lustre/mdc/mdc_reint.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) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 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 #include <linux/kernel.h>
  37
  38 #include <obd_class.h>
  39 #include "mdc_internal.h"
  40 #include <lustre_fid.h>
  41
  42 /* mdc_setattr does its own semaphore handling */
  43 static int mdc_reint(struct ptlrpc_request *request, int level)
  44 {
  45         int rc;
  46
  47         request->rq_send_state = level;
  48
  49         ptlrpc_get_mod_rpc_slot(request);
  50         rc = ptlrpc_queue_wait(request);
  51         ptlrpc_put_mod_rpc_slot(request);
  52         if (rc)
  53                 CDEBUG(D_INFO, "error in handling %d\n", rc);
  54         else if (!req_capsule_server_get(&request->rq_pill, &RMF_MDT_BODY)) {
  55                 rc = -EPROTO;
  56         }
  57         return rc;
  58 }
  59
  60 /* Find and cancel locally locks matched by inode @bits & @mode in the resource
  61  * found by @fid. Found locks are added into @cancel list. Returns the amount of
  62  * locks added to @cancels list. */
  63 int mdc_resource_get_unused_res(struct obd_export *exp,
  64                                 struct ldlm_res_id *res_id,
  65                                 struct list_head *cancels,
  66                                 enum ldlm_mode mode, __u64 bits)
  67 {
  68         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
  69         union ldlm_policy_data policy = { { 0 } };
  70         struct ldlm_resource *res;
  71         int count;
  72
  73         ENTRY;
  74
  75         /* Return, i.e. cancel nothing, only if ELC is supported (flag in
  76          * export) but disabled through procfs (flag in NS).
  77          *
  78          * This distinguishes from a case when ELC is not supported originally,
  79          * when we still want to cancel locks in advance and just cancel them
  80          * locally, without sending any RPC. */
  81         if (exp_connect_cancelset(exp) && !ns_connect_cancelset(ns))
  82                 RETURN(0);
  83
  84         res = ldlm_resource_get(ns, NULL, res_id, 0, 0);
  85         if (IS_ERR(res))
  86                 RETURN(0);
  87         LDLM_RESOURCE_ADDREF(res);
  88         /* Initialize ibits lock policy. */
  89         policy.l_inodebits.bits = bits;
  90         count = ldlm_cancel_resource_local(res, cancels, &policy, mode, 0, 0,
  91                                            NULL);
  92         LDLM_RESOURCE_DELREF(res);
  93         ldlm_resource_putref(res);
  94         RETURN(count);
  95 }
  96
  97 int mdc_resource_get_unused(struct obd_export *exp, const struct lu_fid *fid,
  98                             struct list_head *cancels, enum ldlm_mode mode,
  99                             __u64 bits)
 100 {
 101         struct ldlm_res_id res_id;
 102
 103         fid_build_reg_res_name(fid, &res_id);
 104         return mdc_resource_get_unused_res(exp, &res_id, cancels, mode, bits);
 105 }
 106
 107 int mdc_setattr(struct obd_export *exp, struct md_op_data *op_data,
 108                 void *ea, size_t ealen, struct ptlrpc_request **request)
 109 {
 110         LIST_HEAD(cancels);
 111         struct ptlrpc_request *req;
 112         int count = 0, rc;
 113         __u64 bits;
 114         ENTRY;
 115
 116         LASSERT(op_data != NULL);
 117
 118         bits = MDS_INODELOCK_UPDATE;
 119         if (op_data->op_attr.ia_valid & (ATTR_MODE|ATTR_UID|ATTR_GID))
 120                 bits |= MDS_INODELOCK_LOOKUP;
 121         if ((op_data->op_flags & MF_MDC_CANCEL_FID1) &&
 122             (fid_is_sane(&op_data->op_fid1)))
 123                 count = mdc_resource_get_unused(exp, &op_data->op_fid1,
 124                                                 &cancels, LCK_EX, bits);
 125         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 126                                    &RQF_MDS_REINT_SETATTR);
 127         if (req == NULL) {
 128                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
 129                 RETURN(-ENOMEM);
 130         }
 131
 132         req_capsule_set_size(&req->rq_pill, &RMF_MDT_EPOCH, RCL_CLIENT, 0);
 133         req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT, ealen);
 134         req_capsule_set_size(&req->rq_pill, &RMF_LOGCOOKIES, RCL_CLIENT, 0);
 135
 136         rc = mdc_prep_elc_req(exp, req, MDS_REINT, &cancels, count);
 137         if (rc) {
 138                 ptlrpc_request_free(req);
 139                 RETURN(rc);
 140         }
 141
 142         if (op_data->op_attr.ia_valid & (ATTR_MTIME | ATTR_CTIME))
 143                 CDEBUG(D_INODE, "setting mtime %lld, ctime %lld\n",
 144                        (s64)op_data->op_attr.ia_mtime.tv_sec,
 145                        (s64)op_data->op_attr.ia_ctime.tv_sec);
 146         mdc_setattr_pack(req, op_data, ea, ealen);
 147
 148         req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, 0);
 149
 150         ptlrpc_request_set_replen(req);
 151
 152         rc = mdc_reint(req, LUSTRE_IMP_FULL);
 153         if (rc == -ERESTARTSYS)
 154                 rc = 0;
 155
 156         *request = req;
 157
 158         RETURN(rc);
 159 }
 160
 161 int mdc_create(struct obd_export *exp, struct md_op_data *op_data,
 162                 const void *data, size_t datalen,
 163                 umode_t mode, uid_t uid, gid_t gid,
 164                 cfs_cap_t cap_effective, __u64 rdev,
 165                 struct ptlrpc_request **request)
 166 {
 167         struct ptlrpc_request *req;
 168         int level, rc;
 169         int count, resends = 0;
 170         struct obd_import *import = exp->exp_obd->u.cli.cl_import;
 171         int generation = import->imp_generation;
 172         LIST_HEAD(cancels);
 173         ENTRY;
 174
 175         /* For case if upper layer did not alloc fid, do it now. */
 176         if (!fid_is_sane(&op_data->op_fid2)) {
 177                 /*
 178                  * mdc_fid_alloc() may return errno 1 in case of switch to new
 179                  * sequence, handle this.
 180                  */
 181                 rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
 182                 if (rc < 0)
 183                         RETURN(rc);
 184         }
 185
 186 rebuild:
 187         count = 0;
 188         if ((op_data->op_flags & MF_MDC_CANCEL_FID1) &&
 189             (fid_is_sane(&op_data->op_fid1)))
 190                 count = mdc_resource_get_unused(exp, &op_data->op_fid1,
 191                                                 &cancels, LCK_EX,
 192                                                 MDS_INODELOCK_UPDATE);
 193
 194         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 195                                    &RQF_MDS_REINT_CREATE_ACL);
 196         if (req == NULL) {
 197                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
 198                 RETURN(-ENOMEM);
 199         }
 200
 201         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 202                              op_data->op_namelen + 1);
 203         req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT,
 204                              data && datalen ? datalen : 0);
 205
 206         req_capsule_set_size(&req->rq_pill, &RMF_FILE_SECCTX_NAME,
 207                              RCL_CLIENT, op_data->op_file_secctx_name != NULL ?
 208                              strlen(op_data->op_file_secctx_name) + 1 : 0);
 209
 210         req_capsule_set_size(&req->rq_pill, &RMF_FILE_SECCTX, RCL_CLIENT,
 211                              op_data->op_file_secctx_size);
 212
 213         req_capsule_set_size(&req->rq_pill, &RMF_FILE_ENCCTX, RCL_CLIENT,
 214                              op_data->op_file_encctx_size);
 215
 216         /* get SELinux policy info if any */
 217         rc = sptlrpc_get_sepol(req);
 218         if (rc < 0) {
 219                 ptlrpc_request_free(req);
 220                 RETURN(rc);
 221         }
 222         req_capsule_set_size(&req->rq_pill, &RMF_SELINUX_POL, RCL_CLIENT,
 223                              strlen(req->rq_sepol) ?
 224                              strlen(req->rq_sepol) + 1 : 0);
 225
 226         rc = mdc_prep_elc_req(exp, req, MDS_REINT, &cancels, count);
 227         if (rc) {
 228                 ptlrpc_request_free(req);
 229                 RETURN(rc);
 230         }
 231
 232         /*
 233          * mdc_create_pack() fills msg->bufs[1] with name and msg->bufs[2] with
 234          * tgt, for symlinks or lov MD data.
 235          */
 236         mdc_create_pack(req, op_data, data, datalen, mode, uid,
 237                         gid, cap_effective, rdev);
 238
 239         ptlrpc_request_set_replen(req);
 240
 241         /* ask ptlrpc not to resend on EINPROGRESS since we have our own retry
 242          * logic here */
 243         req->rq_no_retry_einprogress = 1;
 244
 245         if (resends) {
 246                 req->rq_generation_set = 1;
 247                 req->rq_import_generation = generation;
 248                 req->rq_sent = ktime_get_real_seconds() + resends;
 249         }
 250         level = LUSTRE_IMP_FULL;
 251  resend:
 252         rc = mdc_reint(req, level);
 253
 254         /* Resend if we were told to. */
 255         if (rc == -ERESTARTSYS) {
 256                 level = LUSTRE_IMP_RECOVER;
 257                 goto resend;
 258         } else if (rc == -EINPROGRESS) {
 259                 /* Retry create infinitely until succeed or get other
 260                  * error code or interrupted. */
 261                 ptlrpc_req_finished(req);
 262                 if (generation == import->imp_generation) {
 263                         if (signal_pending(current))
 264                                 RETURN(-EINTR);
 265
 266                         resends++;
 267                         CDEBUG(D_HA, "%s: resend:%d create on "DFID"/"DFID"\n",
 268                                exp->exp_obd->obd_name, resends,
 269                                PFID(&op_data->op_fid1),
 270                                PFID(&op_data->op_fid2));
 271                         goto rebuild;
 272                 } else {
 273                         CDEBUG(D_HA, "resend cross eviction\n");
 274                         RETURN(-EIO);
 275                 }
 276         }
 277
 278         *request = req;
 279         RETURN(rc);
 280 }
 281
 282 int mdc_unlink(struct obd_export *exp, struct md_op_data *op_data,
 283                struct ptlrpc_request **request)
 284 {
 285         LIST_HEAD(cancels);
 286         struct obd_device *obd = class_exp2obd(exp);
 287         struct ptlrpc_request *req = *request;
 288         int count = 0, rc;
 289         ENTRY;
 290
 291         LASSERT(req == NULL);
 292
 293         if ((op_data->op_flags & MF_MDC_CANCEL_FID1) &&
 294             (fid_is_sane(&op_data->op_fid1)))
 295                 count = mdc_resource_get_unused(exp, &op_data->op_fid1,
 296                                                 &cancels, LCK_EX,
 297                                                 MDS_INODELOCK_UPDATE);
 298         if ((op_data->op_flags & MF_MDC_CANCEL_FID3) &&
 299             (fid_is_sane(&op_data->op_fid3)))
 300                 /* cancel DOM lock only if it has no data to flush */
 301                 count += mdc_resource_get_unused(exp, &op_data->op_fid3,
 302                                                  &cancels, LCK_EX,
 303                                                  op_data->op_cli_flags &
 304                                                  CLI_DIRTY_DATA ?
 305                                                  MDS_INODELOCK_ELC :
 306                                                  MDS_INODELOCK_FULL);
 307         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 308                                    &RQF_MDS_REINT_UNLINK);
 309         if (req == NULL) {
 310                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
 311                 RETURN(-ENOMEM);
 312         }
 313
 314         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 315                              op_data->op_namelen + 1);
 316
 317         /* get SELinux policy info if any */
 318         rc = sptlrpc_get_sepol(req);
 319         if (rc < 0) {
 320                 ptlrpc_request_free(req);
 321                 RETURN(rc);
 322         }
 323         req_capsule_set_size(&req->rq_pill, &RMF_SELINUX_POL, RCL_CLIENT,
 324                              strlen(req->rq_sepol) ?
 325                              strlen(req->rq_sepol) + 1 : 0);
 326
 327         rc = mdc_prep_elc_req(exp, req, MDS_REINT, &cancels, count);
 328         if (rc) {
 329                 ptlrpc_request_free(req);
 330                 RETURN(rc);
 331         }
 332
 333         mdc_unlink_pack(req, op_data);
 334
 335         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 336                              obd->u.cli.cl_default_mds_easize);
 337         ptlrpc_request_set_replen(req);
 338
 339         *request = req;
 340
 341         rc = mdc_reint(req, LUSTRE_IMP_FULL);
 342         if (rc == -ERESTARTSYS)
 343                 rc = 0;
 344         RETURN(rc);
 345 }
 346
 347 int mdc_link(struct obd_export *exp, struct md_op_data *op_data,
 348              struct ptlrpc_request **request)
 349 {
 350         LIST_HEAD(cancels);
 351         struct ptlrpc_request *req;
 352         int count = 0, rc;
 353         ENTRY;
 354
 355         if ((op_data->op_flags & MF_MDC_CANCEL_FID2) &&
 356             (fid_is_sane(&op_data->op_fid2)))
 357                 count = mdc_resource_get_unused(exp, &op_data->op_fid2,
 358                                                 &cancels, LCK_EX,
 359                                                 MDS_INODELOCK_UPDATE);
 360         if ((op_data->op_flags & MF_MDC_CANCEL_FID1) &&
 361             (fid_is_sane(&op_data->op_fid1)))
 362                 count += mdc_resource_get_unused(exp, &op_data->op_fid1,
 363                                                  &cancels, LCK_EX,
 364                                                  MDS_INODELOCK_UPDATE);
 365
 366         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_REINT_LINK);
 367         if (req == NULL) {
 368                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
 369                 RETURN(-ENOMEM);
 370         }
 371
 372         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 373                              op_data->op_namelen + 1);
 374
 375         /* get SELinux policy info if any */
 376         rc = sptlrpc_get_sepol(req);
 377         if (rc < 0) {
 378                 ptlrpc_request_free(req);
 379                 RETURN(rc);
 380         }
 381         req_capsule_set_size(&req->rq_pill, &RMF_SELINUX_POL, RCL_CLIENT,
 382                              strlen(req->rq_sepol) ?
 383                              strlen(req->rq_sepol) + 1 : 0);
 384
 385         rc = mdc_prep_elc_req(exp, req, MDS_REINT, &cancels, count);
 386         if (rc) {
 387                 ptlrpc_request_free(req);
 388                 RETURN(rc);
 389         }
 390
 391         mdc_link_pack(req, op_data);
 392         ptlrpc_request_set_replen(req);
 393
 394         rc = mdc_reint(req, LUSTRE_IMP_FULL);
 395         *request = req;
 396         if (rc == -ERESTARTSYS)
 397                 rc = 0;
 398
 399         RETURN(rc);
 400 }
 401
 402 int mdc_rename(struct obd_export *exp, struct md_op_data *op_data,
 403                 const char *old, size_t oldlen, const char *new, size_t newlen,
 404                 struct ptlrpc_request **request)
 405 {
 406         LIST_HEAD(cancels);
 407         struct obd_device *obd = exp->exp_obd;
 408         struct ptlrpc_request *req;
 409         int count = 0, rc;
 410
 411         ENTRY;
 412
 413         if ((op_data->op_flags & MF_MDC_CANCEL_FID1) &&
 414             (fid_is_sane(&op_data->op_fid1)))
 415                 count = mdc_resource_get_unused(exp, &op_data->op_fid1,
 416                                                 &cancels, LCK_EX,
 417                                                 MDS_INODELOCK_UPDATE);
 418         if ((op_data->op_flags & MF_MDC_CANCEL_FID2) &&
 419             (fid_is_sane(&op_data->op_fid2)))
 420                 count += mdc_resource_get_unused(exp, &op_data->op_fid2,
 421                                                  &cancels, LCK_EX,
 422                                                  MDS_INODELOCK_UPDATE);
 423         if ((op_data->op_flags & MF_MDC_CANCEL_FID3) &&
 424             (fid_is_sane(&op_data->op_fid3)))
 425                 count += mdc_resource_get_unused(exp, &op_data->op_fid3,
 426                                                  &cancels, LCK_EX,
 427                                                  MDS_INODELOCK_LOOKUP);
 428         if ((op_data->op_flags & MF_MDC_CANCEL_FID4) &&
 429             (fid_is_sane(&op_data->op_fid4)))
 430                 count += mdc_resource_get_unused(exp, &op_data->op_fid4,
 431                                                  &cancels, LCK_EX,
 432                                                  MDS_INODELOCK_ELC);
 433
 434         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 435                            op_data->op_cli_flags & CLI_MIGRATE ?
 436                            &RQF_MDS_REINT_MIGRATE : &RQF_MDS_REINT_RENAME);
 437         if (req == NULL) {
 438                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
 439                 RETURN(-ENOMEM);
 440         }
 441
 442         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT, oldlen + 1);
 443         req_capsule_set_size(&req->rq_pill, &RMF_SYMTGT, RCL_CLIENT, newlen+1);
 444         if (op_data->op_cli_flags & CLI_MIGRATE)
 445                 req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT,
 446                                      op_data->op_data_size);
 447
 448         /* get SELinux policy info if any */
 449         rc = sptlrpc_get_sepol(req);
 450         if (rc < 0) {
 451                 ptlrpc_request_free(req);
 452                 RETURN(rc);
 453         }
 454         req_capsule_set_size(&req->rq_pill, &RMF_SELINUX_POL, RCL_CLIENT,
 455                              strlen(req->rq_sepol) ?
 456                              strlen(req->rq_sepol) + 1 : 0);
 457
 458         rc = mdc_prep_elc_req(exp, req, MDS_REINT, &cancels, count);
 459         if (rc) {
 460                 ptlrpc_request_free(req);
 461                 RETURN(rc);
 462         }
 463
 464         if (exp_connect_cancelset(exp) && req)
 465                 ldlm_cli_cancel_list(&cancels, count, req, 0);
 466
 467         if (op_data->op_cli_flags & CLI_MIGRATE)
 468                 mdc_migrate_pack(req, op_data, old, oldlen);
 469         else
 470                 mdc_rename_pack(req, op_data, old, oldlen, new, newlen);
 471
 472         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 473                              obd->u.cli.cl_default_mds_easize);
 474         ptlrpc_request_set_replen(req);
 475
 476         rc = mdc_reint(req, LUSTRE_IMP_FULL);
 477         *request = req;
 478         if (rc == -ERESTARTSYS)
 479                 rc = 0;
 480
 481         RETURN(rc);
 482 }
 483
 484 int mdc_file_resync(struct obd_export *exp, struct md_op_data *op_data)
 485 {
 486         LIST_HEAD(cancels);
 487         struct ptlrpc_request *req;
 488         struct ldlm_lock *lock;
 489         struct mdt_rec_resync *rec;
 490         int count = 0, rc;
 491         ENTRY;
 492
 493         if (op_data->op_flags & MF_MDC_CANCEL_FID1 &&
 494             fid_is_sane(&op_data->op_fid1))
 495                 count = mdc_resource_get_unused(exp, &op_data->op_fid1,
 496                                                 &cancels, LCK_EX,
 497                                                 MDS_INODELOCK_LAYOUT);
 498
 499         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 500                                    &RQF_MDS_REINT_RESYNC);
 501         if (req == NULL) {
 502                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
 503                 RETURN(-ENOMEM);
 504         }
 505
 506         rc = mdc_prep_elc_req(exp, req, MDS_REINT, &cancels, count);
 507         if (rc) {
 508                 ptlrpc_request_free(req);
 509                 RETURN(rc);
 510         }
 511
 512         BUILD_BUG_ON(sizeof(*rec) != sizeof(struct mdt_rec_reint));
 513         rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
 514         rec->rs_opcode  = REINT_RESYNC;
 515         rec->rs_fsuid   = op_data->op_fsuid;
 516         rec->rs_fsgid   = op_data->op_fsgid;
 517         rec->rs_cap     = op_data->op_cap;
 518         rec->rs_fid     = op_data->op_fid1;
 519         rec->rs_bias    = op_data->op_bias;
 520         rec->rs_mirror_id = op_data->op_mirror_id;
 521
 522         lock = ldlm_handle2lock(&op_data->op_lease_handle);
 523         if (lock != NULL) {
 524                 rec->rs_lease_handle = lock->l_remote_handle;
 525                 LDLM_LOCK_PUT(lock);
 526         }
 527
 528         ptlrpc_request_set_replen(req);
 529
 530         rc = mdc_reint(req, LUSTRE_IMP_FULL);
 531         if (rc == -ERESTARTSYS)
 532                 rc = 0;
 533
 534         ptlrpc_req_finished(req);
 535         RETURN(rc);
 536 }