fs/btrfs/qgroup.c

   1 /*
   2  * Copyright (C) 2011 STRATO.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/sched.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/writeback.h>
  22 #include <linux/blkdev.h>
  23 #include <linux/rbtree.h>
  24 #include <linux/slab.h>
  25 #include <linux/workqueue.h>
  26 #include <linux/btrfs.h>
  27
  28 #include "ctree.h"
  29 #include "transaction.h"
  30 #include "disk-io.h"
  31 #include "locking.h"
  32 #include "ulist.h"
  33 #include "backref.h"
  34 #include "extent_io.h"
  35 #include "qgroup.h"
  36
  37 /* TODO XXX FIXME
  38  *  - subvol delete -> delete when ref goes to 0? delete limits also?
  39  *  - reorganize keys
  40  *  - compressed
  41  *  - sync
  42  *  - copy also limits on subvol creation
  43  *  - limit
  44  *  - caches fuer ulists
  45  *  - performance benchmarks
  46  *  - check all ioctl parameters
  47  */
  48
  49 /*
  50  * one struct for each qgroup, organized in fs_info->qgroup_tree.
  51  */
  52 struct btrfs_qgroup {
  53         u64 qgroupid;
  54
  55         /*
  56          * state
  57          */
  58         u64 rfer;       /* referenced */
  59         u64 rfer_cmpr;  /* referenced compressed */
  60         u64 excl;       /* exclusive */
  61         u64 excl_cmpr;  /* exclusive compressed */
  62
  63         /*
  64          * limits
  65          */
  66         u64 lim_flags;  /* which limits are set */
  67         u64 max_rfer;
  68         u64 max_excl;
  69         u64 rsv_rfer;
  70         u64 rsv_excl;
  71
  72         /*
  73          * reservation tracking
  74          */
  75         u64 reserved;
  76
  77         /*
  78          * lists
  79          */
  80         struct list_head groups;  /* groups this group is member of */
  81         struct list_head members; /* groups that are members of this group */
  82         struct list_head dirty;   /* dirty groups */
  83         struct rb_node node;      /* tree of qgroups */
  84
  85         /*
  86          * temp variables for accounting operations
  87          */
  88         u64 old_refcnt;
  89         u64 new_refcnt;
  90 };
  91
  92 /*
  93  * glue structure to represent the relations between qgroups.
  94  */
  95 struct btrfs_qgroup_list {
  96         struct list_head next_group;
  97         struct list_head next_member;
  98         struct btrfs_qgroup *group;
  99         struct btrfs_qgroup *member;
 100 };
 101
 102 #define ptr_to_u64(x) ((u64)(uintptr_t)x)
 103 #define u64_to_ptr(x) ((struct btrfs_qgroup *)(uintptr_t)x)
 104
 105 static int
 106 qgroup_rescan_init(struct btrfs_fs_info *fs_info, u64 progress_objectid,
 107                    int init_flags);
 108 static void qgroup_rescan_zero_tracking(struct btrfs_fs_info *fs_info);
 109
 110 /* must be called with qgroup_ioctl_lock held */
 111 static struct btrfs_qgroup *find_qgroup_rb(struct btrfs_fs_info *fs_info,
 112                                            u64 qgroupid)
 113 {
 114         struct rb_node *n = fs_info->qgroup_tree.rb_node;
 115         struct btrfs_qgroup *qgroup;
 116
 117         while (n) {
 118                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
 119                 if (qgroup->qgroupid < qgroupid)
 120                         n = n->rb_left;
 121                 else if (qgroup->qgroupid > qgroupid)
 122                         n = n->rb_right;
 123                 else
 124                         return qgroup;
 125         }
 126         return NULL;
 127 }
 128
 129 /* must be called with qgroup_lock held */
 130 static struct btrfs_qgroup *add_qgroup_rb(struct btrfs_fs_info *fs_info,
 131                                           u64 qgroupid)
 132 {
 133         struct rb_node **p = &fs_info->qgroup_tree.rb_node;
 134         struct rb_node *parent = NULL;
 135         struct btrfs_qgroup *qgroup;
 136
 137         while (*p) {
 138                 parent = *p;
 139                 qgroup = rb_entry(parent, struct btrfs_qgroup, node);
 140
 141                 if (qgroup->qgroupid < qgroupid)
 142                         p = &(*p)->rb_left;
 143                 else if (qgroup->qgroupid > qgroupid)
 144                         p = &(*p)->rb_right;
 145                 else
 146                         return qgroup;
 147         }
 148
 149         qgroup = kzalloc(sizeof(*qgroup), GFP_ATOMIC);
 150         if (!qgroup)
 151                 return ERR_PTR(-ENOMEM);
 152
 153         qgroup->qgroupid = qgroupid;
 154         INIT_LIST_HEAD(&qgroup->groups);
 155         INIT_LIST_HEAD(&qgroup->members);
 156         INIT_LIST_HEAD(&qgroup->dirty);
 157
 158         rb_link_node(&qgroup->node, parent, p);
 159         rb_insert_color(&qgroup->node, &fs_info->qgroup_tree);
 160
 161         return qgroup;
 162 }
 163
 164 static void __del_qgroup_rb(struct btrfs_qgroup *qgroup)
 165 {
 166         struct btrfs_qgroup_list *list;
 167
 168         list_del(&qgroup->dirty);
 169         while (!list_empty(&qgroup->groups)) {
 170                 list = list_first_entry(&qgroup->groups,
 171                                         struct btrfs_qgroup_list, next_group);
 172                 list_del(&list->next_group);
 173                 list_del(&list->next_member);
 174                 kfree(list);
 175         }
 176
 177         while (!list_empty(&qgroup->members)) {
 178                 list = list_first_entry(&qgroup->members,
 179                                         struct btrfs_qgroup_list, next_member);
 180                 list_del(&list->next_group);
 181                 list_del(&list->next_member);
 182                 kfree(list);
 183         }
 184         kfree(qgroup);
 185 }
 186
 187 /* must be called with qgroup_lock held */
 188 static int del_qgroup_rb(struct btrfs_fs_info *fs_info, u64 qgroupid)
 189 {
 190         struct btrfs_qgroup *qgroup = find_qgroup_rb(fs_info, qgroupid);
 191
 192         if (!qgroup)
 193                 return -ENOENT;
 194
 195         rb_erase(&qgroup->node, &fs_info->qgroup_tree);
 196         __del_qgroup_rb(qgroup);
 197         return 0;
 198 }
 199
 200 /* must be called with qgroup_lock held */
 201 static int add_relation_rb(struct btrfs_fs_info *fs_info,
 202                            u64 memberid, u64 parentid)
 203 {
 204         struct btrfs_qgroup *member;
 205         struct btrfs_qgroup *parent;
 206         struct btrfs_qgroup_list *list;
 207
 208         member = find_qgroup_rb(fs_info, memberid);
 209         parent = find_qgroup_rb(fs_info, parentid);
 210         if (!member || !parent)
 211                 return -ENOENT;
 212
 213         list = kzalloc(sizeof(*list), GFP_ATOMIC);
 214         if (!list)
 215                 return -ENOMEM;
 216
 217         list->group = parent;
 218         list->member = member;
 219         list_add_tail(&list->next_group, &member->groups);
 220         list_add_tail(&list->next_member, &parent->members);
 221
 222         return 0;
 223 }
 224
 225 /* must be called with qgroup_lock held */
 226 static int del_relation_rb(struct btrfs_fs_info *fs_info,
 227                            u64 memberid, u64 parentid)
 228 {
 229         struct btrfs_qgroup *member;
 230         struct btrfs_qgroup *parent;
 231         struct btrfs_qgroup_list *list;
 232
 233         member = find_qgroup_rb(fs_info, memberid);
 234         parent = find_qgroup_rb(fs_info, parentid);
 235         if (!member || !parent)
 236                 return -ENOENT;
 237
 238         list_for_each_entry(list, &member->groups, next_group) {
 239                 if (list->group == parent) {
 240                         list_del(&list->next_group);
 241                         list_del(&list->next_member);
 242                         kfree(list);
 243                         return 0;
 244                 }
 245         }
 246         return -ENOENT;
 247 }
 248
 249 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 250 int btrfs_verify_qgroup_counts(struct btrfs_fs_info *fs_info, u64 qgroupid,
 251                                u64 rfer, u64 excl)
 252 {
 253         struct btrfs_qgroup *qgroup;
 254
 255         qgroup = find_qgroup_rb(fs_info, qgroupid);
 256         if (!qgroup)
 257                 return -EINVAL;
 258         if (qgroup->rfer != rfer || qgroup->excl != excl)
 259                 return -EINVAL;
 260         return 0;
 261 }
 262 #endif
 263
 264 /*
 265  * The full config is read in one go, only called from open_ctree()
 266  * It doesn't use any locking, as at this point we're still single-threaded
 267  */
 268 int btrfs_read_qgroup_config(struct btrfs_fs_info *fs_info)
 269 {
 270         struct btrfs_key key;
 271         struct btrfs_key found_key;
 272         struct btrfs_root *quota_root = fs_info->quota_root;
 273         struct btrfs_path *path = NULL;
 274         struct extent_buffer *l;
 275         int slot;
 276         int ret = 0;
 277         u64 flags = 0;
 278         u64 rescan_progress = 0;
 279
 280         if (!fs_info->quota_enabled)
 281                 return 0;
 282
 283         fs_info->qgroup_ulist = ulist_alloc(GFP_NOFS);
 284         if (!fs_info->qgroup_ulist) {
 285                 ret = -ENOMEM;
 286                 goto out;
 287         }
 288
 289         path = btrfs_alloc_path();
 290         if (!path) {
 291                 ret = -ENOMEM;
 292                 goto out;
 293         }
 294
 295         /* default this to quota off, in case no status key is found */
 296         fs_info->qgroup_flags = 0;
 297
 298         /*
 299          * pass 1: read status, all qgroup infos and limits
 300          */
 301         key.objectid = 0;
 302         key.type = 0;
 303         key.offset = 0;
 304         ret = btrfs_search_slot_for_read(quota_root, &key, path, 1, 1);
 305         if (ret)
 306                 goto out;
 307
 308         while (1) {
 309                 struct btrfs_qgroup *qgroup;
 310
 311                 slot = path->slots[0];
 312                 l = path->nodes[0];
 313                 btrfs_item_key_to_cpu(l, &found_key, slot);
 314
 315                 if (found_key.type == BTRFS_QGROUP_STATUS_KEY) {
 316                         struct btrfs_qgroup_status_item *ptr;
 317
 318                         ptr = btrfs_item_ptr(l, slot,
 319                                              struct btrfs_qgroup_status_item);
 320
 321                         if (btrfs_qgroup_status_version(l, ptr) !=
 322                             BTRFS_QGROUP_STATUS_VERSION) {
 323                                 btrfs_err(fs_info,
 324                                  "old qgroup version, quota disabled");
 325                                 goto out;
 326                         }
 327                         if (btrfs_qgroup_status_generation(l, ptr) !=
 328                             fs_info->generation) {
 329                                 flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
 330                                 btrfs_err(fs_info,
 331                                         "qgroup generation mismatch, "
 332                                         "marked as inconsistent");
 333                         }
 334                         fs_info->qgroup_flags = btrfs_qgroup_status_flags(l,
 335                                                                           ptr);
 336                         rescan_progress = btrfs_qgroup_status_rescan(l, ptr);
 337                         goto next1;
 338                 }
 339
 340                 if (found_key.type != BTRFS_QGROUP_INFO_KEY &&
 341                     found_key.type != BTRFS_QGROUP_LIMIT_KEY)
 342                         goto next1;
 343
 344                 qgroup = find_qgroup_rb(fs_info, found_key.offset);
 345                 if ((qgroup && found_key.type == BTRFS_QGROUP_INFO_KEY) ||
 346                     (!qgroup && found_key.type == BTRFS_QGROUP_LIMIT_KEY)) {
 347                         btrfs_err(fs_info, "inconsitent qgroup config");
 348                         flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
 349                 }
 350                 if (!qgroup) {
 351                         qgroup = add_qgroup_rb(fs_info, found_key.offset);
 352                         if (IS_ERR(qgroup)) {
 353                                 ret = PTR_ERR(qgroup);
 354                                 goto out;
 355                         }
 356                 }
 357                 switch (found_key.type) {
 358                 case BTRFS_QGROUP_INFO_KEY: {
 359                         struct btrfs_qgroup_info_item *ptr;
 360
 361                         ptr = btrfs_item_ptr(l, slot,
 362                                              struct btrfs_qgroup_info_item);
 363                         qgroup->rfer = btrfs_qgroup_info_rfer(l, ptr);
 364                         qgroup->rfer_cmpr = btrfs_qgroup_info_rfer_cmpr(l, ptr);
 365                         qgroup->excl = btrfs_qgroup_info_excl(l, ptr);
 366                         qgroup->excl_cmpr = btrfs_qgroup_info_excl_cmpr(l, ptr);
 367                         /* generation currently unused */
 368                         break;
 369                 }
 370                 case BTRFS_QGROUP_LIMIT_KEY: {
 371                         struct btrfs_qgroup_limit_item *ptr;
 372
 373                         ptr = btrfs_item_ptr(l, slot,
 374                                              struct btrfs_qgroup_limit_item);
 375                         qgroup->lim_flags = btrfs_qgroup_limit_flags(l, ptr);
 376                         qgroup->max_rfer = btrfs_qgroup_limit_max_rfer(l, ptr);
 377                         qgroup->max_excl = btrfs_qgroup_limit_max_excl(l, ptr);
 378                         qgroup->rsv_rfer = btrfs_qgroup_limit_rsv_rfer(l, ptr);
 379                         qgroup->rsv_excl = btrfs_qgroup_limit_rsv_excl(l, ptr);
 380                         break;
 381                 }
 382                 }
 383 next1:
 384                 ret = btrfs_next_item(quota_root, path);
 385                 if (ret < 0)
 386                         goto out;
 387                 if (ret)
 388                         break;
 389         }
 390         btrfs_release_path(path);
 391
 392         /*
 393          * pass 2: read all qgroup relations
 394          */
 395         key.objectid = 0;
 396         key.type = BTRFS_QGROUP_RELATION_KEY;
 397         key.offset = 0;
 398         ret = btrfs_search_slot_for_read(quota_root, &key, path, 1, 0);
 399         if (ret)
 400                 goto out;
 401         while (1) {
 402                 slot = path->slots[0];
 403                 l = path->nodes[0];
 404                 btrfs_item_key_to_cpu(l, &found_key, slot);
 405
 406                 if (found_key.type != BTRFS_QGROUP_RELATION_KEY)
 407                         goto next2;
 408
 409                 if (found_key.objectid > found_key.offset) {
 410                         /* parent <- member, not needed to build config */
 411                         /* FIXME should we omit the key completely? */
 412                         goto next2;
 413                 }
 414
 415                 ret = add_relation_rb(fs_info, found_key.objectid,
 416                                       found_key.offset);
 417                 if (ret == -ENOENT) {
 418                         btrfs_warn(fs_info,
 419                                 "orphan qgroup relation 0x%llx->0x%llx",
 420                                 found_key.objectid, found_key.offset);
 421                         ret = 0;        /* ignore the error */
 422                 }
 423                 if (ret)
 424                         goto out;
 425 next2:
 426                 ret = btrfs_next_item(quota_root, path);
 427                 if (ret < 0)
 428                         goto out;
 429                 if (ret)
 430                         break;
 431         }
 432 out:
 433         fs_info->qgroup_flags |= flags;
 434         if (!(fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_ON)) {
 435                 fs_info->quota_enabled = 0;
 436                 fs_info->pending_quota_state = 0;
 437         } else if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN &&
 438                    ret >= 0) {
 439                 ret = qgroup_rescan_init(fs_info, rescan_progress, 0);
 440         }
 441         btrfs_free_path(path);
 442
 443         if (ret < 0) {
 444                 ulist_free(fs_info->qgroup_ulist);
 445                 fs_info->qgroup_ulist = NULL;
 446                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
 447         }
 448
 449         return ret < 0 ? ret : 0;
 450 }
 451
 452 /*
 453  * This is called from close_ctree() or open_ctree() or btrfs_quota_disable(),
 454  * first two are in single-threaded paths.And for the third one, we have set
 455  * quota_root to be null with qgroup_lock held before, so it is safe to clean
 456  * up the in-memory structures without qgroup_lock held.
 457  */
 458 void btrfs_free_qgroup_config(struct btrfs_fs_info *fs_info)
 459 {
 460         struct rb_node *n;
 461         struct btrfs_qgroup *qgroup;
 462
 463         while ((n = rb_first(&fs_info->qgroup_tree))) {
 464                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
 465                 rb_erase(n, &fs_info->qgroup_tree);
 466                 __del_qgroup_rb(qgroup);
 467         }
 468         /*
 469          * we call btrfs_free_qgroup_config() when umounting
 470          * filesystem and disabling quota, so we set qgroup_ulit
 471          * to be null here to avoid double free.
 472          */
 473         ulist_free(fs_info->qgroup_ulist);
 474         fs_info->qgroup_ulist = NULL;
 475 }
 476
 477 static int add_qgroup_relation_item(struct btrfs_trans_handle *trans,
 478                                     struct btrfs_root *quota_root,
 479                                     u64 src, u64 dst)
 480 {
 481         int ret;
 482         struct btrfs_path *path;
 483         struct btrfs_key key;
 484
 485         path = btrfs_alloc_path();
 486         if (!path)
 487                 return -ENOMEM;
 488
 489         key.objectid = src;
 490         key.type = BTRFS_QGROUP_RELATION_KEY;
 491         key.offset = dst;
 492
 493         ret = btrfs_insert_empty_item(trans, quota_root, path, &key, 0);
 494
 495         btrfs_mark_buffer_dirty(path->nodes[0]);
 496
 497         btrfs_free_path(path);
 498         return ret;
 499 }
 500
 501 static int del_qgroup_relation_item(struct btrfs_trans_handle *trans,
 502                                     struct btrfs_root *quota_root,
 503                                     u64 src, u64 dst)
 504 {
 505         int ret;
 506         struct btrfs_path *path;
 507         struct btrfs_key key;
 508
 509         path = btrfs_alloc_path();
 510         if (!path)
 511                 return -ENOMEM;
 512
 513         key.objectid = src;
 514         key.type = BTRFS_QGROUP_RELATION_KEY;
 515         key.offset = dst;
 516
 517         ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
 518         if (ret < 0)
 519                 goto out;
 520
 521         if (ret > 0) {
 522                 ret = -ENOENT;
 523                 goto out;
 524         }
 525
 526         ret = btrfs_del_item(trans, quota_root, path);
 527 out:
 528         btrfs_free_path(path);
 529         return ret;
 530 }
 531
 532 static int add_qgroup_item(struct btrfs_trans_handle *trans,
 533                            struct btrfs_root *quota_root, u64 qgroupid)
 534 {
 535         int ret;
 536         struct btrfs_path *path;
 537         struct btrfs_qgroup_info_item *qgroup_info;
 538         struct btrfs_qgroup_limit_item *qgroup_limit;
 539         struct extent_buffer *leaf;
 540         struct btrfs_key key;
 541
 542 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 543         if (unlikely(test_bit(BTRFS_ROOT_DUMMY_ROOT, &quota_root->state)))
 544                 return 0;
 545 #endif
 546         path = btrfs_alloc_path();
 547         if (!path)
 548                 return -ENOMEM;
 549
 550         key.objectid = 0;
 551         key.type = BTRFS_QGROUP_INFO_KEY;
 552         key.offset = qgroupid;
 553
 554         ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
 555                                       sizeof(*qgroup_info));
 556         if (ret)
 557                 goto out;
 558
 559         leaf = path->nodes[0];
 560         qgroup_info = btrfs_item_ptr(leaf, path->slots[0],
 561                                  struct btrfs_qgroup_info_item);
 562         btrfs_set_qgroup_info_generation(leaf, qgroup_info, trans->transid);
 563         btrfs_set_qgroup_info_rfer(leaf, qgroup_info, 0);
 564         btrfs_set_qgroup_info_rfer_cmpr(leaf, qgroup_info, 0);
 565         btrfs_set_qgroup_info_excl(leaf, qgroup_info, 0);
 566         btrfs_set_qgroup_info_excl_cmpr(leaf, qgroup_info, 0);
 567
 568         btrfs_mark_buffer_dirty(leaf);
 569
 570         btrfs_release_path(path);
 571
 572         key.type = BTRFS_QGROUP_LIMIT_KEY;
 573         ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
 574                                       sizeof(*qgroup_limit));
 575         if (ret)
 576                 goto out;
 577
 578         leaf = path->nodes[0];
 579         qgroup_limit = btrfs_item_ptr(leaf, path->slots[0],
 580                                   struct btrfs_qgroup_limit_item);
 581         btrfs_set_qgroup_limit_flags(leaf, qgroup_limit, 0);
 582         btrfs_set_qgroup_limit_max_rfer(leaf, qgroup_limit, 0);
 583         btrfs_set_qgroup_limit_max_excl(leaf, qgroup_limit, 0);
 584         btrfs_set_qgroup_limit_rsv_rfer(leaf, qgroup_limit, 0);
 585         btrfs_set_qgroup_limit_rsv_excl(leaf, qgroup_limit, 0);
 586
 587         btrfs_mark_buffer_dirty(leaf);
 588
 589         ret = 0;
 590 out:
 591         btrfs_free_path(path);
 592         return ret;
 593 }
 594
 595 static int del_qgroup_item(struct btrfs_trans_handle *trans,
 596                            struct btrfs_root *quota_root, u64 qgroupid)
 597 {
 598         int ret;
 599         struct btrfs_path *path;
 600         struct btrfs_key key;
 601
 602         path = btrfs_alloc_path();
 603         if (!path)
 604                 return -ENOMEM;
 605
 606         key.objectid = 0;
 607         key.type = BTRFS_QGROUP_INFO_KEY;
 608         key.offset = qgroupid;
 609         ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
 610         if (ret < 0)
 611                 goto out;
 612
 613         if (ret > 0) {
 614                 ret = -ENOENT;
 615                 goto out;
 616         }
 617
 618         ret = btrfs_del_item(trans, quota_root, path);
 619         if (ret)
 620                 goto out;
 621
 622         btrfs_release_path(path);
 623
 624         key.type = BTRFS_QGROUP_LIMIT_KEY;
 625         ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
 626         if (ret < 0)
 627                 goto out;
 628
 629         if (ret > 0) {
 630                 ret = -ENOENT;
 631                 goto out;
 632         }
 633
 634         ret = btrfs_del_item(trans, quota_root, path);
 635
 636 out:
 637         btrfs_free_path(path);
 638         return ret;
 639 }
 640
 641 static int update_qgroup_limit_item(struct btrfs_trans_handle *trans,
 642                                     struct btrfs_root *root, u64 qgroupid,
 643                                     u64 flags, u64 max_rfer, u64 max_excl,
 644                                     u64 rsv_rfer, u64 rsv_excl)
 645 {
 646         struct btrfs_path *path;
 647         struct btrfs_key key;
 648         struct extent_buffer *l;
 649         struct btrfs_qgroup_limit_item *qgroup_limit;
 650         int ret;
 651         int slot;
 652
 653         key.objectid = 0;
 654         key.type = BTRFS_QGROUP_LIMIT_KEY;
 655         key.offset = qgroupid;
 656
 657         path = btrfs_alloc_path();
 658         if (!path)
 659                 return -ENOMEM;
 660
 661         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
 662         if (ret > 0)
 663                 ret = -ENOENT;
 664
 665         if (ret)
 666                 goto out;
 667
 668         l = path->nodes[0];
 669         slot = path->slots[0];
 670         qgroup_limit = btrfs_item_ptr(l, slot, struct btrfs_qgroup_limit_item);
 671         btrfs_set_qgroup_limit_flags(l, qgroup_limit, flags);
 672         btrfs_set_qgroup_limit_max_rfer(l, qgroup_limit, max_rfer);
 673         btrfs_set_qgroup_limit_max_excl(l, qgroup_limit, max_excl);
 674         btrfs_set_qgroup_limit_rsv_rfer(l, qgroup_limit, rsv_rfer);
 675         btrfs_set_qgroup_limit_rsv_excl(l, qgroup_limit, rsv_excl);
 676
 677         btrfs_mark_buffer_dirty(l);
 678
 679 out:
 680         btrfs_free_path(path);
 681         return ret;
 682 }
 683
 684 static int update_qgroup_info_item(struct btrfs_trans_handle *trans,
 685                                    struct btrfs_root *root,
 686                                    struct btrfs_qgroup *qgroup)
 687 {
 688         struct btrfs_path *path;
 689         struct btrfs_key key;
 690         struct extent_buffer *l;
 691         struct btrfs_qgroup_info_item *qgroup_info;
 692         int ret;
 693         int slot;
 694
 695 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 696         if (unlikely(test_bit(BTRFS_ROOT_DUMMY_ROOT, &root->state)))
 697                 return 0;
 698 #endif
 699         key.objectid = 0;
 700         key.type = BTRFS_QGROUP_INFO_KEY;
 701         key.offset = qgroup->qgroupid;
 702
 703         path = btrfs_alloc_path();
 704         if (!path)
 705                 return -ENOMEM;
 706
 707         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
 708         if (ret > 0)
 709                 ret = -ENOENT;
 710
 711         if (ret)
 712                 goto out;
 713
 714         l = path->nodes[0];
 715         slot = path->slots[0];
 716         qgroup_info = btrfs_item_ptr(l, slot, struct btrfs_qgroup_info_item);
 717         btrfs_set_qgroup_info_generation(l, qgroup_info, trans->transid);
 718         btrfs_set_qgroup_info_rfer(l, qgroup_info, qgroup->rfer);
 719         btrfs_set_qgroup_info_rfer_cmpr(l, qgroup_info, qgroup->rfer_cmpr);
 720         btrfs_set_qgroup_info_excl(l, qgroup_info, qgroup->excl);
 721         btrfs_set_qgroup_info_excl_cmpr(l, qgroup_info, qgroup->excl_cmpr);
 722
 723         btrfs_mark_buffer_dirty(l);
 724
 725 out:
 726         btrfs_free_path(path);
 727         return ret;
 728 }
 729
 730 static int update_qgroup_status_item(struct btrfs_trans_handle *trans,
 731                                      struct btrfs_fs_info *fs_info,
 732                                     struct btrfs_root *root)
 733 {
 734         struct btrfs_path *path;
 735         struct btrfs_key key;
 736         struct extent_buffer *l;
 737         struct btrfs_qgroup_status_item *ptr;
 738         int ret;
 739         int slot;
 740
 741         key.objectid = 0;
 742         key.type = BTRFS_QGROUP_STATUS_KEY;
 743         key.offset = 0;
 744
 745         path = btrfs_alloc_path();
 746         if (!path)
 747                 return -ENOMEM;
 748
 749         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
 750         if (ret > 0)
 751                 ret = -ENOENT;
 752
 753         if (ret)
 754                 goto out;
 755
 756         l = path->nodes[0];
 757         slot = path->slots[0];
 758         ptr = btrfs_item_ptr(l, slot, struct btrfs_qgroup_status_item);
 759         btrfs_set_qgroup_status_flags(l, ptr, fs_info->qgroup_flags);
 760         btrfs_set_qgroup_status_generation(l, ptr, trans->transid);
 761         btrfs_set_qgroup_status_rescan(l, ptr,
 762                                 fs_info->qgroup_rescan_progress.objectid);
 763
 764         btrfs_mark_buffer_dirty(l);
 765
 766 out:
 767         btrfs_free_path(path);
 768         return ret;
 769 }
 770
 771 /*
 772  * called with qgroup_lock held
 773  */
 774 static int btrfs_clean_quota_tree(struct btrfs_trans_handle *trans,
 775                                   struct btrfs_root *root)
 776 {
 777         struct btrfs_path *path;
 778         struct btrfs_key key;
 779         struct extent_buffer *leaf = NULL;
 780         int ret;
 781         int nr = 0;
 782
 783         path = btrfs_alloc_path();
 784         if (!path)
 785                 return -ENOMEM;
 786
 787         path->leave_spinning = 1;
 788
 789         key.objectid = 0;
 790         key.offset = 0;
 791         key.type = 0;
 792
 793         while (1) {
 794                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 795                 if (ret < 0)
 796                         goto out;
 797                 leaf = path->nodes[0];
 798                 nr = btrfs_header_nritems(leaf);
 799                 if (!nr)
 800                         break;
 801                 /*
 802                  * delete the leaf one by one
 803                  * since the whole tree is going
 804                  * to be deleted.
 805                  */
 806                 path->slots[0] = 0;
 807                 ret = btrfs_del_items(trans, root, path, 0, nr);
 808                 if (ret)
 809                         goto out;
 810
 811                 btrfs_release_path(path);
 812         }
 813         ret = 0;
 814 out:
 815         root->fs_info->pending_quota_state = 0;
 816         btrfs_free_path(path);
 817         return ret;
 818 }
 819
 820 int btrfs_quota_enable(struct btrfs_trans_handle *trans,
 821                        struct btrfs_fs_info *fs_info)
 822 {
 823         struct btrfs_root *quota_root;
 824         struct btrfs_root *tree_root = fs_info->tree_root;
 825         struct btrfs_path *path = NULL;
 826         struct btrfs_qgroup_status_item *ptr;
 827         struct extent_buffer *leaf;
 828         struct btrfs_key key;
 829         struct btrfs_key found_key;
 830         struct btrfs_qgroup *qgroup = NULL;
 831         int ret = 0;
 832         int slot;
 833
 834         mutex_lock(&fs_info->qgroup_ioctl_lock);
 835         if (fs_info->quota_root) {
 836                 fs_info->pending_quota_state = 1;
 837                 goto out;
 838         }
 839
 840         fs_info->qgroup_ulist = ulist_alloc(GFP_NOFS);
 841         if (!fs_info->qgroup_ulist) {
 842                 ret = -ENOMEM;
 843                 goto out;
 844         }
 845
 846         /*
 847          * initially create the quota tree
 848          */
 849         quota_root = btrfs_create_tree(trans, fs_info,
 850                                        BTRFS_QUOTA_TREE_OBJECTID);
 851         if (IS_ERR(quota_root)) {
 852                 ret =  PTR_ERR(quota_root);
 853                 goto out;
 854         }
 855
 856         path = btrfs_alloc_path();
 857         if (!path) {
 858                 ret = -ENOMEM;
 859                 goto out_free_root;
 860         }
 861
 862         key.objectid = 0;
 863         key.type = BTRFS_QGROUP_STATUS_KEY;
 864         key.offset = 0;
 865
 866         ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
 867                                       sizeof(*ptr));
 868         if (ret)
 869                 goto out_free_path;
 870
 871         leaf = path->nodes[0];
 872         ptr = btrfs_item_ptr(leaf, path->slots[0],
 873                                  struct btrfs_qgroup_status_item);
 874         btrfs_set_qgroup_status_generation(leaf, ptr, trans->transid);
 875         btrfs_set_qgroup_status_version(leaf, ptr, BTRFS_QGROUP_STATUS_VERSION);
 876         fs_info->qgroup_flags = BTRFS_QGROUP_STATUS_FLAG_ON |
 877                                 BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
 878         btrfs_set_qgroup_status_flags(leaf, ptr, fs_info->qgroup_flags);
 879         btrfs_set_qgroup_status_rescan(leaf, ptr, 0);
 880
 881         btrfs_mark_buffer_dirty(leaf);
 882
 883         key.objectid = 0;
 884         key.type = BTRFS_ROOT_REF_KEY;
 885         key.offset = 0;
 886
 887         btrfs_release_path(path);
 888         ret = btrfs_search_slot_for_read(tree_root, &key, path, 1, 0);
 889         if (ret > 0)
 890                 goto out_add_root;
 891         if (ret < 0)
 892                 goto out_free_path;
 893
 894
 895         while (1) {
 896                 slot = path->slots[0];
 897                 leaf = path->nodes[0];
 898                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
 899
 900                 if (found_key.type == BTRFS_ROOT_REF_KEY) {
 901                         ret = add_qgroup_item(trans, quota_root,
 902                                               found_key.offset);
 903                         if (ret)
 904                                 goto out_free_path;
 905
 906                         qgroup = add_qgroup_rb(fs_info, found_key.offset);
 907                         if (IS_ERR(qgroup)) {
 908                                 ret = PTR_ERR(qgroup);
 909                                 goto out_free_path;
 910                         }
 911                 }
 912                 ret = btrfs_next_item(tree_root, path);
 913                 if (ret < 0)
 914                         goto out_free_path;
 915                 if (ret)
 916                         break;
 917         }
 918
 919 out_add_root:
 920         btrfs_release_path(path);
 921         ret = add_qgroup_item(trans, quota_root, BTRFS_FS_TREE_OBJECTID);
 922         if (ret)
 923                 goto out_free_path;
 924
 925         qgroup = add_qgroup_rb(fs_info, BTRFS_FS_TREE_OBJECTID);
 926         if (IS_ERR(qgroup)) {
 927                 ret = PTR_ERR(qgroup);
 928                 goto out_free_path;
 929         }
 930         spin_lock(&fs_info->qgroup_lock);
 931         fs_info->quota_root = quota_root;
 932         fs_info->pending_quota_state = 1;
 933         spin_unlock(&fs_info->qgroup_lock);
 934 out_free_path:
 935         btrfs_free_path(path);
 936 out_free_root:
 937         if (ret) {
 938                 free_extent_buffer(quota_root->node);
 939                 free_extent_buffer(quota_root->commit_root);
 940                 kfree(quota_root);
 941         }
 942 out:
 943         if (ret) {
 944                 ulist_free(fs_info->qgroup_ulist);
 945                 fs_info->qgroup_ulist = NULL;
 946         }
 947         mutex_unlock(&fs_info->qgroup_ioctl_lock);
 948         return ret;
 949 }
 950
 951 int btrfs_quota_disable(struct btrfs_trans_handle *trans,
 952                         struct btrfs_fs_info *fs_info)
 953 {
 954         struct btrfs_root *tree_root = fs_info->tree_root;
 955         struct btrfs_root *quota_root;
 956         int ret = 0;
 957
 958         mutex_lock(&fs_info->qgroup_ioctl_lock);
 959         if (!fs_info->quota_root)
 960                 goto out;
 961         spin_lock(&fs_info->qgroup_lock);
 962         fs_info->quota_enabled = 0;
 963         fs_info->pending_quota_state = 0;
 964         quota_root = fs_info->quota_root;
 965         fs_info->quota_root = NULL;
 966         spin_unlock(&fs_info->qgroup_lock);
 967
 968         btrfs_free_qgroup_config(fs_info);
 969
 970         ret = btrfs_clean_quota_tree(trans, quota_root);
 971         if (ret)
 972                 goto out;
 973
 974         ret = btrfs_del_root(trans, tree_root, &quota_root->root_key);
 975         if (ret)
 976                 goto out;
 977
 978         list_del(&quota_root->dirty_list);
 979
 980         btrfs_tree_lock(quota_root->node);
 981         clean_tree_block(trans, tree_root, quota_root->node);
 982         btrfs_tree_unlock(quota_root->node);
 983         btrfs_free_tree_block(trans, quota_root, quota_root->node, 0, 1);
 984
 985         free_extent_buffer(quota_root->node);
 986         free_extent_buffer(quota_root->commit_root);
 987         kfree(quota_root);
 988 out:
 989         mutex_unlock(&fs_info->qgroup_ioctl_lock);
 990         return ret;
 991 }
 992
 993 static void qgroup_dirty(struct btrfs_fs_info *fs_info,
 994                          struct btrfs_qgroup *qgroup)
 995 {
 996         if (list_empty(&qgroup->dirty))
 997                 list_add(&qgroup->dirty, &fs_info->dirty_qgroups);
 998 }
 999
1000 int btrfs_add_qgroup_relation(struct btrfs_trans_handle *trans,
1001                               struct btrfs_fs_info *fs_info, u64 src, u64 dst)
1002 {
1003         struct btrfs_root *quota_root;
1004         struct btrfs_qgroup *parent;
1005         struct btrfs_qgroup *member;
1006         struct btrfs_qgroup_list *list;
1007         int ret = 0;
1008
1009         mutex_lock(&fs_info->qgroup_ioctl_lock);
1010         quota_root = fs_info->quota_root;
1011         if (!quota_root) {
1012                 ret = -EINVAL;
1013                 goto out;
1014         }
1015         member = find_qgroup_rb(fs_info, src);
1016         parent = find_qgroup_rb(fs_info, dst);
1017         if (!member || !parent) {
1018                 ret = -EINVAL;
1019                 goto out;
1020         }
1021
1022         /* check if such qgroup relation exist firstly */
1023         list_for_each_entry(list, &member->groups, next_group) {
1024                 if (list->group == parent) {
1025                         ret = -EEXIST;
1026                         goto out;
1027                 }
1028         }
1029
1030         ret = add_qgroup_relation_item(trans, quota_root, src, dst);
1031         if (ret)
1032                 goto out;
1033
1034         ret = add_qgroup_relation_item(trans, quota_root, dst, src);
1035         if (ret) {
1036                 del_qgroup_relation_item(trans, quota_root, src, dst);
1037                 goto out;
1038         }
1039
1040         spin_lock(&fs_info->qgroup_lock);
1041         ret = add_relation_rb(quota_root->fs_info, src, dst);
1042         spin_unlock(&fs_info->qgroup_lock);
1043 out:
1044         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1045         return ret;
1046 }
1047
1048 int btrfs_del_qgroup_relation(struct btrfs_trans_handle *trans,
1049                               struct btrfs_fs_info *fs_info, u64 src, u64 dst)
1050 {
1051         struct btrfs_root *quota_root;
1052         struct btrfs_qgroup *parent;
1053         struct btrfs_qgroup *member;
1054         struct btrfs_qgroup_list *list;
1055         int ret = 0;
1056         int err;
1057
1058         mutex_lock(&fs_info->qgroup_ioctl_lock);
1059         quota_root = fs_info->quota_root;
1060         if (!quota_root) {
1061                 ret = -EINVAL;
1062                 goto out;
1063         }
1064
1065         member = find_qgroup_rb(fs_info, src);
1066         parent = find_qgroup_rb(fs_info, dst);
1067         if (!member || !parent) {
1068                 ret = -EINVAL;
1069                 goto out;
1070         }
1071
1072         /* check if such qgroup relation exist firstly */
1073         list_for_each_entry(list, &member->groups, next_group) {
1074                 if (list->group == parent)
1075                         goto exist;
1076         }
1077         ret = -ENOENT;
1078         goto out;
1079 exist:
1080         ret = del_qgroup_relation_item(trans, quota_root, src, dst);
1081         err = del_qgroup_relation_item(trans, quota_root, dst, src);
1082         if (err && !ret)
1083                 ret = err;
1084
1085         spin_lock(&fs_info->qgroup_lock);
1086         del_relation_rb(fs_info, src, dst);
1087         spin_unlock(&fs_info->qgroup_lock);
1088 out:
1089         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1090         return ret;
1091 }
1092
1093 int btrfs_create_qgroup(struct btrfs_trans_handle *trans,
1094                         struct btrfs_fs_info *fs_info, u64 qgroupid, char *name)
1095 {
1096         struct btrfs_root *quota_root;
1097         struct btrfs_qgroup *qgroup;
1098         int ret = 0;
1099
1100         mutex_lock(&fs_info->qgroup_ioctl_lock);
1101         quota_root = fs_info->quota_root;
1102         if (!quota_root) {
1103                 ret = -EINVAL;
1104                 goto out;
1105         }
1106         qgroup = find_qgroup_rb(fs_info, qgroupid);
1107         if (qgroup) {
1108                 ret = -EEXIST;
1109                 goto out;
1110         }
1111
1112         ret = add_qgroup_item(trans, quota_root, qgroupid);
1113         if (ret)
1114                 goto out;
1115
1116         spin_lock(&fs_info->qgroup_lock);
1117         qgroup = add_qgroup_rb(fs_info, qgroupid);
1118         spin_unlock(&fs_info->qgroup_lock);
1119
1120         if (IS_ERR(qgroup))
1121                 ret = PTR_ERR(qgroup);
1122 out:
1123         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1124         return ret;
1125 }
1126
1127 int btrfs_remove_qgroup(struct btrfs_trans_handle *trans,
1128                         struct btrfs_fs_info *fs_info, u64 qgroupid)
1129 {
1130         struct btrfs_root *quota_root;
1131         struct btrfs_qgroup *qgroup;
1132         int ret = 0;
1133
1134         mutex_lock(&fs_info->qgroup_ioctl_lock);
1135         quota_root = fs_info->quota_root;
1136         if (!quota_root) {
1137                 ret = -EINVAL;
1138                 goto out;
1139         }
1140
1141         qgroup = find_qgroup_rb(fs_info, qgroupid);
1142         if (!qgroup) {
1143                 ret = -ENOENT;
1144                 goto out;
1145         } else {
1146                 /* check if there are no relations to this qgroup */
1147                 if (!list_empty(&qgroup->groups) ||
1148                     !list_empty(&qgroup->members)) {
1149                         ret = -EBUSY;
1150                         goto out;
1151                 }
1152         }
1153         ret = del_qgroup_item(trans, quota_root, qgroupid);
1154
1155         spin_lock(&fs_info->qgroup_lock);
1156         del_qgroup_rb(quota_root->fs_info, qgroupid);
1157         spin_unlock(&fs_info->qgroup_lock);
1158 out:
1159         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1160         return ret;
1161 }
1162
1163 int btrfs_limit_qgroup(struct btrfs_trans_handle *trans,
1164                        struct btrfs_fs_info *fs_info, u64 qgroupid,
1165                        struct btrfs_qgroup_limit *limit)
1166 {
1167         struct btrfs_root *quota_root;
1168         struct btrfs_qgroup *qgroup;
1169         int ret = 0;
1170
1171         mutex_lock(&fs_info->qgroup_ioctl_lock);
1172         quota_root = fs_info->quota_root;
1173         if (!quota_root) {
1174                 ret = -EINVAL;
1175                 goto out;
1176         }
1177
1178         qgroup = find_qgroup_rb(fs_info, qgroupid);
1179         if (!qgroup) {
1180                 ret = -ENOENT;
1181                 goto out;
1182         }
1183         ret = update_qgroup_limit_item(trans, quota_root, qgroupid,
1184                                        limit->flags, limit->max_rfer,
1185                                        limit->max_excl, limit->rsv_rfer,
1186                                        limit->rsv_excl);
1187         if (ret) {
1188                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1189                 btrfs_info(fs_info, "unable to update quota limit for %llu",
1190                        qgroupid);
1191         }
1192
1193         spin_lock(&fs_info->qgroup_lock);
1194         qgroup->lim_flags = limit->flags;
1195         qgroup->max_rfer = limit->max_rfer;
1196         qgroup->max_excl = limit->max_excl;
1197         qgroup->rsv_rfer = limit->rsv_rfer;
1198         qgroup->rsv_excl = limit->rsv_excl;
1199         spin_unlock(&fs_info->qgroup_lock);
1200 out:
1201         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1202         return ret;
1203 }
1204
1205 static int comp_oper_exist(struct btrfs_qgroup_operation *oper1,
1206                            struct btrfs_qgroup_operation *oper2)
1207 {
1208         /*
1209          * Ignore seq and type here, we're looking for any operation
1210          * at all related to this extent on that root.
1211          */
1212         if (oper1->bytenr < oper2->bytenr)
1213                 return -1;
1214         if (oper1->bytenr > oper2->bytenr)
1215                 return 1;
1216         if (oper1->ref_root < oper2->ref_root)
1217                 return -1;
1218         if (oper1->ref_root > oper2->ref_root)
1219                 return 1;
1220         return 0;
1221 }
1222
1223 static int qgroup_oper_exists(struct btrfs_fs_info *fs_info,
1224                               struct btrfs_qgroup_operation *oper)
1225 {
1226         struct rb_node *n;
1227         struct btrfs_qgroup_operation *cur;
1228         int cmp;
1229
1230         spin_lock(&fs_info->qgroup_op_lock);
1231         n = fs_info->qgroup_op_tree.rb_node;
1232         while (n) {
1233                 cur = rb_entry(n, struct btrfs_qgroup_operation, n);
1234                 cmp = comp_oper_exist(cur, oper);
1235                 if (cmp < 0) {
1236                         n = n->rb_right;
1237                 } else if (cmp) {
1238                         n = n->rb_left;
1239                 } else {
1240                         spin_unlock(&fs_info->qgroup_op_lock);
1241                         return -EEXIST;
1242                 }
1243         }
1244         spin_unlock(&fs_info->qgroup_op_lock);
1245         return 0;
1246 }
1247
1248 static int comp_oper(struct btrfs_qgroup_operation *oper1,
1249                      struct btrfs_qgroup_operation *oper2)
1250 {
1251         if (oper1->bytenr < oper2->bytenr)
1252                 return -1;
1253         if (oper1->bytenr > oper2->bytenr)
1254                 return 1;
1255         if (oper1->seq < oper2->seq)
1256                 return -1;
1257         if (oper1->seq > oper2->seq)
1258                 return -1;
1259         if (oper1->ref_root < oper2->ref_root)
1260                 return -1;
1261         if (oper1->ref_root > oper2->ref_root)
1262                 return 1;
1263         if (oper1->type < oper2->type)
1264                 return -1;
1265         if (oper1->type > oper2->type)
1266                 return 1;
1267         return 0;
1268 }
1269
1270 static int insert_qgroup_oper(struct btrfs_fs_info *fs_info,
1271                               struct btrfs_qgroup_operation *oper)
1272 {
1273         struct rb_node **p;
1274         struct rb_node *parent = NULL;
1275         struct btrfs_qgroup_operation *cur;
1276         int cmp;
1277
1278         spin_lock(&fs_info->qgroup_op_lock);
1279         p = &fs_info->qgroup_op_tree.rb_node;
1280         while (*p) {
1281                 parent = *p;
1282                 cur = rb_entry(parent, struct btrfs_qgroup_operation, n);
1283                 cmp = comp_oper(cur, oper);
1284                 if (cmp < 0) {
1285                         p = &(*p)->rb_right;
1286                 } else if (cmp) {
1287                         p = &(*p)->rb_left;
1288                 } else {
1289                         spin_unlock(&fs_info->qgroup_op_lock);
1290                         return -EEXIST;
1291                 }
1292         }
1293         rb_link_node(&oper->n, parent, p);
1294         rb_insert_color(&oper->n, &fs_info->qgroup_op_tree);
1295         spin_unlock(&fs_info->qgroup_op_lock);
1296         return 0;
1297 }
1298
1299 /*
1300  * Record a quota operation for processing later on.
1301  * @trans: the transaction we are adding the delayed op to.
1302  * @fs_info: the fs_info for this fs.
1303  * @ref_root: the root of the reference we are acting on,
1304  * @bytenr: the bytenr we are acting on.
1305  * @num_bytes: the number of bytes in the reference.
1306  * @type: the type of operation this is.
1307  * @mod_seq: do we need to get a sequence number for looking up roots.
1308  *
1309  * We just add it to our trans qgroup_ref_list and carry on and process these
1310  * operations in order at some later point.  If the reference root isn't a fs
1311  * root then we don't bother with doing anything.
1312  *
1313  * MUST BE HOLDING THE REF LOCK.
1314  */
1315 int btrfs_qgroup_record_ref(struct btrfs_trans_handle *trans,
1316                             struct btrfs_fs_info *fs_info, u64 ref_root,
1317                             u64 bytenr, u64 num_bytes,
1318                             enum btrfs_qgroup_operation_type type, int mod_seq)
1319 {
1320         struct btrfs_qgroup_operation *oper;
1321         int ret;
1322
1323         if (!is_fstree(ref_root) || !fs_info->quota_enabled)
1324                 return 0;
1325
1326         oper = kmalloc(sizeof(*oper), GFP_NOFS);
1327         if (!oper)
1328                 return -ENOMEM;
1329
1330         oper->ref_root = ref_root;
1331         oper->bytenr = bytenr;
1332         oper->num_bytes = num_bytes;
1333         oper->type = type;
1334         oper->seq = atomic_inc_return(&fs_info->qgroup_op_seq);
1335         INIT_LIST_HEAD(&oper->elem.list);
1336         oper->elem.seq = 0;
1337
1338         if (type == BTRFS_QGROUP_OPER_SUB_SUBTREE) {
1339                 /*
1340                  * If any operation for this bytenr/ref_root combo
1341                  * exists, then we know it's not exclusively owned and
1342                  * shouldn't be queued up.
1343                  *
1344                  * This also catches the case where we have a cloned
1345                  * extent that gets queued up multiple times during
1346                  * drop snapshot.
1347                  */
1348                 if (qgroup_oper_exists(fs_info, oper)) {
1349                         kfree(oper);
1350                         return 0;
1351                 }
1352         }
1353
1354         ret = insert_qgroup_oper(fs_info, oper);
1355         if (ret) {
1356                 /* Shouldn't happen so have an assert for developers */
1357                 ASSERT(0);
1358                 kfree(oper);
1359                 return ret;
1360         }
1361         list_add_tail(&oper->list, &trans->qgroup_ref_list);
1362
1363         if (mod_seq)
1364                 btrfs_get_tree_mod_seq(fs_info, &oper->elem);
1365
1366         return 0;
1367 }
1368
1369 /*
1370  * The easy accounting, if we are adding/removing the only ref for an extent
1371  * then this qgroup and all of the parent qgroups get their refrence and
1372  * exclusive counts adjusted.
1373  */
1374 static int qgroup_excl_accounting(struct btrfs_fs_info *fs_info,
1375                                   struct btrfs_qgroup_operation *oper)
1376 {
1377         struct btrfs_qgroup *qgroup;
1378         struct ulist *tmp;
1379         struct btrfs_qgroup_list *glist;
1380         struct ulist_node *unode;
1381         struct ulist_iterator uiter;
1382         int sign = 0;
1383         int ret = 0;
1384
1385         tmp = ulist_alloc(GFP_NOFS);
1386         if (!tmp)
1387                 return -ENOMEM;
1388
1389         spin_lock(&fs_info->qgroup_lock);
1390         if (!fs_info->quota_root)
1391                 goto out;
1392         qgroup = find_qgroup_rb(fs_info, oper->ref_root);
1393         if (!qgroup)
1394                 goto out;
1395         switch (oper->type) {
1396         case BTRFS_QGROUP_OPER_ADD_EXCL:
1397                 sign = 1;
1398                 break;
1399         case BTRFS_QGROUP_OPER_SUB_EXCL:
1400                 sign = -1;
1401                 break;
1402         default:
1403                 ASSERT(0);
1404         }
1405         qgroup->rfer += sign * oper->num_bytes;
1406         qgroup->rfer_cmpr += sign * oper->num_bytes;
1407
1408         WARN_ON(sign < 0 && qgroup->excl < oper->num_bytes);
1409         qgroup->excl += sign * oper->num_bytes;
1410         qgroup->excl_cmpr += sign * oper->num_bytes;
1411
1412         qgroup_dirty(fs_info, qgroup);
1413
1414         /* Get all of the parent groups that contain this qgroup */
1415         list_for_each_entry(glist, &qgroup->groups, next_group) {
1416                 ret = ulist_add(tmp, glist->group->qgroupid,
1417                                 ptr_to_u64(glist->group), GFP_ATOMIC);
1418                 if (ret < 0)
1419                         goto out;
1420         }
1421
1422         /* Iterate all of the parents and adjust their reference counts */
1423         ULIST_ITER_INIT(&uiter);
1424         while ((unode = ulist_next(tmp, &uiter))) {
1425                 qgroup = u64_to_ptr(unode->aux);
1426                 qgroup->rfer += sign * oper->num_bytes;
1427                 qgroup->rfer_cmpr += sign * oper->num_bytes;
1428                 qgroup->excl += sign * oper->num_bytes;
1429                 if (sign < 0)
1430                         WARN_ON(qgroup->excl < oper->num_bytes);
1431                 qgroup->excl_cmpr += sign * oper->num_bytes;
1432                 qgroup_dirty(fs_info, qgroup);
1433
1434                 /* Add any parents of the parents */
1435                 list_for_each_entry(glist, &qgroup->groups, next_group) {
1436                         ret = ulist_add(tmp, glist->group->qgroupid,
1437                                         ptr_to_u64(glist->group), GFP_ATOMIC);
1438                         if (ret < 0)
1439                                 goto out;
1440                 }
1441         }
1442         ret = 0;
1443 out:
1444         spin_unlock(&fs_info->qgroup_lock);
1445         ulist_free(tmp);
1446         return ret;
1447 }
1448
1449 /*
1450  * Walk all of the roots that pointed to our bytenr and adjust their refcnts as
1451  * properly.
1452  */
1453 static int qgroup_calc_old_refcnt(struct btrfs_fs_info *fs_info,
1454                                   u64 root_to_skip, struct ulist *tmp,
1455                                   struct ulist *roots, struct ulist *qgroups,
1456                                   u64 seq, int *old_roots, int rescan)
1457 {
1458         struct ulist_node *unode;
1459         struct ulist_iterator uiter;
1460         struct ulist_node *tmp_unode;
1461         struct ulist_iterator tmp_uiter;
1462         struct btrfs_qgroup *qg;
1463         int ret;
1464
1465         ULIST_ITER_INIT(&uiter);
1466         while ((unode = ulist_next(roots, &uiter))) {
1467                 /* We don't count our current root here */
1468                 if (unode->val == root_to_skip)
1469                         continue;
1470                 qg = find_qgroup_rb(fs_info, unode->val);
1471                 if (!qg)
1472                         continue;
1473                 /*
1474                  * We could have a pending removal of this same ref so we may
1475                  * not have actually found our ref root when doing
1476                  * btrfs_find_all_roots, so we need to keep track of how many
1477                  * old roots we find in case we removed ours and added a
1478                  * different one at the same time.  I don't think this could
1479                  * happen in practice but that sort of thinking leads to pain
1480                  * and suffering and to the dark side.
1481                  */
1482                 (*old_roots)++;
1483
1484                 ulist_reinit(tmp);
1485                 ret = ulist_add(qgroups, qg->qgroupid, ptr_to_u64(qg),
1486                                 GFP_ATOMIC);
1487                 if (ret < 0)
1488                         return ret;
1489                 ret = ulist_add(tmp, qg->qgroupid, ptr_to_u64(qg), GFP_ATOMIC);
1490                 if (ret < 0)
1491                         return ret;
1492                 ULIST_ITER_INIT(&tmp_uiter);
1493                 while ((tmp_unode = ulist_next(tmp, &tmp_uiter))) {
1494                         struct btrfs_qgroup_list *glist;
1495
1496                         qg = u64_to_ptr(tmp_unode->aux);
1497                         /*
1498                          * We use this sequence number to keep from having to
1499                          * run the whole list and 0 out the refcnt every time.
1500                          * We basically use sequnce as the known 0 count and
1501                          * then add 1 everytime we see a qgroup.  This is how we
1502                          * get how many of the roots actually point up to the
1503                          * upper level qgroups in order to determine exclusive
1504                          * counts.
1505                          *
1506                          * For rescan we want to set old_refcnt to seq so our
1507                          * exclusive calculations end up correct.
1508                          */
1509                         if (rescan)
1510                                 qg->old_refcnt = seq;
1511                         else if (qg->old_refcnt < seq)
1512                                 qg->old_refcnt = seq + 1;
1513                         else
1514                                 qg->old_refcnt++;
1515
1516                         if (qg->new_refcnt < seq)
1517                                 qg->new_refcnt = seq + 1;
1518                         else
1519                                 qg->new_refcnt++;
1520                         list_for_each_entry(glist, &qg->groups, next_group) {
1521                                 ret = ulist_add(qgroups, glist->group->qgroupid,
1522                                                 ptr_to_u64(glist->group),
1523                                                 GFP_ATOMIC);
1524                                 if (ret < 0)
1525                                         return ret;
1526                                 ret = ulist_add(tmp, glist->group->qgroupid,
1527                                                 ptr_to_u64(glist->group),
1528                                                 GFP_ATOMIC);
1529                                 if (ret < 0)
1530                                         return ret;
1531                         }
1532                 }
1533         }
1534         return 0;
1535 }
1536
1537 /*
1538  * We need to walk forward in our operation tree and account for any roots that
1539  * were deleted after we made this operation.
1540  */
1541 static int qgroup_account_deleted_refs(struct btrfs_fs_info *fs_info,
1542                                        struct btrfs_qgroup_operation *oper,
1543                                        struct ulist *tmp,
1544                                        struct ulist *qgroups, u64 seq,
1545                                        int *old_roots)
1546 {
1547         struct ulist_node *unode;
1548         struct ulist_iterator uiter;
1549         struct btrfs_qgroup *qg;
1550         struct btrfs_qgroup_operation *tmp_oper;
1551         struct rb_node *n;
1552         int ret;
1553
1554         ulist_reinit(tmp);
1555
1556         /*
1557          * We only walk forward in the tree since we're only interested in
1558          * removals that happened _after_  our operation.
1559          */
1560         spin_lock(&fs_info->qgroup_op_lock);
1561         n = rb_next(&oper->n);
1562         spin_unlock(&fs_info->qgroup_op_lock);
1563         if (!n)
1564                 return 0;
1565         tmp_oper = rb_entry(n, struct btrfs_qgroup_operation, n);
1566         while (tmp_oper->bytenr == oper->bytenr) {
1567                 /*
1568                  * If it's not a removal we don't care, additions work out
1569                  * properly with our refcnt tracking.
1570                  */
1571                 if (tmp_oper->type != BTRFS_QGROUP_OPER_SUB_SHARED &&
1572                     tmp_oper->type != BTRFS_QGROUP_OPER_SUB_EXCL)
1573                         goto next;
1574                 qg = find_qgroup_rb(fs_info, tmp_oper->ref_root);
1575                 if (!qg)
1576                         goto next;
1577                 ret = ulist_add(qgroups, qg->qgroupid, ptr_to_u64(qg),
1578                                 GFP_ATOMIC);
1579                 if (ret) {
1580                         if (ret < 0)
1581                                 return ret;
1582                         /*
1583                          * We only want to increase old_roots if this qgroup is
1584                          * not already in the list of qgroups.  If it is already
1585                          * there then that means it must have been re-added or
1586                          * the delete will be discarded because we had an
1587                          * existing ref that we haven't looked up yet.  In this
1588                          * case we don't want to increase old_roots.  So if ret
1589                          * == 1 then we know that this is the first time we've
1590                          * seen this qgroup and we can bump the old_roots.
1591                          */
1592                         (*old_roots)++;
1593                         ret = ulist_add(tmp, qg->qgroupid, ptr_to_u64(qg),
1594                                         GFP_ATOMIC);
1595                         if (ret < 0)
1596                                 return ret;
1597                 }
1598 next:
1599                 spin_lock(&fs_info->qgroup_op_lock);
1600                 n = rb_next(&tmp_oper->n);
1601                 spin_unlock(&fs_info->qgroup_op_lock);
1602                 if (!n)
1603                         break;
1604                 tmp_oper = rb_entry(n, struct btrfs_qgroup_operation, n);
1605         }
1606
1607         /* Ok now process the qgroups we found */
1608         ULIST_ITER_INIT(&uiter);
1609         while ((unode = ulist_next(tmp, &uiter))) {
1610                 struct btrfs_qgroup_list *glist;
1611
1612                 qg = u64_to_ptr(unode->aux);
1613                 if (qg->old_refcnt < seq)
1614                         qg->old_refcnt = seq + 1;
1615                 else
1616                         qg->old_refcnt++;
1617                 if (qg->new_refcnt < seq)
1618                         qg->new_refcnt = seq + 1;
1619                 else
1620                         qg->new_refcnt++;
1621                 list_for_each_entry(glist, &qg->groups, next_group) {
1622                         ret = ulist_add(qgroups, glist->group->qgroupid,
1623                                         ptr_to_u64(glist->group), GFP_ATOMIC);
1624                         if (ret < 0)
1625                                 return ret;
1626                         ret = ulist_add(tmp, glist->group->qgroupid,
1627                                         ptr_to_u64(glist->group), GFP_ATOMIC);
1628                         if (ret < 0)
1629                                 return ret;
1630                 }
1631         }
1632         return 0;
1633 }
1634
1635 /* Add refcnt for the newly added reference. */
1636 static int qgroup_calc_new_refcnt(struct btrfs_fs_info *fs_info,
1637                                   struct btrfs_qgroup_operation *oper,
1638                                   struct btrfs_qgroup *qgroup,
1639                                   struct ulist *tmp, struct ulist *qgroups,
1640                                   u64 seq)
1641 {
1642         struct ulist_node *unode;
1643         struct ulist_iterator uiter;
1644         struct btrfs_qgroup *qg;
1645         int ret;
1646
1647         ulist_reinit(tmp);
1648         ret = ulist_add(qgroups, qgroup->qgroupid, ptr_to_u64(qgroup),
1649                         GFP_ATOMIC);
1650         if (ret < 0)
1651                 return ret;
1652         ret = ulist_add(tmp, qgroup->qgroupid, ptr_to_u64(qgroup),
1653                         GFP_ATOMIC);
1654         if (ret < 0)
1655                 return ret;
1656         ULIST_ITER_INIT(&uiter);
1657         while ((unode = ulist_next(tmp, &uiter))) {
1658                 struct btrfs_qgroup_list *glist;
1659
1660                 qg = u64_to_ptr(unode->aux);
1661                 if (oper->type == BTRFS_QGROUP_OPER_ADD_SHARED) {
1662                         if (qg->new_refcnt < seq)
1663                                 qg->new_refcnt = seq + 1;
1664                         else
1665                                 qg->new_refcnt++;
1666                 } else {
1667                         if (qg->old_refcnt < seq)
1668                                 qg->old_refcnt = seq + 1;
1669                         else
1670                                 qg->old_refcnt++;
1671                 }
1672                 list_for_each_entry(glist, &qg->groups, next_group) {
1673                         ret = ulist_add(tmp, glist->group->qgroupid,
1674                                         ptr_to_u64(glist->group), GFP_ATOMIC);
1675                         if (ret < 0)
1676                                 return ret;
1677                         ret = ulist_add(qgroups, glist->group->qgroupid,
1678                                         ptr_to_u64(glist->group), GFP_ATOMIC);
1679                         if (ret < 0)
1680                                 return ret;
1681                 }
1682         }
1683         return 0;
1684 }
1685
1686 /*
1687  * This adjusts the counters for all referenced qgroups if need be.
1688  */
1689 static int qgroup_adjust_counters(struct btrfs_fs_info *fs_info,
1690                                   u64 root_to_skip, u64 num_bytes,
1691                                   struct ulist *qgroups, u64 seq,
1692                                   int old_roots, int new_roots, int rescan)
1693 {
1694         struct ulist_node *unode;
1695         struct ulist_iterator uiter;
1696         struct btrfs_qgroup *qg;
1697         u64 cur_new_count, cur_old_count;
1698
1699         ULIST_ITER_INIT(&uiter);
1700         while ((unode = ulist_next(qgroups, &uiter))) {
1701                 bool dirty = false;
1702
1703                 qg = u64_to_ptr(unode->aux);
1704                 /*
1705                  * Wasn't referenced before but is now, add to the reference
1706                  * counters.
1707                  */
1708                 if (qg->old_refcnt <= seq && qg->new_refcnt > seq) {
1709                         qg->rfer += num_bytes;
1710                         qg->rfer_cmpr += num_bytes;
1711                         dirty = true;
1712                 }
1713
1714                 /*
1715                  * Was referenced before but isn't now, subtract from the
1716                  * reference counters.
1717                  */
1718                 if (qg->old_refcnt > seq && qg->new_refcnt <= seq) {
1719                         qg->rfer -= num_bytes;
1720                         qg->rfer_cmpr -= num_bytes;
1721                         dirty = true;
1722                 }
1723
1724                 if (qg->old_refcnt < seq)
1725                         cur_old_count = 0;
1726                 else
1727                         cur_old_count = qg->old_refcnt - seq;
1728                 if (qg->new_refcnt < seq)
1729                         cur_new_count = 0;
1730                 else
1731                         cur_new_count = qg->new_refcnt - seq;
1732
1733                 /*
1734                  * If our refcount was the same as the roots previously but our
1735                  * new count isn't the same as the number of roots now then we
1736                  * went from having a exclusive reference on this range to not.
1737                  */
1738                 if (old_roots && cur_old_count == old_roots &&
1739                     (cur_new_count != new_roots || new_roots == 0)) {
1740                         WARN_ON(cur_new_count != new_roots && new_roots == 0);
1741                         qg->excl -= num_bytes;
1742                         qg->excl_cmpr -= num_bytes;
1743                         dirty = true;
1744                 }
1745
1746                 /*
1747                  * If we didn't reference all the roots before but now we do we
1748                  * have an exclusive reference to this range.
1749                  */
1750                 if ((!old_roots || (old_roots && cur_old_count != old_roots))
1751                     && cur_new_count == new_roots) {
1752                         qg->excl += num_bytes;
1753                         qg->excl_cmpr += num_bytes;
1754                         dirty = true;
1755                 }
1756
1757                 if (dirty)
1758                         qgroup_dirty(fs_info, qg);
1759         }
1760         return 0;
1761 }
1762
1763 /*
1764  * If we removed a data extent and there were other references for that bytenr
1765  * then we need to lookup all referenced roots to make sure we still don't
1766  * reference this bytenr.  If we do then we can just discard this operation.
1767  */
1768 static int check_existing_refs(struct btrfs_trans_handle *trans,
1769                                struct btrfs_fs_info *fs_info,
1770                                struct btrfs_qgroup_operation *oper)
1771 {
1772         struct ulist *roots = NULL;
1773         struct ulist_node *unode;
1774         struct ulist_iterator uiter;
1775         int ret = 0;
1776
1777         ret = btrfs_find_all_roots(trans, fs_info, oper->bytenr,
1778                                    oper->elem.seq, &roots);
1779         if (ret < 0)
1780                 return ret;
1781         ret = 0;
1782
1783         ULIST_ITER_INIT(&uiter);
1784         while ((unode = ulist_next(roots, &uiter))) {
1785                 if (unode->val == oper->ref_root) {
1786                         ret = 1;
1787                         break;
1788                 }
1789         }
1790         ulist_free(roots);
1791         btrfs_put_tree_mod_seq(fs_info, &oper->elem);
1792
1793         return ret;
1794 }
1795
1796 /*
1797  * If we share a reference across multiple roots then we may need to adjust
1798  * various qgroups referenced and exclusive counters.  The basic premise is this
1799  *
1800  * 1) We have seq to represent a 0 count.  Instead of looping through all of the
1801  * qgroups and resetting their refcount to 0 we just constantly bump this
1802  * sequence number to act as the base reference count.  This means that if
1803  * anybody is equal to or below this sequence they were never referenced.  We
1804  * jack this sequence up by the number of roots we found each time in order to
1805  * make sure we don't have any overlap.
1806  *
1807  * 2) We first search all the roots that reference the area _except_ the root
1808  * we're acting on currently.  This makes up the old_refcnt of all the qgroups
1809  * before.
1810  *
1811  * 3) We walk all of the qgroups referenced by the root we are currently acting
1812  * on, and will either adjust old_refcnt in the case of a removal or the
1813  * new_refcnt in the case of an addition.
1814  *
1815  * 4) Finally we walk all the qgroups that are referenced by this range
1816  * including the root we are acting on currently.  We will adjust the counters
1817  * based on the number of roots we had and will have after this operation.
1818  *
1819  * Take this example as an illustration
1820  *
1821  *                      [qgroup 1/0]
1822  *                   /         |          \
1823  *              [qg 0/0]   [qg 0/1]     [qg 0/2]
1824  *                 \          |            /
1825  *                [        extent           ]
1826  *
1827  * Say we are adding a reference that is covered by qg 0/0.  The first step
1828  * would give a refcnt of 1 to qg 0/1 and 0/2 and a refcnt of 2 to qg 1/0 with
1829  * old_roots being 2.  Because it is adding new_roots will be 1.  We then go
1830  * through qg 0/0 which will get the new_refcnt set to 1 and add 1 to qg 1/0's
1831  * new_refcnt, bringing it to 3.  We then walk through all of the qgroups, we
1832  * notice that the old refcnt for qg 0/0 < the new refcnt, so we added a
1833  * reference and thus must add the size to the referenced bytes.  Everything
1834  * else is the same so nothing else changes.
1835  */
1836 static int qgroup_shared_accounting(struct btrfs_trans_handle *trans,
1837                                     struct btrfs_fs_info *fs_info,
1838                                     struct btrfs_qgroup_operation *oper)
1839 {
1840         struct ulist *roots = NULL;
1841         struct ulist *qgroups, *tmp;
1842         struct btrfs_qgroup *qgroup;
1843         struct seq_list elem = {};
1844         u64 seq;
1845         int old_roots = 0;
1846         int new_roots = 0;
1847         int ret = 0;
1848
1849         if (oper->elem.seq) {
1850                 ret = check_existing_refs(trans, fs_info, oper);
1851                 if (ret < 0)
1852                         return ret;
1853                 if (ret)
1854                         return 0;
1855         }
1856
1857         qgroups = ulist_alloc(GFP_NOFS);
1858         if (!qgroups)
1859                 return -ENOMEM;
1860
1861         tmp = ulist_alloc(GFP_NOFS);
1862         if (!tmp) {
1863                 ulist_free(qgroups);
1864                 return -ENOMEM;
1865         }
1866
1867         btrfs_get_tree_mod_seq(fs_info, &elem);
1868         ret = btrfs_find_all_roots(trans, fs_info, oper->bytenr, elem.seq,
1869                                    &roots);
1870         btrfs_put_tree_mod_seq(fs_info, &elem);
1871         if (ret < 0) {
1872                 ulist_free(qgroups);
1873                 ulist_free(tmp);
1874                 return ret;
1875         }
1876         spin_lock(&fs_info->qgroup_lock);
1877         qgroup = find_qgroup_rb(fs_info, oper->ref_root);
1878         if (!qgroup)
1879                 goto out;
1880         seq = fs_info->qgroup_seq;
1881
1882         /*
1883          * So roots is the list of all the roots currently pointing at the
1884          * bytenr, including the ref we are adding if we are adding, or not if
1885          * we are removing a ref.  So we pass in the ref_root to skip that root
1886          * in our calculations.  We set old_refnct and new_refcnt cause who the
1887          * hell knows what everything looked like before, and it doesn't matter
1888          * except...
1889          */
1890         ret = qgroup_calc_old_refcnt(fs_info, oper->ref_root, tmp, roots, qgroups,
1891                                      seq, &old_roots, 0);
1892         if (ret < 0)
1893                 goto out;
1894
1895         /*
1896          * Now adjust the refcounts of the qgroups that care about this
1897          * reference, either the old_count in the case of removal or new_count
1898          * in the case of an addition.
1899          */
1900         ret = qgroup_calc_new_refcnt(fs_info, oper, qgroup, tmp, qgroups,
1901                                      seq);
1902         if (ret < 0)
1903                 goto out;
1904
1905         /*
1906          * ...in the case of removals.  If we had a removal before we got around
1907          * to processing this operation then we need to find that guy and count
1908          * his references as if they really existed so we don't end up screwing
1909          * up the exclusive counts.  Then whenever we go to process the delete
1910          * everything will be grand and we can account for whatever exclusive
1911          * changes need to be made there.  We also have to pass in old_roots so
1912          * we have an accurate count of the roots as it pertains to this
1913          * operations view of the world.
1914          */
1915         ret = qgroup_account_deleted_refs(fs_info, oper, tmp, qgroups, seq,
1916                                           &old_roots);
1917         if (ret < 0)
1918                 goto out;
1919
1920         /*
1921          * We are adding our root, need to adjust up the number of roots,
1922          * otherwise old_roots is the number of roots we want.
1923          */
1924         if (oper->type == BTRFS_QGROUP_OPER_ADD_SHARED) {
1925                 new_roots = old_roots + 1;
1926         } else {
1927                 new_roots = old_roots;
1928                 old_roots++;
1929         }
1930         fs_info->qgroup_seq += old_roots + 1;
1931
1932
1933         /*
1934          * And now the magic happens, bless Arne for having a pretty elegant
1935          * solution for this.
1936          */
1937         qgroup_adjust_counters(fs_info, oper->ref_root, oper->num_bytes,
1938                                qgroups, seq, old_roots, new_roots, 0);
1939 out:
1940         spin_unlock(&fs_info->qgroup_lock);
1941         ulist_free(qgroups);
1942         ulist_free(roots);
1943         ulist_free(tmp);
1944         return ret;
1945 }
1946
1947 /*
1948  * Process a reference to a shared subtree. This type of operation is
1949  * queued during snapshot removal when we encounter extents which are
1950  * shared between more than one root.
1951  */
1952 static int qgroup_subtree_accounting(struct btrfs_trans_handle *trans,
1953                                      struct btrfs_fs_info *fs_info,
1954                                      struct btrfs_qgroup_operation *oper)
1955 {
1956         struct ulist *roots = NULL;
1957         struct ulist_node *unode;
1958         struct ulist_iterator uiter;
1959         struct btrfs_qgroup_list *glist;
1960         struct ulist *parents;
1961         int ret = 0;
1962         int err;
1963         struct btrfs_qgroup *qg;
1964         u64 root_obj = 0;
1965         struct seq_list elem = {};
1966
1967         parents = ulist_alloc(GFP_NOFS);
1968         if (!parents)
1969                 return -ENOMEM;
1970
1971         btrfs_get_tree_mod_seq(fs_info, &elem);
1972         ret = btrfs_find_all_roots(trans, fs_info, oper->bytenr,
1973                                    elem.seq, &roots);
1974         btrfs_put_tree_mod_seq(fs_info, &elem);
1975         if (ret < 0)
1976                 return ret;
1977
1978         if (roots->nnodes != 1)
1979                 goto out;
1980
1981         ULIST_ITER_INIT(&uiter);
1982         unode = ulist_next(roots, &uiter); /* Only want 1 so no need to loop */
1983         /*
1984          * If we find our ref root then that means all refs
1985          * this extent has to the root have not yet been
1986          * deleted. In that case, we do nothing and let the
1987          * last ref for this bytenr drive our update.
1988          *
1989          * This can happen for example if an extent is
1990          * referenced multiple times in a snapshot (clone,
1991          * etc). If we are in the middle of snapshot removal,
1992          * queued updates for such an extent will find the
1993          * root if we have not yet finished removing the
1994          * snapshot.
1995          */
1996         if (unode->val == oper->ref_root)
1997                 goto out;
1998
1999         root_obj = unode->val;
2000         BUG_ON(!root_obj);
2001
2002         spin_lock(&fs_info->qgroup_lock);
2003         qg = find_qgroup_rb(fs_info, root_obj);
2004         if (!qg)
2005                 goto out_unlock;
2006
2007         qg->excl += oper->num_bytes;
2008         qg->excl_cmpr += oper->num_bytes;
2009         qgroup_dirty(fs_info, qg);
2010
2011         /*
2012          * Adjust counts for parent groups. First we find all
2013          * parents, then in the 2nd loop we do the adjustment
2014          * while adding parents of the parents to our ulist.
2015          */
2016         list_for_each_entry(glist, &qg->groups, next_group) {
2017                 err = ulist_add(parents, glist->group->qgroupid,
2018                                 ptr_to_u64(glist->group), GFP_ATOMIC);
2019                 if (err < 0) {
2020                         ret = err;
2021                         goto out_unlock;
2022                 }
2023         }
2024
2025         ULIST_ITER_INIT(&uiter);
2026         while ((unode = ulist_next(parents, &uiter))) {
2027                 qg = u64_to_ptr(unode->aux);
2028                 qg->excl += oper->num_bytes;
2029                 qg->excl_cmpr += oper->num_bytes;
2030                 qgroup_dirty(fs_info, qg);
2031
2032                 /* Add any parents of the parents */
2033                 list_for_each_entry(glist, &qg->groups, next_group) {
2034                         err = ulist_add(parents, glist->group->qgroupid,
2035                                         ptr_to_u64(glist->group), GFP_ATOMIC);
2036                         if (err < 0) {
2037                                 ret = err;
2038                                 goto out_unlock;
2039                         }
2040                 }
2041         }
2042
2043 out_unlock:
2044         spin_unlock(&fs_info->qgroup_lock);
2045
2046 out:
2047         ulist_free(roots);
2048         ulist_free(parents);
2049         return ret;
2050 }
2051
2052 /*
2053  * btrfs_qgroup_account_ref is called for every ref that is added to or deleted
2054  * from the fs. First, all roots referencing the extent are searched, and
2055  * then the space is accounted accordingly to the different roots. The
2056  * accounting algorithm works in 3 steps documented inline.
2057  */
2058 static int btrfs_qgroup_account(struct btrfs_trans_handle *trans,
2059                                 struct btrfs_fs_info *fs_info,
2060                                 struct btrfs_qgroup_operation *oper)
2061 {
2062         int ret = 0;
2063
2064         if (!fs_info->quota_enabled)
2065                 return 0;
2066
2067         BUG_ON(!fs_info->quota_root);
2068
2069         mutex_lock(&fs_info->qgroup_rescan_lock);
2070         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
2071                 if (fs_info->qgroup_rescan_progress.objectid <= oper->bytenr) {
2072                         mutex_unlock(&fs_info->qgroup_rescan_lock);
2073                         return 0;
2074                 }
2075         }
2076         mutex_unlock(&fs_info->qgroup_rescan_lock);
2077
2078         ASSERT(is_fstree(oper->ref_root));
2079
2080         switch (oper->type) {
2081         case BTRFS_QGROUP_OPER_ADD_EXCL:
2082         case BTRFS_QGROUP_OPER_SUB_EXCL:
2083                 ret = qgroup_excl_accounting(fs_info, oper);
2084                 break;
2085         case BTRFS_QGROUP_OPER_ADD_SHARED:
2086         case BTRFS_QGROUP_OPER_SUB_SHARED:
2087                 ret = qgroup_shared_accounting(trans, fs_info, oper);
2088                 break;
2089         case BTRFS_QGROUP_OPER_SUB_SUBTREE:
2090                 ret = qgroup_subtree_accounting(trans, fs_info, oper);
2091                 break;
2092         default:
2093                 ASSERT(0);
2094         }
2095         return ret;
2096 }
2097
2098 /*
2099  * Needs to be called everytime we run delayed refs, even if there is an error
2100  * in order to cleanup outstanding operations.
2101  */
2102 int btrfs_delayed_qgroup_accounting(struct btrfs_trans_handle *trans,
2103                                     struct btrfs_fs_info *fs_info)
2104 {
2105         struct btrfs_qgroup_operation *oper;
2106         int ret = 0;
2107
2108         while (!list_empty(&trans->qgroup_ref_list)) {
2109                 oper = list_first_entry(&trans->qgroup_ref_list,
2110                                         struct btrfs_qgroup_operation, list);
2111                 list_del_init(&oper->list);
2112                 if (!ret || !trans->aborted)
2113                         ret = btrfs_qgroup_account(trans, fs_info, oper);
2114                 spin_lock(&fs_info->qgroup_op_lock);
2115                 rb_erase(&oper->n, &fs_info->qgroup_op_tree);
2116                 spin_unlock(&fs_info->qgroup_op_lock);
2117                 btrfs_put_tree_mod_seq(fs_info, &oper->elem);
2118                 kfree(oper);
2119         }
2120         return ret;
2121 }
2122
2123 /*
2124  * called from commit_transaction. Writes all changed qgroups to disk.
2125  */
2126 int btrfs_run_qgroups(struct btrfs_trans_handle *trans,
2127                       struct btrfs_fs_info *fs_info)
2128 {
2129         struct btrfs_root *quota_root = fs_info->quota_root;
2130         int ret = 0;
2131         int start_rescan_worker = 0;
2132
2133         if (!quota_root)
2134                 goto out;
2135
2136         if (!fs_info->quota_enabled && fs_info->pending_quota_state)
2137                 start_rescan_worker = 1;
2138
2139         fs_info->quota_enabled = fs_info->pending_quota_state;
2140
2141         spin_lock(&fs_info->qgroup_lock);
2142         while (!list_empty(&fs_info->dirty_qgroups)) {
2143                 struct btrfs_qgroup *qgroup;
2144                 qgroup = list_first_entry(&fs_info->dirty_qgroups,
2145                                           struct btrfs_qgroup, dirty);
2146                 list_del_init(&qgroup->dirty);
2147                 spin_unlock(&fs_info->qgroup_lock);
2148                 ret = update_qgroup_info_item(trans, quota_root, qgroup);
2149                 if (ret)
2150                         fs_info->qgroup_flags |=
2151                                         BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2152                 spin_lock(&fs_info->qgroup_lock);
2153         }
2154         if (fs_info->quota_enabled)
2155                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_ON;
2156         else
2157                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_ON;
2158         spin_unlock(&fs_info->qgroup_lock);
2159
2160         ret = update_qgroup_status_item(trans, fs_info, quota_root);
2161         if (ret)
2162                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2163
2164         if (!ret && start_rescan_worker) {
2165                 ret = qgroup_rescan_init(fs_info, 0, 1);
2166                 if (!ret) {
2167                         qgroup_rescan_zero_tracking(fs_info);
2168                         btrfs_queue_work(fs_info->qgroup_rescan_workers,
2169                                          &fs_info->qgroup_rescan_work);
2170                 }
2171                 ret = 0;
2172         }
2173
2174 out:
2175
2176         return ret;
2177 }
2178
2179 /*
2180  * copy the acounting information between qgroups. This is necessary when a
2181  * snapshot or a subvolume is created
2182  */
2183 int btrfs_qgroup_inherit(struct btrfs_trans_handle *trans,
2184                          struct btrfs_fs_info *fs_info, u64 srcid, u64 objectid,
2185                          struct btrfs_qgroup_inherit *inherit)
2186 {
2187         int ret = 0;
2188         int i;
2189         u64 *i_qgroups;
2190         struct btrfs_root *quota_root = fs_info->quota_root;
2191         struct btrfs_qgroup *srcgroup;
2192         struct btrfs_qgroup *dstgroup;
2193         u32 level_size = 0;
2194         u64 nums;
2195
2196         mutex_lock(&fs_info->qgroup_ioctl_lock);
2197         if (!fs_info->quota_enabled)
2198                 goto out;
2199
2200         if (!quota_root) {
2201                 ret = -EINVAL;
2202                 goto out;
2203         }
2204
2205         if (inherit) {
2206                 i_qgroups = (u64 *)(inherit + 1);
2207                 nums = inherit->num_qgroups + 2 * inherit->num_ref_copies +
2208                        2 * inherit->num_excl_copies;
2209                 for (i = 0; i < nums; ++i) {
2210                         srcgroup = find_qgroup_rb(fs_info, *i_qgroups);
2211                         if (!srcgroup) {
2212                                 ret = -EINVAL;
2213                                 goto out;
2214                         }
2215                         ++i_qgroups;
2216                 }
2217         }
2218
2219         /*
2220          * create a tracking group for the subvol itself
2221          */
2222         ret = add_qgroup_item(trans, quota_root, objectid);
2223         if (ret)
2224                 goto out;
2225
2226         if (inherit && inherit->flags & BTRFS_QGROUP_INHERIT_SET_LIMITS) {
2227                 ret = update_qgroup_limit_item(trans, quota_root, objectid,
2228                                                inherit->lim.flags,
2229                                                inherit->lim.max_rfer,
2230                                                inherit->lim.max_excl,
2231                                                inherit->lim.rsv_rfer,
2232                                                inherit->lim.rsv_excl);
2233                 if (ret)
2234                         goto out;
2235         }
2236
2237         if (srcid) {
2238                 struct btrfs_root *srcroot;
2239                 struct btrfs_key srckey;
2240                 int srcroot_level;
2241
2242                 srckey.objectid = srcid;
2243                 srckey.type = BTRFS_ROOT_ITEM_KEY;
2244                 srckey.offset = (u64)-1;
2245                 srcroot = btrfs_read_fs_root_no_name(fs_info, &srckey);
2246                 if (IS_ERR(srcroot)) {
2247                         ret = PTR_ERR(srcroot);
2248                         goto out;
2249                 }
2250
2251                 rcu_read_lock();
2252                 srcroot_level = btrfs_header_level(srcroot->node);
2253                 level_size = btrfs_level_size(srcroot, srcroot_level);
2254                 rcu_read_unlock();
2255         }
2256
2257         /*
2258          * add qgroup to all inherited groups
2259          */
2260         if (inherit) {
2261                 i_qgroups = (u64 *)(inherit + 1);
2262                 for (i = 0; i < inherit->num_qgroups; ++i) {
2263                         ret = add_qgroup_relation_item(trans, quota_root,
2264                                                        objectid, *i_qgroups);
2265                         if (ret)
2266                                 goto out;
2267                         ret = add_qgroup_relation_item(trans, quota_root,
2268                                                        *i_qgroups, objectid);
2269                         if (ret)
2270                                 goto out;
2271                         ++i_qgroups;
2272                 }
2273         }
2274
2275
2276         spin_lock(&fs_info->qgroup_lock);
2277
2278         dstgroup = add_qgroup_rb(fs_info, objectid);
2279         if (IS_ERR(dstgroup)) {
2280                 ret = PTR_ERR(dstgroup);
2281                 goto unlock;
2282         }
2283
2284         if (srcid) {
2285                 srcgroup = find_qgroup_rb(fs_info, srcid);
2286                 if (!srcgroup)
2287                         goto unlock;
2288
2289                 /*
2290                  * We call inherit after we clone the root in order to make sure
2291                  * our counts don't go crazy, so at this point the only
2292                  * difference between the two roots should be the root node.
2293                  */
2294                 dstgroup->rfer = srcgroup->rfer;
2295                 dstgroup->rfer_cmpr = srcgroup->rfer_cmpr;
2296                 dstgroup->excl = level_size;
2297                 dstgroup->excl_cmpr = level_size;
2298                 srcgroup->excl = level_size;
2299                 srcgroup->excl_cmpr = level_size;
2300                 qgroup_dirty(fs_info, dstgroup);
2301                 qgroup_dirty(fs_info, srcgroup);
2302         }
2303
2304         if (!inherit)
2305                 goto unlock;
2306
2307         i_qgroups = (u64 *)(inherit + 1);
2308         for (i = 0; i < inherit->num_qgroups; ++i) {
2309                 ret = add_relation_rb(quota_root->fs_info, objectid,
2310                                       *i_qgroups);
2311                 if (ret)
2312                         goto unlock;
2313                 ++i_qgroups;
2314         }
2315
2316         for (i = 0; i <  inherit->num_ref_copies; ++i) {
2317                 struct btrfs_qgroup *src;
2318                 struct btrfs_qgroup *dst;
2319
2320                 src = find_qgroup_rb(fs_info, i_qgroups[0]);
2321                 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
2322
2323                 if (!src || !dst) {
2324                         ret = -EINVAL;
2325                         goto unlock;
2326                 }
2327
2328                 dst->rfer = src->rfer - level_size;
2329                 dst->rfer_cmpr = src->rfer_cmpr - level_size;
2330                 i_qgroups += 2;
2331         }
2332         for (i = 0; i <  inherit->num_excl_copies; ++i) {
2333                 struct btrfs_qgroup *src;
2334                 struct btrfs_qgroup *dst;
2335
2336                 src = find_qgroup_rb(fs_info, i_qgroups[0]);
2337                 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
2338
2339                 if (!src || !dst) {
2340                         ret = -EINVAL;
2341                         goto unlock;
2342                 }
2343
2344                 dst->excl = src->excl + level_size;
2345                 dst->excl_cmpr = src->excl_cmpr + level_size;
2346                 i_qgroups += 2;
2347         }
2348
2349 unlock:
2350         spin_unlock(&fs_info->qgroup_lock);
2351 out:
2352         mutex_unlock(&fs_info->qgroup_ioctl_lock);
2353         return ret;
2354 }
2355
2356 /*
2357  * reserve some space for a qgroup and all its parents. The reservation takes
2358  * place with start_transaction or dealloc_reserve, similar to ENOSPC
2359  * accounting. If not enough space is available, EDQUOT is returned.
2360  * We assume that the requested space is new for all qgroups.
2361  */
2362 int btrfs_qgroup_reserve(struct btrfs_root *root, u64 num_bytes)
2363 {
2364         struct btrfs_root *quota_root;
2365         struct btrfs_qgroup *qgroup;
2366         struct btrfs_fs_info *fs_info = root->fs_info;
2367         u64 ref_root = root->root_key.objectid;
2368         int ret = 0;
2369         struct ulist_node *unode;
2370         struct ulist_iterator uiter;
2371
2372         if (!is_fstree(ref_root))
2373                 return 0;
2374
2375         if (num_bytes == 0)
2376                 return 0;
2377
2378         spin_lock(&fs_info->qgroup_lock);
2379         quota_root = fs_info->quota_root;
2380         if (!quota_root)
2381                 goto out;
2382
2383         qgroup = find_qgroup_rb(fs_info, ref_root);
2384         if (!qgroup)
2385                 goto out;
2386
2387         /*
2388          * in a first step, we check all affected qgroups if any limits would
2389          * be exceeded
2390          */
2391         ulist_reinit(fs_info->qgroup_ulist);
2392         ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
2393                         (uintptr_t)qgroup, GFP_ATOMIC);
2394         if (ret < 0)
2395                 goto out;
2396         ULIST_ITER_INIT(&uiter);
2397         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
2398                 struct btrfs_qgroup *qg;
2399                 struct btrfs_qgroup_list *glist;
2400
2401                 qg = u64_to_ptr(unode->aux);
2402
2403                 if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_RFER) &&
2404                     qg->reserved + (s64)qg->rfer + num_bytes >
2405                     qg->max_rfer) {
2406                         ret = -EDQUOT;
2407                         goto out;
2408                 }
2409
2410                 if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) &&
2411                     qg->reserved + (s64)qg->excl + num_bytes >
2412                     qg->max_excl) {
2413                         ret = -EDQUOT;
2414                         goto out;
2415                 }
2416
2417                 list_for_each_entry(glist, &qg->groups, next_group) {
2418                         ret = ulist_add(fs_info->qgroup_ulist,
2419                                         glist->group->qgroupid,
2420                                         (uintptr_t)glist->group, GFP_ATOMIC);
2421                         if (ret < 0)
2422                                 goto out;
2423                 }
2424         }
2425         ret = 0;
2426         /*
2427          * no limits exceeded, now record the reservation into all qgroups
2428          */
2429         ULIST_ITER_INIT(&uiter);
2430         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
2431                 struct btrfs_qgroup *qg;
2432
2433                 qg = u64_to_ptr(unode->aux);
2434
2435                 qg->reserved += num_bytes;
2436         }
2437
2438 out:
2439         spin_unlock(&fs_info->qgroup_lock);
2440         return ret;
2441 }
2442
2443 void btrfs_qgroup_free(struct btrfs_root *root, u64 num_bytes)
2444 {
2445         struct btrfs_root *quota_root;
2446         struct btrfs_qgroup *qgroup;
2447         struct btrfs_fs_info *fs_info = root->fs_info;
2448         struct ulist_node *unode;
2449         struct ulist_iterator uiter;
2450         u64 ref_root = root->root_key.objectid;
2451         int ret = 0;
2452
2453         if (!is_fstree(ref_root))
2454                 return;
2455
2456         if (num_bytes == 0)
2457                 return;
2458
2459         spin_lock(&fs_info->qgroup_lock);
2460
2461         quota_root = fs_info->quota_root;
2462         if (!quota_root)
2463                 goto out;
2464
2465         qgroup = find_qgroup_rb(fs_info, ref_root);
2466         if (!qgroup)
2467                 goto out;
2468
2469         ulist_reinit(fs_info->qgroup_ulist);
2470         ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
2471                         (uintptr_t)qgroup, GFP_ATOMIC);
2472         if (ret < 0)
2473                 goto out;
2474         ULIST_ITER_INIT(&uiter);
2475         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
2476                 struct btrfs_qgroup *qg;
2477                 struct btrfs_qgroup_list *glist;
2478
2479                 qg = u64_to_ptr(unode->aux);
2480
2481                 qg->reserved -= num_bytes;
2482
2483                 list_for_each_entry(glist, &qg->groups, next_group) {
2484                         ret = ulist_add(fs_info->qgroup_ulist,
2485                                         glist->group->qgroupid,
2486                                         (uintptr_t)glist->group, GFP_ATOMIC);
2487                         if (ret < 0)
2488                                 goto out;
2489                 }
2490         }
2491
2492 out:
2493         spin_unlock(&fs_info->qgroup_lock);
2494 }
2495
2496 void assert_qgroups_uptodate(struct btrfs_trans_handle *trans)
2497 {
2498         if (list_empty(&trans->qgroup_ref_list) && !trans->delayed_ref_elem.seq)
2499                 return;
2500         btrfs_err(trans->root->fs_info,
2501                 "qgroups not uptodate in trans handle %p:  list is%s empty, "
2502                 "seq is %#x.%x",
2503                 trans, list_empty(&trans->qgroup_ref_list) ? "" : " not",
2504                 (u32)(trans->delayed_ref_elem.seq >> 32),
2505                 (u32)trans->delayed_ref_elem.seq);
2506         BUG();
2507 }
2508
2509 /*
2510  * returns < 0 on error, 0 when more leafs are to be scanned.
2511  * returns 1 when done, 2 when done and FLAG_INCONSISTENT was cleared.
2512  */
2513 static int
2514 qgroup_rescan_leaf(struct btrfs_fs_info *fs_info, struct btrfs_path *path,
2515                    struct btrfs_trans_handle *trans, struct ulist *qgroups,
2516                    struct ulist *tmp, struct extent_buffer *scratch_leaf)
2517 {
2518         struct btrfs_key found;
2519         struct ulist *roots = NULL;
2520         struct seq_list tree_mod_seq_elem = {};
2521         u64 num_bytes;
2522         u64 seq;
2523         int new_roots;
2524         int slot;
2525         int ret;
2526
2527         path->leave_spinning = 1;
2528         mutex_lock(&fs_info->qgroup_rescan_lock);
2529         ret = btrfs_search_slot_for_read(fs_info->extent_root,
2530                                          &fs_info->qgroup_rescan_progress,
2531                                          path, 1, 0);
2532
2533         pr_debug("current progress key (%llu %u %llu), search_slot ret %d\n",
2534                  fs_info->qgroup_rescan_progress.objectid,
2535                  fs_info->qgroup_rescan_progress.type,
2536                  fs_info->qgroup_rescan_progress.offset, ret);
2537
2538         if (ret) {
2539                 /*
2540                  * The rescan is about to end, we will not be scanning any
2541                  * further blocks. We cannot unset the RESCAN flag here, because
2542                  * we want to commit the transaction if everything went well.
2543                  * To make the live accounting work in this phase, we set our
2544                  * scan progress pointer such that every real extent objectid
2545                  * will be smaller.
2546                  */
2547                 fs_info->qgroup_rescan_progress.objectid = (u64)-1;
2548                 btrfs_release_path(path);
2549                 mutex_unlock(&fs_info->qgroup_rescan_lock);
2550                 return ret;
2551         }
2552
2553         btrfs_item_key_to_cpu(path->nodes[0], &found,
2554                               btrfs_header_nritems(path->nodes[0]) - 1);
2555         fs_info->qgroup_rescan_progress.objectid = found.objectid + 1;
2556
2557         btrfs_get_tree_mod_seq(fs_info, &tree_mod_seq_elem);
2558         memcpy(scratch_leaf, path->nodes[0], sizeof(*scratch_leaf));
2559         slot = path->slots[0];
2560         btrfs_release_path(path);
2561         mutex_unlock(&fs_info->qgroup_rescan_lock);
2562
2563         for (; slot < btrfs_header_nritems(scratch_leaf); ++slot) {
2564                 btrfs_item_key_to_cpu(scratch_leaf, &found, slot);
2565                 if (found.type != BTRFS_EXTENT_ITEM_KEY &&
2566                     found.type != BTRFS_METADATA_ITEM_KEY)
2567                         continue;
2568                 if (found.type == BTRFS_METADATA_ITEM_KEY)
2569                         num_bytes = fs_info->extent_root->leafsize;
2570                 else
2571                         num_bytes = found.offset;
2572
2573                 ulist_reinit(qgroups);
2574                 ret = btrfs_find_all_roots(NULL, fs_info, found.objectid, 0,
2575                                            &roots);
2576                 if (ret < 0)
2577                         goto out;
2578                 spin_lock(&fs_info->qgroup_lock);
2579                 seq = fs_info->qgroup_seq;
2580                 fs_info->qgroup_seq += roots->nnodes + 1; /* max refcnt */
2581
2582                 new_roots = 0;
2583                 ret = qgroup_calc_old_refcnt(fs_info, 0, tmp, roots, qgroups,
2584                                              seq, &new_roots, 1);
2585                 if (ret < 0) {
2586                         spin_unlock(&fs_info->qgroup_lock);
2587                         ulist_free(roots);
2588                         goto out;
2589                 }
2590
2591                 ret = qgroup_adjust_counters(fs_info, 0, num_bytes, qgroups,
2592                                              seq, 0, new_roots, 1);
2593                 if (ret < 0) {
2594                         spin_unlock(&fs_info->qgroup_lock);
2595                         ulist_free(roots);
2596                         goto out;
2597                 }
2598                 spin_unlock(&fs_info->qgroup_lock);
2599                 ulist_free(roots);
2600         }
2601 out:
2602         btrfs_put_tree_mod_seq(fs_info, &tree_mod_seq_elem);
2603
2604         return ret;
2605 }
2606
2607 static void btrfs_qgroup_rescan_worker(struct btrfs_work *work)
2608 {
2609         struct btrfs_fs_info *fs_info = container_of(work, struct btrfs_fs_info,
2610                                                      qgroup_rescan_work);
2611         struct btrfs_path *path;
2612         struct btrfs_trans_handle *trans = NULL;
2613         struct ulist *tmp = NULL, *qgroups = NULL;
2614         struct extent_buffer *scratch_leaf = NULL;
2615         int err = -ENOMEM;
2616
2617         path = btrfs_alloc_path();
2618         if (!path)
2619                 goto out;
2620         qgroups = ulist_alloc(GFP_NOFS);
2621         if (!qgroups)
2622                 goto out;
2623         tmp = ulist_alloc(GFP_NOFS);
2624         if (!tmp)
2625                 goto out;
2626         scratch_leaf = kmalloc(sizeof(*scratch_leaf), GFP_NOFS);
2627         if (!scratch_leaf)
2628                 goto out;
2629
2630         err = 0;
2631         while (!err) {
2632                 trans = btrfs_start_transaction(fs_info->fs_root, 0);
2633                 if (IS_ERR(trans)) {
2634                         err = PTR_ERR(trans);
2635                         break;
2636                 }
2637                 if (!fs_info->quota_enabled) {
2638                         err = -EINTR;
2639                 } else {
2640                         err = qgroup_rescan_leaf(fs_info, path, trans,
2641                                                  qgroups, tmp, scratch_leaf);
2642                 }
2643                 if (err > 0)
2644                         btrfs_commit_transaction(trans, fs_info->fs_root);
2645                 else
2646                         btrfs_end_transaction(trans, fs_info->fs_root);
2647         }
2648
2649 out:
2650         kfree(scratch_leaf);
2651         ulist_free(qgroups);
2652         ulist_free(tmp);
2653         btrfs_free_path(path);
2654
2655         mutex_lock(&fs_info->qgroup_rescan_lock);
2656         fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
2657
2658         if (err == 2 &&
2659             fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT) {
2660                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2661         } else if (err < 0) {
2662                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2663         }
2664         mutex_unlock(&fs_info->qgroup_rescan_lock);
2665
2666         if (err >= 0) {
2667                 btrfs_info(fs_info, "qgroup scan completed%s",
2668                         err == 2 ? " (inconsistency flag cleared)" : "");
2669         } else {
2670                 btrfs_err(fs_info, "qgroup scan failed with %d", err);
2671         }
2672
2673         complete_all(&fs_info->qgroup_rescan_completion);
2674 }
2675
2676 /*
2677  * Checks that (a) no rescan is running and (b) quota is enabled. Allocates all
2678  * memory required for the rescan context.
2679  */
2680 static int
2681 qgroup_rescan_init(struct btrfs_fs_info *fs_info, u64 progress_objectid,
2682                    int init_flags)
2683 {
2684         int ret = 0;
2685
2686         if (!init_flags &&
2687             (!(fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) ||
2688              !(fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_ON))) {
2689                 ret = -EINVAL;
2690                 goto err;
2691         }
2692
2693         mutex_lock(&fs_info->qgroup_rescan_lock);
2694         spin_lock(&fs_info->qgroup_lock);
2695
2696         if (init_flags) {
2697                 if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN)
2698                         ret = -EINPROGRESS;
2699                 else if (!(fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_ON))
2700                         ret = -EINVAL;
2701
2702                 if (ret) {
2703                         spin_unlock(&fs_info->qgroup_lock);
2704                         mutex_unlock(&fs_info->qgroup_rescan_lock);
2705                         goto err;
2706                 }
2707
2708                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_RESCAN;
2709         }
2710
2711         memset(&fs_info->qgroup_rescan_progress, 0,
2712                 sizeof(fs_info->qgroup_rescan_progress));
2713         fs_info->qgroup_rescan_progress.objectid = progress_objectid;
2714
2715         spin_unlock(&fs_info->qgroup_lock);
2716         mutex_unlock(&fs_info->qgroup_rescan_lock);
2717
2718         init_completion(&fs_info->qgroup_rescan_completion);
2719
2720         memset(&fs_info->qgroup_rescan_work, 0,
2721                sizeof(fs_info->qgroup_rescan_work));
2722         btrfs_init_work(&fs_info->qgroup_rescan_work,
2723                         btrfs_qgroup_rescan_worker, NULL, NULL);
2724
2725         if (ret) {
2726 err:
2727                 btrfs_info(fs_info, "qgroup_rescan_init failed with %d", ret);
2728                 return ret;
2729         }
2730
2731         return 0;
2732 }
2733
2734 static void
2735 qgroup_rescan_zero_tracking(struct btrfs_fs_info *fs_info)
2736 {
2737         struct rb_node *n;
2738         struct btrfs_qgroup *qgroup;
2739
2740         spin_lock(&fs_info->qgroup_lock);
2741         /* clear all current qgroup tracking information */
2742         for (n = rb_first(&fs_info->qgroup_tree); n; n = rb_next(n)) {
2743                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
2744                 qgroup->rfer = 0;
2745                 qgroup->rfer_cmpr = 0;
2746                 qgroup->excl = 0;
2747                 qgroup->excl_cmpr = 0;
2748         }
2749         spin_unlock(&fs_info->qgroup_lock);
2750 }
2751
2752 int
2753 btrfs_qgroup_rescan(struct btrfs_fs_info *fs_info)
2754 {
2755         int ret = 0;
2756         struct btrfs_trans_handle *trans;
2757
2758         ret = qgroup_rescan_init(fs_info, 0, 1);
2759         if (ret)
2760                 return ret;
2761
2762         /*
2763          * We have set the rescan_progress to 0, which means no more
2764          * delayed refs will be accounted by btrfs_qgroup_account_ref.
2765          * However, btrfs_qgroup_account_ref may be right after its call
2766          * to btrfs_find_all_roots, in which case it would still do the
2767          * accounting.
2768          * To solve this, we're committing the transaction, which will
2769          * ensure we run all delayed refs and only after that, we are
2770          * going to clear all tracking information for a clean start.
2771          */
2772
2773         trans = btrfs_join_transaction(fs_info->fs_root);
2774         if (IS_ERR(trans)) {
2775                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
2776                 return PTR_ERR(trans);
2777         }
2778         ret = btrfs_commit_transaction(trans, fs_info->fs_root);
2779         if (ret) {
2780                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
2781                 return ret;
2782         }
2783
2784         qgroup_rescan_zero_tracking(fs_info);
2785
2786         btrfs_queue_work(fs_info->qgroup_rescan_workers,
2787                          &fs_info->qgroup_rescan_work);
2788
2789         return 0;
2790 }
2791
2792 int btrfs_qgroup_wait_for_completion(struct btrfs_fs_info *fs_info)
2793 {
2794         int running;
2795         int ret = 0;
2796
2797         mutex_lock(&fs_info->qgroup_rescan_lock);
2798         spin_lock(&fs_info->qgroup_lock);
2799         running = fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN;
2800         spin_unlock(&fs_info->qgroup_lock);
2801         mutex_unlock(&fs_info->qgroup_rescan_lock);
2802
2803         if (running)
2804                 ret = wait_for_completion_interruptible(
2805                                         &fs_info->qgroup_rescan_completion);
2806
2807         return ret;
2808 }
2809
2810 /*
2811  * this is only called from open_ctree where we're still single threaded, thus
2812  * locking is omitted here.
2813  */
2814 void
2815 btrfs_qgroup_rescan_resume(struct btrfs_fs_info *fs_info)
2816 {
2817         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN)
2818                 btrfs_queue_work(fs_info->qgroup_rescan_workers,
2819                                  &fs_info->qgroup_rescan_work);
2820 }