lustre/mdc/mdc_request.c

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