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 <lustre_swab.h>
  57 #include <obd_class.h>
  58 #include <lustre_osc.h>
  59
  60 #include "mdc_internal.h"
  61
  62 #define REQUEST_MINOR 244
  63
  64 static int mdc_cleanup(struct obd_device *obd);
  65
  66 static inline int mdc_queue_wait(struct ptlrpc_request *req)
  67 {
  68         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
  69         int rc;
  70
  71         /* obd_get_request_slot() ensures that this client has no more
  72          * than cl_max_rpcs_in_flight RPCs simultaneously inf light
  73          * against an MDT. */
  74         rc = obd_get_request_slot(cli);
  75         if (rc != 0)
  76                 return rc;
  77
  78         rc = ptlrpc_queue_wait(req);
  79         obd_put_request_slot(cli);
  80
  81         return rc;
  82 }
  83
  84 /*
  85  * Send MDS_GET_ROOT RPC to fetch root FID.
  86  *
  87  * If \a fileset is not NULL it should contain a subdirectory off
  88  * the ROOT/ directory to be mounted on the client. Return the FID
  89  * of the subdirectory to the client to mount onto its mountpoint.
  90  *
  91  * \param[in]   imp     MDC import
  92  * \param[in]   fileset fileset name, which could be NULL
  93  * \param[out]  rootfid root FID of this mountpoint
  94  * \param[out]  pc      root capa will be unpacked and saved in this pointer
  95  *
  96  * \retval      0 on success, negative errno on failure
  97  */
  98 static int mdc_get_root(struct obd_export *exp, const char *fileset,
  99                          struct lu_fid *rootfid)
 100 {
 101         struct ptlrpc_request   *req;
 102         struct mdt_body         *body;
 103         int                      rc;
 104
 105         ENTRY;
 106
 107         if (fileset && !(exp_connect_flags(exp) & OBD_CONNECT_SUBTREE))
 108                 RETURN(-ENOTSUPP);
 109
 110         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 111                                 &RQF_MDS_GET_ROOT);
 112         if (req == NULL)
 113                 RETURN(-ENOMEM);
 114
 115         if (fileset != NULL)
 116                 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 117                                      strlen(fileset) + 1);
 118         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_ROOT);
 119         if (rc) {
 120                 ptlrpc_request_free(req);
 121                 RETURN(rc);
 122         }
 123         mdc_pack_body(req, NULL, 0, 0, -1, 0);
 124         if (fileset != NULL) {
 125                 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
 126
 127                 memcpy(name, fileset, strlen(fileset));
 128         }
 129         lustre_msg_add_flags(req->rq_reqmsg, LUSTRE_IMP_FULL);
 130         req->rq_send_state = LUSTRE_IMP_FULL;
 131
 132         ptlrpc_request_set_replen(req);
 133
 134         rc = ptlrpc_queue_wait(req);
 135         if (rc)
 136                 GOTO(out, rc);
 137
 138         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 139         if (body == NULL)
 140                 GOTO(out, rc = -EPROTO);
 141
 142         *rootfid = body->mbo_fid1;
 143         CDEBUG(D_NET, "root fid="DFID", last_committed=%llu\n",
 144                PFID(rootfid), lustre_msg_get_last_committed(req->rq_repmsg));
 145         EXIT;
 146 out:
 147         ptlrpc_req_finished(req);
 148
 149         return rc;
 150 }
 151
 152 /*
 153  * This function now is known to always saying that it will receive 4 buffers
 154  * from server. Even for cases when acl_size and md_size is zero, RPC header
 155  * will contain 4 fields and RPC itself will contain zero size fields. This is
 156  * because mdt_getattr*() _always_ returns 4 fields, but if acl is not needed
 157  * and thus zero, it shrinks it, making zero size. The same story about
 158  * md_size. And this is course of problem when client waits for smaller number
 159  * of fields. This issue will be fixed later when client gets aware of RPC
 160  * layouts.  --umka
 161  */
 162 static int mdc_getattr_common(struct obd_export *exp,
 163                               struct ptlrpc_request *req)
 164 {
 165         struct req_capsule *pill = &req->rq_pill;
 166         struct mdt_body    *body;
 167         void               *eadata;
 168         int                 rc;
 169         ENTRY;
 170
 171         /* Request message already built. */
 172         rc = ptlrpc_queue_wait(req);
 173         if (rc != 0)
 174                 RETURN(rc);
 175
 176         /* sanity check for the reply */
 177         body = req_capsule_server_get(pill, &RMF_MDT_BODY);
 178         if (body == NULL)
 179                 RETURN(-EPROTO);
 180
 181         CDEBUG(D_NET, "mode: %o\n", body->mbo_mode);
 182
 183         mdc_update_max_ea_from_body(exp, body);
 184         if (body->mbo_eadatasize != 0) {
 185                 eadata = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
 186                                                       body->mbo_eadatasize);
 187                 if (eadata == NULL)
 188                         RETURN(-EPROTO);
 189         }
 190
 191         RETURN(0);
 192 }
 193
 194 static void mdc_reset_acl_req(struct ptlrpc_request *req)
 195 {
 196         spin_lock(&req->rq_early_free_lock);
 197         sptlrpc_cli_free_repbuf(req);
 198         req->rq_repbuf = NULL;
 199         req->rq_repbuf_len = 0;
 200         req->rq_repdata = NULL;
 201         req->rq_reqdata_len = 0;
 202         spin_unlock(&req->rq_early_free_lock);
 203 }
 204
 205 static int mdc_getattr(struct obd_export *exp, struct md_op_data *op_data,
 206                        struct ptlrpc_request **request)
 207 {
 208         struct ptlrpc_request *req;
 209         struct obd_import *imp = class_exp2cliimp(exp);
 210         __u32 acl_bufsize = LUSTRE_POSIX_ACL_MAX_SIZE_OLD;
 211         int rc;
 212         ENTRY;
 213
 214         /* Single MDS without an LMV case */
 215         if (op_data->op_flags & MF_GET_MDT_IDX) {
 216                 op_data->op_mds = 0;
 217                 RETURN(0);
 218         }
 219
 220         *request = NULL;
 221         req = ptlrpc_request_alloc(imp, &RQF_MDS_GETATTR);
 222         if (req == NULL)
 223                 RETURN(-ENOMEM);
 224
 225         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR);
 226         if (rc) {
 227                 ptlrpc_request_free(req);
 228                 RETURN(rc);
 229         }
 230
 231 again:
 232         mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
 233                       op_data->op_mode, -1, 0);
 234         req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
 235         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 236                              op_data->op_mode);
 237         ptlrpc_request_set_replen(req);
 238
 239         rc = mdc_getattr_common(exp, req);
 240         if (rc) {
 241                 if (rc == -ERANGE) {
 242                         acl_bufsize = MIN(imp->imp_connect_data.ocd_max_easize,
 243                                           XATTR_SIZE_MAX);
 244                         mdc_reset_acl_req(req);
 245                         goto again;
 246                 }
 247
 248                 ptlrpc_req_finished(req);
 249         } else {
 250                 *request = req;
 251         }
 252
 253         RETURN(rc);
 254 }
 255
 256 static int mdc_getattr_name(struct obd_export *exp, struct md_op_data *op_data,
 257                             struct ptlrpc_request **request)
 258 {
 259         struct ptlrpc_request *req;
 260         struct obd_import *imp = class_exp2cliimp(exp);
 261         __u32 acl_bufsize = LUSTRE_POSIX_ACL_MAX_SIZE_OLD;
 262         int rc;
 263         ENTRY;
 264
 265         *request = NULL;
 266         req = ptlrpc_request_alloc(imp, &RQF_MDS_GETATTR_NAME);
 267         if (req == NULL)
 268                 RETURN(-ENOMEM);
 269
 270         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 271                              op_data->op_namelen + 1);
 272
 273         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR_NAME);
 274         if (rc) {
 275                 ptlrpc_request_free(req);
 276                 RETURN(rc);
 277         }
 278
 279         if (op_data->op_name) {
 280                 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
 281                 LASSERT(strnlen(op_data->op_name, op_data->op_namelen) ==
 282                                 op_data->op_namelen);
 283                 memcpy(name, op_data->op_name, op_data->op_namelen);
 284         }
 285
 286 again:
 287         mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
 288                       op_data->op_mode, op_data->op_suppgids[0], 0);
 289         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 290                              op_data->op_mode);
 291         req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
 292         ptlrpc_request_set_replen(req);
 293
 294         rc = mdc_getattr_common(exp, req);
 295         if (rc) {
 296                 if (rc == -ERANGE) {
 297                         acl_bufsize = MIN(imp->imp_connect_data.ocd_max_easize,
 298                                           XATTR_SIZE_MAX);
 299                         mdc_reset_acl_req(req);
 300                         goto again;
 301                 }
 302
 303                 ptlrpc_req_finished(req);
 304         } else {
 305                 *request = req;
 306         }
 307
 308         RETURN(rc);
 309 }
 310
 311 static int mdc_xattr_common(struct obd_export *exp,const struct req_format *fmt,
 312                             const struct lu_fid *fid, int opcode, u64 valid,
 313                             const char *xattr_name, const char *input,
 314                             int input_size, int output_size, int flags,
 315                             __u32 suppgid, struct ptlrpc_request **request)
 316 {
 317         struct ptlrpc_request *req;
 318         int   xattr_namelen = 0;
 319         char *tmp;
 320         int   rc;
 321         ENTRY;
 322
 323         *request = NULL;
 324         req = ptlrpc_request_alloc(class_exp2cliimp(exp), fmt);
 325         if (req == NULL)
 326                 RETURN(-ENOMEM);
 327
 328         if (xattr_name) {
 329                 xattr_namelen = strlen(xattr_name) + 1;
 330                 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 331                                      xattr_namelen);
 332         }
 333         if (input_size)
 334                 LASSERT(input);
 335         req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT,
 336                              input_size);
 337
 338         /* get SELinux policy info if any */
 339         rc = sptlrpc_get_sepol(req);
 340         if (rc < 0) {
 341                 ptlrpc_request_free(req);
 342                 RETURN(rc);
 343         }
 344         req_capsule_set_size(&req->rq_pill, &RMF_SELINUX_POL, RCL_CLIENT,
 345                              strlen(req->rq_sepol) ?
 346                              strlen(req->rq_sepol) + 1 : 0);
 347
 348         /* Flush local XATTR locks to get rid of a possible cancel RPC */
 349         if (opcode == MDS_REINT && fid_is_sane(fid) &&
 350             exp->exp_connect_data.ocd_ibits_known & MDS_INODELOCK_XATTR) {
 351                 struct list_head cancels = LIST_HEAD_INIT(cancels);
 352                 int count;
 353
 354                 /* Without that packing would fail */
 355                 if (input_size == 0)
 356                         req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
 357                                              RCL_CLIENT, 0);
 358
 359                 count = mdc_resource_get_unused(exp, fid,
 360                                                 &cancels, LCK_EX,
 361                                                 MDS_INODELOCK_XATTR);
 362
 363                 rc = mdc_prep_elc_req(exp, req, MDS_REINT, &cancels, count);
 364                 if (rc) {
 365                         ptlrpc_request_free(req);
 366                         RETURN(rc);
 367                 }
 368         } else {
 369                 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, opcode);
 370                 if (rc) {
 371                         ptlrpc_request_free(req);
 372                         RETURN(rc);
 373                 }
 374         }
 375
 376         if (opcode == MDS_REINT) {
 377                 struct mdt_rec_setxattr *rec;
 378
 379                 CLASSERT(sizeof(struct mdt_rec_setxattr) ==
 380                          sizeof(struct mdt_rec_reint));
 381                 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
 382                 rec->sx_opcode = REINT_SETXATTR;
 383                 rec->sx_fsuid  = from_kuid(&init_user_ns, current_fsuid());
 384                 rec->sx_fsgid  = from_kgid(&init_user_ns, current_fsgid());
 385                 rec->sx_cap    = cfs_curproc_cap_pack();
 386                 rec->sx_suppgid1 = suppgid;
 387                 rec->sx_suppgid2 = -1;
 388                 rec->sx_fid    = *fid;
 389                 rec->sx_valid  = valid | OBD_MD_FLCTIME;
 390                 rec->sx_time   = ktime_get_real_seconds();
 391                 rec->sx_size   = output_size;
 392                 rec->sx_flags  = flags;
 393         } else {
 394                 mdc_pack_body(req, fid, valid, output_size, suppgid, flags);
 395         }
 396
 397         if (xattr_name) {
 398                 tmp = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
 399                 memcpy(tmp, xattr_name, xattr_namelen);
 400         }
 401         if (input_size) {
 402                 tmp = req_capsule_client_get(&req->rq_pill, &RMF_EADATA);
 403                 memcpy(tmp, input, input_size);
 404         }
 405
 406         mdc_file_sepol_pack(req);
 407
 408         if (req_capsule_has_field(&req->rq_pill, &RMF_EADATA, RCL_SERVER))
 409                 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
 410                                      RCL_SERVER, output_size);
 411         ptlrpc_request_set_replen(req);
 412
 413         /* make rpc */
 414         if (opcode == MDS_REINT)
 415                 mdc_get_mod_rpc_slot(req, NULL);
 416
 417         rc = ptlrpc_queue_wait(req);
 418
 419         if (opcode == MDS_REINT)
 420                 mdc_put_mod_rpc_slot(req, NULL);
 421
 422         if (rc)
 423                 ptlrpc_req_finished(req);
 424         else
 425                 *request = req;
 426         RETURN(rc);
 427 }
 428
 429 static int mdc_setxattr(struct obd_export *exp, const struct lu_fid *fid,
 430                         u64 obd_md_valid, const char *name,
 431                         const void *value, size_t value_size,
 432                         unsigned int xattr_flags, u32 suppgid,
 433                         struct ptlrpc_request **req)
 434 {
 435         LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
 436                 obd_md_valid == OBD_MD_FLXATTRRM);
 437
 438         return mdc_xattr_common(exp, &RQF_MDS_REINT_SETXATTR,
 439                                 fid, MDS_REINT, obd_md_valid, name,
 440                                 value, value_size, 0, xattr_flags, suppgid,
 441                                 req);
 442 }
 443
 444 static int mdc_getxattr(struct obd_export *exp, const struct lu_fid *fid,
 445                         u64 obd_md_valid, const char *name, size_t buf_size,
 446                         struct ptlrpc_request **req)
 447 {
 448         struct mdt_body *body;
 449         int rc;
 450
 451         LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
 452                 obd_md_valid == OBD_MD_FLXATTRLS);
 453
 454         CDEBUG(D_INFO, "%s: get xattr '%s' for "DFID"\n",
 455                exp->exp_obd->obd_name, name, PFID(fid));
 456         rc = mdc_xattr_common(exp, &RQF_MDS_GETXATTR, fid, MDS_GETXATTR,
 457                               obd_md_valid, name, NULL, 0, buf_size, 0, -1,
 458                               req);
 459         if (rc < 0)
 460                 GOTO(out, rc);
 461
 462         body = req_capsule_server_get(&(*req)->rq_pill, &RMF_MDT_BODY);
 463         if (body == NULL)
 464                 GOTO(out, rc = -EPROTO);
 465
 466         /* only detect the xattr size */
 467         if (buf_size == 0) {
 468                 /* LU-11109: Older MDTs do not distinguish
 469                  * between nonexistent xattrs and zero length
 470                  * values in this case. Newer MDTs will return
 471                  * -ENODATA or set OBD_MD_FLXATTR. */
 472                 GOTO(out, rc = body->mbo_eadatasize);
 473         }
 474
 475         if (body->mbo_eadatasize == 0) {
 476                 /* LU-11109: Newer MDTs set OBD_MD_FLXATTR on
 477                  * success so that we can distinguish between
 478                  * zero length value and nonexistent xattr.
 479                  *
 480                  * If OBD_MD_FLXATTR is not set then we keep
 481                  * the old behavior and return -ENODATA for
 482                  * getxattr() when mbo_eadatasize is 0. But
 483                  * -ENODATA only makes sense for getxattr()
 484                  * and not for listxattr(). */
 485                 if (body->mbo_valid & OBD_MD_FLXATTR)
 486                         GOTO(out, rc = 0);
 487                 else if (obd_md_valid == OBD_MD_FLXATTR)
 488                         GOTO(out, rc = -ENODATA);
 489                 else
 490                         GOTO(out, rc = 0);
 491         }
 492
 493         GOTO(out, rc = body->mbo_eadatasize);
 494 out:
 495         if (rc < 0) {
 496                 ptlrpc_req_finished(*req);
 497                 *req = NULL;
 498         }
 499
 500         return rc;
 501 }
 502
 503 #ifdef CONFIG_FS_POSIX_ACL
 504 static int mdc_unpack_acl(struct ptlrpc_request *req, struct lustre_md *md)
 505 {
 506         struct req_capsule     *pill = &req->rq_pill;
 507         struct mdt_body        *body = md->body;
 508         struct posix_acl       *acl;
 509         void                   *buf;
 510         int                     rc;
 511         ENTRY;
 512
 513         if (!body->mbo_aclsize)
 514                 RETURN(0);
 515
 516         buf = req_capsule_server_sized_get(pill, &RMF_ACL, body->mbo_aclsize);
 517
 518         if (!buf)
 519                 RETURN(-EPROTO);
 520
 521         acl = posix_acl_from_xattr(&init_user_ns, buf, body->mbo_aclsize);
 522         if (acl == NULL)
 523                 RETURN(0);
 524         if (IS_ERR(acl)) {
 525                 rc = PTR_ERR(acl);
 526                 CERROR("convert xattr to acl: %d\n", rc);
 527                 RETURN(rc);
 528         }
 529
 530         rc = posix_acl_valid(&init_user_ns, acl);
 531         if (rc) {
 532                 CERROR("validate acl: %d\n", rc);
 533                 posix_acl_release(acl);
 534                 RETURN(rc);
 535         }
 536
 537         md->posix_acl = acl;
 538         RETURN(0);
 539 }
 540 #else
 541 #define mdc_unpack_acl(req, md) 0
 542 #endif
 543
 544 int mdc_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
 545                       struct obd_export *dt_exp, struct obd_export *md_exp,
 546                       struct lustre_md *md)
 547 {
 548         struct req_capsule *pill = &req->rq_pill;
 549         int rc;
 550         ENTRY;
 551
 552         LASSERT(md);
 553         memset(md, 0, sizeof(*md));
 554
 555         md->body = req_capsule_server_get(pill, &RMF_MDT_BODY);
 556         LASSERT(md->body != NULL);
 557
 558         if (md->body->mbo_valid & OBD_MD_FLEASIZE) {
 559                 if (!S_ISREG(md->body->mbo_mode)) {
 560                         CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, should be a "
 561                                "regular file, but is not\n");
 562                         GOTO(out, rc = -EPROTO);
 563                 }
 564
 565                 if (md->body->mbo_eadatasize == 0) {
 566                         CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, "
 567                                "but eadatasize 0\n");
 568                         GOTO(out, rc = -EPROTO);
 569                 }
 570
 571                 md->layout.lb_len = md->body->mbo_eadatasize;
 572                 md->layout.lb_buf = req_capsule_server_sized_get(pill,
 573                                                         &RMF_MDT_MD,
 574                                                         md->layout.lb_len);
 575                 if (md->layout.lb_buf == NULL)
 576                         GOTO(out, rc = -EPROTO);
 577         } else if (md->body->mbo_valid & OBD_MD_FLDIREA) {
 578                 const union lmv_mds_md *lmv;
 579                 size_t lmv_size;
 580
 581                 if (!S_ISDIR(md->body->mbo_mode)) {
 582                         CDEBUG(D_INFO, "OBD_MD_FLDIREA set, should be a "
 583                                "directory, but is not\n");
 584                         GOTO(out, rc = -EPROTO);
 585                 }
 586
 587                 lmv_size = md->body->mbo_eadatasize;
 588                 if (lmv_size == 0) {
 589                         CDEBUG(D_INFO, "OBD_MD_FLDIREA is set, "
 590                                "but eadatasize 0\n");
 591                         RETURN(-EPROTO);
 592                 }
 593
 594                 if (md->body->mbo_valid & OBD_MD_MEA) {
 595                         lmv = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
 596                                                            lmv_size);
 597                         if (lmv == NULL)
 598                                 GOTO(out, rc = -EPROTO);
 599
 600                         rc = md_unpackmd(md_exp, &md->lmv, lmv, lmv_size);
 601                         if (rc < 0)
 602                                 GOTO(out, rc);
 603
 604                         if (rc < (typeof(rc))sizeof(*md->lmv)) {
 605                                 struct lmv_foreign_md   *lfm = md->lfm;
 606
 607                                 /* short (< sizeof(struct lmv_stripe_md))
 608                                  * foreign LMV case
 609                                  */
 610                                 if (lfm->lfm_magic != LMV_MAGIC_FOREIGN) {
 611                                         CDEBUG(D_INFO,
 612                                                "size too small: rc < sizeof(*md->lmv) (%d < %d)\n",
 613                                                rc, (int)sizeof(*md->lmv));
 614                                         GOTO(out, rc = -EPROTO);
 615                                 }
 616                         }
 617                 }
 618         }
 619         rc = 0;
 620
 621         if (md->body->mbo_valid & OBD_MD_FLACL) {
 622                 /* for ACL, it's possible that FLACL is set but aclsize is zero.
 623                  * only when aclsize != 0 there's an actual segment for ACL
 624                  * in reply buffer.
 625                  */
 626                 if (md->body->mbo_aclsize) {
 627                         rc = mdc_unpack_acl(req, md);
 628                         if (rc)
 629                                 GOTO(out, rc);
 630 #ifdef CONFIG_FS_POSIX_ACL
 631                 } else {
 632                         md->posix_acl = NULL;
 633 #endif
 634                 }
 635         }
 636
 637         EXIT;
 638 out:
 639         if (rc) {
 640 #ifdef CONFIG_FS_POSIX_ACL
 641                 posix_acl_release(md->posix_acl);
 642 #endif
 643         }
 644         return rc;
 645 }
 646
 647 int mdc_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
 648 {
 649         ENTRY;
 650         RETURN(0);
 651 }
 652
 653 void mdc_replay_open(struct ptlrpc_request *req)
 654 {
 655         struct md_open_data *mod = req->rq_cb_data;
 656         struct ptlrpc_request *close_req;
 657         struct obd_client_handle *och;
 658         struct lustre_handle old_open_handle = { };
 659         struct mdt_body *body;
 660         ENTRY;
 661
 662         if (mod == NULL) {
 663                 DEBUG_REQ(D_ERROR, req,
 664                           "Can't properly replay without open data.");
 665                 EXIT;
 666                 return;
 667         }
 668
 669         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 670         LASSERT(body != NULL);
 671
 672         spin_lock(&req->rq_lock);
 673         och = mod->mod_och;
 674         if (och && och->och_open_handle.cookie)
 675                 req->rq_early_free_repbuf = 1;
 676         else
 677                 req->rq_early_free_repbuf = 0;
 678         spin_unlock(&req->rq_lock);
 679
 680         if (req->rq_early_free_repbuf) {
 681                 struct lustre_handle *file_open_handle;
 682
 683                 LASSERT(och->och_magic == OBD_CLIENT_HANDLE_MAGIC);
 684
 685                 file_open_handle = &och->och_open_handle;
 686                 CDEBUG(D_HA, "updating handle from %#llx to %#llx\n",
 687                        file_open_handle->cookie, body->mbo_open_handle.cookie);
 688                 old_open_handle = *file_open_handle;
 689                 *file_open_handle = body->mbo_open_handle;
 690         }
 691
 692         close_req = mod->mod_close_req;
 693         if (close_req) {
 694                 __u32 opc = lustre_msg_get_opc(close_req->rq_reqmsg);
 695                 struct mdt_ioepoch *epoch;
 696
 697                 LASSERT(opc == MDS_CLOSE);
 698                 epoch = req_capsule_client_get(&close_req->rq_pill,
 699                                                &RMF_MDT_EPOCH);
 700                 LASSERT(epoch);
 701
 702                 if (req->rq_early_free_repbuf)
 703                         LASSERT(old_open_handle.cookie ==
 704                                 epoch->mio_open_handle.cookie);
 705
 706                 DEBUG_REQ(D_HA, close_req, "updating close body with new fh");
 707                 epoch->mio_open_handle = body->mbo_open_handle;
 708         }
 709         EXIT;
 710 }
 711
 712 void mdc_commit_open(struct ptlrpc_request *req)
 713 {
 714         struct md_open_data *mod = req->rq_cb_data;
 715         if (mod == NULL)
 716                 return;
 717
 718         /**
 719          * No need to touch md_open_data::mod_och, it holds a reference on
 720          * \var mod and will zero references to each other, \var mod will be
 721          * freed after that when md_open_data::mod_och will put the reference.
 722          */
 723
 724         /**
 725          * Do not let open request to disappear as it still may be needed
 726          * for close rpc to happen (it may happen on evict only, otherwise
 727          * ptlrpc_request::rq_replay does not let mdc_commit_open() to be
 728          * called), just mark this rpc as committed to distinguish these 2
 729          * cases, see mdc_close() for details. The open request reference will
 730          * be put along with freeing \var mod.
 731          */
 732         ptlrpc_request_addref(req);
 733         spin_lock(&req->rq_lock);
 734         req->rq_committed = 1;
 735         spin_unlock(&req->rq_lock);
 736         req->rq_cb_data = NULL;
 737         obd_mod_put(mod);
 738 }
 739
 740 int mdc_set_open_replay_data(struct obd_export *exp,
 741                              struct obd_client_handle *och,
 742                              struct lookup_intent *it)
 743 {
 744         struct md_open_data     *mod;
 745         struct mdt_rec_create   *rec;
 746         struct mdt_body         *body;
 747         struct ptlrpc_request   *open_req = it->it_request;
 748         struct obd_import       *imp = open_req->rq_import;
 749         ENTRY;
 750
 751         if (!open_req->rq_replay)
 752                 RETURN(0);
 753
 754         rec = req_capsule_client_get(&open_req->rq_pill, &RMF_REC_REINT);
 755         body = req_capsule_server_get(&open_req->rq_pill, &RMF_MDT_BODY);
 756         LASSERT(rec != NULL);
 757         /* Incoming message in my byte order (it's been swabbed). */
 758         /* Outgoing messages always in my byte order. */
 759         LASSERT(body != NULL);
 760
 761         /* Only if the import is replayable, we set replay_open data */
 762         if (och && imp->imp_replayable) {
 763                 mod = obd_mod_alloc();
 764                 if (mod == NULL) {
 765                         DEBUG_REQ(D_ERROR, open_req,
 766                                   "Can't allocate md_open_data");
 767                         RETURN(0);
 768                 }
 769
 770                 /**
 771                  * Take a reference on \var mod, to be freed on mdc_close().
 772                  * It protects \var mod from being freed on eviction (commit
 773                  * callback is called despite rq_replay flag).
 774                  * Another reference for \var och.
 775                  */
 776                 obd_mod_get(mod);
 777                 obd_mod_get(mod);
 778
 779                 spin_lock(&open_req->rq_lock);
 780                 och->och_mod = mod;
 781                 mod->mod_och = och;
 782                 mod->mod_is_create = it_disposition(it, DISP_OPEN_CREATE) ||
 783                                      it_disposition(it, DISP_OPEN_STRIPE);
 784                 mod->mod_open_req = open_req;
 785                 open_req->rq_cb_data = mod;
 786                 open_req->rq_commit_cb = mdc_commit_open;
 787                 open_req->rq_early_free_repbuf = 1;
 788                 spin_unlock(&open_req->rq_lock);
 789         }
 790
 791         rec->cr_fid2 = body->mbo_fid1;
 792         rec->cr_open_handle_old = body->mbo_open_handle;
 793         open_req->rq_replay_cb = mdc_replay_open;
 794         if (!fid_is_sane(&body->mbo_fid1)) {
 795                 DEBUG_REQ(D_ERROR, open_req,
 796                           "saving replay request with insane FID " DFID,
 797                           PFID(&body->mbo_fid1));
 798                 LBUG();
 799         }
 800
 801         DEBUG_REQ(D_RPCTRACE, open_req, "Set up open replay data");
 802         RETURN(0);
 803 }
 804
 805 static void mdc_free_open(struct md_open_data *mod)
 806 {
 807         int committed = 0;
 808
 809         if (mod->mod_is_create == 0 &&
 810             imp_connect_disp_stripe(mod->mod_open_req->rq_import))
 811                 committed = 1;
 812
 813         /**
 814          * No reason to asssert here if the open request has
 815          * rq_replay == 1. It means that mdc_close failed, and
 816          * close request wasn`t sent. It is not fatal to client.
 817          * The worst thing is eviction if the client gets open lock
 818          **/
 819
 820         DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "free open request rq_replay"
 821                   "= %d\n", mod->mod_open_req->rq_replay);
 822
 823         ptlrpc_request_committed(mod->mod_open_req, committed);
 824         if (mod->mod_close_req)
 825                 ptlrpc_request_committed(mod->mod_close_req, committed);
 826 }
 827
 828 int mdc_clear_open_replay_data(struct obd_export *exp,
 829                                struct obd_client_handle *och)
 830 {
 831         struct md_open_data *mod = och->och_mod;
 832         ENTRY;
 833
 834         /**
 835          * It is possible to not have \var mod in a case of eviction between
 836          * lookup and ll_file_open().
 837          **/
 838         if (mod == NULL)
 839                 RETURN(0);
 840
 841         LASSERT(mod != LP_POISON);
 842         LASSERT(mod->mod_open_req != NULL);
 843
 844         spin_lock(&mod->mod_open_req->rq_lock);
 845         if (mod->mod_och)
 846                 mod->mod_och->och_open_handle.cookie = 0;
 847         mod->mod_open_req->rq_early_free_repbuf = 0;
 848         spin_unlock(&mod->mod_open_req->rq_lock);
 849         mdc_free_open(mod);
 850
 851         mod->mod_och = NULL;
 852         och->och_mod = NULL;
 853         obd_mod_put(mod);
 854
 855         RETURN(0);
 856 }
 857
 858 static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
 859                      struct md_open_data *mod, struct ptlrpc_request **request)
 860 {
 861         struct obd_device     *obd = class_exp2obd(exp);
 862         struct ptlrpc_request *req;
 863         struct req_format     *req_fmt;
 864         size_t                 u32_count = 0;
 865         int                    rc;
 866         int                    saved_rc = 0;
 867         ENTRY;
 868
 869         CDEBUG(D_INODE, "%s: "DFID" file closed with intent: %x\n",
 870                exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
 871                op_data->op_bias);
 872
 873         if (op_data->op_bias & MDS_CLOSE_INTENT) {
 874                 req_fmt = &RQF_MDS_CLOSE_INTENT;
 875                 if (op_data->op_bias & MDS_HSM_RELEASE) {
 876                         /* allocate a FID for volatile file */
 877                         rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2,
 878                                            op_data);
 879                         if (rc < 0) {
 880                                 CERROR("%s: "DFID" allocating FID: rc = %d\n",
 881                                        obd->obd_name, PFID(&op_data->op_fid1),
 882                                        rc);
 883                                 /* save the errcode and proceed to close */
 884                                 saved_rc = rc;
 885                         }
 886                 }
 887                 if (op_data->op_bias & MDS_CLOSE_RESYNC_DONE) {
 888                         size_t count = op_data->op_data_size / sizeof(__u32);
 889
 890                         if (count > INLINE_RESYNC_ARRAY_SIZE)
 891                                 u32_count = count;
 892                 }
 893         } else {
 894                 req_fmt = &RQF_MDS_CLOSE;
 895         }
 896
 897         *request = NULL;
 898         if (OBD_FAIL_CHECK(OBD_FAIL_MDC_CLOSE))
 899                 req = NULL;
 900         else
 901                 req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);
 902
 903         /* Ensure that this close's handle is fixed up during replay. */
 904         if (likely(mod != NULL)) {
 905                 LASSERTF(mod->mod_open_req != NULL &&
 906                          mod->mod_open_req->rq_type != LI_POISON,
 907                          "POISONED open %p!\n", mod->mod_open_req);
 908
 909                 mod->mod_close_req = req;
 910
 911                 DEBUG_REQ(D_HA, mod->mod_open_req, "matched open");
 912                 /* We no longer want to preserve this open for replay even
 913                  * though the open was committed. b=3632, b=3633 */
 914                 spin_lock(&mod->mod_open_req->rq_lock);
 915                 mod->mod_open_req->rq_replay = 0;
 916                 spin_unlock(&mod->mod_open_req->rq_lock);
 917         } else {
 918                 CDEBUG(D_HA, "couldn't find open req; expecting close error\n");
 919         }
 920         if (req == NULL) {
 921                 /**
 922                  * TODO: repeat close after errors
 923                  */
 924                 CWARN("%s: close of FID "DFID" failed, file reference will be "
 925                       "dropped when this client unmounts or is evicted\n",
 926                       obd->obd_name, PFID(&op_data->op_fid1));
 927                 GOTO(out, rc = -ENOMEM);
 928         }
 929
 930         if (u32_count > 0)
 931                 req_capsule_set_size(&req->rq_pill, &RMF_U32, RCL_CLIENT,
 932                                      u32_count * sizeof(__u32));
 933
 934         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
 935         if (rc) {
 936                 ptlrpc_request_free(req);
 937                 req = NULL;
 938                 GOTO(out, rc);
 939         }
 940
 941         /* To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
 942          * portal whose threads are not taking any DLM locks and are therefore
 943          * always progressing */
 944         req->rq_request_portal = MDS_READPAGE_PORTAL;
 945         ptlrpc_at_set_req_timeout(req);
 946
 947         if (!(exp_connect_flags2(exp) & OBD_CONNECT2_LSOM))
 948                 op_data->op_xvalid &= ~(OP_XVALID_LAZYSIZE |
 949                                         OP_XVALID_LAZYBLOCKS);
 950
 951         mdc_close_pack(req, op_data);
 952
 953         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 954                              obd->u.cli.cl_default_mds_easize);
 955
 956         ptlrpc_request_set_replen(req);
 957
 958         mdc_get_mod_rpc_slot(req, NULL);
 959         rc = ptlrpc_queue_wait(req);
 960         mdc_put_mod_rpc_slot(req, NULL);
 961
 962         if (req->rq_repmsg == NULL) {
 963                 CDEBUG(D_RPCTRACE, "request failed to send: %p, %d\n", req,
 964                        req->rq_status);
 965                 if (rc == 0)
 966                         rc = req->rq_status ?: -EIO;
 967         } else if (rc == 0 || rc == -EAGAIN) {
 968                 struct mdt_body *body;
 969
 970                 rc = lustre_msg_get_status(req->rq_repmsg);
 971                 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
 972                         DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR, err "
 973                                   "= %d", rc);
 974                         if (rc > 0)
 975                                 rc = -rc;
 976                 }
 977                 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 978                 if (body == NULL)
 979                         rc = -EPROTO;
 980         } else if (rc == -ESTALE) {
 981                 /**
 982                  * it can be allowed error after 3633 if open was committed and
 983                  * server failed before close was sent. Let's check if mod
 984                  * exists and return no error in that case
 985                  */
 986                 if (mod) {
 987                         DEBUG_REQ(D_HA, req, "Reset ESTALE = %d", rc);
 988                         LASSERT(mod->mod_open_req != NULL);
 989                         if (mod->mod_open_req->rq_committed)
 990                                 rc = 0;
 991                 }
 992         }
 993
 994 out:
 995         if (mod) {
 996                 if (rc != 0)
 997                         mod->mod_close_req = NULL;
 998                 /* Since now, mod is accessed through open_req only,
 999                  * thus close req does not keep a reference on mod anymore. */
1000                 obd_mod_put(mod);
1001         }
1002         *request = req;
1003
1004         RETURN(rc < 0 ? rc : saved_rc);
1005 }
1006
1007 static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid,
1008                        u64 offset, struct page **pages, int npages,
1009                        struct ptlrpc_request **request)
1010 {
1011         struct ptlrpc_request   *req;
1012         struct ptlrpc_bulk_desc *desc;
1013         int                      i;
1014         wait_queue_head_t        waitq;
1015         int                      resends = 0;
1016         struct l_wait_info       lwi;
1017         int                      rc;
1018         ENTRY;
1019
1020         *request = NULL;
1021         init_waitqueue_head(&waitq);
1022
1023 restart_bulk:
1024         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE);
1025         if (req == NULL)
1026                 RETURN(-ENOMEM);
1027
1028         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE);
1029         if (rc) {
1030                 ptlrpc_request_free(req);
1031                 RETURN(rc);
1032         }
1033
1034         req->rq_request_portal = MDS_READPAGE_PORTAL;
1035         ptlrpc_at_set_req_timeout(req);
1036
1037         desc = ptlrpc_prep_bulk_imp(req, npages, 1,
1038                                     PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
1039                                     MDS_BULK_PORTAL,
1040                                     &ptlrpc_bulk_kiov_pin_ops);
1041         if (desc == NULL) {
1042                 ptlrpc_req_finished(req);
1043                 RETURN(-ENOMEM);
1044         }
1045
1046         /* NB req now owns desc and will free it when it gets freed */
1047         for (i = 0; i < npages; i++)
1048                 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1049                                                  PAGE_SIZE);
1050
1051         mdc_readdir_pack(req, offset, PAGE_SIZE * npages, fid);
1052
1053         ptlrpc_request_set_replen(req);
1054         rc = ptlrpc_queue_wait(req);
1055         if (rc) {
1056                 ptlrpc_req_finished(req);
1057                 if (rc != -ETIMEDOUT)
1058                         RETURN(rc);
1059
1060                 resends++;
1061                 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
1062                         CERROR("%s: too many resend retries: rc = %d\n",
1063                                exp->exp_obd->obd_name, -EIO);
1064                         RETURN(-EIO);
1065                 }
1066                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL,
1067                                        NULL);
1068                 l_wait_event(waitq, 0, &lwi);
1069
1070                 goto restart_bulk;
1071         }
1072
1073         rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1074                                           req->rq_bulk->bd_nob_transferred);
1075         if (rc < 0) {
1076                 ptlrpc_req_finished(req);
1077                 RETURN(rc);
1078         }
1079
1080         if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) {
1081                 CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n",
1082                        exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred,
1083                        PAGE_SIZE * npages);
1084                 ptlrpc_req_finished(req);
1085                 RETURN(-EPROTO);
1086         }
1087
1088         *request = req;
1089         RETURN(0);
1090 }
1091
1092 static void mdc_release_page(struct page *page, int remove)
1093 {
1094         if (remove) {
1095                 lock_page(page);
1096                 if (likely(page->mapping != NULL))
1097                         truncate_complete_page(page->mapping, page);
1098                 unlock_page(page);
1099         }
1100         put_page(page);
1101 }
1102
1103 static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash,
1104                                     __u64 *start, __u64 *end, int hash64)
1105 {
1106         /*
1107          * Complement of hash is used as an index so that
1108          * radix_tree_gang_lookup() can be used to find a page with starting
1109          * hash _smaller_ than one we are looking for.
1110          */
1111         unsigned long offset = hash_x_index(*hash, hash64);
1112         struct page *page;
1113         int found;
1114
1115         spin_lock_irq(&mapping->tree_lock);
1116         found = radix_tree_gang_lookup(&mapping->page_tree,
1117                                        (void **)&page, offset, 1);
1118         if (found > 0 && !radix_tree_exceptional_entry(page)) {
1119                 struct lu_dirpage *dp;
1120
1121                 get_page(page);
1122                 spin_unlock_irq(&mapping->tree_lock);
1123                 /*
1124                  * In contrast to find_lock_page() we are sure that directory
1125                  * page cannot be truncated (while DLM lock is held) and,
1126                  * hence, can avoid restart.
1127                  *
1128                  * In fact, page cannot be locked here at all, because
1129                  * mdc_read_page_remote does synchronous io.
1130                  */
1131                 wait_on_page_locked(page);
1132                 if (PageUptodate(page)) {
1133                         dp = kmap(page);
1134                         if (BITS_PER_LONG == 32 && hash64) {
1135                                 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1136                                 *end   = le64_to_cpu(dp->ldp_hash_end) >> 32;
1137                                 *hash  = *hash >> 32;
1138                         } else {
1139                                 *start = le64_to_cpu(dp->ldp_hash_start);
1140                                 *end   = le64_to_cpu(dp->ldp_hash_end);
1141                         }
1142                         if (unlikely(*start == 1 && *hash == 0))
1143                                 *hash = *start;
1144                         else
1145                                 LASSERTF(*start <= *hash, "start = %#llx"
1146                                          ",end = %#llx,hash = %#llx\n",
1147                                          *start, *end, *hash);
1148                         CDEBUG(D_VFSTRACE, "offset %lx [%#llx %#llx],"
1149                               " hash %#llx\n", offset, *start, *end, *hash);
1150                         if (*hash > *end) {
1151                                 kunmap(page);
1152                                 mdc_release_page(page, 0);
1153                                 page = NULL;
1154                         } else if (*end != *start && *hash == *end) {
1155                                 /*
1156                                  * upon hash collision, remove this page,
1157                                  * otherwise put page reference, and
1158                                  * mdc_read_page_remote() will issue RPC to
1159                                  * fetch the page we want.
1160                                  */
1161                                 kunmap(page);
1162                                 mdc_release_page(page,
1163                                     le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1164                                 page = NULL;
1165                         }
1166                 } else {
1167                         put_page(page);
1168                         page = ERR_PTR(-EIO);
1169                 }
1170         } else {
1171                 spin_unlock_irq(&mapping->tree_lock);
1172                 page = NULL;
1173         }
1174         return page;
1175 }
1176
1177 /*
1178  * Adjust a set of pages, each page containing an array of lu_dirpages,
1179  * so that each page can be used as a single logical lu_dirpage.
1180  *
1181  * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
1182  * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
1183  * struct lu_dirent.  It has size up to LU_PAGE_SIZE. The ldp_hash_end
1184  * value is used as a cookie to request the next lu_dirpage in a
1185  * directory listing that spans multiple pages (two in this example):
1186  *   ________
1187  *  |        |
1188  * .|--------v-------   -----.
1189  * |s|e|f|p|ent|ent| ... |ent|
1190  * '--|--------------   -----'   Each PAGE contains a single
1191  *    '------.                   lu_dirpage.
1192  * .---------v-------   -----.
1193  * |s|e|f|p|ent| 0 | ... | 0 |
1194  * '-----------------   -----'
1195  *
1196  * However, on hosts where the native VM page size (PAGE_SIZE) is
1197  * larger than LU_PAGE_SIZE, a single host page may contain multiple
1198  * lu_dirpages. After reading the lu_dirpages from the MDS, the
1199  * ldp_hash_end of the first lu_dirpage refers to the one immediately
1200  * after it in the same PAGE (arrows simplified for brevity, but
1201  * in general e0==s1, e1==s2, etc.):
1202  *
1203  * .--------------------   -----.
1204  * |s0|e0|f0|p|ent|ent| ... |ent|
1205  * |---v----------------   -----|
1206  * |s1|e1|f1|p|ent|ent| ... |ent|
1207  * |---v----------------   -----|  Here, each PAGE contains
1208  *             ...                 multiple lu_dirpages.
1209  * |---v----------------   -----|
1210  * |s'|e'|f'|p|ent|ent| ... |ent|
1211  * '---|----------------   -----'
1212  *     v
1213  * .----------------------------.
1214  * |        next PAGE           |
1215  *
1216  * This structure is transformed into a single logical lu_dirpage as follows:
1217  *
1218  * - Replace e0 with e' so the request for the next lu_dirpage gets the page
1219  *   labeled 'next PAGE'.
1220  *
1221  * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
1222  *   a hash collision with the next page exists.
1223  *
1224  * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
1225  *   to the first entry of the next lu_dirpage.
1226  */
1227 #if PAGE_SIZE > LU_PAGE_SIZE
1228 static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs)
1229 {
1230         int i;
1231
1232         for (i = 0; i < cfs_pgs; i++) {
1233                 struct lu_dirpage       *dp = kmap(pages[i]);
1234                 struct lu_dirpage       *first = dp;
1235                 struct lu_dirent        *end_dirent = NULL;
1236                 struct lu_dirent        *ent;
1237                 __u64           hash_end = le64_to_cpu(dp->ldp_hash_end);
1238                 __u32           flags = le32_to_cpu(dp->ldp_flags);
1239
1240                 while (--lu_pgs > 0) {
1241                         ent = lu_dirent_start(dp);
1242                         for (end_dirent = ent; ent != NULL;
1243                              end_dirent = ent, ent = lu_dirent_next(ent));
1244
1245                         /* Advance dp to next lu_dirpage. */
1246                         dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
1247
1248                         /* Check if we've reached the end of the PAGE. */
1249                         if (!((unsigned long)dp & ~PAGE_MASK))
1250                                 break;
1251
1252                         /* Save the hash and flags of this lu_dirpage. */
1253                         hash_end = le64_to_cpu(dp->ldp_hash_end);
1254                         flags = le32_to_cpu(dp->ldp_flags);
1255
1256                         /* Check if lu_dirpage contains no entries. */
1257                         if (end_dirent == NULL)
1258                                 break;
1259
1260                         /* Enlarge the end entry lde_reclen from 0 to
1261                          * first entry of next lu_dirpage. */
1262                         LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
1263                         end_dirent->lde_reclen =
1264                                 cpu_to_le16((char *)(dp->ldp_entries) -
1265                                             (char *)end_dirent);
1266                 }
1267
1268                 first->ldp_hash_end = hash_end;
1269                 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
1270                 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
1271
1272                 kunmap(pages[i]);
1273         }
1274         LASSERTF(lu_pgs == 0, "left = %d\n", lu_pgs);
1275 }
1276 #else
1277 #define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0)
1278 #endif  /* PAGE_SIZE > LU_PAGE_SIZE */
1279
1280 /* parameters for readdir page */
1281 struct readpage_param {
1282         struct md_op_data       *rp_mod;
1283         __u64                   rp_off;
1284         int                     rp_hash64;
1285         struct obd_export       *rp_exp;
1286         struct md_callback      *rp_cb;
1287 };
1288
1289 #ifndef HAVE_DELETE_FROM_PAGE_CACHE
1290 static inline void delete_from_page_cache(struct page *page)
1291 {
1292         remove_from_page_cache(page);
1293         put_page(page);
1294 }
1295 #endif
1296
1297 /**
1298  * Read pages from server.
1299  *
1300  * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
1301  * a header lu_dirpage which describes the start/end hash, and whether this
1302  * page is empty (contains no dir entry) or hash collide with next page.
1303  * After client receives reply, several pages will be integrated into dir page
1304  * in PAGE_SIZE (if PAGE_SIZE greater than LU_PAGE_SIZE), and the
1305  * lu_dirpage for this integrated page will be adjusted.
1306  **/
1307 static int mdc_read_page_remote(void *data, struct page *page0)
1308 {
1309         struct readpage_param *rp = data;
1310         struct page **page_pool;
1311         struct page *page;
1312         struct lu_dirpage *dp;
1313         struct md_op_data *op_data = rp->rp_mod;
1314         struct ptlrpc_request *req;
1315         int max_pages;
1316         struct inode *inode;
1317         struct lu_fid *fid;
1318         int rd_pgs = 0; /* number of pages actually read */
1319         int npages;
1320         int i;
1321         int rc;
1322         ENTRY;
1323
1324         max_pages = rp->rp_exp->exp_obd->u.cli.cl_max_pages_per_rpc;
1325         inode = op_data->op_data;
1326         fid = &op_data->op_fid1;
1327         LASSERT(inode != NULL);
1328
1329         OBD_ALLOC(page_pool, sizeof(page_pool[0]) * max_pages);
1330         if (page_pool != NULL) {
1331                 page_pool[0] = page0;
1332         } else {
1333                 page_pool = &page0;
1334                 max_pages = 1;
1335         }
1336
1337         for (npages = 1; npages < max_pages; npages++) {
1338                 page = __page_cache_alloc(mapping_gfp_mask(inode->i_mapping)
1339                                           | __GFP_COLD);
1340                 if (page == NULL)
1341                         break;
1342                 page_pool[npages] = page;
1343         }
1344
1345         rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, page_pool, npages, &req);
1346         if (rc < 0) {
1347                 /* page0 is special, which was added into page cache early */
1348                 delete_from_page_cache(page0);
1349         } else {
1350                 int lu_pgs;
1351
1352                 rd_pgs = (req->rq_bulk->bd_nob_transferred + PAGE_SIZE - 1) >>
1353                         PAGE_SHIFT;
1354                 lu_pgs = req->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
1355                 LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
1356
1357                 CDEBUG(D_INODE, "read %d(%d) pages\n", rd_pgs, lu_pgs);
1358
1359                 mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs);
1360
1361                 SetPageUptodate(page0);
1362         }
1363         unlock_page(page0);
1364
1365         ptlrpc_req_finished(req);
1366         CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages);
1367         for (i = 1; i < npages; i++) {
1368                 unsigned long   offset;
1369                 __u64           hash;
1370                 int ret;
1371
1372                 page = page_pool[i];
1373
1374                 if (rc < 0 || i >= rd_pgs) {
1375                         put_page(page);
1376                         continue;
1377                 }
1378
1379                 SetPageUptodate(page);
1380
1381                 dp = kmap(page);
1382                 hash = le64_to_cpu(dp->ldp_hash_start);
1383                 kunmap(page);
1384
1385                 offset = hash_x_index(hash, rp->rp_hash64);
1386
1387                 prefetchw(&page->flags);
1388                 ret = add_to_page_cache_lru(page, inode->i_mapping, offset,
1389                                             GFP_KERNEL);
1390                 if (ret == 0)
1391                         unlock_page(page);
1392                 else
1393                         CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
1394                                " rc = %d\n", offset, ret);
1395                 put_page(page);
1396         }
1397
1398         if (page_pool != &page0)
1399                 OBD_FREE(page_pool, sizeof(page_pool[0]) * max_pages);
1400
1401         RETURN(rc);
1402 }
1403
1404 /**
1405  * Read dir page from cache first, if it can not find it, read it from
1406  * server and add into the cache.
1407  *
1408  * \param[in] exp       MDC export
1409  * \param[in] op_data   client MD stack parameters, transfering parameters
1410  *                      between different layers on client MD stack.
1411  * \param[in] cb_op     callback required for ldlm lock enqueue during
1412  *                      read page
1413  * \param[in] hash_offset the hash offset of the page to be read
1414  * \param[in] ppage     the page to be read
1415  *
1416  * retval               = 0 get the page successfully
1417  *                      errno(<0) get the page failed
1418  */
1419 static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data,
1420                          struct md_callback *cb_op, __u64 hash_offset,
1421                          struct page **ppage)
1422 {
1423         struct lookup_intent    it = { .it_op = IT_READDIR };
1424         struct page             *page;
1425         struct inode            *dir = op_data->op_data;
1426         struct address_space    *mapping;
1427         struct lu_dirpage       *dp;
1428         __u64                   start = 0;
1429         __u64                   end = 0;
1430         struct lustre_handle    lockh;
1431         struct ptlrpc_request   *enq_req = NULL;
1432         struct readpage_param   rp_param;
1433         int rc;
1434
1435         ENTRY;
1436
1437         *ppage = NULL;
1438
1439         LASSERT(dir != NULL);
1440         mapping = dir->i_mapping;
1441
1442         rc = mdc_intent_lock(exp, op_data, &it, &enq_req,
1443                              cb_op->md_blocking_ast, 0);
1444         if (enq_req != NULL)
1445                 ptlrpc_req_finished(enq_req);
1446
1447         if (rc < 0) {
1448                 CERROR("%s: "DFID" lock enqueue fails: rc = %d\n",
1449                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc);
1450                 RETURN(rc);
1451         }
1452
1453         rc = 0;
1454         lockh.cookie = it.it_lock_handle;
1455         mdc_set_lock_data(exp, &lockh, dir, NULL);
1456
1457         rp_param.rp_off = hash_offset;
1458         rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64;
1459         page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end,
1460                                rp_param.rp_hash64);
1461         if (IS_ERR(page)) {
1462                 CERROR("%s: dir page locate: "DFID" at %llu: rc %ld\n",
1463                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1464                        rp_param.rp_off, PTR_ERR(page));
1465                 GOTO(out_unlock, rc = PTR_ERR(page));
1466         } else if (page != NULL) {
1467                 /*
1468                  * XXX nikita: not entirely correct handling of a corner case:
1469                  * suppose hash chain of entries with hash value HASH crosses
1470                  * border between pages P0 and P1. First both P0 and P1 are
1471                  * cached, seekdir() is called for some entry from the P0 part
1472                  * of the chain. Later P0 goes out of cache. telldir(HASH)
1473                  * happens and finds P1, as it starts with matching hash
1474                  * value. Remaining entries from P0 part of the chain are
1475                  * skipped. (Is that really a bug?)
1476                  *
1477                  * Possible solutions: 0. don't cache P1 is such case, handle
1478                  * it as an "overflow" page. 1. invalidate all pages at
1479                  * once. 2. use HASH|1 as an index for P1.
1480                  */
1481                 GOTO(hash_collision, page);
1482         }
1483
1484         rp_param.rp_exp = exp;
1485         rp_param.rp_mod = op_data;
1486         page = read_cache_page(mapping,
1487                                hash_x_index(rp_param.rp_off,
1488                                             rp_param.rp_hash64),
1489                                mdc_read_page_remote, &rp_param);
1490         if (IS_ERR(page)) {
1491                 CDEBUG(D_INFO, "%s: read cache page: "DFID" at %llu: %ld\n",
1492                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1493                        rp_param.rp_off, PTR_ERR(page));
1494                 GOTO(out_unlock, rc = PTR_ERR(page));
1495         }
1496
1497         wait_on_page_locked(page);
1498         (void)kmap(page);
1499         if (!PageUptodate(page)) {
1500                 CERROR("%s: page not updated: "DFID" at %llu: rc %d\n",
1501                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1502                        rp_param.rp_off, -5);
1503                 goto fail;
1504         }
1505         if (!PageChecked(page))
1506                 SetPageChecked(page);
1507         if (PageError(page)) {
1508                 CERROR("%s: page error: "DFID" at %llu: rc %d\n",
1509                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1510                        rp_param.rp_off, -5);
1511                 goto fail;
1512         }
1513
1514 hash_collision:
1515         dp = page_address(page);
1516         if (BITS_PER_LONG == 32 && rp_param.rp_hash64) {
1517                 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1518                 end   = le64_to_cpu(dp->ldp_hash_end) >> 32;
1519                 rp_param.rp_off = hash_offset >> 32;
1520         } else {
1521                 start = le64_to_cpu(dp->ldp_hash_start);
1522                 end   = le64_to_cpu(dp->ldp_hash_end);
1523                 rp_param.rp_off = hash_offset;
1524         }
1525         if (end == start) {
1526                 LASSERT(start == rp_param.rp_off);
1527                 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
1528 #if BITS_PER_LONG == 32
1529                 CWARN("Real page-wide hash collision at [%llu %llu] with "
1530                       "hash %llu\n", le64_to_cpu(dp->ldp_hash_start),
1531                       le64_to_cpu(dp->ldp_hash_end), hash_offset);
1532 #endif
1533
1534                 /*
1535                  * Fetch whole overflow chain...
1536                  *
1537                  * XXX not yet.
1538                  */
1539                 goto fail;
1540         }
1541         *ppage = page;
1542 out_unlock:
1543         ldlm_lock_decref(&lockh, it.it_lock_mode);
1544         return rc;
1545 fail:
1546         kunmap(page);
1547         mdc_release_page(page, 1);
1548         rc = -EIO;
1549         goto out_unlock;
1550 }
1551
1552 static int mdc_statfs(const struct lu_env *env,
1553                       struct obd_export *exp, struct obd_statfs *osfs,
1554                       time64_t max_age, __u32 flags)
1555 {
1556         struct obd_device *obd = class_exp2obd(exp);
1557         struct req_format *fmt;
1558         struct ptlrpc_request *req;
1559         struct obd_statfs *msfs;
1560         struct obd_import *imp = NULL;
1561         int rc;
1562         ENTRY;
1563
1564         /*
1565          * Since the request might also come from lprocfs, so we need
1566          * sync this with client_disconnect_export Bug15684
1567          */
1568         down_read(&obd->u.cli.cl_sem);
1569         if (obd->u.cli.cl_import)
1570                 imp = class_import_get(obd->u.cli.cl_import);
1571         up_read(&obd->u.cli.cl_sem);
1572         if (!imp)
1573                 RETURN(-ENODEV);
1574
1575         fmt = &RQF_MDS_STATFS;
1576         if ((exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS) &&
1577             (flags & OBD_STATFS_SUM))
1578                 fmt = &RQF_MDS_STATFS_NEW;
1579         req = ptlrpc_request_alloc_pack(imp, fmt, LUSTRE_MDS_VERSION,
1580                                         MDS_STATFS);
1581         if (req == NULL)
1582                 GOTO(output, rc = -ENOMEM);
1583
1584         if ((flags & OBD_STATFS_SUM) &&
1585             (exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS)) {
1586                 /* request aggregated states */
1587                 struct mdt_body *body;
1588
1589                 body = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
1590                 if (body == NULL)
1591                         GOTO(out, rc = -EPROTO);
1592                 body->mbo_valid = OBD_MD_FLAGSTATFS;
1593         }
1594
1595         ptlrpc_request_set_replen(req);
1596
1597         if (flags & OBD_STATFS_NODELAY) {
1598                 /* procfs requests not want stay in wait for avoid deadlock */
1599                 req->rq_no_resend = 1;
1600                 req->rq_no_delay = 1;
1601         }
1602
1603         rc = ptlrpc_queue_wait(req);
1604         if (rc) {
1605                 /* check connection error first */
1606                 if (imp->imp_connect_error)
1607                         rc = imp->imp_connect_error;
1608                 GOTO(out, rc);
1609         }
1610
1611         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1612         if (msfs == NULL)
1613                 GOTO(out, rc = -EPROTO);
1614
1615         *osfs = *msfs;
1616         EXIT;
1617 out:
1618         ptlrpc_req_finished(req);
1619 output:
1620         class_import_put(imp);
1621         return rc;
1622 }
1623
1624 static int mdc_ioc_fid2path(struct obd_export *exp, struct getinfo_fid2path *gf)
1625 {
1626         __u32 keylen, vallen;
1627         void *key;
1628         int rc;
1629
1630         if (gf->gf_pathlen > PATH_MAX)
1631                 RETURN(-ENAMETOOLONG);
1632         if (gf->gf_pathlen < 2)
1633                 RETURN(-EOVERFLOW);
1634
1635         /* Key is KEY_FID2PATH + getinfo_fid2path description */
1636         keylen = cfs_size_round(sizeof(KEY_FID2PATH) + sizeof(*gf) +
1637                                 sizeof(struct lu_fid));
1638         OBD_ALLOC(key, keylen);
1639         if (key == NULL)
1640                 RETURN(-ENOMEM);
1641         memcpy(key, KEY_FID2PATH, sizeof(KEY_FID2PATH));
1642         memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)), gf, sizeof(*gf));
1643         memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*gf),
1644                gf->gf_u.gf_root_fid, sizeof(struct lu_fid));
1645         CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
1646                PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno);
1647
1648         if (!fid_is_sane(&gf->gf_fid))
1649                 GOTO(out, rc = -EINVAL);
1650
1651         /* Val is struct getinfo_fid2path result plus path */
1652         vallen = sizeof(*gf) + gf->gf_pathlen;
1653
1654         rc = obd_get_info(NULL, exp, keylen, key, &vallen, gf);
1655         if (rc != 0 && rc != -EREMOTE)
1656                 GOTO(out, rc);
1657
1658         if (vallen <= sizeof(*gf))
1659                 GOTO(out, rc = -EPROTO);
1660         if (vallen > sizeof(*gf) + gf->gf_pathlen)
1661                 GOTO(out, rc = -EOVERFLOW);
1662
1663         CDEBUG(D_IOCTL, "path got "DFID" from %llu #%d: %s\n",
1664                PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno,
1665                gf->gf_pathlen < 512 ? gf->gf_u.gf_path :
1666                /* only log the last 512 characters of the path */
1667                gf->gf_u.gf_path + gf->gf_pathlen - 512);
1668
1669 out:
1670         OBD_FREE(key, keylen);
1671         return rc;
1672 }
1673
1674 static int mdc_ioc_hsm_progress(struct obd_export *exp,
1675                                 struct hsm_progress_kernel *hpk)
1676 {
1677         struct obd_import               *imp = class_exp2cliimp(exp);
1678         struct hsm_progress_kernel      *req_hpk;
1679         struct ptlrpc_request           *req;
1680         int                              rc;
1681         ENTRY;
1682
1683         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_PROGRESS,
1684                                         LUSTRE_MDS_VERSION, MDS_HSM_PROGRESS);
1685         if (req == NULL)
1686                 GOTO(out, rc = -ENOMEM);
1687
1688         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1689
1690         /* Copy hsm_progress struct */
1691         req_hpk = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_PROGRESS);
1692         if (req_hpk == NULL)
1693                 GOTO(out, rc = -EPROTO);
1694
1695         *req_hpk = *hpk;
1696         req_hpk->hpk_errval = lustre_errno_hton(hpk->hpk_errval);
1697
1698         ptlrpc_request_set_replen(req);
1699
1700         mdc_get_mod_rpc_slot(req, NULL);
1701         rc = ptlrpc_queue_wait(req);
1702         mdc_put_mod_rpc_slot(req, NULL);
1703
1704         GOTO(out, rc);
1705 out:
1706         ptlrpc_req_finished(req);
1707         return rc;
1708 }
1709 /**
1710  * Send hsm_ct_register to MDS
1711  *
1712  * \param[in]   imp             import
1713  * \param[in]   archive_count   if in bitmap format, it is the bitmap,
1714  *                              else it is the count of archive_ids
1715  * \param[in]   archives        if in bitmap format, it is NULL,
1716  *                              else it is archive_id lists
1717  */
1718 static int mdc_ioc_hsm_ct_register(struct obd_import *imp, __u32 archive_count,
1719                                    __u32 *archives)
1720 {
1721         struct ptlrpc_request *req;
1722         __u32 *archive_array;
1723         size_t archives_size;
1724         int rc;
1725         ENTRY;
1726
1727         req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_CT_REGISTER);
1728         if (req == NULL)
1729                 RETURN(-ENOMEM);
1730
1731         if (archives != NULL)
1732                 archives_size = sizeof(*archive_array) * archive_count;
1733         else
1734                 archives_size = sizeof(archive_count);
1735
1736         req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_ARCHIVE,
1737                              RCL_CLIENT, archives_size);
1738
1739         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_CT_REGISTER);
1740         if (rc) {
1741                 ptlrpc_request_free(req);
1742                 RETURN(-ENOMEM);
1743         }
1744
1745         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1746
1747         archive_array = req_capsule_client_get(&req->rq_pill,
1748                                                &RMF_MDS_HSM_ARCHIVE);
1749         if (archive_array == NULL)
1750                 GOTO(out, rc = -EPROTO);
1751
1752         if (archives != NULL)
1753                 memcpy(archive_array, archives, archives_size);
1754         else
1755                 *archive_array = archive_count;
1756
1757         ptlrpc_request_set_replen(req);
1758
1759         rc = mdc_queue_wait(req);
1760         GOTO(out, rc);
1761 out:
1762         ptlrpc_req_finished(req);
1763         return rc;
1764 }
1765
1766 static int mdc_ioc_hsm_current_action(struct obd_export *exp,
1767                                       struct md_op_data *op_data)
1768 {
1769         struct hsm_current_action       *hca = op_data->op_data;
1770         struct hsm_current_action       *req_hca;
1771         struct ptlrpc_request           *req;
1772         int                              rc;
1773         ENTRY;
1774
1775         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1776                                    &RQF_MDS_HSM_ACTION);
1777         if (req == NULL)
1778                 RETURN(-ENOMEM);
1779
1780         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_ACTION);
1781         if (rc) {
1782                 ptlrpc_request_free(req);
1783                 RETURN(rc);
1784         }
1785
1786         mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1787                       op_data->op_suppgids[0], 0);
1788
1789         ptlrpc_request_set_replen(req);
1790
1791         rc = mdc_queue_wait(req);
1792         if (rc)
1793                 GOTO(out, rc);
1794
1795         req_hca = req_capsule_server_get(&req->rq_pill,
1796                                          &RMF_MDS_HSM_CURRENT_ACTION);
1797         if (req_hca == NULL)
1798                 GOTO(out, rc = -EPROTO);
1799
1800         *hca = *req_hca;
1801
1802         EXIT;
1803 out:
1804         ptlrpc_req_finished(req);
1805         return rc;
1806 }
1807
1808 static int mdc_ioc_hsm_ct_unregister(struct obd_import *imp)
1809 {
1810         struct ptlrpc_request   *req;
1811         int                      rc;
1812         ENTRY;
1813
1814         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_UNREGISTER,
1815                                         LUSTRE_MDS_VERSION,
1816                                         MDS_HSM_CT_UNREGISTER);
1817         if (req == NULL)
1818                 GOTO(out, rc = -ENOMEM);
1819
1820         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1821
1822         ptlrpc_request_set_replen(req);
1823
1824         rc = mdc_queue_wait(req);
1825         GOTO(out, rc);
1826 out:
1827         ptlrpc_req_finished(req);
1828         return rc;
1829 }
1830
1831 static int mdc_ioc_hsm_state_get(struct obd_export *exp,
1832                                  struct md_op_data *op_data)
1833 {
1834         struct hsm_user_state   *hus = op_data->op_data;
1835         struct hsm_user_state   *req_hus;
1836         struct ptlrpc_request   *req;
1837         int                      rc;
1838         ENTRY;
1839
1840         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1841                                    &RQF_MDS_HSM_STATE_GET);
1842         if (req == NULL)
1843                 RETURN(-ENOMEM);
1844
1845         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_GET);
1846         if (rc != 0) {
1847                 ptlrpc_request_free(req);
1848                 RETURN(rc);
1849         }
1850
1851         mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1852                       op_data->op_suppgids[0], 0);
1853
1854         ptlrpc_request_set_replen(req);
1855
1856         rc = mdc_queue_wait(req);
1857         if (rc)
1858                 GOTO(out, rc);
1859
1860         req_hus = req_capsule_server_get(&req->rq_pill, &RMF_HSM_USER_STATE);
1861         if (req_hus == NULL)
1862                 GOTO(out, rc = -EPROTO);
1863
1864         *hus = *req_hus;
1865
1866         EXIT;
1867 out:
1868         ptlrpc_req_finished(req);
1869         return rc;
1870 }
1871
1872 static int mdc_ioc_hsm_state_set(struct obd_export *exp,
1873                                  struct md_op_data *op_data)
1874 {
1875         struct hsm_state_set    *hss = op_data->op_data;
1876         struct hsm_state_set    *req_hss;
1877         struct ptlrpc_request   *req;
1878         int                      rc;
1879         ENTRY;
1880
1881         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1882                                    &RQF_MDS_HSM_STATE_SET);
1883         if (req == NULL)
1884                 RETURN(-ENOMEM);
1885
1886         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_SET);
1887         if (rc) {
1888                 ptlrpc_request_free(req);
1889                 RETURN(rc);
1890         }
1891
1892         mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1893                       op_data->op_suppgids[0], 0);
1894
1895         /* Copy states */
1896         req_hss = req_capsule_client_get(&req->rq_pill, &RMF_HSM_STATE_SET);
1897         if (req_hss == NULL)
1898                 GOTO(out, rc = -EPROTO);
1899         *req_hss = *hss;
1900
1901         ptlrpc_request_set_replen(req);
1902
1903         mdc_get_mod_rpc_slot(req, NULL);
1904         rc = ptlrpc_queue_wait(req);
1905         mdc_put_mod_rpc_slot(req, NULL);
1906
1907         GOTO(out, rc);
1908 out:
1909         ptlrpc_req_finished(req);
1910         return rc;
1911 }
1912
1913 static int mdc_ioc_hsm_request(struct obd_export *exp,
1914                                struct hsm_user_request *hur)
1915 {
1916         struct obd_import       *imp = class_exp2cliimp(exp);
1917         struct ptlrpc_request   *req;
1918         struct hsm_request      *req_hr;
1919         struct hsm_user_item    *req_hui;
1920         char                    *req_opaque;
1921         int                      rc;
1922         ENTRY;
1923
1924         req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_REQUEST);
1925         if (req == NULL)
1926                 GOTO(out, rc = -ENOMEM);
1927
1928         req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM, RCL_CLIENT,
1929                              hur->hur_request.hr_itemcount
1930                              * sizeof(struct hsm_user_item));
1931         req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA, RCL_CLIENT,
1932                              hur->hur_request.hr_data_len);
1933
1934         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_REQUEST);
1935         if (rc) {
1936                 ptlrpc_request_free(req);
1937                 RETURN(rc);
1938         }
1939
1940         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1941
1942         /* Copy hsm_request struct */
1943         req_hr = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_REQUEST);
1944         if (req_hr == NULL)
1945                 GOTO(out, rc = -EPROTO);
1946         *req_hr = hur->hur_request;
1947
1948         /* Copy hsm_user_item structs */
1949         req_hui = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM);
1950         if (req_hui == NULL)
1951                 GOTO(out, rc = -EPROTO);
1952         memcpy(req_hui, hur->hur_user_item,
1953                hur->hur_request.hr_itemcount * sizeof(struct hsm_user_item));
1954
1955         /* Copy opaque field */
1956         req_opaque = req_capsule_client_get(&req->rq_pill, &RMF_GENERIC_DATA);
1957         if (req_opaque == NULL)
1958                 GOTO(out, rc = -EPROTO);
1959         memcpy(req_opaque, hur_data(hur), hur->hur_request.hr_data_len);
1960
1961         ptlrpc_request_set_replen(req);
1962
1963         mdc_get_mod_rpc_slot(req, NULL);
1964         rc = ptlrpc_queue_wait(req);
1965         mdc_put_mod_rpc_slot(req, NULL);
1966
1967         GOTO(out, rc);
1968
1969 out:
1970         ptlrpc_req_finished(req);
1971         return rc;
1972 }
1973
1974 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
1975                                 struct lustre_kernelcomm *lk);
1976
1977 static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp,
1978                         struct obd_quotactl *oqctl)
1979 {
1980         struct ptlrpc_request   *req;
1981         struct obd_quotactl     *oqc;
1982         int                      rc;
1983         ENTRY;
1984
1985         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1986                                         &RQF_MDS_QUOTACTL, LUSTRE_MDS_VERSION,
1987                                         MDS_QUOTACTL);
1988         if (req == NULL)
1989                 RETURN(-ENOMEM);
1990
1991         oqc = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1992         *oqc = *oqctl;
1993
1994         ptlrpc_request_set_replen(req);
1995         ptlrpc_at_set_req_timeout(req);
1996
1997         rc = ptlrpc_queue_wait(req);
1998         if (rc)
1999                 CERROR("ptlrpc_queue_wait failed, rc: %d\n", rc);
2000
2001         if (req->rq_repmsg &&
2002             (oqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL))) {
2003                 *oqctl = *oqc;
2004         } else if (!rc) {
2005                 CERROR ("Can't unpack obd_quotactl\n");
2006                 rc = -EPROTO;
2007         }
2008         ptlrpc_req_finished(req);
2009
2010         RETURN(rc);
2011 }
2012
2013 static int mdc_ioc_swap_layouts(struct obd_export *exp,
2014                                 struct md_op_data *op_data)
2015 {
2016         struct list_head cancels = LIST_HEAD_INIT(cancels);
2017         struct ptlrpc_request   *req;
2018         int                      rc, count;
2019         struct mdc_swap_layouts *msl, *payload;
2020         ENTRY;
2021
2022         msl = op_data->op_data;
2023
2024         /* When the MDT will get the MDS_SWAP_LAYOUTS RPC the
2025          * first thing it will do is to cancel the 2 layout
2026          * locks held by this client.
2027          * So the client must cancel its layout locks on the 2 fids
2028          * with the request RPC to avoid extra RPC round trips.
2029          */
2030         count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels,
2031                                         LCK_EX, MDS_INODELOCK_LAYOUT |
2032                                         MDS_INODELOCK_XATTR);
2033         count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels,
2034                                          LCK_EX, MDS_INODELOCK_LAYOUT |
2035                                          MDS_INODELOCK_XATTR);
2036
2037         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2038                                    &RQF_MDS_SWAP_LAYOUTS);
2039         if (req == NULL) {
2040                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
2041                 RETURN(-ENOMEM);
2042         }
2043
2044         rc = mdc_prep_elc_req(exp, req, MDS_SWAP_LAYOUTS, &cancels, count);
2045         if (rc) {
2046                 ptlrpc_request_free(req);
2047                 RETURN(rc);
2048         }
2049
2050         mdc_swap_layouts_pack(req, op_data);
2051
2052         payload = req_capsule_client_get(&req->rq_pill, &RMF_SWAP_LAYOUTS);
2053         LASSERT(payload);
2054
2055         *payload = *msl;
2056
2057         ptlrpc_request_set_replen(req);
2058
2059         rc = ptlrpc_queue_wait(req);
2060         if (rc)
2061                 GOTO(out, rc);
2062         EXIT;
2063
2064 out:
2065         ptlrpc_req_finished(req);
2066         return rc;
2067 }
2068
2069 static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2070                          void *karg, void __user *uarg)
2071 {
2072         struct obd_device *obd = exp->exp_obd;
2073         struct obd_ioctl_data *data = karg;
2074         struct obd_import *imp = obd->u.cli.cl_import;
2075         int rc;
2076         ENTRY;
2077
2078         if (!try_module_get(THIS_MODULE)) {
2079                 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2080                        module_name(THIS_MODULE));
2081                 return -EINVAL;
2082         }
2083         switch (cmd) {
2084         case OBD_IOC_FID2PATH:
2085                 rc = mdc_ioc_fid2path(exp, karg);
2086                 GOTO(out, rc);
2087         case LL_IOC_HSM_CT_START:
2088                 rc = mdc_ioc_hsm_ct_start(exp, karg);
2089                 /* ignore if it was already registered on this MDS. */
2090                 if (rc == -EEXIST)
2091                         rc = 0;
2092                 GOTO(out, rc);
2093         case LL_IOC_HSM_PROGRESS:
2094                 rc = mdc_ioc_hsm_progress(exp, karg);
2095                 GOTO(out, rc);
2096         case LL_IOC_HSM_STATE_GET:
2097                 rc = mdc_ioc_hsm_state_get(exp, karg);
2098                 GOTO(out, rc);
2099         case LL_IOC_HSM_STATE_SET:
2100                 rc = mdc_ioc_hsm_state_set(exp, karg);
2101                 GOTO(out, rc);
2102         case LL_IOC_HSM_ACTION:
2103                 rc = mdc_ioc_hsm_current_action(exp, karg);
2104                 GOTO(out, rc);
2105         case LL_IOC_HSM_REQUEST:
2106                 rc = mdc_ioc_hsm_request(exp, karg);
2107                 GOTO(out, rc);
2108         case OBD_IOC_CLIENT_RECOVER:
2109                 rc = ptlrpc_recover_import(imp, data->ioc_inlbuf1, 0);
2110                 if (rc < 0)
2111                         GOTO(out, rc);
2112                 GOTO(out, rc = 0);
2113         case IOC_OSC_SET_ACTIVE:
2114                 rc = ptlrpc_set_import_active(imp, data->ioc_offset);
2115                 GOTO(out, rc);
2116         /*
2117          * Normally IOC_OBD_STATFS, OBD_IOC_QUOTACTL iocontrol are handled by
2118          * LMV instead of MDC. But when the cluster is upgraded from 1.8,
2119          * there'd be no LMV layer thus we might be called here. Eventually
2120          * this code should be removed.
2121          * bz20731, LU-592.
2122          */
2123         case IOC_OBD_STATFS: {
2124                 struct obd_statfs stat_buf = {0};
2125
2126                 if (*((__u32 *) data->ioc_inlbuf2) != 0)
2127                         GOTO(out, rc = -ENODEV);
2128
2129                 /* copy UUID */
2130                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(obd),
2131                                  min((int)data->ioc_plen2,
2132                                      (int)sizeof(struct obd_uuid))))
2133                         GOTO(out, rc = -EFAULT);
2134
2135                 rc = mdc_statfs(NULL, obd->obd_self_export, &stat_buf,
2136                                 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
2137                                 0);
2138                 if (rc != 0)
2139                         GOTO(out, rc);
2140
2141                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
2142                                      min((int) data->ioc_plen1,
2143                                          (int) sizeof(stat_buf))))
2144                         GOTO(out, rc = -EFAULT);
2145
2146                 GOTO(out, rc = 0);
2147         }
2148         case OBD_IOC_QUOTACTL: {
2149                 struct if_quotactl *qctl = karg;
2150                 struct obd_quotactl *oqctl;
2151
2152                 OBD_ALLOC_PTR(oqctl);
2153                 if (oqctl == NULL)
2154                         GOTO(out, rc = -ENOMEM);
2155
2156                 QCTL_COPY(oqctl, qctl);
2157                 rc = obd_quotactl(exp, oqctl);
2158                 if (rc == 0) {
2159                         QCTL_COPY(qctl, oqctl);
2160                         qctl->qc_valid = QC_MDTIDX;
2161                         qctl->obd_uuid = obd->u.cli.cl_target_uuid;
2162                 }
2163
2164                 OBD_FREE_PTR(oqctl);
2165                 GOTO(out, rc);
2166         }
2167         case LL_IOC_GET_CONNECT_FLAGS:
2168                 if (copy_to_user(uarg, exp_connect_flags_ptr(exp),
2169                                  sizeof(*exp_connect_flags_ptr(exp))))
2170                         GOTO(out, rc = -EFAULT);
2171
2172                 GOTO(out, rc = 0);
2173         case LL_IOC_LOV_SWAP_LAYOUTS:
2174                 rc = mdc_ioc_swap_layouts(exp, karg);
2175                 GOTO(out, rc);
2176         default:
2177                 CERROR("unrecognised ioctl: cmd = %#x\n", cmd);
2178                 GOTO(out, rc = -ENOTTY);
2179         }
2180 out:
2181         module_put(THIS_MODULE);
2182
2183         return rc;
2184 }
2185
2186 static int mdc_get_info_rpc(struct obd_export *exp,
2187                             u32 keylen, void *key,
2188                             u32 vallen, void *val)
2189 {
2190         struct obd_import      *imp = class_exp2cliimp(exp);
2191         struct ptlrpc_request  *req;
2192         char                   *tmp;
2193         int                     rc = -EINVAL;
2194         ENTRY;
2195
2196         req = ptlrpc_request_alloc(imp, &RQF_MDS_GET_INFO);
2197         if (req == NULL)
2198                 RETURN(-ENOMEM);
2199
2200         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY,
2201                              RCL_CLIENT, keylen);
2202         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VALLEN,
2203                              RCL_CLIENT, sizeof(vallen));
2204
2205         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_INFO);
2206         if (rc) {
2207                 ptlrpc_request_free(req);
2208                 RETURN(rc);
2209         }
2210
2211         tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
2212         memcpy(tmp, key, keylen);
2213         tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_VALLEN);
2214         memcpy(tmp, &vallen, sizeof(vallen));
2215
2216         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VAL,
2217                              RCL_SERVER, vallen);
2218         ptlrpc_request_set_replen(req);
2219
2220         rc = ptlrpc_queue_wait(req);
2221         /* -EREMOTE means the get_info result is partial, and it needs to
2222          * continue on another MDT, see fid2path part in lmv_iocontrol */
2223         if (rc == 0 || rc == -EREMOTE) {
2224                 tmp = req_capsule_server_get(&req->rq_pill, &RMF_GETINFO_VAL);
2225                 memcpy(val, tmp, vallen);
2226                 if (ptlrpc_rep_need_swab(req)) {
2227                         if (KEY_IS(KEY_FID2PATH))
2228                                 lustre_swab_fid2path(val);
2229                 }
2230         }
2231         ptlrpc_req_finished(req);
2232
2233         RETURN(rc);
2234 }
2235
2236 static void lustre_swab_hai(struct hsm_action_item *h)
2237 {
2238         __swab32s(&h->hai_len);
2239         __swab32s(&h->hai_action);
2240         lustre_swab_lu_fid(&h->hai_fid);
2241         lustre_swab_lu_fid(&h->hai_dfid);
2242         __swab64s(&h->hai_cookie);
2243         __swab64s(&h->hai_extent.offset);
2244         __swab64s(&h->hai_extent.length);
2245         __swab64s(&h->hai_gid);
2246 }
2247
2248 static void lustre_swab_hal(struct hsm_action_list *h)
2249 {
2250         struct hsm_action_item  *hai;
2251         __u32                    i;
2252
2253         __swab32s(&h->hal_version);
2254         __swab32s(&h->hal_count);
2255         __swab32s(&h->hal_archive_id);
2256         __swab64s(&h->hal_flags);
2257         hai = hai_first(h);
2258         for (i = 0; i < h->hal_count; i++, hai = hai_next(hai))
2259                 lustre_swab_hai(hai);
2260 }
2261
2262 static void lustre_swab_kuch(struct kuc_hdr *l)
2263 {
2264         __swab16s(&l->kuc_magic);
2265         /* __u8 l->kuc_transport */
2266         __swab16s(&l->kuc_msgtype);
2267         __swab16s(&l->kuc_msglen);
2268 }
2269
2270 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2271                                 struct lustre_kernelcomm *lk)
2272 {
2273         struct obd_import *imp = class_exp2cliimp(exp);
2274         int rc = 0;
2275
2276         if (lk->lk_group != KUC_GRP_HSM) {
2277                 CERROR("Bad copytool group %d\n", lk->lk_group);
2278                 return -EINVAL;
2279         }
2280
2281         CDEBUG(D_HSM, "CT start r%d w%d u%d g%d f%#x\n", lk->lk_rfd, lk->lk_wfd,
2282                lk->lk_uid, lk->lk_group, lk->lk_flags);
2283
2284         if (lk->lk_flags & LK_FLG_STOP) {
2285                 /* Unregister with the coordinator */
2286                 rc = mdc_ioc_hsm_ct_unregister(imp);
2287         } else {
2288                 __u32 *archives = NULL;
2289
2290                 if ((lk->lk_flags & LK_FLG_DATANR) && lk->lk_data_count > 0)
2291                         archives = lk->lk_data;
2292
2293                 rc = mdc_ioc_hsm_ct_register(imp, lk->lk_data_count, archives);
2294         }
2295
2296         return rc;
2297 }
2298
2299 /**
2300  * Send a message to any listening copytools
2301  * @param val KUC message (kuc_hdr + hsm_action_list)
2302  * @param len total length of message
2303  */
2304 static int mdc_hsm_copytool_send(const struct obd_uuid *uuid,
2305                                  size_t len, void *val)
2306 {
2307         struct kuc_hdr          *lh = (struct kuc_hdr *)val;
2308         struct hsm_action_list  *hal = (struct hsm_action_list *)(lh + 1);
2309         int                      rc;
2310         ENTRY;
2311
2312         if (len < sizeof(*lh) + sizeof(*hal)) {
2313                 CERROR("Short HSM message %zu < %zu\n", len,
2314                        sizeof(*lh) + sizeof(*hal));
2315                 RETURN(-EPROTO);
2316         }
2317         if (lh->kuc_magic == __swab16(KUC_MAGIC)) {
2318                 lustre_swab_kuch(lh);
2319                 lustre_swab_hal(hal);
2320         } else if (lh->kuc_magic != KUC_MAGIC) {
2321                 CERROR("Bad magic %x!=%x\n", lh->kuc_magic, KUC_MAGIC);
2322                 RETURN(-EPROTO);
2323         }
2324
2325         CDEBUG(D_HSM, " Received message mg=%x t=%d m=%d l=%d actions=%d "
2326                "on %s\n",
2327                lh->kuc_magic, lh->kuc_transport, lh->kuc_msgtype,
2328                lh->kuc_msglen, hal->hal_count, hal->hal_fsname);
2329
2330         /* Broadcast to HSM listeners */
2331         rc = libcfs_kkuc_group_put(uuid, KUC_GRP_HSM, lh);
2332
2333         RETURN(rc);
2334 }
2335
2336 /**
2337  * callback function passed to kuc for re-registering each HSM copytool
2338  * running on MDC, after MDT shutdown/recovery.
2339  * @param data copytool registration data
2340  * @param cb_arg callback argument (obd_import)
2341  */
2342 static int mdc_hsm_ct_reregister(void *data, void *cb_arg)
2343 {
2344         struct obd_import *imp = (struct obd_import *)cb_arg;
2345         struct kkuc_ct_data *kcd = data;
2346         __u32 *archives = NULL;
2347         int rc;
2348
2349         if (kcd == NULL ||
2350             (kcd->kcd_magic != KKUC_CT_DATA_ARRAY_MAGIC &&
2351              kcd->kcd_magic != KKUC_CT_DATA_BITMAP_MAGIC))
2352                 return -EPROTO;
2353
2354         if (kcd->kcd_magic == KKUC_CT_DATA_BITMAP_MAGIC) {
2355                 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2356                        "(archive=%#x)\n", imp->imp_obd->obd_name,
2357                        kcd->kcd_nr_archives);
2358         } else {
2359                 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2360                        "(archive nr = %u)\n",
2361                        imp->imp_obd->obd_name, kcd->kcd_nr_archives);
2362                 if (kcd->kcd_nr_archives != 0)
2363                         archives = kcd->kcd_archives;
2364         }
2365
2366         rc = mdc_ioc_hsm_ct_register(imp, kcd->kcd_nr_archives, archives);
2367         /* ignore error if the copytool is already registered */
2368         return (rc == -EEXIST) ? 0 : rc;
2369 }
2370
2371 /**
2372  * Re-establish all kuc contexts with MDT
2373  * after MDT shutdown/recovery.
2374  */
2375 static int mdc_kuc_reregister(struct obd_import *imp)
2376 {
2377         /* re-register HSM agents */
2378         return libcfs_kkuc_group_foreach(&imp->imp_obd->obd_uuid, KUC_GRP_HSM,
2379                                          mdc_hsm_ct_reregister, imp);
2380 }
2381
2382 static int mdc_set_info_async(const struct lu_env *env,
2383                               struct obd_export *exp,
2384                               u32 keylen, void *key,
2385                               u32 vallen, void *val,
2386                               struct ptlrpc_request_set *set)
2387 {
2388         struct obd_import       *imp = class_exp2cliimp(exp);
2389         int                      rc;
2390         ENTRY;
2391
2392         if (KEY_IS(KEY_READ_ONLY)) {
2393                 if (vallen != sizeof(int))
2394                         RETURN(-EINVAL);
2395
2396                 spin_lock(&imp->imp_lock);
2397                 if (*((int *)val)) {
2398                         imp->imp_connect_flags_orig |= OBD_CONNECT_RDONLY;
2399                         imp->imp_connect_data.ocd_connect_flags |=
2400                                                         OBD_CONNECT_RDONLY;
2401                 } else {
2402                         imp->imp_connect_flags_orig &= ~OBD_CONNECT_RDONLY;
2403                         imp->imp_connect_data.ocd_connect_flags &=
2404                                                         ~OBD_CONNECT_RDONLY;
2405                 }
2406                 spin_unlock(&imp->imp_lock);
2407
2408                 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2409                                        keylen, key, vallen, val, set);
2410                 RETURN(rc);
2411         }
2412         if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2413                 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2414                                        keylen, key, vallen, val, set);
2415                 RETURN(rc);
2416         }
2417         if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) {
2418                 rc = mdc_hsm_copytool_send(&imp->imp_obd->obd_uuid, vallen,
2419                                            val);
2420                 RETURN(rc);
2421         }
2422
2423         if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2424                 __u32 *default_easize = val;
2425
2426                 exp->exp_obd->u.cli.cl_default_mds_easize = *default_easize;
2427                 RETURN(0);
2428         }
2429
2430         rc = osc_set_info_async(env, exp, keylen, key, vallen, val, set);
2431         RETURN(rc);
2432 }
2433
2434 static int mdc_get_info(const struct lu_env *env, struct obd_export *exp,
2435                         __u32 keylen, void *key, __u32 *vallen, void *val)
2436 {
2437         int rc = -EINVAL;
2438
2439         if (KEY_IS(KEY_MAX_EASIZE)) {
2440                 __u32 mdsize, *max_easize;
2441
2442                 if (*vallen != sizeof(int))
2443                         RETURN(-EINVAL);
2444                 mdsize = *(__u32 *)val;
2445                 if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
2446                         exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
2447                 max_easize = val;
2448                 *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
2449                 RETURN(0);
2450         } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2451                 __u32 *default_easize;
2452
2453                 if (*vallen != sizeof(int))
2454                         RETURN(-EINVAL);
2455                 default_easize = val;
2456                 *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
2457                 RETURN(0);
2458         } else if (KEY_IS(KEY_CONN_DATA)) {
2459                 struct obd_import *imp = class_exp2cliimp(exp);
2460                 struct obd_connect_data *data = val;
2461
2462                 if (*vallen != sizeof(*data))
2463                         RETURN(-EINVAL);
2464
2465                 *data = imp->imp_connect_data;
2466                 RETURN(0);
2467         } else if (KEY_IS(KEY_TGT_COUNT)) {
2468                 *((__u32 *)val) = 1;
2469                 RETURN(0);
2470         }
2471
2472         rc = mdc_get_info_rpc(exp, keylen, key, *vallen, val);
2473
2474         RETURN(rc);
2475 }
2476
2477 static int mdc_fsync(struct obd_export *exp, const struct lu_fid *fid,
2478                      struct ptlrpc_request **request)
2479 {
2480         struct ptlrpc_request *req;
2481         int                    rc;
2482         ENTRY;
2483
2484         *request = NULL;
2485         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SYNC);
2486         if (req == NULL)
2487                 RETURN(-ENOMEM);
2488
2489         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SYNC);
2490         if (rc) {
2491                 ptlrpc_request_free(req);
2492                 RETURN(rc);
2493         }
2494
2495         mdc_pack_body(req, fid, 0, 0, -1, 0);
2496
2497         ptlrpc_request_set_replen(req);
2498
2499         rc = ptlrpc_queue_wait(req);
2500         if (rc)
2501                 ptlrpc_req_finished(req);
2502         else
2503                 *request = req;
2504         RETURN(rc);
2505 }
2506
2507 static int mdc_import_event(struct obd_device *obd, struct obd_import *imp,
2508                             enum obd_import_event event)
2509 {
2510         struct client_obd *cli = &obd->u.cli;
2511         int rc = 0;
2512
2513         LASSERT(imp->imp_obd == obd);
2514
2515         switch (event) {
2516         case IMP_EVENT_DISCON:
2517                 spin_lock(&cli->cl_loi_list_lock);
2518                 cli->cl_avail_grant = 0;
2519                 cli->cl_lost_grant = 0;
2520                 spin_unlock(&cli->cl_loi_list_lock);
2521                 break;
2522         case IMP_EVENT_INACTIVE:
2523                 /*
2524                  * Flush current sequence to make client obtain new one
2525                  * from server in case of disconnect/reconnect.
2526                  */
2527                 down_read(&cli->cl_seq_rwsem);
2528                 if (cli->cl_seq)
2529                         seq_client_flush(cli->cl_seq);
2530                 up_read(&cli->cl_seq_rwsem);
2531
2532                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
2533                 break;
2534         case IMP_EVENT_INVALIDATE: {
2535                 struct ldlm_namespace *ns = obd->obd_namespace;
2536                 struct lu_env *env;
2537                 __u16 refcheck;
2538
2539                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2540
2541                 env = cl_env_get(&refcheck);
2542                 if (!IS_ERR(env)) {
2543                         /* Reset grants. All pages go to failing rpcs due to
2544                          * the invalid import.
2545                          */
2546                         osc_io_unplug(env, cli, NULL);
2547
2548                         cfs_hash_for_each_nolock(ns->ns_rs_hash,
2549                                                  osc_ldlm_resource_invalidate,
2550                                                  env, 0);
2551                         cl_env_put(env, &refcheck);
2552                         ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2553                 } else {
2554                         rc = PTR_ERR(env);
2555                 }
2556                 break;
2557         }
2558         case IMP_EVENT_ACTIVE:
2559                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
2560                 /* redo the kuc registration after reconnecting */
2561                 if (rc == 0)
2562                         rc = mdc_kuc_reregister(imp);
2563                 break;
2564         case IMP_EVENT_OCD: {
2565                 struct obd_connect_data *ocd = &imp->imp_connect_data;
2566
2567                 if (OCD_HAS_FLAG(ocd, GRANT))
2568                         osc_init_grant(cli, ocd);
2569
2570                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
2571                 break;
2572         }
2573         case IMP_EVENT_DEACTIVATE:
2574         case IMP_EVENT_ACTIVATE:
2575                 break;
2576         default:
2577                 CERROR("Unknown import event %x\n", event);
2578                 LBUG();
2579         }
2580         RETURN(rc);
2581 }
2582
2583 int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp,
2584                   struct lu_fid *fid, struct md_op_data *op_data)
2585 {
2586         struct client_obd *cli = &exp->exp_obd->u.cli;
2587         int rc = -EIO;
2588
2589         ENTRY;
2590
2591         down_read(&cli->cl_seq_rwsem);
2592         if (cli->cl_seq)
2593                 rc = seq_client_alloc_fid(env, cli->cl_seq, fid);
2594         up_read(&cli->cl_seq_rwsem);
2595
2596         RETURN(rc);
2597 }
2598
2599 static struct obd_uuid *mdc_get_uuid(struct obd_export *exp)
2600 {
2601         struct client_obd *cli = &exp->exp_obd->u.cli;
2602         return &cli->cl_target_uuid;
2603 }
2604
2605 /**
2606  * Determine whether the lock can be canceled before replaying it during
2607  * recovery, non zero value will be return if the lock can be canceled,
2608  * or zero returned for not
2609  */
2610 static int mdc_cancel_weight(struct ldlm_lock *lock)
2611 {
2612         if (lock->l_resource->lr_type != LDLM_IBITS)
2613                 RETURN(0);
2614
2615         /* FIXME: if we ever get into a situation where there are too many
2616          * opened files with open locks on a single node, then we really
2617          * should replay these open locks to reget it */
2618         if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN)
2619                 RETURN(0);
2620
2621         /* Special case for DoM locks, cancel only unused and granted locks */
2622         if (ldlm_has_dom(lock) &&
2623             (lock->l_granted_mode != lock->l_req_mode ||
2624              osc_ldlm_weigh_ast(lock) != 0))
2625                 RETURN(0);
2626
2627         RETURN(1);
2628 }
2629
2630 static int mdc_resource_inode_free(struct ldlm_resource *res)
2631 {
2632         if (res->lr_lvb_inode)
2633                 res->lr_lvb_inode = NULL;
2634
2635         return 0;
2636 }
2637
2638 static struct ldlm_valblock_ops inode_lvbo = {
2639         .lvbo_free = mdc_resource_inode_free
2640 };
2641
2642 static int mdc_llog_init(struct obd_device *obd)
2643 {
2644         struct obd_llog_group   *olg = &obd->obd_olg;
2645         struct llog_ctxt        *ctxt;
2646         int                      rc;
2647
2648         ENTRY;
2649
2650         rc = llog_setup(NULL, obd, olg, LLOG_CHANGELOG_REPL_CTXT, obd,
2651                         &llog_client_ops);
2652         if (rc < 0)
2653                 RETURN(rc);
2654
2655         ctxt = llog_group_get_ctxt(olg, LLOG_CHANGELOG_REPL_CTXT);
2656         llog_initiator_connect(ctxt);
2657         llog_ctxt_put(ctxt);
2658
2659         RETURN(0);
2660 }
2661
2662 static void mdc_llog_finish(struct obd_device *obd)
2663 {
2664         struct llog_ctxt *ctxt;
2665
2666         ENTRY;
2667
2668         ctxt = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
2669         if (ctxt != NULL)
2670                 llog_cleanup(NULL, ctxt);
2671
2672         EXIT;
2673 }
2674
2675 int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg)
2676 {
2677         int rc;
2678
2679         ENTRY;
2680
2681         rc = osc_setup_common(obd, cfg);
2682         if (rc < 0)
2683                 RETURN(rc);
2684
2685         rc = mdc_tunables_init(obd);
2686         if (rc)
2687                 GOTO(err_osc_cleanup, rc);
2688
2689         obd->u.cli.cl_dom_min_inline_repsize = MDC_DOM_DEF_INLINE_REPSIZE;
2690
2691         ns_register_cancel(obd->obd_namespace, mdc_cancel_weight);
2692
2693         obd->obd_namespace->ns_lvbo = &inode_lvbo;
2694
2695         rc = mdc_llog_init(obd);
2696         if (rc) {
2697                 CERROR("%s: failed to setup llogging subsystems: rc = %d\n",
2698                        obd->obd_name, rc);
2699                 GOTO(err_llog_cleanup, rc);
2700         }
2701
2702         rc = mdc_changelog_cdev_init(obd);
2703         if (rc) {
2704                 CERROR("%s: failed to setup changelog char device: rc = %d\n",
2705                        obd->obd_name, rc);
2706                 GOTO(err_changelog_cleanup, rc);
2707         }
2708
2709         RETURN(rc);
2710
2711 err_changelog_cleanup:
2712         mdc_llog_finish(obd);
2713 err_llog_cleanup:
2714         lprocfs_free_md_stats(obd);
2715         ptlrpc_lprocfs_unregister_obd(obd);
2716 err_osc_cleanup:
2717         osc_cleanup_common(obd);
2718         return rc;
2719 }
2720
2721 /* Initialize the default and maximum LOV EA sizes.  This allows
2722  * us to make MDS RPCs with large enough reply buffers to hold a default
2723  * sized EA without having to calculate this (via a call into the
2724  * LOV + OSCs) each time we make an RPC.  The maximum size is also tracked
2725  * but not used to avoid wastefully vmalloc()'ing large reply buffers when
2726  * a large number of stripes is possible.  If a larger reply buffer is
2727  * required it will be reallocated in the ptlrpc layer due to overflow.
2728  */
2729 static int mdc_init_ea_size(struct obd_export *exp, __u32 easize,
2730                             __u32 def_easize)
2731 {
2732         struct obd_device *obd = exp->exp_obd;
2733         struct client_obd *cli = &obd->u.cli;
2734         ENTRY;
2735
2736         if (cli->cl_max_mds_easize < easize)
2737                 cli->cl_max_mds_easize = easize;
2738
2739         if (cli->cl_default_mds_easize < def_easize)
2740                 cli->cl_default_mds_easize = def_easize;
2741
2742         RETURN(0);
2743 }
2744
2745 static int mdc_precleanup(struct obd_device *obd)
2746 {
2747         ENTRY;
2748
2749         osc_precleanup_common(obd);
2750         mdc_changelog_cdev_finish(obd);
2751
2752         obd_cleanup_client_import(obd);
2753         ptlrpc_lprocfs_unregister_obd(obd);
2754         lprocfs_free_md_stats(obd);
2755         mdc_llog_finish(obd);
2756         RETURN(0);
2757 }
2758
2759 static int mdc_cleanup(struct obd_device *obd)
2760 {
2761         return osc_cleanup_common(obd);
2762 }
2763
2764 static struct obd_ops mdc_obd_ops = {
2765         .o_owner            = THIS_MODULE,
2766         .o_setup            = mdc_setup,
2767         .o_precleanup       = mdc_precleanup,
2768         .o_cleanup          = mdc_cleanup,
2769         .o_add_conn         = client_import_add_conn,
2770         .o_del_conn         = client_import_del_conn,
2771         .o_connect          = client_connect_import,
2772         .o_reconnect        = osc_reconnect,
2773         .o_disconnect       = osc_disconnect,
2774         .o_iocontrol        = mdc_iocontrol,
2775         .o_set_info_async   = mdc_set_info_async,
2776         .o_statfs           = mdc_statfs,
2777         .o_fid_init         = client_fid_init,
2778         .o_fid_fini         = client_fid_fini,
2779         .o_fid_alloc        = mdc_fid_alloc,
2780         .o_import_event     = mdc_import_event,
2781         .o_get_info         = mdc_get_info,
2782         .o_get_uuid         = mdc_get_uuid,
2783         .o_quotactl         = mdc_quotactl,
2784 };
2785
2786 static struct md_ops mdc_md_ops = {
2787         .m_get_root         = mdc_get_root,
2788         .m_null_inode       = mdc_null_inode,
2789         .m_close            = mdc_close,
2790         .m_create           = mdc_create,
2791         .m_enqueue          = mdc_enqueue,
2792         .m_getattr          = mdc_getattr,
2793         .m_getattr_name     = mdc_getattr_name,
2794         .m_intent_lock      = mdc_intent_lock,
2795         .m_link             = mdc_link,
2796         .m_rename           = mdc_rename,
2797         .m_setattr          = mdc_setattr,
2798         .m_setxattr         = mdc_setxattr,
2799         .m_getxattr         = mdc_getxattr,
2800         .m_fsync                = mdc_fsync,
2801         .m_file_resync          = mdc_file_resync,
2802         .m_read_page            = mdc_read_page,
2803         .m_unlink           = mdc_unlink,
2804         .m_cancel_unused    = mdc_cancel_unused,
2805         .m_init_ea_size     = mdc_init_ea_size,
2806         .m_set_lock_data    = mdc_set_lock_data,
2807         .m_lock_match       = mdc_lock_match,
2808         .m_get_lustre_md    = mdc_get_lustre_md,
2809         .m_free_lustre_md   = mdc_free_lustre_md,
2810         .m_set_open_replay_data = mdc_set_open_replay_data,
2811         .m_clear_open_replay_data = mdc_clear_open_replay_data,
2812         .m_intent_getattr_async = mdc_intent_getattr_async,
2813         .m_revalidate_lock      = mdc_revalidate_lock
2814 };
2815
2816 static int __init mdc_init(void)
2817 {
2818         return class_register_type(&mdc_obd_ops, &mdc_md_ops, true, NULL,
2819                                    LUSTRE_MDC_NAME, &mdc_device_type);
2820 }
2821
2822 static void __exit mdc_exit(void)
2823 {
2824         class_unregister_type(LUSTRE_MDC_NAME);
2825 }
2826
2827 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2828 MODULE_DESCRIPTION("Lustre Metadata Client");
2829 MODULE_VERSION(LUSTRE_VERSION_STRING);
2830 MODULE_LICENSE("GPL");
2831
2832 module_init(mdc_init);
2833 module_exit(mdc_exit);