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