lustre/mdc/mdc_request.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2016, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 #define DEBUG_SUBSYSTEM S_MDC
  34
  35 #include <linux/init.h>
  36 #include <linux/kthread.h>
  37 #include <linux/module.h>
  38 #include <linux/pagemap.h>
  39 #include <linux/user_namespace.h>
  40 #include <linux/utsname.h>
  41 #ifdef HAVE_UIDGID_HEADER
  42 # include <linux/uidgid.h>
  43 #endif
  44
  45 #include <lustre/lustre_errno.h>
  46
  47 #include <cl_object.h>
  48 #include <llog_swab.h>
  49 #include <lprocfs_status.h>
  50 #include <lustre_acl.h>
  51 #include <lustre_fid.h>
  52 #include <uapi/linux/lustre_ioctl.h>
  53 #include <lustre_kernelcomm.h>
  54 #include <lustre_lmv.h>
  55 #include <lustre_log.h>
  56 #include <lustre_param.h>
  57 #include <lustre_swab.h>
  58 #include <obd_class.h>
  59
  60 #include "mdc_internal.h"
  61
  62 #define REQUEST_MINOR 244
  63
  64 static int mdc_cleanup(struct obd_device *obd);
  65
  66 static inline int mdc_queue_wait(struct ptlrpc_request *req)
  67 {
  68         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
  69         int rc;
  70
  71         /* obd_get_request_slot() ensures that this client has no more
  72          * than cl_max_rpcs_in_flight RPCs simultaneously inf light
  73          * against an MDT. */
  74         rc = obd_get_request_slot(cli);
  75         if (rc != 0)
  76                 return rc;
  77
  78         rc = ptlrpc_queue_wait(req);
  79         obd_put_request_slot(cli);
  80
  81         return rc;
  82 }
  83
  84 /*
  85  * Send MDS_GET_ROOT RPC to fetch root FID.
  86  *
  87  * If \a fileset is not NULL it should contain a subdirectory off
  88  * the ROOT/ directory to be mounted on the client. Return the FID
  89  * of the subdirectory to the client to mount onto its mountpoint.
  90  *
  91  * \param[in]   imp     MDC import
  92  * \param[in]   fileset fileset name, which could be NULL
  93  * \param[out]  rootfid root FID of this mountpoint
  94  * \param[out]  pc      root capa will be unpacked and saved in this pointer
  95  *
  96  * \retval      0 on success, negative errno on failure
  97  */
  98 static int mdc_get_root(struct obd_export *exp, const char *fileset,
  99                          struct lu_fid *rootfid)
 100 {
 101         struct ptlrpc_request   *req;
 102         struct mdt_body         *body;
 103         int                      rc;
 104
 105         ENTRY;
 106
 107         if (fileset && !(exp_connect_flags(exp) & OBD_CONNECT_SUBTREE))
 108                 RETURN(-ENOTSUPP);
 109
 110         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 111                                 &RQF_MDS_GET_ROOT);
 112         if (req == NULL)
 113                 RETURN(-ENOMEM);
 114
 115         if (fileset != NULL)
 116                 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 117                                      strlen(fileset) + 1);
 118         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_ROOT);
 119         if (rc) {
 120                 ptlrpc_request_free(req);
 121                 RETURN(rc);
 122         }
 123         mdc_pack_body(req, NULL, 0, 0, -1, 0);
 124         if (fileset != NULL) {
 125                 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
 126
 127                 memcpy(name, fileset, strlen(fileset));
 128         }
 129         lustre_msg_add_flags(req->rq_reqmsg, LUSTRE_IMP_FULL);
 130         req->rq_send_state = LUSTRE_IMP_FULL;
 131
 132         ptlrpc_request_set_replen(req);
 133
 134         rc = ptlrpc_queue_wait(req);
 135         if (rc)
 136                 GOTO(out, rc);
 137
 138         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 139         if (body == NULL)
 140                 GOTO(out, rc = -EPROTO);
 141
 142         *rootfid = body->mbo_fid1;
 143         CDEBUG(D_NET, "root fid="DFID", last_committed=%llu\n",
 144                PFID(rootfid), lustre_msg_get_last_committed(req->rq_repmsg));
 145         EXIT;
 146 out:
 147         ptlrpc_req_finished(req);
 148
 149         return rc;
 150 }
 151
 152 /*
 153  * This function now is known to always saying that it will receive 4 buffers
 154  * from server. Even for cases when acl_size and md_size is zero, RPC header
 155  * will contain 4 fields and RPC itself will contain zero size fields. This is
 156  * because mdt_getattr*() _always_ returns 4 fields, but if acl is not needed
 157  * and thus zero, it shrinks it, making zero size. The same story about
 158  * md_size. And this is course of problem when client waits for smaller number
 159  * of fields. This issue will be fixed later when client gets aware of RPC
 160  * layouts.  --umka
 161  */
 162 static int mdc_getattr_common(struct obd_export *exp,
 163                               struct ptlrpc_request *req)
 164 {
 165         struct req_capsule *pill = &req->rq_pill;
 166         struct mdt_body    *body;
 167         void               *eadata;
 168         int                 rc;
 169         ENTRY;
 170
 171         /* Request message already built. */
 172         rc = ptlrpc_queue_wait(req);
 173         if (rc != 0)
 174                 RETURN(rc);
 175
 176         /* sanity check for the reply */
 177         body = req_capsule_server_get(pill, &RMF_MDT_BODY);
 178         if (body == NULL)
 179                 RETURN(-EPROTO);
 180
 181         CDEBUG(D_NET, "mode: %o\n", body->mbo_mode);
 182
 183         mdc_update_max_ea_from_body(exp, body);
 184         if (body->mbo_eadatasize != 0) {
 185                 eadata = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
 186                                                       body->mbo_eadatasize);
 187                 if (eadata == NULL)
 188                         RETURN(-EPROTO);
 189         }
 190
 191         RETURN(0);
 192 }
 193
 194 static int mdc_getattr(struct obd_export *exp, struct md_op_data *op_data,
 195                        struct ptlrpc_request **request)
 196 {
 197         struct ptlrpc_request *req;
 198         int                    rc;
 199         ENTRY;
 200
 201         /* Single MDS without an LMV case */
 202         if (op_data->op_flags & MF_GET_MDT_IDX) {
 203                 op_data->op_mds = 0;
 204                 RETURN(0);
 205         }
 206         *request = NULL;
 207         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_GETATTR);
 208         if (req == NULL)
 209                 RETURN(-ENOMEM);
 210
 211         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR);
 212         if (rc) {
 213                 ptlrpc_request_free(req);
 214                 RETURN(rc);
 215         }
 216
 217         mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
 218                       op_data->op_mode, -1, 0);
 219
 220         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 221                              op_data->op_mode);
 222         ptlrpc_request_set_replen(req);
 223
 224         rc = mdc_getattr_common(exp, req);
 225         if (rc)
 226                 ptlrpc_req_finished(req);
 227         else
 228                 *request = req;
 229         RETURN(rc);
 230 }
 231
 232 static int mdc_getattr_name(struct obd_export *exp, struct md_op_data *op_data,
 233                             struct ptlrpc_request **request)
 234 {
 235         struct ptlrpc_request *req;
 236         int                    rc;
 237         ENTRY;
 238
 239         *request = NULL;
 240         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 241                                    &RQF_MDS_GETATTR_NAME);
 242         if (req == NULL)
 243                 RETURN(-ENOMEM);
 244
 245         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 246                              op_data->op_namelen + 1);
 247
 248         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR_NAME);
 249         if (rc) {
 250                 ptlrpc_request_free(req);
 251                 RETURN(rc);
 252         }
 253
 254         mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
 255                       op_data->op_mode, op_data->op_suppgids[0], 0);
 256
 257         if (op_data->op_name) {
 258                 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
 259                 LASSERT(strnlen(op_data->op_name, op_data->op_namelen) ==
 260                                 op_data->op_namelen);
 261                 memcpy(name, op_data->op_name, op_data->op_namelen);
 262         }
 263
 264         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 265                              op_data->op_mode);
 266         ptlrpc_request_set_replen(req);
 267
 268         rc = mdc_getattr_common(exp, req);
 269         if (rc)
 270                 ptlrpc_req_finished(req);
 271         else
 272                 *request = req;
 273         RETURN(rc);
 274 }
 275
 276 static int mdc_xattr_common(struct obd_export *exp,const struct req_format *fmt,
 277                             const struct lu_fid *fid, int opcode, u64 valid,
 278                             const char *xattr_name, const char *input,
 279                             int input_size, int output_size, int flags,
 280                             __u32 suppgid, struct ptlrpc_request **request)
 281 {
 282         struct ptlrpc_request *req;
 283         int   xattr_namelen = 0;
 284         char *tmp;
 285         int   rc;
 286         ENTRY;
 287
 288         *request = NULL;
 289         req = ptlrpc_request_alloc(class_exp2cliimp(exp), fmt);
 290         if (req == NULL)
 291                 RETURN(-ENOMEM);
 292
 293         if (xattr_name) {
 294                 xattr_namelen = strlen(xattr_name) + 1;
 295                 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 296                                      xattr_namelen);
 297         }
 298         if (input_size) {
 299                 LASSERT(input);
 300                 req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT,
 301                                      input_size);
 302         }
 303
 304         /* Flush local XATTR locks to get rid of a possible cancel RPC */
 305         if (opcode == MDS_REINT && fid_is_sane(fid) &&
 306             exp->exp_connect_data.ocd_ibits_known & MDS_INODELOCK_XATTR) {
 307                 struct list_head cancels = LIST_HEAD_INIT(cancels);
 308                 int count;
 309
 310                 /* Without that packing would fail */
 311                 if (input_size == 0)
 312                         req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
 313                                              RCL_CLIENT, 0);
 314
 315                 count = mdc_resource_get_unused(exp, fid,
 316                                                 &cancels, LCK_EX,
 317                                                 MDS_INODELOCK_XATTR);
 318
 319                 rc = mdc_prep_elc_req(exp, req, MDS_REINT, &cancels, count);
 320                 if (rc) {
 321                         ptlrpc_request_free(req);
 322                         RETURN(rc);
 323                 }
 324         } else {
 325                 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, opcode);
 326                 if (rc) {
 327                         ptlrpc_request_free(req);
 328                         RETURN(rc);
 329                 }
 330         }
 331
 332         if (opcode == MDS_REINT) {
 333                 struct mdt_rec_setxattr *rec;
 334
 335                 CLASSERT(sizeof(struct mdt_rec_setxattr) ==
 336                          sizeof(struct mdt_rec_reint));
 337                 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
 338                 rec->sx_opcode = REINT_SETXATTR;
 339                 rec->sx_fsuid  = from_kuid(&init_user_ns, current_fsuid());
 340                 rec->sx_fsgid  = from_kgid(&init_user_ns, current_fsgid());
 341                 rec->sx_cap    = cfs_curproc_cap_pack();
 342                 rec->sx_suppgid1 = suppgid;
 343                 rec->sx_suppgid2 = -1;
 344                 rec->sx_fid    = *fid;
 345                 rec->sx_valid  = valid | OBD_MD_FLCTIME;
 346                 rec->sx_time   = ktime_get_real_seconds();
 347                 rec->sx_size   = output_size;
 348                 rec->sx_flags  = flags;
 349         } else {
 350                 mdc_pack_body(req, fid, valid, output_size, suppgid, flags);
 351         }
 352
 353         if (xattr_name) {
 354                 tmp = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
 355                 memcpy(tmp, xattr_name, xattr_namelen);
 356         }
 357         if (input_size) {
 358                 tmp = req_capsule_client_get(&req->rq_pill, &RMF_EADATA);
 359                 memcpy(tmp, input, input_size);
 360         }
 361
 362         if (req_capsule_has_field(&req->rq_pill, &RMF_EADATA, RCL_SERVER))
 363                 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
 364                                      RCL_SERVER, output_size);
 365         ptlrpc_request_set_replen(req);
 366
 367         /* make rpc */
 368         if (opcode == MDS_REINT)
 369                 mdc_get_mod_rpc_slot(req, NULL);
 370
 371         rc = ptlrpc_queue_wait(req);
 372
 373         if (opcode == MDS_REINT)
 374                 mdc_put_mod_rpc_slot(req, NULL);
 375
 376         if (rc)
 377                 ptlrpc_req_finished(req);
 378         else
 379                 *request = req;
 380         RETURN(rc);
 381 }
 382
 383 static int mdc_setxattr(struct obd_export *exp, const struct lu_fid *fid,
 384                         u64 valid, const char *xattr_name,
 385                         const char *input, int input_size, int output_size,
 386                         int flags, __u32 suppgid,
 387                         struct ptlrpc_request **request)
 388 {
 389         return mdc_xattr_common(exp, &RQF_MDS_REINT_SETXATTR,
 390                                 fid, MDS_REINT, valid, xattr_name,
 391                                 input, input_size, output_size, flags,
 392                                 suppgid, request);
 393 }
 394
 395 static int mdc_getxattr(struct obd_export *exp, const struct lu_fid *fid,
 396                         u64 valid, const char *xattr_name,
 397                         const char *input, int input_size, int output_size,
 398                         int flags, struct ptlrpc_request **request)
 399 {
 400         return mdc_xattr_common(exp, &RQF_MDS_GETXATTR,
 401                                 fid, MDS_GETXATTR, valid, xattr_name,
 402                                 input, input_size, output_size, flags,
 403                                 -1, request);
 404 }
 405
 406 #ifdef CONFIG_FS_POSIX_ACL
 407 static int mdc_unpack_acl(struct ptlrpc_request *req, struct lustre_md *md)
 408 {
 409         struct req_capsule     *pill = &req->rq_pill;
 410         struct mdt_body        *body = md->body;
 411         struct posix_acl       *acl;
 412         void                   *buf;
 413         int                     rc;
 414         ENTRY;
 415
 416         if (!body->mbo_aclsize)
 417                 RETURN(0);
 418
 419         buf = req_capsule_server_sized_get(pill, &RMF_ACL, body->mbo_aclsize);
 420
 421         if (!buf)
 422                 RETURN(-EPROTO);
 423
 424         acl = posix_acl_from_xattr(&init_user_ns, buf, body->mbo_aclsize);
 425         if (acl == NULL)
 426                 RETURN(0);
 427         if (IS_ERR(acl)) {
 428                 rc = PTR_ERR(acl);
 429                 CERROR("convert xattr to acl: %d\n", rc);
 430                 RETURN(rc);
 431         }
 432
 433         rc = posix_acl_valid(&init_user_ns, acl);
 434         if (rc) {
 435                 CERROR("validate acl: %d\n", rc);
 436                 posix_acl_release(acl);
 437                 RETURN(rc);
 438         }
 439
 440         md->posix_acl = acl;
 441         RETURN(0);
 442 }
 443 #else
 444 #define mdc_unpack_acl(req, md) 0
 445 #endif
 446
 447 int mdc_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
 448                       struct obd_export *dt_exp, struct obd_export *md_exp,
 449                       struct lustre_md *md)
 450 {
 451         struct req_capsule *pill = &req->rq_pill;
 452         int rc;
 453         ENTRY;
 454
 455         LASSERT(md);
 456         memset(md, 0, sizeof(*md));
 457
 458         md->body = req_capsule_server_get(pill, &RMF_MDT_BODY);
 459         LASSERT(md->body != NULL);
 460
 461         if (md->body->mbo_valid & OBD_MD_FLEASIZE) {
 462                 if (!S_ISREG(md->body->mbo_mode)) {
 463                         CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, should be a "
 464                                "regular file, but is not\n");
 465                         GOTO(out, rc = -EPROTO);
 466                 }
 467
 468                 if (md->body->mbo_eadatasize == 0) {
 469                         CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, "
 470                                "but eadatasize 0\n");
 471                         GOTO(out, rc = -EPROTO);
 472                 }
 473
 474                 md->layout.lb_len = md->body->mbo_eadatasize;
 475                 md->layout.lb_buf = req_capsule_server_sized_get(pill,
 476                                                         &RMF_MDT_MD,
 477                                                         md->layout.lb_len);
 478                 if (md->layout.lb_buf == NULL)
 479                         GOTO(out, rc = -EPROTO);
 480         } else if (md->body->mbo_valid & OBD_MD_FLDIREA) {
 481                 const union lmv_mds_md *lmv;
 482                 size_t lmv_size;
 483
 484                 if (!S_ISDIR(md->body->mbo_mode)) {
 485                         CDEBUG(D_INFO, "OBD_MD_FLDIREA set, should be a "
 486                                "directory, but is not\n");
 487                         GOTO(out, rc = -EPROTO);
 488                 }
 489
 490                 lmv_size = md->body->mbo_eadatasize;
 491                 if (lmv_size == 0) {
 492                         CDEBUG(D_INFO, "OBD_MD_FLDIREA is set, "
 493                                "but eadatasize 0\n");
 494                         RETURN(-EPROTO);
 495                 }
 496
 497                 if (md->body->mbo_valid & OBD_MD_MEA) {
 498                         lmv = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
 499                                                            lmv_size);
 500                         if (lmv == NULL)
 501                                 GOTO(out, rc = -EPROTO);
 502
 503                         rc = md_unpackmd(md_exp, &md->lmv, lmv, lmv_size);
 504                         if (rc < 0)
 505                                 GOTO(out, rc);
 506
 507                         if (rc < (typeof(rc))sizeof(*md->lmv)) {
 508                                 CDEBUG(D_INFO, "size too small:  "
 509                                        "rc < sizeof(*md->lmv) (%d < %d)\n",
 510                                         rc, (int)sizeof(*md->lmv));
 511                                 GOTO(out, rc = -EPROTO);
 512                         }
 513                 }
 514         }
 515         rc = 0;
 516
 517         if (md->body->mbo_valid & OBD_MD_FLACL) {
 518                 /* for ACL, it's possible that FLACL is set but aclsize is zero.
 519                  * only when aclsize != 0 there's an actual segment for ACL
 520                  * in reply buffer.
 521                  */
 522                 if (md->body->mbo_aclsize) {
 523                         rc = mdc_unpack_acl(req, md);
 524                         if (rc)
 525                                 GOTO(out, rc);
 526 #ifdef CONFIG_FS_POSIX_ACL
 527                 } else {
 528                         md->posix_acl = NULL;
 529 #endif
 530                 }
 531         }
 532
 533         EXIT;
 534 out:
 535         if (rc) {
 536 #ifdef CONFIG_FS_POSIX_ACL
 537                 posix_acl_release(md->posix_acl);
 538 #endif
 539         }
 540         return rc;
 541 }
 542
 543 int mdc_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
 544 {
 545         ENTRY;
 546         RETURN(0);
 547 }
 548
 549 void mdc_replay_open(struct ptlrpc_request *req)
 550 {
 551         struct md_open_data *mod = req->rq_cb_data;
 552         struct ptlrpc_request *close_req;
 553         struct obd_client_handle *och;
 554         struct lustre_handle old;
 555         struct mdt_body *body;
 556         ENTRY;
 557
 558         if (mod == NULL) {
 559                 DEBUG_REQ(D_ERROR, req,
 560                           "Can't properly replay without open data.");
 561                 EXIT;
 562                 return;
 563         }
 564
 565         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 566         LASSERT(body != NULL);
 567
 568         och = mod->mod_och;
 569         if (och != NULL) {
 570                 struct lustre_handle *file_fh;
 571
 572                 LASSERT(och->och_magic == OBD_CLIENT_HANDLE_MAGIC);
 573
 574                 file_fh = &och->och_fh;
 575                 CDEBUG(D_HA, "updating handle from %#llx to %#llx\n",
 576                        file_fh->cookie, body->mbo_handle.cookie);
 577                 old = *file_fh;
 578                 *file_fh = body->mbo_handle;
 579         }
 580         close_req = mod->mod_close_req;
 581         if (close_req != NULL) {
 582                 __u32 opc = lustre_msg_get_opc(close_req->rq_reqmsg);
 583                 struct mdt_ioepoch *epoch;
 584
 585                 LASSERT(opc == MDS_CLOSE);
 586                 epoch = req_capsule_client_get(&close_req->rq_pill,
 587                                                &RMF_MDT_EPOCH);
 588                 LASSERT(epoch);
 589
 590                 if (och != NULL)
 591                         LASSERT(!memcmp(&old, &epoch->mio_handle, sizeof(old)));
 592
 593                 DEBUG_REQ(D_HA, close_req, "updating close body with new fh");
 594                 epoch->mio_handle = body->mbo_handle;
 595         }
 596         EXIT;
 597 }
 598
 599 void mdc_commit_open(struct ptlrpc_request *req)
 600 {
 601         struct md_open_data *mod = req->rq_cb_data;
 602         if (mod == NULL)
 603                 return;
 604
 605         /**
 606          * No need to touch md_open_data::mod_och, it holds a reference on
 607          * \var mod and will zero references to each other, \var mod will be
 608          * freed after that when md_open_data::mod_och will put the reference.
 609          */
 610
 611         /**
 612          * Do not let open request to disappear as it still may be needed
 613          * for close rpc to happen (it may happen on evict only, otherwise
 614          * ptlrpc_request::rq_replay does not let mdc_commit_open() to be
 615          * called), just mark this rpc as committed to distinguish these 2
 616          * cases, see mdc_close() for details. The open request reference will
 617          * be put along with freeing \var mod.
 618          */
 619         ptlrpc_request_addref(req);
 620         spin_lock(&req->rq_lock);
 621         req->rq_committed = 1;
 622         spin_unlock(&req->rq_lock);
 623         req->rq_cb_data = NULL;
 624         obd_mod_put(mod);
 625 }
 626
 627 int mdc_set_open_replay_data(struct obd_export *exp,
 628                              struct obd_client_handle *och,
 629                              struct lookup_intent *it)
 630 {
 631         struct md_open_data     *mod;
 632         struct mdt_rec_create   *rec;
 633         struct mdt_body         *body;
 634         struct ptlrpc_request   *open_req = it->it_request;
 635         struct obd_import       *imp = open_req->rq_import;
 636         ENTRY;
 637
 638         if (!open_req->rq_replay)
 639                 RETURN(0);
 640
 641         rec = req_capsule_client_get(&open_req->rq_pill, &RMF_REC_REINT);
 642         body = req_capsule_server_get(&open_req->rq_pill, &RMF_MDT_BODY);
 643         LASSERT(rec != NULL);
 644         /* Incoming message in my byte order (it's been swabbed). */
 645         /* Outgoing messages always in my byte order. */
 646         LASSERT(body != NULL);
 647
 648         /* Only if the import is replayable, we set replay_open data */
 649         if (och && imp->imp_replayable) {
 650                 mod = obd_mod_alloc();
 651                 if (mod == NULL) {
 652                         DEBUG_REQ(D_ERROR, open_req,
 653                                   "Can't allocate md_open_data");
 654                         RETURN(0);
 655                 }
 656
 657                 /**
 658                  * Take a reference on \var mod, to be freed on mdc_close().
 659                  * It protects \var mod from being freed on eviction (commit
 660                  * callback is called despite rq_replay flag).
 661                  * Another reference for \var och.
 662                  */
 663                 obd_mod_get(mod);
 664                 obd_mod_get(mod);
 665
 666                 spin_lock(&open_req->rq_lock);
 667                 och->och_mod = mod;
 668                 mod->mod_och = och;
 669                 mod->mod_is_create = it_disposition(it, DISP_OPEN_CREATE) ||
 670                                      it_disposition(it, DISP_OPEN_STRIPE);
 671                 mod->mod_open_req = open_req;
 672                 open_req->rq_cb_data = mod;
 673                 open_req->rq_commit_cb = mdc_commit_open;
 674                 spin_unlock(&open_req->rq_lock);
 675         }
 676
 677         rec->cr_fid2 = body->mbo_fid1;
 678         rec->cr_ioepoch = body->mbo_ioepoch;
 679         rec->cr_old_handle.cookie = body->mbo_handle.cookie;
 680         open_req->rq_replay_cb = mdc_replay_open;
 681         if (!fid_is_sane(&body->mbo_fid1)) {
 682                 DEBUG_REQ(D_ERROR, open_req, "Saving replay request with "
 683                           "insane fid");
 684                 LBUG();
 685         }
 686
 687         DEBUG_REQ(D_RPCTRACE, open_req, "Set up open replay data");
 688         RETURN(0);
 689 }
 690
 691 static void mdc_free_open(struct md_open_data *mod)
 692 {
 693         int committed = 0;
 694
 695         if (mod->mod_is_create == 0 &&
 696             imp_connect_disp_stripe(mod->mod_open_req->rq_import))
 697                 committed = 1;
 698
 699         /**
 700          * No reason to asssert here if the open request has
 701          * rq_replay == 1. It means that mdc_close failed, and
 702          * close request wasn`t sent. It is not fatal to client.
 703          * The worst thing is eviction if the client gets open lock
 704          **/
 705
 706         DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "free open request rq_replay"
 707                   "= %d\n", mod->mod_open_req->rq_replay);
 708
 709         ptlrpc_request_committed(mod->mod_open_req, committed);
 710         if (mod->mod_close_req)
 711                 ptlrpc_request_committed(mod->mod_close_req, committed);
 712 }
 713
 714 int mdc_clear_open_replay_data(struct obd_export *exp,
 715                                struct obd_client_handle *och)
 716 {
 717         struct md_open_data *mod = och->och_mod;
 718         ENTRY;
 719
 720         /**
 721          * It is possible to not have \var mod in a case of eviction between
 722          * lookup and ll_file_open().
 723          **/
 724         if (mod == NULL)
 725                 RETURN(0);
 726
 727         LASSERT(mod != LP_POISON);
 728         LASSERT(mod->mod_open_req != NULL);
 729         mdc_free_open(mod);
 730
 731         mod->mod_och = NULL;
 732         och->och_mod = NULL;
 733         obd_mod_put(mod);
 734
 735         RETURN(0);
 736 }
 737
 738 static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
 739                      struct md_open_data *mod, struct ptlrpc_request **request)
 740 {
 741         struct obd_device     *obd = class_exp2obd(exp);
 742         struct ptlrpc_request *req;
 743         struct req_format     *req_fmt;
 744         int                    rc;
 745         int                    saved_rc = 0;
 746         ENTRY;
 747
 748         if (op_data->op_bias & MDS_HSM_RELEASE) {
 749                 req_fmt = &RQF_MDS_INTENT_CLOSE;
 750
 751                 /* allocate a FID for volatile file */
 752                 rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
 753                 if (rc < 0) {
 754                         CERROR("%s: "DFID" failed to allocate FID: %d\n",
 755                                obd->obd_name, PFID(&op_data->op_fid1), rc);
 756                         /* save the errcode and proceed to close */
 757                         saved_rc = rc;
 758                 }
 759         } else if (op_data->op_bias & MDS_CLOSE_LAYOUT_SWAP) {
 760                 req_fmt = &RQF_MDS_INTENT_CLOSE;
 761         } else {
 762                 req_fmt = &RQF_MDS_CLOSE;
 763         }
 764
 765         *request = NULL;
 766         if (OBD_FAIL_CHECK(OBD_FAIL_MDC_CLOSE))
 767                 req = NULL;
 768         else
 769                 req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);
 770
 771         /* Ensure that this close's handle is fixed up during replay. */
 772         if (likely(mod != NULL)) {
 773                 LASSERTF(mod->mod_open_req != NULL &&
 774                          mod->mod_open_req->rq_type != LI_POISON,
 775                          "POISONED open %p!\n", mod->mod_open_req);
 776
 777                 mod->mod_close_req = req;
 778
 779                 DEBUG_REQ(D_HA, mod->mod_open_req, "matched open");
 780                 /* We no longer want to preserve this open for replay even
 781                  * though the open was committed. b=3632, b=3633 */
 782                 spin_lock(&mod->mod_open_req->rq_lock);
 783                 mod->mod_open_req->rq_replay = 0;
 784                 spin_unlock(&mod->mod_open_req->rq_lock);
 785         } else {
 786                 CDEBUG(D_HA, "couldn't find open req; expecting close error\n");
 787         }
 788         if (req == NULL) {
 789                 /**
 790                  * TODO: repeat close after errors
 791                  */
 792                 CWARN("%s: close of FID "DFID" failed, file reference will be "
 793                       "dropped when this client unmounts or is evicted\n",
 794                       obd->obd_name, PFID(&op_data->op_fid1));
 795                 GOTO(out, rc = -ENOMEM);
 796         }
 797
 798         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
 799         if (rc) {
 800                 ptlrpc_request_free(req);
 801                 req = NULL;
 802                 GOTO(out, rc);
 803         }
 804
 805         /* To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
 806          * portal whose threads are not taking any DLM locks and are therefore
 807          * always progressing */
 808         req->rq_request_portal = MDS_READPAGE_PORTAL;
 809         ptlrpc_at_set_req_timeout(req);
 810
 811
 812         mdc_close_pack(req, op_data);
 813
 814         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 815                              obd->u.cli.cl_default_mds_easize);
 816
 817         ptlrpc_request_set_replen(req);
 818
 819         mdc_get_mod_rpc_slot(req, NULL);
 820         rc = ptlrpc_queue_wait(req);
 821         mdc_put_mod_rpc_slot(req, NULL);
 822
 823         if (req->rq_repmsg == NULL) {
 824                 CDEBUG(D_RPCTRACE, "request failed to send: %p, %d\n", req,
 825                        req->rq_status);
 826                 if (rc == 0)
 827                         rc = req->rq_status ?: -EIO;
 828         } else if (rc == 0 || rc == -EAGAIN) {
 829                 struct mdt_body *body;
 830
 831                 rc = lustre_msg_get_status(req->rq_repmsg);
 832                 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
 833                         DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR, err "
 834                                   "= %d", rc);
 835                         if (rc > 0)
 836                                 rc = -rc;
 837                 }
 838                 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 839                 if (body == NULL)
 840                         rc = -EPROTO;
 841         } else if (rc == -ESTALE) {
 842                 /**
 843                  * it can be allowed error after 3633 if open was committed and
 844                  * server failed before close was sent. Let's check if mod
 845                  * exists and return no error in that case
 846                  */
 847                 if (mod) {
 848                         DEBUG_REQ(D_HA, req, "Reset ESTALE = %d", rc);
 849                         LASSERT(mod->mod_open_req != NULL);
 850                         if (mod->mod_open_req->rq_committed)
 851                                 rc = 0;
 852                 }
 853         }
 854
 855 out:
 856         if (mod) {
 857                 if (rc != 0)
 858                         mod->mod_close_req = NULL;
 859                 /* Since now, mod is accessed through open_req only,
 860                  * thus close req does not keep a reference on mod anymore. */
 861                 obd_mod_put(mod);
 862         }
 863         *request = req;
 864
 865         RETURN(rc < 0 ? rc : saved_rc);
 866 }
 867
 868 static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid,
 869                        u64 offset, struct page **pages, int npages,
 870                        struct ptlrpc_request **request)
 871 {
 872         struct ptlrpc_request   *req;
 873         struct ptlrpc_bulk_desc *desc;
 874         int                      i;
 875         wait_queue_head_t        waitq;
 876         int                      resends = 0;
 877         struct l_wait_info       lwi;
 878         int                      rc;
 879         ENTRY;
 880
 881         *request = NULL;
 882         init_waitqueue_head(&waitq);
 883
 884 restart_bulk:
 885         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE);
 886         if (req == NULL)
 887                 RETURN(-ENOMEM);
 888
 889         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE);
 890         if (rc) {
 891                 ptlrpc_request_free(req);
 892                 RETURN(rc);
 893         }
 894
 895         req->rq_request_portal = MDS_READPAGE_PORTAL;
 896         ptlrpc_at_set_req_timeout(req);
 897
 898         desc = ptlrpc_prep_bulk_imp(req, npages, 1,
 899                                     PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
 900                                     MDS_BULK_PORTAL,
 901                                     &ptlrpc_bulk_kiov_pin_ops);
 902         if (desc == NULL) {
 903                 ptlrpc_req_finished(req);
 904                 RETURN(-ENOMEM);
 905         }
 906
 907         /* NB req now owns desc and will free it when it gets freed */
 908         for (i = 0; i < npages; i++)
 909                 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
 910                                                  PAGE_SIZE);
 911
 912         mdc_readdir_pack(req, offset, PAGE_SIZE * npages, fid);
 913
 914         ptlrpc_request_set_replen(req);
 915         rc = ptlrpc_queue_wait(req);
 916         if (rc) {
 917                 ptlrpc_req_finished(req);
 918                 if (rc != -ETIMEDOUT)
 919                         RETURN(rc);
 920
 921                 resends++;
 922                 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
 923                         CERROR("%s: too many resend retries: rc = %d\n",
 924                                exp->exp_obd->obd_name, -EIO);
 925                         RETURN(-EIO);
 926                 }
 927                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL,
 928                                        NULL);
 929                 l_wait_event(waitq, 0, &lwi);
 930
 931                 goto restart_bulk;
 932         }
 933
 934         rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
 935                                           req->rq_bulk->bd_nob_transferred);
 936         if (rc < 0) {
 937                 ptlrpc_req_finished(req);
 938                 RETURN(rc);
 939         }
 940
 941         if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) {
 942                 CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n",
 943                        exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred,
 944                        PAGE_SIZE * npages);
 945                 ptlrpc_req_finished(req);
 946                 RETURN(-EPROTO);
 947         }
 948
 949         *request = req;
 950         RETURN(0);
 951 }
 952
 953 static void mdc_release_page(struct page *page, int remove)
 954 {
 955         if (remove) {
 956                 lock_page(page);
 957                 if (likely(page->mapping != NULL))
 958                         truncate_complete_page(page->mapping, page);
 959                 unlock_page(page);
 960         }
 961         put_page(page);
 962 }
 963
 964 static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash,
 965                                     __u64 *start, __u64 *end, int hash64)
 966 {
 967         /*
 968          * Complement of hash is used as an index so that
 969          * radix_tree_gang_lookup() can be used to find a page with starting
 970          * hash _smaller_ than one we are looking for.
 971          */
 972         unsigned long offset = hash_x_index(*hash, hash64);
 973         struct page *page;
 974         int found;
 975
 976         spin_lock_irq(&mapping->tree_lock);
 977         found = radix_tree_gang_lookup(&mapping->page_tree,
 978                                        (void **)&page, offset, 1);
 979         if (found > 0 && !radix_tree_exceptional_entry(page)) {
 980                 struct lu_dirpage *dp;
 981
 982                 get_page(page);
 983                 spin_unlock_irq(&mapping->tree_lock);
 984                 /*
 985                  * In contrast to find_lock_page() we are sure that directory
 986                  * page cannot be truncated (while DLM lock is held) and,
 987                  * hence, can avoid restart.
 988                  *
 989                  * In fact, page cannot be locked here at all, because
 990                  * mdc_read_page_remote does synchronous io.
 991                  */
 992                 wait_on_page_locked(page);
 993                 if (PageUptodate(page)) {
 994                         dp = kmap(page);
 995                         if (BITS_PER_LONG == 32 && hash64) {
 996                                 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
 997                                 *end   = le64_to_cpu(dp->ldp_hash_end) >> 32;
 998                                 *hash  = *hash >> 32;
 999                         } else {
1000                                 *start = le64_to_cpu(dp->ldp_hash_start);
1001                                 *end   = le64_to_cpu(dp->ldp_hash_end);
1002                         }
1003                         if (unlikely(*start == 1 && *hash == 0))
1004                                 *hash = *start;
1005                         else
1006                                 LASSERTF(*start <= *hash, "start = %#llx"
1007                                          ",end = %#llx,hash = %#llx\n",
1008                                          *start, *end, *hash);
1009                         CDEBUG(D_VFSTRACE, "offset %lx [%#llx %#llx],"
1010                               " hash %#llx\n", offset, *start, *end, *hash);
1011                         if (*hash > *end) {
1012                                 kunmap(page);
1013                                 mdc_release_page(page, 0);
1014                                 page = NULL;
1015                         } else if (*end != *start && *hash == *end) {
1016                                 /*
1017                                  * upon hash collision, remove this page,
1018                                  * otherwise put page reference, and
1019                                  * mdc_read_page_remote() will issue RPC to
1020                                  * fetch the page we want.
1021                                  */
1022                                 kunmap(page);
1023                                 mdc_release_page(page,
1024                                     le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1025                                 page = NULL;
1026                         }
1027                 } else {
1028                         put_page(page);
1029                         page = ERR_PTR(-EIO);
1030                 }
1031         } else {
1032                 spin_unlock_irq(&mapping->tree_lock);
1033                 page = NULL;
1034         }
1035         return page;
1036 }
1037
1038 /*
1039  * Adjust a set of pages, each page containing an array of lu_dirpages,
1040  * so that each page can be used as a single logical lu_dirpage.
1041  *
1042  * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
1043  * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
1044  * struct lu_dirent.  It has size up to LU_PAGE_SIZE. The ldp_hash_end
1045  * value is used as a cookie to request the next lu_dirpage in a
1046  * directory listing that spans multiple pages (two in this example):
1047  *   ________
1048  *  |        |
1049  * .|--------v-------   -----.
1050  * |s|e|f|p|ent|ent| ... |ent|
1051  * '--|--------------   -----'   Each PAGE contains a single
1052  *    '------.                   lu_dirpage.
1053  * .---------v-------   -----.
1054  * |s|e|f|p|ent| 0 | ... | 0 |
1055  * '-----------------   -----'
1056  *
1057  * However, on hosts where the native VM page size (PAGE_SIZE) is
1058  * larger than LU_PAGE_SIZE, a single host page may contain multiple
1059  * lu_dirpages. After reading the lu_dirpages from the MDS, the
1060  * ldp_hash_end of the first lu_dirpage refers to the one immediately
1061  * after it in the same PAGE (arrows simplified for brevity, but
1062  * in general e0==s1, e1==s2, etc.):
1063  *
1064  * .--------------------   -----.
1065  * |s0|e0|f0|p|ent|ent| ... |ent|
1066  * |---v----------------   -----|
1067  * |s1|e1|f1|p|ent|ent| ... |ent|
1068  * |---v----------------   -----|  Here, each PAGE contains
1069  *             ...                 multiple lu_dirpages.
1070  * |---v----------------   -----|
1071  * |s'|e'|f'|p|ent|ent| ... |ent|
1072  * '---|----------------   -----'
1073  *     v
1074  * .----------------------------.
1075  * |        next PAGE           |
1076  *
1077  * This structure is transformed into a single logical lu_dirpage as follows:
1078  *
1079  * - Replace e0 with e' so the request for the next lu_dirpage gets the page
1080  *   labeled 'next PAGE'.
1081  *
1082  * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
1083  *   a hash collision with the next page exists.
1084  *
1085  * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
1086  *   to the first entry of the next lu_dirpage.
1087  */
1088 #if PAGE_SIZE > LU_PAGE_SIZE
1089 static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs)
1090 {
1091         int i;
1092
1093         for (i = 0; i < cfs_pgs; i++) {
1094                 struct lu_dirpage       *dp = kmap(pages[i]);
1095                 struct lu_dirpage       *first = dp;
1096                 struct lu_dirent        *end_dirent = NULL;
1097                 struct lu_dirent        *ent;
1098                 __u64           hash_end = le64_to_cpu(dp->ldp_hash_end);
1099                 __u32           flags = le32_to_cpu(dp->ldp_flags);
1100
1101                 while (--lu_pgs > 0) {
1102                         ent = lu_dirent_start(dp);
1103                         for (end_dirent = ent; ent != NULL;
1104                              end_dirent = ent, ent = lu_dirent_next(ent));
1105
1106                         /* Advance dp to next lu_dirpage. */
1107                         dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
1108
1109                         /* Check if we've reached the end of the PAGE. */
1110                         if (!((unsigned long)dp & ~PAGE_MASK))
1111                                 break;
1112
1113                         /* Save the hash and flags of this lu_dirpage. */
1114                         hash_end = le64_to_cpu(dp->ldp_hash_end);
1115                         flags = le32_to_cpu(dp->ldp_flags);
1116
1117                         /* Check if lu_dirpage contains no entries. */
1118                         if (end_dirent == NULL)
1119                                 break;
1120
1121                         /* Enlarge the end entry lde_reclen from 0 to
1122                          * first entry of next lu_dirpage. */
1123                         LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
1124                         end_dirent->lde_reclen =
1125                                 cpu_to_le16((char *)(dp->ldp_entries) -
1126                                             (char *)end_dirent);
1127                 }
1128
1129                 first->ldp_hash_end = hash_end;
1130                 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
1131                 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
1132
1133                 kunmap(pages[i]);
1134         }
1135         LASSERTF(lu_pgs == 0, "left = %d\n", lu_pgs);
1136 }
1137 #else
1138 #define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0)
1139 #endif  /* PAGE_SIZE > LU_PAGE_SIZE */
1140
1141 /* parameters for readdir page */
1142 struct readpage_param {
1143         struct md_op_data       *rp_mod;
1144         __u64                   rp_off;
1145         int                     rp_hash64;
1146         struct obd_export       *rp_exp;
1147         struct md_callback      *rp_cb;
1148 };
1149
1150 #ifndef HAVE_DELETE_FROM_PAGE_CACHE
1151 static inline void delete_from_page_cache(struct page *page)
1152 {
1153         remove_from_page_cache(page);
1154         put_page(page);
1155 }
1156 #endif
1157
1158 /**
1159  * Read pages from server.
1160  *
1161  * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
1162  * a header lu_dirpage which describes the start/end hash, and whether this
1163  * page is empty (contains no dir entry) or hash collide with next page.
1164  * After client receives reply, several pages will be integrated into dir page
1165  * in PAGE_SIZE (if PAGE_SIZE greater than LU_PAGE_SIZE), and the
1166  * lu_dirpage for this integrated page will be adjusted.
1167  **/
1168 static int mdc_read_page_remote(void *data, struct page *page0)
1169 {
1170         struct readpage_param *rp = data;
1171         struct page **page_pool;
1172         struct page *page;
1173         struct lu_dirpage *dp;
1174         struct md_op_data *op_data = rp->rp_mod;
1175         struct ptlrpc_request *req;
1176         int max_pages;
1177         struct inode *inode;
1178         struct lu_fid *fid;
1179         int rd_pgs = 0; /* number of pages actually read */
1180         int npages;
1181         int i;
1182         int rc;
1183         ENTRY;
1184
1185         max_pages = rp->rp_exp->exp_obd->u.cli.cl_max_pages_per_rpc;
1186         inode = op_data->op_data;
1187         fid = &op_data->op_fid1;
1188         LASSERT(inode != NULL);
1189
1190         OBD_ALLOC(page_pool, sizeof(page_pool[0]) * max_pages);
1191         if (page_pool != NULL) {
1192                 page_pool[0] = page0;
1193         } else {
1194                 page_pool = &page0;
1195                 max_pages = 1;
1196         }
1197
1198         for (npages = 1; npages < max_pages; npages++) {
1199                 page = page_cache_alloc_cold(inode->i_mapping);
1200                 if (page == NULL)
1201                         break;
1202                 page_pool[npages] = page;
1203         }
1204
1205         rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, page_pool, npages, &req);
1206         if (rc < 0) {
1207                 /* page0 is special, which was added into page cache early */
1208                 delete_from_page_cache(page0);
1209         } else {
1210                 int lu_pgs;
1211
1212                 rd_pgs = (req->rq_bulk->bd_nob_transferred + PAGE_SIZE - 1) >>
1213                         PAGE_SHIFT;
1214                 lu_pgs = req->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
1215                 LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
1216
1217                 CDEBUG(D_INODE, "read %d(%d) pages\n", rd_pgs, lu_pgs);
1218
1219                 mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs);
1220
1221                 SetPageUptodate(page0);
1222         }
1223         unlock_page(page0);
1224
1225         ptlrpc_req_finished(req);
1226         CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages);
1227         for (i = 1; i < npages; i++) {
1228                 unsigned long   offset;
1229                 __u64           hash;
1230                 int ret;
1231
1232                 page = page_pool[i];
1233
1234                 if (rc < 0 || i >= rd_pgs) {
1235                         put_page(page);
1236                         continue;
1237                 }
1238
1239                 SetPageUptodate(page);
1240
1241                 dp = kmap(page);
1242                 hash = le64_to_cpu(dp->ldp_hash_start);
1243                 kunmap(page);
1244
1245                 offset = hash_x_index(hash, rp->rp_hash64);
1246
1247                 prefetchw(&page->flags);
1248                 ret = add_to_page_cache_lru(page, inode->i_mapping, offset,
1249                                             GFP_KERNEL);
1250                 if (ret == 0)
1251                         unlock_page(page);
1252                 else
1253                         CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
1254                                " rc = %d\n", offset, ret);
1255                 put_page(page);
1256         }
1257
1258         if (page_pool != &page0)
1259                 OBD_FREE(page_pool, sizeof(page_pool[0]) * max_pages);
1260
1261         RETURN(rc);
1262 }
1263
1264 /**
1265  * Read dir page from cache first, if it can not find it, read it from
1266  * server and add into the cache.
1267  *
1268  * \param[in] exp       MDC export
1269  * \param[in] op_data   client MD stack parameters, transfering parameters
1270  *                      between different layers on client MD stack.
1271  * \param[in] cb_op     callback required for ldlm lock enqueue during
1272  *                      read page
1273  * \param[in] hash_offset the hash offset of the page to be read
1274  * \param[in] ppage     the page to be read
1275  *
1276  * retval               = 0 get the page successfully
1277  *                      errno(<0) get the page failed
1278  */
1279 static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data,
1280                          struct md_callback *cb_op, __u64 hash_offset,
1281                          struct page **ppage)
1282 {
1283         struct lookup_intent    it = { .it_op = IT_READDIR };
1284         struct page             *page;
1285         struct inode            *dir = op_data->op_data;
1286         struct address_space    *mapping;
1287         struct lu_dirpage       *dp;
1288         __u64                   start = 0;
1289         __u64                   end = 0;
1290         struct lustre_handle    lockh;
1291         struct ptlrpc_request   *enq_req = NULL;
1292         struct readpage_param   rp_param;
1293         int rc;
1294
1295         ENTRY;
1296
1297         *ppage = NULL;
1298
1299         LASSERT(dir != NULL);
1300         mapping = dir->i_mapping;
1301
1302         rc = mdc_intent_lock(exp, op_data, &it, &enq_req,
1303                              cb_op->md_blocking_ast, 0);
1304         if (enq_req != NULL)
1305                 ptlrpc_req_finished(enq_req);
1306
1307         if (rc < 0) {
1308                 CERROR("%s: "DFID" lock enqueue fails: rc = %d\n",
1309                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc);
1310                 RETURN(rc);
1311         }
1312
1313         rc = 0;
1314         lockh.cookie = it.it_lock_handle;
1315         mdc_set_lock_data(exp, &lockh, dir, NULL);
1316
1317         rp_param.rp_off = hash_offset;
1318         rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64;
1319         page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end,
1320                                rp_param.rp_hash64);
1321         if (IS_ERR(page)) {
1322                 CERROR("%s: dir page locate: "DFID" at %llu: rc %ld\n",
1323                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1324                        rp_param.rp_off, PTR_ERR(page));
1325                 GOTO(out_unlock, rc = PTR_ERR(page));
1326         } else if (page != NULL) {
1327                 /*
1328                  * XXX nikita: not entirely correct handling of a corner case:
1329                  * suppose hash chain of entries with hash value HASH crosses
1330                  * border between pages P0 and P1. First both P0 and P1 are
1331                  * cached, seekdir() is called for some entry from the P0 part
1332                  * of the chain. Later P0 goes out of cache. telldir(HASH)
1333                  * happens and finds P1, as it starts with matching hash
1334                  * value. Remaining entries from P0 part of the chain are
1335                  * skipped. (Is that really a bug?)
1336                  *
1337                  * Possible solutions: 0. don't cache P1 is such case, handle
1338                  * it as an "overflow" page. 1. invalidate all pages at
1339                  * once. 2. use HASH|1 as an index for P1.
1340                  */
1341                 GOTO(hash_collision, page);
1342         }
1343
1344         rp_param.rp_exp = exp;
1345         rp_param.rp_mod = op_data;
1346         page = read_cache_page(mapping,
1347                                hash_x_index(rp_param.rp_off,
1348                                             rp_param.rp_hash64),
1349                                mdc_read_page_remote, &rp_param);
1350         if (IS_ERR(page)) {
1351                 CDEBUG(D_INFO, "%s: read cache page: "DFID" at %llu: %ld\n",
1352                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1353                        rp_param.rp_off, PTR_ERR(page));
1354                 GOTO(out_unlock, rc = PTR_ERR(page));
1355         }
1356
1357         wait_on_page_locked(page);
1358         (void)kmap(page);
1359         if (!PageUptodate(page)) {
1360                 CERROR("%s: page not updated: "DFID" at %llu: rc %d\n",
1361                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1362                        rp_param.rp_off, -5);
1363                 goto fail;
1364         }
1365         if (!PageChecked(page))
1366                 SetPageChecked(page);
1367         if (PageError(page)) {
1368                 CERROR("%s: page error: "DFID" at %llu: rc %d\n",
1369                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1370                        rp_param.rp_off, -5);
1371                 goto fail;
1372         }
1373
1374 hash_collision:
1375         dp = page_address(page);
1376         if (BITS_PER_LONG == 32 && rp_param.rp_hash64) {
1377                 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1378                 end   = le64_to_cpu(dp->ldp_hash_end) >> 32;
1379                 rp_param.rp_off = hash_offset >> 32;
1380         } else {
1381                 start = le64_to_cpu(dp->ldp_hash_start);
1382                 end   = le64_to_cpu(dp->ldp_hash_end);
1383                 rp_param.rp_off = hash_offset;
1384         }
1385         if (end == start) {
1386                 LASSERT(start == rp_param.rp_off);
1387                 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
1388 #if BITS_PER_LONG == 32
1389                 CWARN("Real page-wide hash collision at [%llu %llu] with "
1390                       "hash %llu\n", le64_to_cpu(dp->ldp_hash_start),
1391                       le64_to_cpu(dp->ldp_hash_end), hash_offset);
1392 #endif
1393
1394                 /*
1395                  * Fetch whole overflow chain...
1396                  *
1397                  * XXX not yet.
1398                  */
1399                 goto fail;
1400         }
1401         *ppage = page;
1402 out_unlock:
1403         ldlm_lock_decref(&lockh, it.it_lock_mode);
1404         return rc;
1405 fail:
1406         kunmap(page);
1407         mdc_release_page(page, 1);
1408         rc = -EIO;
1409         goto out_unlock;
1410 }
1411
1412
1413 static int mdc_statfs(const struct lu_env *env,
1414                       struct obd_export *exp, struct obd_statfs *osfs,
1415                       __u64 max_age, __u32 flags)
1416 {
1417         struct obd_device     *obd = class_exp2obd(exp);
1418         struct ptlrpc_request *req;
1419         struct obd_statfs     *msfs;
1420         struct obd_import     *imp = NULL;
1421         int                    rc;
1422         ENTRY;
1423
1424         /*
1425          * Since the request might also come from lprocfs, so we need
1426          * sync this with client_disconnect_export Bug15684
1427          */
1428         down_read(&obd->u.cli.cl_sem);
1429         if (obd->u.cli.cl_import)
1430                 imp = class_import_get(obd->u.cli.cl_import);
1431         up_read(&obd->u.cli.cl_sem);
1432         if (!imp)
1433                 RETURN(-ENODEV);
1434
1435         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_STATFS,
1436                                         LUSTRE_MDS_VERSION, MDS_STATFS);
1437         if (req == NULL)
1438                 GOTO(output, rc = -ENOMEM);
1439
1440         ptlrpc_request_set_replen(req);
1441
1442         if (flags & OBD_STATFS_NODELAY) {
1443                 /* procfs requests not want stay in wait for avoid deadlock */
1444                 req->rq_no_resend = 1;
1445                 req->rq_no_delay = 1;
1446         }
1447
1448         rc = ptlrpc_queue_wait(req);
1449         if (rc) {
1450                 /* check connection error first */
1451                 if (imp->imp_connect_error)
1452                         rc = imp->imp_connect_error;
1453                 GOTO(out, rc);
1454         }
1455
1456         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1457         if (msfs == NULL)
1458                 GOTO(out, rc = -EPROTO);
1459
1460         *osfs = *msfs;
1461         EXIT;
1462 out:
1463         ptlrpc_req_finished(req);
1464 output:
1465         class_import_put(imp);
1466         return rc;
1467 }
1468
1469 static int mdc_ioc_fid2path(struct obd_export *exp, struct getinfo_fid2path *gf)
1470 {
1471         __u32 keylen, vallen;
1472         void *key;
1473         int rc;
1474
1475         if (gf->gf_pathlen > PATH_MAX)
1476                 RETURN(-ENAMETOOLONG);
1477         if (gf->gf_pathlen < 2)
1478                 RETURN(-EOVERFLOW);
1479
1480         /* Key is KEY_FID2PATH + getinfo_fid2path description */
1481         keylen = cfs_size_round(sizeof(KEY_FID2PATH) + sizeof(*gf) +
1482                                 sizeof(struct lu_fid));
1483         OBD_ALLOC(key, keylen);
1484         if (key == NULL)
1485                 RETURN(-ENOMEM);
1486         memcpy(key, KEY_FID2PATH, sizeof(KEY_FID2PATH));
1487         memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)), gf, sizeof(*gf));
1488         memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*gf),
1489                gf->gf_u.gf_root_fid, sizeof(struct lu_fid));
1490         CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
1491                PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno);
1492
1493         if (!fid_is_sane(&gf->gf_fid))
1494                 GOTO(out, rc = -EINVAL);
1495
1496         /* Val is struct getinfo_fid2path result plus path */
1497         vallen = sizeof(*gf) + gf->gf_pathlen;
1498
1499         rc = obd_get_info(NULL, exp, keylen, key, &vallen, gf);
1500         if (rc != 0 && rc != -EREMOTE)
1501                 GOTO(out, rc);
1502
1503         if (vallen <= sizeof(*gf))
1504                 GOTO(out, rc = -EPROTO);
1505         if (vallen > sizeof(*gf) + gf->gf_pathlen)
1506                 GOTO(out, rc = -EOVERFLOW);
1507
1508         CDEBUG(D_IOCTL, "path got "DFID" from %llu #%d: %s\n",
1509                PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno,
1510                gf->gf_pathlen < 512 ? gf->gf_u.gf_path :
1511                /* only log the last 512 characters of the path */
1512                gf->gf_u.gf_path + gf->gf_pathlen - 512);
1513
1514 out:
1515         OBD_FREE(key, keylen);
1516         return rc;
1517 }
1518
1519 static int mdc_ioc_hsm_progress(struct obd_export *exp,
1520                                 struct hsm_progress_kernel *hpk)
1521 {
1522         struct obd_import               *imp = class_exp2cliimp(exp);
1523         struct hsm_progress_kernel      *req_hpk;
1524         struct ptlrpc_request           *req;
1525         int                              rc;
1526         ENTRY;
1527
1528         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_PROGRESS,
1529                                         LUSTRE_MDS_VERSION, MDS_HSM_PROGRESS);
1530         if (req == NULL)
1531                 GOTO(out, rc = -ENOMEM);
1532
1533         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1534
1535         /* Copy hsm_progress struct */
1536         req_hpk = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_PROGRESS);
1537         if (req_hpk == NULL)
1538                 GOTO(out, rc = -EPROTO);
1539
1540         *req_hpk = *hpk;
1541         req_hpk->hpk_errval = lustre_errno_hton(hpk->hpk_errval);
1542
1543         ptlrpc_request_set_replen(req);
1544
1545         mdc_get_mod_rpc_slot(req, NULL);
1546         rc = ptlrpc_queue_wait(req);
1547         mdc_put_mod_rpc_slot(req, NULL);
1548
1549         GOTO(out, rc);
1550 out:
1551         ptlrpc_req_finished(req);
1552         return rc;
1553 }
1554
1555 static int mdc_ioc_hsm_ct_register(struct obd_import *imp, __u32 archives)
1556 {
1557         __u32                   *archive_mask;
1558         struct ptlrpc_request   *req;
1559         int                      rc;
1560         ENTRY;
1561
1562         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_REGISTER,
1563                                         LUSTRE_MDS_VERSION,
1564                                         MDS_HSM_CT_REGISTER);
1565         if (req == NULL)
1566                 GOTO(out, rc = -ENOMEM);
1567
1568         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1569
1570         /* Copy hsm_progress struct */
1571         archive_mask = req_capsule_client_get(&req->rq_pill,
1572                                               &RMF_MDS_HSM_ARCHIVE);
1573         if (archive_mask == NULL)
1574                 GOTO(out, rc = -EPROTO);
1575
1576         *archive_mask = archives;
1577
1578         ptlrpc_request_set_replen(req);
1579
1580         rc = mdc_queue_wait(req);
1581         GOTO(out, rc);
1582 out:
1583         ptlrpc_req_finished(req);
1584         return rc;
1585 }
1586
1587 static int mdc_ioc_hsm_current_action(struct obd_export *exp,
1588                                       struct md_op_data *op_data)
1589 {
1590         struct hsm_current_action       *hca = op_data->op_data;
1591         struct hsm_current_action       *req_hca;
1592         struct ptlrpc_request           *req;
1593         int                              rc;
1594         ENTRY;
1595
1596         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1597                                    &RQF_MDS_HSM_ACTION);
1598         if (req == NULL)
1599                 RETURN(-ENOMEM);
1600
1601         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_ACTION);
1602         if (rc) {
1603                 ptlrpc_request_free(req);
1604                 RETURN(rc);
1605         }
1606
1607         mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1608                       op_data->op_suppgids[0], 0);
1609
1610         ptlrpc_request_set_replen(req);
1611
1612         rc = mdc_queue_wait(req);
1613         if (rc)
1614                 GOTO(out, rc);
1615
1616         req_hca = req_capsule_server_get(&req->rq_pill,
1617                                          &RMF_MDS_HSM_CURRENT_ACTION);
1618         if (req_hca == NULL)
1619                 GOTO(out, rc = -EPROTO);
1620
1621         *hca = *req_hca;
1622
1623         EXIT;
1624 out:
1625         ptlrpc_req_finished(req);
1626         return rc;
1627 }
1628
1629 static int mdc_ioc_hsm_ct_unregister(struct obd_import *imp)
1630 {
1631         struct ptlrpc_request   *req;
1632         int                      rc;
1633         ENTRY;
1634
1635         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_UNREGISTER,
1636                                         LUSTRE_MDS_VERSION,
1637                                         MDS_HSM_CT_UNREGISTER);
1638         if (req == NULL)
1639                 GOTO(out, rc = -ENOMEM);
1640
1641         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1642
1643         ptlrpc_request_set_replen(req);
1644
1645         rc = mdc_queue_wait(req);
1646         GOTO(out, rc);
1647 out:
1648         ptlrpc_req_finished(req);
1649         return rc;
1650 }
1651
1652 static int mdc_ioc_hsm_state_get(struct obd_export *exp,
1653                                  struct md_op_data *op_data)
1654 {
1655         struct hsm_user_state   *hus = op_data->op_data;
1656         struct hsm_user_state   *req_hus;
1657         struct ptlrpc_request   *req;
1658         int                      rc;
1659         ENTRY;
1660
1661         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1662                                    &RQF_MDS_HSM_STATE_GET);
1663         if (req == NULL)
1664                 RETURN(-ENOMEM);
1665
1666         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_GET);
1667         if (rc != 0) {
1668                 ptlrpc_request_free(req);
1669                 RETURN(rc);
1670         }
1671
1672         mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1673                       op_data->op_suppgids[0], 0);
1674
1675         ptlrpc_request_set_replen(req);
1676
1677         rc = mdc_queue_wait(req);
1678         if (rc)
1679                 GOTO(out, rc);
1680
1681         req_hus = req_capsule_server_get(&req->rq_pill, &RMF_HSM_USER_STATE);
1682         if (req_hus == NULL)
1683                 GOTO(out, rc = -EPROTO);
1684
1685         *hus = *req_hus;
1686
1687         EXIT;
1688 out:
1689         ptlrpc_req_finished(req);
1690         return rc;
1691 }
1692
1693 static int mdc_ioc_hsm_state_set(struct obd_export *exp,
1694                                  struct md_op_data *op_data)
1695 {
1696         struct hsm_state_set    *hss = op_data->op_data;
1697         struct hsm_state_set    *req_hss;
1698         struct ptlrpc_request   *req;
1699         int                      rc;
1700         ENTRY;
1701
1702         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1703                                    &RQF_MDS_HSM_STATE_SET);
1704         if (req == NULL)
1705                 RETURN(-ENOMEM);
1706
1707         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_SET);
1708         if (rc) {
1709                 ptlrpc_request_free(req);
1710                 RETURN(rc);
1711         }
1712
1713         mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1714                       op_data->op_suppgids[0], 0);
1715
1716         /* Copy states */
1717         req_hss = req_capsule_client_get(&req->rq_pill, &RMF_HSM_STATE_SET);
1718         if (req_hss == NULL)
1719                 GOTO(out, rc = -EPROTO);
1720         *req_hss = *hss;
1721
1722         ptlrpc_request_set_replen(req);
1723
1724         mdc_get_mod_rpc_slot(req, NULL);
1725         rc = ptlrpc_queue_wait(req);
1726         mdc_put_mod_rpc_slot(req, NULL);
1727
1728         GOTO(out, rc);
1729 out:
1730         ptlrpc_req_finished(req);
1731         return rc;
1732 }
1733
1734 static int mdc_ioc_hsm_request(struct obd_export *exp,
1735                                struct hsm_user_request *hur)
1736 {
1737         struct obd_import       *imp = class_exp2cliimp(exp);
1738         struct ptlrpc_request   *req;
1739         struct hsm_request      *req_hr;
1740         struct hsm_user_item    *req_hui;
1741         char                    *req_opaque;
1742         int                      rc;
1743         ENTRY;
1744
1745         req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_REQUEST);
1746         if (req == NULL)
1747                 GOTO(out, rc = -ENOMEM);
1748
1749         req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM, RCL_CLIENT,
1750                              hur->hur_request.hr_itemcount
1751                              * sizeof(struct hsm_user_item));
1752         req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA, RCL_CLIENT,
1753                              hur->hur_request.hr_data_len);
1754
1755         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_REQUEST);
1756         if (rc) {
1757                 ptlrpc_request_free(req);
1758                 RETURN(rc);
1759         }
1760
1761         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1762
1763         /* Copy hsm_request struct */
1764         req_hr = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_REQUEST);
1765         if (req_hr == NULL)
1766                 GOTO(out, rc = -EPROTO);
1767         *req_hr = hur->hur_request;
1768
1769         /* Copy hsm_user_item structs */
1770         req_hui = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM);
1771         if (req_hui == NULL)
1772                 GOTO(out, rc = -EPROTO);
1773         memcpy(req_hui, hur->hur_user_item,
1774                hur->hur_request.hr_itemcount * sizeof(struct hsm_user_item));
1775
1776         /* Copy opaque field */
1777         req_opaque = req_capsule_client_get(&req->rq_pill, &RMF_GENERIC_DATA);
1778         if (req_opaque == NULL)
1779                 GOTO(out, rc = -EPROTO);
1780         memcpy(req_opaque, hur_data(hur), hur->hur_request.hr_data_len);
1781
1782         ptlrpc_request_set_replen(req);
1783
1784         mdc_get_mod_rpc_slot(req, NULL);
1785         rc = ptlrpc_queue_wait(req);
1786         mdc_put_mod_rpc_slot(req, NULL);
1787
1788         GOTO(out, rc);
1789
1790 out:
1791         ptlrpc_req_finished(req);
1792         return rc;
1793 }
1794
1795 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
1796                                 struct lustre_kernelcomm *lk);
1797
1798 static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp,
1799                         struct obd_quotactl *oqctl)
1800 {
1801         struct ptlrpc_request   *req;
1802         struct obd_quotactl     *oqc;
1803         int                      rc;
1804         ENTRY;
1805
1806         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1807                                         &RQF_MDS_QUOTACTL, LUSTRE_MDS_VERSION,
1808                                         MDS_QUOTACTL);
1809         if (req == NULL)
1810                 RETURN(-ENOMEM);
1811
1812         oqc = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1813         *oqc = *oqctl;
1814
1815         ptlrpc_request_set_replen(req);
1816         ptlrpc_at_set_req_timeout(req);
1817         req->rq_no_resend = 1;
1818
1819         rc = ptlrpc_queue_wait(req);
1820         if (rc)
1821                 CERROR("ptlrpc_queue_wait failed, rc: %d\n", rc);
1822
1823         if (req->rq_repmsg &&
1824             (oqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL))) {
1825                 *oqctl = *oqc;
1826         } else if (!rc) {
1827                 CERROR ("Can't unpack obd_quotactl\n");
1828                 rc = -EPROTO;
1829         }
1830         ptlrpc_req_finished(req);
1831
1832         RETURN(rc);
1833 }
1834
1835 static int mdc_ioc_swap_layouts(struct obd_export *exp,
1836                                 struct md_op_data *op_data)
1837 {
1838         struct list_head cancels = LIST_HEAD_INIT(cancels);
1839         struct ptlrpc_request   *req;
1840         int                      rc, count;
1841         struct mdc_swap_layouts *msl, *payload;
1842         ENTRY;
1843
1844         msl = op_data->op_data;
1845
1846         /* When the MDT will get the MDS_SWAP_LAYOUTS RPC the
1847          * first thing it will do is to cancel the 2 layout
1848          * locks held by this client.
1849          * So the client must cancel its layout locks on the 2 fids
1850          * with the request RPC to avoid extra RPC round trips.
1851          */
1852         count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels,
1853                                         LCK_EX, MDS_INODELOCK_LAYOUT |
1854                                         MDS_INODELOCK_XATTR);
1855         count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels,
1856                                          LCK_EX, MDS_INODELOCK_LAYOUT |
1857                                          MDS_INODELOCK_XATTR);
1858
1859         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1860                                    &RQF_MDS_SWAP_LAYOUTS);
1861         if (req == NULL) {
1862                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
1863                 RETURN(-ENOMEM);
1864         }
1865
1866         rc = mdc_prep_elc_req(exp, req, MDS_SWAP_LAYOUTS, &cancels, count);
1867         if (rc) {
1868                 ptlrpc_request_free(req);
1869                 RETURN(rc);
1870         }
1871
1872         mdc_swap_layouts_pack(req, op_data);
1873
1874         payload = req_capsule_client_get(&req->rq_pill, &RMF_SWAP_LAYOUTS);
1875         LASSERT(payload);
1876
1877         *payload = *msl;
1878
1879         ptlrpc_request_set_replen(req);
1880
1881         rc = ptlrpc_queue_wait(req);
1882         if (rc)
1883                 GOTO(out, rc);
1884         EXIT;
1885
1886 out:
1887         ptlrpc_req_finished(req);
1888         return rc;
1889 }
1890
1891 static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1892                          void *karg, void __user *uarg)
1893 {
1894         struct obd_device *obd = exp->exp_obd;
1895         struct obd_ioctl_data *data = karg;
1896         struct obd_import *imp = obd->u.cli.cl_import;
1897         int rc;
1898         ENTRY;
1899
1900         if (!try_module_get(THIS_MODULE)) {
1901                 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
1902                        module_name(THIS_MODULE));
1903                 return -EINVAL;
1904         }
1905         switch (cmd) {
1906         case OBD_IOC_FID2PATH:
1907                 rc = mdc_ioc_fid2path(exp, karg);
1908                 GOTO(out, rc);
1909         case LL_IOC_HSM_CT_START:
1910                 rc = mdc_ioc_hsm_ct_start(exp, karg);
1911                 /* ignore if it was already registered on this MDS. */
1912                 if (rc == -EEXIST)
1913                         rc = 0;
1914                 GOTO(out, rc);
1915         case LL_IOC_HSM_PROGRESS:
1916                 rc = mdc_ioc_hsm_progress(exp, karg);
1917                 GOTO(out, rc);
1918         case LL_IOC_HSM_STATE_GET:
1919                 rc = mdc_ioc_hsm_state_get(exp, karg);
1920                 GOTO(out, rc);
1921         case LL_IOC_HSM_STATE_SET:
1922                 rc = mdc_ioc_hsm_state_set(exp, karg);
1923                 GOTO(out, rc);
1924         case LL_IOC_HSM_ACTION:
1925                 rc = mdc_ioc_hsm_current_action(exp, karg);
1926                 GOTO(out, rc);
1927         case LL_IOC_HSM_REQUEST:
1928                 rc = mdc_ioc_hsm_request(exp, karg);
1929                 GOTO(out, rc);
1930         case OBD_IOC_CLIENT_RECOVER:
1931                 rc = ptlrpc_recover_import(imp, data->ioc_inlbuf1, 0);
1932                 if (rc < 0)
1933                         GOTO(out, rc);
1934                 GOTO(out, rc = 0);
1935         case IOC_OSC_SET_ACTIVE:
1936                 rc = ptlrpc_set_import_active(imp, data->ioc_offset);
1937                 GOTO(out, rc);
1938         case OBD_IOC_PING_TARGET:
1939                 rc = ptlrpc_obd_ping(obd);
1940                 GOTO(out, rc);
1941         /*
1942          * Normally IOC_OBD_STATFS, OBD_IOC_QUOTACTL iocontrol are handled by
1943          * LMV instead of MDC. But when the cluster is upgraded from 1.8,
1944          * there'd be no LMV layer thus we might be called here. Eventually
1945          * this code should be removed.
1946          * bz20731, LU-592.
1947          */
1948         case IOC_OBD_STATFS: {
1949                 struct obd_statfs stat_buf = {0};
1950
1951                 if (*((__u32 *) data->ioc_inlbuf2) != 0)
1952                         GOTO(out, rc = -ENODEV);
1953
1954                 /* copy UUID */
1955                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(obd),
1956                                  min((int)data->ioc_plen2,
1957                                      (int)sizeof(struct obd_uuid))))
1958                         GOTO(out, rc = -EFAULT);
1959
1960                 rc = mdc_statfs(NULL, obd->obd_self_export, &stat_buf,
1961                                 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
1962                                 0);
1963                 if (rc != 0)
1964                         GOTO(out, rc);
1965
1966                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
1967                                      min((int) data->ioc_plen1,
1968                                          (int) sizeof(stat_buf))))
1969                         GOTO(out, rc = -EFAULT);
1970
1971                 GOTO(out, rc = 0);
1972         }
1973         case OBD_IOC_QUOTACTL: {
1974                 struct if_quotactl *qctl = karg;
1975                 struct obd_quotactl *oqctl;
1976
1977                 OBD_ALLOC_PTR(oqctl);
1978                 if (oqctl == NULL)
1979                         GOTO(out, rc = -ENOMEM);
1980
1981                 QCTL_COPY(oqctl, qctl);
1982                 rc = obd_quotactl(exp, oqctl);
1983                 if (rc == 0) {
1984                         QCTL_COPY(qctl, oqctl);
1985                         qctl->qc_valid = QC_MDTIDX;
1986                         qctl->obd_uuid = obd->u.cli.cl_target_uuid;
1987                 }
1988
1989                 OBD_FREE_PTR(oqctl);
1990                 GOTO(out, rc);
1991         }
1992         case LL_IOC_GET_CONNECT_FLAGS:
1993                 if (copy_to_user(uarg, exp_connect_flags_ptr(exp),
1994                                  sizeof(*exp_connect_flags_ptr(exp))))
1995                         GOTO(out, rc = -EFAULT);
1996
1997                 GOTO(out, rc = 0);
1998         case LL_IOC_LOV_SWAP_LAYOUTS:
1999                 rc = mdc_ioc_swap_layouts(exp, karg);
2000                 GOTO(out, rc);
2001         default:
2002                 CERROR("unrecognised ioctl: cmd = %#x\n", cmd);
2003                 GOTO(out, rc = -ENOTTY);
2004         }
2005 out:
2006         module_put(THIS_MODULE);
2007
2008         return rc;
2009 }
2010
2011 static int mdc_get_info_rpc(struct obd_export *exp,
2012                             u32 keylen, void *key,
2013                             u32 vallen, void *val)
2014 {
2015         struct obd_import      *imp = class_exp2cliimp(exp);
2016         struct ptlrpc_request  *req;
2017         char                   *tmp;
2018         int                     rc = -EINVAL;
2019         ENTRY;
2020
2021         req = ptlrpc_request_alloc(imp, &RQF_MDS_GET_INFO);
2022         if (req == NULL)
2023                 RETURN(-ENOMEM);
2024
2025         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY,
2026                              RCL_CLIENT, keylen);
2027         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VALLEN,
2028                              RCL_CLIENT, sizeof(vallen));
2029
2030         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_INFO);
2031         if (rc) {
2032                 ptlrpc_request_free(req);
2033                 RETURN(rc);
2034         }
2035
2036         tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
2037         memcpy(tmp, key, keylen);
2038         tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_VALLEN);
2039         memcpy(tmp, &vallen, sizeof(vallen));
2040
2041         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VAL,
2042                              RCL_SERVER, vallen);
2043         ptlrpc_request_set_replen(req);
2044
2045         rc = ptlrpc_queue_wait(req);
2046         /* -EREMOTE means the get_info result is partial, and it needs to
2047          * continue on another MDT, see fid2path part in lmv_iocontrol */
2048         if (rc == 0 || rc == -EREMOTE) {
2049                 tmp = req_capsule_server_get(&req->rq_pill, &RMF_GETINFO_VAL);
2050                 memcpy(val, tmp, vallen);
2051                 if (ptlrpc_rep_need_swab(req)) {
2052                         if (KEY_IS(KEY_FID2PATH))
2053                                 lustre_swab_fid2path(val);
2054                 }
2055         }
2056         ptlrpc_req_finished(req);
2057
2058         RETURN(rc);
2059 }
2060
2061 static void lustre_swab_hai(struct hsm_action_item *h)
2062 {
2063         __swab32s(&h->hai_len);
2064         __swab32s(&h->hai_action);
2065         lustre_swab_lu_fid(&h->hai_fid);
2066         lustre_swab_lu_fid(&h->hai_dfid);
2067         __swab64s(&h->hai_cookie);
2068         __swab64s(&h->hai_extent.offset);
2069         __swab64s(&h->hai_extent.length);
2070         __swab64s(&h->hai_gid);
2071 }
2072
2073 static void lustre_swab_hal(struct hsm_action_list *h)
2074 {
2075         struct hsm_action_item  *hai;
2076         __u32                    i;
2077
2078         __swab32s(&h->hal_version);
2079         __swab32s(&h->hal_count);
2080         __swab32s(&h->hal_archive_id);
2081         __swab64s(&h->hal_flags);
2082         hai = hai_first(h);
2083         for (i = 0; i < h->hal_count; i++, hai = hai_next(hai))
2084                 lustre_swab_hai(hai);
2085 }
2086
2087 static void lustre_swab_kuch(struct kuc_hdr *l)
2088 {
2089         __swab16s(&l->kuc_magic);
2090         /* __u8 l->kuc_transport */
2091         __swab16s(&l->kuc_msgtype);
2092         __swab16s(&l->kuc_msglen);
2093 }
2094
2095 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2096                                 struct lustre_kernelcomm *lk)
2097 {
2098         struct obd_import  *imp = class_exp2cliimp(exp);
2099         __u32               archive = lk->lk_data;
2100         int                 rc = 0;
2101
2102         if (lk->lk_group != KUC_GRP_HSM) {
2103                 CERROR("Bad copytool group %d\n", lk->lk_group);
2104                 return -EINVAL;
2105         }
2106
2107         CDEBUG(D_HSM, "CT start r%d w%d u%d g%d f%#x\n", lk->lk_rfd, lk->lk_wfd,
2108                lk->lk_uid, lk->lk_group, lk->lk_flags);
2109
2110         if (lk->lk_flags & LK_FLG_STOP) {
2111                 /* Unregister with the coordinator */
2112                 rc = mdc_ioc_hsm_ct_unregister(imp);
2113         } else {
2114                 rc = mdc_ioc_hsm_ct_register(imp, archive);
2115         }
2116
2117         return rc;
2118 }
2119
2120 /**
2121  * Send a message to any listening copytools
2122  * @param val KUC message (kuc_hdr + hsm_action_list)
2123  * @param len total length of message
2124  */
2125 static int mdc_hsm_copytool_send(size_t len, void *val)
2126 {
2127         struct kuc_hdr          *lh = (struct kuc_hdr *)val;
2128         struct hsm_action_list  *hal = (struct hsm_action_list *)(lh + 1);
2129         int                      rc;
2130         ENTRY;
2131
2132         if (len < sizeof(*lh) + sizeof(*hal)) {
2133                 CERROR("Short HSM message %zu < %zu\n", len,
2134                        sizeof(*lh) + sizeof(*hal));
2135                 RETURN(-EPROTO);
2136         }
2137         if (lh->kuc_magic == __swab16(KUC_MAGIC)) {
2138                 lustre_swab_kuch(lh);
2139                 lustre_swab_hal(hal);
2140         } else if (lh->kuc_magic != KUC_MAGIC) {
2141                 CERROR("Bad magic %x!=%x\n", lh->kuc_magic, KUC_MAGIC);
2142                 RETURN(-EPROTO);
2143         }
2144
2145         CDEBUG(D_HSM, " Received message mg=%x t=%d m=%d l=%d actions=%d "
2146                "on %s\n",
2147                lh->kuc_magic, lh->kuc_transport, lh->kuc_msgtype,
2148                lh->kuc_msglen, hal->hal_count, hal->hal_fsname);
2149
2150         /* Broadcast to HSM listeners */
2151         rc = libcfs_kkuc_group_put(KUC_GRP_HSM, lh);
2152
2153         RETURN(rc);
2154 }
2155
2156 /**
2157  * callback function passed to kuc for re-registering each HSM copytool
2158  * running on MDC, after MDT shutdown/recovery.
2159  * @param data copytool registration data
2160  * @param cb_arg callback argument (obd_import)
2161  */
2162 static int mdc_hsm_ct_reregister(void *data, void *cb_arg)
2163 {
2164         struct kkuc_ct_data     *kcd = data;
2165         struct obd_import       *imp = (struct obd_import *)cb_arg;
2166         int                      rc;
2167
2168         if (kcd == NULL || kcd->kcd_magic != KKUC_CT_DATA_MAGIC)
2169                 return -EPROTO;
2170
2171         if (!obd_uuid_equals(&kcd->kcd_uuid, &imp->imp_obd->obd_uuid))
2172                 return 0;
2173
2174         CDEBUG(D_HA, "%s: recover copytool registration to MDT (archive=%#x)\n",
2175                imp->imp_obd->obd_name, kcd->kcd_archive);
2176         rc = mdc_ioc_hsm_ct_register(imp, kcd->kcd_archive);
2177
2178         /* ignore error if the copytool is already registered */
2179         return (rc == -EEXIST) ? 0 : rc;
2180 }
2181
2182 /**
2183  * Re-establish all kuc contexts with MDT
2184  * after MDT shutdown/recovery.
2185  */
2186 static int mdc_kuc_reregister(struct obd_import *imp)
2187 {
2188         /* re-register HSM agents */
2189         return libcfs_kkuc_group_foreach(KUC_GRP_HSM, mdc_hsm_ct_reregister,
2190                                          (void *)imp);
2191 }
2192
2193 static int mdc_set_info_async(const struct lu_env *env,
2194                               struct obd_export *exp,
2195                               u32 keylen, void *key,
2196                               u32 vallen, void *val,
2197                               struct ptlrpc_request_set *set)
2198 {
2199         struct obd_import       *imp = class_exp2cliimp(exp);
2200         int                      rc;
2201         ENTRY;
2202
2203         if (KEY_IS(KEY_READ_ONLY)) {
2204                 if (vallen != sizeof(int))
2205                         RETURN(-EINVAL);
2206
2207                 spin_lock(&imp->imp_lock);
2208                 if (*((int *)val)) {
2209                         imp->imp_connect_flags_orig |= OBD_CONNECT_RDONLY;
2210                         imp->imp_connect_data.ocd_connect_flags |=
2211                                                         OBD_CONNECT_RDONLY;
2212                 } else {
2213                         imp->imp_connect_flags_orig &= ~OBD_CONNECT_RDONLY;
2214                         imp->imp_connect_data.ocd_connect_flags &=
2215                                                         ~OBD_CONNECT_RDONLY;
2216                 }
2217                 spin_unlock(&imp->imp_lock);
2218
2219                 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2220                                        keylen, key, vallen, val, set);
2221                 RETURN(rc);
2222         }
2223         if (KEY_IS(KEY_SPTLRPC_CONF)) {
2224                 sptlrpc_conf_client_adapt(exp->exp_obd);
2225                 RETURN(0);
2226         }
2227         if (KEY_IS(KEY_FLUSH_CTX)) {
2228                 sptlrpc_import_flush_my_ctx(imp);
2229                 RETURN(0);
2230         }
2231         if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2232                 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2233                                        keylen, key, vallen, val, set);
2234                 RETURN(rc);
2235         }
2236         if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) {
2237                 rc = mdc_hsm_copytool_send(vallen, val);
2238                 RETURN(rc);
2239         }
2240
2241         if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2242                 __u32 *default_easize = val;
2243
2244                 exp->exp_obd->u.cli.cl_default_mds_easize = *default_easize;
2245                 RETURN(0);
2246         }
2247
2248         CERROR("Unknown key %s\n", (char *)key);
2249         RETURN(-EINVAL);
2250 }
2251
2252 static int mdc_get_info(const struct lu_env *env, struct obd_export *exp,
2253                         __u32 keylen, void *key, __u32 *vallen, void *val)
2254 {
2255         int rc = -EINVAL;
2256
2257         if (KEY_IS(KEY_MAX_EASIZE)) {
2258                 __u32 mdsize, *max_easize;
2259
2260                 if (*vallen != sizeof(int))
2261                         RETURN(-EINVAL);
2262                 mdsize = *(__u32 *)val;
2263                 if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
2264                         exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
2265                 max_easize = val;
2266                 *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
2267                 RETURN(0);
2268         } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2269                 __u32 *default_easize;
2270
2271                 if (*vallen != sizeof(int))
2272                         RETURN(-EINVAL);
2273                 default_easize = val;
2274                 *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
2275                 RETURN(0);
2276         } else if (KEY_IS(KEY_CONN_DATA)) {
2277                 struct obd_import *imp = class_exp2cliimp(exp);
2278                 struct obd_connect_data *data = val;
2279
2280                 if (*vallen != sizeof(*data))
2281                         RETURN(-EINVAL);
2282
2283                 *data = imp->imp_connect_data;
2284                 RETURN(0);
2285         } else if (KEY_IS(KEY_TGT_COUNT)) {
2286                 *((__u32 *)val) = 1;
2287                 RETURN(0);
2288         }
2289
2290         rc = mdc_get_info_rpc(exp, keylen, key, *vallen, val);
2291
2292         RETURN(rc);
2293 }
2294
2295 static int mdc_fsync(struct obd_export *exp, const struct lu_fid *fid,
2296                      struct ptlrpc_request **request)
2297 {
2298         struct ptlrpc_request *req;
2299         int                    rc;
2300         ENTRY;
2301
2302         *request = NULL;
2303         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SYNC);
2304         if (req == NULL)
2305                 RETURN(-ENOMEM);
2306
2307         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SYNC);
2308         if (rc) {
2309                 ptlrpc_request_free(req);
2310                 RETURN(rc);
2311         }
2312
2313         mdc_pack_body(req, fid, 0, 0, -1, 0);
2314
2315         ptlrpc_request_set_replen(req);
2316
2317         rc = ptlrpc_queue_wait(req);
2318         if (rc)
2319                 ptlrpc_req_finished(req);
2320         else
2321                 *request = req;
2322         RETURN(rc);
2323 }
2324
2325 static int mdc_import_event(struct obd_device *obd, struct obd_import *imp,
2326                             enum obd_import_event event)
2327 {
2328         int rc = 0;
2329
2330         LASSERT(imp->imp_obd == obd);
2331
2332         switch (event) {
2333
2334         case IMP_EVENT_INACTIVE: {
2335                 struct client_obd *cli = &obd->u.cli;
2336                 /*
2337                  * Flush current sequence to make client obtain new one
2338                  * from server in case of disconnect/reconnect.
2339                  */
2340                 down_read(&cli->cl_seq_rwsem);
2341                 if (cli->cl_seq)
2342                         seq_client_flush(cli->cl_seq);
2343                 up_read(&cli->cl_seq_rwsem);
2344
2345                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
2346                 break;
2347         }
2348         case IMP_EVENT_INVALIDATE: {
2349                 struct ldlm_namespace *ns = obd->obd_namespace;
2350
2351                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2352
2353                 break;
2354         }
2355         case IMP_EVENT_ACTIVE:
2356                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
2357                 /* redo the kuc registration after reconnecting */
2358                 if (rc == 0)
2359                         rc = mdc_kuc_reregister(imp);
2360                 break;
2361         case IMP_EVENT_OCD:
2362                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
2363                 break;
2364         case IMP_EVENT_DISCON:
2365         case IMP_EVENT_DEACTIVATE:
2366         case IMP_EVENT_ACTIVATE:
2367                 break;
2368         default:
2369                 CERROR("Unknown import event %x\n", event);
2370                 LBUG();
2371         }
2372         RETURN(rc);
2373 }
2374
2375 int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp,
2376                   struct lu_fid *fid, struct md_op_data *op_data)
2377 {
2378         struct client_obd *cli = &exp->exp_obd->u.cli;
2379         int rc = -EIO;
2380
2381         ENTRY;
2382
2383         down_read(&cli->cl_seq_rwsem);
2384         if (cli->cl_seq)
2385                 rc = seq_client_alloc_fid(env, cli->cl_seq, fid);
2386         up_read(&cli->cl_seq_rwsem);
2387
2388         RETURN(rc);
2389 }
2390
2391 static struct obd_uuid *mdc_get_uuid(struct obd_export *exp)
2392 {
2393         struct client_obd *cli = &exp->exp_obd->u.cli;
2394         return &cli->cl_target_uuid;
2395 }
2396
2397 /**
2398  * Determine whether the lock can be canceled before replaying it during
2399  * recovery, non zero value will be return if the lock can be canceled,
2400  * or zero returned for not
2401  */
2402 static int mdc_cancel_weight(struct ldlm_lock *lock)
2403 {
2404         if (lock->l_resource->lr_type != LDLM_IBITS)
2405                 RETURN(0);
2406
2407         /* FIXME: if we ever get into a situation where there are too many
2408          * opened files with open locks on a single node, then we really
2409          * should replay these open locks to reget it */
2410         if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN)
2411                 RETURN(0);
2412
2413         RETURN(1);
2414 }
2415
2416 static int mdc_resource_inode_free(struct ldlm_resource *res)
2417 {
2418         if (res->lr_lvb_inode)
2419                 res->lr_lvb_inode = NULL;
2420
2421         return 0;
2422 }
2423
2424 static struct ldlm_valblock_ops inode_lvbo = {
2425         .lvbo_free = mdc_resource_inode_free
2426 };
2427
2428 static int mdc_llog_init(struct obd_device *obd)
2429 {
2430         struct obd_llog_group   *olg = &obd->obd_olg;
2431         struct llog_ctxt        *ctxt;
2432         int                      rc;
2433
2434         ENTRY;
2435
2436         rc = llog_setup(NULL, obd, olg, LLOG_CHANGELOG_REPL_CTXT, obd,
2437                         &llog_client_ops);
2438         if (rc < 0)
2439                 RETURN(rc);
2440
2441         ctxt = llog_group_get_ctxt(olg, LLOG_CHANGELOG_REPL_CTXT);
2442         llog_initiator_connect(ctxt);
2443         llog_ctxt_put(ctxt);
2444
2445         RETURN(0);
2446 }
2447
2448 static void mdc_llog_finish(struct obd_device *obd)
2449 {
2450         struct llog_ctxt *ctxt;
2451
2452         ENTRY;
2453
2454         ctxt = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
2455         if (ctxt != NULL)
2456                 llog_cleanup(NULL, ctxt);
2457
2458         EXIT;
2459 }
2460
2461 static int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg)
2462 {
2463         int                             rc;
2464         ENTRY;
2465
2466         rc = ptlrpcd_addref();
2467         if (rc < 0)
2468                 RETURN(rc);
2469
2470         rc = client_obd_setup(obd, cfg);
2471         if (rc)
2472                 GOTO(err_ptlrpcd_decref, rc);
2473 #ifdef CONFIG_PROC_FS
2474         obd->obd_vars = lprocfs_mdc_obd_vars;
2475         lprocfs_obd_setup(obd);
2476         lprocfs_alloc_md_stats(obd, 0);
2477 #endif
2478         sptlrpc_lprocfs_cliobd_attach(obd);
2479         ptlrpc_lprocfs_register_obd(obd);
2480
2481         ns_register_cancel(obd->obd_namespace, mdc_cancel_weight);
2482
2483         obd->obd_namespace->ns_lvbo = &inode_lvbo;
2484
2485         rc = mdc_llog_init(obd);
2486         if (rc) {
2487                 CERROR("%s: failed to setup llogging subsystems: rc = %d\n",
2488                        obd->obd_name, rc);
2489                 GOTO(err_mdc_cleanup, rc);
2490         }
2491
2492         rc = mdc_changelog_cdev_init(obd);
2493         if (rc) {
2494                 CERROR("%s: failed to setup changelog char device: rc = %d\n",
2495                        obd->obd_name, rc);
2496                 GOTO(err_mdc_cleanup, rc);
2497         }
2498
2499         EXIT;
2500 err_mdc_cleanup:
2501         if (rc)
2502                 client_obd_cleanup(obd);
2503
2504 err_ptlrpcd_decref:
2505         if (rc)
2506                 ptlrpcd_decref();
2507
2508         return rc;
2509 }
2510
2511 /* Initialize the default and maximum LOV EA sizes.  This allows
2512  * us to make MDS RPCs with large enough reply buffers to hold a default
2513  * sized EA without having to calculate this (via a call into the
2514  * LOV + OSCs) each time we make an RPC.  The maximum size is also tracked
2515  * but not used to avoid wastefully vmalloc()'ing large reply buffers when
2516  * a large number of stripes is possible.  If a larger reply buffer is
2517  * required it will be reallocated in the ptlrpc layer due to overflow.
2518  */
2519 static int mdc_init_ea_size(struct obd_export *exp, __u32 easize,
2520                             __u32 def_easize)
2521 {
2522         struct obd_device *obd = exp->exp_obd;
2523         struct client_obd *cli = &obd->u.cli;
2524         ENTRY;
2525
2526         if (cli->cl_max_mds_easize < easize)
2527                 cli->cl_max_mds_easize = easize;
2528
2529         if (cli->cl_default_mds_easize < def_easize)
2530                 cli->cl_default_mds_easize = def_easize;
2531
2532         RETURN(0);
2533 }
2534
2535 static int mdc_precleanup(struct obd_device *obd)
2536 {
2537         ENTRY;
2538
2539         /* Failsafe, ok if racy */
2540         if (obd->obd_type->typ_refcnt <= 1)
2541                 libcfs_kkuc_group_rem(0, KUC_GRP_HSM);
2542
2543         mdc_changelog_cdev_finish(obd);
2544
2545         obd_cleanup_client_import(obd);
2546         ptlrpc_lprocfs_unregister_obd(obd);
2547         lprocfs_obd_cleanup(obd);
2548         lprocfs_free_md_stats(obd);
2549         mdc_llog_finish(obd);
2550         RETURN(0);
2551 }
2552
2553 static int mdc_cleanup(struct obd_device *obd)
2554 {
2555         ptlrpcd_decref();
2556
2557         return client_obd_cleanup(obd);
2558 }
2559
2560 static int mdc_process_config(struct obd_device *obd, size_t len, void *buf)
2561 {
2562         struct lustre_cfg *lcfg = buf;
2563         int rc = class_process_proc_param(PARAM_MDC, obd->obd_vars, lcfg, obd);
2564         return (rc > 0 ? 0: rc);
2565 }
2566
2567 static struct obd_ops mdc_obd_ops = {
2568         .o_owner            = THIS_MODULE,
2569         .o_setup            = mdc_setup,
2570         .o_precleanup       = mdc_precleanup,
2571         .o_cleanup          = mdc_cleanup,
2572         .o_add_conn         = client_import_add_conn,
2573         .o_del_conn         = client_import_del_conn,
2574         .o_connect          = client_connect_import,
2575         .o_disconnect       = client_disconnect_export,
2576         .o_iocontrol        = mdc_iocontrol,
2577         .o_set_info_async   = mdc_set_info_async,
2578         .o_statfs           = mdc_statfs,
2579         .o_fid_init         = client_fid_init,
2580         .o_fid_fini         = client_fid_fini,
2581         .o_fid_alloc        = mdc_fid_alloc,
2582         .o_import_event     = mdc_import_event,
2583         .o_get_info         = mdc_get_info,
2584         .o_process_config   = mdc_process_config,
2585         .o_get_uuid         = mdc_get_uuid,
2586         .o_quotactl         = mdc_quotactl,
2587 };
2588
2589 static struct md_ops mdc_md_ops = {
2590         .m_get_root         = mdc_get_root,
2591         .m_null_inode       = mdc_null_inode,
2592         .m_close            = mdc_close,
2593         .m_create           = mdc_create,
2594         .m_enqueue          = mdc_enqueue,
2595         .m_getattr          = mdc_getattr,
2596         .m_getattr_name     = mdc_getattr_name,
2597         .m_intent_lock      = mdc_intent_lock,
2598         .m_link             = mdc_link,
2599         .m_rename           = mdc_rename,
2600         .m_setattr          = mdc_setattr,
2601         .m_setxattr         = mdc_setxattr,
2602         .m_getxattr         = mdc_getxattr,
2603         .m_fsync                = mdc_fsync,
2604         .m_read_page            = mdc_read_page,
2605         .m_unlink           = mdc_unlink,
2606         .m_cancel_unused    = mdc_cancel_unused,
2607         .m_init_ea_size     = mdc_init_ea_size,
2608         .m_set_lock_data    = mdc_set_lock_data,
2609         .m_lock_match       = mdc_lock_match,
2610         .m_get_lustre_md    = mdc_get_lustre_md,
2611         .m_free_lustre_md   = mdc_free_lustre_md,
2612         .m_set_open_replay_data = mdc_set_open_replay_data,
2613         .m_clear_open_replay_data = mdc_clear_open_replay_data,
2614         .m_intent_getattr_async = mdc_intent_getattr_async,
2615         .m_revalidate_lock      = mdc_revalidate_lock
2616 };
2617
2618 static int __init mdc_init(void)
2619 {
2620         return class_register_type(&mdc_obd_ops, &mdc_md_ops, true, NULL,
2621                                    LUSTRE_MDC_NAME, NULL);
2622 }
2623
2624 static void __exit mdc_exit(void)
2625 {
2626         class_unregister_type(LUSTRE_MDC_NAME);
2627 }
2628
2629 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2630 MODULE_DESCRIPTION("Lustre Metadata Client");
2631 MODULE_VERSION(LUSTRE_VERSION_STRING);
2632 MODULE_LICENSE("GPL");
2633
2634 module_init(mdc_init);
2635 module_exit(mdc_exit);