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