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 #ifdef HAVE_XA_IS_VALUE
 546 # include <linux/xarray.h>
 547 #endif
 548
 549 int mdc_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
 550                       struct obd_export *dt_exp, struct obd_export *md_exp,
 551                       struct lustre_md *md)
 552 {
 553         struct req_capsule *pill = &req->rq_pill;
 554         int rc;
 555         ENTRY;
 556
 557         LASSERT(md);
 558         memset(md, 0, sizeof(*md));
 559
 560         md->body = req_capsule_server_get(pill, &RMF_MDT_BODY);
 561         LASSERT(md->body != NULL);
 562
 563         if (md->body->mbo_valid & OBD_MD_FLEASIZE) {
 564                 if (!S_ISREG(md->body->mbo_mode)) {
 565                         CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, should be a "
 566                                "regular file, but is not\n");
 567                         GOTO(out, rc = -EPROTO);
 568                 }
 569
 570                 if (md->body->mbo_eadatasize == 0) {
 571                         CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, "
 572                                "but eadatasize 0\n");
 573                         GOTO(out, rc = -EPROTO);
 574                 }
 575
 576                 md->layout.lb_len = md->body->mbo_eadatasize;
 577                 md->layout.lb_buf = req_capsule_server_sized_get(pill,
 578                                                         &RMF_MDT_MD,
 579                                                         md->layout.lb_len);
 580                 if (md->layout.lb_buf == NULL)
 581                         GOTO(out, rc = -EPROTO);
 582         } else if (md->body->mbo_valid & OBD_MD_FLDIREA) {
 583                 const union lmv_mds_md *lmv;
 584                 size_t lmv_size;
 585
 586                 if (!S_ISDIR(md->body->mbo_mode)) {
 587                         CDEBUG(D_INFO, "OBD_MD_FLDIREA set, should be a "
 588                                "directory, but is not\n");
 589                         GOTO(out, rc = -EPROTO);
 590                 }
 591
 592                 if (md->body->mbo_valid & OBD_MD_MEA) {
 593                         lmv_size = md->body->mbo_eadatasize;
 594                         if (lmv_size == 0) {
 595                                 CDEBUG(D_INFO, "OBD_MD_FLDIREA is set, "
 596                                        "but eadatasize 0\n");
 597                                 RETURN(-EPROTO);
 598                         }
 599
 600                         lmv = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
 601                                                            lmv_size);
 602                         if (lmv == NULL)
 603                                 GOTO(out, rc = -EPROTO);
 604
 605                         rc = md_unpackmd(md_exp, &md->lmv, lmv, lmv_size);
 606                         if (rc < 0)
 607                                 GOTO(out, rc);
 608
 609                         if (rc < (int)sizeof(*md->lmv)) {
 610                                 struct lmv_foreign_md *lfm = md->lfm;
 611
 612                                 /* short (< sizeof(struct lmv_stripe_md))
 613                                  * foreign LMV case
 614                                  */
 615                                 if (lfm->lfm_magic != LMV_MAGIC_FOREIGN) {
 616                                         CDEBUG(D_INFO,
 617                                                "lmv size too small: %d < %d\n",
 618                                                rc, (int)sizeof(*md->lmv));
 619                                         GOTO(out, rc = -EPROTO);
 620                                 }
 621                         }
 622                 }
 623
 624                 /* since 2.12.58 intent_getattr fetches default LMV */
 625                 if (md->body->mbo_valid & OBD_MD_DEFAULT_MEA) {
 626                         lmv_size = sizeof(struct lmv_user_md);
 627                         lmv = req_capsule_server_sized_get(pill,
 628                                                            &RMF_DEFAULT_MDT_MD,
 629                                                            lmv_size);
 630                         if (!lmv)
 631                                 GOTO(out, rc = -EPROTO);
 632
 633                         rc = md_unpackmd(md_exp, &md->default_lmv, lmv,
 634                                          lmv_size);
 635                         if (rc < 0)
 636                                 GOTO(out, rc);
 637
 638                         if (rc < (int)sizeof(*md->default_lmv)) {
 639                                 CDEBUG(D_INFO,
 640                                        "default lmv size too small: %d < %d\n",
 641                                         rc, (int)sizeof(*md->default_lmv));
 642                                 GOTO(out, rc = -EPROTO);
 643                         }
 644                 }
 645         }
 646         rc = 0;
 647
 648         if (md->body->mbo_valid & OBD_MD_FLACL) {
 649                 /* for ACL, it's possible that FLACL is set but aclsize is zero.
 650                  * only when aclsize != 0 there's an actual segment for ACL
 651                  * in reply buffer.
 652                  */
 653                 if (md->body->mbo_aclsize) {
 654                         rc = mdc_unpack_acl(req, md);
 655                         if (rc)
 656                                 GOTO(out, rc);
 657 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
 658                 } else {
 659                         md->posix_acl = NULL;
 660 #endif
 661                 }
 662         }
 663
 664         EXIT;
 665 out:
 666         if (rc) {
 667 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
 668                 posix_acl_release(md->posix_acl);
 669 #endif
 670         }
 671         return rc;
 672 }
 673
 674 int mdc_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
 675 {
 676         ENTRY;
 677         RETURN(0);
 678 }
 679
 680 void mdc_replay_open(struct ptlrpc_request *req)
 681 {
 682         struct md_open_data *mod = req->rq_cb_data;
 683         struct ptlrpc_request *close_req;
 684         struct obd_client_handle *och;
 685         struct lustre_handle old_open_handle = { };
 686         struct mdt_body *body;
 687         ENTRY;
 688
 689         if (mod == NULL) {
 690                 DEBUG_REQ(D_ERROR, req,
 691                           "cannot properly replay without open data");
 692                 EXIT;
 693                 return;
 694         }
 695
 696         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 697         LASSERT(body != NULL);
 698
 699         spin_lock(&req->rq_lock);
 700         och = mod->mod_och;
 701         if (och && och->och_open_handle.cookie)
 702                 req->rq_early_free_repbuf = 1;
 703         else
 704                 req->rq_early_free_repbuf = 0;
 705         spin_unlock(&req->rq_lock);
 706
 707         if (req->rq_early_free_repbuf) {
 708                 struct lustre_handle *file_open_handle;
 709
 710                 LASSERT(och->och_magic == OBD_CLIENT_HANDLE_MAGIC);
 711
 712                 file_open_handle = &och->och_open_handle;
 713                 CDEBUG(D_HA, "updating handle from %#llx to %#llx\n",
 714                        file_open_handle->cookie, body->mbo_open_handle.cookie);
 715                 old_open_handle = *file_open_handle;
 716                 *file_open_handle = body->mbo_open_handle;
 717         }
 718
 719         close_req = mod->mod_close_req;
 720         if (close_req) {
 721                 __u32 opc = lustre_msg_get_opc(close_req->rq_reqmsg);
 722                 struct mdt_ioepoch *epoch;
 723
 724                 LASSERT(opc == MDS_CLOSE);
 725                 epoch = req_capsule_client_get(&close_req->rq_pill,
 726                                                &RMF_MDT_EPOCH);
 727                 LASSERT(epoch);
 728
 729                 if (req->rq_early_free_repbuf)
 730                         LASSERT(old_open_handle.cookie ==
 731                                 epoch->mio_open_handle.cookie);
 732
 733                 DEBUG_REQ(D_HA, close_req, "updating close body with new fh");
 734                 epoch->mio_open_handle = body->mbo_open_handle;
 735         }
 736         EXIT;
 737 }
 738
 739 void mdc_commit_open(struct ptlrpc_request *req)
 740 {
 741         struct md_open_data *mod = req->rq_cb_data;
 742         if (mod == NULL)
 743                 return;
 744
 745         /**
 746          * No need to touch md_open_data::mod_och, it holds a reference on
 747          * \var mod and will zero references to each other, \var mod will be
 748          * freed after that when md_open_data::mod_och will put the reference.
 749          */
 750
 751         /**
 752          * Do not let open request to disappear as it still may be needed
 753          * for close rpc to happen (it may happen on evict only, otherwise
 754          * ptlrpc_request::rq_replay does not let mdc_commit_open() to be
 755          * called), just mark this rpc as committed to distinguish these 2
 756          * cases, see mdc_close() for details. The open request reference will
 757          * be put along with freeing \var mod.
 758          */
 759         ptlrpc_request_addref(req);
 760         spin_lock(&req->rq_lock);
 761         req->rq_committed = 1;
 762         spin_unlock(&req->rq_lock);
 763         req->rq_cb_data = NULL;
 764         obd_mod_put(mod);
 765 }
 766
 767 int mdc_set_open_replay_data(struct obd_export *exp,
 768                              struct obd_client_handle *och,
 769                              struct lookup_intent *it)
 770 {
 771         struct md_open_data     *mod;
 772         struct mdt_rec_create   *rec;
 773         struct mdt_body         *body;
 774         struct ptlrpc_request   *open_req = it->it_request;
 775         struct obd_import       *imp = open_req->rq_import;
 776         ENTRY;
 777
 778         if (!open_req->rq_replay)
 779                 RETURN(0);
 780
 781         rec = req_capsule_client_get(&open_req->rq_pill, &RMF_REC_REINT);
 782         body = req_capsule_server_get(&open_req->rq_pill, &RMF_MDT_BODY);
 783         LASSERT(rec != NULL);
 784         /* Incoming message in my byte order (it's been swabbed). */
 785         /* Outgoing messages always in my byte order. */
 786         LASSERT(body != NULL);
 787
 788         /* Only if the import is replayable, we set replay_open data */
 789         if (och && imp->imp_replayable) {
 790                 mod = obd_mod_alloc();
 791                 if (mod == NULL) {
 792                         DEBUG_REQ(D_ERROR, open_req,
 793                                   "cannot allocate md_open_data");
 794                         RETURN(0);
 795                 }
 796
 797                 /**
 798                  * Take a reference on \var mod, to be freed on mdc_close().
 799                  * It protects \var mod from being freed on eviction (commit
 800                  * callback is called despite rq_replay flag).
 801                  * Another reference for \var och.
 802                  */
 803                 obd_mod_get(mod);
 804                 obd_mod_get(mod);
 805
 806                 spin_lock(&open_req->rq_lock);
 807                 och->och_mod = mod;
 808                 mod->mod_och = och;
 809                 mod->mod_is_create = it_disposition(it, DISP_OPEN_CREATE) ||
 810                                      it_disposition(it, DISP_OPEN_STRIPE);
 811                 mod->mod_open_req = open_req;
 812                 open_req->rq_cb_data = mod;
 813                 open_req->rq_commit_cb = mdc_commit_open;
 814                 open_req->rq_early_free_repbuf = 1;
 815                 spin_unlock(&open_req->rq_lock);
 816         }
 817
 818         rec->cr_fid2 = body->mbo_fid1;
 819         rec->cr_open_handle_old = body->mbo_open_handle;
 820         open_req->rq_replay_cb = mdc_replay_open;
 821         if (!fid_is_sane(&body->mbo_fid1)) {
 822                 DEBUG_REQ(D_ERROR, open_req,
 823                           "saving replay request with insane FID " DFID,
 824                           PFID(&body->mbo_fid1));
 825                 LBUG();
 826         }
 827
 828         DEBUG_REQ(D_RPCTRACE, open_req, "Set up open replay data");
 829         RETURN(0);
 830 }
 831
 832 static void mdc_free_open(struct md_open_data *mod)
 833 {
 834         int committed = 0;
 835
 836         if (mod->mod_is_create == 0 &&
 837             imp_connect_disp_stripe(mod->mod_open_req->rq_import))
 838                 committed = 1;
 839
 840         /**
 841          * No reason to asssert here if the open request has
 842          * rq_replay == 1. It means that mdc_close failed, and
 843          * close request wasn`t sent. It is not fatal to client.
 844          * The worst thing is eviction if the client gets open lock
 845          **/
 846
 847         DEBUG_REQ(D_RPCTRACE, mod->mod_open_req,
 848                   "free open request, rq_replay=%d",
 849                   mod->mod_open_req->rq_replay);
 850
 851         ptlrpc_request_committed(mod->mod_open_req, committed);
 852         if (mod->mod_close_req)
 853                 ptlrpc_request_committed(mod->mod_close_req, committed);
 854 }
 855
 856 int mdc_clear_open_replay_data(struct obd_export *exp,
 857                                struct obd_client_handle *och)
 858 {
 859         struct md_open_data *mod = och->och_mod;
 860         ENTRY;
 861
 862         /**
 863          * It is possible to not have \var mod in a case of eviction between
 864          * lookup and ll_file_open().
 865          **/
 866         if (mod == NULL)
 867                 RETURN(0);
 868
 869         LASSERT(mod != LP_POISON);
 870         LASSERT(mod->mod_open_req != NULL);
 871
 872         spin_lock(&mod->mod_open_req->rq_lock);
 873         if (mod->mod_och)
 874                 mod->mod_och->och_open_handle.cookie = 0;
 875         mod->mod_open_req->rq_early_free_repbuf = 0;
 876         spin_unlock(&mod->mod_open_req->rq_lock);
 877         mdc_free_open(mod);
 878
 879         mod->mod_och = NULL;
 880         och->och_mod = NULL;
 881         obd_mod_put(mod);
 882
 883         RETURN(0);
 884 }
 885
 886 static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
 887                      struct md_open_data *mod, struct ptlrpc_request **request)
 888 {
 889         struct obd_device     *obd = class_exp2obd(exp);
 890         struct ptlrpc_request *req;
 891         struct req_format     *req_fmt;
 892         size_t                 u32_count = 0;
 893         int                    rc;
 894         int                    saved_rc = 0;
 895         ENTRY;
 896
 897         CDEBUG(D_INODE, "%s: "DFID" file closed with intent: %x\n",
 898                exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
 899                op_data->op_bias);
 900
 901         if (op_data->op_bias & MDS_CLOSE_INTENT) {
 902                 req_fmt = &RQF_MDS_CLOSE_INTENT;
 903                 if (op_data->op_bias & MDS_HSM_RELEASE) {
 904                         /* allocate a FID for volatile file */
 905                         rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2,
 906                                            op_data);
 907                         if (rc < 0) {
 908                                 CERROR("%s: "DFID" allocating FID: rc = %d\n",
 909                                        obd->obd_name, PFID(&op_data->op_fid1),
 910                                        rc);
 911                                 /* save the errcode and proceed to close */
 912                                 saved_rc = rc;
 913                         }
 914                 }
 915                 if (op_data->op_bias & MDS_CLOSE_RESYNC_DONE) {
 916                         size_t count = op_data->op_data_size / sizeof(__u32);
 917
 918                         if (count > INLINE_RESYNC_ARRAY_SIZE)
 919                                 u32_count = count;
 920                 }
 921         } else {
 922                 req_fmt = &RQF_MDS_CLOSE;
 923         }
 924
 925         *request = NULL;
 926         if (OBD_FAIL_CHECK(OBD_FAIL_MDC_CLOSE))
 927                 req = NULL;
 928         else
 929                 req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);
 930
 931         /* Ensure that this close's handle is fixed up during replay. */
 932         if (likely(mod != NULL)) {
 933                 LASSERTF(mod->mod_open_req != NULL &&
 934                          mod->mod_open_req->rq_type != LI_POISON,
 935                          "POISONED open %p!\n", mod->mod_open_req);
 936
 937                 mod->mod_close_req = req;
 938
 939                 DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "matched open");
 940                 /* We no longer want to preserve this open for replay even
 941                  * though the open was committed. b=3632, b=3633 */
 942                 spin_lock(&mod->mod_open_req->rq_lock);
 943                 mod->mod_open_req->rq_replay = 0;
 944                 spin_unlock(&mod->mod_open_req->rq_lock);
 945         } else {
 946                 CDEBUG(D_HA, "couldn't find open req; expecting close error\n");
 947         }
 948         if (req == NULL) {
 949                 /**
 950                  * TODO: repeat close after errors
 951                  */
 952                 CWARN("%s: close of FID "DFID" failed, file reference will be "
 953                       "dropped when this client unmounts or is evicted\n",
 954                       obd->obd_name, PFID(&op_data->op_fid1));
 955                 GOTO(out, rc = -ENOMEM);
 956         }
 957
 958         if (u32_count > 0)
 959                 req_capsule_set_size(&req->rq_pill, &RMF_U32, RCL_CLIENT,
 960                                      u32_count * sizeof(__u32));
 961
 962         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
 963         if (rc) {
 964                 ptlrpc_request_free(req);
 965                 req = NULL;
 966                 GOTO(out, rc);
 967         }
 968
 969         /* To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
 970          * portal whose threads are not taking any DLM locks and are therefore
 971          * always progressing */
 972         req->rq_request_portal = MDS_READPAGE_PORTAL;
 973         ptlrpc_at_set_req_timeout(req);
 974
 975         if (!(exp_connect_flags2(exp) & OBD_CONNECT2_LSOM))
 976                 op_data->op_xvalid &= ~(OP_XVALID_LAZYSIZE |
 977                                         OP_XVALID_LAZYBLOCKS);
 978
 979         mdc_close_pack(req, op_data);
 980
 981         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 982                              obd->u.cli.cl_default_mds_easize);
 983
 984         ptlrpc_request_set_replen(req);
 985
 986         ptlrpc_get_mod_rpc_slot(req);
 987         rc = ptlrpc_queue_wait(req);
 988         ptlrpc_put_mod_rpc_slot(req);
 989
 990         if (req->rq_repmsg == NULL) {
 991                 CDEBUG(D_RPCTRACE, "request %p failed to send: rc = %d\n", req,
 992                        req->rq_status);
 993                 if (rc == 0)
 994                         rc = req->rq_status ?: -EIO;
 995         } else if (rc == 0 || rc == -EAGAIN) {
 996                 struct mdt_body *body;
 997
 998                 rc = lustre_msg_get_status(req->rq_repmsg);
 999                 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1000                         DEBUG_REQ(D_ERROR, req,
1001                                   "type = PTL_RPC_MSG_ERR: rc = %d", rc);
1002                         if (rc > 0)
1003                                 rc = -rc;
1004                 }
1005                 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1006                 if (body == NULL)
1007                         rc = -EPROTO;
1008         } else if (rc == -ESTALE) {
1009                 /**
1010                  * it can be allowed error after 3633 if open was committed and
1011                  * server failed before close was sent. Let's check if mod
1012                  * exists and return no error in that case
1013                  */
1014                 if (mod) {
1015                         DEBUG_REQ(D_HA, req, "Reset ESTALE = %d", rc);
1016                         LASSERT(mod->mod_open_req != NULL);
1017                         if (mod->mod_open_req->rq_committed)
1018                                 rc = 0;
1019                 }
1020         }
1021
1022 out:
1023         if (mod) {
1024                 if (rc != 0)
1025                         mod->mod_close_req = NULL;
1026                 /* Since now, mod is accessed through open_req only,
1027                  * thus close req does not keep a reference on mod anymore. */
1028                 obd_mod_put(mod);
1029         }
1030         *request = req;
1031
1032         RETURN(rc < 0 ? rc : saved_rc);
1033 }
1034
1035 static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid,
1036                        u64 offset, struct page **pages, int npages,
1037                        struct ptlrpc_request **request)
1038 {
1039         struct ptlrpc_request   *req;
1040         struct ptlrpc_bulk_desc *desc;
1041         int                      i;
1042         int                      resends = 0;
1043         int                      rc;
1044         ENTRY;
1045
1046         *request = NULL;
1047
1048 restart_bulk:
1049         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE);
1050         if (req == NULL)
1051                 RETURN(-ENOMEM);
1052
1053         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE);
1054         if (rc) {
1055                 ptlrpc_request_free(req);
1056                 RETURN(rc);
1057         }
1058
1059         req->rq_request_portal = MDS_READPAGE_PORTAL;
1060         ptlrpc_at_set_req_timeout(req);
1061
1062         desc = ptlrpc_prep_bulk_imp(req, npages, 1,
1063                                     PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
1064                                     MDS_BULK_PORTAL,
1065                                     &ptlrpc_bulk_kiov_pin_ops);
1066         if (desc == NULL) {
1067                 ptlrpc_req_finished(req);
1068                 RETURN(-ENOMEM);
1069         }
1070
1071         /* NB req now owns desc and will free it when it gets freed */
1072         for (i = 0; i < npages; i++)
1073                 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1074                                                  PAGE_SIZE);
1075
1076         mdc_readdir_pack(req, offset, PAGE_SIZE * npages, fid);
1077
1078         ptlrpc_request_set_replen(req);
1079         rc = ptlrpc_queue_wait(req);
1080         if (rc) {
1081                 ptlrpc_req_finished(req);
1082                 if (rc != -ETIMEDOUT)
1083                         RETURN(rc);
1084
1085                 resends++;
1086                 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
1087                         CERROR("%s: too many resend retries: rc = %d\n",
1088                                exp->exp_obd->obd_name, -EIO);
1089                         RETURN(-EIO);
1090                 }
1091                 ssleep(resends);
1092
1093                 goto restart_bulk;
1094         }
1095
1096         rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1097                                           req->rq_bulk->bd_nob_transferred);
1098         if (rc < 0) {
1099                 ptlrpc_req_finished(req);
1100                 RETURN(rc);
1101         }
1102
1103         if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) {
1104                 CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n",
1105                        exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred,
1106                        PAGE_SIZE * npages);
1107                 ptlrpc_req_finished(req);
1108                 RETURN(-EPROTO);
1109         }
1110
1111         *request = req;
1112         RETURN(0);
1113 }
1114
1115 static void mdc_release_page(struct page *page, int remove)
1116 {
1117         if (remove) {
1118                 lock_page(page);
1119                 if (likely(page->mapping != NULL))
1120                         delete_from_page_cache(page);
1121                 unlock_page(page);
1122         }
1123         put_page(page);
1124 }
1125
1126 static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash,
1127                                     __u64 *start, __u64 *end, int hash64)
1128 {
1129         /*
1130          * Complement of hash is used as an index so that
1131          * radix_tree_gang_lookup() can be used to find a page with starting
1132          * hash _smaller_ than one we are looking for.
1133          */
1134         unsigned long offset = hash_x_index(*hash, hash64);
1135         struct page *page;
1136         unsigned long flags;
1137         int found;
1138
1139         xa_lock_irqsave(&mapping->i_pages, flags);
1140         found = radix_tree_gang_lookup(&mapping->page_tree,
1141                                        (void **)&page, offset, 1);
1142         if (found > 0 && !xa_is_value(page)) {
1143                 struct lu_dirpage *dp;
1144
1145                 get_page(page);
1146                 xa_unlock_irqrestore(&mapping->i_pages, flags);
1147                 /*
1148                  * In contrast to find_lock_page() we are sure that directory
1149                  * page cannot be truncated (while DLM lock is held) and,
1150                  * hence, can avoid restart.
1151                  *
1152                  * In fact, page cannot be locked here at all, because
1153                  * mdc_read_page_remote does synchronous io.
1154                  */
1155                 wait_on_page_locked(page);
1156                 if (PageUptodate(page)) {
1157                         dp = kmap(page);
1158                         if (BITS_PER_LONG == 32 && hash64) {
1159                                 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1160                                 *end   = le64_to_cpu(dp->ldp_hash_end) >> 32;
1161                                 *hash  = *hash >> 32;
1162                         } else {
1163                                 *start = le64_to_cpu(dp->ldp_hash_start);
1164                                 *end   = le64_to_cpu(dp->ldp_hash_end);
1165                         }
1166                         if (unlikely(*start == 1 && *hash == 0))
1167                                 *hash = *start;
1168                         else
1169                                 LASSERTF(*start <= *hash, "start = %#llx"
1170                                          ",end = %#llx,hash = %#llx\n",
1171                                          *start, *end, *hash);
1172                         CDEBUG(D_VFSTRACE, "offset %lx [%#llx %#llx],"
1173                               " hash %#llx\n", offset, *start, *end, *hash);
1174                         if (*hash > *end) {
1175                                 kunmap(page);
1176                                 mdc_release_page(page, 0);
1177                                 page = NULL;
1178                         } else if (*end != *start && *hash == *end) {
1179                                 /*
1180                                  * upon hash collision, remove this page,
1181                                  * otherwise put page reference, and
1182                                  * mdc_read_page_remote() will issue RPC to
1183                                  * fetch the page we want.
1184                                  */
1185                                 kunmap(page);
1186                                 mdc_release_page(page,
1187                                     le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1188                                 page = NULL;
1189                         }
1190                 } else {
1191                         put_page(page);
1192                         page = ERR_PTR(-EIO);
1193                 }
1194         } else {
1195                 xa_unlock_irqrestore(&mapping->i_pages, flags);
1196                 page = NULL;
1197         }
1198         return page;
1199 }
1200
1201 /*
1202  * Adjust a set of pages, each page containing an array of lu_dirpages,
1203  * so that each page can be used as a single logical lu_dirpage.
1204  *
1205  * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
1206  * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
1207  * struct lu_dirent.  It has size up to LU_PAGE_SIZE. The ldp_hash_end
1208  * value is used as a cookie to request the next lu_dirpage in a
1209  * directory listing that spans multiple pages (two in this example):
1210  *   ________
1211  *  |        |
1212  * .|--------v-------   -----.
1213  * |s|e|f|p|ent|ent| ... |ent|
1214  * '--|--------------   -----'   Each PAGE contains a single
1215  *    '------.                   lu_dirpage.
1216  * .---------v-------   -----.
1217  * |s|e|f|p|ent| 0 | ... | 0 |
1218  * '-----------------   -----'
1219  *
1220  * However, on hosts where the native VM page size (PAGE_SIZE) is
1221  * larger than LU_PAGE_SIZE, a single host page may contain multiple
1222  * lu_dirpages. After reading the lu_dirpages from the MDS, the
1223  * ldp_hash_end of the first lu_dirpage refers to the one immediately
1224  * after it in the same PAGE (arrows simplified for brevity, but
1225  * in general e0==s1, e1==s2, etc.):
1226  *
1227  * .--------------------   -----.
1228  * |s0|e0|f0|p|ent|ent| ... |ent|
1229  * |---v----------------   -----|
1230  * |s1|e1|f1|p|ent|ent| ... |ent|
1231  * |---v----------------   -----|  Here, each PAGE contains
1232  *             ...                 multiple lu_dirpages.
1233  * |---v----------------   -----|
1234  * |s'|e'|f'|p|ent|ent| ... |ent|
1235  * '---|----------------   -----'
1236  *     v
1237  * .----------------------------.
1238  * |        next PAGE           |
1239  *
1240  * This structure is transformed into a single logical lu_dirpage as follows:
1241  *
1242  * - Replace e0 with e' so the request for the next lu_dirpage gets the page
1243  *   labeled 'next PAGE'.
1244  *
1245  * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
1246  *   a hash collision with the next page exists.
1247  *
1248  * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
1249  *   to the first entry of the next lu_dirpage.
1250  */
1251 #if PAGE_SIZE > LU_PAGE_SIZE
1252 static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs)
1253 {
1254         int i;
1255
1256         for (i = 0; i < cfs_pgs; i++) {
1257                 struct lu_dirpage *dp = kmap(pages[i]);
1258                 struct lu_dirpage *first = dp;
1259                 struct lu_dirent *end_dirent = NULL;
1260                 struct lu_dirent *ent;
1261                 __u64 hash_end = dp->ldp_hash_end;
1262                 __u32 flags = dp->ldp_flags;
1263
1264                 while (--lu_pgs > 0) {
1265                         ent = lu_dirent_start(dp);
1266                         for (end_dirent = ent; ent != NULL;
1267                              end_dirent = ent, ent = lu_dirent_next(ent));
1268
1269                         /* Advance dp to next lu_dirpage. */
1270                         dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
1271
1272                         /* Check if we've reached the end of the PAGE. */
1273                         if (!((unsigned long)dp & ~PAGE_MASK))
1274                                 break;
1275
1276                         /* Save the hash and flags of this lu_dirpage. */
1277                         hash_end = dp->ldp_hash_end;
1278                         flags = dp->ldp_flags;
1279
1280                         /* Check if lu_dirpage contains no entries. */
1281                         if (end_dirent == NULL)
1282                                 break;
1283
1284                         /* Enlarge the end entry lde_reclen from 0 to
1285                          * first entry of next lu_dirpage. */
1286                         LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
1287                         end_dirent->lde_reclen =
1288                                 cpu_to_le16((char *)(dp->ldp_entries) -
1289                                             (char *)end_dirent);
1290                 }
1291
1292                 first->ldp_hash_end = hash_end;
1293                 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
1294                 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
1295
1296                 kunmap(pages[i]);
1297         }
1298         LASSERTF(lu_pgs == 0, "left = %d\n", lu_pgs);
1299 }
1300 #else
1301 #define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0)
1302 #endif  /* PAGE_SIZE > LU_PAGE_SIZE */
1303
1304 /* parameters for readdir page */
1305 struct readpage_param {
1306         struct md_op_data       *rp_mod;
1307         __u64                   rp_off;
1308         int                     rp_hash64;
1309         struct obd_export       *rp_exp;
1310         struct md_callback      *rp_cb;
1311 };
1312
1313 /**
1314  * Read pages from server.
1315  *
1316  * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
1317  * a header lu_dirpage which describes the start/end hash, and whether this
1318  * page is empty (contains no dir entry) or hash collide with next page.
1319  * After client receives reply, several pages will be integrated into dir page
1320  * in PAGE_SIZE (if PAGE_SIZE greater than LU_PAGE_SIZE), and the
1321  * lu_dirpage for this integrated page will be adjusted.
1322  **/
1323 static int mdc_read_page_remote(void *data, struct page *page0)
1324 {
1325         struct readpage_param *rp = data;
1326         struct page **page_pool;
1327         struct page *page;
1328         struct lu_dirpage *dp;
1329         struct md_op_data *op_data = rp->rp_mod;
1330         struct ptlrpc_request *req;
1331         int max_pages;
1332         struct inode *inode;
1333         struct lu_fid *fid;
1334         int rd_pgs = 0; /* number of pages actually read */
1335         int npages;
1336         int i;
1337         int rc;
1338         ENTRY;
1339
1340         max_pages = rp->rp_exp->exp_obd->u.cli.cl_max_pages_per_rpc;
1341         inode = op_data->op_data;
1342         fid = &op_data->op_fid1;
1343         LASSERT(inode != NULL);
1344
1345         OBD_ALLOC(page_pool, sizeof(page_pool[0]) * max_pages);
1346         if (page_pool != NULL) {
1347                 page_pool[0] = page0;
1348         } else {
1349                 page_pool = &page0;
1350                 max_pages = 1;
1351         }
1352
1353         for (npages = 1; npages < max_pages; npages++) {
1354                 page = page_cache_alloc(inode->i_mapping);
1355                 if (page == NULL)
1356                         break;
1357                 page_pool[npages] = page;
1358         }
1359
1360         rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, page_pool, npages, &req);
1361         if (rc < 0) {
1362                 /* page0 is special, which was added into page cache early */
1363                 delete_from_page_cache(page0);
1364         } else {
1365                 int lu_pgs;
1366
1367                 rd_pgs = (req->rq_bulk->bd_nob_transferred + PAGE_SIZE - 1) >>
1368                         PAGE_SHIFT;
1369                 lu_pgs = req->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
1370                 LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
1371
1372                 CDEBUG(D_INODE, "read %d(%d) pages\n", rd_pgs, lu_pgs);
1373
1374                 mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs);
1375
1376                 SetPageUptodate(page0);
1377         }
1378         unlock_page(page0);
1379
1380         ptlrpc_req_finished(req);
1381         CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages);
1382         for (i = 1; i < npages; i++) {
1383                 unsigned long   offset;
1384                 __u64           hash;
1385                 int ret;
1386
1387                 page = page_pool[i];
1388
1389                 if (rc < 0 || i >= rd_pgs) {
1390                         put_page(page);
1391                         continue;
1392                 }
1393
1394                 SetPageUptodate(page);
1395
1396                 dp = kmap(page);
1397                 hash = le64_to_cpu(dp->ldp_hash_start);
1398                 kunmap(page);
1399
1400                 offset = hash_x_index(hash, rp->rp_hash64);
1401
1402                 prefetchw(&page->flags);
1403                 ret = add_to_page_cache_lru(page, inode->i_mapping, offset,
1404                                             GFP_KERNEL);
1405                 if (ret == 0)
1406                         unlock_page(page);
1407                 else
1408                         CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
1409                                " rc = %d\n", offset, ret);
1410                 put_page(page);
1411         }
1412
1413         if (page_pool != &page0)
1414                 OBD_FREE(page_pool, sizeof(page_pool[0]) * max_pages);
1415
1416         RETURN(rc);
1417 }
1418
1419 /**
1420  * Read dir page from cache first, if it can not find it, read it from
1421  * server and add into the cache.
1422  *
1423  * \param[in] exp       MDC export
1424  * \param[in] op_data   client MD stack parameters, transfering parameters
1425  *                      between different layers on client MD stack.
1426  * \param[in] cb_op     callback required for ldlm lock enqueue during
1427  *                      read page
1428  * \param[in] hash_offset the hash offset of the page to be read
1429  * \param[in] ppage     the page to be read
1430  *
1431  * retval               = 0 get the page successfully
1432  *                      errno(<0) get the page failed
1433  */
1434 static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data,
1435                          struct md_callback *cb_op, __u64 hash_offset,
1436                          struct page **ppage)
1437 {
1438         struct lookup_intent    it = { .it_op = IT_READDIR };
1439         struct page             *page;
1440         struct inode            *dir = op_data->op_data;
1441         struct address_space    *mapping;
1442         struct lu_dirpage       *dp;
1443         __u64                   start = 0;
1444         __u64                   end = 0;
1445         struct lustre_handle    lockh;
1446         struct ptlrpc_request   *enq_req = NULL;
1447         struct readpage_param   rp_param;
1448         int rc;
1449
1450         ENTRY;
1451
1452         *ppage = NULL;
1453
1454         LASSERT(dir != NULL);
1455         mapping = dir->i_mapping;
1456
1457         rc = mdc_intent_lock(exp, op_data, &it, &enq_req,
1458                              cb_op->md_blocking_ast, 0);
1459         if (enq_req != NULL)
1460                 ptlrpc_req_finished(enq_req);
1461
1462         if (rc < 0) {
1463                 CERROR("%s: "DFID" lock enqueue fails: rc = %d\n",
1464                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc);
1465                 RETURN(rc);
1466         }
1467
1468         rc = 0;
1469         lockh.cookie = it.it_lock_handle;
1470         mdc_set_lock_data(exp, &lockh, dir, NULL);
1471
1472         rp_param.rp_off = hash_offset;
1473         rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64;
1474         page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end,
1475                                rp_param.rp_hash64);
1476         if (IS_ERR(page)) {
1477                 CERROR("%s: dir page locate: "DFID" at %llu: rc %ld\n",
1478                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1479                        rp_param.rp_off, PTR_ERR(page));
1480                 GOTO(out_unlock, rc = PTR_ERR(page));
1481         } else if (page != NULL) {
1482                 /*
1483                  * XXX nikita: not entirely correct handling of a corner case:
1484                  * suppose hash chain of entries with hash value HASH crosses
1485                  * border between pages P0 and P1. First both P0 and P1 are
1486                  * cached, seekdir() is called for some entry from the P0 part
1487                  * of the chain. Later P0 goes out of cache. telldir(HASH)
1488                  * happens and finds P1, as it starts with matching hash
1489                  * value. Remaining entries from P0 part of the chain are
1490                  * skipped. (Is that really a bug?)
1491                  *
1492                  * Possible solutions: 0. don't cache P1 is such case, handle
1493                  * it as an "overflow" page. 1. invalidate all pages at
1494                  * once. 2. use HASH|1 as an index for P1.
1495                  */
1496                 GOTO(hash_collision, page);
1497         }
1498
1499         rp_param.rp_exp = exp;
1500         rp_param.rp_mod = op_data;
1501         page = read_cache_page(mapping,
1502                                hash_x_index(rp_param.rp_off,
1503                                             rp_param.rp_hash64),
1504                                mdc_read_page_remote, &rp_param);
1505         if (IS_ERR(page)) {
1506                 CDEBUG(D_INFO, "%s: read cache page: "DFID" at %llu: %ld\n",
1507                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1508                        rp_param.rp_off, PTR_ERR(page));
1509                 GOTO(out_unlock, rc = PTR_ERR(page));
1510         }
1511
1512         wait_on_page_locked(page);
1513         (void)kmap(page);
1514         if (!PageUptodate(page)) {
1515                 CERROR("%s: page not updated: "DFID" at %llu: rc %d\n",
1516                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1517                        rp_param.rp_off, -5);
1518                 goto fail;
1519         }
1520         if (!PageChecked(page))
1521                 SetPageChecked(page);
1522         if (PageError(page)) {
1523                 CERROR("%s: page error: "DFID" at %llu: rc %d\n",
1524                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1525                        rp_param.rp_off, -5);
1526                 goto fail;
1527         }
1528
1529 hash_collision:
1530         dp = page_address(page);
1531         if (BITS_PER_LONG == 32 && rp_param.rp_hash64) {
1532                 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1533                 end   = le64_to_cpu(dp->ldp_hash_end) >> 32;
1534                 rp_param.rp_off = hash_offset >> 32;
1535         } else {
1536                 start = le64_to_cpu(dp->ldp_hash_start);
1537                 end   = le64_to_cpu(dp->ldp_hash_end);
1538                 rp_param.rp_off = hash_offset;
1539         }
1540         if (end == start) {
1541                 LASSERT(start == rp_param.rp_off);
1542                 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
1543 #if BITS_PER_LONG == 32
1544                 CWARN("Real page-wide hash collision at [%llu %llu] with "
1545                       "hash %llu\n", le64_to_cpu(dp->ldp_hash_start),
1546                       le64_to_cpu(dp->ldp_hash_end), hash_offset);
1547 #endif
1548
1549                 /*
1550                  * Fetch whole overflow chain...
1551                  *
1552                  * XXX not yet.
1553                  */
1554                 goto fail;
1555         }
1556         *ppage = page;
1557 out_unlock:
1558         ldlm_lock_decref(&lockh, it.it_lock_mode);
1559         return rc;
1560 fail:
1561         kunmap(page);
1562         mdc_release_page(page, 1);
1563         rc = -EIO;
1564         goto out_unlock;
1565 }
1566
1567 static int mdc_statfs_interpret(const struct lu_env *env,
1568                                 struct ptlrpc_request *req, void *args, int rc)
1569 {
1570         struct obd_info *oinfo = args;
1571         struct obd_statfs *osfs;
1572
1573         if (!rc) {
1574                 osfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1575                 if (!osfs)
1576                         return -EPROTO;
1577
1578                 oinfo->oi_osfs = osfs;
1579
1580                 CDEBUG(D_CACHE, "blocks=%llu free=%llu avail=%llu "
1581                        "objects=%llu free=%llu state=%x\n",
1582                         osfs->os_blocks, osfs->os_bfree, osfs->os_bavail,
1583                         osfs->os_files, osfs->os_ffree, osfs->os_state);
1584         }
1585
1586         oinfo->oi_cb_up(oinfo, rc);
1587
1588         return rc;
1589 }
1590
1591 static int mdc_statfs_async(struct obd_export *exp,
1592                             struct obd_info *oinfo, time64_t max_age,
1593                             struct ptlrpc_request_set *unused)
1594 {
1595         struct ptlrpc_request *req;
1596         struct obd_info *aa;
1597
1598         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp), &RQF_MDS_STATFS,
1599                                         LUSTRE_MDS_VERSION, MDS_STATFS);
1600         if (req == NULL)
1601                 return -ENOMEM;
1602
1603         ptlrpc_request_set_replen(req);
1604         req->rq_interpret_reply = mdc_statfs_interpret;
1605
1606         aa = ptlrpc_req_async_args(aa, req);
1607         *aa = *oinfo;
1608
1609         ptlrpcd_add_req(req);
1610
1611         return 0;
1612 }
1613
1614 static int mdc_statfs(const struct lu_env *env,
1615                       struct obd_export *exp, struct obd_statfs *osfs,
1616                       time64_t max_age, __u32 flags)
1617 {
1618         struct obd_device *obd = class_exp2obd(exp);
1619         struct req_format *fmt;
1620         struct ptlrpc_request *req;
1621         struct obd_statfs *msfs;
1622         struct obd_import *imp = NULL;
1623         int rc;
1624         ENTRY;
1625
1626         /*
1627          * Since the request might also come from lprocfs, so we need
1628          * sync this with client_disconnect_export Bug15684
1629          */
1630         down_read(&obd->u.cli.cl_sem);
1631         if (obd->u.cli.cl_import)
1632                 imp = class_import_get(obd->u.cli.cl_import);
1633         up_read(&obd->u.cli.cl_sem);
1634         if (!imp)
1635                 RETURN(-ENODEV);
1636
1637         fmt = &RQF_MDS_STATFS;
1638         if ((exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS) &&
1639             (flags & OBD_STATFS_SUM))
1640                 fmt = &RQF_MDS_STATFS_NEW;
1641         req = ptlrpc_request_alloc_pack(imp, fmt, LUSTRE_MDS_VERSION,
1642                                         MDS_STATFS);
1643         if (req == NULL)
1644                 GOTO(output, rc = -ENOMEM);
1645
1646         if ((flags & OBD_STATFS_SUM) &&
1647             (exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS)) {
1648                 /* request aggregated states */
1649                 struct mdt_body *body;
1650
1651                 body = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
1652                 if (body == NULL)
1653                         GOTO(out, rc = -EPROTO);
1654                 body->mbo_valid = OBD_MD_FLAGSTATFS;
1655         }
1656
1657         ptlrpc_request_set_replen(req);
1658
1659         if (flags & OBD_STATFS_NODELAY) {
1660                 /* procfs requests not want stay in wait for avoid deadlock */
1661                 req->rq_no_resend = 1;
1662                 req->rq_no_delay = 1;
1663         }
1664
1665         rc = ptlrpc_queue_wait(req);
1666         if (rc) {
1667                 /* check connection error first */
1668                 if (imp->imp_connect_error)
1669                         rc = imp->imp_connect_error;
1670                 GOTO(out, rc);
1671         }
1672
1673         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1674         if (msfs == NULL)
1675                 GOTO(out, rc = -EPROTO);
1676
1677         *osfs = *msfs;
1678         EXIT;
1679 out:
1680         ptlrpc_req_finished(req);
1681 output:
1682         class_import_put(imp);
1683         return rc;
1684 }
1685
1686 static int mdc_ioc_fid2path(struct obd_export *exp, struct getinfo_fid2path *gf)
1687 {
1688         __u32 keylen, vallen;
1689         void *key;
1690         int rc;
1691
1692         if (gf->gf_pathlen > PATH_MAX)
1693                 RETURN(-ENAMETOOLONG);
1694         if (gf->gf_pathlen < 2)
1695                 RETURN(-EOVERFLOW);
1696
1697         /* Key is KEY_FID2PATH + getinfo_fid2path description */
1698         keylen = cfs_size_round(sizeof(KEY_FID2PATH) + sizeof(*gf) +
1699                                 sizeof(struct lu_fid));
1700         OBD_ALLOC(key, keylen);
1701         if (key == NULL)
1702                 RETURN(-ENOMEM);
1703         memcpy(key, KEY_FID2PATH, sizeof(KEY_FID2PATH));
1704         memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)), gf, sizeof(*gf));
1705         memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*gf),
1706                gf->gf_u.gf_root_fid, sizeof(struct lu_fid));
1707         CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
1708                PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno);
1709
1710         if (!fid_is_sane(&gf->gf_fid))
1711                 GOTO(out, rc = -EINVAL);
1712
1713         /* Val is struct getinfo_fid2path result plus path */
1714         vallen = sizeof(*gf) + gf->gf_pathlen;
1715
1716         rc = obd_get_info(NULL, exp, keylen, key, &vallen, gf);
1717         if (rc != 0 && rc != -EREMOTE)
1718                 GOTO(out, rc);
1719
1720         if (vallen <= sizeof(*gf))
1721                 GOTO(out, rc = -EPROTO);
1722         if (vallen > sizeof(*gf) + gf->gf_pathlen)
1723                 GOTO(out, rc = -EOVERFLOW);
1724
1725         CDEBUG(D_IOCTL, "path got "DFID" from %llu #%d: %s\n",
1726                PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno,
1727                gf->gf_pathlen < 512 ? gf->gf_u.gf_path :
1728                /* only log the last 512 characters of the path */
1729                gf->gf_u.gf_path + gf->gf_pathlen - 512);
1730
1731 out:
1732         OBD_FREE(key, keylen);
1733         return rc;
1734 }
1735
1736 static int mdc_ioc_hsm_progress(struct obd_export *exp,
1737                                 struct hsm_progress_kernel *hpk)
1738 {
1739         struct obd_import               *imp = class_exp2cliimp(exp);
1740         struct hsm_progress_kernel      *req_hpk;
1741         struct ptlrpc_request           *req;
1742         int                              rc;
1743         ENTRY;
1744
1745         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_PROGRESS,
1746                                         LUSTRE_MDS_VERSION, MDS_HSM_PROGRESS);
1747         if (req == NULL)
1748                 GOTO(out, rc = -ENOMEM);
1749
1750         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1751
1752         /* Copy hsm_progress struct */
1753         req_hpk = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_PROGRESS);
1754         if (req_hpk == NULL)
1755                 GOTO(out, rc = -EPROTO);
1756
1757         *req_hpk = *hpk;
1758         req_hpk->hpk_errval = lustre_errno_hton(hpk->hpk_errval);
1759
1760         ptlrpc_request_set_replen(req);
1761
1762         ptlrpc_get_mod_rpc_slot(req);
1763         rc = ptlrpc_queue_wait(req);
1764         ptlrpc_put_mod_rpc_slot(req);
1765
1766         GOTO(out, rc);
1767 out:
1768         ptlrpc_req_finished(req);
1769         return rc;
1770 }
1771 /**
1772  * Send hsm_ct_register to MDS
1773  *
1774  * \param[in]   imp             import
1775  * \param[in]   archive_count   if in bitmap format, it is the bitmap,
1776  *                              else it is the count of archive_ids
1777  * \param[in]   archives        if in bitmap format, it is NULL,
1778  *                              else it is archive_id lists
1779  */
1780 static int mdc_ioc_hsm_ct_register(struct obd_import *imp, __u32 archive_count,
1781                                    __u32 *archives)
1782 {
1783         struct ptlrpc_request *req;
1784         __u32 *archive_array;
1785         size_t archives_size;
1786         int rc;
1787         ENTRY;
1788
1789         req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_CT_REGISTER);
1790         if (req == NULL)
1791                 RETURN(-ENOMEM);
1792
1793         if (archives != NULL)
1794                 archives_size = sizeof(*archive_array) * archive_count;
1795         else
1796                 archives_size = sizeof(archive_count);
1797
1798         req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_ARCHIVE,
1799                              RCL_CLIENT, archives_size);
1800
1801         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_CT_REGISTER);
1802         if (rc) {
1803                 ptlrpc_request_free(req);
1804                 RETURN(-ENOMEM);
1805         }
1806
1807         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1808
1809         archive_array = req_capsule_client_get(&req->rq_pill,
1810                                                &RMF_MDS_HSM_ARCHIVE);
1811         if (archive_array == NULL)
1812                 GOTO(out, rc = -EPROTO);
1813
1814         if (archives != NULL)
1815                 memcpy(archive_array, archives, archives_size);
1816         else
1817                 *archive_array = archive_count;
1818
1819         ptlrpc_request_set_replen(req);
1820
1821         rc = mdc_queue_wait(req);
1822         GOTO(out, rc);
1823 out:
1824         ptlrpc_req_finished(req);
1825         return rc;
1826 }
1827
1828 static int mdc_ioc_hsm_current_action(struct obd_export *exp,
1829                                       struct md_op_data *op_data)
1830 {
1831         struct hsm_current_action       *hca = op_data->op_data;
1832         struct hsm_current_action       *req_hca;
1833         struct ptlrpc_request           *req;
1834         int                              rc;
1835         ENTRY;
1836
1837         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1838                                    &RQF_MDS_HSM_ACTION);
1839         if (req == NULL)
1840                 RETURN(-ENOMEM);
1841
1842         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_ACTION);
1843         if (rc) {
1844                 ptlrpc_request_free(req);
1845                 RETURN(rc);
1846         }
1847
1848         mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1849                       op_data->op_suppgids[0], 0);
1850
1851         ptlrpc_request_set_replen(req);
1852
1853         rc = mdc_queue_wait(req);
1854         if (rc)
1855                 GOTO(out, rc);
1856
1857         req_hca = req_capsule_server_get(&req->rq_pill,
1858                                          &RMF_MDS_HSM_CURRENT_ACTION);
1859         if (req_hca == NULL)
1860                 GOTO(out, rc = -EPROTO);
1861
1862         *hca = *req_hca;
1863
1864         EXIT;
1865 out:
1866         ptlrpc_req_finished(req);
1867         return rc;
1868 }
1869
1870 static int mdc_ioc_hsm_ct_unregister(struct obd_import *imp)
1871 {
1872         struct ptlrpc_request   *req;
1873         int                      rc;
1874         ENTRY;
1875
1876         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_UNREGISTER,
1877                                         LUSTRE_MDS_VERSION,
1878                                         MDS_HSM_CT_UNREGISTER);
1879         if (req == NULL)
1880                 GOTO(out, rc = -ENOMEM);
1881
1882         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1883
1884         ptlrpc_request_set_replen(req);
1885
1886         rc = mdc_queue_wait(req);
1887         GOTO(out, rc);
1888 out:
1889         ptlrpc_req_finished(req);
1890         return rc;
1891 }
1892
1893 static int mdc_ioc_hsm_state_get(struct obd_export *exp,
1894                                  struct md_op_data *op_data)
1895 {
1896         struct hsm_user_state   *hus = op_data->op_data;
1897         struct hsm_user_state   *req_hus;
1898         struct ptlrpc_request   *req;
1899         int                      rc;
1900         ENTRY;
1901
1902         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1903                                    &RQF_MDS_HSM_STATE_GET);
1904         if (req == NULL)
1905                 RETURN(-ENOMEM);
1906
1907         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_GET);
1908         if (rc != 0) {
1909                 ptlrpc_request_free(req);
1910                 RETURN(rc);
1911         }
1912
1913         mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1914                       op_data->op_suppgids[0], 0);
1915
1916         ptlrpc_request_set_replen(req);
1917
1918         rc = mdc_queue_wait(req);
1919         if (rc)
1920                 GOTO(out, rc);
1921
1922         req_hus = req_capsule_server_get(&req->rq_pill, &RMF_HSM_USER_STATE);
1923         if (req_hus == NULL)
1924                 GOTO(out, rc = -EPROTO);
1925
1926         *hus = *req_hus;
1927
1928         EXIT;
1929 out:
1930         ptlrpc_req_finished(req);
1931         return rc;
1932 }
1933
1934 static int mdc_ioc_hsm_state_set(struct obd_export *exp,
1935                                  struct md_op_data *op_data)
1936 {
1937         struct hsm_state_set    *hss = op_data->op_data;
1938         struct hsm_state_set    *req_hss;
1939         struct ptlrpc_request   *req;
1940         int                      rc;
1941         ENTRY;
1942
1943         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1944                                    &RQF_MDS_HSM_STATE_SET);
1945         if (req == NULL)
1946                 RETURN(-ENOMEM);
1947
1948         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_SET);
1949         if (rc) {
1950                 ptlrpc_request_free(req);
1951                 RETURN(rc);
1952         }
1953
1954         mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1955                       op_data->op_suppgids[0], 0);
1956
1957         /* Copy states */
1958         req_hss = req_capsule_client_get(&req->rq_pill, &RMF_HSM_STATE_SET);
1959         if (req_hss == NULL)
1960                 GOTO(out, rc = -EPROTO);
1961         *req_hss = *hss;
1962
1963         ptlrpc_request_set_replen(req);
1964
1965         ptlrpc_get_mod_rpc_slot(req);
1966         rc = ptlrpc_queue_wait(req);
1967         ptlrpc_put_mod_rpc_slot(req);
1968
1969         GOTO(out, rc);
1970 out:
1971         ptlrpc_req_finished(req);
1972         return rc;
1973 }
1974
1975 static int mdc_ioc_hsm_request(struct obd_export *exp,
1976                                struct hsm_user_request *hur)
1977 {
1978         struct obd_import       *imp = class_exp2cliimp(exp);
1979         struct ptlrpc_request   *req;
1980         struct hsm_request      *req_hr;
1981         struct hsm_user_item    *req_hui;
1982         char                    *req_opaque;
1983         int                      rc;
1984         ENTRY;
1985
1986         req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_REQUEST);
1987         if (req == NULL)
1988                 GOTO(out, rc = -ENOMEM);
1989
1990         req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM, RCL_CLIENT,
1991                              hur->hur_request.hr_itemcount
1992                              * sizeof(struct hsm_user_item));
1993         req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA, RCL_CLIENT,
1994                              hur->hur_request.hr_data_len);
1995
1996         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_REQUEST);
1997         if (rc) {
1998                 ptlrpc_request_free(req);
1999                 RETURN(rc);
2000         }
2001
2002         mdc_pack_body(req, NULL, 0, 0, -1, 0);
2003
2004         /* Copy hsm_request struct */
2005         req_hr = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_REQUEST);
2006         if (req_hr == NULL)
2007                 GOTO(out, rc = -EPROTO);
2008         *req_hr = hur->hur_request;
2009
2010         /* Copy hsm_user_item structs */
2011         req_hui = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM);
2012         if (req_hui == NULL)
2013                 GOTO(out, rc = -EPROTO);
2014         memcpy(req_hui, hur->hur_user_item,
2015                hur->hur_request.hr_itemcount * sizeof(struct hsm_user_item));
2016
2017         /* Copy opaque field */
2018         req_opaque = req_capsule_client_get(&req->rq_pill, &RMF_GENERIC_DATA);
2019         if (req_opaque == NULL)
2020                 GOTO(out, rc = -EPROTO);
2021         memcpy(req_opaque, hur_data(hur), hur->hur_request.hr_data_len);
2022
2023         ptlrpc_request_set_replen(req);
2024
2025         ptlrpc_get_mod_rpc_slot(req);
2026         rc = ptlrpc_queue_wait(req);
2027         ptlrpc_put_mod_rpc_slot(req);
2028
2029         GOTO(out, rc);
2030
2031 out:
2032         ptlrpc_req_finished(req);
2033         return rc;
2034 }
2035
2036 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2037                                 struct lustre_kernelcomm *lk);
2038
2039 static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp,
2040                         struct obd_quotactl *oqctl)
2041 {
2042         struct ptlrpc_request   *req;
2043         struct obd_quotactl     *oqc;
2044         int                      rc;
2045         ENTRY;
2046
2047         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
2048                                         &RQF_MDS_QUOTACTL, LUSTRE_MDS_VERSION,
2049                                         MDS_QUOTACTL);
2050         if (req == NULL)
2051                 RETURN(-ENOMEM);
2052
2053         oqc = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2054         *oqc = *oqctl;
2055
2056         ptlrpc_request_set_replen(req);
2057         ptlrpc_at_set_req_timeout(req);
2058
2059         rc = ptlrpc_queue_wait(req);
2060         if (rc)
2061                 CERROR("ptlrpc_queue_wait failed, rc: %d\n", rc);
2062
2063         if (req->rq_repmsg &&
2064             (oqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL))) {
2065                 *oqctl = *oqc;
2066         } else if (!rc) {
2067                 CERROR ("Can't unpack obd_quotactl\n");
2068                 rc = -EPROTO;
2069         }
2070         ptlrpc_req_finished(req);
2071
2072         RETURN(rc);
2073 }
2074
2075 static int mdc_ioc_swap_layouts(struct obd_export *exp,
2076                                 struct md_op_data *op_data)
2077 {
2078         LIST_HEAD(cancels);
2079         struct ptlrpc_request   *req;
2080         int                      rc, count;
2081         struct mdc_swap_layouts *msl, *payload;
2082         ENTRY;
2083
2084         msl = op_data->op_data;
2085
2086         /* When the MDT will get the MDS_SWAP_LAYOUTS RPC the
2087          * first thing it will do is to cancel the 2 layout
2088          * locks held by this client.
2089          * So the client must cancel its layout locks on the 2 fids
2090          * with the request RPC to avoid extra RPC round trips.
2091          */
2092         count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels,
2093                                         LCK_EX, MDS_INODELOCK_LAYOUT |
2094                                         MDS_INODELOCK_XATTR);
2095         count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels,
2096                                          LCK_EX, MDS_INODELOCK_LAYOUT |
2097                                          MDS_INODELOCK_XATTR);
2098
2099         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2100                                    &RQF_MDS_SWAP_LAYOUTS);
2101         if (req == NULL) {
2102                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
2103                 RETURN(-ENOMEM);
2104         }
2105
2106         rc = mdc_prep_elc_req(exp, req, MDS_SWAP_LAYOUTS, &cancels, count);
2107         if (rc) {
2108                 ptlrpc_request_free(req);
2109                 RETURN(rc);
2110         }
2111
2112         mdc_swap_layouts_pack(req, op_data);
2113
2114         payload = req_capsule_client_get(&req->rq_pill, &RMF_SWAP_LAYOUTS);
2115         LASSERT(payload);
2116
2117         *payload = *msl;
2118
2119         ptlrpc_request_set_replen(req);
2120
2121         rc = ptlrpc_queue_wait(req);
2122         if (rc)
2123                 GOTO(out, rc);
2124         EXIT;
2125
2126 out:
2127         ptlrpc_req_finished(req);
2128         return rc;
2129 }
2130
2131 static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2132                          void *karg, void __user *uarg)
2133 {
2134         struct obd_device *obd = exp->exp_obd;
2135         struct obd_ioctl_data *data = karg;
2136         struct obd_import *imp = obd->u.cli.cl_import;
2137         int rc;
2138         ENTRY;
2139
2140         if (!try_module_get(THIS_MODULE)) {
2141                 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2142                        module_name(THIS_MODULE));
2143                 return -EINVAL;
2144         }
2145         switch (cmd) {
2146         case OBD_IOC_FID2PATH:
2147                 rc = mdc_ioc_fid2path(exp, karg);
2148                 GOTO(out, rc);
2149         case LL_IOC_HSM_CT_START:
2150                 rc = mdc_ioc_hsm_ct_start(exp, karg);
2151                 /* ignore if it was already registered on this MDS. */
2152                 if (rc == -EEXIST)
2153                         rc = 0;
2154                 GOTO(out, rc);
2155         case LL_IOC_HSM_PROGRESS:
2156                 rc = mdc_ioc_hsm_progress(exp, karg);
2157                 GOTO(out, rc);
2158         case LL_IOC_HSM_STATE_GET:
2159                 rc = mdc_ioc_hsm_state_get(exp, karg);
2160                 GOTO(out, rc);
2161         case LL_IOC_HSM_STATE_SET:
2162                 rc = mdc_ioc_hsm_state_set(exp, karg);
2163                 GOTO(out, rc);
2164         case LL_IOC_HSM_ACTION:
2165                 rc = mdc_ioc_hsm_current_action(exp, karg);
2166                 GOTO(out, rc);
2167         case LL_IOC_HSM_REQUEST:
2168                 rc = mdc_ioc_hsm_request(exp, karg);
2169                 GOTO(out, rc);
2170         case OBD_IOC_CLIENT_RECOVER:
2171                 rc = ptlrpc_recover_import(imp, data->ioc_inlbuf1, 0);
2172                 if (rc < 0)
2173                         GOTO(out, rc);
2174                 GOTO(out, rc = 0);
2175         case IOC_OSC_SET_ACTIVE:
2176                 rc = ptlrpc_set_import_active(imp, data->ioc_offset);
2177                 GOTO(out, rc);
2178         /*
2179          * Normally IOC_OBD_STATFS, OBD_IOC_QUOTACTL iocontrol are handled by
2180          * LMV instead of MDC. But when the cluster is upgraded from 1.8,
2181          * there'd be no LMV layer thus we might be called here. Eventually
2182          * this code should be removed.
2183          * bz20731, LU-592.
2184          */
2185         case IOC_OBD_STATFS: {
2186                 struct obd_statfs stat_buf = {0};
2187
2188                 if (*((__u32 *) data->ioc_inlbuf2) != 0)
2189                         GOTO(out, rc = -ENODEV);
2190
2191                 /* copy UUID */
2192                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(obd),
2193                                  min((int)data->ioc_plen2,
2194                                      (int)sizeof(struct obd_uuid))))
2195                         GOTO(out, rc = -EFAULT);
2196
2197                 rc = mdc_statfs(NULL, obd->obd_self_export, &stat_buf,
2198                                 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
2199                                 0);
2200                 if (rc != 0)
2201                         GOTO(out, rc);
2202
2203                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
2204                                      min((int) data->ioc_plen1,
2205                                          (int) sizeof(stat_buf))))
2206                         GOTO(out, rc = -EFAULT);
2207
2208                 GOTO(out, rc = 0);
2209         }
2210         case OBD_IOC_QUOTACTL: {
2211                 struct if_quotactl *qctl = karg;
2212                 struct obd_quotactl *oqctl;
2213
2214                 OBD_ALLOC_PTR(oqctl);
2215                 if (oqctl == NULL)
2216                         GOTO(out, rc = -ENOMEM);
2217
2218                 QCTL_COPY(oqctl, qctl);
2219                 rc = obd_quotactl(exp, oqctl);
2220                 if (rc == 0) {
2221                         QCTL_COPY(qctl, oqctl);
2222                         qctl->qc_valid = QC_MDTIDX;
2223                         qctl->obd_uuid = obd->u.cli.cl_target_uuid;
2224                 }
2225
2226                 OBD_FREE_PTR(oqctl);
2227                 GOTO(out, rc);
2228         }
2229         case LL_IOC_GET_CONNECT_FLAGS:
2230                 if (copy_to_user(uarg, exp_connect_flags_ptr(exp),
2231                                  sizeof(*exp_connect_flags_ptr(exp))))
2232                         GOTO(out, rc = -EFAULT);
2233
2234                 GOTO(out, rc = 0);
2235         case LL_IOC_LOV_SWAP_LAYOUTS:
2236                 rc = mdc_ioc_swap_layouts(exp, karg);
2237                 GOTO(out, rc);
2238         default:
2239                 CERROR("unrecognised ioctl: cmd = %#x\n", cmd);
2240                 GOTO(out, rc = -ENOTTY);
2241         }
2242 out:
2243         module_put(THIS_MODULE);
2244
2245         return rc;
2246 }
2247
2248 static int mdc_get_info_rpc(struct obd_export *exp,
2249                             u32 keylen, void *key,
2250                             u32 vallen, void *val)
2251 {
2252         struct obd_import      *imp = class_exp2cliimp(exp);
2253         struct ptlrpc_request  *req;
2254         char                   *tmp;
2255         int                     rc = -EINVAL;
2256         ENTRY;
2257
2258         req = ptlrpc_request_alloc(imp, &RQF_MDS_GET_INFO);
2259         if (req == NULL)
2260                 RETURN(-ENOMEM);
2261
2262         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY,
2263                              RCL_CLIENT, keylen);
2264         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VALLEN,
2265                              RCL_CLIENT, sizeof(vallen));
2266
2267         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_INFO);
2268         if (rc) {
2269                 ptlrpc_request_free(req);
2270                 RETURN(rc);
2271         }
2272
2273         tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
2274         memcpy(tmp, key, keylen);
2275         tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_VALLEN);
2276         memcpy(tmp, &vallen, sizeof(vallen));
2277
2278         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VAL,
2279                              RCL_SERVER, vallen);
2280         ptlrpc_request_set_replen(req);
2281
2282         rc = ptlrpc_queue_wait(req);
2283         /* -EREMOTE means the get_info result is partial, and it needs to
2284          * continue on another MDT, see fid2path part in lmv_iocontrol */
2285         if (rc == 0 || rc == -EREMOTE) {
2286                 tmp = req_capsule_server_get(&req->rq_pill, &RMF_GETINFO_VAL);
2287                 memcpy(val, tmp, vallen);
2288                 if (ptlrpc_rep_need_swab(req)) {
2289                         if (KEY_IS(KEY_FID2PATH))
2290                                 lustre_swab_fid2path(val);
2291                 }
2292         }
2293         ptlrpc_req_finished(req);
2294
2295         RETURN(rc);
2296 }
2297
2298 static void lustre_swab_hai(struct hsm_action_item *h)
2299 {
2300         __swab32s(&h->hai_len);
2301         __swab32s(&h->hai_action);
2302         lustre_swab_lu_fid(&h->hai_fid);
2303         lustre_swab_lu_fid(&h->hai_dfid);
2304         __swab64s(&h->hai_cookie);
2305         __swab64s(&h->hai_extent.offset);
2306         __swab64s(&h->hai_extent.length);
2307         __swab64s(&h->hai_gid);
2308 }
2309
2310 static void lustre_swab_hal(struct hsm_action_list *h)
2311 {
2312         struct hsm_action_item  *hai;
2313         __u32                    i;
2314
2315         __swab32s(&h->hal_version);
2316         __swab32s(&h->hal_count);
2317         __swab32s(&h->hal_archive_id);
2318         __swab64s(&h->hal_flags);
2319         hai = hai_first(h);
2320         for (i = 0; i < h->hal_count; i++, hai = hai_next(hai))
2321                 lustre_swab_hai(hai);
2322 }
2323
2324 static void lustre_swab_kuch(struct kuc_hdr *l)
2325 {
2326         __swab16s(&l->kuc_magic);
2327         /* __u8 l->kuc_transport */
2328         __swab16s(&l->kuc_msgtype);
2329         __swab16s(&l->kuc_msglen);
2330 }
2331
2332 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2333                                 struct lustre_kernelcomm *lk)
2334 {
2335         struct obd_import *imp = class_exp2cliimp(exp);
2336         int rc = 0;
2337
2338         if (lk->lk_group != KUC_GRP_HSM) {
2339                 CERROR("Bad copytool group %d\n", lk->lk_group);
2340                 return -EINVAL;
2341         }
2342
2343         CDEBUG(D_HSM, "CT start r%d w%d u%d g%d f%#x\n", lk->lk_rfd, lk->lk_wfd,
2344                lk->lk_uid, lk->lk_group, lk->lk_flags);
2345
2346         if (lk->lk_flags & LK_FLG_STOP) {
2347                 /* Unregister with the coordinator */
2348                 rc = mdc_ioc_hsm_ct_unregister(imp);
2349         } else {
2350                 __u32 *archives = NULL;
2351
2352                 if ((lk->lk_flags & LK_FLG_DATANR) && lk->lk_data_count > 0)
2353                         archives = lk->lk_data;
2354
2355                 rc = mdc_ioc_hsm_ct_register(imp, lk->lk_data_count, archives);
2356         }
2357
2358         return rc;
2359 }
2360
2361 /**
2362  * Send a message to any listening copytools
2363  * @param val KUC message (kuc_hdr + hsm_action_list)
2364  * @param len total length of message
2365  */
2366 static int mdc_hsm_copytool_send(const struct obd_uuid *uuid,
2367                                  size_t len, void *val)
2368 {
2369         struct kuc_hdr          *lh = (struct kuc_hdr *)val;
2370         struct hsm_action_list  *hal = (struct hsm_action_list *)(lh + 1);
2371         int                      rc;
2372         ENTRY;
2373
2374         if (len < sizeof(*lh) + sizeof(*hal)) {
2375                 CERROR("Short HSM message %zu < %zu\n", len,
2376                        sizeof(*lh) + sizeof(*hal));
2377                 RETURN(-EPROTO);
2378         }
2379         if (lh->kuc_magic == __swab16(KUC_MAGIC)) {
2380                 lustre_swab_kuch(lh);
2381                 lustre_swab_hal(hal);
2382         } else if (lh->kuc_magic != KUC_MAGIC) {
2383                 CERROR("Bad magic %x!=%x\n", lh->kuc_magic, KUC_MAGIC);
2384                 RETURN(-EPROTO);
2385         }
2386
2387         CDEBUG(D_HSM, " Received message mg=%x t=%d m=%d l=%d actions=%d "
2388                "on %s\n",
2389                lh->kuc_magic, lh->kuc_transport, lh->kuc_msgtype,
2390                lh->kuc_msglen, hal->hal_count, hal->hal_fsname);
2391
2392         /* Broadcast to HSM listeners */
2393         rc = libcfs_kkuc_group_put(uuid, KUC_GRP_HSM, lh);
2394
2395         RETURN(rc);
2396 }
2397
2398 /**
2399  * callback function passed to kuc for re-registering each HSM copytool
2400  * running on MDC, after MDT shutdown/recovery.
2401  * @param data copytool registration data
2402  * @param cb_arg callback argument (obd_import)
2403  */
2404 static int mdc_hsm_ct_reregister(void *data, void *cb_arg)
2405 {
2406         struct obd_import *imp = (struct obd_import *)cb_arg;
2407         struct kkuc_ct_data *kcd = data;
2408         __u32 *archives = NULL;
2409         int rc;
2410
2411         if (kcd == NULL ||
2412             (kcd->kcd_magic != KKUC_CT_DATA_ARRAY_MAGIC &&
2413              kcd->kcd_magic != KKUC_CT_DATA_BITMAP_MAGIC))
2414                 return -EPROTO;
2415
2416         if (kcd->kcd_magic == KKUC_CT_DATA_BITMAP_MAGIC) {
2417                 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2418                        "(archive=%#x)\n", imp->imp_obd->obd_name,
2419                        kcd->kcd_nr_archives);
2420         } else {
2421                 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2422                        "(archive nr = %u)\n",
2423                        imp->imp_obd->obd_name, kcd->kcd_nr_archives);
2424                 if (kcd->kcd_nr_archives != 0)
2425                         archives = kcd->kcd_archives;
2426         }
2427
2428         rc = mdc_ioc_hsm_ct_register(imp, kcd->kcd_nr_archives, archives);
2429         /* ignore error if the copytool is already registered */
2430         return (rc == -EEXIST) ? 0 : rc;
2431 }
2432
2433 /**
2434  * Re-establish all kuc contexts with MDT
2435  * after MDT shutdown/recovery.
2436  */
2437 static int mdc_kuc_reregister(struct obd_import *imp)
2438 {
2439         /* re-register HSM agents */
2440         return libcfs_kkuc_group_foreach(&imp->imp_obd->obd_uuid, KUC_GRP_HSM,
2441                                          mdc_hsm_ct_reregister, imp);
2442 }
2443
2444 static int mdc_set_info_async(const struct lu_env *env,
2445                               struct obd_export *exp,
2446                               u32 keylen, void *key,
2447                               u32 vallen, void *val,
2448                               struct ptlrpc_request_set *set)
2449 {
2450         struct obd_import       *imp = class_exp2cliimp(exp);
2451         int                      rc;
2452         ENTRY;
2453
2454         if (KEY_IS(KEY_READ_ONLY)) {
2455                 if (vallen != sizeof(int))
2456                         RETURN(-EINVAL);
2457
2458                 spin_lock(&imp->imp_lock);
2459                 if (*((int *)val)) {
2460                         imp->imp_connect_flags_orig |= OBD_CONNECT_RDONLY;
2461                         imp->imp_connect_data.ocd_connect_flags |=
2462                                                         OBD_CONNECT_RDONLY;
2463                 } else {
2464                         imp->imp_connect_flags_orig &= ~OBD_CONNECT_RDONLY;
2465                         imp->imp_connect_data.ocd_connect_flags &=
2466                                                         ~OBD_CONNECT_RDONLY;
2467                 }
2468                 spin_unlock(&imp->imp_lock);
2469
2470                 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2471                                        keylen, key, vallen, val, set);
2472                 RETURN(rc);
2473         }
2474         if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2475                 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2476                                        keylen, key, vallen, val, set);
2477                 RETURN(rc);
2478         }
2479         if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) {
2480                 rc = mdc_hsm_copytool_send(&imp->imp_obd->obd_uuid, vallen,
2481                                            val);
2482                 RETURN(rc);
2483         }
2484
2485         if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2486                 __u32 *default_easize = val;
2487
2488                 exp->exp_obd->u.cli.cl_default_mds_easize = *default_easize;
2489                 RETURN(0);
2490         }
2491
2492         rc = osc_set_info_async(env, exp, keylen, key, vallen, val, set);
2493         RETURN(rc);
2494 }
2495
2496 static int mdc_get_info(const struct lu_env *env, struct obd_export *exp,
2497                         __u32 keylen, void *key, __u32 *vallen, void *val)
2498 {
2499         int rc = -EINVAL;
2500
2501         if (KEY_IS(KEY_MAX_EASIZE)) {
2502                 __u32 mdsize, *max_easize;
2503
2504                 if (*vallen != sizeof(int))
2505                         RETURN(-EINVAL);
2506                 mdsize = *(__u32 *)val;
2507                 if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
2508                         exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
2509                 max_easize = val;
2510                 *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
2511                 RETURN(0);
2512         } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2513                 __u32 *default_easize;
2514
2515                 if (*vallen != sizeof(int))
2516                         RETURN(-EINVAL);
2517                 default_easize = val;
2518                 *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
2519                 RETURN(0);
2520         } else if (KEY_IS(KEY_CONN_DATA)) {
2521                 struct obd_import *imp = class_exp2cliimp(exp);
2522                 struct obd_connect_data *data = val;
2523
2524                 if (*vallen != sizeof(*data))
2525                         RETURN(-EINVAL);
2526
2527                 *data = imp->imp_connect_data;
2528                 RETURN(0);
2529         } else if (KEY_IS(KEY_TGT_COUNT)) {
2530                 *((__u32 *)val) = 1;
2531                 RETURN(0);
2532         }
2533
2534         rc = mdc_get_info_rpc(exp, keylen, key, *vallen, val);
2535
2536         RETURN(rc);
2537 }
2538
2539 static int mdc_fsync(struct obd_export *exp, const struct lu_fid *fid,
2540                      struct ptlrpc_request **request)
2541 {
2542         struct ptlrpc_request *req;
2543         int                    rc;
2544         ENTRY;
2545
2546         *request = NULL;
2547         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SYNC);
2548         if (req == NULL)
2549                 RETURN(-ENOMEM);
2550
2551         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SYNC);
2552         if (rc) {
2553                 ptlrpc_request_free(req);
2554                 RETURN(rc);
2555         }
2556
2557         mdc_pack_body(req, fid, 0, 0, -1, 0);
2558
2559         ptlrpc_request_set_replen(req);
2560
2561         rc = ptlrpc_queue_wait(req);
2562         if (rc)
2563                 ptlrpc_req_finished(req);
2564         else
2565                 *request = req;
2566         RETURN(rc);
2567 }
2568
2569 struct mdc_rmfid_args {
2570         int *mra_rcs;
2571         int mra_nr;
2572 };
2573
2574 int mdc_rmfid_interpret(const struct lu_env *env, struct ptlrpc_request *req,
2575                           void *args, int rc)
2576 {
2577         struct mdc_rmfid_args *aa;
2578         int *rcs, size;
2579         ENTRY;
2580
2581         if (!rc) {
2582                 aa = ptlrpc_req_async_args(aa, req);
2583
2584                 size = req_capsule_get_size(&req->rq_pill, &RMF_RCS,
2585                                             RCL_SERVER);
2586                 LASSERT(size == sizeof(int) * aa->mra_nr);
2587                 rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2588                 LASSERT(rcs);
2589                 LASSERT(aa->mra_rcs);
2590                 LASSERT(aa->mra_nr);
2591                 memcpy(aa->mra_rcs, rcs, size);
2592         }
2593
2594         RETURN(rc);
2595 }
2596
2597 static int mdc_rmfid(struct obd_export *exp, struct fid_array *fa,
2598                      int *rcs, struct ptlrpc_request_set *set)
2599 {
2600         struct ptlrpc_request *req;
2601         struct mdc_rmfid_args *aa;
2602         struct mdt_body *b;
2603         struct lu_fid *tmp;
2604         int rc, flen;
2605         ENTRY;
2606
2607         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_RMFID);
2608         if (req == NULL)
2609                 RETURN(-ENOMEM);
2610
2611         flen = fa->fa_nr * sizeof(struct lu_fid);
2612         req_capsule_set_size(&req->rq_pill, &RMF_FID_ARRAY,
2613                              RCL_CLIENT, flen);
2614         req_capsule_set_size(&req->rq_pill, &RMF_FID_ARRAY,
2615                              RCL_SERVER, flen);
2616         req_capsule_set_size(&req->rq_pill, &RMF_RCS,
2617                              RCL_SERVER, fa->fa_nr * sizeof(__u32));
2618         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_RMFID);
2619         if (rc) {
2620                 ptlrpc_request_free(req);
2621                 RETURN(rc);
2622         }
2623         tmp = req_capsule_client_get(&req->rq_pill, &RMF_FID_ARRAY);
2624         memcpy(tmp, fa->fa_fids, flen);
2625
2626         mdc_pack_body(req, NULL, 0, 0, -1, 0);
2627         b = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
2628         b->mbo_ctime = ktime_get_real_seconds();
2629
2630         ptlrpc_request_set_replen(req);
2631
2632         LASSERT(rcs);
2633         aa = ptlrpc_req_async_args(aa, req);
2634         aa->mra_rcs = rcs;
2635         aa->mra_nr = fa->fa_nr;
2636         req->rq_interpret_reply = mdc_rmfid_interpret;
2637
2638         ptlrpc_set_add_req(set, req);
2639         ptlrpc_check_set(NULL, set);
2640
2641         RETURN(rc);
2642 }
2643
2644 static int mdc_import_event(struct obd_device *obd, struct obd_import *imp,
2645                             enum obd_import_event event)
2646 {
2647         struct client_obd *cli = &obd->u.cli;
2648         int rc = 0;
2649
2650         LASSERT(imp->imp_obd == obd);
2651
2652         switch (event) {
2653         case IMP_EVENT_DISCON:
2654                 spin_lock(&cli->cl_loi_list_lock);
2655                 cli->cl_avail_grant = 0;
2656                 cli->cl_lost_grant = 0;
2657                 spin_unlock(&cli->cl_loi_list_lock);
2658                 break;
2659         case IMP_EVENT_INACTIVE:
2660                 /*
2661                  * Flush current sequence to make client obtain new one
2662                  * from server in case of disconnect/reconnect.
2663                  */
2664                 down_read(&cli->cl_seq_rwsem);
2665                 if (cli->cl_seq)
2666                         seq_client_flush(cli->cl_seq);
2667                 up_read(&cli->cl_seq_rwsem);
2668
2669                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
2670                 break;
2671         case IMP_EVENT_INVALIDATE: {
2672                 struct ldlm_namespace *ns = obd->obd_namespace;
2673                 struct lu_env *env;
2674                 __u16 refcheck;
2675
2676                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2677
2678                 env = cl_env_get(&refcheck);
2679                 if (!IS_ERR(env)) {
2680                         /* Reset grants. All pages go to failing rpcs due to
2681                          * the invalid import.
2682                          */
2683                         osc_io_unplug(env, cli, NULL);
2684
2685                         cfs_hash_for_each_nolock(ns->ns_rs_hash,
2686                                                  osc_ldlm_resource_invalidate,
2687                                                  env, 0);
2688                         cl_env_put(env, &refcheck);
2689                         ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2690                 } else {
2691                         rc = PTR_ERR(env);
2692                 }
2693                 break;
2694         }
2695         case IMP_EVENT_ACTIVE:
2696                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
2697                 /* redo the kuc registration after reconnecting */
2698                 if (rc == 0)
2699                         rc = mdc_kuc_reregister(imp);
2700                 break;
2701         case IMP_EVENT_OCD: {
2702                 struct obd_connect_data *ocd = &imp->imp_connect_data;
2703
2704                 if (OCD_HAS_FLAG(ocd, GRANT))
2705                         osc_init_grant(cli, ocd);
2706
2707                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
2708                 break;
2709         }
2710         case IMP_EVENT_DEACTIVATE:
2711         case IMP_EVENT_ACTIVATE:
2712                 break;
2713         default:
2714                 CERROR("Unknown import event %x\n", event);
2715                 LBUG();
2716         }
2717         RETURN(rc);
2718 }
2719
2720 int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp,
2721                   struct lu_fid *fid, struct md_op_data *op_data)
2722 {
2723         struct client_obd *cli = &exp->exp_obd->u.cli;
2724         int rc = -EIO;
2725
2726         ENTRY;
2727
2728         down_read(&cli->cl_seq_rwsem);
2729         if (cli->cl_seq)
2730                 rc = seq_client_alloc_fid(env, cli->cl_seq, fid);
2731         up_read(&cli->cl_seq_rwsem);
2732
2733         RETURN(rc);
2734 }
2735
2736 static struct obd_uuid *mdc_get_uuid(struct obd_export *exp)
2737 {
2738         struct client_obd *cli = &exp->exp_obd->u.cli;
2739         return &cli->cl_target_uuid;
2740 }
2741
2742 /**
2743  * Determine whether the lock can be canceled before replaying it during
2744  * recovery, non zero value will be return if the lock can be canceled,
2745  * or zero returned for not
2746  */
2747 static int mdc_cancel_weight(struct ldlm_lock *lock)
2748 {
2749         if (lock->l_resource->lr_type != LDLM_IBITS)
2750                 RETURN(0);
2751
2752         /* FIXME: if we ever get into a situation where there are too many
2753          * opened files with open locks on a single node, then we really
2754          * should replay these open locks to reget it */
2755         if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN)
2756                 RETURN(0);
2757
2758         /* Special case for DoM locks, cancel only unused and granted locks */
2759         if (ldlm_has_dom(lock) &&
2760             (lock->l_granted_mode != lock->l_req_mode ||
2761              osc_ldlm_weigh_ast(lock) != 0))
2762                 RETURN(0);
2763
2764         RETURN(1);
2765 }
2766
2767 static int mdc_resource_inode_free(struct ldlm_resource *res)
2768 {
2769         if (res->lr_lvb_inode)
2770                 res->lr_lvb_inode = NULL;
2771
2772         return 0;
2773 }
2774
2775 static struct ldlm_valblock_ops inode_lvbo = {
2776         .lvbo_free = mdc_resource_inode_free
2777 };
2778
2779 static int mdc_llog_init(struct obd_device *obd)
2780 {
2781         struct obd_llog_group   *olg = &obd->obd_olg;
2782         struct llog_ctxt        *ctxt;
2783         int                      rc;
2784
2785         ENTRY;
2786
2787         rc = llog_setup(NULL, obd, olg, LLOG_CHANGELOG_REPL_CTXT, obd,
2788                         &llog_client_ops);
2789         if (rc < 0)
2790                 RETURN(rc);
2791
2792         ctxt = llog_group_get_ctxt(olg, LLOG_CHANGELOG_REPL_CTXT);
2793         llog_initiator_connect(ctxt);
2794         llog_ctxt_put(ctxt);
2795
2796         RETURN(0);
2797 }
2798
2799 static void mdc_llog_finish(struct obd_device *obd)
2800 {
2801         struct llog_ctxt *ctxt;
2802
2803         ENTRY;
2804
2805         ctxt = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
2806         if (ctxt != NULL)
2807                 llog_cleanup(NULL, ctxt);
2808
2809         EXIT;
2810 }
2811
2812 int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg)
2813 {
2814         int rc;
2815
2816         ENTRY;
2817
2818         rc = osc_setup_common(obd, cfg);
2819         if (rc < 0)
2820                 RETURN(rc);
2821
2822         rc = mdc_tunables_init(obd);
2823         if (rc)
2824                 GOTO(err_osc_cleanup, rc);
2825
2826         obd->u.cli.cl_dom_min_inline_repsize = MDC_DOM_DEF_INLINE_REPSIZE;
2827
2828         ns_register_cancel(obd->obd_namespace, mdc_cancel_weight);
2829
2830         obd->obd_namespace->ns_lvbo = &inode_lvbo;
2831
2832         rc = mdc_llog_init(obd);
2833         if (rc) {
2834                 CERROR("%s: failed to setup llogging subsystems: rc = %d\n",
2835                        obd->obd_name, rc);
2836                 GOTO(err_llog_cleanup, rc);
2837         }
2838
2839         rc = mdc_changelog_cdev_init(obd);
2840         if (rc) {
2841                 CERROR("%s: failed to setup changelog char device: rc = %d\n",
2842                        obd->obd_name, rc);
2843                 GOTO(err_changelog_cleanup, rc);
2844         }
2845
2846         RETURN(rc);
2847
2848 err_changelog_cleanup:
2849         mdc_llog_finish(obd);
2850 err_llog_cleanup:
2851         lprocfs_free_md_stats(obd);
2852         ptlrpc_lprocfs_unregister_obd(obd);
2853 err_osc_cleanup:
2854         osc_cleanup_common(obd);
2855         return rc;
2856 }
2857
2858 /* Initialize the default and maximum LOV EA sizes.  This allows
2859  * us to make MDS RPCs with large enough reply buffers to hold a default
2860  * sized EA without having to calculate this (via a call into the
2861  * LOV + OSCs) each time we make an RPC.  The maximum size is also tracked
2862  * but not used to avoid wastefully vmalloc()'ing large reply buffers when
2863  * a large number of stripes is possible.  If a larger reply buffer is
2864  * required it will be reallocated in the ptlrpc layer due to overflow.
2865  */
2866 static int mdc_init_ea_size(struct obd_export *exp, __u32 easize,
2867                             __u32 def_easize)
2868 {
2869         struct obd_device *obd = exp->exp_obd;
2870         struct client_obd *cli = &obd->u.cli;
2871         ENTRY;
2872
2873         if (cli->cl_max_mds_easize < easize)
2874                 cli->cl_max_mds_easize = easize;
2875
2876         if (cli->cl_default_mds_easize < def_easize)
2877                 cli->cl_default_mds_easize = def_easize;
2878
2879         RETURN(0);
2880 }
2881
2882 static int mdc_precleanup(struct obd_device *obd)
2883 {
2884         ENTRY;
2885
2886         osc_precleanup_common(obd);
2887         mdc_changelog_cdev_finish(obd);
2888
2889         obd_cleanup_client_import(obd);
2890         ptlrpc_lprocfs_unregister_obd(obd);
2891         lprocfs_free_md_stats(obd);
2892         mdc_llog_finish(obd);
2893         RETURN(0);
2894 }
2895
2896 static int mdc_cleanup(struct obd_device *obd)
2897 {
2898         return osc_cleanup_common(obd);
2899 }
2900
2901 static const struct obd_ops mdc_obd_ops = {
2902         .o_owner            = THIS_MODULE,
2903         .o_setup            = mdc_setup,
2904         .o_precleanup       = mdc_precleanup,
2905         .o_cleanup          = mdc_cleanup,
2906         .o_add_conn         = client_import_add_conn,
2907         .o_del_conn         = client_import_del_conn,
2908         .o_connect          = client_connect_import,
2909         .o_reconnect        = osc_reconnect,
2910         .o_disconnect       = osc_disconnect,
2911         .o_iocontrol        = mdc_iocontrol,
2912         .o_set_info_async   = mdc_set_info_async,
2913         .o_statfs           = mdc_statfs,
2914         .o_statfs_async     = mdc_statfs_async,
2915         .o_fid_init         = client_fid_init,
2916         .o_fid_fini         = client_fid_fini,
2917         .o_fid_alloc        = mdc_fid_alloc,
2918         .o_import_event     = mdc_import_event,
2919         .o_get_info         = mdc_get_info,
2920         .o_get_uuid         = mdc_get_uuid,
2921         .o_quotactl         = mdc_quotactl,
2922 };
2923
2924 static const struct md_ops mdc_md_ops = {
2925         .m_get_root         = mdc_get_root,
2926         .m_null_inode       = mdc_null_inode,
2927         .m_close            = mdc_close,
2928         .m_create           = mdc_create,
2929         .m_enqueue          = mdc_enqueue,
2930         .m_getattr          = mdc_getattr,
2931         .m_getattr_name     = mdc_getattr_name,
2932         .m_intent_lock      = mdc_intent_lock,
2933         .m_link             = mdc_link,
2934         .m_rename           = mdc_rename,
2935         .m_setattr          = mdc_setattr,
2936         .m_setxattr         = mdc_setxattr,
2937         .m_getxattr         = mdc_getxattr,
2938         .m_fsync                = mdc_fsync,
2939         .m_file_resync          = mdc_file_resync,
2940         .m_read_page            = mdc_read_page,
2941         .m_unlink           = mdc_unlink,
2942         .m_cancel_unused    = mdc_cancel_unused,
2943         .m_init_ea_size     = mdc_init_ea_size,
2944         .m_set_lock_data    = mdc_set_lock_data,
2945         .m_lock_match       = mdc_lock_match,
2946         .m_get_lustre_md    = mdc_get_lustre_md,
2947         .m_free_lustre_md   = mdc_free_lustre_md,
2948         .m_set_open_replay_data = mdc_set_open_replay_data,
2949         .m_clear_open_replay_data = mdc_clear_open_replay_data,
2950         .m_intent_getattr_async = mdc_intent_getattr_async,
2951         .m_revalidate_lock      = mdc_revalidate_lock,
2952         .m_rmfid                = mdc_rmfid,
2953 };
2954
2955 static int __init mdc_init(void)
2956 {
2957         return class_register_type(&mdc_obd_ops, &mdc_md_ops, true, NULL,
2958                                    LUSTRE_MDC_NAME, &mdc_device_type);
2959 }
2960
2961 static void __exit mdc_exit(void)
2962 {
2963         class_unregister_type(LUSTRE_MDC_NAME);
2964 }
2965
2966 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2967 MODULE_DESCRIPTION("Lustre Metadata Client");
2968 MODULE_VERSION(LUSTRE_VERSION_STRING);
2969 MODULE_LICENSE("GPL");
2970
2971 module_init(mdc_init);
2972 module_exit(mdc_exit);