lustre/ptlrpc/nodemap_storage.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) 2015, Trustees of Indiana University
  24  *
  25  * Copyright (c) 2017, Intel Corporation.
  26  *
  27  * Author: Joshua Walgenbach <jjw@iu.edu>
  28  * Author: Kit Westneat <cwestnea@iu.edu>
  29  *
  30  * Implements the storage functionality for the nodemap configuration. Functions
  31  * in this file prepare, store, and load nodemap configuration data. Targets
  32  * using nodemap services should register a configuration file object. Nodemap
  33  * configuration changes that need to persist should call the appropriate
  34  * storage function for the data being modified.
  35  *
  36  * There are several index types as defined in enum nodemap_idx_type:
  37  *      NODEMAP_CLUSTER_IDX     stores the data found on the lu_nodemap struct,
  38  *                              like root squash and config flags, as well as
  39  *                              the name.
  40  *      NODEMAP_RANGE_IDX       stores NID range information for a nodemap
  41  *      NODEMAP_UIDMAP_IDX      stores a fs/client UID mapping pair
  42  *      NODEMAP_GIDMAP_IDX      stores a fs/client GID mapping pair
  43  *      NODEMAP_GLOBAL_IDX      stores whether or not nodemaps are active
  44  */
  45
  46 #include <libcfs/libcfs.h>
  47 #include <linux/err.h>
  48 #include <linux/kernel.h>
  49 #include <linux/list.h>
  50 #include <linux/mutex.h>
  51 #include <linux/string.h>
  52 #include <linux/types.h>
  53 #include <uapi/linux/lnet/lnet-types.h>
  54 #include <uapi/linux/lustre/lustre_idl.h>
  55 #include <uapi/linux/lustre/lustre_disk.h>
  56 #include <dt_object.h>
  57 #include <lu_object.h>
  58 #include <lustre_net.h>
  59 #include <lustre_nodemap.h>
  60 #include <obd_class.h>
  61 #include <obd_support.h>
  62 #include "nodemap_internal.h"
  63
  64 /* list of registered nodemap index files, except MGS */
  65 static LIST_HEAD(ncf_list_head);
  66 static DEFINE_MUTEX(ncf_list_lock);
  67
  68 /* MGS index is different than others, others are listeners to MGS idx */
  69 static struct nm_config_file *nodemap_mgs_ncf;
  70
  71 static void nodemap_cluster_key_init(struct nodemap_key *nk, unsigned int nm_id,
  72                                      enum nodemap_cluster_rec_subid subid)
  73 {
  74         nk->nk_nodemap_id = cpu_to_le32(nm_idx_set_type(nm_id,
  75                                                         NODEMAP_CLUSTER_IDX));
  76         nk->nk_cluster_subid = subid;
  77 }
  78
  79 static void nodemap_cluster_rec_init(union nodemap_rec *nr,
  80                                      const struct lu_nodemap *nodemap)
  81 {
  82         BUILD_BUG_ON(sizeof(nr->ncr.ncr_name) != sizeof(nodemap->nm_name));
  83
  84         strncpy(nr->ncr.ncr_name, nodemap->nm_name, sizeof(nr->ncr.ncr_name));
  85         nr->ncr.ncr_squash_uid = cpu_to_le32(nodemap->nm_squash_uid);
  86         nr->ncr.ncr_squash_gid = cpu_to_le32(nodemap->nm_squash_gid);
  87         nr->ncr.ncr_squash_projid = cpu_to_le32(nodemap->nm_squash_projid);
  88         nr->ncr.ncr_flags =
  89                 (nodemap->nmf_trust_client_ids ?
  90                         NM_FL_TRUST_CLIENT_IDS : 0) |
  91                 (nodemap->nmf_allow_root_access ?
  92                         NM_FL_ALLOW_ROOT_ACCESS : 0) |
  93                 (nodemap->nmf_deny_unknown ?
  94                         NM_FL_DENY_UNKNOWN : 0) |
  95                 (nodemap->nmf_map_mode & NODEMAP_MAP_UID ?
  96                         NM_FL_MAP_UID : 0) |
  97                 (nodemap->nmf_map_mode & NODEMAP_MAP_GID ?
  98                         NM_FL_MAP_GID : 0) |
  99                 (nodemap->nmf_map_mode & NODEMAP_MAP_PROJID ?
 100                         NM_FL_MAP_PROJID : 0) |
 101                 (nodemap->nmf_enable_audit ?
 102                         NM_FL_ENABLE_AUDIT : 0) |
 103                 (nodemap->nmf_forbid_encryption ?
 104                         NM_FL_FORBID_ENCRYPT : 0);
 105         nr->ncr.ncr_flags2 =
 106                 (nodemap->nmf_readonly_mount ?
 107                         NM_FL2_READONLY_MOUNT : 0);
 108 }
 109
 110 static void nodemap_cluster_roles_rec_init(union nodemap_rec *nr,
 111                                            const struct lu_nodemap *nodemap)
 112 {
 113         struct nodemap_cluster_roles_rec *ncrr = &nr->ncrr;
 114
 115         memset(ncrr, 0, sizeof(struct nodemap_cluster_roles_rec));
 116         ncrr->ncrr_roles = cpu_to_le64(nodemap->nmf_rbac);
 117 }
 118
 119 static void nodemap_idmap_key_init(struct nodemap_key *nk, unsigned int nm_id,
 120                                    enum nodemap_id_type id_type,
 121                                    u32 id_client)
 122 {
 123         enum nodemap_idx_type idx_type;
 124
 125         if (id_type == NODEMAP_UID)
 126                 idx_type = NODEMAP_UIDMAP_IDX;
 127         else if (id_type == NODEMAP_GID)
 128                 idx_type = NODEMAP_GIDMAP_IDX;
 129         else if (id_type == NODEMAP_PROJID)
 130                 idx_type = NODEMAP_PROJIDMAP_IDX;
 131         else
 132                 idx_type = NODEMAP_EMPTY_IDX;
 133
 134         nk->nk_nodemap_id = cpu_to_le32(nm_idx_set_type(nm_id, idx_type));
 135         nk->nk_id_client = cpu_to_le32(id_client);
 136 }
 137
 138 static void nodemap_idmap_rec_init(union nodemap_rec *nr, u32 id_fs)
 139 {
 140         nr->nir.nir_id_fs = cpu_to_le32(id_fs);
 141 }
 142
 143 static void nodemap_range_key_init(struct nodemap_key *nk, unsigned int nm_id,
 144                                    unsigned int rn_id)
 145 {
 146         nk->nk_nodemap_id = cpu_to_le32(nm_idx_set_type(nm_id,
 147                                                         NODEMAP_RANGE_IDX));
 148         nk->nk_range_id = cpu_to_le32(rn_id);
 149 }
 150
 151 static void nodemap_range_rec_init(union nodemap_rec *nr,
 152                                    const lnet_nid_t nid[2])
 153 {
 154         nr->nrr.nrr_start_nid = cpu_to_le64(nid[0]);
 155         nr->nrr.nrr_end_nid = cpu_to_le64(nid[1]);
 156 }
 157
 158 static void nodemap_global_key_init(struct nodemap_key *nk)
 159 {
 160         nk->nk_nodemap_id = cpu_to_le32(nm_idx_set_type(0, NODEMAP_GLOBAL_IDX));
 161         nk->nk_unused = 0;
 162 }
 163
 164 static void nodemap_global_rec_init(union nodemap_rec *nr, bool active)
 165 {
 166         nr->ngr.ngr_is_active = active;
 167 }
 168
 169 /* should be called with dt_write lock */
 170 static void nodemap_inc_version(const struct lu_env *env,
 171                                 struct dt_object *nodemap_idx,
 172                                 struct thandle *th)
 173 {
 174         u64 ver = dt_version_get(env, nodemap_idx);
 175         dt_version_set(env, nodemap_idx, ver + 1, th);
 176 }
 177
 178 enum ncfc_find_create {
 179         NCFC_CREATE_NEW = 1,
 180 };
 181
 182 static struct dt_object *nodemap_cache_find_create(const struct lu_env *env,
 183                                                    struct dt_device *dev,
 184                                                    struct local_oid_storage *los,
 185                                                    enum ncfc_find_create create_new)
 186 {
 187         struct lu_fid tfid;
 188         struct dt_object *root_obj;
 189         struct dt_object *nm_obj;
 190         int rc = 0;
 191
 192         rc = dt_root_get(env, dev, &tfid);
 193         if (rc < 0)
 194                 GOTO(out, nm_obj = ERR_PTR(rc));
 195
 196         root_obj = dt_locate(env, dev, &tfid);
 197         if (unlikely(IS_ERR(root_obj)))
 198                 GOTO(out, nm_obj = root_obj);
 199
 200         rc = dt_lookup_dir(env, root_obj, LUSTRE_NODEMAP_NAME, &tfid);
 201         if (rc == -ENOENT) {
 202                 if (dev->dd_rdonly)
 203                         GOTO(out_root, nm_obj = ERR_PTR(-EROFS));
 204         } else if (rc) {
 205                 GOTO(out_root, nm_obj = ERR_PTR(rc));
 206         } else if (dev->dd_rdonly && create_new == NCFC_CREATE_NEW) {
 207                 GOTO(out_root, nm_obj = ERR_PTR(-EROFS));
 208         }
 209
 210 again:
 211         /* if loading index fails the first time, create new index */
 212         if (create_new == NCFC_CREATE_NEW && rc != -ENOENT) {
 213                 CDEBUG(D_INFO, "removing old index, creating new one\n");
 214                 rc = local_object_unlink(env, dev, root_obj,
 215                                          LUSTRE_NODEMAP_NAME);
 216                 if (rc < 0) {
 217                         /* XXX not sure the best way to get obd name. */
 218                         CERROR("cannot destroy nodemap index: rc = %d\n",
 219                                rc);
 220                         GOTO(out_root, nm_obj = ERR_PTR(rc));
 221                 }
 222         }
 223
 224         nm_obj = local_index_find_or_create(env, los, root_obj,
 225                                                 LUSTRE_NODEMAP_NAME,
 226                                                 S_IFREG | S_IRUGO | S_IWUSR,
 227                                                 &dt_nodemap_features);
 228         if (IS_ERR(nm_obj))
 229                 GOTO(out_root, nm_obj);
 230
 231         if (nm_obj->do_index_ops == NULL) {
 232                 rc = nm_obj->do_ops->do_index_try(env, nm_obj,
 233                                                       &dt_nodemap_features);
 234                 /* even if loading from tgt fails, connecting to MGS will
 235                  * rewrite the config
 236                  */
 237                 if (rc < 0) {
 238                         dt_object_put(env, nm_obj);
 239
 240                         if (create_new == NCFC_CREATE_NEW)
 241                                 GOTO(out_root, nm_obj = ERR_PTR(rc));
 242
 243                         CERROR("cannot load nodemap index from disk, creating "
 244                                "new index: rc = %d\n", rc);
 245                         create_new = NCFC_CREATE_NEW;
 246                         goto again;
 247                 }
 248         }
 249
 250 out_root:
 251         dt_object_put(env, root_obj);
 252 out:
 253         return nm_obj;
 254 }
 255
 256 static int nodemap_idx_insert(const struct lu_env *env,
 257                               struct dt_object *idx,
 258                               const struct nodemap_key *nk,
 259                               const union nodemap_rec *nr)
 260 {
 261         struct thandle *th;
 262         struct dt_device *dev = lu2dt_dev(idx->do_lu.lo_dev);
 263         int rc;
 264
 265         BUILD_BUG_ON(sizeof(union nodemap_rec) != 32);
 266
 267         th = dt_trans_create(env, dev);
 268
 269         if (IS_ERR(th))
 270                 GOTO(out, rc = PTR_ERR(th));
 271
 272         rc = dt_declare_insert(env, idx,
 273                                (const struct dt_rec *)nr,
 274                                (const struct dt_key *)nk, th);
 275         if (rc != 0)
 276                 GOTO(out, rc);
 277
 278         rc = dt_declare_version_set(env, idx, th);
 279         if (rc != 0)
 280                 GOTO(out, rc);
 281
 282         rc = dt_trans_start_local(env, dev, th);
 283         if (rc != 0)
 284                 GOTO(out, rc);
 285
 286         dt_write_lock(env, idx, 0);
 287
 288         rc = dt_insert(env, idx, (const struct dt_rec *)nr,
 289                        (const struct dt_key *)nk, th);
 290
 291         nodemap_inc_version(env, idx, th);
 292         dt_write_unlock(env, idx);
 293 out:
 294         dt_trans_stop(env, dev, th);
 295
 296         return rc;
 297 }
 298
 299 static int nodemap_idx_update(const struct lu_env *env,
 300                               struct dt_object *idx,
 301                               const struct nodemap_key *nk,
 302                               const union nodemap_rec *nr)
 303 {
 304         struct thandle          *th;
 305         struct dt_device        *dev = lu2dt_dev(idx->do_lu.lo_dev);
 306         int                      rc = 0;
 307
 308         th = dt_trans_create(env, dev);
 309
 310         if (IS_ERR(th))
 311                 GOTO(out, rc = PTR_ERR(th));
 312
 313         rc = dt_declare_delete(env, idx, (const struct dt_key *)nk, th);
 314         if (rc != 0)
 315                 GOTO(out, rc);
 316
 317         rc = dt_declare_insert(env, idx, (const struct dt_rec *)nr,
 318                                (const struct dt_key *)nk, th);
 319         if (rc != 0)
 320                 GOTO(out, rc);
 321
 322         rc = dt_declare_version_set(env, idx, th);
 323         if (rc != 0)
 324                 GOTO(out, rc);
 325
 326         rc = dt_trans_start_local(env, dev, th);
 327         if (rc != 0)
 328                 GOTO(out, rc);
 329
 330         dt_write_lock(env, idx, 0);
 331
 332         rc = dt_delete(env, idx, (const struct dt_key *)nk, th);
 333         if (rc != 0)
 334                 GOTO(out_lock, rc);
 335
 336         rc = dt_insert(env, idx, (const struct dt_rec *)nr,
 337                        (const struct dt_key *)nk, th);
 338         if (rc != 0)
 339                 GOTO(out_lock, rc);
 340
 341         nodemap_inc_version(env, idx, th);
 342 out_lock:
 343         dt_write_unlock(env, idx);
 344 out:
 345         dt_trans_stop(env, dev, th);
 346
 347         return rc;
 348 }
 349
 350 static int nodemap_idx_delete(const struct lu_env *env,
 351                               struct dt_object *idx,
 352                               const struct nodemap_key *nk,
 353                               const union nodemap_rec *unused)
 354 {
 355         struct thandle          *th;
 356         struct dt_device        *dev = lu2dt_dev(idx->do_lu.lo_dev);
 357         int                      rc = 0;
 358
 359         th = dt_trans_create(env, dev);
 360
 361         if (IS_ERR(th))
 362                 GOTO(out, rc = PTR_ERR(th));
 363
 364         rc = dt_declare_delete(env, idx, (const struct dt_key *)nk, th);
 365         if (rc != 0)
 366                 GOTO(out, rc);
 367
 368         rc = dt_declare_version_set(env, idx, th);
 369         if (rc != 0)
 370                 GOTO(out, rc);
 371
 372         rc = dt_trans_start_local(env, dev, th);
 373         if (rc != 0)
 374                 GOTO(out, rc);
 375
 376         dt_write_lock(env, idx, 0);
 377
 378         rc = dt_delete(env, idx, (const struct dt_key *)nk, th);
 379
 380         nodemap_inc_version(env, idx, th);
 381
 382         dt_write_unlock(env, idx);
 383 out:
 384         dt_trans_stop(env, dev, th);
 385
 386         return rc;
 387 }
 388
 389 enum nm_add_update {
 390         NM_ADD = 0,
 391         NM_UPDATE = 1,
 392 };
 393
 394 static int nodemap_idx_cluster_add_update(const struct lu_nodemap *nodemap,
 395                                           struct dt_object *idx,
 396                                           enum nm_add_update update,
 397                                           enum nodemap_cluster_rec_subid subid)
 398 {
 399         struct nodemap_key nk;
 400         union nodemap_rec nr;
 401         struct lu_env env;
 402         int rc = 0;
 403
 404         ENTRY;
 405
 406         if (idx == NULL) {
 407                 if (nodemap_mgs_ncf == NULL) {
 408                         CERROR("cannot add nodemap config to non-existing MGS.\n");
 409                         return -EINVAL;
 410                 }
 411                 idx = nodemap_mgs_ncf->ncf_obj;
 412         }
 413
 414         rc = lu_env_init(&env, LCT_LOCAL);
 415         if (rc)
 416                 RETURN(rc);
 417
 418         nodemap_cluster_key_init(&nk, nodemap->nm_id, subid);
 419         switch (subid) {
 420         case NODEMAP_CLUSTER_REC:
 421                 nodemap_cluster_rec_init(&nr, nodemap);
 422                 break;
 423         case NODEMAP_CLUSTER_ROLES:
 424                 nodemap_cluster_roles_rec_init(&nr, nodemap);
 425                 break;
 426         default:
 427                 CWARN("%s: unknown subtype %u\n", nodemap->nm_name, subid);
 428                 GOTO(fini, rc = -EINVAL);
 429         }
 430
 431         if (update == NM_UPDATE)
 432                 rc = nodemap_idx_update(&env, idx, &nk, &nr);
 433         else
 434                 rc = nodemap_idx_insert(&env, idx, &nk, &nr);
 435
 436 fini:
 437         lu_env_fini(&env);
 438         RETURN(rc);
 439 }
 440
 441 int nodemap_idx_nodemap_add(const struct lu_nodemap *nodemap)
 442 {
 443         return nodemap_idx_cluster_add_update(nodemap, NULL,
 444                                               NM_ADD, NODEMAP_CLUSTER_REC);
 445 }
 446
 447 int nodemap_idx_nodemap_update(const struct lu_nodemap *nodemap)
 448 {
 449         return nodemap_idx_cluster_add_update(nodemap, NULL,
 450                                               NM_UPDATE, NODEMAP_CLUSTER_REC);
 451 }
 452
 453 int nodemap_idx_nodemap_del(const struct lu_nodemap *nodemap)
 454 {
 455         struct rb_root           root;
 456         struct lu_idmap         *idmap;
 457         struct lu_idmap         *temp;
 458         struct lu_nid_range     *range;
 459         struct lu_nid_range     *range_temp;
 460         struct nodemap_key       nk;
 461         struct lu_env            env;
 462         int                      rc = 0;
 463         int                      rc2 = 0;
 464
 465         ENTRY;
 466
 467         if (nodemap_mgs_ncf == NULL) {
 468                 CERROR("cannot add nodemap config to non-existing MGS.\n");
 469                 return -EINVAL;
 470         }
 471
 472         rc = lu_env_init(&env, LCT_LOCAL);
 473         if (rc != 0)
 474                 RETURN(rc);
 475
 476         nodemap_cluster_key_init(&nk, nodemap->nm_id, NODEMAP_CLUSTER_ROLES);
 477         rc2 = nodemap_idx_delete(&env, nodemap_mgs_ncf->ncf_obj, &nk, NULL);
 478         if (rc2 < 0 && rc2 != -ENOENT)
 479                 rc = rc2;
 480
 481         root = nodemap->nm_fs_to_client_uidmap;
 482         nm_rbtree_postorder_for_each_entry_safe(idmap, temp, &root,
 483                                                 id_fs_to_client) {
 484                 nodemap_idmap_key_init(&nk, nodemap->nm_id, NODEMAP_UID,
 485                                        idmap->id_client);
 486                 rc2 = nodemap_idx_delete(&env, nodemap_mgs_ncf->ncf_obj,
 487                                          &nk, NULL);
 488                 if (rc2 < 0)
 489                         rc = rc2;
 490         }
 491
 492         root = nodemap->nm_client_to_fs_gidmap;
 493         nm_rbtree_postorder_for_each_entry_safe(idmap, temp, &root,
 494                                                 id_client_to_fs) {
 495                 nodemap_idmap_key_init(&nk, nodemap->nm_id, NODEMAP_GID,
 496                                        idmap->id_client);
 497                 rc2 = nodemap_idx_delete(&env, nodemap_mgs_ncf->ncf_obj,
 498                                          &nk, NULL);
 499                 if (rc2 < 0)
 500                         rc = rc2;
 501         }
 502
 503         root = nodemap->nm_client_to_fs_projidmap;
 504         nm_rbtree_postorder_for_each_entry_safe(idmap, temp, &root,
 505                                                 id_client_to_fs) {
 506                 nodemap_idmap_key_init(&nk, nodemap->nm_id, NODEMAP_PROJID,
 507                                        idmap->id_client);
 508                 rc2 = nodemap_idx_delete(&env, nodemap_mgs_ncf->ncf_obj,
 509                                          &nk, NULL);
 510                 if (rc2 < 0)
 511                         rc = rc2;
 512         }
 513
 514         list_for_each_entry_safe(range, range_temp, &nodemap->nm_ranges,
 515                                  rn_list) {
 516                 nodemap_range_key_init(&nk, nodemap->nm_id, range->rn_id);
 517                 rc2 = nodemap_idx_delete(&env, nodemap_mgs_ncf->ncf_obj,
 518                                          &nk, NULL);
 519                 if (rc2 < 0)
 520                         rc = rc2;
 521         }
 522
 523         nodemap_cluster_key_init(&nk, nodemap->nm_id, NODEMAP_CLUSTER_REC);
 524         rc2 = nodemap_idx_delete(&env, nodemap_mgs_ncf->ncf_obj, &nk, NULL);
 525         if (rc2 < 0)
 526                 rc = rc2;
 527
 528         lu_env_fini(&env);
 529
 530         RETURN(rc);
 531 }
 532
 533 int nodemap_idx_cluster_roles_add(const struct lu_nodemap *nodemap)
 534 {
 535         return nodemap_idx_cluster_add_update(nodemap, NULL, NM_ADD,
 536                                               NODEMAP_CLUSTER_ROLES);
 537 }
 538
 539 int nodemap_idx_cluster_roles_update(const struct lu_nodemap *nodemap)
 540 {
 541         return nodemap_idx_cluster_add_update(nodemap, NULL, NM_UPDATE,
 542                                               NODEMAP_CLUSTER_ROLES);
 543 }
 544
 545 int nodemap_idx_cluster_roles_del(const struct lu_nodemap *nodemap)
 546 {
 547         struct nodemap_key nk;
 548         struct lu_env env;
 549         int rc = 0;
 550
 551         ENTRY;
 552
 553         if (nodemap_mgs_ncf == NULL) {
 554                 CERROR("cannot add nodemap config to non-existing MGS.\n");
 555                 return -EINVAL;
 556         }
 557
 558         rc = lu_env_init(&env, LCT_LOCAL);
 559         if (rc != 0)
 560                 RETURN(rc);
 561
 562         nodemap_cluster_key_init(&nk, nodemap->nm_id, NODEMAP_CLUSTER_ROLES);
 563         rc = nodemap_idx_delete(&env, nodemap_mgs_ncf->ncf_obj, &nk, NULL);
 564
 565         lu_env_fini(&env);
 566         RETURN(rc);
 567 }
 568
 569 int nodemap_idx_range_add(const struct lu_nid_range *range,
 570                           const lnet_nid_t nid[2])
 571 {
 572         struct nodemap_key       nk;
 573         union nodemap_rec        nr;
 574         struct lu_env            env;
 575         int                      rc = 0;
 576         ENTRY;
 577
 578         if (nodemap_mgs_ncf == NULL) {
 579                 CERROR("cannot add nodemap config to non-existing MGS.\n");
 580                 return -EINVAL;
 581         }
 582
 583         rc = lu_env_init(&env, LCT_LOCAL);
 584         if (rc != 0)
 585                 RETURN(rc);
 586
 587         nodemap_range_key_init(&nk, range->rn_nodemap->nm_id, range->rn_id);
 588         nodemap_range_rec_init(&nr, nid);
 589
 590         rc = nodemap_idx_insert(&env, nodemap_mgs_ncf->ncf_obj, &nk, &nr);
 591         lu_env_fini(&env);
 592
 593         RETURN(rc);
 594 }
 595
 596 int nodemap_idx_range_del(const struct lu_nid_range *range)
 597 {
 598         struct nodemap_key       nk;
 599         struct lu_env            env;
 600         int                      rc = 0;
 601         ENTRY;
 602
 603         if (nodemap_mgs_ncf == NULL) {
 604                 CERROR("cannot add nodemap config to non-existing MGS.\n");
 605                 return -EINVAL;
 606         }
 607
 608         rc = lu_env_init(&env, LCT_LOCAL);
 609         if (rc != 0)
 610                 RETURN(rc);
 611
 612         nodemap_range_key_init(&nk, range->rn_nodemap->nm_id, range->rn_id);
 613
 614         rc = nodemap_idx_delete(&env, nodemap_mgs_ncf->ncf_obj, &nk, NULL);
 615         lu_env_fini(&env);
 616
 617         RETURN(rc);
 618 }
 619
 620 int nodemap_idx_idmap_add(const struct lu_nodemap *nodemap,
 621                           enum nodemap_id_type id_type,
 622                           const u32 map[2])
 623 {
 624         struct nodemap_key       nk;
 625         union nodemap_rec        nr;
 626         struct lu_env            env;
 627         int                      rc = 0;
 628         ENTRY;
 629
 630         if (nodemap_mgs_ncf == NULL) {
 631                 CERROR("cannot add nodemap config to non-existing MGS.\n");
 632                 return -EINVAL;
 633         }
 634
 635         rc = lu_env_init(&env, LCT_LOCAL);
 636         if (rc != 0)
 637                 RETURN(rc);
 638
 639         nodemap_idmap_key_init(&nk, nodemap->nm_id, id_type, map[0]);
 640         nodemap_idmap_rec_init(&nr, map[1]);
 641
 642         rc = nodemap_idx_insert(&env, nodemap_mgs_ncf->ncf_obj, &nk, &nr);
 643         lu_env_fini(&env);
 644
 645         RETURN(rc);
 646 }
 647
 648 int nodemap_idx_idmap_del(const struct lu_nodemap *nodemap,
 649                           enum nodemap_id_type id_type,
 650                           const u32 map[2])
 651 {
 652         struct nodemap_key       nk;
 653         struct lu_env            env;
 654         int                      rc = 0;
 655         ENTRY;
 656
 657         if (nodemap_mgs_ncf == NULL) {
 658                 CERROR("cannot add nodemap config to non-existing MGS.\n");
 659                 return -EINVAL;
 660         }
 661
 662         rc = lu_env_init(&env, LCT_LOCAL);
 663         if (rc != 0)
 664                 RETURN(rc);
 665
 666         nodemap_idmap_key_init(&nk, nodemap->nm_id, id_type, map[0]);
 667
 668         rc = nodemap_idx_delete(&env, nodemap_mgs_ncf->ncf_obj, &nk, NULL);
 669         lu_env_fini(&env);
 670
 671         RETURN(rc);
 672 }
 673
 674 static int nodemap_idx_global_add_update(bool value, enum nm_add_update update)
 675 {
 676         struct nodemap_key       nk;
 677         union nodemap_rec        nr;
 678         struct lu_env            env;
 679         int                      rc = 0;
 680         ENTRY;
 681
 682         if (nodemap_mgs_ncf == NULL) {
 683                 CERROR("cannot add nodemap config to non-existing MGS.\n");
 684                 return -EINVAL;
 685         }
 686
 687         rc = lu_env_init(&env, LCT_LOCAL);
 688         if (rc != 0)
 689                 RETURN(rc);
 690
 691         nodemap_global_key_init(&nk);
 692         nodemap_global_rec_init(&nr, value);
 693
 694         if (update == NM_UPDATE)
 695                 rc = nodemap_idx_update(&env, nodemap_mgs_ncf->ncf_obj,
 696                                         &nk, &nr);
 697         else
 698                 rc = nodemap_idx_insert(&env, nodemap_mgs_ncf->ncf_obj,
 699                                         &nk, &nr);
 700
 701         lu_env_fini(&env);
 702
 703         RETURN(rc);
 704 }
 705
 706 int nodemap_idx_nodemap_activate(bool value)
 707 {
 708         return nodemap_idx_global_add_update(value, NM_UPDATE);
 709 }
 710
 711 static enum nodemap_idx_type nodemap_get_key_type(const struct nodemap_key *key)
 712 {
 713         u32                      nodemap_id;
 714
 715         nodemap_id = le32_to_cpu(key->nk_nodemap_id);
 716         return nm_idx_get_type(nodemap_id);
 717 }
 718
 719 static int nodemap_get_key_subtype(const struct nodemap_key *key)
 720 {
 721         enum nodemap_idx_type type = nodemap_get_key_type(key);
 722
 723         return type == NODEMAP_CLUSTER_IDX ? key->nk_cluster_subid : -1;
 724 }
 725
 726 static int nodemap_cluster_rec_helper(struct nodemap_config *config,
 727                                       u32 nodemap_id,
 728                                       const union nodemap_rec *rec,
 729                                       struct lu_nodemap **recent_nodemap)
 730 {
 731         struct lu_nodemap *nodemap, *old_nm;
 732         enum nm_flag_bits flags;
 733         enum nm_flag2_bits flags2;
 734
 735         nodemap = cfs_hash_lookup(config->nmc_nodemap_hash, rec->ncr.ncr_name);
 736         if (nodemap == NULL) {
 737                 if (nodemap_id == LUSTRE_NODEMAP_DEFAULT_ID)
 738                         nodemap = nodemap_create(rec->ncr.ncr_name, config, 1);
 739                 else
 740                         nodemap = nodemap_create(rec->ncr.ncr_name, config, 0);
 741                 if (IS_ERR(nodemap))
 742                         return PTR_ERR(nodemap);
 743
 744                 /* we need to override the local ID with the saved ID */
 745                 nodemap->nm_id = nodemap_id;
 746                 if (nodemap_id > config->nmc_nodemap_highest_id)
 747                         config->nmc_nodemap_highest_id = nodemap_id;
 748
 749         } else if (nodemap->nm_id != nodemap_id) {
 750                 nodemap_putref(nodemap);
 751                 return -EINVAL;
 752         }
 753
 754         nodemap->nm_squash_uid = le32_to_cpu(rec->ncr.ncr_squash_uid);
 755         nodemap->nm_squash_gid = le32_to_cpu(rec->ncr.ncr_squash_gid);
 756         nodemap->nm_squash_projid = le32_to_cpu(rec->ncr.ncr_squash_projid);
 757
 758         flags = rec->ncr.ncr_flags;
 759         nodemap->nmf_allow_root_access = flags & NM_FL_ALLOW_ROOT_ACCESS;
 760         nodemap->nmf_trust_client_ids = flags & NM_FL_TRUST_CLIENT_IDS;
 761         nodemap->nmf_deny_unknown = flags & NM_FL_DENY_UNKNOWN;
 762         nodemap->nmf_map_mode =
 763                 (flags & NM_FL_MAP_UID ? NODEMAP_MAP_UID : 0) |
 764                 (flags & NM_FL_MAP_GID ? NODEMAP_MAP_GID : 0) |
 765                 (flags & NM_FL_MAP_PROJID ? NODEMAP_MAP_PROJID : 0);
 766         if (nodemap->nmf_map_mode == NODEMAP_MAP_BOTH_LEGACY)
 767                 nodemap->nmf_map_mode = NODEMAP_MAP_BOTH;
 768         nodemap->nmf_enable_audit = flags & NM_FL_ENABLE_AUDIT;
 769         nodemap->nmf_forbid_encryption = flags & NM_FL_FORBID_ENCRYPT;
 770         flags2 = rec->ncr.ncr_flags2;
 771         nodemap->nmf_readonly_mount = flags2 & NM_FL2_READONLY_MOUNT;
 772         /* by default, and in the absence of cluster_roles, grant all roles */
 773         nodemap->nmf_rbac = NODEMAP_RBAC_ALL;
 774
 775         /* The fileset should be saved otherwise it will be empty
 776          * every time in case of "NODEMAP_CLUSTER_IDX".
 777          */
 778         mutex_lock(&active_config_lock);
 779         old_nm = nodemap_lookup(rec->ncr.ncr_name);
 780         if (!IS_ERR(old_nm) && old_nm->nm_fileset[0] != '\0')
 781                 strlcpy(nodemap->nm_fileset, old_nm->nm_fileset,
 782                         sizeof(nodemap->nm_fileset));
 783         mutex_unlock(&active_config_lock);
 784         if (!IS_ERR(old_nm))
 785                 nodemap_putref(old_nm);
 786
 787         if (*recent_nodemap == NULL) {
 788                 *recent_nodemap = nodemap;
 789                 INIT_LIST_HEAD(&nodemap->nm_list);
 790         } else {
 791                 list_add(&nodemap->nm_list, &(*recent_nodemap)->nm_list);
 792         }
 793         nodemap_putref(nodemap);
 794
 795         return 0;
 796 }
 797
 798 static int nodemap_cluster_roles_helper(struct lu_nodemap *nodemap,
 799                                         const union nodemap_rec *rec)
 800 {
 801         nodemap->nmf_rbac = le64_to_cpu(rec->ncrr.ncrr_roles);
 802
 803         return 0;
 804 }
 805
 806 /**
 807  * Process a key/rec pair and modify the new configuration.
 808  *
 809  * \param       config          configuration to update with this key/rec data
 810  * \param       key             key of the record that was loaded
 811  * \param       rec             record that was loaded
 812  * \param       recent_nodemap  last referenced nodemap
 813  * \retval      type of record processed, see enum #nodemap_idx_type
 814  * \retval      -ENOENT         range or map loaded before nodemap record
 815  * \retval      -EINVAL         duplicate nodemap cluster records found with
 816  *                              different IDs, or nodemap has invalid name
 817  * \retval      -ENOMEM
 818  */
 819 static int nodemap_process_keyrec(struct nodemap_config *config,
 820                                   const struct nodemap_key *key,
 821                                   const union nodemap_rec *rec,
 822                                   struct lu_nodemap **recent_nodemap)
 823 {
 824         struct lu_nodemap *nodemap = NULL;
 825         enum nodemap_idx_type type;
 826         enum nodemap_id_type id_type;
 827         int subtype;
 828         u32 nodemap_id;
 829         lnet_nid_t nid[2];
 830         u32 map[2];
 831         int rc;
 832
 833         ENTRY;
 834
 835         BUILD_BUG_ON(sizeof(union nodemap_rec) != 32);
 836
 837         nodemap_id = le32_to_cpu(key->nk_nodemap_id);
 838         type = nodemap_get_key_type(key);
 839         subtype = nodemap_get_key_subtype(key);
 840         nodemap_id = nm_idx_set_type(nodemap_id, 0);
 841
 842         CDEBUG(D_INFO, "found config entry, nm_id %d type %d subtype %d\n",
 843                nodemap_id, type, subtype);
 844
 845         /* find the correct nodemap in the load list */
 846         if (type == NODEMAP_RANGE_IDX || type == NODEMAP_UIDMAP_IDX ||
 847             type == NODEMAP_GIDMAP_IDX || type == NODEMAP_PROJIDMAP_IDX ||
 848             (type == NODEMAP_CLUSTER_IDX && subtype != NODEMAP_CLUSTER_REC)) {
 849                 struct lu_nodemap *tmp = NULL;
 850
 851                 nodemap = *recent_nodemap;
 852
 853                 if (nodemap == NULL)
 854                         GOTO(out, rc = -ENOENT);
 855
 856                 if (nodemap->nm_id != nodemap_id) {
 857                         list_for_each_entry(tmp, &nodemap->nm_list, nm_list)
 858                                 if (tmp->nm_id == nodemap_id) {
 859                                         nodemap = tmp;
 860                                         break;
 861                                 }
 862
 863                         if (nodemap->nm_id != nodemap_id)
 864                                 GOTO(out, rc = -ENOENT);
 865                 }
 866
 867                 /* update most recently used nodemap if necessay */
 868                 if (nodemap != *recent_nodemap)
 869                         *recent_nodemap = nodemap;
 870         }
 871
 872         switch (type) {
 873         case NODEMAP_EMPTY_IDX:
 874                 if (nodemap_id != 0)
 875                         CWARN("Found nodemap config record without type field, "
 876                               " nodemap_id=%d. nodemap config file corrupt?\n",
 877                               nodemap_id);
 878                 break;
 879         case NODEMAP_CLUSTER_IDX:
 880                 switch (nodemap_get_key_subtype(key)) {
 881                 case NODEMAP_CLUSTER_REC:
 882                         rc = nodemap_cluster_rec_helper(config, nodemap_id, rec,
 883                                                         recent_nodemap);
 884                         if (rc != 0)
 885                                 GOTO(out, rc);
 886                         break;
 887                 case NODEMAP_CLUSTER_ROLES:
 888                         rc = nodemap_cluster_roles_helper(nodemap, rec);
 889                         if (rc != 0)
 890                                 GOTO(out, rc);
 891                         break;
 892                 default:
 893                         CWARN("%s: ignoring keyrec of type %d with subtype %u\n",
 894                               nodemap->nm_name, NODEMAP_CLUSTER_IDX,
 895                               nodemap_get_key_subtype(key));
 896                         break;
 897                 }
 898                 break;
 899         case NODEMAP_RANGE_IDX:
 900                 nid[0] = le64_to_cpu(rec->nrr.nrr_start_nid);
 901                 nid[1] = le64_to_cpu(rec->nrr.nrr_end_nid);
 902
 903                 rc = nodemap_add_range_helper(config, nodemap, nid,
 904                                         le32_to_cpu(key->nk_range_id));
 905                 if (rc != 0)
 906                         GOTO(out, rc);
 907                 break;
 908         case NODEMAP_UIDMAP_IDX:
 909         case NODEMAP_GIDMAP_IDX:
 910         case NODEMAP_PROJIDMAP_IDX:
 911                 map[0] = le32_to_cpu(key->nk_id_client);
 912                 map[1] = le32_to_cpu(rec->nir.nir_id_fs);
 913
 914                 if (type == NODEMAP_UIDMAP_IDX)
 915                         id_type = NODEMAP_UID;
 916                 else if (type == NODEMAP_GIDMAP_IDX)
 917                         id_type = NODEMAP_GID;
 918                 else if (type == NODEMAP_PROJIDMAP_IDX)
 919                         id_type = NODEMAP_PROJID;
 920                 else
 921                         GOTO(out, rc = -EINVAL);
 922
 923                 rc = nodemap_add_idmap_helper(nodemap, id_type, map);
 924                 if (rc != 0)
 925                         GOTO(out, rc);
 926                 break;
 927         case NODEMAP_GLOBAL_IDX:
 928                 switch (key->nk_unused) {
 929                 case 0:
 930                         config->nmc_nodemap_is_active = rec->ngr.ngr_is_active;
 931                         break;
 932                 default:
 933                         CWARN("%s: ignoring keyrec of type %d with subtype %u\n",
 934                               recent_nodemap ?
 935                                (*recent_nodemap)->nm_name : "nodemap",
 936                               NODEMAP_GLOBAL_IDX, key->nk_unused);
 937                         break;
 938                 }
 939                 break;
 940         default:
 941                 CWARN("%s: ignoring key %u:%u for unknown type %u\n",
 942                       recent_nodemap ? (*recent_nodemap)->nm_name : "nodemap",
 943                       key->nk_nodemap_id & 0x0FFFFFFF, key->nk_unused, type);
 944                 break;
 945         }
 946
 947         rc = type;
 948
 949         EXIT;
 950
 951 out:
 952         return rc;
 953 }
 954
 955 enum nm_config_passes {
 956         NM_READ_CLUSTERS = 0,
 957         NM_READ_ATTRIBUTES = 1,
 958 };
 959
 960 static int nodemap_load_entries(const struct lu_env *env,
 961                                 struct dt_object *nodemap_idx)
 962 {
 963         const struct dt_it_ops *iops;
 964         struct dt_it *it;
 965         struct lu_nodemap *recent_nodemap = NULL;
 966         struct nodemap_config *new_config = NULL;
 967         u64 hash = 0;
 968         bool activate_nodemap = false;
 969         bool loaded_global_idx = false;
 970         enum nm_config_passes cur_pass = NM_READ_CLUSTERS;
 971         int rc = 0;
 972
 973         ENTRY;
 974
 975         iops = &nodemap_idx->do_index_ops->dio_it;
 976
 977         dt_read_lock(env, nodemap_idx, 0);
 978         it = iops->init(env, nodemap_idx, 0);
 979         if (IS_ERR(it))
 980                 GOTO(out, rc = PTR_ERR(it));
 981
 982         rc = iops->load(env, it, hash);
 983         if (rc < 0)
 984                 GOTO(out_iops_fini, rc);
 985
 986         /* rc == 0 means we need to advance to record */
 987         if (rc == 0) {
 988                 rc = iops->next(env, it);
 989
 990                 if (rc < 0)
 991                         GOTO(out_iops_put, rc);
 992                 /* rc > 0 is eof, will be checked in while below */
 993         } else {
 994                 /* rc == 1, we found initial record and can process below */
 995                 rc = 0;
 996         }
 997
 998         new_config = nodemap_config_alloc();
 999         if (IS_ERR(new_config)) {
1000                 rc = PTR_ERR(new_config);
1001                 new_config = NULL;
1002                 GOTO(out_iops_put, rc);
1003         }
1004
1005         /* rc > 0 is eof, check initial iops->next here as well */
1006         while (rc == 0) {
1007                 struct nodemap_key *key;
1008                 union nodemap_rec rec;
1009                 enum nodemap_idx_type key_type;
1010                 int sub_type;
1011
1012                 key = (struct nodemap_key *)iops->key(env, it);
1013                 key_type = nodemap_get_key_type((struct nodemap_key *)key);
1014                 sub_type = nodemap_get_key_subtype((struct nodemap_key *)key);
1015                 if ((cur_pass == NM_READ_CLUSTERS &&
1016                      key_type == NODEMAP_CLUSTER_IDX &&
1017                      sub_type == NODEMAP_CLUSTER_REC) ||
1018                     (cur_pass == NM_READ_ATTRIBUTES &&
1019                      (key_type != NODEMAP_CLUSTER_IDX ||
1020                       sub_type != NODEMAP_CLUSTER_REC) &&
1021                      key_type != NODEMAP_EMPTY_IDX)) {
1022                         rc = iops->rec(env, it, (struct dt_rec *)&rec, 0);
1023                         if (rc != -ESTALE) {
1024                                 if (rc != 0)
1025                                         GOTO(out_nodemap_config, rc);
1026                                 rc = nodemap_process_keyrec(new_config, key, &rec,
1027                                                             &recent_nodemap);
1028                                 if (rc < 0)
1029                                         GOTO(out_nodemap_config, rc);
1030                                 if (rc == NODEMAP_GLOBAL_IDX)
1031                                         loaded_global_idx = true;
1032                         }
1033                 }
1034
1035                 do
1036                         rc = iops->next(env, it);
1037                 while (rc == -ESTALE);
1038
1039                 /* move to second pass */
1040                 if (rc > 0 && cur_pass == NM_READ_CLUSTERS) {
1041                         cur_pass = NM_READ_ATTRIBUTES;
1042                         rc = iops->load(env, it, 0);
1043                         if (rc == 0)
1044                                 rc = iops->next(env, it);
1045                         else if (rc > 0)
1046                                 rc = 0;
1047                         else
1048                                 GOTO(out, rc);
1049                 }
1050         }
1051
1052         if (rc > 0)
1053                 rc = 0;
1054
1055 out_nodemap_config:
1056         if (rc != 0)
1057                 nodemap_config_dealloc(new_config);
1058         else
1059                 /* creating new default needs to be done outside dt read lock */
1060                 activate_nodemap = true;
1061 out_iops_put:
1062         iops->put(env, it);
1063 out_iops_fini:
1064         iops->fini(env, it);
1065 out:
1066         dt_read_unlock(env, nodemap_idx);
1067
1068         if (rc != 0)
1069                 CWARN("%s: failed to load nodemap configuration: rc = %d\n",
1070                       nodemap_idx->do_lu.lo_dev->ld_obd->obd_name, rc);
1071
1072         if (!activate_nodemap)
1073                 RETURN(rc);
1074
1075         if (new_config->nmc_default_nodemap == NULL) {
1076                 /* new MGS won't have a default nm on disk, so create it here */
1077                 struct lu_nodemap *nodemap =
1078                         nodemap_create(DEFAULT_NODEMAP, new_config, 1);
1079                 if (IS_ERR(nodemap)) {
1080                         rc = PTR_ERR(nodemap);
1081                 } else {
1082                         rc = nodemap_idx_cluster_add_update(
1083                                         new_config->nmc_default_nodemap,
1084                                         nodemap_idx,
1085                                         NM_ADD, NODEMAP_CLUSTER_REC);
1086                         nodemap_putref(new_config->nmc_default_nodemap);
1087                 }
1088         }
1089
1090         /* new nodemap config won't have an active/inactive record */
1091         if (rc == 0 && loaded_global_idx == false) {
1092                 struct nodemap_key       nk;
1093                 union nodemap_rec        nr;
1094
1095                 nodemap_global_key_init(&nk);
1096                 nodemap_global_rec_init(&nr, false);
1097                 rc = nodemap_idx_insert(env, nodemap_idx, &nk, &nr);
1098         }
1099
1100         if (rc == 0)
1101                 nodemap_config_set_active(new_config);
1102         else
1103                 nodemap_config_dealloc(new_config);
1104
1105         RETURN(rc);
1106 }
1107
1108 /**
1109  * Step through active config and write to disk.
1110  */
1111 static struct dt_object *
1112 nodemap_save_config_cache(const struct lu_env *env,
1113                           struct dt_device *dev,
1114                           struct local_oid_storage *los)
1115 {
1116         struct dt_object *o;
1117         struct lu_nodemap *nodemap;
1118         struct lu_nodemap *nm_tmp;
1119         struct lu_nid_range *range;
1120         struct lu_nid_range *range_temp;
1121         struct lu_idmap *idmap;
1122         struct lu_idmap *id_tmp;
1123         struct rb_root root;
1124         struct nodemap_key nk;
1125         union nodemap_rec nr;
1126         LIST_HEAD(nodemap_list_head);
1127         int rc = 0, rc2;
1128
1129         ENTRY;
1130
1131         /* create a new index file to fill with active config */
1132         o = nodemap_cache_find_create(env, dev, los, NCFC_CREATE_NEW);
1133         if (IS_ERR(o))
1134                 RETURN(o);
1135
1136         mutex_lock(&active_config_lock);
1137
1138         /* convert hash to list so we don't spin */
1139         cfs_hash_for_each_safe(active_config->nmc_nodemap_hash,
1140                                nm_hash_list_cb, &nodemap_list_head);
1141
1142         list_for_each_entry_safe(nodemap, nm_tmp, &nodemap_list_head, nm_list) {
1143                 nodemap_cluster_key_init(&nk, nodemap->nm_id,
1144                                          NODEMAP_CLUSTER_REC);
1145                 nodemap_cluster_rec_init(&nr, nodemap);
1146
1147                 rc2 = nodemap_idx_insert(env, o, &nk, &nr);
1148                 if (rc2 < 0) {
1149                         rc = rc2;
1150                         continue;
1151                 }
1152
1153                 /* only insert NODEMAP_CLUSTER_ROLES idx in saved config cache
1154                  * if nmf_rbac is not default value NODEMAP_RBAC_ALL
1155                  */
1156                 if (nodemap->nmf_rbac != NODEMAP_RBAC_ALL) {
1157                         nodemap_cluster_key_init(&nk, nodemap->nm_id,
1158                                                  NODEMAP_CLUSTER_ROLES);
1159                         nodemap_cluster_roles_rec_init(&nr, nodemap);
1160                         rc2 = nodemap_idx_insert(env, o, &nk, &nr);
1161                         if (rc2 < 0)
1162                                 rc = rc2;
1163                 }
1164
1165                 down_read(&active_config->nmc_range_tree_lock);
1166                 list_for_each_entry_safe(range, range_temp, &nodemap->nm_ranges,
1167                                          rn_list) {
1168                         lnet_nid_t nid[2] = {
1169                                 range->rn_start,
1170                                 range->rn_end
1171                         };
1172                         nodemap_range_key_init(&nk, nodemap->nm_id,
1173                                                range->rn_id);
1174                         nodemap_range_rec_init(&nr, nid);
1175                         rc2 = nodemap_idx_insert(env, o, &nk, &nr);
1176                         if (rc2 < 0)
1177                                 rc = rc2;
1178                 }
1179                 up_read(&active_config->nmc_range_tree_lock);
1180
1181                 /* we don't need to take nm_idmap_lock because active config
1182                  * lock prevents changes from happening to nodemaps
1183                  */
1184                 root = nodemap->nm_client_to_fs_uidmap;
1185                 nm_rbtree_postorder_for_each_entry_safe(idmap, id_tmp, &root,
1186                                                         id_client_to_fs) {
1187                         nodemap_idmap_key_init(&nk, nodemap->nm_id, NODEMAP_UID,
1188                                                idmap->id_client);
1189                         nodemap_idmap_rec_init(&nr, idmap->id_fs);
1190                         rc2 = nodemap_idx_insert(env, o, &nk, &nr);
1191                         if (rc2 < 0)
1192                                 rc = rc2;
1193                 }
1194
1195                 root = nodemap->nm_client_to_fs_gidmap;
1196                 nm_rbtree_postorder_for_each_entry_safe(idmap, id_tmp, &root,
1197                                                         id_client_to_fs) {
1198                         nodemap_idmap_key_init(&nk, nodemap->nm_id, NODEMAP_GID,
1199                                                idmap->id_client);
1200                         nodemap_idmap_rec_init(&nr, idmap->id_fs);
1201                         rc2 = nodemap_idx_insert(env, o, &nk, &nr);
1202                         if (rc2 < 0)
1203                                 rc = rc2;
1204                 }
1205
1206                 root = nodemap->nm_client_to_fs_projidmap;
1207                 nm_rbtree_postorder_for_each_entry_safe(idmap, id_tmp, &root,
1208                                                         id_client_to_fs) {
1209                         nodemap_idmap_key_init(&nk, nodemap->nm_id,
1210                                                NODEMAP_PROJID,
1211                                                idmap->id_client);
1212                         nodemap_idmap_rec_init(&nr, idmap->id_fs);
1213                         rc2 = nodemap_idx_insert(env, o, &nk, &nr);
1214                         if (rc2 < 0)
1215                                 rc = rc2;
1216                 }
1217         }
1218         nodemap_global_key_init(&nk);
1219         nodemap_global_rec_init(&nr, active_config->nmc_nodemap_is_active);
1220         rc2 = nodemap_idx_insert(env, o, &nk, &nr);
1221         if (rc2 < 0)
1222                 rc = rc2;
1223
1224         mutex_unlock(&active_config_lock);
1225
1226         if (rc < 0) {
1227                 dt_object_put(env, o);
1228                 o = ERR_PTR(rc);
1229         }
1230
1231         RETURN(o);
1232 }
1233
1234 static void nodemap_save_all_caches(void)
1235 {
1236         struct nm_config_file   *ncf;
1237         struct lu_env            env;
1238         int                      rc = 0;
1239
1240         /* recreating nodemap cache requires fld_thread_key be in env */
1241         rc = lu_env_init(&env, LCT_MD_THREAD | LCT_DT_THREAD | LCT_MG_THREAD);
1242         if (rc != 0) {
1243                 CWARN("cannot init env for nodemap config: rc = %d\n", rc);
1244                 return;
1245         }
1246
1247         mutex_lock(&ncf_list_lock);
1248         list_for_each_entry(ncf, &ncf_list_head, ncf_list) {
1249                 struct dt_device *dev = lu2dt_dev(ncf->ncf_obj->do_lu.lo_dev);
1250                 struct obd_device *obd = ncf->ncf_obj->do_lu.lo_dev->ld_obd;
1251                 struct dt_object *o;
1252
1253                 /* put current config file so save conf can rewrite it */
1254                 dt_object_put_nocache(&env, ncf->ncf_obj);
1255                 ncf->ncf_obj = NULL;
1256
1257                 o = nodemap_save_config_cache(&env, dev, ncf->ncf_los);
1258                 if (IS_ERR(o))
1259                         CWARN("%s: error writing to nodemap config: rc = %d\n",
1260                               obd->obd_name, rc);
1261                 else
1262                         ncf->ncf_obj = o;
1263         }
1264         mutex_unlock(&ncf_list_lock);
1265
1266         lu_env_fini(&env);
1267 }
1268
1269 /* tracks if config still needs to be loaded, either from disk or network */
1270 static bool nodemap_config_loaded;
1271 static DEFINE_MUTEX(nodemap_config_loaded_lock);
1272
1273 /**
1274  * Ensures that configs loaded over the wire are prioritized over those loaded
1275  * from disk.
1276  *
1277  * \param config        config to set as the active config
1278  */
1279 void nodemap_config_set_active_mgc(struct nodemap_config *config)
1280 {
1281         mutex_lock(&nodemap_config_loaded_lock);
1282         nodemap_config_set_active(config);
1283         nodemap_config_loaded = true;
1284         nodemap_save_all_caches();
1285         mutex_unlock(&nodemap_config_loaded_lock);
1286 }
1287 EXPORT_SYMBOL(nodemap_config_set_active_mgc);
1288
1289 /**
1290  * Register a dt_object representing the config index file. This should be
1291  * called by targets in order to load the nodemap configuration from disk. The
1292  * dt_object should be created with local_index_find_or_create and the index
1293  * features should be enabled with do_index_try.
1294  *
1295  * \param obj   dt_object returned by local_index_find_or_create
1296  *
1297  * \retval      on success: nm_config_file handle for later deregistration
1298  * \retval      -ENOMEM         memory allocation failure
1299  * \retval      -ENOENT         error loading nodemap config
1300  * \retval      -EINVAL         error loading nodemap config
1301  * \retval      -EEXIST         nodemap config already registered for MGS
1302  */
1303 struct nm_config_file *nm_config_file_register_mgs(const struct lu_env *env,
1304                                                    struct dt_object *obj,
1305                                                    struct local_oid_storage *los)
1306 {
1307         struct nm_config_file *ncf;
1308         int rc = 0;
1309         ENTRY;
1310
1311         if (nodemap_mgs_ncf != NULL)
1312                 GOTO(out, ncf = ERR_PTR(-EEXIST));
1313
1314         OBD_ALLOC_PTR(ncf);
1315         if (ncf == NULL)
1316                 GOTO(out, ncf = ERR_PTR(-ENOMEM));
1317
1318         /* if loading from cache, prevent activation of MGS config until cache
1319          * loading is done, so disk config is overwritten by MGS config.
1320          */
1321         mutex_lock(&nodemap_config_loaded_lock);
1322         rc = nodemap_load_entries(env, obj);
1323         if (!rc)
1324                 nodemap_config_loaded = true;
1325         mutex_unlock(&nodemap_config_loaded_lock);
1326
1327         if (rc) {
1328                 OBD_FREE_PTR(ncf);
1329                 GOTO(out, ncf = ERR_PTR(rc));
1330         }
1331
1332         lu_object_get(&obj->do_lu);
1333
1334         ncf->ncf_obj = obj;
1335         ncf->ncf_los = los;
1336
1337         nodemap_mgs_ncf = ncf;
1338
1339 out:
1340         return ncf;
1341 }
1342 EXPORT_SYMBOL(nm_config_file_register_mgs);
1343
1344 struct nm_config_file *nm_config_file_register_tgt(const struct lu_env *env,
1345                                                    struct dt_device *dev,
1346                                                    struct local_oid_storage *los)
1347 {
1348         struct nm_config_file *ncf;
1349         struct dt_object *config_obj = NULL;
1350         int rc = 0;
1351
1352         OBD_ALLOC_PTR(ncf);
1353         if (ncf == NULL)
1354                 RETURN(ERR_PTR(-ENOMEM));
1355
1356         /* don't load from cache if config already loaded */
1357         mutex_lock(&nodemap_config_loaded_lock);
1358         if (!nodemap_config_loaded) {
1359                 config_obj = nodemap_cache_find_create(env, dev, los, 0);
1360                 if (IS_ERR(config_obj))
1361                         rc = PTR_ERR(config_obj);
1362                 else
1363                         rc = nodemap_load_entries(env, config_obj);
1364
1365                 if (!rc)
1366                         nodemap_config_loaded = true;
1367         }
1368         mutex_unlock(&nodemap_config_loaded_lock);
1369         if (rc)
1370                 GOTO(out_ncf, rc);
1371
1372         /* sync on disk caches w/ loaded config in memory, ncf_obj may change */
1373         if (!config_obj) {
1374                 config_obj = nodemap_save_config_cache(env, dev, los);
1375                 if (IS_ERR(config_obj))
1376                         GOTO(out_ncf, rc = PTR_ERR(config_obj));
1377         }
1378
1379         ncf->ncf_obj = config_obj;
1380         ncf->ncf_los = los;
1381
1382         mutex_lock(&ncf_list_lock);
1383         list_add(&ncf->ncf_list, &ncf_list_head);
1384         mutex_unlock(&ncf_list_lock);
1385
1386 out_ncf:
1387         if (rc) {
1388                 OBD_FREE_PTR(ncf);
1389                 RETURN(ERR_PTR(rc));
1390         }
1391
1392         RETURN(ncf);
1393 }
1394 EXPORT_SYMBOL(nm_config_file_register_tgt);
1395
1396 /**
1397  * Deregister a nm_config_file. Should be called by targets during cleanup.
1398  *
1399  * \param ncf   config file to deregister
1400  */
1401 void nm_config_file_deregister_mgs(const struct lu_env *env,
1402                                    struct nm_config_file *ncf)
1403 {
1404         ENTRY;
1405         LASSERT(nodemap_mgs_ncf == ncf);
1406
1407         nodemap_mgs_ncf = NULL;
1408         if (ncf->ncf_obj)
1409                 dt_object_put(env, ncf->ncf_obj);
1410
1411         OBD_FREE_PTR(ncf);
1412
1413         EXIT;
1414 }
1415 EXPORT_SYMBOL(nm_config_file_deregister_mgs);
1416
1417 void nm_config_file_deregister_tgt(const struct lu_env *env,
1418                                    struct nm_config_file *ncf)
1419 {
1420         ENTRY;
1421
1422         if (ncf == NULL)
1423                 return;
1424
1425         mutex_lock(&ncf_list_lock);
1426         list_del(&ncf->ncf_list);
1427         mutex_unlock(&ncf_list_lock);
1428
1429         if (ncf->ncf_obj)
1430                 dt_object_put(env, ncf->ncf_obj);
1431
1432         OBD_FREE_PTR(ncf);
1433
1434         EXIT;
1435 }
1436 EXPORT_SYMBOL(nm_config_file_deregister_tgt);
1437
1438 int nodemap_process_idx_pages(struct nodemap_config *config, union lu_page *lip,
1439                               struct lu_nodemap **recent_nodemap)
1440 {
1441         struct nodemap_key *key;
1442         union nodemap_rec *rec;
1443         char *entry;
1444         int j;
1445         int k;
1446         int rc = 0;
1447         int size = dt_nodemap_features.dif_keysize_max +
1448                    dt_nodemap_features.dif_recsize_max;
1449         ENTRY;
1450
1451         for (j = 0; j < LU_PAGE_COUNT; j++) {
1452                 if (lip->lp_idx.lip_magic != LIP_MAGIC)
1453                         return -EINVAL;
1454
1455                 /* get and process keys and records from page */
1456                 for (k = 0; k < lip->lp_idx.lip_nr; k++) {
1457                         entry = lip->lp_idx.lip_entries + k * size;
1458                         key = (struct nodemap_key *)entry;
1459
1460                         entry += dt_nodemap_features.dif_keysize_max;
1461                         rec = (union nodemap_rec *)entry;
1462
1463                         rc = nodemap_process_keyrec(config, key, rec,
1464                                                     recent_nodemap);
1465                         if (rc < 0)
1466                                 return rc;
1467                 }
1468                 lip++;
1469         }
1470
1471         EXIT;
1472         return 0;
1473 }
1474 EXPORT_SYMBOL(nodemap_process_idx_pages);
1475
1476 static int nodemap_page_build(const struct lu_env *env, struct dt_object *obj,
1477                               union lu_page *lp, size_t bytes,
1478                               const struct dt_it_ops *iops,
1479                               struct dt_it *it, __u32 attr, void *arg)
1480 {
1481         struct idx_info *ii = (struct idx_info *)arg;
1482         struct lu_idxpage *lip = &lp->lp_idx;
1483         char *entry;
1484         size_t size = ii->ii_keysize + ii->ii_recsize;
1485         int rc;
1486         ENTRY;
1487
1488         if (bytes < LIP_HDR_SIZE)
1489                 return -EINVAL;
1490
1491         /* initialize the header of the new container */
1492         memset(lip, 0, LIP_HDR_SIZE);
1493         lip->lip_magic = LIP_MAGIC;
1494         bytes -= LIP_HDR_SIZE;
1495
1496         entry = lip->lip_entries;
1497         do {
1498                 char *tmp_entry = entry;
1499                 struct dt_key *key;
1500                 __u64 hash;
1501                 enum nodemap_idx_type key_type;
1502                 int sub_type;
1503
1504                 /* fetch 64-bit hash value */
1505                 hash = iops->store(env, it);
1506                 ii->ii_hash_end = hash;
1507
1508                 if (CFS_FAIL_CHECK(OBD_FAIL_OBD_IDX_READ_BREAK)) {
1509                         if (lip->lip_nr != 0)
1510                                 GOTO(out, rc = 0);
1511                 }
1512
1513                 if (bytes < size) {
1514                         if (lip->lip_nr == 0)
1515                                 GOTO(out, rc = -EINVAL);
1516                         GOTO(out, rc = 0);
1517                 }
1518
1519                 key = iops->key(env, it);
1520                 key_type = nodemap_get_key_type((struct nodemap_key *)key);
1521                 sub_type = nodemap_get_key_subtype((struct nodemap_key *)key);
1522
1523                 /* on the first pass, get only the cluster types. On second
1524                  * pass, get all the rest */
1525                 if ((ii->ii_attrs == NM_READ_CLUSTERS &&
1526                      key_type == NODEMAP_CLUSTER_IDX &&
1527                      sub_type == NODEMAP_CLUSTER_REC) ||
1528                     (ii->ii_attrs == NM_READ_ATTRIBUTES &&
1529                      (key_type != NODEMAP_CLUSTER_IDX ||
1530                       sub_type != NODEMAP_CLUSTER_REC) &&
1531                      key_type != NODEMAP_EMPTY_IDX)) {
1532                         memcpy(tmp_entry, key, ii->ii_keysize);
1533                         tmp_entry += ii->ii_keysize;
1534
1535                         /* and finally the record */
1536                         rc = iops->rec(env, it, (struct dt_rec *)tmp_entry,
1537                                        attr);
1538                         if (rc != -ESTALE) {
1539                                 if (rc != 0)
1540                                         GOTO(out, rc);
1541
1542                                 /* hash/key/record successfully copied! */
1543                                 lip->lip_nr++;
1544                                 if (unlikely(lip->lip_nr == 1 &&
1545                                     ii->ii_count == 0))
1546                                         ii->ii_hash_start = hash;
1547
1548                                 entry = tmp_entry + ii->ii_recsize;
1549                                 bytes -= size;
1550                         }
1551                 }
1552
1553                 /* move on to the next record */
1554                 do {
1555                         rc = iops->next(env, it);
1556                 } while (rc == -ESTALE);
1557
1558                 /* move to second pass */
1559                 if (rc > 0 && ii->ii_attrs == NM_READ_CLUSTERS) {
1560                         ii->ii_attrs = NM_READ_ATTRIBUTES;
1561                         rc = iops->load(env, it, 0);
1562                         if (rc == 0)
1563                                 rc = iops->next(env, it);
1564                         else if (rc > 0)
1565                                 rc = 0;
1566                         else
1567                                 GOTO(out, rc);
1568                 }
1569
1570         } while (rc == 0);
1571
1572         GOTO(out, rc);
1573 out:
1574         if (rc >= 0 && lip->lip_nr > 0)
1575                 /* one more container */
1576                 ii->ii_count++;
1577         if (rc > 0)
1578                 /* no more entries */
1579                 ii->ii_hash_end = II_END_OFF;
1580         return rc;
1581 }
1582
1583 int nodemap_index_read(struct lu_env *env,
1584                        struct nm_config_file *ncf,
1585                        struct idx_info *ii,
1586                        const struct lu_rdpg *rdpg)
1587 {
1588         struct dt_object        *nodemap_idx = ncf->ncf_obj;
1589         __u64                    version;
1590         int                      rc = 0;
1591
1592         ii->ii_keysize = dt_nodemap_features.dif_keysize_max;
1593         ii->ii_recsize = dt_nodemap_features.dif_recsize_max;
1594
1595         dt_read_lock(env, nodemap_idx, 0);
1596         version = dt_version_get(env, nodemap_idx);
1597         if (rdpg->rp_hash != 0 && ii->ii_version != version) {
1598                 CDEBUG(D_INFO, "nodemap config changed inflight, old %llu, new %llu\n",
1599                        ii->ii_version,
1600                        version);
1601                 ii->ii_hash_end = 0;
1602         } else {
1603                 rc = dt_index_walk(env, nodemap_idx, rdpg, nodemap_page_build,
1604                                    ii);
1605                 CDEBUG(D_INFO, "walked index, hashend %llx\n", ii->ii_hash_end);
1606         }
1607
1608         if (rc >= 0)
1609                 ii->ii_version = version;
1610
1611         dt_read_unlock(env, nodemap_idx);
1612         return rc;
1613 }
1614 EXPORT_SYMBOL(nodemap_index_read);
1615
1616 /**
1617  * Returns the current nodemap configuration to MGC by walking the nodemap
1618  * config index and storing it in the response buffer.
1619  *
1620  * \param       req             incoming MGS_CONFIG_READ request
1621  * \retval      0               success
1622  * \retval      -EINVAL         malformed request
1623  * \retval      -ENOTCONN       client evicted/reconnected already
1624  * \retval      -ETIMEDOUT      client timeout or network error
1625  * \retval      -ENOMEM
1626  */
1627 int nodemap_get_config_req(struct obd_device *mgs_obd,
1628                            struct ptlrpc_request *req)
1629 {
1630         const struct ptlrpc_bulk_frag_ops *frag_ops = &ptlrpc_bulk_kiov_pin_ops;
1631         struct mgs_config_body *body;
1632         struct mgs_config_res *res;
1633         struct lu_rdpg rdpg;
1634         struct idx_info nodemap_ii;
1635         struct ptlrpc_bulk_desc *desc;
1636         struct tg_export_data *rqexp_ted = &req->rq_export->exp_target_data;
1637         int i;
1638         int page_count;
1639         int bytes = 0;
1640         int rc = 0;
1641
1642         body = req_capsule_client_get(&req->rq_pill, &RMF_MGS_CONFIG_BODY);
1643         if (!body)
1644                 RETURN(-EINVAL);
1645
1646         if (body->mcb_type != MGS_CFG_T_NODEMAP)
1647                 RETURN(-EINVAL);
1648
1649         rdpg.rp_count = (body->mcb_units << body->mcb_bits);
1650         rdpg.rp_npages = (rdpg.rp_count + PAGE_SIZE - 1) >>
1651                 PAGE_SHIFT;
1652         if (rdpg.rp_npages > PTLRPC_MAX_BRW_PAGES)
1653                 RETURN(-EINVAL);
1654
1655         CDEBUG(D_INFO, "reading nodemap log, name '%s', size = %u\n",
1656                body->mcb_name, rdpg.rp_count);
1657
1658         /* allocate pages to store the containers */
1659         OBD_ALLOC_PTR_ARRAY(rdpg.rp_pages, rdpg.rp_npages);
1660         if (rdpg.rp_pages == NULL)
1661                 RETURN(-ENOMEM);
1662         for (i = 0; i < rdpg.rp_npages; i++) {
1663                 rdpg.rp_pages[i] = alloc_page(GFP_NOFS);
1664                 if (rdpg.rp_pages[i] == NULL)
1665                         GOTO(out, rc = -ENOMEM);
1666         }
1667
1668         rdpg.rp_hash = body->mcb_offset;
1669         nodemap_ii.ii_magic = IDX_INFO_MAGIC;
1670         nodemap_ii.ii_flags = II_FL_NOHASH;
1671         nodemap_ii.ii_version = rqexp_ted->ted_nodemap_version;
1672         nodemap_ii.ii_attrs = body->mcb_nm_cur_pass;
1673
1674         bytes = nodemap_index_read(req->rq_svc_thread->t_env,
1675                                    obd2obt(mgs_obd)->obt_nodemap_config_file,
1676                                    &nodemap_ii, &rdpg);
1677         if (bytes < 0)
1678                 GOTO(out, rc = bytes);
1679
1680         rqexp_ted->ted_nodemap_version = nodemap_ii.ii_version;
1681
1682         res = req_capsule_server_get(&req->rq_pill, &RMF_MGS_CONFIG_RES);
1683         if (res == NULL)
1684                 GOTO(out, rc = -EINVAL);
1685         res->mcr_offset = nodemap_ii.ii_hash_end;
1686         res->mcr_nm_cur_pass = nodemap_ii.ii_attrs;
1687
1688         page_count = (bytes + PAGE_SIZE - 1) >> PAGE_SHIFT;
1689         LASSERT(page_count <= rdpg.rp_count);
1690         desc = ptlrpc_prep_bulk_exp(req, page_count, 1,
1691                                     PTLRPC_BULK_PUT_SOURCE,
1692                                     MGS_BULK_PORTAL, frag_ops);
1693         if (desc == NULL)
1694                 GOTO(out, rc = -ENOMEM);
1695
1696         for (i = 0; i < page_count && bytes > 0; i++) {
1697                 frag_ops->add_kiov_frag(desc, rdpg.rp_pages[i], 0,
1698                                         min_t(int, bytes, PAGE_SIZE));
1699                 bytes -= PAGE_SIZE;
1700         }
1701
1702         rc = target_bulk_io(req->rq_export, desc);
1703         ptlrpc_free_bulk(desc);
1704
1705 out:
1706         if (rdpg.rp_pages != NULL) {
1707                 for (i = 0; i < rdpg.rp_npages; i++)
1708                         if (rdpg.rp_pages[i] != NULL)
1709                                 __free_page(rdpg.rp_pages[i]);
1710                 OBD_FREE_PTR_ARRAY(rdpg.rp_pages, rdpg.rp_npages);
1711         }
1712         return rc;
1713 }
1714 EXPORT_SYMBOL(nodemap_get_config_req);