lustre/mdc/mdc_locks.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) 2003, 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  */
  31
  32 #define DEBUG_SUBSYSTEM S_MDC
  33
  34 #include <linux/module.h>
  35
  36 #include <obd.h>
  37 #include <obd_class.h>
  38 #include <lustre_dlm.h>
  39 #include <lustre_fid.h>
  40 #include <lustre_intent.h>
  41 #include <lustre_mdc.h>
  42 #include <lustre_net.h>
  43 #include <lustre_req_layout.h>
  44 #include <lustre_swab.h>
  45 #include <lustre_acl.h>
  46
  47 #include "mdc_internal.h"
  48
  49 struct mdc_getattr_args {
  50         struct obd_export       *ga_exp;
  51         struct md_op_item       *ga_item;
  52 };
  53
  54 int it_open_error(int phase, struct lookup_intent *it)
  55 {
  56         if (it_disposition(it, DISP_OPEN_LEASE)) {
  57                 if (phase >= DISP_OPEN_LEASE)
  58                         return it->it_status;
  59                 else
  60                         return 0;
  61         }
  62         if (it_disposition(it, DISP_OPEN_OPEN)) {
  63                 if (phase >= DISP_OPEN_OPEN)
  64                         return it->it_status;
  65                 else
  66                         return 0;
  67         }
  68
  69         if (it_disposition(it, DISP_OPEN_CREATE)) {
  70                 if (phase >= DISP_OPEN_CREATE)
  71                         return it->it_status;
  72                 else
  73                         return 0;
  74         }
  75
  76         if (it_disposition(it, DISP_LOOKUP_EXECD)) {
  77                 if (phase >= DISP_LOOKUP_EXECD)
  78                         return it->it_status;
  79                 else
  80                         return 0;
  81         }
  82
  83         if (it_disposition(it, DISP_IT_EXECD)) {
  84                 if (phase >= DISP_IT_EXECD)
  85                         return it->it_status;
  86                 else
  87                         return 0;
  88         }
  89
  90         CERROR("it disp: %X, status: %d\n", it->it_disposition, it->it_status);
  91         LBUG();
  92
  93         return 0;
  94 }
  95 EXPORT_SYMBOL(it_open_error);
  96
  97 /* this must be called on a lockh that is known to have a referenced lock */
  98 int mdc_set_lock_data(struct obd_export *exp, const struct lustre_handle *lockh,
  99                       void *data, __u64 *bits)
 100 {
 101         struct ldlm_lock *lock;
 102         struct inode *new_inode = data;
 103
 104         ENTRY;
 105         if (bits)
 106                 *bits = 0;
 107
 108         if (!lustre_handle_is_used(lockh))
 109                 RETURN(0);
 110
 111         lock = ldlm_handle2lock(lockh);
 112
 113         LASSERT(lock != NULL);
 114         lock_res_and_lock(lock);
 115         if (lock->l_resource->lr_lvb_inode &&
 116             lock->l_resource->lr_lvb_inode != data) {
 117                 struct inode *old_inode = lock->l_resource->lr_lvb_inode;
 118
 119                 LASSERTF(old_inode->i_state & I_FREEING,
 120                          "Found existing inode %px/%lu/%u state %lu in lock: setting data to %px/%lu/%u\n",
 121                          old_inode, old_inode->i_ino, old_inode->i_generation,
 122                          old_inode->i_state,
 123                          new_inode, new_inode->i_ino, new_inode->i_generation);
 124         }
 125         lock->l_resource->lr_lvb_inode = new_inode;
 126         if (bits)
 127                 *bits = lock->l_policy_data.l_inodebits.bits;
 128
 129         unlock_res_and_lock(lock);
 130         LDLM_LOCK_PUT(lock);
 131
 132         RETURN(0);
 133 }
 134
 135 enum ldlm_mode mdc_lock_match(struct obd_export *exp, __u64 flags,
 136                               const struct lu_fid *fid, enum ldlm_type type,
 137                               union ldlm_policy_data *policy,
 138                               enum ldlm_mode mode, struct lustre_handle *lockh)
 139 {
 140         struct ldlm_res_id res_id;
 141         enum ldlm_mode rc;
 142
 143         ENTRY;
 144         fid_build_reg_res_name(fid, &res_id);
 145         /* LU-4405: Clear bits not supported by server */
 146         policy->l_inodebits.bits &= exp_connect_ibits(exp);
 147         rc = ldlm_lock_match(class_exp2obd(exp)->obd_namespace, flags,
 148                              &res_id, type, policy, mode, lockh);
 149         RETURN(rc);
 150 }
 151
 152 int mdc_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
 153                       union ldlm_policy_data *policy, enum ldlm_mode mode,
 154                       enum ldlm_cancel_flags flags, void *opaque)
 155 {
 156         struct obd_device *obd = class_exp2obd(exp);
 157         struct ldlm_res_id res_id;
 158         int rc;
 159
 160         ENTRY;
 161         fid_build_reg_res_name(fid, &res_id);
 162         rc = ldlm_cli_cancel_unused_resource(obd->obd_namespace, &res_id,
 163                                              policy, mode, flags, opaque);
 164         RETURN(rc);
 165 }
 166
 167 int mdc_null_inode(struct obd_export *exp,
 168                    const struct lu_fid *fid)
 169 {
 170         struct ldlm_res_id res_id;
 171         struct ldlm_resource *res;
 172         struct ldlm_namespace *ns = class_exp2obd(exp)->obd_namespace;
 173
 174         ENTRY;
 175         LASSERTF(ns != NULL, "no namespace passed\n");
 176
 177         fid_build_reg_res_name(fid, &res_id);
 178
 179         res = ldlm_resource_get(ns, &res_id, 0, 0);
 180         if (IS_ERR(res))
 181                 RETURN(0);
 182
 183         lock_res(res);
 184         res->lr_lvb_inode = NULL;
 185         unlock_res(res);
 186
 187         ldlm_resource_putref(res);
 188         RETURN(0);
 189 }
 190
 191 static inline void mdc_clear_replay_flag(struct ptlrpc_request *req, int rc)
 192 {
 193         /* Don't hold error requests for replay. */
 194         if (req->rq_replay) {
 195                 spin_lock(&req->rq_lock);
 196                 req->rq_replay = 0;
 197                 spin_unlock(&req->rq_lock);
 198         }
 199         if (rc && req->rq_transno != 0) {
 200                 DEBUG_REQ(D_ERROR, req, "transno returned on error: rc = %d",
 201                           rc);
 202                 LBUG();
 203         }
 204 }
 205
 206 /**
 207  * Save a large LOV/LMV EA into the request buffer so that it is available
 208  * for replay.  We don't do this in the initial request because the
 209  * original request doesn't need this buffer (at most it sends just the
 210  * lov_mds_md) and it is a waste of RAM/bandwidth to send the empty
 211  * buffer and may also be difficult to allocate and save a very large
 212  * request buffer for each open. (b=5707)
 213  *
 214  * OOM here may cause recovery failure if lmm is needed (only for the
 215  * original open if the MDS crashed just when this client also OOM'd)
 216  * but this is incredibly unlikely, and questionable whether the client
 217  * could do MDS recovery under OOM anyways...
 218  */
 219 int mdc_save_lmm(struct ptlrpc_request *req, void *data, u32 size)
 220 {
 221         struct req_capsule *pill = &req->rq_pill;
 222         void *lmm;
 223         int rc = 0;
 224
 225         if (req_capsule_get_size(pill, &RMF_EADATA, RCL_CLIENT) < size) {
 226                 rc = sptlrpc_cli_enlarge_reqbuf(req, &RMF_EADATA, size);
 227                 if (rc) {
 228                         CERROR("%s: Can't enlarge ea size to %d: rc = %d\n",
 229                                req->rq_export->exp_obd->obd_name,
 230                                size, rc);
 231                         return rc;
 232                 }
 233         } else {
 234                 req_capsule_shrink(pill, &RMF_EADATA, size, RCL_CLIENT);
 235         }
 236
 237         req_capsule_set_size(pill, &RMF_EADATA, RCL_CLIENT, size);
 238         lmm = req_capsule_client_get(pill, &RMF_EADATA);
 239         if (lmm) {
 240                 memcpy(lmm, data, size);
 241                 lov_fix_ea_for_replay(lmm);
 242         }
 243
 244         return rc;
 245 }
 246
 247 static struct ptlrpc_request *
 248 mdc_intent_open_pack(struct obd_export *exp, struct lookup_intent *it,
 249                      struct md_op_data *op_data, __u32 acl_bufsize)
 250 {
 251         struct ptlrpc_request *req;
 252         struct obd_device *obd = class_exp2obd(exp);
 253         struct ldlm_intent *lit;
 254         const void *lmm = op_data->op_data;
 255         __u32 lmmsize = op_data->op_data_size;
 256         __u32  mdt_md_capsule_size;
 257         LIST_HEAD(cancels);
 258         int count = 0;
 259         enum ldlm_mode mode;
 260         int repsize, repsize_estimate;
 261         struct sptlrpc_sepol *sepol;
 262         int rc;
 263
 264         ENTRY;
 265
 266         mdt_md_capsule_size = obd->u.cli.cl_default_mds_easize;
 267
 268         it->it_create_mode = (it->it_create_mode & ~S_IFMT) | S_IFREG;
 269
 270         /* XXX: openlock is not cancelled for cross-refs. */
 271         /* If inode is known, cancel conflicting OPEN locks. */
 272         if (fid_is_sane(&op_data->op_fid2)) {
 273                 if (it->it_flags & MDS_OPEN_LEASE) { /* try to get lease */
 274                         if (it->it_flags & MDS_FMODE_WRITE)
 275                                 mode = LCK_EX;
 276                         else
 277                                 mode = LCK_PR;
 278                 } else {
 279                         if (it->it_flags & (MDS_FMODE_WRITE | MDS_OPEN_TRUNC))
 280                                 mode = LCK_CW;
 281 #ifdef FMODE_EXEC
 282                         else if (it->it_flags & FMODE_EXEC)
 283                                 mode = LCK_PR;
 284 #endif
 285                         else
 286                                 mode = LCK_CR;
 287                 }
 288                 count = mdc_resource_get_unused(exp, &op_data->op_fid2,
 289                                                 &cancels, mode,
 290                                                 MDS_INODELOCK_OPEN);
 291         }
 292
 293         /* If CREATE, cancel parent's UPDATE lock. */
 294         if (it->it_op & IT_CREAT)
 295                 mode = LCK_EX;
 296         else
 297                 mode = LCK_CR;
 298         count += mdc_resource_get_unused(exp, &op_data->op_fid1,
 299                                          &cancels, mode,
 300                                          MDS_INODELOCK_UPDATE);
 301
 302         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 303                                    &RQF_LDLM_INTENT_OPEN);
 304         if (req == NULL) {
 305                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
 306                 RETURN(ERR_PTR(-ENOMEM));
 307         }
 308
 309         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 310                              op_data->op_namelen + 1);
 311         if (cl_is_lov_delay_create(it->it_flags)) {
 312                 /* open(O_LOV_DELAY_CREATE) won't pack lmm */
 313                 LASSERT(lmmsize == 0);
 314                 req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT, 0);
 315         } else {
 316                 req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT,
 317                              max(lmmsize, obd->u.cli.cl_default_mds_easize));
 318         }
 319
 320         req_capsule_set_size(&req->rq_pill, &RMF_FILE_SECCTX_NAME,
 321                              RCL_CLIENT, op_data->op_file_secctx_name != NULL ?
 322                              op_data->op_file_secctx_name_size : 0);
 323
 324         req_capsule_set_size(&req->rq_pill, &RMF_FILE_SECCTX, RCL_CLIENT,
 325                              op_data->op_file_secctx_size);
 326
 327         req_capsule_set_size(&req->rq_pill, &RMF_FILE_ENCCTX, RCL_CLIENT,
 328                              op_data->op_file_encctx_size);
 329
 330         /* get SELinux policy info if any */
 331         sepol = sptlrpc_sepol_get(req);
 332         if (IS_ERR(sepol))
 333                 GOTO(err_free_rq, rc = PTR_ERR(sepol));
 334
 335         req_capsule_set_size(&req->rq_pill, &RMF_SELINUX_POL, RCL_CLIENT,
 336                              sptlrpc_sepol_size(sepol));
 337
 338         rc = ldlm_prep_enqueue_req(exp, req, &cancels, count);
 339         if (rc < 0)
 340                 GOTO(err_put_sepol, rc);
 341
 342         spin_lock(&req->rq_lock);
 343         req->rq_replay = req->rq_import->imp_replayable;
 344         spin_unlock(&req->rq_lock);
 345
 346         /* pack the intent */
 347         lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
 348         lit->opc = (__u64)it->it_op;
 349
 350         /* pack the intended request */
 351         mdc_open_pack(&req->rq_pill, op_data, it->it_create_mode, 0,
 352                       it->it_flags, lmm, lmmsize, sepol);
 353
 354         sptlrpc_sepol_put(sepol);
 355
 356         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 357                              mdt_md_capsule_size);
 358         req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
 359
 360         if (!(it->it_op & IT_CREAT) && it->it_op & IT_OPEN &&
 361             req_capsule_has_field(&req->rq_pill, &RMF_FILE_SECCTX_NAME,
 362                                   RCL_CLIENT) &&
 363             op_data->op_file_secctx_name_size > 0 &&
 364             op_data->op_file_secctx_name != NULL) {
 365                 char *secctx_name;
 366
 367                 secctx_name = req_capsule_client_get(&req->rq_pill,
 368                                                      &RMF_FILE_SECCTX_NAME);
 369                 memcpy(secctx_name, op_data->op_file_secctx_name,
 370                        op_data->op_file_secctx_name_size);
 371                 req_capsule_set_size(&req->rq_pill, &RMF_FILE_SECCTX,
 372                                      RCL_SERVER,
 373                                      obd->u.cli.cl_max_mds_easize);
 374
 375                 CDEBUG(D_SEC, "packed '%.*s' as security xattr name\n",
 376                        op_data->op_file_secctx_name_size,
 377                        op_data->op_file_secctx_name);
 378
 379         } else {
 380                 req_capsule_set_size(&req->rq_pill, &RMF_FILE_SECCTX,
 381                                      RCL_SERVER, 0);
 382         }
 383
 384         if (exp_connect_encrypt(exp) && !(it->it_op & IT_CREAT) &&
 385             it->it_op & IT_OPEN)
 386                 req_capsule_set_size(&req->rq_pill, &RMF_FILE_ENCCTX,
 387                                      RCL_SERVER,
 388                                      obd->u.cli.cl_max_mds_easize);
 389         else
 390                 req_capsule_set_size(&req->rq_pill, &RMF_FILE_ENCCTX,
 391                                      RCL_SERVER, 0);
 392
 393         /**
 394          * Inline buffer for possible data from Data-on-MDT files.
 395          */
 396         req_capsule_set_size(&req->rq_pill, &RMF_NIOBUF_INLINE, RCL_SERVER,
 397                              sizeof(struct niobuf_remote));
 398         req_capsule_set_size(&req->rq_pill, &RMF_DEFAULT_MDT_MD, RCL_SERVER,
 399                              sizeof(struct lmv_user_md));
 400         ptlrpc_request_set_replen(req);
 401
 402         /* Get real repbuf allocated size as rounded up power of 2 */
 403         repsize = size_roundup_power2(req->rq_replen +
 404                                       lustre_msg_early_size);
 405         /* Estimate free space for DoM files in repbuf */
 406         repsize_estimate = repsize - (req->rq_replen -
 407                            mdt_md_capsule_size +
 408                            sizeof(struct lov_comp_md_v1) +
 409                            sizeof(struct lov_comp_md_entry_v1) +
 410                            lov_mds_md_size(0, LOV_MAGIC_V3));
 411
 412         if (repsize_estimate < obd->u.cli.cl_dom_min_inline_repsize) {
 413                 repsize = obd->u.cli.cl_dom_min_inline_repsize -
 414                           repsize_estimate + sizeof(struct niobuf_remote);
 415                 req_capsule_set_size(&req->rq_pill, &RMF_NIOBUF_INLINE,
 416                                      RCL_SERVER,
 417                                      sizeof(struct niobuf_remote) + repsize);
 418                 ptlrpc_request_set_replen(req);
 419                 CDEBUG(D_INFO, "Increase repbuf by %d bytes, total: %d\n",
 420                        repsize, req->rq_replen);
 421                 repsize = size_roundup_power2(req->rq_replen +
 422                                               lustre_msg_early_size);
 423         }
 424         /* The only way to report real allocated repbuf size to the server
 425          * is the lm_repsize but it must be set prior buffer allocation itself
 426          * due to security reasons - it is part of buffer used in signature
 427          * calculation (see LU-11414). Therefore the saved size is predicted
 428          * value as rq_replen rounded to the next higher power of 2.
 429          * Such estimation is safe. Though the final allocated buffer might
 430          * be even larger, it is not possible to know that at this point.
 431          */
 432         req->rq_reqmsg->lm_repsize = repsize;
 433         RETURN(req);
 434
 435 err_put_sepol:
 436         sptlrpc_sepol_put(sepol);
 437 err_free_rq:
 438         ptlrpc_request_free(req);
 439         return ERR_PTR(rc);
 440 }
 441
 442 #define GA_DEFAULT_EA_NAME_LEN   20
 443 #define GA_DEFAULT_EA_VAL_LEN   250
 444 #define GA_DEFAULT_EA_NUM        10
 445
 446 static struct ptlrpc_request *
 447 mdc_intent_getxattr_pack(struct obd_export *exp, struct lookup_intent *it,
 448                          struct md_op_data *op_data)
 449 {
 450         struct ptlrpc_request *req;
 451         struct ldlm_intent *lit;
 452         struct sptlrpc_sepol *sepol;
 453         int rc, count = 0;
 454         LIST_HEAD(cancels);
 455         u32 ea_vals_buf_size = GA_DEFAULT_EA_VAL_LEN * GA_DEFAULT_EA_NUM;
 456
 457         ENTRY;
 458         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 459                                         &RQF_LDLM_INTENT_GETXATTR);
 460         if (req == NULL)
 461                 RETURN(ERR_PTR(-ENOMEM));
 462
 463         /* get SELinux policy info if any */
 464         sepol = sptlrpc_sepol_get(req);
 465         if (IS_ERR(sepol))
 466                 GOTO(err_free_rq, rc = PTR_ERR(sepol));
 467
 468         req_capsule_set_size(&req->rq_pill, &RMF_SELINUX_POL, RCL_CLIENT,
 469                              sptlrpc_sepol_size(sepol));
 470
 471         rc = ldlm_prep_enqueue_req(exp, req, &cancels, count);
 472         if (rc)
 473                 GOTO(err_put_sepol, rc);
 474
 475         /* pack the intent */
 476         lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
 477         lit->opc = IT_GETXATTR;
 478         /* Message below is checked in sanity-selinux test_20d
 479          * and sanity-sec test_49
 480          */
 481         CDEBUG(D_INFO, "%s: get xattrs for "DFID"\n",
 482                exp->exp_obd->obd_name, PFID(&op_data->op_fid1));
 483
 484 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
 485         /* If the supplied buffer is too small then the server will return
 486          * -ERANGE and llite will fallback to using non cached xattr
 487          * operations. On servers before 2.10.1 a (non-cached) listxattr RPC
 488          * for an orphan or dead file causes an oops. So let's try to avoid
 489          * sending too small a buffer to too old a server. This is effectively
 490          * undoing the memory conservation of LU-9417 when it would be *more*
 491          * likely to crash the server. See LU-9856.
 492          */
 493         if (exp->exp_connect_data.ocd_version < OBD_OCD_VERSION(2, 10, 1, 0))
 494                 ea_vals_buf_size = max_t(u32, ea_vals_buf_size,
 495                                          exp->exp_connect_data.ocd_max_easize);
 496 #endif
 497
 498         /* pack the intended request */
 499         mdc_pack_body(&req->rq_pill, &op_data->op_fid1, op_data->op_valid,
 500                       ea_vals_buf_size, -1, 0);
 501
 502         /* get SELinux policy info if any */
 503         mdc_file_sepol_pack(&req->rq_pill, sepol);
 504         sptlrpc_sepol_put(sepol);
 505
 506         req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_SERVER,
 507                              GA_DEFAULT_EA_NAME_LEN * GA_DEFAULT_EA_NUM);
 508
 509         req_capsule_set_size(&req->rq_pill, &RMF_EAVALS, RCL_SERVER,
 510                              ea_vals_buf_size);
 511
 512         req_capsule_set_size(&req->rq_pill, &RMF_EAVALS_LENS, RCL_SERVER,
 513                              sizeof(u32) * GA_DEFAULT_EA_NUM);
 514
 515         req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, 0);
 516
 517         ptlrpc_request_set_replen(req);
 518
 519         RETURN(req);
 520
 521 err_put_sepol:
 522         sptlrpc_sepol_put(sepol);
 523 err_free_rq:
 524         ptlrpc_request_free(req);
 525         RETURN(ERR_PTR(rc));
 526 }
 527
 528 static struct ptlrpc_request *
 529 mdc_intent_getattr_pack(struct obd_export *exp, struct lookup_intent *it,
 530                         struct md_op_data *op_data, __u32 acl_bufsize)
 531 {
 532         struct ptlrpc_request *req;
 533         struct obd_device *obd = class_exp2obd(exp);
 534         u64 valid = OBD_MD_FLGETATTR | OBD_MD_FLEASIZE | OBD_MD_FLMODEASIZE |
 535                     OBD_MD_FLDIREA | OBD_MD_MEA | OBD_MD_FLACL |
 536                     OBD_MD_DEFAULT_MEA;
 537         struct ldlm_intent *lit;
 538         __u32 easize;
 539         bool have_secctx = false;
 540         int rc;
 541
 542         ENTRY;
 543         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 544                                    &RQF_LDLM_INTENT_GETATTR);
 545         if (req == NULL)
 546                 RETURN(ERR_PTR(-ENOMEM));
 547
 548         /* send name of security xattr to get upon intent */
 549         if (it->it_op & (IT_LOOKUP | IT_GETATTR) &&
 550             req_capsule_has_field(&req->rq_pill, &RMF_FILE_SECCTX_NAME,
 551                                   RCL_CLIENT) &&
 552             op_data->op_file_secctx_name_size > 0 &&
 553             op_data->op_file_secctx_name != NULL) {
 554                 have_secctx = true;
 555                 req_capsule_set_size(&req->rq_pill, &RMF_FILE_SECCTX_NAME,
 556                                      RCL_CLIENT,
 557                                      op_data->op_file_secctx_name_size);
 558         }
 559
 560         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 561                              op_data->op_namelen + 1);
 562
 563         rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
 564         if (rc) {
 565                 ptlrpc_request_free(req);
 566                 RETURN(ERR_PTR(rc));
 567         }
 568
 569         /* pack the intent */
 570         lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
 571         lit->opc = (__u64)it->it_op;
 572
 573         if (obd->u.cli.cl_default_mds_easize > 0)
 574                 easize = obd->u.cli.cl_default_mds_easize;
 575         else
 576                 easize = obd->u.cli.cl_max_mds_easize;
 577
 578         /* pack the intended request */
 579         mdc_getattr_pack(&req->rq_pill, valid, it->it_flags, op_data, easize);
 580
 581         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER, easize);
 582         req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
 583         req_capsule_set_size(&req->rq_pill, &RMF_DEFAULT_MDT_MD, RCL_SERVER,
 584                              sizeof(struct lmv_user_md));
 585
 586         if (have_secctx) {
 587                 char *secctx_name;
 588
 589                 secctx_name = req_capsule_client_get(&req->rq_pill,
 590                                                      &RMF_FILE_SECCTX_NAME);
 591                 memcpy(secctx_name, op_data->op_file_secctx_name,
 592                        op_data->op_file_secctx_name_size);
 593
 594                 req_capsule_set_size(&req->rq_pill, &RMF_FILE_SECCTX,
 595                                      RCL_SERVER, easize);
 596
 597                 CDEBUG(D_SEC, "packed '%.*s' as security xattr name\n",
 598                        op_data->op_file_secctx_name_size,
 599                        op_data->op_file_secctx_name);
 600         } else {
 601                 req_capsule_set_size(&req->rq_pill, &RMF_FILE_SECCTX,
 602                                      RCL_SERVER, 0);
 603         }
 604
 605         if (exp_connect_encrypt(exp) && it->it_op & (IT_LOOKUP | IT_GETATTR))
 606                 req_capsule_set_size(&req->rq_pill, &RMF_FILE_ENCCTX,
 607                                      RCL_SERVER, easize);
 608         else
 609                 req_capsule_set_size(&req->rq_pill, &RMF_FILE_ENCCTX,
 610                                      RCL_SERVER, 0);
 611
 612         ptlrpc_request_set_replen(req);
 613         RETURN(req);
 614 }
 615
 616 static struct ptlrpc_request *mdc_intent_layout_pack(struct obd_export *exp,
 617                                                      struct lookup_intent *it,
 618                                                      struct md_op_data *op_data)
 619 {
 620         struct obd_device *obd = class_exp2obd(exp);
 621         struct ptlrpc_request *req;
 622         struct ldlm_intent *lit;
 623         struct layout_intent *layout;
 624         LIST_HEAD(cancels);
 625         int count = 0, rc;
 626
 627         ENTRY;
 628         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 629                                 &RQF_LDLM_INTENT_LAYOUT);
 630         if (req == NULL)
 631                 RETURN(ERR_PTR(-ENOMEM));
 632
 633         if (fid_is_sane(&op_data->op_fid2) && (it->it_op & IT_LAYOUT) &&
 634             (it->it_flags & FMODE_WRITE)) {
 635                 count = mdc_resource_get_unused(exp, &op_data->op_fid2,
 636                                                 &cancels, LCK_EX,
 637                                                 MDS_INODELOCK_LAYOUT);
 638         }
 639
 640         req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT, 0);
 641         rc = ldlm_prep_enqueue_req(exp, req, &cancels, count);
 642         if (rc) {
 643                 ptlrpc_request_free(req);
 644                 RETURN(ERR_PTR(rc));
 645         }
 646
 647         /* pack the intent */
 648         lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
 649         lit->opc = (__u64)it->it_op;
 650
 651         /* pack the layout intent request */
 652         layout = req_capsule_client_get(&req->rq_pill, &RMF_LAYOUT_INTENT);
 653         LASSERT(op_data->op_data != NULL);
 654         LASSERT(op_data->op_data_size == sizeof(*layout));
 655         memcpy(layout, op_data->op_data, sizeof(*layout));
 656
 657         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
 658                              obd->u.cli.cl_default_mds_easize);
 659         ptlrpc_request_set_replen(req);
 660         RETURN(req);
 661 }
 662
 663 static struct ptlrpc_request *mdc_enqueue_pack(struct obd_export *exp,
 664                                                int lvb_len)
 665 {
 666         struct ptlrpc_request *req;
 667         int rc;
 668
 669         ENTRY;
 670         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_ENQUEUE);
 671         if (req == NULL)
 672                 RETURN(ERR_PTR(-ENOMEM));
 673
 674         rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
 675         if (rc) {
 676                 ptlrpc_request_free(req);
 677                 RETURN(ERR_PTR(rc));
 678         }
 679
 680         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER, lvb_len);
 681         ptlrpc_request_set_replen(req);
 682         RETURN(req);
 683 }
 684
 685 int mdc_finish_enqueue(struct obd_export *exp,
 686                        struct req_capsule *pill,
 687                        struct ldlm_enqueue_info *einfo,
 688                        struct lookup_intent *it,
 689                        struct lustre_handle *lockh, int rc)
 690 {
 691         struct ptlrpc_request *req = pill->rc_req;
 692         struct ldlm_request *lockreq;
 693         struct ldlm_reply *lockrep;
 694         struct ldlm_lock *lock;
 695         struct mdt_body *body = NULL;
 696         void *lvb_data = NULL;
 697         __u32 lvb_len = 0;
 698
 699         ENTRY;
 700         LASSERT(rc >= 0);
 701         /* Similarly, if we're going to replay this request, we don't want to
 702          * actually get a lock, just perform the intent.
 703          */
 704         if (req->rq_transno || req->rq_replay) {
 705                 lockreq = req_capsule_client_get(pill, &RMF_DLM_REQ);
 706                 lockreq->lock_flags |= ldlm_flags_to_wire(LDLM_FL_INTENT_ONLY);
 707         }
 708
 709         if (rc == ELDLM_LOCK_ABORTED) {
 710                 einfo->ei_mode = 0;
 711                 memset(lockh, 0, sizeof(*lockh));
 712                 rc = 0;
 713         } else { /* rc = 0 */
 714                 lock = ldlm_handle2lock(lockh);
 715                 LASSERT(lock != NULL);
 716
 717                 /* If server returned a different lock mode, fix up variables */
 718                 if (lock->l_req_mode != einfo->ei_mode) {
 719                         ldlm_lock_addref(lockh, lock->l_req_mode);
 720                         ldlm_lock_decref(lockh, einfo->ei_mode);
 721                         einfo->ei_mode = lock->l_req_mode;
 722                 }
 723                 LDLM_LOCK_PUT(lock);
 724         }
 725
 726         lockrep = req_capsule_server_get(pill, &RMF_DLM_REP);
 727         LASSERT(lockrep != NULL); /* checked by ldlm_cli_enqueue() */
 728
 729         it->it_disposition = (int)lockrep->lock_policy_res1;
 730         it->it_status = (int)lockrep->lock_policy_res2;
 731         it->it_lock_mode = einfo->ei_mode;
 732         it->it_lock_handle = lockh->cookie;
 733         it->it_request = req;
 734
 735         /* Technically speaking rq_transno must already be zero if
 736          * it_status is in error, so the check is a bit redundant.
 737          */
 738         if ((!req->rq_transno || it->it_status < 0) && req->rq_replay)
 739                 mdc_clear_replay_flag(req, it->it_status);
 740
 741         /* If we're doing an IT_OPEN which did not result in an actual
 742          * successful open, then we need to remove the bit which saves
 743          * this request for unconditional replay.
 744          *
 745          * It's important that we do this first!  Otherwise we might exit the
 746          * function without doing so, and try to replay a failed create.
 747          * (b=3440)
 748          */
 749         if (it->it_op & IT_OPEN && req->rq_replay &&
 750             (!it_disposition(it, DISP_OPEN_OPEN) || it->it_status != 0))
 751                 mdc_clear_replay_flag(req, it->it_status);
 752
 753         DEBUG_REQ(D_RPCTRACE, req, "op=%x disposition=%x, status=%d",
 754                   it->it_op, it->it_disposition, it->it_status);
 755
 756         /* We know what to expect, so we do any byte flipping required here */
 757         if (it_has_reply_body(it)) {
 758                 body = req_capsule_server_get(pill, &RMF_MDT_BODY);
 759                 if (body == NULL) {
 760                         rc = -EPROTO;
 761                         CERROR("%s: cannot swab mdt_body: rc = %d\n",
 762                                exp->exp_obd->obd_name, rc);
 763                         RETURN(rc);
 764                 }
 765
 766                 if (it_disposition(it, DISP_OPEN_OPEN) &&
 767                     !it_open_error(DISP_OPEN_OPEN, it)) {
 768                         /*
 769                          * If this is a successful OPEN request, we need to set
 770                          * replay handler and data early, so that if replay
 771                          * happens immediately after swabbing below, new reply
 772                          * is swabbed by that handler correctly.
 773                          */
 774                         mdc_set_open_replay_data(NULL, NULL, it);
 775                 }
 776
 777                 if (it_disposition(it, DISP_OPEN_CREATE) &&
 778                     !it_open_error(DISP_OPEN_CREATE, it)) {
 779                         lprocfs_counter_incr(exp->exp_obd->obd_md_stats,
 780                                              LPROC_MD_CREATE);
 781                 }
 782
 783                 if (body->mbo_valid & (OBD_MD_FLDIREA | OBD_MD_FLEASIZE)) {
 784                         void *eadata;
 785
 786                         mdc_update_max_ea_from_body(exp, body);
 787
 788                         /*
 789                          * The eadata is opaque; just check that it is there.
 790                          * Eventually, obd_unpackmd() will check the contents.
 791                          */
 792                         eadata = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
 793                                                         body->mbo_eadatasize);
 794                         if (eadata == NULL)
 795                                 RETURN(-EPROTO);
 796
 797                         /* save LVB data and length if for layout lock */
 798                         lvb_data = eadata;
 799                         lvb_len = body->mbo_eadatasize;
 800
 801                         /*
 802                          * We save the reply LOV EA in case we have to replay a
 803                          * create for recovery.  If we didn't allocate a large
 804                          * enough request buffer above we need to reallocate it
 805                          * here to hold the actual LOV EA.
 806                          *
 807                          * To not save LOV EA if request is not going to replay
 808                          * (for example error one).
 809                          */
 810                         if ((it->it_op & IT_OPEN) && req->rq_replay) {
 811                                 rc = mdc_save_lmm(req, eadata,
 812                                                   body->mbo_eadatasize);
 813                                 if (rc) {
 814                                         body->mbo_valid &= ~OBD_MD_FLEASIZE;
 815                                         body->mbo_eadatasize = 0;
 816                                         rc = 0;
 817                                 }
 818                         }
 819                 }
 820         } else if (it->it_op & IT_LAYOUT) {
 821                 /* maybe the lock was granted right away and layout
 822                  * is packed into RMF_DLM_LVB of req
 823                  */
 824                 lvb_len = req_capsule_get_size(pill, &RMF_DLM_LVB, RCL_SERVER);
 825                 CDEBUG(D_INFO, "%s: layout return lvb %d transno %lld\n",
 826                        class_exp2obd(exp)->obd_name, lvb_len, req->rq_transno);
 827                 if (lvb_len > 0) {
 828                         lvb_data = req_capsule_server_sized_get(pill,
 829                                                         &RMF_DLM_LVB, lvb_len);
 830                         if (lvb_data == NULL)
 831                                 RETURN(-EPROTO);
 832
 833                         /**
 834                          * save replied layout data to the request buffer for
 835                          * recovery consideration (lest MDS reinitialize
 836                          * another set of OST objects).
 837                          */
 838                         if (req->rq_transno)
 839                                 mdc_save_lmm(req, lvb_data, lvb_len);
 840                 }
 841         }
 842
 843         /* fill in stripe data for layout lock.
 844          * LU-6581: trust layout data only if layout lock is granted. The MDT
 845          * has stopped sending layout unless the layout lock is granted. The
 846          * client still does this checking in case it's talking with an old
 847          * server. - Jinshan
 848          */
 849         lock = ldlm_handle2lock(lockh);
 850         if (lock == NULL)
 851                 RETURN(rc);
 852
 853         if (ldlm_has_layout(lock) && lvb_data != NULL &&
 854             !(lockrep->lock_flags & LDLM_FL_BLOCKED_MASK)) {
 855                 void *lmm;
 856
 857                 LDLM_DEBUG(lock, "layout lock returned by: %s, lvb_len: %d",
 858                         ldlm_it2str(it->it_op), lvb_len);
 859
 860                 OBD_ALLOC_LARGE(lmm, lvb_len);
 861                 if (lmm == NULL)
 862                         GOTO(out_lock, rc = -ENOMEM);
 863
 864                 memcpy(lmm, lvb_data, lvb_len);
 865
 866                 /* install lvb_data */
 867                 lock_res_and_lock(lock);
 868                 if (lock->l_lvb_data == NULL) {
 869                         lock->l_lvb_type = LVB_T_LAYOUT;
 870                         lock->l_lvb_data = lmm;
 871                         lock->l_lvb_len = lvb_len;
 872                         lmm = NULL;
 873                 }
 874                 unlock_res_and_lock(lock);
 875                 if (lmm != NULL)
 876                         OBD_FREE_LARGE(lmm, lvb_len);
 877         }
 878
 879         if (ldlm_has_dom(lock)) {
 880                 LASSERT(lock->l_glimpse_ast == mdc_ldlm_glimpse_ast);
 881
 882                 body = req_capsule_server_get(pill, &RMF_MDT_BODY);
 883                 if (!(body->mbo_valid & OBD_MD_DOM_SIZE)) {
 884                         LDLM_ERROR(lock, "%s: DoM lock without size.",
 885                                    exp->exp_obd->obd_name);
 886                         GOTO(out_lock, rc = -EPROTO);
 887                 }
 888
 889                 LDLM_DEBUG(lock, "DoM lock is returned by: %s, size: %llu",
 890                            ldlm_it2str(it->it_op), body->mbo_dom_size);
 891
 892                 lock_res_and_lock(lock);
 893                 mdc_body2lvb(body, &lock->l_ost_lvb);
 894                 ldlm_lock_allow_match_locked(lock);
 895                 unlock_res_and_lock(lock);
 896         }
 897 out_lock:
 898         LDLM_LOCK_PUT(lock);
 899
 900         RETURN(rc);
 901 }
 902
 903 static inline bool mdc_skip_mod_rpc_slot(const struct lookup_intent *it)
 904 {
 905         if (it != NULL &&
 906             (it->it_op == IT_GETATTR || it->it_op == IT_LOOKUP ||
 907              it->it_op == IT_READDIR || it->it_op == IT_GETXATTR ||
 908              (it->it_op == IT_LAYOUT && !(it->it_flags & MDS_FMODE_WRITE))))
 909                 return true;
 910         return false;
 911 }
 912
 913 /* We always reserve enough space in the reply packet for a stripe MD, because
 914  * we don't know in advance the file type.
 915  */
 916 static int mdc_enqueue_base(struct obd_export *exp,
 917                             struct ldlm_enqueue_info *einfo,
 918                             const union ldlm_policy_data *policy,
 919                             struct lookup_intent *it,
 920                             struct md_op_data *op_data,
 921                             struct lustre_handle *lockh,
 922                             __u64 extra_lock_flags)
 923 {
 924         struct obd_device *obd = class_exp2obd(exp);
 925         struct ptlrpc_request *req;
 926         __u64 flags, saved_flags = extra_lock_flags;
 927         struct ldlm_res_id res_id;
 928         static const union ldlm_policy_data lookup_policy = {
 929                                   .l_inodebits = { MDS_INODELOCK_LOOKUP } };
 930         static const union ldlm_policy_data update_policy = {
 931                                   .l_inodebits = { MDS_INODELOCK_UPDATE } };
 932         static const union ldlm_policy_data layout_policy = {
 933                                   .l_inodebits = { MDS_INODELOCK_LAYOUT } };
 934         static const union ldlm_policy_data getxattr_policy = {
 935                                   .l_inodebits = { MDS_INODELOCK_XATTR } };
 936         int generation, resends = 0;
 937         struct ldlm_reply *lockrep;
 938         struct obd_import *imp = class_exp2cliimp(exp);
 939         __u32 acl_bufsize;
 940         enum lvb_type lvb_type = 0;
 941         int rc;
 942
 943         ENTRY;
 944         LASSERTF(!it || einfo->ei_type == LDLM_IBITS, "lock type %d\n",
 945                  einfo->ei_type);
 946         fid_build_reg_res_name(&op_data->op_fid1, &res_id);
 947
 948         if (it != NULL) {
 949                 LASSERT(policy == NULL);
 950
 951                 saved_flags |= LDLM_FL_HAS_INTENT;
 952                 if (it->it_op & (IT_GETATTR | IT_READDIR))
 953                         policy = &update_policy;
 954                 else if (it->it_op & IT_LAYOUT)
 955                         policy = &layout_policy;
 956                 else if (it->it_op & IT_GETXATTR)
 957                         policy = &getxattr_policy;
 958                 else
 959                         policy = &lookup_policy;
 960         }
 961
 962         generation = obd->u.cli.cl_import->imp_generation;
 963         if (!it || (it->it_op & (IT_OPEN | IT_CREAT)))
 964                 acl_bufsize = min_t(__u32, imp->imp_connect_data.ocd_max_easize,
 965                                     XATTR_SIZE_MAX);
 966         else
 967                 acl_bufsize = LUSTRE_POSIX_ACL_MAX_SIZE_OLD;
 968
 969 resend:
 970         flags = saved_flags;
 971         if (it == NULL) {
 972                 /* The only way right now is FLOCK. */
 973                 LASSERTF(einfo->ei_type == LDLM_FLOCK, "lock type %d\n",
 974                          einfo->ei_type);
 975                 res_id.name[3] = LDLM_FLOCK;
 976                 req = ldlm_enqueue_pack(exp, 0);
 977         } else if (it->it_op & IT_OPEN) {
 978                 req = mdc_intent_open_pack(exp, it, op_data, acl_bufsize);
 979         } else if (it->it_op & (IT_GETATTR | IT_LOOKUP)) {
 980                 req = mdc_intent_getattr_pack(exp, it, op_data, acl_bufsize);
 981         } else if (it->it_op & IT_READDIR) {
 982                 req = mdc_enqueue_pack(exp, 0);
 983         } else if (it->it_op & IT_LAYOUT) {
 984                 if (!imp_connect_lvb_type(imp))
 985                         RETURN(-EOPNOTSUPP);
 986                 req = mdc_intent_layout_pack(exp, it, op_data);
 987                 lvb_type = LVB_T_LAYOUT;
 988         } else if (it->it_op & IT_GETXATTR) {
 989                 req = mdc_intent_getxattr_pack(exp, it, op_data);
 990         } else {
 991                 LBUG();
 992                 RETURN(-EINVAL);
 993         }
 994
 995         if (IS_ERR(req))
 996                 RETURN(PTR_ERR(req));
 997
 998         if (resends) {
 999                 req->rq_generation_set = 1;
1000                 req->rq_import_generation = generation;
1001                 req->rq_sent = ktime_get_real_seconds() + resends;
1002         }
1003
1004         einfo->ei_req_slot = !(op_data->op_cli_flags & CLI_NO_SLOT);
1005         einfo->ei_mod_slot = !mdc_skip_mod_rpc_slot(it);
1006
1007         /* With Data-on-MDT the glimpse callback is needed too.
1008          * It is set here in advance but not in mdc_finish_enqueue()
1009          * to avoid possible races. It is safe to have glimpse handler
1010          * for non-DOM locks and costs nothing.
1011          */
1012         if (einfo->ei_cb_gl == NULL)
1013                 einfo->ei_cb_gl = mdc_ldlm_glimpse_ast;
1014
1015         rc = ldlm_cli_enqueue(exp, &req, einfo, &res_id, policy, &flags, NULL,
1016                               0, lvb_type, lockh, 0);
1017
1018         if (!it) {
1019                 /* For flock requests we immediatelly return without further
1020                  * delay and let caller deal with the rest, since rest of
1021                  * this function metadata processing makes no sense for flock
1022                  * requests anyway. But in case of problem during comms with
1023                  * server (-ETIMEDOUT) or any signal/kill attempt (-EINTR),
1024                  * we cannot rely on caller and this mainly for F_UNLCKs
1025                  * (explicits or automatically generated by kernel to clean
1026                  * current flocks upon exit) that can't be trashed.
1027                  */
1028                 ptlrpc_req_finished(req);
1029                 if (((rc == -EINTR) || (rc == -ETIMEDOUT)) &&
1030                     (einfo->ei_type == LDLM_FLOCK) &&
1031                     (einfo->ei_mode == LCK_NL))
1032                         goto resend;
1033                 RETURN(rc);
1034         }
1035
1036         if (rc < 0) {
1037                 CDEBUG(D_INFO,
1038                       "%s: ldlm_cli_enqueue "DFID":"DFID"=%s failed: rc = %d\n",
1039                       obd->obd_name, PFID(&op_data->op_fid1),
1040                       PFID(&op_data->op_fid2), op_data->op_name ?: "", rc);
1041
1042                 mdc_clear_replay_flag(req, rc);
1043                 ptlrpc_req_finished(req);
1044                 RETURN(rc);
1045         }
1046
1047         lockrep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1048         LASSERT(lockrep != NULL);
1049
1050         lockrep->lock_policy_res2 =
1051                 ptlrpc_status_ntoh(lockrep->lock_policy_res2);
1052
1053         /* Retry infinitely when the server returns -EINPROGRESS for the
1054          * intent operation, when server returns -EINPROGRESS for acquiring
1055          * intent lock, we'll retry in after_reply().
1056          */
1057         if (it && (int)lockrep->lock_policy_res2 == -EINPROGRESS) {
1058                 mdc_clear_replay_flag(req, rc);
1059                 ptlrpc_req_finished(req);
1060                 if (generation == obd->u.cli.cl_import->imp_generation) {
1061                         if (signal_pending(current))
1062                                 RETURN(-EINTR);
1063
1064                         resends++;
1065                         CDEBUG(D_HA, "%s: resend:%d op:%d "DFID"/"DFID"\n",
1066                                obd->obd_name, resends, it->it_op,
1067                                PFID(&op_data->op_fid1),
1068                                PFID(&op_data->op_fid2));
1069                         goto resend;
1070                 } else {
1071                         CDEBUG(D_HA, "resend cross eviction\n");
1072                         RETURN(-EIO);
1073                 }
1074         }
1075
1076         if ((int)lockrep->lock_policy_res2 == -ERANGE &&
1077             it->it_op & (IT_OPEN | IT_GETATTR | IT_LOOKUP) &&
1078             acl_bufsize == LUSTRE_POSIX_ACL_MAX_SIZE_OLD) {
1079                 mdc_clear_replay_flag(req, -ERANGE);
1080                 ptlrpc_req_finished(req);
1081                 acl_bufsize = min_t(__u32, imp->imp_connect_data.ocd_max_easize,
1082                                     XATTR_SIZE_MAX);
1083                 goto resend;
1084         }
1085
1086         rc = mdc_finish_enqueue(exp, &req->rq_pill, einfo, it, lockh, rc);
1087         if (rc < 0) {
1088                 if (lustre_handle_is_used(lockh)) {
1089                         ldlm_lock_decref(lockh, einfo->ei_mode);
1090                         memset(lockh, 0, sizeof(*lockh));
1091                 }
1092                 ptlrpc_req_finished(req);
1093
1094                 it->it_lock_handle = 0;
1095                 it->it_lock_mode = 0;
1096                 it->it_request = NULL;
1097         }
1098
1099         RETURN(rc);
1100 }
1101
1102 int mdc_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1103                 const union ldlm_policy_data *policy,
1104                 struct md_op_data *op_data,
1105                 struct lustre_handle *lockh, __u64 extra_lock_flags)
1106 {
1107         return mdc_enqueue_base(exp, einfo, policy, NULL,
1108                                 op_data, lockh, extra_lock_flags);
1109 }
1110
1111 static int mdc_finish_intent_lock(struct obd_export *exp,
1112                                   struct ptlrpc_request *request,
1113                                   struct md_op_data *op_data,
1114                                   struct lookup_intent *it,
1115                                   struct lustre_handle *lockh)
1116 {
1117         struct lustre_handle old_lock;
1118         struct ldlm_lock *lock;
1119         int rc = 0;
1120
1121         ENTRY;
1122         LASSERT(request != NULL);
1123         LASSERT(request != LP_POISON);
1124         LASSERT(request->rq_repmsg != LP_POISON);
1125
1126         if (it->it_op & IT_READDIR)
1127                 RETURN(0);
1128
1129         if (it->it_op & (IT_GETXATTR | IT_LAYOUT)) {
1130                 if (it->it_status != 0)
1131                         GOTO(out, rc = it->it_status);
1132         } else {
1133                 if (!it_disposition(it, DISP_IT_EXECD)) {
1134                         /* The server failed before it even started executing
1135                          * the intent, i.e. because it couldn't unpack the
1136                          * request.
1137                          */
1138                         LASSERT(it->it_status != 0);
1139                         GOTO(out, rc = it->it_status);
1140                 }
1141                 rc = it_open_error(DISP_IT_EXECD, it);
1142                 if (rc)
1143                         GOTO(out, rc);
1144
1145                 rc = it_open_error(DISP_LOOKUP_EXECD, it);
1146                 if (rc)
1147                         GOTO(out, rc);
1148
1149                 /* keep requests around for the multiple phases of the call
1150                  * this shows the DISP_XX must guarantee we make it into the
1151                  * call
1152                  */
1153                 if (!it_disposition(it, DISP_ENQ_CREATE_REF) &&
1154                     it_disposition(it, DISP_OPEN_CREATE) &&
1155                     !it_open_error(DISP_OPEN_CREATE, it)) {
1156                         it_set_disposition(it, DISP_ENQ_CREATE_REF);
1157                         /* balanced in ll_create_node */
1158                         ptlrpc_request_addref(request);
1159                 }
1160                 if (!it_disposition(it, DISP_ENQ_OPEN_REF) &&
1161                     it_disposition(it, DISP_OPEN_OPEN) &&
1162                     !it_open_error(DISP_OPEN_OPEN, it)) {
1163                         it_set_disposition(it, DISP_ENQ_OPEN_REF);
1164                         /* balanced in ll_file_open */
1165                         ptlrpc_request_addref(request);
1166                         /* eviction in middle of open RPC processing b=11546 */
1167                         CFS_FAIL_TIMEOUT(OBD_FAIL_MDC_ENQUEUE_PAUSE,
1168                                          obd_timeout);
1169                 }
1170
1171                 if (it->it_op & IT_CREAT) {
1172                         /* XXX this belongs in ll_create_it */
1173                 } else if (it->it_op == IT_OPEN) {
1174                         LASSERT(!it_disposition(it, DISP_OPEN_CREATE));
1175                 } else {
1176                         LASSERT(it->it_op & (IT_GETATTR | IT_LOOKUP));
1177                 }
1178         }
1179
1180         /* If we already have a matching lock, then cancel the new
1181          * one.  We have to set the data here instead of in
1182          * mdc_enqueue, because we need to use the child's inode as
1183          * the l_ast_data to match, and that's not available until
1184          * intent_finish has performed the iget().
1185          */
1186         lock = ldlm_handle2lock(lockh);
1187         if (lock) {
1188                 union ldlm_policy_data policy = lock->l_policy_data;
1189
1190                 LDLM_DEBUG(lock, "matching against this");
1191
1192                 if (it_has_reply_body(it)) {
1193                         struct mdt_body *body;
1194
1195                         body = req_capsule_server_get(&request->rq_pill,
1196                                                       &RMF_MDT_BODY);
1197                         /* mdc_enqueue checked */
1198                         LASSERT(body != NULL);
1199                         LASSERTF(fid_res_name_eq(&body->mbo_fid1,
1200                                                  &lock->l_resource->lr_name),
1201                                  "Lock res_id: "DLDLMRES", fid: "DFID"\n",
1202                                  PLDLMRES(lock->l_resource),
1203                                  PFID(&body->mbo_fid1));
1204                 }
1205                 LDLM_LOCK_PUT(lock);
1206
1207                 memcpy(&old_lock, lockh, sizeof(*lockh));
1208                 if (ldlm_lock_match(NULL, LDLM_FL_BLOCK_GRANTED, NULL,
1209                                    LDLM_IBITS, &policy, LCK_NL, &old_lock)) {
1210                         ldlm_lock_decref_and_cancel(lockh, it->it_lock_mode);
1211                         memcpy(lockh, &old_lock, sizeof(old_lock));
1212                         it->it_lock_handle = lockh->cookie;
1213                 }
1214         }
1215
1216         EXIT;
1217 out:
1218         CDEBUG(D_DENTRY,
1219                "D_IT dentry=%.*s intent=%s status=%d disp=%x: rc = %d\n",
1220                 (int)op_data->op_namelen, op_data->op_name,
1221                 ldlm_it2str(it->it_op), it->it_status, it->it_disposition, rc);
1222
1223         return rc;
1224 }
1225
1226 int mdc_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
1227                         struct lu_fid *fid, __u64 *bits)
1228 {
1229         /* We could just return 1 immediately, but as we should only be called
1230          * in revalidate_it if we already have a lock, let's verify that.
1231          */
1232         struct ldlm_res_id res_id;
1233         struct lustre_handle lockh;
1234         union ldlm_policy_data policy;
1235         enum ldlm_mode mode;
1236
1237         ENTRY;
1238         if (it->it_lock_handle) {
1239                 lockh.cookie = it->it_lock_handle;
1240                 mode = ldlm_revalidate_lock_handle(&lockh, bits);
1241         } else {
1242                 fid_build_reg_res_name(fid, &res_id);
1243                 switch (it->it_op) {
1244                 case IT_GETATTR:
1245                         /* File attributes are held under multiple bits:
1246                          * nlink is under lookup lock, size and times are
1247                          * under UPDATE lock and recently we've also got
1248                          * a separate permissions lock for owner/group/acl that
1249                          * were protected by lookup lock before.
1250                          * Getattr must provide all of that information,
1251                          * so we need to ensure we have all of those locks.
1252                          * Unfortunately, if the bits are split across multiple
1253                          * locks, there's no easy way to match all of them here,
1254                          * so an extra RPC would be performed to fetch all
1255                          * of those bits at once for now.
1256                          */
1257                         /* For new MDTs(> 2.4), UPDATE|PERM should be enough,
1258                          * but for old MDTs (< 2.4), permission is covered
1259                          * by LOOKUP lock, so it needs to match all bits here.
1260                          */
1261                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE |
1262                                                   MDS_INODELOCK_PERM;
1263                         break;
1264                 case IT_READDIR:
1265                         policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
1266                         break;
1267                 case IT_LAYOUT:
1268                         policy.l_inodebits.bits = MDS_INODELOCK_LAYOUT;
1269                         break;
1270                 default:
1271                         policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1272                         break;
1273                 }
1274
1275                 mode = mdc_lock_match(exp, LDLM_FL_BLOCK_GRANTED, fid,
1276                                       LDLM_IBITS, &policy,
1277                                       LCK_CR | LCK_CW | LCK_PR | LCK_PW,
1278                                       &lockh);
1279         }
1280
1281         if (mode) {
1282                 it->it_lock_handle = lockh.cookie;
1283                 it->it_lock_mode = mode;
1284         } else {
1285                 it->it_lock_handle = 0;
1286                 it->it_lock_mode = 0;
1287         }
1288
1289         RETURN(!!mode);
1290 }
1291
1292 /*
1293  * This long block is all about fixing up the lock and request state
1294  * so that it is correct as of the moment _before_ the operation was
1295  * applied; that way, the VFS will think that everything is normal and
1296  * call Lustre's regular VFS methods.
1297  *
1298  * If we're performing a creation, that means that unless the creation
1299  * failed with EEXIST, we should fake up a negative dentry.
1300  *
1301  * For everything else, we want the lookup to succeed.
1302  *
1303  * One additional note: if CREATE or OPEN succeeded, we add an extra
1304  * reference to the request because we need to keep it around until
1305  * ll_create/ll_open gets called.
1306  *
1307  * The server will return to us, in it_disposition, an indication of
1308  * exactly what it_status refers to.
1309  *
1310  * If DISP_OPEN_OPEN is set, then it_status refers to the open() call,
1311  * otherwise if DISP_OPEN_CREATE is set, then it_status is the
1312  * creation failure mode.  In either case, one of DISP_LOOKUP_NEG or
1313  * DISP_LOOKUP_POS will be set, indicating whether the child lookup
1314  * was successful.
1315  *
1316  * Else, if DISP_LOOKUP_EXECD then it_status is the rc of the
1317  * child lookup.
1318  */
1319 int mdc_intent_lock(struct obd_export *exp, struct md_op_data *op_data,
1320                     struct lookup_intent *it, struct ptlrpc_request **reqp,
1321                     ldlm_blocking_callback cb_blocking, __u64 extra_lock_flags)
1322 {
1323         struct ldlm_enqueue_info einfo = {
1324                 .ei_type        = LDLM_IBITS,
1325                 .ei_mode        = it_to_lock_mode(it),
1326                 .ei_cb_bl       = cb_blocking,
1327                 .ei_cb_cp       = ldlm_completion_ast,
1328                 .ei_cb_gl       = mdc_ldlm_glimpse_ast,
1329         };
1330         struct lustre_handle lockh;
1331         int rc = 0;
1332
1333         ENTRY;
1334         LASSERT(it);
1335         CDEBUG(D_DLMTRACE, "(name: %.*s,"DFID") in obj "DFID
1336                 ", intent: %s flags %#llo\n", (int)op_data->op_namelen,
1337                 op_data->op_name, PFID(&op_data->op_fid2),
1338                 PFID(&op_data->op_fid1), ldlm_it2str(it->it_op),
1339                 it->it_flags);
1340
1341         lockh.cookie = 0;
1342         /* MDS_FID_OP is not a revalidate case */
1343         if (fid_is_sane(&op_data->op_fid2) &&
1344             (it->it_op & (IT_LOOKUP | IT_GETATTR | IT_READDIR)) &&
1345             !(op_data->op_bias & MDS_FID_OP)) {
1346                 /* We could just return 1 immediately, but since we should only
1347                  * be called in revalidate_it if we already have a lock, let's
1348                  * verify that.
1349                  */
1350                 it->it_lock_handle = 0;
1351                 rc = mdc_revalidate_lock(exp, it, &op_data->op_fid2, NULL);
1352                 /* Only return failure if it was not GETATTR by cfid
1353                  * (from inode_revalidate()).
1354                  */
1355                 if (rc || op_data->op_namelen != 0)
1356                         RETURN(rc);
1357         }
1358
1359         /* For case if upper layer did not alloc fid, do it now. */
1360         if (!fid_is_sane(&op_data->op_fid2) && it->it_op & IT_CREAT) {
1361                 rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
1362                 if (rc < 0) {
1363                         CERROR("%s: cannot allocate new FID: rc=%d\n",
1364                                exp->exp_obd->obd_name, rc);
1365                         RETURN(rc);
1366                 }
1367         }
1368
1369         rc = mdc_enqueue_base(exp, &einfo, NULL, it, op_data, &lockh,
1370                               extra_lock_flags);
1371         if (rc < 0)
1372                 RETURN(rc);
1373
1374         *reqp = it->it_request;
1375         rc = mdc_finish_intent_lock(exp, *reqp, op_data, it, &lockh);
1376         RETURN(rc);
1377 }
1378
1379 static int mdc_intent_getattr_async_interpret(const struct lu_env *env,
1380                                               struct ptlrpc_request *req,
1381                                               void *args, int rc)
1382 {
1383         struct mdc_getattr_args *ga = args;
1384         struct obd_export *exp = ga->ga_exp;
1385         struct md_op_item *item = ga->ga_item;
1386         struct ldlm_enqueue_info *einfo = &item->mop_einfo;
1387         struct lookup_intent *it = &item->mop_it;
1388         struct lustre_handle *lockh = &item->mop_lockh;
1389         struct req_capsule *pill = &req->rq_pill;
1390         struct ldlm_reply *lockrep;
1391         __u64 flags = LDLM_FL_HAS_INTENT;
1392
1393         ENTRY;
1394         if (CFS_FAIL_CHECK(OBD_FAIL_MDC_GETATTR_ENQUEUE))
1395                 rc = -ETIMEDOUT;
1396
1397         rc = ldlm_cli_enqueue_fini(exp, pill, einfo, 1, &flags, NULL, 0,
1398                                    lockh, rc, true);
1399         if (rc < 0) {
1400                 CERROR("%s: ldlm_cli_enqueue_fini() failed: rc = %d\n",
1401                        exp->exp_obd->obd_name, rc);
1402                 mdc_clear_replay_flag(req, rc);
1403                 GOTO(out, rc);
1404         }
1405
1406         lockrep = req_capsule_server_get(pill, &RMF_DLM_REP);
1407         LASSERT(lockrep != NULL);
1408
1409         lockrep->lock_policy_res2 =
1410                 ptlrpc_status_ntoh(lockrep->lock_policy_res2);
1411
1412         rc = mdc_finish_enqueue(exp, pill, einfo, it, lockh, rc);
1413         if (rc)
1414                 GOTO(out, rc);
1415
1416         rc = mdc_finish_intent_lock(exp, req, &item->mop_data, it, lockh);
1417         EXIT;
1418
1419 out:
1420         item->mop_pill = pill;
1421         item->mop_cb(item, rc);
1422         return 0;
1423 }
1424
1425 int mdc_intent_getattr_async(struct obd_export *exp,
1426                              struct md_op_item *item)
1427 {
1428         struct md_op_data *op_data = &item->mop_data;
1429         struct lookup_intent *it = &item->mop_it;
1430         struct ptlrpc_request *req;
1431         struct mdc_getattr_args *ga;
1432         struct ldlm_res_id res_id;
1433         union ldlm_policy_data policy = {
1434                 .l_inodebits = { MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE }
1435         };
1436         __u64 flags = LDLM_FL_HAS_INTENT;
1437         int rc = 0;
1438
1439         ENTRY;
1440         CDEBUG(D_DLMTRACE,
1441                "name: %.*s in inode "DFID", intent: %s flags %#llo\n",
1442                (int)op_data->op_namelen, op_data->op_name,
1443                PFID(&op_data->op_fid1), ldlm_it2str(it->it_op), it->it_flags);
1444
1445         fid_build_reg_res_name(&op_data->op_fid1, &res_id);
1446         /* If the MDT return -ERANGE because of large ACL, then the sponsor
1447          * of the async getattr RPC will handle that by itself.
1448          */
1449         req = mdc_intent_getattr_pack(exp, it, op_data,
1450                                       LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
1451         if (IS_ERR(req))
1452                 RETURN(PTR_ERR(req));
1453
1454         /* With Data-on-MDT the glimpse callback is needed too.
1455          * It is set here in advance but not in mdc_finish_enqueue()
1456          * to avoid possible races. It is safe to have glimpse handler
1457          * for non-DOM locks and costs nothing.
1458          */
1459         if (item->mop_einfo.ei_cb_gl == NULL)
1460                 item->mop_einfo.ei_cb_gl = mdc_ldlm_glimpse_ast;
1461
1462         rc = ldlm_cli_enqueue(exp, &req, &item->mop_einfo, &res_id, &policy,
1463                               &flags, NULL, 0, LVB_T_NONE, &item->mop_lockh, 1);
1464         if (rc < 0) {
1465                 ptlrpc_req_finished(req);
1466                 RETURN(rc);
1467         }
1468
1469         ga = ptlrpc_req_async_args(ga, req);
1470         ga->ga_exp = exp;
1471         ga->ga_item = item;
1472
1473         req->rq_interpret_reply = mdc_intent_getattr_async_interpret;
1474         ptlrpcd_add_req(req);
1475
1476         RETURN(0);
1477 }