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