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                         struct btrfs_root *tree_root = fs_info->tree_root;
 472
 473                         prealloc = kzalloc(sizeof(*prealloc), GFP_KERNEL);
 474                         if (!prealloc) {
 475                                 ret = -ENOMEM;
 476                                 goto out;
 477                         }
 478                         qgroup = add_qgroup_rb(fs_info, prealloc, found_key.offset);
 479                         /*
 480                          * If a qgroup exists for a subvolume ID, it is possible
 481                          * that subvolume has been deleted, in which case
 482                          * re-using that ID would lead to incorrect accounting.
 483                          *
 484                          * Ensure that we skip any such subvol ids.
 485                          *
 486                          * We don't need to lock because this is only called
 487                          * during mount before we start doing things like creating
 488                          * subvolumes.
 489                          */
 490                         if (is_fstree(qgroup->qgroupid) &&
 491                             qgroup->qgroupid > tree_root->free_objectid)
 492                                 /*
 493                                  * Don't need to check against BTRFS_LAST_FREE_OBJECTID,
 494                                  * as it will get checked on the next call to
 495                                  * btrfs_get_free_objectid.
 496                                  */
 497                                 tree_root->free_objectid = qgroup->qgroupid + 1;
 498                 }
 499                 ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
 500                 if (ret < 0)
 501                         goto out;
 502
 503                 switch (found_key.type) {
 504                 case BTRFS_QGROUP_INFO_KEY: {
 505                         struct btrfs_qgroup_info_item *ptr;
 506
 507                         ptr = btrfs_item_ptr(l, slot,
 508                                              struct btrfs_qgroup_info_item);
 509                         qgroup->rfer = btrfs_qgroup_info_rfer(l, ptr);
 510                         qgroup->rfer_cmpr = btrfs_qgroup_info_rfer_cmpr(l, ptr);
 511                         qgroup->excl = btrfs_qgroup_info_excl(l, ptr);
 512                         qgroup->excl_cmpr = btrfs_qgroup_info_excl_cmpr(l, ptr);
 513                         /* generation currently unused */
 514                         break;
 515                 }
 516                 case BTRFS_QGROUP_LIMIT_KEY: {
 517                         struct btrfs_qgroup_limit_item *ptr;
 518
 519                         ptr = btrfs_item_ptr(l, slot,
 520                                              struct btrfs_qgroup_limit_item);
 521                         qgroup->lim_flags = btrfs_qgroup_limit_flags(l, ptr);
 522                         qgroup->max_rfer = btrfs_qgroup_limit_max_rfer(l, ptr);
 523                         qgroup->max_excl = btrfs_qgroup_limit_max_excl(l, ptr);
 524                         qgroup->rsv_rfer = btrfs_qgroup_limit_rsv_rfer(l, ptr);
 525                         qgroup->rsv_excl = btrfs_qgroup_limit_rsv_excl(l, ptr);
 526                         break;
 527                 }
 528                 }
 529 next1:
 530                 ret = btrfs_next_item(quota_root, path);
 531                 if (ret < 0)
 532                         goto out;
 533                 if (ret)
 534                         break;
 535         }
 536         btrfs_release_path(path);
 537
 538         /*
 539          * pass 2: read all qgroup relations
 540          */
 541         key.objectid = 0;
 542         key.type = BTRFS_QGROUP_RELATION_KEY;
 543         key.offset = 0;
 544         ret = btrfs_search_slot_for_read(quota_root, &key, path, 1, 0);
 545         if (ret)
 546                 goto out;
 547         while (1) {
 548                 struct btrfs_qgroup_list *list = NULL;
 549
 550                 slot = path->slots[0];
 551                 l = path->nodes[0];
 552                 btrfs_item_key_to_cpu(l, &found_key, slot);
 553
 554                 if (found_key.type != BTRFS_QGROUP_RELATION_KEY)
 555                         goto next2;
 556
 557                 if (found_key.objectid > found_key.offset) {
 558                         /* parent <- member, not needed to build config */
 559                         /* FIXME should we omit the key completely? */
 560                         goto next2;
 561                 }
 562
 563                 list = kzalloc(sizeof(*list), GFP_KERNEL);
 564                 if (!list) {
 565                         ret = -ENOMEM;
 566                         goto out;
 567                 }
 568                 ret = add_relation_rb(fs_info, list, found_key.objectid,
 569                                       found_key.offset);
 570                 list = NULL;
 571                 if (ret == -ENOENT) {
 572                         btrfs_warn(fs_info,
 573                                 "orphan qgroup relation 0x%llx->0x%llx",
 574                                 found_key.objectid, found_key.offset);
 575                         ret = 0;        /* ignore the error */
 576                 }
 577                 if (ret)
 578                         goto out;
 579 next2:
 580                 ret = btrfs_next_item(quota_root, path);
 581                 if (ret < 0)
 582                         goto out;
 583                 if (ret)
 584                         break;
 585         }
 586 out:
 587         btrfs_free_path(path);
 588         fs_info->qgroup_flags |= flags;
 589         if (ret >= 0) {
 590                 if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_ON)
 591                         set_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
 592                 if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN)
 593                         ret = qgroup_rescan_init(fs_info, rescan_progress, 0);
 594         } else {
 595                 ulist_free(fs_info->qgroup_ulist);
 596                 fs_info->qgroup_ulist = NULL;
 597                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
 598                 btrfs_sysfs_del_qgroups(fs_info);
 599         }
 600
 601         return ret < 0 ? ret : 0;
 602 }
 603
 604 /*
 605  * Called in close_ctree() when quota is still enabled.  This verifies we don't
 606  * leak some reserved space.
 607  *
 608  * Return false if no reserved space is left.
 609  * Return true if some reserved space is leaked.
 610  */
 611 bool btrfs_check_quota_leak(struct btrfs_fs_info *fs_info)
 612 {
 613         struct rb_node *node;
 614         bool ret = false;
 615
 616         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_DISABLED)
 617                 return ret;
 618         /*
 619          * Since we're unmounting, there is no race and no need to grab qgroup
 620          * lock.  And here we don't go post-order to provide a more user
 621          * friendly sorted result.
 622          */
 623         for (node = rb_first(&fs_info->qgroup_tree); node; node = rb_next(node)) {
 624                 struct btrfs_qgroup *qgroup;
 625                 int i;
 626
 627                 qgroup = rb_entry(node, struct btrfs_qgroup, node);
 628                 for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++) {
 629                         if (qgroup->rsv.values[i]) {
 630                                 ret = true;
 631                                 btrfs_warn(fs_info,
 632                 "qgroup %hu/%llu has unreleased space, type %d rsv %llu",
 633                                    btrfs_qgroup_level(qgroup->qgroupid),
 634                                    btrfs_qgroup_subvolid(qgroup->qgroupid),
 635                                    i, qgroup->rsv.values[i]);
 636                         }
 637                 }
 638         }
 639         return ret;
 640 }
 641
 642 /*
 643  * This is called from close_ctree() or open_ctree() or btrfs_quota_disable(),
 644  * first two are in single-threaded paths.And for the third one, we have set
 645  * quota_root to be null with qgroup_lock held before, so it is safe to clean
 646  * up the in-memory structures without qgroup_lock held.
 647  */
 648 void btrfs_free_qgroup_config(struct btrfs_fs_info *fs_info)
 649 {
 650         struct rb_node *n;
 651         struct btrfs_qgroup *qgroup;
 652
 653         while ((n = rb_first(&fs_info->qgroup_tree))) {
 654                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
 655                 rb_erase(n, &fs_info->qgroup_tree);
 656                 __del_qgroup_rb(fs_info, qgroup);
 657                 btrfs_sysfs_del_one_qgroup(fs_info, qgroup);
 658                 kfree(qgroup);
 659         }
 660         /*
 661          * We call btrfs_free_qgroup_config() when unmounting
 662          * filesystem and disabling quota, so we set qgroup_ulist
 663          * to be null here to avoid double free.
 664          */
 665         ulist_free(fs_info->qgroup_ulist);
 666         fs_info->qgroup_ulist = NULL;
 667         btrfs_sysfs_del_qgroups(fs_info);
 668 }
 669
 670 static int add_qgroup_relation_item(struct btrfs_trans_handle *trans, u64 src,
 671                                     u64 dst)
 672 {
 673         int ret;
 674         struct btrfs_root *quota_root = trans->fs_info->quota_root;
 675         struct btrfs_path *path;
 676         struct btrfs_key key;
 677
 678         path = btrfs_alloc_path();
 679         if (!path)
 680                 return -ENOMEM;
 681
 682         key.objectid = src;
 683         key.type = BTRFS_QGROUP_RELATION_KEY;
 684         key.offset = dst;
 685
 686         ret = btrfs_insert_empty_item(trans, quota_root, path, &key, 0);
 687
 688         btrfs_mark_buffer_dirty(trans, path->nodes[0]);
 689
 690         btrfs_free_path(path);
 691         return ret;
 692 }
 693
 694 static int del_qgroup_relation_item(struct btrfs_trans_handle *trans, u64 src,
 695                                     u64 dst)
 696 {
 697         int ret;
 698         struct btrfs_root *quota_root = trans->fs_info->quota_root;
 699         struct btrfs_path *path;
 700         struct btrfs_key key;
 701
 702         path = btrfs_alloc_path();
 703         if (!path)
 704                 return -ENOMEM;
 705
 706         key.objectid = src;
 707         key.type = BTRFS_QGROUP_RELATION_KEY;
 708         key.offset = dst;
 709
 710         ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
 711         if (ret < 0)
 712                 goto out;
 713
 714         if (ret > 0) {
 715                 ret = -ENOENT;
 716                 goto out;
 717         }
 718
 719         ret = btrfs_del_item(trans, quota_root, path);
 720 out:
 721         btrfs_free_path(path);
 722         return ret;
 723 }
 724
 725 static int add_qgroup_item(struct btrfs_trans_handle *trans,
 726                            struct btrfs_root *quota_root, u64 qgroupid)
 727 {
 728         int ret;
 729         struct btrfs_path *path;
 730         struct btrfs_qgroup_info_item *qgroup_info;
 731         struct btrfs_qgroup_limit_item *qgroup_limit;
 732         struct extent_buffer *leaf;
 733         struct btrfs_key key;
 734
 735         if (btrfs_is_testing(quota_root->fs_info))
 736                 return 0;
 737
 738         path = btrfs_alloc_path();
 739         if (!path)
 740                 return -ENOMEM;
 741
 742         key.objectid = 0;
 743         key.type = BTRFS_QGROUP_INFO_KEY;
 744         key.offset = qgroupid;
 745
 746         /*
 747          * Avoid a transaction abort by catching -EEXIST here. In that
 748          * case, we proceed by re-initializing the existing structure
 749          * on disk.
 750          */
 751
 752         ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
 753                                       sizeof(*qgroup_info));
 754         if (ret && ret != -EEXIST)
 755                 goto out;
 756
 757         leaf = path->nodes[0];
 758         qgroup_info = btrfs_item_ptr(leaf, path->slots[0],
 759                                  struct btrfs_qgroup_info_item);
 760         btrfs_set_qgroup_info_generation(leaf, qgroup_info, trans->transid);
 761         btrfs_set_qgroup_info_rfer(leaf, qgroup_info, 0);
 762         btrfs_set_qgroup_info_rfer_cmpr(leaf, qgroup_info, 0);
 763         btrfs_set_qgroup_info_excl(leaf, qgroup_info, 0);
 764         btrfs_set_qgroup_info_excl_cmpr(leaf, qgroup_info, 0);
 765
 766         btrfs_mark_buffer_dirty(trans, leaf);
 767
 768         btrfs_release_path(path);
 769
 770         key.type = BTRFS_QGROUP_LIMIT_KEY;
 771         ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
 772                                       sizeof(*qgroup_limit));
 773         if (ret && ret != -EEXIST)
 774                 goto out;
 775
 776         leaf = path->nodes[0];
 777         qgroup_limit = btrfs_item_ptr(leaf, path->slots[0],
 778                                   struct btrfs_qgroup_limit_item);
 779         btrfs_set_qgroup_limit_flags(leaf, qgroup_limit, 0);
 780         btrfs_set_qgroup_limit_max_rfer(leaf, qgroup_limit, 0);
 781         btrfs_set_qgroup_limit_max_excl(leaf, qgroup_limit, 0);
 782         btrfs_set_qgroup_limit_rsv_rfer(leaf, qgroup_limit, 0);
 783         btrfs_set_qgroup_limit_rsv_excl(leaf, qgroup_limit, 0);
 784
 785         btrfs_mark_buffer_dirty(trans, leaf);
 786
 787         ret = 0;
 788 out:
 789         btrfs_free_path(path);
 790         return ret;
 791 }
 792
 793 static int del_qgroup_item(struct btrfs_trans_handle *trans, u64 qgroupid)
 794 {
 795         int ret;
 796         struct btrfs_root *quota_root = trans->fs_info->quota_root;
 797         struct btrfs_path *path;
 798         struct btrfs_key key;
 799
 800         path = btrfs_alloc_path();
 801         if (!path)
 802                 return -ENOMEM;
 803
 804         key.objectid = 0;
 805         key.type = BTRFS_QGROUP_INFO_KEY;
 806         key.offset = qgroupid;
 807         ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
 808         if (ret < 0)
 809                 goto out;
 810
 811         if (ret > 0) {
 812                 ret = -ENOENT;
 813                 goto out;
 814         }
 815
 816         ret = btrfs_del_item(trans, quota_root, path);
 817         if (ret)
 818                 goto out;
 819
 820         btrfs_release_path(path);
 821
 822         key.type = BTRFS_QGROUP_LIMIT_KEY;
 823         ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
 824         if (ret < 0)
 825                 goto out;
 826
 827         if (ret > 0) {
 828                 ret = -ENOENT;
 829                 goto out;
 830         }
 831
 832         ret = btrfs_del_item(trans, quota_root, path);
 833
 834 out:
 835         btrfs_free_path(path);
 836         return ret;
 837 }
 838
 839 static int update_qgroup_limit_item(struct btrfs_trans_handle *trans,
 840                                     struct btrfs_qgroup *qgroup)
 841 {
 842         struct btrfs_root *quota_root = trans->fs_info->quota_root;
 843         struct btrfs_path *path;
 844         struct btrfs_key key;
 845         struct extent_buffer *l;
 846         struct btrfs_qgroup_limit_item *qgroup_limit;
 847         int ret;
 848         int slot;
 849
 850         key.objectid = 0;
 851         key.type = BTRFS_QGROUP_LIMIT_KEY;
 852         key.offset = qgroup->qgroupid;
 853
 854         path = btrfs_alloc_path();
 855         if (!path)
 856                 return -ENOMEM;
 857
 858         ret = btrfs_search_slot(trans, quota_root, &key, path, 0, 1);
 859         if (ret > 0)
 860                 ret = -ENOENT;
 861
 862         if (ret)
 863                 goto out;
 864
 865         l = path->nodes[0];
 866         slot = path->slots[0];
 867         qgroup_limit = btrfs_item_ptr(l, slot, struct btrfs_qgroup_limit_item);
 868         btrfs_set_qgroup_limit_flags(l, qgroup_limit, qgroup->lim_flags);
 869         btrfs_set_qgroup_limit_max_rfer(l, qgroup_limit, qgroup->max_rfer);
 870         btrfs_set_qgroup_limit_max_excl(l, qgroup_limit, qgroup->max_excl);
 871         btrfs_set_qgroup_limit_rsv_rfer(l, qgroup_limit, qgroup->rsv_rfer);
 872         btrfs_set_qgroup_limit_rsv_excl(l, qgroup_limit, qgroup->rsv_excl);
 873
 874         btrfs_mark_buffer_dirty(trans, l);
 875
 876 out:
 877         btrfs_free_path(path);
 878         return ret;
 879 }
 880
 881 static int update_qgroup_info_item(struct btrfs_trans_handle *trans,
 882                                    struct btrfs_qgroup *qgroup)
 883 {
 884         struct btrfs_fs_info *fs_info = trans->fs_info;
 885         struct btrfs_root *quota_root = fs_info->quota_root;
 886         struct btrfs_path *path;
 887         struct btrfs_key key;
 888         struct extent_buffer *l;
 889         struct btrfs_qgroup_info_item *qgroup_info;
 890         int ret;
 891         int slot;
 892
 893         if (btrfs_is_testing(fs_info))
 894                 return 0;
 895
 896         key.objectid = 0;
 897         key.type = BTRFS_QGROUP_INFO_KEY;
 898         key.offset = qgroup->qgroupid;
 899
 900         path = btrfs_alloc_path();
 901         if (!path)
 902                 return -ENOMEM;
 903
 904         ret = btrfs_search_slot(trans, quota_root, &key, path, 0, 1);
 905         if (ret > 0)
 906                 ret = -ENOENT;
 907
 908         if (ret)
 909                 goto out;
 910
 911         l = path->nodes[0];
 912         slot = path->slots[0];
 913         qgroup_info = btrfs_item_ptr(l, slot, struct btrfs_qgroup_info_item);
 914         btrfs_set_qgroup_info_generation(l, qgroup_info, trans->transid);
 915         btrfs_set_qgroup_info_rfer(l, qgroup_info, qgroup->rfer);
 916         btrfs_set_qgroup_info_rfer_cmpr(l, qgroup_info, qgroup->rfer_cmpr);
 917         btrfs_set_qgroup_info_excl(l, qgroup_info, qgroup->excl);
 918         btrfs_set_qgroup_info_excl_cmpr(l, qgroup_info, qgroup->excl_cmpr);
 919
 920         btrfs_mark_buffer_dirty(trans, l);
 921
 922 out:
 923         btrfs_free_path(path);
 924         return ret;
 925 }
 926
 927 static int update_qgroup_status_item(struct btrfs_trans_handle *trans)
 928 {
 929         struct btrfs_fs_info *fs_info = trans->fs_info;
 930         struct btrfs_root *quota_root = fs_info->quota_root;
 931         struct btrfs_path *path;
 932         struct btrfs_key key;
 933         struct extent_buffer *l;
 934         struct btrfs_qgroup_status_item *ptr;
 935         int ret;
 936         int slot;
 937
 938         key.objectid = 0;
 939         key.type = BTRFS_QGROUP_STATUS_KEY;
 940         key.offset = 0;
 941
 942         path = btrfs_alloc_path();
 943         if (!path)
 944                 return -ENOMEM;
 945
 946         ret = btrfs_search_slot(trans, quota_root, &key, path, 0, 1);
 947         if (ret > 0)
 948                 ret = -ENOENT;
 949
 950         if (ret)
 951                 goto out;
 952
 953         l = path->nodes[0];
 954         slot = path->slots[0];
 955         ptr = btrfs_item_ptr(l, slot, struct btrfs_qgroup_status_item);
 956         btrfs_set_qgroup_status_flags(l, ptr, fs_info->qgroup_flags &
 957                                       BTRFS_QGROUP_STATUS_FLAGS_MASK);
 958         btrfs_set_qgroup_status_generation(l, ptr, trans->transid);
 959         btrfs_set_qgroup_status_rescan(l, ptr,
 960                                 fs_info->qgroup_rescan_progress.objectid);
 961
 962         btrfs_mark_buffer_dirty(trans, l);
 963
 964 out:
 965         btrfs_free_path(path);
 966         return ret;
 967 }
 968
 969 /*
 970  * called with qgroup_lock held
 971  */
 972 static int btrfs_clean_quota_tree(struct btrfs_trans_handle *trans,
 973                                   struct btrfs_root *root)
 974 {
 975         struct btrfs_path *path;
 976         struct btrfs_key key;
 977         struct extent_buffer *leaf = NULL;
 978         int ret;
 979         int nr = 0;
 980
 981         path = btrfs_alloc_path();
 982         if (!path)
 983                 return -ENOMEM;
 984
 985         key.objectid = 0;
 986         key.offset = 0;
 987         key.type = 0;
 988
 989         while (1) {
 990                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 991                 if (ret < 0)
 992                         goto out;
 993                 leaf = path->nodes[0];
 994                 nr = btrfs_header_nritems(leaf);
 995                 if (!nr)
 996                         break;
 997                 /*
 998                  * delete the leaf one by one
 999                  * since the whole tree is going
1000                  * to be deleted.
1001                  */
1002                 path->slots[0] = 0;
1003                 ret = btrfs_del_items(trans, root, path, 0, nr);
1004                 if (ret)
1005                         goto out;
1006
1007                 btrfs_release_path(path);
1008         }
1009         ret = 0;
1010 out:
1011         btrfs_free_path(path);
1012         return ret;
1013 }
1014
1015 int btrfs_quota_enable(struct btrfs_fs_info *fs_info,
1016                        struct btrfs_ioctl_quota_ctl_args *quota_ctl_args)
1017 {
1018         struct btrfs_root *quota_root;
1019         struct btrfs_root *tree_root = fs_info->tree_root;
1020         struct btrfs_path *path = NULL;
1021         struct btrfs_qgroup_status_item *ptr;
1022         struct extent_buffer *leaf;
1023         struct btrfs_key key;
1024         struct btrfs_key found_key;
1025         struct btrfs_qgroup *qgroup = NULL;
1026         struct btrfs_qgroup *prealloc = NULL;
1027         struct btrfs_trans_handle *trans = NULL;
1028         struct ulist *ulist = NULL;
1029         const bool simple = (quota_ctl_args->cmd == BTRFS_QUOTA_CTL_ENABLE_SIMPLE_QUOTA);
1030         int ret = 0;
1031         int slot;
1032
1033         /*
1034          * We need to have subvol_sem write locked, to prevent races between
1035          * concurrent tasks trying to enable quotas, because we will unlock
1036          * and relock qgroup_ioctl_lock before setting fs_info->quota_root
1037          * and before setting BTRFS_FS_QUOTA_ENABLED.
1038          */
1039         lockdep_assert_held_write(&fs_info->subvol_sem);
1040
1041         if (btrfs_fs_incompat(fs_info, EXTENT_TREE_V2)) {
1042                 btrfs_err(fs_info,
1043                           "qgroups are currently unsupported in extent tree v2");
1044                 return -EINVAL;
1045         }
1046
1047         mutex_lock(&fs_info->qgroup_ioctl_lock);
1048         if (fs_info->quota_root)
1049                 goto out;
1050
1051         ulist = ulist_alloc(GFP_KERNEL);
1052         if (!ulist) {
1053                 ret = -ENOMEM;
1054                 goto out;
1055         }
1056
1057         ret = btrfs_sysfs_add_qgroups(fs_info);
1058         if (ret < 0)
1059                 goto out;
1060
1061         /*
1062          * Unlock qgroup_ioctl_lock before starting the transaction. This is to
1063          * avoid lock acquisition inversion problems (reported by lockdep) between
1064          * qgroup_ioctl_lock and the vfs freeze semaphores, acquired when we
1065          * start a transaction.
1066          * After we started the transaction lock qgroup_ioctl_lock again and
1067          * check if someone else created the quota root in the meanwhile. If so,
1068          * just return success and release the transaction handle.
1069          *
1070          * Also we don't need to worry about someone else calling
1071          * btrfs_sysfs_add_qgroups() after we unlock and getting an error because
1072          * that function returns 0 (success) when the sysfs entries already exist.
1073          */
1074         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1075
1076         /*
1077          * 1 for quota root item
1078          * 1 for BTRFS_QGROUP_STATUS item
1079          *
1080          * Yet we also need 2*n items for a QGROUP_INFO/QGROUP_LIMIT items
1081          * per subvolume. However those are not currently reserved since it
1082          * would be a lot of overkill.
1083          */
1084         trans = btrfs_start_transaction(tree_root, 2);
1085
1086         mutex_lock(&fs_info->qgroup_ioctl_lock);
1087         if (IS_ERR(trans)) {
1088                 ret = PTR_ERR(trans);
1089                 trans = NULL;
1090                 goto out;
1091         }
1092
1093         if (fs_info->quota_root)
1094                 goto out;
1095
1096         fs_info->qgroup_ulist = ulist;
1097         ulist = NULL;
1098
1099         /*
1100          * initially create the quota tree
1101          */
1102         quota_root = btrfs_create_tree(trans, BTRFS_QUOTA_TREE_OBJECTID);
1103         if (IS_ERR(quota_root)) {
1104                 ret =  PTR_ERR(quota_root);
1105                 btrfs_abort_transaction(trans, ret);
1106                 goto out;
1107         }
1108
1109         path = btrfs_alloc_path();
1110         if (!path) {
1111                 ret = -ENOMEM;
1112                 btrfs_abort_transaction(trans, ret);
1113                 goto out_free_root;
1114         }
1115
1116         key.objectid = 0;
1117         key.type = BTRFS_QGROUP_STATUS_KEY;
1118         key.offset = 0;
1119
1120         ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
1121                                       sizeof(*ptr));
1122         if (ret) {
1123                 btrfs_abort_transaction(trans, ret);
1124                 goto out_free_path;
1125         }
1126
1127         leaf = path->nodes[0];
1128         ptr = btrfs_item_ptr(leaf, path->slots[0],
1129                                  struct btrfs_qgroup_status_item);
1130         btrfs_set_qgroup_status_generation(leaf, ptr, trans->transid);
1131         btrfs_set_qgroup_status_version(leaf, ptr, BTRFS_QGROUP_STATUS_VERSION);
1132         fs_info->qgroup_flags = BTRFS_QGROUP_STATUS_FLAG_ON;
1133         if (simple) {
1134                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_SIMPLE_MODE;
1135                 btrfs_set_qgroup_status_enable_gen(leaf, ptr, trans->transid);
1136         } else {
1137                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1138         }
1139         btrfs_set_qgroup_status_flags(leaf, ptr, fs_info->qgroup_flags &
1140                                       BTRFS_QGROUP_STATUS_FLAGS_MASK);
1141         btrfs_set_qgroup_status_rescan(leaf, ptr, 0);
1142
1143         btrfs_mark_buffer_dirty(trans, leaf);
1144
1145         key.objectid = 0;
1146         key.type = BTRFS_ROOT_REF_KEY;
1147         key.offset = 0;
1148
1149         btrfs_release_path(path);
1150         ret = btrfs_search_slot_for_read(tree_root, &key, path, 1, 0);
1151         if (ret > 0)
1152                 goto out_add_root;
1153         if (ret < 0) {
1154                 btrfs_abort_transaction(trans, ret);
1155                 goto out_free_path;
1156         }
1157
1158         while (1) {
1159                 slot = path->slots[0];
1160                 leaf = path->nodes[0];
1161                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1162
1163                 if (found_key.type == BTRFS_ROOT_REF_KEY) {
1164
1165                         /* Release locks on tree_root before we access quota_root */
1166                         btrfs_release_path(path);
1167
1168                         /* We should not have a stray @prealloc pointer. */
1169                         ASSERT(prealloc == NULL);
1170                         prealloc = kzalloc(sizeof(*prealloc), GFP_NOFS);
1171                         if (!prealloc) {
1172                                 ret = -ENOMEM;
1173                                 btrfs_abort_transaction(trans, ret);
1174                                 goto out_free_path;
1175                         }
1176
1177                         ret = add_qgroup_item(trans, quota_root,
1178                                               found_key.offset);
1179                         if (ret) {
1180                                 btrfs_abort_transaction(trans, ret);
1181                                 goto out_free_path;
1182                         }
1183
1184                         qgroup = add_qgroup_rb(fs_info, prealloc, found_key.offset);
1185                         prealloc = NULL;
1186                         if (IS_ERR(qgroup)) {
1187                                 ret = PTR_ERR(qgroup);
1188                                 btrfs_abort_transaction(trans, ret);
1189                                 goto out_free_path;
1190                         }
1191                         ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1192                         if (ret < 0) {
1193                                 btrfs_abort_transaction(trans, ret);
1194                                 goto out_free_path;
1195                         }
1196                         ret = btrfs_search_slot_for_read(tree_root, &found_key,
1197                                                          path, 1, 0);
1198                         if (ret < 0) {
1199                                 btrfs_abort_transaction(trans, ret);
1200                                 goto out_free_path;
1201                         }
1202                         if (ret > 0) {
1203                                 /*
1204                                  * Shouldn't happen, but in case it does we
1205                                  * don't need to do the btrfs_next_item, just
1206                                  * continue.
1207                                  */
1208                                 continue;
1209                         }
1210                 }
1211                 ret = btrfs_next_item(tree_root, path);
1212                 if (ret < 0) {
1213                         btrfs_abort_transaction(trans, ret);
1214                         goto out_free_path;
1215                 }
1216                 if (ret)
1217                         break;
1218         }
1219
1220 out_add_root:
1221         btrfs_release_path(path);
1222         ret = add_qgroup_item(trans, quota_root, BTRFS_FS_TREE_OBJECTID);
1223         if (ret) {
1224                 btrfs_abort_transaction(trans, ret);
1225                 goto out_free_path;
1226         }
1227
1228         ASSERT(prealloc == NULL);
1229         prealloc = kzalloc(sizeof(*prealloc), GFP_NOFS);
1230         if (!prealloc) {
1231                 ret = -ENOMEM;
1232                 goto out_free_path;
1233         }
1234         qgroup = add_qgroup_rb(fs_info, prealloc, BTRFS_FS_TREE_OBJECTID);
1235         prealloc = NULL;
1236         ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1237         if (ret < 0) {
1238                 btrfs_abort_transaction(trans, ret);
1239                 goto out_free_path;
1240         }
1241
1242         fs_info->qgroup_enable_gen = trans->transid;
1243
1244         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1245         /*
1246          * Commit the transaction while not holding qgroup_ioctl_lock, to avoid
1247          * a deadlock with tasks concurrently doing other qgroup operations, such
1248          * adding/removing qgroups or adding/deleting qgroup relations for example,
1249          * because all qgroup operations first start or join a transaction and then
1250          * lock the qgroup_ioctl_lock mutex.
1251          * We are safe from a concurrent task trying to enable quotas, by calling
1252          * this function, since we are serialized by fs_info->subvol_sem.
1253          */
1254         ret = btrfs_commit_transaction(trans);
1255         trans = NULL;
1256         mutex_lock(&fs_info->qgroup_ioctl_lock);
1257         if (ret)
1258                 goto out_free_path;
1259
1260         /*
1261          * Set quota enabled flag after committing the transaction, to avoid
1262          * deadlocks on fs_info->qgroup_ioctl_lock with concurrent snapshot
1263          * creation.
1264          */
1265         spin_lock(&fs_info->qgroup_lock);
1266         fs_info->quota_root = quota_root;
1267         set_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1268         if (simple)
1269                 btrfs_set_fs_incompat(fs_info, SIMPLE_QUOTA);
1270         spin_unlock(&fs_info->qgroup_lock);
1271
1272         /* Skip rescan for simple qgroups. */
1273         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE)
1274                 goto out_free_path;
1275
1276         ret = qgroup_rescan_init(fs_info, 0, 1);
1277         if (!ret) {
1278                 qgroup_rescan_zero_tracking(fs_info);
1279                 fs_info->qgroup_rescan_running = true;
1280                 btrfs_queue_work(fs_info->qgroup_rescan_workers,
1281                                  &fs_info->qgroup_rescan_work);
1282         } else {
1283                 /*
1284                  * We have set both BTRFS_FS_QUOTA_ENABLED and
1285                  * BTRFS_QGROUP_STATUS_FLAG_ON, so we can only fail with
1286                  * -EINPROGRESS. That can happen because someone started the
1287                  * rescan worker by calling quota rescan ioctl before we
1288                  * attempted to initialize the rescan worker. Failure due to
1289                  * quotas disabled in the meanwhile is not possible, because
1290                  * we are holding a write lock on fs_info->subvol_sem, which
1291                  * is also acquired when disabling quotas.
1292                  * Ignore such error, and any other error would need to undo
1293                  * everything we did in the transaction we just committed.
1294                  */
1295                 ASSERT(ret == -EINPROGRESS);
1296                 ret = 0;
1297         }
1298
1299 out_free_path:
1300         btrfs_free_path(path);
1301 out_free_root:
1302         if (ret)
1303                 btrfs_put_root(quota_root);
1304 out:
1305         if (ret) {
1306                 ulist_free(fs_info->qgroup_ulist);
1307                 fs_info->qgroup_ulist = NULL;
1308                 btrfs_sysfs_del_qgroups(fs_info);
1309         }
1310         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1311         if (ret && trans)
1312                 btrfs_end_transaction(trans);
1313         else if (trans)
1314                 ret = btrfs_end_transaction(trans);
1315         ulist_free(ulist);
1316         kfree(prealloc);
1317         return ret;
1318 }
1319
1320 /*
1321  * It is possible to have outstanding ordered extents which reserved bytes
1322  * before we disabled. We need to fully flush delalloc, ordered extents, and a
1323  * commit to ensure that we don't leak such reservations, only to have them
1324  * come back if we re-enable.
1325  *
1326  * - enable simple quotas
1327  * - reserve space
1328  * - release it, store rsv_bytes in OE
1329  * - disable quotas
1330  * - enable simple quotas (qgroup rsv are all 0)
1331  * - OE finishes
1332  * - run delayed refs
1333  * - free rsv_bytes, resulting in miscounting or even underflow
1334  */
1335 static int flush_reservations(struct btrfs_fs_info *fs_info)
1336 {
1337         struct btrfs_trans_handle *trans;
1338         int ret;
1339
1340         ret = btrfs_start_delalloc_roots(fs_info, LONG_MAX, false);
1341         if (ret)
1342                 return ret;
1343         btrfs_wait_ordered_roots(fs_info, U64_MAX, 0, (u64)-1);
1344         trans = btrfs_join_transaction(fs_info->tree_root);
1345         if (IS_ERR(trans))
1346                 return PTR_ERR(trans);
1347         ret = btrfs_commit_transaction(trans);
1348
1349         return ret;
1350 }
1351
1352 int btrfs_quota_disable(struct btrfs_fs_info *fs_info)
1353 {
1354         struct btrfs_root *quota_root = NULL;
1355         struct btrfs_trans_handle *trans = NULL;
1356         int ret = 0;
1357
1358         /*
1359          * We need to have subvol_sem write locked to prevent races with
1360          * snapshot creation.
1361          */
1362         lockdep_assert_held_write(&fs_info->subvol_sem);
1363
1364         /*
1365          * Relocation will mess with backrefs, so make sure we have the
1366          * cleaner_mutex held to protect us from relocate.
1367          */
1368         lockdep_assert_held(&fs_info->cleaner_mutex);
1369
1370         mutex_lock(&fs_info->qgroup_ioctl_lock);
1371         if (!fs_info->quota_root)
1372                 goto out;
1373
1374         /*
1375          * Unlock the qgroup_ioctl_lock mutex before waiting for the rescan worker to
1376          * complete. Otherwise we can deadlock because btrfs_remove_qgroup() needs
1377          * to lock that mutex while holding a transaction handle and the rescan
1378          * worker needs to commit a transaction.
1379          */
1380         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1381
1382         /*
1383          * Request qgroup rescan worker to complete and wait for it. This wait
1384          * must be done before transaction start for quota disable since it may
1385          * deadlock with transaction by the qgroup rescan worker.
1386          */
1387         clear_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1388         btrfs_qgroup_wait_for_completion(fs_info, false);
1389
1390         /*
1391          * We have nothing held here and no trans handle, just return the error
1392          * if there is one.
1393          */
1394         ret = flush_reservations(fs_info);
1395         if (ret)
1396                 return ret;
1397
1398         /*
1399          * 1 For the root item
1400          *
1401          * We should also reserve enough items for the quota tree deletion in
1402          * btrfs_clean_quota_tree but this is not done.
1403          *
1404          * Also, we must always start a transaction without holding the mutex
1405          * qgroup_ioctl_lock, see btrfs_quota_enable().
1406          */
1407         trans = btrfs_start_transaction(fs_info->tree_root, 1);
1408
1409         mutex_lock(&fs_info->qgroup_ioctl_lock);
1410         if (IS_ERR(trans)) {
1411                 ret = PTR_ERR(trans);
1412                 trans = NULL;
1413                 set_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1414                 goto out;
1415         }
1416
1417         if (!fs_info->quota_root)
1418                 goto out;
1419
1420         spin_lock(&fs_info->qgroup_lock);
1421         quota_root = fs_info->quota_root;
1422         fs_info->quota_root = NULL;
1423         fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_ON;
1424         fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_SIMPLE_MODE;
1425         fs_info->qgroup_drop_subtree_thres = BTRFS_MAX_LEVEL;
1426         spin_unlock(&fs_info->qgroup_lock);
1427
1428         btrfs_free_qgroup_config(fs_info);
1429
1430         ret = btrfs_clean_quota_tree(trans, quota_root);
1431         if (ret) {
1432                 btrfs_abort_transaction(trans, ret);
1433                 goto out;
1434         }
1435
1436         ret = btrfs_del_root(trans, &quota_root->root_key);
1437         if (ret) {
1438                 btrfs_abort_transaction(trans, ret);
1439                 goto out;
1440         }
1441
1442         spin_lock(&fs_info->trans_lock);
1443         list_del(&quota_root->dirty_list);
1444         spin_unlock(&fs_info->trans_lock);
1445
1446         btrfs_tree_lock(quota_root->node);
1447         btrfs_clear_buffer_dirty(trans, quota_root->node);
1448         btrfs_tree_unlock(quota_root->node);
1449         btrfs_free_tree_block(trans, btrfs_root_id(quota_root),
1450                               quota_root->node, 0, 1);
1451
1452
1453 out:
1454         btrfs_put_root(quota_root);
1455         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1456         if (ret && trans)
1457                 btrfs_end_transaction(trans);
1458         else if (trans)
1459                 ret = btrfs_commit_transaction(trans);
1460         return ret;
1461 }
1462
1463 static void qgroup_dirty(struct btrfs_fs_info *fs_info,
1464                          struct btrfs_qgroup *qgroup)
1465 {
1466         if (list_empty(&qgroup->dirty))
1467                 list_add(&qgroup->dirty, &fs_info->dirty_qgroups);
1468 }
1469
1470 static void qgroup_iterator_add(struct list_head *head, struct btrfs_qgroup *qgroup)
1471 {
1472         if (!list_empty(&qgroup->iterator))
1473                 return;
1474
1475         list_add_tail(&qgroup->iterator, head);
1476 }
1477
1478 static void qgroup_iterator_clean(struct list_head *head)
1479 {
1480         while (!list_empty(head)) {
1481                 struct btrfs_qgroup *qgroup;
1482
1483                 qgroup = list_first_entry(head, struct btrfs_qgroup, iterator);
1484                 list_del_init(&qgroup->iterator);
1485         }
1486 }
1487
1488 /*
1489  * The easy accounting, we're updating qgroup relationship whose child qgroup
1490  * only has exclusive extents.
1491  *
1492  * In this case, all exclusive extents will also be exclusive for parent, so
1493  * excl/rfer just get added/removed.
1494  *
1495  * So is qgroup reservation space, which should also be added/removed to
1496  * parent.
1497  * Or when child tries to release reservation space, parent will underflow its
1498  * reservation (for relationship adding case).
1499  *
1500  * Caller should hold fs_info->qgroup_lock.
1501  */
1502 static int __qgroup_excl_accounting(struct btrfs_fs_info *fs_info, u64 ref_root,
1503                                     struct btrfs_qgroup *src, int sign)
1504 {
1505         struct btrfs_qgroup *qgroup;
1506         struct btrfs_qgroup *cur;
1507         LIST_HEAD(qgroup_list);
1508         u64 num_bytes = src->excl;
1509         int ret = 0;
1510
1511         qgroup = find_qgroup_rb(fs_info, ref_root);
1512         if (!qgroup)
1513                 goto out;
1514
1515         qgroup_iterator_add(&qgroup_list, qgroup);
1516         list_for_each_entry(cur, &qgroup_list, iterator) {
1517                 struct btrfs_qgroup_list *glist;
1518
1519                 qgroup->rfer += sign * num_bytes;
1520                 qgroup->rfer_cmpr += sign * num_bytes;
1521
1522                 WARN_ON(sign < 0 && qgroup->excl < num_bytes);
1523                 qgroup->excl += sign * num_bytes;
1524                 qgroup->excl_cmpr += sign * num_bytes;
1525
1526                 if (sign > 0)
1527                         qgroup_rsv_add_by_qgroup(fs_info, qgroup, src);
1528                 else
1529                         qgroup_rsv_release_by_qgroup(fs_info, qgroup, src);
1530                 qgroup_dirty(fs_info, qgroup);
1531
1532                 /* Append parent qgroups to @qgroup_list. */
1533                 list_for_each_entry(glist, &qgroup->groups, next_group)
1534                         qgroup_iterator_add(&qgroup_list, glist->group);
1535         }
1536         ret = 0;
1537 out:
1538         qgroup_iterator_clean(&qgroup_list);
1539         return ret;
1540 }
1541
1542
1543 /*
1544  * Quick path for updating qgroup with only excl refs.
1545  *
1546  * In that case, just update all parent will be enough.
1547  * Or we needs to do a full rescan.
1548  * Caller should also hold fs_info->qgroup_lock.
1549  *
1550  * Return 0 for quick update, return >0 for need to full rescan
1551  * and mark INCONSISTENT flag.
1552  * Return < 0 for other error.
1553  */
1554 static int quick_update_accounting(struct btrfs_fs_info *fs_info,
1555                                    u64 src, u64 dst, int sign)
1556 {
1557         struct btrfs_qgroup *qgroup;
1558         int ret = 1;
1559
1560         qgroup = find_qgroup_rb(fs_info, src);
1561         if (!qgroup)
1562                 goto out;
1563         if (qgroup->excl == qgroup->rfer) {
1564                 ret = __qgroup_excl_accounting(fs_info, dst, qgroup, sign);
1565                 if (ret < 0)
1566                         goto out;
1567                 ret = 0;
1568         }
1569 out:
1570         if (ret)
1571                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1572         return ret;
1573 }
1574
1575 int btrfs_add_qgroup_relation(struct btrfs_trans_handle *trans, u64 src, u64 dst)
1576 {
1577         struct btrfs_fs_info *fs_info = trans->fs_info;
1578         struct btrfs_qgroup *parent;
1579         struct btrfs_qgroup *member;
1580         struct btrfs_qgroup_list *list;
1581         struct btrfs_qgroup_list *prealloc = NULL;
1582         int ret = 0;
1583
1584         /* Check the level of src and dst first */
1585         if (btrfs_qgroup_level(src) >= btrfs_qgroup_level(dst))
1586                 return -EINVAL;
1587
1588         mutex_lock(&fs_info->qgroup_ioctl_lock);
1589         if (!fs_info->quota_root) {
1590                 ret = -ENOTCONN;
1591                 goto out;
1592         }
1593         member = find_qgroup_rb(fs_info, src);
1594         parent = find_qgroup_rb(fs_info, dst);
1595         if (!member || !parent) {
1596                 ret = -EINVAL;
1597                 goto out;
1598         }
1599
1600         /* check if such qgroup relation exist firstly */
1601         list_for_each_entry(list, &member->groups, next_group) {
1602                 if (list->group == parent) {
1603                         ret = -EEXIST;
1604                         goto out;
1605                 }
1606         }
1607
1608         prealloc = kzalloc(sizeof(*list), GFP_NOFS);
1609         if (!prealloc) {
1610                 ret = -ENOMEM;
1611                 goto out;
1612         }
1613         ret = add_qgroup_relation_item(trans, src, dst);
1614         if (ret)
1615                 goto out;
1616
1617         ret = add_qgroup_relation_item(trans, dst, src);
1618         if (ret) {
1619                 del_qgroup_relation_item(trans, src, dst);
1620                 goto out;
1621         }
1622
1623         spin_lock(&fs_info->qgroup_lock);
1624         ret = __add_relation_rb(prealloc, member, parent);
1625         prealloc = NULL;
1626         if (ret < 0) {
1627                 spin_unlock(&fs_info->qgroup_lock);
1628                 goto out;
1629         }
1630         ret = quick_update_accounting(fs_info, src, dst, 1);
1631         spin_unlock(&fs_info->qgroup_lock);
1632 out:
1633         kfree(prealloc);
1634         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1635         return ret;
1636 }
1637
1638 static int __del_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1639                                  u64 dst)
1640 {
1641         struct btrfs_fs_info *fs_info = trans->fs_info;
1642         struct btrfs_qgroup *parent;
1643         struct btrfs_qgroup *member;
1644         struct btrfs_qgroup_list *list;
1645         bool found = false;
1646         int ret = 0;
1647         int ret2;
1648
1649         if (!fs_info->quota_root) {
1650                 ret = -ENOTCONN;
1651                 goto out;
1652         }
1653
1654         member = find_qgroup_rb(fs_info, src);
1655         parent = find_qgroup_rb(fs_info, dst);
1656         /*
1657          * The parent/member pair doesn't exist, then try to delete the dead
1658          * relation items only.
1659          */
1660         if (!member || !parent)
1661                 goto delete_item;
1662
1663         /* check if such qgroup relation exist firstly */
1664         list_for_each_entry(list, &member->groups, next_group) {
1665                 if (list->group == parent) {
1666                         found = true;
1667                         break;
1668                 }
1669         }
1670
1671 delete_item:
1672         ret = del_qgroup_relation_item(trans, src, dst);
1673         if (ret < 0 && ret != -ENOENT)
1674                 goto out;
1675         ret2 = del_qgroup_relation_item(trans, dst, src);
1676         if (ret2 < 0 && ret2 != -ENOENT)
1677                 goto out;
1678
1679         /* At least one deletion succeeded, return 0 */
1680         if (!ret || !ret2)
1681                 ret = 0;
1682
1683         if (found) {
1684                 spin_lock(&fs_info->qgroup_lock);
1685                 del_relation_rb(fs_info, src, dst);
1686                 ret = quick_update_accounting(fs_info, src, dst, -1);
1687                 spin_unlock(&fs_info->qgroup_lock);
1688         }
1689 out:
1690         return ret;
1691 }
1692
1693 int btrfs_del_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1694                               u64 dst)
1695 {
1696         struct btrfs_fs_info *fs_info = trans->fs_info;
1697         int ret = 0;
1698
1699         mutex_lock(&fs_info->qgroup_ioctl_lock);
1700         ret = __del_qgroup_relation(trans, src, dst);
1701         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1702
1703         return ret;
1704 }
1705
1706 int btrfs_create_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid)
1707 {
1708         struct btrfs_fs_info *fs_info = trans->fs_info;
1709         struct btrfs_root *quota_root;
1710         struct btrfs_qgroup *qgroup;
1711         struct btrfs_qgroup *prealloc = NULL;
1712         int ret = 0;
1713
1714         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_DISABLED)
1715                 return 0;
1716
1717         mutex_lock(&fs_info->qgroup_ioctl_lock);
1718         if (!fs_info->quota_root) {
1719                 ret = -ENOTCONN;
1720                 goto out;
1721         }
1722         quota_root = fs_info->quota_root;
1723         qgroup = find_qgroup_rb(fs_info, qgroupid);
1724         if (qgroup) {
1725                 ret = -EEXIST;
1726                 goto out;
1727         }
1728
1729         prealloc = kzalloc(sizeof(*prealloc), GFP_NOFS);
1730         if (!prealloc) {
1731                 ret = -ENOMEM;
1732                 goto out;
1733         }
1734
1735         ret = add_qgroup_item(trans, quota_root, qgroupid);
1736         if (ret)
1737                 goto out;
1738
1739         spin_lock(&fs_info->qgroup_lock);
1740         qgroup = add_qgroup_rb(fs_info, prealloc, qgroupid);
1741         spin_unlock(&fs_info->qgroup_lock);
1742         prealloc = NULL;
1743
1744         ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1745 out:
1746         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1747         kfree(prealloc);
1748         return ret;
1749 }
1750
1751 static bool qgroup_has_usage(struct btrfs_qgroup *qgroup)
1752 {
1753         return (qgroup->rfer > 0 || qgroup->rfer_cmpr > 0 ||
1754                 qgroup->excl > 0 || qgroup->excl_cmpr > 0 ||
1755                 qgroup->rsv.values[BTRFS_QGROUP_RSV_DATA] > 0 ||
1756                 qgroup->rsv.values[BTRFS_QGROUP_RSV_META_PREALLOC] > 0 ||
1757                 qgroup->rsv.values[BTRFS_QGROUP_RSV_META_PERTRANS] > 0);
1758 }
1759
1760 int btrfs_remove_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid)
1761 {
1762         struct btrfs_fs_info *fs_info = trans->fs_info;
1763         struct btrfs_qgroup *qgroup;
1764         struct btrfs_qgroup_list *list;
1765         int ret = 0;
1766
1767         mutex_lock(&fs_info->qgroup_ioctl_lock);
1768         if (!fs_info->quota_root) {
1769                 ret = -ENOTCONN;
1770                 goto out;
1771         }
1772
1773         qgroup = find_qgroup_rb(fs_info, qgroupid);
1774         if (!qgroup) {
1775                 ret = -ENOENT;
1776                 goto out;
1777         }
1778
1779         if (is_fstree(qgroupid) && qgroup_has_usage(qgroup)) {
1780                 ret = -EBUSY;
1781                 goto out;
1782         }
1783
1784         /* Check if there are no children of this qgroup */
1785         if (!list_empty(&qgroup->members)) {
1786                 ret = -EBUSY;
1787                 goto out;
1788         }
1789
1790         ret = del_qgroup_item(trans, qgroupid);
1791         if (ret && ret != -ENOENT)
1792                 goto out;
1793
1794         while (!list_empty(&qgroup->groups)) {
1795                 list = list_first_entry(&qgroup->groups,
1796                                         struct btrfs_qgroup_list, next_group);
1797                 ret = __del_qgroup_relation(trans, qgroupid,
1798                                             list->group->qgroupid);
1799                 if (ret)
1800                         goto out;
1801         }
1802
1803         spin_lock(&fs_info->qgroup_lock);
1804         del_qgroup_rb(fs_info, qgroupid);
1805         spin_unlock(&fs_info->qgroup_lock);
1806
1807         /*
1808          * Remove the qgroup from sysfs now without holding the qgroup_lock
1809          * spinlock, since the sysfs_remove_group() function needs to take
1810          * the mutex kernfs_mutex through kernfs_remove_by_name_ns().
1811          */
1812         btrfs_sysfs_del_one_qgroup(fs_info, qgroup);
1813         kfree(qgroup);
1814 out:
1815         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1816         return ret;
1817 }
1818
1819 int btrfs_limit_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid,
1820                        struct btrfs_qgroup_limit *limit)
1821 {
1822         struct btrfs_fs_info *fs_info = trans->fs_info;
1823         struct btrfs_qgroup *qgroup;
1824         int ret = 0;
1825         /* Sometimes we would want to clear the limit on this qgroup.
1826          * To meet this requirement, we treat the -1 as a special value
1827          * which tell kernel to clear the limit on this qgroup.
1828          */
1829         const u64 CLEAR_VALUE = -1;
1830
1831         mutex_lock(&fs_info->qgroup_ioctl_lock);
1832         if (!fs_info->quota_root) {
1833                 ret = -ENOTCONN;
1834                 goto out;
1835         }
1836
1837         qgroup = find_qgroup_rb(fs_info, qgroupid);
1838         if (!qgroup) {
1839                 ret = -ENOENT;
1840                 goto out;
1841         }
1842
1843         spin_lock(&fs_info->qgroup_lock);
1844         if (limit->flags & BTRFS_QGROUP_LIMIT_MAX_RFER) {
1845                 if (limit->max_rfer == CLEAR_VALUE) {
1846                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_MAX_RFER;
1847                         limit->flags &= ~BTRFS_QGROUP_LIMIT_MAX_RFER;
1848                         qgroup->max_rfer = 0;
1849                 } else {
1850                         qgroup->max_rfer = limit->max_rfer;
1851                 }
1852         }
1853         if (limit->flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) {
1854                 if (limit->max_excl == CLEAR_VALUE) {
1855                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_MAX_EXCL;
1856                         limit->flags &= ~BTRFS_QGROUP_LIMIT_MAX_EXCL;
1857                         qgroup->max_excl = 0;
1858                 } else {
1859                         qgroup->max_excl = limit->max_excl;
1860                 }
1861         }
1862         if (limit->flags & BTRFS_QGROUP_LIMIT_RSV_RFER) {
1863                 if (limit->rsv_rfer == CLEAR_VALUE) {
1864                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_RSV_RFER;
1865                         limit->flags &= ~BTRFS_QGROUP_LIMIT_RSV_RFER;
1866                         qgroup->rsv_rfer = 0;
1867                 } else {
1868                         qgroup->rsv_rfer = limit->rsv_rfer;
1869                 }
1870         }
1871         if (limit->flags & BTRFS_QGROUP_LIMIT_RSV_EXCL) {
1872                 if (limit->rsv_excl == CLEAR_VALUE) {
1873                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_RSV_EXCL;
1874                         limit->flags &= ~BTRFS_QGROUP_LIMIT_RSV_EXCL;
1875                         qgroup->rsv_excl = 0;
1876                 } else {
1877                         qgroup->rsv_excl = limit->rsv_excl;
1878                 }
1879         }
1880         qgroup->lim_flags |= limit->flags;
1881
1882         spin_unlock(&fs_info->qgroup_lock);
1883
1884         ret = update_qgroup_limit_item(trans, qgroup);
1885         if (ret) {
1886                 qgroup_mark_inconsistent(fs_info);
1887                 btrfs_info(fs_info, "unable to update quota limit for %llu",
1888                        qgroupid);
1889         }
1890
1891 out:
1892         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1893         return ret;
1894 }
1895
1896 /*
1897  * Inform qgroup to trace one dirty extent, its info is recorded in @record.
1898  * So qgroup can account it at transaction committing time.
1899  *
1900  * No lock version, caller must acquire delayed ref lock and allocated memory,
1901  * then call btrfs_qgroup_trace_extent_post() after exiting lock context.
1902  *
1903  * Return 0 for success insert
1904  * Return >0 for existing record, caller can free @record safely.
1905  * Error is not possible
1906  */
1907 int btrfs_qgroup_trace_extent_nolock(struct btrfs_fs_info *fs_info,
1908                                 struct btrfs_delayed_ref_root *delayed_refs,
1909                                 struct btrfs_qgroup_extent_record *record)
1910 {
1911         struct rb_node **p = &delayed_refs->dirty_extent_root.rb_node;
1912         struct rb_node *parent_node = NULL;
1913         struct btrfs_qgroup_extent_record *entry;
1914         u64 bytenr = record->bytenr;
1915
1916         if (!btrfs_qgroup_full_accounting(fs_info))
1917                 return 1;
1918
1919         lockdep_assert_held(&delayed_refs->lock);
1920         trace_btrfs_qgroup_trace_extent(fs_info, record);
1921
1922         while (*p) {
1923                 parent_node = *p;
1924                 entry = rb_entry(parent_node, struct btrfs_qgroup_extent_record,
1925                                  node);
1926                 if (bytenr < entry->bytenr) {
1927                         p = &(*p)->rb_left;
1928                 } else if (bytenr > entry->bytenr) {
1929                         p = &(*p)->rb_right;
1930                 } else {
1931                         if (record->data_rsv && !entry->data_rsv) {
1932                                 entry->data_rsv = record->data_rsv;
1933                                 entry->data_rsv_refroot =
1934                                         record->data_rsv_refroot;
1935                         }
1936                         return 1;
1937                 }
1938         }
1939
1940         rb_link_node(&record->node, parent_node, p);
1941         rb_insert_color(&record->node, &delayed_refs->dirty_extent_root);
1942         return 0;
1943 }
1944
1945 /*
1946  * Post handler after qgroup_trace_extent_nolock().
1947  *
1948  * NOTE: Current qgroup does the expensive backref walk at transaction
1949  * committing time with TRANS_STATE_COMMIT_DOING, this blocks incoming
1950  * new transaction.
1951  * This is designed to allow btrfs_find_all_roots() to get correct new_roots
1952  * result.
1953  *
1954  * However for old_roots there is no need to do backref walk at that time,
1955  * since we search commit roots to walk backref and result will always be
1956  * correct.
1957  *
1958  * Due to the nature of no lock version, we can't do backref there.
1959  * So we must call btrfs_qgroup_trace_extent_post() after exiting
1960  * spinlock context.
1961  *
1962  * TODO: If we can fix and prove btrfs_find_all_roots() can get correct result
1963  * using current root, then we can move all expensive backref walk out of
1964  * transaction committing, but not now as qgroup accounting will be wrong again.
1965  */
1966 int btrfs_qgroup_trace_extent_post(struct btrfs_trans_handle *trans,
1967                                    struct btrfs_qgroup_extent_record *qrecord)
1968 {
1969         struct btrfs_backref_walk_ctx ctx = { 0 };
1970         int ret;
1971
1972         if (!btrfs_qgroup_full_accounting(trans->fs_info))
1973                 return 0;
1974         /*
1975          * We are always called in a context where we are already holding a
1976          * transaction handle. Often we are called when adding a data delayed
1977          * reference from btrfs_truncate_inode_items() (truncating or unlinking),
1978          * in which case we will be holding a write lock on extent buffer from a
1979          * subvolume tree. In this case we can't allow btrfs_find_all_roots() to
1980          * acquire fs_info->commit_root_sem, because that is a higher level lock
1981          * that must be acquired before locking any extent buffers.
1982          *
1983          * So we want btrfs_find_all_roots() to not acquire the commit_root_sem
1984          * but we can't pass it a non-NULL transaction handle, because otherwise
1985          * it would not use commit roots and would lock extent buffers, causing
1986          * a deadlock if it ends up trying to read lock the same extent buffer
1987          * that was previously write locked at btrfs_truncate_inode_items().
1988          *
1989          * So pass a NULL transaction handle to btrfs_find_all_roots() and
1990          * explicitly tell it to not acquire the commit_root_sem - if we are
1991          * holding a transaction handle we don't need its protection.
1992          */
1993         ASSERT(trans != NULL);
1994
1995         if (trans->fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING)
1996                 return 0;
1997
1998         ctx.bytenr = qrecord->bytenr;
1999         ctx.fs_info = trans->fs_info;
2000
2001         ret = btrfs_find_all_roots(&ctx, true);
2002         if (ret < 0) {
2003                 qgroup_mark_inconsistent(trans->fs_info);
2004                 btrfs_warn(trans->fs_info,
2005 "error accounting new delayed refs extent (err code: %d), quota inconsistent",
2006                         ret);
2007                 return 0;
2008         }
2009
2010         /*
2011          * Here we don't need to get the lock of
2012          * trans->transaction->delayed_refs, since inserted qrecord won't
2013          * be deleted, only qrecord->node may be modified (new qrecord insert)
2014          *
2015          * So modifying qrecord->old_roots is safe here
2016          */
2017         qrecord->old_roots = ctx.roots;
2018         return 0;
2019 }
2020
2021 /*
2022  * Inform qgroup to trace one dirty extent, specified by @bytenr and
2023  * @num_bytes.
2024  * So qgroup can account it at commit trans time.
2025  *
2026  * Better encapsulated version, with memory allocation and backref walk for
2027  * commit roots.
2028  * So this can sleep.
2029  *
2030  * Return 0 if the operation is done.
2031  * Return <0 for error, like memory allocation failure or invalid parameter
2032  * (NULL trans)
2033  */
2034 int btrfs_qgroup_trace_extent(struct btrfs_trans_handle *trans, u64 bytenr,
2035                               u64 num_bytes)
2036 {
2037         struct btrfs_fs_info *fs_info = trans->fs_info;
2038         struct btrfs_qgroup_extent_record *record;
2039         struct btrfs_delayed_ref_root *delayed_refs;
2040         int ret;
2041
2042         if (!btrfs_qgroup_full_accounting(fs_info) || bytenr == 0 || num_bytes == 0)
2043                 return 0;
2044         record = kzalloc(sizeof(*record), GFP_NOFS);
2045         if (!record)
2046                 return -ENOMEM;
2047
2048         delayed_refs = &trans->transaction->delayed_refs;
2049         record->bytenr = bytenr;
2050         record->num_bytes = num_bytes;
2051         record->old_roots = NULL;
2052
2053         spin_lock(&delayed_refs->lock);
2054         ret = btrfs_qgroup_trace_extent_nolock(fs_info, delayed_refs, record);
2055         spin_unlock(&delayed_refs->lock);
2056         if (ret > 0) {
2057                 kfree(record);
2058                 return 0;
2059         }
2060         return btrfs_qgroup_trace_extent_post(trans, record);
2061 }
2062
2063 /*
2064  * Inform qgroup to trace all leaf items of data
2065  *
2066  * Return 0 for success
2067  * Return <0 for error(ENOMEM)
2068  */
2069 int btrfs_qgroup_trace_leaf_items(struct btrfs_trans_handle *trans,
2070                                   struct extent_buffer *eb)
2071 {
2072         struct btrfs_fs_info *fs_info = trans->fs_info;
2073         int nr = btrfs_header_nritems(eb);
2074         int i, extent_type, ret;
2075         struct btrfs_key key;
2076         struct btrfs_file_extent_item *fi;
2077         u64 bytenr, num_bytes;
2078
2079         /* We can be called directly from walk_up_proc() */
2080         if (!btrfs_qgroup_full_accounting(fs_info))
2081                 return 0;
2082
2083         for (i = 0; i < nr; i++) {
2084                 btrfs_item_key_to_cpu(eb, &key, i);
2085
2086                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2087                         continue;
2088
2089                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
2090                 /* filter out non qgroup-accountable extents  */
2091                 extent_type = btrfs_file_extent_type(eb, fi);
2092
2093                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
2094                         continue;
2095
2096                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
2097                 if (!bytenr)
2098                         continue;
2099
2100                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
2101
2102                 ret = btrfs_qgroup_trace_extent(trans, bytenr, num_bytes);
2103                 if (ret)
2104                         return ret;
2105         }
2106         cond_resched();
2107         return 0;
2108 }
2109
2110 /*
2111  * Walk up the tree from the bottom, freeing leaves and any interior
2112  * nodes which have had all slots visited. If a node (leaf or
2113  * interior) is freed, the node above it will have it's slot
2114  * incremented. The root node will never be freed.
2115  *
2116  * At the end of this function, we should have a path which has all
2117  * slots incremented to the next position for a search. If we need to
2118  * read a new node it will be NULL and the node above it will have the
2119  * correct slot selected for a later read.
2120  *
2121  * If we increment the root nodes slot counter past the number of
2122  * elements, 1 is returned to signal completion of the search.
2123  */
2124 static int adjust_slots_upwards(struct btrfs_path *path, int root_level)
2125 {
2126         int level = 0;
2127         int nr, slot;
2128         struct extent_buffer *eb;
2129
2130         if (root_level == 0)
2131                 return 1;
2132
2133         while (level <= root_level) {
2134                 eb = path->nodes[level];
2135                 nr = btrfs_header_nritems(eb);
2136                 path->slots[level]++;
2137                 slot = path->slots[level];
2138                 if (slot >= nr || level == 0) {
2139                         /*
2140                          * Don't free the root -  we will detect this
2141                          * condition after our loop and return a
2142                          * positive value for caller to stop walking the tree.
2143                          */
2144                         if (level != root_level) {
2145                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
2146                                 path->locks[level] = 0;
2147
2148                                 free_extent_buffer(eb);
2149                                 path->nodes[level] = NULL;
2150                                 path->slots[level] = 0;
2151                         }
2152                 } else {
2153                         /*
2154                          * We have a valid slot to walk back down
2155                          * from. Stop here so caller can process these
2156                          * new nodes.
2157                          */
2158                         break;
2159                 }
2160
2161                 level++;
2162         }
2163
2164         eb = path->nodes[root_level];
2165         if (path->slots[root_level] >= btrfs_header_nritems(eb))
2166                 return 1;
2167
2168         return 0;
2169 }
2170
2171 /*
2172  * Helper function to trace a subtree tree block swap.
2173  *
2174  * The swap will happen in highest tree block, but there may be a lot of
2175  * tree blocks involved.
2176  *
2177  * For example:
2178  *  OO = Old tree blocks
2179  *  NN = New tree blocks allocated during balance
2180  *
2181  *           File tree (257)                  Reloc tree for 257
2182  * L2              OO                                NN
2183  *               /    \                            /    \
2184  * L1          OO      OO (a)                    OO      NN (a)
2185  *            / \     / \                       / \     / \
2186  * L0       OO   OO OO   OO                   OO   OO NN   NN
2187  *                  (b)  (c)                          (b)  (c)
2188  *
2189  * When calling qgroup_trace_extent_swap(), we will pass:
2190  * @src_eb = OO(a)
2191  * @dst_path = [ nodes[1] = NN(a), nodes[0] = NN(c) ]
2192  * @dst_level = 0
2193  * @root_level = 1
2194  *
2195  * In that case, qgroup_trace_extent_swap() will search from OO(a) to
2196  * reach OO(c), then mark both OO(c) and NN(c) as qgroup dirty.
2197  *
2198  * The main work of qgroup_trace_extent_swap() can be split into 3 parts:
2199  *
2200  * 1) Tree search from @src_eb
2201  *    It should acts as a simplified btrfs_search_slot().
2202  *    The key for search can be extracted from @dst_path->nodes[dst_level]
2203  *    (first key).
2204  *
2205  * 2) Mark the final tree blocks in @src_path and @dst_path qgroup dirty
2206  *    NOTE: In above case, OO(a) and NN(a) won't be marked qgroup dirty.
2207  *    They should be marked during previous (@dst_level = 1) iteration.
2208  *
2209  * 3) Mark file extents in leaves dirty
2210  *    We don't have good way to pick out new file extents only.
2211  *    So we still follow the old method by scanning all file extents in
2212  *    the leave.
2213  *
2214  * This function can free us from keeping two paths, thus later we only need
2215  * to care about how to iterate all new tree blocks in reloc tree.
2216  */
2217 static int qgroup_trace_extent_swap(struct btrfs_trans_handle* trans,
2218                                     struct extent_buffer *src_eb,
2219                                     struct btrfs_path *dst_path,
2220                                     int dst_level, int root_level,
2221                                     bool trace_leaf)
2222 {
2223         struct btrfs_key key;
2224         struct btrfs_path *src_path;
2225         struct btrfs_fs_info *fs_info = trans->fs_info;
2226         u32 nodesize = fs_info->nodesize;
2227         int cur_level = root_level;
2228         int ret;
2229
2230         BUG_ON(dst_level > root_level);
2231         /* Level mismatch */
2232         if (btrfs_header_level(src_eb) != root_level)
2233                 return -EINVAL;
2234
2235         src_path = btrfs_alloc_path();
2236         if (!src_path) {
2237                 ret = -ENOMEM;
2238                 goto out;
2239         }
2240
2241         if (dst_level)
2242                 btrfs_node_key_to_cpu(dst_path->nodes[dst_level], &key, 0);
2243         else
2244                 btrfs_item_key_to_cpu(dst_path->nodes[dst_level], &key, 0);
2245
2246         /* For src_path */
2247         atomic_inc(&src_eb->refs);
2248         src_path->nodes[root_level] = src_eb;
2249         src_path->slots[root_level] = dst_path->slots[root_level];
2250         src_path->locks[root_level] = 0;
2251
2252         /* A simplified version of btrfs_search_slot() */
2253         while (cur_level >= dst_level) {
2254                 struct btrfs_key src_key;
2255                 struct btrfs_key dst_key;
2256
2257                 if (src_path->nodes[cur_level] == NULL) {
2258                         struct extent_buffer *eb;
2259                         int parent_slot;
2260
2261                         eb = src_path->nodes[cur_level + 1];
2262                         parent_slot = src_path->slots[cur_level + 1];
2263
2264                         eb = btrfs_read_node_slot(eb, parent_slot);
2265                         if (IS_ERR(eb)) {
2266                                 ret = PTR_ERR(eb);
2267                                 goto out;
2268                         }
2269
2270                         src_path->nodes[cur_level] = eb;
2271
2272                         btrfs_tree_read_lock(eb);
2273                         src_path->locks[cur_level] = BTRFS_READ_LOCK;
2274                 }
2275
2276                 src_path->slots[cur_level] = dst_path->slots[cur_level];
2277                 if (cur_level) {
2278                         btrfs_node_key_to_cpu(dst_path->nodes[cur_level],
2279                                         &dst_key, dst_path->slots[cur_level]);
2280                         btrfs_node_key_to_cpu(src_path->nodes[cur_level],
2281                                         &src_key, src_path->slots[cur_level]);
2282                 } else {
2283                         btrfs_item_key_to_cpu(dst_path->nodes[cur_level],
2284                                         &dst_key, dst_path->slots[cur_level]);
2285                         btrfs_item_key_to_cpu(src_path->nodes[cur_level],
2286                                         &src_key, src_path->slots[cur_level]);
2287                 }
2288                 /* Content mismatch, something went wrong */
2289                 if (btrfs_comp_cpu_keys(&dst_key, &src_key)) {
2290                         ret = -ENOENT;
2291                         goto out;
2292                 }
2293                 cur_level--;
2294         }
2295
2296         /*
2297          * Now both @dst_path and @src_path have been populated, record the tree
2298          * blocks for qgroup accounting.
2299          */
2300         ret = btrfs_qgroup_trace_extent(trans, src_path->nodes[dst_level]->start,
2301                                         nodesize);
2302         if (ret < 0)
2303                 goto out;
2304         ret = btrfs_qgroup_trace_extent(trans, dst_path->nodes[dst_level]->start,
2305                                         nodesize);
2306         if (ret < 0)
2307                 goto out;
2308
2309         /* Record leaf file extents */
2310         if (dst_level == 0 && trace_leaf) {
2311                 ret = btrfs_qgroup_trace_leaf_items(trans, src_path->nodes[0]);
2312                 if (ret < 0)
2313                         goto out;
2314                 ret = btrfs_qgroup_trace_leaf_items(trans, dst_path->nodes[0]);
2315         }
2316 out:
2317         btrfs_free_path(src_path);
2318         return ret;
2319 }
2320
2321 /*
2322  * Helper function to do recursive generation-aware depth-first search, to
2323  * locate all new tree blocks in a subtree of reloc tree.
2324  *
2325  * E.g. (OO = Old tree blocks, NN = New tree blocks, whose gen == last_snapshot)
2326  *         reloc tree
2327  * L2         NN (a)
2328  *          /    \
2329  * L1    OO        NN (b)
2330  *      /  \      /  \
2331  * L0  OO  OO    OO  NN
2332  *               (c) (d)
2333  * If we pass:
2334  * @dst_path = [ nodes[1] = NN(b), nodes[0] = NULL ],
2335  * @cur_level = 1
2336  * @root_level = 1
2337  *
2338  * We will iterate through tree blocks NN(b), NN(d) and info qgroup to trace
2339  * above tree blocks along with their counter parts in file tree.
2340  * While during search, old tree blocks OO(c) will be skipped as tree block swap
2341  * won't affect OO(c).
2342  */
2343 static int qgroup_trace_new_subtree_blocks(struct btrfs_trans_handle* trans,
2344                                            struct extent_buffer *src_eb,
2345                                            struct btrfs_path *dst_path,
2346                                            int cur_level, int root_level,
2347                                            u64 last_snapshot, bool trace_leaf)
2348 {
2349         struct btrfs_fs_info *fs_info = trans->fs_info;
2350         struct extent_buffer *eb;
2351         bool need_cleanup = false;
2352         int ret = 0;
2353         int i;
2354
2355         /* Level sanity check */
2356         if (cur_level < 0 || cur_level >= BTRFS_MAX_LEVEL - 1 ||
2357             root_level < 0 || root_level >= BTRFS_MAX_LEVEL - 1 ||
2358             root_level < cur_level) {
2359                 btrfs_err_rl(fs_info,
2360                         "%s: bad levels, cur_level=%d root_level=%d",
2361                         __func__, cur_level, root_level);
2362                 return -EUCLEAN;
2363         }
2364
2365         /* Read the tree block if needed */
2366         if (dst_path->nodes[cur_level] == NULL) {
2367                 int parent_slot;
2368                 u64 child_gen;
2369
2370                 /*
2371                  * dst_path->nodes[root_level] must be initialized before
2372                  * calling this function.
2373                  */
2374                 if (cur_level == root_level) {
2375                         btrfs_err_rl(fs_info,
2376         "%s: dst_path->nodes[%d] not initialized, root_level=%d cur_level=%d",
2377                                 __func__, root_level, root_level, cur_level);
2378                         return -EUCLEAN;
2379                 }
2380
2381                 /*
2382                  * We need to get child blockptr/gen from parent before we can
2383                  * read it.
2384                   */
2385                 eb = dst_path->nodes[cur_level + 1];
2386                 parent_slot = dst_path->slots[cur_level + 1];
2387                 child_gen = btrfs_node_ptr_generation(eb, parent_slot);
2388
2389                 /* This node is old, no need to trace */
2390                 if (child_gen < last_snapshot)
2391                         goto out;
2392
2393                 eb = btrfs_read_node_slot(eb, parent_slot);
2394                 if (IS_ERR(eb)) {
2395                         ret = PTR_ERR(eb);
2396                         goto out;
2397                 }
2398
2399                 dst_path->nodes[cur_level] = eb;
2400                 dst_path->slots[cur_level] = 0;
2401
2402                 btrfs_tree_read_lock(eb);
2403                 dst_path->locks[cur_level] = BTRFS_READ_LOCK;
2404                 need_cleanup = true;
2405         }
2406
2407         /* Now record this tree block and its counter part for qgroups */
2408         ret = qgroup_trace_extent_swap(trans, src_eb, dst_path, cur_level,
2409                                        root_level, trace_leaf);
2410         if (ret < 0)
2411                 goto cleanup;
2412
2413         eb = dst_path->nodes[cur_level];
2414
2415         if (cur_level > 0) {
2416                 /* Iterate all child tree blocks */
2417                 for (i = 0; i < btrfs_header_nritems(eb); i++) {
2418                         /* Skip old tree blocks as they won't be swapped */
2419                         if (btrfs_node_ptr_generation(eb, i) < last_snapshot)
2420                                 continue;
2421                         dst_path->slots[cur_level] = i;
2422
2423                         /* Recursive call (at most 7 times) */
2424                         ret = qgroup_trace_new_subtree_blocks(trans, src_eb,
2425                                         dst_path, cur_level - 1, root_level,
2426                                         last_snapshot, trace_leaf);
2427                         if (ret < 0)
2428                                 goto cleanup;
2429                 }
2430         }
2431
2432 cleanup:
2433         if (need_cleanup) {
2434                 /* Clean up */
2435                 btrfs_tree_unlock_rw(dst_path->nodes[cur_level],
2436                                      dst_path->locks[cur_level]);
2437                 free_extent_buffer(dst_path->nodes[cur_level]);
2438                 dst_path->nodes[cur_level] = NULL;
2439                 dst_path->slots[cur_level] = 0;
2440                 dst_path->locks[cur_level] = 0;
2441         }
2442 out:
2443         return ret;
2444 }
2445
2446 static int qgroup_trace_subtree_swap(struct btrfs_trans_handle *trans,
2447                                 struct extent_buffer *src_eb,
2448                                 struct extent_buffer *dst_eb,
2449                                 u64 last_snapshot, bool trace_leaf)
2450 {
2451         struct btrfs_fs_info *fs_info = trans->fs_info;
2452         struct btrfs_path *dst_path = NULL;
2453         int level;
2454         int ret;
2455
2456         if (!btrfs_qgroup_full_accounting(fs_info))
2457                 return 0;
2458
2459         /* Wrong parameter order */
2460         if (btrfs_header_generation(src_eb) > btrfs_header_generation(dst_eb)) {
2461                 btrfs_err_rl(fs_info,
2462                 "%s: bad parameter order, src_gen=%llu dst_gen=%llu", __func__,
2463                              btrfs_header_generation(src_eb),
2464                              btrfs_header_generation(dst_eb));
2465                 return -EUCLEAN;
2466         }
2467
2468         if (!extent_buffer_uptodate(src_eb) || !extent_buffer_uptodate(dst_eb)) {
2469                 ret = -EIO;
2470                 goto out;
2471         }
2472
2473         level = btrfs_header_level(dst_eb);
2474         dst_path = btrfs_alloc_path();
2475         if (!dst_path) {
2476                 ret = -ENOMEM;
2477                 goto out;
2478         }
2479         /* For dst_path */
2480         atomic_inc(&dst_eb->refs);
2481         dst_path->nodes[level] = dst_eb;
2482         dst_path->slots[level] = 0;
2483         dst_path->locks[level] = 0;
2484
2485         /* Do the generation aware breadth-first search */
2486         ret = qgroup_trace_new_subtree_blocks(trans, src_eb, dst_path, level,
2487                                               level, last_snapshot, trace_leaf);
2488         if (ret < 0)
2489                 goto out;
2490         ret = 0;
2491
2492 out:
2493         btrfs_free_path(dst_path);
2494         if (ret < 0)
2495                 qgroup_mark_inconsistent(fs_info);
2496         return ret;
2497 }
2498
2499 /*
2500  * Inform qgroup to trace a whole subtree, including all its child tree
2501  * blocks and data.
2502  * The root tree block is specified by @root_eb.
2503  *
2504  * Normally used by relocation(tree block swap) and subvolume deletion.
2505  *
2506  * Return 0 for success
2507  * Return <0 for error(ENOMEM or tree search error)
2508  */
2509 int btrfs_qgroup_trace_subtree(struct btrfs_trans_handle *trans,
2510                                struct extent_buffer *root_eb,
2511                                u64 root_gen, int root_level)
2512 {
2513         struct btrfs_fs_info *fs_info = trans->fs_info;
2514         int ret = 0;
2515         int level;
2516         u8 drop_subptree_thres;
2517         struct extent_buffer *eb = root_eb;
2518         struct btrfs_path *path = NULL;
2519
2520         ASSERT(0 <= root_level && root_level < BTRFS_MAX_LEVEL);
2521         ASSERT(root_eb != NULL);
2522
2523         if (!btrfs_qgroup_full_accounting(fs_info))
2524                 return 0;
2525
2526         spin_lock(&fs_info->qgroup_lock);
2527         drop_subptree_thres = fs_info->qgroup_drop_subtree_thres;
2528         spin_unlock(&fs_info->qgroup_lock);
2529
2530         /*
2531          * This function only gets called for snapshot drop, if we hit a high
2532          * node here, it means we are going to change ownership for quite a lot
2533          * of extents, which will greatly slow down btrfs_commit_transaction().
2534          *
2535          * So here if we find a high tree here, we just skip the accounting and
2536          * mark qgroup inconsistent.
2537          */
2538         if (root_level >= drop_subptree_thres) {
2539                 qgroup_mark_inconsistent(fs_info);
2540                 return 0;
2541         }
2542
2543         if (!extent_buffer_uptodate(root_eb)) {
2544                 struct btrfs_tree_parent_check check = {
2545                         .has_first_key = false,
2546                         .transid = root_gen,
2547                         .level = root_level
2548                 };
2549
2550                 ret = btrfs_read_extent_buffer(root_eb, &check);
2551                 if (ret)
2552                         goto out;
2553         }
2554
2555         if (root_level == 0) {
2556                 ret = btrfs_qgroup_trace_leaf_items(trans, root_eb);
2557                 goto out;
2558         }
2559
2560         path = btrfs_alloc_path();
2561         if (!path)
2562                 return -ENOMEM;
2563
2564         /*
2565          * Walk down the tree.  Missing extent blocks are filled in as
2566          * we go. Metadata is accounted every time we read a new
2567          * extent block.
2568          *
2569          * When we reach a leaf, we account for file extent items in it,
2570          * walk back up the tree (adjusting slot pointers as we go)
2571          * and restart the search process.
2572          */
2573         atomic_inc(&root_eb->refs);     /* For path */
2574         path->nodes[root_level] = root_eb;
2575         path->slots[root_level] = 0;
2576         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
2577 walk_down:
2578         level = root_level;
2579         while (level >= 0) {
2580                 if (path->nodes[level] == NULL) {
2581                         int parent_slot;
2582                         u64 child_bytenr;
2583
2584                         /*
2585                          * We need to get child blockptr from parent before we
2586                          * can read it.
2587                           */
2588                         eb = path->nodes[level + 1];
2589                         parent_slot = path->slots[level + 1];
2590                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
2591
2592                         eb = btrfs_read_node_slot(eb, parent_slot);
2593                         if (IS_ERR(eb)) {
2594                                 ret = PTR_ERR(eb);
2595                                 goto out;
2596                         }
2597
2598                         path->nodes[level] = eb;
2599                         path->slots[level] = 0;
2600
2601                         btrfs_tree_read_lock(eb);
2602                         path->locks[level] = BTRFS_READ_LOCK;
2603
2604                         ret = btrfs_qgroup_trace_extent(trans, child_bytenr,
2605                                                         fs_info->nodesize);
2606                         if (ret)
2607                                 goto out;
2608                 }
2609
2610                 if (level == 0) {
2611                         ret = btrfs_qgroup_trace_leaf_items(trans,
2612                                                             path->nodes[level]);
2613                         if (ret)
2614                                 goto out;
2615
2616                         /* Nonzero return here means we completed our search */
2617                         ret = adjust_slots_upwards(path, root_level);
2618                         if (ret)
2619                                 break;
2620
2621                         /* Restart search with new slots */
2622                         goto walk_down;
2623                 }
2624
2625                 level--;
2626         }
2627
2628         ret = 0;
2629 out:
2630         btrfs_free_path(path);
2631
2632         return ret;
2633 }
2634
2635 static void qgroup_iterator_nested_add(struct list_head *head, struct btrfs_qgroup *qgroup)
2636 {
2637         if (!list_empty(&qgroup->nested_iterator))
2638                 return;
2639
2640         list_add_tail(&qgroup->nested_iterator, head);
2641 }
2642
2643 static void qgroup_iterator_nested_clean(struct list_head *head)
2644 {
2645         while (!list_empty(head)) {
2646                 struct btrfs_qgroup *qgroup;
2647
2648                 qgroup = list_first_entry(head, struct btrfs_qgroup, nested_iterator);
2649                 list_del_init(&qgroup->nested_iterator);
2650         }
2651 }
2652
2653 #define UPDATE_NEW      0
2654 #define UPDATE_OLD      1
2655 /*
2656  * Walk all of the roots that points to the bytenr and adjust their refcnts.
2657  */
2658 static void qgroup_update_refcnt(struct btrfs_fs_info *fs_info,
2659                                  struct ulist *roots, struct list_head *qgroups,
2660                                  u64 seq, int update_old)
2661 {
2662         struct ulist_node *unode;
2663         struct ulist_iterator uiter;
2664         struct btrfs_qgroup *qg;
2665
2666         if (!roots)
2667                 return;
2668         ULIST_ITER_INIT(&uiter);
2669         while ((unode = ulist_next(roots, &uiter))) {
2670                 LIST_HEAD(tmp);
2671
2672                 qg = find_qgroup_rb(fs_info, unode->val);
2673                 if (!qg)
2674                         continue;
2675
2676                 qgroup_iterator_nested_add(qgroups, qg);
2677                 qgroup_iterator_add(&tmp, qg);
2678                 list_for_each_entry(qg, &tmp, iterator) {
2679                         struct btrfs_qgroup_list *glist;
2680
2681                         if (update_old)
2682                                 btrfs_qgroup_update_old_refcnt(qg, seq, 1);
2683                         else
2684                                 btrfs_qgroup_update_new_refcnt(qg, seq, 1);
2685
2686                         list_for_each_entry(glist, &qg->groups, next_group) {
2687                                 qgroup_iterator_nested_add(qgroups, glist->group);
2688                                 qgroup_iterator_add(&tmp, glist->group);
2689                         }
2690                 }
2691                 qgroup_iterator_clean(&tmp);
2692         }
2693 }
2694
2695 /*
2696  * Update qgroup rfer/excl counters.
2697  * Rfer update is easy, codes can explain themselves.
2698  *
2699  * Excl update is tricky, the update is split into 2 parts.
2700  * Part 1: Possible exclusive <-> sharing detect:
2701  *      |       A       |       !A      |
2702  *  -------------------------------------
2703  *  B   |       *       |       -       |
2704  *  -------------------------------------
2705  *  !B  |       +       |       **      |
2706  *  -------------------------------------
2707  *
2708  * Conditions:
2709  * A:   cur_old_roots < nr_old_roots    (not exclusive before)
2710  * !A:  cur_old_roots == nr_old_roots   (possible exclusive before)
2711  * B:   cur_new_roots < nr_new_roots    (not exclusive now)
2712  * !B:  cur_new_roots == nr_new_roots   (possible exclusive now)
2713  *
2714  * Results:
2715  * +: Possible sharing -> exclusive     -: Possible exclusive -> sharing
2716  * *: Definitely not changed.           **: Possible unchanged.
2717  *
2718  * For !A and !B condition, the exception is cur_old/new_roots == 0 case.
2719  *
2720  * To make the logic clear, we first use condition A and B to split
2721  * combination into 4 results.
2722  *
2723  * Then, for result "+" and "-", check old/new_roots == 0 case, as in them
2724  * only on variant maybe 0.
2725  *
2726  * Lastly, check result **, since there are 2 variants maybe 0, split them
2727  * again(2x2).
2728  * But this time we don't need to consider other things, the codes and logic
2729  * is easy to understand now.
2730  */
2731 static void qgroup_update_counters(struct btrfs_fs_info *fs_info,
2732                                    struct list_head *qgroups, u64 nr_old_roots,
2733                                    u64 nr_new_roots, u64 num_bytes, u64 seq)
2734 {
2735         struct btrfs_qgroup *qg;
2736
2737         list_for_each_entry(qg, qgroups, nested_iterator) {
2738                 u64 cur_new_count, cur_old_count;
2739                 bool dirty = false;
2740
2741                 cur_old_count = btrfs_qgroup_get_old_refcnt(qg, seq);
2742                 cur_new_count = btrfs_qgroup_get_new_refcnt(qg, seq);
2743
2744                 trace_qgroup_update_counters(fs_info, qg, cur_old_count,
2745                                              cur_new_count);
2746
2747                 /* Rfer update part */
2748                 if (cur_old_count == 0 && cur_new_count > 0) {
2749                         qg->rfer += num_bytes;
2750                         qg->rfer_cmpr += num_bytes;
2751                         dirty = true;
2752                 }
2753                 if (cur_old_count > 0 && cur_new_count == 0) {
2754                         qg->rfer -= num_bytes;
2755                         qg->rfer_cmpr -= num_bytes;
2756                         dirty = true;
2757                 }
2758
2759                 /* Excl update part */
2760                 /* Exclusive/none -> shared case */
2761                 if (cur_old_count == nr_old_roots &&
2762                     cur_new_count < nr_new_roots) {
2763                         /* Exclusive -> shared */
2764                         if (cur_old_count != 0) {
2765                                 qg->excl -= num_bytes;
2766                                 qg->excl_cmpr -= num_bytes;
2767                                 dirty = true;
2768                         }
2769                 }
2770
2771                 /* Shared -> exclusive/none case */
2772                 if (cur_old_count < nr_old_roots &&
2773                     cur_new_count == nr_new_roots) {
2774                         /* Shared->exclusive */
2775                         if (cur_new_count != 0) {
2776                                 qg->excl += num_bytes;
2777                                 qg->excl_cmpr += num_bytes;
2778                                 dirty = true;
2779                         }
2780                 }
2781
2782                 /* Exclusive/none -> exclusive/none case */
2783                 if (cur_old_count == nr_old_roots &&
2784                     cur_new_count == nr_new_roots) {
2785                         if (cur_old_count == 0) {
2786                                 /* None -> exclusive/none */
2787
2788                                 if (cur_new_count != 0) {
2789                                         /* None -> exclusive */
2790                                         qg->excl += num_bytes;
2791                                         qg->excl_cmpr += num_bytes;
2792                                         dirty = true;
2793                                 }
2794                                 /* None -> none, nothing changed */
2795                         } else {
2796                                 /* Exclusive -> exclusive/none */
2797
2798                                 if (cur_new_count == 0) {
2799                                         /* Exclusive -> none */
2800                                         qg->excl -= num_bytes;
2801                                         qg->excl_cmpr -= num_bytes;
2802                                         dirty = true;
2803                                 }
2804                                 /* Exclusive -> exclusive, nothing changed */
2805                         }
2806                 }
2807
2808                 if (dirty)
2809                         qgroup_dirty(fs_info, qg);
2810         }
2811 }
2812
2813 /*
2814  * Check if the @roots potentially is a list of fs tree roots
2815  *
2816  * Return 0 for definitely not a fs/subvol tree roots ulist
2817  * Return 1 for possible fs/subvol tree roots in the list (considering an empty
2818  *          one as well)
2819  */
2820 static int maybe_fs_roots(struct ulist *roots)
2821 {
2822         struct ulist_node *unode;
2823         struct ulist_iterator uiter;
2824
2825         /* Empty one, still possible for fs roots */
2826         if (!roots || roots->nnodes == 0)
2827                 return 1;
2828
2829         ULIST_ITER_INIT(&uiter);
2830         unode = ulist_next(roots, &uiter);
2831         if (!unode)
2832                 return 1;
2833
2834         /*
2835          * If it contains fs tree roots, then it must belong to fs/subvol
2836          * trees.
2837          * If it contains a non-fs tree, it won't be shared with fs/subvol trees.
2838          */
2839         return is_fstree(unode->val);
2840 }
2841
2842 int btrfs_qgroup_account_extent(struct btrfs_trans_handle *trans, u64 bytenr,
2843                                 u64 num_bytes, struct ulist *old_roots,
2844                                 struct ulist *new_roots)
2845 {
2846         struct btrfs_fs_info *fs_info = trans->fs_info;
2847         LIST_HEAD(qgroups);
2848         u64 seq;
2849         u64 nr_new_roots = 0;
2850         u64 nr_old_roots = 0;
2851         int ret = 0;
2852
2853         /*
2854          * If quotas get disabled meanwhile, the resources need to be freed and
2855          * we can't just exit here.
2856          */
2857         if (!btrfs_qgroup_full_accounting(fs_info) ||
2858             fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING)
2859                 goto out_free;
2860
2861         if (new_roots) {
2862                 if (!maybe_fs_roots(new_roots))
2863                         goto out_free;
2864                 nr_new_roots = new_roots->nnodes;
2865         }
2866         if (old_roots) {
2867                 if (!maybe_fs_roots(old_roots))
2868                         goto out_free;
2869                 nr_old_roots = old_roots->nnodes;
2870         }
2871
2872         /* Quick exit, either not fs tree roots, or won't affect any qgroup */
2873         if (nr_old_roots == 0 && nr_new_roots == 0)
2874                 goto out_free;
2875
2876         trace_btrfs_qgroup_account_extent(fs_info, trans->transid, bytenr,
2877                                         num_bytes, nr_old_roots, nr_new_roots);
2878
2879         mutex_lock(&fs_info->qgroup_rescan_lock);
2880         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
2881                 if (fs_info->qgroup_rescan_progress.objectid <= bytenr) {
2882                         mutex_unlock(&fs_info->qgroup_rescan_lock);
2883                         ret = 0;
2884                         goto out_free;
2885                 }
2886         }
2887         mutex_unlock(&fs_info->qgroup_rescan_lock);
2888
2889         spin_lock(&fs_info->qgroup_lock);
2890         seq = fs_info->qgroup_seq;
2891
2892         /* Update old refcnts using old_roots */
2893         qgroup_update_refcnt(fs_info, old_roots, &qgroups, seq, UPDATE_OLD);
2894
2895         /* Update new refcnts using new_roots */
2896         qgroup_update_refcnt(fs_info, new_roots, &qgroups, seq, UPDATE_NEW);
2897
2898         qgroup_update_counters(fs_info, &qgroups, nr_old_roots, nr_new_roots,
2899                                num_bytes, seq);
2900
2901         /*
2902          * We're done using the iterator, release all its qgroups while holding
2903          * fs_info->qgroup_lock so that we don't race with btrfs_remove_qgroup()
2904          * and trigger use-after-free accesses to qgroups.
2905          */
2906         qgroup_iterator_nested_clean(&qgroups);
2907
2908         /*
2909          * Bump qgroup_seq to avoid seq overlap
2910          */
2911         fs_info->qgroup_seq += max(nr_old_roots, nr_new_roots) + 1;
2912         spin_unlock(&fs_info->qgroup_lock);
2913 out_free:
2914         ulist_free(old_roots);
2915         ulist_free(new_roots);
2916         return ret;
2917 }
2918
2919 int btrfs_qgroup_account_extents(struct btrfs_trans_handle *trans)
2920 {
2921         struct btrfs_fs_info *fs_info = trans->fs_info;
2922         struct btrfs_qgroup_extent_record *record;
2923         struct btrfs_delayed_ref_root *delayed_refs;
2924         struct ulist *new_roots = NULL;
2925         struct rb_node *node;
2926         u64 num_dirty_extents = 0;
2927         u64 qgroup_to_skip;
2928         int ret = 0;
2929
2930         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE)
2931                 return 0;
2932
2933         delayed_refs = &trans->transaction->delayed_refs;
2934         qgroup_to_skip = delayed_refs->qgroup_to_skip;
2935         while ((node = rb_first(&delayed_refs->dirty_extent_root))) {
2936                 record = rb_entry(node, struct btrfs_qgroup_extent_record,
2937                                   node);
2938
2939                 num_dirty_extents++;
2940                 trace_btrfs_qgroup_account_extents(fs_info, record);
2941
2942                 if (!ret && !(fs_info->qgroup_flags &
2943                               BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING)) {
2944                         struct btrfs_backref_walk_ctx ctx = { 0 };
2945
2946                         ctx.bytenr = record->bytenr;
2947                         ctx.fs_info = fs_info;
2948
2949                         /*
2950                          * Old roots should be searched when inserting qgroup
2951                          * extent record.
2952                          *
2953                          * But for INCONSISTENT (NO_ACCOUNTING) -> rescan case,
2954                          * we may have some record inserted during
2955                          * NO_ACCOUNTING (thus no old_roots populated), but
2956                          * later we start rescan, which clears NO_ACCOUNTING,
2957                          * leaving some inserted records without old_roots
2958                          * populated.
2959                          *
2960                          * Those cases are rare and should not cause too much
2961                          * time spent during commit_transaction().
2962                          */
2963                         if (!record->old_roots) {
2964                                 /* Search commit root to find old_roots */
2965                                 ret = btrfs_find_all_roots(&ctx, false);
2966                                 if (ret < 0)
2967                                         goto cleanup;
2968                                 record->old_roots = ctx.roots;
2969                                 ctx.roots = NULL;
2970                         }
2971
2972                         /*
2973                          * Use BTRFS_SEQ_LAST as time_seq to do special search,
2974                          * which doesn't lock tree or delayed_refs and search
2975                          * current root. It's safe inside commit_transaction().
2976                          */
2977                         ctx.trans = trans;
2978                         ctx.time_seq = BTRFS_SEQ_LAST;
2979                         ret = btrfs_find_all_roots(&ctx, false);
2980                         if (ret < 0)
2981                                 goto cleanup;
2982                         new_roots = ctx.roots;
2983                         if (qgroup_to_skip) {
2984                                 ulist_del(new_roots, qgroup_to_skip, 0);
2985                                 ulist_del(record->old_roots, qgroup_to_skip,
2986                                           0);
2987                         }
2988                         ret = btrfs_qgroup_account_extent(trans, record->bytenr,
2989                                                           record->num_bytes,
2990                                                           record->old_roots,
2991                                                           new_roots);
2992                         record->old_roots = NULL;
2993                         new_roots = NULL;
2994                 }
2995                 /* Free the reserved data space */
2996                 btrfs_qgroup_free_refroot(fs_info,
2997                                 record->data_rsv_refroot,
2998                                 record->data_rsv,
2999                                 BTRFS_QGROUP_RSV_DATA);
3000 cleanup:
3001                 ulist_free(record->old_roots);
3002                 ulist_free(new_roots);
3003                 new_roots = NULL;
3004                 rb_erase(node, &delayed_refs->dirty_extent_root);
3005                 kfree(record);
3006
3007         }
3008         trace_qgroup_num_dirty_extents(fs_info, trans->transid,
3009                                        num_dirty_extents);
3010         return ret;
3011 }
3012
3013 /*
3014  * Writes all changed qgroups to disk.
3015  * Called by the transaction commit path and the qgroup assign ioctl.
3016  */
3017 int btrfs_run_qgroups(struct btrfs_trans_handle *trans)
3018 {
3019         struct btrfs_fs_info *fs_info = trans->fs_info;
3020         int ret = 0;
3021
3022         /*
3023          * In case we are called from the qgroup assign ioctl, assert that we
3024          * are holding the qgroup_ioctl_lock, otherwise we can race with a quota
3025          * disable operation (ioctl) and access a freed quota root.
3026          */
3027         if (trans->transaction->state != TRANS_STATE_COMMIT_DOING)
3028                 lockdep_assert_held(&fs_info->qgroup_ioctl_lock);
3029
3030         if (!fs_info->quota_root)
3031                 return ret;
3032
3033         spin_lock(&fs_info->qgroup_lock);
3034         while (!list_empty(&fs_info->dirty_qgroups)) {
3035                 struct btrfs_qgroup *qgroup;
3036                 qgroup = list_first_entry(&fs_info->dirty_qgroups,
3037                                           struct btrfs_qgroup, dirty);
3038                 list_del_init(&qgroup->dirty);
3039                 spin_unlock(&fs_info->qgroup_lock);
3040                 ret = update_qgroup_info_item(trans, qgroup);
3041                 if (ret)
3042                         qgroup_mark_inconsistent(fs_info);
3043                 ret = update_qgroup_limit_item(trans, qgroup);
3044                 if (ret)
3045                         qgroup_mark_inconsistent(fs_info);
3046                 spin_lock(&fs_info->qgroup_lock);
3047         }
3048         if (btrfs_qgroup_enabled(fs_info))
3049                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_ON;
3050         else
3051                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_ON;
3052         spin_unlock(&fs_info->qgroup_lock);
3053
3054         ret = update_qgroup_status_item(trans);
3055         if (ret)
3056                 qgroup_mark_inconsistent(fs_info);
3057
3058         return ret;
3059 }
3060
3061 int btrfs_qgroup_check_inherit(struct btrfs_fs_info *fs_info,
3062                                struct btrfs_qgroup_inherit *inherit,
3063                                size_t size)
3064 {
3065         if (!btrfs_qgroup_enabled(fs_info))
3066                 return 0;
3067         if (inherit->flags & ~BTRFS_QGROUP_INHERIT_FLAGS_SUPP)
3068                 return -EOPNOTSUPP;
3069         if (size < sizeof(*inherit) || size > PAGE_SIZE)
3070                 return -EINVAL;
3071
3072         /*
3073          * In the past we allowed btrfs_qgroup_inherit to specify to copy
3074          * rfer/excl numbers directly from other qgroups.  This behavior has
3075          * been disabled in userspace for a very long time, but here we should
3076          * also disable it in kernel, as this behavior is known to mark qgroup
3077          * inconsistent, and a rescan would wipe out the changes anyway.
3078          *
3079          * Reject any btrfs_qgroup_inherit with num_ref_copies or num_excl_copies.
3080          */
3081         if (inherit->num_ref_copies > 0 || inherit->num_excl_copies > 0)
3082                 return -EINVAL;
3083
3084         if (size != struct_size(inherit, qgroups, inherit->num_qgroups))
3085                 return -EINVAL;
3086
3087         /*
3088          * Now check all the remaining qgroups, they should all:
3089          *
3090          * - Exist
3091          * - Be higher level qgroups.
3092          */
3093         for (int i = 0; i < inherit->num_qgroups; i++) {
3094                 struct btrfs_qgroup *qgroup;
3095                 u64 qgroupid = inherit->qgroups[i];
3096
3097                 if (btrfs_qgroup_level(qgroupid) == 0)
3098                         return -EINVAL;
3099
3100                 spin_lock(&fs_info->qgroup_lock);
3101                 qgroup = find_qgroup_rb(fs_info, qgroupid);
3102                 if (!qgroup) {
3103                         spin_unlock(&fs_info->qgroup_lock);
3104                         return -ENOENT;
3105                 }
3106                 spin_unlock(&fs_info->qgroup_lock);
3107         }
3108         return 0;
3109 }
3110
3111 static int qgroup_auto_inherit(struct btrfs_fs_info *fs_info,
3112                                u64 inode_rootid,
3113                                struct btrfs_qgroup_inherit **inherit)
3114 {
3115         int i = 0;
3116         u64 num_qgroups = 0;
3117         struct btrfs_qgroup *inode_qg;
3118         struct btrfs_qgroup_list *qg_list;
3119         struct btrfs_qgroup_inherit *res;
3120         size_t struct_sz;
3121         u64 *qgids;
3122
3123         if (*inherit)
3124                 return -EEXIST;
3125
3126         inode_qg = find_qgroup_rb(fs_info, inode_rootid);
3127         if (!inode_qg)
3128                 return -ENOENT;
3129
3130         num_qgroups = list_count_nodes(&inode_qg->groups);
3131
3132         if (!num_qgroups)
3133                 return 0;
3134
3135         struct_sz = struct_size(res, qgroups, num_qgroups);
3136         if (struct_sz == SIZE_MAX)
3137                 return -ERANGE;
3138
3139         res = kzalloc(struct_sz, GFP_NOFS);
3140         if (!res)
3141                 return -ENOMEM;
3142         res->num_qgroups = num_qgroups;
3143         qgids = res->qgroups;
3144
3145         list_for_each_entry(qg_list, &inode_qg->groups, next_group)
3146                 qgids[i++] = qg_list->group->qgroupid;
3147
3148         *inherit = res;
3149         return 0;
3150 }
3151
3152 /*
3153  * Check if we can skip rescan when inheriting qgroups.  If @src has a single
3154  * @parent, and that @parent is owning all its bytes exclusively, we can skip
3155  * the full rescan, by just adding nodesize to the @parent's excl/rfer.
3156  *
3157  * Return <0 for fatal errors (like srcid/parentid has no qgroup).
3158  * Return 0 if a quick inherit is done.
3159  * Return >0 if a quick inherit is not possible, and a full rescan is needed.
3160  */
3161 static int qgroup_snapshot_quick_inherit(struct btrfs_fs_info *fs_info,
3162                                          u64 srcid, u64 parentid)
3163 {
3164         struct btrfs_qgroup *src;
3165         struct btrfs_qgroup *parent;
3166         struct btrfs_qgroup_list *list;
3167         int nr_parents = 0;
3168
3169         src = find_qgroup_rb(fs_info, srcid);
3170         if (!src)
3171                 return -ENOENT;
3172         parent = find_qgroup_rb(fs_info, parentid);
3173         if (!parent)
3174                 return -ENOENT;
3175
3176         /*
3177          * Source has no parent qgroup, but our new qgroup would have one.
3178          * Qgroup numbers would become inconsistent.
3179          */
3180         if (list_empty(&src->groups))
3181                 return 1;
3182
3183         list_for_each_entry(list, &src->groups, next_group) {
3184                 /* The parent is not the same, quick update is not possible. */
3185                 if (list->group->qgroupid != parentid)
3186                         return 1;
3187                 nr_parents++;
3188                 /*
3189                  * More than one parent qgroup, we can't be sure about accounting
3190                  * consistency.
3191                  */
3192                 if (nr_parents > 1)
3193                         return 1;
3194         }
3195
3196         /*
3197          * The parent is not exclusively owning all its bytes.  We're not sure
3198          * if the source has any bytes not fully owned by the parent.
3199          */
3200         if (parent->excl != parent->rfer)
3201                 return 1;
3202
3203         parent->excl += fs_info->nodesize;
3204         parent->rfer += fs_info->nodesize;
3205         return 0;
3206 }
3207
3208 /*
3209  * Copy the accounting information between qgroups. This is necessary
3210  * when a snapshot or a subvolume is created. Throwing an error will
3211  * cause a transaction abort so we take extra care here to only error
3212  * when a readonly fs is a reasonable outcome.
3213  */
3214 int btrfs_qgroup_inherit(struct btrfs_trans_handle *trans, u64 srcid,
3215                          u64 objectid, u64 inode_rootid,
3216                          struct btrfs_qgroup_inherit *inherit)
3217 {
3218         int ret = 0;
3219         int i;
3220         u64 *i_qgroups;
3221         bool committing = false;
3222         struct btrfs_fs_info *fs_info = trans->fs_info;
3223         struct btrfs_root *quota_root;
3224         struct btrfs_qgroup *srcgroup;
3225         struct btrfs_qgroup *dstgroup;
3226         struct btrfs_qgroup *prealloc;
3227         struct btrfs_qgroup_list **qlist_prealloc = NULL;
3228         bool free_inherit = false;
3229         bool need_rescan = false;
3230         u32 level_size = 0;
3231         u64 nums;
3232
3233         prealloc = kzalloc(sizeof(*prealloc), GFP_NOFS);
3234         if (!prealloc)
3235                 return -ENOMEM;
3236
3237         /*
3238          * There are only two callers of this function.
3239          *
3240          * One in create_subvol() in the ioctl context, which needs to hold
3241          * the qgroup_ioctl_lock.
3242          *
3243          * The other one in create_pending_snapshot() where no other qgroup
3244          * code can modify the fs as they all need to either start a new trans
3245          * or hold a trans handler, thus we don't need to hold
3246          * qgroup_ioctl_lock.
3247          * This would avoid long and complex lock chain and make lockdep happy.
3248          */
3249         spin_lock(&fs_info->trans_lock);
3250         if (trans->transaction->state == TRANS_STATE_COMMIT_DOING)
3251                 committing = true;
3252         spin_unlock(&fs_info->trans_lock);
3253
3254         if (!committing)
3255                 mutex_lock(&fs_info->qgroup_ioctl_lock);
3256         if (!btrfs_qgroup_enabled(fs_info))
3257                 goto out;
3258
3259         quota_root = fs_info->quota_root;
3260         if (!quota_root) {
3261                 ret = -EINVAL;
3262                 goto out;
3263         }
3264
3265         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE && !inherit) {
3266                 ret = qgroup_auto_inherit(fs_info, inode_rootid, &inherit);
3267                 if (ret)
3268                         goto out;
3269                 free_inherit = true;
3270         }
3271
3272         if (inherit) {
3273                 i_qgroups = (u64 *)(inherit + 1);
3274                 nums = inherit->num_qgroups + 2 * inherit->num_ref_copies +
3275                        2 * inherit->num_excl_copies;
3276                 for (i = 0; i < nums; ++i) {
3277                         srcgroup = find_qgroup_rb(fs_info, *i_qgroups);
3278
3279                         /*
3280                          * Zero out invalid groups so we can ignore
3281                          * them later.
3282                          */
3283                         if (!srcgroup ||
3284                             ((srcgroup->qgroupid >> 48) <= (objectid >> 48)))
3285                                 *i_qgroups = 0ULL;
3286
3287                         ++i_qgroups;
3288                 }
3289         }
3290
3291         /*
3292          * create a tracking group for the subvol itself
3293          */
3294         ret = add_qgroup_item(trans, quota_root, objectid);
3295         if (ret)
3296                 goto out;
3297
3298         /*
3299          * add qgroup to all inherited groups
3300          */
3301         if (inherit) {
3302                 i_qgroups = (u64 *)(inherit + 1);
3303                 for (i = 0; i < inherit->num_qgroups; ++i, ++i_qgroups) {
3304                         if (*i_qgroups == 0)
3305                                 continue;
3306                         ret = add_qgroup_relation_item(trans, objectid,
3307                                                        *i_qgroups);
3308                         if (ret && ret != -EEXIST)
3309                                 goto out;
3310                         ret = add_qgroup_relation_item(trans, *i_qgroups,
3311                                                        objectid);
3312                         if (ret && ret != -EEXIST)
3313                                 goto out;
3314                 }
3315                 ret = 0;
3316
3317                 qlist_prealloc = kcalloc(inherit->num_qgroups,
3318                                          sizeof(struct btrfs_qgroup_list *),
3319                                          GFP_NOFS);
3320                 if (!qlist_prealloc) {
3321                         ret = -ENOMEM;
3322                         goto out;
3323                 }
3324                 for (int i = 0; i < inherit->num_qgroups; i++) {
3325                         qlist_prealloc[i] = kzalloc(sizeof(struct btrfs_qgroup_list),
3326                                                     GFP_NOFS);
3327                         if (!qlist_prealloc[i]) {
3328                                 ret = -ENOMEM;
3329                                 goto out;
3330                         }
3331                 }
3332         }
3333
3334         spin_lock(&fs_info->qgroup_lock);
3335
3336         dstgroup = add_qgroup_rb(fs_info, prealloc, objectid);
3337         prealloc = NULL;
3338
3339         if (inherit && inherit->flags & BTRFS_QGROUP_INHERIT_SET_LIMITS) {
3340                 dstgroup->lim_flags = inherit->lim.flags;
3341                 dstgroup->max_rfer = inherit->lim.max_rfer;
3342                 dstgroup->max_excl = inherit->lim.max_excl;
3343                 dstgroup->rsv_rfer = inherit->lim.rsv_rfer;
3344                 dstgroup->rsv_excl = inherit->lim.rsv_excl;
3345
3346                 qgroup_dirty(fs_info, dstgroup);
3347         }
3348
3349         if (srcid && btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_FULL) {
3350                 srcgroup = find_qgroup_rb(fs_info, srcid);
3351                 if (!srcgroup)
3352                         goto unlock;
3353
3354                 /*
3355                  * We call inherit after we clone the root in order to make sure
3356                  * our counts don't go crazy, so at this point the only
3357                  * difference between the two roots should be the root node.
3358                  */
3359                 level_size = fs_info->nodesize;
3360                 dstgroup->rfer = srcgroup->rfer;
3361                 dstgroup->rfer_cmpr = srcgroup->rfer_cmpr;
3362                 dstgroup->excl = level_size;
3363                 dstgroup->excl_cmpr = level_size;
3364                 srcgroup->excl = level_size;
3365                 srcgroup->excl_cmpr = level_size;
3366
3367                 /* inherit the limit info */
3368                 dstgroup->lim_flags = srcgroup->lim_flags;
3369                 dstgroup->max_rfer = srcgroup->max_rfer;
3370                 dstgroup->max_excl = srcgroup->max_excl;
3371                 dstgroup->rsv_rfer = srcgroup->rsv_rfer;
3372                 dstgroup->rsv_excl = srcgroup->rsv_excl;
3373
3374                 qgroup_dirty(fs_info, dstgroup);
3375                 qgroup_dirty(fs_info, srcgroup);
3376
3377                 /*
3378                  * If the source qgroup has parent but the new one doesn't,
3379                  * we need a full rescan.
3380                  */
3381                 if (!inherit && !list_empty(&srcgroup->groups))
3382                         need_rescan = true;
3383         }
3384
3385         if (!inherit)
3386                 goto unlock;
3387
3388         i_qgroups = (u64 *)(inherit + 1);
3389         for (i = 0; i < inherit->num_qgroups; ++i) {
3390                 if (*i_qgroups) {
3391                         ret = add_relation_rb(fs_info, qlist_prealloc[i], objectid,
3392                                               *i_qgroups);
3393                         qlist_prealloc[i] = NULL;
3394                         if (ret)
3395                                 goto unlock;
3396                 }
3397                 if (srcid) {
3398                         /* Check if we can do a quick inherit. */
3399                         ret = qgroup_snapshot_quick_inherit(fs_info, srcid, *i_qgroups);
3400                         if (ret < 0)
3401                                 goto unlock;
3402                         if (ret > 0)
3403                                 need_rescan = true;
3404                         ret = 0;
3405                 }
3406                 ++i_qgroups;
3407         }
3408
3409         for (i = 0; i <  inherit->num_ref_copies; ++i, i_qgroups += 2) {
3410                 struct btrfs_qgroup *src;
3411                 struct btrfs_qgroup *dst;
3412
3413                 if (!i_qgroups[0] || !i_qgroups[1])
3414                         continue;
3415
3416                 src = find_qgroup_rb(fs_info, i_qgroups[0]);
3417                 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
3418
3419                 if (!src || !dst) {
3420                         ret = -EINVAL;
3421                         goto unlock;
3422                 }
3423
3424                 dst->rfer = src->rfer - level_size;
3425                 dst->rfer_cmpr = src->rfer_cmpr - level_size;
3426
3427                 /* Manually tweaking numbers certainly needs a rescan */
3428                 need_rescan = true;
3429         }
3430         for (i = 0; i <  inherit->num_excl_copies; ++i, i_qgroups += 2) {
3431                 struct btrfs_qgroup *src;
3432                 struct btrfs_qgroup *dst;
3433
3434                 if (!i_qgroups[0] || !i_qgroups[1])
3435                         continue;
3436
3437                 src = find_qgroup_rb(fs_info, i_qgroups[0]);
3438                 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
3439
3440                 if (!src || !dst) {
3441                         ret = -EINVAL;
3442                         goto unlock;
3443                 }
3444
3445                 dst->excl = src->excl + level_size;
3446                 dst->excl_cmpr = src->excl_cmpr + level_size;
3447                 need_rescan = true;
3448         }
3449
3450 unlock:
3451         spin_unlock(&fs_info->qgroup_lock);
3452         if (!ret)
3453                 ret = btrfs_sysfs_add_one_qgroup(fs_info, dstgroup);
3454 out:
3455         if (!committing)
3456                 mutex_unlock(&fs_info->qgroup_ioctl_lock);
3457         if (need_rescan)
3458                 qgroup_mark_inconsistent(fs_info);
3459         if (qlist_prealloc) {
3460                 for (int i = 0; i < inherit->num_qgroups; i++)
3461                         kfree(qlist_prealloc[i]);
3462                 kfree(qlist_prealloc);
3463         }
3464         if (free_inherit)
3465                 kfree(inherit);
3466         kfree(prealloc);
3467         return ret;
3468 }
3469
3470 static bool qgroup_check_limits(const struct btrfs_qgroup *qg, u64 num_bytes)
3471 {
3472         if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_RFER) &&
3473             qgroup_rsv_total(qg) + (s64)qg->rfer + num_bytes > qg->max_rfer)
3474                 return false;
3475
3476         if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) &&
3477             qgroup_rsv_total(qg) + (s64)qg->excl + num_bytes > qg->max_excl)
3478                 return false;
3479
3480         return true;
3481 }
3482
3483 static int qgroup_reserve(struct btrfs_root *root, u64 num_bytes, bool enforce,
3484                           enum btrfs_qgroup_rsv_type type)
3485 {
3486         struct btrfs_qgroup *qgroup;
3487         struct btrfs_fs_info *fs_info = root->fs_info;
3488         u64 ref_root = btrfs_root_id(root);
3489         int ret = 0;
3490         LIST_HEAD(qgroup_list);
3491
3492         if (!is_fstree(ref_root))
3493                 return 0;
3494
3495         if (num_bytes == 0)
3496                 return 0;
3497
3498         if (test_bit(BTRFS_FS_QUOTA_OVERRIDE, &fs_info->flags) &&
3499             capable(CAP_SYS_RESOURCE))
3500                 enforce = false;
3501
3502         spin_lock(&fs_info->qgroup_lock);
3503         if (!fs_info->quota_root)
3504                 goto out;
3505
3506         qgroup = find_qgroup_rb(fs_info, ref_root);
3507         if (!qgroup)
3508                 goto out;
3509
3510         qgroup_iterator_add(&qgroup_list, qgroup);
3511         list_for_each_entry(qgroup, &qgroup_list, iterator) {
3512                 struct btrfs_qgroup_list *glist;
3513
3514                 if (enforce && !qgroup_check_limits(qgroup, num_bytes)) {
3515                         ret = -EDQUOT;
3516                         goto out;
3517                 }
3518
3519                 list_for_each_entry(glist, &qgroup->groups, next_group)
3520                         qgroup_iterator_add(&qgroup_list, glist->group);
3521         }
3522
3523         ret = 0;
3524         /*
3525          * no limits exceeded, now record the reservation into all qgroups
3526          */
3527         list_for_each_entry(qgroup, &qgroup_list, iterator)
3528                 qgroup_rsv_add(fs_info, qgroup, num_bytes, type);
3529
3530 out:
3531         qgroup_iterator_clean(&qgroup_list);
3532         spin_unlock(&fs_info->qgroup_lock);
3533         return ret;
3534 }
3535
3536 /*
3537  * Free @num_bytes of reserved space with @type for qgroup.  (Normally level 0
3538  * qgroup).
3539  *
3540  * Will handle all higher level qgroup too.
3541  *
3542  * NOTE: If @num_bytes is (u64)-1, this means to free all bytes of this qgroup.
3543  * This special case is only used for META_PERTRANS type.
3544  */
3545 void btrfs_qgroup_free_refroot(struct btrfs_fs_info *fs_info,
3546                                u64 ref_root, u64 num_bytes,
3547                                enum btrfs_qgroup_rsv_type type)
3548 {
3549         struct btrfs_qgroup *qgroup;
3550         LIST_HEAD(qgroup_list);
3551
3552         if (!is_fstree(ref_root))
3553                 return;
3554
3555         if (num_bytes == 0)
3556                 return;
3557
3558         if (num_bytes == (u64)-1 && type != BTRFS_QGROUP_RSV_META_PERTRANS) {
3559                 WARN(1, "%s: Invalid type to free", __func__);
3560                 return;
3561         }
3562         spin_lock(&fs_info->qgroup_lock);
3563
3564         if (!fs_info->quota_root)
3565                 goto out;
3566
3567         qgroup = find_qgroup_rb(fs_info, ref_root);
3568         if (!qgroup)
3569                 goto out;
3570
3571         if (num_bytes == (u64)-1)
3572                 /*
3573                  * We're freeing all pertrans rsv, get reserved value from
3574                  * level 0 qgroup as real num_bytes to free.
3575                  */
3576                 num_bytes = qgroup->rsv.values[type];
3577
3578         qgroup_iterator_add(&qgroup_list, qgroup);
3579         list_for_each_entry(qgroup, &qgroup_list, iterator) {
3580                 struct btrfs_qgroup_list *glist;
3581
3582                 qgroup_rsv_release(fs_info, qgroup, num_bytes, type);
3583                 list_for_each_entry(glist, &qgroup->groups, next_group) {
3584                         qgroup_iterator_add(&qgroup_list, glist->group);
3585                 }
3586         }
3587 out:
3588         qgroup_iterator_clean(&qgroup_list);
3589         spin_unlock(&fs_info->qgroup_lock);
3590 }
3591
3592 /*
3593  * Check if the leaf is the last leaf. Which means all node pointers
3594  * are at their last position.
3595  */
3596 static bool is_last_leaf(struct btrfs_path *path)
3597 {
3598         int i;
3599
3600         for (i = 1; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
3601                 if (path->slots[i] != btrfs_header_nritems(path->nodes[i]) - 1)
3602                         return false;
3603         }
3604         return true;
3605 }
3606
3607 /*
3608  * returns < 0 on error, 0 when more leafs are to be scanned.
3609  * returns 1 when done.
3610  */
3611 static int qgroup_rescan_leaf(struct btrfs_trans_handle *trans,
3612                               struct btrfs_path *path)
3613 {
3614         struct btrfs_fs_info *fs_info = trans->fs_info;
3615         struct btrfs_root *extent_root;
3616         struct btrfs_key found;
3617         struct extent_buffer *scratch_leaf = NULL;
3618         u64 num_bytes;
3619         bool done;
3620         int slot;
3621         int ret;
3622
3623         if (!btrfs_qgroup_full_accounting(fs_info))
3624                 return 1;
3625
3626         mutex_lock(&fs_info->qgroup_rescan_lock);
3627         extent_root = btrfs_extent_root(fs_info,
3628                                 fs_info->qgroup_rescan_progress.objectid);
3629         ret = btrfs_search_slot_for_read(extent_root,
3630                                          &fs_info->qgroup_rescan_progress,
3631                                          path, 1, 0);
3632
3633         btrfs_debug(fs_info,
3634                 "current progress key (%llu %u %llu), search_slot ret %d",
3635                 fs_info->qgroup_rescan_progress.objectid,
3636                 fs_info->qgroup_rescan_progress.type,
3637                 fs_info->qgroup_rescan_progress.offset, ret);
3638
3639         if (ret) {
3640                 /*
3641                  * The rescan is about to end, we will not be scanning any
3642                  * further blocks. We cannot unset the RESCAN flag here, because
3643                  * we want to commit the transaction if everything went well.
3644                  * To make the live accounting work in this phase, we set our
3645                  * scan progress pointer such that every real extent objectid
3646                  * will be smaller.
3647                  */
3648                 fs_info->qgroup_rescan_progress.objectid = (u64)-1;
3649                 btrfs_release_path(path);
3650                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3651                 return ret;
3652         }
3653         done = is_last_leaf(path);
3654
3655         btrfs_item_key_to_cpu(path->nodes[0], &found,
3656                               btrfs_header_nritems(path->nodes[0]) - 1);
3657         fs_info->qgroup_rescan_progress.objectid = found.objectid + 1;
3658
3659         scratch_leaf = btrfs_clone_extent_buffer(path->nodes[0]);
3660         if (!scratch_leaf) {
3661                 ret = -ENOMEM;
3662                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3663                 goto out;
3664         }
3665         slot = path->slots[0];
3666         btrfs_release_path(path);
3667         mutex_unlock(&fs_info->qgroup_rescan_lock);
3668
3669         for (; slot < btrfs_header_nritems(scratch_leaf); ++slot) {
3670                 struct btrfs_backref_walk_ctx ctx = { 0 };
3671
3672                 btrfs_item_key_to_cpu(scratch_leaf, &found, slot);
3673                 if (found.type != BTRFS_EXTENT_ITEM_KEY &&
3674                     found.type != BTRFS_METADATA_ITEM_KEY)
3675                         continue;
3676                 if (found.type == BTRFS_METADATA_ITEM_KEY)
3677                         num_bytes = fs_info->nodesize;
3678                 else
3679                         num_bytes = found.offset;
3680
3681                 ctx.bytenr = found.objectid;
3682                 ctx.fs_info = fs_info;
3683
3684                 ret = btrfs_find_all_roots(&ctx, false);
3685                 if (ret < 0)
3686                         goto out;
3687                 /* For rescan, just pass old_roots as NULL */
3688                 ret = btrfs_qgroup_account_extent(trans, found.objectid,
3689                                                   num_bytes, NULL, ctx.roots);
3690                 if (ret < 0)
3691                         goto out;
3692         }
3693 out:
3694         if (scratch_leaf)
3695                 free_extent_buffer(scratch_leaf);
3696
3697         if (done && !ret) {
3698                 ret = 1;
3699                 fs_info->qgroup_rescan_progress.objectid = (u64)-1;
3700         }
3701         return ret;
3702 }
3703
3704 static bool rescan_should_stop(struct btrfs_fs_info *fs_info)
3705 {
3706         if (btrfs_fs_closing(fs_info))
3707                 return true;
3708         if (test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state))
3709                 return true;
3710         if (!btrfs_qgroup_enabled(fs_info))
3711                 return true;
3712         if (fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN)
3713                 return true;
3714         return false;
3715 }
3716
3717 static void btrfs_qgroup_rescan_worker(struct btrfs_work *work)
3718 {
3719         struct btrfs_fs_info *fs_info = container_of(work, struct btrfs_fs_info,
3720                                                      qgroup_rescan_work);
3721         struct btrfs_path *path;
3722         struct btrfs_trans_handle *trans = NULL;
3723         int ret = 0;
3724         bool stopped = false;
3725         bool did_leaf_rescans = false;
3726
3727         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE)
3728                 return;
3729
3730         path = btrfs_alloc_path();
3731         if (!path) {
3732                 ret = -ENOMEM;
3733                 goto out;
3734         }
3735         /*
3736          * Rescan should only search for commit root, and any later difference
3737          * should be recorded by qgroup
3738          */
3739         path->search_commit_root = 1;
3740         path->skip_locking = 1;
3741
3742         while (!ret && !(stopped = rescan_should_stop(fs_info))) {
3743                 trans = btrfs_start_transaction(fs_info->fs_root, 0);
3744                 if (IS_ERR(trans)) {
3745                         ret = PTR_ERR(trans);
3746                         break;
3747                 }
3748
3749                 ret = qgroup_rescan_leaf(trans, path);
3750                 did_leaf_rescans = true;
3751
3752                 if (ret > 0)
3753                         btrfs_commit_transaction(trans);
3754                 else
3755                         btrfs_end_transaction(trans);
3756         }
3757
3758 out:
3759         btrfs_free_path(path);
3760
3761         mutex_lock(&fs_info->qgroup_rescan_lock);
3762         if (ret > 0 &&
3763             fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT) {
3764                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
3765         } else if (ret < 0 || stopped) {
3766                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
3767         }
3768         mutex_unlock(&fs_info->qgroup_rescan_lock);
3769
3770         /*
3771          * Only update status, since the previous part has already updated the
3772          * qgroup info, and only if we did any actual work. This also prevents
3773          * race with a concurrent quota disable, which has already set
3774          * fs_info->quota_root to NULL and cleared BTRFS_FS_QUOTA_ENABLED at
3775          * btrfs_quota_disable().
3776          */
3777         if (did_leaf_rescans) {
3778                 trans = btrfs_start_transaction(fs_info->quota_root, 1);
3779                 if (IS_ERR(trans)) {
3780                         ret = PTR_ERR(trans);
3781                         trans = NULL;
3782                         btrfs_err(fs_info,
3783                                   "fail to start transaction for status update: %d",
3784                                   ret);
3785                 }
3786         } else {
3787                 trans = NULL;
3788         }
3789
3790         mutex_lock(&fs_info->qgroup_rescan_lock);
3791         if (!stopped ||
3792             fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN)
3793                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3794         if (trans) {
3795                 int ret2 = update_qgroup_status_item(trans);
3796
3797                 if (ret2 < 0) {
3798                         ret = ret2;
3799                         btrfs_err(fs_info, "fail to update qgroup status: %d", ret);
3800                 }
3801         }
3802         fs_info->qgroup_rescan_running = false;
3803         fs_info->qgroup_flags &= ~BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN;
3804         complete_all(&fs_info->qgroup_rescan_completion);
3805         mutex_unlock(&fs_info->qgroup_rescan_lock);
3806
3807         if (!trans)
3808                 return;
3809
3810         btrfs_end_transaction(trans);
3811
3812         if (stopped) {
3813                 btrfs_info(fs_info, "qgroup scan paused");
3814         } else if (fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN) {
3815                 btrfs_info(fs_info, "qgroup scan cancelled");
3816         } else if (ret >= 0) {
3817                 btrfs_info(fs_info, "qgroup scan completed%s",
3818                         ret > 0 ? " (inconsistency flag cleared)" : "");
3819         } else {
3820                 btrfs_err(fs_info, "qgroup scan failed with %d", ret);
3821         }
3822 }
3823
3824 /*
3825  * Checks that (a) no rescan is running and (b) quota is enabled. Allocates all
3826  * memory required for the rescan context.
3827  */
3828 static int
3829 qgroup_rescan_init(struct btrfs_fs_info *fs_info, u64 progress_objectid,
3830                    int init_flags)
3831 {
3832         int ret = 0;
3833
3834         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_SIMPLE) {
3835                 btrfs_warn(fs_info, "qgroup rescan init failed, running in simple mode");
3836                 return -EINVAL;
3837         }
3838
3839         if (!init_flags) {
3840                 /* we're resuming qgroup rescan at mount time */
3841                 if (!(fs_info->qgroup_flags &
3842                       BTRFS_QGROUP_STATUS_FLAG_RESCAN)) {
3843                         btrfs_debug(fs_info,
3844                         "qgroup rescan init failed, qgroup rescan is not queued");
3845                         ret = -EINVAL;
3846                 } else if (!(fs_info->qgroup_flags &
3847                              BTRFS_QGROUP_STATUS_FLAG_ON)) {
3848                         btrfs_debug(fs_info,
3849                         "qgroup rescan init failed, qgroup is not enabled");
3850                         ret = -ENOTCONN;
3851                 }
3852
3853                 if (ret)
3854                         return ret;
3855         }
3856
3857         mutex_lock(&fs_info->qgroup_rescan_lock);
3858
3859         if (init_flags) {
3860                 if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
3861                         ret = -EINPROGRESS;
3862                 } else if (!(fs_info->qgroup_flags &
3863                              BTRFS_QGROUP_STATUS_FLAG_ON)) {
3864                         btrfs_debug(fs_info,
3865                         "qgroup rescan init failed, qgroup is not enabled");
3866                         ret = -ENOTCONN;
3867                 } else if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_DISABLED) {
3868                         /* Quota disable is in progress */
3869                         ret = -EBUSY;
3870                 }
3871
3872                 if (ret) {
3873                         mutex_unlock(&fs_info->qgroup_rescan_lock);
3874                         return ret;
3875                 }
3876                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3877         }
3878
3879         memset(&fs_info->qgroup_rescan_progress, 0,
3880                 sizeof(fs_info->qgroup_rescan_progress));
3881         fs_info->qgroup_flags &= ~(BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN |
3882                                    BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING);
3883         fs_info->qgroup_rescan_progress.objectid = progress_objectid;
3884         init_completion(&fs_info->qgroup_rescan_completion);
3885         mutex_unlock(&fs_info->qgroup_rescan_lock);
3886
3887         btrfs_init_work(&fs_info->qgroup_rescan_work,
3888                         btrfs_qgroup_rescan_worker, NULL);
3889         return 0;
3890 }
3891
3892 static void
3893 qgroup_rescan_zero_tracking(struct btrfs_fs_info *fs_info)
3894 {
3895         struct rb_node *n;
3896         struct btrfs_qgroup *qgroup;
3897
3898         spin_lock(&fs_info->qgroup_lock);
3899         /* clear all current qgroup tracking information */
3900         for (n = rb_first(&fs_info->qgroup_tree); n; n = rb_next(n)) {
3901                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
3902                 qgroup->rfer = 0;
3903                 qgroup->rfer_cmpr = 0;
3904                 qgroup->excl = 0;
3905                 qgroup->excl_cmpr = 0;
3906                 qgroup_dirty(fs_info, qgroup);
3907         }
3908         spin_unlock(&fs_info->qgroup_lock);
3909 }
3910
3911 int
3912 btrfs_qgroup_rescan(struct btrfs_fs_info *fs_info)
3913 {
3914         int ret = 0;
3915         struct btrfs_trans_handle *trans;
3916
3917         ret = qgroup_rescan_init(fs_info, 0, 1);
3918         if (ret)
3919                 return ret;
3920
3921         /*
3922          * We have set the rescan_progress to 0, which means no more
3923          * delayed refs will be accounted by btrfs_qgroup_account_ref.
3924          * However, btrfs_qgroup_account_ref may be right after its call
3925          * to btrfs_find_all_roots, in which case it would still do the
3926          * accounting.
3927          * To solve this, we're committing the transaction, which will
3928          * ensure we run all delayed refs and only after that, we are
3929          * going to clear all tracking information for a clean start.
3930          */
3931
3932         trans = btrfs_attach_transaction_barrier(fs_info->fs_root);
3933         if (IS_ERR(trans) && trans != ERR_PTR(-ENOENT)) {
3934                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3935                 return PTR_ERR(trans);
3936         } else if (trans != ERR_PTR(-ENOENT)) {
3937                 ret = btrfs_commit_transaction(trans);
3938                 if (ret) {
3939                         fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3940                         return ret;
3941                 }
3942         }
3943
3944         qgroup_rescan_zero_tracking(fs_info);
3945
3946         mutex_lock(&fs_info->qgroup_rescan_lock);
3947         fs_info->qgroup_rescan_running = true;
3948         btrfs_queue_work(fs_info->qgroup_rescan_workers,
3949                          &fs_info->qgroup_rescan_work);
3950         mutex_unlock(&fs_info->qgroup_rescan_lock);
3951
3952         return 0;
3953 }
3954
3955 int btrfs_qgroup_wait_for_completion(struct btrfs_fs_info *fs_info,
3956                                      bool interruptible)
3957 {
3958         int running;
3959         int ret = 0;
3960
3961         mutex_lock(&fs_info->qgroup_rescan_lock);
3962         running = fs_info->qgroup_rescan_running;
3963         mutex_unlock(&fs_info->qgroup_rescan_lock);
3964
3965         if (!running)
3966                 return 0;
3967
3968         if (interruptible)
3969                 ret = wait_for_completion_interruptible(
3970                                         &fs_info->qgroup_rescan_completion);
3971         else
3972                 wait_for_completion(&fs_info->qgroup_rescan_completion);
3973
3974         return ret;
3975 }
3976
3977 /*
3978  * this is only called from open_ctree where we're still single threaded, thus
3979  * locking is omitted here.
3980  */
3981 void
3982 btrfs_qgroup_rescan_resume(struct btrfs_fs_info *fs_info)
3983 {
3984         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
3985                 mutex_lock(&fs_info->qgroup_rescan_lock);
3986                 fs_info->qgroup_rescan_running = true;
3987                 btrfs_queue_work(fs_info->qgroup_rescan_workers,
3988                                  &fs_info->qgroup_rescan_work);
3989                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3990         }
3991 }
3992
3993 #define rbtree_iterate_from_safe(node, next, start)                             \
3994        for (node = start; node && ({ next = rb_next(node); 1;}); node = next)
3995
3996 static int qgroup_unreserve_range(struct btrfs_inode *inode,
3997                                   struct extent_changeset *reserved, u64 start,
3998                                   u64 len)
3999 {
4000         struct rb_node *node;
4001         struct rb_node *next;
4002         struct ulist_node *entry;
4003         int ret = 0;
4004
4005         node = reserved->range_changed.root.rb_node;
4006         if (!node)
4007                 return 0;
4008         while (node) {
4009                 entry = rb_entry(node, struct ulist_node, rb_node);
4010                 if (entry->val < start)
4011                         node = node->rb_right;
4012                 else
4013                         node = node->rb_left;
4014         }
4015
4016         if (entry->val > start && rb_prev(&entry->rb_node))
4017                 entry = rb_entry(rb_prev(&entry->rb_node), struct ulist_node,
4018                                  rb_node);
4019
4020         rbtree_iterate_from_safe(node, next, &entry->rb_node) {
4021                 u64 entry_start;
4022                 u64 entry_end;
4023                 u64 entry_len;
4024                 int clear_ret;
4025
4026                 entry = rb_entry(node, struct ulist_node, rb_node);
4027                 entry_start = entry->val;
4028                 entry_end = entry->aux;
4029                 entry_len = entry_end - entry_start + 1;
4030
4031                 if (entry_start >= start + len)
4032                         break;
4033                 if (entry_start + entry_len <= start)
4034                         continue;
4035                 /*
4036                  * Now the entry is in [start, start + len), revert the
4037                  * EXTENT_QGROUP_RESERVED bit.
4038                  */
4039                 clear_ret = clear_extent_bits(&inode->io_tree, entry_start,
4040                                               entry_end, EXTENT_QGROUP_RESERVED);
4041                 if (!ret && clear_ret < 0)
4042                         ret = clear_ret;
4043
4044                 ulist_del(&reserved->range_changed, entry->val, entry->aux);
4045                 if (likely(reserved->bytes_changed >= entry_len)) {
4046                         reserved->bytes_changed -= entry_len;
4047                 } else {
4048                         WARN_ON(1);
4049                         reserved->bytes_changed = 0;
4050                 }
4051         }
4052
4053         return ret;
4054 }
4055
4056 /*
4057  * Try to free some space for qgroup.
4058  *
4059  * For qgroup, there are only 3 ways to free qgroup space:
4060  * - Flush nodatacow write
4061  *   Any nodatacow write will free its reserved data space at run_delalloc_range().
4062  *   In theory, we should only flush nodatacow inodes, but it's not yet
4063  *   possible, so we need to flush the whole root.
4064  *
4065  * - Wait for ordered extents
4066  *   When ordered extents are finished, their reserved metadata is finally
4067  *   converted to per_trans status, which can be freed by later commit
4068  *   transaction.
4069  *
4070  * - Commit transaction
4071  *   This would free the meta_per_trans space.
4072  *   In theory this shouldn't provide much space, but any more qgroup space
4073  *   is needed.
4074  */
4075 static int try_flush_qgroup(struct btrfs_root *root)
4076 {
4077         struct btrfs_trans_handle *trans;
4078         int ret;
4079
4080         /* Can't hold an open transaction or we run the risk of deadlocking. */
4081         ASSERT(current->journal_info == NULL);
4082         if (WARN_ON(current->journal_info))
4083                 return 0;
4084
4085         /*
4086          * We don't want to run flush again and again, so if there is a running
4087          * one, we won't try to start a new flush, but exit directly.
4088          */
4089         if (test_and_set_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state)) {
4090                 wait_event(root->qgroup_flush_wait,
4091                         !test_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state));
4092                 return 0;
4093         }
4094
4095         ret = btrfs_start_delalloc_snapshot(root, true);
4096         if (ret < 0)
4097                 goto out;
4098         btrfs_wait_ordered_extents(root, U64_MAX, 0, (u64)-1);
4099
4100         trans = btrfs_attach_transaction_barrier(root);
4101         if (IS_ERR(trans)) {
4102                 ret = PTR_ERR(trans);
4103                 if (ret == -ENOENT)
4104                         ret = 0;
4105                 goto out;
4106         }
4107
4108         ret = btrfs_commit_transaction(trans);
4109 out:
4110         clear_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state);
4111         wake_up(&root->qgroup_flush_wait);
4112         return ret;
4113 }
4114
4115 static int qgroup_reserve_data(struct btrfs_inode *inode,
4116                         struct extent_changeset **reserved_ret, u64 start,
4117                         u64 len)
4118 {
4119         struct btrfs_root *root = inode->root;
4120         struct extent_changeset *reserved;
4121         bool new_reserved = false;
4122         u64 orig_reserved;
4123         u64 to_reserve;
4124         int ret;
4125
4126         if (btrfs_qgroup_mode(root->fs_info) == BTRFS_QGROUP_MODE_DISABLED ||
4127             !is_fstree(btrfs_root_id(root)) || len == 0)
4128                 return 0;
4129
4130         /* @reserved parameter is mandatory for qgroup */
4131         if (WARN_ON(!reserved_ret))
4132                 return -EINVAL;
4133         if (!*reserved_ret) {
4134                 new_reserved = true;
4135                 *reserved_ret = extent_changeset_alloc();
4136                 if (!*reserved_ret)
4137                         return -ENOMEM;
4138         }
4139         reserved = *reserved_ret;
4140         /* Record already reserved space */
4141         orig_reserved = reserved->bytes_changed;
4142         ret = set_record_extent_bits(&inode->io_tree, start,
4143                         start + len -1, EXTENT_QGROUP_RESERVED, reserved);
4144
4145         /* Newly reserved space */
4146         to_reserve = reserved->bytes_changed - orig_reserved;
4147         trace_btrfs_qgroup_reserve_data(&inode->vfs_inode, start, len,
4148                                         to_reserve, QGROUP_RESERVE);
4149         if (ret < 0)
4150                 goto out;
4151         ret = qgroup_reserve(root, to_reserve, true, BTRFS_QGROUP_RSV_DATA);
4152         if (ret < 0)
4153                 goto cleanup;
4154
4155         return ret;
4156
4157 cleanup:
4158         qgroup_unreserve_range(inode, reserved, start, len);
4159 out:
4160         if (new_reserved) {
4161                 extent_changeset_free(reserved);
4162                 *reserved_ret = NULL;
4163         }
4164         return ret;
4165 }
4166
4167 /*
4168  * Reserve qgroup space for range [start, start + len).
4169  *
4170  * This function will either reserve space from related qgroups or do nothing
4171  * if the range is already reserved.
4172  *
4173  * Return 0 for successful reservation
4174  * Return <0 for error (including -EQUOT)
4175  *
4176  * NOTE: This function may sleep for memory allocation, dirty page flushing and
4177  *       commit transaction. So caller should not hold any dirty page locked.
4178  */
4179 int btrfs_qgroup_reserve_data(struct btrfs_inode *inode,
4180                         struct extent_changeset **reserved_ret, u64 start,
4181                         u64 len)
4182 {
4183         int ret;
4184
4185         ret = qgroup_reserve_data(inode, reserved_ret, start, len);
4186         if (ret <= 0 && ret != -EDQUOT)
4187                 return ret;
4188
4189         ret = try_flush_qgroup(inode->root);
4190         if (ret < 0)
4191                 return ret;
4192         return qgroup_reserve_data(inode, reserved_ret, start, len);
4193 }
4194
4195 /* Free ranges specified by @reserved, normally in error path */
4196 static int qgroup_free_reserved_data(struct btrfs_inode *inode,
4197                                      struct extent_changeset *reserved,
4198                                      u64 start, u64 len, u64 *freed_ret)
4199 {
4200         struct btrfs_root *root = inode->root;
4201         struct ulist_node *unode;
4202         struct ulist_iterator uiter;
4203         struct extent_changeset changeset;
4204         u64 freed = 0;
4205         int ret;
4206
4207         extent_changeset_init(&changeset);
4208         len = round_up(start + len, root->fs_info->sectorsize);
4209         start = round_down(start, root->fs_info->sectorsize);
4210
4211         ULIST_ITER_INIT(&uiter);
4212         while ((unode = ulist_next(&reserved->range_changed, &uiter))) {
4213                 u64 range_start = unode->val;
4214                 /* unode->aux is the inclusive end */
4215                 u64 range_len = unode->aux - range_start + 1;
4216                 u64 free_start;
4217                 u64 free_len;
4218
4219                 extent_changeset_release(&changeset);
4220
4221                 /* Only free range in range [start, start + len) */
4222                 if (range_start >= start + len ||
4223                     range_start + range_len <= start)
4224                         continue;
4225                 free_start = max(range_start, start);
4226                 free_len = min(start + len, range_start + range_len) -
4227                            free_start;
4228                 /*
4229                  * TODO: To also modify reserved->ranges_reserved to reflect
4230                  * the modification.
4231                  *
4232                  * However as long as we free qgroup reserved according to
4233                  * EXTENT_QGROUP_RESERVED, we won't double free.
4234                  * So not need to rush.
4235                  */
4236                 ret = clear_record_extent_bits(&inode->io_tree, free_start,
4237                                 free_start + free_len - 1,
4238                                 EXTENT_QGROUP_RESERVED, &changeset);
4239                 if (ret < 0)
4240                         goto out;
4241                 freed += changeset.bytes_changed;
4242         }
4243         btrfs_qgroup_free_refroot(root->fs_info, btrfs_root_id(root), freed,
4244                                   BTRFS_QGROUP_RSV_DATA);
4245         if (freed_ret)
4246                 *freed_ret = freed;
4247         ret = 0;
4248 out:
4249         extent_changeset_release(&changeset);
4250         return ret;
4251 }
4252
4253 static int __btrfs_qgroup_release_data(struct btrfs_inode *inode,
4254                         struct extent_changeset *reserved, u64 start, u64 len,
4255                         u64 *released, int free)
4256 {
4257         struct extent_changeset changeset;
4258         int trace_op = QGROUP_RELEASE;
4259         int ret;
4260
4261         if (btrfs_qgroup_mode(inode->root->fs_info) == BTRFS_QGROUP_MODE_DISABLED) {
4262                 extent_changeset_init(&changeset);
4263                 return clear_record_extent_bits(&inode->io_tree, start,
4264                                                 start + len - 1,
4265                                                 EXTENT_QGROUP_RESERVED, &changeset);
4266         }
4267
4268         /* In release case, we shouldn't have @reserved */
4269         WARN_ON(!free && reserved);
4270         if (free && reserved)
4271                 return qgroup_free_reserved_data(inode, reserved, start, len, released);
4272         extent_changeset_init(&changeset);
4273         ret = clear_record_extent_bits(&inode->io_tree, start, start + len -1,
4274                                        EXTENT_QGROUP_RESERVED, &changeset);
4275         if (ret < 0)
4276                 goto out;
4277
4278         if (free)
4279                 trace_op = QGROUP_FREE;
4280         trace_btrfs_qgroup_release_data(&inode->vfs_inode, start, len,
4281                                         changeset.bytes_changed, trace_op);
4282         if (free)
4283                 btrfs_qgroup_free_refroot(inode->root->fs_info,
4284                                 btrfs_root_id(inode->root),
4285                                 changeset.bytes_changed, BTRFS_QGROUP_RSV_DATA);
4286         if (released)
4287                 *released = changeset.bytes_changed;
4288 out:
4289         extent_changeset_release(&changeset);
4290         return ret;
4291 }
4292
4293 /*
4294  * Free a reserved space range from io_tree and related qgroups
4295  *
4296  * Should be called when a range of pages get invalidated before reaching disk.
4297  * Or for error cleanup case.
4298  * if @reserved is given, only reserved range in [@start, @start + @len) will
4299  * be freed.
4300  *
4301  * For data written to disk, use btrfs_qgroup_release_data().
4302  *
4303  * NOTE: This function may sleep for memory allocation.
4304  */
4305 int btrfs_qgroup_free_data(struct btrfs_inode *inode,
4306                            struct extent_changeset *reserved,
4307                            u64 start, u64 len, u64 *freed)
4308 {
4309         return __btrfs_qgroup_release_data(inode, reserved, start, len, freed, 1);
4310 }
4311
4312 /*
4313  * Release a reserved space range from io_tree only.
4314  *
4315  * Should be called when a range of pages get written to disk and corresponding
4316  * FILE_EXTENT is inserted into corresponding root.
4317  *
4318  * Since new qgroup accounting framework will only update qgroup numbers at
4319  * commit_transaction() time, its reserved space shouldn't be freed from
4320  * related qgroups.
4321  *
4322  * But we should release the range from io_tree, to allow further write to be
4323  * COWed.
4324  *
4325  * NOTE: This function may sleep for memory allocation.
4326  */
4327 int btrfs_qgroup_release_data(struct btrfs_inode *inode, u64 start, u64 len, u64 *released)
4328 {
4329         return __btrfs_qgroup_release_data(inode, NULL, start, len, released, 0);
4330 }
4331
4332 static void add_root_meta_rsv(struct btrfs_root *root, int num_bytes,
4333                               enum btrfs_qgroup_rsv_type type)
4334 {
4335         if (type != BTRFS_QGROUP_RSV_META_PREALLOC &&
4336             type != BTRFS_QGROUP_RSV_META_PERTRANS)
4337                 return;
4338         if (num_bytes == 0)
4339                 return;
4340
4341         spin_lock(&root->qgroup_meta_rsv_lock);
4342         if (type == BTRFS_QGROUP_RSV_META_PREALLOC)
4343                 root->qgroup_meta_rsv_prealloc += num_bytes;
4344         else
4345                 root->qgroup_meta_rsv_pertrans += num_bytes;
4346         spin_unlock(&root->qgroup_meta_rsv_lock);
4347 }
4348
4349 static int sub_root_meta_rsv(struct btrfs_root *root, int num_bytes,
4350                              enum btrfs_qgroup_rsv_type type)
4351 {
4352         if (type != BTRFS_QGROUP_RSV_META_PREALLOC &&
4353             type != BTRFS_QGROUP_RSV_META_PERTRANS)
4354                 return 0;
4355         if (num_bytes == 0)
4356                 return 0;
4357
4358         spin_lock(&root->qgroup_meta_rsv_lock);
4359         if (type == BTRFS_QGROUP_RSV_META_PREALLOC) {
4360                 num_bytes = min_t(u64, root->qgroup_meta_rsv_prealloc,
4361                                   num_bytes);
4362                 root->qgroup_meta_rsv_prealloc -= num_bytes;
4363         } else {
4364                 num_bytes = min_t(u64, root->qgroup_meta_rsv_pertrans,
4365                                   num_bytes);
4366                 root->qgroup_meta_rsv_pertrans -= num_bytes;
4367         }
4368         spin_unlock(&root->qgroup_meta_rsv_lock);
4369         return num_bytes;
4370 }
4371
4372 int btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
4373                               enum btrfs_qgroup_rsv_type type, bool enforce)
4374 {
4375         struct btrfs_fs_info *fs_info = root->fs_info;
4376         int ret;
4377
4378         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_DISABLED ||
4379             !is_fstree(btrfs_root_id(root)) || num_bytes == 0)
4380                 return 0;
4381
4382         BUG_ON(num_bytes != round_down(num_bytes, fs_info->nodesize));
4383         trace_qgroup_meta_reserve(root, (s64)num_bytes, type);
4384         ret = qgroup_reserve(root, num_bytes, enforce, type);
4385         if (ret < 0)
4386                 return ret;
4387         /*
4388          * Record what we have reserved into root.
4389          *
4390          * To avoid quota disabled->enabled underflow.
4391          * In that case, we may try to free space we haven't reserved
4392          * (since quota was disabled), so record what we reserved into root.
4393          * And ensure later release won't underflow this number.
4394          */
4395         add_root_meta_rsv(root, num_bytes, type);
4396         return ret;
4397 }
4398
4399 int __btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
4400                                 enum btrfs_qgroup_rsv_type type, bool enforce,
4401                                 bool noflush)
4402 {
4403         int ret;
4404
4405         ret = btrfs_qgroup_reserve_meta(root, num_bytes, type, enforce);
4406         if ((ret <= 0 && ret != -EDQUOT) || noflush)
4407                 return ret;
4408
4409         ret = try_flush_qgroup(root);
4410         if (ret < 0)
4411                 return ret;
4412         return btrfs_qgroup_reserve_meta(root, num_bytes, type, enforce);
4413 }
4414
4415 /*
4416  * Per-transaction meta reservation should be all freed at transaction commit
4417  * time
4418  */
4419 void btrfs_qgroup_free_meta_all_pertrans(struct btrfs_root *root)
4420 {
4421         struct btrfs_fs_info *fs_info = root->fs_info;
4422
4423         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_DISABLED ||
4424             !is_fstree(btrfs_root_id(root)))
4425                 return;
4426
4427         /* TODO: Update trace point to handle such free */
4428         trace_qgroup_meta_free_all_pertrans(root);
4429         /* Special value -1 means to free all reserved space */
4430         btrfs_qgroup_free_refroot(fs_info, btrfs_root_id(root), (u64)-1,
4431                                   BTRFS_QGROUP_RSV_META_PERTRANS);
4432 }
4433
4434 void __btrfs_qgroup_free_meta(struct btrfs_root *root, int num_bytes,
4435                               enum btrfs_qgroup_rsv_type type)
4436 {
4437         struct btrfs_fs_info *fs_info = root->fs_info;
4438
4439         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_DISABLED ||
4440             !is_fstree(btrfs_root_id(root)))
4441                 return;
4442
4443         /*
4444          * reservation for META_PREALLOC can happen before quota is enabled,
4445          * which can lead to underflow.
4446          * Here ensure we will only free what we really have reserved.
4447          */
4448         num_bytes = sub_root_meta_rsv(root, num_bytes, type);
4449         BUG_ON(num_bytes != round_down(num_bytes, fs_info->nodesize));
4450         trace_qgroup_meta_reserve(root, -(s64)num_bytes, type);
4451         btrfs_qgroup_free_refroot(fs_info, btrfs_root_id(root), num_bytes, type);
4452 }
4453
4454 static void qgroup_convert_meta(struct btrfs_fs_info *fs_info, u64 ref_root,
4455                                 int num_bytes)
4456 {
4457         struct btrfs_qgroup *qgroup;
4458         LIST_HEAD(qgroup_list);
4459
4460         if (num_bytes == 0)
4461                 return;
4462         if (!fs_info->quota_root)
4463                 return;
4464
4465         spin_lock(&fs_info->qgroup_lock);
4466         qgroup = find_qgroup_rb(fs_info, ref_root);
4467         if (!qgroup)
4468                 goto out;
4469
4470         qgroup_iterator_add(&qgroup_list, qgroup);
4471         list_for_each_entry(qgroup, &qgroup_list, iterator) {
4472                 struct btrfs_qgroup_list *glist;
4473
4474                 qgroup_rsv_release(fs_info, qgroup, num_bytes,
4475                                 BTRFS_QGROUP_RSV_META_PREALLOC);
4476                 if (!sb_rdonly(fs_info->sb))
4477                         qgroup_rsv_add(fs_info, qgroup, num_bytes,
4478                                        BTRFS_QGROUP_RSV_META_PERTRANS);
4479
4480                 list_for_each_entry(glist, &qgroup->groups, next_group)
4481                         qgroup_iterator_add(&qgroup_list, glist->group);
4482         }
4483 out:
4484         qgroup_iterator_clean(&qgroup_list);
4485         spin_unlock(&fs_info->qgroup_lock);
4486 }
4487
4488 /*
4489  * Convert @num_bytes of META_PREALLOCATED reservation to META_PERTRANS.
4490  *
4491  * This is called when preallocated meta reservation needs to be used.
4492  * Normally after btrfs_join_transaction() call.
4493  */
4494 void btrfs_qgroup_convert_reserved_meta(struct btrfs_root *root, int num_bytes)
4495 {
4496         struct btrfs_fs_info *fs_info = root->fs_info;
4497
4498         if (btrfs_qgroup_mode(fs_info) == BTRFS_QGROUP_MODE_DISABLED ||
4499             !is_fstree(btrfs_root_id(root)))
4500                 return;
4501         /* Same as btrfs_qgroup_free_meta_prealloc() */
4502         num_bytes = sub_root_meta_rsv(root, num_bytes,
4503                                       BTRFS_QGROUP_RSV_META_PREALLOC);
4504         trace_qgroup_meta_convert(root, num_bytes);
4505         qgroup_convert_meta(fs_info, btrfs_root_id(root), num_bytes);
4506         if (!sb_rdonly(fs_info->sb))
4507                 add_root_meta_rsv(root, num_bytes, BTRFS_QGROUP_RSV_META_PERTRANS);
4508 }
4509
4510 /*
4511  * Check qgroup reserved space leaking, normally at destroy inode
4512  * time
4513  */
4514 void btrfs_qgroup_check_reserved_leak(struct btrfs_inode *inode)
4515 {
4516         struct extent_changeset changeset;
4517         struct ulist_node *unode;
4518         struct ulist_iterator iter;
4519         int ret;
4520
4521         extent_changeset_init(&changeset);
4522         ret = clear_record_extent_bits(&inode->io_tree, 0, (u64)-1,
4523                         EXTENT_QGROUP_RESERVED, &changeset);
4524
4525         WARN_ON(ret < 0);
4526         if (WARN_ON(changeset.bytes_changed)) {
4527                 ULIST_ITER_INIT(&iter);
4528                 while ((unode = ulist_next(&changeset.range_changed, &iter))) {
4529                         btrfs_warn(inode->root->fs_info,
4530                 "leaking qgroup reserved space, ino: %llu, start: %llu, end: %llu",
4531                                 btrfs_ino(inode), unode->val, unode->aux);
4532                 }
4533                 btrfs_qgroup_free_refroot(inode->root->fs_info,
4534                                 btrfs_root_id(inode->root),
4535                                 changeset.bytes_changed, BTRFS_QGROUP_RSV_DATA);
4536
4537         }
4538         extent_changeset_release(&changeset);
4539 }
4540
4541 void btrfs_qgroup_init_swapped_blocks(
4542         struct btrfs_qgroup_swapped_blocks *swapped_blocks)
4543 {
4544         int i;
4545
4546         spin_lock_init(&swapped_blocks->lock);
4547         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
4548                 swapped_blocks->blocks[i] = RB_ROOT;
4549         swapped_blocks->swapped = false;
4550 }
4551
4552 /*
4553  * Delete all swapped blocks record of @root.
4554  * Every record here means we skipped a full subtree scan for qgroup.
4555  *
4556  * Gets called when committing one transaction.
4557  */
4558 void btrfs_qgroup_clean_swapped_blocks(struct btrfs_root *root)
4559 {
4560         struct btrfs_qgroup_swapped_blocks *swapped_blocks;
4561         int i;
4562
4563         swapped_blocks = &root->swapped_blocks;
4564
4565         spin_lock(&swapped_blocks->lock);
4566         if (!swapped_blocks->swapped)
4567                 goto out;
4568         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
4569                 struct rb_root *cur_root = &swapped_blocks->blocks[i];
4570                 struct btrfs_qgroup_swapped_block *entry;
4571                 struct btrfs_qgroup_swapped_block *next;
4572
4573                 rbtree_postorder_for_each_entry_safe(entry, next, cur_root,
4574                                                      node)
4575                         kfree(entry);
4576                 swapped_blocks->blocks[i] = RB_ROOT;
4577         }
4578         swapped_blocks->swapped = false;
4579 out:
4580         spin_unlock(&swapped_blocks->lock);
4581 }
4582
4583 /*
4584  * Add subtree roots record into @subvol_root.
4585  *
4586  * @subvol_root:        tree root of the subvolume tree get swapped
4587  * @bg:                 block group under balance
4588  * @subvol_parent/slot: pointer to the subtree root in subvolume tree
4589  * @reloc_parent/slot:  pointer to the subtree root in reloc tree
4590  *                      BOTH POINTERS ARE BEFORE TREE SWAP
4591  * @last_snapshot:      last snapshot generation of the subvolume tree
4592  */
4593 int btrfs_qgroup_add_swapped_blocks(struct btrfs_trans_handle *trans,
4594                 struct btrfs_root *subvol_root,
4595                 struct btrfs_block_group *bg,
4596                 struct extent_buffer *subvol_parent, int subvol_slot,
4597                 struct extent_buffer *reloc_parent, int reloc_slot,
4598                 u64 last_snapshot)
4599 {
4600         struct btrfs_fs_info *fs_info = subvol_root->fs_info;
4601         struct btrfs_qgroup_swapped_blocks *blocks = &subvol_root->swapped_blocks;
4602         struct btrfs_qgroup_swapped_block *block;
4603         struct rb_node **cur;
4604         struct rb_node *parent = NULL;
4605         int level = btrfs_header_level(subvol_parent) - 1;
4606         int ret = 0;
4607
4608         if (!btrfs_qgroup_full_accounting(fs_info))
4609                 return 0;
4610
4611         if (btrfs_node_ptr_generation(subvol_parent, subvol_slot) >
4612             btrfs_node_ptr_generation(reloc_parent, reloc_slot)) {
4613                 btrfs_err_rl(fs_info,
4614                 "%s: bad parameter order, subvol_gen=%llu reloc_gen=%llu",
4615                         __func__,
4616                         btrfs_node_ptr_generation(subvol_parent, subvol_slot),
4617                         btrfs_node_ptr_generation(reloc_parent, reloc_slot));
4618                 return -EUCLEAN;
4619         }
4620
4621         block = kmalloc(sizeof(*block), GFP_NOFS);
4622         if (!block) {
4623                 ret = -ENOMEM;
4624                 goto out;
4625         }
4626
4627         /*
4628          * @reloc_parent/slot is still before swap, while @block is going to
4629          * record the bytenr after swap, so we do the swap here.
4630          */
4631         block->subvol_bytenr = btrfs_node_blockptr(reloc_parent, reloc_slot);
4632         block->subvol_generation = btrfs_node_ptr_generation(reloc_parent,
4633                                                              reloc_slot);
4634         block->reloc_bytenr = btrfs_node_blockptr(subvol_parent, subvol_slot);
4635         block->reloc_generation = btrfs_node_ptr_generation(subvol_parent,
4636                                                             subvol_slot);
4637         block->last_snapshot = last_snapshot;
4638         block->level = level;
4639
4640         /*
4641          * If we have bg == NULL, we're called from btrfs_recover_relocation(),
4642          * no one else can modify tree blocks thus we qgroup will not change
4643          * no matter the value of trace_leaf.
4644          */
4645         if (bg && bg->flags & BTRFS_BLOCK_GROUP_DATA)
4646                 block->trace_leaf = true;
4647         else
4648                 block->trace_leaf = false;
4649         btrfs_node_key_to_cpu(reloc_parent, &block->first_key, reloc_slot);
4650
4651         /* Insert @block into @blocks */
4652         spin_lock(&blocks->lock);
4653         cur = &blocks->blocks[level].rb_node;
4654         while (*cur) {
4655                 struct btrfs_qgroup_swapped_block *entry;
4656
4657                 parent = *cur;
4658                 entry = rb_entry(parent, struct btrfs_qgroup_swapped_block,
4659                                  node);
4660
4661                 if (entry->subvol_bytenr < block->subvol_bytenr) {
4662                         cur = &(*cur)->rb_left;
4663                 } else if (entry->subvol_bytenr > block->subvol_bytenr) {
4664                         cur = &(*cur)->rb_right;
4665                 } else {
4666                         if (entry->subvol_generation !=
4667                                         block->subvol_generation ||
4668                             entry->reloc_bytenr != block->reloc_bytenr ||
4669                             entry->reloc_generation !=
4670                                         block->reloc_generation) {
4671                                 /*
4672                                  * Duplicated but mismatch entry found.
4673                                  * Shouldn't happen.
4674                                  *
4675                                  * Marking qgroup inconsistent should be enough
4676                                  * for end users.
4677                                  */
4678                                 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
4679                                 ret = -EEXIST;
4680                         }
4681                         kfree(block);
4682                         goto out_unlock;
4683                 }
4684         }
4685         rb_link_node(&block->node, parent, cur);
4686         rb_insert_color(&block->node, &blocks->blocks[level]);
4687         blocks->swapped = true;
4688 out_unlock:
4689         spin_unlock(&blocks->lock);
4690 out:
4691         if (ret < 0)
4692                 qgroup_mark_inconsistent(fs_info);
4693         return ret;
4694 }
4695
4696 /*
4697  * Check if the tree block is a subtree root, and if so do the needed
4698  * delayed subtree trace for qgroup.
4699  *
4700  * This is called during btrfs_cow_block().
4701  */
4702 int btrfs_qgroup_trace_subtree_after_cow(struct btrfs_trans_handle *trans,
4703                                          struct btrfs_root *root,
4704                                          struct extent_buffer *subvol_eb)
4705 {
4706         struct btrfs_fs_info *fs_info = root->fs_info;
4707         struct btrfs_tree_parent_check check = { 0 };
4708         struct btrfs_qgroup_swapped_blocks *blocks = &root->swapped_blocks;
4709         struct btrfs_qgroup_swapped_block *block;
4710         struct extent_buffer *reloc_eb = NULL;
4711         struct rb_node *node;
4712         bool found = false;
4713         bool swapped = false;
4714         int level = btrfs_header_level(subvol_eb);
4715         int ret = 0;
4716         int i;
4717
4718         if (!btrfs_qgroup_full_accounting(fs_info))
4719                 return 0;
4720         if (!is_fstree(btrfs_root_id(root)) || !root->reloc_root)
4721                 return 0;
4722
4723         spin_lock(&blocks->lock);
4724         if (!blocks->swapped) {
4725                 spin_unlock(&blocks->lock);
4726                 return 0;
4727         }
4728         node = blocks->blocks[level].rb_node;
4729
4730         while (node) {
4731                 block = rb_entry(node, struct btrfs_qgroup_swapped_block, node);
4732                 if (block->subvol_bytenr < subvol_eb->start) {
4733                         node = node->rb_left;
4734                 } else if (block->subvol_bytenr > subvol_eb->start) {
4735                         node = node->rb_right;
4736                 } else {
4737                         found = true;
4738                         break;
4739                 }
4740         }
4741         if (!found) {
4742                 spin_unlock(&blocks->lock);
4743                 goto out;
4744         }
4745         /* Found one, remove it from @blocks first and update blocks->swapped */
4746         rb_erase(&block->node, &blocks->blocks[level]);
4747         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
4748                 if (RB_EMPTY_ROOT(&blocks->blocks[i])) {
4749                         swapped = true;
4750                         break;
4751                 }
4752         }
4753         blocks->swapped = swapped;
4754         spin_unlock(&blocks->lock);
4755
4756         check.level = block->level;
4757         check.transid = block->reloc_generation;
4758         check.has_first_key = true;
4759         memcpy(&check.first_key, &block->first_key, sizeof(check.first_key));
4760
4761         /* Read out reloc subtree root */
4762         reloc_eb = read_tree_block(fs_info, block->reloc_bytenr, &check);
4763         if (IS_ERR(reloc_eb)) {
4764                 ret = PTR_ERR(reloc_eb);
4765                 reloc_eb = NULL;
4766                 goto free_out;
4767         }
4768         if (!extent_buffer_uptodate(reloc_eb)) {
4769                 ret = -EIO;
4770                 goto free_out;
4771         }
4772
4773         ret = qgroup_trace_subtree_swap(trans, reloc_eb, subvol_eb,
4774                         block->last_snapshot, block->trace_leaf);
4775 free_out:
4776         kfree(block);
4777         free_extent_buffer(reloc_eb);
4778 out:
4779         if (ret < 0) {
4780                 btrfs_err_rl(fs_info,
4781                              "failed to account subtree at bytenr %llu: %d",
4782                              subvol_eb->start, ret);
4783                 qgroup_mark_inconsistent(fs_info);
4784         }
4785         return ret;
4786 }
4787
4788 void btrfs_qgroup_destroy_extent_records(struct btrfs_transaction *trans)
4789 {
4790         struct btrfs_qgroup_extent_record *entry;
4791         struct btrfs_qgroup_extent_record *next;
4792         struct rb_root *root;
4793
4794         root = &trans->delayed_refs.dirty_extent_root;
4795         rbtree_postorder_for_each_entry_safe(entry, next, root, node) {
4796                 ulist_free(entry->old_roots);
4797                 kfree(entry);
4798         }
4799         *root = RB_ROOT;
4800 }
4801
4802 void btrfs_free_squota_rsv(struct btrfs_fs_info *fs_info, u64 root, u64 rsv_bytes)
4803 {
4804         if (btrfs_qgroup_mode(fs_info) != BTRFS_QGROUP_MODE_SIMPLE)
4805                 return;
4806
4807         if (!is_fstree(root))
4808                 return;
4809
4810         btrfs_qgroup_free_refroot(fs_info, root, rsv_bytes, BTRFS_QGROUP_RSV_DATA);
4811 }
4812
4813 int btrfs_record_squota_delta(struct btrfs_fs_info *fs_info,
4814                               struct btrfs_squota_delta *delta)
4815 {
4816         int ret;
4817         struct btrfs_qgroup *qgroup;
4818         struct btrfs_qgroup *qg;
4819         LIST_HEAD(qgroup_list);
4820         u64 root = delta->root;
4821         u64 num_bytes = delta->num_bytes;
4822         const int sign = (delta->is_inc ? 1 : -1);
4823
4824         if (btrfs_qgroup_mode(fs_info) != BTRFS_QGROUP_MODE_SIMPLE)
4825                 return 0;
4826
4827         if (!is_fstree(root))
4828                 return 0;
4829
4830         /* If the extent predates enabling quotas, don't count it. */
4831         if (delta->generation < fs_info->qgroup_enable_gen)
4832                 return 0;
4833
4834         spin_lock(&fs_info->qgroup_lock);
4835         qgroup = find_qgroup_rb(fs_info, root);
4836         if (!qgroup) {
4837                 ret = -ENOENT;
4838                 goto out;
4839         }
4840
4841         ret = 0;
4842         qgroup_iterator_add(&qgroup_list, qgroup);
4843         list_for_each_entry(qg, &qgroup_list, iterator) {
4844                 struct btrfs_qgroup_list *glist;
4845
4846                 qg->excl += num_bytes * sign;
4847                 qg->rfer += num_bytes * sign;
4848                 qgroup_dirty(fs_info, qg);
4849
4850                 list_for_each_entry(glist, &qg->groups, next_group)
4851                         qgroup_iterator_add(&qgroup_list, glist->group);
4852         }
4853         qgroup_iterator_clean(&qgroup_list);
4854
4855 out:
4856         spin_unlock(&fs_info->qgroup_lock);
4857         return ret;
4858 }