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