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                 WARN_ON(sign < 0 && qgroup->excl < oper->num_bytes);
1435                 qgroup->excl += sign * oper->num_bytes;
1436                 qgroup->excl_cmpr += sign * oper->num_bytes;
1437                 qgroup_dirty(fs_info, qgroup);
1438
1439                 /* Add any parents of the parents */
1440                 list_for_each_entry(glist, &qgroup->groups, next_group) {
1441                         ret = ulist_add(tmp, glist->group->qgroupid,
1442                                         ptr_to_u64(glist->group), GFP_ATOMIC);
1443                         if (ret < 0)
1444                                 goto out;
1445                 }
1446         }
1447         ret = 0;
1448 out:
1449         spin_unlock(&fs_info->qgroup_lock);
1450         ulist_free(tmp);
1451         return ret;
1452 }
1453
1454 /*
1455  * Walk all of the roots that pointed to our bytenr and adjust their refcnts as
1456  * properly.
1457  */
1458 static int qgroup_calc_old_refcnt(struct btrfs_fs_info *fs_info,
1459                                   u64 root_to_skip, struct ulist *tmp,
1460                                   struct ulist *roots, struct ulist *qgroups,
1461                                   u64 seq, int *old_roots, int rescan)
1462 {
1463         struct ulist_node *unode;
1464         struct ulist_iterator uiter;
1465         struct ulist_node *tmp_unode;
1466         struct ulist_iterator tmp_uiter;
1467         struct btrfs_qgroup *qg;
1468         int ret;
1469
1470         ULIST_ITER_INIT(&uiter);
1471         while ((unode = ulist_next(roots, &uiter))) {
1472                 /* We don't count our current root here */
1473                 if (unode->val == root_to_skip)
1474                         continue;
1475                 qg = find_qgroup_rb(fs_info, unode->val);
1476                 if (!qg)
1477                         continue;
1478                 /*
1479                  * We could have a pending removal of this same ref so we may
1480                  * not have actually found our ref root when doing
1481                  * btrfs_find_all_roots, so we need to keep track of how many
1482                  * old roots we find in case we removed ours and added a
1483                  * different one at the same time.  I don't think this could
1484                  * happen in practice but that sort of thinking leads to pain
1485                  * and suffering and to the dark side.
1486                  */
1487                 (*old_roots)++;
1488
1489                 ulist_reinit(tmp);
1490                 ret = ulist_add(qgroups, qg->qgroupid, ptr_to_u64(qg),
1491                                 GFP_ATOMIC);
1492                 if (ret < 0)
1493                         return ret;
1494                 ret = ulist_add(tmp, qg->qgroupid, ptr_to_u64(qg), GFP_ATOMIC);
1495                 if (ret < 0)
1496                         return ret;
1497                 ULIST_ITER_INIT(&tmp_uiter);
1498                 while ((tmp_unode = ulist_next(tmp, &tmp_uiter))) {
1499                         struct btrfs_qgroup_list *glist;
1500
1501                         qg = u64_to_ptr(tmp_unode->aux);
1502                         /*
1503                          * We use this sequence number to keep from having to
1504                          * run the whole list and 0 out the refcnt every time.
1505                          * We basically use sequnce as the known 0 count and
1506                          * then add 1 everytime we see a qgroup.  This is how we
1507                          * get how many of the roots actually point up to the
1508                          * upper level qgroups in order to determine exclusive
1509                          * counts.
1510                          *
1511                          * For rescan we want to set old_refcnt to seq so our
1512                          * exclusive calculations end up correct.
1513                          */
1514                         if (rescan)
1515                                 qg->old_refcnt = seq;
1516                         else if (qg->old_refcnt < seq)
1517                                 qg->old_refcnt = seq + 1;
1518                         else
1519                                 qg->old_refcnt++;
1520
1521                         if (qg->new_refcnt < seq)
1522                                 qg->new_refcnt = seq + 1;
1523                         else
1524                                 qg->new_refcnt++;
1525                         list_for_each_entry(glist, &qg->groups, next_group) {
1526                                 ret = ulist_add(qgroups, glist->group->qgroupid,
1527                                                 ptr_to_u64(glist->group),
1528                                                 GFP_ATOMIC);
1529                                 if (ret < 0)
1530                                         return ret;
1531                                 ret = ulist_add(tmp, glist->group->qgroupid,
1532                                                 ptr_to_u64(glist->group),
1533                                                 GFP_ATOMIC);
1534                                 if (ret < 0)
1535                                         return ret;
1536                         }
1537                 }
1538         }
1539         return 0;
1540 }
1541
1542 /*
1543  * We need to walk forward in our operation tree and account for any roots that
1544  * were deleted after we made this operation.
1545  */
1546 static int qgroup_account_deleted_refs(struct btrfs_fs_info *fs_info,
1547                                        struct btrfs_qgroup_operation *oper,
1548                                        struct ulist *tmp,
1549                                        struct ulist *qgroups, u64 seq,
1550                                        int *old_roots)
1551 {
1552         struct ulist_node *unode;
1553         struct ulist_iterator uiter;
1554         struct btrfs_qgroup *qg;
1555         struct btrfs_qgroup_operation *tmp_oper;
1556         struct rb_node *n;
1557         int ret;
1558
1559         ulist_reinit(tmp);
1560
1561         /*
1562          * We only walk forward in the tree since we're only interested in
1563          * removals that happened _after_  our operation.
1564          */
1565         spin_lock(&fs_info->qgroup_op_lock);
1566         n = rb_next(&oper->n);
1567         spin_unlock(&fs_info->qgroup_op_lock);
1568         if (!n)
1569                 return 0;
1570         tmp_oper = rb_entry(n, struct btrfs_qgroup_operation, n);
1571         while (tmp_oper->bytenr == oper->bytenr) {
1572                 /*
1573                  * If it's not a removal we don't care, additions work out
1574                  * properly with our refcnt tracking.
1575                  */
1576                 if (tmp_oper->type != BTRFS_QGROUP_OPER_SUB_SHARED &&
1577                     tmp_oper->type != BTRFS_QGROUP_OPER_SUB_EXCL)
1578                         goto next;
1579                 qg = find_qgroup_rb(fs_info, tmp_oper->ref_root);
1580                 if (!qg)
1581                         goto next;
1582                 ret = ulist_add(qgroups, qg->qgroupid, ptr_to_u64(qg),
1583                                 GFP_ATOMIC);
1584                 if (ret) {
1585                         if (ret < 0)
1586                                 return ret;
1587                         /*
1588                          * We only want to increase old_roots if this qgroup is
1589                          * not already in the list of qgroups.  If it is already
1590                          * there then that means it must have been re-added or
1591                          * the delete will be discarded because we had an
1592                          * existing ref that we haven't looked up yet.  In this
1593                          * case we don't want to increase old_roots.  So if ret
1594                          * == 1 then we know that this is the first time we've
1595                          * seen this qgroup and we can bump the old_roots.
1596                          */
1597                         (*old_roots)++;
1598                         ret = ulist_add(tmp, qg->qgroupid, ptr_to_u64(qg),
1599                                         GFP_ATOMIC);
1600                         if (ret < 0)
1601                                 return ret;
1602                 }
1603 next:
1604                 spin_lock(&fs_info->qgroup_op_lock);
1605                 n = rb_next(&tmp_oper->n);
1606                 spin_unlock(&fs_info->qgroup_op_lock);
1607                 if (!n)
1608                         break;
1609                 tmp_oper = rb_entry(n, struct btrfs_qgroup_operation, n);
1610         }
1611
1612         /* Ok now process the qgroups we found */
1613         ULIST_ITER_INIT(&uiter);
1614         while ((unode = ulist_next(tmp, &uiter))) {
1615                 struct btrfs_qgroup_list *glist;
1616
1617                 qg = u64_to_ptr(unode->aux);
1618                 if (qg->old_refcnt < seq)
1619                         qg->old_refcnt = seq + 1;
1620                 else
1621                         qg->old_refcnt++;
1622                 if (qg->new_refcnt < seq)
1623                         qg->new_refcnt = seq + 1;
1624                 else
1625                         qg->new_refcnt++;
1626                 list_for_each_entry(glist, &qg->groups, next_group) {
1627                         ret = ulist_add(qgroups, glist->group->qgroupid,
1628                                         ptr_to_u64(glist->group), GFP_ATOMIC);
1629                         if (ret < 0)
1630                                 return ret;
1631                         ret = ulist_add(tmp, glist->group->qgroupid,
1632                                         ptr_to_u64(glist->group), GFP_ATOMIC);
1633                         if (ret < 0)
1634                                 return ret;
1635                 }
1636         }
1637         return 0;
1638 }
1639
1640 /* Add refcnt for the newly added reference. */
1641 static int qgroup_calc_new_refcnt(struct btrfs_fs_info *fs_info,
1642                                   struct btrfs_qgroup_operation *oper,
1643                                   struct btrfs_qgroup *qgroup,
1644                                   struct ulist *tmp, struct ulist *qgroups,
1645                                   u64 seq)
1646 {
1647         struct ulist_node *unode;
1648         struct ulist_iterator uiter;
1649         struct btrfs_qgroup *qg;
1650         int ret;
1651
1652         ulist_reinit(tmp);
1653         ret = ulist_add(qgroups, qgroup->qgroupid, ptr_to_u64(qgroup),
1654                         GFP_ATOMIC);
1655         if (ret < 0)
1656                 return ret;
1657         ret = ulist_add(tmp, qgroup->qgroupid, ptr_to_u64(qgroup),
1658                         GFP_ATOMIC);
1659         if (ret < 0)
1660                 return ret;
1661         ULIST_ITER_INIT(&uiter);
1662         while ((unode = ulist_next(tmp, &uiter))) {
1663                 struct btrfs_qgroup_list *glist;
1664
1665                 qg = u64_to_ptr(unode->aux);
1666                 if (oper->type == BTRFS_QGROUP_OPER_ADD_SHARED) {
1667                         if (qg->new_refcnt < seq)
1668                                 qg->new_refcnt = seq + 1;
1669                         else
1670                                 qg->new_refcnt++;
1671                 } else {
1672                         if (qg->old_refcnt < seq)
1673                                 qg->old_refcnt = seq + 1;
1674                         else
1675                                 qg->old_refcnt++;
1676                 }
1677                 list_for_each_entry(glist, &qg->groups, next_group) {
1678                         ret = ulist_add(tmp, glist->group->qgroupid,
1679                                         ptr_to_u64(glist->group), GFP_ATOMIC);
1680                         if (ret < 0)
1681                                 return ret;
1682                         ret = ulist_add(qgroups, glist->group->qgroupid,
1683                                         ptr_to_u64(glist->group), GFP_ATOMIC);
1684                         if (ret < 0)
1685                                 return ret;
1686                 }
1687         }
1688         return 0;
1689 }
1690
1691 /*
1692  * This adjusts the counters for all referenced qgroups if need be.
1693  */
1694 static int qgroup_adjust_counters(struct btrfs_fs_info *fs_info,
1695                                   u64 root_to_skip, u64 num_bytes,
1696                                   struct ulist *qgroups, u64 seq,
1697                                   int old_roots, int new_roots, int rescan)
1698 {
1699         struct ulist_node *unode;
1700         struct ulist_iterator uiter;
1701         struct btrfs_qgroup *qg;
1702         u64 cur_new_count, cur_old_count;
1703
1704         ULIST_ITER_INIT(&uiter);
1705         while ((unode = ulist_next(qgroups, &uiter))) {
1706                 bool dirty = false;
1707
1708                 qg = u64_to_ptr(unode->aux);
1709                 /*
1710                  * Wasn't referenced before but is now, add to the reference
1711                  * counters.
1712                  */
1713                 if (qg->old_refcnt <= seq && qg->new_refcnt > seq) {
1714                         qg->rfer += num_bytes;
1715                         qg->rfer_cmpr += num_bytes;
1716                         dirty = true;
1717                 }
1718
1719                 /*
1720                  * Was referenced before but isn't now, subtract from the
1721                  * reference counters.
1722                  */
1723                 if (qg->old_refcnt > seq && qg->new_refcnt <= seq) {
1724                         qg->rfer -= num_bytes;
1725                         qg->rfer_cmpr -= num_bytes;
1726                         dirty = true;
1727                 }
1728
1729                 if (qg->old_refcnt < seq)
1730                         cur_old_count = 0;
1731                 else
1732                         cur_old_count = qg->old_refcnt - seq;
1733                 if (qg->new_refcnt < seq)
1734                         cur_new_count = 0;
1735                 else
1736                         cur_new_count = qg->new_refcnt - seq;
1737
1738                 /*
1739                  * If our refcount was the same as the roots previously but our
1740                  * new count isn't the same as the number of roots now then we
1741                  * went from having a exclusive reference on this range to not.
1742                  */
1743                 if (old_roots && cur_old_count == old_roots &&
1744                     (cur_new_count != new_roots || new_roots == 0)) {
1745                         WARN_ON(cur_new_count != new_roots && new_roots == 0);
1746                         qg->excl -= num_bytes;
1747                         qg->excl_cmpr -= num_bytes;
1748                         dirty = true;
1749                 }
1750
1751                 /*
1752                  * If we didn't reference all the roots before but now we do we
1753                  * have an exclusive reference to this range.
1754                  */
1755                 if ((!old_roots || (old_roots && cur_old_count != old_roots))
1756                     && cur_new_count == new_roots) {
1757                         qg->excl += num_bytes;
1758                         qg->excl_cmpr += num_bytes;
1759                         dirty = true;
1760                 }
1761
1762                 if (dirty)
1763                         qgroup_dirty(fs_info, qg);
1764         }
1765         return 0;
1766 }
1767
1768 /*
1769  * If we removed a data extent and there were other references for that bytenr
1770  * then we need to lookup all referenced roots to make sure we still don't
1771  * reference this bytenr.  If we do then we can just discard this operation.
1772  */
1773 static int check_existing_refs(struct btrfs_trans_handle *trans,
1774                                struct btrfs_fs_info *fs_info,
1775                                struct btrfs_qgroup_operation *oper)
1776 {
1777         struct ulist *roots = NULL;
1778         struct ulist_node *unode;
1779         struct ulist_iterator uiter;
1780         int ret = 0;
1781
1782         ret = btrfs_find_all_roots(trans, fs_info, oper->bytenr,
1783                                    oper->elem.seq, &roots);
1784         if (ret < 0)
1785                 return ret;
1786         ret = 0;
1787
1788         ULIST_ITER_INIT(&uiter);
1789         while ((unode = ulist_next(roots, &uiter))) {
1790                 if (unode->val == oper->ref_root) {
1791                         ret = 1;
1792                         break;
1793                 }
1794         }
1795         ulist_free(roots);
1796         btrfs_put_tree_mod_seq(fs_info, &oper->elem);
1797
1798         return ret;
1799 }
1800
1801 /*
1802  * If we share a reference across multiple roots then we may need to adjust
1803  * various qgroups referenced and exclusive counters.  The basic premise is this
1804  *
1805  * 1) We have seq to represent a 0 count.  Instead of looping through all of the
1806  * qgroups and resetting their refcount to 0 we just constantly bump this
1807  * sequence number to act as the base reference count.  This means that if
1808  * anybody is equal to or below this sequence they were never referenced.  We
1809  * jack this sequence up by the number of roots we found each time in order to
1810  * make sure we don't have any overlap.
1811  *
1812  * 2) We first search all the roots that reference the area _except_ the root
1813  * we're acting on currently.  This makes up the old_refcnt of all the qgroups
1814  * before.
1815  *
1816  * 3) We walk all of the qgroups referenced by the root we are currently acting
1817  * on, and will either adjust old_refcnt in the case of a removal or the
1818  * new_refcnt in the case of an addition.
1819  *
1820  * 4) Finally we walk all the qgroups that are referenced by this range
1821  * including the root we are acting on currently.  We will adjust the counters
1822  * based on the number of roots we had and will have after this operation.
1823  *
1824  * Take this example as an illustration
1825  *
1826  *                      [qgroup 1/0]
1827  *                   /         |          \
1828  *              [qg 0/0]   [qg 0/1]     [qg 0/2]
1829  *                 \          |            /
1830  *                [        extent           ]
1831  *
1832  * Say we are adding a reference that is covered by qg 0/0.  The first step
1833  * would give a refcnt of 1 to qg 0/1 and 0/2 and a refcnt of 2 to qg 1/0 with
1834  * old_roots being 2.  Because it is adding new_roots will be 1.  We then go
1835  * through qg 0/0 which will get the new_refcnt set to 1 and add 1 to qg 1/0's
1836  * new_refcnt, bringing it to 3.  We then walk through all of the qgroups, we
1837  * notice that the old refcnt for qg 0/0 < the new refcnt, so we added a
1838  * reference and thus must add the size to the referenced bytes.  Everything
1839  * else is the same so nothing else changes.
1840  */
1841 static int qgroup_shared_accounting(struct btrfs_trans_handle *trans,
1842                                     struct btrfs_fs_info *fs_info,
1843                                     struct btrfs_qgroup_operation *oper)
1844 {
1845         struct ulist *roots = NULL;
1846         struct ulist *qgroups, *tmp;
1847         struct btrfs_qgroup *qgroup;
1848         struct seq_list elem = {};
1849         u64 seq;
1850         int old_roots = 0;
1851         int new_roots = 0;
1852         int ret = 0;
1853
1854         if (oper->elem.seq) {
1855                 ret = check_existing_refs(trans, fs_info, oper);
1856                 if (ret < 0)
1857                         return ret;
1858                 if (ret)
1859                         return 0;
1860         }
1861
1862         qgroups = ulist_alloc(GFP_NOFS);
1863         if (!qgroups)
1864                 return -ENOMEM;
1865
1866         tmp = ulist_alloc(GFP_NOFS);
1867         if (!tmp) {
1868                 ulist_free(qgroups);
1869                 return -ENOMEM;
1870         }
1871
1872         btrfs_get_tree_mod_seq(fs_info, &elem);
1873         ret = btrfs_find_all_roots(trans, fs_info, oper->bytenr, elem.seq,
1874                                    &roots);
1875         btrfs_put_tree_mod_seq(fs_info, &elem);
1876         if (ret < 0) {
1877                 ulist_free(qgroups);
1878                 ulist_free(tmp);
1879                 return ret;
1880         }
1881         spin_lock(&fs_info->qgroup_lock);
1882         qgroup = find_qgroup_rb(fs_info, oper->ref_root);
1883         if (!qgroup)
1884                 goto out;
1885         seq = fs_info->qgroup_seq;
1886
1887         /*
1888          * So roots is the list of all the roots currently pointing at the
1889          * bytenr, including the ref we are adding if we are adding, or not if
1890          * we are removing a ref.  So we pass in the ref_root to skip that root
1891          * in our calculations.  We set old_refnct and new_refcnt cause who the
1892          * hell knows what everything looked like before, and it doesn't matter
1893          * except...
1894          */
1895         ret = qgroup_calc_old_refcnt(fs_info, oper->ref_root, tmp, roots, qgroups,
1896                                      seq, &old_roots, 0);
1897         if (ret < 0)
1898                 goto out;
1899
1900         /*
1901          * Now adjust the refcounts of the qgroups that care about this
1902          * reference, either the old_count in the case of removal or new_count
1903          * in the case of an addition.
1904          */
1905         ret = qgroup_calc_new_refcnt(fs_info, oper, qgroup, tmp, qgroups,
1906                                      seq);
1907         if (ret < 0)
1908                 goto out;
1909
1910         /*
1911          * ...in the case of removals.  If we had a removal before we got around
1912          * to processing this operation then we need to find that guy and count
1913          * his references as if they really existed so we don't end up screwing
1914          * up the exclusive counts.  Then whenever we go to process the delete
1915          * everything will be grand and we can account for whatever exclusive
1916          * changes need to be made there.  We also have to pass in old_roots so
1917          * we have an accurate count of the roots as it pertains to this
1918          * operations view of the world.
1919          */
1920         ret = qgroup_account_deleted_refs(fs_info, oper, tmp, qgroups, seq,
1921                                           &old_roots);
1922         if (ret < 0)
1923                 goto out;
1924
1925         /*
1926          * We are adding our root, need to adjust up the number of roots,
1927          * otherwise old_roots is the number of roots we want.
1928          */
1929         if (oper->type == BTRFS_QGROUP_OPER_ADD_SHARED) {
1930                 new_roots = old_roots + 1;
1931         } else {
1932                 new_roots = old_roots;
1933                 old_roots++;
1934         }
1935         fs_info->qgroup_seq += old_roots + 1;
1936
1937
1938         /*
1939          * And now the magic happens, bless Arne for having a pretty elegant
1940          * solution for this.
1941          */
1942         qgroup_adjust_counters(fs_info, oper->ref_root, oper->num_bytes,
1943                                qgroups, seq, old_roots, new_roots, 0);
1944 out:
1945         spin_unlock(&fs_info->qgroup_lock);
1946         ulist_free(qgroups);
1947         ulist_free(roots);
1948         ulist_free(tmp);
1949         return ret;
1950 }
1951
1952 /*
1953  * Process a reference to a shared subtree. This type of operation is
1954  * queued during snapshot removal when we encounter extents which are
1955  * shared between more than one root.
1956  */
1957 static int qgroup_subtree_accounting(struct btrfs_trans_handle *trans,
1958                                      struct btrfs_fs_info *fs_info,
1959                                      struct btrfs_qgroup_operation *oper)
1960 {
1961         struct ulist *roots = NULL;
1962         struct ulist_node *unode;
1963         struct ulist_iterator uiter;
1964         struct btrfs_qgroup_list *glist;
1965         struct ulist *parents;
1966         int ret = 0;
1967         int err;
1968         struct btrfs_qgroup *qg;
1969         u64 root_obj = 0;
1970         struct seq_list elem = {};
1971
1972         parents = ulist_alloc(GFP_NOFS);
1973         if (!parents)
1974                 return -ENOMEM;
1975
1976         btrfs_get_tree_mod_seq(fs_info, &elem);
1977         ret = btrfs_find_all_roots(trans, fs_info, oper->bytenr,
1978                                    elem.seq, &roots);
1979         btrfs_put_tree_mod_seq(fs_info, &elem);
1980         if (ret < 0)
1981                 goto out;
1982
1983         if (roots->nnodes != 1)
1984                 goto out;
1985
1986         ULIST_ITER_INIT(&uiter);
1987         unode = ulist_next(roots, &uiter); /* Only want 1 so no need to loop */
1988         /*
1989          * If we find our ref root then that means all refs
1990          * this extent has to the root have not yet been
1991          * deleted. In that case, we do nothing and let the
1992          * last ref for this bytenr drive our update.
1993          *
1994          * This can happen for example if an extent is
1995          * referenced multiple times in a snapshot (clone,
1996          * etc). If we are in the middle of snapshot removal,
1997          * queued updates for such an extent will find the
1998          * root if we have not yet finished removing the
1999          * snapshot.
2000          */
2001         if (unode->val == oper->ref_root)
2002                 goto out;
2003
2004         root_obj = unode->val;
2005         BUG_ON(!root_obj);
2006
2007         spin_lock(&fs_info->qgroup_lock);
2008         qg = find_qgroup_rb(fs_info, root_obj);
2009         if (!qg)
2010                 goto out_unlock;
2011
2012         qg->excl += oper->num_bytes;
2013         qg->excl_cmpr += oper->num_bytes;
2014         qgroup_dirty(fs_info, qg);
2015
2016         /*
2017          * Adjust counts for parent groups. First we find all
2018          * parents, then in the 2nd loop we do the adjustment
2019          * while adding parents of the parents to our ulist.
2020          */
2021         list_for_each_entry(glist, &qg->groups, next_group) {
2022                 err = ulist_add(parents, glist->group->qgroupid,
2023                                 ptr_to_u64(glist->group), GFP_ATOMIC);
2024                 if (err < 0) {
2025                         ret = err;
2026                         goto out_unlock;
2027                 }
2028         }
2029
2030         ULIST_ITER_INIT(&uiter);
2031         while ((unode = ulist_next(parents, &uiter))) {
2032                 qg = u64_to_ptr(unode->aux);
2033                 qg->excl += oper->num_bytes;
2034                 qg->excl_cmpr += oper->num_bytes;
2035                 qgroup_dirty(fs_info, qg);
2036
2037                 /* Add any parents of the parents */
2038                 list_for_each_entry(glist, &qg->groups, next_group) {
2039                         err = ulist_add(parents, glist->group->qgroupid,
2040                                         ptr_to_u64(glist->group), GFP_ATOMIC);
2041                         if (err < 0) {
2042                                 ret = err;
2043                                 goto out_unlock;
2044                         }
2045                 }
2046         }
2047
2048 out_unlock:
2049         spin_unlock(&fs_info->qgroup_lock);
2050
2051 out:
2052         ulist_free(roots);
2053         ulist_free(parents);
2054         return ret;
2055 }
2056
2057 /*
2058  * btrfs_qgroup_account_ref is called for every ref that is added to or deleted
2059  * from the fs. First, all roots referencing the extent are searched, and
2060  * then the space is accounted accordingly to the different roots. The
2061  * accounting algorithm works in 3 steps documented inline.
2062  */
2063 static int btrfs_qgroup_account(struct btrfs_trans_handle *trans,
2064                                 struct btrfs_fs_info *fs_info,
2065                                 struct btrfs_qgroup_operation *oper)
2066 {
2067         int ret = 0;
2068
2069         if (!fs_info->quota_enabled)
2070                 return 0;
2071
2072         BUG_ON(!fs_info->quota_root);
2073
2074         mutex_lock(&fs_info->qgroup_rescan_lock);
2075         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
2076                 if (fs_info->qgroup_rescan_progress.objectid <= oper->bytenr) {
2077                         mutex_unlock(&fs_info->qgroup_rescan_lock);
2078                         return 0;
2079                 }
2080         }
2081         mutex_unlock(&fs_info->qgroup_rescan_lock);
2082
2083         ASSERT(is_fstree(oper->ref_root));
2084
2085         trace_btrfs_qgroup_account(oper);
2086
2087         switch (oper->type) {
2088         case BTRFS_QGROUP_OPER_ADD_EXCL:
2089         case BTRFS_QGROUP_OPER_SUB_EXCL:
2090                 ret = qgroup_excl_accounting(fs_info, oper);
2091                 break;
2092         case BTRFS_QGROUP_OPER_ADD_SHARED:
2093         case BTRFS_QGROUP_OPER_SUB_SHARED:
2094                 ret = qgroup_shared_accounting(trans, fs_info, oper);
2095                 break;
2096         case BTRFS_QGROUP_OPER_SUB_SUBTREE:
2097                 ret = qgroup_subtree_accounting(trans, fs_info, oper);
2098                 break;
2099         default:
2100                 ASSERT(0);
2101         }
2102         return ret;
2103 }
2104
2105 /*
2106  * Needs to be called everytime we run delayed refs, even if there is an error
2107  * in order to cleanup outstanding operations.
2108  */
2109 int btrfs_delayed_qgroup_accounting(struct btrfs_trans_handle *trans,
2110                                     struct btrfs_fs_info *fs_info)
2111 {
2112         struct btrfs_qgroup_operation *oper;
2113         int ret = 0;
2114
2115         while (!list_empty(&trans->qgroup_ref_list)) {
2116                 oper = list_first_entry(&trans->qgroup_ref_list,
2117                                         struct btrfs_qgroup_operation, list);
2118                 list_del_init(&oper->list);
2119                 if (!ret || !trans->aborted)
2120                         ret = btrfs_qgroup_account(trans, fs_info, oper);
2121                 spin_lock(&fs_info->qgroup_op_lock);
2122                 rb_erase(&oper->n, &fs_info->qgroup_op_tree);
2123                 spin_unlock(&fs_info->qgroup_op_lock);
2124                 btrfs_put_tree_mod_seq(fs_info, &oper->elem);
2125                 kfree(oper);
2126         }
2127         return ret;
2128 }
2129
2130 /*
2131  * called from commit_transaction. Writes all changed qgroups to disk.
2132  */
2133 int btrfs_run_qgroups(struct btrfs_trans_handle *trans,
2134                       struct btrfs_fs_info *fs_info)
2135 {
2136         struct btrfs_root *quota_root = fs_info->quota_root;
2137         int ret = 0;
2138         int start_rescan_worker = 0;
2139
2140         if (!quota_root)
2141                 goto out;
2142
2143         if (!fs_info->quota_enabled && fs_info->pending_quota_state)
2144                 start_rescan_worker = 1;
2145
2146         fs_info->quota_enabled = fs_info->pending_quota_state;
2147
2148         spin_lock(&fs_info->qgroup_lock);
2149         while (!list_empty(&fs_info->dirty_qgroups)) {
2150                 struct btrfs_qgroup *qgroup;
2151                 qgroup = list_first_entry(&fs_info->dirty_qgroups,
2152                                           struct btrfs_qgroup, dirty);
2153                 list_del_init(&qgroup->dirty);
2154                 spin_unlock(&fs_info->qgroup_lock);
2155                 ret = update_qgroup_info_item(trans, quota_root, qgroup);
2156                 if (ret)
2157                         fs_info->qgroup_flags |=
2158                                         BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2159                 spin_lock(&fs_info->qgroup_lock);
2160         }
2161         if (fs_info->quota_enabled)
2162                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_ON;
2163         else
2164                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_ON;
2165         spin_unlock(&fs_info->qgroup_lock);
2166
2167         ret = update_qgroup_status_item(trans, fs_info, quota_root);
2168         if (ret)
2169                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2170
2171         if (!ret && start_rescan_worker) {
2172                 ret = qgroup_rescan_init(fs_info, 0, 1);
2173                 if (!ret) {
2174                         qgroup_rescan_zero_tracking(fs_info);
2175                         btrfs_queue_work(fs_info->qgroup_rescan_workers,
2176                                          &fs_info->qgroup_rescan_work);
2177                 }
2178                 ret = 0;
2179         }
2180
2181 out:
2182
2183         return ret;
2184 }
2185
2186 /*
2187  * copy the acounting information between qgroups. This is necessary when a
2188  * snapshot or a subvolume is created
2189  */
2190 int btrfs_qgroup_inherit(struct btrfs_trans_handle *trans,
2191                          struct btrfs_fs_info *fs_info, u64 srcid, u64 objectid,
2192                          struct btrfs_qgroup_inherit *inherit)
2193 {
2194         int ret = 0;
2195         int i;
2196         u64 *i_qgroups;
2197         struct btrfs_root *quota_root = fs_info->quota_root;
2198         struct btrfs_qgroup *srcgroup;
2199         struct btrfs_qgroup *dstgroup;
2200         u32 level_size = 0;
2201         u64 nums;
2202
2203         mutex_lock(&fs_info->qgroup_ioctl_lock);
2204         if (!fs_info->quota_enabled)
2205                 goto out;
2206
2207         if (!quota_root) {
2208                 ret = -EINVAL;
2209                 goto out;
2210         }
2211
2212         if (inherit) {
2213                 i_qgroups = (u64 *)(inherit + 1);
2214                 nums = inherit->num_qgroups + 2 * inherit->num_ref_copies +
2215                        2 * inherit->num_excl_copies;
2216                 for (i = 0; i < nums; ++i) {
2217                         srcgroup = find_qgroup_rb(fs_info, *i_qgroups);
2218                         if (!srcgroup) {
2219                                 ret = -EINVAL;
2220                                 goto out;
2221                         }
2222                         ++i_qgroups;
2223                 }
2224         }
2225
2226         /*
2227          * create a tracking group for the subvol itself
2228          */
2229         ret = add_qgroup_item(trans, quota_root, objectid);
2230         if (ret)
2231                 goto out;
2232
2233         if (inherit && inherit->flags & BTRFS_QGROUP_INHERIT_SET_LIMITS) {
2234                 ret = update_qgroup_limit_item(trans, quota_root, objectid,
2235                                                inherit->lim.flags,
2236                                                inherit->lim.max_rfer,
2237                                                inherit->lim.max_excl,
2238                                                inherit->lim.rsv_rfer,
2239                                                inherit->lim.rsv_excl);
2240                 if (ret)
2241                         goto out;
2242         }
2243
2244         if (srcid) {
2245                 struct btrfs_root *srcroot;
2246                 struct btrfs_key srckey;
2247
2248                 srckey.objectid = srcid;
2249                 srckey.type = BTRFS_ROOT_ITEM_KEY;
2250                 srckey.offset = (u64)-1;
2251                 srcroot = btrfs_read_fs_root_no_name(fs_info, &srckey);
2252                 if (IS_ERR(srcroot)) {
2253                         ret = PTR_ERR(srcroot);
2254                         goto out;
2255                 }
2256
2257                 rcu_read_lock();
2258                 level_size = srcroot->nodesize;
2259                 rcu_read_unlock();
2260         }
2261
2262         /*
2263          * add qgroup to all inherited groups
2264          */
2265         if (inherit) {
2266                 i_qgroups = (u64 *)(inherit + 1);
2267                 for (i = 0; i < inherit->num_qgroups; ++i) {
2268                         ret = add_qgroup_relation_item(trans, quota_root,
2269                                                        objectid, *i_qgroups);
2270                         if (ret)
2271                                 goto out;
2272                         ret = add_qgroup_relation_item(trans, quota_root,
2273                                                        *i_qgroups, objectid);
2274                         if (ret)
2275                                 goto out;
2276                         ++i_qgroups;
2277                 }
2278         }
2279
2280
2281         spin_lock(&fs_info->qgroup_lock);
2282
2283         dstgroup = add_qgroup_rb(fs_info, objectid);
2284         if (IS_ERR(dstgroup)) {
2285                 ret = PTR_ERR(dstgroup);
2286                 goto unlock;
2287         }
2288
2289         if (srcid) {
2290                 srcgroup = find_qgroup_rb(fs_info, srcid);
2291                 if (!srcgroup)
2292                         goto unlock;
2293
2294                 /*
2295                  * We call inherit after we clone the root in order to make sure
2296                  * our counts don't go crazy, so at this point the only
2297                  * difference between the two roots should be the root node.
2298                  */
2299                 dstgroup->rfer = srcgroup->rfer;
2300                 dstgroup->rfer_cmpr = srcgroup->rfer_cmpr;
2301                 dstgroup->excl = level_size;
2302                 dstgroup->excl_cmpr = level_size;
2303                 srcgroup->excl = level_size;
2304                 srcgroup->excl_cmpr = level_size;
2305                 qgroup_dirty(fs_info, dstgroup);
2306                 qgroup_dirty(fs_info, srcgroup);
2307         }
2308
2309         if (!inherit)
2310                 goto unlock;
2311
2312         i_qgroups = (u64 *)(inherit + 1);
2313         for (i = 0; i < inherit->num_qgroups; ++i) {
2314                 ret = add_relation_rb(quota_root->fs_info, objectid,
2315                                       *i_qgroups);
2316                 if (ret)
2317                         goto unlock;
2318                 ++i_qgroups;
2319         }
2320
2321         for (i = 0; i <  inherit->num_ref_copies; ++i) {
2322                 struct btrfs_qgroup *src;
2323                 struct btrfs_qgroup *dst;
2324
2325                 src = find_qgroup_rb(fs_info, i_qgroups[0]);
2326                 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
2327
2328                 if (!src || !dst) {
2329                         ret = -EINVAL;
2330                         goto unlock;
2331                 }
2332
2333                 dst->rfer = src->rfer - level_size;
2334                 dst->rfer_cmpr = src->rfer_cmpr - level_size;
2335                 i_qgroups += 2;
2336         }
2337         for (i = 0; i <  inherit->num_excl_copies; ++i) {
2338                 struct btrfs_qgroup *src;
2339                 struct btrfs_qgroup *dst;
2340
2341                 src = find_qgroup_rb(fs_info, i_qgroups[0]);
2342                 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
2343
2344                 if (!src || !dst) {
2345                         ret = -EINVAL;
2346                         goto unlock;
2347                 }
2348
2349                 dst->excl = src->excl + level_size;
2350                 dst->excl_cmpr = src->excl_cmpr + level_size;
2351                 i_qgroups += 2;
2352         }
2353
2354 unlock:
2355         spin_unlock(&fs_info->qgroup_lock);
2356 out:
2357         mutex_unlock(&fs_info->qgroup_ioctl_lock);
2358         return ret;
2359 }
2360
2361 /*
2362  * reserve some space for a qgroup and all its parents. The reservation takes
2363  * place with start_transaction or dealloc_reserve, similar to ENOSPC
2364  * accounting. If not enough space is available, EDQUOT is returned.
2365  * We assume that the requested space is new for all qgroups.
2366  */
2367 int btrfs_qgroup_reserve(struct btrfs_root *root, u64 num_bytes)
2368 {
2369         struct btrfs_root *quota_root;
2370         struct btrfs_qgroup *qgroup;
2371         struct btrfs_fs_info *fs_info = root->fs_info;
2372         u64 ref_root = root->root_key.objectid;
2373         int ret = 0;
2374         struct ulist_node *unode;
2375         struct ulist_iterator uiter;
2376
2377         if (!is_fstree(ref_root))
2378                 return 0;
2379
2380         if (num_bytes == 0)
2381                 return 0;
2382
2383         spin_lock(&fs_info->qgroup_lock);
2384         quota_root = fs_info->quota_root;
2385         if (!quota_root)
2386                 goto out;
2387
2388         qgroup = find_qgroup_rb(fs_info, ref_root);
2389         if (!qgroup)
2390                 goto out;
2391
2392         /*
2393          * in a first step, we check all affected qgroups if any limits would
2394          * be exceeded
2395          */
2396         ulist_reinit(fs_info->qgroup_ulist);
2397         ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
2398                         (uintptr_t)qgroup, GFP_ATOMIC);
2399         if (ret < 0)
2400                 goto out;
2401         ULIST_ITER_INIT(&uiter);
2402         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
2403                 struct btrfs_qgroup *qg;
2404                 struct btrfs_qgroup_list *glist;
2405
2406                 qg = u64_to_ptr(unode->aux);
2407
2408                 if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_RFER) &&
2409                     qg->reserved + (s64)qg->rfer + num_bytes >
2410                     qg->max_rfer) {
2411                         ret = -EDQUOT;
2412                         goto out;
2413                 }
2414
2415                 if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) &&
2416                     qg->reserved + (s64)qg->excl + num_bytes >
2417                     qg->max_excl) {
2418                         ret = -EDQUOT;
2419                         goto out;
2420                 }
2421
2422                 list_for_each_entry(glist, &qg->groups, next_group) {
2423                         ret = ulist_add(fs_info->qgroup_ulist,
2424                                         glist->group->qgroupid,
2425                                         (uintptr_t)glist->group, GFP_ATOMIC);
2426                         if (ret < 0)
2427                                 goto out;
2428                 }
2429         }
2430         ret = 0;
2431         /*
2432          * no limits exceeded, now record the reservation into all qgroups
2433          */
2434         ULIST_ITER_INIT(&uiter);
2435         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
2436                 struct btrfs_qgroup *qg;
2437
2438                 qg = u64_to_ptr(unode->aux);
2439
2440                 qg->reserved += num_bytes;
2441         }
2442
2443 out:
2444         spin_unlock(&fs_info->qgroup_lock);
2445         return ret;
2446 }
2447
2448 void btrfs_qgroup_free(struct btrfs_root *root, u64 num_bytes)
2449 {
2450         struct btrfs_root *quota_root;
2451         struct btrfs_qgroup *qgroup;
2452         struct btrfs_fs_info *fs_info = root->fs_info;
2453         struct ulist_node *unode;
2454         struct ulist_iterator uiter;
2455         u64 ref_root = root->root_key.objectid;
2456         int ret = 0;
2457
2458         if (!is_fstree(ref_root))
2459                 return;
2460
2461         if (num_bytes == 0)
2462                 return;
2463
2464         spin_lock(&fs_info->qgroup_lock);
2465
2466         quota_root = fs_info->quota_root;
2467         if (!quota_root)
2468                 goto out;
2469
2470         qgroup = find_qgroup_rb(fs_info, ref_root);
2471         if (!qgroup)
2472                 goto out;
2473
2474         ulist_reinit(fs_info->qgroup_ulist);
2475         ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
2476                         (uintptr_t)qgroup, GFP_ATOMIC);
2477         if (ret < 0)
2478                 goto out;
2479         ULIST_ITER_INIT(&uiter);
2480         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
2481                 struct btrfs_qgroup *qg;
2482                 struct btrfs_qgroup_list *glist;
2483
2484                 qg = u64_to_ptr(unode->aux);
2485
2486                 qg->reserved -= num_bytes;
2487
2488                 list_for_each_entry(glist, &qg->groups, next_group) {
2489                         ret = ulist_add(fs_info->qgroup_ulist,
2490                                         glist->group->qgroupid,
2491                                         (uintptr_t)glist->group, GFP_ATOMIC);
2492                         if (ret < 0)
2493                                 goto out;
2494                 }
2495         }
2496
2497 out:
2498         spin_unlock(&fs_info->qgroup_lock);
2499 }
2500
2501 void assert_qgroups_uptodate(struct btrfs_trans_handle *trans)
2502 {
2503         if (list_empty(&trans->qgroup_ref_list) && !trans->delayed_ref_elem.seq)
2504                 return;
2505         btrfs_err(trans->root->fs_info,
2506                 "qgroups not uptodate in trans handle %p:  list is%s empty, "
2507                 "seq is %#x.%x",
2508                 trans, list_empty(&trans->qgroup_ref_list) ? "" : " not",
2509                 (u32)(trans->delayed_ref_elem.seq >> 32),
2510                 (u32)trans->delayed_ref_elem.seq);
2511         BUG();
2512 }
2513
2514 /*
2515  * returns < 0 on error, 0 when more leafs are to be scanned.
2516  * returns 1 when done, 2 when done and FLAG_INCONSISTENT was cleared.
2517  */
2518 static int
2519 qgroup_rescan_leaf(struct btrfs_fs_info *fs_info, struct btrfs_path *path,
2520                    struct btrfs_trans_handle *trans, struct ulist *qgroups,
2521                    struct ulist *tmp, struct extent_buffer *scratch_leaf)
2522 {
2523         struct btrfs_key found;
2524         struct ulist *roots = NULL;
2525         struct seq_list tree_mod_seq_elem = {};
2526         u64 num_bytes;
2527         u64 seq;
2528         int new_roots;
2529         int slot;
2530         int ret;
2531
2532         path->leave_spinning = 1;
2533         mutex_lock(&fs_info->qgroup_rescan_lock);
2534         ret = btrfs_search_slot_for_read(fs_info->extent_root,
2535                                          &fs_info->qgroup_rescan_progress,
2536                                          path, 1, 0);
2537
2538         pr_debug("current progress key (%llu %u %llu), search_slot ret %d\n",
2539                  fs_info->qgroup_rescan_progress.objectid,
2540                  fs_info->qgroup_rescan_progress.type,
2541                  fs_info->qgroup_rescan_progress.offset, ret);
2542
2543         if (ret) {
2544                 /*
2545                  * The rescan is about to end, we will not be scanning any
2546                  * further blocks. We cannot unset the RESCAN flag here, because
2547                  * we want to commit the transaction if everything went well.
2548                  * To make the live accounting work in this phase, we set our
2549                  * scan progress pointer such that every real extent objectid
2550                  * will be smaller.
2551                  */
2552                 fs_info->qgroup_rescan_progress.objectid = (u64)-1;
2553                 btrfs_release_path(path);
2554                 mutex_unlock(&fs_info->qgroup_rescan_lock);
2555                 return ret;
2556         }
2557
2558         btrfs_item_key_to_cpu(path->nodes[0], &found,
2559                               btrfs_header_nritems(path->nodes[0]) - 1);
2560         fs_info->qgroup_rescan_progress.objectid = found.objectid + 1;
2561
2562         btrfs_get_tree_mod_seq(fs_info, &tree_mod_seq_elem);
2563         memcpy(scratch_leaf, path->nodes[0], sizeof(*scratch_leaf));
2564         slot = path->slots[0];
2565         btrfs_release_path(path);
2566         mutex_unlock(&fs_info->qgroup_rescan_lock);
2567
2568         for (; slot < btrfs_header_nritems(scratch_leaf); ++slot) {
2569                 btrfs_item_key_to_cpu(scratch_leaf, &found, slot);
2570                 if (found.type != BTRFS_EXTENT_ITEM_KEY &&
2571                     found.type != BTRFS_METADATA_ITEM_KEY)
2572                         continue;
2573                 if (found.type == BTRFS_METADATA_ITEM_KEY)
2574                         num_bytes = fs_info->extent_root->nodesize;
2575                 else
2576                         num_bytes = found.offset;
2577
2578                 ulist_reinit(qgroups);
2579                 ret = btrfs_find_all_roots(NULL, fs_info, found.objectid, 0,
2580                                            &roots);
2581                 if (ret < 0)
2582                         goto out;
2583                 spin_lock(&fs_info->qgroup_lock);
2584                 seq = fs_info->qgroup_seq;
2585                 fs_info->qgroup_seq += roots->nnodes + 1; /* max refcnt */
2586
2587                 new_roots = 0;
2588                 ret = qgroup_calc_old_refcnt(fs_info, 0, tmp, roots, qgroups,
2589                                              seq, &new_roots, 1);
2590                 if (ret < 0) {
2591                         spin_unlock(&fs_info->qgroup_lock);
2592                         ulist_free(roots);
2593                         goto out;
2594                 }
2595
2596                 ret = qgroup_adjust_counters(fs_info, 0, num_bytes, qgroups,
2597                                              seq, 0, new_roots, 1);
2598                 if (ret < 0) {
2599                         spin_unlock(&fs_info->qgroup_lock);
2600                         ulist_free(roots);
2601                         goto out;
2602                 }
2603                 spin_unlock(&fs_info->qgroup_lock);
2604                 ulist_free(roots);
2605         }
2606 out:
2607         btrfs_put_tree_mod_seq(fs_info, &tree_mod_seq_elem);
2608
2609         return ret;
2610 }
2611
2612 static void btrfs_qgroup_rescan_worker(struct btrfs_work *work)
2613 {
2614         struct btrfs_fs_info *fs_info = container_of(work, struct btrfs_fs_info,
2615                                                      qgroup_rescan_work);
2616         struct btrfs_path *path;
2617         struct btrfs_trans_handle *trans = NULL;
2618         struct ulist *tmp = NULL, *qgroups = NULL;
2619         struct extent_buffer *scratch_leaf = NULL;
2620         int err = -ENOMEM;
2621
2622         path = btrfs_alloc_path();
2623         if (!path)
2624                 goto out;
2625         qgroups = ulist_alloc(GFP_NOFS);
2626         if (!qgroups)
2627                 goto out;
2628         tmp = ulist_alloc(GFP_NOFS);
2629         if (!tmp)
2630                 goto out;
2631         scratch_leaf = kmalloc(sizeof(*scratch_leaf), GFP_NOFS);
2632         if (!scratch_leaf)
2633                 goto out;
2634
2635         err = 0;
2636         while (!err) {
2637                 trans = btrfs_start_transaction(fs_info->fs_root, 0);
2638                 if (IS_ERR(trans)) {
2639                         err = PTR_ERR(trans);
2640                         break;
2641                 }
2642                 if (!fs_info->quota_enabled) {
2643                         err = -EINTR;
2644                 } else {
2645                         err = qgroup_rescan_leaf(fs_info, path, trans,
2646                                                  qgroups, tmp, scratch_leaf);
2647                 }
2648                 if (err > 0)
2649                         btrfs_commit_transaction(trans, fs_info->fs_root);
2650                 else
2651                         btrfs_end_transaction(trans, fs_info->fs_root);
2652         }
2653
2654 out:
2655         kfree(scratch_leaf);
2656         ulist_free(qgroups);
2657         ulist_free(tmp);
2658         btrfs_free_path(path);
2659
2660         mutex_lock(&fs_info->qgroup_rescan_lock);
2661         fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
2662
2663         if (err == 2 &&
2664             fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT) {
2665                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2666         } else if (err < 0) {
2667                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2668         }
2669         mutex_unlock(&fs_info->qgroup_rescan_lock);
2670
2671         if (err >= 0) {
2672                 btrfs_info(fs_info, "qgroup scan completed%s",
2673                         err == 2 ? " (inconsistency flag cleared)" : "");
2674         } else {
2675                 btrfs_err(fs_info, "qgroup scan failed with %d", err);
2676         }
2677
2678         complete_all(&fs_info->qgroup_rescan_completion);
2679 }
2680
2681 /*
2682  * Checks that (a) no rescan is running and (b) quota is enabled. Allocates all
2683  * memory required for the rescan context.
2684  */
2685 static int
2686 qgroup_rescan_init(struct btrfs_fs_info *fs_info, u64 progress_objectid,
2687                    int init_flags)
2688 {
2689         int ret = 0;
2690
2691         if (!init_flags &&
2692             (!(fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) ||
2693              !(fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_ON))) {
2694                 ret = -EINVAL;
2695                 goto err;
2696         }
2697
2698         mutex_lock(&fs_info->qgroup_rescan_lock);
2699         spin_lock(&fs_info->qgroup_lock);
2700
2701         if (init_flags) {
2702                 if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN)
2703                         ret = -EINPROGRESS;
2704                 else if (!(fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_ON))
2705                         ret = -EINVAL;
2706
2707                 if (ret) {
2708                         spin_unlock(&fs_info->qgroup_lock);
2709                         mutex_unlock(&fs_info->qgroup_rescan_lock);
2710                         goto err;
2711                 }
2712
2713                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_RESCAN;
2714         }
2715
2716         memset(&fs_info->qgroup_rescan_progress, 0,
2717                 sizeof(fs_info->qgroup_rescan_progress));
2718         fs_info->qgroup_rescan_progress.objectid = progress_objectid;
2719
2720         spin_unlock(&fs_info->qgroup_lock);
2721         mutex_unlock(&fs_info->qgroup_rescan_lock);
2722
2723         init_completion(&fs_info->qgroup_rescan_completion);
2724
2725         memset(&fs_info->qgroup_rescan_work, 0,
2726                sizeof(fs_info->qgroup_rescan_work));
2727         btrfs_init_work(&fs_info->qgroup_rescan_work,
2728                         btrfs_qgroup_rescan_helper,
2729                         btrfs_qgroup_rescan_worker, NULL, NULL);
2730
2731         if (ret) {
2732 err:
2733                 btrfs_info(fs_info, "qgroup_rescan_init failed with %d", ret);
2734                 return ret;
2735         }
2736
2737         return 0;
2738 }
2739
2740 static void
2741 qgroup_rescan_zero_tracking(struct btrfs_fs_info *fs_info)
2742 {
2743         struct rb_node *n;
2744         struct btrfs_qgroup *qgroup;
2745
2746         spin_lock(&fs_info->qgroup_lock);
2747         /* clear all current qgroup tracking information */
2748         for (n = rb_first(&fs_info->qgroup_tree); n; n = rb_next(n)) {
2749                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
2750                 qgroup->rfer = 0;
2751                 qgroup->rfer_cmpr = 0;
2752                 qgroup->excl = 0;
2753                 qgroup->excl_cmpr = 0;
2754         }
2755         spin_unlock(&fs_info->qgroup_lock);
2756 }
2757
2758 int
2759 btrfs_qgroup_rescan(struct btrfs_fs_info *fs_info)
2760 {
2761         int ret = 0;
2762         struct btrfs_trans_handle *trans;
2763
2764         ret = qgroup_rescan_init(fs_info, 0, 1);
2765         if (ret)
2766                 return ret;
2767
2768         /*
2769          * We have set the rescan_progress to 0, which means no more
2770          * delayed refs will be accounted by btrfs_qgroup_account_ref.
2771          * However, btrfs_qgroup_account_ref may be right after its call
2772          * to btrfs_find_all_roots, in which case it would still do the
2773          * accounting.
2774          * To solve this, we're committing the transaction, which will
2775          * ensure we run all delayed refs and only after that, we are
2776          * going to clear all tracking information for a clean start.
2777          */
2778
2779         trans = btrfs_join_transaction(fs_info->fs_root);
2780         if (IS_ERR(trans)) {
2781                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
2782                 return PTR_ERR(trans);
2783         }
2784         ret = btrfs_commit_transaction(trans, fs_info->fs_root);
2785         if (ret) {
2786                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
2787                 return ret;
2788         }
2789
2790         qgroup_rescan_zero_tracking(fs_info);
2791
2792         btrfs_queue_work(fs_info->qgroup_rescan_workers,
2793                          &fs_info->qgroup_rescan_work);
2794
2795         return 0;
2796 }
2797
2798 int btrfs_qgroup_wait_for_completion(struct btrfs_fs_info *fs_info)
2799 {
2800         int running;
2801         int ret = 0;
2802
2803         mutex_lock(&fs_info->qgroup_rescan_lock);
2804         spin_lock(&fs_info->qgroup_lock);
2805         running = fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN;
2806         spin_unlock(&fs_info->qgroup_lock);
2807         mutex_unlock(&fs_info->qgroup_rescan_lock);
2808
2809         if (running)
2810                 ret = wait_for_completion_interruptible(
2811                                         &fs_info->qgroup_rescan_completion);
2812
2813         return ret;
2814 }
2815
2816 /*
2817  * this is only called from open_ctree where we're still single threaded, thus
2818  * locking is omitted here.
2819  */
2820 void
2821 btrfs_qgroup_rescan_resume(struct btrfs_fs_info *fs_info)
2822 {
2823         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN)
2824                 btrfs_queue_work(fs_info->qgroup_rescan_workers,
2825                                  &fs_info->qgroup_rescan_work);
2826 }