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         if (btrfs_test_is_dummy_root(quota_root))
 543                 return 0;
 544
 545         path = btrfs_alloc_path();
 546         if (!path)
 547                 return -ENOMEM;
 548
 549         key.objectid = 0;
 550         key.type = BTRFS_QGROUP_INFO_KEY;
 551         key.offset = qgroupid;
 552
 553         /*
 554          * Avoid a transaction abort by catching -EEXIST here. In that
 555          * case, we proceed by re-initializing the existing structure
 556          * on disk.
 557          */
 558
 559         ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
 560                                       sizeof(*qgroup_info));
 561         if (ret && ret != -EEXIST)
 562                 goto out;
 563
 564         leaf = path->nodes[0];
 565         qgroup_info = btrfs_item_ptr(leaf, path->slots[0],
 566                                  struct btrfs_qgroup_info_item);
 567         btrfs_set_qgroup_info_generation(leaf, qgroup_info, trans->transid);
 568         btrfs_set_qgroup_info_rfer(leaf, qgroup_info, 0);
 569         btrfs_set_qgroup_info_rfer_cmpr(leaf, qgroup_info, 0);
 570         btrfs_set_qgroup_info_excl(leaf, qgroup_info, 0);
 571         btrfs_set_qgroup_info_excl_cmpr(leaf, qgroup_info, 0);
 572
 573         btrfs_mark_buffer_dirty(leaf);
 574
 575         btrfs_release_path(path);
 576
 577         key.type = BTRFS_QGROUP_LIMIT_KEY;
 578         ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
 579                                       sizeof(*qgroup_limit));
 580         if (ret && ret != -EEXIST)
 581                 goto out;
 582
 583         leaf = path->nodes[0];
 584         qgroup_limit = btrfs_item_ptr(leaf, path->slots[0],
 585                                   struct btrfs_qgroup_limit_item);
 586         btrfs_set_qgroup_limit_flags(leaf, qgroup_limit, 0);
 587         btrfs_set_qgroup_limit_max_rfer(leaf, qgroup_limit, 0);
 588         btrfs_set_qgroup_limit_max_excl(leaf, qgroup_limit, 0);
 589         btrfs_set_qgroup_limit_rsv_rfer(leaf, qgroup_limit, 0);
 590         btrfs_set_qgroup_limit_rsv_excl(leaf, qgroup_limit, 0);
 591
 592         btrfs_mark_buffer_dirty(leaf);
 593
 594         ret = 0;
 595 out:
 596         btrfs_free_path(path);
 597         return ret;
 598 }
 599
 600 static int del_qgroup_item(struct btrfs_trans_handle *trans,
 601                            struct btrfs_root *quota_root, u64 qgroupid)
 602 {
 603         int ret;
 604         struct btrfs_path *path;
 605         struct btrfs_key key;
 606
 607         path = btrfs_alloc_path();
 608         if (!path)
 609                 return -ENOMEM;
 610
 611         key.objectid = 0;
 612         key.type = BTRFS_QGROUP_INFO_KEY;
 613         key.offset = qgroupid;
 614         ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
 615         if (ret < 0)
 616                 goto out;
 617
 618         if (ret > 0) {
 619                 ret = -ENOENT;
 620                 goto out;
 621         }
 622
 623         ret = btrfs_del_item(trans, quota_root, path);
 624         if (ret)
 625                 goto out;
 626
 627         btrfs_release_path(path);
 628
 629         key.type = BTRFS_QGROUP_LIMIT_KEY;
 630         ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
 631         if (ret < 0)
 632                 goto out;
 633
 634         if (ret > 0) {
 635                 ret = -ENOENT;
 636                 goto out;
 637         }
 638
 639         ret = btrfs_del_item(trans, quota_root, path);
 640
 641 out:
 642         btrfs_free_path(path);
 643         return ret;
 644 }
 645
 646 static int update_qgroup_limit_item(struct btrfs_trans_handle *trans,
 647                                     struct btrfs_root *root, u64 qgroupid,
 648                                     u64 flags, u64 max_rfer, u64 max_excl,
 649                                     u64 rsv_rfer, u64 rsv_excl)
 650 {
 651         struct btrfs_path *path;
 652         struct btrfs_key key;
 653         struct extent_buffer *l;
 654         struct btrfs_qgroup_limit_item *qgroup_limit;
 655         int ret;
 656         int slot;
 657
 658         key.objectid = 0;
 659         key.type = BTRFS_QGROUP_LIMIT_KEY;
 660         key.offset = qgroupid;
 661
 662         path = btrfs_alloc_path();
 663         if (!path)
 664                 return -ENOMEM;
 665
 666         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
 667         if (ret > 0)
 668                 ret = -ENOENT;
 669
 670         if (ret)
 671                 goto out;
 672
 673         l = path->nodes[0];
 674         slot = path->slots[0];
 675         qgroup_limit = btrfs_item_ptr(l, slot, struct btrfs_qgroup_limit_item);
 676         btrfs_set_qgroup_limit_flags(l, qgroup_limit, flags);
 677         btrfs_set_qgroup_limit_max_rfer(l, qgroup_limit, max_rfer);
 678         btrfs_set_qgroup_limit_max_excl(l, qgroup_limit, max_excl);
 679         btrfs_set_qgroup_limit_rsv_rfer(l, qgroup_limit, rsv_rfer);
 680         btrfs_set_qgroup_limit_rsv_excl(l, qgroup_limit, rsv_excl);
 681
 682         btrfs_mark_buffer_dirty(l);
 683
 684 out:
 685         btrfs_free_path(path);
 686         return ret;
 687 }
 688
 689 static int update_qgroup_info_item(struct btrfs_trans_handle *trans,
 690                                    struct btrfs_root *root,
 691                                    struct btrfs_qgroup *qgroup)
 692 {
 693         struct btrfs_path *path;
 694         struct btrfs_key key;
 695         struct extent_buffer *l;
 696         struct btrfs_qgroup_info_item *qgroup_info;
 697         int ret;
 698         int slot;
 699
 700         if (btrfs_test_is_dummy_root(root))
 701                 return 0;
 702
 703         key.objectid = 0;
 704         key.type = BTRFS_QGROUP_INFO_KEY;
 705         key.offset = qgroup->qgroupid;
 706
 707         path = btrfs_alloc_path();
 708         if (!path)
 709                 return -ENOMEM;
 710
 711         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
 712         if (ret > 0)
 713                 ret = -ENOENT;
 714
 715         if (ret)
 716                 goto out;
 717
 718         l = path->nodes[0];
 719         slot = path->slots[0];
 720         qgroup_info = btrfs_item_ptr(l, slot, struct btrfs_qgroup_info_item);
 721         btrfs_set_qgroup_info_generation(l, qgroup_info, trans->transid);
 722         btrfs_set_qgroup_info_rfer(l, qgroup_info, qgroup->rfer);
 723         btrfs_set_qgroup_info_rfer_cmpr(l, qgroup_info, qgroup->rfer_cmpr);
 724         btrfs_set_qgroup_info_excl(l, qgroup_info, qgroup->excl);
 725         btrfs_set_qgroup_info_excl_cmpr(l, qgroup_info, qgroup->excl_cmpr);
 726
 727         btrfs_mark_buffer_dirty(l);
 728
 729 out:
 730         btrfs_free_path(path);
 731         return ret;
 732 }
 733
 734 static int update_qgroup_status_item(struct btrfs_trans_handle *trans,
 735                                      struct btrfs_fs_info *fs_info,
 736                                     struct btrfs_root *root)
 737 {
 738         struct btrfs_path *path;
 739         struct btrfs_key key;
 740         struct extent_buffer *l;
 741         struct btrfs_qgroup_status_item *ptr;
 742         int ret;
 743         int slot;
 744
 745         key.objectid = 0;
 746         key.type = BTRFS_QGROUP_STATUS_KEY;
 747         key.offset = 0;
 748
 749         path = btrfs_alloc_path();
 750         if (!path)
 751                 return -ENOMEM;
 752
 753         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
 754         if (ret > 0)
 755                 ret = -ENOENT;
 756
 757         if (ret)
 758                 goto out;
 759
 760         l = path->nodes[0];
 761         slot = path->slots[0];
 762         ptr = btrfs_item_ptr(l, slot, struct btrfs_qgroup_status_item);
 763         btrfs_set_qgroup_status_flags(l, ptr, fs_info->qgroup_flags);
 764         btrfs_set_qgroup_status_generation(l, ptr, trans->transid);
 765         btrfs_set_qgroup_status_rescan(l, ptr,
 766                                 fs_info->qgroup_rescan_progress.objectid);
 767
 768         btrfs_mark_buffer_dirty(l);
 769
 770 out:
 771         btrfs_free_path(path);
 772         return ret;
 773 }
 774
 775 /*
 776  * called with qgroup_lock held
 777  */
 778 static int btrfs_clean_quota_tree(struct btrfs_trans_handle *trans,
 779                                   struct btrfs_root *root)
 780 {
 781         struct btrfs_path *path;
 782         struct btrfs_key key;
 783         struct extent_buffer *leaf = NULL;
 784         int ret;
 785         int nr = 0;
 786
 787         path = btrfs_alloc_path();
 788         if (!path)
 789                 return -ENOMEM;
 790
 791         path->leave_spinning = 1;
 792
 793         key.objectid = 0;
 794         key.offset = 0;
 795         key.type = 0;
 796
 797         while (1) {
 798                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 799                 if (ret < 0)
 800                         goto out;
 801                 leaf = path->nodes[0];
 802                 nr = btrfs_header_nritems(leaf);
 803                 if (!nr)
 804                         break;
 805                 /*
 806                  * delete the leaf one by one
 807                  * since the whole tree is going
 808                  * to be deleted.
 809                  */
 810                 path->slots[0] = 0;
 811                 ret = btrfs_del_items(trans, root, path, 0, nr);
 812                 if (ret)
 813                         goto out;
 814
 815                 btrfs_release_path(path);
 816         }
 817         ret = 0;
 818 out:
 819         root->fs_info->pending_quota_state = 0;
 820         btrfs_free_path(path);
 821         return ret;
 822 }
 823
 824 int btrfs_quota_enable(struct btrfs_trans_handle *trans,
 825                        struct btrfs_fs_info *fs_info)
 826 {
 827         struct btrfs_root *quota_root;
 828         struct btrfs_root *tree_root = fs_info->tree_root;
 829         struct btrfs_path *path = NULL;
 830         struct btrfs_qgroup_status_item *ptr;
 831         struct extent_buffer *leaf;
 832         struct btrfs_key key;
 833         struct btrfs_key found_key;
 834         struct btrfs_qgroup *qgroup = NULL;
 835         int ret = 0;
 836         int slot;
 837
 838         mutex_lock(&fs_info->qgroup_ioctl_lock);
 839         if (fs_info->quota_root) {
 840                 fs_info->pending_quota_state = 1;
 841                 goto out;
 842         }
 843
 844         fs_info->qgroup_ulist = ulist_alloc(GFP_NOFS);
 845         if (!fs_info->qgroup_ulist) {
 846                 ret = -ENOMEM;
 847                 goto out;
 848         }
 849
 850         /*
 851          * initially create the quota tree
 852          */
 853         quota_root = btrfs_create_tree(trans, fs_info,
 854                                        BTRFS_QUOTA_TREE_OBJECTID);
 855         if (IS_ERR(quota_root)) {
 856                 ret =  PTR_ERR(quota_root);
 857                 goto out;
 858         }
 859
 860         path = btrfs_alloc_path();
 861         if (!path) {
 862                 ret = -ENOMEM;
 863                 goto out_free_root;
 864         }
 865
 866         key.objectid = 0;
 867         key.type = BTRFS_QGROUP_STATUS_KEY;
 868         key.offset = 0;
 869
 870         ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
 871                                       sizeof(*ptr));
 872         if (ret)
 873                 goto out_free_path;
 874
 875         leaf = path->nodes[0];
 876         ptr = btrfs_item_ptr(leaf, path->slots[0],
 877                                  struct btrfs_qgroup_status_item);
 878         btrfs_set_qgroup_status_generation(leaf, ptr, trans->transid);
 879         btrfs_set_qgroup_status_version(leaf, ptr, BTRFS_QGROUP_STATUS_VERSION);
 880         fs_info->qgroup_flags = BTRFS_QGROUP_STATUS_FLAG_ON |
 881                                 BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
 882         btrfs_set_qgroup_status_flags(leaf, ptr, fs_info->qgroup_flags);
 883         btrfs_set_qgroup_status_rescan(leaf, ptr, 0);
 884
 885         btrfs_mark_buffer_dirty(leaf);
 886
 887         key.objectid = 0;
 888         key.type = BTRFS_ROOT_REF_KEY;
 889         key.offset = 0;
 890
 891         btrfs_release_path(path);
 892         ret = btrfs_search_slot_for_read(tree_root, &key, path, 1, 0);
 893         if (ret > 0)
 894                 goto out_add_root;
 895         if (ret < 0)
 896                 goto out_free_path;
 897
 898
 899         while (1) {
 900                 slot = path->slots[0];
 901                 leaf = path->nodes[0];
 902                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
 903
 904                 if (found_key.type == BTRFS_ROOT_REF_KEY) {
 905                         ret = add_qgroup_item(trans, quota_root,
 906                                               found_key.offset);
 907                         if (ret)
 908                                 goto out_free_path;
 909
 910                         qgroup = add_qgroup_rb(fs_info, found_key.offset);
 911                         if (IS_ERR(qgroup)) {
 912                                 ret = PTR_ERR(qgroup);
 913                                 goto out_free_path;
 914                         }
 915                 }
 916                 ret = btrfs_next_item(tree_root, path);
 917                 if (ret < 0)
 918                         goto out_free_path;
 919                 if (ret)
 920                         break;
 921         }
 922
 923 out_add_root:
 924         btrfs_release_path(path);
 925         ret = add_qgroup_item(trans, quota_root, BTRFS_FS_TREE_OBJECTID);
 926         if (ret)
 927                 goto out_free_path;
 928
 929         qgroup = add_qgroup_rb(fs_info, BTRFS_FS_TREE_OBJECTID);
 930         if (IS_ERR(qgroup)) {
 931                 ret = PTR_ERR(qgroup);
 932                 goto out_free_path;
 933         }
 934         spin_lock(&fs_info->qgroup_lock);
 935         fs_info->quota_root = quota_root;
 936         fs_info->pending_quota_state = 1;
 937         spin_unlock(&fs_info->qgroup_lock);
 938 out_free_path:
 939         btrfs_free_path(path);
 940 out_free_root:
 941         if (ret) {
 942                 free_extent_buffer(quota_root->node);
 943                 free_extent_buffer(quota_root->commit_root);
 944                 kfree(quota_root);
 945         }
 946 out:
 947         if (ret) {
 948                 ulist_free(fs_info->qgroup_ulist);
 949                 fs_info->qgroup_ulist = NULL;
 950         }
 951         mutex_unlock(&fs_info->qgroup_ioctl_lock);
 952         return ret;
 953 }
 954
 955 int btrfs_quota_disable(struct btrfs_trans_handle *trans,
 956                         struct btrfs_fs_info *fs_info)
 957 {
 958         struct btrfs_root *tree_root = fs_info->tree_root;
 959         struct btrfs_root *quota_root;
 960         int ret = 0;
 961
 962         mutex_lock(&fs_info->qgroup_ioctl_lock);
 963         if (!fs_info->quota_root)
 964                 goto out;
 965         spin_lock(&fs_info->qgroup_lock);
 966         fs_info->quota_enabled = 0;
 967         fs_info->pending_quota_state = 0;
 968         quota_root = fs_info->quota_root;
 969         fs_info->quota_root = NULL;
 970         spin_unlock(&fs_info->qgroup_lock);
 971
 972         btrfs_free_qgroup_config(fs_info);
 973
 974         ret = btrfs_clean_quota_tree(trans, quota_root);
 975         if (ret)
 976                 goto out;
 977
 978         ret = btrfs_del_root(trans, tree_root, &quota_root->root_key);
 979         if (ret)
 980                 goto out;
 981
 982         list_del(&quota_root->dirty_list);
 983
 984         btrfs_tree_lock(quota_root->node);
 985         clean_tree_block(trans, tree_root, quota_root->node);
 986         btrfs_tree_unlock(quota_root->node);
 987         btrfs_free_tree_block(trans, quota_root, quota_root->node, 0, 1);
 988
 989         free_extent_buffer(quota_root->node);
 990         free_extent_buffer(quota_root->commit_root);
 991         kfree(quota_root);
 992 out:
 993         mutex_unlock(&fs_info->qgroup_ioctl_lock);
 994         return ret;
 995 }
 996
 997 static void qgroup_dirty(struct btrfs_fs_info *fs_info,
 998                          struct btrfs_qgroup *qgroup)
 999 {
1000         if (list_empty(&qgroup->dirty))
1001                 list_add(&qgroup->dirty, &fs_info->dirty_qgroups);
1002 }
1003
1004 int btrfs_add_qgroup_relation(struct btrfs_trans_handle *trans,
1005                               struct btrfs_fs_info *fs_info, u64 src, u64 dst)
1006 {
1007         struct btrfs_root *quota_root;
1008         struct btrfs_qgroup *parent;
1009         struct btrfs_qgroup *member;
1010         struct btrfs_qgroup_list *list;
1011         int ret = 0;
1012
1013         mutex_lock(&fs_info->qgroup_ioctl_lock);
1014         quota_root = fs_info->quota_root;
1015         if (!quota_root) {
1016                 ret = -EINVAL;
1017                 goto out;
1018         }
1019         member = find_qgroup_rb(fs_info, src);
1020         parent = find_qgroup_rb(fs_info, dst);
1021         if (!member || !parent) {
1022                 ret = -EINVAL;
1023                 goto out;
1024         }
1025
1026         /* check if such qgroup relation exist firstly */
1027         list_for_each_entry(list, &member->groups, next_group) {
1028                 if (list->group == parent) {
1029                         ret = -EEXIST;
1030                         goto out;
1031                 }
1032         }
1033
1034         ret = add_qgroup_relation_item(trans, quota_root, src, dst);
1035         if (ret)
1036                 goto out;
1037
1038         ret = add_qgroup_relation_item(trans, quota_root, dst, src);
1039         if (ret) {
1040                 del_qgroup_relation_item(trans, quota_root, src, dst);
1041                 goto out;
1042         }
1043
1044         spin_lock(&fs_info->qgroup_lock);
1045         ret = add_relation_rb(quota_root->fs_info, src, dst);
1046         spin_unlock(&fs_info->qgroup_lock);
1047 out:
1048         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1049         return ret;
1050 }
1051
1052 int btrfs_del_qgroup_relation(struct btrfs_trans_handle *trans,
1053                               struct btrfs_fs_info *fs_info, u64 src, u64 dst)
1054 {
1055         struct btrfs_root *quota_root;
1056         struct btrfs_qgroup *parent;
1057         struct btrfs_qgroup *member;
1058         struct btrfs_qgroup_list *list;
1059         int ret = 0;
1060         int err;
1061
1062         mutex_lock(&fs_info->qgroup_ioctl_lock);
1063         quota_root = fs_info->quota_root;
1064         if (!quota_root) {
1065                 ret = -EINVAL;
1066                 goto out;
1067         }
1068
1069         member = find_qgroup_rb(fs_info, src);
1070         parent = find_qgroup_rb(fs_info, dst);
1071         if (!member || !parent) {
1072                 ret = -EINVAL;
1073                 goto out;
1074         }
1075
1076         /* check if such qgroup relation exist firstly */
1077         list_for_each_entry(list, &member->groups, next_group) {
1078                 if (list->group == parent)
1079                         goto exist;
1080         }
1081         ret = -ENOENT;
1082         goto out;
1083 exist:
1084         ret = del_qgroup_relation_item(trans, quota_root, src, dst);
1085         err = del_qgroup_relation_item(trans, quota_root, dst, src);
1086         if (err && !ret)
1087                 ret = err;
1088
1089         spin_lock(&fs_info->qgroup_lock);
1090         del_relation_rb(fs_info, src, dst);
1091         spin_unlock(&fs_info->qgroup_lock);
1092 out:
1093         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1094         return ret;
1095 }
1096
1097 int btrfs_create_qgroup(struct btrfs_trans_handle *trans,
1098                         struct btrfs_fs_info *fs_info, u64 qgroupid, char *name)
1099 {
1100         struct btrfs_root *quota_root;
1101         struct btrfs_qgroup *qgroup;
1102         int ret = 0;
1103
1104         mutex_lock(&fs_info->qgroup_ioctl_lock);
1105         quota_root = fs_info->quota_root;
1106         if (!quota_root) {
1107                 ret = -EINVAL;
1108                 goto out;
1109         }
1110         qgroup = find_qgroup_rb(fs_info, qgroupid);
1111         if (qgroup) {
1112                 ret = -EEXIST;
1113                 goto out;
1114         }
1115
1116         ret = add_qgroup_item(trans, quota_root, qgroupid);
1117         if (ret)
1118                 goto out;
1119
1120         spin_lock(&fs_info->qgroup_lock);
1121         qgroup = add_qgroup_rb(fs_info, qgroupid);
1122         spin_unlock(&fs_info->qgroup_lock);
1123
1124         if (IS_ERR(qgroup))
1125                 ret = PTR_ERR(qgroup);
1126 out:
1127         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1128         return ret;
1129 }
1130
1131 int btrfs_remove_qgroup(struct btrfs_trans_handle *trans,
1132                         struct btrfs_fs_info *fs_info, u64 qgroupid)
1133 {
1134         struct btrfs_root *quota_root;
1135         struct btrfs_qgroup *qgroup;
1136         int ret = 0;
1137
1138         mutex_lock(&fs_info->qgroup_ioctl_lock);
1139         quota_root = fs_info->quota_root;
1140         if (!quota_root) {
1141                 ret = -EINVAL;
1142                 goto out;
1143         }
1144
1145         qgroup = find_qgroup_rb(fs_info, qgroupid);
1146         if (!qgroup) {
1147                 ret = -ENOENT;
1148                 goto out;
1149         } else {
1150                 /* check if there are no relations to this qgroup */
1151                 if (!list_empty(&qgroup->groups) ||
1152                     !list_empty(&qgroup->members)) {
1153                         ret = -EBUSY;
1154                         goto out;
1155                 }
1156         }
1157         ret = del_qgroup_item(trans, quota_root, qgroupid);
1158
1159         spin_lock(&fs_info->qgroup_lock);
1160         del_qgroup_rb(quota_root->fs_info, qgroupid);
1161         spin_unlock(&fs_info->qgroup_lock);
1162 out:
1163         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1164         return ret;
1165 }
1166
1167 int btrfs_limit_qgroup(struct btrfs_trans_handle *trans,
1168                        struct btrfs_fs_info *fs_info, u64 qgroupid,
1169                        struct btrfs_qgroup_limit *limit)
1170 {
1171         struct btrfs_root *quota_root;
1172         struct btrfs_qgroup *qgroup;
1173         int ret = 0;
1174
1175         mutex_lock(&fs_info->qgroup_ioctl_lock);
1176         quota_root = fs_info->quota_root;
1177         if (!quota_root) {
1178                 ret = -EINVAL;
1179                 goto out;
1180         }
1181
1182         qgroup = find_qgroup_rb(fs_info, qgroupid);
1183         if (!qgroup) {
1184                 ret = -ENOENT;
1185                 goto out;
1186         }
1187         ret = update_qgroup_limit_item(trans, quota_root, qgroupid,
1188                                        limit->flags, limit->max_rfer,
1189                                        limit->max_excl, limit->rsv_rfer,
1190                                        limit->rsv_excl);
1191         if (ret) {
1192                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1193                 btrfs_info(fs_info, "unable to update quota limit for %llu",
1194                        qgroupid);
1195         }
1196
1197         spin_lock(&fs_info->qgroup_lock);
1198         qgroup->lim_flags = limit->flags;
1199         qgroup->max_rfer = limit->max_rfer;
1200         qgroup->max_excl = limit->max_excl;
1201         qgroup->rsv_rfer = limit->rsv_rfer;
1202         qgroup->rsv_excl = limit->rsv_excl;
1203         spin_unlock(&fs_info->qgroup_lock);
1204 out:
1205         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1206         return ret;
1207 }
1208
1209 static int comp_oper_exist(struct btrfs_qgroup_operation *oper1,
1210                            struct btrfs_qgroup_operation *oper2)
1211 {
1212         /*
1213          * Ignore seq and type here, we're looking for any operation
1214          * at all related to this extent on that root.
1215          */
1216         if (oper1->bytenr < oper2->bytenr)
1217                 return -1;
1218         if (oper1->bytenr > oper2->bytenr)
1219                 return 1;
1220         if (oper1->ref_root < oper2->ref_root)
1221                 return -1;
1222         if (oper1->ref_root > oper2->ref_root)
1223                 return 1;
1224         return 0;
1225 }
1226
1227 static int qgroup_oper_exists(struct btrfs_fs_info *fs_info,
1228                               struct btrfs_qgroup_operation *oper)
1229 {
1230         struct rb_node *n;
1231         struct btrfs_qgroup_operation *cur;
1232         int cmp;
1233
1234         spin_lock(&fs_info->qgroup_op_lock);
1235         n = fs_info->qgroup_op_tree.rb_node;
1236         while (n) {
1237                 cur = rb_entry(n, struct btrfs_qgroup_operation, n);
1238                 cmp = comp_oper_exist(cur, oper);
1239                 if (cmp < 0) {
1240                         n = n->rb_right;
1241                 } else if (cmp) {
1242                         n = n->rb_left;
1243                 } else {
1244                         spin_unlock(&fs_info->qgroup_op_lock);
1245                         return -EEXIST;
1246                 }
1247         }
1248         spin_unlock(&fs_info->qgroup_op_lock);
1249         return 0;
1250 }
1251
1252 static int comp_oper(struct btrfs_qgroup_operation *oper1,
1253                      struct btrfs_qgroup_operation *oper2)
1254 {
1255         if (oper1->bytenr < oper2->bytenr)
1256                 return -1;
1257         if (oper1->bytenr > oper2->bytenr)
1258                 return 1;
1259         if (oper1->seq < oper2->seq)
1260                 return -1;
1261         if (oper1->seq > oper2->seq)
1262                 return -1;
1263         if (oper1->ref_root < oper2->ref_root)
1264                 return -1;
1265         if (oper1->ref_root > oper2->ref_root)
1266                 return 1;
1267         if (oper1->type < oper2->type)
1268                 return -1;
1269         if (oper1->type > oper2->type)
1270                 return 1;
1271         return 0;
1272 }
1273
1274 static int insert_qgroup_oper(struct btrfs_fs_info *fs_info,
1275                               struct btrfs_qgroup_operation *oper)
1276 {
1277         struct rb_node **p;
1278         struct rb_node *parent = NULL;
1279         struct btrfs_qgroup_operation *cur;
1280         int cmp;
1281
1282         spin_lock(&fs_info->qgroup_op_lock);
1283         p = &fs_info->qgroup_op_tree.rb_node;
1284         while (*p) {
1285                 parent = *p;
1286                 cur = rb_entry(parent, struct btrfs_qgroup_operation, n);
1287                 cmp = comp_oper(cur, oper);
1288                 if (cmp < 0) {
1289                         p = &(*p)->rb_right;
1290                 } else if (cmp) {
1291                         p = &(*p)->rb_left;
1292                 } else {
1293                         spin_unlock(&fs_info->qgroup_op_lock);
1294                         return -EEXIST;
1295                 }
1296         }
1297         rb_link_node(&oper->n, parent, p);
1298         rb_insert_color(&oper->n, &fs_info->qgroup_op_tree);
1299         spin_unlock(&fs_info->qgroup_op_lock);
1300         return 0;
1301 }
1302
1303 /*
1304  * Record a quota operation for processing later on.
1305  * @trans: the transaction we are adding the delayed op to.
1306  * @fs_info: the fs_info for this fs.
1307  * @ref_root: the root of the reference we are acting on,
1308  * @bytenr: the bytenr we are acting on.
1309  * @num_bytes: the number of bytes in the reference.
1310  * @type: the type of operation this is.
1311  * @mod_seq: do we need to get a sequence number for looking up roots.
1312  *
1313  * We just add it to our trans qgroup_ref_list and carry on and process these
1314  * operations in order at some later point.  If the reference root isn't a fs
1315  * root then we don't bother with doing anything.
1316  *
1317  * MUST BE HOLDING THE REF LOCK.
1318  */
1319 int btrfs_qgroup_record_ref(struct btrfs_trans_handle *trans,
1320                             struct btrfs_fs_info *fs_info, u64 ref_root,
1321                             u64 bytenr, u64 num_bytes,
1322                             enum btrfs_qgroup_operation_type type, int mod_seq)
1323 {
1324         struct btrfs_qgroup_operation *oper;
1325         int ret;
1326
1327         if (!is_fstree(ref_root) || !fs_info->quota_enabled)
1328                 return 0;
1329
1330         oper = kmalloc(sizeof(*oper), GFP_NOFS);
1331         if (!oper)
1332                 return -ENOMEM;
1333
1334         oper->ref_root = ref_root;
1335         oper->bytenr = bytenr;
1336         oper->num_bytes = num_bytes;
1337         oper->type = type;
1338         oper->seq = atomic_inc_return(&fs_info->qgroup_op_seq);
1339         INIT_LIST_HEAD(&oper->elem.list);
1340         oper->elem.seq = 0;
1341
1342         trace_btrfs_qgroup_record_ref(oper);
1343
1344         if (type == BTRFS_QGROUP_OPER_SUB_SUBTREE) {
1345                 /*
1346                  * If any operation for this bytenr/ref_root combo
1347                  * exists, then we know it's not exclusively owned and
1348                  * shouldn't be queued up.
1349                  *
1350                  * This also catches the case where we have a cloned
1351                  * extent that gets queued up multiple times during
1352                  * drop snapshot.
1353                  */
1354                 if (qgroup_oper_exists(fs_info, oper)) {
1355                         kfree(oper);
1356                         return 0;
1357                 }
1358         }
1359
1360         ret = insert_qgroup_oper(fs_info, oper);
1361         if (ret) {
1362                 /* Shouldn't happen so have an assert for developers */
1363                 ASSERT(0);
1364                 kfree(oper);
1365                 return ret;
1366         }
1367         list_add_tail(&oper->list, &trans->qgroup_ref_list);
1368
1369         if (mod_seq)
1370                 btrfs_get_tree_mod_seq(fs_info, &oper->elem);
1371
1372         return 0;
1373 }
1374
1375 /*
1376  * The easy accounting, if we are adding/removing the only ref for an extent
1377  * then this qgroup and all of the parent qgroups get their refrence and
1378  * exclusive counts adjusted.
1379  */
1380 static int qgroup_excl_accounting(struct btrfs_fs_info *fs_info,
1381                                   struct btrfs_qgroup_operation *oper)
1382 {
1383         struct btrfs_qgroup *qgroup;
1384         struct ulist *tmp;
1385         struct btrfs_qgroup_list *glist;
1386         struct ulist_node *unode;
1387         struct ulist_iterator uiter;
1388         int sign = 0;
1389         int ret = 0;
1390
1391         tmp = ulist_alloc(GFP_NOFS);
1392         if (!tmp)
1393                 return -ENOMEM;
1394
1395         spin_lock(&fs_info->qgroup_lock);
1396         if (!fs_info->quota_root)
1397                 goto out;
1398         qgroup = find_qgroup_rb(fs_info, oper->ref_root);
1399         if (!qgroup)
1400                 goto out;
1401         switch (oper->type) {
1402         case BTRFS_QGROUP_OPER_ADD_EXCL:
1403                 sign = 1;
1404                 break;
1405         case BTRFS_QGROUP_OPER_SUB_EXCL:
1406                 sign = -1;
1407                 break;
1408         default:
1409                 ASSERT(0);
1410         }
1411         qgroup->rfer += sign * oper->num_bytes;
1412         qgroup->rfer_cmpr += sign * oper->num_bytes;
1413
1414         WARN_ON(sign < 0 && qgroup->excl < oper->num_bytes);
1415         qgroup->excl += sign * oper->num_bytes;
1416         qgroup->excl_cmpr += sign * oper->num_bytes;
1417
1418         qgroup_dirty(fs_info, qgroup);
1419
1420         /* Get all of the parent groups that contain this qgroup */
1421         list_for_each_entry(glist, &qgroup->groups, next_group) {
1422                 ret = ulist_add(tmp, glist->group->qgroupid,
1423                                 ptr_to_u64(glist->group), GFP_ATOMIC);
1424                 if (ret < 0)
1425                         goto out;
1426         }
1427
1428         /* Iterate all of the parents and adjust their reference counts */
1429         ULIST_ITER_INIT(&uiter);
1430         while ((unode = ulist_next(tmp, &uiter))) {
1431                 qgroup = u64_to_ptr(unode->aux);
1432                 qgroup->rfer += sign * oper->num_bytes;
1433                 qgroup->rfer_cmpr += sign * oper->num_bytes;
1434                 qgroup->excl += sign * oper->num_bytes;
1435                 if (sign < 0)
1436                         WARN_ON(qgroup->excl < oper->num_bytes);
1437                 qgroup->excl_cmpr += sign * oper->num_bytes;
1438                 qgroup_dirty(fs_info, qgroup);
1439
1440                 /* Add any parents of the parents */
1441                 list_for_each_entry(glist, &qgroup->groups, next_group) {
1442                         ret = ulist_add(tmp, glist->group->qgroupid,
1443                                         ptr_to_u64(glist->group), GFP_ATOMIC);
1444                         if (ret < 0)
1445                                 goto out;
1446                 }
1447         }
1448         ret = 0;
1449 out:
1450         spin_unlock(&fs_info->qgroup_lock);
1451         ulist_free(tmp);
1452         return ret;
1453 }
1454
1455 /*
1456  * Walk all of the roots that pointed to our bytenr and adjust their refcnts as
1457  * properly.
1458  */
1459 static int qgroup_calc_old_refcnt(struct btrfs_fs_info *fs_info,
1460                                   u64 root_to_skip, struct ulist *tmp,
1461                                   struct ulist *roots, struct ulist *qgroups,
1462                                   u64 seq, int *old_roots, int rescan)
1463 {
1464         struct ulist_node *unode;
1465         struct ulist_iterator uiter;
1466         struct ulist_node *tmp_unode;
1467         struct ulist_iterator tmp_uiter;
1468         struct btrfs_qgroup *qg;
1469         int ret;
1470
1471         ULIST_ITER_INIT(&uiter);
1472         while ((unode = ulist_next(roots, &uiter))) {
1473                 /* We don't count our current root here */
1474                 if (unode->val == root_to_skip)
1475                         continue;
1476                 qg = find_qgroup_rb(fs_info, unode->val);
1477                 if (!qg)
1478                         continue;
1479                 /*
1480                  * We could have a pending removal of this same ref so we may
1481                  * not have actually found our ref root when doing
1482                  * btrfs_find_all_roots, so we need to keep track of how many
1483                  * old roots we find in case we removed ours and added a
1484                  * different one at the same time.  I don't think this could
1485                  * happen in practice but that sort of thinking leads to pain
1486                  * and suffering and to the dark side.
1487                  */
1488                 (*old_roots)++;
1489
1490                 ulist_reinit(tmp);
1491                 ret = ulist_add(qgroups, qg->qgroupid, ptr_to_u64(qg),
1492                                 GFP_ATOMIC);
1493                 if (ret < 0)
1494                         return ret;
1495                 ret = ulist_add(tmp, qg->qgroupid, ptr_to_u64(qg), GFP_ATOMIC);
1496                 if (ret < 0)
1497                         return ret;
1498                 ULIST_ITER_INIT(&tmp_uiter);
1499                 while ((tmp_unode = ulist_next(tmp, &tmp_uiter))) {
1500                         struct btrfs_qgroup_list *glist;
1501
1502                         qg = u64_to_ptr(tmp_unode->aux);
1503                         /*
1504                          * We use this sequence number to keep from having to
1505                          * run the whole list and 0 out the refcnt every time.
1506                          * We basically use sequnce as the known 0 count and
1507                          * then add 1 everytime we see a qgroup.  This is how we
1508                          * get how many of the roots actually point up to the
1509                          * upper level qgroups in order to determine exclusive
1510                          * counts.
1511                          *
1512                          * For rescan we want to set old_refcnt to seq so our
1513                          * exclusive calculations end up correct.
1514                          */
1515                         if (rescan)
1516                                 qg->old_refcnt = seq;
1517                         else if (qg->old_refcnt < seq)
1518                                 qg->old_refcnt = seq + 1;
1519                         else
1520                                 qg->old_refcnt++;
1521
1522                         if (qg->new_refcnt < seq)
1523                                 qg->new_refcnt = seq + 1;
1524                         else
1525                                 qg->new_refcnt++;
1526                         list_for_each_entry(glist, &qg->groups, next_group) {
1527                                 ret = ulist_add(qgroups, glist->group->qgroupid,
1528                                                 ptr_to_u64(glist->group),
1529                                                 GFP_ATOMIC);
1530                                 if (ret < 0)
1531                                         return ret;
1532                                 ret = ulist_add(tmp, glist->group->qgroupid,
1533                                                 ptr_to_u64(glist->group),
1534                                                 GFP_ATOMIC);
1535                                 if (ret < 0)
1536                                         return ret;
1537                         }
1538                 }
1539         }
1540         return 0;
1541 }
1542
1543 /*
1544  * We need to walk forward in our operation tree and account for any roots that
1545  * were deleted after we made this operation.
1546  */
1547 static int qgroup_account_deleted_refs(struct btrfs_fs_info *fs_info,
1548                                        struct btrfs_qgroup_operation *oper,
1549                                        struct ulist *tmp,
1550                                        struct ulist *qgroups, u64 seq,
1551                                        int *old_roots)
1552 {
1553         struct ulist_node *unode;
1554         struct ulist_iterator uiter;
1555         struct btrfs_qgroup *qg;
1556         struct btrfs_qgroup_operation *tmp_oper;
1557         struct rb_node *n;
1558         int ret;
1559
1560         ulist_reinit(tmp);
1561
1562         /*
1563          * We only walk forward in the tree since we're only interested in
1564          * removals that happened _after_  our operation.
1565          */
1566         spin_lock(&fs_info->qgroup_op_lock);
1567         n = rb_next(&oper->n);
1568         spin_unlock(&fs_info->qgroup_op_lock);
1569         if (!n)
1570                 return 0;
1571         tmp_oper = rb_entry(n, struct btrfs_qgroup_operation, n);
1572         while (tmp_oper->bytenr == oper->bytenr) {
1573                 /*
1574                  * If it's not a removal we don't care, additions work out
1575                  * properly with our refcnt tracking.
1576                  */
1577                 if (tmp_oper->type != BTRFS_QGROUP_OPER_SUB_SHARED &&
1578                     tmp_oper->type != BTRFS_QGROUP_OPER_SUB_EXCL)
1579                         goto next;
1580                 qg = find_qgroup_rb(fs_info, tmp_oper->ref_root);
1581                 if (!qg)
1582                         goto next;
1583                 ret = ulist_add(qgroups, qg->qgroupid, ptr_to_u64(qg),
1584                                 GFP_ATOMIC);
1585                 if (ret) {
1586                         if (ret < 0)
1587                                 return ret;
1588                         /*
1589                          * We only want to increase old_roots if this qgroup is
1590                          * not already in the list of qgroups.  If it is already
1591                          * there then that means it must have been re-added or
1592                          * the delete will be discarded because we had an
1593                          * existing ref that we haven't looked up yet.  In this
1594                          * case we don't want to increase old_roots.  So if ret
1595                          * == 1 then we know that this is the first time we've
1596                          * seen this qgroup and we can bump the old_roots.
1597                          */
1598                         (*old_roots)++;
1599                         ret = ulist_add(tmp, qg->qgroupid, ptr_to_u64(qg),
1600                                         GFP_ATOMIC);
1601                         if (ret < 0)
1602                                 return ret;
1603                 }
1604 next:
1605                 spin_lock(&fs_info->qgroup_op_lock);
1606                 n = rb_next(&tmp_oper->n);
1607                 spin_unlock(&fs_info->qgroup_op_lock);
1608                 if (!n)
1609                         break;
1610                 tmp_oper = rb_entry(n, struct btrfs_qgroup_operation, n);
1611         }
1612
1613         /* Ok now process the qgroups we found */
1614         ULIST_ITER_INIT(&uiter);
1615         while ((unode = ulist_next(tmp, &uiter))) {
1616                 struct btrfs_qgroup_list *glist;
1617
1618                 qg = u64_to_ptr(unode->aux);
1619                 if (qg->old_refcnt < seq)
1620                         qg->old_refcnt = seq + 1;
1621                 else
1622                         qg->old_refcnt++;
1623                 if (qg->new_refcnt < seq)
1624                         qg->new_refcnt = seq + 1;
1625                 else
1626                         qg->new_refcnt++;
1627                 list_for_each_entry(glist, &qg->groups, next_group) {
1628                         ret = ulist_add(qgroups, glist->group->qgroupid,
1629                                         ptr_to_u64(glist->group), GFP_ATOMIC);
1630                         if (ret < 0)
1631                                 return ret;
1632                         ret = ulist_add(tmp, glist->group->qgroupid,
1633                                         ptr_to_u64(glist->group), GFP_ATOMIC);
1634                         if (ret < 0)
1635                                 return ret;
1636                 }
1637         }
1638         return 0;
1639 }
1640
1641 /* Add refcnt for the newly added reference. */
1642 static int qgroup_calc_new_refcnt(struct btrfs_fs_info *fs_info,
1643                                   struct btrfs_qgroup_operation *oper,
1644                                   struct btrfs_qgroup *qgroup,
1645                                   struct ulist *tmp, struct ulist *qgroups,
1646                                   u64 seq)
1647 {
1648         struct ulist_node *unode;
1649         struct ulist_iterator uiter;
1650         struct btrfs_qgroup *qg;
1651         int ret;
1652
1653         ulist_reinit(tmp);
1654         ret = ulist_add(qgroups, qgroup->qgroupid, ptr_to_u64(qgroup),
1655                         GFP_ATOMIC);
1656         if (ret < 0)
1657                 return ret;
1658         ret = ulist_add(tmp, qgroup->qgroupid, ptr_to_u64(qgroup),
1659                         GFP_ATOMIC);
1660         if (ret < 0)
1661                 return ret;
1662         ULIST_ITER_INIT(&uiter);
1663         while ((unode = ulist_next(tmp, &uiter))) {
1664                 struct btrfs_qgroup_list *glist;
1665
1666                 qg = u64_to_ptr(unode->aux);
1667                 if (oper->type == BTRFS_QGROUP_OPER_ADD_SHARED) {
1668                         if (qg->new_refcnt < seq)
1669                                 qg->new_refcnt = seq + 1;
1670                         else
1671                                 qg->new_refcnt++;
1672                 } else {
1673                         if (qg->old_refcnt < seq)
1674                                 qg->old_refcnt = seq + 1;
1675                         else
1676                                 qg->old_refcnt++;
1677                 }
1678                 list_for_each_entry(glist, &qg->groups, next_group) {
1679                         ret = ulist_add(tmp, glist->group->qgroupid,
1680                                         ptr_to_u64(glist->group), GFP_ATOMIC);
1681                         if (ret < 0)
1682                                 return ret;
1683                         ret = ulist_add(qgroups, glist->group->qgroupid,
1684                                         ptr_to_u64(glist->group), GFP_ATOMIC);
1685                         if (ret < 0)
1686                                 return ret;
1687                 }
1688         }
1689         return 0;
1690 }
1691
1692 /*
1693  * This adjusts the counters for all referenced qgroups if need be.
1694  */
1695 static int qgroup_adjust_counters(struct btrfs_fs_info *fs_info,
1696                                   u64 root_to_skip, u64 num_bytes,
1697                                   struct ulist *qgroups, u64 seq,
1698                                   int old_roots, int new_roots, int rescan)
1699 {
1700         struct ulist_node *unode;
1701         struct ulist_iterator uiter;
1702         struct btrfs_qgroup *qg;
1703         u64 cur_new_count, cur_old_count;
1704
1705         ULIST_ITER_INIT(&uiter);
1706         while ((unode = ulist_next(qgroups, &uiter))) {
1707                 bool dirty = false;
1708
1709                 qg = u64_to_ptr(unode->aux);
1710                 /*
1711                  * Wasn't referenced before but is now, add to the reference
1712                  * counters.
1713                  */
1714                 if (qg->old_refcnt <= seq && qg->new_refcnt > seq) {
1715                         qg->rfer += num_bytes;
1716                         qg->rfer_cmpr += num_bytes;
1717                         dirty = true;
1718                 }
1719
1720                 /*
1721                  * Was referenced before but isn't now, subtract from the
1722                  * reference counters.
1723                  */
1724                 if (qg->old_refcnt > seq && qg->new_refcnt <= seq) {
1725                         qg->rfer -= num_bytes;
1726                         qg->rfer_cmpr -= num_bytes;
1727                         dirty = true;
1728                 }
1729
1730                 if (qg->old_refcnt < seq)
1731                         cur_old_count = 0;
1732                 else
1733                         cur_old_count = qg->old_refcnt - seq;
1734                 if (qg->new_refcnt < seq)
1735                         cur_new_count = 0;
1736                 else
1737                         cur_new_count = qg->new_refcnt - seq;
1738
1739                 /*
1740                  * If our refcount was the same as the roots previously but our
1741                  * new count isn't the same as the number of roots now then we
1742                  * went from having a exclusive reference on this range to not.
1743                  */
1744                 if (old_roots && cur_old_count == old_roots &&
1745                     (cur_new_count != new_roots || new_roots == 0)) {
1746                         WARN_ON(cur_new_count != new_roots && new_roots == 0);
1747                         qg->excl -= num_bytes;
1748                         qg->excl_cmpr -= num_bytes;
1749                         dirty = true;
1750                 }
1751
1752                 /*
1753                  * If we didn't reference all the roots before but now we do we
1754                  * have an exclusive reference to this range.
1755                  */
1756                 if ((!old_roots || (old_roots && cur_old_count != old_roots))
1757                     && cur_new_count == new_roots) {
1758                         qg->excl += num_bytes;
1759                         qg->excl_cmpr += num_bytes;
1760                         dirty = true;
1761                 }
1762
1763                 if (dirty)
1764                         qgroup_dirty(fs_info, qg);
1765         }
1766         return 0;
1767 }
1768
1769 /*
1770  * If we removed a data extent and there were other references for that bytenr
1771  * then we need to lookup all referenced roots to make sure we still don't
1772  * reference this bytenr.  If we do then we can just discard this operation.
1773  */
1774 static int check_existing_refs(struct btrfs_trans_handle *trans,
1775                                struct btrfs_fs_info *fs_info,
1776                                struct btrfs_qgroup_operation *oper)
1777 {
1778         struct ulist *roots = NULL;
1779         struct ulist_node *unode;
1780         struct ulist_iterator uiter;
1781         int ret = 0;
1782
1783         ret = btrfs_find_all_roots(trans, fs_info, oper->bytenr,
1784                                    oper->elem.seq, &roots);
1785         if (ret < 0)
1786                 return ret;
1787         ret = 0;
1788
1789         ULIST_ITER_INIT(&uiter);
1790         while ((unode = ulist_next(roots, &uiter))) {
1791                 if (unode->val == oper->ref_root) {
1792                         ret = 1;
1793                         break;
1794                 }
1795         }
1796         ulist_free(roots);
1797         btrfs_put_tree_mod_seq(fs_info, &oper->elem);
1798
1799         return ret;
1800 }
1801
1802 /*
1803  * If we share a reference across multiple roots then we may need to adjust
1804  * various qgroups referenced and exclusive counters.  The basic premise is this
1805  *
1806  * 1) We have seq to represent a 0 count.  Instead of looping through all of the
1807  * qgroups and resetting their refcount to 0 we just constantly bump this
1808  * sequence number to act as the base reference count.  This means that if
1809  * anybody is equal to or below this sequence they were never referenced.  We
1810  * jack this sequence up by the number of roots we found each time in order to
1811  * make sure we don't have any overlap.
1812  *
1813  * 2) We first search all the roots that reference the area _except_ the root
1814  * we're acting on currently.  This makes up the old_refcnt of all the qgroups
1815  * before.
1816  *
1817  * 3) We walk all of the qgroups referenced by the root we are currently acting
1818  * on, and will either adjust old_refcnt in the case of a removal or the
1819  * new_refcnt in the case of an addition.
1820  *
1821  * 4) Finally we walk all the qgroups that are referenced by this range
1822  * including the root we are acting on currently.  We will adjust the counters
1823  * based on the number of roots we had and will have after this operation.
1824  *
1825  * Take this example as an illustration
1826  *
1827  *                      [qgroup 1/0]
1828  *                   /         |          \
1829  *              [qg 0/0]   [qg 0/1]     [qg 0/2]
1830  *                 \          |            /
1831  *                [        extent           ]
1832  *
1833  * Say we are adding a reference that is covered by qg 0/0.  The first step
1834  * would give a refcnt of 1 to qg 0/1 and 0/2 and a refcnt of 2 to qg 1/0 with
1835  * old_roots being 2.  Because it is adding new_roots will be 1.  We then go
1836  * through qg 0/0 which will get the new_refcnt set to 1 and add 1 to qg 1/0's
1837  * new_refcnt, bringing it to 3.  We then walk through all of the qgroups, we
1838  * notice that the old refcnt for qg 0/0 < the new refcnt, so we added a
1839  * reference and thus must add the size to the referenced bytes.  Everything
1840  * else is the same so nothing else changes.
1841  */
1842 static int qgroup_shared_accounting(struct btrfs_trans_handle *trans,
1843                                     struct btrfs_fs_info *fs_info,
1844                                     struct btrfs_qgroup_operation *oper)
1845 {
1846         struct ulist *roots = NULL;
1847         struct ulist *qgroups, *tmp;
1848         struct btrfs_qgroup *qgroup;
1849         struct seq_list elem = {};
1850         u64 seq;
1851         int old_roots = 0;
1852         int new_roots = 0;
1853         int ret = 0;
1854
1855         if (oper->elem.seq) {
1856                 ret = check_existing_refs(trans, fs_info, oper);
1857                 if (ret < 0)
1858                         return ret;
1859                 if (ret)
1860                         return 0;
1861         }
1862
1863         qgroups = ulist_alloc(GFP_NOFS);
1864         if (!qgroups)
1865                 return -ENOMEM;
1866
1867         tmp = ulist_alloc(GFP_NOFS);
1868         if (!tmp) {
1869                 ulist_free(qgroups);
1870                 return -ENOMEM;
1871         }
1872
1873         btrfs_get_tree_mod_seq(fs_info, &elem);
1874         ret = btrfs_find_all_roots(trans, fs_info, oper->bytenr, elem.seq,
1875                                    &roots);
1876         btrfs_put_tree_mod_seq(fs_info, &elem);
1877         if (ret < 0) {
1878                 ulist_free(qgroups);
1879                 ulist_free(tmp);
1880                 return ret;
1881         }
1882         spin_lock(&fs_info->qgroup_lock);
1883         qgroup = find_qgroup_rb(fs_info, oper->ref_root);
1884         if (!qgroup)
1885                 goto out;
1886         seq = fs_info->qgroup_seq;
1887
1888         /*
1889          * So roots is the list of all the roots currently pointing at the
1890          * bytenr, including the ref we are adding if we are adding, or not if
1891          * we are removing a ref.  So we pass in the ref_root to skip that root
1892          * in our calculations.  We set old_refnct and new_refcnt cause who the
1893          * hell knows what everything looked like before, and it doesn't matter
1894          * except...
1895          */
1896         ret = qgroup_calc_old_refcnt(fs_info, oper->ref_root, tmp, roots, qgroups,
1897                                      seq, &old_roots, 0);
1898         if (ret < 0)
1899                 goto out;
1900
1901         /*
1902          * Now adjust the refcounts of the qgroups that care about this
1903          * reference, either the old_count in the case of removal or new_count
1904          * in the case of an addition.
1905          */
1906         ret = qgroup_calc_new_refcnt(fs_info, oper, qgroup, tmp, qgroups,
1907                                      seq);
1908         if (ret < 0)
1909                 goto out;
1910
1911         /*
1912          * ...in the case of removals.  If we had a removal before we got around
1913          * to processing this operation then we need to find that guy and count
1914          * his references as if they really existed so we don't end up screwing
1915          * up the exclusive counts.  Then whenever we go to process the delete
1916          * everything will be grand and we can account for whatever exclusive
1917          * changes need to be made there.  We also have to pass in old_roots so
1918          * we have an accurate count of the roots as it pertains to this
1919          * operations view of the world.
1920          */
1921         ret = qgroup_account_deleted_refs(fs_info, oper, tmp, qgroups, seq,
1922                                           &old_roots);
1923         if (ret < 0)
1924                 goto out;
1925
1926         /*
1927          * We are adding our root, need to adjust up the number of roots,
1928          * otherwise old_roots is the number of roots we want.
1929          */
1930         if (oper->type == BTRFS_QGROUP_OPER_ADD_SHARED) {
1931                 new_roots = old_roots + 1;
1932         } else {
1933                 new_roots = old_roots;
1934                 old_roots++;
1935         }
1936         fs_info->qgroup_seq += old_roots + 1;
1937
1938
1939         /*
1940          * And now the magic happens, bless Arne for having a pretty elegant
1941          * solution for this.
1942          */
1943         qgroup_adjust_counters(fs_info, oper->ref_root, oper->num_bytes,
1944                                qgroups, seq, old_roots, new_roots, 0);
1945 out:
1946         spin_unlock(&fs_info->qgroup_lock);
1947         ulist_free(qgroups);
1948         ulist_free(roots);
1949         ulist_free(tmp);
1950         return ret;
1951 }
1952
1953 /*
1954  * Process a reference to a shared subtree. This type of operation is
1955  * queued during snapshot removal when we encounter extents which are
1956  * shared between more than one root.
1957  */
1958 static int qgroup_subtree_accounting(struct btrfs_trans_handle *trans,
1959                                      struct btrfs_fs_info *fs_info,
1960                                      struct btrfs_qgroup_operation *oper)
1961 {
1962         struct ulist *roots = NULL;
1963         struct ulist_node *unode;
1964         struct ulist_iterator uiter;
1965         struct btrfs_qgroup_list *glist;
1966         struct ulist *parents;
1967         int ret = 0;
1968         int err;
1969         struct btrfs_qgroup *qg;
1970         u64 root_obj = 0;
1971         struct seq_list elem = {};
1972
1973         parents = ulist_alloc(GFP_NOFS);
1974         if (!parents)
1975                 return -ENOMEM;
1976
1977         btrfs_get_tree_mod_seq(fs_info, &elem);
1978         ret = btrfs_find_all_roots(trans, fs_info, oper->bytenr,
1979                                    elem.seq, &roots);
1980         btrfs_put_tree_mod_seq(fs_info, &elem);
1981         if (ret < 0)
1982                 goto out;
1983
1984         if (roots->nnodes != 1)
1985                 goto out;
1986
1987         ULIST_ITER_INIT(&uiter);
1988         unode = ulist_next(roots, &uiter); /* Only want 1 so no need to loop */
1989         /*
1990          * If we find our ref root then that means all refs
1991          * this extent has to the root have not yet been
1992          * deleted. In that case, we do nothing and let the
1993          * last ref for this bytenr drive our update.
1994          *
1995          * This can happen for example if an extent is
1996          * referenced multiple times in a snapshot (clone,
1997          * etc). If we are in the middle of snapshot removal,
1998          * queued updates for such an extent will find the
1999          * root if we have not yet finished removing the
2000          * snapshot.
2001          */
2002         if (unode->val == oper->ref_root)
2003                 goto out;
2004
2005         root_obj = unode->val;
2006         BUG_ON(!root_obj);
2007
2008         spin_lock(&fs_info->qgroup_lock);
2009         qg = find_qgroup_rb(fs_info, root_obj);
2010         if (!qg)
2011                 goto out_unlock;
2012
2013         qg->excl += oper->num_bytes;
2014         qg->excl_cmpr += oper->num_bytes;
2015         qgroup_dirty(fs_info, qg);
2016
2017         /*
2018          * Adjust counts for parent groups. First we find all
2019          * parents, then in the 2nd loop we do the adjustment
2020          * while adding parents of the parents to our ulist.
2021          */
2022         list_for_each_entry(glist, &qg->groups, next_group) {
2023                 err = ulist_add(parents, glist->group->qgroupid,
2024                                 ptr_to_u64(glist->group), GFP_ATOMIC);
2025                 if (err < 0) {
2026                         ret = err;
2027                         goto out_unlock;
2028                 }
2029         }
2030
2031         ULIST_ITER_INIT(&uiter);
2032         while ((unode = ulist_next(parents, &uiter))) {
2033                 qg = u64_to_ptr(unode->aux);
2034                 qg->excl += oper->num_bytes;
2035                 qg->excl_cmpr += oper->num_bytes;
2036                 qgroup_dirty(fs_info, qg);
2037
2038                 /* Add any parents of the parents */
2039                 list_for_each_entry(glist, &qg->groups, next_group) {
2040                         err = ulist_add(parents, glist->group->qgroupid,
2041                                         ptr_to_u64(glist->group), GFP_ATOMIC);
2042                         if (err < 0) {
2043                                 ret = err;
2044                                 goto out_unlock;
2045                         }
2046                 }
2047         }
2048
2049 out_unlock:
2050         spin_unlock(&fs_info->qgroup_lock);
2051
2052 out:
2053         ulist_free(roots);
2054         ulist_free(parents);
2055         return ret;
2056 }
2057
2058 /*
2059  * btrfs_qgroup_account_ref is called for every ref that is added to or deleted
2060  * from the fs. First, all roots referencing the extent are searched, and
2061  * then the space is accounted accordingly to the different roots. The
2062  * accounting algorithm works in 3 steps documented inline.
2063  */
2064 static int btrfs_qgroup_account(struct btrfs_trans_handle *trans,
2065                                 struct btrfs_fs_info *fs_info,
2066                                 struct btrfs_qgroup_operation *oper)
2067 {
2068         int ret = 0;
2069
2070         if (!fs_info->quota_enabled)
2071                 return 0;
2072
2073         BUG_ON(!fs_info->quota_root);
2074
2075         mutex_lock(&fs_info->qgroup_rescan_lock);
2076         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
2077                 if (fs_info->qgroup_rescan_progress.objectid <= oper->bytenr) {
2078                         mutex_unlock(&fs_info->qgroup_rescan_lock);
2079                         return 0;
2080                 }
2081         }
2082         mutex_unlock(&fs_info->qgroup_rescan_lock);
2083
2084         ASSERT(is_fstree(oper->ref_root));
2085
2086         trace_btrfs_qgroup_account(oper);
2087
2088         switch (oper->type) {
2089         case BTRFS_QGROUP_OPER_ADD_EXCL:
2090         case BTRFS_QGROUP_OPER_SUB_EXCL:
2091                 ret = qgroup_excl_accounting(fs_info, oper);
2092                 break;
2093         case BTRFS_QGROUP_OPER_ADD_SHARED:
2094         case BTRFS_QGROUP_OPER_SUB_SHARED:
2095                 ret = qgroup_shared_accounting(trans, fs_info, oper);
2096                 break;
2097         case BTRFS_QGROUP_OPER_SUB_SUBTREE:
2098                 ret = qgroup_subtree_accounting(trans, fs_info, oper);
2099                 break;
2100         default:
2101                 ASSERT(0);
2102         }
2103         return ret;
2104 }
2105
2106 /*
2107  * Needs to be called everytime we run delayed refs, even if there is an error
2108  * in order to cleanup outstanding operations.
2109  */
2110 int btrfs_delayed_qgroup_accounting(struct btrfs_trans_handle *trans,
2111                                     struct btrfs_fs_info *fs_info)
2112 {
2113         struct btrfs_qgroup_operation *oper;
2114         int ret = 0;
2115
2116         while (!list_empty(&trans->qgroup_ref_list)) {
2117                 oper = list_first_entry(&trans->qgroup_ref_list,
2118                                         struct btrfs_qgroup_operation, list);
2119                 list_del_init(&oper->list);
2120                 if (!ret || !trans->aborted)
2121                         ret = btrfs_qgroup_account(trans, fs_info, oper);
2122                 spin_lock(&fs_info->qgroup_op_lock);
2123                 rb_erase(&oper->n, &fs_info->qgroup_op_tree);
2124                 spin_unlock(&fs_info->qgroup_op_lock);
2125                 btrfs_put_tree_mod_seq(fs_info, &oper->elem);
2126                 kfree(oper);
2127         }
2128         return ret;
2129 }
2130
2131 /*
2132  * called from commit_transaction. Writes all changed qgroups to disk.
2133  */
2134 int btrfs_run_qgroups(struct btrfs_trans_handle *trans,
2135                       struct btrfs_fs_info *fs_info)
2136 {
2137         struct btrfs_root *quota_root = fs_info->quota_root;
2138         int ret = 0;
2139         int start_rescan_worker = 0;
2140
2141         if (!quota_root)
2142                 goto out;
2143
2144         if (!fs_info->quota_enabled && fs_info->pending_quota_state)
2145                 start_rescan_worker = 1;
2146
2147         fs_info->quota_enabled = fs_info->pending_quota_state;
2148
2149         spin_lock(&fs_info->qgroup_lock);
2150         while (!list_empty(&fs_info->dirty_qgroups)) {
2151                 struct btrfs_qgroup *qgroup;
2152                 qgroup = list_first_entry(&fs_info->dirty_qgroups,
2153                                           struct btrfs_qgroup, dirty);
2154                 list_del_init(&qgroup->dirty);
2155                 spin_unlock(&fs_info->qgroup_lock);
2156                 ret = update_qgroup_info_item(trans, quota_root, qgroup);
2157                 if (ret)
2158                         fs_info->qgroup_flags |=
2159                                         BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2160                 spin_lock(&fs_info->qgroup_lock);
2161         }
2162         if (fs_info->quota_enabled)
2163                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_ON;
2164         else
2165                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_ON;
2166         spin_unlock(&fs_info->qgroup_lock);
2167
2168         ret = update_qgroup_status_item(trans, fs_info, quota_root);
2169         if (ret)
2170                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2171
2172         if (!ret && start_rescan_worker) {
2173                 ret = qgroup_rescan_init(fs_info, 0, 1);
2174                 if (!ret) {
2175                         qgroup_rescan_zero_tracking(fs_info);
2176                         btrfs_queue_work(fs_info->qgroup_rescan_workers,
2177                                          &fs_info->qgroup_rescan_work);
2178                 }
2179                 ret = 0;
2180         }
2181
2182 out:
2183
2184         return ret;
2185 }
2186
2187 /*
2188  * copy the acounting information between qgroups. This is necessary when a
2189  * snapshot or a subvolume is created
2190  */
2191 int btrfs_qgroup_inherit(struct btrfs_trans_handle *trans,
2192                          struct btrfs_fs_info *fs_info, u64 srcid, u64 objectid,
2193                          struct btrfs_qgroup_inherit *inherit)
2194 {
2195         int ret = 0;
2196         int i;
2197         u64 *i_qgroups;
2198         struct btrfs_root *quota_root = fs_info->quota_root;
2199         struct btrfs_qgroup *srcgroup;
2200         struct btrfs_qgroup *dstgroup;
2201         u32 level_size = 0;
2202         u64 nums;
2203
2204         mutex_lock(&fs_info->qgroup_ioctl_lock);
2205         if (!fs_info->quota_enabled)
2206                 goto out;
2207
2208         if (!quota_root) {
2209                 ret = -EINVAL;
2210                 goto out;
2211         }
2212
2213         if (inherit) {
2214                 i_qgroups = (u64 *)(inherit + 1);
2215                 nums = inherit->num_qgroups + 2 * inherit->num_ref_copies +
2216                        2 * inherit->num_excl_copies;
2217                 for (i = 0; i < nums; ++i) {
2218                         srcgroup = find_qgroup_rb(fs_info, *i_qgroups);
2219                         if (!srcgroup) {
2220                                 ret = -EINVAL;
2221                                 goto out;
2222                         }
2223                         ++i_qgroups;
2224                 }
2225         }
2226
2227         /*
2228          * create a tracking group for the subvol itself
2229          */
2230         ret = add_qgroup_item(trans, quota_root, objectid);
2231         if (ret)
2232                 goto out;
2233
2234         if (inherit && inherit->flags & BTRFS_QGROUP_INHERIT_SET_LIMITS) {
2235                 ret = update_qgroup_limit_item(trans, quota_root, objectid,
2236                                                inherit->lim.flags,
2237                                                inherit->lim.max_rfer,
2238                                                inherit->lim.max_excl,
2239                                                inherit->lim.rsv_rfer,
2240                                                inherit->lim.rsv_excl);
2241                 if (ret)
2242                         goto out;
2243         }
2244
2245         if (srcid) {
2246                 struct btrfs_root *srcroot;
2247                 struct btrfs_key srckey;
2248
2249                 srckey.objectid = srcid;
2250                 srckey.type = BTRFS_ROOT_ITEM_KEY;
2251                 srckey.offset = (u64)-1;
2252                 srcroot = btrfs_read_fs_root_no_name(fs_info, &srckey);
2253                 if (IS_ERR(srcroot)) {
2254                         ret = PTR_ERR(srcroot);
2255                         goto out;
2256                 }
2257
2258                 rcu_read_lock();
2259                 level_size = srcroot->nodesize;
2260                 rcu_read_unlock();
2261         }
2262
2263         /*
2264          * add qgroup to all inherited groups
2265          */
2266         if (inherit) {
2267                 i_qgroups = (u64 *)(inherit + 1);
2268                 for (i = 0; i < inherit->num_qgroups; ++i) {
2269                         ret = add_qgroup_relation_item(trans, quota_root,
2270                                                        objectid, *i_qgroups);
2271                         if (ret)
2272                                 goto out;
2273                         ret = add_qgroup_relation_item(trans, quota_root,
2274                                                        *i_qgroups, objectid);
2275                         if (ret)
2276                                 goto out;
2277                         ++i_qgroups;
2278                 }
2279         }
2280
2281
2282         spin_lock(&fs_info->qgroup_lock);
2283
2284         dstgroup = add_qgroup_rb(fs_info, objectid);
2285         if (IS_ERR(dstgroup)) {
2286                 ret = PTR_ERR(dstgroup);
2287                 goto unlock;
2288         }
2289
2290         if (srcid) {
2291                 srcgroup = find_qgroup_rb(fs_info, srcid);
2292                 if (!srcgroup)
2293                         goto unlock;
2294
2295                 /*
2296                  * We call inherit after we clone the root in order to make sure
2297                  * our counts don't go crazy, so at this point the only
2298                  * difference between the two roots should be the root node.
2299                  */
2300                 dstgroup->rfer = srcgroup->rfer;
2301                 dstgroup->rfer_cmpr = srcgroup->rfer_cmpr;
2302                 dstgroup->excl = level_size;
2303                 dstgroup->excl_cmpr = level_size;
2304                 srcgroup->excl = level_size;
2305                 srcgroup->excl_cmpr = level_size;
2306                 qgroup_dirty(fs_info, dstgroup);
2307                 qgroup_dirty(fs_info, srcgroup);
2308         }
2309
2310         if (!inherit)
2311                 goto unlock;
2312
2313         i_qgroups = (u64 *)(inherit + 1);
2314         for (i = 0; i < inherit->num_qgroups; ++i) {
2315                 ret = add_relation_rb(quota_root->fs_info, objectid,
2316                                       *i_qgroups);
2317                 if (ret)
2318                         goto unlock;
2319                 ++i_qgroups;
2320         }
2321
2322         for (i = 0; i <  inherit->num_ref_copies; ++i) {
2323                 struct btrfs_qgroup *src;
2324                 struct btrfs_qgroup *dst;
2325
2326                 src = find_qgroup_rb(fs_info, i_qgroups[0]);
2327                 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
2328
2329                 if (!src || !dst) {
2330                         ret = -EINVAL;
2331                         goto unlock;
2332                 }
2333
2334                 dst->rfer = src->rfer - level_size;
2335                 dst->rfer_cmpr = src->rfer_cmpr - level_size;
2336                 i_qgroups += 2;
2337         }
2338         for (i = 0; i <  inherit->num_excl_copies; ++i) {
2339                 struct btrfs_qgroup *src;
2340                 struct btrfs_qgroup *dst;
2341
2342                 src = find_qgroup_rb(fs_info, i_qgroups[0]);
2343                 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
2344
2345                 if (!src || !dst) {
2346                         ret = -EINVAL;
2347                         goto unlock;
2348                 }
2349
2350                 dst->excl = src->excl + level_size;
2351                 dst->excl_cmpr = src->excl_cmpr + level_size;
2352                 i_qgroups += 2;
2353         }
2354
2355 unlock:
2356         spin_unlock(&fs_info->qgroup_lock);
2357 out:
2358         mutex_unlock(&fs_info->qgroup_ioctl_lock);
2359         return ret;
2360 }
2361
2362 /*
2363  * reserve some space for a qgroup and all its parents. The reservation takes
2364  * place with start_transaction or dealloc_reserve, similar to ENOSPC
2365  * accounting. If not enough space is available, EDQUOT is returned.
2366  * We assume that the requested space is new for all qgroups.
2367  */
2368 int btrfs_qgroup_reserve(struct btrfs_root *root, u64 num_bytes)
2369 {
2370         struct btrfs_root *quota_root;
2371         struct btrfs_qgroup *qgroup;
2372         struct btrfs_fs_info *fs_info = root->fs_info;
2373         u64 ref_root = root->root_key.objectid;
2374         int ret = 0;
2375         struct ulist_node *unode;
2376         struct ulist_iterator uiter;
2377
2378         if (!is_fstree(ref_root))
2379                 return 0;
2380
2381         if (num_bytes == 0)
2382                 return 0;
2383
2384         spin_lock(&fs_info->qgroup_lock);
2385         quota_root = fs_info->quota_root;
2386         if (!quota_root)
2387                 goto out;
2388
2389         qgroup = find_qgroup_rb(fs_info, ref_root);
2390         if (!qgroup)
2391                 goto out;
2392
2393         /*
2394          * in a first step, we check all affected qgroups if any limits would
2395          * be exceeded
2396          */
2397         ulist_reinit(fs_info->qgroup_ulist);
2398         ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
2399                         (uintptr_t)qgroup, GFP_ATOMIC);
2400         if (ret < 0)
2401                 goto out;
2402         ULIST_ITER_INIT(&uiter);
2403         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
2404                 struct btrfs_qgroup *qg;
2405                 struct btrfs_qgroup_list *glist;
2406
2407                 qg = u64_to_ptr(unode->aux);
2408
2409                 if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_RFER) &&
2410                     qg->reserved + (s64)qg->rfer + num_bytes >
2411                     qg->max_rfer) {
2412                         ret = -EDQUOT;
2413                         goto out;
2414                 }
2415
2416                 if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) &&
2417                     qg->reserved + (s64)qg->excl + num_bytes >
2418                     qg->max_excl) {
2419                         ret = -EDQUOT;
2420                         goto out;
2421                 }
2422
2423                 list_for_each_entry(glist, &qg->groups, next_group) {
2424                         ret = ulist_add(fs_info->qgroup_ulist,
2425                                         glist->group->qgroupid,
2426                                         (uintptr_t)glist->group, GFP_ATOMIC);
2427                         if (ret < 0)
2428                                 goto out;
2429                 }
2430         }
2431         ret = 0;
2432         /*
2433          * no limits exceeded, now record the reservation into all qgroups
2434          */
2435         ULIST_ITER_INIT(&uiter);
2436         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
2437                 struct btrfs_qgroup *qg;
2438
2439                 qg = u64_to_ptr(unode->aux);
2440
2441                 qg->reserved += num_bytes;
2442         }
2443
2444 out:
2445         spin_unlock(&fs_info->qgroup_lock);
2446         return ret;
2447 }
2448
2449 void btrfs_qgroup_free(struct btrfs_root *root, u64 num_bytes)
2450 {
2451         struct btrfs_root *quota_root;
2452         struct btrfs_qgroup *qgroup;
2453         struct btrfs_fs_info *fs_info = root->fs_info;
2454         struct ulist_node *unode;
2455         struct ulist_iterator uiter;
2456         u64 ref_root = root->root_key.objectid;
2457         int ret = 0;
2458
2459         if (!is_fstree(ref_root))
2460                 return;
2461
2462         if (num_bytes == 0)
2463                 return;
2464
2465         spin_lock(&fs_info->qgroup_lock);
2466
2467         quota_root = fs_info->quota_root;
2468         if (!quota_root)
2469                 goto out;
2470
2471         qgroup = find_qgroup_rb(fs_info, ref_root);
2472         if (!qgroup)
2473                 goto out;
2474
2475         ulist_reinit(fs_info->qgroup_ulist);
2476         ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
2477                         (uintptr_t)qgroup, GFP_ATOMIC);
2478         if (ret < 0)
2479                 goto out;
2480         ULIST_ITER_INIT(&uiter);
2481         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
2482                 struct btrfs_qgroup *qg;
2483                 struct btrfs_qgroup_list *glist;
2484
2485                 qg = u64_to_ptr(unode->aux);
2486
2487                 qg->reserved -= num_bytes;
2488
2489                 list_for_each_entry(glist, &qg->groups, next_group) {
2490                         ret = ulist_add(fs_info->qgroup_ulist,
2491                                         glist->group->qgroupid,
2492                                         (uintptr_t)glist->group, GFP_ATOMIC);
2493                         if (ret < 0)
2494                                 goto out;
2495                 }
2496         }
2497
2498 out:
2499         spin_unlock(&fs_info->qgroup_lock);
2500 }
2501
2502 void assert_qgroups_uptodate(struct btrfs_trans_handle *trans)
2503 {
2504         if (list_empty(&trans->qgroup_ref_list) && !trans->delayed_ref_elem.seq)
2505                 return;
2506         btrfs_err(trans->root->fs_info,
2507                 "qgroups not uptodate in trans handle %p:  list is%s empty, "
2508                 "seq is %#x.%x",
2509                 trans, list_empty(&trans->qgroup_ref_list) ? "" : " not",
2510                 (u32)(trans->delayed_ref_elem.seq >> 32),
2511                 (u32)trans->delayed_ref_elem.seq);
2512         BUG();
2513 }
2514
2515 /*
2516  * returns < 0 on error, 0 when more leafs are to be scanned.
2517  * returns 1 when done, 2 when done and FLAG_INCONSISTENT was cleared.
2518  */
2519 static int
2520 qgroup_rescan_leaf(struct btrfs_fs_info *fs_info, struct btrfs_path *path,
2521                    struct btrfs_trans_handle *trans, struct ulist *qgroups,
2522                    struct ulist *tmp, struct extent_buffer *scratch_leaf)
2523 {
2524         struct btrfs_key found;
2525         struct ulist *roots = NULL;
2526         struct seq_list tree_mod_seq_elem = {};
2527         u64 num_bytes;
2528         u64 seq;
2529         int new_roots;
2530         int slot;
2531         int ret;
2532
2533         path->leave_spinning = 1;
2534         mutex_lock(&fs_info->qgroup_rescan_lock);
2535         ret = btrfs_search_slot_for_read(fs_info->extent_root,
2536                                          &fs_info->qgroup_rescan_progress,
2537                                          path, 1, 0);
2538
2539         pr_debug("current progress key (%llu %u %llu), search_slot ret %d\n",
2540                  fs_info->qgroup_rescan_progress.objectid,
2541                  fs_info->qgroup_rescan_progress.type,
2542                  fs_info->qgroup_rescan_progress.offset, ret);
2543
2544         if (ret) {
2545                 /*
2546                  * The rescan is about to end, we will not be scanning any
2547                  * further blocks. We cannot unset the RESCAN flag here, because
2548                  * we want to commit the transaction if everything went well.
2549                  * To make the live accounting work in this phase, we set our
2550                  * scan progress pointer such that every real extent objectid
2551                  * will be smaller.
2552                  */
2553                 fs_info->qgroup_rescan_progress.objectid = (u64)-1;
2554                 btrfs_release_path(path);
2555                 mutex_unlock(&fs_info->qgroup_rescan_lock);
2556                 return ret;
2557         }
2558
2559         btrfs_item_key_to_cpu(path->nodes[0], &found,
2560                               btrfs_header_nritems(path->nodes[0]) - 1);
2561         fs_info->qgroup_rescan_progress.objectid = found.objectid + 1;
2562
2563         btrfs_get_tree_mod_seq(fs_info, &tree_mod_seq_elem);
2564         memcpy(scratch_leaf, path->nodes[0], sizeof(*scratch_leaf));
2565         slot = path->slots[0];
2566         btrfs_release_path(path);
2567         mutex_unlock(&fs_info->qgroup_rescan_lock);
2568
2569         for (; slot < btrfs_header_nritems(scratch_leaf); ++slot) {
2570                 btrfs_item_key_to_cpu(scratch_leaf, &found, slot);
2571                 if (found.type != BTRFS_EXTENT_ITEM_KEY &&
2572                     found.type != BTRFS_METADATA_ITEM_KEY)
2573                         continue;
2574                 if (found.type == BTRFS_METADATA_ITEM_KEY)
2575                         num_bytes = fs_info->extent_root->nodesize;
2576                 else
2577                         num_bytes = found.offset;
2578
2579                 ulist_reinit(qgroups);
2580                 ret = btrfs_find_all_roots(NULL, fs_info, found.objectid, 0,
2581                                            &roots);
2582                 if (ret < 0)
2583                         goto out;
2584                 spin_lock(&fs_info->qgroup_lock);
2585                 seq = fs_info->qgroup_seq;
2586                 fs_info->qgroup_seq += roots->nnodes + 1; /* max refcnt */
2587
2588                 new_roots = 0;
2589                 ret = qgroup_calc_old_refcnt(fs_info, 0, tmp, roots, qgroups,
2590                                              seq, &new_roots, 1);
2591                 if (ret < 0) {
2592                         spin_unlock(&fs_info->qgroup_lock);
2593                         ulist_free(roots);
2594                         goto out;
2595                 }
2596
2597                 ret = qgroup_adjust_counters(fs_info, 0, num_bytes, qgroups,
2598                                              seq, 0, new_roots, 1);
2599                 if (ret < 0) {
2600                         spin_unlock(&fs_info->qgroup_lock);
2601                         ulist_free(roots);
2602                         goto out;
2603                 }
2604                 spin_unlock(&fs_info->qgroup_lock);
2605                 ulist_free(roots);
2606         }
2607 out:
2608         btrfs_put_tree_mod_seq(fs_info, &tree_mod_seq_elem);
2609
2610         return ret;
2611 }
2612
2613 static void btrfs_qgroup_rescan_worker(struct btrfs_work *work)
2614 {
2615         struct btrfs_fs_info *fs_info = container_of(work, struct btrfs_fs_info,
2616                                                      qgroup_rescan_work);
2617         struct btrfs_path *path;
2618         struct btrfs_trans_handle *trans = NULL;
2619         struct ulist *tmp = NULL, *qgroups = NULL;
2620         struct extent_buffer *scratch_leaf = NULL;
2621         int err = -ENOMEM;
2622
2623         path = btrfs_alloc_path();
2624         if (!path)
2625                 goto out;
2626         qgroups = ulist_alloc(GFP_NOFS);
2627         if (!qgroups)
2628                 goto out;
2629         tmp = ulist_alloc(GFP_NOFS);
2630         if (!tmp)
2631                 goto out;
2632         scratch_leaf = kmalloc(sizeof(*scratch_leaf), GFP_NOFS);
2633         if (!scratch_leaf)
2634                 goto out;
2635
2636         err = 0;
2637         while (!err) {
2638                 trans = btrfs_start_transaction(fs_info->fs_root, 0);
2639                 if (IS_ERR(trans)) {
2640                         err = PTR_ERR(trans);
2641                         break;
2642                 }
2643                 if (!fs_info->quota_enabled) {
2644                         err = -EINTR;
2645                 } else {
2646                         err = qgroup_rescan_leaf(fs_info, path, trans,
2647                                                  qgroups, tmp, scratch_leaf);
2648                 }
2649                 if (err > 0)
2650                         btrfs_commit_transaction(trans, fs_info->fs_root);
2651                 else
2652                         btrfs_end_transaction(trans, fs_info->fs_root);
2653         }
2654
2655 out:
2656         kfree(scratch_leaf);
2657         ulist_free(qgroups);
2658         ulist_free(tmp);
2659         btrfs_free_path(path);
2660
2661         mutex_lock(&fs_info->qgroup_rescan_lock);
2662         fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
2663
2664         if (err == 2 &&
2665             fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT) {
2666                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2667         } else if (err < 0) {
2668                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2669         }
2670         mutex_unlock(&fs_info->qgroup_rescan_lock);
2671
2672         if (err >= 0) {
2673                 btrfs_info(fs_info, "qgroup scan completed%s",
2674                         err == 2 ? " (inconsistency flag cleared)" : "");
2675         } else {
2676                 btrfs_err(fs_info, "qgroup scan failed with %d", err);
2677         }
2678
2679         complete_all(&fs_info->qgroup_rescan_completion);
2680 }
2681
2682 /*
2683  * Checks that (a) no rescan is running and (b) quota is enabled. Allocates all
2684  * memory required for the rescan context.
2685  */
2686 static int
2687 qgroup_rescan_init(struct btrfs_fs_info *fs_info, u64 progress_objectid,
2688                    int init_flags)
2689 {
2690         int ret = 0;
2691
2692         if (!init_flags &&
2693             (!(fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) ||
2694              !(fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_ON))) {
2695                 ret = -EINVAL;
2696                 goto err;
2697         }
2698
2699         mutex_lock(&fs_info->qgroup_rescan_lock);
2700         spin_lock(&fs_info->qgroup_lock);
2701
2702         if (init_flags) {
2703                 if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN)
2704                         ret = -EINPROGRESS;
2705                 else if (!(fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_ON))
2706                         ret = -EINVAL;
2707
2708                 if (ret) {
2709                         spin_unlock(&fs_info->qgroup_lock);
2710                         mutex_unlock(&fs_info->qgroup_rescan_lock);
2711                         goto err;
2712                 }
2713
2714                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_RESCAN;
2715         }
2716
2717         memset(&fs_info->qgroup_rescan_progress, 0,
2718                 sizeof(fs_info->qgroup_rescan_progress));
2719         fs_info->qgroup_rescan_progress.objectid = progress_objectid;
2720
2721         spin_unlock(&fs_info->qgroup_lock);
2722         mutex_unlock(&fs_info->qgroup_rescan_lock);
2723
2724         init_completion(&fs_info->qgroup_rescan_completion);
2725
2726         memset(&fs_info->qgroup_rescan_work, 0,
2727                sizeof(fs_info->qgroup_rescan_work));
2728         btrfs_init_work(&fs_info->qgroup_rescan_work,
2729                         btrfs_qgroup_rescan_helper,
2730                         btrfs_qgroup_rescan_worker, NULL, NULL);
2731
2732         if (ret) {
2733 err:
2734                 btrfs_info(fs_info, "qgroup_rescan_init failed with %d", ret);
2735                 return ret;
2736         }
2737
2738         return 0;
2739 }
2740
2741 static void
2742 qgroup_rescan_zero_tracking(struct btrfs_fs_info *fs_info)
2743 {
2744         struct rb_node *n;
2745         struct btrfs_qgroup *qgroup;
2746
2747         spin_lock(&fs_info->qgroup_lock);
2748         /* clear all current qgroup tracking information */
2749         for (n = rb_first(&fs_info->qgroup_tree); n; n = rb_next(n)) {
2750                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
2751                 qgroup->rfer = 0;
2752                 qgroup->rfer_cmpr = 0;
2753                 qgroup->excl = 0;
2754                 qgroup->excl_cmpr = 0;
2755         }
2756         spin_unlock(&fs_info->qgroup_lock);
2757 }
2758
2759 int
2760 btrfs_qgroup_rescan(struct btrfs_fs_info *fs_info)
2761 {
2762         int ret = 0;
2763         struct btrfs_trans_handle *trans;
2764
2765         ret = qgroup_rescan_init(fs_info, 0, 1);
2766         if (ret)
2767                 return ret;
2768
2769         /*
2770          * We have set the rescan_progress to 0, which means no more
2771          * delayed refs will be accounted by btrfs_qgroup_account_ref.
2772          * However, btrfs_qgroup_account_ref may be right after its call
2773          * to btrfs_find_all_roots, in which case it would still do the
2774          * accounting.
2775          * To solve this, we're committing the transaction, which will
2776          * ensure we run all delayed refs and only after that, we are
2777          * going to clear all tracking information for a clean start.
2778          */
2779
2780         trans = btrfs_join_transaction(fs_info->fs_root);
2781         if (IS_ERR(trans)) {
2782                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
2783                 return PTR_ERR(trans);
2784         }
2785         ret = btrfs_commit_transaction(trans, fs_info->fs_root);
2786         if (ret) {
2787                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
2788                 return ret;
2789         }
2790
2791         qgroup_rescan_zero_tracking(fs_info);
2792
2793         btrfs_queue_work(fs_info->qgroup_rescan_workers,
2794                          &fs_info->qgroup_rescan_work);
2795
2796         return 0;
2797 }
2798
2799 int btrfs_qgroup_wait_for_completion(struct btrfs_fs_info *fs_info)
2800 {
2801         int running;
2802         int ret = 0;
2803
2804         mutex_lock(&fs_info->qgroup_rescan_lock);
2805         spin_lock(&fs_info->qgroup_lock);
2806         running = fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN;
2807         spin_unlock(&fs_info->qgroup_lock);
2808         mutex_unlock(&fs_info->qgroup_rescan_lock);
2809
2810         if (running)
2811                 ret = wait_for_completion_interruptible(
2812                                         &fs_info->qgroup_rescan_completion);
2813
2814         return ret;
2815 }
2816
2817 /*
2818  * this is only called from open_ctree where we're still single threaded, thus
2819  * locking is omitted here.
2820  */
2821 void
2822 btrfs_qgroup_rescan_resume(struct btrfs_fs_info *fs_info)
2823 {
2824         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN)
2825                 btrfs_queue_work(fs_info->qgroup_rescan_workers,
2826                                  &fs_info->qgroup_rescan_work);
2827 }