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 transferred 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         ret = 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          * Relocation will mess with backrefs, so make sure we have the
1346          * cleaner_mutex held to protect us from relocate.
1347          */
1348         lockdep_assert_held(&fs_info->cleaner_mutex);
1349
1350         mutex_lock(&fs_info->qgroup_ioctl_lock);
1351         if (!fs_info->quota_root)
1352                 goto out;
1353
1354         /*
1355          * Unlock the qgroup_ioctl_lock mutex before waiting for the rescan worker to
1356          * complete. Otherwise we can deadlock because btrfs_remove_qgroup() needs
1357          * to lock that mutex while holding a transaction handle and the rescan
1358          * worker needs to commit a transaction.
1359          */
1360         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1361
1362         /*
1363          * Request qgroup rescan worker to complete and wait for it. This wait
1364          * must be done before transaction start for quota disable since it may
1365          * deadlock with transaction by the qgroup rescan worker.
1366          */
1367         clear_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1368         btrfs_qgroup_wait_for_completion(fs_info, false);
1369
1370         /*
1371          * We have nothing held here and no trans handle, just return the error
1372          * if there is one.
1373          */
1374         ret = flush_reservations(fs_info);
1375         if (ret)
1376                 return ret;
1377
1378         /*
1379          * 1 For the root item
1380          *
1381          * We should also reserve enough items for the quota tree deletion in
1382          * btrfs_clean_quota_tree but this is not done.
1383          *
1384          * Also, we must always start a transaction without holding the mutex
1385          * qgroup_ioctl_lock, see btrfs_quota_enable().
1386          */
1387         trans = btrfs_start_transaction(fs_info->tree_root, 1);
1388
1389         mutex_lock(&fs_info->qgroup_ioctl_lock);
1390         if (IS_ERR(trans)) {
1391                 ret = PTR_ERR(trans);
1392                 trans = NULL;
1393                 set_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1394                 goto out;
1395         }
1396
1397         if (!fs_info->quota_root)
1398                 goto out;
1399
1400         spin_lock(&fs_info->qgroup_lock);
1401         quota_root = fs_info->quota_root;
1402         fs_info->quota_root = NULL;
1403         fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_ON;
1404         fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_SIMPLE_MODE;
1405         fs_info->qgroup_drop_subtree_thres = BTRFS_MAX_LEVEL;
1406         spin_unlock(&fs_info->qgroup_lock);
1407
1408         btrfs_free_qgroup_config(fs_info);
1409
1410         ret = btrfs_clean_quota_tree(trans, quota_root);
1411         if (ret) {
1412                 btrfs_abort_transaction(trans, ret);
1413                 goto out;
1414         }
1415
1416         ret = btrfs_del_root(trans, &quota_root->root_key);
1417         if (ret) {
1418                 btrfs_abort_transaction(trans, ret);
1419                 goto out;
1420         }
1421
1422         spin_lock(&fs_info->trans_lock);
1423         list_del(&quota_root->dirty_list);
1424         spin_unlock(&fs_info->trans_lock);
1425
1426         btrfs_tree_lock(quota_root->node);
1427         btrfs_clear_buffer_dirty(trans, quota_root->node);
1428         btrfs_tree_unlock(quota_root->node);
1429         btrfs_free_tree_block(trans, btrfs_root_id(quota_root),
1430                               quota_root->node, 0, 1);
1431
1432         btrfs_put_root(quota_root);
1433
1434 out:
1435         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1436         if (ret && trans)
1437                 btrfs_end_transaction(trans);
1438         else if (trans)
1439                 ret = btrfs_commit_transaction(trans);
1440         return ret;
1441 }
1442
1443 static void qgroup_dirty(struct btrfs_fs_info *fs_info,
1444                          struct btrfs_qgroup *qgroup)
1445 {
1446         if (list_empty(&qgroup->dirty))
1447                 list_add(&qgroup->dirty, &fs_info->dirty_qgroups);
1448 }
1449
1450 static void qgroup_iterator_add(struct list_head *head, struct btrfs_qgroup *qgroup)
1451 {
1452         if (!list_empty(&qgroup->iterator))
1453                 return;
1454
1455         list_add_tail(&qgroup->iterator, head);
1456 }
1457
1458 static void qgroup_iterator_clean(struct list_head *head)
1459 {
1460         while (!list_empty(head)) {
1461                 struct btrfs_qgroup *qgroup;
1462
1463                 qgroup = list_first_entry(head, struct btrfs_qgroup, iterator);
1464                 list_del_init(&qgroup->iterator);
1465         }
1466 }
1467
1468 /*
1469  * The easy accounting, we're updating qgroup relationship whose child qgroup
1470  * only has exclusive extents.
1471  *
1472  * In this case, all exclusive extents will also be exclusive for parent, so
1473  * excl/rfer just get added/removed.
1474  *
1475  * So is qgroup reservation space, which should also be added/removed to
1476  * parent.
1477  * Or when child tries to release reservation space, parent will underflow its
1478  * reservation (for relationship adding case).
1479  *
1480  * Caller should hold fs_info->qgroup_lock.
1481  */
1482 static int __qgroup_excl_accounting(struct btrfs_fs_info *fs_info, u64 ref_root,
1483                                     struct btrfs_qgroup *src, int sign)
1484 {
1485         struct btrfs_qgroup *qgroup;
1486         struct btrfs_qgroup *cur;
1487         LIST_HEAD(qgroup_list);
1488         u64 num_bytes = src->excl;
1489         int ret = 0;
1490
1491         qgroup = find_qgroup_rb(fs_info, ref_root);
1492         if (!qgroup)
1493                 goto out;
1494
1495         qgroup_iterator_add(&qgroup_list, qgroup);
1496         list_for_each_entry(cur, &qgroup_list, iterator) {
1497                 struct btrfs_qgroup_list *glist;
1498
1499                 qgroup->rfer += sign * num_bytes;
1500                 qgroup->rfer_cmpr += sign * num_bytes;
1501
1502                 WARN_ON(sign < 0 && qgroup->excl < num_bytes);
1503                 qgroup->excl += sign * num_bytes;
1504                 qgroup->excl_cmpr += sign * num_bytes;
1505
1506                 if (sign > 0)
1507                         qgroup_rsv_add_by_qgroup(fs_info, qgroup, src);
1508                 else
1509                         qgroup_rsv_release_by_qgroup(fs_info, qgroup, src);
1510                 qgroup_dirty(fs_info, qgroup);
1511
1512                 /* Append parent qgroups to @qgroup_list. */
1513                 list_for_each_entry(glist, &qgroup->groups, next_group)
1514                         qgroup_iterator_add(&qgroup_list, glist->group);
1515         }
1516         ret = 0;
1517 out:
1518         qgroup_iterator_clean(&qgroup_list);
1519         return ret;
1520 }
1521
1522
1523 /*
1524  * Quick path for updating qgroup with only excl refs.
1525  *
1526  * In that case, just update all parent will be enough.
1527  * Or we needs to do a full rescan.
1528  * Caller should also hold fs_info->qgroup_lock.
1529  *
1530  * Return 0 for quick update, return >0 for need to full rescan
1531  * and mark INCONSISTENT flag.
1532  * Return < 0 for other error.
1533  */
1534 static int quick_update_accounting(struct btrfs_fs_info *fs_info,
1535                                    u64 src, u64 dst, int sign)
1536 {
1537         struct btrfs_qgroup *qgroup;
1538         int ret = 1;
1539         int err = 0;
1540
1541         qgroup = find_qgroup_rb(fs_info, src);
1542         if (!qgroup)
1543                 goto out;
1544         if (qgroup->excl == qgroup->rfer) {
1545                 ret = 0;
1546                 err = __qgroup_excl_accounting(fs_info, dst, qgroup, sign);
1547                 if (err < 0) {
1548                         ret = err;
1549                         goto out;
1550                 }
1551         }
1552 out:
1553         if (ret)
1554                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1555         return ret;
1556 }
1557
1558 int btrfs_add_qgroup_relation(struct btrfs_trans_handle *trans, u64 src, u64 dst)
1559 {
1560         struct btrfs_fs_info *fs_info = trans->fs_info;
1561         struct btrfs_qgroup *parent;
1562         struct btrfs_qgroup *member;
1563         struct btrfs_qgroup_list *list;
1564         struct btrfs_qgroup_list *prealloc = NULL;
1565         int ret = 0;
1566
1567         /* Check the level of src and dst first */
1568         if (btrfs_qgroup_level(src) >= btrfs_qgroup_level(dst))
1569                 return -EINVAL;
1570
1571         mutex_lock(&fs_info->qgroup_ioctl_lock);
1572         if (!fs_info->quota_root) {
1573                 ret = -ENOTCONN;
1574                 goto out;
1575         }
1576         member = find_qgroup_rb(fs_info, src);
1577         parent = find_qgroup_rb(fs_info, dst);
1578         if (!member || !parent) {
1579                 ret = -EINVAL;
1580                 goto out;
1581         }
1582
1583         /* check if such qgroup relation exist firstly */
1584         list_for_each_entry(list, &member->groups, next_group) {
1585                 if (list->group == parent) {
1586                         ret = -EEXIST;
1587                         goto out;
1588                 }
1589         }
1590
1591         prealloc = kzalloc(sizeof(*list), GFP_NOFS);
1592         if (!prealloc) {
1593                 ret = -ENOMEM;
1594                 goto out;
1595         }
1596         ret = add_qgroup_relation_item(trans, src, dst);
1597         if (ret)
1598                 goto out;
1599
1600         ret = add_qgroup_relation_item(trans, dst, src);
1601         if (ret) {
1602                 del_qgroup_relation_item(trans, src, dst);
1603                 goto out;
1604         }
1605
1606         spin_lock(&fs_info->qgroup_lock);
1607         ret = __add_relation_rb(prealloc, member, parent);
1608         prealloc = NULL;
1609         if (ret < 0) {
1610                 spin_unlock(&fs_info->qgroup_lock);
1611                 goto out;
1612         }
1613         ret = quick_update_accounting(fs_info, src, dst, 1);
1614         spin_unlock(&fs_info->qgroup_lock);
1615 out:
1616         kfree(prealloc);
1617         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1618         return ret;
1619 }
1620
1621 static int __del_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1622                                  u64 dst)
1623 {
1624         struct btrfs_fs_info *fs_info = trans->fs_info;
1625         struct btrfs_qgroup *parent;
1626         struct btrfs_qgroup *member;
1627         struct btrfs_qgroup_list *list;
1628         bool found = false;
1629         int ret = 0;
1630         int ret2;
1631
1632         if (!fs_info->quota_root) {
1633                 ret = -ENOTCONN;
1634                 goto out;
1635         }
1636
1637         member = find_qgroup_rb(fs_info, src);
1638         parent = find_qgroup_rb(fs_info, dst);
1639         /*
1640          * The parent/member pair doesn't exist, then try to delete the dead
1641          * relation items only.
1642          */
1643         if (!member || !parent)
1644                 goto delete_item;
1645
1646         /* check if such qgroup relation exist firstly */
1647         list_for_each_entry(list, &member->groups, next_group) {
1648                 if (list->group == parent) {
1649                         found = true;
1650                         break;
1651                 }
1652         }
1653
1654 delete_item:
1655         ret = del_qgroup_relation_item(trans, src, dst);
1656         if (ret < 0 && ret != -ENOENT)
1657                 goto out;
1658         ret2 = del_qgroup_relation_item(trans, dst, src);
1659         if (ret2 < 0 && ret2 != -ENOENT)
1660                 goto out;
1661
1662         /* At least one deletion succeeded, return 0 */
1663         if (!ret || !ret2)
1664                 ret = 0;
1665
1666         if (found) {
1667                 spin_lock(&fs_info->qgroup_lock);
1668                 del_relation_rb(fs_info, src, dst);
1669                 ret = quick_update_accounting(fs_info, src, dst, -1);
1670                 spin_unlock(&fs_info->qgroup_lock);
1671         }
1672 out:
1673         return ret;
1674 }
1675
1676 int btrfs_del_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1677                               u64 dst)
1678 {
1679         struct btrfs_fs_info *fs_info = trans->fs_info;
1680         int ret = 0;
1681
1682         mutex_lock(&fs_info->qgroup_ioctl_lock);
1683         ret = __del_qgroup_relation(trans, src, dst);
1684         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1685
1686         return ret;
1687 }
1688
1689 int btrfs_create_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid)
1690 {
1691         struct btrfs_fs_info *fs_info = trans->fs_info;
1692         struct btrfs_root *quota_root;
1693         struct btrfs_qgroup *qgroup;
1694         struct btrfs_qgroup *prealloc = NULL;
1695         int ret = 0;
1696
1697         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_DISABLED)
1698                 return 0;
1699
1700         mutex_lock(&fs_info->qgroup_ioctl_lock);
1701         if (!fs_info->quota_root) {
1702                 ret = -ENOTCONN;
1703                 goto out;
1704         }
1705         quota_root = fs_info->quota_root;
1706         qgroup = find_qgroup_rb(fs_info, qgroupid);
1707         if (qgroup) {
1708                 ret = -EEXIST;
1709                 goto out;
1710         }
1711
1712         prealloc = kzalloc(sizeof(*prealloc), GFP_NOFS);
1713         if (!prealloc) {
1714                 ret = -ENOMEM;
1715                 goto out;
1716         }
1717
1718         ret = add_qgroup_item(trans, quota_root, qgroupid);
1719         if (ret)
1720                 goto out;
1721
1722         spin_lock(&fs_info->qgroup_lock);
1723         qgroup = add_qgroup_rb(fs_info, prealloc, qgroupid);
1724         spin_unlock(&fs_info->qgroup_lock);
1725         prealloc = NULL;
1726
1727         ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1728 out:
1729         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1730         kfree(prealloc);
1731         return ret;
1732 }
1733
1734 static bool qgroup_has_usage(struct btrfs_qgroup *qgroup)
1735 {
1736         return (qgroup->rfer > 0 || qgroup->rfer_cmpr > 0 ||
1737                 qgroup->excl > 0 || qgroup->excl_cmpr > 0 ||
1738                 qgroup->rsv.values[BTRFS_QGROUP_RSV_DATA] > 0 ||
1739                 qgroup->rsv.values[BTRFS_QGROUP_RSV_META_PREALLOC] > 0 ||
1740                 qgroup->rsv.values[BTRFS_QGROUP_RSV_META_PERTRANS] > 0);
1741 }
1742
1743 int btrfs_remove_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid)
1744 {
1745         struct btrfs_fs_info *fs_info = trans->fs_info;
1746         struct btrfs_qgroup *qgroup;
1747         struct btrfs_qgroup_list *list;
1748         int ret = 0;
1749
1750         mutex_lock(&fs_info->qgroup_ioctl_lock);
1751         if (!fs_info->quota_root) {
1752                 ret = -ENOTCONN;
1753                 goto out;
1754         }
1755
1756         qgroup = find_qgroup_rb(fs_info, qgroupid);
1757         if (!qgroup) {
1758                 ret = -ENOENT;
1759                 goto out;
1760         }
1761
1762         if (is_fstree(qgroupid) && qgroup_has_usage(qgroup)) {
1763                 ret = -EBUSY;
1764                 goto out;
1765         }
1766
1767         /* Check if there are no children of this qgroup */
1768         if (!list_empty(&qgroup->members)) {
1769                 ret = -EBUSY;
1770                 goto out;
1771         }
1772
1773         ret = del_qgroup_item(trans, qgroupid);
1774         if (ret && ret != -ENOENT)
1775                 goto out;
1776
1777         while (!list_empty(&qgroup->groups)) {
1778                 list = list_first_entry(&qgroup->groups,
1779                                         struct btrfs_qgroup_list, next_group);
1780                 ret = __del_qgroup_relation(trans, qgroupid,
1781                                             list->group->qgroupid);
1782                 if (ret)
1783                         goto out;
1784         }
1785
1786         spin_lock(&fs_info->qgroup_lock);
1787         del_qgroup_rb(fs_info, qgroupid);
1788         spin_unlock(&fs_info->qgroup_lock);
1789
1790         /*
1791          * Remove the qgroup from sysfs now without holding the qgroup_lock
1792          * spinlock, since the sysfs_remove_group() function needs to take
1793          * the mutex kernfs_mutex through kernfs_remove_by_name_ns().
1794          */
1795         btrfs_sysfs_del_one_qgroup(fs_info, qgroup);
1796         kfree(qgroup);
1797 out:
1798         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1799         return ret;
1800 }
1801
1802 int btrfs_limit_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid,
1803                        struct btrfs_qgroup_limit *limit)
1804 {
1805         struct btrfs_fs_info *fs_info = trans->fs_info;
1806         struct btrfs_qgroup *qgroup;
1807         int ret = 0;
1808         /* Sometimes we would want to clear the limit on this qgroup.
1809          * To meet this requirement, we treat the -1 as a special value
1810          * which tell kernel to clear the limit on this qgroup.
1811          */
1812         const u64 CLEAR_VALUE = -1;
1813
1814         mutex_lock(&fs_info->qgroup_ioctl_lock);
1815         if (!fs_info->quota_root) {
1816                 ret = -ENOTCONN;
1817                 goto out;
1818         }
1819
1820         qgroup = find_qgroup_rb(fs_info, qgroupid);
1821         if (!qgroup) {
1822                 ret = -ENOENT;
1823                 goto out;
1824         }
1825
1826         spin_lock(&fs_info->qgroup_lock);
1827         if (limit->flags & BTRFS_QGROUP_LIMIT_MAX_RFER) {
1828                 if (limit->max_rfer == CLEAR_VALUE) {
1829                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_MAX_RFER;
1830                         limit->flags &= ~BTRFS_QGROUP_LIMIT_MAX_RFER;
1831                         qgroup->max_rfer = 0;
1832                 } else {
1833                         qgroup->max_rfer = limit->max_rfer;
1834                 }
1835         }
1836         if (limit->flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) {
1837                 if (limit->max_excl == CLEAR_VALUE) {
1838                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_MAX_EXCL;
1839                         limit->flags &= ~BTRFS_QGROUP_LIMIT_MAX_EXCL;
1840                         qgroup->max_excl = 0;
1841                 } else {
1842                         qgroup->max_excl = limit->max_excl;
1843                 }
1844         }
1845         if (limit->flags & BTRFS_QGROUP_LIMIT_RSV_RFER) {
1846                 if (limit->rsv_rfer == CLEAR_VALUE) {
1847                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_RSV_RFER;
1848                         limit->flags &= ~BTRFS_QGROUP_LIMIT_RSV_RFER;
1849                         qgroup->rsv_rfer = 0;
1850                 } else {
1851                         qgroup->rsv_rfer = limit->rsv_rfer;
1852                 }
1853         }
1854         if (limit->flags & BTRFS_QGROUP_LIMIT_RSV_EXCL) {
1855                 if (limit->rsv_excl == CLEAR_VALUE) {
1856                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_RSV_EXCL;
1857                         limit->flags &= ~BTRFS_QGROUP_LIMIT_RSV_EXCL;
1858                         qgroup->rsv_excl = 0;
1859                 } else {
1860                         qgroup->rsv_excl = limit->rsv_excl;
1861                 }
1862         }
1863         qgroup->lim_flags |= limit->flags;
1864
1865         spin_unlock(&fs_info->qgroup_lock);
1866
1867         ret = update_qgroup_limit_item(trans, qgroup);
1868         if (ret) {
1869                 qgroup_mark_inconsistent(fs_info);
1870                 btrfs_info(fs_info, "unable to update quota limit for %llu",
1871                        qgroupid);
1872         }
1873
1874 out:
1875         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1876         return ret;
1877 }
1878
1879 /*
1880  * Inform qgroup to trace one dirty extent, its info is recorded in @record.
1881  * So qgroup can account it at transaction committing time.
1882  *
1883  * No lock version, caller must acquire delayed ref lock and allocated memory,
1884  * then call btrfs_qgroup_trace_extent_post() after exiting lock context.
1885  *
1886  * Return 0 for success insert
1887  * Return >0 for existing record, caller can free @record safely.
1888  * Error is not possible
1889  */
1890 int btrfs_qgroup_trace_extent_nolock(struct btrfs_fs_info *fs_info,
1891                                 struct btrfs_delayed_ref_root *delayed_refs,
1892                                 struct btrfs_qgroup_extent_record *record)
1893 {
1894         struct rb_node **p = &delayed_refs->dirty_extent_root.rb_node;
1895         struct rb_node *parent_node = NULL;
1896         struct btrfs_qgroup_extent_record *entry;
1897         u64 bytenr = record->bytenr;
1898
1899         if (!btrfs_qgroup_full_accounting(fs_info))
1900                 return 1;
1901
1902         lockdep_assert_held(&delayed_refs->lock);
1903         trace_btrfs_qgroup_trace_extent(fs_info, record);
1904
1905         while (*p) {
1906                 parent_node = *p;
1907                 entry = rb_entry(parent_node, struct btrfs_qgroup_extent_record,
1908                                  node);
1909                 if (bytenr < entry->bytenr) {
1910                         p = &(*p)->rb_left;
1911                 } else if (bytenr > entry->bytenr) {
1912                         p = &(*p)->rb_right;
1913                 } else {
1914                         if (record->data_rsv && !entry->data_rsv) {
1915                                 entry->data_rsv = record->data_rsv;
1916                                 entry->data_rsv_refroot =
1917                                         record->data_rsv_refroot;
1918                         }
1919                         return 1;
1920                 }
1921         }
1922
1923         rb_link_node(&record->node, parent_node, p);
1924         rb_insert_color(&record->node, &delayed_refs->dirty_extent_root);
1925         return 0;
1926 }
1927
1928 /*
1929  * Post handler after qgroup_trace_extent_nolock().
1930  *
1931  * NOTE: Current qgroup does the expensive backref walk at transaction
1932  * committing time with TRANS_STATE_COMMIT_DOING, this blocks incoming
1933  * new transaction.
1934  * This is designed to allow btrfs_find_all_roots() to get correct new_roots
1935  * result.
1936  *
1937  * However for old_roots there is no need to do backref walk at that time,
1938  * since we search commit roots to walk backref and result will always be
1939  * correct.
1940  *
1941  * Due to the nature of no lock version, we can't do backref there.
1942  * So we must call btrfs_qgroup_trace_extent_post() after exiting
1943  * spinlock context.
1944  *
1945  * TODO: If we can fix and prove btrfs_find_all_roots() can get correct result
1946  * using current root, then we can move all expensive backref walk out of
1947  * transaction committing, but not now as qgroup accounting will be wrong again.
1948  */
1949 int btrfs_qgroup_trace_extent_post(struct btrfs_trans_handle *trans,
1950                                    struct btrfs_qgroup_extent_record *qrecord)
1951 {
1952         struct btrfs_backref_walk_ctx ctx = { 0 };
1953         int ret;
1954
1955         if (!btrfs_qgroup_full_accounting(trans->fs_info))
1956                 return 0;
1957         /*
1958          * We are always called in a context where we are already holding a
1959          * transaction handle. Often we are called when adding a data delayed
1960          * reference from btrfs_truncate_inode_items() (truncating or unlinking),
1961          * in which case we will be holding a write lock on extent buffer from a
1962          * subvolume tree. In this case we can't allow btrfs_find_all_roots() to
1963          * acquire fs_info->commit_root_sem, because that is a higher level lock
1964          * that must be acquired before locking any extent buffers.
1965          *
1966          * So we want btrfs_find_all_roots() to not acquire the commit_root_sem
1967          * but we can't pass it a non-NULL transaction handle, because otherwise
1968          * it would not use commit roots and would lock extent buffers, causing
1969          * a deadlock if it ends up trying to read lock the same extent buffer
1970          * that was previously write locked at btrfs_truncate_inode_items().
1971          *
1972          * So pass a NULL transaction handle to btrfs_find_all_roots() and
1973          * explicitly tell it to not acquire the commit_root_sem - if we are
1974          * holding a transaction handle we don't need its protection.
1975          */
1976         ASSERT(trans != NULL);
1977
1978         if (trans->fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING)
1979                 return 0;
1980
1981         ctx.bytenr = qrecord->bytenr;
1982         ctx.fs_info = trans->fs_info;
1983
1984         ret = btrfs_find_all_roots(&ctx, true);
1985         if (ret < 0) {
1986                 qgroup_mark_inconsistent(trans->fs_info);
1987                 btrfs_warn(trans->fs_info,
1988 "error accounting new delayed refs extent (err code: %d), quota inconsistent",
1989                         ret);
1990                 return 0;
1991         }
1992
1993         /*
1994          * Here we don't need to get the lock of
1995          * trans->transaction->delayed_refs, since inserted qrecord won't
1996          * be deleted, only qrecord->node may be modified (new qrecord insert)
1997          *
1998          * So modifying qrecord->old_roots is safe here
1999          */
2000         qrecord->old_roots = ctx.roots;
2001         return 0;
2002 }
2003
2004 /*
2005  * Inform qgroup to trace one dirty extent, specified by @bytenr and
2006  * @num_bytes.
2007  * So qgroup can account it at commit trans time.
2008  *
2009  * Better encapsulated version, with memory allocation and backref walk for
2010  * commit roots.
2011  * So this can sleep.
2012  *
2013  * Return 0 if the operation is done.
2014  * Return <0 for error, like memory allocation failure or invalid parameter
2015  * (NULL trans)
2016  */
2017 int btrfs_qgroup_trace_extent(struct btrfs_trans_handle *trans, u64 bytenr,
2018                               u64 num_bytes)
2019 {
2020         struct btrfs_fs_info *fs_info = trans->fs_info;
2021         struct btrfs_qgroup_extent_record *record;
2022         struct btrfs_delayed_ref_root *delayed_refs;
2023         int ret;
2024
2025         if (!btrfs_qgroup_full_accounting(fs_info) || bytenr == 0 || num_bytes == 0)
2026                 return 0;
2027         record = kzalloc(sizeof(*record), GFP_NOFS);
2028         if (!record)
2029                 return -ENOMEM;
2030
2031         delayed_refs = &trans->transaction->delayed_refs;
2032         record->bytenr = bytenr;
2033         record->num_bytes = num_bytes;
2034         record->old_roots = NULL;
2035
2036         spin_lock(&delayed_refs->lock);
2037         ret = btrfs_qgroup_trace_extent_nolock(fs_info, delayed_refs, record);
2038         spin_unlock(&delayed_refs->lock);
2039         if (ret > 0) {
2040                 kfree(record);
2041                 return 0;
2042         }
2043         return btrfs_qgroup_trace_extent_post(trans, record);
2044 }
2045
2046 /*
2047  * Inform qgroup to trace all leaf items of data
2048  *
2049  * Return 0 for success
2050  * Return <0 for error(ENOMEM)
2051  */
2052 int btrfs_qgroup_trace_leaf_items(struct btrfs_trans_handle *trans,
2053                                   struct extent_buffer *eb)
2054 {
2055         struct btrfs_fs_info *fs_info = trans->fs_info;
2056         int nr = btrfs_header_nritems(eb);
2057         int i, extent_type, ret;
2058         struct btrfs_key key;
2059         struct btrfs_file_extent_item *fi;
2060         u64 bytenr, num_bytes;
2061
2062         /* We can be called directly from walk_up_proc() */
2063         if (!btrfs_qgroup_full_accounting(fs_info))
2064                 return 0;
2065
2066         for (i = 0; i < nr; i++) {
2067                 btrfs_item_key_to_cpu(eb, &key, i);
2068
2069                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2070                         continue;
2071
2072                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
2073                 /* filter out non qgroup-accountable extents  */
2074                 extent_type = btrfs_file_extent_type(eb, fi);
2075
2076                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
2077                         continue;
2078
2079                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
2080                 if (!bytenr)
2081                         continue;
2082
2083                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
2084
2085                 ret = btrfs_qgroup_trace_extent(trans, bytenr, num_bytes);
2086                 if (ret)
2087                         return ret;
2088         }
2089         cond_resched();
2090         return 0;
2091 }
2092
2093 /*
2094  * Walk up the tree from the bottom, freeing leaves and any interior
2095  * nodes which have had all slots visited. If a node (leaf or
2096  * interior) is freed, the node above it will have it's slot
2097  * incremented. The root node will never be freed.
2098  *
2099  * At the end of this function, we should have a path which has all
2100  * slots incremented to the next position for a search. If we need to
2101  * read a new node it will be NULL and the node above it will have the
2102  * correct slot selected for a later read.
2103  *
2104  * If we increment the root nodes slot counter past the number of
2105  * elements, 1 is returned to signal completion of the search.
2106  */
2107 static int adjust_slots_upwards(struct btrfs_path *path, int root_level)
2108 {
2109         int level = 0;
2110         int nr, slot;
2111         struct extent_buffer *eb;
2112
2113         if (root_level == 0)
2114                 return 1;
2115
2116         while (level <= root_level) {
2117                 eb = path->nodes[level];
2118                 nr = btrfs_header_nritems(eb);
2119                 path->slots[level]++;
2120                 slot = path->slots[level];
2121                 if (slot >= nr || level == 0) {
2122                         /*
2123                          * Don't free the root -  we will detect this
2124                          * condition after our loop and return a
2125                          * positive value for caller to stop walking the tree.
2126                          */
2127                         if (level != root_level) {
2128                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
2129                                 path->locks[level] = 0;
2130
2131                                 free_extent_buffer(eb);
2132                                 path->nodes[level] = NULL;
2133                                 path->slots[level] = 0;
2134                         }
2135                 } else {
2136                         /*
2137                          * We have a valid slot to walk back down
2138                          * from. Stop here so caller can process these
2139                          * new nodes.
2140                          */
2141                         break;
2142                 }
2143
2144                 level++;
2145         }
2146
2147         eb = path->nodes[root_level];
2148         if (path->slots[root_level] >= btrfs_header_nritems(eb))
2149                 return 1;
2150
2151         return 0;
2152 }
2153
2154 /*
2155  * Helper function to trace a subtree tree block swap.
2156  *
2157  * The swap will happen in highest tree block, but there may be a lot of
2158  * tree blocks involved.
2159  *
2160  * For example:
2161  *  OO = Old tree blocks
2162  *  NN = New tree blocks allocated during balance
2163  *
2164  *           File tree (257)                  Reloc tree for 257
2165  * L2              OO                                NN
2166  *               /    \                            /    \
2167  * L1          OO      OO (a)                    OO      NN (a)
2168  *            / \     / \                       / \     / \
2169  * L0       OO   OO OO   OO                   OO   OO NN   NN
2170  *                  (b)  (c)                          (b)  (c)
2171  *
2172  * When calling qgroup_trace_extent_swap(), we will pass:
2173  * @src_eb = OO(a)
2174  * @dst_path = [ nodes[1] = NN(a), nodes[0] = NN(c) ]
2175  * @dst_level = 0
2176  * @root_level = 1
2177  *
2178  * In that case, qgroup_trace_extent_swap() will search from OO(a) to
2179  * reach OO(c), then mark both OO(c) and NN(c) as qgroup dirty.
2180  *
2181  * The main work of qgroup_trace_extent_swap() can be split into 3 parts:
2182  *
2183  * 1) Tree search from @src_eb
2184  *    It should acts as a simplified btrfs_search_slot().
2185  *    The key for search can be extracted from @dst_path->nodes[dst_level]
2186  *    (first key).
2187  *
2188  * 2) Mark the final tree blocks in @src_path and @dst_path qgroup dirty
2189  *    NOTE: In above case, OO(a) and NN(a) won't be marked qgroup dirty.
2190  *    They should be marked during previous (@dst_level = 1) iteration.
2191  *
2192  * 3) Mark file extents in leaves dirty
2193  *    We don't have good way to pick out new file extents only.
2194  *    So we still follow the old method by scanning all file extents in
2195  *    the leave.
2196  *
2197  * This function can free us from keeping two paths, thus later we only need
2198  * to care about how to iterate all new tree blocks in reloc tree.
2199  */
2200 static int qgroup_trace_extent_swap(struct btrfs_trans_handle* trans,
2201                                     struct extent_buffer *src_eb,
2202                                     struct btrfs_path *dst_path,
2203                                     int dst_level, int root_level,
2204                                     bool trace_leaf)
2205 {
2206         struct btrfs_key key;
2207         struct btrfs_path *src_path;
2208         struct btrfs_fs_info *fs_info = trans->fs_info;
2209         u32 nodesize = fs_info->nodesize;
2210         int cur_level = root_level;
2211         int ret;
2212
2213         BUG_ON(dst_level > root_level);
2214         /* Level mismatch */
2215         if (btrfs_header_level(src_eb) != root_level)
2216                 return -EINVAL;
2217
2218         src_path = btrfs_alloc_path();
2219         if (!src_path) {
2220                 ret = -ENOMEM;
2221                 goto out;
2222         }
2223
2224         if (dst_level)
2225                 btrfs_node_key_to_cpu(dst_path->nodes[dst_level], &key, 0);
2226         else
2227                 btrfs_item_key_to_cpu(dst_path->nodes[dst_level], &key, 0);
2228
2229         /* For src_path */
2230         atomic_inc(&src_eb->refs);
2231         src_path->nodes[root_level] = src_eb;
2232         src_path->slots[root_level] = dst_path->slots[root_level];
2233         src_path->locks[root_level] = 0;
2234
2235         /* A simplified version of btrfs_search_slot() */
2236         while (cur_level >= dst_level) {
2237                 struct btrfs_key src_key;
2238                 struct btrfs_key dst_key;
2239
2240                 if (src_path->nodes[cur_level] == NULL) {
2241                         struct extent_buffer *eb;
2242                         int parent_slot;
2243
2244                         eb = src_path->nodes[cur_level + 1];
2245                         parent_slot = src_path->slots[cur_level + 1];
2246
2247                         eb = btrfs_read_node_slot(eb, parent_slot);
2248                         if (IS_ERR(eb)) {
2249                                 ret = PTR_ERR(eb);
2250                                 goto out;
2251                         }
2252
2253                         src_path->nodes[cur_level] = eb;
2254
2255                         btrfs_tree_read_lock(eb);
2256                         src_path->locks[cur_level] = BTRFS_READ_LOCK;
2257                 }
2258
2259                 src_path->slots[cur_level] = dst_path->slots[cur_level];
2260                 if (cur_level) {
2261                         btrfs_node_key_to_cpu(dst_path->nodes[cur_level],
2262                                         &dst_key, dst_path->slots[cur_level]);
2263                         btrfs_node_key_to_cpu(src_path->nodes[cur_level],
2264                                         &src_key, src_path->slots[cur_level]);
2265                 } else {
2266                         btrfs_item_key_to_cpu(dst_path->nodes[cur_level],
2267                                         &dst_key, dst_path->slots[cur_level]);
2268                         btrfs_item_key_to_cpu(src_path->nodes[cur_level],
2269                                         &src_key, src_path->slots[cur_level]);
2270                 }
2271                 /* Content mismatch, something went wrong */
2272                 if (btrfs_comp_cpu_keys(&dst_key, &src_key)) {
2273                         ret = -ENOENT;
2274                         goto out;
2275                 }
2276                 cur_level--;
2277         }
2278
2279         /*
2280          * Now both @dst_path and @src_path have been populated, record the tree
2281          * blocks for qgroup accounting.
2282          */
2283         ret = btrfs_qgroup_trace_extent(trans, src_path->nodes[dst_level]->start,
2284                                         nodesize);
2285         if (ret < 0)
2286                 goto out;
2287         ret = btrfs_qgroup_trace_extent(trans, dst_path->nodes[dst_level]->start,
2288                                         nodesize);
2289         if (ret < 0)
2290                 goto out;
2291
2292         /* Record leaf file extents */
2293         if (dst_level == 0 && trace_leaf) {
2294                 ret = btrfs_qgroup_trace_leaf_items(trans, src_path->nodes[0]);
2295                 if (ret < 0)
2296                         goto out;
2297                 ret = btrfs_qgroup_trace_leaf_items(trans, dst_path->nodes[0]);
2298         }
2299 out:
2300         btrfs_free_path(src_path);
2301         return ret;
2302 }
2303
2304 /*
2305  * Helper function to do recursive generation-aware depth-first search, to
2306  * locate all new tree blocks in a subtree of reloc tree.
2307  *
2308  * E.g. (OO = Old tree blocks, NN = New tree blocks, whose gen == last_snapshot)
2309  *         reloc tree
2310  * L2         NN (a)
2311  *          /    \
2312  * L1    OO        NN (b)
2313  *      /  \      /  \
2314  * L0  OO  OO    OO  NN
2315  *               (c) (d)
2316  * If we pass:
2317  * @dst_path = [ nodes[1] = NN(b), nodes[0] = NULL ],
2318  * @cur_level = 1
2319  * @root_level = 1
2320  *
2321  * We will iterate through tree blocks NN(b), NN(d) and info qgroup to trace
2322  * above tree blocks along with their counter parts in file tree.
2323  * While during search, old tree blocks OO(c) will be skipped as tree block swap
2324  * won't affect OO(c).
2325  */
2326 static int qgroup_trace_new_subtree_blocks(struct btrfs_trans_handle* trans,
2327                                            struct extent_buffer *src_eb,
2328                                            struct btrfs_path *dst_path,
2329                                            int cur_level, int root_level,
2330                                            u64 last_snapshot, bool trace_leaf)
2331 {
2332         struct btrfs_fs_info *fs_info = trans->fs_info;
2333         struct extent_buffer *eb;
2334         bool need_cleanup = false;
2335         int ret = 0;
2336         int i;
2337
2338         /* Level sanity check */
2339         if (cur_level < 0 || cur_level >= BTRFS_MAX_LEVEL - 1 ||
2340             root_level < 0 || root_level >= BTRFS_MAX_LEVEL - 1 ||
2341             root_level < cur_level) {
2342                 btrfs_err_rl(fs_info,
2343                         "%s: bad levels, cur_level=%d root_level=%d",
2344                         __func__, cur_level, root_level);
2345                 return -EUCLEAN;
2346         }
2347
2348         /* Read the tree block if needed */
2349         if (dst_path->nodes[cur_level] == NULL) {
2350                 int parent_slot;
2351                 u64 child_gen;
2352
2353                 /*
2354                  * dst_path->nodes[root_level] must be initialized before
2355                  * calling this function.
2356                  */
2357                 if (cur_level == root_level) {
2358                         btrfs_err_rl(fs_info,
2359         "%s: dst_path->nodes[%d] not initialized, root_level=%d cur_level=%d",
2360                                 __func__, root_level, root_level, cur_level);
2361                         return -EUCLEAN;
2362                 }
2363
2364                 /*
2365                  * We need to get child blockptr/gen from parent before we can
2366                  * read it.
2367                   */
2368                 eb = dst_path->nodes[cur_level + 1];
2369                 parent_slot = dst_path->slots[cur_level + 1];
2370                 child_gen = btrfs_node_ptr_generation(eb, parent_slot);
2371
2372                 /* This node is old, no need to trace */
2373                 if (child_gen < last_snapshot)
2374                         goto out;
2375
2376                 eb = btrfs_read_node_slot(eb, parent_slot);
2377                 if (IS_ERR(eb)) {
2378                         ret = PTR_ERR(eb);
2379                         goto out;
2380                 }
2381
2382                 dst_path->nodes[cur_level] = eb;
2383                 dst_path->slots[cur_level] = 0;
2384
2385                 btrfs_tree_read_lock(eb);
2386                 dst_path->locks[cur_level] = BTRFS_READ_LOCK;
2387                 need_cleanup = true;
2388         }
2389
2390         /* Now record this tree block and its counter part for qgroups */
2391         ret = qgroup_trace_extent_swap(trans, src_eb, dst_path, cur_level,
2392                                        root_level, trace_leaf);
2393         if (ret < 0)
2394                 goto cleanup;
2395
2396         eb = dst_path->nodes[cur_level];
2397
2398         if (cur_level > 0) {
2399                 /* Iterate all child tree blocks */
2400                 for (i = 0; i < btrfs_header_nritems(eb); i++) {
2401                         /* Skip old tree blocks as they won't be swapped */
2402                         if (btrfs_node_ptr_generation(eb, i) < last_snapshot)
2403                                 continue;
2404                         dst_path->slots[cur_level] = i;
2405
2406                         /* Recursive call (at most 7 times) */
2407                         ret = qgroup_trace_new_subtree_blocks(trans, src_eb,
2408                                         dst_path, cur_level - 1, root_level,
2409                                         last_snapshot, trace_leaf);
2410                         if (ret < 0)
2411                                 goto cleanup;
2412                 }
2413         }
2414
2415 cleanup:
2416         if (need_cleanup) {
2417                 /* Clean up */
2418                 btrfs_tree_unlock_rw(dst_path->nodes[cur_level],
2419                                      dst_path->locks[cur_level]);
2420                 free_extent_buffer(dst_path->nodes[cur_level]);
2421                 dst_path->nodes[cur_level] = NULL;
2422                 dst_path->slots[cur_level] = 0;
2423                 dst_path->locks[cur_level] = 0;
2424         }
2425 out:
2426         return ret;
2427 }
2428
2429 static int qgroup_trace_subtree_swap(struct btrfs_trans_handle *trans,
2430                                 struct extent_buffer *src_eb,
2431                                 struct extent_buffer *dst_eb,
2432                                 u64 last_snapshot, bool trace_leaf)
2433 {
2434         struct btrfs_fs_info *fs_info = trans->fs_info;
2435         struct btrfs_path *dst_path = NULL;
2436         int level;
2437         int ret;
2438
2439         if (!btrfs_qgroup_full_accounting(fs_info))
2440                 return 0;
2441
2442         /* Wrong parameter order */
2443         if (btrfs_header_generation(src_eb) > btrfs_header_generation(dst_eb)) {
2444                 btrfs_err_rl(fs_info,
2445                 "%s: bad parameter order, src_gen=%llu dst_gen=%llu", __func__,
2446                              btrfs_header_generation(src_eb),
2447                              btrfs_header_generation(dst_eb));
2448                 return -EUCLEAN;
2449         }
2450
2451         if (!extent_buffer_uptodate(src_eb) || !extent_buffer_uptodate(dst_eb)) {
2452                 ret = -EIO;
2453                 goto out;
2454         }
2455
2456         level = btrfs_header_level(dst_eb);
2457         dst_path = btrfs_alloc_path();
2458         if (!dst_path) {
2459                 ret = -ENOMEM;
2460                 goto out;
2461         }
2462         /* For dst_path */
2463         atomic_inc(&dst_eb->refs);
2464         dst_path->nodes[level] = dst_eb;
2465         dst_path->slots[level] = 0;
2466         dst_path->locks[level] = 0;
2467
2468         /* Do the generation aware breadth-first search */
2469         ret = qgroup_trace_new_subtree_blocks(trans, src_eb, dst_path, level,
2470                                               level, last_snapshot, trace_leaf);
2471         if (ret < 0)
2472                 goto out;
2473         ret = 0;
2474
2475 out:
2476         btrfs_free_path(dst_path);
2477         if (ret < 0)
2478                 qgroup_mark_inconsistent(fs_info);
2479         return ret;
2480 }
2481
2482 /*
2483  * Inform qgroup to trace a whole subtree, including all its child tree
2484  * blocks and data.
2485  * The root tree block is specified by @root_eb.
2486  *
2487  * Normally used by relocation(tree block swap) and subvolume deletion.
2488  *
2489  * Return 0 for success
2490  * Return <0 for error(ENOMEM or tree search error)
2491  */
2492 int btrfs_qgroup_trace_subtree(struct btrfs_trans_handle *trans,
2493                                struct extent_buffer *root_eb,
2494                                u64 root_gen, int root_level)
2495 {
2496         struct btrfs_fs_info *fs_info = trans->fs_info;
2497         int ret = 0;
2498         int level;
2499         u8 drop_subptree_thres;
2500         struct extent_buffer *eb = root_eb;
2501         struct btrfs_path *path = NULL;
2502
2503         ASSERT(0 <= root_level && root_level < BTRFS_MAX_LEVEL);
2504         ASSERT(root_eb != NULL);
2505
2506         if (!btrfs_qgroup_full_accounting(fs_info))
2507                 return 0;
2508
2509         spin_lock(&fs_info->qgroup_lock);
2510         drop_subptree_thres = fs_info->qgroup_drop_subtree_thres;
2511         spin_unlock(&fs_info->qgroup_lock);
2512
2513         /*
2514          * This function only gets called for snapshot drop, if we hit a high
2515          * node here, it means we are going to change ownership for quite a lot
2516          * of extents, which will greatly slow down btrfs_commit_transaction().
2517          *
2518          * So here if we find a high tree here, we just skip the accounting and
2519          * mark qgroup inconsistent.
2520          */
2521         if (root_level >= drop_subptree_thres) {
2522                 qgroup_mark_inconsistent(fs_info);
2523                 return 0;
2524         }
2525
2526         if (!extent_buffer_uptodate(root_eb)) {
2527                 struct btrfs_tree_parent_check check = {
2528                         .has_first_key = false,
2529                         .transid = root_gen,
2530                         .level = root_level
2531                 };
2532
2533                 ret = btrfs_read_extent_buffer(root_eb, &check);
2534                 if (ret)
2535                         goto out;
2536         }
2537
2538         if (root_level == 0) {
2539                 ret = btrfs_qgroup_trace_leaf_items(trans, root_eb);
2540                 goto out;
2541         }
2542
2543         path = btrfs_alloc_path();
2544         if (!path)
2545                 return -ENOMEM;
2546
2547         /*
2548          * Walk down the tree.  Missing extent blocks are filled in as
2549          * we go. Metadata is accounted every time we read a new
2550          * extent block.
2551          *
2552          * When we reach a leaf, we account for file extent items in it,
2553          * walk back up the tree (adjusting slot pointers as we go)
2554          * and restart the search process.
2555          */
2556         atomic_inc(&root_eb->refs);     /* For path */
2557         path->nodes[root_level] = root_eb;
2558         path->slots[root_level] = 0;
2559         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
2560 walk_down:
2561         level = root_level;
2562         while (level >= 0) {
2563                 if (path->nodes[level] == NULL) {
2564                         int parent_slot;
2565                         u64 child_bytenr;
2566
2567                         /*
2568                          * We need to get child blockptr from parent before we
2569                          * can read it.
2570                           */
2571                         eb = path->nodes[level + 1];
2572                         parent_slot = path->slots[level + 1];
2573                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
2574
2575                         eb = btrfs_read_node_slot(eb, parent_slot);
2576                         if (IS_ERR(eb)) {
2577                                 ret = PTR_ERR(eb);
2578                                 goto out;
2579                         }
2580
2581                         path->nodes[level] = eb;
2582                         path->slots[level] = 0;
2583
2584                         btrfs_tree_read_lock(eb);
2585                         path->locks[level] = BTRFS_READ_LOCK;
2586
2587                         ret = btrfs_qgroup_trace_extent(trans, child_bytenr,
2588                                                         fs_info->nodesize);
2589                         if (ret)
2590                                 goto out;
2591                 }
2592
2593                 if (level == 0) {
2594                         ret = btrfs_qgroup_trace_leaf_items(trans,
2595                                                             path->nodes[level]);
2596                         if (ret)
2597                                 goto out;
2598
2599                         /* Nonzero return here means we completed our search */
2600                         ret = adjust_slots_upwards(path, root_level);
2601                         if (ret)
2602                                 break;
2603
2604                         /* Restart search with new slots */
2605                         goto walk_down;
2606                 }
2607
2608                 level--;
2609         }
2610
2611         ret = 0;
2612 out:
2613         btrfs_free_path(path);
2614
2615         return ret;
2616 }
2617
2618 static void qgroup_iterator_nested_add(struct list_head *head, struct btrfs_qgroup *qgroup)
2619 {
2620         if (!list_empty(&qgroup->nested_iterator))
2621                 return;
2622
2623         list_add_tail(&qgroup->nested_iterator, head);
2624 }
2625
2626 static void qgroup_iterator_nested_clean(struct list_head *head)
2627 {
2628         while (!list_empty(head)) {
2629                 struct btrfs_qgroup *qgroup;
2630
2631                 qgroup = list_first_entry(head, struct btrfs_qgroup, nested_iterator);
2632                 list_del_init(&qgroup->nested_iterator);
2633         }
2634 }
2635
2636 #define UPDATE_NEW      0
2637 #define UPDATE_OLD      1
2638 /*
2639  * Walk all of the roots that points to the bytenr and adjust their refcnts.
2640  */
2641 static void qgroup_update_refcnt(struct btrfs_fs_info *fs_info,
2642                                  struct ulist *roots, struct list_head *qgroups,
2643                                  u64 seq, int update_old)
2644 {
2645         struct ulist_node *unode;
2646         struct ulist_iterator uiter;
2647         struct btrfs_qgroup *qg;
2648
2649         if (!roots)
2650                 return;
2651         ULIST_ITER_INIT(&uiter);
2652         while ((unode = ulist_next(roots, &uiter))) {
2653                 LIST_HEAD(tmp);
2654
2655                 qg = find_qgroup_rb(fs_info, unode->val);
2656                 if (!qg)
2657                         continue;
2658
2659                 qgroup_iterator_nested_add(qgroups, qg);
2660                 qgroup_iterator_add(&tmp, qg);
2661                 list_for_each_entry(qg, &tmp, iterator) {
2662                         struct btrfs_qgroup_list *glist;
2663
2664                         if (update_old)
2665                                 btrfs_qgroup_update_old_refcnt(qg, seq, 1);
2666                         else
2667                                 btrfs_qgroup_update_new_refcnt(qg, seq, 1);
2668
2669                         list_for_each_entry(glist, &qg->groups, next_group) {
2670                                 qgroup_iterator_nested_add(qgroups, glist->group);
2671                                 qgroup_iterator_add(&tmp, glist->group);
2672                         }
2673                 }
2674                 qgroup_iterator_clean(&tmp);
2675         }
2676 }
2677
2678 /*
2679  * Update qgroup rfer/excl counters.
2680  * Rfer update is easy, codes can explain themselves.
2681  *
2682  * Excl update is tricky, the update is split into 2 parts.
2683  * Part 1: Possible exclusive <-> sharing detect:
2684  *      |       A       |       !A      |
2685  *  -------------------------------------
2686  *  B   |       *       |       -       |
2687  *  -------------------------------------
2688  *  !B  |       +       |       **      |
2689  *  -------------------------------------
2690  *
2691  * Conditions:
2692  * A:   cur_old_roots < nr_old_roots    (not exclusive before)
2693  * !A:  cur_old_roots == nr_old_roots   (possible exclusive before)
2694  * B:   cur_new_roots < nr_new_roots    (not exclusive now)
2695  * !B:  cur_new_roots == nr_new_roots   (possible exclusive now)
2696  *
2697  * Results:
2698  * +: Possible sharing -> exclusive     -: Possible exclusive -> sharing
2699  * *: Definitely not changed.           **: Possible unchanged.
2700  *
2701  * For !A and !B condition, the exception is cur_old/new_roots == 0 case.
2702  *
2703  * To make the logic clear, we first use condition A and B to split
2704  * combination into 4 results.
2705  *
2706  * Then, for result "+" and "-", check old/new_roots == 0 case, as in them
2707  * only on variant maybe 0.
2708  *
2709  * Lastly, check result **, since there are 2 variants maybe 0, split them
2710  * again(2x2).
2711  * But this time we don't need to consider other things, the codes and logic
2712  * is easy to understand now.
2713  */
2714 static void qgroup_update_counters(struct btrfs_fs_info *fs_info,
2715                                    struct list_head *qgroups, u64 nr_old_roots,
2716                                    u64 nr_new_roots, u64 num_bytes, u64 seq)
2717 {
2718         struct btrfs_qgroup *qg;
2719
2720         list_for_each_entry(qg, qgroups, nested_iterator) {
2721                 u64 cur_new_count, cur_old_count;
2722                 bool dirty = false;
2723
2724                 cur_old_count = btrfs_qgroup_get_old_refcnt(qg, seq);
2725                 cur_new_count = btrfs_qgroup_get_new_refcnt(qg, seq);
2726
2727                 trace_qgroup_update_counters(fs_info, qg, cur_old_count,
2728                                              cur_new_count);
2729
2730                 /* Rfer update part */
2731                 if (cur_old_count == 0 && cur_new_count > 0) {
2732                         qg->rfer += num_bytes;
2733                         qg->rfer_cmpr += num_bytes;
2734                         dirty = true;
2735                 }
2736                 if (cur_old_count > 0 && cur_new_count == 0) {
2737                         qg->rfer -= num_bytes;
2738                         qg->rfer_cmpr -= num_bytes;
2739                         dirty = true;
2740                 }
2741
2742                 /* Excl update part */
2743                 /* Exclusive/none -> shared case */
2744                 if (cur_old_count == nr_old_roots &&
2745                     cur_new_count < nr_new_roots) {
2746                         /* Exclusive -> shared */
2747                         if (cur_old_count != 0) {
2748                                 qg->excl -= num_bytes;
2749                                 qg->excl_cmpr -= num_bytes;
2750                                 dirty = true;
2751                         }
2752                 }
2753
2754                 /* Shared -> exclusive/none case */
2755                 if (cur_old_count < nr_old_roots &&
2756                     cur_new_count == nr_new_roots) {
2757                         /* Shared->exclusive */
2758                         if (cur_new_count != 0) {
2759                                 qg->excl += num_bytes;
2760                                 qg->excl_cmpr += num_bytes;
2761                                 dirty = true;
2762                         }
2763                 }
2764
2765                 /* Exclusive/none -> exclusive/none case */
2766                 if (cur_old_count == nr_old_roots &&
2767                     cur_new_count == nr_new_roots) {
2768                         if (cur_old_count == 0) {
2769                                 /* None -> exclusive/none */
2770
2771                                 if (cur_new_count != 0) {
2772                                         /* None -> exclusive */
2773                                         qg->excl += num_bytes;
2774                                         qg->excl_cmpr += num_bytes;
2775                                         dirty = true;
2776                                 }
2777                                 /* None -> none, nothing changed */
2778                         } else {
2779                                 /* Exclusive -> exclusive/none */
2780
2781                                 if (cur_new_count == 0) {
2782                                         /* Exclusive -> none */
2783                                         qg->excl -= num_bytes;
2784                                         qg->excl_cmpr -= num_bytes;
2785                                         dirty = true;
2786                                 }
2787                                 /* Exclusive -> exclusive, nothing changed */
2788                         }
2789                 }
2790
2791                 if (dirty)
2792                         qgroup_dirty(fs_info, qg);
2793         }
2794 }
2795
2796 /*
2797  * Check if the @roots potentially is a list of fs tree roots
2798  *
2799  * Return 0 for definitely not a fs/subvol tree roots ulist
2800  * Return 1 for possible fs/subvol tree roots in the list (considering an empty
2801  *          one as well)
2802  */
2803 static int maybe_fs_roots(struct ulist *roots)
2804 {
2805         struct ulist_node *unode;
2806         struct ulist_iterator uiter;
2807
2808         /* Empty one, still possible for fs roots */
2809         if (!roots || roots->nnodes == 0)
2810                 return 1;
2811
2812         ULIST_ITER_INIT(&uiter);
2813         unode = ulist_next(roots, &uiter);
2814         if (!unode)
2815                 return 1;
2816
2817         /*
2818          * If it contains fs tree roots, then it must belong to fs/subvol
2819          * trees.
2820          * If it contains a non-fs tree, it won't be shared with fs/subvol trees.
2821          */
2822         return is_fstree(unode->val);
2823 }
2824
2825 int btrfs_qgroup_account_extent(struct btrfs_trans_handle *trans, u64 bytenr,
2826                                 u64 num_bytes, struct ulist *old_roots,
2827                                 struct ulist *new_roots)
2828 {
2829         struct btrfs_fs_info *fs_info = trans->fs_info;
2830         LIST_HEAD(qgroups);
2831         u64 seq;
2832         u64 nr_new_roots = 0;
2833         u64 nr_old_roots = 0;
2834         int ret = 0;
2835
2836         /*
2837          * If quotas get disabled meanwhile, the resources need to be freed and
2838          * we can't just exit here.
2839          */
2840         if (!btrfs_qgroup_full_accounting(fs_info) ||
2841             fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING)
2842                 goto out_free;
2843
2844         if (new_roots) {
2845                 if (!maybe_fs_roots(new_roots))
2846                         goto out_free;
2847                 nr_new_roots = new_roots->nnodes;
2848         }
2849         if (old_roots) {
2850                 if (!maybe_fs_roots(old_roots))
2851                         goto out_free;
2852                 nr_old_roots = old_roots->nnodes;
2853         }
2854
2855         /* Quick exit, either not fs tree roots, or won't affect any qgroup */
2856         if (nr_old_roots == 0 && nr_new_roots == 0)
2857                 goto out_free;
2858
2859         trace_btrfs_qgroup_account_extent(fs_info, trans->transid, bytenr,
2860                                         num_bytes, nr_old_roots, nr_new_roots);
2861
2862         mutex_lock(&fs_info->qgroup_rescan_lock);
2863         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
2864                 if (fs_info->qgroup_rescan_progress.objectid <= bytenr) {
2865                         mutex_unlock(&fs_info->qgroup_rescan_lock);
2866                         ret = 0;
2867                         goto out_free;
2868                 }
2869         }
2870         mutex_unlock(&fs_info->qgroup_rescan_lock);
2871
2872         spin_lock(&fs_info->qgroup_lock);
2873         seq = fs_info->qgroup_seq;
2874
2875         /* Update old refcnts using old_roots */
2876         qgroup_update_refcnt(fs_info, old_roots, &qgroups, seq, UPDATE_OLD);
2877
2878         /* Update new refcnts using new_roots */
2879         qgroup_update_refcnt(fs_info, new_roots, &qgroups, seq, UPDATE_NEW);
2880
2881         qgroup_update_counters(fs_info, &qgroups, nr_old_roots, nr_new_roots,
2882                                num_bytes, seq);
2883
2884         /*
2885          * We're done using the iterator, release all its qgroups while holding
2886          * fs_info->qgroup_lock so that we don't race with btrfs_remove_qgroup()
2887          * and trigger use-after-free accesses to qgroups.
2888          */
2889         qgroup_iterator_nested_clean(&qgroups);
2890
2891         /*
2892          * Bump qgroup_seq to avoid seq overlap
2893          */
2894         fs_info->qgroup_seq += max(nr_old_roots, nr_new_roots) + 1;
2895         spin_unlock(&fs_info->qgroup_lock);
2896 out_free:
2897         ulist_free(old_roots);
2898         ulist_free(new_roots);
2899         return ret;
2900 }
2901
2902 int btrfs_qgroup_account_extents(struct btrfs_trans_handle *trans)
2903 {
2904         struct btrfs_fs_info *fs_info = trans->fs_info;
2905         struct btrfs_qgroup_extent_record *record;
2906         struct btrfs_delayed_ref_root *delayed_refs;
2907         struct ulist *new_roots = NULL;
2908         struct rb_node *node;
2909         u64 num_dirty_extents = 0;
2910         u64 qgroup_to_skip;
2911         int ret = 0;
2912
2913         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE)
2914                 return 0;
2915
2916         delayed_refs = &trans->transaction->delayed_refs;
2917         qgroup_to_skip = delayed_refs->qgroup_to_skip;
2918         while ((node = rb_first(&delayed_refs->dirty_extent_root))) {
2919                 record = rb_entry(node, struct btrfs_qgroup_extent_record,
2920                                   node);
2921
2922                 num_dirty_extents++;
2923                 trace_btrfs_qgroup_account_extents(fs_info, record);
2924
2925                 if (!ret && !(fs_info->qgroup_flags &
2926                               BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING)) {
2927                         struct btrfs_backref_walk_ctx ctx = { 0 };
2928
2929                         ctx.bytenr = record->bytenr;
2930                         ctx.fs_info = fs_info;
2931
2932                         /*
2933                          * Old roots should be searched when inserting qgroup
2934                          * extent record.
2935                          *
2936                          * But for INCONSISTENT (NO_ACCOUNTING) -> rescan case,
2937                          * we may have some record inserted during
2938                          * NO_ACCOUNTING (thus no old_roots populated), but
2939                          * later we start rescan, which clears NO_ACCOUNTING,
2940                          * leaving some inserted records without old_roots
2941                          * populated.
2942                          *
2943                          * Those cases are rare and should not cause too much
2944                          * time spent during commit_transaction().
2945                          */
2946                         if (!record->old_roots) {
2947                                 /* Search commit root to find old_roots */
2948                                 ret = btrfs_find_all_roots(&ctx, false);
2949                                 if (ret < 0)
2950                                         goto cleanup;
2951                                 record->old_roots = ctx.roots;
2952                                 ctx.roots = NULL;
2953                         }
2954
2955                         /*
2956                          * Use BTRFS_SEQ_LAST as time_seq to do special search,
2957                          * which doesn't lock tree or delayed_refs and search
2958                          * current root. It's safe inside commit_transaction().
2959                          */
2960                         ctx.trans = trans;
2961                         ctx.time_seq = BTRFS_SEQ_LAST;
2962                         ret = btrfs_find_all_roots(&ctx, false);
2963                         if (ret < 0)
2964                                 goto cleanup;
2965                         new_roots = ctx.roots;
2966                         if (qgroup_to_skip) {
2967                                 ulist_del(new_roots, qgroup_to_skip, 0);
2968                                 ulist_del(record->old_roots, qgroup_to_skip,
2969                                           0);
2970                         }
2971                         ret = btrfs_qgroup_account_extent(trans, record->bytenr,
2972                                                           record->num_bytes,
2973                                                           record->old_roots,
2974                                                           new_roots);
2975                         record->old_roots = NULL;
2976                         new_roots = NULL;
2977                 }
2978                 /* Free the reserved data space */
2979                 btrfs_qgroup_free_refroot(fs_info,
2980                                 record->data_rsv_refroot,
2981                                 record->data_rsv,
2982                                 BTRFS_QGROUP_RSV_DATA);
2983 cleanup:
2984                 ulist_free(record->old_roots);
2985                 ulist_free(new_roots);
2986                 new_roots = NULL;
2987                 rb_erase(node, &delayed_refs->dirty_extent_root);
2988                 kfree(record);
2989
2990         }
2991         trace_qgroup_num_dirty_extents(fs_info, trans->transid,
2992                                        num_dirty_extents);
2993         return ret;
2994 }
2995
2996 /*
2997  * Writes all changed qgroups to disk.
2998  * Called by the transaction commit path and the qgroup assign ioctl.
2999  */
3000 int btrfs_run_qgroups(struct btrfs_trans_handle *trans)
3001 {
3002         struct btrfs_fs_info *fs_info = trans->fs_info;
3003         int ret = 0;
3004
3005         /*
3006          * In case we are called from the qgroup assign ioctl, assert that we
3007          * are holding the qgroup_ioctl_lock, otherwise we can race with a quota
3008          * disable operation (ioctl) and access a freed quota root.
3009          */
3010         if (trans->transaction->state != TRANS_STATE_COMMIT_DOING)
3011                 lockdep_assert_held(&fs_info->qgroup_ioctl_lock);
3012
3013         if (!fs_info->quota_root)
3014                 return ret;
3015
3016         spin_lock(&fs_info->qgroup_lock);
3017         while (!list_empty(&fs_info->dirty_qgroups)) {
3018                 struct btrfs_qgroup *qgroup;
3019                 qgroup = list_first_entry(&fs_info->dirty_qgroups,
3020                                           struct btrfs_qgroup, dirty);
3021                 list_del_init(&qgroup->dirty);
3022                 spin_unlock(&fs_info->qgroup_lock);
3023                 ret = update_qgroup_info_item(trans, qgroup);
3024                 if (ret)
3025                         qgroup_mark_inconsistent(fs_info);
3026                 ret = update_qgroup_limit_item(trans, qgroup);
3027                 if (ret)
3028                         qgroup_mark_inconsistent(fs_info);
3029                 spin_lock(&fs_info->qgroup_lock);
3030         }
3031         if (btrfs_qgroup_enabled(fs_info))
3032                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_ON;
3033         else
3034                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_ON;
3035         spin_unlock(&fs_info->qgroup_lock);
3036
3037         ret = update_qgroup_status_item(trans);
3038         if (ret)
3039                 qgroup_mark_inconsistent(fs_info);
3040
3041         return ret;
3042 }
3043
3044 int btrfs_qgroup_check_inherit(struct btrfs_fs_info *fs_info,
3045                                struct btrfs_qgroup_inherit *inherit,
3046                                size_t size)
3047 {
3048         if (!btrfs_qgroup_enabled(fs_info))
3049                 return 0;
3050         if (inherit->flags & ~BTRFS_QGROUP_INHERIT_FLAGS_SUPP)
3051                 return -EOPNOTSUPP;
3052         if (size < sizeof(*inherit) || size > PAGE_SIZE)
3053                 return -EINVAL;
3054
3055         /*
3056          * In the past we allowed btrfs_qgroup_inherit to specify to copy
3057          * rfer/excl numbers directly from other qgroups.  This behavior has
3058          * been disabled in userspace for a very long time, but here we should
3059          * also disable it in kernel, as this behavior is known to mark qgroup
3060          * inconsistent, and a rescan would wipe out the changes anyway.
3061          *
3062          * Reject any btrfs_qgroup_inherit with num_ref_copies or num_excl_copies.
3063          */
3064         if (inherit->num_ref_copies > 0 || inherit->num_excl_copies > 0)
3065                 return -EINVAL;
3066
3067         if (inherit->num_qgroups > PAGE_SIZE)
3068                 return -EINVAL;
3069
3070         if (size != struct_size(inherit, qgroups, inherit->num_qgroups))
3071                 return -EINVAL;
3072
3073         /*
3074          * Now check all the remaining qgroups, they should all:
3075          *
3076          * - Exist
3077          * - Be higher level qgroups.
3078          */
3079         for (int i = 0; i < inherit->num_qgroups; i++) {
3080                 struct btrfs_qgroup *qgroup;
3081                 u64 qgroupid = inherit->qgroups[i];
3082
3083                 if (btrfs_qgroup_level(qgroupid) == 0)
3084                         return -EINVAL;
3085
3086                 spin_lock(&fs_info->qgroup_lock);
3087                 qgroup = find_qgroup_rb(fs_info, qgroupid);
3088                 if (!qgroup) {
3089                         spin_unlock(&fs_info->qgroup_lock);
3090                         return -ENOENT;
3091                 }
3092                 spin_unlock(&fs_info->qgroup_lock);
3093         }
3094         return 0;
3095 }
3096
3097 static int qgroup_auto_inherit(struct btrfs_fs_info *fs_info,
3098                                u64 inode_rootid,
3099                                struct btrfs_qgroup_inherit **inherit)
3100 {
3101         int i = 0;
3102         u64 num_qgroups = 0;
3103         struct btrfs_qgroup *inode_qg;
3104         struct btrfs_qgroup_list *qg_list;
3105         struct btrfs_qgroup_inherit *res;
3106         size_t struct_sz;
3107         u64 *qgids;
3108
3109         if (*inherit)
3110                 return -EEXIST;
3111
3112         inode_qg = find_qgroup_rb(fs_info, inode_rootid);
3113         if (!inode_qg)
3114                 return -ENOENT;
3115
3116         num_qgroups = list_count_nodes(&inode_qg->groups);
3117
3118         if (!num_qgroups)
3119                 return 0;
3120
3121         struct_sz = struct_size(res, qgroups, num_qgroups);
3122         if (struct_sz == SIZE_MAX)
3123                 return -ERANGE;
3124
3125         res = kzalloc(struct_sz, GFP_NOFS);
3126         if (!res)
3127                 return -ENOMEM;
3128         res->num_qgroups = num_qgroups;
3129         qgids = res->qgroups;
3130
3131         list_for_each_entry(qg_list, &inode_qg->groups, next_group)
3132                 qgids[i] = qg_list->group->qgroupid;
3133
3134         *inherit = res;
3135         return 0;
3136 }
3137
3138 /*
3139  * Check if we can skip rescan when inheriting qgroups.  If @src has a single
3140  * @parent, and that @parent is owning all its bytes exclusively, we can skip
3141  * the full rescan, by just adding nodesize to the @parent's excl/rfer.
3142  *
3143  * Return <0 for fatal errors (like srcid/parentid has no qgroup).
3144  * Return 0 if a quick inherit is done.
3145  * Return >0 if a quick inherit is not possible, and a full rescan is needed.
3146  */
3147 static int qgroup_snapshot_quick_inherit(struct btrfs_fs_info *fs_info,
3148                                          u64 srcid, u64 parentid)
3149 {
3150         struct btrfs_qgroup *src;
3151         struct btrfs_qgroup *parent;
3152         struct btrfs_qgroup_list *list;
3153         int nr_parents = 0;
3154
3155         src = find_qgroup_rb(fs_info, srcid);
3156         if (!src)
3157                 return -ENOENT;
3158         parent = find_qgroup_rb(fs_info, parentid);
3159         if (!parent)
3160                 return -ENOENT;
3161
3162         /*
3163          * Source has no parent qgroup, but our new qgroup would have one.
3164          * Qgroup numbers would become inconsistent.
3165          */
3166         if (list_empty(&src->groups))
3167                 return 1;
3168
3169         list_for_each_entry(list, &src->groups, next_group) {
3170                 /* The parent is not the same, quick update is not possible. */
3171                 if (list->group->qgroupid != parentid)
3172                         return 1;
3173                 nr_parents++;
3174                 /*
3175                  * More than one parent qgroup, we can't be sure about accounting
3176                  * consistency.
3177                  */
3178                 if (nr_parents > 1)
3179                         return 1;
3180         }
3181
3182         /*
3183          * The parent is not exclusively owning all its bytes.  We're not sure
3184          * if the source has any bytes not fully owned by the parent.
3185          */
3186         if (parent->excl != parent->rfer)
3187                 return 1;
3188
3189         parent->excl += fs_info->nodesize;
3190         parent->rfer += fs_info->nodesize;
3191         return 0;
3192 }
3193
3194 /*
3195  * Copy the accounting information between qgroups. This is necessary
3196  * when a snapshot or a subvolume is created. Throwing an error will
3197  * cause a transaction abort so we take extra care here to only error
3198  * when a readonly fs is a reasonable outcome.
3199  */
3200 int btrfs_qgroup_inherit(struct btrfs_trans_handle *trans, u64 srcid,
3201                          u64 objectid, u64 inode_rootid,
3202                          struct btrfs_qgroup_inherit *inherit)
3203 {
3204         int ret = 0;
3205         int i;
3206         u64 *i_qgroups;
3207         bool committing = false;
3208         struct btrfs_fs_info *fs_info = trans->fs_info;
3209         struct btrfs_root *quota_root;
3210         struct btrfs_qgroup *srcgroup;
3211         struct btrfs_qgroup *dstgroup;
3212         struct btrfs_qgroup *prealloc;
3213         struct btrfs_qgroup_list **qlist_prealloc = NULL;
3214         bool free_inherit = false;
3215         bool need_rescan = false;
3216         u32 level_size = 0;
3217         u64 nums;
3218
3219         prealloc = kzalloc(sizeof(*prealloc), GFP_NOFS);
3220         if (!prealloc)
3221                 return -ENOMEM;
3222
3223         /*
3224          * There are only two callers of this function.
3225          *
3226          * One in create_subvol() in the ioctl context, which needs to hold
3227          * the qgroup_ioctl_lock.
3228          *
3229          * The other one in create_pending_snapshot() where no other qgroup
3230          * code can modify the fs as they all need to either start a new trans
3231          * or hold a trans handler, thus we don't need to hold
3232          * qgroup_ioctl_lock.
3233          * This would avoid long and complex lock chain and make lockdep happy.
3234          */
3235         spin_lock(&fs_info->trans_lock);
3236         if (trans->transaction->state == TRANS_STATE_COMMIT_DOING)
3237                 committing = true;
3238         spin_unlock(&fs_info->trans_lock);
3239
3240         if (!committing)
3241                 mutex_lock(&fs_info->qgroup_ioctl_lock);
3242         if (!btrfs_qgroup_enabled(fs_info))
3243                 goto out;
3244
3245         quota_root = fs_info->quota_root;
3246         if (!quota_root) {
3247                 ret = -EINVAL;
3248                 goto out;
3249         }
3250
3251         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE && !inherit) {
3252                 ret = qgroup_auto_inherit(fs_info, inode_rootid, &inherit);
3253                 if (ret)
3254                         goto out;
3255                 free_inherit = true;
3256         }
3257
3258         if (inherit) {
3259                 i_qgroups = (u64 *)(inherit + 1);
3260                 nums = inherit->num_qgroups + 2 * inherit->num_ref_copies +
3261                        2 * inherit->num_excl_copies;
3262                 for (i = 0; i < nums; ++i) {
3263                         srcgroup = find_qgroup_rb(fs_info, *i_qgroups);
3264
3265                         /*
3266                          * Zero out invalid groups so we can ignore
3267                          * them later.
3268                          */
3269                         if (!srcgroup ||
3270                             ((srcgroup->qgroupid >> 48) <= (objectid >> 48)))
3271                                 *i_qgroups = 0ULL;
3272
3273                         ++i_qgroups;
3274                 }
3275         }
3276
3277         /*
3278          * create a tracking group for the subvol itself
3279          */
3280         ret = add_qgroup_item(trans, quota_root, objectid);
3281         if (ret)
3282                 goto out;
3283
3284         /*
3285          * add qgroup to all inherited groups
3286          */
3287         if (inherit) {
3288                 i_qgroups = (u64 *)(inherit + 1);
3289                 for (i = 0; i < inherit->num_qgroups; ++i, ++i_qgroups) {
3290                         if (*i_qgroups == 0)
3291                                 continue;
3292                         ret = add_qgroup_relation_item(trans, objectid,
3293                                                        *i_qgroups);
3294                         if (ret && ret != -EEXIST)
3295                                 goto out;
3296                         ret = add_qgroup_relation_item(trans, *i_qgroups,
3297                                                        objectid);
3298                         if (ret && ret != -EEXIST)
3299                                 goto out;
3300                 }
3301                 ret = 0;
3302
3303                 qlist_prealloc = kcalloc(inherit->num_qgroups,
3304                                          sizeof(struct btrfs_qgroup_list *),
3305                                          GFP_NOFS);
3306                 if (!qlist_prealloc) {
3307                         ret = -ENOMEM;
3308                         goto out;
3309                 }
3310                 for (int i = 0; i < inherit->num_qgroups; i++) {
3311                         qlist_prealloc[i] = kzalloc(sizeof(struct btrfs_qgroup_list),
3312                                                     GFP_NOFS);
3313                         if (!qlist_prealloc[i]) {
3314                                 ret = -ENOMEM;
3315                                 goto out;
3316                         }
3317                 }
3318         }
3319
3320         spin_lock(&fs_info->qgroup_lock);
3321
3322         dstgroup = add_qgroup_rb(fs_info, prealloc, objectid);
3323         prealloc = NULL;
3324
3325         if (inherit && inherit->flags & BTRFS_QGROUP_INHERIT_SET_LIMITS) {
3326                 dstgroup->lim_flags = inherit->lim.flags;
3327                 dstgroup->max_rfer = inherit->lim.max_rfer;
3328                 dstgroup->max_excl = inherit->lim.max_excl;
3329                 dstgroup->rsv_rfer = inherit->lim.rsv_rfer;
3330                 dstgroup->rsv_excl = inherit->lim.rsv_excl;
3331
3332                 qgroup_dirty(fs_info, dstgroup);
3333         }
3334
3335         if (srcid && btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_FULL) {
3336                 srcgroup = find_qgroup_rb(fs_info, srcid);
3337                 if (!srcgroup)
3338                         goto unlock;
3339
3340                 /*
3341                  * We call inherit after we clone the root in order to make sure
3342                  * our counts don't go crazy, so at this point the only
3343                  * difference between the two roots should be the root node.
3344                  */
3345                 level_size = fs_info->nodesize;
3346                 dstgroup->rfer = srcgroup->rfer;
3347                 dstgroup->rfer_cmpr = srcgroup->rfer_cmpr;
3348                 dstgroup->excl = level_size;
3349                 dstgroup->excl_cmpr = level_size;
3350                 srcgroup->excl = level_size;
3351                 srcgroup->excl_cmpr = level_size;
3352
3353                 /* inherit the limit info */
3354                 dstgroup->lim_flags = srcgroup->lim_flags;
3355                 dstgroup->max_rfer = srcgroup->max_rfer;
3356                 dstgroup->max_excl = srcgroup->max_excl;
3357                 dstgroup->rsv_rfer = srcgroup->rsv_rfer;
3358                 dstgroup->rsv_excl = srcgroup->rsv_excl;
3359
3360                 qgroup_dirty(fs_info, dstgroup);
3361                 qgroup_dirty(fs_info, srcgroup);
3362
3363                 /*
3364                  * If the source qgroup has parent but the new one doesn't,
3365                  * we need a full rescan.
3366                  */
3367                 if (!inherit && !list_empty(&srcgroup->groups))
3368                         need_rescan = true;
3369         }
3370
3371         if (!inherit)
3372                 goto unlock;
3373
3374         i_qgroups = (u64 *)(inherit + 1);
3375         for (i = 0; i < inherit->num_qgroups; ++i) {
3376                 if (*i_qgroups) {
3377                         ret = add_relation_rb(fs_info, qlist_prealloc[i], objectid,
3378                                               *i_qgroups);
3379                         qlist_prealloc[i] = NULL;
3380                         if (ret)
3381                                 goto unlock;
3382                 }
3383                 if (srcid) {
3384                         /* Check if we can do a quick inherit. */
3385                         ret = qgroup_snapshot_quick_inherit(fs_info, srcid, *i_qgroups);
3386                         if (ret < 0)
3387                                 goto unlock;
3388                         if (ret > 0)
3389                                 need_rescan = true;
3390                         ret = 0;
3391                 }
3392                 ++i_qgroups;
3393         }
3394
3395         for (i = 0; i <  inherit->num_ref_copies; ++i, i_qgroups += 2) {
3396                 struct btrfs_qgroup *src;
3397                 struct btrfs_qgroup *dst;
3398
3399                 if (!i_qgroups[0] || !i_qgroups[1])
3400                         continue;
3401
3402                 src = find_qgroup_rb(fs_info, i_qgroups[0]);
3403                 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
3404
3405                 if (!src || !dst) {
3406                         ret = -EINVAL;
3407                         goto unlock;
3408                 }
3409
3410                 dst->rfer = src->rfer - level_size;
3411                 dst->rfer_cmpr = src->rfer_cmpr - level_size;
3412
3413                 /* Manually tweaking numbers certainly needs a rescan */
3414                 need_rescan = true;
3415         }
3416         for (i = 0; i <  inherit->num_excl_copies; ++i, i_qgroups += 2) {
3417                 struct btrfs_qgroup *src;
3418                 struct btrfs_qgroup *dst;
3419
3420                 if (!i_qgroups[0] || !i_qgroups[1])
3421                         continue;
3422
3423                 src = find_qgroup_rb(fs_info, i_qgroups[0]);
3424                 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
3425
3426                 if (!src || !dst) {
3427                         ret = -EINVAL;
3428                         goto unlock;
3429                 }
3430
3431                 dst->excl = src->excl + level_size;
3432                 dst->excl_cmpr = src->excl_cmpr + level_size;
3433                 need_rescan = true;
3434         }
3435
3436 unlock:
3437         spin_unlock(&fs_info->qgroup_lock);
3438         if (!ret)
3439                 ret = btrfs_sysfs_add_one_qgroup(fs_info, dstgroup);
3440 out:
3441         if (!committing)
3442                 mutex_unlock(&fs_info->qgroup_ioctl_lock);
3443         if (need_rescan)
3444                 qgroup_mark_inconsistent(fs_info);
3445         if (qlist_prealloc) {
3446                 for (int i = 0; i < inherit->num_qgroups; i++)
3447                         kfree(qlist_prealloc[i]);
3448                 kfree(qlist_prealloc);
3449         }
3450         if (free_inherit)
3451                 kfree(inherit);
3452         kfree(prealloc);
3453         return ret;
3454 }
3455
3456 static bool qgroup_check_limits(const struct btrfs_qgroup *qg, u64 num_bytes)
3457 {
3458         if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_RFER) &&
3459             qgroup_rsv_total(qg) + (s64)qg->rfer + num_bytes > qg->max_rfer)
3460                 return false;
3461
3462         if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) &&
3463             qgroup_rsv_total(qg) + (s64)qg->excl + num_bytes > qg->max_excl)
3464                 return false;
3465
3466         return true;
3467 }
3468
3469 static int qgroup_reserve(struct btrfs_root *root, u64 num_bytes, bool enforce,
3470                           enum btrfs_qgroup_rsv_type type)
3471 {
3472         struct btrfs_qgroup *qgroup;
3473         struct btrfs_fs_info *fs_info = root->fs_info;
3474         u64 ref_root = root->root_key.objectid;
3475         int ret = 0;
3476         LIST_HEAD(qgroup_list);
3477
3478         if (!is_fstree(ref_root))
3479                 return 0;
3480
3481         if (num_bytes == 0)
3482                 return 0;
3483
3484         if (test_bit(BTRFS_FS_QUOTA_OVERRIDE, &fs_info->flags) &&
3485             capable(CAP_SYS_RESOURCE))
3486                 enforce = false;
3487
3488         spin_lock(&fs_info->qgroup_lock);
3489         if (!fs_info->quota_root)
3490                 goto out;
3491
3492         qgroup = find_qgroup_rb(fs_info, ref_root);
3493         if (!qgroup)
3494                 goto out;
3495
3496         qgroup_iterator_add(&qgroup_list, qgroup);
3497         list_for_each_entry(qgroup, &qgroup_list, iterator) {
3498                 struct btrfs_qgroup_list *glist;
3499
3500                 if (enforce && !qgroup_check_limits(qgroup, num_bytes)) {
3501                         ret = -EDQUOT;
3502                         goto out;
3503                 }
3504
3505                 list_for_each_entry(glist, &qgroup->groups, next_group)
3506                         qgroup_iterator_add(&qgroup_list, glist->group);
3507         }
3508
3509         ret = 0;
3510         /*
3511          * no limits exceeded, now record the reservation into all qgroups
3512          */
3513         list_for_each_entry(qgroup, &qgroup_list, iterator)
3514                 qgroup_rsv_add(fs_info, qgroup, num_bytes, type);
3515
3516 out:
3517         qgroup_iterator_clean(&qgroup_list);
3518         spin_unlock(&fs_info->qgroup_lock);
3519         return ret;
3520 }
3521
3522 /*
3523  * Free @num_bytes of reserved space with @type for qgroup.  (Normally level 0
3524  * qgroup).
3525  *
3526  * Will handle all higher level qgroup too.
3527  *
3528  * NOTE: If @num_bytes is (u64)-1, this means to free all bytes of this qgroup.
3529  * This special case is only used for META_PERTRANS type.
3530  */
3531 void btrfs_qgroup_free_refroot(struct btrfs_fs_info *fs_info,
3532                                u64 ref_root, u64 num_bytes,
3533                                enum btrfs_qgroup_rsv_type type)
3534 {
3535         struct btrfs_qgroup *qgroup;
3536         LIST_HEAD(qgroup_list);
3537
3538         if (!is_fstree(ref_root))
3539                 return;
3540
3541         if (num_bytes == 0)
3542                 return;
3543
3544         if (num_bytes == (u64)-1 && type != BTRFS_QGROUP_RSV_META_PERTRANS) {
3545                 WARN(1, "%s: Invalid type to free", __func__);
3546                 return;
3547         }
3548         spin_lock(&fs_info->qgroup_lock);
3549
3550         if (!fs_info->quota_root)
3551                 goto out;
3552
3553         qgroup = find_qgroup_rb(fs_info, ref_root);
3554         if (!qgroup)
3555                 goto out;
3556
3557         if (num_bytes == (u64)-1)
3558                 /*
3559                  * We're freeing all pertrans rsv, get reserved value from
3560                  * level 0 qgroup as real num_bytes to free.
3561                  */
3562                 num_bytes = qgroup->rsv.values[type];
3563
3564         qgroup_iterator_add(&qgroup_list, qgroup);
3565         list_for_each_entry(qgroup, &qgroup_list, iterator) {
3566                 struct btrfs_qgroup_list *glist;
3567
3568                 qgroup_rsv_release(fs_info, qgroup, num_bytes, type);
3569                 list_for_each_entry(glist, &qgroup->groups, next_group) {
3570                         qgroup_iterator_add(&qgroup_list, glist->group);
3571                 }
3572         }
3573 out:
3574         qgroup_iterator_clean(&qgroup_list);
3575         spin_unlock(&fs_info->qgroup_lock);
3576 }
3577
3578 /*
3579  * Check if the leaf is the last leaf. Which means all node pointers
3580  * are at their last position.
3581  */
3582 static bool is_last_leaf(struct btrfs_path *path)
3583 {
3584         int i;
3585
3586         for (i = 1; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
3587                 if (path->slots[i] != btrfs_header_nritems(path->nodes[i]) - 1)
3588                         return false;
3589         }
3590         return true;
3591 }
3592
3593 /*
3594  * returns < 0 on error, 0 when more leafs are to be scanned.
3595  * returns 1 when done.
3596  */
3597 static int qgroup_rescan_leaf(struct btrfs_trans_handle *trans,
3598                               struct btrfs_path *path)
3599 {
3600         struct btrfs_fs_info *fs_info = trans->fs_info;
3601         struct btrfs_root *extent_root;
3602         struct btrfs_key found;
3603         struct extent_buffer *scratch_leaf = NULL;
3604         u64 num_bytes;
3605         bool done;
3606         int slot;
3607         int ret;
3608
3609         if (!btrfs_qgroup_full_accounting(fs_info))
3610                 return 1;
3611
3612         mutex_lock(&fs_info->qgroup_rescan_lock);
3613         extent_root = btrfs_extent_root(fs_info,
3614                                 fs_info->qgroup_rescan_progress.objectid);
3615         ret = btrfs_search_slot_for_read(extent_root,
3616                                          &fs_info->qgroup_rescan_progress,
3617                                          path, 1, 0);
3618
3619         btrfs_debug(fs_info,
3620                 "current progress key (%llu %u %llu), search_slot ret %d",
3621                 fs_info->qgroup_rescan_progress.objectid,
3622                 fs_info->qgroup_rescan_progress.type,
3623                 fs_info->qgroup_rescan_progress.offset, ret);
3624
3625         if (ret) {
3626                 /*
3627                  * The rescan is about to end, we will not be scanning any
3628                  * further blocks. We cannot unset the RESCAN flag here, because
3629                  * we want to commit the transaction if everything went well.
3630                  * To make the live accounting work in this phase, we set our
3631                  * scan progress pointer such that every real extent objectid
3632                  * will be smaller.
3633                  */
3634                 fs_info->qgroup_rescan_progress.objectid = (u64)-1;
3635                 btrfs_release_path(path);
3636                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3637                 return ret;
3638         }
3639         done = is_last_leaf(path);
3640
3641         btrfs_item_key_to_cpu(path->nodes[0], &found,
3642                               btrfs_header_nritems(path->nodes[0]) - 1);
3643         fs_info->qgroup_rescan_progress.objectid = found.objectid + 1;
3644
3645         scratch_leaf = btrfs_clone_extent_buffer(path->nodes[0]);
3646         if (!scratch_leaf) {
3647                 ret = -ENOMEM;
3648                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3649                 goto out;
3650         }
3651         slot = path->slots[0];
3652         btrfs_release_path(path);
3653         mutex_unlock(&fs_info->qgroup_rescan_lock);
3654
3655         for (; slot < btrfs_header_nritems(scratch_leaf); ++slot) {
3656                 struct btrfs_backref_walk_ctx ctx = { 0 };
3657
3658                 btrfs_item_key_to_cpu(scratch_leaf, &found, slot);
3659                 if (found.type != BTRFS_EXTENT_ITEM_KEY &&
3660                     found.type != BTRFS_METADATA_ITEM_KEY)
3661                         continue;
3662                 if (found.type == BTRFS_METADATA_ITEM_KEY)
3663                         num_bytes = fs_info->nodesize;
3664                 else
3665                         num_bytes = found.offset;
3666
3667                 ctx.bytenr = found.objectid;
3668                 ctx.fs_info = fs_info;
3669
3670                 ret = btrfs_find_all_roots(&ctx, false);
3671                 if (ret < 0)
3672                         goto out;
3673                 /* For rescan, just pass old_roots as NULL */
3674                 ret = btrfs_qgroup_account_extent(trans, found.objectid,
3675                                                   num_bytes, NULL, ctx.roots);
3676                 if (ret < 0)
3677                         goto out;
3678         }
3679 out:
3680         if (scratch_leaf)
3681                 free_extent_buffer(scratch_leaf);
3682
3683         if (done && !ret) {
3684                 ret = 1;
3685                 fs_info->qgroup_rescan_progress.objectid = (u64)-1;
3686         }
3687         return ret;
3688 }
3689
3690 static bool rescan_should_stop(struct btrfs_fs_info *fs_info)
3691 {
3692         if (btrfs_fs_closing(fs_info))
3693                 return true;
3694         if (test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state))
3695                 return true;
3696         if (!btrfs_qgroup_enabled(fs_info))
3697                 return true;
3698         if (fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN)
3699                 return true;
3700         return false;
3701 }
3702
3703 static void btrfs_qgroup_rescan_worker(struct btrfs_work *work)
3704 {
3705         struct btrfs_fs_info *fs_info = container_of(work, struct btrfs_fs_info,
3706                                                      qgroup_rescan_work);
3707         struct btrfs_path *path;
3708         struct btrfs_trans_handle *trans = NULL;
3709         int err = -ENOMEM;
3710         int ret = 0;
3711         bool stopped = false;
3712         bool did_leaf_rescans = false;
3713
3714         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE)
3715                 return;
3716
3717         path = btrfs_alloc_path();
3718         if (!path)
3719                 goto out;
3720         /*
3721          * Rescan should only search for commit root, and any later difference
3722          * should be recorded by qgroup
3723          */
3724         path->search_commit_root = 1;
3725         path->skip_locking = 1;
3726
3727         err = 0;
3728         while (!err && !(stopped = rescan_should_stop(fs_info))) {
3729                 trans = btrfs_start_transaction(fs_info->fs_root, 0);
3730                 if (IS_ERR(trans)) {
3731                         err = PTR_ERR(trans);
3732                         break;
3733                 }
3734
3735                 err = qgroup_rescan_leaf(trans, path);
3736                 did_leaf_rescans = true;
3737
3738                 if (err > 0)
3739                         btrfs_commit_transaction(trans);
3740                 else
3741                         btrfs_end_transaction(trans);
3742         }
3743
3744 out:
3745         btrfs_free_path(path);
3746
3747         mutex_lock(&fs_info->qgroup_rescan_lock);
3748         if (err > 0 &&
3749             fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT) {
3750                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
3751         } else if (err < 0 || stopped) {
3752                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
3753         }
3754         mutex_unlock(&fs_info->qgroup_rescan_lock);
3755
3756         /*
3757          * Only update status, since the previous part has already updated the
3758          * qgroup info, and only if we did any actual work. This also prevents
3759          * race with a concurrent quota disable, which has already set
3760          * fs_info->quota_root to NULL and cleared BTRFS_FS_QUOTA_ENABLED at
3761          * btrfs_quota_disable().
3762          */
3763         if (did_leaf_rescans) {
3764                 trans = btrfs_start_transaction(fs_info->quota_root, 1);
3765                 if (IS_ERR(trans)) {
3766                         err = PTR_ERR(trans);
3767                         trans = NULL;
3768                         btrfs_err(fs_info,
3769                                   "fail to start transaction for status update: %d",
3770                                   err);
3771                 }
3772         } else {
3773                 trans = NULL;
3774         }
3775
3776         mutex_lock(&fs_info->qgroup_rescan_lock);
3777         if (!stopped ||
3778             fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN)
3779                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3780         if (trans) {
3781                 ret = update_qgroup_status_item(trans);
3782                 if (ret < 0) {
3783                         err = ret;
3784                         btrfs_err(fs_info, "fail to update qgroup status: %d",
3785                                   err);
3786                 }
3787         }
3788         fs_info->qgroup_rescan_running = false;
3789         fs_info->qgroup_flags &= ~BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN;
3790         complete_all(&fs_info->qgroup_rescan_completion);
3791         mutex_unlock(&fs_info->qgroup_rescan_lock);
3792
3793         if (!trans)
3794                 return;
3795
3796         btrfs_end_transaction(trans);
3797
3798         if (stopped) {
3799                 btrfs_info(fs_info, "qgroup scan paused");
3800         } else if (fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN) {
3801                 btrfs_info(fs_info, "qgroup scan cancelled");
3802         } else if (err >= 0) {
3803                 btrfs_info(fs_info, "qgroup scan completed%s",
3804                         err > 0 ? " (inconsistency flag cleared)" : "");
3805         } else {
3806                 btrfs_err(fs_info, "qgroup scan failed with %d", err);
3807         }
3808 }
3809
3810 /*
3811  * Checks that (a) no rescan is running and (b) quota is enabled. Allocates all
3812  * memory required for the rescan context.
3813  */
3814 static int
3815 qgroup_rescan_init(struct btrfs_fs_info *fs_info, u64 progress_objectid,
3816                    int init_flags)
3817 {
3818         int ret = 0;
3819
3820         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE) {
3821                 btrfs_warn(fs_info, "qgroup rescan init failed, running in simple mode");
3822                 return -EINVAL;
3823         }
3824
3825         if (!init_flags) {
3826                 /* we're resuming qgroup rescan at mount time */
3827                 if (!(fs_info->qgroup_flags &
3828                       BTRFS_QGROUP_STATUS_FLAG_RESCAN)) {
3829                         btrfs_warn(fs_info,
3830                         "qgroup rescan init failed, qgroup rescan is not queued");
3831                         ret = -EINVAL;
3832                 } else if (!(fs_info->qgroup_flags &
3833                              BTRFS_QGROUP_STATUS_FLAG_ON)) {
3834                         btrfs_warn(fs_info,
3835                         "qgroup rescan init failed, qgroup is not enabled");
3836                         ret = -EINVAL;
3837                 }
3838
3839                 if (ret)
3840                         return ret;
3841         }
3842
3843         mutex_lock(&fs_info->qgroup_rescan_lock);
3844
3845         if (init_flags) {
3846                 if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
3847                         btrfs_warn(fs_info,
3848                                    "qgroup rescan is already in progress");
3849                         ret = -EINPROGRESS;
3850                 } else if (!(fs_info->qgroup_flags &
3851                              BTRFS_QGROUP_STATUS_FLAG_ON)) {
3852                         btrfs_warn(fs_info,
3853                         "qgroup rescan init failed, qgroup is not enabled");
3854                         ret = -EINVAL;
3855                 } else if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_DISABLED) {
3856                         /* Quota disable is in progress */
3857                         ret = -EBUSY;
3858                 }
3859
3860                 if (ret) {
3861                         mutex_unlock(&fs_info->qgroup_rescan_lock);
3862                         return ret;
3863                 }
3864                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3865         }
3866
3867         memset(&fs_info->qgroup_rescan_progress, 0,
3868                 sizeof(fs_info->qgroup_rescan_progress));
3869         fs_info->qgroup_flags &= ~(BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN |
3870                                    BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING);
3871         fs_info->qgroup_rescan_progress.objectid = progress_objectid;
3872         init_completion(&fs_info->qgroup_rescan_completion);
3873         mutex_unlock(&fs_info->qgroup_rescan_lock);
3874
3875         btrfs_init_work(&fs_info->qgroup_rescan_work,
3876                         btrfs_qgroup_rescan_worker, NULL);
3877         return 0;
3878 }
3879
3880 static void
3881 qgroup_rescan_zero_tracking(struct btrfs_fs_info *fs_info)
3882 {
3883         struct rb_node *n;
3884         struct btrfs_qgroup *qgroup;
3885
3886         spin_lock(&fs_info->qgroup_lock);
3887         /* clear all current qgroup tracking information */
3888         for (n = rb_first(&fs_info->qgroup_tree); n; n = rb_next(n)) {
3889                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
3890                 qgroup->rfer = 0;
3891                 qgroup->rfer_cmpr = 0;
3892                 qgroup->excl = 0;
3893                 qgroup->excl_cmpr = 0;
3894                 qgroup_dirty(fs_info, qgroup);
3895         }
3896         spin_unlock(&fs_info->qgroup_lock);
3897 }
3898
3899 int
3900 btrfs_qgroup_rescan(struct btrfs_fs_info *fs_info)
3901 {
3902         int ret = 0;
3903         struct btrfs_trans_handle *trans;
3904
3905         ret = qgroup_rescan_init(fs_info, 0, 1);
3906         if (ret)
3907                 return ret;
3908
3909         /*
3910          * We have set the rescan_progress to 0, which means no more
3911          * delayed refs will be accounted by btrfs_qgroup_account_ref.
3912          * However, btrfs_qgroup_account_ref may be right after its call
3913          * to btrfs_find_all_roots, in which case it would still do the
3914          * accounting.
3915          * To solve this, we're committing the transaction, which will
3916          * ensure we run all delayed refs and only after that, we are
3917          * going to clear all tracking information for a clean start.
3918          */
3919
3920         trans = btrfs_attach_transaction_barrier(fs_info->fs_root);
3921         if (IS_ERR(trans) && trans != ERR_PTR(-ENOENT)) {
3922                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3923                 return PTR_ERR(trans);
3924         } else if (trans != ERR_PTR(-ENOENT)) {
3925                 ret = btrfs_commit_transaction(trans);
3926                 if (ret) {
3927                         fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3928                         return ret;
3929                 }
3930         }
3931
3932         qgroup_rescan_zero_tracking(fs_info);
3933
3934         mutex_lock(&fs_info->qgroup_rescan_lock);
3935         fs_info->qgroup_rescan_running = true;
3936         btrfs_queue_work(fs_info->qgroup_rescan_workers,
3937                          &fs_info->qgroup_rescan_work);
3938         mutex_unlock(&fs_info->qgroup_rescan_lock);
3939
3940         return 0;
3941 }
3942
3943 int btrfs_qgroup_wait_for_completion(struct btrfs_fs_info *fs_info,
3944                                      bool interruptible)
3945 {
3946         int running;
3947         int ret = 0;
3948
3949         mutex_lock(&fs_info->qgroup_rescan_lock);
3950         running = fs_info->qgroup_rescan_running;
3951         mutex_unlock(&fs_info->qgroup_rescan_lock);
3952
3953         if (!running)
3954                 return 0;
3955
3956         if (interruptible)
3957                 ret = wait_for_completion_interruptible(
3958                                         &fs_info->qgroup_rescan_completion);
3959         else
3960                 wait_for_completion(&fs_info->qgroup_rescan_completion);
3961
3962         return ret;
3963 }
3964
3965 /*
3966  * this is only called from open_ctree where we're still single threaded, thus
3967  * locking is omitted here.
3968  */
3969 void
3970 btrfs_qgroup_rescan_resume(struct btrfs_fs_info *fs_info)
3971 {
3972         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
3973                 mutex_lock(&fs_info->qgroup_rescan_lock);
3974                 fs_info->qgroup_rescan_running = true;
3975                 btrfs_queue_work(fs_info->qgroup_rescan_workers,
3976                                  &fs_info->qgroup_rescan_work);
3977                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3978         }
3979 }
3980
3981 #define rbtree_iterate_from_safe(node, next, start)                             \
3982        for (node = start; node && ({ next = rb_next(node); 1;}); node = next)
3983
3984 static int qgroup_unreserve_range(struct btrfs_inode *inode,
3985                                   struct extent_changeset *reserved, u64 start,
3986                                   u64 len)
3987 {
3988         struct rb_node *node;
3989         struct rb_node *next;
3990         struct ulist_node *entry;
3991         int ret = 0;
3992
3993         node = reserved->range_changed.root.rb_node;
3994         if (!node)
3995                 return 0;
3996         while (node) {
3997                 entry = rb_entry(node, struct ulist_node, rb_node);
3998                 if (entry->val < start)
3999                         node = node->rb_right;
4000                 else
4001                         node = node->rb_left;
4002         }
4003
4004         if (entry->val > start && rb_prev(&entry->rb_node))
4005                 entry = rb_entry(rb_prev(&entry->rb_node), struct ulist_node,
4006                                  rb_node);
4007
4008         rbtree_iterate_from_safe(node, next, &entry->rb_node) {
4009                 u64 entry_start;
4010                 u64 entry_end;
4011                 u64 entry_len;
4012                 int clear_ret;
4013
4014                 entry = rb_entry(node, struct ulist_node, rb_node);
4015                 entry_start = entry->val;
4016                 entry_end = entry->aux;
4017                 entry_len = entry_end - entry_start + 1;
4018
4019                 if (entry_start >= start + len)
4020                         break;
4021                 if (entry_start + entry_len <= start)
4022                         continue;
4023                 /*
4024                  * Now the entry is in [start, start + len), revert the
4025                  * EXTENT_QGROUP_RESERVED bit.
4026                  */
4027                 clear_ret = clear_extent_bits(&inode->io_tree, entry_start,
4028                                               entry_end, EXTENT_QGROUP_RESERVED);
4029                 if (!ret && clear_ret < 0)
4030                         ret = clear_ret;
4031
4032                 ulist_del(&reserved->range_changed, entry->val, entry->aux);
4033                 if (likely(reserved->bytes_changed >= entry_len)) {
4034                         reserved->bytes_changed -= entry_len;
4035                 } else {
4036                         WARN_ON(1);
4037                         reserved->bytes_changed = 0;
4038                 }
4039         }
4040
4041         return ret;
4042 }
4043
4044 /*
4045  * Try to free some space for qgroup.
4046  *
4047  * For qgroup, there are only 3 ways to free qgroup space:
4048  * - Flush nodatacow write
4049  *   Any nodatacow write will free its reserved data space at run_delalloc_range().
4050  *   In theory, we should only flush nodatacow inodes, but it's not yet
4051  *   possible, so we need to flush the whole root.
4052  *
4053  * - Wait for ordered extents
4054  *   When ordered extents are finished, their reserved metadata is finally
4055  *   converted to per_trans status, which can be freed by later commit
4056  *   transaction.
4057  *
4058  * - Commit transaction
4059  *   This would free the meta_per_trans space.
4060  *   In theory this shouldn't provide much space, but any more qgroup space
4061  *   is needed.
4062  */
4063 static int try_flush_qgroup(struct btrfs_root *root)
4064 {
4065         struct btrfs_trans_handle *trans;
4066         int ret;
4067
4068         /* Can't hold an open transaction or we run the risk of deadlocking. */
4069         ASSERT(current->journal_info == NULL);
4070         if (WARN_ON(current->journal_info))
4071                 return 0;
4072
4073         /*
4074          * We don't want to run flush again and again, so if there is a running
4075          * one, we won't try to start a new flush, but exit directly.
4076          */
4077         if (test_and_set_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state)) {
4078                 wait_event(root->qgroup_flush_wait,
4079                         !test_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state));
4080                 return 0;
4081         }
4082
4083         ret = btrfs_start_delalloc_snapshot(root, true);
4084         if (ret < 0)
4085                 goto out;
4086         btrfs_wait_ordered_extents(root, U64_MAX, 0, (u64)-1);
4087
4088         trans = btrfs_attach_transaction_barrier(root);
4089         if (IS_ERR(trans)) {
4090                 ret = PTR_ERR(trans);
4091                 if (ret == -ENOENT)
4092                         ret = 0;
4093                 goto out;
4094         }
4095
4096         ret = btrfs_commit_transaction(trans);
4097 out:
4098         clear_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state);
4099         wake_up(&root->qgroup_flush_wait);
4100         return ret;
4101 }
4102
4103 static int qgroup_reserve_data(struct btrfs_inode *inode,
4104                         struct extent_changeset **reserved_ret, u64 start,
4105                         u64 len)
4106 {
4107         struct btrfs_root *root = inode->root;
4108         struct extent_changeset *reserved;
4109         bool new_reserved = false;
4110         u64 orig_reserved;
4111         u64 to_reserve;
4112         int ret;
4113
4114         if (btrfs_qgroup_mode(root->fs_info) == BTRFS_QGROUP_MODE_DISABLED ||
4115             !is_fstree(root->root_key.objectid) || len == 0)
4116                 return 0;
4117
4118         /* @reserved parameter is mandatory for qgroup */
4119         if (WARN_ON(!reserved_ret))
4120                 return -EINVAL;
4121         if (!*reserved_ret) {
4122                 new_reserved = true;
4123                 *reserved_ret = extent_changeset_alloc();
4124                 if (!*reserved_ret)
4125                         return -ENOMEM;
4126         }
4127         reserved = *reserved_ret;
4128         /* Record already reserved space */
4129         orig_reserved = reserved->bytes_changed;
4130         ret = set_record_extent_bits(&inode->io_tree, start,
4131                         start + len -1, EXTENT_QGROUP_RESERVED, reserved);
4132
4133         /* Newly reserved space */
4134         to_reserve = reserved->bytes_changed - orig_reserved;
4135         trace_btrfs_qgroup_reserve_data(&inode->vfs_inode, start, len,
4136                                         to_reserve, QGROUP_RESERVE);
4137         if (ret < 0)
4138                 goto out;
4139         ret = qgroup_reserve(root, to_reserve, true, BTRFS_QGROUP_RSV_DATA);
4140         if (ret < 0)
4141                 goto cleanup;
4142
4143         return ret;
4144
4145 cleanup:
4146         qgroup_unreserve_range(inode, reserved, start, len);
4147 out:
4148         if (new_reserved) {
4149                 extent_changeset_free(reserved);
4150                 *reserved_ret = NULL;
4151         }
4152         return ret;
4153 }
4154
4155 /*
4156  * Reserve qgroup space for range [start, start + len).
4157  *
4158  * This function will either reserve space from related qgroups or do nothing
4159  * if the range is already reserved.
4160  *
4161  * Return 0 for successful reservation
4162  * Return <0 for error (including -EQUOT)
4163  *
4164  * NOTE: This function may sleep for memory allocation, dirty page flushing and
4165  *       commit transaction. So caller should not hold any dirty page locked.
4166  */
4167 int btrfs_qgroup_reserve_data(struct btrfs_inode *inode,
4168                         struct extent_changeset **reserved_ret, u64 start,
4169                         u64 len)
4170 {
4171         int ret;
4172
4173         ret = qgroup_reserve_data(inode, reserved_ret, start, len);
4174         if (ret <= 0 && ret != -EDQUOT)
4175                 return ret;
4176
4177         ret = try_flush_qgroup(inode->root);
4178         if (ret < 0)
4179                 return ret;
4180         return qgroup_reserve_data(inode, reserved_ret, start, len);
4181 }
4182
4183 /* Free ranges specified by @reserved, normally in error path */
4184 static int qgroup_free_reserved_data(struct btrfs_inode *inode,
4185                                      struct extent_changeset *reserved,
4186                                      u64 start, u64 len, u64 *freed_ret)
4187 {
4188         struct btrfs_root *root = inode->root;
4189         struct ulist_node *unode;
4190         struct ulist_iterator uiter;
4191         struct extent_changeset changeset;
4192         u64 freed = 0;
4193         int ret;
4194
4195         extent_changeset_init(&changeset);
4196         len = round_up(start + len, root->fs_info->sectorsize);
4197         start = round_down(start, root->fs_info->sectorsize);
4198
4199         ULIST_ITER_INIT(&uiter);
4200         while ((unode = ulist_next(&reserved->range_changed, &uiter))) {
4201                 u64 range_start = unode->val;
4202                 /* unode->aux is the inclusive end */
4203                 u64 range_len = unode->aux - range_start + 1;
4204                 u64 free_start;
4205                 u64 free_len;
4206
4207                 extent_changeset_release(&changeset);
4208
4209                 /* Only free range in range [start, start + len) */
4210                 if (range_start >= start + len ||
4211                     range_start + range_len <= start)
4212                         continue;
4213                 free_start = max(range_start, start);
4214                 free_len = min(start + len, range_start + range_len) -
4215                            free_start;
4216                 /*
4217                  * TODO: To also modify reserved->ranges_reserved to reflect
4218                  * the modification.
4219                  *
4220                  * However as long as we free qgroup reserved according to
4221                  * EXTENT_QGROUP_RESERVED, we won't double free.
4222                  * So not need to rush.
4223                  */
4224                 ret = clear_record_extent_bits(&inode->io_tree, free_start,
4225                                 free_start + free_len - 1,
4226                                 EXTENT_QGROUP_RESERVED, &changeset);
4227                 if (ret < 0)
4228                         goto out;
4229                 freed += changeset.bytes_changed;
4230         }
4231         btrfs_qgroup_free_refroot(root->fs_info, root->root_key.objectid, freed,
4232                                   BTRFS_QGROUP_RSV_DATA);
4233         if (freed_ret)
4234                 *freed_ret = freed;
4235         ret = 0;
4236 out:
4237         extent_changeset_release(&changeset);
4238         return ret;
4239 }
4240
4241 static int __btrfs_qgroup_release_data(struct btrfs_inode *inode,
4242                         struct extent_changeset *reserved, u64 start, u64 len,
4243                         u64 *released, int free)
4244 {
4245         struct extent_changeset changeset;
4246         int trace_op = QGROUP_RELEASE;
4247         int ret;
4248
4249         if (btrfs_qgroup_mode(inode->root->fs_info) == BTRFS_QGROUP_MODE_DISABLED) {
4250                 extent_changeset_init(&changeset);
4251                 return clear_record_extent_bits(&inode->io_tree, start,
4252                                                 start + len - 1,
4253                                                 EXTENT_QGROUP_RESERVED, &changeset);
4254         }
4255
4256         /* In release case, we shouldn't have @reserved */
4257         WARN_ON(!free && reserved);
4258         if (free && reserved)
4259                 return qgroup_free_reserved_data(inode, reserved, start, len, released);
4260         extent_changeset_init(&changeset);
4261         ret = clear_record_extent_bits(&inode->io_tree, start, start + len -1,
4262                                        EXTENT_QGROUP_RESERVED, &changeset);
4263         if (ret < 0)
4264                 goto out;
4265
4266         if (free)
4267                 trace_op = QGROUP_FREE;
4268         trace_btrfs_qgroup_release_data(&inode->vfs_inode, start, len,
4269                                         changeset.bytes_changed, trace_op);
4270         if (free)
4271                 btrfs_qgroup_free_refroot(inode->root->fs_info,
4272                                 inode->root->root_key.objectid,
4273                                 changeset.bytes_changed, BTRFS_QGROUP_RSV_DATA);
4274         if (released)
4275                 *released = changeset.bytes_changed;
4276 out:
4277         extent_changeset_release(&changeset);
4278         return ret;
4279 }
4280
4281 /*
4282  * Free a reserved space range from io_tree and related qgroups
4283  *
4284  * Should be called when a range of pages get invalidated before reaching disk.
4285  * Or for error cleanup case.
4286  * if @reserved is given, only reserved range in [@start, @start + @len) will
4287  * be freed.
4288  *
4289  * For data written to disk, use btrfs_qgroup_release_data().
4290  *
4291  * NOTE: This function may sleep for memory allocation.
4292  */
4293 int btrfs_qgroup_free_data(struct btrfs_inode *inode,
4294                            struct extent_changeset *reserved,
4295                            u64 start, u64 len, u64 *freed)
4296 {
4297         return __btrfs_qgroup_release_data(inode, reserved, start, len, freed, 1);
4298 }
4299
4300 /*
4301  * Release a reserved space range from io_tree only.
4302  *
4303  * Should be called when a range of pages get written to disk and corresponding
4304  * FILE_EXTENT is inserted into corresponding root.
4305  *
4306  * Since new qgroup accounting framework will only update qgroup numbers at
4307  * commit_transaction() time, its reserved space shouldn't be freed from
4308  * related qgroups.
4309  *
4310  * But we should release the range from io_tree, to allow further write to be
4311  * COWed.
4312  *
4313  * NOTE: This function may sleep for memory allocation.
4314  */
4315 int btrfs_qgroup_release_data(struct btrfs_inode *inode, u64 start, u64 len, u64 *released)
4316 {
4317         return __btrfs_qgroup_release_data(inode, NULL, start, len, released, 0);
4318 }
4319
4320 static void add_root_meta_rsv(struct btrfs_root *root, int num_bytes,
4321                               enum btrfs_qgroup_rsv_type type)
4322 {
4323         if (type != BTRFS_QGROUP_RSV_META_PREALLOC &&
4324             type != BTRFS_QGROUP_RSV_META_PERTRANS)
4325                 return;
4326         if (num_bytes == 0)
4327                 return;
4328
4329         spin_lock(&root->qgroup_meta_rsv_lock);
4330         if (type == BTRFS_QGROUP_RSV_META_PREALLOC)
4331                 root->qgroup_meta_rsv_prealloc += num_bytes;
4332         else
4333                 root->qgroup_meta_rsv_pertrans += num_bytes;
4334         spin_unlock(&root->qgroup_meta_rsv_lock);
4335 }
4336
4337 static int sub_root_meta_rsv(struct btrfs_root *root, int num_bytes,
4338                              enum btrfs_qgroup_rsv_type type)
4339 {
4340         if (type != BTRFS_QGROUP_RSV_META_PREALLOC &&
4341             type != BTRFS_QGROUP_RSV_META_PERTRANS)
4342                 return 0;
4343         if (num_bytes == 0)
4344                 return 0;
4345
4346         spin_lock(&root->qgroup_meta_rsv_lock);
4347         if (type == BTRFS_QGROUP_RSV_META_PREALLOC) {
4348                 num_bytes = min_t(u64, root->qgroup_meta_rsv_prealloc,
4349                                   num_bytes);
4350                 root->qgroup_meta_rsv_prealloc -= num_bytes;
4351         } else {
4352                 num_bytes = min_t(u64, root->qgroup_meta_rsv_pertrans,
4353                                   num_bytes);
4354                 root->qgroup_meta_rsv_pertrans -= num_bytes;
4355         }
4356         spin_unlock(&root->qgroup_meta_rsv_lock);
4357         return num_bytes;
4358 }
4359
4360 int btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
4361                               enum btrfs_qgroup_rsv_type type, bool enforce)
4362 {
4363         struct btrfs_fs_info *fs_info = root->fs_info;
4364         int ret;
4365
4366         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_DISABLED ||
4367             !is_fstree(root->root_key.objectid) || num_bytes == 0)
4368                 return 0;
4369
4370         BUG_ON(num_bytes != round_down(num_bytes, fs_info->nodesize));
4371         trace_qgroup_meta_reserve(root, (s64)num_bytes, type);
4372         ret = qgroup_reserve(root, num_bytes, enforce, type);
4373         if (ret < 0)
4374                 return ret;
4375         /*
4376          * Record what we have reserved into root.
4377          *
4378          * To avoid quota disabled->enabled underflow.
4379          * In that case, we may try to free space we haven't reserved
4380          * (since quota was disabled), so record what we reserved into root.
4381          * And ensure later release won't underflow this number.
4382          */
4383         add_root_meta_rsv(root, num_bytes, type);
4384         return ret;
4385 }
4386
4387 int __btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
4388                                 enum btrfs_qgroup_rsv_type type, bool enforce,
4389                                 bool noflush)
4390 {
4391         int ret;
4392
4393         ret = btrfs_qgroup_reserve_meta(root, num_bytes, type, enforce);
4394         if ((ret <= 0 && ret != -EDQUOT) || noflush)
4395                 return ret;
4396
4397         ret = try_flush_qgroup(root);
4398         if (ret < 0)
4399                 return ret;
4400         return btrfs_qgroup_reserve_meta(root, num_bytes, type, enforce);
4401 }
4402
4403 /*
4404  * Per-transaction meta reservation should be all freed at transaction commit
4405  * time
4406  */
4407 void btrfs_qgroup_free_meta_all_pertrans(struct btrfs_root *root)
4408 {
4409         struct btrfs_fs_info *fs_info = root->fs_info;
4410
4411         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_DISABLED ||
4412             !is_fstree(root->root_key.objectid))
4413                 return;
4414
4415         /* TODO: Update trace point to handle such free */
4416         trace_qgroup_meta_free_all_pertrans(root);
4417         /* Special value -1 means to free all reserved space */
4418         btrfs_qgroup_free_refroot(fs_info, root->root_key.objectid, (u64)-1,
4419                                   BTRFS_QGROUP_RSV_META_PERTRANS);
4420 }
4421
4422 void __btrfs_qgroup_free_meta(struct btrfs_root *root, int num_bytes,
4423                               enum btrfs_qgroup_rsv_type type)
4424 {
4425         struct btrfs_fs_info *fs_info = root->fs_info;
4426
4427         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_DISABLED ||
4428             !is_fstree(root->root_key.objectid))
4429                 return;
4430
4431         /*
4432          * reservation for META_PREALLOC can happen before quota is enabled,
4433          * which can lead to underflow.
4434          * Here ensure we will only free what we really have reserved.
4435          */
4436         num_bytes = sub_root_meta_rsv(root, num_bytes, type);
4437         BUG_ON(num_bytes != round_down(num_bytes, fs_info->nodesize));
4438         trace_qgroup_meta_reserve(root, -(s64)num_bytes, type);
4439         btrfs_qgroup_free_refroot(fs_info, root->root_key.objectid,
4440                                   num_bytes, type);
4441 }
4442
4443 static void qgroup_convert_meta(struct btrfs_fs_info *fs_info, u64 ref_root,
4444                                 int num_bytes)
4445 {
4446         struct btrfs_qgroup *qgroup;
4447         LIST_HEAD(qgroup_list);
4448
4449         if (num_bytes == 0)
4450                 return;
4451         if (!fs_info->quota_root)
4452                 return;
4453
4454         spin_lock(&fs_info->qgroup_lock);
4455         qgroup = find_qgroup_rb(fs_info, ref_root);
4456         if (!qgroup)
4457                 goto out;
4458
4459         qgroup_iterator_add(&qgroup_list, qgroup);
4460         list_for_each_entry(qgroup, &qgroup_list, iterator) {
4461                 struct btrfs_qgroup_list *glist;
4462
4463                 qgroup_rsv_release(fs_info, qgroup, num_bytes,
4464                                 BTRFS_QGROUP_RSV_META_PREALLOC);
4465                 if (!sb_rdonly(fs_info->sb))
4466                         qgroup_rsv_add(fs_info, qgroup, num_bytes,
4467                                        BTRFS_QGROUP_RSV_META_PERTRANS);
4468
4469                 list_for_each_entry(glist, &qgroup->groups, next_group)
4470                         qgroup_iterator_add(&qgroup_list, glist->group);
4471         }
4472 out:
4473         qgroup_iterator_clean(&qgroup_list);
4474         spin_unlock(&fs_info->qgroup_lock);
4475 }
4476
4477 /*
4478  * Convert @num_bytes of META_PREALLOCATED reservation to META_PERTRANS.
4479  *
4480  * This is called when preallocated meta reservation needs to be used.
4481  * Normally after btrfs_join_transaction() call.
4482  */
4483 void btrfs_qgroup_convert_reserved_meta(struct btrfs_root *root, int num_bytes)
4484 {
4485         struct btrfs_fs_info *fs_info = root->fs_info;
4486
4487         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_DISABLED ||
4488             !is_fstree(root->root_key.objectid))
4489                 return;
4490         /* Same as btrfs_qgroup_free_meta_prealloc() */
4491         num_bytes = sub_root_meta_rsv(root, num_bytes,
4492                                       BTRFS_QGROUP_RSV_META_PREALLOC);
4493         trace_qgroup_meta_convert(root, num_bytes);
4494         qgroup_convert_meta(fs_info, root->root_key.objectid, num_bytes);
4495         if (!sb_rdonly(fs_info->sb))
4496                 add_root_meta_rsv(root, num_bytes, BTRFS_QGROUP_RSV_META_PERTRANS);
4497 }
4498
4499 /*
4500  * Check qgroup reserved space leaking, normally at destroy inode
4501  * time
4502  */
4503 void btrfs_qgroup_check_reserved_leak(struct btrfs_inode *inode)
4504 {
4505         struct extent_changeset changeset;
4506         struct ulist_node *unode;
4507         struct ulist_iterator iter;
4508         int ret;
4509
4510         extent_changeset_init(&changeset);
4511         ret = clear_record_extent_bits(&inode->io_tree, 0, (u64)-1,
4512                         EXTENT_QGROUP_RESERVED, &changeset);
4513
4514         WARN_ON(ret < 0);
4515         if (WARN_ON(changeset.bytes_changed)) {
4516                 ULIST_ITER_INIT(&iter);
4517                 while ((unode = ulist_next(&changeset.range_changed, &iter))) {
4518                         btrfs_warn(inode->root->fs_info,
4519                 "leaking qgroup reserved space, ino: %llu, start: %llu, end: %llu",
4520                                 btrfs_ino(inode), unode->val, unode->aux);
4521                 }
4522                 btrfs_qgroup_free_refroot(inode->root->fs_info,
4523                                 inode->root->root_key.objectid,
4524                                 changeset.bytes_changed, BTRFS_QGROUP_RSV_DATA);
4525
4526         }
4527         extent_changeset_release(&changeset);
4528 }
4529
4530 void btrfs_qgroup_init_swapped_blocks(
4531         struct btrfs_qgroup_swapped_blocks *swapped_blocks)
4532 {
4533         int i;
4534
4535         spin_lock_init(&swapped_blocks->lock);
4536         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
4537                 swapped_blocks->blocks[i] = RB_ROOT;
4538         swapped_blocks->swapped = false;
4539 }
4540
4541 /*
4542  * Delete all swapped blocks record of @root.
4543  * Every record here means we skipped a full subtree scan for qgroup.
4544  *
4545  * Gets called when committing one transaction.
4546  */
4547 void btrfs_qgroup_clean_swapped_blocks(struct btrfs_root *root)
4548 {
4549         struct btrfs_qgroup_swapped_blocks *swapped_blocks;
4550         int i;
4551
4552         swapped_blocks = &root->swapped_blocks;
4553
4554         spin_lock(&swapped_blocks->lock);
4555         if (!swapped_blocks->swapped)
4556                 goto out;
4557         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
4558                 struct rb_root *cur_root = &swapped_blocks->blocks[i];
4559                 struct btrfs_qgroup_swapped_block *entry;
4560                 struct btrfs_qgroup_swapped_block *next;
4561
4562                 rbtree_postorder_for_each_entry_safe(entry, next, cur_root,
4563                                                      node)
4564                         kfree(entry);
4565                 swapped_blocks->blocks[i] = RB_ROOT;
4566         }
4567         swapped_blocks->swapped = false;
4568 out:
4569         spin_unlock(&swapped_blocks->lock);
4570 }
4571
4572 /*
4573  * Add subtree roots record into @subvol_root.
4574  *
4575  * @subvol_root:        tree root of the subvolume tree get swapped
4576  * @bg:                 block group under balance
4577  * @subvol_parent/slot: pointer to the subtree root in subvolume tree
4578  * @reloc_parent/slot:  pointer to the subtree root in reloc tree
4579  *                      BOTH POINTERS ARE BEFORE TREE SWAP
4580  * @last_snapshot:      last snapshot generation of the subvolume tree
4581  */
4582 int btrfs_qgroup_add_swapped_blocks(struct btrfs_trans_handle *trans,
4583                 struct btrfs_root *subvol_root,
4584                 struct btrfs_block_group *bg,
4585                 struct extent_buffer *subvol_parent, int subvol_slot,
4586                 struct extent_buffer *reloc_parent, int reloc_slot,
4587                 u64 last_snapshot)
4588 {
4589         struct btrfs_fs_info *fs_info = subvol_root->fs_info;
4590         struct btrfs_qgroup_swapped_blocks *blocks = &subvol_root->swapped_blocks;
4591         struct btrfs_qgroup_swapped_block *block;
4592         struct rb_node **cur;
4593         struct rb_node *parent = NULL;
4594         int level = btrfs_header_level(subvol_parent) - 1;
4595         int ret = 0;
4596
4597         if (!btrfs_qgroup_full_accounting(fs_info))
4598                 return 0;
4599
4600         if (btrfs_node_ptr_generation(subvol_parent, subvol_slot) >
4601             btrfs_node_ptr_generation(reloc_parent, reloc_slot)) {
4602                 btrfs_err_rl(fs_info,
4603                 "%s: bad parameter order, subvol_gen=%llu reloc_gen=%llu",
4604                         __func__,
4605                         btrfs_node_ptr_generation(subvol_parent, subvol_slot),
4606                         btrfs_node_ptr_generation(reloc_parent, reloc_slot));
4607                 return -EUCLEAN;
4608         }
4609
4610         block = kmalloc(sizeof(*block), GFP_NOFS);
4611         if (!block) {
4612                 ret = -ENOMEM;
4613                 goto out;
4614         }
4615
4616         /*
4617          * @reloc_parent/slot is still before swap, while @block is going to
4618          * record the bytenr after swap, so we do the swap here.
4619          */
4620         block->subvol_bytenr = btrfs_node_blockptr(reloc_parent, reloc_slot);
4621         block->subvol_generation = btrfs_node_ptr_generation(reloc_parent,
4622                                                              reloc_slot);
4623         block->reloc_bytenr = btrfs_node_blockptr(subvol_parent, subvol_slot);
4624         block->reloc_generation = btrfs_node_ptr_generation(subvol_parent,
4625                                                             subvol_slot);
4626         block->last_snapshot = last_snapshot;
4627         block->level = level;
4628
4629         /*
4630          * If we have bg == NULL, we're called from btrfs_recover_relocation(),
4631          * no one else can modify tree blocks thus we qgroup will not change
4632          * no matter the value of trace_leaf.
4633          */
4634         if (bg && bg->flags & BTRFS_BLOCK_GROUP_DATA)
4635                 block->trace_leaf = true;
4636         else
4637                 block->trace_leaf = false;
4638         btrfs_node_key_to_cpu(reloc_parent, &block->first_key, reloc_slot);
4639
4640         /* Insert @block into @blocks */
4641         spin_lock(&blocks->lock);
4642         cur = &blocks->blocks[level].rb_node;
4643         while (*cur) {
4644                 struct btrfs_qgroup_swapped_block *entry;
4645
4646                 parent = *cur;
4647                 entry = rb_entry(parent, struct btrfs_qgroup_swapped_block,
4648                                  node);
4649
4650                 if (entry->subvol_bytenr < block->subvol_bytenr) {
4651                         cur = &(*cur)->rb_left;
4652                 } else if (entry->subvol_bytenr > block->subvol_bytenr) {
4653                         cur = &(*cur)->rb_right;
4654                 } else {
4655                         if (entry->subvol_generation !=
4656                                         block->subvol_generation ||
4657                             entry->reloc_bytenr != block->reloc_bytenr ||
4658                             entry->reloc_generation !=
4659                                         block->reloc_generation) {
4660                                 /*
4661                                  * Duplicated but mismatch entry found.
4662                                  * Shouldn't happen.
4663                                  *
4664                                  * Marking qgroup inconsistent should be enough
4665                                  * for end users.
4666                                  */
4667                                 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
4668                                 ret = -EEXIST;
4669                         }
4670                         kfree(block);
4671                         goto out_unlock;
4672                 }
4673         }
4674         rb_link_node(&block->node, parent, cur);
4675         rb_insert_color(&block->node, &blocks->blocks[level]);
4676         blocks->swapped = true;
4677 out_unlock:
4678         spin_unlock(&blocks->lock);
4679 out:
4680         if (ret < 0)
4681                 qgroup_mark_inconsistent(fs_info);
4682         return ret;
4683 }
4684
4685 /*
4686  * Check if the tree block is a subtree root, and if so do the needed
4687  * delayed subtree trace for qgroup.
4688  *
4689  * This is called during btrfs_cow_block().
4690  */
4691 int btrfs_qgroup_trace_subtree_after_cow(struct btrfs_trans_handle *trans,
4692                                          struct btrfs_root *root,
4693                                          struct extent_buffer *subvol_eb)
4694 {
4695         struct btrfs_fs_info *fs_info = root->fs_info;
4696         struct btrfs_tree_parent_check check = { 0 };
4697         struct btrfs_qgroup_swapped_blocks *blocks = &root->swapped_blocks;
4698         struct btrfs_qgroup_swapped_block *block;
4699         struct extent_buffer *reloc_eb = NULL;
4700         struct rb_node *node;
4701         bool found = false;
4702         bool swapped = false;
4703         int level = btrfs_header_level(subvol_eb);
4704         int ret = 0;
4705         int i;
4706
4707         if (!btrfs_qgroup_full_accounting(fs_info))
4708                 return 0;
4709         if (!is_fstree(root->root_key.objectid) || !root->reloc_root)
4710                 return 0;
4711
4712         spin_lock(&blocks->lock);
4713         if (!blocks->swapped) {
4714                 spin_unlock(&blocks->lock);
4715                 return 0;
4716         }
4717         node = blocks->blocks[level].rb_node;
4718
4719         while (node) {
4720                 block = rb_entry(node, struct btrfs_qgroup_swapped_block, node);
4721                 if (block->subvol_bytenr < subvol_eb->start) {
4722                         node = node->rb_left;
4723                 } else if (block->subvol_bytenr > subvol_eb->start) {
4724                         node = node->rb_right;
4725                 } else {
4726                         found = true;
4727                         break;
4728                 }
4729         }
4730         if (!found) {
4731                 spin_unlock(&blocks->lock);
4732                 goto out;
4733         }
4734         /* Found one, remove it from @blocks first and update blocks->swapped */
4735         rb_erase(&block->node, &blocks->blocks[level]);
4736         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
4737                 if (RB_EMPTY_ROOT(&blocks->blocks[i])) {
4738                         swapped = true;
4739                         break;
4740                 }
4741         }
4742         blocks->swapped = swapped;
4743         spin_unlock(&blocks->lock);
4744
4745         check.level = block->level;
4746         check.transid = block->reloc_generation;
4747         check.has_first_key = true;
4748         memcpy(&check.first_key, &block->first_key, sizeof(check.first_key));
4749
4750         /* Read out reloc subtree root */
4751         reloc_eb = read_tree_block(fs_info, block->reloc_bytenr, &check);
4752         if (IS_ERR(reloc_eb)) {
4753                 ret = PTR_ERR(reloc_eb);
4754                 reloc_eb = NULL;
4755                 goto free_out;
4756         }
4757         if (!extent_buffer_uptodate(reloc_eb)) {
4758                 ret = -EIO;
4759                 goto free_out;
4760         }
4761
4762         ret = qgroup_trace_subtree_swap(trans, reloc_eb, subvol_eb,
4763                         block->last_snapshot, block->trace_leaf);
4764 free_out:
4765         kfree(block);
4766         free_extent_buffer(reloc_eb);
4767 out:
4768         if (ret < 0) {
4769                 btrfs_err_rl(fs_info,
4770                              "failed to account subtree at bytenr %llu: %d",
4771                              subvol_eb->start, ret);
4772                 qgroup_mark_inconsistent(fs_info);
4773         }
4774         return ret;
4775 }
4776
4777 void btrfs_qgroup_destroy_extent_records(struct btrfs_transaction *trans)
4778 {
4779         struct btrfs_qgroup_extent_record *entry;
4780         struct btrfs_qgroup_extent_record *next;
4781         struct rb_root *root;
4782
4783         root = &trans->delayed_refs.dirty_extent_root;
4784         rbtree_postorder_for_each_entry_safe(entry, next, root, node) {
4785                 ulist_free(entry->old_roots);
4786                 kfree(entry);
4787         }
4788         *root = RB_ROOT;
4789 }
4790
4791 void btrfs_free_squota_rsv(struct btrfs_fs_info *fs_info, u64 root, u64 rsv_bytes)
4792 {
4793         if (btrfs_qgroup_mode(fs_info) != BTRFS_QGROUP_MODE_SIMPLE)
4794                 return;
4795
4796         if (!is_fstree(root))
4797                 return;
4798
4799         btrfs_qgroup_free_refroot(fs_info, root, rsv_bytes, BTRFS_QGROUP_RSV_DATA);
4800 }
4801
4802 int btrfs_record_squota_delta(struct btrfs_fs_info *fs_info,
4803                               struct btrfs_squota_delta *delta)
4804 {
4805         int ret;
4806         struct btrfs_qgroup *qgroup;
4807         struct btrfs_qgroup *qg;
4808         LIST_HEAD(qgroup_list);
4809         u64 root = delta->root;
4810         u64 num_bytes = delta->num_bytes;
4811         const int sign = (delta->is_inc ? 1 : -1);
4812
4813         if (btrfs_qgroup_mode(fs_info) != BTRFS_QGROUP_MODE_SIMPLE)
4814                 return 0;
4815
4816         if (!is_fstree(root))
4817                 return 0;
4818
4819         /* If the extent predates enabling quotas, don't count it. */
4820         if (delta->generation < fs_info->qgroup_enable_gen)
4821                 return 0;
4822
4823         spin_lock(&fs_info->qgroup_lock);
4824         qgroup = find_qgroup_rb(fs_info, root);
4825         if (!qgroup) {
4826                 ret = -ENOENT;
4827                 goto out;
4828         }
4829
4830         ret = 0;
4831         qgroup_iterator_add(&qgroup_list, qgroup);
4832         list_for_each_entry(qg, &qgroup_list, iterator) {
4833                 struct btrfs_qgroup_list *glist;
4834
4835                 qg->excl += num_bytes * sign;
4836                 qg->rfer += num_bytes * sign;
4837                 qgroup_dirty(fs_info, qg);
4838
4839                 list_for_each_entry(glist, &qg->groups, next_group)
4840                         qgroup_iterator_add(&qgroup_list, glist->group);
4841         }
4842         qgroup_iterator_clean(&qgroup_list);
4843
4844 out:
4845         spin_unlock(&fs_info->qgroup_lock);
4846         return ret;
4847 }