fs/btrfs/qgroup.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2011 STRATO.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/writeback.h>
   9 #include <linux/blkdev.h>
  10 #include <linux/rbtree.h>
  11 #include <linux/slab.h>
  12 #include <linux/workqueue.h>
  13 #include <linux/btrfs.h>
  14 #include <linux/sched/mm.h>
  15
  16 #include "ctree.h"
  17 #include "transaction.h"
  18 #include "disk-io.h"
  19 #include "locking.h"
  20 #include "ulist.h"
  21 #include "backref.h"
  22 #include "extent_io.h"
  23 #include "qgroup.h"
  24 #include "block-group.h"
  25 #include "sysfs.h"
  26 #include "tree-mod-log.h"
  27 #include "fs.h"
  28 #include "accessors.h"
  29 #include "extent-tree.h"
  30 #include "root-tree.h"
  31 #include "tree-checker.h"
  32
  33 enum btrfs_qgroup_mode btrfs_qgroup_mode(struct btrfs_fs_info *fs_info)
  34 {
  35         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
  36                 return BTRFS_QGROUP_MODE_DISABLED;
  37         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_SIMPLE_MODE)
  38                 return BTRFS_QGROUP_MODE_SIMPLE;
  39         return BTRFS_QGROUP_MODE_FULL;
  40 }
  41
  42 bool btrfs_qgroup_enabled(struct btrfs_fs_info *fs_info)
  43 {
  44         return btrfs_qgroup_mode(fs_info) != BTRFS_QGROUP_MODE_DISABLED;
  45 }
  46
  47 bool btrfs_qgroup_full_accounting(struct btrfs_fs_info *fs_info)
  48 {
  49         return btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_FULL;
  50 }
  51
  52 /*
  53  * Helpers to access qgroup reservation
  54  *
  55  * Callers should ensure the lock context and type are valid
  56  */
  57
  58 static u64 qgroup_rsv_total(const struct btrfs_qgroup *qgroup)
  59 {
  60         u64 ret = 0;
  61         int i;
  62
  63         for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++)
  64                 ret += qgroup->rsv.values[i];
  65
  66         return ret;
  67 }
  68
  69 #ifdef CONFIG_BTRFS_DEBUG
  70 static const char *qgroup_rsv_type_str(enum btrfs_qgroup_rsv_type type)
  71 {
  72         if (type == BTRFS_QGROUP_RSV_DATA)
  73                 return "data";
  74         if (type == BTRFS_QGROUP_RSV_META_PERTRANS)
  75                 return "meta_pertrans";
  76         if (type == BTRFS_QGROUP_RSV_META_PREALLOC)
  77                 return "meta_prealloc";
  78         return NULL;
  79 }
  80 #endif
  81
  82 static void qgroup_rsv_add(struct btrfs_fs_info *fs_info,
  83                            struct btrfs_qgroup *qgroup, u64 num_bytes,
  84                            enum btrfs_qgroup_rsv_type type)
  85 {
  86         trace_qgroup_update_reserve(fs_info, qgroup, num_bytes, type);
  87         qgroup->rsv.values[type] += num_bytes;
  88 }
  89
  90 static void qgroup_rsv_release(struct btrfs_fs_info *fs_info,
  91                                struct btrfs_qgroup *qgroup, u64 num_bytes,
  92                                enum btrfs_qgroup_rsv_type type)
  93 {
  94         trace_qgroup_update_reserve(fs_info, qgroup, -(s64)num_bytes, type);
  95         if (qgroup->rsv.values[type] >= num_bytes) {
  96                 qgroup->rsv.values[type] -= num_bytes;
  97                 return;
  98         }
  99 #ifdef CONFIG_BTRFS_DEBUG
 100         WARN_RATELIMIT(1,
 101                 "qgroup %llu %s reserved space underflow, have %llu to free %llu",
 102                 qgroup->qgroupid, qgroup_rsv_type_str(type),
 103                 qgroup->rsv.values[type], num_bytes);
 104 #endif
 105         qgroup->rsv.values[type] = 0;
 106 }
 107
 108 static void qgroup_rsv_add_by_qgroup(struct btrfs_fs_info *fs_info,
 109                                      struct btrfs_qgroup *dest,
 110                                      struct btrfs_qgroup *src)
 111 {
 112         int i;
 113
 114         for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++)
 115                 qgroup_rsv_add(fs_info, dest, src->rsv.values[i], i);
 116 }
 117
 118 static void qgroup_rsv_release_by_qgroup(struct btrfs_fs_info *fs_info,
 119                                          struct btrfs_qgroup *dest,
 120                                           struct btrfs_qgroup *src)
 121 {
 122         int i;
 123
 124         for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++)
 125                 qgroup_rsv_release(fs_info, dest, src->rsv.values[i], i);
 126 }
 127
 128 static void btrfs_qgroup_update_old_refcnt(struct btrfs_qgroup *qg, u64 seq,
 129                                            int mod)
 130 {
 131         if (qg->old_refcnt < seq)
 132                 qg->old_refcnt = seq;
 133         qg->old_refcnt += mod;
 134 }
 135
 136 static void btrfs_qgroup_update_new_refcnt(struct btrfs_qgroup *qg, u64 seq,
 137                                            int mod)
 138 {
 139         if (qg->new_refcnt < seq)
 140                 qg->new_refcnt = seq;
 141         qg->new_refcnt += mod;
 142 }
 143
 144 static inline u64 btrfs_qgroup_get_old_refcnt(struct btrfs_qgroup *qg, u64 seq)
 145 {
 146         if (qg->old_refcnt < seq)
 147                 return 0;
 148         return qg->old_refcnt - seq;
 149 }
 150
 151 static inline u64 btrfs_qgroup_get_new_refcnt(struct btrfs_qgroup *qg, u64 seq)
 152 {
 153         if (qg->new_refcnt < seq)
 154                 return 0;
 155         return qg->new_refcnt - seq;
 156 }
 157
 158 /*
 159  * glue structure to represent the relations between qgroups.
 160  */
 161 struct btrfs_qgroup_list {
 162         struct list_head next_group;
 163         struct list_head next_member;
 164         struct btrfs_qgroup *group;
 165         struct btrfs_qgroup *member;
 166 };
 167
 168 static int
 169 qgroup_rescan_init(struct btrfs_fs_info *fs_info, u64 progress_objectid,
 170                    int init_flags);
 171 static void qgroup_rescan_zero_tracking(struct btrfs_fs_info *fs_info);
 172
 173 /* must be called with qgroup_ioctl_lock held */
 174 static struct btrfs_qgroup *find_qgroup_rb(struct btrfs_fs_info *fs_info,
 175                                            u64 qgroupid)
 176 {
 177         struct rb_node *n = fs_info->qgroup_tree.rb_node;
 178         struct btrfs_qgroup *qgroup;
 179
 180         while (n) {
 181                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
 182                 if (qgroup->qgroupid < qgroupid)
 183                         n = n->rb_left;
 184                 else if (qgroup->qgroupid > qgroupid)
 185                         n = n->rb_right;
 186                 else
 187                         return qgroup;
 188         }
 189         return NULL;
 190 }
 191
 192 /*
 193  * Add qgroup to the filesystem's qgroup tree.
 194  *
 195  * Must be called with qgroup_lock held and @prealloc preallocated.
 196  *
 197  * The control on the lifespan of @prealloc would be transfered to this
 198  * function, thus caller should no longer touch @prealloc.
 199  */
 200 static struct btrfs_qgroup *add_qgroup_rb(struct btrfs_fs_info *fs_info,
 201                                           struct btrfs_qgroup *prealloc,
 202                                           u64 qgroupid)
 203 {
 204         struct rb_node **p = &fs_info->qgroup_tree.rb_node;
 205         struct rb_node *parent = NULL;
 206         struct btrfs_qgroup *qgroup;
 207
 208         /* Caller must have pre-allocated @prealloc. */
 209         ASSERT(prealloc);
 210
 211         while (*p) {
 212                 parent = *p;
 213                 qgroup = rb_entry(parent, struct btrfs_qgroup, node);
 214
 215                 if (qgroup->qgroupid < qgroupid) {
 216                         p = &(*p)->rb_left;
 217                 } else if (qgroup->qgroupid > qgroupid) {
 218                         p = &(*p)->rb_right;
 219                 } else {
 220                         kfree(prealloc);
 221                         return qgroup;
 222                 }
 223         }
 224
 225         qgroup = prealloc;
 226         qgroup->qgroupid = qgroupid;
 227         INIT_LIST_HEAD(&qgroup->groups);
 228         INIT_LIST_HEAD(&qgroup->members);
 229         INIT_LIST_HEAD(&qgroup->dirty);
 230         INIT_LIST_HEAD(&qgroup->iterator);
 231         INIT_LIST_HEAD(&qgroup->nested_iterator);
 232
 233         rb_link_node(&qgroup->node, parent, p);
 234         rb_insert_color(&qgroup->node, &fs_info->qgroup_tree);
 235
 236         return qgroup;
 237 }
 238
 239 static void __del_qgroup_rb(struct btrfs_fs_info *fs_info,
 240                             struct btrfs_qgroup *qgroup)
 241 {
 242         struct btrfs_qgroup_list *list;
 243
 244         list_del(&qgroup->dirty);
 245         while (!list_empty(&qgroup->groups)) {
 246                 list = list_first_entry(&qgroup->groups,
 247                                         struct btrfs_qgroup_list, next_group);
 248                 list_del(&list->next_group);
 249                 list_del(&list->next_member);
 250                 kfree(list);
 251         }
 252
 253         while (!list_empty(&qgroup->members)) {
 254                 list = list_first_entry(&qgroup->members,
 255                                         struct btrfs_qgroup_list, next_member);
 256                 list_del(&list->next_group);
 257                 list_del(&list->next_member);
 258                 kfree(list);
 259         }
 260 }
 261
 262 /* must be called with qgroup_lock held */
 263 static int del_qgroup_rb(struct btrfs_fs_info *fs_info, u64 qgroupid)
 264 {
 265         struct btrfs_qgroup *qgroup = find_qgroup_rb(fs_info, qgroupid);
 266
 267         if (!qgroup)
 268                 return -ENOENT;
 269
 270         rb_erase(&qgroup->node, &fs_info->qgroup_tree);
 271         __del_qgroup_rb(fs_info, qgroup);
 272         return 0;
 273 }
 274
 275 /*
 276  * Add relation specified by two qgroups.
 277  *
 278  * Must be called with qgroup_lock held, the ownership of @prealloc is
 279  * transferred to this function and caller should not touch it anymore.
 280  *
 281  * Return: 0        on success
 282  *         -ENOENT  if one of the qgroups is NULL
 283  *         <0       other errors
 284  */
 285 static int __add_relation_rb(struct btrfs_qgroup_list *prealloc,
 286                              struct btrfs_qgroup *member,
 287                              struct btrfs_qgroup *parent)
 288 {
 289         if (!member || !parent) {
 290                 kfree(prealloc);
 291                 return -ENOENT;
 292         }
 293
 294         prealloc->group = parent;
 295         prealloc->member = member;
 296         list_add_tail(&prealloc->next_group, &member->groups);
 297         list_add_tail(&prealloc->next_member, &parent->members);
 298
 299         return 0;
 300 }
 301
 302 /*
 303  * Add relation specified by two qgroup ids.
 304  *
 305  * Must be called with qgroup_lock held.
 306  *
 307  * Return: 0        on success
 308  *         -ENOENT  if one of the ids does not exist
 309  *         <0       other errors
 310  */
 311 static int add_relation_rb(struct btrfs_fs_info *fs_info,
 312                            struct btrfs_qgroup_list *prealloc,
 313                            u64 memberid, u64 parentid)
 314 {
 315         struct btrfs_qgroup *member;
 316         struct btrfs_qgroup *parent;
 317
 318         member = find_qgroup_rb(fs_info, memberid);
 319         parent = find_qgroup_rb(fs_info, parentid);
 320
 321         return __add_relation_rb(prealloc, member, parent);
 322 }
 323
 324 /* Must be called with qgroup_lock held */
 325 static int del_relation_rb(struct btrfs_fs_info *fs_info,
 326                            u64 memberid, u64 parentid)
 327 {
 328         struct btrfs_qgroup *member;
 329         struct btrfs_qgroup *parent;
 330         struct btrfs_qgroup_list *list;
 331
 332         member = find_qgroup_rb(fs_info, memberid);
 333         parent = find_qgroup_rb(fs_info, parentid);
 334         if (!member || !parent)
 335                 return -ENOENT;
 336
 337         list_for_each_entry(list, &member->groups, next_group) {
 338                 if (list->group == parent) {
 339                         list_del(&list->next_group);
 340                         list_del(&list->next_member);
 341                         kfree(list);
 342                         return 0;
 343                 }
 344         }
 345         return -ENOENT;
 346 }
 347
 348 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 349 int btrfs_verify_qgroup_counts(struct btrfs_fs_info *fs_info, u64 qgroupid,
 350                                u64 rfer, u64 excl)
 351 {
 352         struct btrfs_qgroup *qgroup;
 353
 354         qgroup = find_qgroup_rb(fs_info, qgroupid);
 355         if (!qgroup)
 356                 return -EINVAL;
 357         if (qgroup->rfer != rfer || qgroup->excl != excl)
 358                 return -EINVAL;
 359         return 0;
 360 }
 361 #endif
 362
 363 static void qgroup_mark_inconsistent(struct btrfs_fs_info *fs_info)
 364 {
 365         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE)
 366                 return;
 367         fs_info->qgroup_flags |= (BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT |
 368                                   BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN |
 369                                   BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING);
 370 }
 371
 372 static void qgroup_read_enable_gen(struct btrfs_fs_info *fs_info,
 373                                    struct extent_buffer *leaf, int slot,
 374                                    struct btrfs_qgroup_status_item *ptr)
 375 {
 376         ASSERT(btrfs_fs_incompat(fs_info, SIMPLE_QUOTA));
 377         ASSERT(btrfs_item_size(leaf, slot) >= sizeof(*ptr));
 378         fs_info->qgroup_enable_gen = btrfs_qgroup_status_enable_gen(leaf, ptr);
 379 }
 380
 381 /*
 382  * The full config is read in one go, only called from open_ctree()
 383  * It doesn't use any locking, as at this point we're still single-threaded
 384  */
 385 int btrfs_read_qgroup_config(struct btrfs_fs_info *fs_info)
 386 {
 387         struct btrfs_key key;
 388         struct btrfs_key found_key;
 389         struct btrfs_root *quota_root = fs_info->quota_root;
 390         struct btrfs_path *path = NULL;
 391         struct extent_buffer *l;
 392         int slot;
 393         int ret = 0;
 394         u64 flags = 0;
 395         u64 rescan_progress = 0;
 396
 397         if (!fs_info->quota_root)
 398                 return 0;
 399
 400         fs_info->qgroup_ulist = ulist_alloc(GFP_KERNEL);
 401         if (!fs_info->qgroup_ulist) {
 402                 ret = -ENOMEM;
 403                 goto out;
 404         }
 405
 406         path = btrfs_alloc_path();
 407         if (!path) {
 408                 ret = -ENOMEM;
 409                 goto out;
 410         }
 411
 412         ret = btrfs_sysfs_add_qgroups(fs_info);
 413         if (ret < 0)
 414                 goto out;
 415         /* default this to quota off, in case no status key is found */
 416         fs_info->qgroup_flags = 0;
 417
 418         /*
 419          * pass 1: read status, all qgroup infos and limits
 420          */
 421         key.objectid = 0;
 422         key.type = 0;
 423         key.offset = 0;
 424         ret = btrfs_search_slot_for_read(quota_root, &key, path, 1, 1);
 425         if (ret)
 426                 goto out;
 427
 428         while (1) {
 429                 struct btrfs_qgroup *qgroup;
 430
 431                 slot = path->slots[0];
 432                 l = path->nodes[0];
 433                 btrfs_item_key_to_cpu(l, &found_key, slot);
 434
 435                 if (found_key.type == BTRFS_QGROUP_STATUS_KEY) {
 436                         struct btrfs_qgroup_status_item *ptr;
 437
 438                         ptr = btrfs_item_ptr(l, slot,
 439                                              struct btrfs_qgroup_status_item);
 440
 441                         if (btrfs_qgroup_status_version(l, ptr) !=
 442                             BTRFS_QGROUP_STATUS_VERSION) {
 443                                 btrfs_err(fs_info,
 444                                  "old qgroup version, quota disabled");
 445                                 goto out;
 446                         }
 447                         fs_info->qgroup_flags = btrfs_qgroup_status_flags(l, ptr);
 448                         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_SIMPLE_MODE) {
 449                                 qgroup_read_enable_gen(fs_info, l, slot, ptr);
 450                         } else if (btrfs_qgroup_status_generation(l, ptr) != fs_info->generation) {
 451                                 qgroup_mark_inconsistent(fs_info);
 452                                 btrfs_err(fs_info,
 453                                         "qgroup generation mismatch, marked as inconsistent");
 454                         }
 455                         rescan_progress = btrfs_qgroup_status_rescan(l, ptr);
 456                         goto next1;
 457                 }
 458
 459                 if (found_key.type != BTRFS_QGROUP_INFO_KEY &&
 460                     found_key.type != BTRFS_QGROUP_LIMIT_KEY)
 461                         goto next1;
 462
 463                 qgroup = find_qgroup_rb(fs_info, found_key.offset);
 464                 if ((qgroup && found_key.type == BTRFS_QGROUP_INFO_KEY) ||
 465                     (!qgroup && found_key.type == BTRFS_QGROUP_LIMIT_KEY)) {
 466                         btrfs_err(fs_info, "inconsistent qgroup config");
 467                         qgroup_mark_inconsistent(fs_info);
 468                 }
 469                 if (!qgroup) {
 470                         struct btrfs_qgroup *prealloc;
 471
 472                         prealloc = kzalloc(sizeof(*prealloc), GFP_KERNEL);
 473                         if (!prealloc) {
 474                                 ret = -ENOMEM;
 475                                 goto out;
 476                         }
 477                         qgroup = add_qgroup_rb(fs_info, prealloc, found_key.offset);
 478                 }
 479                 ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
 480                 if (ret < 0)
 481                         goto out;
 482
 483                 switch (found_key.type) {
 484                 case BTRFS_QGROUP_INFO_KEY: {
 485                         struct btrfs_qgroup_info_item *ptr;
 486
 487                         ptr = btrfs_item_ptr(l, slot,
 488                                              struct btrfs_qgroup_info_item);
 489                         qgroup->rfer = btrfs_qgroup_info_rfer(l, ptr);
 490                         qgroup->rfer_cmpr = btrfs_qgroup_info_rfer_cmpr(l, ptr);
 491                         qgroup->excl = btrfs_qgroup_info_excl(l, ptr);
 492                         qgroup->excl_cmpr = btrfs_qgroup_info_excl_cmpr(l, ptr);
 493                         /* generation currently unused */
 494                         break;
 495                 }
 496                 case BTRFS_QGROUP_LIMIT_KEY: {
 497                         struct btrfs_qgroup_limit_item *ptr;
 498
 499                         ptr = btrfs_item_ptr(l, slot,
 500                                              struct btrfs_qgroup_limit_item);
 501                         qgroup->lim_flags = btrfs_qgroup_limit_flags(l, ptr);
 502                         qgroup->max_rfer = btrfs_qgroup_limit_max_rfer(l, ptr);
 503                         qgroup->max_excl = btrfs_qgroup_limit_max_excl(l, ptr);
 504                         qgroup->rsv_rfer = btrfs_qgroup_limit_rsv_rfer(l, ptr);
 505                         qgroup->rsv_excl = btrfs_qgroup_limit_rsv_excl(l, ptr);
 506                         break;
 507                 }
 508                 }
 509 next1:
 510                 ret = btrfs_next_item(quota_root, path);
 511                 if (ret < 0)
 512                         goto out;
 513                 if (ret)
 514                         break;
 515         }
 516         btrfs_release_path(path);
 517
 518         /*
 519          * pass 2: read all qgroup relations
 520          */
 521         key.objectid = 0;
 522         key.type = BTRFS_QGROUP_RELATION_KEY;
 523         key.offset = 0;
 524         ret = btrfs_search_slot_for_read(quota_root, &key, path, 1, 0);
 525         if (ret)
 526                 goto out;
 527         while (1) {
 528                 struct btrfs_qgroup_list *list = NULL;
 529
 530                 slot = path->slots[0];
 531                 l = path->nodes[0];
 532                 btrfs_item_key_to_cpu(l, &found_key, slot);
 533
 534                 if (found_key.type != BTRFS_QGROUP_RELATION_KEY)
 535                         goto next2;
 536
 537                 if (found_key.objectid > found_key.offset) {
 538                         /* parent <- member, not needed to build config */
 539                         /* FIXME should we omit the key completely? */
 540                         goto next2;
 541                 }
 542
 543                 list = kzalloc(sizeof(*list), GFP_KERNEL);
 544                 if (!list) {
 545                         ret = -ENOMEM;
 546                         goto out;
 547                 }
 548                 ret = add_relation_rb(fs_info, list, found_key.objectid,
 549                                       found_key.offset);
 550                 list = NULL;
 551                 if (ret == -ENOENT) {
 552                         btrfs_warn(fs_info,
 553                                 "orphan qgroup relation 0x%llx->0x%llx",
 554                                 found_key.objectid, found_key.offset);
 555                         ret = 0;        /* ignore the error */
 556                 }
 557                 if (ret)
 558                         goto out;
 559 next2:
 560                 ret = btrfs_next_item(quota_root, path);
 561                 if (ret < 0)
 562                         goto out;
 563                 if (ret)
 564                         break;
 565         }
 566 out:
 567         btrfs_free_path(path);
 568         fs_info->qgroup_flags |= flags;
 569         if (ret >= 0) {
 570                 if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_ON)
 571                         set_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
 572                 if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN)
 573                         ret = qgroup_rescan_init(fs_info, rescan_progress, 0);
 574         } else {
 575                 ulist_free(fs_info->qgroup_ulist);
 576                 fs_info->qgroup_ulist = NULL;
 577                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
 578                 btrfs_sysfs_del_qgroups(fs_info);
 579         }
 580
 581         return ret < 0 ? ret : 0;
 582 }
 583
 584 /*
 585  * Called in close_ctree() when quota is still enabled.  This verifies we don't
 586  * leak some reserved space.
 587  *
 588  * Return false if no reserved space is left.
 589  * Return true if some reserved space is leaked.
 590  */
 591 bool btrfs_check_quota_leak(struct btrfs_fs_info *fs_info)
 592 {
 593         struct rb_node *node;
 594         bool ret = false;
 595
 596         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_DISABLED)
 597                 return ret;
 598         /*
 599          * Since we're unmounting, there is no race and no need to grab qgroup
 600          * lock.  And here we don't go post-order to provide a more user
 601          * friendly sorted result.
 602          */
 603         for (node = rb_first(&fs_info->qgroup_tree); node; node = rb_next(node)) {
 604                 struct btrfs_qgroup *qgroup;
 605                 int i;
 606
 607                 qgroup = rb_entry(node, struct btrfs_qgroup, node);
 608                 for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++) {
 609                         if (qgroup->rsv.values[i]) {
 610                                 ret = true;
 611                                 btrfs_warn(fs_info,
 612                 "qgroup %hu/%llu has unreleased space, type %d rsv %llu",
 613                                    btrfs_qgroup_level(qgroup->qgroupid),
 614                                    btrfs_qgroup_subvolid(qgroup->qgroupid),
 615                                    i, qgroup->rsv.values[i]);
 616                         }
 617                 }
 618         }
 619         return ret;
 620 }
 621
 622 /*
 623  * This is called from close_ctree() or open_ctree() or btrfs_quota_disable(),
 624  * first two are in single-threaded paths.And for the third one, we have set
 625  * quota_root to be null with qgroup_lock held before, so it is safe to clean
 626  * up the in-memory structures without qgroup_lock held.
 627  */
 628 void btrfs_free_qgroup_config(struct btrfs_fs_info *fs_info)
 629 {
 630         struct rb_node *n;
 631         struct btrfs_qgroup *qgroup;
 632
 633         while ((n = rb_first(&fs_info->qgroup_tree))) {
 634                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
 635                 rb_erase(n, &fs_info->qgroup_tree);
 636                 __del_qgroup_rb(fs_info, qgroup);
 637                 btrfs_sysfs_del_one_qgroup(fs_info, qgroup);
 638                 kfree(qgroup);
 639         }
 640         /*
 641          * We call btrfs_free_qgroup_config() when unmounting
 642          * filesystem and disabling quota, so we set qgroup_ulist
 643          * to be null here to avoid double free.
 644          */
 645         ulist_free(fs_info->qgroup_ulist);
 646         fs_info->qgroup_ulist = NULL;
 647         btrfs_sysfs_del_qgroups(fs_info);
 648 }
 649
 650 static int add_qgroup_relation_item(struct btrfs_trans_handle *trans, u64 src,
 651                                     u64 dst)
 652 {
 653         int ret;
 654         struct btrfs_root *quota_root = trans->fs_info->quota_root;
 655         struct btrfs_path *path;
 656         struct btrfs_key key;
 657
 658         path = btrfs_alloc_path();
 659         if (!path)
 660                 return -ENOMEM;
 661
 662         key.objectid = src;
 663         key.type = BTRFS_QGROUP_RELATION_KEY;
 664         key.offset = dst;
 665
 666         ret = btrfs_insert_empty_item(trans, quota_root, path, &key, 0);
 667
 668         btrfs_mark_buffer_dirty(trans, path->nodes[0]);
 669
 670         btrfs_free_path(path);
 671         return ret;
 672 }
 673
 674 static int del_qgroup_relation_item(struct btrfs_trans_handle *trans, u64 src,
 675                                     u64 dst)
 676 {
 677         int ret;
 678         struct btrfs_root *quota_root = trans->fs_info->quota_root;
 679         struct btrfs_path *path;
 680         struct btrfs_key key;
 681
 682         path = btrfs_alloc_path();
 683         if (!path)
 684                 return -ENOMEM;
 685
 686         key.objectid = src;
 687         key.type = BTRFS_QGROUP_RELATION_KEY;
 688         key.offset = dst;
 689
 690         ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
 691         if (ret < 0)
 692                 goto out;
 693
 694         if (ret > 0) {
 695                 ret = -ENOENT;
 696                 goto out;
 697         }
 698
 699         ret = btrfs_del_item(trans, quota_root, path);
 700 out:
 701         btrfs_free_path(path);
 702         return ret;
 703 }
 704
 705 static int add_qgroup_item(struct btrfs_trans_handle *trans,
 706                            struct btrfs_root *quota_root, u64 qgroupid)
 707 {
 708         int ret;
 709         struct btrfs_path *path;
 710         struct btrfs_qgroup_info_item *qgroup_info;
 711         struct btrfs_qgroup_limit_item *qgroup_limit;
 712         struct extent_buffer *leaf;
 713         struct btrfs_key key;
 714
 715         if (btrfs_is_testing(quota_root->fs_info))
 716                 return 0;
 717
 718         path = btrfs_alloc_path();
 719         if (!path)
 720                 return -ENOMEM;
 721
 722         key.objectid = 0;
 723         key.type = BTRFS_QGROUP_INFO_KEY;
 724         key.offset = qgroupid;
 725
 726         /*
 727          * Avoid a transaction abort by catching -EEXIST here. In that
 728          * case, we proceed by re-initializing the existing structure
 729          * on disk.
 730          */
 731
 732         ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
 733                                       sizeof(*qgroup_info));
 734         if (ret && ret != -EEXIST)
 735                 goto out;
 736
 737         leaf = path->nodes[0];
 738         qgroup_info = btrfs_item_ptr(leaf, path->slots[0],
 739                                  struct btrfs_qgroup_info_item);
 740         btrfs_set_qgroup_info_generation(leaf, qgroup_info, trans->transid);
 741         btrfs_set_qgroup_info_rfer(leaf, qgroup_info, 0);
 742         btrfs_set_qgroup_info_rfer_cmpr(leaf, qgroup_info, 0);
 743         btrfs_set_qgroup_info_excl(leaf, qgroup_info, 0);
 744         btrfs_set_qgroup_info_excl_cmpr(leaf, qgroup_info, 0);
 745
 746         btrfs_mark_buffer_dirty(trans, leaf);
 747
 748         btrfs_release_path(path);
 749
 750         key.type = BTRFS_QGROUP_LIMIT_KEY;
 751         ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
 752                                       sizeof(*qgroup_limit));
 753         if (ret && ret != -EEXIST)
 754                 goto out;
 755
 756         leaf = path->nodes[0];
 757         qgroup_limit = btrfs_item_ptr(leaf, path->slots[0],
 758                                   struct btrfs_qgroup_limit_item);
 759         btrfs_set_qgroup_limit_flags(leaf, qgroup_limit, 0);
 760         btrfs_set_qgroup_limit_max_rfer(leaf, qgroup_limit, 0);
 761         btrfs_set_qgroup_limit_max_excl(leaf, qgroup_limit, 0);
 762         btrfs_set_qgroup_limit_rsv_rfer(leaf, qgroup_limit, 0);
 763         btrfs_set_qgroup_limit_rsv_excl(leaf, qgroup_limit, 0);
 764
 765         btrfs_mark_buffer_dirty(trans, leaf);
 766
 767         ret = 0;
 768 out:
 769         btrfs_free_path(path);
 770         return ret;
 771 }
 772
 773 static int del_qgroup_item(struct btrfs_trans_handle *trans, u64 qgroupid)
 774 {
 775         int ret;
 776         struct btrfs_root *quota_root = trans->fs_info->quota_root;
 777         struct btrfs_path *path;
 778         struct btrfs_key key;
 779
 780         path = btrfs_alloc_path();
 781         if (!path)
 782                 return -ENOMEM;
 783
 784         key.objectid = 0;
 785         key.type = BTRFS_QGROUP_INFO_KEY;
 786         key.offset = qgroupid;
 787         ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
 788         if (ret < 0)
 789                 goto out;
 790
 791         if (ret > 0) {
 792                 ret = -ENOENT;
 793                 goto out;
 794         }
 795
 796         ret = btrfs_del_item(trans, quota_root, path);
 797         if (ret)
 798                 goto out;
 799
 800         btrfs_release_path(path);
 801
 802         key.type = BTRFS_QGROUP_LIMIT_KEY;
 803         ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
 804         if (ret < 0)
 805                 goto out;
 806
 807         if (ret > 0) {
 808                 ret = -ENOENT;
 809                 goto out;
 810         }
 811
 812         ret = btrfs_del_item(trans, quota_root, path);
 813
 814 out:
 815         btrfs_free_path(path);
 816         return ret;
 817 }
 818
 819 static int update_qgroup_limit_item(struct btrfs_trans_handle *trans,
 820                                     struct btrfs_qgroup *qgroup)
 821 {
 822         struct btrfs_root *quota_root = trans->fs_info->quota_root;
 823         struct btrfs_path *path;
 824         struct btrfs_key key;
 825         struct extent_buffer *l;
 826         struct btrfs_qgroup_limit_item *qgroup_limit;
 827         int ret;
 828         int slot;
 829
 830         key.objectid = 0;
 831         key.type = BTRFS_QGROUP_LIMIT_KEY;
 832         key.offset = qgroup->qgroupid;
 833
 834         path = btrfs_alloc_path();
 835         if (!path)
 836                 return -ENOMEM;
 837
 838         ret = btrfs_search_slot(trans, quota_root, &key, path, 0, 1);
 839         if (ret > 0)
 840                 ret = -ENOENT;
 841
 842         if (ret)
 843                 goto out;
 844
 845         l = path->nodes[0];
 846         slot = path->slots[0];
 847         qgroup_limit = btrfs_item_ptr(l, slot, struct btrfs_qgroup_limit_item);
 848         btrfs_set_qgroup_limit_flags(l, qgroup_limit, qgroup->lim_flags);
 849         btrfs_set_qgroup_limit_max_rfer(l, qgroup_limit, qgroup->max_rfer);
 850         btrfs_set_qgroup_limit_max_excl(l, qgroup_limit, qgroup->max_excl);
 851         btrfs_set_qgroup_limit_rsv_rfer(l, qgroup_limit, qgroup->rsv_rfer);
 852         btrfs_set_qgroup_limit_rsv_excl(l, qgroup_limit, qgroup->rsv_excl);
 853
 854         btrfs_mark_buffer_dirty(trans, l);
 855
 856 out:
 857         btrfs_free_path(path);
 858         return ret;
 859 }
 860
 861 static int update_qgroup_info_item(struct btrfs_trans_handle *trans,
 862                                    struct btrfs_qgroup *qgroup)
 863 {
 864         struct btrfs_fs_info *fs_info = trans->fs_info;
 865         struct btrfs_root *quota_root = fs_info->quota_root;
 866         struct btrfs_path *path;
 867         struct btrfs_key key;
 868         struct extent_buffer *l;
 869         struct btrfs_qgroup_info_item *qgroup_info;
 870         int ret;
 871         int slot;
 872
 873         if (btrfs_is_testing(fs_info))
 874                 return 0;
 875
 876         key.objectid = 0;
 877         key.type = BTRFS_QGROUP_INFO_KEY;
 878         key.offset = qgroup->qgroupid;
 879
 880         path = btrfs_alloc_path();
 881         if (!path)
 882                 return -ENOMEM;
 883
 884         ret = btrfs_search_slot(trans, quota_root, &key, path, 0, 1);
 885         if (ret > 0)
 886                 ret = -ENOENT;
 887
 888         if (ret)
 889                 goto out;
 890
 891         l = path->nodes[0];
 892         slot = path->slots[0];
 893         qgroup_info = btrfs_item_ptr(l, slot, struct btrfs_qgroup_info_item);
 894         btrfs_set_qgroup_info_generation(l, qgroup_info, trans->transid);
 895         btrfs_set_qgroup_info_rfer(l, qgroup_info, qgroup->rfer);
 896         btrfs_set_qgroup_info_rfer_cmpr(l, qgroup_info, qgroup->rfer_cmpr);
 897         btrfs_set_qgroup_info_excl(l, qgroup_info, qgroup->excl);
 898         btrfs_set_qgroup_info_excl_cmpr(l, qgroup_info, qgroup->excl_cmpr);
 899
 900         btrfs_mark_buffer_dirty(trans, l);
 901
 902 out:
 903         btrfs_free_path(path);
 904         return ret;
 905 }
 906
 907 static int update_qgroup_status_item(struct btrfs_trans_handle *trans)
 908 {
 909         struct btrfs_fs_info *fs_info = trans->fs_info;
 910         struct btrfs_root *quota_root = fs_info->quota_root;
 911         struct btrfs_path *path;
 912         struct btrfs_key key;
 913         struct extent_buffer *l;
 914         struct btrfs_qgroup_status_item *ptr;
 915         int ret;
 916         int slot;
 917
 918         key.objectid = 0;
 919         key.type = BTRFS_QGROUP_STATUS_KEY;
 920         key.offset = 0;
 921
 922         path = btrfs_alloc_path();
 923         if (!path)
 924                 return -ENOMEM;
 925
 926         ret = btrfs_search_slot(trans, quota_root, &key, path, 0, 1);
 927         if (ret > 0)
 928                 ret = -ENOENT;
 929
 930         if (ret)
 931                 goto out;
 932
 933         l = path->nodes[0];
 934         slot = path->slots[0];
 935         ptr = btrfs_item_ptr(l, slot, struct btrfs_qgroup_status_item);
 936         btrfs_set_qgroup_status_flags(l, ptr, fs_info->qgroup_flags &
 937                                       BTRFS_QGROUP_STATUS_FLAGS_MASK);
 938         btrfs_set_qgroup_status_generation(l, ptr, trans->transid);
 939         btrfs_set_qgroup_status_rescan(l, ptr,
 940                                 fs_info->qgroup_rescan_progress.objectid);
 941
 942         btrfs_mark_buffer_dirty(trans, l);
 943
 944 out:
 945         btrfs_free_path(path);
 946         return ret;
 947 }
 948
 949 /*
 950  * called with qgroup_lock held
 951  */
 952 static int btrfs_clean_quota_tree(struct btrfs_trans_handle *trans,
 953                                   struct btrfs_root *root)
 954 {
 955         struct btrfs_path *path;
 956         struct btrfs_key key;
 957         struct extent_buffer *leaf = NULL;
 958         int ret;
 959         int nr = 0;
 960
 961         path = btrfs_alloc_path();
 962         if (!path)
 963                 return -ENOMEM;
 964
 965         key.objectid = 0;
 966         key.offset = 0;
 967         key.type = 0;
 968
 969         while (1) {
 970                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 971                 if (ret < 0)
 972                         goto out;
 973                 leaf = path->nodes[0];
 974                 nr = btrfs_header_nritems(leaf);
 975                 if (!nr)
 976                         break;
 977                 /*
 978                  * delete the leaf one by one
 979                  * since the whole tree is going
 980                  * to be deleted.
 981                  */
 982                 path->slots[0] = 0;
 983                 ret = btrfs_del_items(trans, root, path, 0, nr);
 984                 if (ret)
 985                         goto out;
 986
 987                 btrfs_release_path(path);
 988         }
 989         ret = 0;
 990 out:
 991         btrfs_free_path(path);
 992         return ret;
 993 }
 994
 995 int btrfs_quota_enable(struct btrfs_fs_info *fs_info,
 996                        struct btrfs_ioctl_quota_ctl_args *quota_ctl_args)
 997 {
 998         struct btrfs_root *quota_root;
 999         struct btrfs_root *tree_root = fs_info->tree_root;
1000         struct btrfs_path *path = NULL;
1001         struct btrfs_qgroup_status_item *ptr;
1002         struct extent_buffer *leaf;
1003         struct btrfs_key key;
1004         struct btrfs_key found_key;
1005         struct btrfs_qgroup *qgroup = NULL;
1006         struct btrfs_qgroup *prealloc = NULL;
1007         struct btrfs_trans_handle *trans = NULL;
1008         struct ulist *ulist = NULL;
1009         const bool simple = (quota_ctl_args->cmd == BTRFS_QUOTA_CTL_ENABLE_SIMPLE_QUOTA);
1010         int ret = 0;
1011         int slot;
1012
1013         /*
1014          * We need to have subvol_sem write locked, to prevent races between
1015          * concurrent tasks trying to enable quotas, because we will unlock
1016          * and relock qgroup_ioctl_lock before setting fs_info->quota_root
1017          * and before setting BTRFS_FS_QUOTA_ENABLED.
1018          */
1019         lockdep_assert_held_write(&fs_info->subvol_sem);
1020
1021         if (btrfs_fs_incompat(fs_info, EXTENT_TREE_V2)) {
1022                 btrfs_err(fs_info,
1023                           "qgroups are currently unsupported in extent tree v2");
1024                 return -EINVAL;
1025         }
1026
1027         mutex_lock(&fs_info->qgroup_ioctl_lock);
1028         if (fs_info->quota_root)
1029                 goto out;
1030
1031         ulist = ulist_alloc(GFP_KERNEL);
1032         if (!ulist) {
1033                 ret = -ENOMEM;
1034                 goto out;
1035         }
1036
1037         ret = btrfs_sysfs_add_qgroups(fs_info);
1038         if (ret < 0)
1039                 goto out;
1040
1041         /*
1042          * Unlock qgroup_ioctl_lock before starting the transaction. This is to
1043          * avoid lock acquisition inversion problems (reported by lockdep) between
1044          * qgroup_ioctl_lock and the vfs freeze semaphores, acquired when we
1045          * start a transaction.
1046          * After we started the transaction lock qgroup_ioctl_lock again and
1047          * check if someone else created the quota root in the meanwhile. If so,
1048          * just return success and release the transaction handle.
1049          *
1050          * Also we don't need to worry about someone else calling
1051          * btrfs_sysfs_add_qgroups() after we unlock and getting an error because
1052          * that function returns 0 (success) when the sysfs entries already exist.
1053          */
1054         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1055
1056         /*
1057          * 1 for quota root item
1058          * 1 for BTRFS_QGROUP_STATUS item
1059          *
1060          * Yet we also need 2*n items for a QGROUP_INFO/QGROUP_LIMIT items
1061          * per subvolume. However those are not currently reserved since it
1062          * would be a lot of overkill.
1063          */
1064         trans = btrfs_start_transaction(tree_root, 2);
1065
1066         mutex_lock(&fs_info->qgroup_ioctl_lock);
1067         if (IS_ERR(trans)) {
1068                 ret = PTR_ERR(trans);
1069                 trans = NULL;
1070                 goto out;
1071         }
1072
1073         if (fs_info->quota_root)
1074                 goto out;
1075
1076         fs_info->qgroup_ulist = ulist;
1077         ulist = NULL;
1078
1079         /*
1080          * initially create the quota tree
1081          */
1082         quota_root = btrfs_create_tree(trans, BTRFS_QUOTA_TREE_OBJECTID);
1083         if (IS_ERR(quota_root)) {
1084                 ret =  PTR_ERR(quota_root);
1085                 btrfs_abort_transaction(trans, ret);
1086                 goto out;
1087         }
1088
1089         path = btrfs_alloc_path();
1090         if (!path) {
1091                 ret = -ENOMEM;
1092                 btrfs_abort_transaction(trans, ret);
1093                 goto out_free_root;
1094         }
1095
1096         key.objectid = 0;
1097         key.type = BTRFS_QGROUP_STATUS_KEY;
1098         key.offset = 0;
1099
1100         ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
1101                                       sizeof(*ptr));
1102         if (ret) {
1103                 btrfs_abort_transaction(trans, ret);
1104                 goto out_free_path;
1105         }
1106
1107         leaf = path->nodes[0];
1108         ptr = btrfs_item_ptr(leaf, path->slots[0],
1109                                  struct btrfs_qgroup_status_item);
1110         btrfs_set_qgroup_status_generation(leaf, ptr, trans->transid);
1111         btrfs_set_qgroup_status_version(leaf, ptr, BTRFS_QGROUP_STATUS_VERSION);
1112         fs_info->qgroup_flags = BTRFS_QGROUP_STATUS_FLAG_ON;
1113         if (simple) {
1114                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_SIMPLE_MODE;
1115                 btrfs_set_qgroup_status_enable_gen(leaf, ptr, trans->transid);
1116         } else {
1117                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1118         }
1119         btrfs_set_qgroup_status_flags(leaf, ptr, fs_info->qgroup_flags &
1120                                       BTRFS_QGROUP_STATUS_FLAGS_MASK);
1121         btrfs_set_qgroup_status_rescan(leaf, ptr, 0);
1122
1123         btrfs_mark_buffer_dirty(trans, leaf);
1124
1125         key.objectid = 0;
1126         key.type = BTRFS_ROOT_REF_KEY;
1127         key.offset = 0;
1128
1129         btrfs_release_path(path);
1130         ret = btrfs_search_slot_for_read(tree_root, &key, path, 1, 0);
1131         if (ret > 0)
1132                 goto out_add_root;
1133         if (ret < 0) {
1134                 btrfs_abort_transaction(trans, ret);
1135                 goto out_free_path;
1136         }
1137
1138         while (1) {
1139                 slot = path->slots[0];
1140                 leaf = path->nodes[0];
1141                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1142
1143                 if (found_key.type == BTRFS_ROOT_REF_KEY) {
1144
1145                         /* Release locks on tree_root before we access quota_root */
1146                         btrfs_release_path(path);
1147
1148                         /* We should not have a stray @prealloc pointer. */
1149                         ASSERT(prealloc == NULL);
1150                         prealloc = kzalloc(sizeof(*prealloc), GFP_NOFS);
1151                         if (!prealloc) {
1152                                 ret = -ENOMEM;
1153                                 btrfs_abort_transaction(trans, ret);
1154                                 goto out_free_path;
1155                         }
1156
1157                         ret = add_qgroup_item(trans, quota_root,
1158                                               found_key.offset);
1159                         if (ret) {
1160                                 btrfs_abort_transaction(trans, ret);
1161                                 goto out_free_path;
1162                         }
1163
1164                         qgroup = add_qgroup_rb(fs_info, prealloc, found_key.offset);
1165                         prealloc = NULL;
1166                         if (IS_ERR(qgroup)) {
1167                                 ret = PTR_ERR(qgroup);
1168                                 btrfs_abort_transaction(trans, ret);
1169                                 goto out_free_path;
1170                         }
1171                         ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1172                         if (ret < 0) {
1173                                 btrfs_abort_transaction(trans, ret);
1174                                 goto out_free_path;
1175                         }
1176                         ret = btrfs_search_slot_for_read(tree_root, &found_key,
1177                                                          path, 1, 0);
1178                         if (ret < 0) {
1179                                 btrfs_abort_transaction(trans, ret);
1180                                 goto out_free_path;
1181                         }
1182                         if (ret > 0) {
1183                                 /*
1184                                  * Shouldn't happen, but in case it does we
1185                                  * don't need to do the btrfs_next_item, just
1186                                  * continue.
1187                                  */
1188                                 continue;
1189                         }
1190                 }
1191                 ret = btrfs_next_item(tree_root, path);
1192                 if (ret < 0) {
1193                         btrfs_abort_transaction(trans, ret);
1194                         goto out_free_path;
1195                 }
1196                 if (ret)
1197                         break;
1198         }
1199
1200 out_add_root:
1201         btrfs_release_path(path);
1202         ret = add_qgroup_item(trans, quota_root, BTRFS_FS_TREE_OBJECTID);
1203         if (ret) {
1204                 btrfs_abort_transaction(trans, ret);
1205                 goto out_free_path;
1206         }
1207
1208         ASSERT(prealloc == NULL);
1209         prealloc = kzalloc(sizeof(*prealloc), GFP_NOFS);
1210         if (!prealloc) {
1211                 ret = -ENOMEM;
1212                 goto out_free_path;
1213         }
1214         qgroup = add_qgroup_rb(fs_info, prealloc, BTRFS_FS_TREE_OBJECTID);
1215         prealloc = NULL;
1216         ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1217         if (ret < 0) {
1218                 btrfs_abort_transaction(trans, ret);
1219                 goto out_free_path;
1220         }
1221
1222         fs_info->qgroup_enable_gen = trans->transid;
1223
1224         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1225         /*
1226          * Commit the transaction while not holding qgroup_ioctl_lock, to avoid
1227          * a deadlock with tasks concurrently doing other qgroup operations, such
1228          * adding/removing qgroups or adding/deleting qgroup relations for example,
1229          * because all qgroup operations first start or join a transaction and then
1230          * lock the qgroup_ioctl_lock mutex.
1231          * We are safe from a concurrent task trying to enable quotas, by calling
1232          * this function, since we are serialized by fs_info->subvol_sem.
1233          */
1234         ret = btrfs_commit_transaction(trans);
1235         trans = NULL;
1236         mutex_lock(&fs_info->qgroup_ioctl_lock);
1237         if (ret)
1238                 goto out_free_path;
1239
1240         /*
1241          * Set quota enabled flag after committing the transaction, to avoid
1242          * deadlocks on fs_info->qgroup_ioctl_lock with concurrent snapshot
1243          * creation.
1244          */
1245         spin_lock(&fs_info->qgroup_lock);
1246         fs_info->quota_root = quota_root;
1247         set_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1248         if (simple)
1249                 btrfs_set_fs_incompat(fs_info, SIMPLE_QUOTA);
1250         spin_unlock(&fs_info->qgroup_lock);
1251
1252         /* Skip rescan for simple qgroups. */
1253         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE)
1254                 goto out_free_path;
1255
1256         ret = qgroup_rescan_init(fs_info, 0, 1);
1257         if (!ret) {
1258                 qgroup_rescan_zero_tracking(fs_info);
1259                 fs_info->qgroup_rescan_running = true;
1260                 btrfs_queue_work(fs_info->qgroup_rescan_workers,
1261                                  &fs_info->qgroup_rescan_work);
1262         } else {
1263                 /*
1264                  * We have set both BTRFS_FS_QUOTA_ENABLED and
1265                  * BTRFS_QGROUP_STATUS_FLAG_ON, so we can only fail with
1266                  * -EINPROGRESS. That can happen because someone started the
1267                  * rescan worker by calling quota rescan ioctl before we
1268                  * attempted to initialize the rescan worker. Failure due to
1269                  * quotas disabled in the meanwhile is not possible, because
1270                  * we are holding a write lock on fs_info->subvol_sem, which
1271                  * is also acquired when disabling quotas.
1272                  * Ignore such error, and any other error would need to undo
1273                  * everything we did in the transaction we just committed.
1274                  */
1275                 ASSERT(ret == -EINPROGRESS);
1276                 ret = 0;
1277         }
1278
1279 out_free_path:
1280         btrfs_free_path(path);
1281 out_free_root:
1282         if (ret)
1283                 btrfs_put_root(quota_root);
1284 out:
1285         if (ret) {
1286                 ulist_free(fs_info->qgroup_ulist);
1287                 fs_info->qgroup_ulist = NULL;
1288                 btrfs_sysfs_del_qgroups(fs_info);
1289         }
1290         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1291         if (ret && trans)
1292                 btrfs_end_transaction(trans);
1293         else if (trans)
1294                 ret = btrfs_end_transaction(trans);
1295         ulist_free(ulist);
1296         kfree(prealloc);
1297         return ret;
1298 }
1299
1300 /*
1301  * It is possible to have outstanding ordered extents which reserved bytes
1302  * before we disabled. We need to fully flush delalloc, ordered extents, and a
1303  * commit to ensure that we don't leak such reservations, only to have them
1304  * come back if we re-enable.
1305  *
1306  * - enable simple quotas
1307  * - reserve space
1308  * - release it, store rsv_bytes in OE
1309  * - disable quotas
1310  * - enable simple quotas (qgroup rsv are all 0)
1311  * - OE finishes
1312  * - run delayed refs
1313  * - free rsv_bytes, resulting in miscounting or even underflow
1314  */
1315 static int flush_reservations(struct btrfs_fs_info *fs_info)
1316 {
1317         struct btrfs_trans_handle *trans;
1318         int ret;
1319
1320         ret = btrfs_start_delalloc_roots(fs_info, LONG_MAX, false);
1321         if (ret)
1322                 return ret;
1323         btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
1324         trans = btrfs_join_transaction(fs_info->tree_root);
1325         if (IS_ERR(trans))
1326                 return PTR_ERR(trans);
1327         btrfs_commit_transaction(trans);
1328
1329         return ret;
1330 }
1331
1332 int btrfs_quota_disable(struct btrfs_fs_info *fs_info)
1333 {
1334         struct btrfs_root *quota_root;
1335         struct btrfs_trans_handle *trans = NULL;
1336         int ret = 0;
1337
1338         /*
1339          * We need to have subvol_sem write locked to prevent races with
1340          * snapshot creation.
1341          */
1342         lockdep_assert_held_write(&fs_info->subvol_sem);
1343
1344         /*
1345          * Lock the cleaner mutex to prevent races with concurrent relocation,
1346          * because relocation may be building backrefs for blocks of the quota
1347          * root while we are deleting the root. This is like dropping fs roots
1348          * of deleted snapshots/subvolumes, we need the same protection.
1349          *
1350          * This also prevents races between concurrent tasks trying to disable
1351          * quotas, because we will unlock and relock qgroup_ioctl_lock across
1352          * BTRFS_FS_QUOTA_ENABLED changes.
1353          */
1354         mutex_lock(&fs_info->cleaner_mutex);
1355
1356         mutex_lock(&fs_info->qgroup_ioctl_lock);
1357         if (!fs_info->quota_root)
1358                 goto out;
1359
1360         /*
1361          * Unlock the qgroup_ioctl_lock mutex before waiting for the rescan worker to
1362          * complete. Otherwise we can deadlock because btrfs_remove_qgroup() needs
1363          * to lock that mutex while holding a transaction handle and the rescan
1364          * worker needs to commit a transaction.
1365          */
1366         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1367
1368         /*
1369          * Request qgroup rescan worker to complete and wait for it. This wait
1370          * must be done before transaction start for quota disable since it may
1371          * deadlock with transaction by the qgroup rescan worker.
1372          */
1373         clear_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1374         btrfs_qgroup_wait_for_completion(fs_info, false);
1375
1376         ret = flush_reservations(fs_info);
1377         if (ret)
1378                 goto out_unlock_cleaner;
1379
1380         /*
1381          * 1 For the root item
1382          *
1383          * We should also reserve enough items for the quota tree deletion in
1384          * btrfs_clean_quota_tree but this is not done.
1385          *
1386          * Also, we must always start a transaction without holding the mutex
1387          * qgroup_ioctl_lock, see btrfs_quota_enable().
1388          */
1389         trans = btrfs_start_transaction(fs_info->tree_root, 1);
1390
1391         mutex_lock(&fs_info->qgroup_ioctl_lock);
1392         if (IS_ERR(trans)) {
1393                 ret = PTR_ERR(trans);
1394                 trans = NULL;
1395                 set_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1396                 goto out;
1397         }
1398
1399         if (!fs_info->quota_root)
1400                 goto out;
1401
1402         spin_lock(&fs_info->qgroup_lock);
1403         quota_root = fs_info->quota_root;
1404         fs_info->quota_root = NULL;
1405         fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_ON;
1406         fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_SIMPLE_MODE;
1407         fs_info->qgroup_drop_subtree_thres = BTRFS_MAX_LEVEL;
1408         spin_unlock(&fs_info->qgroup_lock);
1409
1410         btrfs_free_qgroup_config(fs_info);
1411
1412         ret = btrfs_clean_quota_tree(trans, quota_root);
1413         if (ret) {
1414                 btrfs_abort_transaction(trans, ret);
1415                 goto out;
1416         }
1417
1418         ret = btrfs_del_root(trans, &quota_root->root_key);
1419         if (ret) {
1420                 btrfs_abort_transaction(trans, ret);
1421                 goto out;
1422         }
1423
1424         spin_lock(&fs_info->trans_lock);
1425         list_del(&quota_root->dirty_list);
1426         spin_unlock(&fs_info->trans_lock);
1427
1428         btrfs_tree_lock(quota_root->node);
1429         btrfs_clear_buffer_dirty(trans, quota_root->node);
1430         btrfs_tree_unlock(quota_root->node);
1431         btrfs_free_tree_block(trans, btrfs_root_id(quota_root),
1432                               quota_root->node, 0, 1);
1433
1434         btrfs_put_root(quota_root);
1435
1436 out:
1437         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1438         if (ret && trans)
1439                 btrfs_end_transaction(trans);
1440         else if (trans)
1441                 ret = btrfs_commit_transaction(trans);
1442 out_unlock_cleaner:
1443         mutex_unlock(&fs_info->cleaner_mutex);
1444
1445         return ret;
1446 }
1447
1448 static void qgroup_dirty(struct btrfs_fs_info *fs_info,
1449                          struct btrfs_qgroup *qgroup)
1450 {
1451         if (list_empty(&qgroup->dirty))
1452                 list_add(&qgroup->dirty, &fs_info->dirty_qgroups);
1453 }
1454
1455 static void qgroup_iterator_add(struct list_head *head, struct btrfs_qgroup *qgroup)
1456 {
1457         if (!list_empty(&qgroup->iterator))
1458                 return;
1459
1460         list_add_tail(&qgroup->iterator, head);
1461 }
1462
1463 static void qgroup_iterator_clean(struct list_head *head)
1464 {
1465         while (!list_empty(head)) {
1466                 struct btrfs_qgroup *qgroup;
1467
1468                 qgroup = list_first_entry(head, struct btrfs_qgroup, iterator);
1469                 list_del_init(&qgroup->iterator);
1470         }
1471 }
1472
1473 /*
1474  * The easy accounting, we're updating qgroup relationship whose child qgroup
1475  * only has exclusive extents.
1476  *
1477  * In this case, all exclusive extents will also be exclusive for parent, so
1478  * excl/rfer just get added/removed.
1479  *
1480  * So is qgroup reservation space, which should also be added/removed to
1481  * parent.
1482  * Or when child tries to release reservation space, parent will underflow its
1483  * reservation (for relationship adding case).
1484  *
1485  * Caller should hold fs_info->qgroup_lock.
1486  */
1487 static int __qgroup_excl_accounting(struct btrfs_fs_info *fs_info, u64 ref_root,
1488                                     struct btrfs_qgroup *src, int sign)
1489 {
1490         struct btrfs_qgroup *qgroup;
1491         struct btrfs_qgroup *cur;
1492         LIST_HEAD(qgroup_list);
1493         u64 num_bytes = src->excl;
1494         int ret = 0;
1495
1496         qgroup = find_qgroup_rb(fs_info, ref_root);
1497         if (!qgroup)
1498                 goto out;
1499
1500         qgroup_iterator_add(&qgroup_list, qgroup);
1501         list_for_each_entry(cur, &qgroup_list, iterator) {
1502                 struct btrfs_qgroup_list *glist;
1503
1504                 qgroup->rfer += sign * num_bytes;
1505                 qgroup->rfer_cmpr += sign * num_bytes;
1506
1507                 WARN_ON(sign < 0 && qgroup->excl < num_bytes);
1508                 qgroup->excl += sign * num_bytes;
1509                 qgroup->excl_cmpr += sign * num_bytes;
1510
1511                 if (sign > 0)
1512                         qgroup_rsv_add_by_qgroup(fs_info, qgroup, src);
1513                 else
1514                         qgroup_rsv_release_by_qgroup(fs_info, qgroup, src);
1515                 qgroup_dirty(fs_info, qgroup);
1516
1517                 /* Append parent qgroups to @qgroup_list. */
1518                 list_for_each_entry(glist, &qgroup->groups, next_group)
1519                         qgroup_iterator_add(&qgroup_list, glist->group);
1520         }
1521         ret = 0;
1522 out:
1523         qgroup_iterator_clean(&qgroup_list);
1524         return ret;
1525 }
1526
1527
1528 /*
1529  * Quick path for updating qgroup with only excl refs.
1530  *
1531  * In that case, just update all parent will be enough.
1532  * Or we needs to do a full rescan.
1533  * Caller should also hold fs_info->qgroup_lock.
1534  *
1535  * Return 0 for quick update, return >0 for need to full rescan
1536  * and mark INCONSISTENT flag.
1537  * Return < 0 for other error.
1538  */
1539 static int quick_update_accounting(struct btrfs_fs_info *fs_info,
1540                                    u64 src, u64 dst, int sign)
1541 {
1542         struct btrfs_qgroup *qgroup;
1543         int ret = 1;
1544         int err = 0;
1545
1546         qgroup = find_qgroup_rb(fs_info, src);
1547         if (!qgroup)
1548                 goto out;
1549         if (qgroup->excl == qgroup->rfer) {
1550                 ret = 0;
1551                 err = __qgroup_excl_accounting(fs_info, dst, qgroup, sign);
1552                 if (err < 0) {
1553                         ret = err;
1554                         goto out;
1555                 }
1556         }
1557 out:
1558         if (ret)
1559                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1560         return ret;
1561 }
1562
1563 int btrfs_add_qgroup_relation(struct btrfs_trans_handle *trans, u64 src, u64 dst)
1564 {
1565         struct btrfs_fs_info *fs_info = trans->fs_info;
1566         struct btrfs_qgroup *parent;
1567         struct btrfs_qgroup *member;
1568         struct btrfs_qgroup_list *list;
1569         struct btrfs_qgroup_list *prealloc = NULL;
1570         int ret = 0;
1571
1572         /* Check the level of src and dst first */
1573         if (btrfs_qgroup_level(src) >= btrfs_qgroup_level(dst))
1574                 return -EINVAL;
1575
1576         mutex_lock(&fs_info->qgroup_ioctl_lock);
1577         if (!fs_info->quota_root) {
1578                 ret = -ENOTCONN;
1579                 goto out;
1580         }
1581         member = find_qgroup_rb(fs_info, src);
1582         parent = find_qgroup_rb(fs_info, dst);
1583         if (!member || !parent) {
1584                 ret = -EINVAL;
1585                 goto out;
1586         }
1587
1588         /* check if such qgroup relation exist firstly */
1589         list_for_each_entry(list, &member->groups, next_group) {
1590                 if (list->group == parent) {
1591                         ret = -EEXIST;
1592                         goto out;
1593                 }
1594         }
1595
1596         prealloc = kzalloc(sizeof(*list), GFP_NOFS);
1597         if (!prealloc) {
1598                 ret = -ENOMEM;
1599                 goto out;
1600         }
1601         ret = add_qgroup_relation_item(trans, src, dst);
1602         if (ret)
1603                 goto out;
1604
1605         ret = add_qgroup_relation_item(trans, dst, src);
1606         if (ret) {
1607                 del_qgroup_relation_item(trans, src, dst);
1608                 goto out;
1609         }
1610
1611         spin_lock(&fs_info->qgroup_lock);
1612         ret = __add_relation_rb(prealloc, member, parent);
1613         prealloc = NULL;
1614         if (ret < 0) {
1615                 spin_unlock(&fs_info->qgroup_lock);
1616                 goto out;
1617         }
1618         ret = quick_update_accounting(fs_info, src, dst, 1);
1619         spin_unlock(&fs_info->qgroup_lock);
1620 out:
1621         kfree(prealloc);
1622         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1623         return ret;
1624 }
1625
1626 static int __del_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1627                                  u64 dst)
1628 {
1629         struct btrfs_fs_info *fs_info = trans->fs_info;
1630         struct btrfs_qgroup *parent;
1631         struct btrfs_qgroup *member;
1632         struct btrfs_qgroup_list *list;
1633         bool found = false;
1634         int ret = 0;
1635         int ret2;
1636
1637         if (!fs_info->quota_root) {
1638                 ret = -ENOTCONN;
1639                 goto out;
1640         }
1641
1642         member = find_qgroup_rb(fs_info, src);
1643         parent = find_qgroup_rb(fs_info, dst);
1644         /*
1645          * The parent/member pair doesn't exist, then try to delete the dead
1646          * relation items only.
1647          */
1648         if (!member || !parent)
1649                 goto delete_item;
1650
1651         /* check if such qgroup relation exist firstly */
1652         list_for_each_entry(list, &member->groups, next_group) {
1653                 if (list->group == parent) {
1654                         found = true;
1655                         break;
1656                 }
1657         }
1658
1659 delete_item:
1660         ret = del_qgroup_relation_item(trans, src, dst);
1661         if (ret < 0 && ret != -ENOENT)
1662                 goto out;
1663         ret2 = del_qgroup_relation_item(trans, dst, src);
1664         if (ret2 < 0 && ret2 != -ENOENT)
1665                 goto out;
1666
1667         /* At least one deletion succeeded, return 0 */
1668         if (!ret || !ret2)
1669                 ret = 0;
1670
1671         if (found) {
1672                 spin_lock(&fs_info->qgroup_lock);
1673                 del_relation_rb(fs_info, src, dst);
1674                 ret = quick_update_accounting(fs_info, src, dst, -1);
1675                 spin_unlock(&fs_info->qgroup_lock);
1676         }
1677 out:
1678         return ret;
1679 }
1680
1681 int btrfs_del_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1682                               u64 dst)
1683 {
1684         struct btrfs_fs_info *fs_info = trans->fs_info;
1685         int ret = 0;
1686
1687         mutex_lock(&fs_info->qgroup_ioctl_lock);
1688         ret = __del_qgroup_relation(trans, src, dst);
1689         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1690
1691         return ret;
1692 }
1693
1694 int btrfs_create_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid)
1695 {
1696         struct btrfs_fs_info *fs_info = trans->fs_info;
1697         struct btrfs_root *quota_root;
1698         struct btrfs_qgroup *qgroup;
1699         struct btrfs_qgroup *prealloc = NULL;
1700         int ret = 0;
1701
1702         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_DISABLED)
1703                 return 0;
1704
1705         mutex_lock(&fs_info->qgroup_ioctl_lock);
1706         if (!fs_info->quota_root) {
1707                 ret = -ENOTCONN;
1708                 goto out;
1709         }
1710         quota_root = fs_info->quota_root;
1711         qgroup = find_qgroup_rb(fs_info, qgroupid);
1712         if (qgroup) {
1713                 ret = -EEXIST;
1714                 goto out;
1715         }
1716
1717         prealloc = kzalloc(sizeof(*prealloc), GFP_NOFS);
1718         if (!prealloc) {
1719                 ret = -ENOMEM;
1720                 goto out;
1721         }
1722
1723         ret = add_qgroup_item(trans, quota_root, qgroupid);
1724         if (ret)
1725                 goto out;
1726
1727         spin_lock(&fs_info->qgroup_lock);
1728         qgroup = add_qgroup_rb(fs_info, prealloc, qgroupid);
1729         spin_unlock(&fs_info->qgroup_lock);
1730         prealloc = NULL;
1731
1732         ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1733 out:
1734         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1735         kfree(prealloc);
1736         return ret;
1737 }
1738
1739 int btrfs_remove_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid)
1740 {
1741         struct btrfs_fs_info *fs_info = trans->fs_info;
1742         struct btrfs_qgroup *qgroup;
1743         struct btrfs_qgroup_list *list;
1744         int ret = 0;
1745
1746         mutex_lock(&fs_info->qgroup_ioctl_lock);
1747         if (!fs_info->quota_root) {
1748                 ret = -ENOTCONN;
1749                 goto out;
1750         }
1751
1752         qgroup = find_qgroup_rb(fs_info, qgroupid);
1753         if (!qgroup) {
1754                 ret = -ENOENT;
1755                 goto out;
1756         }
1757
1758         /* Check if there are no children of this qgroup */
1759         if (!list_empty(&qgroup->members)) {
1760                 ret = -EBUSY;
1761                 goto out;
1762         }
1763
1764         ret = del_qgroup_item(trans, qgroupid);
1765         if (ret && ret != -ENOENT)
1766                 goto out;
1767
1768         while (!list_empty(&qgroup->groups)) {
1769                 list = list_first_entry(&qgroup->groups,
1770                                         struct btrfs_qgroup_list, next_group);
1771                 ret = __del_qgroup_relation(trans, qgroupid,
1772                                             list->group->qgroupid);
1773                 if (ret)
1774                         goto out;
1775         }
1776
1777         spin_lock(&fs_info->qgroup_lock);
1778         del_qgroup_rb(fs_info, qgroupid);
1779         spin_unlock(&fs_info->qgroup_lock);
1780
1781         /*
1782          * Remove the qgroup from sysfs now without holding the qgroup_lock
1783          * spinlock, since the sysfs_remove_group() function needs to take
1784          * the mutex kernfs_mutex through kernfs_remove_by_name_ns().
1785          */
1786         btrfs_sysfs_del_one_qgroup(fs_info, qgroup);
1787         kfree(qgroup);
1788 out:
1789         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1790         return ret;
1791 }
1792
1793 int btrfs_limit_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid,
1794                        struct btrfs_qgroup_limit *limit)
1795 {
1796         struct btrfs_fs_info *fs_info = trans->fs_info;
1797         struct btrfs_qgroup *qgroup;
1798         int ret = 0;
1799         /* Sometimes we would want to clear the limit on this qgroup.
1800          * To meet this requirement, we treat the -1 as a special value
1801          * which tell kernel to clear the limit on this qgroup.
1802          */
1803         const u64 CLEAR_VALUE = -1;
1804
1805         mutex_lock(&fs_info->qgroup_ioctl_lock);
1806         if (!fs_info->quota_root) {
1807                 ret = -ENOTCONN;
1808                 goto out;
1809         }
1810
1811         qgroup = find_qgroup_rb(fs_info, qgroupid);
1812         if (!qgroup) {
1813                 ret = -ENOENT;
1814                 goto out;
1815         }
1816
1817         spin_lock(&fs_info->qgroup_lock);
1818         if (limit->flags & BTRFS_QGROUP_LIMIT_MAX_RFER) {
1819                 if (limit->max_rfer == CLEAR_VALUE) {
1820                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_MAX_RFER;
1821                         limit->flags &= ~BTRFS_QGROUP_LIMIT_MAX_RFER;
1822                         qgroup->max_rfer = 0;
1823                 } else {
1824                         qgroup->max_rfer = limit->max_rfer;
1825                 }
1826         }
1827         if (limit->flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) {
1828                 if (limit->max_excl == CLEAR_VALUE) {
1829                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_MAX_EXCL;
1830                         limit->flags &= ~BTRFS_QGROUP_LIMIT_MAX_EXCL;
1831                         qgroup->max_excl = 0;
1832                 } else {
1833                         qgroup->max_excl = limit->max_excl;
1834                 }
1835         }
1836         if (limit->flags & BTRFS_QGROUP_LIMIT_RSV_RFER) {
1837                 if (limit->rsv_rfer == CLEAR_VALUE) {
1838                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_RSV_RFER;
1839                         limit->flags &= ~BTRFS_QGROUP_LIMIT_RSV_RFER;
1840                         qgroup->rsv_rfer = 0;
1841                 } else {
1842                         qgroup->rsv_rfer = limit->rsv_rfer;
1843                 }
1844         }
1845         if (limit->flags & BTRFS_QGROUP_LIMIT_RSV_EXCL) {
1846                 if (limit->rsv_excl == CLEAR_VALUE) {
1847                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_RSV_EXCL;
1848                         limit->flags &= ~BTRFS_QGROUP_LIMIT_RSV_EXCL;
1849                         qgroup->rsv_excl = 0;
1850                 } else {
1851                         qgroup->rsv_excl = limit->rsv_excl;
1852                 }
1853         }
1854         qgroup->lim_flags |= limit->flags;
1855
1856         spin_unlock(&fs_info->qgroup_lock);
1857
1858         ret = update_qgroup_limit_item(trans, qgroup);
1859         if (ret) {
1860                 qgroup_mark_inconsistent(fs_info);
1861                 btrfs_info(fs_info, "unable to update quota limit for %llu",
1862                        qgroupid);
1863         }
1864
1865 out:
1866         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1867         return ret;
1868 }
1869
1870 /*
1871  * Inform qgroup to trace one dirty extent, its info is recorded in @record.
1872  * So qgroup can account it at transaction committing time.
1873  *
1874  * No lock version, caller must acquire delayed ref lock and allocated memory,
1875  * then call btrfs_qgroup_trace_extent_post() after exiting lock context.
1876  *
1877  * Return 0 for success insert
1878  * Return >0 for existing record, caller can free @record safely.
1879  * Error is not possible
1880  */
1881 int btrfs_qgroup_trace_extent_nolock(struct btrfs_fs_info *fs_info,
1882                                 struct btrfs_delayed_ref_root *delayed_refs,
1883                                 struct btrfs_qgroup_extent_record *record)
1884 {
1885         struct rb_node **p = &delayed_refs->dirty_extent_root.rb_node;
1886         struct rb_node *parent_node = NULL;
1887         struct btrfs_qgroup_extent_record *entry;
1888         u64 bytenr = record->bytenr;
1889
1890         if (!btrfs_qgroup_full_accounting(fs_info))
1891                 return 1;
1892
1893         lockdep_assert_held(&delayed_refs->lock);
1894         trace_btrfs_qgroup_trace_extent(fs_info, record);
1895
1896         while (*p) {
1897                 parent_node = *p;
1898                 entry = rb_entry(parent_node, struct btrfs_qgroup_extent_record,
1899                                  node);
1900                 if (bytenr < entry->bytenr) {
1901                         p = &(*p)->rb_left;
1902                 } else if (bytenr > entry->bytenr) {
1903                         p = &(*p)->rb_right;
1904                 } else {
1905                         if (record->data_rsv && !entry->data_rsv) {
1906                                 entry->data_rsv = record->data_rsv;
1907                                 entry->data_rsv_refroot =
1908                                         record->data_rsv_refroot;
1909                         }
1910                         return 1;
1911                 }
1912         }
1913
1914         rb_link_node(&record->node, parent_node, p);
1915         rb_insert_color(&record->node, &delayed_refs->dirty_extent_root);
1916         return 0;
1917 }
1918
1919 /*
1920  * Post handler after qgroup_trace_extent_nolock().
1921  *
1922  * NOTE: Current qgroup does the expensive backref walk at transaction
1923  * committing time with TRANS_STATE_COMMIT_DOING, this blocks incoming
1924  * new transaction.
1925  * This is designed to allow btrfs_find_all_roots() to get correct new_roots
1926  * result.
1927  *
1928  * However for old_roots there is no need to do backref walk at that time,
1929  * since we search commit roots to walk backref and result will always be
1930  * correct.
1931  *
1932  * Due to the nature of no lock version, we can't do backref there.
1933  * So we must call btrfs_qgroup_trace_extent_post() after exiting
1934  * spinlock context.
1935  *
1936  * TODO: If we can fix and prove btrfs_find_all_roots() can get correct result
1937  * using current root, then we can move all expensive backref walk out of
1938  * transaction committing, but not now as qgroup accounting will be wrong again.
1939  */
1940 int btrfs_qgroup_trace_extent_post(struct btrfs_trans_handle *trans,
1941                                    struct btrfs_qgroup_extent_record *qrecord)
1942 {
1943         struct btrfs_backref_walk_ctx ctx = { 0 };
1944         int ret;
1945
1946         if (!btrfs_qgroup_full_accounting(trans->fs_info))
1947                 return 0;
1948         /*
1949          * We are always called in a context where we are already holding a
1950          * transaction handle. Often we are called when adding a data delayed
1951          * reference from btrfs_truncate_inode_items() (truncating or unlinking),
1952          * in which case we will be holding a write lock on extent buffer from a
1953          * subvolume tree. In this case we can't allow btrfs_find_all_roots() to
1954          * acquire fs_info->commit_root_sem, because that is a higher level lock
1955          * that must be acquired before locking any extent buffers.
1956          *
1957          * So we want btrfs_find_all_roots() to not acquire the commit_root_sem
1958          * but we can't pass it a non-NULL transaction handle, because otherwise
1959          * it would not use commit roots and would lock extent buffers, causing
1960          * a deadlock if it ends up trying to read lock the same extent buffer
1961          * that was previously write locked at btrfs_truncate_inode_items().
1962          *
1963          * So pass a NULL transaction handle to btrfs_find_all_roots() and
1964          * explicitly tell it to not acquire the commit_root_sem - if we are
1965          * holding a transaction handle we don't need its protection.
1966          */
1967         ASSERT(trans != NULL);
1968
1969         if (trans->fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING)
1970                 return 0;
1971
1972         ctx.bytenr = qrecord->bytenr;
1973         ctx.fs_info = trans->fs_info;
1974
1975         ret = btrfs_find_all_roots(&ctx, true);
1976         if (ret < 0) {
1977                 qgroup_mark_inconsistent(trans->fs_info);
1978                 btrfs_warn(trans->fs_info,
1979 "error accounting new delayed refs extent (err code: %d), quota inconsistent",
1980                         ret);
1981                 return 0;
1982         }
1983
1984         /*
1985          * Here we don't need to get the lock of
1986          * trans->transaction->delayed_refs, since inserted qrecord won't
1987          * be deleted, only qrecord->node may be modified (new qrecord insert)
1988          *
1989          * So modifying qrecord->old_roots is safe here
1990          */
1991         qrecord->old_roots = ctx.roots;
1992         return 0;
1993 }
1994
1995 /*
1996  * Inform qgroup to trace one dirty extent, specified by @bytenr and
1997  * @num_bytes.
1998  * So qgroup can account it at commit trans time.
1999  *
2000  * Better encapsulated version, with memory allocation and backref walk for
2001  * commit roots.
2002  * So this can sleep.
2003  *
2004  * Return 0 if the operation is done.
2005  * Return <0 for error, like memory allocation failure or invalid parameter
2006  * (NULL trans)
2007  */
2008 int btrfs_qgroup_trace_extent(struct btrfs_trans_handle *trans, u64 bytenr,
2009                               u64 num_bytes)
2010 {
2011         struct btrfs_fs_info *fs_info = trans->fs_info;
2012         struct btrfs_qgroup_extent_record *record;
2013         struct btrfs_delayed_ref_root *delayed_refs;
2014         int ret;
2015
2016         if (!btrfs_qgroup_full_accounting(fs_info) || bytenr == 0 || num_bytes == 0)
2017                 return 0;
2018         record = kzalloc(sizeof(*record), GFP_NOFS);
2019         if (!record)
2020                 return -ENOMEM;
2021
2022         delayed_refs = &trans->transaction->delayed_refs;
2023         record->bytenr = bytenr;
2024         record->num_bytes = num_bytes;
2025         record->old_roots = NULL;
2026
2027         spin_lock(&delayed_refs->lock);
2028         ret = btrfs_qgroup_trace_extent_nolock(fs_info, delayed_refs, record);
2029         spin_unlock(&delayed_refs->lock);
2030         if (ret > 0) {
2031                 kfree(record);
2032                 return 0;
2033         }
2034         return btrfs_qgroup_trace_extent_post(trans, record);
2035 }
2036
2037 /*
2038  * Inform qgroup to trace all leaf items of data
2039  *
2040  * Return 0 for success
2041  * Return <0 for error(ENOMEM)
2042  */
2043 int btrfs_qgroup_trace_leaf_items(struct btrfs_trans_handle *trans,
2044                                   struct extent_buffer *eb)
2045 {
2046         struct btrfs_fs_info *fs_info = trans->fs_info;
2047         int nr = btrfs_header_nritems(eb);
2048         int i, extent_type, ret;
2049         struct btrfs_key key;
2050         struct btrfs_file_extent_item *fi;
2051         u64 bytenr, num_bytes;
2052
2053         /* We can be called directly from walk_up_proc() */
2054         if (!btrfs_qgroup_full_accounting(fs_info))
2055                 return 0;
2056
2057         for (i = 0; i < nr; i++) {
2058                 btrfs_item_key_to_cpu(eb, &key, i);
2059
2060                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2061                         continue;
2062
2063                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
2064                 /* filter out non qgroup-accountable extents  */
2065                 extent_type = btrfs_file_extent_type(eb, fi);
2066
2067                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
2068                         continue;
2069
2070                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
2071                 if (!bytenr)
2072                         continue;
2073
2074                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
2075
2076                 ret = btrfs_qgroup_trace_extent(trans, bytenr, num_bytes);
2077                 if (ret)
2078                         return ret;
2079         }
2080         cond_resched();
2081         return 0;
2082 }
2083
2084 /*
2085  * Walk up the tree from the bottom, freeing leaves and any interior
2086  * nodes which have had all slots visited. If a node (leaf or
2087  * interior) is freed, the node above it will have it's slot
2088  * incremented. The root node will never be freed.
2089  *
2090  * At the end of this function, we should have a path which has all
2091  * slots incremented to the next position for a search. If we need to
2092  * read a new node it will be NULL and the node above it will have the
2093  * correct slot selected for a later read.
2094  *
2095  * If we increment the root nodes slot counter past the number of
2096  * elements, 1 is returned to signal completion of the search.
2097  */
2098 static int adjust_slots_upwards(struct btrfs_path *path, int root_level)
2099 {
2100         int level = 0;
2101         int nr, slot;
2102         struct extent_buffer *eb;
2103
2104         if (root_level == 0)
2105                 return 1;
2106
2107         while (level <= root_level) {
2108                 eb = path->nodes[level];
2109                 nr = btrfs_header_nritems(eb);
2110                 path->slots[level]++;
2111                 slot = path->slots[level];
2112                 if (slot >= nr || level == 0) {
2113                         /*
2114                          * Don't free the root -  we will detect this
2115                          * condition after our loop and return a
2116                          * positive value for caller to stop walking the tree.
2117                          */
2118                         if (level != root_level) {
2119                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
2120                                 path->locks[level] = 0;
2121
2122                                 free_extent_buffer(eb);
2123                                 path->nodes[level] = NULL;
2124                                 path->slots[level] = 0;
2125                         }
2126                 } else {
2127                         /*
2128                          * We have a valid slot to walk back down
2129                          * from. Stop here so caller can process these
2130                          * new nodes.
2131                          */
2132                         break;
2133                 }
2134
2135                 level++;
2136         }
2137
2138         eb = path->nodes[root_level];
2139         if (path->slots[root_level] >= btrfs_header_nritems(eb))
2140                 return 1;
2141
2142         return 0;
2143 }
2144
2145 /*
2146  * Helper function to trace a subtree tree block swap.
2147  *
2148  * The swap will happen in highest tree block, but there may be a lot of
2149  * tree blocks involved.
2150  *
2151  * For example:
2152  *  OO = Old tree blocks
2153  *  NN = New tree blocks allocated during balance
2154  *
2155  *           File tree (257)                  Reloc tree for 257
2156  * L2              OO                                NN
2157  *               /    \                            /    \
2158  * L1          OO      OO (a)                    OO      NN (a)
2159  *            / \     / \                       / \     / \
2160  * L0       OO   OO OO   OO                   OO   OO NN   NN
2161  *                  (b)  (c)                          (b)  (c)
2162  *
2163  * When calling qgroup_trace_extent_swap(), we will pass:
2164  * @src_eb = OO(a)
2165  * @dst_path = [ nodes[1] = NN(a), nodes[0] = NN(c) ]
2166  * @dst_level = 0
2167  * @root_level = 1
2168  *
2169  * In that case, qgroup_trace_extent_swap() will search from OO(a) to
2170  * reach OO(c), then mark both OO(c) and NN(c) as qgroup dirty.
2171  *
2172  * The main work of qgroup_trace_extent_swap() can be split into 3 parts:
2173  *
2174  * 1) Tree search from @src_eb
2175  *    It should acts as a simplified btrfs_search_slot().
2176  *    The key for search can be extracted from @dst_path->nodes[dst_level]
2177  *    (first key).
2178  *
2179  * 2) Mark the final tree blocks in @src_path and @dst_path qgroup dirty
2180  *    NOTE: In above case, OO(a) and NN(a) won't be marked qgroup dirty.
2181  *    They should be marked during previous (@dst_level = 1) iteration.
2182  *
2183  * 3) Mark file extents in leaves dirty
2184  *    We don't have good way to pick out new file extents only.
2185  *    So we still follow the old method by scanning all file extents in
2186  *    the leave.
2187  *
2188  * This function can free us from keeping two paths, thus later we only need
2189  * to care about how to iterate all new tree blocks in reloc tree.
2190  */
2191 static int qgroup_trace_extent_swap(struct btrfs_trans_handle* trans,
2192                                     struct extent_buffer *src_eb,
2193                                     struct btrfs_path *dst_path,
2194                                     int dst_level, int root_level,
2195                                     bool trace_leaf)
2196 {
2197         struct btrfs_key key;
2198         struct btrfs_path *src_path;
2199         struct btrfs_fs_info *fs_info = trans->fs_info;
2200         u32 nodesize = fs_info->nodesize;
2201         int cur_level = root_level;
2202         int ret;
2203
2204         BUG_ON(dst_level > root_level);
2205         /* Level mismatch */
2206         if (btrfs_header_level(src_eb) != root_level)
2207                 return -EINVAL;
2208
2209         src_path = btrfs_alloc_path();
2210         if (!src_path) {
2211                 ret = -ENOMEM;
2212                 goto out;
2213         }
2214
2215         if (dst_level)
2216                 btrfs_node_key_to_cpu(dst_path->nodes[dst_level], &key, 0);
2217         else
2218                 btrfs_item_key_to_cpu(dst_path->nodes[dst_level], &key, 0);
2219
2220         /* For src_path */
2221         atomic_inc(&src_eb->refs);
2222         src_path->nodes[root_level] = src_eb;
2223         src_path->slots[root_level] = dst_path->slots[root_level];
2224         src_path->locks[root_level] = 0;
2225
2226         /* A simplified version of btrfs_search_slot() */
2227         while (cur_level >= dst_level) {
2228                 struct btrfs_key src_key;
2229                 struct btrfs_key dst_key;
2230
2231                 if (src_path->nodes[cur_level] == NULL) {
2232                         struct extent_buffer *eb;
2233                         int parent_slot;
2234
2235                         eb = src_path->nodes[cur_level + 1];
2236                         parent_slot = src_path->slots[cur_level + 1];
2237
2238                         eb = btrfs_read_node_slot(eb, parent_slot);
2239                         if (IS_ERR(eb)) {
2240                                 ret = PTR_ERR(eb);
2241                                 goto out;
2242                         }
2243
2244                         src_path->nodes[cur_level] = eb;
2245
2246                         btrfs_tree_read_lock(eb);
2247                         src_path->locks[cur_level] = BTRFS_READ_LOCK;
2248                 }
2249
2250                 src_path->slots[cur_level] = dst_path->slots[cur_level];
2251                 if (cur_level) {
2252                         btrfs_node_key_to_cpu(dst_path->nodes[cur_level],
2253                                         &dst_key, dst_path->slots[cur_level]);
2254                         btrfs_node_key_to_cpu(src_path->nodes[cur_level],
2255                                         &src_key, src_path->slots[cur_level]);
2256                 } else {
2257                         btrfs_item_key_to_cpu(dst_path->nodes[cur_level],
2258                                         &dst_key, dst_path->slots[cur_level]);
2259                         btrfs_item_key_to_cpu(src_path->nodes[cur_level],
2260                                         &src_key, src_path->slots[cur_level]);
2261                 }
2262                 /* Content mismatch, something went wrong */
2263                 if (btrfs_comp_cpu_keys(&dst_key, &src_key)) {
2264                         ret = -ENOENT;
2265                         goto out;
2266                 }
2267                 cur_level--;
2268         }
2269
2270         /*
2271          * Now both @dst_path and @src_path have been populated, record the tree
2272          * blocks for qgroup accounting.
2273          */
2274         ret = btrfs_qgroup_trace_extent(trans, src_path->nodes[dst_level]->start,
2275                                         nodesize);
2276         if (ret < 0)
2277                 goto out;
2278         ret = btrfs_qgroup_trace_extent(trans, dst_path->nodes[dst_level]->start,
2279                                         nodesize);
2280         if (ret < 0)
2281                 goto out;
2282
2283         /* Record leaf file extents */
2284         if (dst_level == 0 && trace_leaf) {
2285                 ret = btrfs_qgroup_trace_leaf_items(trans, src_path->nodes[0]);
2286                 if (ret < 0)
2287                         goto out;
2288                 ret = btrfs_qgroup_trace_leaf_items(trans, dst_path->nodes[0]);
2289         }
2290 out:
2291         btrfs_free_path(src_path);
2292         return ret;
2293 }
2294
2295 /*
2296  * Helper function to do recursive generation-aware depth-first search, to
2297  * locate all new tree blocks in a subtree of reloc tree.
2298  *
2299  * E.g. (OO = Old tree blocks, NN = New tree blocks, whose gen == last_snapshot)
2300  *         reloc tree
2301  * L2         NN (a)
2302  *          /    \
2303  * L1    OO        NN (b)
2304  *      /  \      /  \
2305  * L0  OO  OO    OO  NN
2306  *               (c) (d)
2307  * If we pass:
2308  * @dst_path = [ nodes[1] = NN(b), nodes[0] = NULL ],
2309  * @cur_level = 1
2310  * @root_level = 1
2311  *
2312  * We will iterate through tree blocks NN(b), NN(d) and info qgroup to trace
2313  * above tree blocks along with their counter parts in file tree.
2314  * While during search, old tree blocks OO(c) will be skipped as tree block swap
2315  * won't affect OO(c).
2316  */
2317 static int qgroup_trace_new_subtree_blocks(struct btrfs_trans_handle* trans,
2318                                            struct extent_buffer *src_eb,
2319                                            struct btrfs_path *dst_path,
2320                                            int cur_level, int root_level,
2321                                            u64 last_snapshot, bool trace_leaf)
2322 {
2323         struct btrfs_fs_info *fs_info = trans->fs_info;
2324         struct extent_buffer *eb;
2325         bool need_cleanup = false;
2326         int ret = 0;
2327         int i;
2328
2329         /* Level sanity check */
2330         if (cur_level < 0 || cur_level >= BTRFS_MAX_LEVEL - 1 ||
2331             root_level < 0 || root_level >= BTRFS_MAX_LEVEL - 1 ||
2332             root_level < cur_level) {
2333                 btrfs_err_rl(fs_info,
2334                         "%s: bad levels, cur_level=%d root_level=%d",
2335                         __func__, cur_level, root_level);
2336                 return -EUCLEAN;
2337         }
2338
2339         /* Read the tree block if needed */
2340         if (dst_path->nodes[cur_level] == NULL) {
2341                 int parent_slot;
2342                 u64 child_gen;
2343
2344                 /*
2345                  * dst_path->nodes[root_level] must be initialized before
2346                  * calling this function.
2347                  */
2348                 if (cur_level == root_level) {
2349                         btrfs_err_rl(fs_info,
2350         "%s: dst_path->nodes[%d] not initialized, root_level=%d cur_level=%d",
2351                                 __func__, root_level, root_level, cur_level);
2352                         return -EUCLEAN;
2353                 }
2354
2355                 /*
2356                  * We need to get child blockptr/gen from parent before we can
2357                  * read it.
2358                   */
2359                 eb = dst_path->nodes[cur_level + 1];
2360                 parent_slot = dst_path->slots[cur_level + 1];
2361                 child_gen = btrfs_node_ptr_generation(eb, parent_slot);
2362
2363                 /* This node is old, no need to trace */
2364                 if (child_gen < last_snapshot)
2365                         goto out;
2366
2367                 eb = btrfs_read_node_slot(eb, parent_slot);
2368                 if (IS_ERR(eb)) {
2369                         ret = PTR_ERR(eb);
2370                         goto out;
2371                 }
2372
2373                 dst_path->nodes[cur_level] = eb;
2374                 dst_path->slots[cur_level] = 0;
2375
2376                 btrfs_tree_read_lock(eb);
2377                 dst_path->locks[cur_level] = BTRFS_READ_LOCK;
2378                 need_cleanup = true;
2379         }
2380
2381         /* Now record this tree block and its counter part for qgroups */
2382         ret = qgroup_trace_extent_swap(trans, src_eb, dst_path, cur_level,
2383                                        root_level, trace_leaf);
2384         if (ret < 0)
2385                 goto cleanup;
2386
2387         eb = dst_path->nodes[cur_level];
2388
2389         if (cur_level > 0) {
2390                 /* Iterate all child tree blocks */
2391                 for (i = 0; i < btrfs_header_nritems(eb); i++) {
2392                         /* Skip old tree blocks as they won't be swapped */
2393                         if (btrfs_node_ptr_generation(eb, i) < last_snapshot)
2394                                 continue;
2395                         dst_path->slots[cur_level] = i;
2396
2397                         /* Recursive call (at most 7 times) */
2398                         ret = qgroup_trace_new_subtree_blocks(trans, src_eb,
2399                                         dst_path, cur_level - 1, root_level,
2400                                         last_snapshot, trace_leaf);
2401                         if (ret < 0)
2402                                 goto cleanup;
2403                 }
2404         }
2405
2406 cleanup:
2407         if (need_cleanup) {
2408                 /* Clean up */
2409                 btrfs_tree_unlock_rw(dst_path->nodes[cur_level],
2410                                      dst_path->locks[cur_level]);
2411                 free_extent_buffer(dst_path->nodes[cur_level]);
2412                 dst_path->nodes[cur_level] = NULL;
2413                 dst_path->slots[cur_level] = 0;
2414                 dst_path->locks[cur_level] = 0;
2415         }
2416 out:
2417         return ret;
2418 }
2419
2420 static int qgroup_trace_subtree_swap(struct btrfs_trans_handle *trans,
2421                                 struct extent_buffer *src_eb,
2422                                 struct extent_buffer *dst_eb,
2423                                 u64 last_snapshot, bool trace_leaf)
2424 {
2425         struct btrfs_fs_info *fs_info = trans->fs_info;
2426         struct btrfs_path *dst_path = NULL;
2427         int level;
2428         int ret;
2429
2430         if (!btrfs_qgroup_full_accounting(fs_info))
2431                 return 0;
2432
2433         /* Wrong parameter order */
2434         if (btrfs_header_generation(src_eb) > btrfs_header_generation(dst_eb)) {
2435                 btrfs_err_rl(fs_info,
2436                 "%s: bad parameter order, src_gen=%llu dst_gen=%llu", __func__,
2437                              btrfs_header_generation(src_eb),
2438                              btrfs_header_generation(dst_eb));
2439                 return -EUCLEAN;
2440         }
2441
2442         if (!extent_buffer_uptodate(src_eb) || !extent_buffer_uptodate(dst_eb)) {
2443                 ret = -EIO;
2444                 goto out;
2445         }
2446
2447         level = btrfs_header_level(dst_eb);
2448         dst_path = btrfs_alloc_path();
2449         if (!dst_path) {
2450                 ret = -ENOMEM;
2451                 goto out;
2452         }
2453         /* For dst_path */
2454         atomic_inc(&dst_eb->refs);
2455         dst_path->nodes[level] = dst_eb;
2456         dst_path->slots[level] = 0;
2457         dst_path->locks[level] = 0;
2458
2459         /* Do the generation aware breadth-first search */
2460         ret = qgroup_trace_new_subtree_blocks(trans, src_eb, dst_path, level,
2461                                               level, last_snapshot, trace_leaf);
2462         if (ret < 0)
2463                 goto out;
2464         ret = 0;
2465
2466 out:
2467         btrfs_free_path(dst_path);
2468         if (ret < 0)
2469                 qgroup_mark_inconsistent(fs_info);
2470         return ret;
2471 }
2472
2473 /*
2474  * Inform qgroup to trace a whole subtree, including all its child tree
2475  * blocks and data.
2476  * The root tree block is specified by @root_eb.
2477  *
2478  * Normally used by relocation(tree block swap) and subvolume deletion.
2479  *
2480  * Return 0 for success
2481  * Return <0 for error(ENOMEM or tree search error)
2482  */
2483 int btrfs_qgroup_trace_subtree(struct btrfs_trans_handle *trans,
2484                                struct extent_buffer *root_eb,
2485                                u64 root_gen, int root_level)
2486 {
2487         struct btrfs_fs_info *fs_info = trans->fs_info;
2488         int ret = 0;
2489         int level;
2490         u8 drop_subptree_thres;
2491         struct extent_buffer *eb = root_eb;
2492         struct btrfs_path *path = NULL;
2493
2494         BUG_ON(root_level < 0 || root_level >= BTRFS_MAX_LEVEL);
2495         BUG_ON(root_eb == NULL);
2496
2497         if (!btrfs_qgroup_full_accounting(fs_info))
2498                 return 0;
2499
2500         spin_lock(&fs_info->qgroup_lock);
2501         drop_subptree_thres = fs_info->qgroup_drop_subtree_thres;
2502         spin_unlock(&fs_info->qgroup_lock);
2503
2504         /*
2505          * This function only gets called for snapshot drop, if we hit a high
2506          * node here, it means we are going to change ownership for quite a lot
2507          * of extents, which will greatly slow down btrfs_commit_transaction().
2508          *
2509          * So here if we find a high tree here, we just skip the accounting and
2510          * mark qgroup inconsistent.
2511          */
2512         if (root_level >= drop_subptree_thres) {
2513                 qgroup_mark_inconsistent(fs_info);
2514                 return 0;
2515         }
2516
2517         if (!extent_buffer_uptodate(root_eb)) {
2518                 struct btrfs_tree_parent_check check = {
2519                         .has_first_key = false,
2520                         .transid = root_gen,
2521                         .level = root_level
2522                 };
2523
2524                 ret = btrfs_read_extent_buffer(root_eb, &check);
2525                 if (ret)
2526                         goto out;
2527         }
2528
2529         if (root_level == 0) {
2530                 ret = btrfs_qgroup_trace_leaf_items(trans, root_eb);
2531                 goto out;
2532         }
2533
2534         path = btrfs_alloc_path();
2535         if (!path)
2536                 return -ENOMEM;
2537
2538         /*
2539          * Walk down the tree.  Missing extent blocks are filled in as
2540          * we go. Metadata is accounted every time we read a new
2541          * extent block.
2542          *
2543          * When we reach a leaf, we account for file extent items in it,
2544          * walk back up the tree (adjusting slot pointers as we go)
2545          * and restart the search process.
2546          */
2547         atomic_inc(&root_eb->refs);     /* For path */
2548         path->nodes[root_level] = root_eb;
2549         path->slots[root_level] = 0;
2550         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
2551 walk_down:
2552         level = root_level;
2553         while (level >= 0) {
2554                 if (path->nodes[level] == NULL) {
2555                         int parent_slot;
2556                         u64 child_bytenr;
2557
2558                         /*
2559                          * We need to get child blockptr from parent before we
2560                          * can read it.
2561                           */
2562                         eb = path->nodes[level + 1];
2563                         parent_slot = path->slots[level + 1];
2564                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
2565
2566                         eb = btrfs_read_node_slot(eb, parent_slot);
2567                         if (IS_ERR(eb)) {
2568                                 ret = PTR_ERR(eb);
2569                                 goto out;
2570                         }
2571
2572                         path->nodes[level] = eb;
2573                         path->slots[level] = 0;
2574
2575                         btrfs_tree_read_lock(eb);
2576                         path->locks[level] = BTRFS_READ_LOCK;
2577
2578                         ret = btrfs_qgroup_trace_extent(trans, child_bytenr,
2579                                                         fs_info->nodesize);
2580                         if (ret)
2581                                 goto out;
2582                 }
2583
2584                 if (level == 0) {
2585                         ret = btrfs_qgroup_trace_leaf_items(trans,
2586                                                             path->nodes[level]);
2587                         if (ret)
2588                                 goto out;
2589
2590                         /* Nonzero return here means we completed our search */
2591                         ret = adjust_slots_upwards(path, root_level);
2592                         if (ret)
2593                                 break;
2594
2595                         /* Restart search with new slots */
2596                         goto walk_down;
2597                 }
2598
2599                 level--;
2600         }
2601
2602         ret = 0;
2603 out:
2604         btrfs_free_path(path);
2605
2606         return ret;
2607 }
2608
2609 static void qgroup_iterator_nested_add(struct list_head *head, struct btrfs_qgroup *qgroup)
2610 {
2611         if (!list_empty(&qgroup->nested_iterator))
2612                 return;
2613
2614         list_add_tail(&qgroup->nested_iterator, head);
2615 }
2616
2617 static void qgroup_iterator_nested_clean(struct list_head *head)
2618 {
2619         while (!list_empty(head)) {
2620                 struct btrfs_qgroup *qgroup;
2621
2622                 qgroup = list_first_entry(head, struct btrfs_qgroup, nested_iterator);
2623                 list_del_init(&qgroup->nested_iterator);
2624         }
2625 }
2626
2627 #define UPDATE_NEW      0
2628 #define UPDATE_OLD      1
2629 /*
2630  * Walk all of the roots that points to the bytenr and adjust their refcnts.
2631  */
2632 static void qgroup_update_refcnt(struct btrfs_fs_info *fs_info,
2633                                  struct ulist *roots, struct list_head *qgroups,
2634                                  u64 seq, int update_old)
2635 {
2636         struct ulist_node *unode;
2637         struct ulist_iterator uiter;
2638         struct btrfs_qgroup *qg;
2639
2640         if (!roots)
2641                 return;
2642         ULIST_ITER_INIT(&uiter);
2643         while ((unode = ulist_next(roots, &uiter))) {
2644                 LIST_HEAD(tmp);
2645
2646                 qg = find_qgroup_rb(fs_info, unode->val);
2647                 if (!qg)
2648                         continue;
2649
2650                 qgroup_iterator_nested_add(qgroups, qg);
2651                 qgroup_iterator_add(&tmp, qg);
2652                 list_for_each_entry(qg, &tmp, iterator) {
2653                         struct btrfs_qgroup_list *glist;
2654
2655                         if (update_old)
2656                                 btrfs_qgroup_update_old_refcnt(qg, seq, 1);
2657                         else
2658                                 btrfs_qgroup_update_new_refcnt(qg, seq, 1);
2659
2660                         list_for_each_entry(glist, &qg->groups, next_group) {
2661                                 qgroup_iterator_nested_add(qgroups, glist->group);
2662                                 qgroup_iterator_add(&tmp, glist->group);
2663                         }
2664                 }
2665                 qgroup_iterator_clean(&tmp);
2666         }
2667 }
2668
2669 /*
2670  * Update qgroup rfer/excl counters.
2671  * Rfer update is easy, codes can explain themselves.
2672  *
2673  * Excl update is tricky, the update is split into 2 parts.
2674  * Part 1: Possible exclusive <-> sharing detect:
2675  *      |       A       |       !A      |
2676  *  -------------------------------------
2677  *  B   |       *       |       -       |
2678  *  -------------------------------------
2679  *  !B  |       +       |       **      |
2680  *  -------------------------------------
2681  *
2682  * Conditions:
2683  * A:   cur_old_roots < nr_old_roots    (not exclusive before)
2684  * !A:  cur_old_roots == nr_old_roots   (possible exclusive before)
2685  * B:   cur_new_roots < nr_new_roots    (not exclusive now)
2686  * !B:  cur_new_roots == nr_new_roots   (possible exclusive now)
2687  *
2688  * Results:
2689  * +: Possible sharing -> exclusive     -: Possible exclusive -> sharing
2690  * *: Definitely not changed.           **: Possible unchanged.
2691  *
2692  * For !A and !B condition, the exception is cur_old/new_roots == 0 case.
2693  *
2694  * To make the logic clear, we first use condition A and B to split
2695  * combination into 4 results.
2696  *
2697  * Then, for result "+" and "-", check old/new_roots == 0 case, as in them
2698  * only on variant maybe 0.
2699  *
2700  * Lastly, check result **, since there are 2 variants maybe 0, split them
2701  * again(2x2).
2702  * But this time we don't need to consider other things, the codes and logic
2703  * is easy to understand now.
2704  */
2705 static void qgroup_update_counters(struct btrfs_fs_info *fs_info,
2706                                    struct list_head *qgroups, u64 nr_old_roots,
2707                                    u64 nr_new_roots, u64 num_bytes, u64 seq)
2708 {
2709         struct btrfs_qgroup *qg;
2710
2711         list_for_each_entry(qg, qgroups, nested_iterator) {
2712                 u64 cur_new_count, cur_old_count;
2713                 bool dirty = false;
2714
2715                 cur_old_count = btrfs_qgroup_get_old_refcnt(qg, seq);
2716                 cur_new_count = btrfs_qgroup_get_new_refcnt(qg, seq);
2717
2718                 trace_qgroup_update_counters(fs_info, qg, cur_old_count,
2719                                              cur_new_count);
2720
2721                 /* Rfer update part */
2722                 if (cur_old_count == 0 && cur_new_count > 0) {
2723                         qg->rfer += num_bytes;
2724                         qg->rfer_cmpr += num_bytes;
2725                         dirty = true;
2726                 }
2727                 if (cur_old_count > 0 && cur_new_count == 0) {
2728                         qg->rfer -= num_bytes;
2729                         qg->rfer_cmpr -= num_bytes;
2730                         dirty = true;
2731                 }
2732
2733                 /* Excl update part */
2734                 /* Exclusive/none -> shared case */
2735                 if (cur_old_count == nr_old_roots &&
2736                     cur_new_count < nr_new_roots) {
2737                         /* Exclusive -> shared */
2738                         if (cur_old_count != 0) {
2739                                 qg->excl -= num_bytes;
2740                                 qg->excl_cmpr -= num_bytes;
2741                                 dirty = true;
2742                         }
2743                 }
2744
2745                 /* Shared -> exclusive/none case */
2746                 if (cur_old_count < nr_old_roots &&
2747                     cur_new_count == nr_new_roots) {
2748                         /* Shared->exclusive */
2749                         if (cur_new_count != 0) {
2750                                 qg->excl += num_bytes;
2751                                 qg->excl_cmpr += num_bytes;
2752                                 dirty = true;
2753                         }
2754                 }
2755
2756                 /* Exclusive/none -> exclusive/none case */
2757                 if (cur_old_count == nr_old_roots &&
2758                     cur_new_count == nr_new_roots) {
2759                         if (cur_old_count == 0) {
2760                                 /* None -> exclusive/none */
2761
2762                                 if (cur_new_count != 0) {
2763                                         /* None -> exclusive */
2764                                         qg->excl += num_bytes;
2765                                         qg->excl_cmpr += num_bytes;
2766                                         dirty = true;
2767                                 }
2768                                 /* None -> none, nothing changed */
2769                         } else {
2770                                 /* Exclusive -> exclusive/none */
2771
2772                                 if (cur_new_count == 0) {
2773                                         /* Exclusive -> none */
2774                                         qg->excl -= num_bytes;
2775                                         qg->excl_cmpr -= num_bytes;
2776                                         dirty = true;
2777                                 }
2778                                 /* Exclusive -> exclusive, nothing changed */
2779                         }
2780                 }
2781
2782                 if (dirty)
2783                         qgroup_dirty(fs_info, qg);
2784         }
2785 }
2786
2787 /*
2788  * Check if the @roots potentially is a list of fs tree roots
2789  *
2790  * Return 0 for definitely not a fs/subvol tree roots ulist
2791  * Return 1 for possible fs/subvol tree roots in the list (considering an empty
2792  *          one as well)
2793  */
2794 static int maybe_fs_roots(struct ulist *roots)
2795 {
2796         struct ulist_node *unode;
2797         struct ulist_iterator uiter;
2798
2799         /* Empty one, still possible for fs roots */
2800         if (!roots || roots->nnodes == 0)
2801                 return 1;
2802
2803         ULIST_ITER_INIT(&uiter);
2804         unode = ulist_next(roots, &uiter);
2805         if (!unode)
2806                 return 1;
2807
2808         /*
2809          * If it contains fs tree roots, then it must belong to fs/subvol
2810          * trees.
2811          * If it contains a non-fs tree, it won't be shared with fs/subvol trees.
2812          */
2813         return is_fstree(unode->val);
2814 }
2815
2816 int btrfs_qgroup_account_extent(struct btrfs_trans_handle *trans, u64 bytenr,
2817                                 u64 num_bytes, struct ulist *old_roots,
2818                                 struct ulist *new_roots)
2819 {
2820         struct btrfs_fs_info *fs_info = trans->fs_info;
2821         LIST_HEAD(qgroups);
2822         u64 seq;
2823         u64 nr_new_roots = 0;
2824         u64 nr_old_roots = 0;
2825         int ret = 0;
2826
2827         /*
2828          * If quotas get disabled meanwhile, the resources need to be freed and
2829          * we can't just exit here.
2830          */
2831         if (!btrfs_qgroup_full_accounting(fs_info) ||
2832             fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING)
2833                 goto out_free;
2834
2835         if (new_roots) {
2836                 if (!maybe_fs_roots(new_roots))
2837                         goto out_free;
2838                 nr_new_roots = new_roots->nnodes;
2839         }
2840         if (old_roots) {
2841                 if (!maybe_fs_roots(old_roots))
2842                         goto out_free;
2843                 nr_old_roots = old_roots->nnodes;
2844         }
2845
2846         /* Quick exit, either not fs tree roots, or won't affect any qgroup */
2847         if (nr_old_roots == 0 && nr_new_roots == 0)
2848                 goto out_free;
2849
2850         BUG_ON(!fs_info->quota_root);
2851
2852         trace_btrfs_qgroup_account_extent(fs_info, trans->transid, bytenr,
2853                                         num_bytes, nr_old_roots, nr_new_roots);
2854
2855         mutex_lock(&fs_info->qgroup_rescan_lock);
2856         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
2857                 if (fs_info->qgroup_rescan_progress.objectid <= bytenr) {
2858                         mutex_unlock(&fs_info->qgroup_rescan_lock);
2859                         ret = 0;
2860                         goto out_free;
2861                 }
2862         }
2863         mutex_unlock(&fs_info->qgroup_rescan_lock);
2864
2865         spin_lock(&fs_info->qgroup_lock);
2866         seq = fs_info->qgroup_seq;
2867
2868         /* Update old refcnts using old_roots */
2869         qgroup_update_refcnt(fs_info, old_roots, &qgroups, seq, UPDATE_OLD);
2870
2871         /* Update new refcnts using new_roots */
2872         qgroup_update_refcnt(fs_info, new_roots, &qgroups, seq, UPDATE_NEW);
2873
2874         qgroup_update_counters(fs_info, &qgroups, nr_old_roots, nr_new_roots,
2875                                num_bytes, seq);
2876
2877         /*
2878          * We're done using the iterator, release all its qgroups while holding
2879          * fs_info->qgroup_lock so that we don't race with btrfs_remove_qgroup()
2880          * and trigger use-after-free accesses to qgroups.
2881          */
2882         qgroup_iterator_nested_clean(&qgroups);
2883
2884         /*
2885          * Bump qgroup_seq to avoid seq overlap
2886          */
2887         fs_info->qgroup_seq += max(nr_old_roots, nr_new_roots) + 1;
2888         spin_unlock(&fs_info->qgroup_lock);
2889 out_free:
2890         ulist_free(old_roots);
2891         ulist_free(new_roots);
2892         return ret;
2893 }
2894
2895 int btrfs_qgroup_account_extents(struct btrfs_trans_handle *trans)
2896 {
2897         struct btrfs_fs_info *fs_info = trans->fs_info;
2898         struct btrfs_qgroup_extent_record *record;
2899         struct btrfs_delayed_ref_root *delayed_refs;
2900         struct ulist *new_roots = NULL;
2901         struct rb_node *node;
2902         u64 num_dirty_extents = 0;
2903         u64 qgroup_to_skip;
2904         int ret = 0;
2905
2906         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE)
2907                 return 0;
2908
2909         delayed_refs = &trans->transaction->delayed_refs;
2910         qgroup_to_skip = delayed_refs->qgroup_to_skip;
2911         while ((node = rb_first(&delayed_refs->dirty_extent_root))) {
2912                 record = rb_entry(node, struct btrfs_qgroup_extent_record,
2913                                   node);
2914
2915                 num_dirty_extents++;
2916                 trace_btrfs_qgroup_account_extents(fs_info, record);
2917
2918                 if (!ret && !(fs_info->qgroup_flags &
2919                               BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING)) {
2920                         struct btrfs_backref_walk_ctx ctx = { 0 };
2921
2922                         ctx.bytenr = record->bytenr;
2923                         ctx.fs_info = fs_info;
2924
2925                         /*
2926                          * Old roots should be searched when inserting qgroup
2927                          * extent record.
2928                          *
2929                          * But for INCONSISTENT (NO_ACCOUNTING) -> rescan case,
2930                          * we may have some record inserted during
2931                          * NO_ACCOUNTING (thus no old_roots populated), but
2932                          * later we start rescan, which clears NO_ACCOUNTING,
2933                          * leaving some inserted records without old_roots
2934                          * populated.
2935                          *
2936                          * Those cases are rare and should not cause too much
2937                          * time spent during commit_transaction().
2938                          */
2939                         if (!record->old_roots) {
2940                                 /* Search commit root to find old_roots */
2941                                 ret = btrfs_find_all_roots(&ctx, false);
2942                                 if (ret < 0)
2943                                         goto cleanup;
2944                                 record->old_roots = ctx.roots;
2945                                 ctx.roots = NULL;
2946                         }
2947
2948                         /* Free the reserved data space */
2949                         btrfs_qgroup_free_refroot(fs_info,
2950                                         record->data_rsv_refroot,
2951                                         record->data_rsv,
2952                                         BTRFS_QGROUP_RSV_DATA);
2953                         /*
2954                          * Use BTRFS_SEQ_LAST as time_seq to do special search,
2955                          * which doesn't lock tree or delayed_refs and search
2956                          * current root. It's safe inside commit_transaction().
2957                          */
2958                         ctx.trans = trans;
2959                         ctx.time_seq = BTRFS_SEQ_LAST;
2960                         ret = btrfs_find_all_roots(&ctx, false);
2961                         if (ret < 0)
2962                                 goto cleanup;
2963                         new_roots = ctx.roots;
2964                         if (qgroup_to_skip) {
2965                                 ulist_del(new_roots, qgroup_to_skip, 0);
2966                                 ulist_del(record->old_roots, qgroup_to_skip,
2967                                           0);
2968                         }
2969                         ret = btrfs_qgroup_account_extent(trans, record->bytenr,
2970                                                           record->num_bytes,
2971                                                           record->old_roots,
2972                                                           new_roots);
2973                         record->old_roots = NULL;
2974                         new_roots = NULL;
2975                 }
2976 cleanup:
2977                 ulist_free(record->old_roots);
2978                 ulist_free(new_roots);
2979                 new_roots = NULL;
2980                 rb_erase(node, &delayed_refs->dirty_extent_root);
2981                 kfree(record);
2982
2983         }
2984         trace_qgroup_num_dirty_extents(fs_info, trans->transid,
2985                                        num_dirty_extents);
2986         return ret;
2987 }
2988
2989 /*
2990  * Writes all changed qgroups to disk.
2991  * Called by the transaction commit path and the qgroup assign ioctl.
2992  */
2993 int btrfs_run_qgroups(struct btrfs_trans_handle *trans)
2994 {
2995         struct btrfs_fs_info *fs_info = trans->fs_info;
2996         int ret = 0;
2997
2998         /*
2999          * In case we are called from the qgroup assign ioctl, assert that we
3000          * are holding the qgroup_ioctl_lock, otherwise we can race with a quota
3001          * disable operation (ioctl) and access a freed quota root.
3002          */
3003         if (trans->transaction->state != TRANS_STATE_COMMIT_DOING)
3004                 lockdep_assert_held(&fs_info->qgroup_ioctl_lock);
3005
3006         if (!fs_info->quota_root)
3007                 return ret;
3008
3009         spin_lock(&fs_info->qgroup_lock);
3010         while (!list_empty(&fs_info->dirty_qgroups)) {
3011                 struct btrfs_qgroup *qgroup;
3012                 qgroup = list_first_entry(&fs_info->dirty_qgroups,
3013                                           struct btrfs_qgroup, dirty);
3014                 list_del_init(&qgroup->dirty);
3015                 spin_unlock(&fs_info->qgroup_lock);
3016                 ret = update_qgroup_info_item(trans, qgroup);
3017                 if (ret)
3018                         qgroup_mark_inconsistent(fs_info);
3019                 ret = update_qgroup_limit_item(trans, qgroup);
3020                 if (ret)
3021                         qgroup_mark_inconsistent(fs_info);
3022                 spin_lock(&fs_info->qgroup_lock);
3023         }
3024         if (btrfs_qgroup_enabled(fs_info))
3025                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_ON;
3026         else
3027                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_ON;
3028         spin_unlock(&fs_info->qgroup_lock);
3029
3030         ret = update_qgroup_status_item(trans);
3031         if (ret)
3032                 qgroup_mark_inconsistent(fs_info);
3033
3034         return ret;
3035 }
3036
3037 static int qgroup_auto_inherit(struct btrfs_fs_info *fs_info,
3038                                u64 inode_rootid,
3039                                struct btrfs_qgroup_inherit **inherit)
3040 {
3041         int i = 0;
3042         u64 num_qgroups = 0;
3043         struct btrfs_qgroup *inode_qg;
3044         struct btrfs_qgroup_list *qg_list;
3045         struct btrfs_qgroup_inherit *res;
3046         size_t struct_sz;
3047         u64 *qgids;
3048
3049         if (*inherit)
3050                 return -EEXIST;
3051
3052         inode_qg = find_qgroup_rb(fs_info, inode_rootid);
3053         if (!inode_qg)
3054                 return -ENOENT;
3055
3056         num_qgroups = list_count_nodes(&inode_qg->groups);
3057
3058         if (!num_qgroups)
3059                 return 0;
3060
3061         struct_sz = struct_size(res, qgroups, num_qgroups);
3062         if (struct_sz == SIZE_MAX)
3063                 return -ERANGE;
3064
3065         res = kzalloc(struct_sz, GFP_NOFS);
3066         if (!res)
3067                 return -ENOMEM;
3068         res->num_qgroups = num_qgroups;
3069         qgids = res->qgroups;
3070
3071         list_for_each_entry(qg_list, &inode_qg->groups, next_group)
3072                 qgids[i] = qg_list->group->qgroupid;
3073
3074         *inherit = res;
3075         return 0;
3076 }
3077
3078 /*
3079  * Copy the accounting information between qgroups. This is necessary
3080  * when a snapshot or a subvolume is created. Throwing an error will
3081  * cause a transaction abort so we take extra care here to only error
3082  * when a readonly fs is a reasonable outcome.
3083  */
3084 int btrfs_qgroup_inherit(struct btrfs_trans_handle *trans, u64 srcid,
3085                          u64 objectid, u64 inode_rootid,
3086                          struct btrfs_qgroup_inherit *inherit)
3087 {
3088         int ret = 0;
3089         int i;
3090         u64 *i_qgroups;
3091         bool committing = false;
3092         struct btrfs_fs_info *fs_info = trans->fs_info;
3093         struct btrfs_root *quota_root;
3094         struct btrfs_qgroup *srcgroup;
3095         struct btrfs_qgroup *dstgroup;
3096         struct btrfs_qgroup *prealloc;
3097         struct btrfs_qgroup_list **qlist_prealloc = NULL;
3098         bool free_inherit = false;
3099         bool need_rescan = false;
3100         u32 level_size = 0;
3101         u64 nums;
3102
3103         prealloc = kzalloc(sizeof(*prealloc), GFP_NOFS);
3104         if (!prealloc)
3105                 return -ENOMEM;
3106
3107         /*
3108          * There are only two callers of this function.
3109          *
3110          * One in create_subvol() in the ioctl context, which needs to hold
3111          * the qgroup_ioctl_lock.
3112          *
3113          * The other one in create_pending_snapshot() where no other qgroup
3114          * code can modify the fs as they all need to either start a new trans
3115          * or hold a trans handler, thus we don't need to hold
3116          * qgroup_ioctl_lock.
3117          * This would avoid long and complex lock chain and make lockdep happy.
3118          */
3119         spin_lock(&fs_info->trans_lock);
3120         if (trans->transaction->state == TRANS_STATE_COMMIT_DOING)
3121                 committing = true;
3122         spin_unlock(&fs_info->trans_lock);
3123
3124         if (!committing)
3125                 mutex_lock(&fs_info->qgroup_ioctl_lock);
3126         if (!btrfs_qgroup_enabled(fs_info))
3127                 goto out;
3128
3129         quota_root = fs_info->quota_root;
3130         if (!quota_root) {
3131                 ret = -EINVAL;
3132                 goto out;
3133         }
3134
3135         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE && !inherit) {
3136                 ret = qgroup_auto_inherit(fs_info, inode_rootid, &inherit);
3137                 if (ret)
3138                         goto out;
3139                 free_inherit = true;
3140         }
3141
3142         if (inherit) {
3143                 i_qgroups = (u64 *)(inherit + 1);
3144                 nums = inherit->num_qgroups + 2 * inherit->num_ref_copies +
3145                        2 * inherit->num_excl_copies;
3146                 for (i = 0; i < nums; ++i) {
3147                         srcgroup = find_qgroup_rb(fs_info, *i_qgroups);
3148
3149                         /*
3150                          * Zero out invalid groups so we can ignore
3151                          * them later.
3152                          */
3153                         if (!srcgroup ||
3154                             ((srcgroup->qgroupid >> 48) <= (objectid >> 48)))
3155                                 *i_qgroups = 0ULL;
3156
3157                         ++i_qgroups;
3158                 }
3159         }
3160
3161         /*
3162          * create a tracking group for the subvol itself
3163          */
3164         ret = add_qgroup_item(trans, quota_root, objectid);
3165         if (ret)
3166                 goto out;
3167
3168         /*
3169          * add qgroup to all inherited groups
3170          */
3171         if (inherit) {
3172                 i_qgroups = (u64 *)(inherit + 1);
3173                 for (i = 0; i < inherit->num_qgroups; ++i, ++i_qgroups) {
3174                         if (*i_qgroups == 0)
3175                                 continue;
3176                         ret = add_qgroup_relation_item(trans, objectid,
3177                                                        *i_qgroups);
3178                         if (ret && ret != -EEXIST)
3179                                 goto out;
3180                         ret = add_qgroup_relation_item(trans, *i_qgroups,
3181                                                        objectid);
3182                         if (ret && ret != -EEXIST)
3183                                 goto out;
3184                 }
3185                 ret = 0;
3186
3187                 qlist_prealloc = kcalloc(inherit->num_qgroups,
3188                                          sizeof(struct btrfs_qgroup_list *),
3189                                          GFP_NOFS);
3190                 if (!qlist_prealloc) {
3191                         ret = -ENOMEM;
3192                         goto out;
3193                 }
3194                 for (int i = 0; i < inherit->num_qgroups; i++) {
3195                         qlist_prealloc[i] = kzalloc(sizeof(struct btrfs_qgroup_list),
3196                                                     GFP_NOFS);
3197                         if (!qlist_prealloc[i]) {
3198                                 ret = -ENOMEM;
3199                                 goto out;
3200                         }
3201                 }
3202         }
3203
3204         spin_lock(&fs_info->qgroup_lock);
3205
3206         dstgroup = add_qgroup_rb(fs_info, prealloc, objectid);
3207         prealloc = NULL;
3208
3209         if (inherit && inherit->flags & BTRFS_QGROUP_INHERIT_SET_LIMITS) {
3210                 dstgroup->lim_flags = inherit->lim.flags;
3211                 dstgroup->max_rfer = inherit->lim.max_rfer;
3212                 dstgroup->max_excl = inherit->lim.max_excl;
3213                 dstgroup->rsv_rfer = inherit->lim.rsv_rfer;
3214                 dstgroup->rsv_excl = inherit->lim.rsv_excl;
3215
3216                 qgroup_dirty(fs_info, dstgroup);
3217         }
3218
3219         if (srcid && btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_FULL) {
3220                 srcgroup = find_qgroup_rb(fs_info, srcid);
3221                 if (!srcgroup)
3222                         goto unlock;
3223
3224                 /*
3225                  * We call inherit after we clone the root in order to make sure
3226                  * our counts don't go crazy, so at this point the only
3227                  * difference between the two roots should be the root node.
3228                  */
3229                 level_size = fs_info->nodesize;
3230                 dstgroup->rfer = srcgroup->rfer;
3231                 dstgroup->rfer_cmpr = srcgroup->rfer_cmpr;
3232                 dstgroup->excl = level_size;
3233                 dstgroup->excl_cmpr = level_size;
3234                 srcgroup->excl = level_size;
3235                 srcgroup->excl_cmpr = level_size;
3236
3237                 /* inherit the limit info */
3238                 dstgroup->lim_flags = srcgroup->lim_flags;
3239                 dstgroup->max_rfer = srcgroup->max_rfer;
3240                 dstgroup->max_excl = srcgroup->max_excl;
3241                 dstgroup->rsv_rfer = srcgroup->rsv_rfer;
3242                 dstgroup->rsv_excl = srcgroup->rsv_excl;
3243
3244                 qgroup_dirty(fs_info, dstgroup);
3245                 qgroup_dirty(fs_info, srcgroup);
3246         }
3247
3248         if (!inherit)
3249                 goto unlock;
3250
3251         i_qgroups = (u64 *)(inherit + 1);
3252         for (i = 0; i < inherit->num_qgroups; ++i) {
3253                 if (*i_qgroups) {
3254                         ret = add_relation_rb(fs_info, qlist_prealloc[i], objectid,
3255                                               *i_qgroups);
3256                         qlist_prealloc[i] = NULL;
3257                         if (ret)
3258                                 goto unlock;
3259                 }
3260                 ++i_qgroups;
3261
3262                 /*
3263                  * If we're doing a snapshot, and adding the snapshot to a new
3264                  * qgroup, the numbers are guaranteed to be incorrect.
3265                  */
3266                 if (srcid)
3267                         need_rescan = true;
3268         }
3269
3270         for (i = 0; i <  inherit->num_ref_copies; ++i, i_qgroups += 2) {
3271                 struct btrfs_qgroup *src;
3272                 struct btrfs_qgroup *dst;
3273
3274                 if (!i_qgroups[0] || !i_qgroups[1])
3275                         continue;
3276
3277                 src = find_qgroup_rb(fs_info, i_qgroups[0]);
3278                 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
3279
3280                 if (!src || !dst) {
3281                         ret = -EINVAL;
3282                         goto unlock;
3283                 }
3284
3285                 dst->rfer = src->rfer - level_size;
3286                 dst->rfer_cmpr = src->rfer_cmpr - level_size;
3287
3288                 /* Manually tweaking numbers certainly needs a rescan */
3289                 need_rescan = true;
3290         }
3291         for (i = 0; i <  inherit->num_excl_copies; ++i, i_qgroups += 2) {
3292                 struct btrfs_qgroup *src;
3293                 struct btrfs_qgroup *dst;
3294
3295                 if (!i_qgroups[0] || !i_qgroups[1])
3296                         continue;
3297
3298                 src = find_qgroup_rb(fs_info, i_qgroups[0]);
3299                 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
3300
3301                 if (!src || !dst) {
3302                         ret = -EINVAL;
3303                         goto unlock;
3304                 }
3305
3306                 dst->excl = src->excl + level_size;
3307                 dst->excl_cmpr = src->excl_cmpr + level_size;
3308                 need_rescan = true;
3309         }
3310
3311 unlock:
3312         spin_unlock(&fs_info->qgroup_lock);
3313         if (!ret)
3314                 ret = btrfs_sysfs_add_one_qgroup(fs_info, dstgroup);
3315 out:
3316         if (!committing)
3317                 mutex_unlock(&fs_info->qgroup_ioctl_lock);
3318         if (need_rescan)
3319                 qgroup_mark_inconsistent(fs_info);
3320         if (qlist_prealloc) {
3321                 for (int i = 0; i < inherit->num_qgroups; i++)
3322                         kfree(qlist_prealloc[i]);
3323                 kfree(qlist_prealloc);
3324         }
3325         if (free_inherit)
3326                 kfree(inherit);
3327         kfree(prealloc);
3328         return ret;
3329 }
3330
3331 static bool qgroup_check_limits(const struct btrfs_qgroup *qg, u64 num_bytes)
3332 {
3333         if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_RFER) &&
3334             qgroup_rsv_total(qg) + (s64)qg->rfer + num_bytes > qg->max_rfer)
3335                 return false;
3336
3337         if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) &&
3338             qgroup_rsv_total(qg) + (s64)qg->excl + num_bytes > qg->max_excl)
3339                 return false;
3340
3341         return true;
3342 }
3343
3344 static int qgroup_reserve(struct btrfs_root *root, u64 num_bytes, bool enforce,
3345                           enum btrfs_qgroup_rsv_type type)
3346 {
3347         struct btrfs_qgroup *qgroup;
3348         struct btrfs_fs_info *fs_info = root->fs_info;
3349         u64 ref_root = root->root_key.objectid;
3350         int ret = 0;
3351         LIST_HEAD(qgroup_list);
3352
3353         if (!is_fstree(ref_root))
3354                 return 0;
3355
3356         if (num_bytes == 0)
3357                 return 0;
3358
3359         if (test_bit(BTRFS_FS_QUOTA_OVERRIDE, &fs_info->flags) &&
3360             capable(CAP_SYS_RESOURCE))
3361                 enforce = false;
3362
3363         spin_lock(&fs_info->qgroup_lock);
3364         if (!fs_info->quota_root)
3365                 goto out;
3366
3367         qgroup = find_qgroup_rb(fs_info, ref_root);
3368         if (!qgroup)
3369                 goto out;
3370
3371         qgroup_iterator_add(&qgroup_list, qgroup);
3372         list_for_each_entry(qgroup, &qgroup_list, iterator) {
3373                 struct btrfs_qgroup_list *glist;
3374
3375                 if (enforce && !qgroup_check_limits(qgroup, num_bytes)) {
3376                         ret = -EDQUOT;
3377                         goto out;
3378                 }
3379
3380                 list_for_each_entry(glist, &qgroup->groups, next_group)
3381                         qgroup_iterator_add(&qgroup_list, glist->group);
3382         }
3383
3384         ret = 0;
3385         /*
3386          * no limits exceeded, now record the reservation into all qgroups
3387          */
3388         list_for_each_entry(qgroup, &qgroup_list, iterator)
3389                 qgroup_rsv_add(fs_info, qgroup, num_bytes, type);
3390
3391 out:
3392         qgroup_iterator_clean(&qgroup_list);
3393         spin_unlock(&fs_info->qgroup_lock);
3394         return ret;
3395 }
3396
3397 /*
3398  * Free @num_bytes of reserved space with @type for qgroup.  (Normally level 0
3399  * qgroup).
3400  *
3401  * Will handle all higher level qgroup too.
3402  *
3403  * NOTE: If @num_bytes is (u64)-1, this means to free all bytes of this qgroup.
3404  * This special case is only used for META_PERTRANS type.
3405  */
3406 void btrfs_qgroup_free_refroot(struct btrfs_fs_info *fs_info,
3407                                u64 ref_root, u64 num_bytes,
3408                                enum btrfs_qgroup_rsv_type type)
3409 {
3410         struct btrfs_qgroup *qgroup;
3411         LIST_HEAD(qgroup_list);
3412
3413         if (!is_fstree(ref_root))
3414                 return;
3415
3416         if (num_bytes == 0)
3417                 return;
3418
3419         if (num_bytes == (u64)-1 && type != BTRFS_QGROUP_RSV_META_PERTRANS) {
3420                 WARN(1, "%s: Invalid type to free", __func__);
3421                 return;
3422         }
3423         spin_lock(&fs_info->qgroup_lock);
3424
3425         if (!fs_info->quota_root)
3426                 goto out;
3427
3428         qgroup = find_qgroup_rb(fs_info, ref_root);
3429         if (!qgroup)
3430                 goto out;
3431
3432         if (num_bytes == (u64)-1)
3433                 /*
3434                  * We're freeing all pertrans rsv, get reserved value from
3435                  * level 0 qgroup as real num_bytes to free.
3436                  */
3437                 num_bytes = qgroup->rsv.values[type];
3438
3439         qgroup_iterator_add(&qgroup_list, qgroup);
3440         list_for_each_entry(qgroup, &qgroup_list, iterator) {
3441                 struct btrfs_qgroup_list *glist;
3442
3443                 qgroup_rsv_release(fs_info, qgroup, num_bytes, type);
3444                 list_for_each_entry(glist, &qgroup->groups, next_group) {
3445                         qgroup_iterator_add(&qgroup_list, glist->group);
3446                 }
3447         }
3448 out:
3449         qgroup_iterator_clean(&qgroup_list);
3450         spin_unlock(&fs_info->qgroup_lock);
3451 }
3452
3453 /*
3454  * Check if the leaf is the last leaf. Which means all node pointers
3455  * are at their last position.
3456  */
3457 static bool is_last_leaf(struct btrfs_path *path)
3458 {
3459         int i;
3460
3461         for (i = 1; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
3462                 if (path->slots[i] != btrfs_header_nritems(path->nodes[i]) - 1)
3463                         return false;
3464         }
3465         return true;
3466 }
3467
3468 /*
3469  * returns < 0 on error, 0 when more leafs are to be scanned.
3470  * returns 1 when done.
3471  */
3472 static int qgroup_rescan_leaf(struct btrfs_trans_handle *trans,
3473                               struct btrfs_path *path)
3474 {
3475         struct btrfs_fs_info *fs_info = trans->fs_info;
3476         struct btrfs_root *extent_root;
3477         struct btrfs_key found;
3478         struct extent_buffer *scratch_leaf = NULL;
3479         u64 num_bytes;
3480         bool done;
3481         int slot;
3482         int ret;
3483
3484         if (!btrfs_qgroup_full_accounting(fs_info))
3485                 return 1;
3486
3487         mutex_lock(&fs_info->qgroup_rescan_lock);
3488         extent_root = btrfs_extent_root(fs_info,
3489                                 fs_info->qgroup_rescan_progress.objectid);
3490         ret = btrfs_search_slot_for_read(extent_root,
3491                                          &fs_info->qgroup_rescan_progress,
3492                                          path, 1, 0);
3493
3494         btrfs_debug(fs_info,
3495                 "current progress key (%llu %u %llu), search_slot ret %d",
3496                 fs_info->qgroup_rescan_progress.objectid,
3497                 fs_info->qgroup_rescan_progress.type,
3498                 fs_info->qgroup_rescan_progress.offset, ret);
3499
3500         if (ret) {
3501                 /*
3502                  * The rescan is about to end, we will not be scanning any
3503                  * further blocks. We cannot unset the RESCAN flag here, because
3504                  * we want to commit the transaction if everything went well.
3505                  * To make the live accounting work in this phase, we set our
3506                  * scan progress pointer such that every real extent objectid
3507                  * will be smaller.
3508                  */
3509                 fs_info->qgroup_rescan_progress.objectid = (u64)-1;
3510                 btrfs_release_path(path);
3511                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3512                 return ret;
3513         }
3514         done = is_last_leaf(path);
3515
3516         btrfs_item_key_to_cpu(path->nodes[0], &found,
3517                               btrfs_header_nritems(path->nodes[0]) - 1);
3518         fs_info->qgroup_rescan_progress.objectid = found.objectid + 1;
3519
3520         scratch_leaf = btrfs_clone_extent_buffer(path->nodes[0]);
3521         if (!scratch_leaf) {
3522                 ret = -ENOMEM;
3523                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3524                 goto out;
3525         }
3526         slot = path->slots[0];
3527         btrfs_release_path(path);
3528         mutex_unlock(&fs_info->qgroup_rescan_lock);
3529
3530         for (; slot < btrfs_header_nritems(scratch_leaf); ++slot) {
3531                 struct btrfs_backref_walk_ctx ctx = { 0 };
3532
3533                 btrfs_item_key_to_cpu(scratch_leaf, &found, slot);
3534                 if (found.type != BTRFS_EXTENT_ITEM_KEY &&
3535                     found.type != BTRFS_METADATA_ITEM_KEY)
3536                         continue;
3537                 if (found.type == BTRFS_METADATA_ITEM_KEY)
3538                         num_bytes = fs_info->nodesize;
3539                 else
3540                         num_bytes = found.offset;
3541
3542                 ctx.bytenr = found.objectid;
3543                 ctx.fs_info = fs_info;
3544
3545                 ret = btrfs_find_all_roots(&ctx, false);
3546                 if (ret < 0)
3547                         goto out;
3548                 /* For rescan, just pass old_roots as NULL */
3549                 ret = btrfs_qgroup_account_extent(trans, found.objectid,
3550                                                   num_bytes, NULL, ctx.roots);
3551                 if (ret < 0)
3552                         goto out;
3553         }
3554 out:
3555         if (scratch_leaf)
3556                 free_extent_buffer(scratch_leaf);
3557
3558         if (done && !ret) {
3559                 ret = 1;
3560                 fs_info->qgroup_rescan_progress.objectid = (u64)-1;
3561         }
3562         return ret;
3563 }
3564
3565 static bool rescan_should_stop(struct btrfs_fs_info *fs_info)
3566 {
3567         if (btrfs_fs_closing(fs_info))
3568                 return true;
3569         if (test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state))
3570                 return true;
3571         if (!btrfs_qgroup_enabled(fs_info))
3572                 return true;
3573         if (fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN)
3574                 return true;
3575         return false;
3576 }
3577
3578 static void btrfs_qgroup_rescan_worker(struct btrfs_work *work)
3579 {
3580         struct btrfs_fs_info *fs_info = container_of(work, struct btrfs_fs_info,
3581                                                      qgroup_rescan_work);
3582         struct btrfs_path *path;
3583         struct btrfs_trans_handle *trans = NULL;
3584         int err = -ENOMEM;
3585         int ret = 0;
3586         bool stopped = false;
3587         bool did_leaf_rescans = false;
3588
3589         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE)
3590                 return;
3591
3592         path = btrfs_alloc_path();
3593         if (!path)
3594                 goto out;
3595         /*
3596          * Rescan should only search for commit root, and any later difference
3597          * should be recorded by qgroup
3598          */
3599         path->search_commit_root = 1;
3600         path->skip_locking = 1;
3601
3602         err = 0;
3603         while (!err && !(stopped = rescan_should_stop(fs_info))) {
3604                 trans = btrfs_start_transaction(fs_info->fs_root, 0);
3605                 if (IS_ERR(trans)) {
3606                         err = PTR_ERR(trans);
3607                         break;
3608                 }
3609
3610                 err = qgroup_rescan_leaf(trans, path);
3611                 did_leaf_rescans = true;
3612
3613                 if (err > 0)
3614                         btrfs_commit_transaction(trans);
3615                 else
3616                         btrfs_end_transaction(trans);
3617         }
3618
3619 out:
3620         btrfs_free_path(path);
3621
3622         mutex_lock(&fs_info->qgroup_rescan_lock);
3623         if (err > 0 &&
3624             fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT) {
3625                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
3626         } else if (err < 0 || stopped) {
3627                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
3628         }
3629         mutex_unlock(&fs_info->qgroup_rescan_lock);
3630
3631         /*
3632          * Only update status, since the previous part has already updated the
3633          * qgroup info, and only if we did any actual work. This also prevents
3634          * race with a concurrent quota disable, which has already set
3635          * fs_info->quota_root to NULL and cleared BTRFS_FS_QUOTA_ENABLED at
3636          * btrfs_quota_disable().
3637          */
3638         if (did_leaf_rescans) {
3639                 trans = btrfs_start_transaction(fs_info->quota_root, 1);
3640                 if (IS_ERR(trans)) {
3641                         err = PTR_ERR(trans);
3642                         trans = NULL;
3643                         btrfs_err(fs_info,
3644                                   "fail to start transaction for status update: %d",
3645                                   err);
3646                 }
3647         } else {
3648                 trans = NULL;
3649         }
3650
3651         mutex_lock(&fs_info->qgroup_rescan_lock);
3652         if (!stopped ||
3653             fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN)
3654                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3655         if (trans) {
3656                 ret = update_qgroup_status_item(trans);
3657                 if (ret < 0) {
3658                         err = ret;
3659                         btrfs_err(fs_info, "fail to update qgroup status: %d",
3660                                   err);
3661                 }
3662         }
3663         fs_info->qgroup_rescan_running = false;
3664         fs_info->qgroup_flags &= ~BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN;
3665         complete_all(&fs_info->qgroup_rescan_completion);
3666         mutex_unlock(&fs_info->qgroup_rescan_lock);
3667
3668         if (!trans)
3669                 return;
3670
3671         btrfs_end_transaction(trans);
3672
3673         if (stopped) {
3674                 btrfs_info(fs_info, "qgroup scan paused");
3675         } else if (fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN) {
3676                 btrfs_info(fs_info, "qgroup scan cancelled");
3677         } else if (err >= 0) {
3678                 btrfs_info(fs_info, "qgroup scan completed%s",
3679                         err > 0 ? " (inconsistency flag cleared)" : "");
3680         } else {
3681                 btrfs_err(fs_info, "qgroup scan failed with %d", err);
3682         }
3683 }
3684
3685 /*
3686  * Checks that (a) no rescan is running and (b) quota is enabled. Allocates all
3687  * memory required for the rescan context.
3688  */
3689 static int
3690 qgroup_rescan_init(struct btrfs_fs_info *fs_info, u64 progress_objectid,
3691                    int init_flags)
3692 {
3693         int ret = 0;
3694
3695         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE) {
3696                 btrfs_warn(fs_info, "qgroup rescan init failed, running in simple mode");
3697                 return -EINVAL;
3698         }
3699
3700         if (!init_flags) {
3701                 /* we're resuming qgroup rescan at mount time */
3702                 if (!(fs_info->qgroup_flags &
3703                       BTRFS_QGROUP_STATUS_FLAG_RESCAN)) {
3704                         btrfs_warn(fs_info,
3705                         "qgroup rescan init failed, qgroup rescan is not queued");
3706                         ret = -EINVAL;
3707                 } else if (!(fs_info->qgroup_flags &
3708                              BTRFS_QGROUP_STATUS_FLAG_ON)) {
3709                         btrfs_warn(fs_info,
3710                         "qgroup rescan init failed, qgroup is not enabled");
3711                         ret = -EINVAL;
3712                 }
3713
3714                 if (ret)
3715                         return ret;
3716         }
3717
3718         mutex_lock(&fs_info->qgroup_rescan_lock);
3719
3720         if (init_flags) {
3721                 if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
3722                         btrfs_warn(fs_info,
3723                                    "qgroup rescan is already in progress");
3724                         ret = -EINPROGRESS;
3725                 } else if (!(fs_info->qgroup_flags &
3726                              BTRFS_QGROUP_STATUS_FLAG_ON)) {
3727                         btrfs_warn(fs_info,
3728                         "qgroup rescan init failed, qgroup is not enabled");
3729                         ret = -EINVAL;
3730                 } else if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_DISABLED) {
3731                         /* Quota disable is in progress */
3732                         ret = -EBUSY;
3733                 }
3734
3735                 if (ret) {
3736                         mutex_unlock(&fs_info->qgroup_rescan_lock);
3737                         return ret;
3738                 }
3739                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3740         }
3741
3742         memset(&fs_info->qgroup_rescan_progress, 0,
3743                 sizeof(fs_info->qgroup_rescan_progress));
3744         fs_info->qgroup_flags &= ~(BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN |
3745                                    BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING);
3746         fs_info->qgroup_rescan_progress.objectid = progress_objectid;
3747         init_completion(&fs_info->qgroup_rescan_completion);
3748         mutex_unlock(&fs_info->qgroup_rescan_lock);
3749
3750         btrfs_init_work(&fs_info->qgroup_rescan_work,
3751                         btrfs_qgroup_rescan_worker, NULL);
3752         return 0;
3753 }
3754
3755 static void
3756 qgroup_rescan_zero_tracking(struct btrfs_fs_info *fs_info)
3757 {
3758         struct rb_node *n;
3759         struct btrfs_qgroup *qgroup;
3760
3761         spin_lock(&fs_info->qgroup_lock);
3762         /* clear all current qgroup tracking information */
3763         for (n = rb_first(&fs_info->qgroup_tree); n; n = rb_next(n)) {
3764                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
3765                 qgroup->rfer = 0;
3766                 qgroup->rfer_cmpr = 0;
3767                 qgroup->excl = 0;
3768                 qgroup->excl_cmpr = 0;
3769                 qgroup_dirty(fs_info, qgroup);
3770         }
3771         spin_unlock(&fs_info->qgroup_lock);
3772 }
3773
3774 int
3775 btrfs_qgroup_rescan(struct btrfs_fs_info *fs_info)
3776 {
3777         int ret = 0;
3778         struct btrfs_trans_handle *trans;
3779
3780         ret = qgroup_rescan_init(fs_info, 0, 1);
3781         if (ret)
3782                 return ret;
3783
3784         /*
3785          * We have set the rescan_progress to 0, which means no more
3786          * delayed refs will be accounted by btrfs_qgroup_account_ref.
3787          * However, btrfs_qgroup_account_ref may be right after its call
3788          * to btrfs_find_all_roots, in which case it would still do the
3789          * accounting.
3790          * To solve this, we're committing the transaction, which will
3791          * ensure we run all delayed refs and only after that, we are
3792          * going to clear all tracking information for a clean start.
3793          */
3794
3795         trans = btrfs_attach_transaction_barrier(fs_info->fs_root);
3796         if (IS_ERR(trans) && trans != ERR_PTR(-ENOENT)) {
3797                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3798                 return PTR_ERR(trans);
3799         } else if (trans != ERR_PTR(-ENOENT)) {
3800                 ret = btrfs_commit_transaction(trans);
3801                 if (ret) {
3802                         fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3803                         return ret;
3804                 }
3805         }
3806
3807         qgroup_rescan_zero_tracking(fs_info);
3808
3809         mutex_lock(&fs_info->qgroup_rescan_lock);
3810         fs_info->qgroup_rescan_running = true;
3811         btrfs_queue_work(fs_info->qgroup_rescan_workers,
3812                          &fs_info->qgroup_rescan_work);
3813         mutex_unlock(&fs_info->qgroup_rescan_lock);
3814
3815         return 0;
3816 }
3817
3818 int btrfs_qgroup_wait_for_completion(struct btrfs_fs_info *fs_info,
3819                                      bool interruptible)
3820 {
3821         int running;
3822         int ret = 0;
3823
3824         mutex_lock(&fs_info->qgroup_rescan_lock);
3825         running = fs_info->qgroup_rescan_running;
3826         mutex_unlock(&fs_info->qgroup_rescan_lock);
3827
3828         if (!running)
3829                 return 0;
3830
3831         if (interruptible)
3832                 ret = wait_for_completion_interruptible(
3833                                         &fs_info->qgroup_rescan_completion);
3834         else
3835                 wait_for_completion(&fs_info->qgroup_rescan_completion);
3836
3837         return ret;
3838 }
3839
3840 /*
3841  * this is only called from open_ctree where we're still single threaded, thus
3842  * locking is omitted here.
3843  */
3844 void
3845 btrfs_qgroup_rescan_resume(struct btrfs_fs_info *fs_info)
3846 {
3847         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
3848                 mutex_lock(&fs_info->qgroup_rescan_lock);
3849                 fs_info->qgroup_rescan_running = true;
3850                 btrfs_queue_work(fs_info->qgroup_rescan_workers,
3851                                  &fs_info->qgroup_rescan_work);
3852                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3853         }
3854 }
3855
3856 #define rbtree_iterate_from_safe(node, next, start)                             \
3857        for (node = start; node && ({ next = rb_next(node); 1;}); node = next)
3858
3859 static int qgroup_unreserve_range(struct btrfs_inode *inode,
3860                                   struct extent_changeset *reserved, u64 start,
3861                                   u64 len)
3862 {
3863         struct rb_node *node;
3864         struct rb_node *next;
3865         struct ulist_node *entry;
3866         int ret = 0;
3867
3868         node = reserved->range_changed.root.rb_node;
3869         if (!node)
3870                 return 0;
3871         while (node) {
3872                 entry = rb_entry(node, struct ulist_node, rb_node);
3873                 if (entry->val < start)
3874                         node = node->rb_right;
3875                 else
3876                         node = node->rb_left;
3877         }
3878
3879         if (entry->val > start && rb_prev(&entry->rb_node))
3880                 entry = rb_entry(rb_prev(&entry->rb_node), struct ulist_node,
3881                                  rb_node);
3882
3883         rbtree_iterate_from_safe(node, next, &entry->rb_node) {
3884                 u64 entry_start;
3885                 u64 entry_end;
3886                 u64 entry_len;
3887                 int clear_ret;
3888
3889                 entry = rb_entry(node, struct ulist_node, rb_node);
3890                 entry_start = entry->val;
3891                 entry_end = entry->aux;
3892                 entry_len = entry_end - entry_start + 1;
3893
3894                 if (entry_start >= start + len)
3895                         break;
3896                 if (entry_start + entry_len <= start)
3897                         continue;
3898                 /*
3899                  * Now the entry is in [start, start + len), revert the
3900                  * EXTENT_QGROUP_RESERVED bit.
3901                  */
3902                 clear_ret = clear_extent_bits(&inode->io_tree, entry_start,
3903                                               entry_end, EXTENT_QGROUP_RESERVED);
3904                 if (!ret && clear_ret < 0)
3905                         ret = clear_ret;
3906
3907                 ulist_del(&reserved->range_changed, entry->val, entry->aux);
3908                 if (likely(reserved->bytes_changed >= entry_len)) {
3909                         reserved->bytes_changed -= entry_len;
3910                 } else {
3911                         WARN_ON(1);
3912                         reserved->bytes_changed = 0;
3913                 }
3914         }
3915
3916         return ret;
3917 }
3918
3919 /*
3920  * Try to free some space for qgroup.
3921  *
3922  * For qgroup, there are only 3 ways to free qgroup space:
3923  * - Flush nodatacow write
3924  *   Any nodatacow write will free its reserved data space at run_delalloc_range().
3925  *   In theory, we should only flush nodatacow inodes, but it's not yet
3926  *   possible, so we need to flush the whole root.
3927  *
3928  * - Wait for ordered extents
3929  *   When ordered extents are finished, their reserved metadata is finally
3930  *   converted to per_trans status, which can be freed by later commit
3931  *   transaction.
3932  *
3933  * - Commit transaction
3934  *   This would free the meta_per_trans space.
3935  *   In theory this shouldn't provide much space, but any more qgroup space
3936  *   is needed.
3937  */
3938 static int try_flush_qgroup(struct btrfs_root *root)
3939 {
3940         struct btrfs_trans_handle *trans;
3941         int ret;
3942
3943         /* Can't hold an open transaction or we run the risk of deadlocking. */
3944         ASSERT(current->journal_info == NULL);
3945         if (WARN_ON(current->journal_info))
3946                 return 0;
3947
3948         /*
3949          * We don't want to run flush again and again, so if there is a running
3950          * one, we won't try to start a new flush, but exit directly.
3951          */
3952         if (test_and_set_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state)) {
3953                 wait_event(root->qgroup_flush_wait,
3954                         !test_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state));
3955                 return 0;
3956         }
3957
3958         ret = btrfs_start_delalloc_snapshot(root, true);
3959         if (ret < 0)
3960                 goto out;
3961         btrfs_wait_ordered_extents(root, U64_MAX, 0, (u64)-1);
3962
3963         trans = btrfs_attach_transaction_barrier(root);
3964         if (IS_ERR(trans)) {
3965                 ret = PTR_ERR(trans);
3966                 if (ret == -ENOENT)
3967                         ret = 0;
3968                 goto out;
3969         }
3970
3971         ret = btrfs_commit_transaction(trans);
3972 out:
3973         clear_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state);
3974         wake_up(&root->qgroup_flush_wait);
3975         return ret;
3976 }
3977
3978 static int qgroup_reserve_data(struct btrfs_inode *inode,
3979                         struct extent_changeset **reserved_ret, u64 start,
3980                         u64 len)
3981 {
3982         struct btrfs_root *root = inode->root;
3983         struct extent_changeset *reserved;
3984         bool new_reserved = false;
3985         u64 orig_reserved;
3986         u64 to_reserve;
3987         int ret;
3988
3989         if (btrfs_qgroup_mode(root->fs_info) == BTRFS_QGROUP_MODE_DISABLED ||
3990             !is_fstree(root->root_key.objectid) || len == 0)
3991                 return 0;
3992
3993         /* @reserved parameter is mandatory for qgroup */
3994         if (WARN_ON(!reserved_ret))
3995                 return -EINVAL;
3996         if (!*reserved_ret) {
3997                 new_reserved = true;
3998                 *reserved_ret = extent_changeset_alloc();
3999                 if (!*reserved_ret)
4000                         return -ENOMEM;
4001         }
4002         reserved = *reserved_ret;
4003         /* Record already reserved space */
4004         orig_reserved = reserved->bytes_changed;
4005         ret = set_record_extent_bits(&inode->io_tree, start,
4006                         start + len -1, EXTENT_QGROUP_RESERVED, reserved);
4007
4008         /* Newly reserved space */
4009         to_reserve = reserved->bytes_changed - orig_reserved;
4010         trace_btrfs_qgroup_reserve_data(&inode->vfs_inode, start, len,
4011                                         to_reserve, QGROUP_RESERVE);
4012         if (ret < 0)
4013                 goto out;
4014         ret = qgroup_reserve(root, to_reserve, true, BTRFS_QGROUP_RSV_DATA);
4015         if (ret < 0)
4016                 goto cleanup;
4017
4018         return ret;
4019
4020 cleanup:
4021         qgroup_unreserve_range(inode, reserved, start, len);
4022 out:
4023         if (new_reserved) {
4024                 extent_changeset_free(reserved);
4025                 *reserved_ret = NULL;
4026         }
4027         return ret;
4028 }
4029
4030 /*
4031  * Reserve qgroup space for range [start, start + len).
4032  *
4033  * This function will either reserve space from related qgroups or do nothing
4034  * if the range is already reserved.
4035  *
4036  * Return 0 for successful reservation
4037  * Return <0 for error (including -EQUOT)
4038  *
4039  * NOTE: This function may sleep for memory allocation, dirty page flushing and
4040  *       commit transaction. So caller should not hold any dirty page locked.
4041  */
4042 int btrfs_qgroup_reserve_data(struct btrfs_inode *inode,
4043                         struct extent_changeset **reserved_ret, u64 start,
4044                         u64 len)
4045 {
4046         int ret;
4047
4048         ret = qgroup_reserve_data(inode, reserved_ret, start, len);
4049         if (ret <= 0 && ret != -EDQUOT)
4050                 return ret;
4051
4052         ret = try_flush_qgroup(inode->root);
4053         if (ret < 0)
4054                 return ret;
4055         return qgroup_reserve_data(inode, reserved_ret, start, len);
4056 }
4057
4058 /* Free ranges specified by @reserved, normally in error path */
4059 static int qgroup_free_reserved_data(struct btrfs_inode *inode,
4060                                      struct extent_changeset *reserved,
4061                                      u64 start, u64 len, u64 *freed_ret)
4062 {
4063         struct btrfs_root *root = inode->root;
4064         struct ulist_node *unode;
4065         struct ulist_iterator uiter;
4066         struct extent_changeset changeset;
4067         u64 freed = 0;
4068         int ret;
4069
4070         extent_changeset_init(&changeset);
4071         len = round_up(start + len, root->fs_info->sectorsize);
4072         start = round_down(start, root->fs_info->sectorsize);
4073
4074         ULIST_ITER_INIT(&uiter);
4075         while ((unode = ulist_next(&reserved->range_changed, &uiter))) {
4076                 u64 range_start = unode->val;
4077                 /* unode->aux is the inclusive end */
4078                 u64 range_len = unode->aux - range_start + 1;
4079                 u64 free_start;
4080                 u64 free_len;
4081
4082                 extent_changeset_release(&changeset);
4083
4084                 /* Only free range in range [start, start + len) */
4085                 if (range_start >= start + len ||
4086                     range_start + range_len <= start)
4087                         continue;
4088                 free_start = max(range_start, start);
4089                 free_len = min(start + len, range_start + range_len) -
4090                            free_start;
4091                 /*
4092                  * TODO: To also modify reserved->ranges_reserved to reflect
4093                  * the modification.
4094                  *
4095                  * However as long as we free qgroup reserved according to
4096                  * EXTENT_QGROUP_RESERVED, we won't double free.
4097                  * So not need to rush.
4098                  */
4099                 ret = clear_record_extent_bits(&inode->io_tree, free_start,
4100                                 free_start + free_len - 1,
4101                                 EXTENT_QGROUP_RESERVED, &changeset);
4102                 if (ret < 0)
4103                         goto out;
4104                 freed += changeset.bytes_changed;
4105         }
4106         btrfs_qgroup_free_refroot(root->fs_info, root->root_key.objectid, freed,
4107                                   BTRFS_QGROUP_RSV_DATA);
4108         if (freed_ret)
4109                 *freed_ret = freed;
4110         ret = 0;
4111 out:
4112         extent_changeset_release(&changeset);
4113         return ret;
4114 }
4115
4116 static int __btrfs_qgroup_release_data(struct btrfs_inode *inode,
4117                         struct extent_changeset *reserved, u64 start, u64 len,
4118                         u64 *released, int free)
4119 {
4120         struct extent_changeset changeset;
4121         int trace_op = QGROUP_RELEASE;
4122         int ret;
4123
4124         if (btrfs_qgroup_mode(inode->root->fs_info) == BTRFS_QGROUP_MODE_DISABLED) {
4125                 extent_changeset_init(&changeset);
4126                 return clear_record_extent_bits(&inode->io_tree, start,
4127                                                 start + len - 1,
4128                                                 EXTENT_QGROUP_RESERVED, &changeset);
4129         }
4130
4131         /* In release case, we shouldn't have @reserved */
4132         WARN_ON(!free && reserved);
4133         if (free && reserved)
4134                 return qgroup_free_reserved_data(inode, reserved, start, len, released);
4135         extent_changeset_init(&changeset);
4136         ret = clear_record_extent_bits(&inode->io_tree, start, start + len -1,
4137                                        EXTENT_QGROUP_RESERVED, &changeset);
4138         if (ret < 0)
4139                 goto out;
4140
4141         if (free)
4142                 trace_op = QGROUP_FREE;
4143         trace_btrfs_qgroup_release_data(&inode->vfs_inode, start, len,
4144                                         changeset.bytes_changed, trace_op);
4145         if (free)
4146                 btrfs_qgroup_free_refroot(inode->root->fs_info,
4147                                 inode->root->root_key.objectid,
4148                                 changeset.bytes_changed, BTRFS_QGROUP_RSV_DATA);
4149         if (released)
4150                 *released = changeset.bytes_changed;
4151 out:
4152         extent_changeset_release(&changeset);
4153         return ret;
4154 }
4155
4156 /*
4157  * Free a reserved space range from io_tree and related qgroups
4158  *
4159  * Should be called when a range of pages get invalidated before reaching disk.
4160  * Or for error cleanup case.
4161  * if @reserved is given, only reserved range in [@start, @start + @len) will
4162  * be freed.
4163  *
4164  * For data written to disk, use btrfs_qgroup_release_data().
4165  *
4166  * NOTE: This function may sleep for memory allocation.
4167  */
4168 int btrfs_qgroup_free_data(struct btrfs_inode *inode,
4169                            struct extent_changeset *reserved,
4170                            u64 start, u64 len, u64 *freed)
4171 {
4172         return __btrfs_qgroup_release_data(inode, reserved, start, len, freed, 1);
4173 }
4174
4175 /*
4176  * Release a reserved space range from io_tree only.
4177  *
4178  * Should be called when a range of pages get written to disk and corresponding
4179  * FILE_EXTENT is inserted into corresponding root.
4180  *
4181  * Since new qgroup accounting framework will only update qgroup numbers at
4182  * commit_transaction() time, its reserved space shouldn't be freed from
4183  * related qgroups.
4184  *
4185  * But we should release the range from io_tree, to allow further write to be
4186  * COWed.
4187  *
4188  * NOTE: This function may sleep for memory allocation.
4189  */
4190 int btrfs_qgroup_release_data(struct btrfs_inode *inode, u64 start, u64 len, u64 *released)
4191 {
4192         return __btrfs_qgroup_release_data(inode, NULL, start, len, released, 0);
4193 }
4194
4195 static void add_root_meta_rsv(struct btrfs_root *root, int num_bytes,
4196                               enum btrfs_qgroup_rsv_type type)
4197 {
4198         if (type != BTRFS_QGROUP_RSV_META_PREALLOC &&
4199             type != BTRFS_QGROUP_RSV_META_PERTRANS)
4200                 return;
4201         if (num_bytes == 0)
4202                 return;
4203
4204         spin_lock(&root->qgroup_meta_rsv_lock);
4205         if (type == BTRFS_QGROUP_RSV_META_PREALLOC)
4206                 root->qgroup_meta_rsv_prealloc += num_bytes;
4207         else
4208                 root->qgroup_meta_rsv_pertrans += num_bytes;
4209         spin_unlock(&root->qgroup_meta_rsv_lock);
4210 }
4211
4212 static int sub_root_meta_rsv(struct btrfs_root *root, int num_bytes,
4213                              enum btrfs_qgroup_rsv_type type)
4214 {
4215         if (type != BTRFS_QGROUP_RSV_META_PREALLOC &&
4216             type != BTRFS_QGROUP_RSV_META_PERTRANS)
4217                 return 0;
4218         if (num_bytes == 0)
4219                 return 0;
4220
4221         spin_lock(&root->qgroup_meta_rsv_lock);
4222         if (type == BTRFS_QGROUP_RSV_META_PREALLOC) {
4223                 num_bytes = min_t(u64, root->qgroup_meta_rsv_prealloc,
4224                                   num_bytes);
4225                 root->qgroup_meta_rsv_prealloc -= num_bytes;
4226         } else {
4227                 num_bytes = min_t(u64, root->qgroup_meta_rsv_pertrans,
4228                                   num_bytes);
4229                 root->qgroup_meta_rsv_pertrans -= num_bytes;
4230         }
4231         spin_unlock(&root->qgroup_meta_rsv_lock);
4232         return num_bytes;
4233 }
4234
4235 int btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
4236                               enum btrfs_qgroup_rsv_type type, bool enforce)
4237 {
4238         struct btrfs_fs_info *fs_info = root->fs_info;
4239         int ret;
4240
4241         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_DISABLED ||
4242             !is_fstree(root->root_key.objectid) || num_bytes == 0)
4243                 return 0;
4244
4245         BUG_ON(num_bytes != round_down(num_bytes, fs_info->nodesize));
4246         trace_qgroup_meta_reserve(root, (s64)num_bytes, type);
4247         ret = qgroup_reserve(root, num_bytes, enforce, type);
4248         if (ret < 0)
4249                 return ret;
4250         /*
4251          * Record what we have reserved into root.
4252          *
4253          * To avoid quota disabled->enabled underflow.
4254          * In that case, we may try to free space we haven't reserved
4255          * (since quota was disabled), so record what we reserved into root.
4256          * And ensure later release won't underflow this number.
4257          */
4258         add_root_meta_rsv(root, num_bytes, type);
4259         return ret;
4260 }
4261
4262 int __btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
4263                                 enum btrfs_qgroup_rsv_type type, bool enforce,
4264                                 bool noflush)
4265 {
4266         int ret;
4267
4268         ret = btrfs_qgroup_reserve_meta(root, num_bytes, type, enforce);
4269         if ((ret <= 0 && ret != -EDQUOT) || noflush)
4270                 return ret;
4271
4272         ret = try_flush_qgroup(root);
4273         if (ret < 0)
4274                 return ret;
4275         return btrfs_qgroup_reserve_meta(root, num_bytes, type, enforce);
4276 }
4277
4278 /*
4279  * Per-transaction meta reservation should be all freed at transaction commit
4280  * time
4281  */
4282 void btrfs_qgroup_free_meta_all_pertrans(struct btrfs_root *root)
4283 {
4284         struct btrfs_fs_info *fs_info = root->fs_info;
4285
4286         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_DISABLED ||
4287             !is_fstree(root->root_key.objectid))
4288                 return;
4289
4290         /* TODO: Update trace point to handle such free */
4291         trace_qgroup_meta_free_all_pertrans(root);
4292         /* Special value -1 means to free all reserved space */
4293         btrfs_qgroup_free_refroot(fs_info, root->root_key.objectid, (u64)-1,
4294                                   BTRFS_QGROUP_RSV_META_PERTRANS);
4295 }
4296
4297 void __btrfs_qgroup_free_meta(struct btrfs_root *root, int num_bytes,
4298                               enum btrfs_qgroup_rsv_type type)
4299 {
4300         struct btrfs_fs_info *fs_info = root->fs_info;
4301
4302         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_DISABLED ||
4303             !is_fstree(root->root_key.objectid))
4304                 return;
4305
4306         /*
4307          * reservation for META_PREALLOC can happen before quota is enabled,
4308          * which can lead to underflow.
4309          * Here ensure we will only free what we really have reserved.
4310          */
4311         num_bytes = sub_root_meta_rsv(root, num_bytes, type);
4312         BUG_ON(num_bytes != round_down(num_bytes, fs_info->nodesize));
4313         trace_qgroup_meta_reserve(root, -(s64)num_bytes, type);
4314         btrfs_qgroup_free_refroot(fs_info, root->root_key.objectid,
4315                                   num_bytes, type);
4316 }
4317
4318 static void qgroup_convert_meta(struct btrfs_fs_info *fs_info, u64 ref_root,
4319                                 int num_bytes)
4320 {
4321         struct btrfs_qgroup *qgroup;
4322         LIST_HEAD(qgroup_list);
4323
4324         if (num_bytes == 0)
4325                 return;
4326         if (!fs_info->quota_root)
4327                 return;
4328
4329         spin_lock(&fs_info->qgroup_lock);
4330         qgroup = find_qgroup_rb(fs_info, ref_root);
4331         if (!qgroup)
4332                 goto out;
4333
4334         qgroup_iterator_add(&qgroup_list, qgroup);
4335         list_for_each_entry(qgroup, &qgroup_list, iterator) {
4336                 struct btrfs_qgroup_list *glist;
4337
4338                 qgroup_rsv_release(fs_info, qgroup, num_bytes,
4339                                 BTRFS_QGROUP_RSV_META_PREALLOC);
4340                 if (!sb_rdonly(fs_info->sb))
4341                         qgroup_rsv_add(fs_info, qgroup, num_bytes,
4342                                        BTRFS_QGROUP_RSV_META_PERTRANS);
4343
4344                 list_for_each_entry(glist, &qgroup->groups, next_group)
4345                         qgroup_iterator_add(&qgroup_list, glist->group);
4346         }
4347 out:
4348         qgroup_iterator_clean(&qgroup_list);
4349         spin_unlock(&fs_info->qgroup_lock);
4350 }
4351
4352 /*
4353  * Convert @num_bytes of META_PREALLOCATED reservation to META_PERTRANS.
4354  *
4355  * This is called when preallocated meta reservation needs to be used.
4356  * Normally after btrfs_join_transaction() call.
4357  */
4358 void btrfs_qgroup_convert_reserved_meta(struct btrfs_root *root, int num_bytes)
4359 {
4360         struct btrfs_fs_info *fs_info = root->fs_info;
4361
4362         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_DISABLED ||
4363             !is_fstree(root->root_key.objectid))
4364                 return;
4365         /* Same as btrfs_qgroup_free_meta_prealloc() */
4366         num_bytes = sub_root_meta_rsv(root, num_bytes,
4367                                       BTRFS_QGROUP_RSV_META_PREALLOC);
4368         trace_qgroup_meta_convert(root, num_bytes);
4369         qgroup_convert_meta(fs_info, root->root_key.objectid, num_bytes);
4370 }
4371
4372 /*
4373  * Check qgroup reserved space leaking, normally at destroy inode
4374  * time
4375  */
4376 void btrfs_qgroup_check_reserved_leak(struct btrfs_inode *inode)
4377 {
4378         struct extent_changeset changeset;
4379         struct ulist_node *unode;
4380         struct ulist_iterator iter;
4381         int ret;
4382
4383         extent_changeset_init(&changeset);
4384         ret = clear_record_extent_bits(&inode->io_tree, 0, (u64)-1,
4385                         EXTENT_QGROUP_RESERVED, &changeset);
4386
4387         WARN_ON(ret < 0);
4388         if (WARN_ON(changeset.bytes_changed)) {
4389                 ULIST_ITER_INIT(&iter);
4390                 while ((unode = ulist_next(&changeset.range_changed, &iter))) {
4391                         btrfs_warn(inode->root->fs_info,
4392                 "leaking qgroup reserved space, ino: %llu, start: %llu, end: %llu",
4393                                 btrfs_ino(inode), unode->val, unode->aux);
4394                 }
4395                 btrfs_qgroup_free_refroot(inode->root->fs_info,
4396                                 inode->root->root_key.objectid,
4397                                 changeset.bytes_changed, BTRFS_QGROUP_RSV_DATA);
4398
4399         }
4400         extent_changeset_release(&changeset);
4401 }
4402
4403 void btrfs_qgroup_init_swapped_blocks(
4404         struct btrfs_qgroup_swapped_blocks *swapped_blocks)
4405 {
4406         int i;
4407
4408         spin_lock_init(&swapped_blocks->lock);
4409         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
4410                 swapped_blocks->blocks[i] = RB_ROOT;
4411         swapped_blocks->swapped = false;
4412 }
4413
4414 /*
4415  * Delete all swapped blocks record of @root.
4416  * Every record here means we skipped a full subtree scan for qgroup.
4417  *
4418  * Gets called when committing one transaction.
4419  */
4420 void btrfs_qgroup_clean_swapped_blocks(struct btrfs_root *root)
4421 {
4422         struct btrfs_qgroup_swapped_blocks *swapped_blocks;
4423         int i;
4424
4425         swapped_blocks = &root->swapped_blocks;
4426
4427         spin_lock(&swapped_blocks->lock);
4428         if (!swapped_blocks->swapped)
4429                 goto out;
4430         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
4431                 struct rb_root *cur_root = &swapped_blocks->blocks[i];
4432                 struct btrfs_qgroup_swapped_block *entry;
4433                 struct btrfs_qgroup_swapped_block *next;
4434
4435                 rbtree_postorder_for_each_entry_safe(entry, next, cur_root,
4436                                                      node)
4437                         kfree(entry);
4438                 swapped_blocks->blocks[i] = RB_ROOT;
4439         }
4440         swapped_blocks->swapped = false;
4441 out:
4442         spin_unlock(&swapped_blocks->lock);
4443 }
4444
4445 /*
4446  * Add subtree roots record into @subvol_root.
4447  *
4448  * @subvol_root:        tree root of the subvolume tree get swapped
4449  * @bg:                 block group under balance
4450  * @subvol_parent/slot: pointer to the subtree root in subvolume tree
4451  * @reloc_parent/slot:  pointer to the subtree root in reloc tree
4452  *                      BOTH POINTERS ARE BEFORE TREE SWAP
4453  * @last_snapshot:      last snapshot generation of the subvolume tree
4454  */
4455 int btrfs_qgroup_add_swapped_blocks(struct btrfs_trans_handle *trans,
4456                 struct btrfs_root *subvol_root,
4457                 struct btrfs_block_group *bg,
4458                 struct extent_buffer *subvol_parent, int subvol_slot,
4459                 struct extent_buffer *reloc_parent, int reloc_slot,
4460                 u64 last_snapshot)
4461 {
4462         struct btrfs_fs_info *fs_info = subvol_root->fs_info;
4463         struct btrfs_qgroup_swapped_blocks *blocks = &subvol_root->swapped_blocks;
4464         struct btrfs_qgroup_swapped_block *block;
4465         struct rb_node **cur;
4466         struct rb_node *parent = NULL;
4467         int level = btrfs_header_level(subvol_parent) - 1;
4468         int ret = 0;
4469
4470         if (!btrfs_qgroup_full_accounting(fs_info))
4471                 return 0;
4472
4473         if (btrfs_node_ptr_generation(subvol_parent, subvol_slot) >
4474             btrfs_node_ptr_generation(reloc_parent, reloc_slot)) {
4475                 btrfs_err_rl(fs_info,
4476                 "%s: bad parameter order, subvol_gen=%llu reloc_gen=%llu",
4477                         __func__,
4478                         btrfs_node_ptr_generation(subvol_parent, subvol_slot),
4479                         btrfs_node_ptr_generation(reloc_parent, reloc_slot));
4480                 return -EUCLEAN;
4481         }
4482
4483         block = kmalloc(sizeof(*block), GFP_NOFS);
4484         if (!block) {
4485                 ret = -ENOMEM;
4486                 goto out;
4487         }
4488
4489         /*
4490          * @reloc_parent/slot is still before swap, while @block is going to
4491          * record the bytenr after swap, so we do the swap here.
4492          */
4493         block->subvol_bytenr = btrfs_node_blockptr(reloc_parent, reloc_slot);
4494         block->subvol_generation = btrfs_node_ptr_generation(reloc_parent,
4495                                                              reloc_slot);
4496         block->reloc_bytenr = btrfs_node_blockptr(subvol_parent, subvol_slot);
4497         block->reloc_generation = btrfs_node_ptr_generation(subvol_parent,
4498                                                             subvol_slot);
4499         block->last_snapshot = last_snapshot;
4500         block->level = level;
4501
4502         /*
4503          * If we have bg == NULL, we're called from btrfs_recover_relocation(),
4504          * no one else can modify tree blocks thus we qgroup will not change
4505          * no matter the value of trace_leaf.
4506          */
4507         if (bg && bg->flags & BTRFS_BLOCK_GROUP_DATA)
4508                 block->trace_leaf = true;
4509         else
4510                 block->trace_leaf = false;
4511         btrfs_node_key_to_cpu(reloc_parent, &block->first_key, reloc_slot);
4512
4513         /* Insert @block into @blocks */
4514         spin_lock(&blocks->lock);
4515         cur = &blocks->blocks[level].rb_node;
4516         while (*cur) {
4517                 struct btrfs_qgroup_swapped_block *entry;
4518
4519                 parent = *cur;
4520                 entry = rb_entry(parent, struct btrfs_qgroup_swapped_block,
4521                                  node);
4522
4523                 if (entry->subvol_bytenr < block->subvol_bytenr) {
4524                         cur = &(*cur)->rb_left;
4525                 } else if (entry->subvol_bytenr > block->subvol_bytenr) {
4526                         cur = &(*cur)->rb_right;
4527                 } else {
4528                         if (entry->subvol_generation !=
4529                                         block->subvol_generation ||
4530                             entry->reloc_bytenr != block->reloc_bytenr ||
4531                             entry->reloc_generation !=
4532                                         block->reloc_generation) {
4533                                 /*
4534                                  * Duplicated but mismatch entry found.
4535                                  * Shouldn't happen.
4536                                  *
4537                                  * Marking qgroup inconsistent should be enough
4538                                  * for end users.
4539                                  */
4540                                 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
4541                                 ret = -EEXIST;
4542                         }
4543                         kfree(block);
4544                         goto out_unlock;
4545                 }
4546         }
4547         rb_link_node(&block->node, parent, cur);
4548         rb_insert_color(&block->node, &blocks->blocks[level]);
4549         blocks->swapped = true;
4550 out_unlock:
4551         spin_unlock(&blocks->lock);
4552 out:
4553         if (ret < 0)
4554                 qgroup_mark_inconsistent(fs_info);
4555         return ret;
4556 }
4557
4558 /*
4559  * Check if the tree block is a subtree root, and if so do the needed
4560  * delayed subtree trace for qgroup.
4561  *
4562  * This is called during btrfs_cow_block().
4563  */
4564 int btrfs_qgroup_trace_subtree_after_cow(struct btrfs_trans_handle *trans,
4565                                          struct btrfs_root *root,
4566                                          struct extent_buffer *subvol_eb)
4567 {
4568         struct btrfs_fs_info *fs_info = root->fs_info;
4569         struct btrfs_tree_parent_check check = { 0 };
4570         struct btrfs_qgroup_swapped_blocks *blocks = &root->swapped_blocks;
4571         struct btrfs_qgroup_swapped_block *block;
4572         struct extent_buffer *reloc_eb = NULL;
4573         struct rb_node *node;
4574         bool found = false;
4575         bool swapped = false;
4576         int level = btrfs_header_level(subvol_eb);
4577         int ret = 0;
4578         int i;
4579
4580         if (!btrfs_qgroup_full_accounting(fs_info))
4581                 return 0;
4582         if (!is_fstree(root->root_key.objectid) || !root->reloc_root)
4583                 return 0;
4584
4585         spin_lock(&blocks->lock);
4586         if (!blocks->swapped) {
4587                 spin_unlock(&blocks->lock);
4588                 return 0;
4589         }
4590         node = blocks->blocks[level].rb_node;
4591
4592         while (node) {
4593                 block = rb_entry(node, struct btrfs_qgroup_swapped_block, node);
4594                 if (block->subvol_bytenr < subvol_eb->start) {
4595                         node = node->rb_left;
4596                 } else if (block->subvol_bytenr > subvol_eb->start) {
4597                         node = node->rb_right;
4598                 } else {
4599                         found = true;
4600                         break;
4601                 }
4602         }
4603         if (!found) {
4604                 spin_unlock(&blocks->lock);
4605                 goto out;
4606         }
4607         /* Found one, remove it from @blocks first and update blocks->swapped */
4608         rb_erase(&block->node, &blocks->blocks[level]);
4609         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
4610                 if (RB_EMPTY_ROOT(&blocks->blocks[i])) {
4611                         swapped = true;
4612                         break;
4613                 }
4614         }
4615         blocks->swapped = swapped;
4616         spin_unlock(&blocks->lock);
4617
4618         check.level = block->level;
4619         check.transid = block->reloc_generation;
4620         check.has_first_key = true;
4621         memcpy(&check.first_key, &block->first_key, sizeof(check.first_key));
4622
4623         /* Read out reloc subtree root */
4624         reloc_eb = read_tree_block(fs_info, block->reloc_bytenr, &check);
4625         if (IS_ERR(reloc_eb)) {
4626                 ret = PTR_ERR(reloc_eb);
4627                 reloc_eb = NULL;
4628                 goto free_out;
4629         }
4630         if (!extent_buffer_uptodate(reloc_eb)) {
4631                 ret = -EIO;
4632                 goto free_out;
4633         }
4634
4635         ret = qgroup_trace_subtree_swap(trans, reloc_eb, subvol_eb,
4636                         block->last_snapshot, block->trace_leaf);
4637 free_out:
4638         kfree(block);
4639         free_extent_buffer(reloc_eb);
4640 out:
4641         if (ret < 0) {
4642                 btrfs_err_rl(fs_info,
4643                              "failed to account subtree at bytenr %llu: %d",
4644                              subvol_eb->start, ret);
4645                 qgroup_mark_inconsistent(fs_info);
4646         }
4647         return ret;
4648 }
4649
4650 void btrfs_qgroup_destroy_extent_records(struct btrfs_transaction *trans)
4651 {
4652         struct btrfs_qgroup_extent_record *entry;
4653         struct btrfs_qgroup_extent_record *next;
4654         struct rb_root *root;
4655
4656         root = &trans->delayed_refs.dirty_extent_root;
4657         rbtree_postorder_for_each_entry_safe(entry, next, root, node) {
4658                 ulist_free(entry->old_roots);
4659                 kfree(entry);
4660         }
4661         *root = RB_ROOT;
4662 }
4663
4664 void btrfs_free_squota_rsv(struct btrfs_fs_info *fs_info, u64 root, u64 rsv_bytes)
4665 {
4666         if (btrfs_qgroup_mode(fs_info) != BTRFS_QGROUP_MODE_SIMPLE)
4667                 return;
4668
4669         if (!is_fstree(root))
4670                 return;
4671
4672         btrfs_qgroup_free_refroot(fs_info, root, rsv_bytes, BTRFS_QGROUP_RSV_DATA);
4673 }
4674
4675 int btrfs_record_squota_delta(struct btrfs_fs_info *fs_info,
4676                               struct btrfs_squota_delta *delta)
4677 {
4678         int ret;
4679         struct btrfs_qgroup *qgroup;
4680         struct btrfs_qgroup *qg;
4681         LIST_HEAD(qgroup_list);
4682         u64 root = delta->root;
4683         u64 num_bytes = delta->num_bytes;
4684         const int sign = (delta->is_inc ? 1 : -1);
4685
4686         if (btrfs_qgroup_mode(fs_info) != BTRFS_QGROUP_MODE_SIMPLE)
4687                 return 0;
4688
4689         if (!is_fstree(root))
4690                 return 0;
4691
4692         /* If the extent predates enabling quotas, don't count it. */
4693         if (delta->generation < fs_info->qgroup_enable_gen)
4694                 return 0;
4695
4696         spin_lock(&fs_info->qgroup_lock);
4697         qgroup = find_qgroup_rb(fs_info, root);
4698         if (!qgroup) {
4699                 ret = -ENOENT;
4700                 goto out;
4701         }
4702
4703         ret = 0;
4704         qgroup_iterator_add(&qgroup_list, qgroup);
4705         list_for_each_entry(qg, &qgroup_list, iterator) {
4706                 struct btrfs_qgroup_list *glist;
4707
4708                 qg->excl += num_bytes * sign;
4709                 qg->rfer += num_bytes * sign;
4710                 qgroup_dirty(fs_info, qg);
4711
4712                 list_for_each_entry(glist, &qg->groups, next_group)
4713                         qgroup_iterator_add(&qgroup_list, glist->group);
4714         }
4715         qgroup_iterator_clean(&qgroup_list);
4716
4717 out:
4718         spin_unlock(&fs_info->qgroup_lock);
4719         return ret;
4720 }