lustre/mdc/mdc_request.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 #define DEBUG_SUBSYSTEM S_MDC
  34
  35 #include <linux/init.h>
  36 #include <linux/kthread.h>
  37 #include <linux/module.h>
  38 #include <linux/pagemap.h>
  39 #include <linux/user_namespace.h>
  40 #include <linux/utsname.h>
  41 #ifdef HAVE_UIDGID_HEADER
  42 # include <linux/uidgid.h>
  43 #endif
  44
  45 #include <lustre_errno.h>
  46
  47 #include <cl_object.h>
  48 #include <llog_swab.h>
  49 #include <lprocfs_status.h>
  50 #include <lustre_acl.h>
  51 #include <lustre_fid.h>
  52 #include <uapi/linux/lustre/lustre_ioctl.h>
  53 #include <lustre_kernelcomm.h>
  54 #include <lustre_lmv.h>
  55 #include <lustre_log.h>
  56 #include <uapi/linux/lustre/lustre_param.h>
  57 #include <lustre_swab.h>
  58 #include <obd_class.h>
  59 #include <lustre_osc.h>
  60
  61 #include "mdc_internal.h"
  62
  63 #define REQUEST_MINOR 244
  64
  65 static int mdc_cleanup(struct obd_device *obd);
  66
  67 static inline int mdc_queue_wait(struct ptlrpc_request *req)
  68 {
  69         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
  70         int rc;
  71
  72         /* obd_get_request_slot() ensures that this client has no more
  73          * than cl_max_rpcs_in_flight RPCs simultaneously inf light
  74          * against an MDT. */
  75         rc = obd_get_request_slot(cli);
  76         if (rc != 0)
  77                 return rc;
  78
  79         rc = ptlrpc_queue_wait(req);
  80         obd_put_request_slot(cli);
  81
  82         return rc;
  83 }
  84
  85 /*
  86  * Send MDS_GET_ROOT RPC to fetch root FID.
  87  *
  88  * If \a fileset is not NULL it should contain a subdirectory off
  89  * the ROOT/ directory to be mounted on the client. Return the FID
  90  * of the subdirectory to the client to mount onto its mountpoint.
  91  *
  92  * \param[in]   imp     MDC import
  93  * \param[in]   fileset fileset name, which could be NULL
  94  * \param[out]  rootfid root FID of this mountpoint
  95  * \param[out]  pc      root capa will be unpacked and saved in this pointer
  96  *
  97  * \retval      0 on success, negative errno on failure
  98  */
  99 static int mdc_get_root(struct obd_export *exp, const char *fileset,
 100                          struct lu_fid *rootfid)
 101 {
 102         struct ptlrpc_request   *req;
 103         struct mdt_body         *body;
 104         int                      rc;
 105
 106         ENTRY;
 107
 108         if (fileset && !(exp_connect_flags(exp) & OBD_CONNECT_SUBTREE))
 109                 RETURN(-ENOTSUPP);
 110
 111         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 112                                 &RQF_MDS_GET_ROOT);
 113         if (req == NULL)
 114                 RETURN(-ENOMEM);
 115
 116         if (fileset != NULL)
 117                 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 118                                      strlen(fileset) + 1);
 119         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_ROOT);
 120         if (rc) {
 121                 ptlrpc_request_free(req);
 122                 RETURN(rc);
 123         }
 124         mdc_pack_body(req, NULL, 0, 0, -1, 0);
 125         if (fileset != NULL) {
 126                 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
 127
 128                 memcpy(name, fileset, strlen(fileset));
 129         }
 130         lustre_msg_add_flags(req->rq_reqmsg, LUSTRE_IMP_FULL);
 131         req->rq_send_state = LUSTRE_IMP_FULL;
 132
 133         ptlrpc_request_set_replen(req);
 134
 135         rc = ptlrpc_queue_wait(req);
 136         if (rc)
 137                 GOTO(out, rc);
 138
 139         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 140         if (body == NULL)
 141                 GOTO(out, rc = -EPROTO);
 142
 143         *rootfid = body->mbo_fid1;
 144         CDEBUG(D_NET, "root fid="DFID", last_committed=%llu\n",
 145                PFID(rootfid), lustre_msg_get_last_committed(req->rq_repmsg));
 146         EXIT;
 147 out:
 148         ptlrpc_req_finished(req);
 149
 150         return rc;
 151 }
 152
 153 /*
 154  * This function now is known to always saying that it will receive 4 buffers
 155  * from server. Even for cases when acl_size and md_size is zero, RPC header
 156  * will contain 4 fields and RPC itself will contain zero size fields. This is
 157  * because mdt_getattr*() _always_ returns 4 fields, but if acl is not needed
 158  * and thus zero, it shrinks it, making zero size. The same story about
 159  * md_size. And this is course of problem when client waits for smaller number
 160  * of fields. This issue will be fixed later when client gets aware of RPC
 161  * layouts.  --umka
 162  */
 163 static int mdc_getattr_common(struct obd_export *exp,
 164                               struct ptlrpc_request *req)
 165 {
 166         struct req_capsule *pill = &req->rq_pill;
 167         struct mdt_body    *body;
 168         void               *eadata;
 169         int                 rc;
 170         ENTRY;
 171
 172         /* Request message already built. */
 173         rc = ptlrpc_queue_wait(req);
 174         if (rc != 0)
 175                 RETURN(rc);
 176
 177         /* sanity check for the reply */
 178         body = req_capsule_server_get(pill, &RMF_MDT_BODY);
 179         if (body == NULL)
 180                 RETURN(-EPROTO);
 181
 182         CDEBUG(D_NET, "mode: %o\n", body->mbo_mode);
 183
 184         mdc_update_max_ea_from_body(exp, body);
 185         if (body->mbo_eadatasize != 0) {
 186                 eadata = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
 187                                                       body->mbo_eadatasize);
 188                 if (eadata == NULL)
 189                         RETURN(-EPROTO);
 190         }
 191
 192         RETURN(0);
 193 }
 194
 195 static void mdc_reset_acl_req(struct ptlrpc_request *req)
 196 {
 197         spin_lock(&req->rq_early_free_lock);
 198         sptlrpc_cli_free_repbuf(req);
 199         req->rq_repbuf = NULL;
 200         req->rq_repbuf_len = 0;
 201         req->rq_repdata = NULL;
 202         req->rq_reqdata_len = 0;
 203         spin_unlock(&req->rq_early_free_lock);
 204 }
 205
 206 static int mdc_getattr(struct obd_export *exp, struct md_op_data *op_data,
 207                        struct ptlrpc_request **request)
 208 {
 209         struct ptlrpc_request *req;
 210         struct obd_import *imp = class_exp2cliimp(exp);
 211         __u32 acl_bufsize = LUSTRE_POSIX_ACL_MAX_SIZE_OLD;
 212         int rc;
 213         ENTRY;
 214
 215         /* Single MDS without an LMV case */
 216         if (op_data->op_flags & MF_GET_MDT_IDX) {
 217                 op_data->op_mds = 0;
 218                 RETURN(0);
 219         }
 220
 221         *request = NULL;
 222         req = ptlrpc_request_alloc(imp, &RQF_MDS_GETATTR);
 223         if (req == NULL)
 224                 RETURN(-ENOMEM);
 225
 226         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR);
 227         if (rc) {
 228                 ptlrpc_request_free(req);
 229                 RETURN(rc);
 230         }
 231
 232 again:
 233         mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
 234                       op_data->op_mode, -1, 0);
 235         req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
 236         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 237                              op_data->op_mode);
 238         ptlrpc_request_set_replen(req);
 239
 240         rc = mdc_getattr_common(exp, req);
 241         if (rc) {
 242                 if (rc == -ERANGE &&
 243                     acl_bufsize != imp->imp_connect_data.ocd_max_easize) {
 244                         acl_bufsize = imp->imp_connect_data.ocd_max_easize;
 245                         mdc_reset_acl_req(req);
 246                         goto again;
 247                 }
 248
 249                 ptlrpc_req_finished(req);
 250         } else {
 251                 *request = req;
 252         }
 253
 254         RETURN(rc);
 255 }
 256
 257 static int mdc_getattr_name(struct obd_export *exp, struct md_op_data *op_data,
 258                             struct ptlrpc_request **request)
 259 {
 260         struct ptlrpc_request *req;
 261         struct obd_import *imp = class_exp2cliimp(exp);
 262         __u32 acl_bufsize = LUSTRE_POSIX_ACL_MAX_SIZE_OLD;
 263         int rc;
 264         ENTRY;
 265
 266         *request = NULL;
 267         req = ptlrpc_request_alloc(imp, &RQF_MDS_GETATTR_NAME);
 268         if (req == NULL)
 269                 RETURN(-ENOMEM);
 270
 271         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 272                              op_data->op_namelen + 1);
 273
 274         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR_NAME);
 275         if (rc) {
 276                 ptlrpc_request_free(req);
 277                 RETURN(rc);
 278         }
 279
 280         if (op_data->op_name) {
 281                 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
 282                 LASSERT(strnlen(op_data->op_name, op_data->op_namelen) ==
 283                                 op_data->op_namelen);
 284                 memcpy(name, op_data->op_name, op_data->op_namelen);
 285         }
 286
 287 again:
 288         mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
 289                       op_data->op_mode, op_data->op_suppgids[0], 0);
 290         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 291                              op_data->op_mode);
 292         req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
 293         ptlrpc_request_set_replen(req);
 294
 295         rc = mdc_getattr_common(exp, req);
 296         if (rc) {
 297                 if (rc == -ERANGE &&
 298                     acl_bufsize != imp->imp_connect_data.ocd_max_easize) {
 299                         acl_bufsize = imp->imp_connect_data.ocd_max_easize;
 300                         mdc_reset_acl_req(req);
 301                         goto again;
 302                 }
 303
 304                 ptlrpc_req_finished(req);
 305         } else {
 306                 *request = req;
 307         }
 308
 309         RETURN(rc);
 310 }
 311
 312 static int mdc_xattr_common(struct obd_export *exp,const struct req_format *fmt,
 313                             const struct lu_fid *fid, int opcode, u64 valid,
 314                             const char *xattr_name, const char *input,
 315                             int input_size, int output_size, int flags,
 316                             __u32 suppgid, struct ptlrpc_request **request)
 317 {
 318         struct ptlrpc_request *req;
 319         int   xattr_namelen = 0;
 320         char *tmp;
 321         int   rc;
 322         ENTRY;
 323
 324         *request = NULL;
 325         req = ptlrpc_request_alloc(class_exp2cliimp(exp), fmt);
 326         if (req == NULL)
 327                 RETURN(-ENOMEM);
 328
 329         if (xattr_name) {
 330                 xattr_namelen = strlen(xattr_name) + 1;
 331                 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 332                                      xattr_namelen);
 333         }
 334         if (input_size) {
 335                 LASSERT(input);
 336                 req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT,
 337                                      input_size);
 338         }
 339
 340         /* Flush local XATTR locks to get rid of a possible cancel RPC */
 341         if (opcode == MDS_REINT && fid_is_sane(fid) &&
 342             exp->exp_connect_data.ocd_ibits_known & MDS_INODELOCK_XATTR) {
 343                 struct list_head cancels = LIST_HEAD_INIT(cancels);
 344                 int count;
 345
 346                 /* Without that packing would fail */
 347                 if (input_size == 0)
 348                         req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
 349                                              RCL_CLIENT, 0);
 350
 351                 count = mdc_resource_get_unused(exp, fid,
 352                                                 &cancels, LCK_EX,
 353                                                 MDS_INODELOCK_XATTR);
 354
 355                 rc = mdc_prep_elc_req(exp, req, MDS_REINT, &cancels, count);
 356                 if (rc) {
 357                         ptlrpc_request_free(req);
 358                         RETURN(rc);
 359                 }
 360         } else {
 361                 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, opcode);
 362                 if (rc) {
 363                         ptlrpc_request_free(req);
 364                         RETURN(rc);
 365                 }
 366         }
 367
 368         if (opcode == MDS_REINT) {
 369                 struct mdt_rec_setxattr *rec;
 370
 371                 CLASSERT(sizeof(struct mdt_rec_setxattr) ==
 372                          sizeof(struct mdt_rec_reint));
 373                 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
 374                 rec->sx_opcode = REINT_SETXATTR;
 375                 rec->sx_fsuid  = from_kuid(&init_user_ns, current_fsuid());
 376                 rec->sx_fsgid  = from_kgid(&init_user_ns, current_fsgid());
 377                 rec->sx_cap    = cfs_curproc_cap_pack();
 378                 rec->sx_suppgid1 = suppgid;
 379                 rec->sx_suppgid2 = -1;
 380                 rec->sx_fid    = *fid;
 381                 rec->sx_valid  = valid | OBD_MD_FLCTIME;
 382                 rec->sx_time   = ktime_get_real_seconds();
 383                 rec->sx_size   = output_size;
 384                 rec->sx_flags  = flags;
 385         } else {
 386                 mdc_pack_body(req, fid, valid, output_size, suppgid, flags);
 387         }
 388
 389         if (xattr_name) {
 390                 tmp = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
 391                 memcpy(tmp, xattr_name, xattr_namelen);
 392         }
 393         if (input_size) {
 394                 tmp = req_capsule_client_get(&req->rq_pill, &RMF_EADATA);
 395                 memcpy(tmp, input, input_size);
 396         }
 397
 398         if (req_capsule_has_field(&req->rq_pill, &RMF_EADATA, RCL_SERVER))
 399                 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
 400                                      RCL_SERVER, output_size);
 401         ptlrpc_request_set_replen(req);
 402
 403         /* make rpc */
 404         if (opcode == MDS_REINT)
 405                 mdc_get_mod_rpc_slot(req, NULL);
 406
 407         rc = ptlrpc_queue_wait(req);
 408
 409         if (opcode == MDS_REINT)
 410                 mdc_put_mod_rpc_slot(req, NULL);
 411
 412         if (rc)
 413                 ptlrpc_req_finished(req);
 414         else
 415                 *request = req;
 416         RETURN(rc);
 417 }
 418
 419 static int mdc_setxattr(struct obd_export *exp, const struct lu_fid *fid,
 420                         u64 obd_md_valid, const char *name,
 421                         const void *value, size_t value_size,
 422                         unsigned int xattr_flags, u32 suppgid,
 423                         struct ptlrpc_request **req)
 424 {
 425         LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
 426                 obd_md_valid == OBD_MD_FLXATTRRM);
 427
 428         return mdc_xattr_common(exp, &RQF_MDS_REINT_SETXATTR,
 429                                 fid, MDS_REINT, obd_md_valid, name,
 430                                 value, value_size, 0, xattr_flags, suppgid,
 431                                 req);
 432 }
 433
 434 static int mdc_getxattr(struct obd_export *exp, const struct lu_fid *fid,
 435                         u64 obd_md_valid, const char *name, size_t buf_size,
 436                         struct ptlrpc_request **req)
 437 {
 438         LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
 439                 obd_md_valid == OBD_MD_FLXATTRLS);
 440
 441         return mdc_xattr_common(exp, &RQF_MDS_GETXATTR, fid, MDS_GETXATTR,
 442                                 obd_md_valid, name, NULL, 0, buf_size, 0, -1,
 443                                 req);
 444 }
 445
 446 #ifdef CONFIG_FS_POSIX_ACL
 447 static int mdc_unpack_acl(struct ptlrpc_request *req, struct lustre_md *md)
 448 {
 449         struct req_capsule     *pill = &req->rq_pill;
 450         struct mdt_body        *body = md->body;
 451         struct posix_acl       *acl;
 452         void                   *buf;
 453         int                     rc;
 454         ENTRY;
 455
 456         if (!body->mbo_aclsize)
 457                 RETURN(0);
 458
 459         buf = req_capsule_server_sized_get(pill, &RMF_ACL, body->mbo_aclsize);
 460
 461         if (!buf)
 462                 RETURN(-EPROTO);
 463
 464         acl = posix_acl_from_xattr(&init_user_ns, buf, body->mbo_aclsize);
 465         if (acl == NULL)
 466                 RETURN(0);
 467         if (IS_ERR(acl)) {
 468                 rc = PTR_ERR(acl);
 469                 CERROR("convert xattr to acl: %d\n", rc);
 470                 RETURN(rc);
 471         }
 472
 473         rc = posix_acl_valid(&init_user_ns, acl);
 474         if (rc) {
 475                 CERROR("validate acl: %d\n", rc);
 476                 posix_acl_release(acl);
 477                 RETURN(rc);
 478         }
 479
 480         md->posix_acl = acl;
 481         RETURN(0);
 482 }
 483 #else
 484 #define mdc_unpack_acl(req, md) 0
 485 #endif
 486
 487 int mdc_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
 488                       struct obd_export *dt_exp, struct obd_export *md_exp,
 489                       struct lustre_md *md)
 490 {
 491         struct req_capsule *pill = &req->rq_pill;
 492         int rc;
 493         ENTRY;
 494
 495         LASSERT(md);
 496         memset(md, 0, sizeof(*md));
 497
 498         md->body = req_capsule_server_get(pill, &RMF_MDT_BODY);
 499         LASSERT(md->body != NULL);
 500
 501         if (md->body->mbo_valid & OBD_MD_FLEASIZE) {
 502                 if (!S_ISREG(md->body->mbo_mode)) {
 503                         CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, should be a "
 504                                "regular file, but is not\n");
 505                         GOTO(out, rc = -EPROTO);
 506                 }
 507
 508                 if (md->body->mbo_eadatasize == 0) {
 509                         CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, "
 510                                "but eadatasize 0\n");
 511                         GOTO(out, rc = -EPROTO);
 512                 }
 513
 514                 md->layout.lb_len = md->body->mbo_eadatasize;
 515                 md->layout.lb_buf = req_capsule_server_sized_get(pill,
 516                                                         &RMF_MDT_MD,
 517                                                         md->layout.lb_len);
 518                 if (md->layout.lb_buf == NULL)
 519                         GOTO(out, rc = -EPROTO);
 520         } else if (md->body->mbo_valid & OBD_MD_FLDIREA) {
 521                 const union lmv_mds_md *lmv;
 522                 size_t lmv_size;
 523
 524                 if (!S_ISDIR(md->body->mbo_mode)) {
 525                         CDEBUG(D_INFO, "OBD_MD_FLDIREA set, should be a "
 526                                "directory, but is not\n");
 527                         GOTO(out, rc = -EPROTO);
 528                 }
 529
 530                 lmv_size = md->body->mbo_eadatasize;
 531                 if (lmv_size == 0) {
 532                         CDEBUG(D_INFO, "OBD_MD_FLDIREA is set, "
 533                                "but eadatasize 0\n");
 534                         RETURN(-EPROTO);
 535                 }
 536
 537                 if (md->body->mbo_valid & OBD_MD_MEA) {
 538                         lmv = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
 539                                                            lmv_size);
 540                         if (lmv == NULL)
 541                                 GOTO(out, rc = -EPROTO);
 542
 543                         rc = md_unpackmd(md_exp, &md->lmv, lmv, lmv_size);
 544                         if (rc < 0)
 545                                 GOTO(out, rc);
 546
 547                         if (rc < (typeof(rc))sizeof(*md->lmv)) {
 548                                 CDEBUG(D_INFO, "size too small:  "
 549                                        "rc < sizeof(*md->lmv) (%d < %d)\n",
 550                                         rc, (int)sizeof(*md->lmv));
 551                                 GOTO(out, rc = -EPROTO);
 552                         }
 553                 }
 554         }
 555         rc = 0;
 556
 557         if (md->body->mbo_valid & OBD_MD_FLACL) {
 558                 /* for ACL, it's possible that FLACL is set but aclsize is zero.
 559                  * only when aclsize != 0 there's an actual segment for ACL
 560                  * in reply buffer.
 561                  */
 562                 if (md->body->mbo_aclsize) {
 563                         rc = mdc_unpack_acl(req, md);
 564                         if (rc)
 565                                 GOTO(out, rc);
 566 #ifdef CONFIG_FS_POSIX_ACL
 567                 } else {
 568                         md->posix_acl = NULL;
 569 #endif
 570                 }
 571         }
 572
 573         EXIT;
 574 out:
 575         if (rc) {
 576 #ifdef CONFIG_FS_POSIX_ACL
 577                 posix_acl_release(md->posix_acl);
 578 #endif
 579         }
 580         return rc;
 581 }
 582
 583 int mdc_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
 584 {
 585         ENTRY;
 586         RETURN(0);
 587 }
 588
 589 void mdc_replay_open(struct ptlrpc_request *req)
 590 {
 591         struct md_open_data *mod = req->rq_cb_data;
 592         struct ptlrpc_request *close_req;
 593         struct obd_client_handle *och;
 594         struct lustre_handle old;
 595         struct mdt_body *body;
 596         ENTRY;
 597
 598         if (mod == NULL) {
 599                 DEBUG_REQ(D_ERROR, req,
 600                           "Can't properly replay without open data.");
 601                 EXIT;
 602                 return;
 603         }
 604
 605         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 606         LASSERT(body != NULL);
 607
 608         spin_lock(&req->rq_lock);
 609         och = mod->mod_och;
 610         if (och && och->och_fh.cookie)
 611                 req->rq_early_free_repbuf = 1;
 612         else
 613                 req->rq_early_free_repbuf = 0;
 614         spin_unlock(&req->rq_lock);
 615
 616         if (req->rq_early_free_repbuf) {
 617                 struct lustre_handle *file_fh;
 618
 619                 LASSERT(och->och_magic == OBD_CLIENT_HANDLE_MAGIC);
 620
 621                 file_fh = &och->och_fh;
 622                 CDEBUG(D_HA, "updating handle from %#llx to %#llx\n",
 623                        file_fh->cookie, body->mbo_handle.cookie);
 624                 old = *file_fh;
 625                 *file_fh = body->mbo_handle;
 626         }
 627
 628         close_req = mod->mod_close_req;
 629         if (close_req) {
 630                 __u32 opc = lustre_msg_get_opc(close_req->rq_reqmsg);
 631                 struct mdt_ioepoch *epoch;
 632
 633                 LASSERT(opc == MDS_CLOSE);
 634                 epoch = req_capsule_client_get(&close_req->rq_pill,
 635                                                &RMF_MDT_EPOCH);
 636                 LASSERT(epoch);
 637
 638                 if (req->rq_early_free_repbuf)
 639                         LASSERT(!memcmp(&old, &epoch->mio_handle, sizeof(old)));
 640
 641                 DEBUG_REQ(D_HA, close_req, "updating close body with new fh");
 642                 epoch->mio_handle = body->mbo_handle;
 643         }
 644         EXIT;
 645 }
 646
 647 void mdc_commit_open(struct ptlrpc_request *req)
 648 {
 649         struct md_open_data *mod = req->rq_cb_data;
 650         if (mod == NULL)
 651                 return;
 652
 653         /**
 654          * No need to touch md_open_data::mod_och, it holds a reference on
 655          * \var mod and will zero references to each other, \var mod will be
 656          * freed after that when md_open_data::mod_och will put the reference.
 657          */
 658
 659         /**
 660          * Do not let open request to disappear as it still may be needed
 661          * for close rpc to happen (it may happen on evict only, otherwise
 662          * ptlrpc_request::rq_replay does not let mdc_commit_open() to be
 663          * called), just mark this rpc as committed to distinguish these 2
 664          * cases, see mdc_close() for details. The open request reference will
 665          * be put along with freeing \var mod.
 666          */
 667         ptlrpc_request_addref(req);
 668         spin_lock(&req->rq_lock);
 669         req->rq_committed = 1;
 670         spin_unlock(&req->rq_lock);
 671         req->rq_cb_data = NULL;
 672         obd_mod_put(mod);
 673 }
 674
 675 int mdc_set_open_replay_data(struct obd_export *exp,
 676                              struct obd_client_handle *och,
 677                              struct lookup_intent *it)
 678 {
 679         struct md_open_data     *mod;
 680         struct mdt_rec_create   *rec;
 681         struct mdt_body         *body;
 682         struct ptlrpc_request   *open_req = it->it_request;
 683         struct obd_import       *imp = open_req->rq_import;
 684         ENTRY;
 685
 686         if (!open_req->rq_replay)
 687                 RETURN(0);
 688
 689         rec = req_capsule_client_get(&open_req->rq_pill, &RMF_REC_REINT);
 690         body = req_capsule_server_get(&open_req->rq_pill, &RMF_MDT_BODY);
 691         LASSERT(rec != NULL);
 692         /* Incoming message in my byte order (it's been swabbed). */
 693         /* Outgoing messages always in my byte order. */
 694         LASSERT(body != NULL);
 695
 696         /* Only if the import is replayable, we set replay_open data */
 697         if (och && imp->imp_replayable) {
 698                 mod = obd_mod_alloc();
 699                 if (mod == NULL) {
 700                         DEBUG_REQ(D_ERROR, open_req,
 701                                   "Can't allocate md_open_data");
 702                         RETURN(0);
 703                 }
 704
 705                 /**
 706                  * Take a reference on \var mod, to be freed on mdc_close().
 707                  * It protects \var mod from being freed on eviction (commit
 708                  * callback is called despite rq_replay flag).
 709                  * Another reference for \var och.
 710                  */
 711                 obd_mod_get(mod);
 712                 obd_mod_get(mod);
 713
 714                 spin_lock(&open_req->rq_lock);
 715                 och->och_mod = mod;
 716                 mod->mod_och = och;
 717                 mod->mod_is_create = it_disposition(it, DISP_OPEN_CREATE) ||
 718                                      it_disposition(it, DISP_OPEN_STRIPE);
 719                 mod->mod_open_req = open_req;
 720                 open_req->rq_cb_data = mod;
 721                 open_req->rq_commit_cb = mdc_commit_open;
 722                 open_req->rq_early_free_repbuf = 1;
 723                 spin_unlock(&open_req->rq_lock);
 724         }
 725
 726         rec->cr_fid2 = body->mbo_fid1;
 727         rec->cr_old_handle.cookie = body->mbo_handle.cookie;
 728         open_req->rq_replay_cb = mdc_replay_open;
 729         if (!fid_is_sane(&body->mbo_fid1)) {
 730                 DEBUG_REQ(D_ERROR, open_req, "Saving replay request with "
 731                           "insane fid");
 732                 LBUG();
 733         }
 734
 735         DEBUG_REQ(D_RPCTRACE, open_req, "Set up open replay data");
 736         RETURN(0);
 737 }
 738
 739 static void mdc_free_open(struct md_open_data *mod)
 740 {
 741         int committed = 0;
 742
 743         if (mod->mod_is_create == 0 &&
 744             imp_connect_disp_stripe(mod->mod_open_req->rq_import))
 745                 committed = 1;
 746
 747         /**
 748          * No reason to asssert here if the open request has
 749          * rq_replay == 1. It means that mdc_close failed, and
 750          * close request wasn`t sent. It is not fatal to client.
 751          * The worst thing is eviction if the client gets open lock
 752          **/
 753
 754         DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "free open request rq_replay"
 755                   "= %d\n", mod->mod_open_req->rq_replay);
 756
 757         ptlrpc_request_committed(mod->mod_open_req, committed);
 758         if (mod->mod_close_req)
 759                 ptlrpc_request_committed(mod->mod_close_req, committed);
 760 }
 761
 762 int mdc_clear_open_replay_data(struct obd_export *exp,
 763                                struct obd_client_handle *och)
 764 {
 765         struct md_open_data *mod = och->och_mod;
 766         ENTRY;
 767
 768         /**
 769          * It is possible to not have \var mod in a case of eviction between
 770          * lookup and ll_file_open().
 771          **/
 772         if (mod == NULL)
 773                 RETURN(0);
 774
 775         LASSERT(mod != LP_POISON);
 776         LASSERT(mod->mod_open_req != NULL);
 777
 778         spin_lock(&mod->mod_open_req->rq_lock);
 779         if (mod->mod_och)
 780                 mod->mod_och->och_fh.cookie = 0;
 781         mod->mod_open_req->rq_early_free_repbuf = 0;
 782         spin_unlock(&mod->mod_open_req->rq_lock);
 783         mdc_free_open(mod);
 784
 785         mod->mod_och = NULL;
 786         och->och_mod = NULL;
 787         obd_mod_put(mod);
 788
 789         RETURN(0);
 790 }
 791
 792 static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
 793                      struct md_open_data *mod, struct ptlrpc_request **request)
 794 {
 795         struct obd_device     *obd = class_exp2obd(exp);
 796         struct ptlrpc_request *req;
 797         struct req_format     *req_fmt;
 798         size_t                 u32_count = 0;
 799         int                    rc;
 800         int                    saved_rc = 0;
 801         ENTRY;
 802
 803         CDEBUG(D_INODE, "%s: "DFID" file closed with intent: %x\n",
 804                exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
 805                op_data->op_bias);
 806
 807         if (op_data->op_bias & MDS_CLOSE_INTENT) {
 808                 req_fmt = &RQF_MDS_CLOSE_INTENT;
 809                 if (op_data->op_bias & MDS_HSM_RELEASE) {
 810                         /* allocate a FID for volatile file */
 811                         rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2,
 812                                            op_data);
 813                         if (rc < 0) {
 814                                 CERROR("%s: "DFID" allocating FID: rc = %d\n",
 815                                        obd->obd_name, PFID(&op_data->op_fid1),
 816                                        rc);
 817                                 /* save the errcode and proceed to close */
 818                                 saved_rc = rc;
 819                         }
 820                 }
 821                 if (op_data->op_bias & MDS_CLOSE_RESYNC_DONE) {
 822                         size_t count = op_data->op_data_size / sizeof(__u32);
 823
 824                         if (count > INLINE_RESYNC_ARRAY_SIZE)
 825                                 u32_count = count;
 826                 }
 827         } else {
 828                 req_fmt = &RQF_MDS_CLOSE;
 829         }
 830
 831         *request = NULL;
 832         if (OBD_FAIL_CHECK(OBD_FAIL_MDC_CLOSE))
 833                 req = NULL;
 834         else
 835                 req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);
 836
 837         /* Ensure that this close's handle is fixed up during replay. */
 838         if (likely(mod != NULL)) {
 839                 LASSERTF(mod->mod_open_req != NULL &&
 840                          mod->mod_open_req->rq_type != LI_POISON,
 841                          "POISONED open %p!\n", mod->mod_open_req);
 842
 843                 mod->mod_close_req = req;
 844
 845                 DEBUG_REQ(D_HA, mod->mod_open_req, "matched open");
 846                 /* We no longer want to preserve this open for replay even
 847                  * though the open was committed. b=3632, b=3633 */
 848                 spin_lock(&mod->mod_open_req->rq_lock);
 849                 mod->mod_open_req->rq_replay = 0;
 850                 spin_unlock(&mod->mod_open_req->rq_lock);
 851         } else {
 852                 CDEBUG(D_HA, "couldn't find open req; expecting close error\n");
 853         }
 854         if (req == NULL) {
 855                 /**
 856                  * TODO: repeat close after errors
 857                  */
 858                 CWARN("%s: close of FID "DFID" failed, file reference will be "
 859                       "dropped when this client unmounts or is evicted\n",
 860                       obd->obd_name, PFID(&op_data->op_fid1));
 861                 GOTO(out, rc = -ENOMEM);
 862         }
 863
 864         if (u32_count > 0)
 865                 req_capsule_set_size(&req->rq_pill, &RMF_U32, RCL_CLIENT,
 866                                      u32_count * sizeof(__u32));
 867
 868         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
 869         if (rc) {
 870                 ptlrpc_request_free(req);
 871                 req = NULL;
 872                 GOTO(out, rc);
 873         }
 874
 875         /* To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
 876          * portal whose threads are not taking any DLM locks and are therefore
 877          * always progressing */
 878         req->rq_request_portal = MDS_READPAGE_PORTAL;
 879         ptlrpc_at_set_req_timeout(req);
 880
 881
 882         mdc_close_pack(req, op_data);
 883
 884         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 885                              obd->u.cli.cl_default_mds_easize);
 886
 887         ptlrpc_request_set_replen(req);
 888
 889         mdc_get_mod_rpc_slot(req, NULL);
 890         rc = ptlrpc_queue_wait(req);
 891         mdc_put_mod_rpc_slot(req, NULL);
 892
 893         if (req->rq_repmsg == NULL) {
 894                 CDEBUG(D_RPCTRACE, "request failed to send: %p, %d\n", req,
 895                        req->rq_status);
 896                 if (rc == 0)
 897                         rc = req->rq_status ?: -EIO;
 898         } else if (rc == 0 || rc == -EAGAIN) {
 899                 struct mdt_body *body;
 900
 901                 rc = lustre_msg_get_status(req->rq_repmsg);
 902                 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
 903                         DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR, err "
 904                                   "= %d", rc);
 905                         if (rc > 0)
 906                                 rc = -rc;
 907                 }
 908                 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 909                 if (body == NULL)
 910                         rc = -EPROTO;
 911         } else if (rc == -ESTALE) {
 912                 /**
 913                  * it can be allowed error after 3633 if open was committed and
 914                  * server failed before close was sent. Let's check if mod
 915                  * exists and return no error in that case
 916                  */
 917                 if (mod) {
 918                         DEBUG_REQ(D_HA, req, "Reset ESTALE = %d", rc);
 919                         LASSERT(mod->mod_open_req != NULL);
 920                         if (mod->mod_open_req->rq_committed)
 921                                 rc = 0;
 922                 }
 923         }
 924
 925 out:
 926         if (mod) {
 927                 if (rc != 0)
 928                         mod->mod_close_req = NULL;
 929                 /* Since now, mod is accessed through open_req only,
 930                  * thus close req does not keep a reference on mod anymore. */
 931                 obd_mod_put(mod);
 932         }
 933         *request = req;
 934
 935         RETURN(rc < 0 ? rc : saved_rc);
 936 }
 937
 938 static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid,
 939                        u64 offset, struct page **pages, int npages,
 940                        struct ptlrpc_request **request)
 941 {
 942         struct ptlrpc_request   *req;
 943         struct ptlrpc_bulk_desc *desc;
 944         int                      i;
 945         wait_queue_head_t        waitq;
 946         int                      resends = 0;
 947         struct l_wait_info       lwi;
 948         int                      rc;
 949         ENTRY;
 950
 951         *request = NULL;
 952         init_waitqueue_head(&waitq);
 953
 954 restart_bulk:
 955         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE);
 956         if (req == NULL)
 957                 RETURN(-ENOMEM);
 958
 959         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE);
 960         if (rc) {
 961                 ptlrpc_request_free(req);
 962                 RETURN(rc);
 963         }
 964
 965         req->rq_request_portal = MDS_READPAGE_PORTAL;
 966         ptlrpc_at_set_req_timeout(req);
 967
 968         desc = ptlrpc_prep_bulk_imp(req, npages, 1,
 969                                     PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
 970                                     MDS_BULK_PORTAL,
 971                                     &ptlrpc_bulk_kiov_pin_ops);
 972         if (desc == NULL) {
 973                 ptlrpc_req_finished(req);
 974                 RETURN(-ENOMEM);
 975         }
 976
 977         /* NB req now owns desc and will free it when it gets freed */
 978         for (i = 0; i < npages; i++)
 979                 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
 980                                                  PAGE_SIZE);
 981
 982         mdc_readdir_pack(req, offset, PAGE_SIZE * npages, fid);
 983
 984         ptlrpc_request_set_replen(req);
 985         rc = ptlrpc_queue_wait(req);
 986         if (rc) {
 987                 ptlrpc_req_finished(req);
 988                 if (rc != -ETIMEDOUT)
 989                         RETURN(rc);
 990
 991                 resends++;
 992                 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
 993                         CERROR("%s: too many resend retries: rc = %d\n",
 994                                exp->exp_obd->obd_name, -EIO);
 995                         RETURN(-EIO);
 996                 }
 997                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL,
 998                                        NULL);
 999                 l_wait_event(waitq, 0, &lwi);
1000
1001                 goto restart_bulk;
1002         }
1003
1004         rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1005                                           req->rq_bulk->bd_nob_transferred);
1006         if (rc < 0) {
1007                 ptlrpc_req_finished(req);
1008                 RETURN(rc);
1009         }
1010
1011         if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) {
1012                 CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n",
1013                        exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred,
1014                        PAGE_SIZE * npages);
1015                 ptlrpc_req_finished(req);
1016                 RETURN(-EPROTO);
1017         }
1018
1019         *request = req;
1020         RETURN(0);
1021 }
1022
1023 static void mdc_release_page(struct page *page, int remove)
1024 {
1025         if (remove) {
1026                 lock_page(page);
1027                 if (likely(page->mapping != NULL))
1028                         truncate_complete_page(page->mapping, page);
1029                 unlock_page(page);
1030         }
1031         put_page(page);
1032 }
1033
1034 static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash,
1035                                     __u64 *start, __u64 *end, int hash64)
1036 {
1037         /*
1038          * Complement of hash is used as an index so that
1039          * radix_tree_gang_lookup() can be used to find a page with starting
1040          * hash _smaller_ than one we are looking for.
1041          */
1042         unsigned long offset = hash_x_index(*hash, hash64);
1043         struct page *page;
1044         int found;
1045
1046         spin_lock_irq(&mapping->tree_lock);
1047         found = radix_tree_gang_lookup(&mapping->page_tree,
1048                                        (void **)&page, offset, 1);
1049         if (found > 0 && !radix_tree_exceptional_entry(page)) {
1050                 struct lu_dirpage *dp;
1051
1052                 get_page(page);
1053                 spin_unlock_irq(&mapping->tree_lock);
1054                 /*
1055                  * In contrast to find_lock_page() we are sure that directory
1056                  * page cannot be truncated (while DLM lock is held) and,
1057                  * hence, can avoid restart.
1058                  *
1059                  * In fact, page cannot be locked here at all, because
1060                  * mdc_read_page_remote does synchronous io.
1061                  */
1062                 wait_on_page_locked(page);
1063                 if (PageUptodate(page)) {
1064                         dp = kmap(page);
1065                         if (BITS_PER_LONG == 32 && hash64) {
1066                                 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1067                                 *end   = le64_to_cpu(dp->ldp_hash_end) >> 32;
1068                                 *hash  = *hash >> 32;
1069                         } else {
1070                                 *start = le64_to_cpu(dp->ldp_hash_start);
1071                                 *end   = le64_to_cpu(dp->ldp_hash_end);
1072                         }
1073                         if (unlikely(*start == 1 && *hash == 0))
1074                                 *hash = *start;
1075                         else
1076                                 LASSERTF(*start <= *hash, "start = %#llx"
1077                                          ",end = %#llx,hash = %#llx\n",
1078                                          *start, *end, *hash);
1079                         CDEBUG(D_VFSTRACE, "offset %lx [%#llx %#llx],"
1080                               " hash %#llx\n", offset, *start, *end, *hash);
1081                         if (*hash > *end) {
1082                                 kunmap(page);
1083                                 mdc_release_page(page, 0);
1084                                 page = NULL;
1085                         } else if (*end != *start && *hash == *end) {
1086                                 /*
1087                                  * upon hash collision, remove this page,
1088                                  * otherwise put page reference, and
1089                                  * mdc_read_page_remote() will issue RPC to
1090                                  * fetch the page we want.
1091                                  */
1092                                 kunmap(page);
1093                                 mdc_release_page(page,
1094                                     le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1095                                 page = NULL;
1096                         }
1097                 } else {
1098                         put_page(page);
1099                         page = ERR_PTR(-EIO);
1100                 }
1101         } else {
1102                 spin_unlock_irq(&mapping->tree_lock);
1103                 page = NULL;
1104         }
1105         return page;
1106 }
1107
1108 /*
1109  * Adjust a set of pages, each page containing an array of lu_dirpages,
1110  * so that each page can be used as a single logical lu_dirpage.
1111  *
1112  * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
1113  * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
1114  * struct lu_dirent.  It has size up to LU_PAGE_SIZE. The ldp_hash_end
1115  * value is used as a cookie to request the next lu_dirpage in a
1116  * directory listing that spans multiple pages (two in this example):
1117  *   ________
1118  *  |        |
1119  * .|--------v-------   -----.
1120  * |s|e|f|p|ent|ent| ... |ent|
1121  * '--|--------------   -----'   Each PAGE contains a single
1122  *    '------.                   lu_dirpage.
1123  * .---------v-------   -----.
1124  * |s|e|f|p|ent| 0 | ... | 0 |
1125  * '-----------------   -----'
1126  *
1127  * However, on hosts where the native VM page size (PAGE_SIZE) is
1128  * larger than LU_PAGE_SIZE, a single host page may contain multiple
1129  * lu_dirpages. After reading the lu_dirpages from the MDS, the
1130  * ldp_hash_end of the first lu_dirpage refers to the one immediately
1131  * after it in the same PAGE (arrows simplified for brevity, but
1132  * in general e0==s1, e1==s2, etc.):
1133  *
1134  * .--------------------   -----.
1135  * |s0|e0|f0|p|ent|ent| ... |ent|
1136  * |---v----------------   -----|
1137  * |s1|e1|f1|p|ent|ent| ... |ent|
1138  * |---v----------------   -----|  Here, each PAGE contains
1139  *             ...                 multiple lu_dirpages.
1140  * |---v----------------   -----|
1141  * |s'|e'|f'|p|ent|ent| ... |ent|
1142  * '---|----------------   -----'
1143  *     v
1144  * .----------------------------.
1145  * |        next PAGE           |
1146  *
1147  * This structure is transformed into a single logical lu_dirpage as follows:
1148  *
1149  * - Replace e0 with e' so the request for the next lu_dirpage gets the page
1150  *   labeled 'next PAGE'.
1151  *
1152  * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
1153  *   a hash collision with the next page exists.
1154  *
1155  * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
1156  *   to the first entry of the next lu_dirpage.
1157  */
1158 #if PAGE_SIZE > LU_PAGE_SIZE
1159 static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs)
1160 {
1161         int i;
1162
1163         for (i = 0; i < cfs_pgs; i++) {
1164                 struct lu_dirpage       *dp = kmap(pages[i]);
1165                 struct lu_dirpage       *first = dp;
1166                 struct lu_dirent        *end_dirent = NULL;
1167                 struct lu_dirent        *ent;
1168                 __u64           hash_end = le64_to_cpu(dp->ldp_hash_end);
1169                 __u32           flags = le32_to_cpu(dp->ldp_flags);
1170
1171                 while (--lu_pgs > 0) {
1172                         ent = lu_dirent_start(dp);
1173                         for (end_dirent = ent; ent != NULL;
1174                              end_dirent = ent, ent = lu_dirent_next(ent));
1175
1176                         /* Advance dp to next lu_dirpage. */
1177                         dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
1178
1179                         /* Check if we've reached the end of the PAGE. */
1180                         if (!((unsigned long)dp & ~PAGE_MASK))
1181                                 break;
1182
1183                         /* Save the hash and flags of this lu_dirpage. */
1184                         hash_end = le64_to_cpu(dp->ldp_hash_end);
1185                         flags = le32_to_cpu(dp->ldp_flags);
1186
1187                         /* Check if lu_dirpage contains no entries. */
1188                         if (end_dirent == NULL)
1189                                 break;
1190
1191                         /* Enlarge the end entry lde_reclen from 0 to
1192                          * first entry of next lu_dirpage. */
1193                         LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
1194                         end_dirent->lde_reclen =
1195                                 cpu_to_le16((char *)(dp->ldp_entries) -
1196                                             (char *)end_dirent);
1197                 }
1198
1199                 first->ldp_hash_end = hash_end;
1200                 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
1201                 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
1202
1203                 kunmap(pages[i]);
1204         }
1205         LASSERTF(lu_pgs == 0, "left = %d\n", lu_pgs);
1206 }
1207 #else
1208 #define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0)
1209 #endif  /* PAGE_SIZE > LU_PAGE_SIZE */
1210
1211 /* parameters for readdir page */
1212 struct readpage_param {
1213         struct md_op_data       *rp_mod;
1214         __u64                   rp_off;
1215         int                     rp_hash64;
1216         struct obd_export       *rp_exp;
1217         struct md_callback      *rp_cb;
1218 };
1219
1220 #ifndef HAVE_DELETE_FROM_PAGE_CACHE
1221 static inline void delete_from_page_cache(struct page *page)
1222 {
1223         remove_from_page_cache(page);
1224         put_page(page);
1225 }
1226 #endif
1227
1228 /**
1229  * Read pages from server.
1230  *
1231  * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
1232  * a header lu_dirpage which describes the start/end hash, and whether this
1233  * page is empty (contains no dir entry) or hash collide with next page.
1234  * After client receives reply, several pages will be integrated into dir page
1235  * in PAGE_SIZE (if PAGE_SIZE greater than LU_PAGE_SIZE), and the
1236  * lu_dirpage for this integrated page will be adjusted.
1237  **/
1238 static int mdc_read_page_remote(void *data, struct page *page0)
1239 {
1240         struct readpage_param *rp = data;
1241         struct page **page_pool;
1242         struct page *page;
1243         struct lu_dirpage *dp;
1244         struct md_op_data *op_data = rp->rp_mod;
1245         struct ptlrpc_request *req;
1246         int max_pages;
1247         struct inode *inode;
1248         struct lu_fid *fid;
1249         int rd_pgs = 0; /* number of pages actually read */
1250         int npages;
1251         int i;
1252         int rc;
1253         ENTRY;
1254
1255         max_pages = rp->rp_exp->exp_obd->u.cli.cl_max_pages_per_rpc;
1256         inode = op_data->op_data;
1257         fid = &op_data->op_fid1;
1258         LASSERT(inode != NULL);
1259
1260         OBD_ALLOC(page_pool, sizeof(page_pool[0]) * max_pages);
1261         if (page_pool != NULL) {
1262                 page_pool[0] = page0;
1263         } else {
1264                 page_pool = &page0;
1265                 max_pages = 1;
1266         }
1267
1268         for (npages = 1; npages < max_pages; npages++) {
1269                 page = __page_cache_alloc(mapping_gfp_mask(inode->i_mapping)
1270                                           | __GFP_COLD);
1271                 if (page == NULL)
1272                         break;
1273                 page_pool[npages] = page;
1274         }
1275
1276         rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, page_pool, npages, &req);
1277         if (rc < 0) {
1278                 /* page0 is special, which was added into page cache early */
1279                 delete_from_page_cache(page0);
1280         } else {
1281                 int lu_pgs;
1282
1283                 rd_pgs = (req->rq_bulk->bd_nob_transferred + PAGE_SIZE - 1) >>
1284                         PAGE_SHIFT;
1285                 lu_pgs = req->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
1286                 LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
1287
1288                 CDEBUG(D_INODE, "read %d(%d) pages\n", rd_pgs, lu_pgs);
1289
1290                 mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs);
1291
1292                 SetPageUptodate(page0);
1293         }
1294         unlock_page(page0);
1295
1296         ptlrpc_req_finished(req);
1297         CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages);
1298         for (i = 1; i < npages; i++) {
1299                 unsigned long   offset;
1300                 __u64           hash;
1301                 int ret;
1302
1303                 page = page_pool[i];
1304
1305                 if (rc < 0 || i >= rd_pgs) {
1306                         put_page(page);
1307                         continue;
1308                 }
1309
1310                 SetPageUptodate(page);
1311
1312                 dp = kmap(page);
1313                 hash = le64_to_cpu(dp->ldp_hash_start);
1314                 kunmap(page);
1315
1316                 offset = hash_x_index(hash, rp->rp_hash64);
1317
1318                 prefetchw(&page->flags);
1319                 ret = add_to_page_cache_lru(page, inode->i_mapping, offset,
1320                                             GFP_KERNEL);
1321                 if (ret == 0)
1322                         unlock_page(page);
1323                 else
1324                         CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
1325                                " rc = %d\n", offset, ret);
1326                 put_page(page);
1327         }
1328
1329         if (page_pool != &page0)
1330                 OBD_FREE(page_pool, sizeof(page_pool[0]) * max_pages);
1331
1332         RETURN(rc);
1333 }
1334
1335 /**
1336  * Read dir page from cache first, if it can not find it, read it from
1337  * server and add into the cache.
1338  *
1339  * \param[in] exp       MDC export
1340  * \param[in] op_data   client MD stack parameters, transfering parameters
1341  *                      between different layers on client MD stack.
1342  * \param[in] cb_op     callback required for ldlm lock enqueue during
1343  *                      read page
1344  * \param[in] hash_offset the hash offset of the page to be read
1345  * \param[in] ppage     the page to be read
1346  *
1347  * retval               = 0 get the page successfully
1348  *                      errno(<0) get the page failed
1349  */
1350 static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data,
1351                          struct md_callback *cb_op, __u64 hash_offset,
1352                          struct page **ppage)
1353 {
1354         struct lookup_intent    it = { .it_op = IT_READDIR };
1355         struct page             *page;
1356         struct inode            *dir = op_data->op_data;
1357         struct address_space    *mapping;
1358         struct lu_dirpage       *dp;
1359         __u64                   start = 0;
1360         __u64                   end = 0;
1361         struct lustre_handle    lockh;
1362         struct ptlrpc_request   *enq_req = NULL;
1363         struct readpage_param   rp_param;
1364         int rc;
1365
1366         ENTRY;
1367
1368         *ppage = NULL;
1369
1370         LASSERT(dir != NULL);
1371         mapping = dir->i_mapping;
1372
1373         rc = mdc_intent_lock(exp, op_data, &it, &enq_req,
1374                              cb_op->md_blocking_ast, 0);
1375         if (enq_req != NULL)
1376                 ptlrpc_req_finished(enq_req);
1377
1378         if (rc < 0) {
1379                 CERROR("%s: "DFID" lock enqueue fails: rc = %d\n",
1380                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc);
1381                 RETURN(rc);
1382         }
1383
1384         rc = 0;
1385         lockh.cookie = it.it_lock_handle;
1386         mdc_set_lock_data(exp, &lockh, dir, NULL);
1387
1388         rp_param.rp_off = hash_offset;
1389         rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64;
1390         page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end,
1391                                rp_param.rp_hash64);
1392         if (IS_ERR(page)) {
1393                 CERROR("%s: dir page locate: "DFID" at %llu: rc %ld\n",
1394                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1395                        rp_param.rp_off, PTR_ERR(page));
1396                 GOTO(out_unlock, rc = PTR_ERR(page));
1397         } else if (page != NULL) {
1398                 /*
1399                  * XXX nikita: not entirely correct handling of a corner case:
1400                  * suppose hash chain of entries with hash value HASH crosses
1401                  * border between pages P0 and P1. First both P0 and P1 are
1402                  * cached, seekdir() is called for some entry from the P0 part
1403                  * of the chain. Later P0 goes out of cache. telldir(HASH)
1404                  * happens and finds P1, as it starts with matching hash
1405                  * value. Remaining entries from P0 part of the chain are
1406                  * skipped. (Is that really a bug?)
1407                  *
1408                  * Possible solutions: 0. don't cache P1 is such case, handle
1409                  * it as an "overflow" page. 1. invalidate all pages at
1410                  * once. 2. use HASH|1 as an index for P1.
1411                  */
1412                 GOTO(hash_collision, page);
1413         }
1414
1415         rp_param.rp_exp = exp;
1416         rp_param.rp_mod = op_data;
1417         page = read_cache_page(mapping,
1418                                hash_x_index(rp_param.rp_off,
1419                                             rp_param.rp_hash64),
1420                                mdc_read_page_remote, &rp_param);
1421         if (IS_ERR(page)) {
1422                 CDEBUG(D_INFO, "%s: read cache page: "DFID" at %llu: %ld\n",
1423                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1424                        rp_param.rp_off, PTR_ERR(page));
1425                 GOTO(out_unlock, rc = PTR_ERR(page));
1426         }
1427
1428         wait_on_page_locked(page);
1429         (void)kmap(page);
1430         if (!PageUptodate(page)) {
1431                 CERROR("%s: page not updated: "DFID" at %llu: rc %d\n",
1432                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1433                        rp_param.rp_off, -5);
1434                 goto fail;
1435         }
1436         if (!PageChecked(page))
1437                 SetPageChecked(page);
1438         if (PageError(page)) {
1439                 CERROR("%s: page error: "DFID" at %llu: rc %d\n",
1440                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1441                        rp_param.rp_off, -5);
1442                 goto fail;
1443         }
1444
1445 hash_collision:
1446         dp = page_address(page);
1447         if (BITS_PER_LONG == 32 && rp_param.rp_hash64) {
1448                 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1449                 end   = le64_to_cpu(dp->ldp_hash_end) >> 32;
1450                 rp_param.rp_off = hash_offset >> 32;
1451         } else {
1452                 start = le64_to_cpu(dp->ldp_hash_start);
1453                 end   = le64_to_cpu(dp->ldp_hash_end);
1454                 rp_param.rp_off = hash_offset;
1455         }
1456         if (end == start) {
1457                 LASSERT(start == rp_param.rp_off);
1458                 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
1459 #if BITS_PER_LONG == 32
1460                 CWARN("Real page-wide hash collision at [%llu %llu] with "
1461                       "hash %llu\n", le64_to_cpu(dp->ldp_hash_start),
1462                       le64_to_cpu(dp->ldp_hash_end), hash_offset);
1463 #endif
1464
1465                 /*
1466                  * Fetch whole overflow chain...
1467                  *
1468                  * XXX not yet.
1469                  */
1470                 goto fail;
1471         }
1472         *ppage = page;
1473 out_unlock:
1474         ldlm_lock_decref(&lockh, it.it_lock_mode);
1475         return rc;
1476 fail:
1477         kunmap(page);
1478         mdc_release_page(page, 1);
1479         rc = -EIO;
1480         goto out_unlock;
1481 }
1482
1483 static int mdc_statfs(const struct lu_env *env,
1484                       struct obd_export *exp, struct obd_statfs *osfs,
1485                       time64_t max_age, __u32 flags)
1486 {
1487         struct obd_device     *obd = class_exp2obd(exp);
1488         struct ptlrpc_request *req;
1489         struct obd_statfs     *msfs;
1490         struct obd_import     *imp = NULL;
1491         int                    rc;
1492         ENTRY;
1493
1494         /*
1495          * Since the request might also come from lprocfs, so we need
1496          * sync this with client_disconnect_export Bug15684
1497          */
1498         down_read(&obd->u.cli.cl_sem);
1499         if (obd->u.cli.cl_import)
1500                 imp = class_import_get(obd->u.cli.cl_import);
1501         up_read(&obd->u.cli.cl_sem);
1502         if (!imp)
1503                 RETURN(-ENODEV);
1504
1505         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_STATFS,
1506                                         LUSTRE_MDS_VERSION, MDS_STATFS);
1507         if (req == NULL)
1508                 GOTO(output, rc = -ENOMEM);
1509
1510         ptlrpc_request_set_replen(req);
1511
1512         if (flags & OBD_STATFS_NODELAY) {
1513                 /* procfs requests not want stay in wait for avoid deadlock */
1514                 req->rq_no_resend = 1;
1515                 req->rq_no_delay = 1;
1516         }
1517
1518         rc = ptlrpc_queue_wait(req);
1519         if (rc) {
1520                 /* check connection error first */
1521                 if (imp->imp_connect_error)
1522                         rc = imp->imp_connect_error;
1523                 GOTO(out, rc);
1524         }
1525
1526         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1527         if (msfs == NULL)
1528                 GOTO(out, rc = -EPROTO);
1529
1530         *osfs = *msfs;
1531         EXIT;
1532 out:
1533         ptlrpc_req_finished(req);
1534 output:
1535         class_import_put(imp);
1536         return rc;
1537 }
1538
1539 static int mdc_ioc_fid2path(struct obd_export *exp, struct getinfo_fid2path *gf)
1540 {
1541         __u32 keylen, vallen;
1542         void *key;
1543         int rc;
1544
1545         if (gf->gf_pathlen > PATH_MAX)
1546                 RETURN(-ENAMETOOLONG);
1547         if (gf->gf_pathlen < 2)
1548                 RETURN(-EOVERFLOW);
1549
1550         /* Key is KEY_FID2PATH + getinfo_fid2path description */
1551         keylen = cfs_size_round(sizeof(KEY_FID2PATH) + sizeof(*gf) +
1552                                 sizeof(struct lu_fid));
1553         OBD_ALLOC(key, keylen);
1554         if (key == NULL)
1555                 RETURN(-ENOMEM);
1556         memcpy(key, KEY_FID2PATH, sizeof(KEY_FID2PATH));
1557         memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)), gf, sizeof(*gf));
1558         memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*gf),
1559                gf->gf_u.gf_root_fid, sizeof(struct lu_fid));
1560         CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
1561                PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno);
1562
1563         if (!fid_is_sane(&gf->gf_fid))
1564                 GOTO(out, rc = -EINVAL);
1565
1566         /* Val is struct getinfo_fid2path result plus path */
1567         vallen = sizeof(*gf) + gf->gf_pathlen;
1568
1569         rc = obd_get_info(NULL, exp, keylen, key, &vallen, gf);
1570         if (rc != 0 && rc != -EREMOTE)
1571                 GOTO(out, rc);
1572
1573         if (vallen <= sizeof(*gf))
1574                 GOTO(out, rc = -EPROTO);
1575         if (vallen > sizeof(*gf) + gf->gf_pathlen)
1576                 GOTO(out, rc = -EOVERFLOW);
1577
1578         CDEBUG(D_IOCTL, "path got "DFID" from %llu #%d: %s\n",
1579                PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno,
1580                gf->gf_pathlen < 512 ? gf->gf_u.gf_path :
1581                /* only log the last 512 characters of the path */
1582                gf->gf_u.gf_path + gf->gf_pathlen - 512);
1583
1584 out:
1585         OBD_FREE(key, keylen);
1586         return rc;
1587 }
1588
1589 static int mdc_ioc_hsm_progress(struct obd_export *exp,
1590                                 struct hsm_progress_kernel *hpk)
1591 {
1592         struct obd_import               *imp = class_exp2cliimp(exp);
1593         struct hsm_progress_kernel      *req_hpk;
1594         struct ptlrpc_request           *req;
1595         int                              rc;
1596         ENTRY;
1597
1598         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_PROGRESS,
1599                                         LUSTRE_MDS_VERSION, MDS_HSM_PROGRESS);
1600         if (req == NULL)
1601                 GOTO(out, rc = -ENOMEM);
1602
1603         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1604
1605         /* Copy hsm_progress struct */
1606         req_hpk = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_PROGRESS);
1607         if (req_hpk == NULL)
1608                 GOTO(out, rc = -EPROTO);
1609
1610         *req_hpk = *hpk;
1611         req_hpk->hpk_errval = lustre_errno_hton(hpk->hpk_errval);
1612
1613         ptlrpc_request_set_replen(req);
1614
1615         mdc_get_mod_rpc_slot(req, NULL);
1616         rc = ptlrpc_queue_wait(req);
1617         mdc_put_mod_rpc_slot(req, NULL);
1618
1619         GOTO(out, rc);
1620 out:
1621         ptlrpc_req_finished(req);
1622         return rc;
1623 }
1624
1625 static int mdc_ioc_hsm_ct_register(struct obd_import *imp, __u32 archives)
1626 {
1627         __u32                   *archive_mask;
1628         struct ptlrpc_request   *req;
1629         int                      rc;
1630         ENTRY;
1631
1632         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_REGISTER,
1633                                         LUSTRE_MDS_VERSION,
1634                                         MDS_HSM_CT_REGISTER);
1635         if (req == NULL)
1636                 GOTO(out, rc = -ENOMEM);
1637
1638         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1639
1640         /* Copy hsm_progress struct */
1641         archive_mask = req_capsule_client_get(&req->rq_pill,
1642                                               &RMF_MDS_HSM_ARCHIVE);
1643         if (archive_mask == NULL)
1644                 GOTO(out, rc = -EPROTO);
1645
1646         *archive_mask = archives;
1647
1648         ptlrpc_request_set_replen(req);
1649
1650         rc = mdc_queue_wait(req);
1651         GOTO(out, rc);
1652 out:
1653         ptlrpc_req_finished(req);
1654         return rc;
1655 }
1656
1657 static int mdc_ioc_hsm_current_action(struct obd_export *exp,
1658                                       struct md_op_data *op_data)
1659 {
1660         struct hsm_current_action       *hca = op_data->op_data;
1661         struct hsm_current_action       *req_hca;
1662         struct ptlrpc_request           *req;
1663         int                              rc;
1664         ENTRY;
1665
1666         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1667                                    &RQF_MDS_HSM_ACTION);
1668         if (req == NULL)
1669                 RETURN(-ENOMEM);
1670
1671         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_ACTION);
1672         if (rc) {
1673                 ptlrpc_request_free(req);
1674                 RETURN(rc);
1675         }
1676
1677         mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1678                       op_data->op_suppgids[0], 0);
1679
1680         ptlrpc_request_set_replen(req);
1681
1682         rc = mdc_queue_wait(req);
1683         if (rc)
1684                 GOTO(out, rc);
1685
1686         req_hca = req_capsule_server_get(&req->rq_pill,
1687                                          &RMF_MDS_HSM_CURRENT_ACTION);
1688         if (req_hca == NULL)
1689                 GOTO(out, rc = -EPROTO);
1690
1691         *hca = *req_hca;
1692
1693         EXIT;
1694 out:
1695         ptlrpc_req_finished(req);
1696         return rc;
1697 }
1698
1699 static int mdc_ioc_hsm_ct_unregister(struct obd_import *imp)
1700 {
1701         struct ptlrpc_request   *req;
1702         int                      rc;
1703         ENTRY;
1704
1705         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_UNREGISTER,
1706                                         LUSTRE_MDS_VERSION,
1707                                         MDS_HSM_CT_UNREGISTER);
1708         if (req == NULL)
1709                 GOTO(out, rc = -ENOMEM);
1710
1711         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1712
1713         ptlrpc_request_set_replen(req);
1714
1715         rc = mdc_queue_wait(req);
1716         GOTO(out, rc);
1717 out:
1718         ptlrpc_req_finished(req);
1719         return rc;
1720 }
1721
1722 static int mdc_ioc_hsm_state_get(struct obd_export *exp,
1723                                  struct md_op_data *op_data)
1724 {
1725         struct hsm_user_state   *hus = op_data->op_data;
1726         struct hsm_user_state   *req_hus;
1727         struct ptlrpc_request   *req;
1728         int                      rc;
1729         ENTRY;
1730
1731         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1732                                    &RQF_MDS_HSM_STATE_GET);
1733         if (req == NULL)
1734                 RETURN(-ENOMEM);
1735
1736         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_GET);
1737         if (rc != 0) {
1738                 ptlrpc_request_free(req);
1739                 RETURN(rc);
1740         }
1741
1742         mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1743                       op_data->op_suppgids[0], 0);
1744
1745         ptlrpc_request_set_replen(req);
1746
1747         rc = mdc_queue_wait(req);
1748         if (rc)
1749                 GOTO(out, rc);
1750
1751         req_hus = req_capsule_server_get(&req->rq_pill, &RMF_HSM_USER_STATE);
1752         if (req_hus == NULL)
1753                 GOTO(out, rc = -EPROTO);
1754
1755         *hus = *req_hus;
1756
1757         EXIT;
1758 out:
1759         ptlrpc_req_finished(req);
1760         return rc;
1761 }
1762
1763 static int mdc_ioc_hsm_state_set(struct obd_export *exp,
1764                                  struct md_op_data *op_data)
1765 {
1766         struct hsm_state_set    *hss = op_data->op_data;
1767         struct hsm_state_set    *req_hss;
1768         struct ptlrpc_request   *req;
1769         int                      rc;
1770         ENTRY;
1771
1772         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1773                                    &RQF_MDS_HSM_STATE_SET);
1774         if (req == NULL)
1775                 RETURN(-ENOMEM);
1776
1777         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_SET);
1778         if (rc) {
1779                 ptlrpc_request_free(req);
1780                 RETURN(rc);
1781         }
1782
1783         mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1784                       op_data->op_suppgids[0], 0);
1785
1786         /* Copy states */
1787         req_hss = req_capsule_client_get(&req->rq_pill, &RMF_HSM_STATE_SET);
1788         if (req_hss == NULL)
1789                 GOTO(out, rc = -EPROTO);
1790         *req_hss = *hss;
1791
1792         ptlrpc_request_set_replen(req);
1793
1794         mdc_get_mod_rpc_slot(req, NULL);
1795         rc = ptlrpc_queue_wait(req);
1796         mdc_put_mod_rpc_slot(req, NULL);
1797
1798         GOTO(out, rc);
1799 out:
1800         ptlrpc_req_finished(req);
1801         return rc;
1802 }
1803
1804 static int mdc_ioc_hsm_request(struct obd_export *exp,
1805                                struct hsm_user_request *hur)
1806 {
1807         struct obd_import       *imp = class_exp2cliimp(exp);
1808         struct ptlrpc_request   *req;
1809         struct hsm_request      *req_hr;
1810         struct hsm_user_item    *req_hui;
1811         char                    *req_opaque;
1812         int                      rc;
1813         ENTRY;
1814
1815         req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_REQUEST);
1816         if (req == NULL)
1817                 GOTO(out, rc = -ENOMEM);
1818
1819         req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM, RCL_CLIENT,
1820                              hur->hur_request.hr_itemcount
1821                              * sizeof(struct hsm_user_item));
1822         req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA, RCL_CLIENT,
1823                              hur->hur_request.hr_data_len);
1824
1825         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_REQUEST);
1826         if (rc) {
1827                 ptlrpc_request_free(req);
1828                 RETURN(rc);
1829         }
1830
1831         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1832
1833         /* Copy hsm_request struct */
1834         req_hr = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_REQUEST);
1835         if (req_hr == NULL)
1836                 GOTO(out, rc = -EPROTO);
1837         *req_hr = hur->hur_request;
1838
1839         /* Copy hsm_user_item structs */
1840         req_hui = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM);
1841         if (req_hui == NULL)
1842                 GOTO(out, rc = -EPROTO);
1843         memcpy(req_hui, hur->hur_user_item,
1844                hur->hur_request.hr_itemcount * sizeof(struct hsm_user_item));
1845
1846         /* Copy opaque field */
1847         req_opaque = req_capsule_client_get(&req->rq_pill, &RMF_GENERIC_DATA);
1848         if (req_opaque == NULL)
1849                 GOTO(out, rc = -EPROTO);
1850         memcpy(req_opaque, hur_data(hur), hur->hur_request.hr_data_len);
1851
1852         ptlrpc_request_set_replen(req);
1853
1854         mdc_get_mod_rpc_slot(req, NULL);
1855         rc = ptlrpc_queue_wait(req);
1856         mdc_put_mod_rpc_slot(req, NULL);
1857
1858         GOTO(out, rc);
1859
1860 out:
1861         ptlrpc_req_finished(req);
1862         return rc;
1863 }
1864
1865 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
1866                                 struct lustre_kernelcomm *lk);
1867
1868 static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp,
1869                         struct obd_quotactl *oqctl)
1870 {
1871         struct ptlrpc_request   *req;
1872         struct obd_quotactl     *oqc;
1873         int                      rc;
1874         ENTRY;
1875
1876         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1877                                         &RQF_MDS_QUOTACTL, LUSTRE_MDS_VERSION,
1878                                         MDS_QUOTACTL);
1879         if (req == NULL)
1880                 RETURN(-ENOMEM);
1881
1882         oqc = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1883         *oqc = *oqctl;
1884
1885         ptlrpc_request_set_replen(req);
1886         ptlrpc_at_set_req_timeout(req);
1887
1888         rc = ptlrpc_queue_wait(req);
1889         if (rc)
1890                 CERROR("ptlrpc_queue_wait failed, rc: %d\n", rc);
1891
1892         if (req->rq_repmsg &&
1893             (oqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL))) {
1894                 *oqctl = *oqc;
1895         } else if (!rc) {
1896                 CERROR ("Can't unpack obd_quotactl\n");
1897                 rc = -EPROTO;
1898         }
1899         ptlrpc_req_finished(req);
1900
1901         RETURN(rc);
1902 }
1903
1904 static int mdc_ioc_swap_layouts(struct obd_export *exp,
1905                                 struct md_op_data *op_data)
1906 {
1907         struct list_head cancels = LIST_HEAD_INIT(cancels);
1908         struct ptlrpc_request   *req;
1909         int                      rc, count;
1910         struct mdc_swap_layouts *msl, *payload;
1911         ENTRY;
1912
1913         msl = op_data->op_data;
1914
1915         /* When the MDT will get the MDS_SWAP_LAYOUTS RPC the
1916          * first thing it will do is to cancel the 2 layout
1917          * locks held by this client.
1918          * So the client must cancel its layout locks on the 2 fids
1919          * with the request RPC to avoid extra RPC round trips.
1920          */
1921         count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels,
1922                                         LCK_EX, MDS_INODELOCK_LAYOUT |
1923                                         MDS_INODELOCK_XATTR);
1924         count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels,
1925                                          LCK_EX, MDS_INODELOCK_LAYOUT |
1926                                          MDS_INODELOCK_XATTR);
1927
1928         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1929                                    &RQF_MDS_SWAP_LAYOUTS);
1930         if (req == NULL) {
1931                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
1932                 RETURN(-ENOMEM);
1933         }
1934
1935         rc = mdc_prep_elc_req(exp, req, MDS_SWAP_LAYOUTS, &cancels, count);
1936         if (rc) {
1937                 ptlrpc_request_free(req);
1938                 RETURN(rc);
1939         }
1940
1941         mdc_swap_layouts_pack(req, op_data);
1942
1943         payload = req_capsule_client_get(&req->rq_pill, &RMF_SWAP_LAYOUTS);
1944         LASSERT(payload);
1945
1946         *payload = *msl;
1947
1948         ptlrpc_request_set_replen(req);
1949
1950         rc = ptlrpc_queue_wait(req);
1951         if (rc)
1952                 GOTO(out, rc);
1953         EXIT;
1954
1955 out:
1956         ptlrpc_req_finished(req);
1957         return rc;
1958 }
1959
1960 static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1961                          void *karg, void __user *uarg)
1962 {
1963         struct obd_device *obd = exp->exp_obd;
1964         struct obd_ioctl_data *data = karg;
1965         struct obd_import *imp = obd->u.cli.cl_import;
1966         int rc;
1967         ENTRY;
1968
1969         if (!try_module_get(THIS_MODULE)) {
1970                 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
1971                        module_name(THIS_MODULE));
1972                 return -EINVAL;
1973         }
1974         switch (cmd) {
1975         case OBD_IOC_FID2PATH:
1976                 rc = mdc_ioc_fid2path(exp, karg);
1977                 GOTO(out, rc);
1978         case LL_IOC_HSM_CT_START:
1979                 rc = mdc_ioc_hsm_ct_start(exp, karg);
1980                 /* ignore if it was already registered on this MDS. */
1981                 if (rc == -EEXIST)
1982                         rc = 0;
1983                 GOTO(out, rc);
1984         case LL_IOC_HSM_PROGRESS:
1985                 rc = mdc_ioc_hsm_progress(exp, karg);
1986                 GOTO(out, rc);
1987         case LL_IOC_HSM_STATE_GET:
1988                 rc = mdc_ioc_hsm_state_get(exp, karg);
1989                 GOTO(out, rc);
1990         case LL_IOC_HSM_STATE_SET:
1991                 rc = mdc_ioc_hsm_state_set(exp, karg);
1992                 GOTO(out, rc);
1993         case LL_IOC_HSM_ACTION:
1994                 rc = mdc_ioc_hsm_current_action(exp, karg);
1995                 GOTO(out, rc);
1996         case LL_IOC_HSM_REQUEST:
1997                 rc = mdc_ioc_hsm_request(exp, karg);
1998                 GOTO(out, rc);
1999         case OBD_IOC_CLIENT_RECOVER:
2000                 rc = ptlrpc_recover_import(imp, data->ioc_inlbuf1, 0);
2001                 if (rc < 0)
2002                         GOTO(out, rc);
2003                 GOTO(out, rc = 0);
2004         case IOC_OSC_SET_ACTIVE:
2005                 rc = ptlrpc_set_import_active(imp, data->ioc_offset);
2006                 GOTO(out, rc);
2007         case OBD_IOC_PING_TARGET:
2008                 rc = ptlrpc_obd_ping(obd);
2009                 GOTO(out, rc);
2010         /*
2011          * Normally IOC_OBD_STATFS, OBD_IOC_QUOTACTL iocontrol are handled by
2012          * LMV instead of MDC. But when the cluster is upgraded from 1.8,
2013          * there'd be no LMV layer thus we might be called here. Eventually
2014          * this code should be removed.
2015          * bz20731, LU-592.
2016          */
2017         case IOC_OBD_STATFS: {
2018                 struct obd_statfs stat_buf = {0};
2019
2020                 if (*((__u32 *) data->ioc_inlbuf2) != 0)
2021                         GOTO(out, rc = -ENODEV);
2022
2023                 /* copy UUID */
2024                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(obd),
2025                                  min((int)data->ioc_plen2,
2026                                      (int)sizeof(struct obd_uuid))))
2027                         GOTO(out, rc = -EFAULT);
2028
2029                 rc = mdc_statfs(NULL, obd->obd_self_export, &stat_buf,
2030                                 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
2031                                 0);
2032                 if (rc != 0)
2033                         GOTO(out, rc);
2034
2035                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
2036                                      min((int) data->ioc_plen1,
2037                                          (int) sizeof(stat_buf))))
2038                         GOTO(out, rc = -EFAULT);
2039
2040                 GOTO(out, rc = 0);
2041         }
2042         case OBD_IOC_QUOTACTL: {
2043                 struct if_quotactl *qctl = karg;
2044                 struct obd_quotactl *oqctl;
2045
2046                 OBD_ALLOC_PTR(oqctl);
2047                 if (oqctl == NULL)
2048                         GOTO(out, rc = -ENOMEM);
2049
2050                 QCTL_COPY(oqctl, qctl);
2051                 rc = obd_quotactl(exp, oqctl);
2052                 if (rc == 0) {
2053                         QCTL_COPY(qctl, oqctl);
2054                         qctl->qc_valid = QC_MDTIDX;
2055                         qctl->obd_uuid = obd->u.cli.cl_target_uuid;
2056                 }
2057
2058                 OBD_FREE_PTR(oqctl);
2059                 GOTO(out, rc);
2060         }
2061         case LL_IOC_GET_CONNECT_FLAGS:
2062                 if (copy_to_user(uarg, exp_connect_flags_ptr(exp),
2063                                  sizeof(*exp_connect_flags_ptr(exp))))
2064                         GOTO(out, rc = -EFAULT);
2065
2066                 GOTO(out, rc = 0);
2067         case LL_IOC_LOV_SWAP_LAYOUTS:
2068                 rc = mdc_ioc_swap_layouts(exp, karg);
2069                 GOTO(out, rc);
2070         default:
2071                 CERROR("unrecognised ioctl: cmd = %#x\n", cmd);
2072                 GOTO(out, rc = -ENOTTY);
2073         }
2074 out:
2075         module_put(THIS_MODULE);
2076
2077         return rc;
2078 }
2079
2080 static int mdc_get_info_rpc(struct obd_export *exp,
2081                             u32 keylen, void *key,
2082                             u32 vallen, void *val)
2083 {
2084         struct obd_import      *imp = class_exp2cliimp(exp);
2085         struct ptlrpc_request  *req;
2086         char                   *tmp;
2087         int                     rc = -EINVAL;
2088         ENTRY;
2089
2090         req = ptlrpc_request_alloc(imp, &RQF_MDS_GET_INFO);
2091         if (req == NULL)
2092                 RETURN(-ENOMEM);
2093
2094         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY,
2095                              RCL_CLIENT, keylen);
2096         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VALLEN,
2097                              RCL_CLIENT, sizeof(vallen));
2098
2099         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_INFO);
2100         if (rc) {
2101                 ptlrpc_request_free(req);
2102                 RETURN(rc);
2103         }
2104
2105         tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
2106         memcpy(tmp, key, keylen);
2107         tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_VALLEN);
2108         memcpy(tmp, &vallen, sizeof(vallen));
2109
2110         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VAL,
2111                              RCL_SERVER, vallen);
2112         ptlrpc_request_set_replen(req);
2113
2114         rc = ptlrpc_queue_wait(req);
2115         /* -EREMOTE means the get_info result is partial, and it needs to
2116          * continue on another MDT, see fid2path part in lmv_iocontrol */
2117         if (rc == 0 || rc == -EREMOTE) {
2118                 tmp = req_capsule_server_get(&req->rq_pill, &RMF_GETINFO_VAL);
2119                 memcpy(val, tmp, vallen);
2120                 if (ptlrpc_rep_need_swab(req)) {
2121                         if (KEY_IS(KEY_FID2PATH))
2122                                 lustre_swab_fid2path(val);
2123                 }
2124         }
2125         ptlrpc_req_finished(req);
2126
2127         RETURN(rc);
2128 }
2129
2130 static void lustre_swab_hai(struct hsm_action_item *h)
2131 {
2132         __swab32s(&h->hai_len);
2133         __swab32s(&h->hai_action);
2134         lustre_swab_lu_fid(&h->hai_fid);
2135         lustre_swab_lu_fid(&h->hai_dfid);
2136         __swab64s(&h->hai_cookie);
2137         __swab64s(&h->hai_extent.offset);
2138         __swab64s(&h->hai_extent.length);
2139         __swab64s(&h->hai_gid);
2140 }
2141
2142 static void lustre_swab_hal(struct hsm_action_list *h)
2143 {
2144         struct hsm_action_item  *hai;
2145         __u32                    i;
2146
2147         __swab32s(&h->hal_version);
2148         __swab32s(&h->hal_count);
2149         __swab32s(&h->hal_archive_id);
2150         __swab64s(&h->hal_flags);
2151         hai = hai_first(h);
2152         for (i = 0; i < h->hal_count; i++, hai = hai_next(hai))
2153                 lustre_swab_hai(hai);
2154 }
2155
2156 static void lustre_swab_kuch(struct kuc_hdr *l)
2157 {
2158         __swab16s(&l->kuc_magic);
2159         /* __u8 l->kuc_transport */
2160         __swab16s(&l->kuc_msgtype);
2161         __swab16s(&l->kuc_msglen);
2162 }
2163
2164 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2165                                 struct lustre_kernelcomm *lk)
2166 {
2167         struct obd_import  *imp = class_exp2cliimp(exp);
2168         __u32               archive = lk->lk_data;
2169         int                 rc = 0;
2170
2171         if (lk->lk_group != KUC_GRP_HSM) {
2172                 CERROR("Bad copytool group %d\n", lk->lk_group);
2173                 return -EINVAL;
2174         }
2175
2176         CDEBUG(D_HSM, "CT start r%d w%d u%d g%d f%#x\n", lk->lk_rfd, lk->lk_wfd,
2177                lk->lk_uid, lk->lk_group, lk->lk_flags);
2178
2179         if (lk->lk_flags & LK_FLG_STOP) {
2180                 /* Unregister with the coordinator */
2181                 rc = mdc_ioc_hsm_ct_unregister(imp);
2182         } else {
2183                 rc = mdc_ioc_hsm_ct_register(imp, archive);
2184         }
2185
2186         return rc;
2187 }
2188
2189 /**
2190  * Send a message to any listening copytools
2191  * @param val KUC message (kuc_hdr + hsm_action_list)
2192  * @param len total length of message
2193  */
2194 static int mdc_hsm_copytool_send(const struct obd_uuid *uuid,
2195                                  size_t len, void *val)
2196 {
2197         struct kuc_hdr          *lh = (struct kuc_hdr *)val;
2198         struct hsm_action_list  *hal = (struct hsm_action_list *)(lh + 1);
2199         int                      rc;
2200         ENTRY;
2201
2202         if (len < sizeof(*lh) + sizeof(*hal)) {
2203                 CERROR("Short HSM message %zu < %zu\n", len,
2204                        sizeof(*lh) + sizeof(*hal));
2205                 RETURN(-EPROTO);
2206         }
2207         if (lh->kuc_magic == __swab16(KUC_MAGIC)) {
2208                 lustre_swab_kuch(lh);
2209                 lustre_swab_hal(hal);
2210         } else if (lh->kuc_magic != KUC_MAGIC) {
2211                 CERROR("Bad magic %x!=%x\n", lh->kuc_magic, KUC_MAGIC);
2212                 RETURN(-EPROTO);
2213         }
2214
2215         CDEBUG(D_HSM, " Received message mg=%x t=%d m=%d l=%d actions=%d "
2216                "on %s\n",
2217                lh->kuc_magic, lh->kuc_transport, lh->kuc_msgtype,
2218                lh->kuc_msglen, hal->hal_count, hal->hal_fsname);
2219
2220         /* Broadcast to HSM listeners */
2221         rc = libcfs_kkuc_group_put(uuid, KUC_GRP_HSM, lh);
2222
2223         RETURN(rc);
2224 }
2225
2226 /**
2227  * callback function passed to kuc for re-registering each HSM copytool
2228  * running on MDC, after MDT shutdown/recovery.
2229  * @param data copytool registration data
2230  * @param cb_arg callback argument (obd_import)
2231  */
2232 static int mdc_hsm_ct_reregister(void *data, void *cb_arg)
2233 {
2234         struct kkuc_ct_data     *kcd = data;
2235         struct obd_import       *imp = (struct obd_import *)cb_arg;
2236         int                      rc;
2237
2238         if (kcd == NULL || kcd->kcd_magic != KKUC_CT_DATA_MAGIC)
2239                 return -EPROTO;
2240
2241         CDEBUG(D_HA, "%s: recover copytool registration to MDT (archive=%#x)\n",
2242                imp->imp_obd->obd_name, kcd->kcd_archive);
2243         rc = mdc_ioc_hsm_ct_register(imp, kcd->kcd_archive);
2244
2245         /* ignore error if the copytool is already registered */
2246         return (rc == -EEXIST) ? 0 : rc;
2247 }
2248
2249 /**
2250  * Re-establish all kuc contexts with MDT
2251  * after MDT shutdown/recovery.
2252  */
2253 static int mdc_kuc_reregister(struct obd_import *imp)
2254 {
2255         /* re-register HSM agents */
2256         return libcfs_kkuc_group_foreach(&imp->imp_obd->obd_uuid, KUC_GRP_HSM,
2257                                          mdc_hsm_ct_reregister, imp);
2258 }
2259
2260 static int mdc_set_info_async(const struct lu_env *env,
2261                               struct obd_export *exp,
2262                               u32 keylen, void *key,
2263                               u32 vallen, void *val,
2264                               struct ptlrpc_request_set *set)
2265 {
2266         struct obd_import       *imp = class_exp2cliimp(exp);
2267         int                      rc;
2268         ENTRY;
2269
2270         if (KEY_IS(KEY_READ_ONLY)) {
2271                 if (vallen != sizeof(int))
2272                         RETURN(-EINVAL);
2273
2274                 spin_lock(&imp->imp_lock);
2275                 if (*((int *)val)) {
2276                         imp->imp_connect_flags_orig |= OBD_CONNECT_RDONLY;
2277                         imp->imp_connect_data.ocd_connect_flags |=
2278                                                         OBD_CONNECT_RDONLY;
2279                 } else {
2280                         imp->imp_connect_flags_orig &= ~OBD_CONNECT_RDONLY;
2281                         imp->imp_connect_data.ocd_connect_flags &=
2282                                                         ~OBD_CONNECT_RDONLY;
2283                 }
2284                 spin_unlock(&imp->imp_lock);
2285
2286                 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2287                                        keylen, key, vallen, val, set);
2288                 RETURN(rc);
2289         }
2290         if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2291                 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2292                                        keylen, key, vallen, val, set);
2293                 RETURN(rc);
2294         }
2295         if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) {
2296                 rc = mdc_hsm_copytool_send(&imp->imp_obd->obd_uuid, vallen,
2297                                            val);
2298                 RETURN(rc);
2299         }
2300
2301         if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2302                 __u32 *default_easize = val;
2303
2304                 exp->exp_obd->u.cli.cl_default_mds_easize = *default_easize;
2305                 RETURN(0);
2306         }
2307
2308         rc = osc_set_info_async(env, exp, keylen, key, vallen, val, set);
2309         RETURN(rc);
2310 }
2311
2312 static int mdc_get_info(const struct lu_env *env, struct obd_export *exp,
2313                         __u32 keylen, void *key, __u32 *vallen, void *val)
2314 {
2315         int rc = -EINVAL;
2316
2317         if (KEY_IS(KEY_MAX_EASIZE)) {
2318                 __u32 mdsize, *max_easize;
2319
2320                 if (*vallen != sizeof(int))
2321                         RETURN(-EINVAL);
2322                 mdsize = *(__u32 *)val;
2323                 if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
2324                         exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
2325                 max_easize = val;
2326                 *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
2327                 RETURN(0);
2328         } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2329                 __u32 *default_easize;
2330
2331                 if (*vallen != sizeof(int))
2332                         RETURN(-EINVAL);
2333                 default_easize = val;
2334                 *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
2335                 RETURN(0);
2336         } else if (KEY_IS(KEY_CONN_DATA)) {
2337                 struct obd_import *imp = class_exp2cliimp(exp);
2338                 struct obd_connect_data *data = val;
2339
2340                 if (*vallen != sizeof(*data))
2341                         RETURN(-EINVAL);
2342
2343                 *data = imp->imp_connect_data;
2344                 RETURN(0);
2345         } else if (KEY_IS(KEY_TGT_COUNT)) {
2346                 *((__u32 *)val) = 1;
2347                 RETURN(0);
2348         }
2349
2350         rc = mdc_get_info_rpc(exp, keylen, key, *vallen, val);
2351
2352         RETURN(rc);
2353 }
2354
2355 static int mdc_fsync(struct obd_export *exp, const struct lu_fid *fid,
2356                      struct ptlrpc_request **request)
2357 {
2358         struct ptlrpc_request *req;
2359         int                    rc;
2360         ENTRY;
2361
2362         *request = NULL;
2363         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SYNC);
2364         if (req == NULL)
2365                 RETURN(-ENOMEM);
2366
2367         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SYNC);
2368         if (rc) {
2369                 ptlrpc_request_free(req);
2370                 RETURN(rc);
2371         }
2372
2373         mdc_pack_body(req, fid, 0, 0, -1, 0);
2374
2375         ptlrpc_request_set_replen(req);
2376
2377         rc = ptlrpc_queue_wait(req);
2378         if (rc)
2379                 ptlrpc_req_finished(req);
2380         else
2381                 *request = req;
2382         RETURN(rc);
2383 }
2384
2385 static int mdc_import_event(struct obd_device *obd, struct obd_import *imp,
2386                             enum obd_import_event event)
2387 {
2388         struct client_obd *cli = &obd->u.cli;
2389         int rc = 0;
2390
2391         LASSERT(imp->imp_obd == obd);
2392
2393         switch (event) {
2394         case IMP_EVENT_DISCON:
2395                 spin_lock(&cli->cl_loi_list_lock);
2396                 cli->cl_avail_grant = 0;
2397                 cli->cl_lost_grant = 0;
2398                 spin_unlock(&cli->cl_loi_list_lock);
2399                 break;
2400         case IMP_EVENT_INACTIVE:
2401                 /*
2402                  * Flush current sequence to make client obtain new one
2403                  * from server in case of disconnect/reconnect.
2404                  */
2405                 down_read(&cli->cl_seq_rwsem);
2406                 if (cli->cl_seq)
2407                         seq_client_flush(cli->cl_seq);
2408                 up_read(&cli->cl_seq_rwsem);
2409
2410                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
2411                 break;
2412         case IMP_EVENT_INVALIDATE: {
2413                 struct ldlm_namespace *ns = obd->obd_namespace;
2414                 struct lu_env *env;
2415                 __u16 refcheck;
2416
2417                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2418
2419                 env = cl_env_get(&refcheck);
2420                 if (!IS_ERR(env)) {
2421                         /* Reset grants. All pages go to failing rpcs due to
2422                          * the invalid import.
2423                          */
2424                         osc_io_unplug(env, cli, NULL);
2425
2426                         cfs_hash_for_each_nolock(ns->ns_rs_hash,
2427                                                  osc_ldlm_resource_invalidate,
2428                                                  env, 0);
2429                         cl_env_put(env, &refcheck);
2430                         ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2431                 } else {
2432                         rc = PTR_ERR(env);
2433                 }
2434                 break;
2435         }
2436         case IMP_EVENT_ACTIVE:
2437                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
2438                 /* redo the kuc registration after reconnecting */
2439                 if (rc == 0)
2440                         rc = mdc_kuc_reregister(imp);
2441                 break;
2442         case IMP_EVENT_OCD: {
2443                 struct obd_connect_data *ocd = &imp->imp_connect_data;
2444
2445                 if (OCD_HAS_FLAG(ocd, GRANT))
2446                         osc_init_grant(cli, ocd);
2447
2448                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
2449                 break;
2450         }
2451         case IMP_EVENT_DEACTIVATE:
2452         case IMP_EVENT_ACTIVATE:
2453                 break;
2454         default:
2455                 CERROR("Unknown import event %x\n", event);
2456                 LBUG();
2457         }
2458         RETURN(rc);
2459 }
2460
2461 int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp,
2462                   struct lu_fid *fid, struct md_op_data *op_data)
2463 {
2464         struct client_obd *cli = &exp->exp_obd->u.cli;
2465         int rc = -EIO;
2466
2467         ENTRY;
2468
2469         down_read(&cli->cl_seq_rwsem);
2470         if (cli->cl_seq)
2471                 rc = seq_client_alloc_fid(env, cli->cl_seq, fid);
2472         up_read(&cli->cl_seq_rwsem);
2473
2474         RETURN(rc);
2475 }
2476
2477 static struct obd_uuid *mdc_get_uuid(struct obd_export *exp)
2478 {
2479         struct client_obd *cli = &exp->exp_obd->u.cli;
2480         return &cli->cl_target_uuid;
2481 }
2482
2483 /**
2484  * Determine whether the lock can be canceled before replaying it during
2485  * recovery, non zero value will be return if the lock can be canceled,
2486  * or zero returned for not
2487  */
2488 static int mdc_cancel_weight(struct ldlm_lock *lock)
2489 {
2490         if (lock->l_resource->lr_type != LDLM_IBITS)
2491                 RETURN(0);
2492
2493         /* FIXME: if we ever get into a situation where there are too many
2494          * opened files with open locks on a single node, then we really
2495          * should replay these open locks to reget it */
2496         if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN)
2497                 RETURN(0);
2498
2499         RETURN(1);
2500 }
2501
2502 static int mdc_resource_inode_free(struct ldlm_resource *res)
2503 {
2504         if (res->lr_lvb_inode)
2505                 res->lr_lvb_inode = NULL;
2506
2507         return 0;
2508 }
2509
2510 static struct ldlm_valblock_ops inode_lvbo = {
2511         .lvbo_free = mdc_resource_inode_free
2512 };
2513
2514 static int mdc_llog_init(struct obd_device *obd)
2515 {
2516         struct obd_llog_group   *olg = &obd->obd_olg;
2517         struct llog_ctxt        *ctxt;
2518         int                      rc;
2519
2520         ENTRY;
2521
2522         rc = llog_setup(NULL, obd, olg, LLOG_CHANGELOG_REPL_CTXT, obd,
2523                         &llog_client_ops);
2524         if (rc < 0)
2525                 RETURN(rc);
2526
2527         ctxt = llog_group_get_ctxt(olg, LLOG_CHANGELOG_REPL_CTXT);
2528         llog_initiator_connect(ctxt);
2529         llog_ctxt_put(ctxt);
2530
2531         RETURN(0);
2532 }
2533
2534 static void mdc_llog_finish(struct obd_device *obd)
2535 {
2536         struct llog_ctxt *ctxt;
2537
2538         ENTRY;
2539
2540         ctxt = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
2541         if (ctxt != NULL)
2542                 llog_cleanup(NULL, ctxt);
2543
2544         EXIT;
2545 }
2546
2547 int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg)
2548 {
2549         int rc;
2550
2551         ENTRY;
2552
2553         rc = osc_setup_common(obd, cfg);
2554         if (rc < 0)
2555                 RETURN(rc);
2556
2557         rc = mdc_tunables_init(obd);
2558         if (rc)
2559                 GOTO(err_osc_cleanup, rc);
2560
2561         ns_register_cancel(obd->obd_namespace, mdc_cancel_weight);
2562
2563         obd->obd_namespace->ns_lvbo = &inode_lvbo;
2564
2565         rc = mdc_llog_init(obd);
2566         if (rc) {
2567                 CERROR("%s: failed to setup llogging subsystems: rc = %d\n",
2568                        obd->obd_name, rc);
2569                 GOTO(err_llog_cleanup, rc);
2570         }
2571
2572         rc = mdc_changelog_cdev_init(obd);
2573         if (rc) {
2574                 CERROR("%s: failed to setup changelog char device: rc = %d\n",
2575                        obd->obd_name, rc);
2576                 GOTO(err_changelog_cleanup, rc);
2577         }
2578
2579         RETURN(rc);
2580
2581 err_changelog_cleanup:
2582         mdc_llog_finish(obd);
2583 err_llog_cleanup:
2584         lprocfs_free_md_stats(obd);
2585         ptlrpc_lprocfs_unregister_obd(obd);
2586 err_osc_cleanup:
2587         osc_cleanup_common(obd);
2588         return rc;
2589 }
2590
2591 /* Initialize the default and maximum LOV EA sizes.  This allows
2592  * us to make MDS RPCs with large enough reply buffers to hold a default
2593  * sized EA without having to calculate this (via a call into the
2594  * LOV + OSCs) each time we make an RPC.  The maximum size is also tracked
2595  * but not used to avoid wastefully vmalloc()'ing large reply buffers when
2596  * a large number of stripes is possible.  If a larger reply buffer is
2597  * required it will be reallocated in the ptlrpc layer due to overflow.
2598  */
2599 static int mdc_init_ea_size(struct obd_export *exp, __u32 easize,
2600                             __u32 def_easize)
2601 {
2602         struct obd_device *obd = exp->exp_obd;
2603         struct client_obd *cli = &obd->u.cli;
2604         ENTRY;
2605
2606         if (cli->cl_max_mds_easize < easize)
2607                 cli->cl_max_mds_easize = easize;
2608
2609         if (cli->cl_default_mds_easize < def_easize)
2610                 cli->cl_default_mds_easize = def_easize;
2611
2612         RETURN(0);
2613 }
2614
2615 static int mdc_precleanup(struct obd_device *obd)
2616 {
2617         ENTRY;
2618
2619         osc_precleanup_common(obd);
2620         mdc_changelog_cdev_finish(obd);
2621
2622         obd_cleanup_client_import(obd);
2623         ptlrpc_lprocfs_unregister_obd(obd);
2624         lprocfs_free_md_stats(obd);
2625         mdc_llog_finish(obd);
2626         RETURN(0);
2627 }
2628
2629 static int mdc_cleanup(struct obd_device *obd)
2630 {
2631         return osc_cleanup_common(obd);
2632 }
2633
2634 int mdc_process_config(struct obd_device *obd, size_t len, void *buf)
2635 {
2636         struct lustre_cfg *lcfg = buf;
2637         size_t count  = class_modify_config(lcfg, PARAM_MDC,
2638                                             &obd->obd_kset.kobj);
2639
2640         return count > 0 ? 0 : count;
2641 }
2642
2643 static struct obd_ops mdc_obd_ops = {
2644         .o_owner            = THIS_MODULE,
2645         .o_setup            = mdc_setup,
2646         .o_precleanup       = mdc_precleanup,
2647         .o_cleanup          = mdc_cleanup,
2648         .o_add_conn         = client_import_add_conn,
2649         .o_del_conn         = client_import_del_conn,
2650         .o_connect          = client_connect_import,
2651         .o_reconnect        = osc_reconnect,
2652         .o_disconnect       = osc_disconnect,
2653         .o_iocontrol        = mdc_iocontrol,
2654         .o_set_info_async   = mdc_set_info_async,
2655         .o_statfs           = mdc_statfs,
2656         .o_fid_init         = client_fid_init,
2657         .o_fid_fini         = client_fid_fini,
2658         .o_fid_alloc        = mdc_fid_alloc,
2659         .o_import_event     = mdc_import_event,
2660         .o_get_info         = mdc_get_info,
2661         .o_process_config   = mdc_process_config,
2662         .o_get_uuid         = mdc_get_uuid,
2663         .o_quotactl         = mdc_quotactl,
2664 };
2665
2666 static struct md_ops mdc_md_ops = {
2667         .m_get_root         = mdc_get_root,
2668         .m_null_inode       = mdc_null_inode,
2669         .m_close            = mdc_close,
2670         .m_create           = mdc_create,
2671         .m_enqueue          = mdc_enqueue,
2672         .m_getattr          = mdc_getattr,
2673         .m_getattr_name     = mdc_getattr_name,
2674         .m_intent_lock      = mdc_intent_lock,
2675         .m_link             = mdc_link,
2676         .m_rename           = mdc_rename,
2677         .m_setattr          = mdc_setattr,
2678         .m_setxattr         = mdc_setxattr,
2679         .m_getxattr         = mdc_getxattr,
2680         .m_fsync                = mdc_fsync,
2681         .m_file_resync          = mdc_file_resync,
2682         .m_read_page            = mdc_read_page,
2683         .m_unlink           = mdc_unlink,
2684         .m_cancel_unused    = mdc_cancel_unused,
2685         .m_init_ea_size     = mdc_init_ea_size,
2686         .m_set_lock_data    = mdc_set_lock_data,
2687         .m_lock_match       = mdc_lock_match,
2688         .m_get_lustre_md    = mdc_get_lustre_md,
2689         .m_free_lustre_md   = mdc_free_lustre_md,
2690         .m_set_open_replay_data = mdc_set_open_replay_data,
2691         .m_clear_open_replay_data = mdc_clear_open_replay_data,
2692         .m_intent_getattr_async = mdc_intent_getattr_async,
2693         .m_revalidate_lock      = mdc_revalidate_lock
2694 };
2695
2696 static int __init mdc_init(void)
2697 {
2698         return class_register_type(&mdc_obd_ops, &mdc_md_ops, true, NULL,
2699                                    LUSTRE_MDC_NAME, &mdc_device_type);
2700 }
2701
2702 static void __exit mdc_exit(void)
2703 {
2704         class_unregister_type(LUSTRE_MDC_NAME);
2705 }
2706
2707 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2708 MODULE_DESCRIPTION("Lustre Metadata Client");
2709 MODULE_VERSION(LUSTRE_VERSION_STRING);
2710 MODULE_LICENSE("GPL");
2711
2712 module_init(mdc_init);
2713 module_exit(mdc_exit);