lustre/mdc/mdc_request.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 #define DEBUG_SUBSYSTEM S_MDC
  34
  35 #include <linux/init.h>
  36 #include <linux/kthread.h>
  37 #include <linux/module.h>
  38 #include <linux/pagemap.h>
  39 #include <linux/user_namespace.h>
  40 #include <linux/utsname.h>
  41 #include <linux/delay.h>
  42 #include <linux/uidgid.h>
  43
  44 #include <lustre_errno.h>
  45
  46 #include <cl_object.h>
  47 #include <llog_swab.h>
  48 #include <lprocfs_status.h>
  49 #include <lustre_acl.h>
  50 #include <lustre_fid.h>
  51 #include <uapi/linux/lustre/lustre_ioctl.h>
  52 #include <lustre_kernelcomm.h>
  53 #include <lustre_lmv.h>
  54 #include <lustre_log.h>
  55 #include <lustre_swab.h>
  56 #include <obd_class.h>
  57 #include <lustre_osc.h>
  58
  59 #include "mdc_internal.h"
  60
  61 #define REQUEST_MINOR 244
  62
  63 static int mdc_cleanup(struct obd_device *obd);
  64
  65 static inline int mdc_queue_wait(struct ptlrpc_request *req)
  66 {
  67         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
  68         int rc;
  69
  70         /* obd_get_request_slot() ensures that this client has no more
  71          * than cl_max_rpcs_in_flight RPCs simultaneously inf light
  72          * against an MDT. */
  73         rc = obd_get_request_slot(cli);
  74         if (rc != 0)
  75                 return rc;
  76
  77         rc = ptlrpc_queue_wait(req);
  78         obd_put_request_slot(cli);
  79
  80         return rc;
  81 }
  82
  83 /*
  84  * Send MDS_GET_ROOT RPC to fetch root FID.
  85  *
  86  * If \a fileset is not NULL it should contain a subdirectory off
  87  * the ROOT/ directory to be mounted on the client. Return the FID
  88  * of the subdirectory to the client to mount onto its mountpoint.
  89  *
  90  * \param[in]   imp     MDC import
  91  * \param[in]   fileset fileset name, which could be NULL
  92  * \param[out]  rootfid root FID of this mountpoint
  93  * \param[out]  pc      root capa will be unpacked and saved in this pointer
  94  *
  95  * \retval      0 on success, negative errno on failure
  96  */
  97 static int mdc_get_root(struct obd_export *exp, const char *fileset,
  98                          struct lu_fid *rootfid)
  99 {
 100         struct ptlrpc_request   *req;
 101         struct mdt_body         *body;
 102         int                      rc;
 103
 104         ENTRY;
 105
 106         if (fileset && !(exp_connect_flags(exp) & OBD_CONNECT_SUBTREE))
 107                 RETURN(-ENOTSUPP);
 108
 109         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 110                                 &RQF_MDS_GET_ROOT);
 111         if (req == NULL)
 112                 RETURN(-ENOMEM);
 113
 114         if (fileset != NULL)
 115                 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 116                                      strlen(fileset) + 1);
 117         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_ROOT);
 118         if (rc) {
 119                 ptlrpc_request_free(req);
 120                 RETURN(rc);
 121         }
 122         mdc_pack_body(req, NULL, 0, 0, -1, 0);
 123         if (fileset != NULL) {
 124                 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
 125
 126                 memcpy(name, fileset, strlen(fileset));
 127         }
 128         lustre_msg_add_flags(req->rq_reqmsg, LUSTRE_IMP_FULL);
 129         req->rq_send_state = LUSTRE_IMP_FULL;
 130
 131         ptlrpc_request_set_replen(req);
 132
 133         rc = ptlrpc_queue_wait(req);
 134         if (rc)
 135                 GOTO(out, rc);
 136
 137         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 138         if (body == NULL)
 139                 GOTO(out, rc = -EPROTO);
 140
 141         *rootfid = body->mbo_fid1;
 142         CDEBUG(D_NET, "root fid="DFID", last_committed=%llu\n",
 143                PFID(rootfid), lustre_msg_get_last_committed(req->rq_repmsg));
 144         EXIT;
 145 out:
 146         ptlrpc_req_finished(req);
 147
 148         return rc;
 149 }
 150
 151 /*
 152  * This function now is known to always saying that it will receive 4 buffers
 153  * from server. Even for cases when acl_size and md_size is zero, RPC header
 154  * will contain 4 fields and RPC itself will contain zero size fields. This is
 155  * because mdt_getattr*() _always_ returns 4 fields, but if acl is not needed
 156  * and thus zero, it shrinks it, making zero size. The same story about
 157  * md_size. And this is course of problem when client waits for smaller number
 158  * of fields. This issue will be fixed later when client gets aware of RPC
 159  * layouts.  --umka
 160  */
 161 static int mdc_getattr_common(struct obd_export *exp,
 162                               struct ptlrpc_request *req)
 163 {
 164         struct req_capsule *pill = &req->rq_pill;
 165         struct mdt_body    *body;
 166         void               *eadata;
 167         int                 rc;
 168         ENTRY;
 169
 170         /* Request message already built. */
 171         rc = ptlrpc_queue_wait(req);
 172         if (rc != 0)
 173                 RETURN(rc);
 174
 175         /* sanity check for the reply */
 176         body = req_capsule_server_get(pill, &RMF_MDT_BODY);
 177         if (body == NULL)
 178                 RETURN(-EPROTO);
 179
 180         CDEBUG(D_NET, "mode: %o\n", body->mbo_mode);
 181
 182         mdc_update_max_ea_from_body(exp, body);
 183         if (body->mbo_eadatasize != 0) {
 184                 eadata = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
 185                                                       body->mbo_eadatasize);
 186                 if (eadata == NULL)
 187                         RETURN(-EPROTO);
 188         }
 189
 190         RETURN(0);
 191 }
 192
 193 static void mdc_reset_acl_req(struct ptlrpc_request *req)
 194 {
 195         spin_lock(&req->rq_early_free_lock);
 196         sptlrpc_cli_free_repbuf(req);
 197         req->rq_repbuf = NULL;
 198         req->rq_repbuf_len = 0;
 199         req->rq_repdata = NULL;
 200         req->rq_reqdata_len = 0;
 201         spin_unlock(&req->rq_early_free_lock);
 202 }
 203
 204 static int mdc_getattr(struct obd_export *exp, struct md_op_data *op_data,
 205                        struct ptlrpc_request **request)
 206 {
 207         struct ptlrpc_request *req;
 208         struct obd_import *imp = class_exp2cliimp(exp);
 209         __u32 acl_bufsize = LUSTRE_POSIX_ACL_MAX_SIZE_OLD;
 210         int rc;
 211         ENTRY;
 212
 213         /* Single MDS without an LMV case */
 214         if (op_data->op_flags & MF_GET_MDT_IDX) {
 215                 op_data->op_mds = 0;
 216                 RETURN(0);
 217         }
 218
 219         *request = NULL;
 220         req = ptlrpc_request_alloc(imp, &RQF_MDS_GETATTR);
 221         if (req == NULL)
 222                 RETURN(-ENOMEM);
 223
 224         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR);
 225         if (rc) {
 226                 ptlrpc_request_free(req);
 227                 RETURN(rc);
 228         }
 229
 230 again:
 231         mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
 232                       op_data->op_mode, -1, 0);
 233         req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
 234         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 235                              op_data->op_mode);
 236         ptlrpc_request_set_replen(req);
 237
 238         rc = mdc_getattr_common(exp, req);
 239         if (rc) {
 240                 if (rc == -ERANGE) {
 241                         acl_bufsize = min_t(__u32,
 242                                             imp->imp_connect_data.ocd_max_easize,
 243                                             XATTR_SIZE_MAX);
 244                         mdc_reset_acl_req(req);
 245                         goto again;
 246                 }
 247
 248                 ptlrpc_req_finished(req);
 249         } else {
 250                 *request = req;
 251         }
 252
 253         RETURN(rc);
 254 }
 255
 256 static int mdc_getattr_name(struct obd_export *exp, struct md_op_data *op_data,
 257                             struct ptlrpc_request **request)
 258 {
 259         struct ptlrpc_request *req;
 260         struct obd_import *imp = class_exp2cliimp(exp);
 261         __u32 acl_bufsize = LUSTRE_POSIX_ACL_MAX_SIZE_OLD;
 262         int rc;
 263         ENTRY;
 264
 265         *request = NULL;
 266         req = ptlrpc_request_alloc(imp, &RQF_MDS_GETATTR_NAME);
 267         if (req == NULL)
 268                 RETURN(-ENOMEM);
 269
 270         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 271                              op_data->op_namelen + 1);
 272
 273         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR_NAME);
 274         if (rc) {
 275                 ptlrpc_request_free(req);
 276                 RETURN(rc);
 277         }
 278
 279         if (op_data->op_name) {
 280                 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
 281                 LASSERT(strnlen(op_data->op_name, op_data->op_namelen) ==
 282                                 op_data->op_namelen);
 283                 memcpy(name, op_data->op_name, op_data->op_namelen);
 284         }
 285
 286 again:
 287         mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
 288                       op_data->op_mode, op_data->op_suppgids[0], 0);
 289         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 290                              op_data->op_mode);
 291         req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
 292         ptlrpc_request_set_replen(req);
 293
 294         rc = mdc_getattr_common(exp, req);
 295         if (rc) {
 296                 if (rc == -ERANGE) {
 297                         acl_bufsize = min_t(__u32,
 298                                             imp->imp_connect_data.ocd_max_easize,
 299                                             XATTR_SIZE_MAX);
 300                         mdc_reset_acl_req(req);
 301                         goto again;
 302                 }
 303
 304                 ptlrpc_req_finished(req);
 305         } else {
 306                 *request = req;
 307         }
 308
 309         RETURN(rc);
 310 }
 311
 312 static int mdc_xattr_common(struct obd_export *exp,const struct req_format *fmt,
 313                             const struct lu_fid *fid, int opcode, u64 valid,
 314                             const char *xattr_name, const char *input,
 315                             int input_size, int output_size, int flags,
 316                             __u32 suppgid, struct ptlrpc_request **request)
 317 {
 318         struct ptlrpc_request *req;
 319         int   xattr_namelen = 0;
 320         char *tmp;
 321         int   rc;
 322         ENTRY;
 323
 324         *request = NULL;
 325         req = ptlrpc_request_alloc(class_exp2cliimp(exp), fmt);
 326         if (req == NULL)
 327                 RETURN(-ENOMEM);
 328
 329         if (xattr_name) {
 330                 xattr_namelen = strlen(xattr_name) + 1;
 331                 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 332                                      xattr_namelen);
 333         }
 334         if (input_size)
 335                 LASSERT(input);
 336         req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT,
 337                              input_size);
 338
 339         /* get SELinux policy info if any */
 340         rc = sptlrpc_get_sepol(req);
 341         if (rc < 0) {
 342                 ptlrpc_request_free(req);
 343                 RETURN(rc);
 344         }
 345         req_capsule_set_size(&req->rq_pill, &RMF_SELINUX_POL, RCL_CLIENT,
 346                              strlen(req->rq_sepol) ?
 347                              strlen(req->rq_sepol) + 1 : 0);
 348
 349         /* Flush local XATTR locks to get rid of a possible cancel RPC */
 350         if (opcode == MDS_REINT && fid_is_sane(fid) &&
 351             exp->exp_connect_data.ocd_ibits_known & MDS_INODELOCK_XATTR) {
 352                 LIST_HEAD(cancels);
 353                 int count;
 354
 355                 /* Without that packing would fail */
 356                 if (input_size == 0)
 357                         req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
 358                                              RCL_CLIENT, 0);
 359
 360                 count = mdc_resource_get_unused(exp, fid,
 361                                                 &cancels, LCK_EX,
 362                                                 MDS_INODELOCK_XATTR);
 363
 364                 rc = mdc_prep_elc_req(exp, req, MDS_REINT, &cancels, count);
 365                 if (rc) {
 366                         ptlrpc_request_free(req);
 367                         RETURN(rc);
 368                 }
 369         } else {
 370                 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, opcode);
 371                 if (rc) {
 372                         ptlrpc_request_free(req);
 373                         RETURN(rc);
 374                 }
 375         }
 376
 377         if (opcode == MDS_REINT) {
 378                 struct mdt_rec_setxattr *rec;
 379
 380                 BUILD_BUG_ON(sizeof(struct mdt_rec_setxattr) !=
 381                              sizeof(struct mdt_rec_reint));
 382                 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
 383                 rec->sx_opcode = REINT_SETXATTR;
 384                 rec->sx_fsuid  = from_kuid(&init_user_ns, current_fsuid());
 385                 rec->sx_fsgid  = from_kgid(&init_user_ns, current_fsgid());
 386                 rec->sx_cap    = cfs_curproc_cap_pack();
 387                 rec->sx_suppgid1 = suppgid;
 388                 rec->sx_suppgid2 = -1;
 389                 rec->sx_fid    = *fid;
 390                 rec->sx_valid  = valid | OBD_MD_FLCTIME;
 391                 rec->sx_time   = ktime_get_real_seconds();
 392                 rec->sx_size   = output_size;
 393                 rec->sx_flags  = flags;
 394         } else {
 395                 mdc_pack_body(req, fid, valid, output_size, suppgid, flags);
 396         }
 397
 398         if (xattr_name) {
 399                 tmp = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
 400                 memcpy(tmp, xattr_name, xattr_namelen);
 401         }
 402         if (input_size) {
 403                 tmp = req_capsule_client_get(&req->rq_pill, &RMF_EADATA);
 404                 memcpy(tmp, input, input_size);
 405         }
 406
 407         mdc_file_sepol_pack(req);
 408
 409         if (req_capsule_has_field(&req->rq_pill, &RMF_EADATA, RCL_SERVER))
 410                 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
 411                                      RCL_SERVER, output_size);
 412         ptlrpc_request_set_replen(req);
 413
 414         /* make rpc */
 415         if (opcode == MDS_REINT)
 416                 ptlrpc_get_mod_rpc_slot(req);
 417
 418         rc = ptlrpc_queue_wait(req);
 419
 420         if (opcode == MDS_REINT)
 421                 ptlrpc_put_mod_rpc_slot(req);
 422
 423         if (rc)
 424                 ptlrpc_req_finished(req);
 425         else
 426                 *request = req;
 427         RETURN(rc);
 428 }
 429
 430 static int mdc_setxattr(struct obd_export *exp, const struct lu_fid *fid,
 431                         u64 obd_md_valid, const char *name,
 432                         const void *value, size_t value_size,
 433                         unsigned int xattr_flags, u32 suppgid,
 434                         struct ptlrpc_request **req)
 435 {
 436         LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
 437                 obd_md_valid == OBD_MD_FLXATTRRM);
 438
 439         return mdc_xattr_common(exp, &RQF_MDS_REINT_SETXATTR,
 440                                 fid, MDS_REINT, obd_md_valid, name,
 441                                 value, value_size, 0, xattr_flags, suppgid,
 442                                 req);
 443 }
 444
 445 static int mdc_getxattr(struct obd_export *exp, const struct lu_fid *fid,
 446                         u64 obd_md_valid, const char *name, size_t buf_size,
 447                         struct ptlrpc_request **req)
 448 {
 449         struct mdt_body *body;
 450         int rc;
 451
 452         LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
 453                 obd_md_valid == OBD_MD_FLXATTRLS);
 454
 455         /* The below message is checked in sanity-selinux.sh test_20d */
 456         CDEBUG(D_INFO, "%s: get xattr '%s' for "DFID"\n",
 457                exp->exp_obd->obd_name, name, PFID(fid));
 458         rc = mdc_xattr_common(exp, &RQF_MDS_GETXATTR, fid, MDS_GETXATTR,
 459                               obd_md_valid, name, NULL, 0, buf_size, 0, -1,
 460                               req);
 461         if (rc < 0)
 462                 GOTO(out, rc);
 463
 464         body = req_capsule_server_get(&(*req)->rq_pill, &RMF_MDT_BODY);
 465         if (body == NULL)
 466                 GOTO(out, rc = -EPROTO);
 467
 468         /* only detect the xattr size */
 469         if (buf_size == 0) {
 470                 /* LU-11109: Older MDTs do not distinguish
 471                  * between nonexistent xattrs and zero length
 472                  * values in this case. Newer MDTs will return
 473                  * -ENODATA or set OBD_MD_FLXATTR. */
 474                 GOTO(out, rc = body->mbo_eadatasize);
 475         }
 476
 477         if (body->mbo_eadatasize == 0) {
 478                 /* LU-11109: Newer MDTs set OBD_MD_FLXATTR on
 479                  * success so that we can distinguish between
 480                  * zero length value and nonexistent xattr.
 481                  *
 482                  * If OBD_MD_FLXATTR is not set then we keep
 483                  * the old behavior and return -ENODATA for
 484                  * getxattr() when mbo_eadatasize is 0. But
 485                  * -ENODATA only makes sense for getxattr()
 486                  * and not for listxattr(). */
 487                 if (body->mbo_valid & OBD_MD_FLXATTR)
 488                         GOTO(out, rc = 0);
 489                 else if (obd_md_valid == OBD_MD_FLXATTR)
 490                         GOTO(out, rc = -ENODATA);
 491                 else
 492                         GOTO(out, rc = 0);
 493         }
 494
 495         GOTO(out, rc = body->mbo_eadatasize);
 496 out:
 497         if (rc < 0) {
 498                 ptlrpc_req_finished(*req);
 499                 *req = NULL;
 500         }
 501
 502         return rc;
 503 }
 504
 505 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
 506 static int mdc_unpack_acl(struct ptlrpc_request *req, struct lustre_md *md)
 507 {
 508         struct req_capsule     *pill = &req->rq_pill;
 509         struct mdt_body        *body = md->body;
 510         struct posix_acl       *acl;
 511         void                   *buf;
 512         int                     rc;
 513         ENTRY;
 514
 515         if (!body->mbo_aclsize)
 516                 RETURN(0);
 517
 518         buf = req_capsule_server_sized_get(pill, &RMF_ACL, body->mbo_aclsize);
 519
 520         if (!buf)
 521                 RETURN(-EPROTO);
 522
 523         acl = posix_acl_from_xattr(&init_user_ns, buf, body->mbo_aclsize);
 524         if (acl == NULL)
 525                 RETURN(0);
 526         if (IS_ERR(acl)) {
 527                 rc = PTR_ERR(acl);
 528                 CERROR("convert xattr to acl: %d\n", rc);
 529                 RETURN(rc);
 530         }
 531
 532         rc = posix_acl_valid(&init_user_ns, acl);
 533         if (rc) {
 534                 CERROR("validate acl: %d\n", rc);
 535                 posix_acl_release(acl);
 536                 RETURN(rc);
 537         }
 538
 539         md->posix_acl = acl;
 540         RETURN(0);
 541 }
 542 #else
 543 #define mdc_unpack_acl(req, md) 0
 544 #endif
 545
 546 int mdc_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
 547                       struct obd_export *dt_exp, struct obd_export *md_exp,
 548                       struct lustre_md *md)
 549 {
 550         struct req_capsule *pill = &req->rq_pill;
 551         int rc;
 552         ENTRY;
 553
 554         LASSERT(md);
 555         memset(md, 0, sizeof(*md));
 556
 557         md->body = req_capsule_server_get(pill, &RMF_MDT_BODY);
 558         LASSERT(md->body != NULL);
 559
 560         if (md->body->mbo_valid & OBD_MD_FLEASIZE) {
 561                 if (!S_ISREG(md->body->mbo_mode)) {
 562                         CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, should be a "
 563                                "regular file, but is not\n");
 564                         GOTO(out, rc = -EPROTO);
 565                 }
 566
 567                 if (md->body->mbo_eadatasize == 0) {
 568                         CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, "
 569                                "but eadatasize 0\n");
 570                         GOTO(out, rc = -EPROTO);
 571                 }
 572
 573                 md->layout.lb_len = md->body->mbo_eadatasize;
 574                 md->layout.lb_buf = req_capsule_server_sized_get(pill,
 575                                                         &RMF_MDT_MD,
 576                                                         md->layout.lb_len);
 577                 if (md->layout.lb_buf == NULL)
 578                         GOTO(out, rc = -EPROTO);
 579         } else if (md->body->mbo_valid & OBD_MD_FLDIREA) {
 580                 const union lmv_mds_md *lmv;
 581                 size_t lmv_size;
 582
 583                 if (!S_ISDIR(md->body->mbo_mode)) {
 584                         CDEBUG(D_INFO, "OBD_MD_FLDIREA set, should be a "
 585                                "directory, but is not\n");
 586                         GOTO(out, rc = -EPROTO);
 587                 }
 588
 589                 if (md->body->mbo_valid & OBD_MD_MEA) {
 590                         lmv_size = md->body->mbo_eadatasize;
 591                         if (lmv_size == 0) {
 592                                 CDEBUG(D_INFO, "OBD_MD_FLDIREA is set, "
 593                                        "but eadatasize 0\n");
 594                                 RETURN(-EPROTO);
 595                         }
 596
 597                         lmv = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
 598                                                            lmv_size);
 599                         if (lmv == NULL)
 600                                 GOTO(out, rc = -EPROTO);
 601
 602                         rc = md_unpackmd(md_exp, &md->lmv, lmv, lmv_size);
 603                         if (rc < 0)
 604                                 GOTO(out, rc);
 605
 606                         if (rc < (int)sizeof(*md->lmv)) {
 607                                 struct lmv_foreign_md *lfm = md->lfm;
 608
 609                                 /* short (< sizeof(struct lmv_stripe_md))
 610                                  * foreign LMV case
 611                                  */
 612                                 if (lfm->lfm_magic != LMV_MAGIC_FOREIGN) {
 613                                         CDEBUG(D_INFO,
 614                                                "lmv size too small: %d < %d\n",
 615                                                rc, (int)sizeof(*md->lmv));
 616                                         GOTO(out, rc = -EPROTO);
 617                                 }
 618                         }
 619                 }
 620
 621                 /* since 2.12.58 intent_getattr fetches default LMV */
 622                 if (md->body->mbo_valid & OBD_MD_DEFAULT_MEA) {
 623                         lmv_size = sizeof(struct lmv_user_md);
 624                         lmv = req_capsule_server_sized_get(pill,
 625                                                            &RMF_DEFAULT_MDT_MD,
 626                                                            lmv_size);
 627                         if (!lmv)
 628                                 GOTO(out, rc = -EPROTO);
 629
 630                         rc = md_unpackmd(md_exp, &md->default_lmv, lmv,
 631                                          lmv_size);
 632                         if (rc < 0)
 633                                 GOTO(out, rc);
 634
 635                         if (rc < (int)sizeof(*md->default_lmv)) {
 636                                 CDEBUG(D_INFO,
 637                                        "default lmv size too small: %d < %d\n",
 638                                         rc, (int)sizeof(*md->default_lmv));
 639                                 GOTO(out, rc = -EPROTO);
 640                         }
 641                 }
 642         }
 643         rc = 0;
 644
 645         if (md->body->mbo_valid & OBD_MD_FLACL) {
 646                 /* for ACL, it's possible that FLACL is set but aclsize is zero.
 647                  * only when aclsize != 0 there's an actual segment for ACL
 648                  * in reply buffer.
 649                  */
 650                 if (md->body->mbo_aclsize) {
 651                         rc = mdc_unpack_acl(req, md);
 652                         if (rc)
 653                                 GOTO(out, rc);
 654 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
 655                 } else {
 656                         md->posix_acl = NULL;
 657 #endif
 658                 }
 659         }
 660
 661         EXIT;
 662 out:
 663         if (rc) {
 664 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
 665                 posix_acl_release(md->posix_acl);
 666 #endif
 667         }
 668         return rc;
 669 }
 670
 671 int mdc_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
 672 {
 673         ENTRY;
 674         RETURN(0);
 675 }
 676
 677 void mdc_replay_open(struct ptlrpc_request *req)
 678 {
 679         struct md_open_data *mod = req->rq_cb_data;
 680         struct ptlrpc_request *close_req;
 681         struct obd_client_handle *och;
 682         struct lustre_handle old_open_handle = { };
 683         struct mdt_body *body;
 684         ENTRY;
 685
 686         if (mod == NULL) {
 687                 DEBUG_REQ(D_ERROR, req,
 688                           "cannot properly replay without open data");
 689                 EXIT;
 690                 return;
 691         }
 692
 693         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 694         LASSERT(body != NULL);
 695
 696         spin_lock(&req->rq_lock);
 697         och = mod->mod_och;
 698         if (och && och->och_open_handle.cookie)
 699                 req->rq_early_free_repbuf = 1;
 700         else
 701                 req->rq_early_free_repbuf = 0;
 702         spin_unlock(&req->rq_lock);
 703
 704         if (req->rq_early_free_repbuf) {
 705                 struct lustre_handle *file_open_handle;
 706
 707                 LASSERT(och->och_magic == OBD_CLIENT_HANDLE_MAGIC);
 708
 709                 file_open_handle = &och->och_open_handle;
 710                 CDEBUG(D_HA, "updating handle from %#llx to %#llx\n",
 711                        file_open_handle->cookie, body->mbo_open_handle.cookie);
 712                 old_open_handle = *file_open_handle;
 713                 *file_open_handle = body->mbo_open_handle;
 714         }
 715
 716         close_req = mod->mod_close_req;
 717         if (close_req) {
 718                 __u32 opc = lustre_msg_get_opc(close_req->rq_reqmsg);
 719                 struct mdt_ioepoch *epoch;
 720
 721                 LASSERT(opc == MDS_CLOSE);
 722                 epoch = req_capsule_client_get(&close_req->rq_pill,
 723                                                &RMF_MDT_EPOCH);
 724                 LASSERT(epoch);
 725
 726                 if (req->rq_early_free_repbuf)
 727                         LASSERT(old_open_handle.cookie ==
 728                                 epoch->mio_open_handle.cookie);
 729
 730                 DEBUG_REQ(D_HA, close_req, "updating close body with new fh");
 731                 epoch->mio_open_handle = body->mbo_open_handle;
 732         }
 733         EXIT;
 734 }
 735
 736 void mdc_commit_open(struct ptlrpc_request *req)
 737 {
 738         struct md_open_data *mod = req->rq_cb_data;
 739         if (mod == NULL)
 740                 return;
 741
 742         /**
 743          * No need to touch md_open_data::mod_och, it holds a reference on
 744          * \var mod and will zero references to each other, \var mod will be
 745          * freed after that when md_open_data::mod_och will put the reference.
 746          */
 747
 748         /**
 749          * Do not let open request to disappear as it still may be needed
 750          * for close rpc to happen (it may happen on evict only, otherwise
 751          * ptlrpc_request::rq_replay does not let mdc_commit_open() to be
 752          * called), just mark this rpc as committed to distinguish these 2
 753          * cases, see mdc_close() for details. The open request reference will
 754          * be put along with freeing \var mod.
 755          */
 756         ptlrpc_request_addref(req);
 757         spin_lock(&req->rq_lock);
 758         req->rq_committed = 1;
 759         spin_unlock(&req->rq_lock);
 760         req->rq_cb_data = NULL;
 761         obd_mod_put(mod);
 762 }
 763
 764 int mdc_set_open_replay_data(struct obd_export *exp,
 765                              struct obd_client_handle *och,
 766                              struct lookup_intent *it)
 767 {
 768         struct md_open_data     *mod;
 769         struct mdt_rec_create   *rec;
 770         struct mdt_body         *body;
 771         struct ptlrpc_request   *open_req = it->it_request;
 772         struct obd_import       *imp = open_req->rq_import;
 773         ENTRY;
 774
 775         if (!open_req->rq_replay)
 776                 RETURN(0);
 777
 778         rec = req_capsule_client_get(&open_req->rq_pill, &RMF_REC_REINT);
 779         body = req_capsule_server_get(&open_req->rq_pill, &RMF_MDT_BODY);
 780         LASSERT(rec != NULL);
 781         /* Incoming message in my byte order (it's been swabbed). */
 782         /* Outgoing messages always in my byte order. */
 783         LASSERT(body != NULL);
 784
 785         /* Only if the import is replayable, we set replay_open data */
 786         if (och && imp->imp_replayable) {
 787                 mod = obd_mod_alloc();
 788                 if (mod == NULL) {
 789                         DEBUG_REQ(D_ERROR, open_req,
 790                                   "cannot allocate md_open_data");
 791                         RETURN(0);
 792                 }
 793
 794                 /**
 795                  * Take a reference on \var mod, to be freed on mdc_close().
 796                  * It protects \var mod from being freed on eviction (commit
 797                  * callback is called despite rq_replay flag).
 798                  * Another reference for \var och.
 799                  */
 800                 obd_mod_get(mod);
 801                 obd_mod_get(mod);
 802
 803                 spin_lock(&open_req->rq_lock);
 804                 och->och_mod = mod;
 805                 mod->mod_och = och;
 806                 mod->mod_is_create = it_disposition(it, DISP_OPEN_CREATE) ||
 807                                      it_disposition(it, DISP_OPEN_STRIPE);
 808                 mod->mod_open_req = open_req;
 809                 open_req->rq_cb_data = mod;
 810                 open_req->rq_commit_cb = mdc_commit_open;
 811                 open_req->rq_early_free_repbuf = 1;
 812                 spin_unlock(&open_req->rq_lock);
 813         }
 814
 815         rec->cr_fid2 = body->mbo_fid1;
 816         rec->cr_open_handle_old = body->mbo_open_handle;
 817         open_req->rq_replay_cb = mdc_replay_open;
 818         if (!fid_is_sane(&body->mbo_fid1)) {
 819                 DEBUG_REQ(D_ERROR, open_req,
 820                           "saving replay request with insane FID " DFID,
 821                           PFID(&body->mbo_fid1));
 822                 LBUG();
 823         }
 824
 825         DEBUG_REQ(D_RPCTRACE, open_req, "Set up open replay data");
 826         RETURN(0);
 827 }
 828
 829 static void mdc_free_open(struct md_open_data *mod)
 830 {
 831         int committed = 0;
 832
 833         if (mod->mod_is_create == 0 &&
 834             imp_connect_disp_stripe(mod->mod_open_req->rq_import))
 835                 committed = 1;
 836
 837         /**
 838          * No reason to asssert here if the open request has
 839          * rq_replay == 1. It means that mdc_close failed, and
 840          * close request wasn`t sent. It is not fatal to client.
 841          * The worst thing is eviction if the client gets open lock
 842          **/
 843
 844         DEBUG_REQ(D_RPCTRACE, mod->mod_open_req,
 845                   "free open request, rq_replay=%d",
 846                   mod->mod_open_req->rq_replay);
 847
 848         ptlrpc_request_committed(mod->mod_open_req, committed);
 849         if (mod->mod_close_req)
 850                 ptlrpc_request_committed(mod->mod_close_req, committed);
 851 }
 852
 853 int mdc_clear_open_replay_data(struct obd_export *exp,
 854                                struct obd_client_handle *och)
 855 {
 856         struct md_open_data *mod = och->och_mod;
 857         ENTRY;
 858
 859         /**
 860          * It is possible to not have \var mod in a case of eviction between
 861          * lookup and ll_file_open().
 862          **/
 863         if (mod == NULL)
 864                 RETURN(0);
 865
 866         LASSERT(mod != LP_POISON);
 867         LASSERT(mod->mod_open_req != NULL);
 868
 869         spin_lock(&mod->mod_open_req->rq_lock);
 870         if (mod->mod_och)
 871                 mod->mod_och->och_open_handle.cookie = 0;
 872         mod->mod_open_req->rq_early_free_repbuf = 0;
 873         spin_unlock(&mod->mod_open_req->rq_lock);
 874         mdc_free_open(mod);
 875
 876         mod->mod_och = NULL;
 877         och->och_mod = NULL;
 878         obd_mod_put(mod);
 879
 880         RETURN(0);
 881 }
 882
 883 static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
 884                      struct md_open_data *mod, struct ptlrpc_request **request)
 885 {
 886         struct obd_device     *obd = class_exp2obd(exp);
 887         struct ptlrpc_request *req;
 888         struct req_format     *req_fmt;
 889         size_t                 u32_count = 0;
 890         int                    rc;
 891         int                    saved_rc = 0;
 892         ENTRY;
 893
 894         CDEBUG(D_INODE, "%s: "DFID" file closed with intent: %x\n",
 895                exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
 896                op_data->op_bias);
 897
 898         if (op_data->op_bias & MDS_CLOSE_INTENT) {
 899                 req_fmt = &RQF_MDS_CLOSE_INTENT;
 900                 if (op_data->op_bias & MDS_HSM_RELEASE) {
 901                         /* allocate a FID for volatile file */
 902                         rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2,
 903                                            op_data);
 904                         if (rc < 0) {
 905                                 CERROR("%s: "DFID" allocating FID: rc = %d\n",
 906                                        obd->obd_name, PFID(&op_data->op_fid1),
 907                                        rc);
 908                                 /* save the errcode and proceed to close */
 909                                 saved_rc = rc;
 910                         }
 911                 }
 912                 if (op_data->op_bias & MDS_CLOSE_RESYNC_DONE) {
 913                         size_t count = op_data->op_data_size / sizeof(__u32);
 914
 915                         if (count > INLINE_RESYNC_ARRAY_SIZE)
 916                                 u32_count = count;
 917                 }
 918         } else {
 919                 req_fmt = &RQF_MDS_CLOSE;
 920         }
 921
 922         *request = NULL;
 923         if (OBD_FAIL_CHECK(OBD_FAIL_MDC_CLOSE))
 924                 req = NULL;
 925         else
 926                 req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);
 927
 928         /* Ensure that this close's handle is fixed up during replay. */
 929         if (likely(mod != NULL)) {
 930                 LASSERTF(mod->mod_open_req != NULL &&
 931                          mod->mod_open_req->rq_type != LI_POISON,
 932                          "POISONED open %p!\n", mod->mod_open_req);
 933
 934                 mod->mod_close_req = req;
 935
 936                 DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "matched open");
 937                 /* We no longer want to preserve this open for replay even
 938                  * though the open was committed. b=3632, b=3633 */
 939                 spin_lock(&mod->mod_open_req->rq_lock);
 940                 mod->mod_open_req->rq_replay = 0;
 941                 spin_unlock(&mod->mod_open_req->rq_lock);
 942         } else {
 943                 CDEBUG(D_HA, "couldn't find open req; expecting close error\n");
 944         }
 945         if (req == NULL) {
 946                 /**
 947                  * TODO: repeat close after errors
 948                  */
 949                 CWARN("%s: close of FID "DFID" failed, file reference will be "
 950                       "dropped when this client unmounts or is evicted\n",
 951                       obd->obd_name, PFID(&op_data->op_fid1));
 952                 GOTO(out, rc = -ENOMEM);
 953         }
 954
 955         if (u32_count > 0)
 956                 req_capsule_set_size(&req->rq_pill, &RMF_U32, RCL_CLIENT,
 957                                      u32_count * sizeof(__u32));
 958
 959         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
 960         if (rc) {
 961                 ptlrpc_request_free(req);
 962                 req = NULL;
 963                 GOTO(out, rc);
 964         }
 965
 966         /* To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
 967          * portal whose threads are not taking any DLM locks and are therefore
 968          * always progressing */
 969         req->rq_request_portal = MDS_READPAGE_PORTAL;
 970         ptlrpc_at_set_req_timeout(req);
 971
 972         if (!(exp_connect_flags2(exp) & OBD_CONNECT2_LSOM))
 973                 op_data->op_xvalid &= ~(OP_XVALID_LAZYSIZE |
 974                                         OP_XVALID_LAZYBLOCKS);
 975
 976         mdc_close_pack(req, op_data);
 977
 978         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 979                              obd->u.cli.cl_default_mds_easize);
 980
 981         ptlrpc_request_set_replen(req);
 982
 983         ptlrpc_get_mod_rpc_slot(req);
 984         rc = ptlrpc_queue_wait(req);
 985         ptlrpc_put_mod_rpc_slot(req);
 986
 987         if (req->rq_repmsg == NULL) {
 988                 CDEBUG(D_RPCTRACE, "request %p failed to send: rc = %d\n", req,
 989                        req->rq_status);
 990                 if (rc == 0)
 991                         rc = req->rq_status ?: -EIO;
 992         } else if (rc == 0 || rc == -EAGAIN) {
 993                 struct mdt_body *body;
 994
 995                 rc = lustre_msg_get_status(req->rq_repmsg);
 996                 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
 997                         DEBUG_REQ(D_ERROR, req,
 998                                   "type = PTL_RPC_MSG_ERR: rc = %d", rc);
 999                         if (rc > 0)
1000                                 rc = -rc;
1001                 }
1002                 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1003                 if (body == NULL)
1004                         rc = -EPROTO;
1005         } else if (rc == -ESTALE) {
1006                 /**
1007                  * it can be allowed error after 3633 if open was committed and
1008                  * server failed before close was sent. Let's check if mod
1009                  * exists and return no error in that case
1010                  */
1011                 if (mod) {
1012                         DEBUG_REQ(D_HA, req, "Reset ESTALE = %d", rc);
1013                         LASSERT(mod->mod_open_req != NULL);
1014                         if (mod->mod_open_req->rq_committed)
1015                                 rc = 0;
1016                 }
1017         }
1018
1019 out:
1020         if (mod) {
1021                 if (rc != 0)
1022                         mod->mod_close_req = NULL;
1023                 /* Since now, mod is accessed through open_req only,
1024                  * thus close req does not keep a reference on mod anymore. */
1025                 obd_mod_put(mod);
1026         }
1027         *request = req;
1028
1029         RETURN(rc < 0 ? rc : saved_rc);
1030 }
1031
1032 static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid,
1033                        u64 offset, struct page **pages, int npages,
1034                        struct ptlrpc_request **request)
1035 {
1036         struct ptlrpc_request   *req;
1037         struct ptlrpc_bulk_desc *desc;
1038         int                      i;
1039         int                      resends = 0;
1040         int                      rc;
1041         ENTRY;
1042
1043         *request = NULL;
1044
1045 restart_bulk:
1046         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE);
1047         if (req == NULL)
1048                 RETURN(-ENOMEM);
1049
1050         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE);
1051         if (rc) {
1052                 ptlrpc_request_free(req);
1053                 RETURN(rc);
1054         }
1055
1056         req->rq_request_portal = MDS_READPAGE_PORTAL;
1057         ptlrpc_at_set_req_timeout(req);
1058
1059         desc = ptlrpc_prep_bulk_imp(req, npages, 1,
1060                                     PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
1061                                     MDS_BULK_PORTAL,
1062                                     &ptlrpc_bulk_kiov_pin_ops);
1063         if (desc == NULL) {
1064                 ptlrpc_req_finished(req);
1065                 RETURN(-ENOMEM);
1066         }
1067
1068         /* NB req now owns desc and will free it when it gets freed */
1069         for (i = 0; i < npages; i++)
1070                 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1071                                                  PAGE_SIZE);
1072
1073         mdc_readdir_pack(req, offset, PAGE_SIZE * npages, fid);
1074
1075         ptlrpc_request_set_replen(req);
1076         rc = ptlrpc_queue_wait(req);
1077         if (rc) {
1078                 ptlrpc_req_finished(req);
1079                 if (rc != -ETIMEDOUT)
1080                         RETURN(rc);
1081
1082                 resends++;
1083                 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
1084                         CERROR("%s: too many resend retries: rc = %d\n",
1085                                exp->exp_obd->obd_name, -EIO);
1086                         RETURN(-EIO);
1087                 }
1088                 ssleep(resends);
1089
1090                 goto restart_bulk;
1091         }
1092
1093         rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1094                                           req->rq_bulk->bd_nob_transferred);
1095         if (rc < 0) {
1096                 ptlrpc_req_finished(req);
1097                 RETURN(rc);
1098         }
1099
1100         if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) {
1101                 CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n",
1102                        exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred,
1103                        PAGE_SIZE * npages);
1104                 ptlrpc_req_finished(req);
1105                 RETURN(-EPROTO);
1106         }
1107
1108         *request = req;
1109         RETURN(0);
1110 }
1111
1112 static void mdc_release_page(struct page *page, int remove)
1113 {
1114         if (remove) {
1115                 lock_page(page);
1116                 if (likely(page->mapping != NULL))
1117                         delete_from_page_cache(page);
1118                 unlock_page(page);
1119         }
1120         put_page(page);
1121 }
1122
1123 static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash,
1124                                     __u64 *start, __u64 *end, int hash64)
1125 {
1126         /*
1127          * Complement of hash is used as an index so that
1128          * radix_tree_gang_lookup() can be used to find a page with starting
1129          * hash _smaller_ than one we are looking for.
1130          */
1131         unsigned long offset = hash_x_index(*hash, hash64);
1132         struct page *page;
1133         unsigned long flags;
1134         int found;
1135
1136         xa_lock_irqsave(&mapping->i_pages, flags);
1137         found = radix_tree_gang_lookup(&mapping->page_tree,
1138                                        (void **)&page, offset, 1);
1139         if (found > 0 && !xa_is_value(page)) {
1140                 struct lu_dirpage *dp;
1141
1142                 get_page(page);
1143                 xa_unlock_irqrestore(&mapping->i_pages, flags);
1144                 /*
1145                  * In contrast to find_lock_page() we are sure that directory
1146                  * page cannot be truncated (while DLM lock is held) and,
1147                  * hence, can avoid restart.
1148                  *
1149                  * In fact, page cannot be locked here at all, because
1150                  * mdc_read_page_remote does synchronous io.
1151                  */
1152                 wait_on_page_locked(page);
1153                 if (PageUptodate(page)) {
1154                         dp = kmap(page);
1155                         if (BITS_PER_LONG == 32 && hash64) {
1156                                 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1157                                 *end   = le64_to_cpu(dp->ldp_hash_end) >> 32;
1158                                 *hash  = *hash >> 32;
1159                         } else {
1160                                 *start = le64_to_cpu(dp->ldp_hash_start);
1161                                 *end   = le64_to_cpu(dp->ldp_hash_end);
1162                         }
1163                         if (unlikely(*start == 1 && *hash == 0))
1164                                 *hash = *start;
1165                         else
1166                                 LASSERTF(*start <= *hash, "start = %#llx"
1167                                          ",end = %#llx,hash = %#llx\n",
1168                                          *start, *end, *hash);
1169                         CDEBUG(D_VFSTRACE, "offset %lx [%#llx %#llx],"
1170                               " hash %#llx\n", offset, *start, *end, *hash);
1171                         if (*hash > *end) {
1172                                 kunmap(page);
1173                                 mdc_release_page(page, 0);
1174                                 page = NULL;
1175                         } else if (*end != *start && *hash == *end) {
1176                                 /*
1177                                  * upon hash collision, remove this page,
1178                                  * otherwise put page reference, and
1179                                  * mdc_read_page_remote() will issue RPC to
1180                                  * fetch the page we want.
1181                                  */
1182                                 kunmap(page);
1183                                 mdc_release_page(page,
1184                                     le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1185                                 page = NULL;
1186                         }
1187                 } else {
1188                         put_page(page);
1189                         page = ERR_PTR(-EIO);
1190                 }
1191         } else {
1192                 xa_unlock_irqrestore(&mapping->i_pages, flags);
1193                 page = NULL;
1194         }
1195         return page;
1196 }
1197
1198 /*
1199  * Adjust a set of pages, each page containing an array of lu_dirpages,
1200  * so that each page can be used as a single logical lu_dirpage.
1201  *
1202  * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
1203  * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
1204  * struct lu_dirent.  It has size up to LU_PAGE_SIZE. The ldp_hash_end
1205  * value is used as a cookie to request the next lu_dirpage in a
1206  * directory listing that spans multiple pages (two in this example):
1207  *   ________
1208  *  |        |
1209  * .|--------v-------   -----.
1210  * |s|e|f|p|ent|ent| ... |ent|
1211  * '--|--------------   -----'   Each PAGE contains a single
1212  *    '------.                   lu_dirpage.
1213  * .---------v-------   -----.
1214  * |s|e|f|p|ent| 0 | ... | 0 |
1215  * '-----------------   -----'
1216  *
1217  * However, on hosts where the native VM page size (PAGE_SIZE) is
1218  * larger than LU_PAGE_SIZE, a single host page may contain multiple
1219  * lu_dirpages. After reading the lu_dirpages from the MDS, the
1220  * ldp_hash_end of the first lu_dirpage refers to the one immediately
1221  * after it in the same PAGE (arrows simplified for brevity, but
1222  * in general e0==s1, e1==s2, etc.):
1223  *
1224  * .--------------------   -----.
1225  * |s0|e0|f0|p|ent|ent| ... |ent|
1226  * |---v----------------   -----|
1227  * |s1|e1|f1|p|ent|ent| ... |ent|
1228  * |---v----------------   -----|  Here, each PAGE contains
1229  *             ...                 multiple lu_dirpages.
1230  * |---v----------------   -----|
1231  * |s'|e'|f'|p|ent|ent| ... |ent|
1232  * '---|----------------   -----'
1233  *     v
1234  * .----------------------------.
1235  * |        next PAGE           |
1236  *
1237  * This structure is transformed into a single logical lu_dirpage as follows:
1238  *
1239  * - Replace e0 with e' so the request for the next lu_dirpage gets the page
1240  *   labeled 'next PAGE'.
1241  *
1242  * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
1243  *   a hash collision with the next page exists.
1244  *
1245  * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
1246  *   to the first entry of the next lu_dirpage.
1247  */
1248 #if PAGE_SIZE > LU_PAGE_SIZE
1249 static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs)
1250 {
1251         int i;
1252
1253         for (i = 0; i < cfs_pgs; i++) {
1254                 struct lu_dirpage *dp = kmap(pages[i]);
1255                 struct lu_dirpage *first = dp;
1256                 struct lu_dirent *end_dirent = NULL;
1257                 struct lu_dirent *ent;
1258                 __u64 hash_end = dp->ldp_hash_end;
1259                 __u32 flags = dp->ldp_flags;
1260
1261                 while (--lu_pgs > 0) {
1262                         ent = lu_dirent_start(dp);
1263                         for (end_dirent = ent; ent != NULL;
1264                              end_dirent = ent, ent = lu_dirent_next(ent));
1265
1266                         /* Advance dp to next lu_dirpage. */
1267                         dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
1268
1269                         /* Check if we've reached the end of the PAGE. */
1270                         if (!((unsigned long)dp & ~PAGE_MASK))
1271                                 break;
1272
1273                         /* Save the hash and flags of this lu_dirpage. */
1274                         hash_end = dp->ldp_hash_end;
1275                         flags = dp->ldp_flags;
1276
1277                         /* Check if lu_dirpage contains no entries. */
1278                         if (end_dirent == NULL)
1279                                 break;
1280
1281                         /* Enlarge the end entry lde_reclen from 0 to
1282                          * first entry of next lu_dirpage. */
1283                         LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
1284                         end_dirent->lde_reclen =
1285                                 cpu_to_le16((char *)(dp->ldp_entries) -
1286                                             (char *)end_dirent);
1287                 }
1288
1289                 first->ldp_hash_end = hash_end;
1290                 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
1291                 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
1292
1293                 kunmap(pages[i]);
1294         }
1295         LASSERTF(lu_pgs == 0, "left = %d\n", lu_pgs);
1296 }
1297 #else
1298 #define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0)
1299 #endif  /* PAGE_SIZE > LU_PAGE_SIZE */
1300
1301 /* parameters for readdir page */
1302 struct readpage_param {
1303         struct md_op_data       *rp_mod;
1304         __u64                   rp_off;
1305         int                     rp_hash64;
1306         struct obd_export       *rp_exp;
1307         struct md_callback      *rp_cb;
1308 };
1309
1310 /**
1311  * Read pages from server.
1312  *
1313  * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
1314  * a header lu_dirpage which describes the start/end hash, and whether this
1315  * page is empty (contains no dir entry) or hash collide with next page.
1316  * After client receives reply, several pages will be integrated into dir page
1317  * in PAGE_SIZE (if PAGE_SIZE greater than LU_PAGE_SIZE), and the
1318  * lu_dirpage for this integrated page will be adjusted.
1319  **/
1320 static int mdc_read_page_remote(void *data, struct page *page0)
1321 {
1322         struct readpage_param *rp = data;
1323         struct page **page_pool;
1324         struct page *page;
1325         struct lu_dirpage *dp;
1326         struct md_op_data *op_data = rp->rp_mod;
1327         struct ptlrpc_request *req;
1328         int max_pages;
1329         struct inode *inode;
1330         struct lu_fid *fid;
1331         int rd_pgs = 0; /* number of pages actually read */
1332         int npages;
1333         int i;
1334         int rc;
1335         ENTRY;
1336
1337         max_pages = rp->rp_exp->exp_obd->u.cli.cl_max_pages_per_rpc;
1338         inode = op_data->op_data;
1339         fid = &op_data->op_fid1;
1340         LASSERT(inode != NULL);
1341
1342         OBD_ALLOC(page_pool, sizeof(page_pool[0]) * max_pages);
1343         if (page_pool != NULL) {
1344                 page_pool[0] = page0;
1345         } else {
1346                 page_pool = &page0;
1347                 max_pages = 1;
1348         }
1349
1350         for (npages = 1; npages < max_pages; npages++) {
1351                 page = __page_cache_alloc(mapping_gfp_mask(inode->i_mapping)
1352                                           | __GFP_COLD);
1353                 if (page == NULL)
1354                         break;
1355                 page_pool[npages] = page;
1356         }
1357
1358         rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, page_pool, npages, &req);
1359         if (rc < 0) {
1360                 /* page0 is special, which was added into page cache early */
1361                 delete_from_page_cache(page0);
1362         } else {
1363                 int lu_pgs;
1364
1365                 rd_pgs = (req->rq_bulk->bd_nob_transferred + PAGE_SIZE - 1) >>
1366                         PAGE_SHIFT;
1367                 lu_pgs = req->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
1368                 LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
1369
1370                 CDEBUG(D_INODE, "read %d(%d) pages\n", rd_pgs, lu_pgs);
1371
1372                 mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs);
1373
1374                 SetPageUptodate(page0);
1375         }
1376         unlock_page(page0);
1377
1378         ptlrpc_req_finished(req);
1379         CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages);
1380         for (i = 1; i < npages; i++) {
1381                 unsigned long   offset;
1382                 __u64           hash;
1383                 int ret;
1384
1385                 page = page_pool[i];
1386
1387                 if (rc < 0 || i >= rd_pgs) {
1388                         put_page(page);
1389                         continue;
1390                 }
1391
1392                 SetPageUptodate(page);
1393
1394                 dp = kmap(page);
1395                 hash = le64_to_cpu(dp->ldp_hash_start);
1396                 kunmap(page);
1397
1398                 offset = hash_x_index(hash, rp->rp_hash64);
1399
1400                 prefetchw(&page->flags);
1401                 ret = add_to_page_cache_lru(page, inode->i_mapping, offset,
1402                                             GFP_KERNEL);
1403                 if (ret == 0)
1404                         unlock_page(page);
1405                 else
1406                         CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
1407                                " rc = %d\n", offset, ret);
1408                 put_page(page);
1409         }
1410
1411         if (page_pool != &page0)
1412                 OBD_FREE(page_pool, sizeof(page_pool[0]) * max_pages);
1413
1414         RETURN(rc);
1415 }
1416
1417 /**
1418  * Read dir page from cache first, if it can not find it, read it from
1419  * server and add into the cache.
1420  *
1421  * \param[in] exp       MDC export
1422  * \param[in] op_data   client MD stack parameters, transfering parameters
1423  *                      between different layers on client MD stack.
1424  * \param[in] cb_op     callback required for ldlm lock enqueue during
1425  *                      read page
1426  * \param[in] hash_offset the hash offset of the page to be read
1427  * \param[in] ppage     the page to be read
1428  *
1429  * retval               = 0 get the page successfully
1430  *                      errno(<0) get the page failed
1431  */
1432 static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data,
1433                          struct md_callback *cb_op, __u64 hash_offset,
1434                          struct page **ppage)
1435 {
1436         struct lookup_intent    it = { .it_op = IT_READDIR };
1437         struct page             *page;
1438         struct inode            *dir = op_data->op_data;
1439         struct address_space    *mapping;
1440         struct lu_dirpage       *dp;
1441         __u64                   start = 0;
1442         __u64                   end = 0;
1443         struct lustre_handle    lockh;
1444         struct ptlrpc_request   *enq_req = NULL;
1445         struct readpage_param   rp_param;
1446         int rc;
1447
1448         ENTRY;
1449
1450         *ppage = NULL;
1451
1452         LASSERT(dir != NULL);
1453         mapping = dir->i_mapping;
1454
1455         rc = mdc_intent_lock(exp, op_data, &it, &enq_req,
1456                              cb_op->md_blocking_ast, 0);
1457         if (enq_req != NULL)
1458                 ptlrpc_req_finished(enq_req);
1459
1460         if (rc < 0) {
1461                 CERROR("%s: "DFID" lock enqueue fails: rc = %d\n",
1462                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc);
1463                 RETURN(rc);
1464         }
1465
1466         rc = 0;
1467         lockh.cookie = it.it_lock_handle;
1468         mdc_set_lock_data(exp, &lockh, dir, NULL);
1469
1470         rp_param.rp_off = hash_offset;
1471         rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64;
1472         page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end,
1473                                rp_param.rp_hash64);
1474         if (IS_ERR(page)) {
1475                 CERROR("%s: dir page locate: "DFID" at %llu: rc %ld\n",
1476                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1477                        rp_param.rp_off, PTR_ERR(page));
1478                 GOTO(out_unlock, rc = PTR_ERR(page));
1479         } else if (page != NULL) {
1480                 /*
1481                  * XXX nikita: not entirely correct handling of a corner case:
1482                  * suppose hash chain of entries with hash value HASH crosses
1483                  * border between pages P0 and P1. First both P0 and P1 are
1484                  * cached, seekdir() is called for some entry from the P0 part
1485                  * of the chain. Later P0 goes out of cache. telldir(HASH)
1486                  * happens and finds P1, as it starts with matching hash
1487                  * value. Remaining entries from P0 part of the chain are
1488                  * skipped. (Is that really a bug?)
1489                  *
1490                  * Possible solutions: 0. don't cache P1 is such case, handle
1491                  * it as an "overflow" page. 1. invalidate all pages at
1492                  * once. 2. use HASH|1 as an index for P1.
1493                  */
1494                 GOTO(hash_collision, page);
1495         }
1496
1497         rp_param.rp_exp = exp;
1498         rp_param.rp_mod = op_data;
1499         page = read_cache_page(mapping,
1500                                hash_x_index(rp_param.rp_off,
1501                                             rp_param.rp_hash64),
1502                                mdc_read_page_remote, &rp_param);
1503         if (IS_ERR(page)) {
1504                 CDEBUG(D_INFO, "%s: read cache page: "DFID" at %llu: %ld\n",
1505                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1506                        rp_param.rp_off, PTR_ERR(page));
1507                 GOTO(out_unlock, rc = PTR_ERR(page));
1508         }
1509
1510         wait_on_page_locked(page);
1511         (void)kmap(page);
1512         if (!PageUptodate(page)) {
1513                 CERROR("%s: page not updated: "DFID" at %llu: rc %d\n",
1514                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1515                        rp_param.rp_off, -5);
1516                 goto fail;
1517         }
1518         if (!PageChecked(page))
1519                 SetPageChecked(page);
1520         if (PageError(page)) {
1521                 CERROR("%s: page error: "DFID" at %llu: rc %d\n",
1522                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1523                        rp_param.rp_off, -5);
1524                 goto fail;
1525         }
1526
1527 hash_collision:
1528         dp = page_address(page);
1529         if (BITS_PER_LONG == 32 && rp_param.rp_hash64) {
1530                 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1531                 end   = le64_to_cpu(dp->ldp_hash_end) >> 32;
1532                 rp_param.rp_off = hash_offset >> 32;
1533         } else {
1534                 start = le64_to_cpu(dp->ldp_hash_start);
1535                 end   = le64_to_cpu(dp->ldp_hash_end);
1536                 rp_param.rp_off = hash_offset;
1537         }
1538         if (end == start) {
1539                 LASSERT(start == rp_param.rp_off);
1540                 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
1541 #if BITS_PER_LONG == 32
1542                 CWARN("Real page-wide hash collision at [%llu %llu] with "
1543                       "hash %llu\n", le64_to_cpu(dp->ldp_hash_start),
1544                       le64_to_cpu(dp->ldp_hash_end), hash_offset);
1545 #endif
1546
1547                 /*
1548                  * Fetch whole overflow chain...
1549                  *
1550                  * XXX not yet.
1551                  */
1552                 goto fail;
1553         }
1554         *ppage = page;
1555 out_unlock:
1556         ldlm_lock_decref(&lockh, it.it_lock_mode);
1557         return rc;
1558 fail:
1559         kunmap(page);
1560         mdc_release_page(page, 1);
1561         rc = -EIO;
1562         goto out_unlock;
1563 }
1564
1565 static int mdc_statfs_interpret(const struct lu_env *env,
1566                                 struct ptlrpc_request *req, void *args, int rc)
1567 {
1568         struct obd_info *oinfo = args;
1569         struct obd_statfs *osfs;
1570
1571         if (!rc) {
1572                 osfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1573                 if (!osfs)
1574                         return -EPROTO;
1575
1576                 oinfo->oi_osfs = osfs;
1577
1578                 CDEBUG(D_CACHE, "blocks=%llu free=%llu avail=%llu "
1579                        "objects=%llu free=%llu state=%x\n",
1580                         osfs->os_blocks, osfs->os_bfree, osfs->os_bavail,
1581                         osfs->os_files, osfs->os_ffree, osfs->os_state);
1582         }
1583
1584         oinfo->oi_cb_up(oinfo, rc);
1585
1586         return rc;
1587 }
1588
1589 static int mdc_statfs_async(struct obd_export *exp,
1590                             struct obd_info *oinfo, time64_t max_age,
1591                             struct ptlrpc_request_set *unused)
1592 {
1593         struct ptlrpc_request *req;
1594         struct obd_info *aa;
1595
1596         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp), &RQF_MDS_STATFS,
1597                                         LUSTRE_MDS_VERSION, MDS_STATFS);
1598         if (req == NULL)
1599                 return -ENOMEM;
1600
1601         ptlrpc_request_set_replen(req);
1602         req->rq_interpret_reply = mdc_statfs_interpret;
1603
1604         aa = ptlrpc_req_async_args(aa, req);
1605         *aa = *oinfo;
1606
1607         ptlrpcd_add_req(req);
1608
1609         return 0;
1610 }
1611
1612 static int mdc_statfs(const struct lu_env *env,
1613                       struct obd_export *exp, struct obd_statfs *osfs,
1614                       time64_t max_age, __u32 flags)
1615 {
1616         struct obd_device *obd = class_exp2obd(exp);
1617         struct req_format *fmt;
1618         struct ptlrpc_request *req;
1619         struct obd_statfs *msfs;
1620         struct obd_import *imp = NULL;
1621         int rc;
1622         ENTRY;
1623
1624         /*
1625          * Since the request might also come from lprocfs, so we need
1626          * sync this with client_disconnect_export Bug15684
1627          */
1628         down_read(&obd->u.cli.cl_sem);
1629         if (obd->u.cli.cl_import)
1630                 imp = class_import_get(obd->u.cli.cl_import);
1631         up_read(&obd->u.cli.cl_sem);
1632         if (!imp)
1633                 RETURN(-ENODEV);
1634
1635         fmt = &RQF_MDS_STATFS;
1636         if ((exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS) &&
1637             (flags & OBD_STATFS_SUM))
1638                 fmt = &RQF_MDS_STATFS_NEW;
1639         req = ptlrpc_request_alloc_pack(imp, fmt, LUSTRE_MDS_VERSION,
1640                                         MDS_STATFS);
1641         if (req == NULL)
1642                 GOTO(output, rc = -ENOMEM);
1643
1644         if ((flags & OBD_STATFS_SUM) &&
1645             (exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS)) {
1646                 /* request aggregated states */
1647                 struct mdt_body *body;
1648
1649                 body = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
1650                 if (body == NULL)
1651                         GOTO(out, rc = -EPROTO);
1652                 body->mbo_valid = OBD_MD_FLAGSTATFS;
1653         }
1654
1655         ptlrpc_request_set_replen(req);
1656
1657         if (flags & OBD_STATFS_NODELAY) {
1658                 /* procfs requests not want stay in wait for avoid deadlock */
1659                 req->rq_no_resend = 1;
1660                 req->rq_no_delay = 1;
1661         }
1662
1663         rc = ptlrpc_queue_wait(req);
1664         if (rc) {
1665                 /* check connection error first */
1666                 if (imp->imp_connect_error)
1667                         rc = imp->imp_connect_error;
1668                 GOTO(out, rc);
1669         }
1670
1671         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1672         if (msfs == NULL)
1673                 GOTO(out, rc = -EPROTO);
1674
1675         *osfs = *msfs;
1676         EXIT;
1677 out:
1678         ptlrpc_req_finished(req);
1679 output:
1680         class_import_put(imp);
1681         return rc;
1682 }
1683
1684 static int mdc_ioc_fid2path(struct obd_export *exp, struct getinfo_fid2path *gf)
1685 {
1686         __u32 keylen, vallen;
1687         void *key;
1688         int rc;
1689
1690         if (gf->gf_pathlen > PATH_MAX)
1691                 RETURN(-ENAMETOOLONG);
1692         if (gf->gf_pathlen < 2)
1693                 RETURN(-EOVERFLOW);
1694
1695         /* Key is KEY_FID2PATH + getinfo_fid2path description */
1696         keylen = cfs_size_round(sizeof(KEY_FID2PATH) + sizeof(*gf) +
1697                                 sizeof(struct lu_fid));
1698         OBD_ALLOC(key, keylen);
1699         if (key == NULL)
1700                 RETURN(-ENOMEM);
1701         memcpy(key, KEY_FID2PATH, sizeof(KEY_FID2PATH));
1702         memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)), gf, sizeof(*gf));
1703         memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*gf),
1704                gf->gf_u.gf_root_fid, sizeof(struct lu_fid));
1705         CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
1706                PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno);
1707
1708         if (!fid_is_sane(&gf->gf_fid))
1709                 GOTO(out, rc = -EINVAL);
1710
1711         /* Val is struct getinfo_fid2path result plus path */
1712         vallen = sizeof(*gf) + gf->gf_pathlen;
1713
1714         rc = obd_get_info(NULL, exp, keylen, key, &vallen, gf);
1715         if (rc != 0 && rc != -EREMOTE)
1716                 GOTO(out, rc);
1717
1718         if (vallen <= sizeof(*gf))
1719                 GOTO(out, rc = -EPROTO);
1720         if (vallen > sizeof(*gf) + gf->gf_pathlen)
1721                 GOTO(out, rc = -EOVERFLOW);
1722
1723         CDEBUG(D_IOCTL, "path got "DFID" from %llu #%d: %s\n",
1724                PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno,
1725                gf->gf_pathlen < 512 ? gf->gf_u.gf_path :
1726                /* only log the last 512 characters of the path */
1727                gf->gf_u.gf_path + gf->gf_pathlen - 512);
1728
1729 out:
1730         OBD_FREE(key, keylen);
1731         return rc;
1732 }
1733
1734 static int mdc_ioc_hsm_progress(struct obd_export *exp,
1735                                 struct hsm_progress_kernel *hpk)
1736 {
1737         struct obd_import               *imp = class_exp2cliimp(exp);
1738         struct hsm_progress_kernel      *req_hpk;
1739         struct ptlrpc_request           *req;
1740         int                              rc;
1741         ENTRY;
1742
1743         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_PROGRESS,
1744                                         LUSTRE_MDS_VERSION, MDS_HSM_PROGRESS);
1745         if (req == NULL)
1746                 GOTO(out, rc = -ENOMEM);
1747
1748         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1749
1750         /* Copy hsm_progress struct */
1751         req_hpk = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_PROGRESS);
1752         if (req_hpk == NULL)
1753                 GOTO(out, rc = -EPROTO);
1754
1755         *req_hpk = *hpk;
1756         req_hpk->hpk_errval = lustre_errno_hton(hpk->hpk_errval);
1757
1758         ptlrpc_request_set_replen(req);
1759
1760         ptlrpc_get_mod_rpc_slot(req);
1761         rc = ptlrpc_queue_wait(req);
1762         ptlrpc_put_mod_rpc_slot(req);
1763
1764         GOTO(out, rc);
1765 out:
1766         ptlrpc_req_finished(req);
1767         return rc;
1768 }
1769 /**
1770  * Send hsm_ct_register to MDS
1771  *
1772  * \param[in]   imp             import
1773  * \param[in]   archive_count   if in bitmap format, it is the bitmap,
1774  *                              else it is the count of archive_ids
1775  * \param[in]   archives        if in bitmap format, it is NULL,
1776  *                              else it is archive_id lists
1777  */
1778 static int mdc_ioc_hsm_ct_register(struct obd_import *imp, __u32 archive_count,
1779                                    __u32 *archives)
1780 {
1781         struct ptlrpc_request *req;
1782         __u32 *archive_array;
1783         size_t archives_size;
1784         int rc;
1785         ENTRY;
1786
1787         req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_CT_REGISTER);
1788         if (req == NULL)
1789                 RETURN(-ENOMEM);
1790
1791         if (archives != NULL)
1792                 archives_size = sizeof(*archive_array) * archive_count;
1793         else
1794                 archives_size = sizeof(archive_count);
1795
1796         req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_ARCHIVE,
1797                              RCL_CLIENT, archives_size);
1798
1799         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_CT_REGISTER);
1800         if (rc) {
1801                 ptlrpc_request_free(req);
1802                 RETURN(-ENOMEM);
1803         }
1804
1805         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1806
1807         archive_array = req_capsule_client_get(&req->rq_pill,
1808                                                &RMF_MDS_HSM_ARCHIVE);
1809         if (archive_array == NULL)
1810                 GOTO(out, rc = -EPROTO);
1811
1812         if (archives != NULL)
1813                 memcpy(archive_array, archives, archives_size);
1814         else
1815                 *archive_array = archive_count;
1816
1817         ptlrpc_request_set_replen(req);
1818
1819         rc = mdc_queue_wait(req);
1820         GOTO(out, rc);
1821 out:
1822         ptlrpc_req_finished(req);
1823         return rc;
1824 }
1825
1826 static int mdc_ioc_hsm_current_action(struct obd_export *exp,
1827                                       struct md_op_data *op_data)
1828 {
1829         struct hsm_current_action       *hca = op_data->op_data;
1830         struct hsm_current_action       *req_hca;
1831         struct ptlrpc_request           *req;
1832         int                              rc;
1833         ENTRY;
1834
1835         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1836                                    &RQF_MDS_HSM_ACTION);
1837         if (req == NULL)
1838                 RETURN(-ENOMEM);
1839
1840         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_ACTION);
1841         if (rc) {
1842                 ptlrpc_request_free(req);
1843                 RETURN(rc);
1844         }
1845
1846         mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1847                       op_data->op_suppgids[0], 0);
1848
1849         ptlrpc_request_set_replen(req);
1850
1851         rc = mdc_queue_wait(req);
1852         if (rc)
1853                 GOTO(out, rc);
1854
1855         req_hca = req_capsule_server_get(&req->rq_pill,
1856                                          &RMF_MDS_HSM_CURRENT_ACTION);
1857         if (req_hca == NULL)
1858                 GOTO(out, rc = -EPROTO);
1859
1860         *hca = *req_hca;
1861
1862         EXIT;
1863 out:
1864         ptlrpc_req_finished(req);
1865         return rc;
1866 }
1867
1868 static int mdc_ioc_hsm_ct_unregister(struct obd_import *imp)
1869 {
1870         struct ptlrpc_request   *req;
1871         int                      rc;
1872         ENTRY;
1873
1874         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_UNREGISTER,
1875                                         LUSTRE_MDS_VERSION,
1876                                         MDS_HSM_CT_UNREGISTER);
1877         if (req == NULL)
1878                 GOTO(out, rc = -ENOMEM);
1879
1880         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1881
1882         ptlrpc_request_set_replen(req);
1883
1884         rc = mdc_queue_wait(req);
1885         GOTO(out, rc);
1886 out:
1887         ptlrpc_req_finished(req);
1888         return rc;
1889 }
1890
1891 static int mdc_ioc_hsm_state_get(struct obd_export *exp,
1892                                  struct md_op_data *op_data)
1893 {
1894         struct hsm_user_state   *hus = op_data->op_data;
1895         struct hsm_user_state   *req_hus;
1896         struct ptlrpc_request   *req;
1897         int                      rc;
1898         ENTRY;
1899
1900         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1901                                    &RQF_MDS_HSM_STATE_GET);
1902         if (req == NULL)
1903                 RETURN(-ENOMEM);
1904
1905         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_GET);
1906         if (rc != 0) {
1907                 ptlrpc_request_free(req);
1908                 RETURN(rc);
1909         }
1910
1911         mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1912                       op_data->op_suppgids[0], 0);
1913
1914         ptlrpc_request_set_replen(req);
1915
1916         rc = mdc_queue_wait(req);
1917         if (rc)
1918                 GOTO(out, rc);
1919
1920         req_hus = req_capsule_server_get(&req->rq_pill, &RMF_HSM_USER_STATE);
1921         if (req_hus == NULL)
1922                 GOTO(out, rc = -EPROTO);
1923
1924         *hus = *req_hus;
1925
1926         EXIT;
1927 out:
1928         ptlrpc_req_finished(req);
1929         return rc;
1930 }
1931
1932 static int mdc_ioc_hsm_state_set(struct obd_export *exp,
1933                                  struct md_op_data *op_data)
1934 {
1935         struct hsm_state_set    *hss = op_data->op_data;
1936         struct hsm_state_set    *req_hss;
1937         struct ptlrpc_request   *req;
1938         int                      rc;
1939         ENTRY;
1940
1941         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1942                                    &RQF_MDS_HSM_STATE_SET);
1943         if (req == NULL)
1944                 RETURN(-ENOMEM);
1945
1946         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_SET);
1947         if (rc) {
1948                 ptlrpc_request_free(req);
1949                 RETURN(rc);
1950         }
1951
1952         mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1953                       op_data->op_suppgids[0], 0);
1954
1955         /* Copy states */
1956         req_hss = req_capsule_client_get(&req->rq_pill, &RMF_HSM_STATE_SET);
1957         if (req_hss == NULL)
1958                 GOTO(out, rc = -EPROTO);
1959         *req_hss = *hss;
1960
1961         ptlrpc_request_set_replen(req);
1962
1963         ptlrpc_get_mod_rpc_slot(req);
1964         rc = ptlrpc_queue_wait(req);
1965         ptlrpc_put_mod_rpc_slot(req);
1966
1967         GOTO(out, rc);
1968 out:
1969         ptlrpc_req_finished(req);
1970         return rc;
1971 }
1972
1973 static int mdc_ioc_hsm_request(struct obd_export *exp,
1974                                struct hsm_user_request *hur)
1975 {
1976         struct obd_import       *imp = class_exp2cliimp(exp);
1977         struct ptlrpc_request   *req;
1978         struct hsm_request      *req_hr;
1979         struct hsm_user_item    *req_hui;
1980         char                    *req_opaque;
1981         int                      rc;
1982         ENTRY;
1983
1984         req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_REQUEST);
1985         if (req == NULL)
1986                 GOTO(out, rc = -ENOMEM);
1987
1988         req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM, RCL_CLIENT,
1989                              hur->hur_request.hr_itemcount
1990                              * sizeof(struct hsm_user_item));
1991         req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA, RCL_CLIENT,
1992                              hur->hur_request.hr_data_len);
1993
1994         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_REQUEST);
1995         if (rc) {
1996                 ptlrpc_request_free(req);
1997                 RETURN(rc);
1998         }
1999
2000         mdc_pack_body(req, NULL, 0, 0, -1, 0);
2001
2002         /* Copy hsm_request struct */
2003         req_hr = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_REQUEST);
2004         if (req_hr == NULL)
2005                 GOTO(out, rc = -EPROTO);
2006         *req_hr = hur->hur_request;
2007
2008         /* Copy hsm_user_item structs */
2009         req_hui = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM);
2010         if (req_hui == NULL)
2011                 GOTO(out, rc = -EPROTO);
2012         memcpy(req_hui, hur->hur_user_item,
2013                hur->hur_request.hr_itemcount * sizeof(struct hsm_user_item));
2014
2015         /* Copy opaque field */
2016         req_opaque = req_capsule_client_get(&req->rq_pill, &RMF_GENERIC_DATA);
2017         if (req_opaque == NULL)
2018                 GOTO(out, rc = -EPROTO);
2019         memcpy(req_opaque, hur_data(hur), hur->hur_request.hr_data_len);
2020
2021         ptlrpc_request_set_replen(req);
2022
2023         ptlrpc_get_mod_rpc_slot(req);
2024         rc = ptlrpc_queue_wait(req);
2025         ptlrpc_put_mod_rpc_slot(req);
2026
2027         GOTO(out, rc);
2028
2029 out:
2030         ptlrpc_req_finished(req);
2031         return rc;
2032 }
2033
2034 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2035                                 struct lustre_kernelcomm *lk);
2036
2037 static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp,
2038                         struct obd_quotactl *oqctl)
2039 {
2040         struct ptlrpc_request   *req;
2041         struct obd_quotactl     *oqc;
2042         int                      rc;
2043         ENTRY;
2044
2045         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
2046                                         &RQF_MDS_QUOTACTL, LUSTRE_MDS_VERSION,
2047                                         MDS_QUOTACTL);
2048         if (req == NULL)
2049                 RETURN(-ENOMEM);
2050
2051         oqc = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2052         *oqc = *oqctl;
2053
2054         ptlrpc_request_set_replen(req);
2055         ptlrpc_at_set_req_timeout(req);
2056
2057         rc = ptlrpc_queue_wait(req);
2058         if (rc)
2059                 CERROR("ptlrpc_queue_wait failed, rc: %d\n", rc);
2060
2061         if (req->rq_repmsg &&
2062             (oqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL))) {
2063                 *oqctl = *oqc;
2064         } else if (!rc) {
2065                 CERROR ("Can't unpack obd_quotactl\n");
2066                 rc = -EPROTO;
2067         }
2068         ptlrpc_req_finished(req);
2069
2070         RETURN(rc);
2071 }
2072
2073 static int mdc_ioc_swap_layouts(struct obd_export *exp,
2074                                 struct md_op_data *op_data)
2075 {
2076         LIST_HEAD(cancels);
2077         struct ptlrpc_request   *req;
2078         int                      rc, count;
2079         struct mdc_swap_layouts *msl, *payload;
2080         ENTRY;
2081
2082         msl = op_data->op_data;
2083
2084         /* When the MDT will get the MDS_SWAP_LAYOUTS RPC the
2085          * first thing it will do is to cancel the 2 layout
2086          * locks held by this client.
2087          * So the client must cancel its layout locks on the 2 fids
2088          * with the request RPC to avoid extra RPC round trips.
2089          */
2090         count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels,
2091                                         LCK_EX, MDS_INODELOCK_LAYOUT |
2092                                         MDS_INODELOCK_XATTR);
2093         count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels,
2094                                          LCK_EX, MDS_INODELOCK_LAYOUT |
2095                                          MDS_INODELOCK_XATTR);
2096
2097         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2098                                    &RQF_MDS_SWAP_LAYOUTS);
2099         if (req == NULL) {
2100                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
2101                 RETURN(-ENOMEM);
2102         }
2103
2104         rc = mdc_prep_elc_req(exp, req, MDS_SWAP_LAYOUTS, &cancels, count);
2105         if (rc) {
2106                 ptlrpc_request_free(req);
2107                 RETURN(rc);
2108         }
2109
2110         mdc_swap_layouts_pack(req, op_data);
2111
2112         payload = req_capsule_client_get(&req->rq_pill, &RMF_SWAP_LAYOUTS);
2113         LASSERT(payload);
2114
2115         *payload = *msl;
2116
2117         ptlrpc_request_set_replen(req);
2118
2119         rc = ptlrpc_queue_wait(req);
2120         if (rc)
2121                 GOTO(out, rc);
2122         EXIT;
2123
2124 out:
2125         ptlrpc_req_finished(req);
2126         return rc;
2127 }
2128
2129 static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2130                          void *karg, void __user *uarg)
2131 {
2132         struct obd_device *obd = exp->exp_obd;
2133         struct obd_ioctl_data *data = karg;
2134         struct obd_import *imp = obd->u.cli.cl_import;
2135         int rc;
2136         ENTRY;
2137
2138         if (!try_module_get(THIS_MODULE)) {
2139                 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2140                        module_name(THIS_MODULE));
2141                 return -EINVAL;
2142         }
2143         switch (cmd) {
2144         case OBD_IOC_FID2PATH:
2145                 rc = mdc_ioc_fid2path(exp, karg);
2146                 GOTO(out, rc);
2147         case LL_IOC_HSM_CT_START:
2148                 rc = mdc_ioc_hsm_ct_start(exp, karg);
2149                 /* ignore if it was already registered on this MDS. */
2150                 if (rc == -EEXIST)
2151                         rc = 0;
2152                 GOTO(out, rc);
2153         case LL_IOC_HSM_PROGRESS:
2154                 rc = mdc_ioc_hsm_progress(exp, karg);
2155                 GOTO(out, rc);
2156         case LL_IOC_HSM_STATE_GET:
2157                 rc = mdc_ioc_hsm_state_get(exp, karg);
2158                 GOTO(out, rc);
2159         case LL_IOC_HSM_STATE_SET:
2160                 rc = mdc_ioc_hsm_state_set(exp, karg);
2161                 GOTO(out, rc);
2162         case LL_IOC_HSM_ACTION:
2163                 rc = mdc_ioc_hsm_current_action(exp, karg);
2164                 GOTO(out, rc);
2165         case LL_IOC_HSM_REQUEST:
2166                 rc = mdc_ioc_hsm_request(exp, karg);
2167                 GOTO(out, rc);
2168         case OBD_IOC_CLIENT_RECOVER:
2169                 rc = ptlrpc_recover_import(imp, data->ioc_inlbuf1, 0);
2170                 if (rc < 0)
2171                         GOTO(out, rc);
2172                 GOTO(out, rc = 0);
2173         case IOC_OSC_SET_ACTIVE:
2174                 rc = ptlrpc_set_import_active(imp, data->ioc_offset);
2175                 GOTO(out, rc);
2176         /*
2177          * Normally IOC_OBD_STATFS, OBD_IOC_QUOTACTL iocontrol are handled by
2178          * LMV instead of MDC. But when the cluster is upgraded from 1.8,
2179          * there'd be no LMV layer thus we might be called here. Eventually
2180          * this code should be removed.
2181          * bz20731, LU-592.
2182          */
2183         case IOC_OBD_STATFS: {
2184                 struct obd_statfs stat_buf = {0};
2185
2186                 if (*((__u32 *) data->ioc_inlbuf2) != 0)
2187                         GOTO(out, rc = -ENODEV);
2188
2189                 /* copy UUID */
2190                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(obd),
2191                                  min((int)data->ioc_plen2,
2192                                      (int)sizeof(struct obd_uuid))))
2193                         GOTO(out, rc = -EFAULT);
2194
2195                 rc = mdc_statfs(NULL, obd->obd_self_export, &stat_buf,
2196                                 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
2197                                 0);
2198                 if (rc != 0)
2199                         GOTO(out, rc);
2200
2201                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
2202                                      min((int) data->ioc_plen1,
2203                                          (int) sizeof(stat_buf))))
2204                         GOTO(out, rc = -EFAULT);
2205
2206                 GOTO(out, rc = 0);
2207         }
2208         case OBD_IOC_QUOTACTL: {
2209                 struct if_quotactl *qctl = karg;
2210                 struct obd_quotactl *oqctl;
2211
2212                 OBD_ALLOC_PTR(oqctl);
2213                 if (oqctl == NULL)
2214                         GOTO(out, rc = -ENOMEM);
2215
2216                 QCTL_COPY(oqctl, qctl);
2217                 rc = obd_quotactl(exp, oqctl);
2218                 if (rc == 0) {
2219                         QCTL_COPY(qctl, oqctl);
2220                         qctl->qc_valid = QC_MDTIDX;
2221                         qctl->obd_uuid = obd->u.cli.cl_target_uuid;
2222                 }
2223
2224                 OBD_FREE_PTR(oqctl);
2225                 GOTO(out, rc);
2226         }
2227         case LL_IOC_GET_CONNECT_FLAGS:
2228                 if (copy_to_user(uarg, exp_connect_flags_ptr(exp),
2229                                  sizeof(*exp_connect_flags_ptr(exp))))
2230                         GOTO(out, rc = -EFAULT);
2231
2232                 GOTO(out, rc = 0);
2233         case LL_IOC_LOV_SWAP_LAYOUTS:
2234                 rc = mdc_ioc_swap_layouts(exp, karg);
2235                 GOTO(out, rc);
2236         default:
2237                 CERROR("unrecognised ioctl: cmd = %#x\n", cmd);
2238                 GOTO(out, rc = -ENOTTY);
2239         }
2240 out:
2241         module_put(THIS_MODULE);
2242
2243         return rc;
2244 }
2245
2246 static int mdc_get_info_rpc(struct obd_export *exp,
2247                             u32 keylen, void *key,
2248                             u32 vallen, void *val)
2249 {
2250         struct obd_import      *imp = class_exp2cliimp(exp);
2251         struct ptlrpc_request  *req;
2252         char                   *tmp;
2253         int                     rc = -EINVAL;
2254         ENTRY;
2255
2256         req = ptlrpc_request_alloc(imp, &RQF_MDS_GET_INFO);
2257         if (req == NULL)
2258                 RETURN(-ENOMEM);
2259
2260         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY,
2261                              RCL_CLIENT, keylen);
2262         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VALLEN,
2263                              RCL_CLIENT, sizeof(vallen));
2264
2265         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_INFO);
2266         if (rc) {
2267                 ptlrpc_request_free(req);
2268                 RETURN(rc);
2269         }
2270
2271         tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
2272         memcpy(tmp, key, keylen);
2273         tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_VALLEN);
2274         memcpy(tmp, &vallen, sizeof(vallen));
2275
2276         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VAL,
2277                              RCL_SERVER, vallen);
2278         ptlrpc_request_set_replen(req);
2279
2280         rc = ptlrpc_queue_wait(req);
2281         /* -EREMOTE means the get_info result is partial, and it needs to
2282          * continue on another MDT, see fid2path part in lmv_iocontrol */
2283         if (rc == 0 || rc == -EREMOTE) {
2284                 tmp = req_capsule_server_get(&req->rq_pill, &RMF_GETINFO_VAL);
2285                 memcpy(val, tmp, vallen);
2286                 if (ptlrpc_rep_need_swab(req)) {
2287                         if (KEY_IS(KEY_FID2PATH))
2288                                 lustre_swab_fid2path(val);
2289                 }
2290         }
2291         ptlrpc_req_finished(req);
2292
2293         RETURN(rc);
2294 }
2295
2296 static void lustre_swab_hai(struct hsm_action_item *h)
2297 {
2298         __swab32s(&h->hai_len);
2299         __swab32s(&h->hai_action);
2300         lustre_swab_lu_fid(&h->hai_fid);
2301         lustre_swab_lu_fid(&h->hai_dfid);
2302         __swab64s(&h->hai_cookie);
2303         __swab64s(&h->hai_extent.offset);
2304         __swab64s(&h->hai_extent.length);
2305         __swab64s(&h->hai_gid);
2306 }
2307
2308 static void lustre_swab_hal(struct hsm_action_list *h)
2309 {
2310         struct hsm_action_item  *hai;
2311         __u32                    i;
2312
2313         __swab32s(&h->hal_version);
2314         __swab32s(&h->hal_count);
2315         __swab32s(&h->hal_archive_id);
2316         __swab64s(&h->hal_flags);
2317         hai = hai_first(h);
2318         for (i = 0; i < h->hal_count; i++, hai = hai_next(hai))
2319                 lustre_swab_hai(hai);
2320 }
2321
2322 static void lustre_swab_kuch(struct kuc_hdr *l)
2323 {
2324         __swab16s(&l->kuc_magic);
2325         /* __u8 l->kuc_transport */
2326         __swab16s(&l->kuc_msgtype);
2327         __swab16s(&l->kuc_msglen);
2328 }
2329
2330 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2331                                 struct lustre_kernelcomm *lk)
2332 {
2333         struct obd_import *imp = class_exp2cliimp(exp);
2334         int rc = 0;
2335
2336         if (lk->lk_group != KUC_GRP_HSM) {
2337                 CERROR("Bad copytool group %d\n", lk->lk_group);
2338                 return -EINVAL;
2339         }
2340
2341         CDEBUG(D_HSM, "CT start r%d w%d u%d g%d f%#x\n", lk->lk_rfd, lk->lk_wfd,
2342                lk->lk_uid, lk->lk_group, lk->lk_flags);
2343
2344         if (lk->lk_flags & LK_FLG_STOP) {
2345                 /* Unregister with the coordinator */
2346                 rc = mdc_ioc_hsm_ct_unregister(imp);
2347         } else {
2348                 __u32 *archives = NULL;
2349
2350                 if ((lk->lk_flags & LK_FLG_DATANR) && lk->lk_data_count > 0)
2351                         archives = lk->lk_data;
2352
2353                 rc = mdc_ioc_hsm_ct_register(imp, lk->lk_data_count, archives);
2354         }
2355
2356         return rc;
2357 }
2358
2359 /**
2360  * Send a message to any listening copytools
2361  * @param val KUC message (kuc_hdr + hsm_action_list)
2362  * @param len total length of message
2363  */
2364 static int mdc_hsm_copytool_send(const struct obd_uuid *uuid,
2365                                  size_t len, void *val)
2366 {
2367         struct kuc_hdr          *lh = (struct kuc_hdr *)val;
2368         struct hsm_action_list  *hal = (struct hsm_action_list *)(lh + 1);
2369         int                      rc;
2370         ENTRY;
2371
2372         if (len < sizeof(*lh) + sizeof(*hal)) {
2373                 CERROR("Short HSM message %zu < %zu\n", len,
2374                        sizeof(*lh) + sizeof(*hal));
2375                 RETURN(-EPROTO);
2376         }
2377         if (lh->kuc_magic == __swab16(KUC_MAGIC)) {
2378                 lustre_swab_kuch(lh);
2379                 lustre_swab_hal(hal);
2380         } else if (lh->kuc_magic != KUC_MAGIC) {
2381                 CERROR("Bad magic %x!=%x\n", lh->kuc_magic, KUC_MAGIC);
2382                 RETURN(-EPROTO);
2383         }
2384
2385         CDEBUG(D_HSM, " Received message mg=%x t=%d m=%d l=%d actions=%d "
2386                "on %s\n",
2387                lh->kuc_magic, lh->kuc_transport, lh->kuc_msgtype,
2388                lh->kuc_msglen, hal->hal_count, hal->hal_fsname);
2389
2390         /* Broadcast to HSM listeners */
2391         rc = libcfs_kkuc_group_put(uuid, KUC_GRP_HSM, lh);
2392
2393         RETURN(rc);
2394 }
2395
2396 /**
2397  * callback function passed to kuc for re-registering each HSM copytool
2398  * running on MDC, after MDT shutdown/recovery.
2399  * @param data copytool registration data
2400  * @param cb_arg callback argument (obd_import)
2401  */
2402 static int mdc_hsm_ct_reregister(void *data, void *cb_arg)
2403 {
2404         struct obd_import *imp = (struct obd_import *)cb_arg;
2405         struct kkuc_ct_data *kcd = data;
2406         __u32 *archives = NULL;
2407         int rc;
2408
2409         if (kcd == NULL ||
2410             (kcd->kcd_magic != KKUC_CT_DATA_ARRAY_MAGIC &&
2411              kcd->kcd_magic != KKUC_CT_DATA_BITMAP_MAGIC))
2412                 return -EPROTO;
2413
2414         if (kcd->kcd_magic == KKUC_CT_DATA_BITMAP_MAGIC) {
2415                 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2416                        "(archive=%#x)\n", imp->imp_obd->obd_name,
2417                        kcd->kcd_nr_archives);
2418         } else {
2419                 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2420                        "(archive nr = %u)\n",
2421                        imp->imp_obd->obd_name, kcd->kcd_nr_archives);
2422                 if (kcd->kcd_nr_archives != 0)
2423                         archives = kcd->kcd_archives;
2424         }
2425
2426         rc = mdc_ioc_hsm_ct_register(imp, kcd->kcd_nr_archives, archives);
2427         /* ignore error if the copytool is already registered */
2428         return (rc == -EEXIST) ? 0 : rc;
2429 }
2430
2431 /**
2432  * Re-establish all kuc contexts with MDT
2433  * after MDT shutdown/recovery.
2434  */
2435 static int mdc_kuc_reregister(struct obd_import *imp)
2436 {
2437         /* re-register HSM agents */
2438         return libcfs_kkuc_group_foreach(&imp->imp_obd->obd_uuid, KUC_GRP_HSM,
2439                                          mdc_hsm_ct_reregister, imp);
2440 }
2441
2442 static int mdc_set_info_async(const struct lu_env *env,
2443                               struct obd_export *exp,
2444                               u32 keylen, void *key,
2445                               u32 vallen, void *val,
2446                               struct ptlrpc_request_set *set)
2447 {
2448         struct obd_import       *imp = class_exp2cliimp(exp);
2449         int                      rc;
2450         ENTRY;
2451
2452         if (KEY_IS(KEY_READ_ONLY)) {
2453                 if (vallen != sizeof(int))
2454                         RETURN(-EINVAL);
2455
2456                 spin_lock(&imp->imp_lock);
2457                 if (*((int *)val)) {
2458                         imp->imp_connect_flags_orig |= OBD_CONNECT_RDONLY;
2459                         imp->imp_connect_data.ocd_connect_flags |=
2460                                                         OBD_CONNECT_RDONLY;
2461                 } else {
2462                         imp->imp_connect_flags_orig &= ~OBD_CONNECT_RDONLY;
2463                         imp->imp_connect_data.ocd_connect_flags &=
2464                                                         ~OBD_CONNECT_RDONLY;
2465                 }
2466                 spin_unlock(&imp->imp_lock);
2467
2468                 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2469                                        keylen, key, vallen, val, set);
2470                 RETURN(rc);
2471         }
2472         if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2473                 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2474                                        keylen, key, vallen, val, set);
2475                 RETURN(rc);
2476         }
2477         if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) {
2478                 rc = mdc_hsm_copytool_send(&imp->imp_obd->obd_uuid, vallen,
2479                                            val);
2480                 RETURN(rc);
2481         }
2482
2483         if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2484                 __u32 *default_easize = val;
2485
2486                 exp->exp_obd->u.cli.cl_default_mds_easize = *default_easize;
2487                 RETURN(0);
2488         }
2489
2490         rc = osc_set_info_async(env, exp, keylen, key, vallen, val, set);
2491         RETURN(rc);
2492 }
2493
2494 static int mdc_get_info(const struct lu_env *env, struct obd_export *exp,
2495                         __u32 keylen, void *key, __u32 *vallen, void *val)
2496 {
2497         int rc = -EINVAL;
2498
2499         if (KEY_IS(KEY_MAX_EASIZE)) {
2500                 __u32 mdsize, *max_easize;
2501
2502                 if (*vallen != sizeof(int))
2503                         RETURN(-EINVAL);
2504                 mdsize = *(__u32 *)val;
2505                 if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
2506                         exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
2507                 max_easize = val;
2508                 *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
2509                 RETURN(0);
2510         } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2511                 __u32 *default_easize;
2512
2513                 if (*vallen != sizeof(int))
2514                         RETURN(-EINVAL);
2515                 default_easize = val;
2516                 *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
2517                 RETURN(0);
2518         } else if (KEY_IS(KEY_CONN_DATA)) {
2519                 struct obd_import *imp = class_exp2cliimp(exp);
2520                 struct obd_connect_data *data = val;
2521
2522                 if (*vallen != sizeof(*data))
2523                         RETURN(-EINVAL);
2524
2525                 *data = imp->imp_connect_data;
2526                 RETURN(0);
2527         } else if (KEY_IS(KEY_TGT_COUNT)) {
2528                 *((__u32 *)val) = 1;
2529                 RETURN(0);
2530         }
2531
2532         rc = mdc_get_info_rpc(exp, keylen, key, *vallen, val);
2533
2534         RETURN(rc);
2535 }
2536
2537 static int mdc_fsync(struct obd_export *exp, const struct lu_fid *fid,
2538                      struct ptlrpc_request **request)
2539 {
2540         struct ptlrpc_request *req;
2541         int                    rc;
2542         ENTRY;
2543
2544         *request = NULL;
2545         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SYNC);
2546         if (req == NULL)
2547                 RETURN(-ENOMEM);
2548
2549         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SYNC);
2550         if (rc) {
2551                 ptlrpc_request_free(req);
2552                 RETURN(rc);
2553         }
2554
2555         mdc_pack_body(req, fid, 0, 0, -1, 0);
2556
2557         ptlrpc_request_set_replen(req);
2558
2559         rc = ptlrpc_queue_wait(req);
2560         if (rc)
2561                 ptlrpc_req_finished(req);
2562         else
2563                 *request = req;
2564         RETURN(rc);
2565 }
2566
2567 struct mdc_rmfid_args {
2568         int *mra_rcs;
2569         int mra_nr;
2570 };
2571
2572 int mdc_rmfid_interpret(const struct lu_env *env, struct ptlrpc_request *req,
2573                           void *args, int rc)
2574 {
2575         struct mdc_rmfid_args *aa;
2576         int *rcs, size;
2577         ENTRY;
2578
2579         if (!rc) {
2580                 aa = ptlrpc_req_async_args(aa, req);
2581
2582                 size = req_capsule_get_size(&req->rq_pill, &RMF_RCS,
2583                                             RCL_SERVER);
2584                 LASSERT(size == sizeof(int) * aa->mra_nr);
2585                 rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2586                 LASSERT(rcs);
2587                 LASSERT(aa->mra_rcs);
2588                 LASSERT(aa->mra_nr);
2589                 memcpy(aa->mra_rcs, rcs, size);
2590         }
2591
2592         RETURN(rc);
2593 }
2594
2595 static int mdc_rmfid(struct obd_export *exp, struct fid_array *fa,
2596                      int *rcs, struct ptlrpc_request_set *set)
2597 {
2598         struct ptlrpc_request *req;
2599         struct mdc_rmfid_args *aa;
2600         struct mdt_body *b;
2601         struct lu_fid *tmp;
2602         int rc, flen;
2603         ENTRY;
2604
2605         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_RMFID);
2606         if (req == NULL)
2607                 RETURN(-ENOMEM);
2608
2609         flen = fa->fa_nr * sizeof(struct lu_fid);
2610         req_capsule_set_size(&req->rq_pill, &RMF_FID_ARRAY,
2611                              RCL_CLIENT, flen);
2612         req_capsule_set_size(&req->rq_pill, &RMF_FID_ARRAY,
2613                              RCL_SERVER, flen);
2614         req_capsule_set_size(&req->rq_pill, &RMF_RCS,
2615                              RCL_SERVER, fa->fa_nr * sizeof(__u32));
2616         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_RMFID);
2617         if (rc) {
2618                 ptlrpc_request_free(req);
2619                 RETURN(rc);
2620         }
2621         tmp = req_capsule_client_get(&req->rq_pill, &RMF_FID_ARRAY);
2622         memcpy(tmp, fa->fa_fids, flen);
2623
2624         mdc_pack_body(req, NULL, 0, 0, -1, 0);
2625         b = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
2626         b->mbo_ctime = ktime_get_real_seconds();
2627
2628         ptlrpc_request_set_replen(req);
2629
2630         LASSERT(rcs);
2631         aa = ptlrpc_req_async_args(aa, req);
2632         aa->mra_rcs = rcs;
2633         aa->mra_nr = fa->fa_nr;
2634         req->rq_interpret_reply = mdc_rmfid_interpret;
2635
2636         ptlrpc_set_add_req(set, req);
2637         ptlrpc_check_set(NULL, set);
2638
2639         RETURN(rc);
2640 }
2641
2642 static int mdc_import_event(struct obd_device *obd, struct obd_import *imp,
2643                             enum obd_import_event event)
2644 {
2645         struct client_obd *cli = &obd->u.cli;
2646         int rc = 0;
2647
2648         LASSERT(imp->imp_obd == obd);
2649
2650         switch (event) {
2651         case IMP_EVENT_DISCON:
2652                 spin_lock(&cli->cl_loi_list_lock);
2653                 cli->cl_avail_grant = 0;
2654                 cli->cl_lost_grant = 0;
2655                 spin_unlock(&cli->cl_loi_list_lock);
2656                 break;
2657         case IMP_EVENT_INACTIVE:
2658                 /*
2659                  * Flush current sequence to make client obtain new one
2660                  * from server in case of disconnect/reconnect.
2661                  */
2662                 down_read(&cli->cl_seq_rwsem);
2663                 if (cli->cl_seq)
2664                         seq_client_flush(cli->cl_seq);
2665                 up_read(&cli->cl_seq_rwsem);
2666
2667                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
2668                 break;
2669         case IMP_EVENT_INVALIDATE: {
2670                 struct ldlm_namespace *ns = obd->obd_namespace;
2671                 struct lu_env *env;
2672                 __u16 refcheck;
2673
2674                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2675
2676                 env = cl_env_get(&refcheck);
2677                 if (!IS_ERR(env)) {
2678                         /* Reset grants. All pages go to failing rpcs due to
2679                          * the invalid import.
2680                          */
2681                         osc_io_unplug(env, cli, NULL);
2682
2683                         cfs_hash_for_each_nolock(ns->ns_rs_hash,
2684                                                  osc_ldlm_resource_invalidate,
2685                                                  env, 0);
2686                         cl_env_put(env, &refcheck);
2687                         ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2688                 } else {
2689                         rc = PTR_ERR(env);
2690                 }
2691                 break;
2692         }
2693         case IMP_EVENT_ACTIVE:
2694                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
2695                 /* redo the kuc registration after reconnecting */
2696                 if (rc == 0)
2697                         rc = mdc_kuc_reregister(imp);
2698                 break;
2699         case IMP_EVENT_OCD: {
2700                 struct obd_connect_data *ocd = &imp->imp_connect_data;
2701
2702                 if (OCD_HAS_FLAG(ocd, GRANT))
2703                         osc_init_grant(cli, ocd);
2704
2705                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
2706                 break;
2707         }
2708         case IMP_EVENT_DEACTIVATE:
2709         case IMP_EVENT_ACTIVATE:
2710                 break;
2711         default:
2712                 CERROR("Unknown import event %x\n", event);
2713                 LBUG();
2714         }
2715         RETURN(rc);
2716 }
2717
2718 int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp,
2719                   struct lu_fid *fid, struct md_op_data *op_data)
2720 {
2721         struct client_obd *cli = &exp->exp_obd->u.cli;
2722         int rc = -EIO;
2723
2724         ENTRY;
2725
2726         down_read(&cli->cl_seq_rwsem);
2727         if (cli->cl_seq)
2728                 rc = seq_client_alloc_fid(env, cli->cl_seq, fid);
2729         up_read(&cli->cl_seq_rwsem);
2730
2731         RETURN(rc);
2732 }
2733
2734 static struct obd_uuid *mdc_get_uuid(struct obd_export *exp)
2735 {
2736         struct client_obd *cli = &exp->exp_obd->u.cli;
2737         return &cli->cl_target_uuid;
2738 }
2739
2740 /**
2741  * Determine whether the lock can be canceled before replaying it during
2742  * recovery, non zero value will be return if the lock can be canceled,
2743  * or zero returned for not
2744  */
2745 static int mdc_cancel_weight(struct ldlm_lock *lock)
2746 {
2747         if (lock->l_resource->lr_type != LDLM_IBITS)
2748                 RETURN(0);
2749
2750         /* FIXME: if we ever get into a situation where there are too many
2751          * opened files with open locks on a single node, then we really
2752          * should replay these open locks to reget it */
2753         if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN)
2754                 RETURN(0);
2755
2756         /* Special case for DoM locks, cancel only unused and granted locks */
2757         if (ldlm_has_dom(lock) &&
2758             (lock->l_granted_mode != lock->l_req_mode ||
2759              osc_ldlm_weigh_ast(lock) != 0))
2760                 RETURN(0);
2761
2762         RETURN(1);
2763 }
2764
2765 static int mdc_resource_inode_free(struct ldlm_resource *res)
2766 {
2767         if (res->lr_lvb_inode)
2768                 res->lr_lvb_inode = NULL;
2769
2770         return 0;
2771 }
2772
2773 static struct ldlm_valblock_ops inode_lvbo = {
2774         .lvbo_free = mdc_resource_inode_free
2775 };
2776
2777 static int mdc_llog_init(struct obd_device *obd)
2778 {
2779         struct obd_llog_group   *olg = &obd->obd_olg;
2780         struct llog_ctxt        *ctxt;
2781         int                      rc;
2782
2783         ENTRY;
2784
2785         rc = llog_setup(NULL, obd, olg, LLOG_CHANGELOG_REPL_CTXT, obd,
2786                         &llog_client_ops);
2787         if (rc < 0)
2788                 RETURN(rc);
2789
2790         ctxt = llog_group_get_ctxt(olg, LLOG_CHANGELOG_REPL_CTXT);
2791         llog_initiator_connect(ctxt);
2792         llog_ctxt_put(ctxt);
2793
2794         RETURN(0);
2795 }
2796
2797 static void mdc_llog_finish(struct obd_device *obd)
2798 {
2799         struct llog_ctxt *ctxt;
2800
2801         ENTRY;
2802
2803         ctxt = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
2804         if (ctxt != NULL)
2805                 llog_cleanup(NULL, ctxt);
2806
2807         EXIT;
2808 }
2809
2810 int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg)
2811 {
2812         int rc;
2813
2814         ENTRY;
2815
2816         rc = osc_setup_common(obd, cfg);
2817         if (rc < 0)
2818                 RETURN(rc);
2819
2820         rc = mdc_tunables_init(obd);
2821         if (rc)
2822                 GOTO(err_osc_cleanup, rc);
2823
2824         obd->u.cli.cl_dom_min_inline_repsize = MDC_DOM_DEF_INLINE_REPSIZE;
2825
2826         ns_register_cancel(obd->obd_namespace, mdc_cancel_weight);
2827
2828         obd->obd_namespace->ns_lvbo = &inode_lvbo;
2829
2830         rc = mdc_llog_init(obd);
2831         if (rc) {
2832                 CERROR("%s: failed to setup llogging subsystems: rc = %d\n",
2833                        obd->obd_name, rc);
2834                 GOTO(err_llog_cleanup, rc);
2835         }
2836
2837         rc = mdc_changelog_cdev_init(obd);
2838         if (rc) {
2839                 CERROR("%s: failed to setup changelog char device: rc = %d\n",
2840                        obd->obd_name, rc);
2841                 GOTO(err_changelog_cleanup, rc);
2842         }
2843
2844         RETURN(rc);
2845
2846 err_changelog_cleanup:
2847         mdc_llog_finish(obd);
2848 err_llog_cleanup:
2849         lprocfs_free_md_stats(obd);
2850         ptlrpc_lprocfs_unregister_obd(obd);
2851 err_osc_cleanup:
2852         osc_cleanup_common(obd);
2853         return rc;
2854 }
2855
2856 /* Initialize the default and maximum LOV EA sizes.  This allows
2857  * us to make MDS RPCs with large enough reply buffers to hold a default
2858  * sized EA without having to calculate this (via a call into the
2859  * LOV + OSCs) each time we make an RPC.  The maximum size is also tracked
2860  * but not used to avoid wastefully vmalloc()'ing large reply buffers when
2861  * a large number of stripes is possible.  If a larger reply buffer is
2862  * required it will be reallocated in the ptlrpc layer due to overflow.
2863  */
2864 static int mdc_init_ea_size(struct obd_export *exp, __u32 easize,
2865                             __u32 def_easize)
2866 {
2867         struct obd_device *obd = exp->exp_obd;
2868         struct client_obd *cli = &obd->u.cli;
2869         ENTRY;
2870
2871         if (cli->cl_max_mds_easize < easize)
2872                 cli->cl_max_mds_easize = easize;
2873
2874         if (cli->cl_default_mds_easize < def_easize)
2875                 cli->cl_default_mds_easize = def_easize;
2876
2877         RETURN(0);
2878 }
2879
2880 static int mdc_precleanup(struct obd_device *obd)
2881 {
2882         ENTRY;
2883
2884         osc_precleanup_common(obd);
2885         mdc_changelog_cdev_finish(obd);
2886
2887         obd_cleanup_client_import(obd);
2888         ptlrpc_lprocfs_unregister_obd(obd);
2889         lprocfs_free_md_stats(obd);
2890         mdc_llog_finish(obd);
2891         RETURN(0);
2892 }
2893
2894 static int mdc_cleanup(struct obd_device *obd)
2895 {
2896         return osc_cleanup_common(obd);
2897 }
2898
2899 static const struct obd_ops mdc_obd_ops = {
2900         .o_owner            = THIS_MODULE,
2901         .o_setup            = mdc_setup,
2902         .o_precleanup       = mdc_precleanup,
2903         .o_cleanup          = mdc_cleanup,
2904         .o_add_conn         = client_import_add_conn,
2905         .o_del_conn         = client_import_del_conn,
2906         .o_connect          = client_connect_import,
2907         .o_reconnect        = osc_reconnect,
2908         .o_disconnect       = osc_disconnect,
2909         .o_iocontrol        = mdc_iocontrol,
2910         .o_set_info_async   = mdc_set_info_async,
2911         .o_statfs           = mdc_statfs,
2912         .o_statfs_async     = mdc_statfs_async,
2913         .o_fid_init         = client_fid_init,
2914         .o_fid_fini         = client_fid_fini,
2915         .o_fid_alloc        = mdc_fid_alloc,
2916         .o_import_event     = mdc_import_event,
2917         .o_get_info         = mdc_get_info,
2918         .o_get_uuid         = mdc_get_uuid,
2919         .o_quotactl         = mdc_quotactl,
2920 };
2921
2922 static const struct md_ops mdc_md_ops = {
2923         .m_get_root         = mdc_get_root,
2924         .m_null_inode       = mdc_null_inode,
2925         .m_close            = mdc_close,
2926         .m_create           = mdc_create,
2927         .m_enqueue          = mdc_enqueue,
2928         .m_getattr          = mdc_getattr,
2929         .m_getattr_name     = mdc_getattr_name,
2930         .m_intent_lock      = mdc_intent_lock,
2931         .m_link             = mdc_link,
2932         .m_rename           = mdc_rename,
2933         .m_setattr          = mdc_setattr,
2934         .m_setxattr         = mdc_setxattr,
2935         .m_getxattr         = mdc_getxattr,
2936         .m_fsync                = mdc_fsync,
2937         .m_file_resync          = mdc_file_resync,
2938         .m_read_page            = mdc_read_page,
2939         .m_unlink           = mdc_unlink,
2940         .m_cancel_unused    = mdc_cancel_unused,
2941         .m_init_ea_size     = mdc_init_ea_size,
2942         .m_set_lock_data    = mdc_set_lock_data,
2943         .m_lock_match       = mdc_lock_match,
2944         .m_get_lustre_md    = mdc_get_lustre_md,
2945         .m_free_lustre_md   = mdc_free_lustre_md,
2946         .m_set_open_replay_data = mdc_set_open_replay_data,
2947         .m_clear_open_replay_data = mdc_clear_open_replay_data,
2948         .m_intent_getattr_async = mdc_intent_getattr_async,
2949         .m_revalidate_lock      = mdc_revalidate_lock,
2950         .m_rmfid                = mdc_rmfid,
2951 };
2952
2953 static int __init mdc_init(void)
2954 {
2955         return class_register_type(&mdc_obd_ops, &mdc_md_ops, true, NULL,
2956                                    LUSTRE_MDC_NAME, &mdc_device_type);
2957 }
2958
2959 static void __exit mdc_exit(void)
2960 {
2961         class_unregister_type(LUSTRE_MDC_NAME);
2962 }
2963
2964 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2965 MODULE_DESCRIPTION("Lustre Metadata Client");
2966 MODULE_VERSION(LUSTRE_VERSION_STRING);
2967 MODULE_LICENSE("GPL");
2968
2969 module_init(mdc_init);
2970 module_exit(mdc_exit);