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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, Whamcloud, Inc.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_MDC
  38
  39 #ifdef __KERNEL__
  40 # include <linux/module.h>
  41 # include <linux/kernel.h>
  42 #else
  43 # include <liblustre.h>
  44 #endif
  45
  46 #include <obd_class.h>
  47 #include "mdc_internal.h"
  48 #include <lustre_fid.h>
  49
  50 /* mdc_setattr does its own semaphore handling */
  51 static int mdc_reint(struct ptlrpc_request *request,
  52                      struct mdc_rpc_lock *rpc_lock,
  53                      int level)
  54 {
  55         int rc;
  56
  57         request->rq_send_state = level;
  58
  59         mdc_get_rpc_lock(rpc_lock, NULL);
  60         rc = ptlrpc_queue_wait(request);
  61         mdc_put_rpc_lock(rpc_lock, NULL);
  62         if (rc)
  63                 CDEBUG(D_INFO, "error in handling %d\n", rc);
  64         else if (!req_capsule_server_get(&request->rq_pill, &RMF_MDT_BODY)) {
  65                 rc = -EPROTO;
  66         }
  67         return rc;
  68 }
  69
  70 /* Find and cancel locally locks matched by inode @bits & @mode in the resource
  71  * found by @fid. Found locks are added into @cancel list. Returns the amount of
  72  * locks added to @cancels list. */
  73 int mdc_resource_get_unused(struct obd_export *exp, struct lu_fid *fid,
  74                             cfs_list_t *cancels, ldlm_mode_t mode,
  75                             __u64 bits)
  76 {
  77         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
  78         ldlm_policy_data_t policy = {{0}};
  79         struct ldlm_res_id res_id;
  80         struct ldlm_resource *res;
  81         int count;
  82         ENTRY;
  83
  84         /* Return, i.e. cancel nothing, only if ELC is supported (flag in
  85          * export) but disabled through procfs (flag in NS).
  86          *
  87          * This distinguishes from a case when ELC is not supported originally,
  88          * when we still want to cancel locks in advance and just cancel them
  89          * locally, without sending any RPC. */
  90         if (exp_connect_cancelset(exp) && !ns_connect_cancelset(ns))
  91                 RETURN(0);
  92
  93         fid_build_reg_res_name(fid, &res_id);
  94         res = ldlm_resource_get(exp->exp_obd->obd_namespace,
  95                                 NULL, &res_id, 0, 0);
  96         if (res == NULL)
  97                 RETURN(0);
  98         LDLM_RESOURCE_ADDREF(res);
  99         /* Initialize ibits lock policy. */
 100         policy.l_inodebits.bits = bits;
 101         count = ldlm_cancel_resource_local(res, cancels, &policy,
 102                                            mode, 0, 0, NULL);
 103         LDLM_RESOURCE_DELREF(res);
 104         ldlm_resource_putref(res);
 105         RETURN(count);
 106 }
 107
 108 static int mdc_prep_elc_req(struct obd_export *exp, struct ptlrpc_request *req,
 109                             cfs_list_t *cancels, int count)
 110 {
 111         return ldlm_prep_elc_req(exp, req, LUSTRE_MDS_VERSION, MDS_REINT,
 112                                  0, cancels, count);
 113 }
 114
 115 int mdc_setattr(struct obd_export *exp, struct md_op_data *op_data,
 116                 void *ea, int ealen, void *ea2, int ea2len,
 117                 struct ptlrpc_request **request, struct md_open_data **mod)
 118 {
 119         CFS_LIST_HEAD(cancels);
 120         struct ptlrpc_request *req;
 121         struct mdc_rpc_lock *rpc_lock;
 122         struct obd_device *obd = exp->exp_obd;
 123         int count = 0, rc;
 124         __u64 bits;
 125         ENTRY;
 126
 127         LASSERT(op_data != NULL);
 128
 129         bits = MDS_INODELOCK_UPDATE;
 130         if (op_data->op_attr.ia_valid & (ATTR_MODE|ATTR_UID|ATTR_GID))
 131                 bits |= MDS_INODELOCK_LOOKUP;
 132         if ((op_data->op_flags & MF_MDC_CANCEL_FID1) &&
 133             (fid_is_sane(&op_data->op_fid1)) &&
 134             !OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET))
 135                 count = mdc_resource_get_unused(exp, &op_data->op_fid1,
 136                                                 &cancels, LCK_EX, bits);
 137         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 138                                    &RQF_MDS_REINT_SETATTR);
 139         if (req == NULL) {
 140                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
 141                 RETURN(-ENOMEM);
 142         }
 143         mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1);
 144         if ((op_data->op_flags & (MF_SOM_CHANGE | MF_EPOCH_OPEN)) == 0)
 145                 req_capsule_set_size(&req->rq_pill, &RMF_MDT_EPOCH, RCL_CLIENT,
 146                                      0);
 147         req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT, ealen);
 148         req_capsule_set_size(&req->rq_pill, &RMF_LOGCOOKIES, RCL_CLIENT,
 149                              ea2len);
 150
 151         rc = mdc_prep_elc_req(exp, req, &cancels, count);
 152         if (rc) {
 153                 ptlrpc_request_free(req);
 154                 RETURN(rc);
 155         }
 156
 157         rpc_lock = obd->u.cli.cl_rpc_lock;
 158
 159         if (op_data->op_attr.ia_valid & (ATTR_MTIME | ATTR_CTIME))
 160                 CDEBUG(D_INODE, "setting mtime "CFS_TIME_T
 161                        ", ctime "CFS_TIME_T"\n",
 162                        LTIME_S(op_data->op_attr.ia_mtime),
 163                        LTIME_S(op_data->op_attr.ia_ctime));
 164         mdc_setattr_pack(req, op_data, ea, ealen, ea2, ea2len);
 165
 166         ptlrpc_request_set_replen(req);
 167         if (mod && (op_data->op_flags & MF_EPOCH_OPEN) &&
 168             req->rq_import->imp_replayable)
 169         {
 170                 LASSERT(*mod == NULL);
 171
 172                 *mod = obd_mod_alloc();
 173                 if (*mod == NULL) {
 174                         DEBUG_REQ(D_ERROR, req, "Can't allocate "
 175                                   "md_open_data");
 176                 } else {
 177                         req->rq_replay = 1;
 178                         req->rq_cb_data = *mod;
 179                         (*mod)->mod_open_req = req;
 180                         req->rq_commit_cb = mdc_commit_open;
 181                         /**
 182                          * Take an extra reference on \var mod, it protects \var
 183                          * mod from being freed on eviction (commit callback is
 184                          * called despite rq_replay flag).
 185                          * Will be put on mdc_done_writing().
 186                          */
 187                         obd_mod_get(*mod);
 188                 }
 189         }
 190
 191         rc = mdc_reint(req, rpc_lock, LUSTRE_IMP_FULL);
 192
 193         /* Save the obtained info in the original RPC for the replay case. */
 194         if (rc == 0 && (op_data->op_flags & MF_EPOCH_OPEN)) {
 195                 struct mdt_ioepoch *epoch;
 196                 struct mdt_body  *body;
 197
 198                 epoch = req_capsule_client_get(&req->rq_pill, &RMF_MDT_EPOCH);
 199                 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 200                 LASSERT(epoch != NULL);
 201                 LASSERT(body != NULL);
 202                 epoch->handle = body->handle;
 203                 epoch->ioepoch = body->ioepoch;
 204                 req->rq_replay_cb = mdc_replay_open;
 205         /** bug 3633, open may be committed and estale answer is not error */
 206         } else if (rc == -ESTALE && (op_data->op_flags & MF_SOM_CHANGE)) {
 207                 rc = 0;
 208         } else if (rc == -ERESTARTSYS) {
 209                 rc = 0;
 210         }
 211         *request = req;
 212         if (rc && req->rq_commit_cb) {
 213                 /* Put an extra reference on \var mod on error case. */
 214                 obd_mod_put(*mod);
 215                 req->rq_commit_cb(req);
 216         }
 217         RETURN(rc);
 218 }
 219
 220 int mdc_create(struct obd_export *exp, struct md_op_data *op_data,
 221                const void *data, int datalen, int mode, __u32 uid, __u32 gid,
 222                cfs_cap_t cap_effective, __u64 rdev,
 223                struct ptlrpc_request **request)
 224 {
 225         struct ptlrpc_request *req;
 226         int level, rc;
 227         int count, resends = 0;
 228         struct obd_import *import = exp->exp_obd->u.cli.cl_import;
 229         int generation = import->imp_generation;
 230         CFS_LIST_HEAD(cancels);
 231         ENTRY;
 232
 233         /* For case if upper layer did not alloc fid, do it now. */
 234         if (!fid_is_sane(&op_data->op_fid2)) {
 235                 /*
 236                  * mdc_fid_alloc() may return errno 1 in case of switch to new
 237                  * sequence, handle this.
 238                  */
 239                 rc = mdc_fid_alloc(exp, &op_data->op_fid2, op_data);
 240                 if (rc < 0) {
 241                         CERROR("Can't alloc new fid, rc %d\n", rc);
 242                         RETURN(rc);
 243                 }
 244         }
 245
 246 rebuild:
 247         count = 0;
 248         if ((op_data->op_flags & MF_MDC_CANCEL_FID1) &&
 249             (fid_is_sane(&op_data->op_fid1)))
 250                 count = mdc_resource_get_unused(exp, &op_data->op_fid1,
 251                                                 &cancels, LCK_EX,
 252                                                 MDS_INODELOCK_UPDATE);
 253
 254         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 255                                    &RQF_MDS_REINT_CREATE_RMT_ACL);
 256         if (req == NULL) {
 257                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
 258                 RETURN(-ENOMEM);
 259         }
 260         mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1);
 261         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 262                              op_data->op_namelen + 1);
 263         req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT,
 264                              data && datalen ? datalen : 0);
 265
 266         rc = mdc_prep_elc_req(exp, req, &cancels, count);
 267         if (rc) {
 268                 ptlrpc_request_free(req);
 269                 RETURN(rc);
 270         }
 271
 272         /*
 273          * mdc_create_pack() fills msg->bufs[1] with name and msg->bufs[2] with
 274          * tgt, for symlinks or lov MD data.
 275          */
 276         mdc_create_pack(req, op_data, data, datalen, mode, uid,
 277                         gid, cap_effective, rdev);
 278
 279         ptlrpc_request_set_replen(req);
 280
 281         if (resends) {
 282                 req->rq_generation_set = 1;
 283                 req->rq_import_generation = generation;
 284                 req->rq_sent = cfs_time_current_sec() + resends;
 285         }
 286         level = LUSTRE_IMP_FULL;
 287  resend:
 288         rc = mdc_reint(req, exp->exp_obd->u.cli.cl_rpc_lock, level);
 289
 290         /* Resend if we were told to. */
 291         if (rc == -ERESTARTSYS) {
 292                 level = LUSTRE_IMP_RECOVER;
 293                 goto resend;
 294         } else if (rc == -EINPROGRESS) {
 295                 /* Retry create infinitely until succeed or get other
 296                  * error code. */
 297                 ptlrpc_req_finished(req);
 298                 resends++;
 299
 300                 CDEBUG(D_HA, "%s: resend:%d create on "DFID"/"DFID"\n",
 301                        exp->exp_obd->obd_name, resends,
 302                        PFID(&op_data->op_fid1), PFID(&op_data->op_fid2));
 303
 304                 if (generation == import->imp_generation) {
 305                         goto rebuild;
 306                 } else {
 307                         CDEBUG(D_HA, "resend cross eviction\n");
 308                         RETURN(-EIO);
 309                 }
 310         } else if (rc == 0) {
 311                 struct mdt_body *body;
 312                 struct lustre_capa *capa;
 313
 314                 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
 315                 LASSERT(body);
 316                 if (body->valid & OBD_MD_FLMDSCAPA) {
 317                         capa = req_capsule_server_get(&req->rq_pill,
 318                                                       &RMF_CAPA1);
 319                         if (capa == NULL)
 320                                 rc = -EPROTO;
 321                 }
 322         }
 323
 324         *request = req;
 325         RETURN(rc);
 326 }
 327
 328 int mdc_unlink(struct obd_export *exp, struct md_op_data *op_data,
 329                struct ptlrpc_request **request)
 330 {
 331         CFS_LIST_HEAD(cancels);
 332         struct obd_device *obd = class_exp2obd(exp);
 333         struct ptlrpc_request *req = *request;
 334         int count = 0, rc;
 335         ENTRY;
 336
 337         LASSERT(req == NULL);
 338
 339         if ((op_data->op_flags & MF_MDC_CANCEL_FID1) &&
 340             (fid_is_sane(&op_data->op_fid1)) &&
 341             !OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET))
 342                 count = mdc_resource_get_unused(exp, &op_data->op_fid1,
 343                                                 &cancels, LCK_EX,
 344                                                 MDS_INODELOCK_UPDATE);
 345         if ((op_data->op_flags & MF_MDC_CANCEL_FID3) &&
 346             (fid_is_sane(&op_data->op_fid3)) &&
 347             !OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET))
 348                 count += mdc_resource_get_unused(exp, &op_data->op_fid3,
 349                                                  &cancels, LCK_EX,
 350                                                  MDS_INODELOCK_FULL);
 351         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 352                                    &RQF_MDS_REINT_UNLINK);
 353         if (req == NULL) {
 354                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
 355                 RETURN(-ENOMEM);
 356         }
 357         mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1);
 358         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 359                              op_data->op_namelen + 1);
 360
 361         rc = mdc_prep_elc_req(exp, req, &cancels, count);
 362         if (rc) {
 363                 ptlrpc_request_free(req);
 364                 RETURN(rc);
 365         }
 366
 367         mdc_unlink_pack(req, op_data);
 368
 369         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 370                              obd->u.cli.cl_max_mds_easize);
 371         req_capsule_set_size(&req->rq_pill, &RMF_LOGCOOKIES, RCL_SERVER,
 372                              obd->u.cli.cl_max_mds_cookiesize);
 373         ptlrpc_request_set_replen(req);
 374
 375         *request = req;
 376
 377         rc = mdc_reint(req, obd->u.cli.cl_rpc_lock, LUSTRE_IMP_FULL);
 378         if (rc == -ERESTARTSYS)
 379                 rc = 0;
 380         RETURN(rc);
 381 }
 382
 383 int mdc_link(struct obd_export *exp, struct md_op_data *op_data,
 384              struct ptlrpc_request **request)
 385 {
 386         CFS_LIST_HEAD(cancels);
 387         struct obd_device *obd = exp->exp_obd;
 388         struct ptlrpc_request *req;
 389         int count = 0, rc;
 390         ENTRY;
 391
 392         if ((op_data->op_flags & MF_MDC_CANCEL_FID2) &&
 393             (fid_is_sane(&op_data->op_fid2)))
 394                 count = mdc_resource_get_unused(exp, &op_data->op_fid2,
 395                                                 &cancels, LCK_EX,
 396                                                 MDS_INODELOCK_UPDATE);
 397         if ((op_data->op_flags & MF_MDC_CANCEL_FID1) &&
 398             (fid_is_sane(&op_data->op_fid1)))
 399                 count += mdc_resource_get_unused(exp, &op_data->op_fid1,
 400                                                  &cancels, LCK_EX,
 401                                                  MDS_INODELOCK_UPDATE);
 402
 403         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_REINT_LINK);
 404         if (req == NULL) {
 405                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
 406                 RETURN(-ENOMEM);
 407         }
 408         mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1);
 409         mdc_set_capa_size(req, &RMF_CAPA2, op_data->op_capa2);
 410         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
 411                              op_data->op_namelen + 1);
 412
 413         rc = mdc_prep_elc_req(exp, req, &cancels, count);
 414         if (rc) {
 415                 ptlrpc_request_free(req);
 416                 RETURN(rc);
 417         }
 418
 419         mdc_link_pack(req, op_data);
 420         ptlrpc_request_set_replen(req);
 421
 422         rc = mdc_reint(req, obd->u.cli.cl_rpc_lock, LUSTRE_IMP_FULL);
 423         *request = req;
 424         if (rc == -ERESTARTSYS)
 425                 rc = 0;
 426
 427         RETURN(rc);
 428 }
 429
 430 int mdc_rename(struct obd_export *exp, struct md_op_data *op_data,
 431                const char *old, int oldlen, const char *new, int newlen,
 432                struct ptlrpc_request **request)
 433 {
 434         CFS_LIST_HEAD(cancels);
 435         struct obd_device *obd = exp->exp_obd;
 436         struct ptlrpc_request *req;
 437         int count = 0, rc;
 438         ENTRY;
 439
 440         if ((op_data->op_flags & MF_MDC_CANCEL_FID1) &&
 441             (fid_is_sane(&op_data->op_fid1)))
 442                 count = mdc_resource_get_unused(exp, &op_data->op_fid1,
 443                                                 &cancels, LCK_EX,
 444                                                 MDS_INODELOCK_UPDATE);
 445         if ((op_data->op_flags & MF_MDC_CANCEL_FID2) &&
 446             (fid_is_sane(&op_data->op_fid2)))
 447                 count += mdc_resource_get_unused(exp, &op_data->op_fid2,
 448                                                  &cancels, LCK_EX,
 449                                                  MDS_INODELOCK_UPDATE);
 450         if ((op_data->op_flags & MF_MDC_CANCEL_FID3) &&
 451             (fid_is_sane(&op_data->op_fid3)))
 452                 count += mdc_resource_get_unused(exp, &op_data->op_fid3,
 453                                                  &cancels, LCK_EX,
 454                                                  MDS_INODELOCK_LOOKUP);
 455         if ((op_data->op_flags & MF_MDC_CANCEL_FID4) &&
 456              (fid_is_sane(&op_data->op_fid4)))
 457                 count += mdc_resource_get_unused(exp, &op_data->op_fid4,
 458                                                  &cancels, LCK_EX,
 459                                                  MDS_INODELOCK_FULL);
 460
 461         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
 462                                    &RQF_MDS_REINT_RENAME);
 463         if (req == NULL) {
 464                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
 465                 RETURN(-ENOMEM);
 466         }
 467
 468         mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1);
 469         mdc_set_capa_size(req, &RMF_CAPA2, op_data->op_capa2);
 470         req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT, oldlen + 1);
 471         req_capsule_set_size(&req->rq_pill, &RMF_SYMTGT, RCL_CLIENT, newlen+1);
 472
 473         rc = mdc_prep_elc_req(exp, req, &cancels, count);
 474         if (rc) {
 475                 ptlrpc_request_free(req);
 476                 RETURN(rc);
 477         }
 478
 479         if (exp_connect_cancelset(exp) && req)
 480                 ldlm_cli_cancel_list(&cancels, count, req, 0);
 481
 482         mdc_rename_pack(req, op_data, old, oldlen, new, newlen);
 483
 484         req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
 485                              obd->u.cli.cl_max_mds_easize);
 486         req_capsule_set_size(&req->rq_pill, &RMF_LOGCOOKIES, RCL_SERVER,
 487                              obd->u.cli.cl_max_mds_cookiesize);
 488         ptlrpc_request_set_replen(req);
 489
 490         rc = mdc_reint(req, obd->u.cli.cl_rpc_lock, LUSTRE_IMP_FULL);
 491         *request = req;
 492         if (rc == -ERESTARTSYS)
 493                 rc = 0;
 494
 495         RETURN(rc);
 496 }