lustre/mdc/mdc_request.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 #define DEBUG_SUBSYSTEM S_MDC
  34
  35 #include <linux/init.h>
  36 #include <linux/kthread.h>
  37 #include <linux/module.h>
  38 #include <linux/pagemap.h>
  39 #include <linux/user_namespace.h>
  40 #include <linux/utsname.h>
  41 #ifdef HAVE_UIDGID_HEADER
  42 # include <linux/uidgid.h>
  43 #endif
  44
  45 #include <lustre_errno.h>
  46
  47 #include <cl_object.h>
  48 #include <llog_swab.h>
  49 #include <lprocfs_status.h>
  50 #include <lustre_acl.h>
  51 #include <lustre_fid.h>
  52 #include <uapi/linux/lustre/lustre_ioctl.h>
  53 #include <lustre_kernelcomm.h>
  54 #include <lustre_lmv.h>
  55 #include <lustre_log.h>
  56 #include <lustre_swab.h>
  57 #include <obd_class.h>
  58 #include <lustre_osc.h>
  59
  60 #include "mdc_internal.h"
  61
  62 #define REQUEST_MINOR 244
  63
  64 static int mdc_cleanup(struct obd_device *obd);
  65
  66 static inline int mdc_queue_wait(struct ptlrpc_request *req)
  67 {
  68         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
  69         int rc;
  70
  71         /* obd_get_request_slot() ensures that this client has no more
  72          * than cl_max_rpcs_in_flight RPCs simultaneously inf light
  73          * against an MDT. */
  74         rc = obd_get_request_slot(cli);
  75         if (rc != 0)
  76                 return rc;
  77
  78         rc = ptlrpc_queue_wait(req);
  79         obd_put_request_slot(cli);
  80
  81         return rc;
  82 }
  83
  84 /*
  85  * Send MDS_GET_ROOT RPC to fetch root FID.
  86  *
  87  * If \a fileset is not NULL it should contain a subdirectory off
  88  * the ROOT/ directory to be mounted on the client. Return the FID
  89  * of the subdirectory to the client to mount onto its mountpoint.
  90  *
  91  * \param[in]   imp     MDC import
  92  * \param[in]   fileset fileset name, which could be NULL
  93  * \param[out]  rootfid root FID of this mountpoint
  94  * \param[out]  pc      root capa will be unpacked and saved in this pointer
  95  *
  96  * \retval      0 on success, negative errno on failure
  97  */
  98 static int mdc_get_root(struct obd_export *exp, const char *fileset,
  99                          struct lu_fid *rootfid)
 100 {
 101         struct ptlrpc_request   *req;
 102         struct mdt_body         *body;
 103         int                      rc;
 104
 105         ENTRY;
 106
 107         if (fileset && !(exp_connect_flags(exp) & OBD_CONNECT_SUBTREE))
 108                 RETURN(-ENOTSUPP);
 109
 110         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 111                                 &RQF_MDS_GET_ROOT);
 112         if (req == NULL)
 113                 RETURN(-ENOMEM);
 114
 115         if (fileset != NULL)
 116                 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 117                                      strlen(fileset) + 1);
 118         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_ROOT);
 119         if (rc) {
 120                 ptlrpc_request_free(req);
 121                 RETURN(rc);
 122         }
 123         mdc_pack_body(req, NULL, 0, 0, -1, 0);
 124         if (fileset != NULL) {
 125                 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
 126
 127                 memcpy(name, fileset, strlen(fileset));
 128         }
 129         lustre_msg_add_flags(req->rq_reqmsg, LUSTRE_IMP_FULL);
 130         req->rq_send_state = LUSTRE_IMP_FULL;
 131
 132         ptlrpc_request_set_replen(req);
 133
 134         rc = ptlrpc_queue_wait(req);
 135         if (rc)
 136                 GOTO(out, rc);
 137
 138         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 139         if (body == NULL)
 140                 GOTO(out, rc = -EPROTO);
 141
 142         *rootfid = body->mbo_fid1;
 143         CDEBUG(D_NET, "root fid="DFID", last_committed=%llu\n",
 144                PFID(rootfid), lustre_msg_get_last_committed(req->rq_repmsg));
 145         EXIT;
 146 out:
 147         ptlrpc_req_finished(req);
 148
 149         return rc;
 150 }
 151
 152 /*
 153  * This function now is known to always saying that it will receive 4 buffers
 154  * from server. Even for cases when acl_size and md_size is zero, RPC header
 155  * will contain 4 fields and RPC itself will contain zero size fields. This is
 156  * because mdt_getattr*() _always_ returns 4 fields, but if acl is not needed
 157  * and thus zero, it shrinks it, making zero size. The same story about
 158  * md_size. And this is course of problem when client waits for smaller number
 159  * of fields. This issue will be fixed later when client gets aware of RPC
 160  * layouts.  --umka
 161  */
 162 static int mdc_getattr_common(struct obd_export *exp,
 163                               struct ptlrpc_request *req)
 164 {
 165         struct req_capsule *pill = &req->rq_pill;
 166         struct mdt_body    *body;
 167         void               *eadata;
 168         int                 rc;
 169         ENTRY;
 170
 171         /* Request message already built. */
 172         rc = ptlrpc_queue_wait(req);
 173         if (rc != 0)
 174                 RETURN(rc);
 175
 176         /* sanity check for the reply */
 177         body = req_capsule_server_get(pill, &RMF_MDT_BODY);
 178         if (body == NULL)
 179                 RETURN(-EPROTO);
 180
 181         CDEBUG(D_NET, "mode: %o\n", body->mbo_mode);
 182
 183         mdc_update_max_ea_from_body(exp, body);
 184         if (body->mbo_eadatasize != 0) {
 185                 eadata = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
 186                                                       body->mbo_eadatasize);
 187                 if (eadata == NULL)
 188                         RETURN(-EPROTO);
 189         }
 190
 191         RETURN(0);
 192 }
 193
 194 static void mdc_reset_acl_req(struct ptlrpc_request *req)
 195 {
 196         spin_lock(&req->rq_early_free_lock);
 197         sptlrpc_cli_free_repbuf(req);
 198         req->rq_repbuf = NULL;
 199         req->rq_repbuf_len = 0;
 200         req->rq_repdata = NULL;
 201         req->rq_reqdata_len = 0;
 202         spin_unlock(&req->rq_early_free_lock);
 203 }
 204
 205 static int mdc_getattr(struct obd_export *exp, struct md_op_data *op_data,
 206                        struct ptlrpc_request **request)
 207 {
 208         struct ptlrpc_request *req;
 209         struct obd_import *imp = class_exp2cliimp(exp);
 210         __u32 acl_bufsize = LUSTRE_POSIX_ACL_MAX_SIZE_OLD;
 211         int rc;
 212         ENTRY;
 213
 214         /* Single MDS without an LMV case */
 215         if (op_data->op_flags & MF_GET_MDT_IDX) {
 216                 op_data->op_mds = 0;
 217                 RETURN(0);
 218         }
 219
 220         *request = NULL;
 221         req = ptlrpc_request_alloc(imp, &RQF_MDS_GETATTR);
 222         if (req == NULL)
 223                 RETURN(-ENOMEM);
 224
 225         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR);
 226         if (rc) {
 227                 ptlrpc_request_free(req);
 228                 RETURN(rc);
 229         }
 230
 231 again:
 232         mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
 233                       op_data->op_mode, -1, 0);
 234         req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
 235         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 236                              op_data->op_mode);
 237         ptlrpc_request_set_replen(req);
 238
 239         rc = mdc_getattr_common(exp, req);
 240         if (rc) {
 241                 if (rc == -ERANGE) {
 242                         acl_bufsize = MIN(imp->imp_connect_data.ocd_max_easize,
 243                                           XATTR_SIZE_MAX);
 244                         mdc_reset_acl_req(req);
 245                         goto again;
 246                 }
 247
 248                 ptlrpc_req_finished(req);
 249         } else {
 250                 *request = req;
 251         }
 252
 253         RETURN(rc);
 254 }
 255
 256 static int mdc_getattr_name(struct obd_export *exp, struct md_op_data *op_data,
 257                             struct ptlrpc_request **request)
 258 {
 259         struct ptlrpc_request *req;
 260         struct obd_import *imp = class_exp2cliimp(exp);
 261         __u32 acl_bufsize = LUSTRE_POSIX_ACL_MAX_SIZE_OLD;
 262         int rc;
 263         ENTRY;
 264
 265         *request = NULL;
 266         req = ptlrpc_request_alloc(imp, &RQF_MDS_GETATTR_NAME);
 267         if (req == NULL)
 268                 RETURN(-ENOMEM);
 269
 270         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 271                              op_data->op_namelen + 1);
 272
 273         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR_NAME);
 274         if (rc) {
 275                 ptlrpc_request_free(req);
 276                 RETURN(rc);
 277         }
 278
 279         if (op_data->op_name) {
 280                 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
 281                 LASSERT(strnlen(op_data->op_name, op_data->op_namelen) ==
 282                                 op_data->op_namelen);
 283                 memcpy(name, op_data->op_name, op_data->op_namelen);
 284         }
 285
 286 again:
 287         mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
 288                       op_data->op_mode, op_data->op_suppgids[0], 0);
 289         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 290                              op_data->op_mode);
 291         req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
 292         ptlrpc_request_set_replen(req);
 293
 294         rc = mdc_getattr_common(exp, req);
 295         if (rc) {
 296                 if (rc == -ERANGE) {
 297                         acl_bufsize = MIN(imp->imp_connect_data.ocd_max_easize,
 298                                           XATTR_SIZE_MAX);
 299                         mdc_reset_acl_req(req);
 300                         goto again;
 301                 }
 302
 303                 ptlrpc_req_finished(req);
 304         } else {
 305                 *request = req;
 306         }
 307
 308         RETURN(rc);
 309 }
 310
 311 static int mdc_xattr_common(struct obd_export *exp,const struct req_format *fmt,
 312                             const struct lu_fid *fid, int opcode, u64 valid,
 313                             const char *xattr_name, const char *input,
 314                             int input_size, int output_size, int flags,
 315                             __u32 suppgid, struct ptlrpc_request **request)
 316 {
 317         struct ptlrpc_request *req;
 318         int   xattr_namelen = 0;
 319         char *tmp;
 320         int   rc;
 321         ENTRY;
 322
 323         *request = NULL;
 324         req = ptlrpc_request_alloc(class_exp2cliimp(exp), fmt);
 325         if (req == NULL)
 326                 RETURN(-ENOMEM);
 327
 328         if (xattr_name) {
 329                 xattr_namelen = strlen(xattr_name) + 1;
 330                 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 331                                      xattr_namelen);
 332         }
 333         if (input_size)
 334                 LASSERT(input);
 335         req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT,
 336                              input_size);
 337
 338         /* get SELinux policy info if any */
 339         rc = sptlrpc_get_sepol(req);
 340         if (rc < 0) {
 341                 ptlrpc_request_free(req);
 342                 RETURN(rc);
 343         }
 344         req_capsule_set_size(&req->rq_pill, &RMF_SELINUX_POL, RCL_CLIENT,
 345                              strlen(req->rq_sepol) ?
 346                              strlen(req->rq_sepol) + 1 : 0);
 347
 348         /* Flush local XATTR locks to get rid of a possible cancel RPC */
 349         if (opcode == MDS_REINT && fid_is_sane(fid) &&
 350             exp->exp_connect_data.ocd_ibits_known & MDS_INODELOCK_XATTR) {
 351                 struct list_head cancels = LIST_HEAD_INIT(cancels);
 352                 int count;
 353
 354                 /* Without that packing would fail */
 355                 if (input_size == 0)
 356                         req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
 357                                              RCL_CLIENT, 0);
 358
 359                 count = mdc_resource_get_unused(exp, fid,
 360                                                 &cancels, LCK_EX,
 361                                                 MDS_INODELOCK_XATTR);
 362
 363                 rc = mdc_prep_elc_req(exp, req, MDS_REINT, &cancels, count);
 364                 if (rc) {
 365                         ptlrpc_request_free(req);
 366                         RETURN(rc);
 367                 }
 368         } else {
 369                 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, opcode);
 370                 if (rc) {
 371                         ptlrpc_request_free(req);
 372                         RETURN(rc);
 373                 }
 374         }
 375
 376         if (opcode == MDS_REINT) {
 377                 struct mdt_rec_setxattr *rec;
 378
 379                 CLASSERT(sizeof(struct mdt_rec_setxattr) ==
 380                          sizeof(struct mdt_rec_reint));
 381                 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
 382                 rec->sx_opcode = REINT_SETXATTR;
 383                 rec->sx_fsuid  = from_kuid(&init_user_ns, current_fsuid());
 384                 rec->sx_fsgid  = from_kgid(&init_user_ns, current_fsgid());
 385                 rec->sx_cap    = cfs_curproc_cap_pack();
 386                 rec->sx_suppgid1 = suppgid;
 387                 rec->sx_suppgid2 = -1;
 388                 rec->sx_fid    = *fid;
 389                 rec->sx_valid  = valid | OBD_MD_FLCTIME;
 390                 rec->sx_time   = ktime_get_real_seconds();
 391                 rec->sx_size   = output_size;
 392                 rec->sx_flags  = flags;
 393         } else {
 394                 mdc_pack_body(req, fid, valid, output_size, suppgid, flags);
 395         }
 396
 397         if (xattr_name) {
 398                 tmp = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
 399                 memcpy(tmp, xattr_name, xattr_namelen);
 400         }
 401         if (input_size) {
 402                 tmp = req_capsule_client_get(&req->rq_pill, &RMF_EADATA);
 403                 memcpy(tmp, input, input_size);
 404         }
 405
 406         mdc_file_sepol_pack(req);
 407
 408         if (req_capsule_has_field(&req->rq_pill, &RMF_EADATA, RCL_SERVER))
 409                 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
 410                                      RCL_SERVER, output_size);
 411         ptlrpc_request_set_replen(req);
 412
 413         /* make rpc */
 414         if (opcode == MDS_REINT)
 415                 mdc_get_mod_rpc_slot(req, NULL);
 416
 417         rc = ptlrpc_queue_wait(req);
 418
 419         if (opcode == MDS_REINT)
 420                 mdc_put_mod_rpc_slot(req, NULL);
 421
 422         if (rc)
 423                 ptlrpc_req_finished(req);
 424         else
 425                 *request = req;
 426         RETURN(rc);
 427 }
 428
 429 static int mdc_setxattr(struct obd_export *exp, const struct lu_fid *fid,
 430                         u64 obd_md_valid, const char *name,
 431                         const void *value, size_t value_size,
 432                         unsigned int xattr_flags, u32 suppgid,
 433                         struct ptlrpc_request **req)
 434 {
 435         LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
 436                 obd_md_valid == OBD_MD_FLXATTRRM);
 437
 438         return mdc_xattr_common(exp, &RQF_MDS_REINT_SETXATTR,
 439                                 fid, MDS_REINT, obd_md_valid, name,
 440                                 value, value_size, 0, xattr_flags, suppgid,
 441                                 req);
 442 }
 443
 444 static int mdc_getxattr(struct obd_export *exp, const struct lu_fid *fid,
 445                         u64 obd_md_valid, const char *name, size_t buf_size,
 446                         struct ptlrpc_request **req)
 447 {
 448         struct mdt_body *body;
 449         int rc;
 450
 451         LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
 452                 obd_md_valid == OBD_MD_FLXATTRLS);
 453
 454         CDEBUG(D_INFO, "%s: get xattr '%s' for "DFID"\n",
 455                exp->exp_obd->obd_name, name, PFID(fid));
 456         rc = mdc_xattr_common(exp, &RQF_MDS_GETXATTR, fid, MDS_GETXATTR,
 457                               obd_md_valid, name, NULL, 0, buf_size, 0, -1,
 458                               req);
 459         if (rc < 0)
 460                 GOTO(out, rc);
 461
 462         body = req_capsule_server_get(&(*req)->rq_pill, &RMF_MDT_BODY);
 463         if (body == NULL)
 464                 GOTO(out, rc = -EPROTO);
 465
 466         /* only detect the xattr size */
 467         if (buf_size == 0) {
 468                 /* LU-11109: Older MDTs do not distinguish
 469                  * between nonexistent xattrs and zero length
 470                  * values in this case. Newer MDTs will return
 471                  * -ENODATA or set OBD_MD_FLXATTR. */
 472                 GOTO(out, rc = body->mbo_eadatasize);
 473         }
 474
 475         if (body->mbo_eadatasize == 0) {
 476                 /* LU-11109: Newer MDTs set OBD_MD_FLXATTR on
 477                  * success so that we can distinguish between
 478                  * zero length value and nonexistent xattr.
 479                  *
 480                  * If OBD_MD_FLXATTR is not set then we keep
 481                  * the old behavior and return -ENODATA for
 482                  * getxattr() when mbo_eadatasize is 0. But
 483                  * -ENODATA only makes sense for getxattr()
 484                  * and not for listxattr(). */
 485                 if (body->mbo_valid & OBD_MD_FLXATTR)
 486                         GOTO(out, rc = 0);
 487                 else if (obd_md_valid == OBD_MD_FLXATTR)
 488                         GOTO(out, rc = -ENODATA);
 489                 else
 490                         GOTO(out, rc = 0);
 491         }
 492
 493         GOTO(out, rc = body->mbo_eadatasize);
 494 out:
 495         if (rc < 0) {
 496                 ptlrpc_req_finished(*req);
 497                 *req = NULL;
 498         }
 499
 500         return rc;
 501 }
 502
 503 #ifdef CONFIG_FS_POSIX_ACL
 504 static int mdc_unpack_acl(struct ptlrpc_request *req, struct lustre_md *md)
 505 {
 506         struct req_capsule     *pill = &req->rq_pill;
 507         struct mdt_body        *body = md->body;
 508         struct posix_acl       *acl;
 509         void                   *buf;
 510         int                     rc;
 511         ENTRY;
 512
 513         if (!body->mbo_aclsize)
 514                 RETURN(0);
 515
 516         buf = req_capsule_server_sized_get(pill, &RMF_ACL, body->mbo_aclsize);
 517
 518         if (!buf)
 519                 RETURN(-EPROTO);
 520
 521         acl = posix_acl_from_xattr(&init_user_ns, buf, body->mbo_aclsize);
 522         if (acl == NULL)
 523                 RETURN(0);
 524         if (IS_ERR(acl)) {
 525                 rc = PTR_ERR(acl);
 526                 CERROR("convert xattr to acl: %d\n", rc);
 527                 RETURN(rc);
 528         }
 529
 530         rc = posix_acl_valid(&init_user_ns, acl);
 531         if (rc) {
 532                 CERROR("validate acl: %d\n", rc);
 533                 posix_acl_release(acl);
 534                 RETURN(rc);
 535         }
 536
 537         md->posix_acl = acl;
 538         RETURN(0);
 539 }
 540 #else
 541 #define mdc_unpack_acl(req, md) 0
 542 #endif
 543
 544 int mdc_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
 545                       struct obd_export *dt_exp, struct obd_export *md_exp,
 546                       struct lustre_md *md)
 547 {
 548         struct req_capsule *pill = &req->rq_pill;
 549         int rc;
 550         ENTRY;
 551
 552         LASSERT(md);
 553         memset(md, 0, sizeof(*md));
 554
 555         md->body = req_capsule_server_get(pill, &RMF_MDT_BODY);
 556         LASSERT(md->body != NULL);
 557
 558         if (md->body->mbo_valid & OBD_MD_FLEASIZE) {
 559                 if (!S_ISREG(md->body->mbo_mode)) {
 560                         CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, should be a "
 561                                "regular file, but is not\n");
 562                         GOTO(out, rc = -EPROTO);
 563                 }
 564
 565                 if (md->body->mbo_eadatasize == 0) {
 566                         CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, "
 567                                "but eadatasize 0\n");
 568                         GOTO(out, rc = -EPROTO);
 569                 }
 570
 571                 md->layout.lb_len = md->body->mbo_eadatasize;
 572                 md->layout.lb_buf = req_capsule_server_sized_get(pill,
 573                                                         &RMF_MDT_MD,
 574                                                         md->layout.lb_len);
 575                 if (md->layout.lb_buf == NULL)
 576                         GOTO(out, rc = -EPROTO);
 577         } else if (md->body->mbo_valid & OBD_MD_FLDIREA) {
 578                 const union lmv_mds_md *lmv;
 579                 size_t lmv_size;
 580
 581                 if (!S_ISDIR(md->body->mbo_mode)) {
 582                         CDEBUG(D_INFO, "OBD_MD_FLDIREA set, should be a "
 583                                "directory, but is not\n");
 584                         GOTO(out, rc = -EPROTO);
 585                 }
 586
 587                 if (md->body->mbo_valid & OBD_MD_MEA) {
 588                         lmv_size = md->body->mbo_eadatasize;
 589                         if (lmv_size == 0) {
 590                                 CDEBUG(D_INFO, "OBD_MD_FLDIREA is set, "
 591                                        "but eadatasize 0\n");
 592                                 RETURN(-EPROTO);
 593                         }
 594
 595                         lmv = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
 596                                                            lmv_size);
 597                         if (lmv == NULL)
 598                                 GOTO(out, rc = -EPROTO);
 599
 600                         rc = md_unpackmd(md_exp, &md->lmv, lmv, lmv_size);
 601                         if (rc < 0)
 602                                 GOTO(out, rc);
 603
 604                         if (rc < (int)sizeof(*md->lmv)) {
 605                                 struct lmv_foreign_md *lfm = md->lfm;
 606
 607                                 /* short (< sizeof(struct lmv_stripe_md))
 608                                  * foreign LMV case
 609                                  */
 610                                 if (lfm->lfm_magic != LMV_MAGIC_FOREIGN) {
 611                                         CDEBUG(D_INFO,
 612                                                "lmv size too small: %d < %d\n",
 613                                                rc, (int)sizeof(*md->lmv));
 614                                         GOTO(out, rc = -EPROTO);
 615                                 }
 616                         }
 617                 }
 618
 619                 /* since 2.12.58 intent_getattr fetches default LMV */
 620                 if (md->body->mbo_valid & OBD_MD_DEFAULT_MEA) {
 621                         lmv_size = sizeof(struct lmv_user_md);
 622                         lmv = req_capsule_server_sized_get(pill,
 623                                                            &RMF_DEFAULT_MDT_MD,
 624                                                            lmv_size);
 625                         if (!lmv)
 626                                 GOTO(out, rc = -EPROTO);
 627
 628                         rc = md_unpackmd(md_exp, &md->default_lmv, lmv,
 629                                          lmv_size);
 630                         if (rc < 0)
 631                                 GOTO(out, rc);
 632
 633                         if (rc < (int)sizeof(*md->default_lmv)) {
 634                                 CDEBUG(D_INFO,
 635                                        "default lmv size too small: %d < %d\n",
 636                                         rc, (int)sizeof(*md->default_lmv));
 637                                 GOTO(out, rc = -EPROTO);
 638                         }
 639                 }
 640         }
 641         rc = 0;
 642
 643         if (md->body->mbo_valid & OBD_MD_FLACL) {
 644                 /* for ACL, it's possible that FLACL is set but aclsize is zero.
 645                  * only when aclsize != 0 there's an actual segment for ACL
 646                  * in reply buffer.
 647                  */
 648                 if (md->body->mbo_aclsize) {
 649                         rc = mdc_unpack_acl(req, md);
 650                         if (rc)
 651                                 GOTO(out, rc);
 652 #ifdef CONFIG_FS_POSIX_ACL
 653                 } else {
 654                         md->posix_acl = NULL;
 655 #endif
 656                 }
 657         }
 658
 659         EXIT;
 660 out:
 661         if (rc) {
 662 #ifdef CONFIG_FS_POSIX_ACL
 663                 posix_acl_release(md->posix_acl);
 664 #endif
 665         }
 666         return rc;
 667 }
 668
 669 int mdc_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
 670 {
 671         ENTRY;
 672         RETURN(0);
 673 }
 674
 675 void mdc_replay_open(struct ptlrpc_request *req)
 676 {
 677         struct md_open_data *mod = req->rq_cb_data;
 678         struct ptlrpc_request *close_req;
 679         struct obd_client_handle *och;
 680         struct lustre_handle old_open_handle = { };
 681         struct mdt_body *body;
 682         ENTRY;
 683
 684         if (mod == NULL) {
 685                 DEBUG_REQ(D_ERROR, req,
 686                           "Can't properly replay without open data.");
 687                 EXIT;
 688                 return;
 689         }
 690
 691         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 692         LASSERT(body != NULL);
 693
 694         spin_lock(&req->rq_lock);
 695         och = mod->mod_och;
 696         if (och && och->och_open_handle.cookie)
 697                 req->rq_early_free_repbuf = 1;
 698         else
 699                 req->rq_early_free_repbuf = 0;
 700         spin_unlock(&req->rq_lock);
 701
 702         if (req->rq_early_free_repbuf) {
 703                 struct lustre_handle *file_open_handle;
 704
 705                 LASSERT(och->och_magic == OBD_CLIENT_HANDLE_MAGIC);
 706
 707                 file_open_handle = &och->och_open_handle;
 708                 CDEBUG(D_HA, "updating handle from %#llx to %#llx\n",
 709                        file_open_handle->cookie, body->mbo_open_handle.cookie);
 710                 old_open_handle = *file_open_handle;
 711                 *file_open_handle = body->mbo_open_handle;
 712         }
 713
 714         close_req = mod->mod_close_req;
 715         if (close_req) {
 716                 __u32 opc = lustre_msg_get_opc(close_req->rq_reqmsg);
 717                 struct mdt_ioepoch *epoch;
 718
 719                 LASSERT(opc == MDS_CLOSE);
 720                 epoch = req_capsule_client_get(&close_req->rq_pill,
 721                                                &RMF_MDT_EPOCH);
 722                 LASSERT(epoch);
 723
 724                 if (req->rq_early_free_repbuf)
 725                         LASSERT(old_open_handle.cookie ==
 726                                 epoch->mio_open_handle.cookie);
 727
 728                 DEBUG_REQ(D_HA, close_req, "updating close body with new fh");
 729                 epoch->mio_open_handle = body->mbo_open_handle;
 730         }
 731         EXIT;
 732 }
 733
 734 void mdc_commit_open(struct ptlrpc_request *req)
 735 {
 736         struct md_open_data *mod = req->rq_cb_data;
 737         if (mod == NULL)
 738                 return;
 739
 740         /**
 741          * No need to touch md_open_data::mod_och, it holds a reference on
 742          * \var mod and will zero references to each other, \var mod will be
 743          * freed after that when md_open_data::mod_och will put the reference.
 744          */
 745
 746         /**
 747          * Do not let open request to disappear as it still may be needed
 748          * for close rpc to happen (it may happen on evict only, otherwise
 749          * ptlrpc_request::rq_replay does not let mdc_commit_open() to be
 750          * called), just mark this rpc as committed to distinguish these 2
 751          * cases, see mdc_close() for details. The open request reference will
 752          * be put along with freeing \var mod.
 753          */
 754         ptlrpc_request_addref(req);
 755         spin_lock(&req->rq_lock);
 756         req->rq_committed = 1;
 757         spin_unlock(&req->rq_lock);
 758         req->rq_cb_data = NULL;
 759         obd_mod_put(mod);
 760 }
 761
 762 int mdc_set_open_replay_data(struct obd_export *exp,
 763                              struct obd_client_handle *och,
 764                              struct lookup_intent *it)
 765 {
 766         struct md_open_data     *mod;
 767         struct mdt_rec_create   *rec;
 768         struct mdt_body         *body;
 769         struct ptlrpc_request   *open_req = it->it_request;
 770         struct obd_import       *imp = open_req->rq_import;
 771         ENTRY;
 772
 773         if (!open_req->rq_replay)
 774                 RETURN(0);
 775
 776         rec = req_capsule_client_get(&open_req->rq_pill, &RMF_REC_REINT);
 777         body = req_capsule_server_get(&open_req->rq_pill, &RMF_MDT_BODY);
 778         LASSERT(rec != NULL);
 779         /* Incoming message in my byte order (it's been swabbed). */
 780         /* Outgoing messages always in my byte order. */
 781         LASSERT(body != NULL);
 782
 783         /* Only if the import is replayable, we set replay_open data */
 784         if (och && imp->imp_replayable) {
 785                 mod = obd_mod_alloc();
 786                 if (mod == NULL) {
 787                         DEBUG_REQ(D_ERROR, open_req,
 788                                   "Can't allocate md_open_data");
 789                         RETURN(0);
 790                 }
 791
 792                 /**
 793                  * Take a reference on \var mod, to be freed on mdc_close().
 794                  * It protects \var mod from being freed on eviction (commit
 795                  * callback is called despite rq_replay flag).
 796                  * Another reference for \var och.
 797                  */
 798                 obd_mod_get(mod);
 799                 obd_mod_get(mod);
 800
 801                 spin_lock(&open_req->rq_lock);
 802                 och->och_mod = mod;
 803                 mod->mod_och = och;
 804                 mod->mod_is_create = it_disposition(it, DISP_OPEN_CREATE) ||
 805                                      it_disposition(it, DISP_OPEN_STRIPE);
 806                 mod->mod_open_req = open_req;
 807                 open_req->rq_cb_data = mod;
 808                 open_req->rq_commit_cb = mdc_commit_open;
 809                 open_req->rq_early_free_repbuf = 1;
 810                 spin_unlock(&open_req->rq_lock);
 811         }
 812
 813         rec->cr_fid2 = body->mbo_fid1;
 814         rec->cr_open_handle_old = body->mbo_open_handle;
 815         open_req->rq_replay_cb = mdc_replay_open;
 816         if (!fid_is_sane(&body->mbo_fid1)) {
 817                 DEBUG_REQ(D_ERROR, open_req,
 818                           "saving replay request with insane FID " DFID,
 819                           PFID(&body->mbo_fid1));
 820                 LBUG();
 821         }
 822
 823         DEBUG_REQ(D_RPCTRACE, open_req, "Set up open replay data");
 824         RETURN(0);
 825 }
 826
 827 static void mdc_free_open(struct md_open_data *mod)
 828 {
 829         int committed = 0;
 830
 831         if (mod->mod_is_create == 0 &&
 832             imp_connect_disp_stripe(mod->mod_open_req->rq_import))
 833                 committed = 1;
 834
 835         /**
 836          * No reason to asssert here if the open request has
 837          * rq_replay == 1. It means that mdc_close failed, and
 838          * close request wasn`t sent. It is not fatal to client.
 839          * The worst thing is eviction if the client gets open lock
 840          **/
 841
 842         DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "free open request rq_replay"
 843                   "= %d\n", mod->mod_open_req->rq_replay);
 844
 845         ptlrpc_request_committed(mod->mod_open_req, committed);
 846         if (mod->mod_close_req)
 847                 ptlrpc_request_committed(mod->mod_close_req, committed);
 848 }
 849
 850 int mdc_clear_open_replay_data(struct obd_export *exp,
 851                                struct obd_client_handle *och)
 852 {
 853         struct md_open_data *mod = och->och_mod;
 854         ENTRY;
 855
 856         /**
 857          * It is possible to not have \var mod in a case of eviction between
 858          * lookup and ll_file_open().
 859          **/
 860         if (mod == NULL)
 861                 RETURN(0);
 862
 863         LASSERT(mod != LP_POISON);
 864         LASSERT(mod->mod_open_req != NULL);
 865
 866         spin_lock(&mod->mod_open_req->rq_lock);
 867         if (mod->mod_och)
 868                 mod->mod_och->och_open_handle.cookie = 0;
 869         mod->mod_open_req->rq_early_free_repbuf = 0;
 870         spin_unlock(&mod->mod_open_req->rq_lock);
 871         mdc_free_open(mod);
 872
 873         mod->mod_och = NULL;
 874         och->och_mod = NULL;
 875         obd_mod_put(mod);
 876
 877         RETURN(0);
 878 }
 879
 880 static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
 881                      struct md_open_data *mod, struct ptlrpc_request **request)
 882 {
 883         struct obd_device     *obd = class_exp2obd(exp);
 884         struct ptlrpc_request *req;
 885         struct req_format     *req_fmt;
 886         size_t                 u32_count = 0;
 887         int                    rc;
 888         int                    saved_rc = 0;
 889         ENTRY;
 890
 891         CDEBUG(D_INODE, "%s: "DFID" file closed with intent: %x\n",
 892                exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
 893                op_data->op_bias);
 894
 895         if (op_data->op_bias & MDS_CLOSE_INTENT) {
 896                 req_fmt = &RQF_MDS_CLOSE_INTENT;
 897                 if (op_data->op_bias & MDS_HSM_RELEASE) {
 898                         /* allocate a FID for volatile file */
 899                         rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2,
 900                                            op_data);
 901                         if (rc < 0) {
 902                                 CERROR("%s: "DFID" allocating FID: rc = %d\n",
 903                                        obd->obd_name, PFID(&op_data->op_fid1),
 904                                        rc);
 905                                 /* save the errcode and proceed to close */
 906                                 saved_rc = rc;
 907                         }
 908                 }
 909                 if (op_data->op_bias & MDS_CLOSE_RESYNC_DONE) {
 910                         size_t count = op_data->op_data_size / sizeof(__u32);
 911
 912                         if (count > INLINE_RESYNC_ARRAY_SIZE)
 913                                 u32_count = count;
 914                 }
 915         } else {
 916                 req_fmt = &RQF_MDS_CLOSE;
 917         }
 918
 919         *request = NULL;
 920         if (OBD_FAIL_CHECK(OBD_FAIL_MDC_CLOSE))
 921                 req = NULL;
 922         else
 923                 req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);
 924
 925         /* Ensure that this close's handle is fixed up during replay. */
 926         if (likely(mod != NULL)) {
 927                 LASSERTF(mod->mod_open_req != NULL &&
 928                          mod->mod_open_req->rq_type != LI_POISON,
 929                          "POISONED open %p!\n", mod->mod_open_req);
 930
 931                 mod->mod_close_req = req;
 932
 933                 DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "matched open");
 934                 /* We no longer want to preserve this open for replay even
 935                  * though the open was committed. b=3632, b=3633 */
 936                 spin_lock(&mod->mod_open_req->rq_lock);
 937                 mod->mod_open_req->rq_replay = 0;
 938                 spin_unlock(&mod->mod_open_req->rq_lock);
 939         } else {
 940                 CDEBUG(D_HA, "couldn't find open req; expecting close error\n");
 941         }
 942         if (req == NULL) {
 943                 /**
 944                  * TODO: repeat close after errors
 945                  */
 946                 CWARN("%s: close of FID "DFID" failed, file reference will be "
 947                       "dropped when this client unmounts or is evicted\n",
 948                       obd->obd_name, PFID(&op_data->op_fid1));
 949                 GOTO(out, rc = -ENOMEM);
 950         }
 951
 952         if (u32_count > 0)
 953                 req_capsule_set_size(&req->rq_pill, &RMF_U32, RCL_CLIENT,
 954                                      u32_count * sizeof(__u32));
 955
 956         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
 957         if (rc) {
 958                 ptlrpc_request_free(req);
 959                 req = NULL;
 960                 GOTO(out, rc);
 961         }
 962
 963         /* To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
 964          * portal whose threads are not taking any DLM locks and are therefore
 965          * always progressing */
 966         req->rq_request_portal = MDS_READPAGE_PORTAL;
 967         ptlrpc_at_set_req_timeout(req);
 968
 969         if (!(exp_connect_flags2(exp) & OBD_CONNECT2_LSOM))
 970                 op_data->op_xvalid &= ~(OP_XVALID_LAZYSIZE |
 971                                         OP_XVALID_LAZYBLOCKS);
 972
 973         mdc_close_pack(req, op_data);
 974
 975         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 976                              obd->u.cli.cl_default_mds_easize);
 977
 978         ptlrpc_request_set_replen(req);
 979
 980         mdc_get_mod_rpc_slot(req, NULL);
 981         rc = ptlrpc_queue_wait(req);
 982         mdc_put_mod_rpc_slot(req, NULL);
 983
 984         if (req->rq_repmsg == NULL) {
 985                 CDEBUG(D_RPCTRACE, "request failed to send: %p, %d\n", req,
 986                        req->rq_status);
 987                 if (rc == 0)
 988                         rc = req->rq_status ?: -EIO;
 989         } else if (rc == 0 || rc == -EAGAIN) {
 990                 struct mdt_body *body;
 991
 992                 rc = lustre_msg_get_status(req->rq_repmsg);
 993                 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
 994                         DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR, err "
 995                                   "= %d", rc);
 996                         if (rc > 0)
 997                                 rc = -rc;
 998                 }
 999                 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1000                 if (body == NULL)
1001                         rc = -EPROTO;
1002         } else if (rc == -ESTALE) {
1003                 /**
1004                  * it can be allowed error after 3633 if open was committed and
1005                  * server failed before close was sent. Let's check if mod
1006                  * exists and return no error in that case
1007                  */
1008                 if (mod) {
1009                         DEBUG_REQ(D_HA, req, "Reset ESTALE = %d", rc);
1010                         LASSERT(mod->mod_open_req != NULL);
1011                         if (mod->mod_open_req->rq_committed)
1012                                 rc = 0;
1013                 }
1014         }
1015
1016 out:
1017         if (mod) {
1018                 if (rc != 0)
1019                         mod->mod_close_req = NULL;
1020                 /* Since now, mod is accessed through open_req only,
1021                  * thus close req does not keep a reference on mod anymore. */
1022                 obd_mod_put(mod);
1023         }
1024         *request = req;
1025
1026         RETURN(rc < 0 ? rc : saved_rc);
1027 }
1028
1029 static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid,
1030                        u64 offset, struct page **pages, int npages,
1031                        struct ptlrpc_request **request)
1032 {
1033         struct ptlrpc_request   *req;
1034         struct ptlrpc_bulk_desc *desc;
1035         int                      i;
1036         wait_queue_head_t        waitq;
1037         int                      resends = 0;
1038         struct l_wait_info       lwi;
1039         int                      rc;
1040         ENTRY;
1041
1042         *request = NULL;
1043         init_waitqueue_head(&waitq);
1044
1045 restart_bulk:
1046         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE);
1047         if (req == NULL)
1048                 RETURN(-ENOMEM);
1049
1050         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE);
1051         if (rc) {
1052                 ptlrpc_request_free(req);
1053                 RETURN(rc);
1054         }
1055
1056         req->rq_request_portal = MDS_READPAGE_PORTAL;
1057         ptlrpc_at_set_req_timeout(req);
1058
1059         desc = ptlrpc_prep_bulk_imp(req, npages, 1,
1060                                     PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
1061                                     MDS_BULK_PORTAL,
1062                                     &ptlrpc_bulk_kiov_pin_ops);
1063         if (desc == NULL) {
1064                 ptlrpc_req_finished(req);
1065                 RETURN(-ENOMEM);
1066         }
1067
1068         /* NB req now owns desc and will free it when it gets freed */
1069         for (i = 0; i < npages; i++)
1070                 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1071                                                  PAGE_SIZE);
1072
1073         mdc_readdir_pack(req, offset, PAGE_SIZE * npages, fid);
1074
1075         ptlrpc_request_set_replen(req);
1076         rc = ptlrpc_queue_wait(req);
1077         if (rc) {
1078                 ptlrpc_req_finished(req);
1079                 if (rc != -ETIMEDOUT)
1080                         RETURN(rc);
1081
1082                 resends++;
1083                 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
1084                         CERROR("%s: too many resend retries: rc = %d\n",
1085                                exp->exp_obd->obd_name, -EIO);
1086                         RETURN(-EIO);
1087                 }
1088                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL,
1089                                        NULL);
1090                 l_wait_event(waitq, 0, &lwi);
1091
1092                 goto restart_bulk;
1093         }
1094
1095         rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1096                                           req->rq_bulk->bd_nob_transferred);
1097         if (rc < 0) {
1098                 ptlrpc_req_finished(req);
1099                 RETURN(rc);
1100         }
1101
1102         if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) {
1103                 CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n",
1104                        exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred,
1105                        PAGE_SIZE * npages);
1106                 ptlrpc_req_finished(req);
1107                 RETURN(-EPROTO);
1108         }
1109
1110         *request = req;
1111         RETURN(0);
1112 }
1113
1114 static void mdc_release_page(struct page *page, int remove)
1115 {
1116         if (remove) {
1117                 lock_page(page);
1118                 if (likely(page->mapping != NULL))
1119                         delete_from_page_cache(page);
1120                 unlock_page(page);
1121         }
1122         put_page(page);
1123 }
1124
1125 static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash,
1126                                     __u64 *start, __u64 *end, int hash64)
1127 {
1128         /*
1129          * Complement of hash is used as an index so that
1130          * radix_tree_gang_lookup() can be used to find a page with starting
1131          * hash _smaller_ than one we are looking for.
1132          */
1133         unsigned long offset = hash_x_index(*hash, hash64);
1134         struct page *page;
1135         int found;
1136
1137         xa_lock_irq(&mapping->i_pages);
1138         found = radix_tree_gang_lookup(&mapping->page_tree,
1139                                        (void **)&page, offset, 1);
1140         if (found > 0 && !radix_tree_exceptional_entry(page)) {
1141                 struct lu_dirpage *dp;
1142
1143                 get_page(page);
1144                 xa_unlock_irq(&mapping->i_pages);
1145                 /*
1146                  * In contrast to find_lock_page() we are sure that directory
1147                  * page cannot be truncated (while DLM lock is held) and,
1148                  * hence, can avoid restart.
1149                  *
1150                  * In fact, page cannot be locked here at all, because
1151                  * mdc_read_page_remote does synchronous io.
1152                  */
1153                 wait_on_page_locked(page);
1154                 if (PageUptodate(page)) {
1155                         dp = kmap(page);
1156                         if (BITS_PER_LONG == 32 && hash64) {
1157                                 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1158                                 *end   = le64_to_cpu(dp->ldp_hash_end) >> 32;
1159                                 *hash  = *hash >> 32;
1160                         } else {
1161                                 *start = le64_to_cpu(dp->ldp_hash_start);
1162                                 *end   = le64_to_cpu(dp->ldp_hash_end);
1163                         }
1164                         if (unlikely(*start == 1 && *hash == 0))
1165                                 *hash = *start;
1166                         else
1167                                 LASSERTF(*start <= *hash, "start = %#llx"
1168                                          ",end = %#llx,hash = %#llx\n",
1169                                          *start, *end, *hash);
1170                         CDEBUG(D_VFSTRACE, "offset %lx [%#llx %#llx],"
1171                               " hash %#llx\n", offset, *start, *end, *hash);
1172                         if (*hash > *end) {
1173                                 kunmap(page);
1174                                 mdc_release_page(page, 0);
1175                                 page = NULL;
1176                         } else if (*end != *start && *hash == *end) {
1177                                 /*
1178                                  * upon hash collision, remove this page,
1179                                  * otherwise put page reference, and
1180                                  * mdc_read_page_remote() will issue RPC to
1181                                  * fetch the page we want.
1182                                  */
1183                                 kunmap(page);
1184                                 mdc_release_page(page,
1185                                     le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1186                                 page = NULL;
1187                         }
1188                 } else {
1189                         put_page(page);
1190                         page = ERR_PTR(-EIO);
1191                 }
1192         } else {
1193                 xa_unlock_irq(&mapping->i_pages);
1194                 page = NULL;
1195         }
1196         return page;
1197 }
1198
1199 /*
1200  * Adjust a set of pages, each page containing an array of lu_dirpages,
1201  * so that each page can be used as a single logical lu_dirpage.
1202  *
1203  * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
1204  * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
1205  * struct lu_dirent.  It has size up to LU_PAGE_SIZE. The ldp_hash_end
1206  * value is used as a cookie to request the next lu_dirpage in a
1207  * directory listing that spans multiple pages (two in this example):
1208  *   ________
1209  *  |        |
1210  * .|--------v-------   -----.
1211  * |s|e|f|p|ent|ent| ... |ent|
1212  * '--|--------------   -----'   Each PAGE contains a single
1213  *    '------.                   lu_dirpage.
1214  * .---------v-------   -----.
1215  * |s|e|f|p|ent| 0 | ... | 0 |
1216  * '-----------------   -----'
1217  *
1218  * However, on hosts where the native VM page size (PAGE_SIZE) is
1219  * larger than LU_PAGE_SIZE, a single host page may contain multiple
1220  * lu_dirpages. After reading the lu_dirpages from the MDS, the
1221  * ldp_hash_end of the first lu_dirpage refers to the one immediately
1222  * after it in the same PAGE (arrows simplified for brevity, but
1223  * in general e0==s1, e1==s2, etc.):
1224  *
1225  * .--------------------   -----.
1226  * |s0|e0|f0|p|ent|ent| ... |ent|
1227  * |---v----------------   -----|
1228  * |s1|e1|f1|p|ent|ent| ... |ent|
1229  * |---v----------------   -----|  Here, each PAGE contains
1230  *             ...                 multiple lu_dirpages.
1231  * |---v----------------   -----|
1232  * |s'|e'|f'|p|ent|ent| ... |ent|
1233  * '---|----------------   -----'
1234  *     v
1235  * .----------------------------.
1236  * |        next PAGE           |
1237  *
1238  * This structure is transformed into a single logical lu_dirpage as follows:
1239  *
1240  * - Replace e0 with e' so the request for the next lu_dirpage gets the page
1241  *   labeled 'next PAGE'.
1242  *
1243  * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
1244  *   a hash collision with the next page exists.
1245  *
1246  * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
1247  *   to the first entry of the next lu_dirpage.
1248  */
1249 #if PAGE_SIZE > LU_PAGE_SIZE
1250 static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs)
1251 {
1252         int i;
1253
1254         for (i = 0; i < cfs_pgs; i++) {
1255                 struct lu_dirpage *dp = kmap(pages[i]);
1256                 struct lu_dirpage *first = dp;
1257                 struct lu_dirent *end_dirent = NULL;
1258                 struct lu_dirent *ent;
1259                 __u64 hash_end = dp->ldp_hash_end;
1260                 __u32 flags = dp->ldp_flags;
1261
1262                 while (--lu_pgs > 0) {
1263                         ent = lu_dirent_start(dp);
1264                         for (end_dirent = ent; ent != NULL;
1265                              end_dirent = ent, ent = lu_dirent_next(ent));
1266
1267                         /* Advance dp to next lu_dirpage. */
1268                         dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
1269
1270                         /* Check if we've reached the end of the PAGE. */
1271                         if (!((unsigned long)dp & ~PAGE_MASK))
1272                                 break;
1273
1274                         /* Save the hash and flags of this lu_dirpage. */
1275                         hash_end = dp->ldp_hash_end;
1276                         flags = dp->ldp_flags;
1277
1278                         /* Check if lu_dirpage contains no entries. */
1279                         if (end_dirent == NULL)
1280                                 break;
1281
1282                         /* Enlarge the end entry lde_reclen from 0 to
1283                          * first entry of next lu_dirpage. */
1284                         LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
1285                         end_dirent->lde_reclen =
1286                                 cpu_to_le16((char *)(dp->ldp_entries) -
1287                                             (char *)end_dirent);
1288                 }
1289
1290                 first->ldp_hash_end = hash_end;
1291                 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
1292                 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
1293
1294                 kunmap(pages[i]);
1295         }
1296         LASSERTF(lu_pgs == 0, "left = %d\n", lu_pgs);
1297 }
1298 #else
1299 #define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0)
1300 #endif  /* PAGE_SIZE > LU_PAGE_SIZE */
1301
1302 /* parameters for readdir page */
1303 struct readpage_param {
1304         struct md_op_data       *rp_mod;
1305         __u64                   rp_off;
1306         int                     rp_hash64;
1307         struct obd_export       *rp_exp;
1308         struct md_callback      *rp_cb;
1309 };
1310
1311 /**
1312  * Read pages from server.
1313  *
1314  * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
1315  * a header lu_dirpage which describes the start/end hash, and whether this
1316  * page is empty (contains no dir entry) or hash collide with next page.
1317  * After client receives reply, several pages will be integrated into dir page
1318  * in PAGE_SIZE (if PAGE_SIZE greater than LU_PAGE_SIZE), and the
1319  * lu_dirpage for this integrated page will be adjusted.
1320  **/
1321 static int mdc_read_page_remote(void *data, struct page *page0)
1322 {
1323         struct readpage_param *rp = data;
1324         struct page **page_pool;
1325         struct page *page;
1326         struct lu_dirpage *dp;
1327         struct md_op_data *op_data = rp->rp_mod;
1328         struct ptlrpc_request *req;
1329         int max_pages;
1330         struct inode *inode;
1331         struct lu_fid *fid;
1332         int rd_pgs = 0; /* number of pages actually read */
1333         int npages;
1334         int i;
1335         int rc;
1336         ENTRY;
1337
1338         max_pages = rp->rp_exp->exp_obd->u.cli.cl_max_pages_per_rpc;
1339         inode = op_data->op_data;
1340         fid = &op_data->op_fid1;
1341         LASSERT(inode != NULL);
1342
1343         OBD_ALLOC(page_pool, sizeof(page_pool[0]) * max_pages);
1344         if (page_pool != NULL) {
1345                 page_pool[0] = page0;
1346         } else {
1347                 page_pool = &page0;
1348                 max_pages = 1;
1349         }
1350
1351         for (npages = 1; npages < max_pages; npages++) {
1352                 page = __page_cache_alloc(mapping_gfp_mask(inode->i_mapping)
1353                                           | __GFP_COLD);
1354                 if (page == NULL)
1355                         break;
1356                 page_pool[npages] = page;
1357         }
1358
1359         rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, page_pool, npages, &req);
1360         if (rc < 0) {
1361                 /* page0 is special, which was added into page cache early */
1362                 delete_from_page_cache(page0);
1363         } else {
1364                 int lu_pgs;
1365
1366                 rd_pgs = (req->rq_bulk->bd_nob_transferred + PAGE_SIZE - 1) >>
1367                         PAGE_SHIFT;
1368                 lu_pgs = req->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
1369                 LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
1370
1371                 CDEBUG(D_INODE, "read %d(%d) pages\n", rd_pgs, lu_pgs);
1372
1373                 mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs);
1374
1375                 SetPageUptodate(page0);
1376         }
1377         unlock_page(page0);
1378
1379         ptlrpc_req_finished(req);
1380         CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages);
1381         for (i = 1; i < npages; i++) {
1382                 unsigned long   offset;
1383                 __u64           hash;
1384                 int ret;
1385
1386                 page = page_pool[i];
1387
1388                 if (rc < 0 || i >= rd_pgs) {
1389                         put_page(page);
1390                         continue;
1391                 }
1392
1393                 SetPageUptodate(page);
1394
1395                 dp = kmap(page);
1396                 hash = le64_to_cpu(dp->ldp_hash_start);
1397                 kunmap(page);
1398
1399                 offset = hash_x_index(hash, rp->rp_hash64);
1400
1401                 prefetchw(&page->flags);
1402                 ret = add_to_page_cache_lru(page, inode->i_mapping, offset,
1403                                             GFP_KERNEL);
1404                 if (ret == 0)
1405                         unlock_page(page);
1406                 else
1407                         CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
1408                                " rc = %d\n", offset, ret);
1409                 put_page(page);
1410         }
1411
1412         if (page_pool != &page0)
1413                 OBD_FREE(page_pool, sizeof(page_pool[0]) * max_pages);
1414
1415         RETURN(rc);
1416 }
1417
1418 /**
1419  * Read dir page from cache first, if it can not find it, read it from
1420  * server and add into the cache.
1421  *
1422  * \param[in] exp       MDC export
1423  * \param[in] op_data   client MD stack parameters, transfering parameters
1424  *                      between different layers on client MD stack.
1425  * \param[in] cb_op     callback required for ldlm lock enqueue during
1426  *                      read page
1427  * \param[in] hash_offset the hash offset of the page to be read
1428  * \param[in] ppage     the page to be read
1429  *
1430  * retval               = 0 get the page successfully
1431  *                      errno(<0) get the page failed
1432  */
1433 static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data,
1434                          struct md_callback *cb_op, __u64 hash_offset,
1435                          struct page **ppage)
1436 {
1437         struct lookup_intent    it = { .it_op = IT_READDIR };
1438         struct page             *page;
1439         struct inode            *dir = op_data->op_data;
1440         struct address_space    *mapping;
1441         struct lu_dirpage       *dp;
1442         __u64                   start = 0;
1443         __u64                   end = 0;
1444         struct lustre_handle    lockh;
1445         struct ptlrpc_request   *enq_req = NULL;
1446         struct readpage_param   rp_param;
1447         int rc;
1448
1449         ENTRY;
1450
1451         *ppage = NULL;
1452
1453         LASSERT(dir != NULL);
1454         mapping = dir->i_mapping;
1455
1456         rc = mdc_intent_lock(exp, op_data, &it, &enq_req,
1457                              cb_op->md_blocking_ast, 0);
1458         if (enq_req != NULL)
1459                 ptlrpc_req_finished(enq_req);
1460
1461         if (rc < 0) {
1462                 CERROR("%s: "DFID" lock enqueue fails: rc = %d\n",
1463                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc);
1464                 RETURN(rc);
1465         }
1466
1467         rc = 0;
1468         lockh.cookie = it.it_lock_handle;
1469         mdc_set_lock_data(exp, &lockh, dir, NULL);
1470
1471         rp_param.rp_off = hash_offset;
1472         rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64;
1473         page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end,
1474                                rp_param.rp_hash64);
1475         if (IS_ERR(page)) {
1476                 CERROR("%s: dir page locate: "DFID" at %llu: rc %ld\n",
1477                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1478                        rp_param.rp_off, PTR_ERR(page));
1479                 GOTO(out_unlock, rc = PTR_ERR(page));
1480         } else if (page != NULL) {
1481                 /*
1482                  * XXX nikita: not entirely correct handling of a corner case:
1483                  * suppose hash chain of entries with hash value HASH crosses
1484                  * border between pages P0 and P1. First both P0 and P1 are
1485                  * cached, seekdir() is called for some entry from the P0 part
1486                  * of the chain. Later P0 goes out of cache. telldir(HASH)
1487                  * happens and finds P1, as it starts with matching hash
1488                  * value. Remaining entries from P0 part of the chain are
1489                  * skipped. (Is that really a bug?)
1490                  *
1491                  * Possible solutions: 0. don't cache P1 is such case, handle
1492                  * it as an "overflow" page. 1. invalidate all pages at
1493                  * once. 2. use HASH|1 as an index for P1.
1494                  */
1495                 GOTO(hash_collision, page);
1496         }
1497
1498         rp_param.rp_exp = exp;
1499         rp_param.rp_mod = op_data;
1500         page = read_cache_page(mapping,
1501                                hash_x_index(rp_param.rp_off,
1502                                             rp_param.rp_hash64),
1503                                mdc_read_page_remote, &rp_param);
1504         if (IS_ERR(page)) {
1505                 CDEBUG(D_INFO, "%s: read cache page: "DFID" at %llu: %ld\n",
1506                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1507                        rp_param.rp_off, PTR_ERR(page));
1508                 GOTO(out_unlock, rc = PTR_ERR(page));
1509         }
1510
1511         wait_on_page_locked(page);
1512         (void)kmap(page);
1513         if (!PageUptodate(page)) {
1514                 CERROR("%s: page not updated: "DFID" at %llu: rc %d\n",
1515                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1516                        rp_param.rp_off, -5);
1517                 goto fail;
1518         }
1519         if (!PageChecked(page))
1520                 SetPageChecked(page);
1521         if (PageError(page)) {
1522                 CERROR("%s: page error: "DFID" at %llu: rc %d\n",
1523                        exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1524                        rp_param.rp_off, -5);
1525                 goto fail;
1526         }
1527
1528 hash_collision:
1529         dp = page_address(page);
1530         if (BITS_PER_LONG == 32 && rp_param.rp_hash64) {
1531                 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1532                 end   = le64_to_cpu(dp->ldp_hash_end) >> 32;
1533                 rp_param.rp_off = hash_offset >> 32;
1534         } else {
1535                 start = le64_to_cpu(dp->ldp_hash_start);
1536                 end   = le64_to_cpu(dp->ldp_hash_end);
1537                 rp_param.rp_off = hash_offset;
1538         }
1539         if (end == start) {
1540                 LASSERT(start == rp_param.rp_off);
1541                 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
1542 #if BITS_PER_LONG == 32
1543                 CWARN("Real page-wide hash collision at [%llu %llu] with "
1544                       "hash %llu\n", le64_to_cpu(dp->ldp_hash_start),
1545                       le64_to_cpu(dp->ldp_hash_end), hash_offset);
1546 #endif
1547
1548                 /*
1549                  * Fetch whole overflow chain...
1550                  *
1551                  * XXX not yet.
1552                  */
1553                 goto fail;
1554         }
1555         *ppage = page;
1556 out_unlock:
1557         ldlm_lock_decref(&lockh, it.it_lock_mode);
1558         return rc;
1559 fail:
1560         kunmap(page);
1561         mdc_release_page(page, 1);
1562         rc = -EIO;
1563         goto out_unlock;
1564 }
1565
1566 static int mdc_statfs_interpret(const struct lu_env *env,
1567                                 struct ptlrpc_request *req, void *args, int rc)
1568 {
1569         struct obd_info *oinfo = args;
1570         struct obd_statfs *osfs;
1571
1572         if (!rc) {
1573                 osfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1574                 if (!osfs)
1575                         return -EPROTO;
1576
1577                 oinfo->oi_osfs = osfs;
1578
1579                 CDEBUG(D_CACHE, "blocks=%llu free=%llu avail=%llu "
1580                        "objects=%llu free=%llu state=%x\n",
1581                         osfs->os_blocks, osfs->os_bfree, osfs->os_bavail,
1582                         osfs->os_files, osfs->os_ffree, osfs->os_state);
1583         }
1584
1585         oinfo->oi_cb_up(oinfo, rc);
1586
1587         return rc;
1588 }
1589
1590 static int mdc_statfs_async(struct obd_export *exp,
1591                             struct obd_info *oinfo, time64_t max_age,
1592                             struct ptlrpc_request_set *unused)
1593 {
1594         struct ptlrpc_request *req;
1595         struct obd_info *aa;
1596
1597         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp), &RQF_MDS_STATFS,
1598                                         LUSTRE_MDS_VERSION, MDS_STATFS);
1599         if (req == NULL)
1600                 return -ENOMEM;
1601
1602         ptlrpc_request_set_replen(req);
1603         req->rq_interpret_reply = mdc_statfs_interpret;
1604
1605         aa = ptlrpc_req_async_args(aa, req);
1606         *aa = *oinfo;
1607
1608         ptlrpcd_add_req(req);
1609
1610         return 0;
1611 }
1612
1613 static int mdc_statfs(const struct lu_env *env,
1614                       struct obd_export *exp, struct obd_statfs *osfs,
1615                       time64_t max_age, __u32 flags)
1616 {
1617         struct obd_device *obd = class_exp2obd(exp);
1618         struct req_format *fmt;
1619         struct ptlrpc_request *req;
1620         struct obd_statfs *msfs;
1621         struct obd_import *imp = NULL;
1622         int rc;
1623         ENTRY;
1624
1625         /*
1626          * Since the request might also come from lprocfs, so we need
1627          * sync this with client_disconnect_export Bug15684
1628          */
1629         down_read(&obd->u.cli.cl_sem);
1630         if (obd->u.cli.cl_import)
1631                 imp = class_import_get(obd->u.cli.cl_import);
1632         up_read(&obd->u.cli.cl_sem);
1633         if (!imp)
1634                 RETURN(-ENODEV);
1635
1636         fmt = &RQF_MDS_STATFS;
1637         if ((exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS) &&
1638             (flags & OBD_STATFS_SUM))
1639                 fmt = &RQF_MDS_STATFS_NEW;
1640         req = ptlrpc_request_alloc_pack(imp, fmt, LUSTRE_MDS_VERSION,
1641                                         MDS_STATFS);
1642         if (req == NULL)
1643                 GOTO(output, rc = -ENOMEM);
1644
1645         if ((flags & OBD_STATFS_SUM) &&
1646             (exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS)) {
1647                 /* request aggregated states */
1648                 struct mdt_body *body;
1649
1650                 body = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
1651                 if (body == NULL)
1652                         GOTO(out, rc = -EPROTO);
1653                 body->mbo_valid = OBD_MD_FLAGSTATFS;
1654         }
1655
1656         ptlrpc_request_set_replen(req);
1657
1658         if (flags & OBD_STATFS_NODELAY) {
1659                 /* procfs requests not want stay in wait for avoid deadlock */
1660                 req->rq_no_resend = 1;
1661                 req->rq_no_delay = 1;
1662         }
1663
1664         rc = ptlrpc_queue_wait(req);
1665         if (rc) {
1666                 /* check connection error first */
1667                 if (imp->imp_connect_error)
1668                         rc = imp->imp_connect_error;
1669                 GOTO(out, rc);
1670         }
1671
1672         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1673         if (msfs == NULL)
1674                 GOTO(out, rc = -EPROTO);
1675
1676         *osfs = *msfs;
1677         EXIT;
1678 out:
1679         ptlrpc_req_finished(req);
1680 output:
1681         class_import_put(imp);
1682         return rc;
1683 }
1684
1685 static int mdc_ioc_fid2path(struct obd_export *exp, struct getinfo_fid2path *gf)
1686 {
1687         __u32 keylen, vallen;
1688         void *key;
1689         int rc;
1690
1691         if (gf->gf_pathlen > PATH_MAX)
1692                 RETURN(-ENAMETOOLONG);
1693         if (gf->gf_pathlen < 2)
1694                 RETURN(-EOVERFLOW);
1695
1696         /* Key is KEY_FID2PATH + getinfo_fid2path description */
1697         keylen = cfs_size_round(sizeof(KEY_FID2PATH) + sizeof(*gf) +
1698                                 sizeof(struct lu_fid));
1699         OBD_ALLOC(key, keylen);
1700         if (key == NULL)
1701                 RETURN(-ENOMEM);
1702         memcpy(key, KEY_FID2PATH, sizeof(KEY_FID2PATH));
1703         memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)), gf, sizeof(*gf));
1704         memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*gf),
1705                gf->gf_u.gf_root_fid, sizeof(struct lu_fid));
1706         CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
1707                PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno);
1708
1709         if (!fid_is_sane(&gf->gf_fid))
1710                 GOTO(out, rc = -EINVAL);
1711
1712         /* Val is struct getinfo_fid2path result plus path */
1713         vallen = sizeof(*gf) + gf->gf_pathlen;
1714
1715         rc = obd_get_info(NULL, exp, keylen, key, &vallen, gf);
1716         if (rc != 0 && rc != -EREMOTE)
1717                 GOTO(out, rc);
1718
1719         if (vallen <= sizeof(*gf))
1720                 GOTO(out, rc = -EPROTO);
1721         if (vallen > sizeof(*gf) + gf->gf_pathlen)
1722                 GOTO(out, rc = -EOVERFLOW);
1723
1724         CDEBUG(D_IOCTL, "path got "DFID" from %llu #%d: %s\n",
1725                PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno,
1726                gf->gf_pathlen < 512 ? gf->gf_u.gf_path :
1727                /* only log the last 512 characters of the path */
1728                gf->gf_u.gf_path + gf->gf_pathlen - 512);
1729
1730 out:
1731         OBD_FREE(key, keylen);
1732         return rc;
1733 }
1734
1735 static int mdc_ioc_hsm_progress(struct obd_export *exp,
1736                                 struct hsm_progress_kernel *hpk)
1737 {
1738         struct obd_import               *imp = class_exp2cliimp(exp);
1739         struct hsm_progress_kernel      *req_hpk;
1740         struct ptlrpc_request           *req;
1741         int                              rc;
1742         ENTRY;
1743
1744         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_PROGRESS,
1745                                         LUSTRE_MDS_VERSION, MDS_HSM_PROGRESS);
1746         if (req == NULL)
1747                 GOTO(out, rc = -ENOMEM);
1748
1749         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1750
1751         /* Copy hsm_progress struct */
1752         req_hpk = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_PROGRESS);
1753         if (req_hpk == NULL)
1754                 GOTO(out, rc = -EPROTO);
1755
1756         *req_hpk = *hpk;
1757         req_hpk->hpk_errval = lustre_errno_hton(hpk->hpk_errval);
1758
1759         ptlrpc_request_set_replen(req);
1760
1761         mdc_get_mod_rpc_slot(req, NULL);
1762         rc = ptlrpc_queue_wait(req);
1763         mdc_put_mod_rpc_slot(req, NULL);
1764
1765         GOTO(out, rc);
1766 out:
1767         ptlrpc_req_finished(req);
1768         return rc;
1769 }
1770 /**
1771  * Send hsm_ct_register to MDS
1772  *
1773  * \param[in]   imp             import
1774  * \param[in]   archive_count   if in bitmap format, it is the bitmap,
1775  *                              else it is the count of archive_ids
1776  * \param[in]   archives        if in bitmap format, it is NULL,
1777  *                              else it is archive_id lists
1778  */
1779 static int mdc_ioc_hsm_ct_register(struct obd_import *imp, __u32 archive_count,
1780                                    __u32 *archives)
1781 {
1782         struct ptlrpc_request *req;
1783         __u32 *archive_array;
1784         size_t archives_size;
1785         int rc;
1786         ENTRY;
1787
1788         req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_CT_REGISTER);
1789         if (req == NULL)
1790                 RETURN(-ENOMEM);
1791
1792         if (archives != NULL)
1793                 archives_size = sizeof(*archive_array) * archive_count;
1794         else
1795                 archives_size = sizeof(archive_count);
1796
1797         req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_ARCHIVE,
1798                              RCL_CLIENT, archives_size);
1799
1800         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_CT_REGISTER);
1801         if (rc) {
1802                 ptlrpc_request_free(req);
1803                 RETURN(-ENOMEM);
1804         }
1805
1806         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1807
1808         archive_array = req_capsule_client_get(&req->rq_pill,
1809                                                &RMF_MDS_HSM_ARCHIVE);
1810         if (archive_array == NULL)
1811                 GOTO(out, rc = -EPROTO);
1812
1813         if (archives != NULL)
1814                 memcpy(archive_array, archives, archives_size);
1815         else
1816                 *archive_array = archive_count;
1817
1818         ptlrpc_request_set_replen(req);
1819
1820         rc = mdc_queue_wait(req);
1821         GOTO(out, rc);
1822 out:
1823         ptlrpc_req_finished(req);
1824         return rc;
1825 }
1826
1827 static int mdc_ioc_hsm_current_action(struct obd_export *exp,
1828                                       struct md_op_data *op_data)
1829 {
1830         struct hsm_current_action       *hca = op_data->op_data;
1831         struct hsm_current_action       *req_hca;
1832         struct ptlrpc_request           *req;
1833         int                              rc;
1834         ENTRY;
1835
1836         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1837                                    &RQF_MDS_HSM_ACTION);
1838         if (req == NULL)
1839                 RETURN(-ENOMEM);
1840
1841         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_ACTION);
1842         if (rc) {
1843                 ptlrpc_request_free(req);
1844                 RETURN(rc);
1845         }
1846
1847         mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1848                       op_data->op_suppgids[0], 0);
1849
1850         ptlrpc_request_set_replen(req);
1851
1852         rc = mdc_queue_wait(req);
1853         if (rc)
1854                 GOTO(out, rc);
1855
1856         req_hca = req_capsule_server_get(&req->rq_pill,
1857                                          &RMF_MDS_HSM_CURRENT_ACTION);
1858         if (req_hca == NULL)
1859                 GOTO(out, rc = -EPROTO);
1860
1861         *hca = *req_hca;
1862
1863         EXIT;
1864 out:
1865         ptlrpc_req_finished(req);
1866         return rc;
1867 }
1868
1869 static int mdc_ioc_hsm_ct_unregister(struct obd_import *imp)
1870 {
1871         struct ptlrpc_request   *req;
1872         int                      rc;
1873         ENTRY;
1874
1875         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_UNREGISTER,
1876                                         LUSTRE_MDS_VERSION,
1877                                         MDS_HSM_CT_UNREGISTER);
1878         if (req == NULL)
1879                 GOTO(out, rc = -ENOMEM);
1880
1881         mdc_pack_body(req, NULL, 0, 0, -1, 0);
1882
1883         ptlrpc_request_set_replen(req);
1884
1885         rc = mdc_queue_wait(req);
1886         GOTO(out, rc);
1887 out:
1888         ptlrpc_req_finished(req);
1889         return rc;
1890 }
1891
1892 static int mdc_ioc_hsm_state_get(struct obd_export *exp,
1893                                  struct md_op_data *op_data)
1894 {
1895         struct hsm_user_state   *hus = op_data->op_data;
1896         struct hsm_user_state   *req_hus;
1897         struct ptlrpc_request   *req;
1898         int                      rc;
1899         ENTRY;
1900
1901         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1902                                    &RQF_MDS_HSM_STATE_GET);
1903         if (req == NULL)
1904                 RETURN(-ENOMEM);
1905
1906         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_GET);
1907         if (rc != 0) {
1908                 ptlrpc_request_free(req);
1909                 RETURN(rc);
1910         }
1911
1912         mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1913                       op_data->op_suppgids[0], 0);
1914
1915         ptlrpc_request_set_replen(req);
1916
1917         rc = mdc_queue_wait(req);
1918         if (rc)
1919                 GOTO(out, rc);
1920
1921         req_hus = req_capsule_server_get(&req->rq_pill, &RMF_HSM_USER_STATE);
1922         if (req_hus == NULL)
1923                 GOTO(out, rc = -EPROTO);
1924
1925         *hus = *req_hus;
1926
1927         EXIT;
1928 out:
1929         ptlrpc_req_finished(req);
1930         return rc;
1931 }
1932
1933 static int mdc_ioc_hsm_state_set(struct obd_export *exp,
1934                                  struct md_op_data *op_data)
1935 {
1936         struct hsm_state_set    *hss = op_data->op_data;
1937         struct hsm_state_set    *req_hss;
1938         struct ptlrpc_request   *req;
1939         int                      rc;
1940         ENTRY;
1941
1942         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1943                                    &RQF_MDS_HSM_STATE_SET);
1944         if (req == NULL)
1945                 RETURN(-ENOMEM);
1946
1947         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_SET);
1948         if (rc) {
1949                 ptlrpc_request_free(req);
1950                 RETURN(rc);
1951         }
1952
1953         mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1954                       op_data->op_suppgids[0], 0);
1955
1956         /* Copy states */
1957         req_hss = req_capsule_client_get(&req->rq_pill, &RMF_HSM_STATE_SET);
1958         if (req_hss == NULL)
1959                 GOTO(out, rc = -EPROTO);
1960         *req_hss = *hss;
1961
1962         ptlrpc_request_set_replen(req);
1963
1964         mdc_get_mod_rpc_slot(req, NULL);
1965         rc = ptlrpc_queue_wait(req);
1966         mdc_put_mod_rpc_slot(req, NULL);
1967
1968         GOTO(out, rc);
1969 out:
1970         ptlrpc_req_finished(req);
1971         return rc;
1972 }
1973
1974 static int mdc_ioc_hsm_request(struct obd_export *exp,
1975                                struct hsm_user_request *hur)
1976 {
1977         struct obd_import       *imp = class_exp2cliimp(exp);
1978         struct ptlrpc_request   *req;
1979         struct hsm_request      *req_hr;
1980         struct hsm_user_item    *req_hui;
1981         char                    *req_opaque;
1982         int                      rc;
1983         ENTRY;
1984
1985         req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_REQUEST);
1986         if (req == NULL)
1987                 GOTO(out, rc = -ENOMEM);
1988
1989         req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM, RCL_CLIENT,
1990                              hur->hur_request.hr_itemcount
1991                              * sizeof(struct hsm_user_item));
1992         req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA, RCL_CLIENT,
1993                              hur->hur_request.hr_data_len);
1994
1995         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_REQUEST);
1996         if (rc) {
1997                 ptlrpc_request_free(req);
1998                 RETURN(rc);
1999         }
2000
2001         mdc_pack_body(req, NULL, 0, 0, -1, 0);
2002
2003         /* Copy hsm_request struct */
2004         req_hr = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_REQUEST);
2005         if (req_hr == NULL)
2006                 GOTO(out, rc = -EPROTO);
2007         *req_hr = hur->hur_request;
2008
2009         /* Copy hsm_user_item structs */
2010         req_hui = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM);
2011         if (req_hui == NULL)
2012                 GOTO(out, rc = -EPROTO);
2013         memcpy(req_hui, hur->hur_user_item,
2014                hur->hur_request.hr_itemcount * sizeof(struct hsm_user_item));
2015
2016         /* Copy opaque field */
2017         req_opaque = req_capsule_client_get(&req->rq_pill, &RMF_GENERIC_DATA);
2018         if (req_opaque == NULL)
2019                 GOTO(out, rc = -EPROTO);
2020         memcpy(req_opaque, hur_data(hur), hur->hur_request.hr_data_len);
2021
2022         ptlrpc_request_set_replen(req);
2023
2024         mdc_get_mod_rpc_slot(req, NULL);
2025         rc = ptlrpc_queue_wait(req);
2026         mdc_put_mod_rpc_slot(req, NULL);
2027
2028         GOTO(out, rc);
2029
2030 out:
2031         ptlrpc_req_finished(req);
2032         return rc;
2033 }
2034
2035 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2036                                 struct lustre_kernelcomm *lk);
2037
2038 static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp,
2039                         struct obd_quotactl *oqctl)
2040 {
2041         struct ptlrpc_request   *req;
2042         struct obd_quotactl     *oqc;
2043         int                      rc;
2044         ENTRY;
2045
2046         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
2047                                         &RQF_MDS_QUOTACTL, LUSTRE_MDS_VERSION,
2048                                         MDS_QUOTACTL);
2049         if (req == NULL)
2050                 RETURN(-ENOMEM);
2051
2052         oqc = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2053         *oqc = *oqctl;
2054
2055         ptlrpc_request_set_replen(req);
2056         ptlrpc_at_set_req_timeout(req);
2057
2058         rc = ptlrpc_queue_wait(req);
2059         if (rc)
2060                 CERROR("ptlrpc_queue_wait failed, rc: %d\n", rc);
2061
2062         if (req->rq_repmsg &&
2063             (oqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL))) {
2064                 *oqctl = *oqc;
2065         } else if (!rc) {
2066                 CERROR ("Can't unpack obd_quotactl\n");
2067                 rc = -EPROTO;
2068         }
2069         ptlrpc_req_finished(req);
2070
2071         RETURN(rc);
2072 }
2073
2074 static int mdc_ioc_swap_layouts(struct obd_export *exp,
2075                                 struct md_op_data *op_data)
2076 {
2077         struct list_head cancels = LIST_HEAD_INIT(cancels);
2078         struct ptlrpc_request   *req;
2079         int                      rc, count;
2080         struct mdc_swap_layouts *msl, *payload;
2081         ENTRY;
2082
2083         msl = op_data->op_data;
2084
2085         /* When the MDT will get the MDS_SWAP_LAYOUTS RPC the
2086          * first thing it will do is to cancel the 2 layout
2087          * locks held by this client.
2088          * So the client must cancel its layout locks on the 2 fids
2089          * with the request RPC to avoid extra RPC round trips.
2090          */
2091         count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels,
2092                                         LCK_EX, MDS_INODELOCK_LAYOUT |
2093                                         MDS_INODELOCK_XATTR);
2094         count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels,
2095                                          LCK_EX, MDS_INODELOCK_LAYOUT |
2096                                          MDS_INODELOCK_XATTR);
2097
2098         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2099                                    &RQF_MDS_SWAP_LAYOUTS);
2100         if (req == NULL) {
2101                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
2102                 RETURN(-ENOMEM);
2103         }
2104
2105         rc = mdc_prep_elc_req(exp, req, MDS_SWAP_LAYOUTS, &cancels, count);
2106         if (rc) {
2107                 ptlrpc_request_free(req);
2108                 RETURN(rc);
2109         }
2110
2111         mdc_swap_layouts_pack(req, op_data);
2112
2113         payload = req_capsule_client_get(&req->rq_pill, &RMF_SWAP_LAYOUTS);
2114         LASSERT(payload);
2115
2116         *payload = *msl;
2117
2118         ptlrpc_request_set_replen(req);
2119
2120         rc = ptlrpc_queue_wait(req);
2121         if (rc)
2122                 GOTO(out, rc);
2123         EXIT;
2124
2125 out:
2126         ptlrpc_req_finished(req);
2127         return rc;
2128 }
2129
2130 static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2131                          void *karg, void __user *uarg)
2132 {
2133         struct obd_device *obd = exp->exp_obd;
2134         struct obd_ioctl_data *data = karg;
2135         struct obd_import *imp = obd->u.cli.cl_import;
2136         int rc;
2137         ENTRY;
2138
2139         if (!try_module_get(THIS_MODULE)) {
2140                 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2141                        module_name(THIS_MODULE));
2142                 return -EINVAL;
2143         }
2144         switch (cmd) {
2145         case OBD_IOC_FID2PATH:
2146                 rc = mdc_ioc_fid2path(exp, karg);
2147                 GOTO(out, rc);
2148         case LL_IOC_HSM_CT_START:
2149                 rc = mdc_ioc_hsm_ct_start(exp, karg);
2150                 /* ignore if it was already registered on this MDS. */
2151                 if (rc == -EEXIST)
2152                         rc = 0;
2153                 GOTO(out, rc);
2154         case LL_IOC_HSM_PROGRESS:
2155                 rc = mdc_ioc_hsm_progress(exp, karg);
2156                 GOTO(out, rc);
2157         case LL_IOC_HSM_STATE_GET:
2158                 rc = mdc_ioc_hsm_state_get(exp, karg);
2159                 GOTO(out, rc);
2160         case LL_IOC_HSM_STATE_SET:
2161                 rc = mdc_ioc_hsm_state_set(exp, karg);
2162                 GOTO(out, rc);
2163         case LL_IOC_HSM_ACTION:
2164                 rc = mdc_ioc_hsm_current_action(exp, karg);
2165                 GOTO(out, rc);
2166         case LL_IOC_HSM_REQUEST:
2167                 rc = mdc_ioc_hsm_request(exp, karg);
2168                 GOTO(out, rc);
2169         case OBD_IOC_CLIENT_RECOVER:
2170                 rc = ptlrpc_recover_import(imp, data->ioc_inlbuf1, 0);
2171                 if (rc < 0)
2172                         GOTO(out, rc);
2173                 GOTO(out, rc = 0);
2174         case IOC_OSC_SET_ACTIVE:
2175                 rc = ptlrpc_set_import_active(imp, data->ioc_offset);
2176                 GOTO(out, rc);
2177         /*
2178          * Normally IOC_OBD_STATFS, OBD_IOC_QUOTACTL iocontrol are handled by
2179          * LMV instead of MDC. But when the cluster is upgraded from 1.8,
2180          * there'd be no LMV layer thus we might be called here. Eventually
2181          * this code should be removed.
2182          * bz20731, LU-592.
2183          */
2184         case IOC_OBD_STATFS: {
2185                 struct obd_statfs stat_buf = {0};
2186
2187                 if (*((__u32 *) data->ioc_inlbuf2) != 0)
2188                         GOTO(out, rc = -ENODEV);
2189
2190                 /* copy UUID */
2191                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(obd),
2192                                  min((int)data->ioc_plen2,
2193                                      (int)sizeof(struct obd_uuid))))
2194                         GOTO(out, rc = -EFAULT);
2195
2196                 rc = mdc_statfs(NULL, obd->obd_self_export, &stat_buf,
2197                                 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
2198                                 0);
2199                 if (rc != 0)
2200                         GOTO(out, rc);
2201
2202                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
2203                                      min((int) data->ioc_plen1,
2204                                          (int) sizeof(stat_buf))))
2205                         GOTO(out, rc = -EFAULT);
2206
2207                 GOTO(out, rc = 0);
2208         }
2209         case OBD_IOC_QUOTACTL: {
2210                 struct if_quotactl *qctl = karg;
2211                 struct obd_quotactl *oqctl;
2212
2213                 OBD_ALLOC_PTR(oqctl);
2214                 if (oqctl == NULL)
2215                         GOTO(out, rc = -ENOMEM);
2216
2217                 QCTL_COPY(oqctl, qctl);
2218                 rc = obd_quotactl(exp, oqctl);
2219                 if (rc == 0) {
2220                         QCTL_COPY(qctl, oqctl);
2221                         qctl->qc_valid = QC_MDTIDX;
2222                         qctl->obd_uuid = obd->u.cli.cl_target_uuid;
2223                 }
2224
2225                 OBD_FREE_PTR(oqctl);
2226                 GOTO(out, rc);
2227         }
2228         case LL_IOC_GET_CONNECT_FLAGS:
2229                 if (copy_to_user(uarg, exp_connect_flags_ptr(exp),
2230                                  sizeof(*exp_connect_flags_ptr(exp))))
2231                         GOTO(out, rc = -EFAULT);
2232
2233                 GOTO(out, rc = 0);
2234         case LL_IOC_LOV_SWAP_LAYOUTS:
2235                 rc = mdc_ioc_swap_layouts(exp, karg);
2236                 GOTO(out, rc);
2237         default:
2238                 CERROR("unrecognised ioctl: cmd = %#x\n", cmd);
2239                 GOTO(out, rc = -ENOTTY);
2240         }
2241 out:
2242         module_put(THIS_MODULE);
2243
2244         return rc;
2245 }
2246
2247 static int mdc_get_info_rpc(struct obd_export *exp,
2248                             u32 keylen, void *key,
2249                             u32 vallen, void *val)
2250 {
2251         struct obd_import      *imp = class_exp2cliimp(exp);
2252         struct ptlrpc_request  *req;
2253         char                   *tmp;
2254         int                     rc = -EINVAL;
2255         ENTRY;
2256
2257         req = ptlrpc_request_alloc(imp, &RQF_MDS_GET_INFO);
2258         if (req == NULL)
2259                 RETURN(-ENOMEM);
2260
2261         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY,
2262                              RCL_CLIENT, keylen);
2263         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VALLEN,
2264                              RCL_CLIENT, sizeof(vallen));
2265
2266         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_INFO);
2267         if (rc) {
2268                 ptlrpc_request_free(req);
2269                 RETURN(rc);
2270         }
2271
2272         tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
2273         memcpy(tmp, key, keylen);
2274         tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_VALLEN);
2275         memcpy(tmp, &vallen, sizeof(vallen));
2276
2277         req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VAL,
2278                              RCL_SERVER, vallen);
2279         ptlrpc_request_set_replen(req);
2280
2281         rc = ptlrpc_queue_wait(req);
2282         /* -EREMOTE means the get_info result is partial, and it needs to
2283          * continue on another MDT, see fid2path part in lmv_iocontrol */
2284         if (rc == 0 || rc == -EREMOTE) {
2285                 tmp = req_capsule_server_get(&req->rq_pill, &RMF_GETINFO_VAL);
2286                 memcpy(val, tmp, vallen);
2287                 if (ptlrpc_rep_need_swab(req)) {
2288                         if (KEY_IS(KEY_FID2PATH))
2289                                 lustre_swab_fid2path(val);
2290                 }
2291         }
2292         ptlrpc_req_finished(req);
2293
2294         RETURN(rc);
2295 }
2296
2297 static void lustre_swab_hai(struct hsm_action_item *h)
2298 {
2299         __swab32s(&h->hai_len);
2300         __swab32s(&h->hai_action);
2301         lustre_swab_lu_fid(&h->hai_fid);
2302         lustre_swab_lu_fid(&h->hai_dfid);
2303         __swab64s(&h->hai_cookie);
2304         __swab64s(&h->hai_extent.offset);
2305         __swab64s(&h->hai_extent.length);
2306         __swab64s(&h->hai_gid);
2307 }
2308
2309 static void lustre_swab_hal(struct hsm_action_list *h)
2310 {
2311         struct hsm_action_item  *hai;
2312         __u32                    i;
2313
2314         __swab32s(&h->hal_version);
2315         __swab32s(&h->hal_count);
2316         __swab32s(&h->hal_archive_id);
2317         __swab64s(&h->hal_flags);
2318         hai = hai_first(h);
2319         for (i = 0; i < h->hal_count; i++, hai = hai_next(hai))
2320                 lustre_swab_hai(hai);
2321 }
2322
2323 static void lustre_swab_kuch(struct kuc_hdr *l)
2324 {
2325         __swab16s(&l->kuc_magic);
2326         /* __u8 l->kuc_transport */
2327         __swab16s(&l->kuc_msgtype);
2328         __swab16s(&l->kuc_msglen);
2329 }
2330
2331 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2332                                 struct lustre_kernelcomm *lk)
2333 {
2334         struct obd_import *imp = class_exp2cliimp(exp);
2335         int rc = 0;
2336
2337         if (lk->lk_group != KUC_GRP_HSM) {
2338                 CERROR("Bad copytool group %d\n", lk->lk_group);
2339                 return -EINVAL;
2340         }
2341
2342         CDEBUG(D_HSM, "CT start r%d w%d u%d g%d f%#x\n", lk->lk_rfd, lk->lk_wfd,
2343                lk->lk_uid, lk->lk_group, lk->lk_flags);
2344
2345         if (lk->lk_flags & LK_FLG_STOP) {
2346                 /* Unregister with the coordinator */
2347                 rc = mdc_ioc_hsm_ct_unregister(imp);
2348         } else {
2349                 __u32 *archives = NULL;
2350
2351                 if ((lk->lk_flags & LK_FLG_DATANR) && lk->lk_data_count > 0)
2352                         archives = lk->lk_data;
2353
2354                 rc = mdc_ioc_hsm_ct_register(imp, lk->lk_data_count, archives);
2355         }
2356
2357         return rc;
2358 }
2359
2360 /**
2361  * Send a message to any listening copytools
2362  * @param val KUC message (kuc_hdr + hsm_action_list)
2363  * @param len total length of message
2364  */
2365 static int mdc_hsm_copytool_send(const struct obd_uuid *uuid,
2366                                  size_t len, void *val)
2367 {
2368         struct kuc_hdr          *lh = (struct kuc_hdr *)val;
2369         struct hsm_action_list  *hal = (struct hsm_action_list *)(lh + 1);
2370         int                      rc;
2371         ENTRY;
2372
2373         if (len < sizeof(*lh) + sizeof(*hal)) {
2374                 CERROR("Short HSM message %zu < %zu\n", len,
2375                        sizeof(*lh) + sizeof(*hal));
2376                 RETURN(-EPROTO);
2377         }
2378         if (lh->kuc_magic == __swab16(KUC_MAGIC)) {
2379                 lustre_swab_kuch(lh);
2380                 lustre_swab_hal(hal);
2381         } else if (lh->kuc_magic != KUC_MAGIC) {
2382                 CERROR("Bad magic %x!=%x\n", lh->kuc_magic, KUC_MAGIC);
2383                 RETURN(-EPROTO);
2384         }
2385
2386         CDEBUG(D_HSM, " Received message mg=%x t=%d m=%d l=%d actions=%d "
2387                "on %s\n",
2388                lh->kuc_magic, lh->kuc_transport, lh->kuc_msgtype,
2389                lh->kuc_msglen, hal->hal_count, hal->hal_fsname);
2390
2391         /* Broadcast to HSM listeners */
2392         rc = libcfs_kkuc_group_put(uuid, KUC_GRP_HSM, lh);
2393
2394         RETURN(rc);
2395 }
2396
2397 /**
2398  * callback function passed to kuc for re-registering each HSM copytool
2399  * running on MDC, after MDT shutdown/recovery.
2400  * @param data copytool registration data
2401  * @param cb_arg callback argument (obd_import)
2402  */
2403 static int mdc_hsm_ct_reregister(void *data, void *cb_arg)
2404 {
2405         struct obd_import *imp = (struct obd_import *)cb_arg;
2406         struct kkuc_ct_data *kcd = data;
2407         __u32 *archives = NULL;
2408         int rc;
2409
2410         if (kcd == NULL ||
2411             (kcd->kcd_magic != KKUC_CT_DATA_ARRAY_MAGIC &&
2412              kcd->kcd_magic != KKUC_CT_DATA_BITMAP_MAGIC))
2413                 return -EPROTO;
2414
2415         if (kcd->kcd_magic == KKUC_CT_DATA_BITMAP_MAGIC) {
2416                 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2417                        "(archive=%#x)\n", imp->imp_obd->obd_name,
2418                        kcd->kcd_nr_archives);
2419         } else {
2420                 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2421                        "(archive nr = %u)\n",
2422                        imp->imp_obd->obd_name, kcd->kcd_nr_archives);
2423                 if (kcd->kcd_nr_archives != 0)
2424                         archives = kcd->kcd_archives;
2425         }
2426
2427         rc = mdc_ioc_hsm_ct_register(imp, kcd->kcd_nr_archives, archives);
2428         /* ignore error if the copytool is already registered */
2429         return (rc == -EEXIST) ? 0 : rc;
2430 }
2431
2432 /**
2433  * Re-establish all kuc contexts with MDT
2434  * after MDT shutdown/recovery.
2435  */
2436 static int mdc_kuc_reregister(struct obd_import *imp)
2437 {
2438         /* re-register HSM agents */
2439         return libcfs_kkuc_group_foreach(&imp->imp_obd->obd_uuid, KUC_GRP_HSM,
2440                                          mdc_hsm_ct_reregister, imp);
2441 }
2442
2443 static int mdc_set_info_async(const struct lu_env *env,
2444                               struct obd_export *exp,
2445                               u32 keylen, void *key,
2446                               u32 vallen, void *val,
2447                               struct ptlrpc_request_set *set)
2448 {
2449         struct obd_import       *imp = class_exp2cliimp(exp);
2450         int                      rc;
2451         ENTRY;
2452
2453         if (KEY_IS(KEY_READ_ONLY)) {
2454                 if (vallen != sizeof(int))
2455                         RETURN(-EINVAL);
2456
2457                 spin_lock(&imp->imp_lock);
2458                 if (*((int *)val)) {
2459                         imp->imp_connect_flags_orig |= OBD_CONNECT_RDONLY;
2460                         imp->imp_connect_data.ocd_connect_flags |=
2461                                                         OBD_CONNECT_RDONLY;
2462                 } else {
2463                         imp->imp_connect_flags_orig &= ~OBD_CONNECT_RDONLY;
2464                         imp->imp_connect_data.ocd_connect_flags &=
2465                                                         ~OBD_CONNECT_RDONLY;
2466                 }
2467                 spin_unlock(&imp->imp_lock);
2468
2469                 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2470                                        keylen, key, vallen, val, set);
2471                 RETURN(rc);
2472         }
2473         if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2474                 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2475                                        keylen, key, vallen, val, set);
2476                 RETURN(rc);
2477         }
2478         if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) {
2479                 rc = mdc_hsm_copytool_send(&imp->imp_obd->obd_uuid, vallen,
2480                                            val);
2481                 RETURN(rc);
2482         }
2483
2484         if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2485                 __u32 *default_easize = val;
2486
2487                 exp->exp_obd->u.cli.cl_default_mds_easize = *default_easize;
2488                 RETURN(0);
2489         }
2490
2491         rc = osc_set_info_async(env, exp, keylen, key, vallen, val, set);
2492         RETURN(rc);
2493 }
2494
2495 static int mdc_get_info(const struct lu_env *env, struct obd_export *exp,
2496                         __u32 keylen, void *key, __u32 *vallen, void *val)
2497 {
2498         int rc = -EINVAL;
2499
2500         if (KEY_IS(KEY_MAX_EASIZE)) {
2501                 __u32 mdsize, *max_easize;
2502
2503                 if (*vallen != sizeof(int))
2504                         RETURN(-EINVAL);
2505                 mdsize = *(__u32 *)val;
2506                 if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
2507                         exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
2508                 max_easize = val;
2509                 *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
2510                 RETURN(0);
2511         } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2512                 __u32 *default_easize;
2513
2514                 if (*vallen != sizeof(int))
2515                         RETURN(-EINVAL);
2516                 default_easize = val;
2517                 *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
2518                 RETURN(0);
2519         } else if (KEY_IS(KEY_CONN_DATA)) {
2520                 struct obd_import *imp = class_exp2cliimp(exp);
2521                 struct obd_connect_data *data = val;
2522
2523                 if (*vallen != sizeof(*data))
2524                         RETURN(-EINVAL);
2525
2526                 *data = imp->imp_connect_data;
2527                 RETURN(0);
2528         } else if (KEY_IS(KEY_TGT_COUNT)) {
2529                 *((__u32 *)val) = 1;
2530                 RETURN(0);
2531         }
2532
2533         rc = mdc_get_info_rpc(exp, keylen, key, *vallen, val);
2534
2535         RETURN(rc);
2536 }
2537
2538 static int mdc_fsync(struct obd_export *exp, const struct lu_fid *fid,
2539                      struct ptlrpc_request **request)
2540 {
2541         struct ptlrpc_request *req;
2542         int                    rc;
2543         ENTRY;
2544
2545         *request = NULL;
2546         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SYNC);
2547         if (req == NULL)
2548                 RETURN(-ENOMEM);
2549
2550         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SYNC);
2551         if (rc) {
2552                 ptlrpc_request_free(req);
2553                 RETURN(rc);
2554         }
2555
2556         mdc_pack_body(req, fid, 0, 0, -1, 0);
2557
2558         ptlrpc_request_set_replen(req);
2559
2560         rc = ptlrpc_queue_wait(req);
2561         if (rc)
2562                 ptlrpc_req_finished(req);
2563         else
2564                 *request = req;
2565         RETURN(rc);
2566 }
2567
2568 struct mdc_rmfid_args {
2569         int *mra_rcs;
2570         int mra_nr;
2571 };
2572
2573 int mdc_rmfid_interpret(const struct lu_env *env, struct ptlrpc_request *req,
2574                           void *args, int rc)
2575 {
2576         struct mdc_rmfid_args *aa;
2577         int *rcs, size;
2578         ENTRY;
2579
2580         if (!rc) {
2581                 aa = ptlrpc_req_async_args(aa, req);
2582
2583                 size = req_capsule_get_size(&req->rq_pill, &RMF_RCS,
2584                                             RCL_SERVER);
2585                 LASSERT(size == sizeof(int) * aa->mra_nr);
2586                 rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2587                 LASSERT(rcs);
2588                 LASSERT(aa->mra_rcs);
2589                 LASSERT(aa->mra_nr);
2590                 memcpy(aa->mra_rcs, rcs, size);
2591         }
2592
2593         RETURN(rc);
2594 }
2595
2596 static int mdc_rmfid(struct obd_export *exp, struct fid_array *fa,
2597                      int *rcs, struct ptlrpc_request_set *set)
2598 {
2599         struct ptlrpc_request *req;
2600         struct mdc_rmfid_args *aa;
2601         struct mdt_body *b;
2602         struct lu_fid *tmp;
2603         int rc, flen;
2604         ENTRY;
2605
2606         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_RMFID);
2607         if (req == NULL)
2608                 RETURN(-ENOMEM);
2609
2610         flen = fa->fa_nr * sizeof(struct lu_fid);
2611         req_capsule_set_size(&req->rq_pill, &RMF_FID_ARRAY,
2612                              RCL_CLIENT, flen);
2613         req_capsule_set_size(&req->rq_pill, &RMF_FID_ARRAY,
2614                              RCL_SERVER, flen);
2615         req_capsule_set_size(&req->rq_pill, &RMF_RCS,
2616                              RCL_SERVER, fa->fa_nr * sizeof(__u32));
2617         rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_RMFID);
2618         if (rc) {
2619                 ptlrpc_request_free(req);
2620                 RETURN(rc);
2621         }
2622         tmp = req_capsule_client_get(&req->rq_pill, &RMF_FID_ARRAY);
2623         memcpy(tmp, fa->fa_fids, flen);
2624
2625         mdc_pack_body(req, NULL, 0, 0, -1, 0);
2626         b = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
2627         b->mbo_ctime = ktime_get_real_seconds();
2628
2629         ptlrpc_request_set_replen(req);
2630
2631         LASSERT(rcs);
2632         aa = ptlrpc_req_async_args(aa, req);
2633         aa->mra_rcs = rcs;
2634         aa->mra_nr = fa->fa_nr;
2635         req->rq_interpret_reply = mdc_rmfid_interpret;
2636
2637         ptlrpc_set_add_req(set, req);
2638         ptlrpc_check_set(NULL, set);
2639
2640         RETURN(rc);
2641 }
2642
2643 static int mdc_import_event(struct obd_device *obd, struct obd_import *imp,
2644                             enum obd_import_event event)
2645 {
2646         struct client_obd *cli = &obd->u.cli;
2647         int rc = 0;
2648
2649         LASSERT(imp->imp_obd == obd);
2650
2651         switch (event) {
2652         case IMP_EVENT_DISCON:
2653                 spin_lock(&cli->cl_loi_list_lock);
2654                 cli->cl_avail_grant = 0;
2655                 cli->cl_lost_grant = 0;
2656                 spin_unlock(&cli->cl_loi_list_lock);
2657                 break;
2658         case IMP_EVENT_INACTIVE:
2659                 /*
2660                  * Flush current sequence to make client obtain new one
2661                  * from server in case of disconnect/reconnect.
2662                  */
2663                 down_read(&cli->cl_seq_rwsem);
2664                 if (cli->cl_seq)
2665                         seq_client_flush(cli->cl_seq);
2666                 up_read(&cli->cl_seq_rwsem);
2667
2668                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
2669                 break;
2670         case IMP_EVENT_INVALIDATE: {
2671                 struct ldlm_namespace *ns = obd->obd_namespace;
2672                 struct lu_env *env;
2673                 __u16 refcheck;
2674
2675                 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2676
2677                 env = cl_env_get(&refcheck);
2678                 if (!IS_ERR(env)) {
2679                         /* Reset grants. All pages go to failing rpcs due to
2680                          * the invalid import.
2681                          */
2682                         osc_io_unplug(env, cli, NULL);
2683
2684                         cfs_hash_for_each_nolock(ns->ns_rs_hash,
2685                                                  osc_ldlm_resource_invalidate,
2686                                                  env, 0);
2687                         cl_env_put(env, &refcheck);
2688                         ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2689                 } else {
2690                         rc = PTR_ERR(env);
2691                 }
2692                 break;
2693         }
2694         case IMP_EVENT_ACTIVE:
2695                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
2696                 /* redo the kuc registration after reconnecting */
2697                 if (rc == 0)
2698                         rc = mdc_kuc_reregister(imp);
2699                 break;
2700         case IMP_EVENT_OCD: {
2701                 struct obd_connect_data *ocd = &imp->imp_connect_data;
2702
2703                 if (OCD_HAS_FLAG(ocd, GRANT))
2704                         osc_init_grant(cli, ocd);
2705
2706                 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
2707                 break;
2708         }
2709         case IMP_EVENT_DEACTIVATE:
2710         case IMP_EVENT_ACTIVATE:
2711                 break;
2712         default:
2713                 CERROR("Unknown import event %x\n", event);
2714                 LBUG();
2715         }
2716         RETURN(rc);
2717 }
2718
2719 int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp,
2720                   struct lu_fid *fid, struct md_op_data *op_data)
2721 {
2722         struct client_obd *cli = &exp->exp_obd->u.cli;
2723         int rc = -EIO;
2724
2725         ENTRY;
2726
2727         down_read(&cli->cl_seq_rwsem);
2728         if (cli->cl_seq)
2729                 rc = seq_client_alloc_fid(env, cli->cl_seq, fid);
2730         up_read(&cli->cl_seq_rwsem);
2731
2732         RETURN(rc);
2733 }
2734
2735 static struct obd_uuid *mdc_get_uuid(struct obd_export *exp)
2736 {
2737         struct client_obd *cli = &exp->exp_obd->u.cli;
2738         return &cli->cl_target_uuid;
2739 }
2740
2741 /**
2742  * Determine whether the lock can be canceled before replaying it during
2743  * recovery, non zero value will be return if the lock can be canceled,
2744  * or zero returned for not
2745  */
2746 static int mdc_cancel_weight(struct ldlm_lock *lock)
2747 {
2748         if (lock->l_resource->lr_type != LDLM_IBITS)
2749                 RETURN(0);
2750
2751         /* FIXME: if we ever get into a situation where there are too many
2752          * opened files with open locks on a single node, then we really
2753          * should replay these open locks to reget it */
2754         if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN)
2755                 RETURN(0);
2756
2757         /* Special case for DoM locks, cancel only unused and granted locks */
2758         if (ldlm_has_dom(lock) &&
2759             (lock->l_granted_mode != lock->l_req_mode ||
2760              osc_ldlm_weigh_ast(lock) != 0))
2761                 RETURN(0);
2762
2763         RETURN(1);
2764 }
2765
2766 static int mdc_resource_inode_free(struct ldlm_resource *res)
2767 {
2768         if (res->lr_lvb_inode)
2769                 res->lr_lvb_inode = NULL;
2770
2771         return 0;
2772 }
2773
2774 static struct ldlm_valblock_ops inode_lvbo = {
2775         .lvbo_free = mdc_resource_inode_free
2776 };
2777
2778 static int mdc_llog_init(struct obd_device *obd)
2779 {
2780         struct obd_llog_group   *olg = &obd->obd_olg;
2781         struct llog_ctxt        *ctxt;
2782         int                      rc;
2783
2784         ENTRY;
2785
2786         rc = llog_setup(NULL, obd, olg, LLOG_CHANGELOG_REPL_CTXT, obd,
2787                         &llog_client_ops);
2788         if (rc < 0)
2789                 RETURN(rc);
2790
2791         ctxt = llog_group_get_ctxt(olg, LLOG_CHANGELOG_REPL_CTXT);
2792         llog_initiator_connect(ctxt);
2793         llog_ctxt_put(ctxt);
2794
2795         RETURN(0);
2796 }
2797
2798 static void mdc_llog_finish(struct obd_device *obd)
2799 {
2800         struct llog_ctxt *ctxt;
2801
2802         ENTRY;
2803
2804         ctxt = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
2805         if (ctxt != NULL)
2806                 llog_cleanup(NULL, ctxt);
2807
2808         EXIT;
2809 }
2810
2811 int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg)
2812 {
2813         int rc;
2814
2815         ENTRY;
2816
2817         rc = osc_setup_common(obd, cfg);
2818         if (rc < 0)
2819                 RETURN(rc);
2820
2821         rc = mdc_tunables_init(obd);
2822         if (rc)
2823                 GOTO(err_osc_cleanup, rc);
2824
2825         obd->u.cli.cl_dom_min_inline_repsize = MDC_DOM_DEF_INLINE_REPSIZE;
2826
2827         ns_register_cancel(obd->obd_namespace, mdc_cancel_weight);
2828
2829         obd->obd_namespace->ns_lvbo = &inode_lvbo;
2830
2831         rc = mdc_llog_init(obd);
2832         if (rc) {
2833                 CERROR("%s: failed to setup llogging subsystems: rc = %d\n",
2834                        obd->obd_name, rc);
2835                 GOTO(err_llog_cleanup, rc);
2836         }
2837
2838         rc = mdc_changelog_cdev_init(obd);
2839         if (rc) {
2840                 CERROR("%s: failed to setup changelog char device: rc = %d\n",
2841                        obd->obd_name, rc);
2842                 GOTO(err_changelog_cleanup, rc);
2843         }
2844
2845         RETURN(rc);
2846
2847 err_changelog_cleanup:
2848         mdc_llog_finish(obd);
2849 err_llog_cleanup:
2850         lprocfs_free_md_stats(obd);
2851         ptlrpc_lprocfs_unregister_obd(obd);
2852 err_osc_cleanup:
2853         osc_cleanup_common(obd);
2854         return rc;
2855 }
2856
2857 /* Initialize the default and maximum LOV EA sizes.  This allows
2858  * us to make MDS RPCs with large enough reply buffers to hold a default
2859  * sized EA without having to calculate this (via a call into the
2860  * LOV + OSCs) each time we make an RPC.  The maximum size is also tracked
2861  * but not used to avoid wastefully vmalloc()'ing large reply buffers when
2862  * a large number of stripes is possible.  If a larger reply buffer is
2863  * required it will be reallocated in the ptlrpc layer due to overflow.
2864  */
2865 static int mdc_init_ea_size(struct obd_export *exp, __u32 easize,
2866                             __u32 def_easize)
2867 {
2868         struct obd_device *obd = exp->exp_obd;
2869         struct client_obd *cli = &obd->u.cli;
2870         ENTRY;
2871
2872         if (cli->cl_max_mds_easize < easize)
2873                 cli->cl_max_mds_easize = easize;
2874
2875         if (cli->cl_default_mds_easize < def_easize)
2876                 cli->cl_default_mds_easize = def_easize;
2877
2878         RETURN(0);
2879 }
2880
2881 static int mdc_precleanup(struct obd_device *obd)
2882 {
2883         ENTRY;
2884
2885         osc_precleanup_common(obd);
2886         mdc_changelog_cdev_finish(obd);
2887
2888         obd_cleanup_client_import(obd);
2889         ptlrpc_lprocfs_unregister_obd(obd);
2890         lprocfs_free_md_stats(obd);
2891         mdc_llog_finish(obd);
2892         RETURN(0);
2893 }
2894
2895 static int mdc_cleanup(struct obd_device *obd)
2896 {
2897         return osc_cleanup_common(obd);
2898 }
2899
2900 static struct obd_ops mdc_obd_ops = {
2901         .o_owner            = THIS_MODULE,
2902         .o_setup            = mdc_setup,
2903         .o_precleanup       = mdc_precleanup,
2904         .o_cleanup          = mdc_cleanup,
2905         .o_add_conn         = client_import_add_conn,
2906         .o_del_conn         = client_import_del_conn,
2907         .o_connect          = client_connect_import,
2908         .o_reconnect        = osc_reconnect,
2909         .o_disconnect       = osc_disconnect,
2910         .o_iocontrol        = mdc_iocontrol,
2911         .o_set_info_async   = mdc_set_info_async,
2912         .o_statfs           = mdc_statfs,
2913         .o_statfs_async     = mdc_statfs_async,
2914         .o_fid_init         = client_fid_init,
2915         .o_fid_fini         = client_fid_fini,
2916         .o_fid_alloc        = mdc_fid_alloc,
2917         .o_import_event     = mdc_import_event,
2918         .o_get_info         = mdc_get_info,
2919         .o_get_uuid         = mdc_get_uuid,
2920         .o_quotactl         = mdc_quotactl,
2921 };
2922
2923 static struct md_ops mdc_md_ops = {
2924         .m_get_root         = mdc_get_root,
2925         .m_null_inode       = mdc_null_inode,
2926         .m_close            = mdc_close,
2927         .m_create           = mdc_create,
2928         .m_enqueue          = mdc_enqueue,
2929         .m_getattr          = mdc_getattr,
2930         .m_getattr_name     = mdc_getattr_name,
2931         .m_intent_lock      = mdc_intent_lock,
2932         .m_link             = mdc_link,
2933         .m_rename           = mdc_rename,
2934         .m_setattr          = mdc_setattr,
2935         .m_setxattr         = mdc_setxattr,
2936         .m_getxattr         = mdc_getxattr,
2937         .m_fsync                = mdc_fsync,
2938         .m_file_resync          = mdc_file_resync,
2939         .m_read_page            = mdc_read_page,
2940         .m_unlink           = mdc_unlink,
2941         .m_cancel_unused    = mdc_cancel_unused,
2942         .m_init_ea_size     = mdc_init_ea_size,
2943         .m_set_lock_data    = mdc_set_lock_data,
2944         .m_lock_match       = mdc_lock_match,
2945         .m_get_lustre_md    = mdc_get_lustre_md,
2946         .m_free_lustre_md   = mdc_free_lustre_md,
2947         .m_set_open_replay_data = mdc_set_open_replay_data,
2948         .m_clear_open_replay_data = mdc_clear_open_replay_data,
2949         .m_intent_getattr_async = mdc_intent_getattr_async,
2950         .m_revalidate_lock      = mdc_revalidate_lock,
2951         .m_rmfid                = mdc_rmfid,
2952 };
2953
2954 static int __init mdc_init(void)
2955 {
2956         return class_register_type(&mdc_obd_ops, &mdc_md_ops, true, NULL,
2957                                    LUSTRE_MDC_NAME, &mdc_device_type);
2958 }
2959
2960 static void __exit mdc_exit(void)
2961 {
2962         class_unregister_type(LUSTRE_MDC_NAME);
2963 }
2964
2965 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2966 MODULE_DESCRIPTION("Lustre Metadata Client");
2967 MODULE_VERSION(LUSTRE_VERSION_STRING);
2968 MODULE_LICENSE("GPL");
2969
2970 module_init(mdc_init);
2971 module_exit(mdc_exit);