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, 2017, 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_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/lustre_ioctl.h>
  53 #include <lustre_kernelcomm.h>
  54 #include <lustre_lmv.h>
  55 #include <lustre_log.h>
  56 #include <uapi/linux/lustre/lustre_param.h>
  57 #include <lustre_swab.h>
  58 #include <obd_class.h>
  59 #include <lustre_osc.h>
  60
  61 #include "mdc_internal.h"
  62
  63 #define REQUEST_MINOR 244
  64
  65 static int mdc_cleanup(struct obd_device *obd);
  66
  67 static inline int mdc_queue_wait(struct ptlrpc_request *req)
  68 {
  69         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
  70         int rc;
  71
  72         /* obd_get_request_slot() ensures that this client has no more
  73          * than cl_max_rpcs_in_flight RPCs simultaneously inf light
  74          * against an MDT. */
  75         rc = obd_get_request_slot(cli);
  76         if (rc != 0)
  77                 return rc;
  78
  79         rc = ptlrpc_queue_wait(req);
  80         obd_put_request_slot(cli);
  81
  82         return rc;
  83 }
  84
  85 /*
  86  * Send MDS_GET_ROOT RPC to fetch root FID.
  87  *
  88  * If \a fileset is not NULL it should contain a subdirectory off
  89  * the ROOT/ directory to be mounted on the client. Return the FID
  90  * of the subdirectory to the client to mount onto its mountpoint.
  91  *
  92  * \param[in]   imp     MDC import
  93  * \param[in]   fileset fileset name, which could be NULL
  94  * \param[out]  rootfid root FID of this mountpoint
  95  * \param[out]  pc      root capa will be unpacked and saved in this pointer
  96  *
  97  * \retval      0 on success, negative errno on failure
  98  */
  99 static int mdc_get_root(struct obd_export *exp, const char *fileset,
 100                          struct lu_fid *rootfid)
 101 {
 102         struct ptlrpc_request   *req;
 103         struct mdt_body         *body;
 104         int                      rc;
 105
 106         ENTRY;
 107
 108         if (fileset && !(exp_connect_flags(exp) & OBD_CONNECT_SUBTREE))
 109                 RETURN(-ENOTSUPP);
 110
 111         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 112                                 &RQF_MDS_GET_ROOT);
 113         if (req == NULL)
 114                 RETURN(-ENOMEM);
 115
 116         if (fileset != NULL)
 117                 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 118                                      strlen(fileset) + 1);
 119         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_ROOT);
 120         if (rc) {
 121                 ptlrpc_request_free(req);
 122                 RETURN(rc);
 123         }
 124         mdc_pack_body(req, NULL, 0, 0, -1, 0);
 125         if (fileset != NULL) {
 126                 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
 127
 128                 memcpy(name, fileset, strlen(fileset));
 129         }
 130         lustre_msg_add_flags(req->rq_reqmsg, LUSTRE_IMP_FULL);
 131         req->rq_send_state = LUSTRE_IMP_FULL;
 132
 133         ptlrpc_request_set_replen(req);
 134
 135         rc = ptlrpc_queue_wait(req);
 136         if (rc)
 137                 GOTO(out, rc);
 138
 139         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 140         if (body == NULL)
 141                 GOTO(out, rc = -EPROTO);
 142
 143         *rootfid = body->mbo_fid1;
 144         CDEBUG(D_NET, "root fid="DFID", last_committed=%llu\n",
 145                PFID(rootfid), lustre_msg_get_last_committed(req->rq_repmsg));
 146         EXIT;
 147 out:
 148         ptlrpc_req_finished(req);
 149
 150         return rc;
 151 }
 152
 153 /*
 154  * This function now is known to always saying that it will receive 4 buffers
 155  * from server. Even for cases when acl_size and md_size is zero, RPC header
 156  * will contain 4 fields and RPC itself will contain zero size fields. This is
 157  * because mdt_getattr*() _always_ returns 4 fields, but if acl is not needed
 158  * and thus zero, it shrinks it, making zero size. The same story about
 159  * md_size. And this is course of problem when client waits for smaller number
 160  * of fields. This issue will be fixed later when client gets aware of RPC
 161  * layouts.  --umka
 162  */
 163 static int mdc_getattr_common(struct obd_export *exp,
 164                               struct ptlrpc_request *req)
 165 {
 166         struct req_capsule *pill = &req->rq_pill;
 167         struct mdt_body    *body;
 168         void               *eadata;
 169         int                 rc;
 170         ENTRY;
 171
 172         /* Request message already built. */
 173         rc = ptlrpc_queue_wait(req);
 174         if (rc != 0)
 175                 RETURN(rc);
 176
 177         /* sanity check for the reply */
 178         body = req_capsule_server_get(pill, &RMF_MDT_BODY);
 179         if (body == NULL)
 180                 RETURN(-EPROTO);
 181
 182         CDEBUG(D_NET, "mode: %o\n", body->mbo_mode);
 183
 184         mdc_update_max_ea_from_body(exp, body);
 185         if (body->mbo_eadatasize != 0) {
 186                 eadata = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
 187                                                       body->mbo_eadatasize);
 188                 if (eadata == NULL)
 189                         RETURN(-EPROTO);
 190         }
 191
 192         RETURN(0);
 193 }
 194
 195 static void mdc_reset_acl_req(struct ptlrpc_request *req)
 196 {
 197         spin_lock(&req->rq_early_free_lock);
 198         sptlrpc_cli_free_repbuf(req);
 199         req->rq_repbuf = NULL;
 200         req->rq_repbuf_len = 0;
 201         req->rq_repdata = NULL;
 202         req->rq_reqdata_len = 0;
 203         spin_unlock(&req->rq_early_free_lock);
 204 }
 205
 206 static int mdc_getattr(struct obd_export *exp, struct md_op_data *op_data,
 207                        struct ptlrpc_request **request)
 208 {
 209         struct ptlrpc_request *req;
 210         struct obd_import *imp = class_exp2cliimp(exp);
 211         __u32 acl_bufsize = LUSTRE_POSIX_ACL_MAX_SIZE_OLD;
 212         int rc;
 213         ENTRY;
 214
 215         /* Single MDS without an LMV case */
 216         if (op_data->op_flags & MF_GET_MDT_IDX) {
 217                 op_data->op_mds = 0;
 218                 RETURN(0);
 219         }
 220
 221         *request = NULL;
 222         req = ptlrpc_request_alloc(imp, &RQF_MDS_GETATTR);
 223         if (req == NULL)
 224                 RETURN(-ENOMEM);
 225
 226         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR);
 227         if (rc) {
 228                 ptlrpc_request_free(req);
 229                 RETURN(rc);
 230         }
 231
 232 again:
 233         mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
 234                       op_data->op_mode, -1, 0);
 235         req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
 236         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 237                              op_data->op_mode);
 238         ptlrpc_request_set_replen(req);
 239
 240         rc = mdc_getattr_common(exp, req);
 241         if (rc) {
 242                 if (rc == -ERANGE &&
 243                     acl_bufsize != imp->imp_connect_data.ocd_max_easize) {
 244                         acl_bufsize = imp->imp_connect_data.ocd_max_easize;
 245                         mdc_reset_acl_req(req);
 246                         goto again;
 247                 }
 248
 249                 ptlrpc_req_finished(req);
 250         } else {
 251                 *request = req;
 252         }
 253
 254         RETURN(rc);
 255 }
 256
 257 static int mdc_getattr_name(struct obd_export *exp, struct md_op_data *op_data,
 258                             struct ptlrpc_request **request)
 259 {
 260         struct ptlrpc_request *req;
 261         struct obd_import *imp = class_exp2cliimp(exp);
 262         __u32 acl_bufsize = LUSTRE_POSIX_ACL_MAX_SIZE_OLD;
 263         int rc;
 264         ENTRY;
 265
 266         *request = NULL;
 267         req = ptlrpc_request_alloc(imp, &RQF_MDS_GETATTR_NAME);
 268         if (req == NULL)
 269                 RETURN(-ENOMEM);
 270
 271         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 272                              op_data->op_namelen + 1);
 273
 274         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR_NAME);
 275         if (rc) {
 276                 ptlrpc_request_free(req);
 277                 RETURN(rc);
 278         }
 279
 280         if (op_data->op_name) {
 281                 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
 282                 LASSERT(strnlen(op_data->op_name, op_data->op_namelen) ==
 283                                 op_data->op_namelen);
 284                 memcpy(name, op_data->op_name, op_data->op_namelen);
 285         }
 286
 287 again:
 288         mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
 289                       op_data->op_mode, op_data->op_suppgids[0], 0);
 290         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 291                              op_data->op_mode);
 292         req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
 293         ptlrpc_request_set_replen(req);
 294
 295         rc = mdc_getattr_common(exp, req);
 296         if (rc) {
 297                 if (rc == -ERANGE &&
 298                     acl_bufsize != imp->imp_connect_data.ocd_max_easize) {
 299                         acl_bufsize = imp->imp_connect_data.ocd_max_easize;
 300                         mdc_reset_acl_req(req);
 301                         goto again;
 302                 }
 303
 304                 ptlrpc_req_finished(req);
 305         } else {
 306                 *request = req;
 307         }
 308
 309         RETURN(rc);
 310 }
 311
 312 static int mdc_xattr_common(struct obd_export *exp,const struct req_format *fmt,
 313                             const struct lu_fid *fid, int opcode, u64 valid,
 314                             const char *xattr_name, const char *input,
 315                             int input_size, int output_size, int flags,
 316                             __u32 suppgid, struct ptlrpc_request **request)
 317 {
 318         struct ptlrpc_request *req;
 319         int   xattr_namelen = 0;
 320         char *tmp;
 321         int   rc;
 322         ENTRY;
 323
 324         *request = NULL;
 325         req = ptlrpc_request_alloc(class_exp2cliimp(exp), fmt);
 326         if (req == NULL)
 327                 RETURN(-ENOMEM);
 328
 329         if (xattr_name) {
 330                 xattr_namelen = strlen(xattr_name) + 1;
 331                 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 332                                      xattr_namelen);
 333         }
 334         if (input_size) {
 335                 LASSERT(input);
 336                 req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT,
 337                                      input_size);
 338         }
 339
 340         /* Flush local XATTR locks to get rid of a possible cancel RPC */
 341         if (opcode == MDS_REINT && fid_is_sane(fid) &&
 342             exp->exp_connect_data.ocd_ibits_known & MDS_INODELOCK_XATTR) {
 343                 struct list_head cancels = LIST_HEAD_INIT(cancels);
 344                 int count;
 345
 346                 /* Without that packing would fail */
 347                 if (input_size == 0)
 348                         req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
 349                                              RCL_CLIENT, 0);
 350
 351                 count = mdc_resource_get_unused(exp, fid,
 352                                                 &cancels, LCK_EX,
 353                                                 MDS_INODELOCK_XATTR);
 354
 355                 rc = mdc_prep_elc_req(exp, req, MDS_REINT, &cancels, count);
 356                 if (rc) {
 357                         ptlrpc_request_free(req);
 358                         RETURN(rc);
 359                 }
 360         } else {
 361                 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, opcode);
 362                 if (rc) {
 363                         ptlrpc_request_free(req);
 364                         RETURN(rc);
 365                 }
 366         }
 367
 368         if (opcode == MDS_REINT) {
 369                 struct mdt_rec_setxattr *rec;
 370
 371                 CLASSERT(sizeof(struct mdt_rec_setxattr) ==
 372                          sizeof(struct mdt_rec_reint));
 373                 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
 374                 rec->sx_opcode = REINT_SETXATTR;
 375                 rec->sx_fsuid  = from_kuid(&init_user_ns, current_fsuid());
 376                 rec->sx_fsgid  = from_kgid(&init_user_ns, current_fsgid());
 377                 rec->sx_cap    = cfs_curproc_cap_pack();
 378                 rec->sx_suppgid1 = suppgid;
 379                 rec->sx_suppgid2 = -1;
 380                 rec->sx_fid    = *fid;
 381                 rec->sx_valid  = valid | OBD_MD_FLCTIME;
 382                 rec->sx_time   = ktime_get_real_seconds();
 383                 rec->sx_size   = output_size;
 384                 rec->sx_flags  = flags;
 385         } else {
 386                 mdc_pack_body(req, fid, valid, output_size, suppgid, flags);
 387         }
 388
 389         if (xattr_name) {
 390                 tmp = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
 391                 memcpy(tmp, xattr_name, xattr_namelen);
 392         }
 393         if (input_size) {
 394                 tmp = req_capsule_client_get(&req->rq_pill, &RMF_EADATA);
 395                 memcpy(tmp, input, input_size);
 396         }
 397
 398         if (req_capsule_has_field(&req->rq_pill, &RMF_EADATA, RCL_SERVER))
 399                 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
 400                                      RCL_SERVER, output_size);
 401         ptlrpc_request_set_replen(req);
 402
 403         /* make rpc */
 404         if (opcode == MDS_REINT)
 405                 mdc_get_mod_rpc_slot(req, NULL);
 406
 407         rc = ptlrpc_queue_wait(req);
 408
 409         if (opcode == MDS_REINT)
 410                 mdc_put_mod_rpc_slot(req, NULL);
 411
 412         if (rc)
 413                 ptlrpc_req_finished(req);
 414         else
 415                 *request = req;
 416         RETURN(rc);
 417 }
 418
 419 static int mdc_setxattr(struct obd_export *exp, const struct lu_fid *fid,
 420                         u64 valid, const char *xattr_name,
 421                         const char *input, int input_size, int output_size,
 422                         int flags, __u32 suppgid,
 423                         struct ptlrpc_request **request)
 424 {
 425         return mdc_xattr_common(exp, &RQF_MDS_REINT_SETXATTR,
 426                                 fid, MDS_REINT, valid, xattr_name,
 427                                 input, input_size, output_size, flags,
 428                                 suppgid, request);
 429 }
 430
 431 static int mdc_getxattr(struct obd_export *exp, const struct lu_fid *fid,
 432                         u64 valid, const char *xattr_name,
 433                         const char *input, int input_size, int output_size,
 434                         int flags, struct ptlrpc_request **request)
 435 {
 436         return mdc_xattr_common(exp, &RQF_MDS_GETXATTR,
 437                                 fid, MDS_GETXATTR, valid, xattr_name,
 438                                 input, input_size, output_size, flags,
 439                                 -1, request);
 440 }
 441
 442 #ifdef CONFIG_FS_POSIX_ACL
 443 static int mdc_unpack_acl(struct ptlrpc_request *req, struct lustre_md *md)
 444 {
 445         struct req_capsule     *pill = &req->rq_pill;
 446         struct mdt_body        *body = md->body;
 447         struct posix_acl       *acl;
 448         void                   *buf;
 449         int                     rc;
 450         ENTRY;
 451
 452         if (!body->mbo_aclsize)
 453                 RETURN(0);
 454
 455         buf = req_capsule_server_sized_get(pill, &RMF_ACL, body->mbo_aclsize);
 456
 457         if (!buf)
 458                 RETURN(-EPROTO);
 459
 460         acl = posix_acl_from_xattr(&init_user_ns, buf, body->mbo_aclsize);
 461         if (acl == NULL)
 462                 RETURN(0);
 463         if (IS_ERR(acl)) {
 464                 rc = PTR_ERR(acl);
 465                 CERROR("convert xattr to acl: %d\n", rc);
 466                 RETURN(rc);
 467         }
 468
 469         rc = posix_acl_valid(&init_user_ns, acl);
 470         if (rc) {
 471                 CERROR("validate acl: %d\n", rc);
 472                 posix_acl_release(acl);
 473                 RETURN(rc);
 474         }
 475
 476         md->posix_acl = acl;
 477         RETURN(0);
 478 }
 479 #else
 480 #define mdc_unpack_acl(req, md) 0
 481 #endif
 482
 483 int mdc_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
 484                       struct obd_export *dt_exp, struct obd_export *md_exp,
 485                       struct lustre_md *md)
 486 {
 487         struct req_capsule *pill = &req->rq_pill;
 488         int rc;
 489         ENTRY;
 490
 491         LASSERT(md);
 492         memset(md, 0, sizeof(*md));
 493
 494         md->body = req_capsule_server_get(pill, &RMF_MDT_BODY);
 495         LASSERT(md->body != NULL);
 496
 497         if (md->body->mbo_valid & OBD_MD_FLEASIZE) {
 498                 if (!S_ISREG(md->body->mbo_mode)) {
 499                         CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, should be a "
 500                                "regular file, but is not\n");
 501                         GOTO(out, rc = -EPROTO);
 502                 }
 503
 504                 if (md->body->mbo_eadatasize == 0) {
 505                         CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, "
 506                                "but eadatasize 0\n");
 507                         GOTO(out, rc = -EPROTO);
 508                 }
 509
 510                 md->layout.lb_len = md->body->mbo_eadatasize;
 511                 md->layout.lb_buf = req_capsule_server_sized_get(pill,
 512                                                         &RMF_MDT_MD,
 513                                                         md->layout.lb_len);
 514                 if (md->layout.lb_buf == NULL)
 515                         GOTO(out, rc = -EPROTO);
 516         } else if (md->body->mbo_valid & OBD_MD_FLDIREA) {
 517                 const union lmv_mds_md *lmv;
 518                 size_t lmv_size;
 519
 520                 if (!S_ISDIR(md->body->mbo_mode)) {
 521                         CDEBUG(D_INFO, "OBD_MD_FLDIREA set, should be a "
 522                                "directory, but is not\n");
 523                         GOTO(out, rc = -EPROTO);
 524                 }
 525
 526                 lmv_size = md->body->mbo_eadatasize;
 527                 if (lmv_size == 0) {
 528                         CDEBUG(D_INFO, "OBD_MD_FLDIREA is set, "
 529                                "but eadatasize 0\n");
 530                         RETURN(-EPROTO);
 531                 }
 532
 533                 if (md->body->mbo_valid & OBD_MD_MEA) {
 534                         lmv = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
 535                                                            lmv_size);
 536                         if (lmv == NULL)
 537                                 GOTO(out, rc = -EPROTO);
 538
 539                         rc = md_unpackmd(md_exp, &md->lmv, lmv, lmv_size);
 540                         if (rc < 0)
 541                                 GOTO(out, rc);
 542
 543                         if (rc < (typeof(rc))sizeof(*md->lmv)) {
 544                                 CDEBUG(D_INFO, "size too small:  "
 545                                        "rc < sizeof(*md->lmv) (%d < %d)\n",
 546                                         rc, (int)sizeof(*md->lmv));
 547                                 GOTO(out, rc = -EPROTO);
 548                         }
 549                 }
 550         }
 551         rc = 0;
 552
 553         if (md->body->mbo_valid & OBD_MD_FLACL) {
 554                 /* for ACL, it's possible that FLACL is set but aclsize is zero.
 555                  * only when aclsize != 0 there's an actual segment for ACL
 556                  * in reply buffer.
 557                  */
 558                 if (md->body->mbo_aclsize) {
 559                         rc = mdc_unpack_acl(req, md);
 560                         if (rc)
 561                                 GOTO(out, rc);
 562 #ifdef CONFIG_FS_POSIX_ACL
 563                 } else {
 564                         md->posix_acl = NULL;
 565 #endif
 566                 }
 567         }
 568
 569         EXIT;
 570 out:
 571         if (rc) {
 572 #ifdef CONFIG_FS_POSIX_ACL
 573                 posix_acl_release(md->posix_acl);
 574 #endif
 575         }
 576         return rc;
 577 }
 578
 579 int mdc_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
 580 {
 581         ENTRY;
 582         RETURN(0);
 583 }
 584
 585 void mdc_replay_open(struct ptlrpc_request *req)
 586 {
 587         struct md_open_data *mod = req->rq_cb_data;
 588         struct ptlrpc_request *close_req;
 589         struct obd_client_handle *och;
 590         struct lustre_handle old;
 591         struct mdt_body *body;
 592         ENTRY;
 593
 594         if (mod == NULL) {
 595                 DEBUG_REQ(D_ERROR, req,
 596                           "Can't properly replay without open data.");
 597                 EXIT;
 598                 return;
 599         }
 600
 601         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 602         LASSERT(body != NULL);
 603
 604         spin_lock(&req->rq_lock);
 605         och = mod->mod_och;
 606         if (och && och->och_fh.cookie)
 607                 req->rq_early_free_repbuf = 1;
 608         else
 609                 req->rq_early_free_repbuf = 0;
 610         spin_unlock(&req->rq_lock);
 611
 612         if (req->rq_early_free_repbuf) {
 613                 struct lustre_handle *file_fh;
 614
 615                 LASSERT(och->och_magic == OBD_CLIENT_HANDLE_MAGIC);
 616
 617                 file_fh = &och->och_fh;
 618                 CDEBUG(D_HA, "updating handle from %#llx to %#llx\n",
 619                        file_fh->cookie, body->mbo_handle.cookie);
 620                 old = *file_fh;
 621                 *file_fh = body->mbo_handle;
 622         }
 623
 624         close_req = mod->mod_close_req;
 625         if (close_req) {
 626                 __u32 opc = lustre_msg_get_opc(close_req->rq_reqmsg);
 627                 struct mdt_ioepoch *epoch;
 628
 629                 LASSERT(opc == MDS_CLOSE);
 630                 epoch = req_capsule_client_get(&close_req->rq_pill,
 631                                                &RMF_MDT_EPOCH);
 632                 LASSERT(epoch);
 633
 634                 if (req->rq_early_free_repbuf)
 635                         LASSERT(!memcmp(&old, &epoch->mio_handle, sizeof(old)));
 636
 637                 DEBUG_REQ(D_HA, close_req, "updating close body with new fh");
 638                 epoch->mio_handle = body->mbo_handle;
 639         }
 640         EXIT;
 641 }
 642
 643 void mdc_commit_open(struct ptlrpc_request *req)
 644 {
 645         struct md_open_data *mod = req->rq_cb_data;
 646         if (mod == NULL)
 647                 return;
 648
 649         /**
 650          * No need to touch md_open_data::mod_och, it holds a reference on
 651          * \var mod and will zero references to each other, \var mod will be
 652          * freed after that when md_open_data::mod_och will put the reference.
 653          */
 654
 655         /**
 656          * Do not let open request to disappear as it still may be needed
 657          * for close rpc to happen (it may happen on evict only, otherwise
 658          * ptlrpc_request::rq_replay does not let mdc_commit_open() to be
 659          * called), just mark this rpc as committed to distinguish these 2
 660          * cases, see mdc_close() for details. The open request reference will
 661          * be put along with freeing \var mod.
 662          */
 663         ptlrpc_request_addref(req);
 664         spin_lock(&req->rq_lock);
 665         req->rq_committed = 1;
 666         spin_unlock(&req->rq_lock);
 667         req->rq_cb_data = NULL;
 668         obd_mod_put(mod);
 669 }
 670
 671 int mdc_set_open_replay_data(struct obd_export *exp,
 672                              struct obd_client_handle *och,
 673                              struct lookup_intent *it)
 674 {
 675         struct md_open_data     *mod;
 676         struct mdt_rec_create   *rec;
 677         struct mdt_body         *body;
 678         struct ptlrpc_request   *open_req = it->it_request;
 679         struct obd_import       *imp = open_req->rq_import;
 680         ENTRY;
 681
 682         if (!open_req->rq_replay)
 683                 RETURN(0);
 684
 685         rec = req_capsule_client_get(&open_req->rq_pill, &RMF_REC_REINT);
 686         body = req_capsule_server_get(&open_req->rq_pill, &RMF_MDT_BODY);
 687         LASSERT(rec != NULL);
 688         /* Incoming message in my byte order (it's been swabbed). */
 689         /* Outgoing messages always in my byte order. */
 690         LASSERT(body != NULL);
 691
 692         /* Only if the import is replayable, we set replay_open data */
 693         if (och && imp->imp_replayable) {
 694                 mod = obd_mod_alloc();
 695                 if (mod == NULL) {
 696                         DEBUG_REQ(D_ERROR, open_req,
 697                                   "Can't allocate md_open_data");
 698                         RETURN(0);
 699                 }
 700
 701                 /**
 702                  * Take a reference on \var mod, to be freed on mdc_close().
 703                  * It protects \var mod from being freed on eviction (commit
 704                  * callback is called despite rq_replay flag).
 705                  * Another reference for \var och.
 706                  */
 707                 obd_mod_get(mod);
 708                 obd_mod_get(mod);
 709
 710                 spin_lock(&open_req->rq_lock);
 711                 och->och_mod = mod;
 712                 mod->mod_och = och;
 713                 mod->mod_is_create = it_disposition(it, DISP_OPEN_CREATE) ||
 714                                      it_disposition(it, DISP_OPEN_STRIPE);
 715                 mod->mod_open_req = open_req;
 716                 open_req->rq_cb_data = mod;
 717                 open_req->rq_commit_cb = mdc_commit_open;
 718                 open_req->rq_early_free_repbuf = 1;
 719                 spin_unlock(&open_req->rq_lock);
 720         }
 721
 722         rec->cr_fid2 = body->mbo_fid1;
 723         rec->cr_ioepoch = body->mbo_ioepoch;
 724         rec->cr_old_handle.cookie = body->mbo_handle.cookie;
 725         open_req->rq_replay_cb = mdc_replay_open;
 726         if (!fid_is_sane(&body->mbo_fid1)) {
 727                 DEBUG_REQ(D_ERROR, open_req, "Saving replay request with "
 728                           "insane fid");
 729                 LBUG();
 730         }
 731
 732         DEBUG_REQ(D_RPCTRACE, open_req, "Set up open replay data");
 733         RETURN(0);
 734 }
 735
 736 static void mdc_free_open(struct md_open_data *mod)
 737 {
 738         int committed = 0;
 739
 740         if (mod->mod_is_create == 0 &&
 741             imp_connect_disp_stripe(mod->mod_open_req->rq_import))
 742                 committed = 1;
 743
 744         /**
 745          * No reason to asssert here if the open request has
 746          * rq_replay == 1. It means that mdc_close failed, and
 747          * close request wasn`t sent. It is not fatal to client.
 748          * The worst thing is eviction if the client gets open lock
 749          **/
 750
 751         DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "free open request rq_replay"
 752                   "= %d\n", mod->mod_open_req->rq_replay);
 753
 754         ptlrpc_request_committed(mod->mod_open_req, committed);
 755         if (mod->mod_close_req)
 756                 ptlrpc_request_committed(mod->mod_close_req, committed);
 757 }
 758
 759 int mdc_clear_open_replay_data(struct obd_export *exp,
 760                                struct obd_client_handle *och)
 761 {
 762         struct md_open_data *mod = och->och_mod;
 763         ENTRY;
 764
 765         /**
 766          * It is possible to not have \var mod in a case of eviction between
 767          * lookup and ll_file_open().
 768          **/
 769         if (mod == NULL)
 770                 RETURN(0);
 771
 772         LASSERT(mod != LP_POISON);
 773         LASSERT(mod->mod_open_req != NULL);
 774
 775         spin_lock(&mod->mod_open_req->rq_lock);
 776         if (mod->mod_och)
 777                 mod->mod_och->och_fh.cookie = 0;
 778         mod->mod_open_req->rq_early_free_repbuf = 0;
 779         spin_unlock(&mod->mod_open_req->rq_lock);
 780         mdc_free_open(mod);
 781
 782         mod->mod_och = NULL;
 783         och->och_mod = NULL;
 784         obd_mod_put(mod);
 785
 786         RETURN(0);
 787 }
 788
 789 static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
 790                      struct md_open_data *mod, struct ptlrpc_request **request)
 791 {
 792         struct obd_device     *obd = class_exp2obd(exp);
 793         struct ptlrpc_request *req;
 794         struct req_format     *req_fmt;
 795         size_t                 u32_count = 0;
 796         int                    rc;
 797         int                    saved_rc = 0;
 798         ENTRY;
 799
 800         CDEBUG(D_INODE, "%s: "DFID" file closed with intent: %x\n",
 801                exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
 802                op_data->op_bias);
 803
 804         if (op_data->op_bias & MDS_CLOSE_INTENT) {
 805                 req_fmt = &RQF_MDS_CLOSE_INTENT;
 806                 if (op_data->op_bias & MDS_HSM_RELEASE) {
 807                         /* allocate a FID for volatile file */
 808                         rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2,
 809                                            op_data);
 810                         if (rc < 0) {
 811                                 CERROR("%s: "DFID" allocating FID: rc = %d\n",
 812                                        obd->obd_name, PFID(&op_data->op_fid1),
 813                                        rc);
 814                                 /* save the errcode and proceed to close */
 815                                 saved_rc = rc;
 816                         }
 817                 }
 818                 if (op_data->op_bias & MDS_CLOSE_RESYNC_DONE) {
 819                         size_t count = op_data->op_data_size / sizeof(__u32);
 820
 821                         if (count > INLINE_RESYNC_ARRAY_SIZE)
 822                                 u32_count = count;
 823                 }
 824         } else {
 825                 req_fmt = &RQF_MDS_CLOSE;
 826         }
 827
 828         *request = NULL;
 829         if (OBD_FAIL_CHECK(OBD_FAIL_MDC_CLOSE))
 830                 req = NULL;
 831         else
 832                 req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);
 833
 834         /* Ensure that this close's handle is fixed up during replay. */
 835         if (likely(mod != NULL)) {
 836                 LASSERTF(mod->mod_open_req != NULL &&
 837                          mod->mod_open_req->rq_type != LI_POISON,
 838                          "POISONED open %p!\n", mod->mod_open_req);
 839
 840                 mod->mod_close_req = req;
 841
 842                 DEBUG_REQ(D_HA, mod->mod_open_req, "matched open");
 843                 /* We no longer want to preserve this open for replay even
 844                  * though the open was committed. b=3632, b=3633 */
 845                 spin_lock(&mod->mod_open_req->rq_lock);
 846                 mod->mod_open_req->rq_replay = 0;
 847                 spin_unlock(&mod->mod_open_req->rq_lock);
 848         } else {
 849                 CDEBUG(D_HA, "couldn't find open req; expecting close error\n");
 850         }
 851         if (req == NULL) {
 852                 /**
 853                  * TODO: repeat close after errors
 854                  */
 855                 CWARN("%s: close of FID "DFID" failed, file reference will be "
 856                       "dropped when this client unmounts or is evicted\n",
 857                       obd->obd_name, PFID(&op_data->op_fid1));
 858                 GOTO(out, rc = -ENOMEM);
 859         }
 860
 861         if (u32_count > 0)
 862                 req_capsule_set_size(&req->rq_pill, &RMF_U32, RCL_CLIENT,
 863                                      u32_count * sizeof(__u32));
 864
 865         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
 866         if (rc) {
 867                 ptlrpc_request_free(req);
 868                 req = NULL;
 869                 GOTO(out, rc);
 870         }
 871
 872         /* To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
 873          * portal whose threads are not taking any DLM locks and are therefore
 874          * always progressing */
 875         req->rq_request_portal = MDS_READPAGE_PORTAL;
 876         ptlrpc_at_set_req_timeout(req);
 877
 878
 879         mdc_close_pack(req, op_data);
 880
 881         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 882                              obd->u.cli.cl_default_mds_easize);
 883
 884         ptlrpc_request_set_replen(req);
 885
 886         mdc_get_mod_rpc_slot(req, NULL);
 887         rc = ptlrpc_queue_wait(req);
 888         mdc_put_mod_rpc_slot(req, NULL);
 889
 890         if (req->rq_repmsg == NULL) {
 891                 CDEBUG(D_RPCTRACE, "request failed to send: %p, %d\n", req,
 892                        req->rq_status);
 893                 if (rc == 0)
 894                         rc = req->rq_status ?: -EIO;
 895         } else if (rc == 0 || rc == -EAGAIN) {
 896                 struct mdt_body *body;
 897
 898                 rc = lustre_msg_get_status(req->rq_repmsg);
 899                 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
 900                         DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR, err "
 901                                   "= %d", rc);
 902                         if (rc > 0)
 903                                 rc = -rc;
 904                 }
 905                 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 906                 if (body == NULL)
 907                         rc = -EPROTO;
 908         } else if (rc == -ESTALE) {
 909                 /**
 910                  * it can be allowed error after 3633 if open was committed and
 911                  * server failed before close was sent. Let's check if mod
 912                  * exists and return no error in that case
 913                  */
 914                 if (mod) {
 915                         DEBUG_REQ(D_HA, req, "Reset ESTALE = %d", rc);
 916                         LASSERT(mod->mod_open_req != NULL);
 917                         if (mod->mod_open_req->rq_committed)
 918                                 rc = 0;
 919                 }
 920         }
 921
 922 out:
 923         if (mod) {
 924                 if (rc != 0)
 925                         mod->mod_close_req = NULL;
 926                 /* Since now, mod is accessed through open_req only,
 927                  * thus close req does not keep a reference on mod anymore. */
 928                 obd_mod_put(mod);
 929         }
 930         *request = req;
 931
 932         RETURN(rc < 0 ? rc : saved_rc);
 933 }
 934
 935 static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid,
 936                        u64 offset, struct page **pages, int npages,
 937                        struct ptlrpc_request **request)
 938 {
 939         struct ptlrpc_request   *req;
 940         struct ptlrpc_bulk_desc *desc;
 941         int                      i;
 942         wait_queue_head_t        waitq;
 943         int                      resends = 0;
 944         struct l_wait_info       lwi;
 945         int                      rc;
 946         ENTRY;
 947
 948         *request = NULL;
 949         init_waitqueue_head(&waitq);
 950
 951 restart_bulk:
 952         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE);
 953         if (req == NULL)
 954                 RETURN(-ENOMEM);
 955
 956         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE);
 957         if (rc) {
 958                 ptlrpc_request_free(req);
 959                 RETURN(rc);
 960         }
 961
 962         req->rq_request_portal = MDS_READPAGE_PORTAL;
 963         ptlrpc_at_set_req_timeout(req);
 964
 965         desc = ptlrpc_prep_bulk_imp(req, npages, 1,
 966                                     PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
 967                                     MDS_BULK_PORTAL,
 968                                     &ptlrpc_bulk_kiov_pin_ops);
 969         if (desc == NULL) {
 970                 ptlrpc_req_finished(req);
 971                 RETURN(-ENOMEM);
 972         }
 973
 974         /* NB req now owns desc and will free it when it gets freed */
 975         for (i = 0; i < npages; i++)
 976                 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
 977                                                  PAGE_SIZE);
 978
 979         mdc_readdir_pack(req, offset, PAGE_SIZE * npages, fid);
 980
 981         ptlrpc_request_set_replen(req);
 982         rc = ptlrpc_queue_wait(req);
 983         if (rc) {
 984                 ptlrpc_req_finished(req);
 985                 if (rc != -ETIMEDOUT)
 986                         RETURN(rc);
 987
 988                 resends++;
 989                 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
 990                         CERROR("%s: too many resend retries: rc = %d\n",
 991                                exp->exp_obd->obd_name, -EIO);
 992                         RETURN(-EIO);
 993                 }
 994                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL,
 995                                        NULL);
 996                 l_wait_event(waitq, 0, &lwi);
 997
 998                 goto restart_bulk;
 999         }
1000
1001         rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1002                                           req->rq_bulk->bd_nob_transferred);
1003         if (rc < 0) {
1004                 ptlrpc_req_finished(req);
1005                 RETURN(rc);
1006         }
1007
1008         if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) {
1009                 CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n",
1010                        exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred,
1011                        PAGE_SIZE * npages);
1012                 ptlrpc_req_finished(req);
1013                 RETURN(-EPROTO);
1014         }
1015
1016         *request = req;
1017         RETURN(0);
1018 }
1019
1020 static void mdc_release_page(struct page *page, int remove)
1021 {
1022         if (remove) {
1023                 lock_page(page);
1024                 if (likely(page->mapping != NULL))
1025                         truncate_complete_page(page->mapping, page);
1026                 unlock_page(page);
1027         }
1028         put_page(page);
1029 }
1030
1031 static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash,
1032                                     __u64 *start, __u64 *end, int hash64)
1033 {
1034         /*
1035          * Complement of hash is used as an index so that
1036          * radix_tree_gang_lookup() can be used to find a page with starting
1037          * hash _smaller_ than one we are looking for.
1038          */
1039         unsigned long offset = hash_x_index(*hash, hash64);
1040         struct page *page;
1041         int found;
1042
1043         spin_lock_irq(&mapping->tree_lock);
1044         found = radix_tree_gang_lookup(&mapping->page_tree,
1045                                        (void **)&page, offset, 1);
1046         if (found > 0 && !radix_tree_exceptional_entry(page)) {
1047                 struct lu_dirpage *dp;
1048
1049                 get_page(page);
1050                 spin_unlock_irq(&mapping->tree_lock);
1051                 /*
1052                  * In contrast to find_lock_page() we are sure that directory
1053                  * page cannot be truncated (while DLM lock is held) and,
1054                  * hence, can avoid restart.
1055                  *
1056                  * In fact, page cannot be locked here at all, because
1057                  * mdc_read_page_remote does synchronous io.
1058                  */
1059                 wait_on_page_locked(page);
1060                 if (PageUptodate(page)) {
1061                         dp = kmap(page);
1062                         if (BITS_PER_LONG == 32 && hash64) {
1063                                 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1064                                 *end   = le64_to_cpu(dp->ldp_hash_end) >> 32;
1065                                 *hash  = *hash >> 32;
1066                         } else {
1067                                 *start = le64_to_cpu(dp->ldp_hash_start);
1068                                 *end   = le64_to_cpu(dp->ldp_hash_end);
1069                         }
1070                         if (unlikely(*start == 1 && *hash == 0))
1071                                 *hash = *start;
1072                         else
1073                                 LASSERTF(*start <= *hash, "start = %#llx"
1074                                          ",end = %#llx,hash = %#llx\n",
1075                                          *start, *end, *hash);
1076                         CDEBUG(D_VFSTRACE, "offset %lx [%#llx %#llx],"
1077                               " hash %#llx\n", offset, *start, *end, *hash);
1078                         if (*hash > *end) {
1079                                 kunmap(page);
1080                                 mdc_release_page(page, 0);
1081                                 page = NULL;
1082                         } else if (*end != *start && *hash == *end) {
1083                                 /*
1084                                  * upon hash collision, remove this page,
1085                                  * otherwise put page reference, and
1086                                  * mdc_read_page_remote() will issue RPC to
1087                                  * fetch the page we want.
1088                                  */
1089                                 kunmap(page);
1090                                 mdc_release_page(page,
1091                                     le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1092                                 page = NULL;
1093                         }
1094                 } else {
1095                         put_page(page);
1096                         page = ERR_PTR(-EIO);
1097                 }
1098         } else {
1099                 spin_unlock_irq(&mapping->tree_lock);
1100                 page = NULL;
1101         }
1102         return page;
1103 }
1104
1105 /*
1106  * Adjust a set of pages, each page containing an array of lu_dirpages,
1107  * so that each page can be used as a single logical lu_dirpage.
1108  *
1109  * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
1110  * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
1111  * struct lu_dirent.  It has size up to LU_PAGE_SIZE. The ldp_hash_end
1112  * value is used as a cookie to request the next lu_dirpage in a
1113  * directory listing that spans multiple pages (two in this example):
1114  *   ________
1115  *  |        |
1116  * .|--------v-------   -----.
1117  * |s|e|f|p|ent|ent| ... |ent|
1118  * '--|--------------   -----'   Each PAGE contains a single
1119  *    '------.                   lu_dirpage.
1120  * .---------v-------   -----.
1121  * |s|e|f|p|ent| 0 | ... | 0 |
1122  * '-----------------   -----'
1123  *
1124  * However, on hosts where the native VM page size (PAGE_SIZE) is
1125  * larger than LU_PAGE_SIZE, a single host page may contain multiple
1126  * lu_dirpages. After reading the lu_dirpages from the MDS, the
1127  * ldp_hash_end of the first lu_dirpage refers to the one immediately
1128  * after it in the same PAGE (arrows simplified for brevity, but
1129  * in general e0==s1, e1==s2, etc.):
1130  *
1131  * .--------------------   -----.
1132  * |s0|e0|f0|p|ent|ent| ... |ent|
1133  * |---v----------------   -----|
1134  * |s1|e1|f1|p|ent|ent| ... |ent|
1135  * |---v----------------   -----|  Here, each PAGE contains
1136  *             ...                 multiple lu_dirpages.
1137  * |---v----------------   -----|
1138  * |s'|e'|f'|p|ent|ent| ... |ent|
1139  * '---|----------------   -----'
1140  *     v
1141  * .----------------------------.
1142  * |        next PAGE           |
1143  *
1144  * This structure is transformed into a single logical lu_dirpage as follows:
1145  *
1146  * - Replace e0 with e' so the request for the next lu_dirpage gets the page
1147  *   labeled 'next PAGE'.
1148  *
1149  * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
1150  *   a hash collision with the next page exists.
1151  *
1152  * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
1153  *   to the first entry of the next lu_dirpage.
1154  */
1155 #if PAGE_SIZE > LU_PAGE_SIZE
1156 static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs)
1157 {
1158         int i;
1159
1160         for (i = 0; i < cfs_pgs; i++) {
1161                 struct lu_dirpage       *dp = kmap(pages[i]);
1162                 struct lu_dirpage       *first = dp;
1163                 struct lu_dirent        *end_dirent = NULL;
1164                 struct lu_dirent        *ent;
1165                 __u64           hash_end = le64_to_cpu(dp->ldp_hash_end);
1166                 __u32           flags = le32_to_cpu(dp->ldp_flags);
1167
1168                 while (--lu_pgs > 0) {
1169                         ent = lu_dirent_start(dp);
1170                         for (end_dirent = ent; ent != NULL;
1171                              end_dirent = ent, ent = lu_dirent_next(ent));
1172
1173                         /* Advance dp to next lu_dirpage. */
1174                         dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
1175
1176                         /* Check if we've reached the end of the PAGE. */
1177                         if (!((unsigned long)dp & ~PAGE_MASK))
1178                                 break;
1179
1180                         /* Save the hash and flags of this lu_dirpage. */
1181                         hash_end = le64_to_cpu(dp->ldp_hash_end);
1182                         flags = le32_to_cpu(dp->ldp_flags);
1183
1184                         /* Check if lu_dirpage contains no entries. */
1185                         if (end_dirent == NULL)
1186                                 break;
1187
1188                         /* Enlarge the end entry lde_reclen from 0 to
1189                          * first entry of next lu_dirpage. */
1190                         LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
1191                         end_dirent->lde_reclen =
1192                                 cpu_to_le16((char *)(dp->ldp_entries) -
1193                                             (char *)end_dirent);
1194                 }
1195
1196                 first->ldp_hash_end = hash_end;
1197                 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
1198                 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
1199
1200                 kunmap(pages[i]);
1201         }
1202         LASSERTF(lu_pgs == 0, "left = %d\n", lu_pgs);
1203 }
1204 #else
1205 #define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0)
1206 #endif  /* PAGE_SIZE > LU_PAGE_SIZE */
1207
1208 /* parameters for readdir page */
1209 struct readpage_param {
1210         struct md_op_data       *rp_mod;
1211         __u64                   rp_off;
1212         int                     rp_hash64;
1213         struct obd_export       *rp_exp;
1214         struct md_callback      *rp_cb;
1215 };
1216
1217 #ifndef HAVE_DELETE_FROM_PAGE_CACHE
1218 static inline void delete_from_page_cache(struct page *page)
1219 {
1220         remove_from_page_cache(page);
1221         put_page(page);
1222 }
1223 #endif
1224
1225 /**
1226  * Read pages from server.
1227  *
1228  * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
1229  * a header lu_dirpage which describes the start/end hash, and whether this
1230  * page is empty (contains no dir entry) or hash collide with next page.
1231  * After client receives reply, several pages will be integrated into dir page
1232  * in PAGE_SIZE (if PAGE_SIZE greater than LU_PAGE_SIZE), and the
1233  * lu_dirpage for this integrated page will be adjusted.
1234  **/
1235 static int mdc_read_page_remote(void *data, struct page *page0)
1236 {
1237         struct readpage_param *rp = data;
1238         struct page **page_pool;
1239         struct page *page;
1240         struct lu_dirpage *dp;
1241         struct md_op_data *op_data = rp->rp_mod;
1242         struct ptlrpc_request *req;
1243         int max_pages;
1244         struct inode *inode;
1245         struct lu_fid *fid;
1246         int rd_pgs = 0; /* number of pages actually read */
1247         int npages;
1248         int i;
1249         int rc;
1250         ENTRY;
1251
1252         max_pages = rp->rp_exp->exp_obd->u.cli.cl_max_pages_per_rpc;
1253         inode = op_data->op_data;
1254         fid = &op_data->op_fid1;
1255         LASSERT(inode != NULL);
1256
1257         OBD_ALLOC(page_pool, sizeof(page_pool[0]) * max_pages);
1258         if (page_pool != NULL) {
1259                 page_pool[0] = page0;
1260         } else {
1261                 page_pool = &page0;
1262                 max_pages = 1;
1263         }
1264
1265         for (npages = 1; npages < max_pages; npages++) {
1266                 page = __page_cache_alloc(mapping_gfp_mask(inode->i_mapping)
1267                                           | __GFP_COLD);
1268                 if (page == NULL)
1269                         break;
1270                 page_pool[npages] = page;
1271         }
1272
1273         rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, page_pool, npages, &req);
1274         if (rc < 0) {
1275                 /* page0 is special, which was added into page cache early */
1276                 delete_from_page_cache(page0);
1277         } else {
1278                 int lu_pgs;
1279
1280                 rd_pgs = (req->rq_bulk->bd_nob_transferred + PAGE_SIZE - 1) >>
1281                         PAGE_SHIFT;
1282                 lu_pgs = req->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
1283                 LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
1284
1285                 CDEBUG(D_INODE, "read %d(%d) pages\n", rd_pgs, lu_pgs);
1286
1287                 mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs);
1288
1289                 SetPageUptodate(page0);
1290         }
1291         unlock_page(page0);
1292
1293         ptlrpc_req_finished(req);
1294         CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages);
1295         for (i = 1; i < npages; i++) {
1296                 unsigned long   offset;
1297                 __u64           hash;
1298                 int ret;
1299
1300                 page = page_pool[i];
1301
1302                 if (rc < 0 || i >= rd_pgs) {
1303                         put_page(page);
1304                         continue;
1305                 }
1306
1307                 SetPageUptodate(page);
1308
1309                 dp = kmap(page);
1310                 hash = le64_to_cpu(dp->ldp_hash_start);
1311                 kunmap(page);
1312
1313                 offset = hash_x_index(hash, rp->rp_hash64);
1314
1315                 prefetchw(&page->flags);
1316                 ret = add_to_page_cache_lru(page, inode->i_mapping, offset,
1317                                             GFP_KERNEL);
1318                 if (ret == 0)
1319                         unlock_page(page);
1320                 else
1321                         CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
1322                                " rc = %d\n", offset, ret);
1323                 put_page(page);
1324         }
1325
1326         if (page_pool != &page0)
1327                 OBD_FREE(page_pool, sizeof(page_pool[0]) * max_pages);
1328
1329         RETURN(rc);
1330 }
1331
1332 /**
1333  * Read dir page from cache first, if it can not find it, read it from
1334  * server and add into the cache.
1335  *
1336  * \param[in] exp       MDC export
1337  * \param[in] op_data   client MD stack parameters, transfering parameters
1338  *                      between different layers on client MD stack.
1339  * \param[in] cb_op     callback required for ldlm lock enqueue during
1340  *                      read page
1341  * \param[in] hash_offset the hash offset of the page to be read
1342  * \param[in] ppage     the page to be read
1343  *
1344  * retval               = 0 get the page successfully
1345  *                      errno(<0) get the page failed
1346  */
1347 static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data,
1348                          struct md_callback *cb_op, __u64 hash_offset,
1349                          struct page **ppage)
1350 {
1351         struct lookup_intent    it = { .it_op = IT_READDIR };
1352         struct page             *page;
1353         struct inode            *dir = op_data->op_data;
1354         struct address_space    *mapping;
1355         struct lu_dirpage       *dp;
1356         __u64                   start = 0;
1357         __u64                   end = 0;
1358         struct lustre_handle    lockh;
1359         struct ptlrpc_request   *enq_req = NULL;
1360         struct readpage_param   rp_param;
1361         int rc;
1362
1363         ENTRY;
1364
1365         *ppage = NULL;
1366
1367         LASSERT(dir != NULL);
1368         mapping = dir->i_mapping;
1369
1370         rc = mdc_intent_lock(exp, op_data, &it, &enq_req,
1371                              cb_op->md_blocking_ast, 0);
1372         if (enq_req != NULL)
1373                 ptlrpc_req_finished(enq_req);
1374
1375         if (rc < 0) {
1376                 CERROR("%s: "DFID" lock enqueue fails: rc = %d\n",
1377                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc);
1378                 RETURN(rc);
1379         }
1380
1381         rc = 0;
1382         lockh.cookie = it.it_lock_handle;
1383         mdc_set_lock_data(exp, &lockh, dir, NULL);
1384
1385         rp_param.rp_off = hash_offset;
1386         rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64;
1387         page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end,
1388                                rp_param.rp_hash64);
1389         if (IS_ERR(page)) {
1390                 CERROR("%s: dir page locate: "DFID" at %llu: rc %ld\n",
1391                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1392                        rp_param.rp_off, PTR_ERR(page));
1393                 GOTO(out_unlock, rc = PTR_ERR(page));
1394         } else if (page != NULL) {
1395                 /*
1396                  * XXX nikita: not entirely correct handling of a corner case:
1397                  * suppose hash chain of entries with hash value HASH crosses
1398                  * border between pages P0 and P1. First both P0 and P1 are
1399                  * cached, seekdir() is called for some entry from the P0 part
1400                  * of the chain. Later P0 goes out of cache. telldir(HASH)
1401                  * happens and finds P1, as it starts with matching hash
1402                  * value. Remaining entries from P0 part of the chain are
1403                  * skipped. (Is that really a bug?)
1404                  *
1405                  * Possible solutions: 0. don't cache P1 is such case, handle
1406                  * it as an "overflow" page. 1. invalidate all pages at
1407                  * once. 2. use HASH|1 as an index for P1.
1408                  */
1409                 GOTO(hash_collision, page);
1410         }
1411
1412         rp_param.rp_exp = exp;
1413         rp_param.rp_mod = op_data;
1414         page = read_cache_page(mapping,
1415                                hash_x_index(rp_param.rp_off,
1416                                             rp_param.rp_hash64),
1417                                mdc_read_page_remote, &rp_param);
1418         if (IS_ERR(page)) {
1419                 CDEBUG(D_INFO, "%s: read cache page: "DFID" at %llu: %ld\n",
1420                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1421                        rp_param.rp_off, PTR_ERR(page));
1422                 GOTO(out_unlock, rc = PTR_ERR(page));
1423         }
1424
1425         wait_on_page_locked(page);
1426         (void)kmap(page);
1427         if (!PageUptodate(page)) {
1428                 CERROR("%s: page not updated: "DFID" at %llu: rc %d\n",
1429                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1430                        rp_param.rp_off, -5);
1431                 goto fail;
1432         }
1433         if (!PageChecked(page))
1434                 SetPageChecked(page);
1435         if (PageError(page)) {
1436                 CERROR("%s: page error: "DFID" at %llu: rc %d\n",
1437                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1438                        rp_param.rp_off, -5);
1439                 goto fail;
1440         }
1441
1442 hash_collision:
1443         dp = page_address(page);
1444         if (BITS_PER_LONG == 32 && rp_param.rp_hash64) {
1445                 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1446                 end   = le64_to_cpu(dp->ldp_hash_end) >> 32;
1447                 rp_param.rp_off = hash_offset >> 32;
1448         } else {
1449                 start = le64_to_cpu(dp->ldp_hash_start);
1450                 end   = le64_to_cpu(dp->ldp_hash_end);
1451                 rp_param.rp_off = hash_offset;
1452         }
1453         if (end == start) {
1454                 LASSERT(start == rp_param.rp_off);
1455                 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
1456 #if BITS_PER_LONG == 32
1457                 CWARN("Real page-wide hash collision at [%llu %llu] with "
1458                       "hash %llu\n", le64_to_cpu(dp->ldp_hash_start),
1459                       le64_to_cpu(dp->ldp_hash_end), hash_offset);
1460 #endif
1461
1462                 /*
1463                  * Fetch whole overflow chain...
1464                  *
1465                  * XXX not yet.
1466                  */
1467                 goto fail;
1468         }
1469         *ppage = page;
1470 out_unlock:
1471         ldlm_lock_decref(&lockh, it.it_lock_mode);
1472         return rc;
1473 fail:
1474         kunmap(page);
1475         mdc_release_page(page, 1);
1476         rc = -EIO;
1477         goto out_unlock;
1478 }
1479
1480 static int mdc_statfs(const struct lu_env *env,
1481                       struct obd_export *exp, struct obd_statfs *osfs,
1482                       time64_t max_age, __u32 flags)
1483 {
1484         struct obd_device     *obd = class_exp2obd(exp);
1485         struct ptlrpc_request *req;
1486         struct obd_statfs     *msfs;
1487         struct obd_import     *imp = NULL;
1488         int                    rc;
1489         ENTRY;
1490
1491         /*
1492          * Since the request might also come from lprocfs, so we need
1493          * sync this with client_disconnect_export Bug15684
1494          */
1495         down_read(&obd->u.cli.cl_sem);
1496         if (obd->u.cli.cl_import)
1497                 imp = class_import_get(obd->u.cli.cl_import);
1498         up_read(&obd->u.cli.cl_sem);
1499         if (!imp)
1500                 RETURN(-ENODEV);
1501
1502         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_STATFS,
1503                                         LUSTRE_MDS_VERSION, MDS_STATFS);
1504         if (req == NULL)
1505                 GOTO(output, rc = -ENOMEM);
1506
1507         ptlrpc_request_set_replen(req);
1508
1509         if (flags & OBD_STATFS_NODELAY) {
1510                 /* procfs requests not want stay in wait for avoid deadlock */
1511                 req->rq_no_resend = 1;
1512                 req->rq_no_delay = 1;
1513         }
1514
1515         rc = ptlrpc_queue_wait(req);
1516         if (rc) {
1517                 /* check connection error first */
1518                 if (imp->imp_connect_error)
1519                         rc = imp->imp_connect_error;
1520                 GOTO(out, rc);
1521         }
1522
1523         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1524         if (msfs == NULL)
1525                 GOTO(out, rc = -EPROTO);
1526
1527         *osfs = *msfs;
1528         EXIT;
1529 out:
1530         ptlrpc_req_finished(req);
1531 output:
1532         class_import_put(imp);
1533         return rc;
1534 }
1535
1536 static int mdc_ioc_fid2path(struct obd_export *exp, struct getinfo_fid2path *gf)
1537 {
1538         __u32 keylen, vallen;
1539         void *key;
1540         int rc;
1541
1542         if (gf->gf_pathlen > PATH_MAX)
1543                 RETURN(-ENAMETOOLONG);
1544         if (gf->gf_pathlen < 2)
1545                 RETURN(-EOVERFLOW);
1546
1547         /* Key is KEY_FID2PATH + getinfo_fid2path description */
1548         keylen = cfs_size_round(sizeof(KEY_FID2PATH) + sizeof(*gf) +
1549                                 sizeof(struct lu_fid));
1550         OBD_ALLOC(key, keylen);
1551         if (key == NULL)
1552                 RETURN(-ENOMEM);
1553         memcpy(key, KEY_FID2PATH, sizeof(KEY_FID2PATH));
1554         memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)), gf, sizeof(*gf));
1555         memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*gf),
1556                gf->gf_u.gf_root_fid, sizeof(struct lu_fid));
1557         CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
1558                PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno);
1559
1560         if (!fid_is_sane(&gf->gf_fid))
1561                 GOTO(out, rc = -EINVAL);
1562
1563         /* Val is struct getinfo_fid2path result plus path */
1564         vallen = sizeof(*gf) + gf->gf_pathlen;
1565
1566         rc = obd_get_info(NULL, exp, keylen, key, &vallen, gf);
1567         if (rc != 0 && rc != -EREMOTE)
1568                 GOTO(out, rc);
1569
1570         if (vallen <= sizeof(*gf))
1571                 GOTO(out, rc = -EPROTO);
1572         if (vallen > sizeof(*gf) + gf->gf_pathlen)
1573                 GOTO(out, rc = -EOVERFLOW);
1574
1575         CDEBUG(D_IOCTL, "path got "DFID" from %llu #%d: %s\n",
1576                PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno,
1577                gf->gf_pathlen < 512 ? gf->gf_u.gf_path :
1578                /* only log the last 512 characters of the path */
1579                gf->gf_u.gf_path + gf->gf_pathlen - 512);
1580
1581 out:
1582         OBD_FREE(key, keylen);
1583         return rc;
1584 }
1585
1586 static int mdc_ioc_hsm_progress(struct obd_export *exp,
1587                                 struct hsm_progress_kernel *hpk)
1588 {
1589         struct obd_import               *imp = class_exp2cliimp(exp);
1590         struct hsm_progress_kernel      *req_hpk;
1591         struct ptlrpc_request           *req;
1592         int                              rc;
1593         ENTRY;
1594
1595         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_PROGRESS,
1596                                         LUSTRE_MDS_VERSION, MDS_HSM_PROGRESS);
1597         if (req == NULL)
1598                 GOTO(out, rc = -ENOMEM);
1599
1600         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1601
1602         /* Copy hsm_progress struct */
1603         req_hpk = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_PROGRESS);
1604         if (req_hpk == NULL)
1605                 GOTO(out, rc = -EPROTO);
1606
1607         *req_hpk = *hpk;
1608         req_hpk->hpk_errval = lustre_errno_hton(hpk->hpk_errval);
1609
1610         ptlrpc_request_set_replen(req);
1611
1612         mdc_get_mod_rpc_slot(req, NULL);
1613         rc = ptlrpc_queue_wait(req);
1614         mdc_put_mod_rpc_slot(req, NULL);
1615
1616         GOTO(out, rc);
1617 out:
1618         ptlrpc_req_finished(req);
1619         return rc;
1620 }
1621
1622 static int mdc_ioc_hsm_ct_register(struct obd_import *imp, __u32 archives)
1623 {
1624         __u32                   *archive_mask;
1625         struct ptlrpc_request   *req;
1626         int                      rc;
1627         ENTRY;
1628
1629         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_REGISTER,
1630                                         LUSTRE_MDS_VERSION,
1631                                         MDS_HSM_CT_REGISTER);
1632         if (req == NULL)
1633                 GOTO(out, rc = -ENOMEM);
1634
1635         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1636
1637         /* Copy hsm_progress struct */
1638         archive_mask = req_capsule_client_get(&req->rq_pill,
1639                                               &RMF_MDS_HSM_ARCHIVE);
1640         if (archive_mask == NULL)
1641                 GOTO(out, rc = -EPROTO);
1642
1643         *archive_mask = archives;
1644
1645         ptlrpc_request_set_replen(req);
1646
1647         rc = mdc_queue_wait(req);
1648         GOTO(out, rc);
1649 out:
1650         ptlrpc_req_finished(req);
1651         return rc;
1652 }
1653
1654 static int mdc_ioc_hsm_current_action(struct obd_export *exp,
1655                                       struct md_op_data *op_data)
1656 {
1657         struct hsm_current_action       *hca = op_data->op_data;
1658         struct hsm_current_action       *req_hca;
1659         struct ptlrpc_request           *req;
1660         int                              rc;
1661         ENTRY;
1662
1663         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1664                                    &RQF_MDS_HSM_ACTION);
1665         if (req == NULL)
1666                 RETURN(-ENOMEM);
1667
1668         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_ACTION);
1669         if (rc) {
1670                 ptlrpc_request_free(req);
1671                 RETURN(rc);
1672         }
1673
1674         mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1675                       op_data->op_suppgids[0], 0);
1676
1677         ptlrpc_request_set_replen(req);
1678
1679         rc = mdc_queue_wait(req);
1680         if (rc)
1681                 GOTO(out, rc);
1682
1683         req_hca = req_capsule_server_get(&req->rq_pill,
1684                                          &RMF_MDS_HSM_CURRENT_ACTION);
1685         if (req_hca == NULL)
1686                 GOTO(out, rc = -EPROTO);
1687
1688         *hca = *req_hca;
1689
1690         EXIT;
1691 out:
1692         ptlrpc_req_finished(req);
1693         return rc;
1694 }
1695
1696 static int mdc_ioc_hsm_ct_unregister(struct obd_import *imp)
1697 {
1698         struct ptlrpc_request   *req;
1699         int                      rc;
1700         ENTRY;
1701
1702         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_UNREGISTER,
1703                                         LUSTRE_MDS_VERSION,
1704                                         MDS_HSM_CT_UNREGISTER);
1705         if (req == NULL)
1706                 GOTO(out, rc = -ENOMEM);
1707
1708         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1709
1710         ptlrpc_request_set_replen(req);
1711
1712         rc = mdc_queue_wait(req);
1713         GOTO(out, rc);
1714 out:
1715         ptlrpc_req_finished(req);
1716         return rc;
1717 }
1718
1719 static int mdc_ioc_hsm_state_get(struct obd_export *exp,
1720                                  struct md_op_data *op_data)
1721 {
1722         struct hsm_user_state   *hus = op_data->op_data;
1723         struct hsm_user_state   *req_hus;
1724         struct ptlrpc_request   *req;
1725         int                      rc;
1726         ENTRY;
1727
1728         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1729                                    &RQF_MDS_HSM_STATE_GET);
1730         if (req == NULL)
1731                 RETURN(-ENOMEM);
1732
1733         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_GET);
1734         if (rc != 0) {
1735                 ptlrpc_request_free(req);
1736                 RETURN(rc);
1737         }
1738
1739         mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1740                       op_data->op_suppgids[0], 0);
1741
1742         ptlrpc_request_set_replen(req);
1743
1744         rc = mdc_queue_wait(req);
1745         if (rc)
1746                 GOTO(out, rc);
1747
1748         req_hus = req_capsule_server_get(&req->rq_pill, &RMF_HSM_USER_STATE);
1749         if (req_hus == NULL)
1750                 GOTO(out, rc = -EPROTO);
1751
1752         *hus = *req_hus;
1753
1754         EXIT;
1755 out:
1756         ptlrpc_req_finished(req);
1757         return rc;
1758 }
1759
1760 static int mdc_ioc_hsm_state_set(struct obd_export *exp,
1761                                  struct md_op_data *op_data)
1762 {
1763         struct hsm_state_set    *hss = op_data->op_data;
1764         struct hsm_state_set    *req_hss;
1765         struct ptlrpc_request   *req;
1766         int                      rc;
1767         ENTRY;
1768
1769         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1770                                    &RQF_MDS_HSM_STATE_SET);
1771         if (req == NULL)
1772                 RETURN(-ENOMEM);
1773
1774         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_SET);
1775         if (rc) {
1776                 ptlrpc_request_free(req);
1777                 RETURN(rc);
1778         }
1779
1780         mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1781                       op_data->op_suppgids[0], 0);
1782
1783         /* Copy states */
1784         req_hss = req_capsule_client_get(&req->rq_pill, &RMF_HSM_STATE_SET);
1785         if (req_hss == NULL)
1786                 GOTO(out, rc = -EPROTO);
1787         *req_hss = *hss;
1788
1789         ptlrpc_request_set_replen(req);
1790
1791         mdc_get_mod_rpc_slot(req, NULL);
1792         rc = ptlrpc_queue_wait(req);
1793         mdc_put_mod_rpc_slot(req, NULL);
1794
1795         GOTO(out, rc);
1796 out:
1797         ptlrpc_req_finished(req);
1798         return rc;
1799 }
1800
1801 static int mdc_ioc_hsm_request(struct obd_export *exp,
1802                                struct hsm_user_request *hur)
1803 {
1804         struct obd_import       *imp = class_exp2cliimp(exp);
1805         struct ptlrpc_request   *req;
1806         struct hsm_request      *req_hr;
1807         struct hsm_user_item    *req_hui;
1808         char                    *req_opaque;
1809         int                      rc;
1810         ENTRY;
1811
1812         req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_REQUEST);
1813         if (req == NULL)
1814                 GOTO(out, rc = -ENOMEM);
1815
1816         req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM, RCL_CLIENT,
1817                              hur->hur_request.hr_itemcount
1818                              * sizeof(struct hsm_user_item));
1819         req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA, RCL_CLIENT,
1820                              hur->hur_request.hr_data_len);
1821
1822         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_REQUEST);
1823         if (rc) {
1824                 ptlrpc_request_free(req);
1825                 RETURN(rc);
1826         }
1827
1828         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1829
1830         /* Copy hsm_request struct */
1831         req_hr = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_REQUEST);
1832         if (req_hr == NULL)
1833                 GOTO(out, rc = -EPROTO);
1834         *req_hr = hur->hur_request;
1835
1836         /* Copy hsm_user_item structs */
1837         req_hui = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM);
1838         if (req_hui == NULL)
1839                 GOTO(out, rc = -EPROTO);
1840         memcpy(req_hui, hur->hur_user_item,
1841                hur->hur_request.hr_itemcount * sizeof(struct hsm_user_item));
1842
1843         /* Copy opaque field */
1844         req_opaque = req_capsule_client_get(&req->rq_pill, &RMF_GENERIC_DATA);
1845         if (req_opaque == NULL)
1846                 GOTO(out, rc = -EPROTO);
1847         memcpy(req_opaque, hur_data(hur), hur->hur_request.hr_data_len);
1848
1849         ptlrpc_request_set_replen(req);
1850
1851         mdc_get_mod_rpc_slot(req, NULL);
1852         rc = ptlrpc_queue_wait(req);
1853         mdc_put_mod_rpc_slot(req, NULL);
1854
1855         GOTO(out, rc);
1856
1857 out:
1858         ptlrpc_req_finished(req);
1859         return rc;
1860 }
1861
1862 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
1863                                 struct lustre_kernelcomm *lk);
1864
1865 static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp,
1866                         struct obd_quotactl *oqctl)
1867 {
1868         struct ptlrpc_request   *req;
1869         struct obd_quotactl     *oqc;
1870         int                      rc;
1871         ENTRY;
1872
1873         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1874                                         &RQF_MDS_QUOTACTL, LUSTRE_MDS_VERSION,
1875                                         MDS_QUOTACTL);
1876         if (req == NULL)
1877                 RETURN(-ENOMEM);
1878
1879         oqc = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1880         *oqc = *oqctl;
1881
1882         ptlrpc_request_set_replen(req);
1883         ptlrpc_at_set_req_timeout(req);
1884         req->rq_no_resend = 1;
1885
1886         rc = ptlrpc_queue_wait(req);
1887         if (rc)
1888                 CERROR("ptlrpc_queue_wait failed, rc: %d\n", rc);
1889
1890         if (req->rq_repmsg &&
1891             (oqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL))) {
1892                 *oqctl = *oqc;
1893         } else if (!rc) {
1894                 CERROR ("Can't unpack obd_quotactl\n");
1895                 rc = -EPROTO;
1896         }
1897         ptlrpc_req_finished(req);
1898
1899         RETURN(rc);
1900 }
1901
1902 static int mdc_ioc_swap_layouts(struct obd_export *exp,
1903                                 struct md_op_data *op_data)
1904 {
1905         struct list_head cancels = LIST_HEAD_INIT(cancels);
1906         struct ptlrpc_request   *req;
1907         int                      rc, count;
1908         struct mdc_swap_layouts *msl, *payload;
1909         ENTRY;
1910
1911         msl = op_data->op_data;
1912
1913         /* When the MDT will get the MDS_SWAP_LAYOUTS RPC the
1914          * first thing it will do is to cancel the 2 layout
1915          * locks held by this client.
1916          * So the client must cancel its layout locks on the 2 fids
1917          * with the request RPC to avoid extra RPC round trips.
1918          */
1919         count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels,
1920                                         LCK_EX, MDS_INODELOCK_LAYOUT |
1921                                         MDS_INODELOCK_XATTR);
1922         count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels,
1923                                          LCK_EX, MDS_INODELOCK_LAYOUT |
1924                                          MDS_INODELOCK_XATTR);
1925
1926         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1927                                    &RQF_MDS_SWAP_LAYOUTS);
1928         if (req == NULL) {
1929                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
1930                 RETURN(-ENOMEM);
1931         }
1932
1933         rc = mdc_prep_elc_req(exp, req, MDS_SWAP_LAYOUTS, &cancels, count);
1934         if (rc) {
1935                 ptlrpc_request_free(req);
1936                 RETURN(rc);
1937         }
1938
1939         mdc_swap_layouts_pack(req, op_data);
1940
1941         payload = req_capsule_client_get(&req->rq_pill, &RMF_SWAP_LAYOUTS);
1942         LASSERT(payload);
1943
1944         *payload = *msl;
1945
1946         ptlrpc_request_set_replen(req);
1947
1948         rc = ptlrpc_queue_wait(req);
1949         if (rc)
1950                 GOTO(out, rc);
1951         EXIT;
1952
1953 out:
1954         ptlrpc_req_finished(req);
1955         return rc;
1956 }
1957
1958 static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1959                          void *karg, void __user *uarg)
1960 {
1961         struct obd_device *obd = exp->exp_obd;
1962         struct obd_ioctl_data *data = karg;
1963         struct obd_import *imp = obd->u.cli.cl_import;
1964         int rc;
1965         ENTRY;
1966
1967         if (!try_module_get(THIS_MODULE)) {
1968                 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
1969                        module_name(THIS_MODULE));
1970                 return -EINVAL;
1971         }
1972         switch (cmd) {
1973         case OBD_IOC_FID2PATH:
1974                 rc = mdc_ioc_fid2path(exp, karg);
1975                 GOTO(out, rc);
1976         case LL_IOC_HSM_CT_START:
1977                 rc = mdc_ioc_hsm_ct_start(exp, karg);
1978                 /* ignore if it was already registered on this MDS. */
1979                 if (rc == -EEXIST)
1980                         rc = 0;
1981                 GOTO(out, rc);
1982         case LL_IOC_HSM_PROGRESS:
1983                 rc = mdc_ioc_hsm_progress(exp, karg);
1984                 GOTO(out, rc);
1985         case LL_IOC_HSM_STATE_GET:
1986                 rc = mdc_ioc_hsm_state_get(exp, karg);
1987                 GOTO(out, rc);
1988         case LL_IOC_HSM_STATE_SET:
1989                 rc = mdc_ioc_hsm_state_set(exp, karg);
1990                 GOTO(out, rc);
1991         case LL_IOC_HSM_ACTION:
1992                 rc = mdc_ioc_hsm_current_action(exp, karg);
1993                 GOTO(out, rc);
1994         case LL_IOC_HSM_REQUEST:
1995                 rc = mdc_ioc_hsm_request(exp, karg);
1996                 GOTO(out, rc);
1997         case OBD_IOC_CLIENT_RECOVER:
1998                 rc = ptlrpc_recover_import(imp, data->ioc_inlbuf1, 0);
1999                 if (rc < 0)
2000                         GOTO(out, rc);
2001                 GOTO(out, rc = 0);
2002         case IOC_OSC_SET_ACTIVE:
2003                 rc = ptlrpc_set_import_active(imp, data->ioc_offset);
2004                 GOTO(out, rc);
2005         case OBD_IOC_PING_TARGET:
2006                 rc = ptlrpc_obd_ping(obd);
2007                 GOTO(out, rc);
2008         /*
2009          * Normally IOC_OBD_STATFS, OBD_IOC_QUOTACTL iocontrol are handled by
2010          * LMV instead of MDC. But when the cluster is upgraded from 1.8,
2011          * there'd be no LMV layer thus we might be called here. Eventually
2012          * this code should be removed.
2013          * bz20731, LU-592.
2014          */
2015         case IOC_OBD_STATFS: {
2016                 struct obd_statfs stat_buf = {0};
2017
2018                 if (*((__u32 *) data->ioc_inlbuf2) != 0)
2019                         GOTO(out, rc = -ENODEV);
2020
2021                 /* copy UUID */
2022                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(obd),
2023                                  min((int)data->ioc_plen2,
2024                                      (int)sizeof(struct obd_uuid))))
2025                         GOTO(out, rc = -EFAULT);
2026
2027                 rc = mdc_statfs(NULL, obd->obd_self_export, &stat_buf,
2028                                 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
2029                                 0);
2030                 if (rc != 0)
2031                         GOTO(out, rc);
2032
2033                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
2034                                      min((int) data->ioc_plen1,
2035                                          (int) sizeof(stat_buf))))
2036                         GOTO(out, rc = -EFAULT);
2037
2038                 GOTO(out, rc = 0);
2039         }
2040         case OBD_IOC_QUOTACTL: {
2041                 struct if_quotactl *qctl = karg;
2042                 struct obd_quotactl *oqctl;
2043
2044                 OBD_ALLOC_PTR(oqctl);
2045                 if (oqctl == NULL)
2046                         GOTO(out, rc = -ENOMEM);
2047
2048                 QCTL_COPY(oqctl, qctl);
2049                 rc = obd_quotactl(exp, oqctl);
2050                 if (rc == 0) {
2051                         QCTL_COPY(qctl, oqctl);
2052                         qctl->qc_valid = QC_MDTIDX;
2053                         qctl->obd_uuid = obd->u.cli.cl_target_uuid;
2054                 }
2055
2056                 OBD_FREE_PTR(oqctl);
2057                 GOTO(out, rc);
2058         }
2059         case LL_IOC_GET_CONNECT_FLAGS:
2060                 if (copy_to_user(uarg, exp_connect_flags_ptr(exp),
2061                                  sizeof(*exp_connect_flags_ptr(exp))))
2062                         GOTO(out, rc = -EFAULT);
2063
2064                 GOTO(out, rc = 0);
2065         case LL_IOC_LOV_SWAP_LAYOUTS:
2066                 rc = mdc_ioc_swap_layouts(exp, karg);
2067                 GOTO(out, rc);
2068         default:
2069                 CERROR("unrecognised ioctl: cmd = %#x\n", cmd);
2070                 GOTO(out, rc = -ENOTTY);
2071         }
2072 out:
2073         module_put(THIS_MODULE);
2074
2075         return rc;
2076 }
2077
2078 static int mdc_get_info_rpc(struct obd_export *exp,
2079                             u32 keylen, void *key,
2080                             u32 vallen, void *val)
2081 {
2082         struct obd_import      *imp = class_exp2cliimp(exp);
2083         struct ptlrpc_request  *req;
2084         char                   *tmp;
2085         int                     rc = -EINVAL;
2086         ENTRY;
2087
2088         req = ptlrpc_request_alloc(imp, &RQF_MDS_GET_INFO);
2089         if (req == NULL)
2090                 RETURN(-ENOMEM);
2091
2092         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY,
2093                              RCL_CLIENT, keylen);
2094         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VALLEN,
2095                              RCL_CLIENT, sizeof(vallen));
2096
2097         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_INFO);
2098         if (rc) {
2099                 ptlrpc_request_free(req);
2100                 RETURN(rc);
2101         }
2102
2103         tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
2104         memcpy(tmp, key, keylen);
2105         tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_VALLEN);
2106         memcpy(tmp, &vallen, sizeof(vallen));
2107
2108         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VAL,
2109                              RCL_SERVER, vallen);
2110         ptlrpc_request_set_replen(req);
2111
2112         rc = ptlrpc_queue_wait(req);
2113         /* -EREMOTE means the get_info result is partial, and it needs to
2114          * continue on another MDT, see fid2path part in lmv_iocontrol */
2115         if (rc == 0 || rc == -EREMOTE) {
2116                 tmp = req_capsule_server_get(&req->rq_pill, &RMF_GETINFO_VAL);
2117                 memcpy(val, tmp, vallen);
2118                 if (ptlrpc_rep_need_swab(req)) {
2119                         if (KEY_IS(KEY_FID2PATH))
2120                                 lustre_swab_fid2path(val);
2121                 }
2122         }
2123         ptlrpc_req_finished(req);
2124
2125         RETURN(rc);
2126 }
2127
2128 static void lustre_swab_hai(struct hsm_action_item *h)
2129 {
2130         __swab32s(&h->hai_len);
2131         __swab32s(&h->hai_action);
2132         lustre_swab_lu_fid(&h->hai_fid);
2133         lustre_swab_lu_fid(&h->hai_dfid);
2134         __swab64s(&h->hai_cookie);
2135         __swab64s(&h->hai_extent.offset);
2136         __swab64s(&h->hai_extent.length);
2137         __swab64s(&h->hai_gid);
2138 }
2139
2140 static void lustre_swab_hal(struct hsm_action_list *h)
2141 {
2142         struct hsm_action_item  *hai;
2143         __u32                    i;
2144
2145         __swab32s(&h->hal_version);
2146         __swab32s(&h->hal_count);
2147         __swab32s(&h->hal_archive_id);
2148         __swab64s(&h->hal_flags);
2149         hai = hai_first(h);
2150         for (i = 0; i < h->hal_count; i++, hai = hai_next(hai))
2151                 lustre_swab_hai(hai);
2152 }
2153
2154 static void lustre_swab_kuch(struct kuc_hdr *l)
2155 {
2156         __swab16s(&l->kuc_magic);
2157         /* __u8 l->kuc_transport */
2158         __swab16s(&l->kuc_msgtype);
2159         __swab16s(&l->kuc_msglen);
2160 }
2161
2162 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2163                                 struct lustre_kernelcomm *lk)
2164 {
2165         struct obd_import  *imp = class_exp2cliimp(exp);
2166         __u32               archive = lk->lk_data;
2167         int                 rc = 0;
2168
2169         if (lk->lk_group != KUC_GRP_HSM) {
2170                 CERROR("Bad copytool group %d\n", lk->lk_group);
2171                 return -EINVAL;
2172         }
2173
2174         CDEBUG(D_HSM, "CT start r%d w%d u%d g%d f%#x\n", lk->lk_rfd, lk->lk_wfd,
2175                lk->lk_uid, lk->lk_group, lk->lk_flags);
2176
2177         if (lk->lk_flags & LK_FLG_STOP) {
2178                 /* Unregister with the coordinator */
2179                 rc = mdc_ioc_hsm_ct_unregister(imp);
2180         } else {
2181                 rc = mdc_ioc_hsm_ct_register(imp, archive);
2182         }
2183
2184         return rc;
2185 }
2186
2187 /**
2188  * Send a message to any listening copytools
2189  * @param val KUC message (kuc_hdr + hsm_action_list)
2190  * @param len total length of message
2191  */
2192 static int mdc_hsm_copytool_send(const struct obd_uuid *uuid,
2193                                  size_t len, void *val)
2194 {
2195         struct kuc_hdr          *lh = (struct kuc_hdr *)val;
2196         struct hsm_action_list  *hal = (struct hsm_action_list *)(lh + 1);
2197         int                      rc;
2198         ENTRY;
2199
2200         if (len < sizeof(*lh) + sizeof(*hal)) {
2201                 CERROR("Short HSM message %zu < %zu\n", len,
2202                        sizeof(*lh) + sizeof(*hal));
2203                 RETURN(-EPROTO);
2204         }
2205         if (lh->kuc_magic == __swab16(KUC_MAGIC)) {
2206                 lustre_swab_kuch(lh);
2207                 lustre_swab_hal(hal);
2208         } else if (lh->kuc_magic != KUC_MAGIC) {
2209                 CERROR("Bad magic %x!=%x\n", lh->kuc_magic, KUC_MAGIC);
2210                 RETURN(-EPROTO);
2211         }
2212
2213         CDEBUG(D_HSM, " Received message mg=%x t=%d m=%d l=%d actions=%d "
2214                "on %s\n",
2215                lh->kuc_magic, lh->kuc_transport, lh->kuc_msgtype,
2216                lh->kuc_msglen, hal->hal_count, hal->hal_fsname);
2217
2218         /* Broadcast to HSM listeners */
2219         rc = libcfs_kkuc_group_put(uuid, KUC_GRP_HSM, lh);
2220
2221         RETURN(rc);
2222 }
2223
2224 /**
2225  * callback function passed to kuc for re-registering each HSM copytool
2226  * running on MDC, after MDT shutdown/recovery.
2227  * @param data copytool registration data
2228  * @param cb_arg callback argument (obd_import)
2229  */
2230 static int mdc_hsm_ct_reregister(void *data, void *cb_arg)
2231 {
2232         struct kkuc_ct_data     *kcd = data;
2233         struct obd_import       *imp = (struct obd_import *)cb_arg;
2234         int                      rc;
2235
2236         if (kcd == NULL || kcd->kcd_magic != KKUC_CT_DATA_MAGIC)
2237                 return -EPROTO;
2238
2239         CDEBUG(D_HA, "%s: recover copytool registration to MDT (archive=%#x)\n",
2240                imp->imp_obd->obd_name, kcd->kcd_archive);
2241         rc = mdc_ioc_hsm_ct_register(imp, kcd->kcd_archive);
2242
2243         /* ignore error if the copytool is already registered */
2244         return (rc == -EEXIST) ? 0 : rc;
2245 }
2246
2247 /**
2248  * Re-establish all kuc contexts with MDT
2249  * after MDT shutdown/recovery.
2250  */
2251 static int mdc_kuc_reregister(struct obd_import *imp)
2252 {
2253         /* re-register HSM agents */
2254         return libcfs_kkuc_group_foreach(&imp->imp_obd->obd_uuid, KUC_GRP_HSM,
2255                                          mdc_hsm_ct_reregister, imp);
2256 }
2257
2258 static int mdc_set_info_async(const struct lu_env *env,
2259                               struct obd_export *exp,
2260                               u32 keylen, void *key,
2261                               u32 vallen, void *val,
2262                               struct ptlrpc_request_set *set)
2263 {
2264         struct obd_import       *imp = class_exp2cliimp(exp);
2265         int                      rc;
2266         ENTRY;
2267
2268         if (KEY_IS(KEY_READ_ONLY)) {
2269                 if (vallen != sizeof(int))
2270                         RETURN(-EINVAL);
2271
2272                 spin_lock(&imp->imp_lock);
2273                 if (*((int *)val)) {
2274                         imp->imp_connect_flags_orig |= OBD_CONNECT_RDONLY;
2275                         imp->imp_connect_data.ocd_connect_flags |=
2276                                                         OBD_CONNECT_RDONLY;
2277                 } else {
2278                         imp->imp_connect_flags_orig &= ~OBD_CONNECT_RDONLY;
2279                         imp->imp_connect_data.ocd_connect_flags &=
2280                                                         ~OBD_CONNECT_RDONLY;
2281                 }
2282                 spin_unlock(&imp->imp_lock);
2283
2284                 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2285                                        keylen, key, vallen, val, set);
2286                 RETURN(rc);
2287         }
2288         if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2289                 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2290                                        keylen, key, vallen, val, set);
2291                 RETURN(rc);
2292         }
2293         if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) {
2294                 rc = mdc_hsm_copytool_send(&imp->imp_obd->obd_uuid, vallen,
2295                                            val);
2296                 RETURN(rc);
2297         }
2298
2299         if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2300                 __u32 *default_easize = val;
2301
2302                 exp->exp_obd->u.cli.cl_default_mds_easize = *default_easize;
2303                 RETURN(0);
2304         }
2305
2306         rc = osc_set_info_async(env, exp, keylen, key, vallen, val, set);
2307         RETURN(rc);
2308 }
2309
2310 static int mdc_get_info(const struct lu_env *env, struct obd_export *exp,
2311                         __u32 keylen, void *key, __u32 *vallen, void *val)
2312 {
2313         int rc = -EINVAL;
2314
2315         if (KEY_IS(KEY_MAX_EASIZE)) {
2316                 __u32 mdsize, *max_easize;
2317
2318                 if (*vallen != sizeof(int))
2319                         RETURN(-EINVAL);
2320                 mdsize = *(__u32 *)val;
2321                 if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
2322                         exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
2323                 max_easize = val;
2324                 *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
2325                 RETURN(0);
2326         } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2327                 __u32 *default_easize;
2328
2329                 if (*vallen != sizeof(int))
2330                         RETURN(-EINVAL);
2331                 default_easize = val;
2332                 *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
2333                 RETURN(0);
2334         } else if (KEY_IS(KEY_CONN_DATA)) {
2335                 struct obd_import *imp = class_exp2cliimp(exp);
2336                 struct obd_connect_data *data = val;
2337
2338                 if (*vallen != sizeof(*data))
2339                         RETURN(-EINVAL);
2340
2341                 *data = imp->imp_connect_data;
2342                 RETURN(0);
2343         } else if (KEY_IS(KEY_TGT_COUNT)) {
2344                 *((__u32 *)val) = 1;
2345                 RETURN(0);
2346         }
2347
2348         rc = mdc_get_info_rpc(exp, keylen, key, *vallen, val);
2349
2350         RETURN(rc);
2351 }
2352
2353 static int mdc_fsync(struct obd_export *exp, const struct lu_fid *fid,
2354                      struct ptlrpc_request **request)
2355 {
2356         struct ptlrpc_request *req;
2357         int                    rc;
2358         ENTRY;
2359
2360         *request = NULL;
2361         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SYNC);
2362         if (req == NULL)
2363                 RETURN(-ENOMEM);
2364
2365         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SYNC);
2366         if (rc) {
2367                 ptlrpc_request_free(req);
2368                 RETURN(rc);
2369         }
2370
2371         mdc_pack_body(req, fid, 0, 0, -1, 0);
2372
2373         ptlrpc_request_set_replen(req);
2374
2375         rc = ptlrpc_queue_wait(req);
2376         if (rc)
2377                 ptlrpc_req_finished(req);
2378         else
2379                 *request = req;
2380         RETURN(rc);
2381 }
2382
2383 static int mdc_import_event(struct obd_device *obd, struct obd_import *imp,
2384                             enum obd_import_event event)
2385 {
2386         struct client_obd *cli = &obd->u.cli;
2387         int rc = 0;
2388
2389         LASSERT(imp->imp_obd == obd);
2390
2391         switch (event) {
2392         case IMP_EVENT_DISCON:
2393                 spin_lock(&cli->cl_loi_list_lock);
2394                 cli->cl_avail_grant = 0;
2395                 cli->cl_lost_grant = 0;
2396                 spin_unlock(&cli->cl_loi_list_lock);
2397                 break;
2398         case IMP_EVENT_INACTIVE:
2399                 /*
2400                  * Flush current sequence to make client obtain new one
2401                  * from server in case of disconnect/reconnect.
2402                  */
2403                 down_read(&cli->cl_seq_rwsem);
2404                 if (cli->cl_seq)
2405                         seq_client_flush(cli->cl_seq);
2406                 up_read(&cli->cl_seq_rwsem);
2407
2408                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
2409                 break;
2410         case IMP_EVENT_INVALIDATE: {
2411                 struct ldlm_namespace *ns = obd->obd_namespace;
2412                 struct lu_env *env;
2413                 __u16 refcheck;
2414
2415                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2416
2417                 env = cl_env_get(&refcheck);
2418                 if (!IS_ERR(env)) {
2419                         /* Reset grants. All pages go to failing rpcs due to
2420                          * the invalid import.
2421                          */
2422                         osc_io_unplug(env, cli, NULL);
2423
2424                         cfs_hash_for_each_nolock(ns->ns_rs_hash,
2425                                                  osc_ldlm_resource_invalidate,
2426                                                  env, 0);
2427                         cl_env_put(env, &refcheck);
2428                         ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2429                 } else {
2430                         rc = PTR_ERR(env);
2431                 }
2432                 break;
2433         }
2434         case IMP_EVENT_ACTIVE:
2435                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
2436                 /* redo the kuc registration after reconnecting */
2437                 if (rc == 0)
2438                         rc = mdc_kuc_reregister(imp);
2439                 break;
2440         case IMP_EVENT_OCD: {
2441                 struct obd_connect_data *ocd = &imp->imp_connect_data;
2442
2443                 if (OCD_HAS_FLAG(ocd, GRANT))
2444                         osc_init_grant(cli, ocd);
2445
2446                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
2447                 break;
2448         }
2449         case IMP_EVENT_DEACTIVATE:
2450         case IMP_EVENT_ACTIVATE:
2451                 break;
2452         default:
2453                 CERROR("Unknown import event %x\n", event);
2454                 LBUG();
2455         }
2456         RETURN(rc);
2457 }
2458
2459 int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp,
2460                   struct lu_fid *fid, struct md_op_data *op_data)
2461 {
2462         struct client_obd *cli = &exp->exp_obd->u.cli;
2463         int rc = -EIO;
2464
2465         ENTRY;
2466
2467         down_read(&cli->cl_seq_rwsem);
2468         if (cli->cl_seq)
2469                 rc = seq_client_alloc_fid(env, cli->cl_seq, fid);
2470         up_read(&cli->cl_seq_rwsem);
2471
2472         RETURN(rc);
2473 }
2474
2475 static struct obd_uuid *mdc_get_uuid(struct obd_export *exp)
2476 {
2477         struct client_obd *cli = &exp->exp_obd->u.cli;
2478         return &cli->cl_target_uuid;
2479 }
2480
2481 /**
2482  * Determine whether the lock can be canceled before replaying it during
2483  * recovery, non zero value will be return if the lock can be canceled,
2484  * or zero returned for not
2485  */
2486 static int mdc_cancel_weight(struct ldlm_lock *lock)
2487 {
2488         if (lock->l_resource->lr_type != LDLM_IBITS)
2489                 RETURN(0);
2490
2491         /* FIXME: if we ever get into a situation where there are too many
2492          * opened files with open locks on a single node, then we really
2493          * should replay these open locks to reget it */
2494         if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN)
2495                 RETURN(0);
2496
2497         RETURN(1);
2498 }
2499
2500 static int mdc_resource_inode_free(struct ldlm_resource *res)
2501 {
2502         if (res->lr_lvb_inode)
2503                 res->lr_lvb_inode = NULL;
2504
2505         return 0;
2506 }
2507
2508 static struct ldlm_valblock_ops inode_lvbo = {
2509         .lvbo_free = mdc_resource_inode_free
2510 };
2511
2512 static int mdc_llog_init(struct obd_device *obd)
2513 {
2514         struct obd_llog_group   *olg = &obd->obd_olg;
2515         struct llog_ctxt        *ctxt;
2516         int                      rc;
2517
2518         ENTRY;
2519
2520         rc = llog_setup(NULL, obd, olg, LLOG_CHANGELOG_REPL_CTXT, obd,
2521                         &llog_client_ops);
2522         if (rc < 0)
2523                 RETURN(rc);
2524
2525         ctxt = llog_group_get_ctxt(olg, LLOG_CHANGELOG_REPL_CTXT);
2526         llog_initiator_connect(ctxt);
2527         llog_ctxt_put(ctxt);
2528
2529         RETURN(0);
2530 }
2531
2532 static void mdc_llog_finish(struct obd_device *obd)
2533 {
2534         struct llog_ctxt *ctxt;
2535
2536         ENTRY;
2537
2538         ctxt = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
2539         if (ctxt != NULL)
2540                 llog_cleanup(NULL, ctxt);
2541
2542         EXIT;
2543 }
2544
2545 int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg)
2546 {
2547         int rc;
2548
2549         ENTRY;
2550
2551         rc = osc_setup_common(obd, cfg);
2552         if (rc < 0)
2553                 RETURN(rc);
2554
2555 #ifdef CONFIG_PROC_FS
2556         obd->obd_vars = lprocfs_mdc_obd_vars;
2557         lprocfs_obd_setup(obd, false);
2558         lprocfs_alloc_md_stats(obd, 0);
2559 #endif
2560
2561         sptlrpc_lprocfs_cliobd_attach(obd);
2562         ptlrpc_lprocfs_register_obd(obd);
2563
2564         ns_register_cancel(obd->obd_namespace, mdc_cancel_weight);
2565
2566         obd->obd_namespace->ns_lvbo = &inode_lvbo;
2567
2568         rc = mdc_llog_init(obd);
2569         if (rc) {
2570                 CERROR("%s: failed to setup llogging subsystems: rc = %d\n",
2571                        obd->obd_name, rc);
2572                 GOTO(err_llog_cleanup, rc);
2573         }
2574
2575         rc = mdc_changelog_cdev_init(obd);
2576         if (rc) {
2577                 CERROR("%s: failed to setup changelog char device: rc = %d\n",
2578                        obd->obd_name, rc);
2579                 GOTO(err_changelog_cleanup, rc);
2580         }
2581
2582         RETURN(rc);
2583
2584 err_changelog_cleanup:
2585         mdc_llog_finish(obd);
2586 err_llog_cleanup:
2587         ptlrpc_lprocfs_unregister_obd(obd);
2588         lprocfs_free_md_stats(obd);
2589
2590         osc_cleanup_common(obd);
2591         return rc;
2592 }
2593
2594 /* Initialize the default and maximum LOV EA sizes.  This allows
2595  * us to make MDS RPCs with large enough reply buffers to hold a default
2596  * sized EA without having to calculate this (via a call into the
2597  * LOV + OSCs) each time we make an RPC.  The maximum size is also tracked
2598  * but not used to avoid wastefully vmalloc()'ing large reply buffers when
2599  * a large number of stripes is possible.  If a larger reply buffer is
2600  * required it will be reallocated in the ptlrpc layer due to overflow.
2601  */
2602 static int mdc_init_ea_size(struct obd_export *exp, __u32 easize,
2603                             __u32 def_easize)
2604 {
2605         struct obd_device *obd = exp->exp_obd;
2606         struct client_obd *cli = &obd->u.cli;
2607         ENTRY;
2608
2609         if (cli->cl_max_mds_easize < easize)
2610                 cli->cl_max_mds_easize = easize;
2611
2612         if (cli->cl_default_mds_easize < def_easize)
2613                 cli->cl_default_mds_easize = def_easize;
2614
2615         RETURN(0);
2616 }
2617
2618 static int mdc_precleanup(struct obd_device *obd)
2619 {
2620         ENTRY;
2621
2622         osc_precleanup_common(obd);
2623         mdc_changelog_cdev_finish(obd);
2624
2625         obd_cleanup_client_import(obd);
2626         ptlrpc_lprocfs_unregister_obd(obd);
2627         lprocfs_free_md_stats(obd);
2628         mdc_llog_finish(obd);
2629         RETURN(0);
2630 }
2631
2632 static int mdc_cleanup(struct obd_device *obd)
2633 {
2634         return osc_cleanup_common(obd);
2635 }
2636
2637 int mdc_process_config(struct obd_device *obd, size_t len, void *buf)
2638 {
2639         struct lustre_cfg *lcfg = buf;
2640         int rc;
2641
2642         rc = class_process_proc_param(PARAM_MDC, obd->obd_vars, lcfg, obd);
2643         return (rc > 0 ? 0: rc);
2644 }
2645
2646 static struct obd_ops mdc_obd_ops = {
2647         .o_owner            = THIS_MODULE,
2648         .o_setup            = mdc_setup,
2649         .o_precleanup       = mdc_precleanup,
2650         .o_cleanup          = mdc_cleanup,
2651         .o_add_conn         = client_import_add_conn,
2652         .o_del_conn         = client_import_del_conn,
2653         .o_connect          = client_connect_import,
2654         .o_reconnect        = osc_reconnect,
2655         .o_disconnect       = osc_disconnect,
2656         .o_iocontrol        = mdc_iocontrol,
2657         .o_set_info_async   = mdc_set_info_async,
2658         .o_statfs           = mdc_statfs,
2659         .o_fid_init         = client_fid_init,
2660         .o_fid_fini         = client_fid_fini,
2661         .o_fid_alloc        = mdc_fid_alloc,
2662         .o_import_event     = mdc_import_event,
2663         .o_get_info         = mdc_get_info,
2664         .o_process_config   = mdc_process_config,
2665         .o_get_uuid         = mdc_get_uuid,
2666         .o_quotactl         = mdc_quotactl,
2667 };
2668
2669 static struct md_ops mdc_md_ops = {
2670         .m_get_root         = mdc_get_root,
2671         .m_null_inode       = mdc_null_inode,
2672         .m_close            = mdc_close,
2673         .m_create           = mdc_create,
2674         .m_enqueue          = mdc_enqueue,
2675         .m_getattr          = mdc_getattr,
2676         .m_getattr_name     = mdc_getattr_name,
2677         .m_intent_lock      = mdc_intent_lock,
2678         .m_link             = mdc_link,
2679         .m_rename           = mdc_rename,
2680         .m_setattr          = mdc_setattr,
2681         .m_setxattr         = mdc_setxattr,
2682         .m_getxattr         = mdc_getxattr,
2683         .m_fsync                = mdc_fsync,
2684         .m_file_resync          = mdc_file_resync,
2685         .m_read_page            = mdc_read_page,
2686         .m_unlink           = mdc_unlink,
2687         .m_cancel_unused    = mdc_cancel_unused,
2688         .m_init_ea_size     = mdc_init_ea_size,
2689         .m_set_lock_data    = mdc_set_lock_data,
2690         .m_lock_match       = mdc_lock_match,
2691         .m_get_lustre_md    = mdc_get_lustre_md,
2692         .m_free_lustre_md   = mdc_free_lustre_md,
2693         .m_set_open_replay_data = mdc_set_open_replay_data,
2694         .m_clear_open_replay_data = mdc_clear_open_replay_data,
2695         .m_intent_getattr_async = mdc_intent_getattr_async,
2696         .m_revalidate_lock      = mdc_revalidate_lock
2697 };
2698
2699 static int __init mdc_init(void)
2700 {
2701         return class_register_type(&mdc_obd_ops, &mdc_md_ops, true, NULL,
2702                                    LUSTRE_MDC_NAME, &mdc_device_type);
2703 }
2704
2705 static void __exit mdc_exit(void)
2706 {
2707         class_unregister_type(LUSTRE_MDC_NAME);
2708 }
2709
2710 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2711 MODULE_DESCRIPTION("Lustre Metadata Client");
2712 MODULE_VERSION(LUSTRE_VERSION_STRING);
2713 MODULE_LICENSE("GPL");
2714
2715 module_init(mdc_init);
2716 module_exit(mdc_exit);