lustre/mdc/mdc_request.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2016, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 #define DEBUG_SUBSYSTEM S_MDC
  34
  35 #include <linux/init.h>
  36 #include <linux/kthread.h>
  37 #include <linux/miscdevice.h>
  38 #include <linux/module.h>
  39 #include <linux/pagemap.h>
  40 #include <linux/user_namespace.h>
  41 #include <linux/utsname.h>
  42 #ifdef HAVE_UIDGID_HEADER
  43 # include <linux/uidgid.h>
  44 #endif
  45
  46 #include <lustre/lustre_errno.h>
  47
  48 #include <cl_object.h>
  49 #include <llog_swab.h>
  50 #include <lprocfs_status.h>
  51 #include <lustre_acl.h>
  52 #include <lustre_fid.h>
  53 #include <lustre_ioctl.h>
  54 #include <lustre_kernelcomm.h>
  55 #include <lustre_lmv.h>
  56 #include <lustre_log.h>
  57 #include <lustre_param.h>
  58 #include <lustre_swab.h>
  59 #include <obd_class.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 int mdc_getattr(struct obd_export *exp, struct md_op_data *op_data,
 196                        struct ptlrpc_request **request)
 197 {
 198         struct ptlrpc_request *req;
 199         int                    rc;
 200         ENTRY;
 201
 202         /* Single MDS without an LMV case */
 203         if (op_data->op_flags & MF_GET_MDT_IDX) {
 204                 op_data->op_mds = 0;
 205                 RETURN(0);
 206         }
 207         *request = NULL;
 208         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_GETATTR);
 209         if (req == NULL)
 210                 RETURN(-ENOMEM);
 211
 212         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR);
 213         if (rc) {
 214                 ptlrpc_request_free(req);
 215                 RETURN(rc);
 216         }
 217
 218         mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
 219                       op_data->op_mode, -1, 0);
 220
 221         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 222                              op_data->op_mode);
 223         ptlrpc_request_set_replen(req);
 224
 225         rc = mdc_getattr_common(exp, req);
 226         if (rc)
 227                 ptlrpc_req_finished(req);
 228         else
 229                 *request = req;
 230         RETURN(rc);
 231 }
 232
 233 static int mdc_getattr_name(struct obd_export *exp, struct md_op_data *op_data,
 234                             struct ptlrpc_request **request)
 235 {
 236         struct ptlrpc_request *req;
 237         int                    rc;
 238         ENTRY;
 239
 240         *request = NULL;
 241         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 242                                    &RQF_MDS_GETATTR_NAME);
 243         if (req == NULL)
 244                 RETURN(-ENOMEM);
 245
 246         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 247                              op_data->op_namelen + 1);
 248
 249         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR_NAME);
 250         if (rc) {
 251                 ptlrpc_request_free(req);
 252                 RETURN(rc);
 253         }
 254
 255         mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
 256                       op_data->op_mode, op_data->op_suppgids[0], 0);
 257
 258         if (op_data->op_name) {
 259                 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
 260                 LASSERT(strnlen(op_data->op_name, op_data->op_namelen) ==
 261                                 op_data->op_namelen);
 262                 memcpy(name, op_data->op_name, op_data->op_namelen);
 263         }
 264
 265         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 266                              op_data->op_mode);
 267         ptlrpc_request_set_replen(req);
 268
 269         rc = mdc_getattr_common(exp, req);
 270         if (rc)
 271                 ptlrpc_req_finished(req);
 272         else
 273                 *request = req;
 274         RETURN(rc);
 275 }
 276
 277 static int mdc_xattr_common(struct obd_export *exp,const struct req_format *fmt,
 278                             const struct lu_fid *fid, int opcode, u64 valid,
 279                             const char *xattr_name, const char *input,
 280                             int input_size, int output_size, int flags,
 281                             __u32 suppgid, struct ptlrpc_request **request)
 282 {
 283         struct ptlrpc_request *req;
 284         int   xattr_namelen = 0;
 285         char *tmp;
 286         int   rc;
 287         ENTRY;
 288
 289         *request = NULL;
 290         req = ptlrpc_request_alloc(class_exp2cliimp(exp), fmt);
 291         if (req == NULL)
 292                 RETURN(-ENOMEM);
 293
 294         if (xattr_name) {
 295                 xattr_namelen = strlen(xattr_name) + 1;
 296                 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 297                                      xattr_namelen);
 298         }
 299         if (input_size) {
 300                 LASSERT(input);
 301                 req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT,
 302                                      input_size);
 303         }
 304
 305         /* Flush local XATTR locks to get rid of a possible cancel RPC */
 306         if (opcode == MDS_REINT && fid_is_sane(fid) &&
 307             exp->exp_connect_data.ocd_ibits_known & MDS_INODELOCK_XATTR) {
 308                 struct list_head cancels = LIST_HEAD_INIT(cancels);
 309                 int count;
 310
 311                 /* Without that packing would fail */
 312                 if (input_size == 0)
 313                         req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
 314                                              RCL_CLIENT, 0);
 315
 316                 count = mdc_resource_get_unused(exp, fid,
 317                                                 &cancels, LCK_EX,
 318                                                 MDS_INODELOCK_XATTR);
 319
 320                 rc = mdc_prep_elc_req(exp, req, MDS_REINT, &cancels, count);
 321                 if (rc) {
 322                         ptlrpc_request_free(req);
 323                         RETURN(rc);
 324                 }
 325         } else {
 326                 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, opcode);
 327                 if (rc) {
 328                         ptlrpc_request_free(req);
 329                         RETURN(rc);
 330                 }
 331         }
 332
 333         if (opcode == MDS_REINT) {
 334                 struct mdt_rec_setxattr *rec;
 335
 336                 CLASSERT(sizeof(struct mdt_rec_setxattr) ==
 337                          sizeof(struct mdt_rec_reint));
 338                 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
 339                 rec->sx_opcode = REINT_SETXATTR;
 340                 rec->sx_fsuid  = from_kuid(&init_user_ns, current_fsuid());
 341                 rec->sx_fsgid  = from_kgid(&init_user_ns, current_fsgid());
 342                 rec->sx_cap    = cfs_curproc_cap_pack();
 343                 rec->sx_suppgid1 = suppgid;
 344                 rec->sx_suppgid2 = -1;
 345                 rec->sx_fid    = *fid;
 346                 rec->sx_valid  = valid | OBD_MD_FLCTIME;
 347                 rec->sx_time   = ktime_get_real_seconds();
 348                 rec->sx_size   = output_size;
 349                 rec->sx_flags  = flags;
 350         } else {
 351                 mdc_pack_body(req, fid, valid, output_size, suppgid, flags);
 352         }
 353
 354         if (xattr_name) {
 355                 tmp = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
 356                 memcpy(tmp, xattr_name, xattr_namelen);
 357         }
 358         if (input_size) {
 359                 tmp = req_capsule_client_get(&req->rq_pill, &RMF_EADATA);
 360                 memcpy(tmp, input, input_size);
 361         }
 362
 363         if (req_capsule_has_field(&req->rq_pill, &RMF_EADATA, RCL_SERVER))
 364                 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
 365                                      RCL_SERVER, output_size);
 366         ptlrpc_request_set_replen(req);
 367
 368         /* make rpc */
 369         if (opcode == MDS_REINT)
 370                 mdc_get_mod_rpc_slot(req, NULL);
 371
 372         rc = ptlrpc_queue_wait(req);
 373
 374         if (opcode == MDS_REINT)
 375                 mdc_put_mod_rpc_slot(req, NULL);
 376
 377         /* For XATTR_LUSTRE_LOV.add, we'd save the LOVEA for replay. */
 378         if (opcode == MDS_REINT && rc == 0) {
 379                 struct mdt_body *body;
 380                 struct req_capsule *pill = &req->rq_pill;
 381
 382                 body = req_capsule_server_get(pill, &RMF_MDT_BODY);
 383                 if (body == NULL)
 384                         GOTO(out, rc = -EPROTO);
 385
 386                 if (body->mbo_valid & OBD_MD_FLEASIZE) {
 387                         void *eadata;
 388
 389                         eadata = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
 390                                                         body->mbo_eadatasize);
 391                         if (eadata == NULL)
 392                                 GOTO(out, rc = -EPROTO);
 393
 394                         rc = mdc_save_lovea(req, &RMF_EADATA, eadata,
 395                                             body->mbo_eadatasize);
 396                         if (rc)
 397                                 GOTO(out, rc);
 398                 }
 399         }
 400 out:
 401         if (rc)
 402                 ptlrpc_req_finished(req);
 403         else
 404                 *request = req;
 405         RETURN(rc);
 406 }
 407
 408 static int mdc_setxattr(struct obd_export *exp, const struct lu_fid *fid,
 409                         u64 valid, const char *xattr_name,
 410                         const char *input, int input_size, int output_size,
 411                         int flags, __u32 suppgid,
 412                         struct ptlrpc_request **request)
 413 {
 414         return mdc_xattr_common(exp, &RQF_MDS_REINT_SETXATTR,
 415                                 fid, MDS_REINT, valid, xattr_name,
 416                                 input, input_size, output_size, flags,
 417                                 suppgid, request);
 418 }
 419
 420 static int mdc_getxattr(struct obd_export *exp, const struct lu_fid *fid,
 421                         u64 valid, const char *xattr_name,
 422                         const char *input, int input_size, int output_size,
 423                         int flags, struct ptlrpc_request **request)
 424 {
 425         return mdc_xattr_common(exp, &RQF_MDS_GETXATTR,
 426                                 fid, MDS_GETXATTR, valid, xattr_name,
 427                                 input, input_size, output_size, flags,
 428                                 -1, request);
 429 }
 430
 431 #ifdef CONFIG_FS_POSIX_ACL
 432 static int mdc_unpack_acl(struct ptlrpc_request *req, struct lustre_md *md)
 433 {
 434         struct req_capsule     *pill = &req->rq_pill;
 435         struct mdt_body        *body = md->body;
 436         struct posix_acl       *acl;
 437         void                   *buf;
 438         int                     rc;
 439         ENTRY;
 440
 441         if (!body->mbo_aclsize)
 442                 RETURN(0);
 443
 444         buf = req_capsule_server_sized_get(pill, &RMF_ACL, body->mbo_aclsize);
 445
 446         if (!buf)
 447                 RETURN(-EPROTO);
 448
 449         acl = posix_acl_from_xattr(&init_user_ns, buf, body->mbo_aclsize);
 450         if (acl == NULL)
 451                 RETURN(0);
 452         if (IS_ERR(acl)) {
 453                 rc = PTR_ERR(acl);
 454                 CERROR("convert xattr to acl: %d\n", rc);
 455                 RETURN(rc);
 456         }
 457
 458         rc = posix_acl_valid(&init_user_ns, acl);
 459         if (rc) {
 460                 CERROR("validate acl: %d\n", rc);
 461                 posix_acl_release(acl);
 462                 RETURN(rc);
 463         }
 464
 465         md->posix_acl = acl;
 466         RETURN(0);
 467 }
 468 #else
 469 #define mdc_unpack_acl(req, md) 0
 470 #endif
 471
 472 int mdc_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
 473                       struct obd_export *dt_exp, struct obd_export *md_exp,
 474                       struct lustre_md *md)
 475 {
 476         struct req_capsule *pill = &req->rq_pill;
 477         int rc;
 478         ENTRY;
 479
 480         LASSERT(md);
 481         memset(md, 0, sizeof(*md));
 482
 483         md->body = req_capsule_server_get(pill, &RMF_MDT_BODY);
 484         LASSERT(md->body != NULL);
 485
 486         if (md->body->mbo_valid & OBD_MD_FLEASIZE) {
 487                 if (!S_ISREG(md->body->mbo_mode)) {
 488                         CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, should be a "
 489                                "regular file, but is not\n");
 490                         GOTO(out, rc = -EPROTO);
 491                 }
 492
 493                 if (md->body->mbo_eadatasize == 0) {
 494                         CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, "
 495                                "but eadatasize 0\n");
 496                         GOTO(out, rc = -EPROTO);
 497                 }
 498
 499                 md->layout.lb_len = md->body->mbo_eadatasize;
 500                 md->layout.lb_buf = req_capsule_server_sized_get(pill,
 501                                                         &RMF_MDT_MD,
 502                                                         md->layout.lb_len);
 503                 if (md->layout.lb_buf == NULL)
 504                         GOTO(out, rc = -EPROTO);
 505         } else if (md->body->mbo_valid & OBD_MD_FLDIREA) {
 506                 const union lmv_mds_md *lmv;
 507                 size_t lmv_size;
 508
 509                 if (!S_ISDIR(md->body->mbo_mode)) {
 510                         CDEBUG(D_INFO, "OBD_MD_FLDIREA set, should be a "
 511                                "directory, but is not\n");
 512                         GOTO(out, rc = -EPROTO);
 513                 }
 514
 515                 lmv_size = md->body->mbo_eadatasize;
 516                 if (lmv_size == 0) {
 517                         CDEBUG(D_INFO, "OBD_MD_FLDIREA is set, "
 518                                "but eadatasize 0\n");
 519                         RETURN(-EPROTO);
 520                 }
 521
 522                 if (md->body->mbo_valid & OBD_MD_MEA) {
 523                         lmv = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
 524                                                            lmv_size);
 525                         if (lmv == NULL)
 526                                 GOTO(out, rc = -EPROTO);
 527
 528                         rc = md_unpackmd(md_exp, &md->lmv, lmv, lmv_size);
 529                         if (rc < 0)
 530                                 GOTO(out, rc);
 531
 532                         if (rc < (typeof(rc))sizeof(*md->lmv)) {
 533                                 CDEBUG(D_INFO, "size too small:  "
 534                                        "rc < sizeof(*md->lmv) (%d < %d)\n",
 535                                         rc, (int)sizeof(*md->lmv));
 536                                 GOTO(out, rc = -EPROTO);
 537                         }
 538                 }
 539         }
 540         rc = 0;
 541
 542         if (md->body->mbo_valid & OBD_MD_FLACL) {
 543                 /* for ACL, it's possible that FLACL is set but aclsize is zero.
 544                  * only when aclsize != 0 there's an actual segment for ACL
 545                  * in reply buffer.
 546                  */
 547                 if (md->body->mbo_aclsize) {
 548                         rc = mdc_unpack_acl(req, md);
 549                         if (rc)
 550                                 GOTO(out, rc);
 551 #ifdef CONFIG_FS_POSIX_ACL
 552                 } else {
 553                         md->posix_acl = NULL;
 554 #endif
 555                 }
 556         }
 557
 558         EXIT;
 559 out:
 560         if (rc) {
 561 #ifdef CONFIG_FS_POSIX_ACL
 562                 posix_acl_release(md->posix_acl);
 563 #endif
 564         }
 565         return rc;
 566 }
 567
 568 int mdc_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
 569 {
 570         ENTRY;
 571         RETURN(0);
 572 }
 573
 574 void mdc_replay_open(struct ptlrpc_request *req)
 575 {
 576         struct md_open_data *mod = req->rq_cb_data;
 577         struct ptlrpc_request *close_req;
 578         struct obd_client_handle *och;
 579         struct lustre_handle old;
 580         struct mdt_body *body;
 581         ENTRY;
 582
 583         if (mod == NULL) {
 584                 DEBUG_REQ(D_ERROR, req,
 585                           "Can't properly replay without open data.");
 586                 EXIT;
 587                 return;
 588         }
 589
 590         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 591         LASSERT(body != NULL);
 592
 593         och = mod->mod_och;
 594         if (och != NULL) {
 595                 struct lustre_handle *file_fh;
 596
 597                 LASSERT(och->och_magic == OBD_CLIENT_HANDLE_MAGIC);
 598
 599                 file_fh = &och->och_fh;
 600                 CDEBUG(D_HA, "updating handle from %#llx to %#llx\n",
 601                        file_fh->cookie, body->mbo_handle.cookie);
 602                 old = *file_fh;
 603                 *file_fh = body->mbo_handle;
 604         }
 605         close_req = mod->mod_close_req;
 606         if (close_req != NULL) {
 607                 __u32 opc = lustre_msg_get_opc(close_req->rq_reqmsg);
 608                 struct mdt_ioepoch *epoch;
 609
 610                 LASSERT(opc == MDS_CLOSE);
 611                 epoch = req_capsule_client_get(&close_req->rq_pill,
 612                                                &RMF_MDT_EPOCH);
 613                 LASSERT(epoch);
 614
 615                 if (och != NULL)
 616                         LASSERT(!memcmp(&old, &epoch->mio_handle, sizeof(old)));
 617
 618                 DEBUG_REQ(D_HA, close_req, "updating close body with new fh");
 619                 epoch->mio_handle = body->mbo_handle;
 620         }
 621         EXIT;
 622 }
 623
 624 void mdc_commit_open(struct ptlrpc_request *req)
 625 {
 626         struct md_open_data *mod = req->rq_cb_data;
 627         if (mod == NULL)
 628                 return;
 629
 630         /**
 631          * No need to touch md_open_data::mod_och, it holds a reference on
 632          * \var mod and will zero references to each other, \var mod will be
 633          * freed after that when md_open_data::mod_och will put the reference.
 634          */
 635
 636         /**
 637          * Do not let open request to disappear as it still may be needed
 638          * for close rpc to happen (it may happen on evict only, otherwise
 639          * ptlrpc_request::rq_replay does not let mdc_commit_open() to be
 640          * called), just mark this rpc as committed to distinguish these 2
 641          * cases, see mdc_close() for details. The open request reference will
 642          * be put along with freeing \var mod.
 643          */
 644         ptlrpc_request_addref(req);
 645         spin_lock(&req->rq_lock);
 646         req->rq_committed = 1;
 647         spin_unlock(&req->rq_lock);
 648         req->rq_cb_data = NULL;
 649         obd_mod_put(mod);
 650 }
 651
 652 int mdc_set_open_replay_data(struct obd_export *exp,
 653                              struct obd_client_handle *och,
 654                              struct lookup_intent *it)
 655 {
 656         struct md_open_data     *mod;
 657         struct mdt_rec_create   *rec;
 658         struct mdt_body         *body;
 659         struct ptlrpc_request   *open_req = it->it_request;
 660         struct obd_import       *imp = open_req->rq_import;
 661         ENTRY;
 662
 663         if (!open_req->rq_replay)
 664                 RETURN(0);
 665
 666         rec = req_capsule_client_get(&open_req->rq_pill, &RMF_REC_REINT);
 667         body = req_capsule_server_get(&open_req->rq_pill, &RMF_MDT_BODY);
 668         LASSERT(rec != NULL);
 669         /* Incoming message in my byte order (it's been swabbed). */
 670         /* Outgoing messages always in my byte order. */
 671         LASSERT(body != NULL);
 672
 673         /* Only if the import is replayable, we set replay_open data */
 674         if (och && imp->imp_replayable) {
 675                 mod = obd_mod_alloc();
 676                 if (mod == NULL) {
 677                         DEBUG_REQ(D_ERROR, open_req,
 678                                   "Can't allocate md_open_data");
 679                         RETURN(0);
 680                 }
 681
 682                 /**
 683                  * Take a reference on \var mod, to be freed on mdc_close().
 684                  * It protects \var mod from being freed on eviction (commit
 685                  * callback is called despite rq_replay flag).
 686                  * Another reference for \var och.
 687                  */
 688                 obd_mod_get(mod);
 689                 obd_mod_get(mod);
 690
 691                 spin_lock(&open_req->rq_lock);
 692                 och->och_mod = mod;
 693                 mod->mod_och = och;
 694                 mod->mod_is_create = it_disposition(it, DISP_OPEN_CREATE) ||
 695                                      it_disposition(it, DISP_OPEN_STRIPE);
 696                 mod->mod_open_req = open_req;
 697                 open_req->rq_cb_data = mod;
 698                 open_req->rq_commit_cb = mdc_commit_open;
 699                 spin_unlock(&open_req->rq_lock);
 700         }
 701
 702         rec->cr_fid2 = body->mbo_fid1;
 703         rec->cr_ioepoch = body->mbo_ioepoch;
 704         rec->cr_old_handle.cookie = body->mbo_handle.cookie;
 705         open_req->rq_replay_cb = mdc_replay_open;
 706         if (!fid_is_sane(&body->mbo_fid1)) {
 707                 DEBUG_REQ(D_ERROR, open_req, "Saving replay request with "
 708                           "insane fid");
 709                 LBUG();
 710         }
 711
 712         DEBUG_REQ(D_RPCTRACE, open_req, "Set up open replay data");
 713         RETURN(0);
 714 }
 715
 716 static void mdc_free_open(struct md_open_data *mod)
 717 {
 718         int committed = 0;
 719
 720         if (mod->mod_is_create == 0 &&
 721             imp_connect_disp_stripe(mod->mod_open_req->rq_import))
 722                 committed = 1;
 723
 724         /**
 725          * No reason to asssert here if the open request has
 726          * rq_replay == 1. It means that mdc_close failed, and
 727          * close request wasn`t sent. It is not fatal to client.
 728          * The worst thing is eviction if the client gets open lock
 729          **/
 730
 731         DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "free open request rq_replay"
 732                   "= %d\n", mod->mod_open_req->rq_replay);
 733
 734         ptlrpc_request_committed(mod->mod_open_req, committed);
 735         if (mod->mod_close_req)
 736                 ptlrpc_request_committed(mod->mod_close_req, committed);
 737 }
 738
 739 int mdc_clear_open_replay_data(struct obd_export *exp,
 740                                struct obd_client_handle *och)
 741 {
 742         struct md_open_data *mod = och->och_mod;
 743         ENTRY;
 744
 745         /**
 746          * It is possible to not have \var mod in a case of eviction between
 747          * lookup and ll_file_open().
 748          **/
 749         if (mod == NULL)
 750                 RETURN(0);
 751
 752         LASSERT(mod != LP_POISON);
 753         LASSERT(mod->mod_open_req != NULL);
 754         mdc_free_open(mod);
 755
 756         mod->mod_och = NULL;
 757         och->och_mod = NULL;
 758         obd_mod_put(mod);
 759
 760         RETURN(0);
 761 }
 762
 763 static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
 764                      struct md_open_data *mod, struct ptlrpc_request **request)
 765 {
 766         struct obd_device     *obd = class_exp2obd(exp);
 767         struct ptlrpc_request *req;
 768         struct req_format     *req_fmt;
 769         int                    rc;
 770         int                    saved_rc = 0;
 771         ENTRY;
 772
 773         if (op_data->op_bias & MDS_HSM_RELEASE) {
 774                 req_fmt = &RQF_MDS_INTENT_CLOSE;
 775
 776                 /* allocate a FID for volatile file */
 777                 rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
 778                 if (rc < 0) {
 779                         CERROR("%s: "DFID" failed to allocate FID: %d\n",
 780                                obd->obd_name, PFID(&op_data->op_fid1), rc);
 781                         /* save the errcode and proceed to close */
 782                         saved_rc = rc;
 783                 }
 784         } else if (op_data->op_bias & MDS_CLOSE_LAYOUT_SWAP) {
 785                 req_fmt = &RQF_MDS_INTENT_CLOSE;
 786         } else {
 787                 req_fmt = &RQF_MDS_CLOSE;
 788         }
 789
 790         *request = NULL;
 791         if (OBD_FAIL_CHECK(OBD_FAIL_MDC_CLOSE))
 792                 req = NULL;
 793         else
 794                 req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);
 795
 796         /* Ensure that this close's handle is fixed up during replay. */
 797         if (likely(mod != NULL)) {
 798                 LASSERTF(mod->mod_open_req != NULL &&
 799                          mod->mod_open_req->rq_type != LI_POISON,
 800                          "POISONED open %p!\n", mod->mod_open_req);
 801
 802                 mod->mod_close_req = req;
 803
 804                 DEBUG_REQ(D_HA, mod->mod_open_req, "matched open");
 805                 /* We no longer want to preserve this open for replay even
 806                  * though the open was committed. b=3632, b=3633 */
 807                 spin_lock(&mod->mod_open_req->rq_lock);
 808                 mod->mod_open_req->rq_replay = 0;
 809                 spin_unlock(&mod->mod_open_req->rq_lock);
 810         } else {
 811                 CDEBUG(D_HA, "couldn't find open req; expecting close error\n");
 812         }
 813         if (req == NULL) {
 814                 /**
 815                  * TODO: repeat close after errors
 816                  */
 817                 CWARN("%s: close of FID "DFID" failed, file reference will be "
 818                       "dropped when this client unmounts or is evicted\n",
 819                       obd->obd_name, PFID(&op_data->op_fid1));
 820                 GOTO(out, rc = -ENOMEM);
 821         }
 822
 823         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
 824         if (rc) {
 825                 ptlrpc_request_free(req);
 826                 req = NULL;
 827                 GOTO(out, rc);
 828         }
 829
 830         /* To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
 831          * portal whose threads are not taking any DLM locks and are therefore
 832          * always progressing */
 833         req->rq_request_portal = MDS_READPAGE_PORTAL;
 834         ptlrpc_at_set_req_timeout(req);
 835
 836
 837         mdc_close_pack(req, op_data);
 838
 839         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 840                              obd->u.cli.cl_default_mds_easize);
 841
 842         ptlrpc_request_set_replen(req);
 843
 844         mdc_get_mod_rpc_slot(req, NULL);
 845         rc = ptlrpc_queue_wait(req);
 846         mdc_put_mod_rpc_slot(req, NULL);
 847
 848         if (req->rq_repmsg == NULL) {
 849                 CDEBUG(D_RPCTRACE, "request failed to send: %p, %d\n", req,
 850                        req->rq_status);
 851                 if (rc == 0)
 852                         rc = req->rq_status ?: -EIO;
 853         } else if (rc == 0 || rc == -EAGAIN) {
 854                 struct mdt_body *body;
 855
 856                 rc = lustre_msg_get_status(req->rq_repmsg);
 857                 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
 858                         DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR, err "
 859                                   "= %d", rc);
 860                         if (rc > 0)
 861                                 rc = -rc;
 862                 }
 863                 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 864                 if (body == NULL)
 865                         rc = -EPROTO;
 866         } else if (rc == -ESTALE) {
 867                 /**
 868                  * it can be allowed error after 3633 if open was committed and
 869                  * server failed before close was sent. Let's check if mod
 870                  * exists and return no error in that case
 871                  */
 872                 if (mod) {
 873                         DEBUG_REQ(D_HA, req, "Reset ESTALE = %d", rc);
 874                         LASSERT(mod->mod_open_req != NULL);
 875                         if (mod->mod_open_req->rq_committed)
 876                                 rc = 0;
 877                 }
 878         }
 879
 880 out:
 881         if (mod) {
 882                 if (rc != 0)
 883                         mod->mod_close_req = NULL;
 884                 /* Since now, mod is accessed through open_req only,
 885                  * thus close req does not keep a reference on mod anymore. */
 886                 obd_mod_put(mod);
 887         }
 888         *request = req;
 889
 890         RETURN(rc < 0 ? rc : saved_rc);
 891 }
 892
 893 static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid,
 894                        u64 offset, struct page **pages, int npages,
 895                        struct ptlrpc_request **request)
 896 {
 897         struct ptlrpc_request   *req;
 898         struct ptlrpc_bulk_desc *desc;
 899         int                      i;
 900         wait_queue_head_t        waitq;
 901         int                      resends = 0;
 902         struct l_wait_info       lwi;
 903         int                      rc;
 904         ENTRY;
 905
 906         *request = NULL;
 907         init_waitqueue_head(&waitq);
 908
 909 restart_bulk:
 910         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE);
 911         if (req == NULL)
 912                 RETURN(-ENOMEM);
 913
 914         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE);
 915         if (rc) {
 916                 ptlrpc_request_free(req);
 917                 RETURN(rc);
 918         }
 919
 920         req->rq_request_portal = MDS_READPAGE_PORTAL;
 921         ptlrpc_at_set_req_timeout(req);
 922
 923         desc = ptlrpc_prep_bulk_imp(req, npages, 1,
 924                                     PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
 925                                     MDS_BULK_PORTAL,
 926                                     &ptlrpc_bulk_kiov_pin_ops);
 927         if (desc == NULL) {
 928                 ptlrpc_request_free(req);
 929                 RETURN(-ENOMEM);
 930         }
 931
 932         /* NB req now owns desc and will free it when it gets freed */
 933         for (i = 0; i < npages; i++)
 934                 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
 935                                                  PAGE_SIZE);
 936
 937         mdc_readdir_pack(req, offset, PAGE_SIZE * npages, fid);
 938
 939         ptlrpc_request_set_replen(req);
 940         rc = ptlrpc_queue_wait(req);
 941         if (rc) {
 942                 ptlrpc_req_finished(req);
 943                 if (rc != -ETIMEDOUT)
 944                         RETURN(rc);
 945
 946                 resends++;
 947                 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
 948                         CERROR("%s: too many resend retries: rc = %d\n",
 949                                exp->exp_obd->obd_name, -EIO);
 950                         RETURN(-EIO);
 951                 }
 952                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL,
 953                                        NULL);
 954                 l_wait_event(waitq, 0, &lwi);
 955
 956                 goto restart_bulk;
 957         }
 958
 959         rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
 960                                           req->rq_bulk->bd_nob_transferred);
 961         if (rc < 0) {
 962                 ptlrpc_req_finished(req);
 963                 RETURN(rc);
 964         }
 965
 966         if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) {
 967                 CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n",
 968                        exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred,
 969                        PAGE_SIZE * npages);
 970                 ptlrpc_req_finished(req);
 971                 RETURN(-EPROTO);
 972         }
 973
 974         *request = req;
 975         RETURN(0);
 976 }
 977
 978 static void mdc_release_page(struct page *page, int remove)
 979 {
 980         if (remove) {
 981                 lock_page(page);
 982                 if (likely(page->mapping != NULL))
 983                         truncate_complete_page(page->mapping, page);
 984                 unlock_page(page);
 985         }
 986         put_page(page);
 987 }
 988
 989 static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash,
 990                                     __u64 *start, __u64 *end, int hash64)
 991 {
 992         /*
 993          * Complement of hash is used as an index so that
 994          * radix_tree_gang_lookup() can be used to find a page with starting
 995          * hash _smaller_ than one we are looking for.
 996          */
 997         unsigned long offset = hash_x_index(*hash, hash64);
 998         struct page *page;
 999         int found;
1000
1001         spin_lock_irq(&mapping->tree_lock);
1002         found = radix_tree_gang_lookup(&mapping->page_tree,
1003                                        (void **)&page, offset, 1);
1004         if (found > 0 && !radix_tree_exceptional_entry(page)) {
1005                 struct lu_dirpage *dp;
1006
1007                 get_page(page);
1008                 spin_unlock_irq(&mapping->tree_lock);
1009                 /*
1010                  * In contrast to find_lock_page() we are sure that directory
1011                  * page cannot be truncated (while DLM lock is held) and,
1012                  * hence, can avoid restart.
1013                  *
1014                  * In fact, page cannot be locked here at all, because
1015                  * mdc_read_page_remote does synchronous io.
1016                  */
1017                 wait_on_page_locked(page);
1018                 if (PageUptodate(page)) {
1019                         dp = kmap(page);
1020                         if (BITS_PER_LONG == 32 && hash64) {
1021                                 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1022                                 *end   = le64_to_cpu(dp->ldp_hash_end) >> 32;
1023                                 *hash  = *hash >> 32;
1024                         } else {
1025                                 *start = le64_to_cpu(dp->ldp_hash_start);
1026                                 *end   = le64_to_cpu(dp->ldp_hash_end);
1027                         }
1028                         if (unlikely(*start == 1 && *hash == 0))
1029                                 *hash = *start;
1030                         else
1031                                 LASSERTF(*start <= *hash, "start = %#llx"
1032                                          ",end = %#llx,hash = %#llx\n",
1033                                          *start, *end, *hash);
1034                         CDEBUG(D_VFSTRACE, "offset %lx [%#llx %#llx],"
1035                               " hash %#llx\n", offset, *start, *end, *hash);
1036                         if (*hash > *end) {
1037                                 kunmap(page);
1038                                 mdc_release_page(page, 0);
1039                                 page = NULL;
1040                         } else if (*end != *start && *hash == *end) {
1041                                 /*
1042                                  * upon hash collision, remove this page,
1043                                  * otherwise put page reference, and
1044                                  * mdc_read_page_remote() will issue RPC to
1045                                  * fetch the page we want.
1046                                  */
1047                                 kunmap(page);
1048                                 mdc_release_page(page,
1049                                     le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1050                                 page = NULL;
1051                         }
1052                 } else {
1053                         put_page(page);
1054                         page = ERR_PTR(-EIO);
1055                 }
1056         } else {
1057                 spin_unlock_irq(&mapping->tree_lock);
1058                 page = NULL;
1059         }
1060         return page;
1061 }
1062
1063 /*
1064  * Adjust a set of pages, each page containing an array of lu_dirpages,
1065  * so that each page can be used as a single logical lu_dirpage.
1066  *
1067  * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
1068  * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
1069  * struct lu_dirent.  It has size up to LU_PAGE_SIZE. The ldp_hash_end
1070  * value is used as a cookie to request the next lu_dirpage in a
1071  * directory listing that spans multiple pages (two in this example):
1072  *   ________
1073  *  |        |
1074  * .|--------v-------   -----.
1075  * |s|e|f|p|ent|ent| ... |ent|
1076  * '--|--------------   -----'   Each PAGE contains a single
1077  *    '------.                   lu_dirpage.
1078  * .---------v-------   -----.
1079  * |s|e|f|p|ent| 0 | ... | 0 |
1080  * '-----------------   -----'
1081  *
1082  * However, on hosts where the native VM page size (PAGE_SIZE) is
1083  * larger than LU_PAGE_SIZE, a single host page may contain multiple
1084  * lu_dirpages. After reading the lu_dirpages from the MDS, the
1085  * ldp_hash_end of the first lu_dirpage refers to the one immediately
1086  * after it in the same PAGE (arrows simplified for brevity, but
1087  * in general e0==s1, e1==s2, etc.):
1088  *
1089  * .--------------------   -----.
1090  * |s0|e0|f0|p|ent|ent| ... |ent|
1091  * |---v----------------   -----|
1092  * |s1|e1|f1|p|ent|ent| ... |ent|
1093  * |---v----------------   -----|  Here, each PAGE contains
1094  *             ...                 multiple lu_dirpages.
1095  * |---v----------------   -----|
1096  * |s'|e'|f'|p|ent|ent| ... |ent|
1097  * '---|----------------   -----'
1098  *     v
1099  * .----------------------------.
1100  * |        next PAGE           |
1101  *
1102  * This structure is transformed into a single logical lu_dirpage as follows:
1103  *
1104  * - Replace e0 with e' so the request for the next lu_dirpage gets the page
1105  *   labeled 'next PAGE'.
1106  *
1107  * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
1108  *   a hash collision with the next page exists.
1109  *
1110  * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
1111  *   to the first entry of the next lu_dirpage.
1112  */
1113 #if PAGE_SIZE > LU_PAGE_SIZE
1114 static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs)
1115 {
1116         int i;
1117
1118         for (i = 0; i < cfs_pgs; i++) {
1119                 struct lu_dirpage       *dp = kmap(pages[i]);
1120                 struct lu_dirpage       *first = dp;
1121                 struct lu_dirent        *end_dirent = NULL;
1122                 struct lu_dirent        *ent;
1123                 __u64           hash_end = le64_to_cpu(dp->ldp_hash_end);
1124                 __u32           flags = le32_to_cpu(dp->ldp_flags);
1125
1126                 while (--lu_pgs > 0) {
1127                         ent = lu_dirent_start(dp);
1128                         for (end_dirent = ent; ent != NULL;
1129                              end_dirent = ent, ent = lu_dirent_next(ent));
1130
1131                         /* Advance dp to next lu_dirpage. */
1132                         dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
1133
1134                         /* Check if we've reached the end of the PAGE. */
1135                         if (!((unsigned long)dp & ~PAGE_MASK))
1136                                 break;
1137
1138                         /* Save the hash and flags of this lu_dirpage. */
1139                         hash_end = le64_to_cpu(dp->ldp_hash_end);
1140                         flags = le32_to_cpu(dp->ldp_flags);
1141
1142                         /* Check if lu_dirpage contains no entries. */
1143                         if (end_dirent == NULL)
1144                                 break;
1145
1146                         /* Enlarge the end entry lde_reclen from 0 to
1147                          * first entry of next lu_dirpage. */
1148                         LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
1149                         end_dirent->lde_reclen =
1150                                 cpu_to_le16((char *)(dp->ldp_entries) -
1151                                             (char *)end_dirent);
1152                 }
1153
1154                 first->ldp_hash_end = hash_end;
1155                 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
1156                 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
1157
1158                 kunmap(pages[i]);
1159         }
1160         LASSERTF(lu_pgs == 0, "left = %d\n", lu_pgs);
1161 }
1162 #else
1163 #define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0)
1164 #endif  /* PAGE_SIZE > LU_PAGE_SIZE */
1165
1166 /* parameters for readdir page */
1167 struct readpage_param {
1168         struct md_op_data       *rp_mod;
1169         __u64                   rp_off;
1170         int                     rp_hash64;
1171         struct obd_export       *rp_exp;
1172         struct md_callback      *rp_cb;
1173 };
1174
1175 #ifndef HAVE_DELETE_FROM_PAGE_CACHE
1176 static inline void delete_from_page_cache(struct page *page)
1177 {
1178         remove_from_page_cache(page);
1179         put_page(page);
1180 }
1181 #endif
1182
1183 /**
1184  * Read pages from server.
1185  *
1186  * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
1187  * a header lu_dirpage which describes the start/end hash, and whether this
1188  * page is empty (contains no dir entry) or hash collide with next page.
1189  * After client receives reply, several pages will be integrated into dir page
1190  * in PAGE_SIZE (if PAGE_SIZE greater than LU_PAGE_SIZE), and the
1191  * lu_dirpage for this integrated page will be adjusted.
1192  **/
1193 static int mdc_read_page_remote(void *data, struct page *page0)
1194 {
1195         struct readpage_param   *rp = data;
1196         struct page             **page_pool;
1197         struct page             *page;
1198         struct lu_dirpage       *dp;
1199         int                     rd_pgs = 0; /* number of pages read actually */
1200         int                     npages;
1201         struct md_op_data       *op_data = rp->rp_mod;
1202         struct ptlrpc_request   *req;
1203         int                     max_pages = op_data->op_max_pages;
1204         struct inode            *inode;
1205         struct lu_fid           *fid;
1206         int                     i;
1207         int                     rc;
1208         ENTRY;
1209
1210         LASSERT(max_pages > 0 && max_pages <= PTLRPC_MAX_BRW_PAGES);
1211         inode = op_data->op_data;
1212         fid = &op_data->op_fid1;
1213         LASSERT(inode != NULL);
1214
1215         OBD_ALLOC(page_pool, sizeof(page_pool[0]) * max_pages);
1216         if (page_pool != NULL) {
1217                 page_pool[0] = page0;
1218         } else {
1219                 page_pool = &page0;
1220                 max_pages = 1;
1221         }
1222
1223         for (npages = 1; npages < max_pages; npages++) {
1224                 page = page_cache_alloc_cold(inode->i_mapping);
1225                 if (page == NULL)
1226                         break;
1227                 page_pool[npages] = page;
1228         }
1229
1230         rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, page_pool, npages, &req);
1231         if (rc < 0) {
1232                 /* page0 is special, which was added into page cache early */
1233                 delete_from_page_cache(page0);
1234         } else {
1235                 int lu_pgs;
1236
1237                 rd_pgs = (req->rq_bulk->bd_nob_transferred +
1238                             PAGE_SIZE - 1) >> PAGE_SHIFT;
1239                 lu_pgs = req->rq_bulk->bd_nob_transferred >>
1240                                                         LU_PAGE_SHIFT;
1241                 LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
1242
1243                 CDEBUG(D_INODE, "read %d(%d) pages\n", rd_pgs, lu_pgs);
1244
1245                 mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs);
1246
1247                 SetPageUptodate(page0);
1248         }
1249         unlock_page(page0);
1250
1251         ptlrpc_req_finished(req);
1252         CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages);
1253         for (i = 1; i < npages; i++) {
1254                 unsigned long   offset;
1255                 __u64           hash;
1256                 int ret;
1257
1258                 page = page_pool[i];
1259
1260                 if (rc < 0 || i >= rd_pgs) {
1261                         put_page(page);
1262                         continue;
1263                 }
1264
1265                 SetPageUptodate(page);
1266
1267                 dp = kmap(page);
1268                 hash = le64_to_cpu(dp->ldp_hash_start);
1269                 kunmap(page);
1270
1271                 offset = hash_x_index(hash, rp->rp_hash64);
1272
1273                 prefetchw(&page->flags);
1274                 ret = add_to_page_cache_lru(page, inode->i_mapping, offset,
1275                                             GFP_KERNEL);
1276                 if (ret == 0)
1277                         unlock_page(page);
1278                 else
1279                         CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
1280                                " rc = %d\n", offset, ret);
1281                 put_page(page);
1282         }
1283
1284         if (page_pool != &page0)
1285                 OBD_FREE(page_pool, sizeof(page_pool[0]) * max_pages);
1286
1287         RETURN(rc);
1288 }
1289
1290 /**
1291  * Read dir page from cache first, if it can not find it, read it from
1292  * server and add into the cache.
1293  *
1294  * \param[in] exp       MDC export
1295  * \param[in] op_data   client MD stack parameters, transfering parameters
1296  *                      between different layers on client MD stack.
1297  * \param[in] cb_op     callback required for ldlm lock enqueue during
1298  *                      read page
1299  * \param[in] hash_offset the hash offset of the page to be read
1300  * \param[in] ppage     the page to be read
1301  *
1302  * retval               = 0 get the page successfully
1303  *                      errno(<0) get the page failed
1304  */
1305 static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data,
1306                          struct md_callback *cb_op, __u64 hash_offset,
1307                          struct page **ppage)
1308 {
1309         struct lookup_intent    it = { .it_op = IT_READDIR };
1310         struct page             *page;
1311         struct inode            *dir = op_data->op_data;
1312         struct address_space    *mapping;
1313         struct lu_dirpage       *dp;
1314         __u64                   start = 0;
1315         __u64                   end = 0;
1316         struct lustre_handle    lockh;
1317         struct ptlrpc_request   *enq_req = NULL;
1318         struct readpage_param   rp_param;
1319         int rc;
1320
1321         ENTRY;
1322
1323         *ppage = NULL;
1324
1325         LASSERT(dir != NULL);
1326         mapping = dir->i_mapping;
1327
1328         rc = mdc_intent_lock(exp, op_data, &it, &enq_req,
1329                              cb_op->md_blocking_ast, 0);
1330         if (enq_req != NULL)
1331                 ptlrpc_req_finished(enq_req);
1332
1333         if (rc < 0) {
1334                 CERROR("%s: "DFID" lock enqueue fails: rc = %d\n",
1335                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc);
1336                 RETURN(rc);
1337         }
1338
1339         rc = 0;
1340         lockh.cookie = it.it_lock_handle;
1341         mdc_set_lock_data(exp, &lockh, dir, NULL);
1342
1343         rp_param.rp_off = hash_offset;
1344         rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64;
1345         page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end,
1346                                rp_param.rp_hash64);
1347         if (IS_ERR(page)) {
1348                 CERROR("%s: dir page locate: "DFID" at %llu: rc %ld\n",
1349                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1350                        rp_param.rp_off, PTR_ERR(page));
1351                 GOTO(out_unlock, rc = PTR_ERR(page));
1352         } else if (page != NULL) {
1353                 /*
1354                  * XXX nikita: not entirely correct handling of a corner case:
1355                  * suppose hash chain of entries with hash value HASH crosses
1356                  * border between pages P0 and P1. First both P0 and P1 are
1357                  * cached, seekdir() is called for some entry from the P0 part
1358                  * of the chain. Later P0 goes out of cache. telldir(HASH)
1359                  * happens and finds P1, as it starts with matching hash
1360                  * value. Remaining entries from P0 part of the chain are
1361                  * skipped. (Is that really a bug?)
1362                  *
1363                  * Possible solutions: 0. don't cache P1 is such case, handle
1364                  * it as an "overflow" page. 1. invalidate all pages at
1365                  * once. 2. use HASH|1 as an index for P1.
1366                  */
1367                 GOTO(hash_collision, page);
1368         }
1369
1370         rp_param.rp_exp = exp;
1371         rp_param.rp_mod = op_data;
1372         page = read_cache_page(mapping,
1373                                hash_x_index(rp_param.rp_off,
1374                                             rp_param.rp_hash64),
1375                                mdc_read_page_remote, &rp_param);
1376         if (IS_ERR(page)) {
1377                 CDEBUG(D_INFO, "%s: read cache page: "DFID" at %llu: %ld\n",
1378                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1379                        rp_param.rp_off, PTR_ERR(page));
1380                 GOTO(out_unlock, rc = PTR_ERR(page));
1381         }
1382
1383         wait_on_page_locked(page);
1384         (void)kmap(page);
1385         if (!PageUptodate(page)) {
1386                 CERROR("%s: page not updated: "DFID" at %llu: rc %d\n",
1387                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1388                        rp_param.rp_off, -5);
1389                 goto fail;
1390         }
1391         if (!PageChecked(page))
1392                 SetPageChecked(page);
1393         if (PageError(page)) {
1394                 CERROR("%s: page error: "DFID" at %llu: rc %d\n",
1395                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1396                        rp_param.rp_off, -5);
1397                 goto fail;
1398         }
1399
1400 hash_collision:
1401         dp = page_address(page);
1402         if (BITS_PER_LONG == 32 && rp_param.rp_hash64) {
1403                 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1404                 end   = le64_to_cpu(dp->ldp_hash_end) >> 32;
1405                 rp_param.rp_off = hash_offset >> 32;
1406         } else {
1407                 start = le64_to_cpu(dp->ldp_hash_start);
1408                 end   = le64_to_cpu(dp->ldp_hash_end);
1409                 rp_param.rp_off = hash_offset;
1410         }
1411         if (end == start) {
1412                 LASSERT(start == rp_param.rp_off);
1413                 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
1414 #if BITS_PER_LONG == 32
1415                 CWARN("Real page-wide hash collision at [%llu %llu] with "
1416                       "hash %llu\n", le64_to_cpu(dp->ldp_hash_start),
1417                       le64_to_cpu(dp->ldp_hash_end), hash_offset);
1418 #endif
1419
1420                 /*
1421                  * Fetch whole overflow chain...
1422                  *
1423                  * XXX not yet.
1424                  */
1425                 goto fail;
1426         }
1427         *ppage = page;
1428 out_unlock:
1429         ldlm_lock_decref(&lockh, it.it_lock_mode);
1430         return rc;
1431 fail:
1432         kunmap(page);
1433         mdc_release_page(page, 1);
1434         rc = -EIO;
1435         goto out_unlock;
1436 }
1437
1438
1439 static int mdc_statfs(const struct lu_env *env,
1440                       struct obd_export *exp, struct obd_statfs *osfs,
1441                       __u64 max_age, __u32 flags)
1442 {
1443         struct obd_device     *obd = class_exp2obd(exp);
1444         struct ptlrpc_request *req;
1445         struct obd_statfs     *msfs;
1446         struct obd_import     *imp = NULL;
1447         int                    rc;
1448         ENTRY;
1449
1450         /*
1451          * Since the request might also come from lprocfs, so we need
1452          * sync this with client_disconnect_export Bug15684
1453          */
1454         down_read(&obd->u.cli.cl_sem);
1455         if (obd->u.cli.cl_import)
1456                 imp = class_import_get(obd->u.cli.cl_import);
1457         up_read(&obd->u.cli.cl_sem);
1458         if (!imp)
1459                 RETURN(-ENODEV);
1460
1461         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_STATFS,
1462                                         LUSTRE_MDS_VERSION, MDS_STATFS);
1463         if (req == NULL)
1464                 GOTO(output, rc = -ENOMEM);
1465
1466         ptlrpc_request_set_replen(req);
1467
1468         if (flags & OBD_STATFS_NODELAY) {
1469                 /* procfs requests not want stay in wait for avoid deadlock */
1470                 req->rq_no_resend = 1;
1471                 req->rq_no_delay = 1;
1472         }
1473
1474         rc = ptlrpc_queue_wait(req);
1475         if (rc) {
1476                 /* check connection error first */
1477                 if (imp->imp_connect_error)
1478                         rc = imp->imp_connect_error;
1479                 GOTO(out, rc);
1480         }
1481
1482         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1483         if (msfs == NULL)
1484                 GOTO(out, rc = -EPROTO);
1485
1486         *osfs = *msfs;
1487         EXIT;
1488 out:
1489         ptlrpc_req_finished(req);
1490 output:
1491         class_import_put(imp);
1492         return rc;
1493 }
1494
1495 static int mdc_ioc_fid2path(struct obd_export *exp, struct getinfo_fid2path *gf)
1496 {
1497         __u32 keylen, vallen;
1498         void *key;
1499         int rc;
1500
1501         if (gf->gf_pathlen > PATH_MAX)
1502                 RETURN(-ENAMETOOLONG);
1503         if (gf->gf_pathlen < 2)
1504                 RETURN(-EOVERFLOW);
1505
1506         /* Key is KEY_FID2PATH + getinfo_fid2path description */
1507         keylen = cfs_size_round(sizeof(KEY_FID2PATH) + sizeof(*gf) +
1508                                 sizeof(struct lu_fid));
1509         OBD_ALLOC(key, keylen);
1510         if (key == NULL)
1511                 RETURN(-ENOMEM);
1512         memcpy(key, KEY_FID2PATH, sizeof(KEY_FID2PATH));
1513         memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)), gf, sizeof(*gf));
1514         memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*gf),
1515                gf->gf_u.gf_root_fid, sizeof(struct lu_fid));
1516         CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
1517                PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno);
1518
1519         if (!fid_is_sane(&gf->gf_fid))
1520                 GOTO(out, rc = -EINVAL);
1521
1522         /* Val is struct getinfo_fid2path result plus path */
1523         vallen = sizeof(*gf) + gf->gf_pathlen;
1524
1525         rc = obd_get_info(NULL, exp, keylen, key, &vallen, gf);
1526         if (rc != 0 && rc != -EREMOTE)
1527                 GOTO(out, rc);
1528
1529         if (vallen <= sizeof(*gf))
1530                 GOTO(out, rc = -EPROTO);
1531         if (vallen > sizeof(*gf) + gf->gf_pathlen)
1532                 GOTO(out, rc = -EOVERFLOW);
1533
1534         CDEBUG(D_IOCTL, "path got "DFID" from %llu #%d: %s\n",
1535                PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno,
1536                gf->gf_pathlen < 512 ? gf->gf_u.gf_path :
1537                /* only log the last 512 characters of the path */
1538                gf->gf_u.gf_path + gf->gf_pathlen - 512);
1539
1540 out:
1541         OBD_FREE(key, keylen);
1542         return rc;
1543 }
1544
1545 static int mdc_ioc_hsm_progress(struct obd_export *exp,
1546                                 struct hsm_progress_kernel *hpk)
1547 {
1548         struct obd_import               *imp = class_exp2cliimp(exp);
1549         struct hsm_progress_kernel      *req_hpk;
1550         struct ptlrpc_request           *req;
1551         int                              rc;
1552         ENTRY;
1553
1554         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_PROGRESS,
1555                                         LUSTRE_MDS_VERSION, MDS_HSM_PROGRESS);
1556         if (req == NULL)
1557                 GOTO(out, rc = -ENOMEM);
1558
1559         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1560
1561         /* Copy hsm_progress struct */
1562         req_hpk = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_PROGRESS);
1563         if (req_hpk == NULL)
1564                 GOTO(out, rc = -EPROTO);
1565
1566         *req_hpk = *hpk;
1567         req_hpk->hpk_errval = lustre_errno_hton(hpk->hpk_errval);
1568
1569         ptlrpc_request_set_replen(req);
1570
1571         mdc_get_mod_rpc_slot(req, NULL);
1572         rc = ptlrpc_queue_wait(req);
1573         mdc_put_mod_rpc_slot(req, NULL);
1574
1575         GOTO(out, rc);
1576 out:
1577         ptlrpc_req_finished(req);
1578         return rc;
1579 }
1580
1581 static int mdc_ioc_hsm_ct_register(struct obd_import *imp, __u32 archives)
1582 {
1583         __u32                   *archive_mask;
1584         struct ptlrpc_request   *req;
1585         int                      rc;
1586         ENTRY;
1587
1588         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_REGISTER,
1589                                         LUSTRE_MDS_VERSION,
1590                                         MDS_HSM_CT_REGISTER);
1591         if (req == NULL)
1592                 GOTO(out, rc = -ENOMEM);
1593
1594         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1595
1596         /* Copy hsm_progress struct */
1597         archive_mask = req_capsule_client_get(&req->rq_pill,
1598                                               &RMF_MDS_HSM_ARCHIVE);
1599         if (archive_mask == NULL)
1600                 GOTO(out, rc = -EPROTO);
1601
1602         *archive_mask = archives;
1603
1604         ptlrpc_request_set_replen(req);
1605
1606         rc = mdc_queue_wait(req);
1607         GOTO(out, rc);
1608 out:
1609         ptlrpc_req_finished(req);
1610         return rc;
1611 }
1612
1613 static int mdc_ioc_hsm_current_action(struct obd_export *exp,
1614                                       struct md_op_data *op_data)
1615 {
1616         struct hsm_current_action       *hca = op_data->op_data;
1617         struct hsm_current_action       *req_hca;
1618         struct ptlrpc_request           *req;
1619         int                              rc;
1620         ENTRY;
1621
1622         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1623                                    &RQF_MDS_HSM_ACTION);
1624         if (req == NULL)
1625                 RETURN(-ENOMEM);
1626
1627         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_ACTION);
1628         if (rc) {
1629                 ptlrpc_request_free(req);
1630                 RETURN(rc);
1631         }
1632
1633         mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1634                       op_data->op_suppgids[0], 0);
1635
1636         ptlrpc_request_set_replen(req);
1637
1638         rc = mdc_queue_wait(req);
1639         if (rc)
1640                 GOTO(out, rc);
1641
1642         req_hca = req_capsule_server_get(&req->rq_pill,
1643                                          &RMF_MDS_HSM_CURRENT_ACTION);
1644         if (req_hca == NULL)
1645                 GOTO(out, rc = -EPROTO);
1646
1647         *hca = *req_hca;
1648
1649         EXIT;
1650 out:
1651         ptlrpc_req_finished(req);
1652         return rc;
1653 }
1654
1655 static int mdc_ioc_hsm_ct_unregister(struct obd_import *imp)
1656 {
1657         struct ptlrpc_request   *req;
1658         int                      rc;
1659         ENTRY;
1660
1661         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_UNREGISTER,
1662                                         LUSTRE_MDS_VERSION,
1663                                         MDS_HSM_CT_UNREGISTER);
1664         if (req == NULL)
1665                 GOTO(out, rc = -ENOMEM);
1666
1667         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1668
1669         ptlrpc_request_set_replen(req);
1670
1671         rc = mdc_queue_wait(req);
1672         GOTO(out, rc);
1673 out:
1674         ptlrpc_req_finished(req);
1675         return rc;
1676 }
1677
1678 static int mdc_ioc_hsm_state_get(struct obd_export *exp,
1679                                  struct md_op_data *op_data)
1680 {
1681         struct hsm_user_state   *hus = op_data->op_data;
1682         struct hsm_user_state   *req_hus;
1683         struct ptlrpc_request   *req;
1684         int                      rc;
1685         ENTRY;
1686
1687         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1688                                    &RQF_MDS_HSM_STATE_GET);
1689         if (req == NULL)
1690                 RETURN(-ENOMEM);
1691
1692         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_GET);
1693         if (rc != 0) {
1694                 ptlrpc_request_free(req);
1695                 RETURN(rc);
1696         }
1697
1698         mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1699                       op_data->op_suppgids[0], 0);
1700
1701         ptlrpc_request_set_replen(req);
1702
1703         rc = mdc_queue_wait(req);
1704         if (rc)
1705                 GOTO(out, rc);
1706
1707         req_hus = req_capsule_server_get(&req->rq_pill, &RMF_HSM_USER_STATE);
1708         if (req_hus == NULL)
1709                 GOTO(out, rc = -EPROTO);
1710
1711         *hus = *req_hus;
1712
1713         EXIT;
1714 out:
1715         ptlrpc_req_finished(req);
1716         return rc;
1717 }
1718
1719 static int mdc_ioc_hsm_state_set(struct obd_export *exp,
1720                                  struct md_op_data *op_data)
1721 {
1722         struct hsm_state_set    *hss = op_data->op_data;
1723         struct hsm_state_set    *req_hss;
1724         struct ptlrpc_request   *req;
1725         int                      rc;
1726         ENTRY;
1727
1728         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1729                                    &RQF_MDS_HSM_STATE_SET);
1730         if (req == NULL)
1731                 RETURN(-ENOMEM);
1732
1733         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_SET);
1734         if (rc) {
1735                 ptlrpc_request_free(req);
1736                 RETURN(rc);
1737         }
1738
1739         mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1740                       op_data->op_suppgids[0], 0);
1741
1742         /* Copy states */
1743         req_hss = req_capsule_client_get(&req->rq_pill, &RMF_HSM_STATE_SET);
1744         if (req_hss == NULL)
1745                 GOTO(out, rc = -EPROTO);
1746         *req_hss = *hss;
1747
1748         ptlrpc_request_set_replen(req);
1749
1750         mdc_get_mod_rpc_slot(req, NULL);
1751         rc = ptlrpc_queue_wait(req);
1752         mdc_put_mod_rpc_slot(req, NULL);
1753
1754         GOTO(out, rc);
1755 out:
1756         ptlrpc_req_finished(req);
1757         return rc;
1758 }
1759
1760 static int mdc_ioc_hsm_request(struct obd_export *exp,
1761                                struct hsm_user_request *hur)
1762 {
1763         struct obd_import       *imp = class_exp2cliimp(exp);
1764         struct ptlrpc_request   *req;
1765         struct hsm_request      *req_hr;
1766         struct hsm_user_item    *req_hui;
1767         char                    *req_opaque;
1768         int                      rc;
1769         ENTRY;
1770
1771         req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_REQUEST);
1772         if (req == NULL)
1773                 GOTO(out, rc = -ENOMEM);
1774
1775         req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM, RCL_CLIENT,
1776                              hur->hur_request.hr_itemcount
1777                              * sizeof(struct hsm_user_item));
1778         req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA, RCL_CLIENT,
1779                              hur->hur_request.hr_data_len);
1780
1781         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_REQUEST);
1782         if (rc) {
1783                 ptlrpc_request_free(req);
1784                 RETURN(rc);
1785         }
1786
1787         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1788
1789         /* Copy hsm_request struct */
1790         req_hr = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_REQUEST);
1791         if (req_hr == NULL)
1792                 GOTO(out, rc = -EPROTO);
1793         *req_hr = hur->hur_request;
1794
1795         /* Copy hsm_user_item structs */
1796         req_hui = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM);
1797         if (req_hui == NULL)
1798                 GOTO(out, rc = -EPROTO);
1799         memcpy(req_hui, hur->hur_user_item,
1800                hur->hur_request.hr_itemcount * sizeof(struct hsm_user_item));
1801
1802         /* Copy opaque field */
1803         req_opaque = req_capsule_client_get(&req->rq_pill, &RMF_GENERIC_DATA);
1804         if (req_opaque == NULL)
1805                 GOTO(out, rc = -EPROTO);
1806         memcpy(req_opaque, hur_data(hur), hur->hur_request.hr_data_len);
1807
1808         ptlrpc_request_set_replen(req);
1809
1810         mdc_get_mod_rpc_slot(req, NULL);
1811         rc = ptlrpc_queue_wait(req);
1812         mdc_put_mod_rpc_slot(req, NULL);
1813
1814         GOTO(out, rc);
1815
1816 out:
1817         ptlrpc_req_finished(req);
1818         return rc;
1819 }
1820
1821 static struct kuc_hdr *changelog_kuc_hdr(char *buf, size_t len, __u32 flags)
1822 {
1823         struct kuc_hdr *lh = (struct kuc_hdr *)buf;
1824
1825         LASSERT(len <= KUC_CHANGELOG_MSG_MAXSIZE);
1826
1827         lh->kuc_magic = KUC_MAGIC;
1828         lh->kuc_transport = KUC_TRANSPORT_CHANGELOG;
1829         lh->kuc_flags = flags;
1830         lh->kuc_msgtype = CL_RECORD;
1831         lh->kuc_msglen = len;
1832         return lh;
1833 }
1834
1835 struct changelog_show {
1836         __u64                            cs_startrec;
1837         enum changelog_send_flag         cs_flags;
1838         struct file                     *cs_fp;
1839         char                            *cs_buf;
1840         struct obd_device               *cs_obd;
1841 };
1842
1843 static inline char *cs_obd_name(struct changelog_show *cs)
1844 {
1845         return cs->cs_obd->obd_name;
1846 }
1847
1848 static int changelog_kkuc_cb(const struct lu_env *env, struct llog_handle *llh,
1849                              struct llog_rec_hdr *hdr, void *data)
1850 {
1851         struct changelog_show           *cs = data;
1852         struct llog_changelog_rec       *rec = (struct llog_changelog_rec *)hdr;
1853         struct kuc_hdr                  *lh;
1854         size_t                           len;
1855         int                              rc;
1856         ENTRY;
1857
1858         if (rec->cr_hdr.lrh_type != CHANGELOG_REC) {
1859                 rc = -EINVAL;
1860                 CERROR("%s: not a changelog rec %x/%d: rc = %d\n",
1861                        cs_obd_name(cs), rec->cr_hdr.lrh_type,
1862                        rec->cr.cr_type, rc);
1863                 RETURN(rc);
1864         }
1865
1866         if (rec->cr.cr_index < cs->cs_startrec) {
1867                 /* Skip entries earlier than what we are interested in */
1868                 CDEBUG(D_HSM, "rec=%llu start=%llu\n",
1869                        rec->cr.cr_index, cs->cs_startrec);
1870                 RETURN(0);
1871         }
1872
1873         CDEBUG(D_HSM, "%llu %02d%-5s %llu 0x%x t="DFID" p="DFID" %.*s\n",
1874                rec->cr.cr_index, rec->cr.cr_type,
1875                changelog_type2str(rec->cr.cr_type), rec->cr.cr_time,
1876                rec->cr.cr_flags & CLF_FLAGMASK,
1877                PFID(&rec->cr.cr_tfid), PFID(&rec->cr.cr_pfid),
1878                rec->cr.cr_namelen, changelog_rec_name(&rec->cr));
1879
1880         len = sizeof(*lh) + changelog_rec_size(&rec->cr) + rec->cr.cr_namelen;
1881
1882         /* Set up the message */
1883         lh = changelog_kuc_hdr(cs->cs_buf, len, cs->cs_flags);
1884         memcpy(lh + 1, &rec->cr, len - sizeof(*lh));
1885
1886         rc = libcfs_kkuc_msg_put(cs->cs_fp, lh);
1887         CDEBUG(D_HSM, "kucmsg fp %p len %zu rc %d\n", cs->cs_fp, len, rc);
1888
1889         RETURN(rc);
1890 }
1891
1892 static int mdc_changelog_send_thread(void *csdata)
1893 {
1894         struct changelog_show   *cs = csdata;
1895         struct llog_ctxt        *ctxt = NULL;
1896         struct llog_handle      *llh = NULL;
1897         struct kuc_hdr          *kuch;
1898         enum llog_flag           flags = LLOG_F_IS_CAT;
1899         int                      rc;
1900
1901         CDEBUG(D_HSM, "changelog to fp=%p start %llu\n",
1902                cs->cs_fp, cs->cs_startrec);
1903
1904         OBD_ALLOC(cs->cs_buf, KUC_CHANGELOG_MSG_MAXSIZE);
1905         if (cs->cs_buf == NULL)
1906                 GOTO(out, rc = -ENOMEM);
1907
1908         /* Set up the remote catalog handle */
1909         ctxt = llog_get_context(cs->cs_obd, LLOG_CHANGELOG_REPL_CTXT);
1910         if (ctxt == NULL)
1911                 GOTO(out, rc = -ENOENT);
1912         rc = llog_open(NULL, ctxt, &llh, NULL, CHANGELOG_CATALOG,
1913                        LLOG_OPEN_EXISTS);
1914         if (rc) {
1915                 CERROR("%s: fail to open changelog catalog: rc = %d\n",
1916                        cs_obd_name(cs), rc);
1917                 GOTO(out, rc);
1918         }
1919
1920         if (cs->cs_flags & CHANGELOG_FLAG_JOBID)
1921                 flags |= LLOG_F_EXT_JOBID;
1922
1923         rc = llog_init_handle(NULL, llh, flags, NULL);
1924         if (rc) {
1925                 CERROR("llog_init_handle failed %d\n", rc);
1926                 GOTO(out, rc);
1927         }
1928
1929         rc = llog_cat_process(NULL, llh, changelog_kkuc_cb, cs, 0, 0);
1930
1931         /* Send EOF no matter what our result */
1932         kuch = changelog_kuc_hdr(cs->cs_buf, sizeof(*kuch), cs->cs_flags);
1933         kuch->kuc_msgtype = CL_EOF;
1934         libcfs_kkuc_msg_put(cs->cs_fp, kuch);
1935
1936 out:
1937         fput(cs->cs_fp);
1938         if (llh)
1939                 llog_cat_close(NULL, llh);
1940         if (ctxt)
1941                 llog_ctxt_put(ctxt);
1942         if (cs->cs_buf)
1943                 OBD_FREE(cs->cs_buf, KUC_CHANGELOG_MSG_MAXSIZE);
1944         OBD_FREE_PTR(cs);
1945         return rc;
1946 }
1947
1948 static int mdc_ioc_changelog_send(struct obd_device *obd,
1949                                   struct ioc_changelog *icc)
1950 {
1951         struct changelog_show *cs;
1952         struct task_struct *task;
1953         int rc;
1954
1955         /* Freed in mdc_changelog_send_thread */
1956         OBD_ALLOC_PTR(cs);
1957         if (!cs)
1958                 return -ENOMEM;
1959
1960         cs->cs_obd = obd;
1961         cs->cs_startrec = icc->icc_recno;
1962         /* matching fput in mdc_changelog_send_thread */
1963         cs->cs_fp = fget(icc->icc_id);
1964         cs->cs_flags = icc->icc_flags;
1965
1966         /*
1967          * New thread because we should return to user app before
1968          * writing into our pipe
1969          */
1970         task = kthread_run(mdc_changelog_send_thread, cs,
1971                            "mdc_clg_send_thread");
1972         if (IS_ERR(task)) {
1973                 rc = PTR_ERR(task);
1974                 CERROR("%s: cannot start changelog thread: rc = %d\n",
1975                        cs_obd_name(cs), rc);
1976                 OBD_FREE_PTR(cs);
1977         } else {
1978                 rc = 0;
1979                 CDEBUG(D_HSM, "%s: started changelog thread\n",
1980                        cs_obd_name(cs));
1981         }
1982
1983         return rc;
1984 }
1985
1986 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
1987                                 struct lustre_kernelcomm *lk);
1988
1989 static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp,
1990                         struct obd_quotactl *oqctl)
1991 {
1992         struct ptlrpc_request   *req;
1993         struct obd_quotactl     *oqc;
1994         int                      rc;
1995         ENTRY;
1996
1997         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1998                                         &RQF_MDS_QUOTACTL, LUSTRE_MDS_VERSION,
1999                                         MDS_QUOTACTL);
2000         if (req == NULL)
2001                 RETURN(-ENOMEM);
2002
2003         oqc = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2004         *oqc = *oqctl;
2005
2006         ptlrpc_request_set_replen(req);
2007         ptlrpc_at_set_req_timeout(req);
2008         req->rq_no_resend = 1;
2009
2010         rc = ptlrpc_queue_wait(req);
2011         if (rc)
2012                 CERROR("ptlrpc_queue_wait failed, rc: %d\n", rc);
2013
2014         if (req->rq_repmsg &&
2015             (oqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL))) {
2016                 *oqctl = *oqc;
2017         } else if (!rc) {
2018                 CERROR ("Can't unpack obd_quotactl\n");
2019                 rc = -EPROTO;
2020         }
2021         ptlrpc_req_finished(req);
2022
2023         RETURN(rc);
2024 }
2025
2026 static int mdc_ioc_swap_layouts(struct obd_export *exp,
2027                                 struct md_op_data *op_data)
2028 {
2029         struct list_head cancels = LIST_HEAD_INIT(cancels);
2030         struct ptlrpc_request   *req;
2031         int                      rc, count;
2032         struct mdc_swap_layouts *msl, *payload;
2033         ENTRY;
2034
2035         msl = op_data->op_data;
2036
2037         /* When the MDT will get the MDS_SWAP_LAYOUTS RPC the
2038          * first thing it will do is to cancel the 2 layout
2039          * locks held by this client.
2040          * So the client must cancel its layout locks on the 2 fids
2041          * with the request RPC to avoid extra RPC round trips.
2042          */
2043         count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels,
2044                                         LCK_EX, MDS_INODELOCK_LAYOUT |
2045                                         MDS_INODELOCK_XATTR);
2046         count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels,
2047                                          LCK_EX, MDS_INODELOCK_LAYOUT |
2048                                          MDS_INODELOCK_XATTR);
2049
2050         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2051                                    &RQF_MDS_SWAP_LAYOUTS);
2052         if (req == NULL) {
2053                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
2054                 RETURN(-ENOMEM);
2055         }
2056
2057         rc = mdc_prep_elc_req(exp, req, MDS_SWAP_LAYOUTS, &cancels, count);
2058         if (rc) {
2059                 ptlrpc_request_free(req);
2060                 RETURN(rc);
2061         }
2062
2063         mdc_swap_layouts_pack(req, op_data);
2064
2065         payload = req_capsule_client_get(&req->rq_pill, &RMF_SWAP_LAYOUTS);
2066         LASSERT(payload);
2067
2068         *payload = *msl;
2069
2070         ptlrpc_request_set_replen(req);
2071
2072         rc = ptlrpc_queue_wait(req);
2073         if (rc)
2074                 GOTO(out, rc);
2075         EXIT;
2076
2077 out:
2078         ptlrpc_req_finished(req);
2079         return rc;
2080 }
2081
2082 static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2083                          void *karg, void __user *uarg)
2084 {
2085         struct obd_device *obd = exp->exp_obd;
2086         struct obd_ioctl_data *data = karg;
2087         struct obd_import *imp = obd->u.cli.cl_import;
2088         int rc;
2089         ENTRY;
2090
2091         if (!try_module_get(THIS_MODULE)) {
2092                 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2093                        module_name(THIS_MODULE));
2094                 return -EINVAL;
2095         }
2096         switch (cmd) {
2097         case OBD_IOC_CHANGELOG_SEND:
2098                 rc = mdc_ioc_changelog_send(obd, karg);
2099                 GOTO(out, rc);
2100         case OBD_IOC_CHANGELOG_CLEAR: {
2101                 struct ioc_changelog *icc = karg;
2102                 struct changelog_setinfo cs =
2103                         {.cs_recno = icc->icc_recno, .cs_id = icc->icc_id};
2104                 rc = obd_set_info_async(NULL, exp, strlen(KEY_CHANGELOG_CLEAR),
2105                                         KEY_CHANGELOG_CLEAR, sizeof(cs), &cs,
2106                                         NULL);
2107                 GOTO(out, rc);
2108         }
2109         case OBD_IOC_FID2PATH:
2110                 rc = mdc_ioc_fid2path(exp, karg);
2111                 GOTO(out, rc);
2112         case LL_IOC_HSM_CT_START:
2113                 rc = mdc_ioc_hsm_ct_start(exp, karg);
2114                 /* ignore if it was already registered on this MDS. */
2115                 if (rc == -EEXIST)
2116                         rc = 0;
2117                 GOTO(out, rc);
2118         case LL_IOC_HSM_PROGRESS:
2119                 rc = mdc_ioc_hsm_progress(exp, karg);
2120                 GOTO(out, rc);
2121         case LL_IOC_HSM_STATE_GET:
2122                 rc = mdc_ioc_hsm_state_get(exp, karg);
2123                 GOTO(out, rc);
2124         case LL_IOC_HSM_STATE_SET:
2125                 rc = mdc_ioc_hsm_state_set(exp, karg);
2126                 GOTO(out, rc);
2127         case LL_IOC_HSM_ACTION:
2128                 rc = mdc_ioc_hsm_current_action(exp, karg);
2129                 GOTO(out, rc);
2130         case LL_IOC_HSM_REQUEST:
2131                 rc = mdc_ioc_hsm_request(exp, karg);
2132                 GOTO(out, rc);
2133         case OBD_IOC_CLIENT_RECOVER:
2134                 rc = ptlrpc_recover_import(imp, data->ioc_inlbuf1, 0);
2135                 if (rc < 0)
2136                         GOTO(out, rc);
2137                 GOTO(out, rc = 0);
2138         case IOC_OSC_SET_ACTIVE:
2139                 rc = ptlrpc_set_import_active(imp, data->ioc_offset);
2140                 GOTO(out, rc);
2141         case OBD_IOC_PING_TARGET:
2142                 rc = ptlrpc_obd_ping(obd);
2143                 GOTO(out, rc);
2144         /*
2145          * Normally IOC_OBD_STATFS, OBD_IOC_QUOTACTL iocontrol are handled by
2146          * LMV instead of MDC. But when the cluster is upgraded from 1.8,
2147          * there'd be no LMV layer thus we might be called here. Eventually
2148          * this code should be removed.
2149          * bz20731, LU-592.
2150          */
2151         case IOC_OBD_STATFS: {
2152                 struct obd_statfs stat_buf = {0};
2153
2154                 if (*((__u32 *) data->ioc_inlbuf2) != 0)
2155                         GOTO(out, rc = -ENODEV);
2156
2157                 /* copy UUID */
2158                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(obd),
2159                                  min((int)data->ioc_plen2,
2160                                      (int)sizeof(struct obd_uuid))))
2161                         GOTO(out, rc = -EFAULT);
2162
2163                 rc = mdc_statfs(NULL, obd->obd_self_export, &stat_buf,
2164                                 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
2165                                 0);
2166                 if (rc != 0)
2167                         GOTO(out, rc);
2168
2169                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
2170                                      min((int) data->ioc_plen1,
2171                                          (int) sizeof(stat_buf))))
2172                         GOTO(out, rc = -EFAULT);
2173
2174                 GOTO(out, rc = 0);
2175         }
2176         case OBD_IOC_QUOTACTL: {
2177                 struct if_quotactl *qctl = karg;
2178                 struct obd_quotactl *oqctl;
2179
2180                 OBD_ALLOC_PTR(oqctl);
2181                 if (oqctl == NULL)
2182                         GOTO(out, rc = -ENOMEM);
2183
2184                 QCTL_COPY(oqctl, qctl);
2185                 rc = obd_quotactl(exp, oqctl);
2186                 if (rc == 0) {
2187                         QCTL_COPY(qctl, oqctl);
2188                         qctl->qc_valid = QC_MDTIDX;
2189                         qctl->obd_uuid = obd->u.cli.cl_target_uuid;
2190                 }
2191
2192                 OBD_FREE_PTR(oqctl);
2193                 GOTO(out, rc);
2194         }
2195         case LL_IOC_GET_CONNECT_FLAGS:
2196                 if (copy_to_user(uarg, exp_connect_flags_ptr(exp),
2197                                  sizeof(*exp_connect_flags_ptr(exp))))
2198                         GOTO(out, rc = -EFAULT);
2199
2200                 GOTO(out, rc = 0);
2201         case LL_IOC_LOV_SWAP_LAYOUTS:
2202                 rc = mdc_ioc_swap_layouts(exp, karg);
2203                 GOTO(out, rc);
2204         default:
2205                 CERROR("unrecognised ioctl: cmd = %#x\n", cmd);
2206                 GOTO(out, rc = -ENOTTY);
2207         }
2208 out:
2209         module_put(THIS_MODULE);
2210
2211         return rc;
2212 }
2213
2214 static int mdc_get_info_rpc(struct obd_export *exp,
2215                             u32 keylen, void *key,
2216                             u32 vallen, void *val)
2217 {
2218         struct obd_import      *imp = class_exp2cliimp(exp);
2219         struct ptlrpc_request  *req;
2220         char                   *tmp;
2221         int                     rc = -EINVAL;
2222         ENTRY;
2223
2224         req = ptlrpc_request_alloc(imp, &RQF_MDS_GET_INFO);
2225         if (req == NULL)
2226                 RETURN(-ENOMEM);
2227
2228         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY,
2229                              RCL_CLIENT, keylen);
2230         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VALLEN,
2231                              RCL_CLIENT, sizeof(vallen));
2232
2233         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_INFO);
2234         if (rc) {
2235                 ptlrpc_request_free(req);
2236                 RETURN(rc);
2237         }
2238
2239         tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
2240         memcpy(tmp, key, keylen);
2241         tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_VALLEN);
2242         memcpy(tmp, &vallen, sizeof(vallen));
2243
2244         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VAL,
2245                              RCL_SERVER, vallen);
2246         ptlrpc_request_set_replen(req);
2247
2248         rc = ptlrpc_queue_wait(req);
2249         /* -EREMOTE means the get_info result is partial, and it needs to
2250          * continue on another MDT, see fid2path part in lmv_iocontrol */
2251         if (rc == 0 || rc == -EREMOTE) {
2252                 tmp = req_capsule_server_get(&req->rq_pill, &RMF_GETINFO_VAL);
2253                 memcpy(val, tmp, vallen);
2254                 if (ptlrpc_rep_need_swab(req)) {
2255                         if (KEY_IS(KEY_FID2PATH))
2256                                 lustre_swab_fid2path(val);
2257                 }
2258         }
2259         ptlrpc_req_finished(req);
2260
2261         RETURN(rc);
2262 }
2263
2264 static void lustre_swab_hai(struct hsm_action_item *h)
2265 {
2266         __swab32s(&h->hai_len);
2267         __swab32s(&h->hai_action);
2268         lustre_swab_lu_fid(&h->hai_fid);
2269         lustre_swab_lu_fid(&h->hai_dfid);
2270         __swab64s(&h->hai_cookie);
2271         __swab64s(&h->hai_extent.offset);
2272         __swab64s(&h->hai_extent.length);
2273         __swab64s(&h->hai_gid);
2274 }
2275
2276 static void lustre_swab_hal(struct hsm_action_list *h)
2277 {
2278         struct hsm_action_item  *hai;
2279         __u32                    i;
2280
2281         __swab32s(&h->hal_version);
2282         __swab32s(&h->hal_count);
2283         __swab32s(&h->hal_archive_id);
2284         __swab64s(&h->hal_flags);
2285         hai = hai_first(h);
2286         for (i = 0; i < h->hal_count; i++, hai = hai_next(hai))
2287                 lustre_swab_hai(hai);
2288 }
2289
2290 static void lustre_swab_kuch(struct kuc_hdr *l)
2291 {
2292         __swab16s(&l->kuc_magic);
2293         /* __u8 l->kuc_transport */
2294         __swab16s(&l->kuc_msgtype);
2295         __swab16s(&l->kuc_msglen);
2296 }
2297
2298 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2299                                 struct lustre_kernelcomm *lk)
2300 {
2301         struct obd_import  *imp = class_exp2cliimp(exp);
2302         __u32               archive = lk->lk_data;
2303         int                 rc = 0;
2304
2305         if (lk->lk_group != KUC_GRP_HSM) {
2306                 CERROR("Bad copytool group %d\n", lk->lk_group);
2307                 return -EINVAL;
2308         }
2309
2310         CDEBUG(D_HSM, "CT start r%d w%d u%d g%d f%#x\n", lk->lk_rfd, lk->lk_wfd,
2311                lk->lk_uid, lk->lk_group, lk->lk_flags);
2312
2313         if (lk->lk_flags & LK_FLG_STOP) {
2314                 /* Unregister with the coordinator */
2315                 rc = mdc_ioc_hsm_ct_unregister(imp);
2316         } else {
2317                 rc = mdc_ioc_hsm_ct_register(imp, archive);
2318         }
2319
2320         return rc;
2321 }
2322
2323 /**
2324  * Send a message to any listening copytools
2325  * @param val KUC message (kuc_hdr + hsm_action_list)
2326  * @param len total length of message
2327  */
2328 static int mdc_hsm_copytool_send(size_t len, void *val)
2329 {
2330         struct kuc_hdr          *lh = (struct kuc_hdr *)val;
2331         struct hsm_action_list  *hal = (struct hsm_action_list *)(lh + 1);
2332         int                      rc;
2333         ENTRY;
2334
2335         if (len < sizeof(*lh) + sizeof(*hal)) {
2336                 CERROR("Short HSM message %zu < %zu\n", len,
2337                        sizeof(*lh) + sizeof(*hal));
2338                 RETURN(-EPROTO);
2339         }
2340         if (lh->kuc_magic == __swab16(KUC_MAGIC)) {
2341                 lustre_swab_kuch(lh);
2342                 lustre_swab_hal(hal);
2343         } else if (lh->kuc_magic != KUC_MAGIC) {
2344                 CERROR("Bad magic %x!=%x\n", lh->kuc_magic, KUC_MAGIC);
2345                 RETURN(-EPROTO);
2346         }
2347
2348         CDEBUG(D_HSM, " Received message mg=%x t=%d m=%d l=%d actions=%d "
2349                "on %s\n",
2350                lh->kuc_magic, lh->kuc_transport, lh->kuc_msgtype,
2351                lh->kuc_msglen, hal->hal_count, hal->hal_fsname);
2352
2353         /* Broadcast to HSM listeners */
2354         rc = libcfs_kkuc_group_put(KUC_GRP_HSM, lh);
2355
2356         RETURN(rc);
2357 }
2358
2359 /**
2360  * callback function passed to kuc for re-registering each HSM copytool
2361  * running on MDC, after MDT shutdown/recovery.
2362  * @param data copytool registration data
2363  * @param cb_arg callback argument (obd_import)
2364  */
2365 static int mdc_hsm_ct_reregister(void *data, void *cb_arg)
2366 {
2367         struct kkuc_ct_data     *kcd = data;
2368         struct obd_import       *imp = (struct obd_import *)cb_arg;
2369         int                      rc;
2370
2371         if (kcd == NULL || kcd->kcd_magic != KKUC_CT_DATA_MAGIC)
2372                 return -EPROTO;
2373
2374         if (!obd_uuid_equals(&kcd->kcd_uuid, &imp->imp_obd->obd_uuid))
2375                 return 0;
2376
2377         CDEBUG(D_HA, "%s: recover copytool registration to MDT (archive=%#x)\n",
2378                imp->imp_obd->obd_name, kcd->kcd_archive);
2379         rc = mdc_ioc_hsm_ct_register(imp, kcd->kcd_archive);
2380
2381         /* ignore error if the copytool is already registered */
2382         return (rc == -EEXIST) ? 0 : rc;
2383 }
2384
2385 /**
2386  * Re-establish all kuc contexts with MDT
2387  * after MDT shutdown/recovery.
2388  */
2389 static int mdc_kuc_reregister(struct obd_import *imp)
2390 {
2391         /* re-register HSM agents */
2392         return libcfs_kkuc_group_foreach(KUC_GRP_HSM, mdc_hsm_ct_reregister,
2393                                          (void *)imp);
2394 }
2395
2396 static int mdc_set_info_async(const struct lu_env *env,
2397                               struct obd_export *exp,
2398                               u32 keylen, void *key,
2399                               u32 vallen, void *val,
2400                               struct ptlrpc_request_set *set)
2401 {
2402         struct obd_import       *imp = class_exp2cliimp(exp);
2403         int                      rc;
2404         ENTRY;
2405
2406         if (KEY_IS(KEY_READ_ONLY)) {
2407                 if (vallen != sizeof(int))
2408                         RETURN(-EINVAL);
2409
2410                 spin_lock(&imp->imp_lock);
2411                 if (*((int *)val)) {
2412                         imp->imp_connect_flags_orig |= OBD_CONNECT_RDONLY;
2413                         imp->imp_connect_data.ocd_connect_flags |=
2414                                                         OBD_CONNECT_RDONLY;
2415                 } else {
2416                         imp->imp_connect_flags_orig &= ~OBD_CONNECT_RDONLY;
2417                         imp->imp_connect_data.ocd_connect_flags &=
2418                                                         ~OBD_CONNECT_RDONLY;
2419                 }
2420                 spin_unlock(&imp->imp_lock);
2421
2422                 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2423                                        keylen, key, vallen, val, set);
2424                 RETURN(rc);
2425         }
2426         if (KEY_IS(KEY_SPTLRPC_CONF)) {
2427                 sptlrpc_conf_client_adapt(exp->exp_obd);
2428                 RETURN(0);
2429         }
2430         if (KEY_IS(KEY_FLUSH_CTX)) {
2431                 sptlrpc_import_flush_my_ctx(imp);
2432                 RETURN(0);
2433         }
2434         if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2435                 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2436                                        keylen, key, vallen, val, set);
2437                 RETURN(rc);
2438         }
2439         if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) {
2440                 rc = mdc_hsm_copytool_send(vallen, val);
2441                 RETURN(rc);
2442         }
2443
2444         if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2445                 __u32 *default_easize = val;
2446
2447                 exp->exp_obd->u.cli.cl_default_mds_easize = *default_easize;
2448                 RETURN(0);
2449         }
2450
2451         CERROR("Unknown key %s\n", (char *)key);
2452         RETURN(-EINVAL);
2453 }
2454
2455 static int mdc_get_info(const struct lu_env *env, struct obd_export *exp,
2456                         __u32 keylen, void *key, __u32 *vallen, void *val)
2457 {
2458         int rc = -EINVAL;
2459
2460         if (KEY_IS(KEY_MAX_EASIZE)) {
2461                 __u32 mdsize, *max_easize;
2462
2463                 if (*vallen != sizeof(int))
2464                         RETURN(-EINVAL);
2465                 mdsize = *(__u32 *)val;
2466                 if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
2467                         exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
2468                 max_easize = val;
2469                 *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
2470                 RETURN(0);
2471         } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2472                 __u32 *default_easize;
2473
2474                 if (*vallen != sizeof(int))
2475                         RETURN(-EINVAL);
2476                 default_easize = val;
2477                 *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
2478                 RETURN(0);
2479         } else if (KEY_IS(KEY_CONN_DATA)) {
2480                 struct obd_import *imp = class_exp2cliimp(exp);
2481                 struct obd_connect_data *data = val;
2482
2483                 if (*vallen != sizeof(*data))
2484                         RETURN(-EINVAL);
2485
2486                 *data = imp->imp_connect_data;
2487                 RETURN(0);
2488         } else if (KEY_IS(KEY_TGT_COUNT)) {
2489                 *((__u32 *)val) = 1;
2490                 RETURN(0);
2491         }
2492
2493         rc = mdc_get_info_rpc(exp, keylen, key, *vallen, val);
2494
2495         RETURN(rc);
2496 }
2497
2498 static int mdc_fsync(struct obd_export *exp, const struct lu_fid *fid,
2499                      struct ptlrpc_request **request)
2500 {
2501         struct ptlrpc_request *req;
2502         int                    rc;
2503         ENTRY;
2504
2505         *request = NULL;
2506         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SYNC);
2507         if (req == NULL)
2508                 RETURN(-ENOMEM);
2509
2510         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SYNC);
2511         if (rc) {
2512                 ptlrpc_request_free(req);
2513                 RETURN(rc);
2514         }
2515
2516         mdc_pack_body(req, fid, 0, 0, -1, 0);
2517
2518         ptlrpc_request_set_replen(req);
2519
2520         rc = ptlrpc_queue_wait(req);
2521         if (rc)
2522                 ptlrpc_req_finished(req);
2523         else
2524                 *request = req;
2525         RETURN(rc);
2526 }
2527
2528 static int mdc_import_event(struct obd_device *obd, struct obd_import *imp,
2529                             enum obd_import_event event)
2530 {
2531         int rc = 0;
2532
2533         LASSERT(imp->imp_obd == obd);
2534
2535         switch (event) {
2536
2537         case IMP_EVENT_INACTIVE: {
2538                 struct client_obd *cli = &obd->u.cli;
2539                 /*
2540                  * Flush current sequence to make client obtain new one
2541                  * from server in case of disconnect/reconnect.
2542                  */
2543                 down_read(&cli->cl_seq_rwsem);
2544                 if (cli->cl_seq)
2545                         seq_client_flush(cli->cl_seq);
2546                 up_read(&cli->cl_seq_rwsem);
2547
2548                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
2549                 break;
2550         }
2551         case IMP_EVENT_INVALIDATE: {
2552                 struct ldlm_namespace *ns = obd->obd_namespace;
2553
2554                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2555
2556                 break;
2557         }
2558         case IMP_EVENT_ACTIVE:
2559                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
2560                 /* redo the kuc registration after reconnecting */
2561                 if (rc == 0)
2562                         rc = mdc_kuc_reregister(imp);
2563                 break;
2564         case IMP_EVENT_OCD:
2565                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
2566                 break;
2567         case IMP_EVENT_DISCON:
2568         case IMP_EVENT_DEACTIVATE:
2569         case IMP_EVENT_ACTIVATE:
2570                 break;
2571         default:
2572                 CERROR("Unknown import event %x\n", event);
2573                 LBUG();
2574         }
2575         RETURN(rc);
2576 }
2577
2578 int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp,
2579                   struct lu_fid *fid, struct md_op_data *op_data)
2580 {
2581         struct client_obd *cli = &exp->exp_obd->u.cli;
2582         int rc = -EIO;
2583
2584         ENTRY;
2585
2586         down_read(&cli->cl_seq_rwsem);
2587         if (cli->cl_seq)
2588                 rc = seq_client_alloc_fid(env, cli->cl_seq, fid);
2589         up_read(&cli->cl_seq_rwsem);
2590
2591         RETURN(rc);
2592 }
2593
2594 static struct obd_uuid *mdc_get_uuid(struct obd_export *exp)
2595 {
2596         struct client_obd *cli = &exp->exp_obd->u.cli;
2597         return &cli->cl_target_uuid;
2598 }
2599
2600 /**
2601  * Determine whether the lock can be canceled before replaying it during
2602  * recovery, non zero value will be return if the lock can be canceled,
2603  * or zero returned for not
2604  */
2605 static int mdc_cancel_weight(struct ldlm_lock *lock)
2606 {
2607         if (lock->l_resource->lr_type != LDLM_IBITS)
2608                 RETURN(0);
2609
2610         /* FIXME: if we ever get into a situation where there are too many
2611          * opened files with open locks on a single node, then we really
2612          * should replay these open locks to reget it */
2613         if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN)
2614                 RETURN(0);
2615
2616         RETURN(1);
2617 }
2618
2619 static int mdc_resource_inode_free(struct ldlm_resource *res)
2620 {
2621         if (res->lr_lvb_inode)
2622                 res->lr_lvb_inode = NULL;
2623
2624         return 0;
2625 }
2626
2627 static struct ldlm_valblock_ops inode_lvbo = {
2628         .lvbo_free = mdc_resource_inode_free
2629 };
2630
2631 static int mdc_llog_init(struct obd_device *obd)
2632 {
2633         struct obd_llog_group   *olg = &obd->obd_olg;
2634         struct llog_ctxt        *ctxt;
2635         int                      rc;
2636
2637         ENTRY;
2638
2639         rc = llog_setup(NULL, obd, olg, LLOG_CHANGELOG_REPL_CTXT, obd,
2640                         &llog_client_ops);
2641         if (rc < 0)
2642                 RETURN(rc);
2643
2644         ctxt = llog_group_get_ctxt(olg, LLOG_CHANGELOG_REPL_CTXT);
2645         llog_initiator_connect(ctxt);
2646         llog_ctxt_put(ctxt);
2647
2648         RETURN(0);
2649 }
2650
2651 static void mdc_llog_finish(struct obd_device *obd)
2652 {
2653         struct llog_ctxt *ctxt;
2654
2655         ENTRY;
2656
2657         ctxt = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
2658         if (ctxt != NULL)
2659                 llog_cleanup(NULL, ctxt);
2660
2661         EXIT;
2662 }
2663
2664 static int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg)
2665 {
2666         int                             rc;
2667         ENTRY;
2668
2669         rc = ptlrpcd_addref();
2670         if (rc < 0)
2671                 RETURN(rc);
2672
2673         rc = client_obd_setup(obd, cfg);
2674         if (rc)
2675                 GOTO(err_ptlrpcd_decref, rc);
2676 #ifdef CONFIG_PROC_FS
2677         obd->obd_vars = lprocfs_mdc_obd_vars;
2678         lprocfs_obd_setup(obd);
2679         lprocfs_alloc_md_stats(obd, 0);
2680 #endif
2681         sptlrpc_lprocfs_cliobd_attach(obd);
2682         ptlrpc_lprocfs_register_obd(obd);
2683
2684         ns_register_cancel(obd->obd_namespace, mdc_cancel_weight);
2685
2686         obd->obd_namespace->ns_lvbo = &inode_lvbo;
2687
2688         rc = mdc_llog_init(obd);
2689         if (rc) {
2690                 mdc_cleanup(obd);
2691                 CERROR("failed to setup llogging subsystems\n");
2692                 RETURN(rc);
2693         }
2694
2695         RETURN(rc);
2696
2697 err_ptlrpcd_decref:
2698         ptlrpcd_decref();
2699         RETURN(rc);
2700 }
2701
2702 /* Initialize the default and maximum LOV EA sizes.  This allows
2703  * us to make MDS RPCs with large enough reply buffers to hold a default
2704  * sized EA without having to calculate this (via a call into the
2705  * LOV + OSCs) each time we make an RPC.  The maximum size is also tracked
2706  * but not used to avoid wastefully vmalloc()'ing large reply buffers when
2707  * a large number of stripes is possible.  If a larger reply buffer is
2708  * required it will be reallocated in the ptlrpc layer due to overflow.
2709  */
2710 static int mdc_init_ea_size(struct obd_export *exp, __u32 easize,
2711                             __u32 def_easize)
2712 {
2713         struct obd_device *obd = exp->exp_obd;
2714         struct client_obd *cli = &obd->u.cli;
2715         ENTRY;
2716
2717         if (cli->cl_max_mds_easize < easize)
2718                 cli->cl_max_mds_easize = easize;
2719
2720         if (cli->cl_default_mds_easize < def_easize)
2721                 cli->cl_default_mds_easize = def_easize;
2722
2723         RETURN(0);
2724 }
2725
2726 static int mdc_precleanup(struct obd_device *obd)
2727 {
2728         ENTRY;
2729
2730         /* Failsafe, ok if racy */
2731         if (obd->obd_type->typ_refcnt <= 1)
2732                 libcfs_kkuc_group_rem(0, KUC_GRP_HSM);
2733
2734         obd_cleanup_client_import(obd);
2735         ptlrpc_lprocfs_unregister_obd(obd);
2736         lprocfs_obd_cleanup(obd);
2737         lprocfs_free_md_stats(obd);
2738         mdc_llog_finish(obd);
2739         RETURN(0);
2740 }
2741
2742 static int mdc_cleanup(struct obd_device *obd)
2743 {
2744         ptlrpcd_decref();
2745
2746         return client_obd_cleanup(obd);
2747 }
2748
2749 static int mdc_process_config(struct obd_device *obd, size_t len, void *buf)
2750 {
2751         struct lustre_cfg *lcfg = buf;
2752         int rc = class_process_proc_param(PARAM_MDC, obd->obd_vars, lcfg, obd);
2753         return (rc > 0 ? 0: rc);
2754 }
2755
2756 static struct obd_ops mdc_obd_ops = {
2757         .o_owner            = THIS_MODULE,
2758         .o_setup            = mdc_setup,
2759         .o_precleanup       = mdc_precleanup,
2760         .o_cleanup          = mdc_cleanup,
2761         .o_add_conn         = client_import_add_conn,
2762         .o_del_conn         = client_import_del_conn,
2763         .o_connect          = client_connect_import,
2764         .o_disconnect       = client_disconnect_export,
2765         .o_iocontrol        = mdc_iocontrol,
2766         .o_set_info_async   = mdc_set_info_async,
2767         .o_statfs           = mdc_statfs,
2768         .o_fid_init         = client_fid_init,
2769         .o_fid_fini         = client_fid_fini,
2770         .o_fid_alloc        = mdc_fid_alloc,
2771         .o_import_event     = mdc_import_event,
2772         .o_get_info         = mdc_get_info,
2773         .o_process_config   = mdc_process_config,
2774         .o_get_uuid         = mdc_get_uuid,
2775         .o_quotactl         = mdc_quotactl,
2776 };
2777
2778 static struct md_ops mdc_md_ops = {
2779         .m_get_root         = mdc_get_root,
2780         .m_null_inode       = mdc_null_inode,
2781         .m_close            = mdc_close,
2782         .m_create           = mdc_create,
2783         .m_enqueue          = mdc_enqueue,
2784         .m_getattr          = mdc_getattr,
2785         .m_getattr_name     = mdc_getattr_name,
2786         .m_intent_lock      = mdc_intent_lock,
2787         .m_link             = mdc_link,
2788         .m_rename           = mdc_rename,
2789         .m_setattr          = mdc_setattr,
2790         .m_setxattr         = mdc_setxattr,
2791         .m_getxattr         = mdc_getxattr,
2792         .m_fsync                = mdc_fsync,
2793         .m_read_page            = mdc_read_page,
2794         .m_unlink           = mdc_unlink,
2795         .m_cancel_unused    = mdc_cancel_unused,
2796         .m_init_ea_size     = mdc_init_ea_size,
2797         .m_set_lock_data    = mdc_set_lock_data,
2798         .m_lock_match       = mdc_lock_match,
2799         .m_get_lustre_md    = mdc_get_lustre_md,
2800         .m_free_lustre_md   = mdc_free_lustre_md,
2801         .m_set_open_replay_data = mdc_set_open_replay_data,
2802         .m_clear_open_replay_data = mdc_clear_open_replay_data,
2803         .m_intent_getattr_async = mdc_intent_getattr_async,
2804         .m_revalidate_lock      = mdc_revalidate_lock
2805 };
2806
2807 static int __init mdc_init(void)
2808 {
2809         return class_register_type(&mdc_obd_ops, &mdc_md_ops, true, NULL,
2810                                    LUSTRE_MDC_NAME, NULL);
2811 }
2812
2813 static void __exit mdc_exit(void)
2814 {
2815         class_unregister_type(LUSTRE_MDC_NAME);
2816 }
2817
2818 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2819 MODULE_DESCRIPTION("Lustre Metadata Client");
2820 MODULE_VERSION(LUSTRE_VERSION_STRING);
2821 MODULE_LICENSE("GPL");
2822
2823 module_init(mdc_init);
2824 module_exit(mdc_exit);