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