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