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