retval other ok to cancel
3118  */
3119 static int osc_cancel_for_recovery(struct ldlm_lock *lock)
3120 {
3121         check_res_locked(lock->l_resource);
3122
3123         /*
3124          * Cancel all unused extent lock in granted mode LCK_PR or LCK_CR.
3125          *
3126          * XXX as a future improvement, we can also cancel unused write lock
3127          * if it doesn't have dirty data and active mmaps.
3128          */
3129         if (lock->l_resource->lr_type == LDLM_EXTENT &&
3130             (lock->l_granted_mode == LCK_PR ||
3131              lock->l_granted_mode == LCK_CR) &&
3132             (osc_dlm_lock_pageref(lock) == 0))
3133                 return 1;
3134
3135         return 0;
3136 }
3137
3138 static int brw_queue_work(const struct lu_env *env, void *data)
3139 {
3140         struct client_obd *cli = data;
3141
3142         CDEBUG(D_CACHE, "Run writeback work for client obd %p.\n", cli);
3143
3144         osc_io_unplug(env, cli, NULL);
3145         return 0;
3146 }
3147
3148 int osc_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
3149 {
3150         struct lprocfs_static_vars lvars = { NULL };
3151         struct client_obd *cli = &obd->u.cli;
3152         void *handler;
3153         int rc;
3154         int adding;
3155         int added;
3156         int req_count;
3157
3158         rc = ptlrpcd_addref();
3159         if (rc)
3160                 return rc;
3161
3162         rc = client_obd_setup(obd, lcfg);
3163         if (rc)
3164                 goto out_ptlrpcd;
3165
3166         handler = ptlrpcd_alloc_work(cli->cl_import, brw_queue_work, cli);
3167         if (IS_ERR(handler)) {
3168                 rc = PTR_ERR(handler);
3169                 goto out_client_setup;
3170         }
3171         cli->cl_writeback_work = handler;
3172
3173         rc = osc_quota_setup(obd);
3174         if (rc)
3175                 goto out_ptlrpcd_work;
3176
3177         cli->cl_grant_shrink_interval = GRANT_SHRINK_INTERVAL;
3178         lprocfs_osc_init_vars(&lvars);
3179         if (lprocfs_obd_setup(obd, lvars.obd_vars, lvars.sysfs_vars) == 0) {
3180                 lproc_osc_attach_seqstat(obd);
3181                 sptlrpc_lprocfs_cliobd_attach(obd);
3182                 ptlrpc_lprocfs_register_obd(obd);
3183         }
3184
3185         /*
3186          * We try to control the total number of requests with a upper limit
3187          * osc_reqpool_maxreqcount. There might be some race which will cause
3188          * over-limit allocation, but it is fine.
3189          */
3190         req_count = atomic_read(&osc_pool_req_count);
3191         if (req_count < osc_reqpool_maxreqcount) {
3192                 adding = cli->cl_max_rpcs_in_flight + 2;
3193                 if (req_count + adding > osc_reqpool_maxreqcount)
3194                         adding = osc_reqpool_maxreqcount - req_count;
3195
3196                 added = ptlrpc_add_rqs_to_pool(osc_rq_pool, adding);
3197                 atomic_add(added, &osc_pool_req_count);
3198         }
3199
3200         INIT_LIST_HEAD(&cli->cl_grant_shrink_list);
3201         ns_register_cancel(obd->obd_namespace, osc_cancel_for_recovery);
3202         return rc;
3203
3204 out_ptlrpcd_work:
3205         ptlrpcd_destroy_work(handler);
3206 out_client_setup:
3207         client_obd_cleanup(obd);
3208 out_ptlrpcd:
3209         ptlrpcd_decref();
3210         return rc;
3211 }
3212
3213 static int osc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
3214 {
3215         switch (stage) {
3216         case OBD_CLEANUP_EARLY: {
3217                 struct obd_import *imp;
3218
3219                 imp = obd->u.cli.cl_import;
3220                 CDEBUG(D_HA, "Deactivating import %s\n", obd->obd_name);
3221                 /* ptlrpc_abort_inflight to stop an mds_lov_synchronize */
3222                 ptlrpc_deactivate_import(imp);
3223                 spin_lock(&imp->imp_lock);
3224                 imp->imp_pingable = 0;
3225                 spin_unlock(&imp->imp_lock);
3226                 break;
3227         }
3228         case OBD_CLEANUP_EXPORTS: {
3229                 struct client_obd *cli = &obd->u.cli;
3230                 /* LU-464
3231                  * for echo client, export may be on zombie list, wait for
3232                  * zombie thread to cull it, because cli.cl_import will be
3233                  * cleared in client_disconnect_export():
3234                  *   class_export_destroy() -> obd_cleanup() ->
3235                  *   echo_device_free() -> echo_client_cleanup() ->
3236                  *   obd_disconnect() -> osc_disconnect() ->
3237                  *   client_disconnect_export()
3238                  */
3239                 obd_zombie_barrier();
3240                 if (cli->cl_writeback_work) {
3241                         ptlrpcd_destroy_work(cli->cl_writeback_work);
3242                         cli->cl_writeback_work = NULL;
3243                 }
3244                 obd_cleanup_client_import(obd);
3245                 ptlrpc_lprocfs_unregister_obd(obd);
3246                 lprocfs_obd_cleanup(obd);
3247                 break;
3248                 }
3249         }
3250         return 0;
3251 }
3252
3253 static int osc_cleanup(struct obd_device *obd)
3254 {
3255         struct client_obd *cli = &obd->u.cli;
3256         int rc;
3257
3258         /* lru cleanup */
3259         if (cli->cl_cache) {
3260                 LASSERT(atomic_read(&cli->cl_cache->ccc_users) > 0);
3261                 spin_lock(&cli->cl_cache->ccc_lru_lock);
3262                 list_del_init(&cli->cl_lru_osc);
3263                 spin_unlock(&cli->cl_cache->ccc_lru_lock);
3264                 cli->cl_lru_left = NULL;
3265                 atomic_dec(&cli->cl_cache->ccc_users);
3266                 cli->cl_cache = NULL;
3267         }
3268
3269         /* free memory of osc quota cache */
3270         osc_quota_cleanup(obd);
3271
3272         rc = client_obd_cleanup(obd);
3273
3274         ptlrpcd_decref();
3275         return rc;
3276 }
3277
3278 int osc_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg)
3279 {
3280         struct lprocfs_static_vars lvars = { NULL };
3281         int rc = 0;
3282
3283         lprocfs_osc_init_vars(&lvars);
3284
3285         switch (lcfg->lcfg_command) {
3286         default:
3287                 rc = class_process_proc_param(PARAM_OSC, lvars.obd_vars,
3288                                               lcfg, obd);
3289                 if (rc > 0)
3290                         rc = 0;
3291                 break;
3292         }
3293
3294         return rc;
3295 }
3296
3297 static int osc_process_config(struct obd_device *obd, u32 len, void *buf)
3298 {
3299         return osc_process_config_base(obd, buf);
3300 }
3301
3302 static struct obd_ops osc_obd_ops = {
3303         .owner          = THIS_MODULE,
3304         .setup          = osc_setup,
3305         .precleanup     = osc_precleanup,
3306         .cleanup        = osc_cleanup,
3307         .add_conn       = client_import_add_conn,
3308         .del_conn       = client_import_del_conn,
3309         .connect        = client_connect_import,
3310         .reconnect      = osc_reconnect,
3311         .disconnect     = osc_disconnect,
3312         .statfs         = osc_statfs,
3313         .statfs_async   = osc_statfs_async,
3314         .packmd         = osc_packmd,
3315         .unpackmd       = osc_unpackmd,
3316         .create         = osc_create,
3317         .destroy        = osc_destroy,
3318         .getattr        = osc_getattr,
3319         .getattr_async  = osc_getattr_async,
3320         .setattr        = osc_setattr,
3321         .setattr_async  = osc_setattr_async,
3322         .find_cbdata    = osc_find_cbdata,
3323         .iocontrol      = osc_iocontrol,
3324         .get_info       = osc_get_info,
3325         .set_info_async = osc_set_info_async,
3326         .import_event   = osc_import_event,
3327         .process_config = osc_process_config,
3328         .quotactl       = osc_quotactl,
3329         .quotacheck     = osc_quotacheck,
3330 };
3331
3332 extern struct lu_kmem_descr osc_caches[];
3333 extern spinlock_t osc_ast_guard;
3334 extern struct lock_class_key osc_ast_guard_class;
3335
3336 static int __init osc_init(void)
3337 {
3338         struct lprocfs_static_vars lvars = { NULL };
3339         unsigned int reqpool_size;
3340         unsigned int reqsize;
3341         int rc;
3342
3343         /* print an address of _any_ initialized kernel symbol from this
3344          * module, to allow debugging with gdb that doesn't support data
3345          * symbols from modules.
3346          */
3347         CDEBUG(D_INFO, "Lustre OSC module (%p).\n", &osc_caches);
3348
3349         rc = lu_kmem_init(osc_caches);
3350         if (rc)
3351                 return rc;
3352
3353         lprocfs_osc_init_vars(&lvars);
3354
3355         rc = class_register_type(&osc_obd_ops, NULL,
3356                                  LUSTRE_OSC_NAME, &osc_device_type);
3357         if (rc)
3358                 goto out_kmem;
3359
3360         spin_lock_init(&osc_ast_guard);
3361         lockdep_set_class(&osc_ast_guard, &osc_ast_guard_class);
3362
3363         /* This is obviously too much memory, only prevent overflow here */
3364         if (osc_reqpool_mem_max >= 1 << 12 || osc_reqpool_mem_max == 0) {
3365                 rc = -EINVAL;
3366                 goto out_type;
3367         }
3368
3369         reqpool_size = osc_reqpool_mem_max << 20;
3370
3371         reqsize = 1;
3372         while (reqsize < OST_MAXREQSIZE)
3373                 reqsize = reqsize << 1;
3374
3375         /*
3376          * We don't enlarge the request count in OSC pool according to
3377          * cl_max_rpcs_in_flight. The allocation from the pool will only be
3378          * tried after normal allocation failed. So a small OSC pool won't
3379          * cause much performance degression in most of cases.
3380          */
3381         osc_reqpool_maxreqcount = reqpool_size / reqsize;
3382
3383         atomic_set(&osc_pool_req_count, 0);
3384         osc_rq_pool = ptlrpc_init_rq_pool(0, OST_MAXREQSIZE,
3385                                           ptlrpc_add_rqs_to_pool);
3386
3387         if (osc_rq_pool)
3388                 return 0;
3389
3390         rc = -ENOMEM;
3391
3392 out_type:
3393         class_unregister_type(LUSTRE_OSC_NAME);
3394 out_kmem:
3395         lu_kmem_fini(osc_caches);
3396         return rc;
3397 }
3398
3399 static void /*__exit*/ osc_exit(void)
3400 {
3401         class_unregister_type(LUSTRE_OSC_NAME);
3402         lu_kmem_fini(osc_caches);
3403         ptlrpc_free_rq_pool(osc_rq_pool);
3404 }
3405
3406 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3407 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3408 MODULE_LICENSE("GPL");
3409 MODULE_VERSION(LUSTRE_VERSION_STRING);
3410
3411 module_init(osc_init);
3412 module_exit(osc_exit);