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 with
1236          * snapshot creation.
1237          */
1238         lockdep_assert_held_write(&fs_info->subvol_sem);
1239
1240         /*
1241          * Lock the cleaner mutex to prevent races with concurrent relocation,
1242          * because relocation may be building backrefs for blocks of the quota
1243          * root while we are deleting the root. This is like dropping fs roots
1244          * of deleted snapshots/subvolumes, we need the same protection.
1245          *
1246          * This also prevents races between concurrent tasks trying to disable
1247          * quotas, because we will unlock and relock qgroup_ioctl_lock across
1248          * BTRFS_FS_QUOTA_ENABLED changes.
1249          */
1250         mutex_lock(&fs_info->cleaner_mutex);
1251
1252         mutex_lock(&fs_info->qgroup_ioctl_lock);
1253         if (!fs_info->quota_root)
1254                 goto out;
1255
1256         /*
1257          * Unlock the qgroup_ioctl_lock mutex before waiting for the rescan worker to
1258          * complete. Otherwise we can deadlock because btrfs_remove_qgroup() needs
1259          * to lock that mutex while holding a transaction handle and the rescan
1260          * worker needs to commit a transaction.
1261          */
1262         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1263
1264         /*
1265          * Request qgroup rescan worker to complete and wait for it. This wait
1266          * must be done before transaction start for quota disable since it may
1267          * deadlock with transaction by the qgroup rescan worker.
1268          */
1269         clear_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1270         btrfs_qgroup_wait_for_completion(fs_info, false);
1271
1272         /*
1273          * 1 For the root item
1274          *
1275          * We should also reserve enough items for the quota tree deletion in
1276          * btrfs_clean_quota_tree but this is not done.
1277          *
1278          * Also, we must always start a transaction without holding the mutex
1279          * qgroup_ioctl_lock, see btrfs_quota_enable().
1280          */
1281         trans = btrfs_start_transaction(fs_info->tree_root, 1);
1282
1283         mutex_lock(&fs_info->qgroup_ioctl_lock);
1284         if (IS_ERR(trans)) {
1285                 ret = PTR_ERR(trans);
1286                 trans = NULL;
1287                 set_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1288                 goto out;
1289         }
1290
1291         if (!fs_info->quota_root)
1292                 goto out;
1293
1294         spin_lock(&fs_info->qgroup_lock);
1295         quota_root = fs_info->quota_root;
1296         fs_info->quota_root = NULL;
1297         fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_ON;
1298         fs_info->qgroup_drop_subtree_thres = BTRFS_MAX_LEVEL;
1299         spin_unlock(&fs_info->qgroup_lock);
1300
1301         btrfs_free_qgroup_config(fs_info);
1302
1303         ret = btrfs_clean_quota_tree(trans, quota_root);
1304         if (ret) {
1305                 btrfs_abort_transaction(trans, ret);
1306                 goto out;
1307         }
1308
1309         ret = btrfs_del_root(trans, &quota_root->root_key);
1310         if (ret) {
1311                 btrfs_abort_transaction(trans, ret);
1312                 goto out;
1313         }
1314
1315         spin_lock(&fs_info->trans_lock);
1316         list_del(&quota_root->dirty_list);
1317         spin_unlock(&fs_info->trans_lock);
1318
1319         btrfs_tree_lock(quota_root->node);
1320         btrfs_clear_buffer_dirty(trans, quota_root->node);
1321         btrfs_tree_unlock(quota_root->node);
1322         btrfs_free_tree_block(trans, btrfs_root_id(quota_root),
1323                               quota_root->node, 0, 1);
1324
1325         btrfs_put_root(quota_root);
1326
1327 out:
1328         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1329         if (ret && trans)
1330                 btrfs_end_transaction(trans);
1331         else if (trans)
1332                 ret = btrfs_end_transaction(trans);
1333         mutex_unlock(&fs_info->cleaner_mutex);
1334
1335         return ret;
1336 }
1337
1338 static void qgroup_dirty(struct btrfs_fs_info *fs_info,
1339                          struct btrfs_qgroup *qgroup)
1340 {
1341         if (list_empty(&qgroup->dirty))
1342                 list_add(&qgroup->dirty, &fs_info->dirty_qgroups);
1343 }
1344
1345 /*
1346  * The easy accounting, we're updating qgroup relationship whose child qgroup
1347  * only has exclusive extents.
1348  *
1349  * In this case, all exclusive extents will also be exclusive for parent, so
1350  * excl/rfer just get added/removed.
1351  *
1352  * So is qgroup reservation space, which should also be added/removed to
1353  * parent.
1354  * Or when child tries to release reservation space, parent will underflow its
1355  * reservation (for relationship adding case).
1356  *
1357  * Caller should hold fs_info->qgroup_lock.
1358  */
1359 static int __qgroup_excl_accounting(struct btrfs_fs_info *fs_info,
1360                                     struct ulist *tmp, u64 ref_root,
1361                                     struct btrfs_qgroup *src, int sign)
1362 {
1363         struct btrfs_qgroup *qgroup;
1364         struct btrfs_qgroup_list *glist;
1365         struct ulist_node *unode;
1366         struct ulist_iterator uiter;
1367         u64 num_bytes = src->excl;
1368         int ret = 0;
1369
1370         qgroup = find_qgroup_rb(fs_info, ref_root);
1371         if (!qgroup)
1372                 goto out;
1373
1374         qgroup->rfer += sign * num_bytes;
1375         qgroup->rfer_cmpr += sign * num_bytes;
1376
1377         WARN_ON(sign < 0 && qgroup->excl < num_bytes);
1378         qgroup->excl += sign * num_bytes;
1379         qgroup->excl_cmpr += sign * num_bytes;
1380
1381         if (sign > 0)
1382                 qgroup_rsv_add_by_qgroup(fs_info, qgroup, src);
1383         else
1384                 qgroup_rsv_release_by_qgroup(fs_info, qgroup, src);
1385
1386         qgroup_dirty(fs_info, qgroup);
1387
1388         /* Get all of the parent groups that contain this qgroup */
1389         list_for_each_entry(glist, &qgroup->groups, next_group) {
1390                 ret = ulist_add(tmp, glist->group->qgroupid,
1391                                 qgroup_to_aux(glist->group), GFP_ATOMIC);
1392                 if (ret < 0)
1393                         goto out;
1394         }
1395
1396         /* Iterate all of the parents and adjust their reference counts */
1397         ULIST_ITER_INIT(&uiter);
1398         while ((unode = ulist_next(tmp, &uiter))) {
1399                 qgroup = unode_aux_to_qgroup(unode);
1400                 qgroup->rfer += sign * num_bytes;
1401                 qgroup->rfer_cmpr += sign * num_bytes;
1402                 WARN_ON(sign < 0 && qgroup->excl < num_bytes);
1403                 qgroup->excl += sign * num_bytes;
1404                 if (sign > 0)
1405                         qgroup_rsv_add_by_qgroup(fs_info, qgroup, src);
1406                 else
1407                         qgroup_rsv_release_by_qgroup(fs_info, qgroup, src);
1408                 qgroup->excl_cmpr += sign * num_bytes;
1409                 qgroup_dirty(fs_info, qgroup);
1410
1411                 /* Add any parents of the parents */
1412                 list_for_each_entry(glist, &qgroup->groups, next_group) {
1413                         ret = ulist_add(tmp, glist->group->qgroupid,
1414                                         qgroup_to_aux(glist->group), GFP_ATOMIC);
1415                         if (ret < 0)
1416                                 goto out;
1417                 }
1418         }
1419         ret = 0;
1420 out:
1421         return ret;
1422 }
1423
1424
1425 /*
1426  * Quick path for updating qgroup with only excl refs.
1427  *
1428  * In that case, just update all parent will be enough.
1429  * Or we needs to do a full rescan.
1430  * Caller should also hold fs_info->qgroup_lock.
1431  *
1432  * Return 0 for quick update, return >0 for need to full rescan
1433  * and mark INCONSISTENT flag.
1434  * Return < 0 for other error.
1435  */
1436 static int quick_update_accounting(struct btrfs_fs_info *fs_info,
1437                                    struct ulist *tmp, u64 src, u64 dst,
1438                                    int sign)
1439 {
1440         struct btrfs_qgroup *qgroup;
1441         int ret = 1;
1442         int err = 0;
1443
1444         qgroup = find_qgroup_rb(fs_info, src);
1445         if (!qgroup)
1446                 goto out;
1447         if (qgroup->excl == qgroup->rfer) {
1448                 ret = 0;
1449                 err = __qgroup_excl_accounting(fs_info, tmp, dst,
1450                                                qgroup, sign);
1451                 if (err < 0) {
1452                         ret = err;
1453                         goto out;
1454                 }
1455         }
1456 out:
1457         if (ret)
1458                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1459         return ret;
1460 }
1461
1462 int btrfs_add_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1463                               u64 dst)
1464 {
1465         struct btrfs_fs_info *fs_info = trans->fs_info;
1466         struct btrfs_qgroup *parent;
1467         struct btrfs_qgroup *member;
1468         struct btrfs_qgroup_list *list;
1469         struct ulist *tmp;
1470         unsigned int nofs_flag;
1471         int ret = 0;
1472
1473         /* Check the level of src and dst first */
1474         if (btrfs_qgroup_level(src) >= btrfs_qgroup_level(dst))
1475                 return -EINVAL;
1476
1477         /* We hold a transaction handle open, must do a NOFS allocation. */
1478         nofs_flag = memalloc_nofs_save();
1479         tmp = ulist_alloc(GFP_KERNEL);
1480         memalloc_nofs_restore(nofs_flag);
1481         if (!tmp)
1482                 return -ENOMEM;
1483
1484         mutex_lock(&fs_info->qgroup_ioctl_lock);
1485         if (!fs_info->quota_root) {
1486                 ret = -ENOTCONN;
1487                 goto out;
1488         }
1489         member = find_qgroup_rb(fs_info, src);
1490         parent = find_qgroup_rb(fs_info, dst);
1491         if (!member || !parent) {
1492                 ret = -EINVAL;
1493                 goto out;
1494         }
1495
1496         /* check if such qgroup relation exist firstly */
1497         list_for_each_entry(list, &member->groups, next_group) {
1498                 if (list->group == parent) {
1499                         ret = -EEXIST;
1500                         goto out;
1501                 }
1502         }
1503
1504         ret = add_qgroup_relation_item(trans, src, dst);
1505         if (ret)
1506                 goto out;
1507
1508         ret = add_qgroup_relation_item(trans, dst, src);
1509         if (ret) {
1510                 del_qgroup_relation_item(trans, src, dst);
1511                 goto out;
1512         }
1513
1514         spin_lock(&fs_info->qgroup_lock);
1515         ret = __add_relation_rb(member, parent);
1516         if (ret < 0) {
1517                 spin_unlock(&fs_info->qgroup_lock);
1518                 goto out;
1519         }
1520         ret = quick_update_accounting(fs_info, tmp, src, dst, 1);
1521         spin_unlock(&fs_info->qgroup_lock);
1522 out:
1523         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1524         ulist_free(tmp);
1525         return ret;
1526 }
1527
1528 static int __del_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1529                                  u64 dst)
1530 {
1531         struct btrfs_fs_info *fs_info = trans->fs_info;
1532         struct btrfs_qgroup *parent;
1533         struct btrfs_qgroup *member;
1534         struct btrfs_qgroup_list *list;
1535         struct ulist *tmp;
1536         bool found = false;
1537         unsigned int nofs_flag;
1538         int ret = 0;
1539         int ret2;
1540
1541         /* We hold a transaction handle open, must do a NOFS allocation. */
1542         nofs_flag = memalloc_nofs_save();
1543         tmp = ulist_alloc(GFP_KERNEL);
1544         memalloc_nofs_restore(nofs_flag);
1545         if (!tmp)
1546                 return -ENOMEM;
1547
1548         if (!fs_info->quota_root) {
1549                 ret = -ENOTCONN;
1550                 goto out;
1551         }
1552
1553         member = find_qgroup_rb(fs_info, src);
1554         parent = find_qgroup_rb(fs_info, dst);
1555         /*
1556          * The parent/member pair doesn't exist, then try to delete the dead
1557          * relation items only.
1558          */
1559         if (!member || !parent)
1560                 goto delete_item;
1561
1562         /* check if such qgroup relation exist firstly */
1563         list_for_each_entry(list, &member->groups, next_group) {
1564                 if (list->group == parent) {
1565                         found = true;
1566                         break;
1567                 }
1568         }
1569
1570 delete_item:
1571         ret = del_qgroup_relation_item(trans, src, dst);
1572         if (ret < 0 && ret != -ENOENT)
1573                 goto out;
1574         ret2 = del_qgroup_relation_item(trans, dst, src);
1575         if (ret2 < 0 && ret2 != -ENOENT)
1576                 goto out;
1577
1578         /* At least one deletion succeeded, return 0 */
1579         if (!ret || !ret2)
1580                 ret = 0;
1581
1582         if (found) {
1583                 spin_lock(&fs_info->qgroup_lock);
1584                 del_relation_rb(fs_info, src, dst);
1585                 ret = quick_update_accounting(fs_info, tmp, src, dst, -1);
1586                 spin_unlock(&fs_info->qgroup_lock);
1587         }
1588 out:
1589         ulist_free(tmp);
1590         return ret;
1591 }
1592
1593 int btrfs_del_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1594                               u64 dst)
1595 {
1596         struct btrfs_fs_info *fs_info = trans->fs_info;
1597         int ret = 0;
1598
1599         mutex_lock(&fs_info->qgroup_ioctl_lock);
1600         ret = __del_qgroup_relation(trans, src, dst);
1601         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1602
1603         return ret;
1604 }
1605
1606 int btrfs_create_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid)
1607 {
1608         struct btrfs_fs_info *fs_info = trans->fs_info;
1609         struct btrfs_root *quota_root;
1610         struct btrfs_qgroup *qgroup;
1611         int ret = 0;
1612
1613         mutex_lock(&fs_info->qgroup_ioctl_lock);
1614         if (!fs_info->quota_root) {
1615                 ret = -ENOTCONN;
1616                 goto out;
1617         }
1618         quota_root = fs_info->quota_root;
1619         qgroup = find_qgroup_rb(fs_info, qgroupid);
1620         if (qgroup) {
1621                 ret = -EEXIST;
1622                 goto out;
1623         }
1624
1625         ret = add_qgroup_item(trans, quota_root, qgroupid);
1626         if (ret)
1627                 goto out;
1628
1629         spin_lock(&fs_info->qgroup_lock);
1630         qgroup = add_qgroup_rb(fs_info, qgroupid);
1631         spin_unlock(&fs_info->qgroup_lock);
1632
1633         if (IS_ERR(qgroup)) {
1634                 ret = PTR_ERR(qgroup);
1635                 goto out;
1636         }
1637         ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1638 out:
1639         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1640         return ret;
1641 }
1642
1643 int btrfs_remove_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid)
1644 {
1645         struct btrfs_fs_info *fs_info = trans->fs_info;
1646         struct btrfs_qgroup *qgroup;
1647         struct btrfs_qgroup_list *list;
1648         int ret = 0;
1649
1650         mutex_lock(&fs_info->qgroup_ioctl_lock);
1651         if (!fs_info->quota_root) {
1652                 ret = -ENOTCONN;
1653                 goto out;
1654         }
1655
1656         qgroup = find_qgroup_rb(fs_info, qgroupid);
1657         if (!qgroup) {
1658                 ret = -ENOENT;
1659                 goto out;
1660         }
1661
1662         /* Check if there are no children of this qgroup */
1663         if (!list_empty(&qgroup->members)) {
1664                 ret = -EBUSY;
1665                 goto out;
1666         }
1667
1668         ret = del_qgroup_item(trans, qgroupid);
1669         if (ret && ret != -ENOENT)
1670                 goto out;
1671
1672         while (!list_empty(&qgroup->groups)) {
1673                 list = list_first_entry(&qgroup->groups,
1674                                         struct btrfs_qgroup_list, next_group);
1675                 ret = __del_qgroup_relation(trans, qgroupid,
1676                                             list->group->qgroupid);
1677                 if (ret)
1678                         goto out;
1679         }
1680
1681         spin_lock(&fs_info->qgroup_lock);
1682         del_qgroup_rb(fs_info, qgroupid);
1683         spin_unlock(&fs_info->qgroup_lock);
1684
1685         /*
1686          * Remove the qgroup from sysfs now without holding the qgroup_lock
1687          * spinlock, since the sysfs_remove_group() function needs to take
1688          * the mutex kernfs_mutex through kernfs_remove_by_name_ns().
1689          */
1690         btrfs_sysfs_del_one_qgroup(fs_info, qgroup);
1691         kfree(qgroup);
1692 out:
1693         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1694         return ret;
1695 }
1696
1697 int btrfs_limit_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid,
1698                        struct btrfs_qgroup_limit *limit)
1699 {
1700         struct btrfs_fs_info *fs_info = trans->fs_info;
1701         struct btrfs_qgroup *qgroup;
1702         int ret = 0;
1703         /* Sometimes we would want to clear the limit on this qgroup.
1704          * To meet this requirement, we treat the -1 as a special value
1705          * which tell kernel to clear the limit on this qgroup.
1706          */
1707         const u64 CLEAR_VALUE = -1;
1708
1709         mutex_lock(&fs_info->qgroup_ioctl_lock);
1710         if (!fs_info->quota_root) {
1711                 ret = -ENOTCONN;
1712                 goto out;
1713         }
1714
1715         qgroup = find_qgroup_rb(fs_info, qgroupid);
1716         if (!qgroup) {
1717                 ret = -ENOENT;
1718                 goto out;
1719         }
1720
1721         spin_lock(&fs_info->qgroup_lock);
1722         if (limit->flags & BTRFS_QGROUP_LIMIT_MAX_RFER) {
1723                 if (limit->max_rfer == CLEAR_VALUE) {
1724                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_MAX_RFER;
1725                         limit->flags &= ~BTRFS_QGROUP_LIMIT_MAX_RFER;
1726                         qgroup->max_rfer = 0;
1727                 } else {
1728                         qgroup->max_rfer = limit->max_rfer;
1729                 }
1730         }
1731         if (limit->flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) {
1732                 if (limit->max_excl == CLEAR_VALUE) {
1733                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_MAX_EXCL;
1734                         limit->flags &= ~BTRFS_QGROUP_LIMIT_MAX_EXCL;
1735                         qgroup->max_excl = 0;
1736                 } else {
1737                         qgroup->max_excl = limit->max_excl;
1738                 }
1739         }
1740         if (limit->flags & BTRFS_QGROUP_LIMIT_RSV_RFER) {
1741                 if (limit->rsv_rfer == CLEAR_VALUE) {
1742                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_RSV_RFER;
1743                         limit->flags &= ~BTRFS_QGROUP_LIMIT_RSV_RFER;
1744                         qgroup->rsv_rfer = 0;
1745                 } else {
1746                         qgroup->rsv_rfer = limit->rsv_rfer;
1747                 }
1748         }
1749         if (limit->flags & BTRFS_QGROUP_LIMIT_RSV_EXCL) {
1750                 if (limit->rsv_excl == CLEAR_VALUE) {
1751                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_RSV_EXCL;
1752                         limit->flags &= ~BTRFS_QGROUP_LIMIT_RSV_EXCL;
1753                         qgroup->rsv_excl = 0;
1754                 } else {
1755                         qgroup->rsv_excl = limit->rsv_excl;
1756                 }
1757         }
1758         qgroup->lim_flags |= limit->flags;
1759
1760         spin_unlock(&fs_info->qgroup_lock);
1761
1762         ret = update_qgroup_limit_item(trans, qgroup);
1763         if (ret) {
1764                 qgroup_mark_inconsistent(fs_info);
1765                 btrfs_info(fs_info, "unable to update quota limit for %llu",
1766                        qgroupid);
1767         }
1768
1769 out:
1770         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1771         return ret;
1772 }
1773
1774 int btrfs_qgroup_trace_extent_nolock(struct btrfs_fs_info *fs_info,
1775                                 struct btrfs_delayed_ref_root *delayed_refs,
1776                                 struct btrfs_qgroup_extent_record *record)
1777 {
1778         struct rb_node **p = &delayed_refs->dirty_extent_root.rb_node;
1779         struct rb_node *parent_node = NULL;
1780         struct btrfs_qgroup_extent_record *entry;
1781         u64 bytenr = record->bytenr;
1782
1783         lockdep_assert_held(&delayed_refs->lock);
1784         trace_btrfs_qgroup_trace_extent(fs_info, record);
1785
1786         while (*p) {
1787                 parent_node = *p;
1788                 entry = rb_entry(parent_node, struct btrfs_qgroup_extent_record,
1789                                  node);
1790                 if (bytenr < entry->bytenr) {
1791                         p = &(*p)->rb_left;
1792                 } else if (bytenr > entry->bytenr) {
1793                         p = &(*p)->rb_right;
1794                 } else {
1795                         if (record->data_rsv && !entry->data_rsv) {
1796                                 entry->data_rsv = record->data_rsv;
1797                                 entry->data_rsv_refroot =
1798                                         record->data_rsv_refroot;
1799                         }
1800                         return 1;
1801                 }
1802         }
1803
1804         rb_link_node(&record->node, parent_node, p);
1805         rb_insert_color(&record->node, &delayed_refs->dirty_extent_root);
1806         return 0;
1807 }
1808
1809 int btrfs_qgroup_trace_extent_post(struct btrfs_trans_handle *trans,
1810                                    struct btrfs_qgroup_extent_record *qrecord)
1811 {
1812         struct btrfs_backref_walk_ctx ctx = { 0 };
1813         int ret;
1814
1815         /*
1816          * We are always called in a context where we are already holding a
1817          * transaction handle. Often we are called when adding a data delayed
1818          * reference from btrfs_truncate_inode_items() (truncating or unlinking),
1819          * in which case we will be holding a write lock on extent buffer from a
1820          * subvolume tree. In this case we can't allow btrfs_find_all_roots() to
1821          * acquire fs_info->commit_root_sem, because that is a higher level lock
1822          * that must be acquired before locking any extent buffers.
1823          *
1824          * So we want btrfs_find_all_roots() to not acquire the commit_root_sem
1825          * but we can't pass it a non-NULL transaction handle, because otherwise
1826          * it would not use commit roots and would lock extent buffers, causing
1827          * a deadlock if it ends up trying to read lock the same extent buffer
1828          * that was previously write locked at btrfs_truncate_inode_items().
1829          *
1830          * So pass a NULL transaction handle to btrfs_find_all_roots() and
1831          * explicitly tell it to not acquire the commit_root_sem - if we are
1832          * holding a transaction handle we don't need its protection.
1833          */
1834         ASSERT(trans != NULL);
1835
1836         if (trans->fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING)
1837                 return 0;
1838
1839         ctx.bytenr = qrecord->bytenr;
1840         ctx.fs_info = trans->fs_info;
1841
1842         ret = btrfs_find_all_roots(&ctx, true);
1843         if (ret < 0) {
1844                 qgroup_mark_inconsistent(trans->fs_info);
1845                 btrfs_warn(trans->fs_info,
1846 "error accounting new delayed refs extent (err code: %d), quota inconsistent",
1847                         ret);
1848                 return 0;
1849         }
1850
1851         /*
1852          * Here we don't need to get the lock of
1853          * trans->transaction->delayed_refs, since inserted qrecord won't
1854          * be deleted, only qrecord->node may be modified (new qrecord insert)
1855          *
1856          * So modifying qrecord->old_roots is safe here
1857          */
1858         qrecord->old_roots = ctx.roots;
1859         return 0;
1860 }
1861
1862 int btrfs_qgroup_trace_extent(struct btrfs_trans_handle *trans, u64 bytenr,
1863                               u64 num_bytes)
1864 {
1865         struct btrfs_fs_info *fs_info = trans->fs_info;
1866         struct btrfs_qgroup_extent_record *record;
1867         struct btrfs_delayed_ref_root *delayed_refs;
1868         int ret;
1869
1870         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)
1871             || bytenr == 0 || num_bytes == 0)
1872                 return 0;
1873         record = kzalloc(sizeof(*record), GFP_NOFS);
1874         if (!record)
1875                 return -ENOMEM;
1876
1877         delayed_refs = &trans->transaction->delayed_refs;
1878         record->bytenr = bytenr;
1879         record->num_bytes = num_bytes;
1880         record->old_roots = NULL;
1881
1882         spin_lock(&delayed_refs->lock);
1883         ret = btrfs_qgroup_trace_extent_nolock(fs_info, delayed_refs, record);
1884         spin_unlock(&delayed_refs->lock);
1885         if (ret > 0) {
1886                 kfree(record);
1887                 return 0;
1888         }
1889         return btrfs_qgroup_trace_extent_post(trans, record);
1890 }
1891
1892 int btrfs_qgroup_trace_leaf_items(struct btrfs_trans_handle *trans,
1893                                   struct extent_buffer *eb)
1894 {
1895         struct btrfs_fs_info *fs_info = trans->fs_info;
1896         int nr = btrfs_header_nritems(eb);
1897         int i, extent_type, ret;
1898         struct btrfs_key key;
1899         struct btrfs_file_extent_item *fi;
1900         u64 bytenr, num_bytes;
1901
1902         /* We can be called directly from walk_up_proc() */
1903         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
1904                 return 0;
1905
1906         for (i = 0; i < nr; i++) {
1907                 btrfs_item_key_to_cpu(eb, &key, i);
1908
1909                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1910                         continue;
1911
1912                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
1913                 /* filter out non qgroup-accountable extents  */
1914                 extent_type = btrfs_file_extent_type(eb, fi);
1915
1916                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
1917                         continue;
1918
1919                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
1920                 if (!bytenr)
1921                         continue;
1922
1923                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
1924
1925                 ret = btrfs_qgroup_trace_extent(trans, bytenr, num_bytes);
1926                 if (ret)
1927                         return ret;
1928         }
1929         cond_resched();
1930         return 0;
1931 }
1932
1933 /*
1934  * Walk up the tree from the bottom, freeing leaves and any interior
1935  * nodes which have had all slots visited. If a node (leaf or
1936  * interior) is freed, the node above it will have it's slot
1937  * incremented. The root node will never be freed.
1938  *
1939  * At the end of this function, we should have a path which has all
1940  * slots incremented to the next position for a search. If we need to
1941  * read a new node it will be NULL and the node above it will have the
1942  * correct slot selected for a later read.
1943  *
1944  * If we increment the root nodes slot counter past the number of
1945  * elements, 1 is returned to signal completion of the search.
1946  */
1947 static int adjust_slots_upwards(struct btrfs_path *path, int root_level)
1948 {
1949         int level = 0;
1950         int nr, slot;
1951         struct extent_buffer *eb;
1952
1953         if (root_level == 0)
1954                 return 1;
1955
1956         while (level <= root_level) {
1957                 eb = path->nodes[level];
1958                 nr = btrfs_header_nritems(eb);
1959                 path->slots[level]++;
1960                 slot = path->slots[level];
1961                 if (slot >= nr || level == 0) {
1962                         /*
1963                          * Don't free the root -  we will detect this
1964                          * condition after our loop and return a
1965                          * positive value for caller to stop walking the tree.
1966                          */
1967                         if (level != root_level) {
1968                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
1969                                 path->locks[level] = 0;
1970
1971                                 free_extent_buffer(eb);
1972                                 path->nodes[level] = NULL;
1973                                 path->slots[level] = 0;
1974                         }
1975                 } else {
1976                         /*
1977                          * We have a valid slot to walk back down
1978                          * from. Stop here so caller can process these
1979                          * new nodes.
1980                          */
1981                         break;
1982                 }
1983
1984                 level++;
1985         }
1986
1987         eb = path->nodes[root_level];
1988         if (path->slots[root_level] >= btrfs_header_nritems(eb))
1989                 return 1;
1990
1991         return 0;
1992 }
1993
1994 /*
1995  * Helper function to trace a subtree tree block swap.
1996  *
1997  * The swap will happen in highest tree block, but there may be a lot of
1998  * tree blocks involved.
1999  *
2000  * For example:
2001  *  OO = Old tree blocks
2002  *  NN = New tree blocks allocated during balance
2003  *
2004  *           File tree (257)                  Reloc tree for 257
2005  * L2              OO                                NN
2006  *               /    \                            /    \
2007  * L1          OO      OO (a)                    OO      NN (a)
2008  *            / \     / \                       / \     / \
2009  * L0       OO   OO OO   OO                   OO   OO NN   NN
2010  *                  (b)  (c)                          (b)  (c)
2011  *
2012  * When calling qgroup_trace_extent_swap(), we will pass:
2013  * @src_eb = OO(a)
2014  * @dst_path = [ nodes[1] = NN(a), nodes[0] = NN(c) ]
2015  * @dst_level = 0
2016  * @root_level = 1
2017  *
2018  * In that case, qgroup_trace_extent_swap() will search from OO(a) to
2019  * reach OO(c), then mark both OO(c) and NN(c) as qgroup dirty.
2020  *
2021  * The main work of qgroup_trace_extent_swap() can be split into 3 parts:
2022  *
2023  * 1) Tree search from @src_eb
2024  *    It should acts as a simplified btrfs_search_slot().
2025  *    The key for search can be extracted from @dst_path->nodes[dst_level]
2026  *    (first key).
2027  *
2028  * 2) Mark the final tree blocks in @src_path and @dst_path qgroup dirty
2029  *    NOTE: In above case, OO(a) and NN(a) won't be marked qgroup dirty.
2030  *    They should be marked during previous (@dst_level = 1) iteration.
2031  *
2032  * 3) Mark file extents in leaves dirty
2033  *    We don't have good way to pick out new file extents only.
2034  *    So we still follow the old method by scanning all file extents in
2035  *    the leave.
2036  *
2037  * This function can free us from keeping two paths, thus later we only need
2038  * to care about how to iterate all new tree blocks in reloc tree.
2039  */
2040 static int qgroup_trace_extent_swap(struct btrfs_trans_handle* trans,
2041                                     struct extent_buffer *src_eb,
2042                                     struct btrfs_path *dst_path,
2043                                     int dst_level, int root_level,
2044                                     bool trace_leaf)
2045 {
2046         struct btrfs_key key;
2047         struct btrfs_path *src_path;
2048         struct btrfs_fs_info *fs_info = trans->fs_info;
2049         u32 nodesize = fs_info->nodesize;
2050         int cur_level = root_level;
2051         int ret;
2052
2053         BUG_ON(dst_level > root_level);
2054         /* Level mismatch */
2055         if (btrfs_header_level(src_eb) != root_level)
2056                 return -EINVAL;
2057
2058         src_path = btrfs_alloc_path();
2059         if (!src_path) {
2060                 ret = -ENOMEM;
2061                 goto out;
2062         }
2063
2064         if (dst_level)
2065                 btrfs_node_key_to_cpu(dst_path->nodes[dst_level], &key, 0);
2066         else
2067                 btrfs_item_key_to_cpu(dst_path->nodes[dst_level], &key, 0);
2068
2069         /* For src_path */
2070         atomic_inc(&src_eb->refs);
2071         src_path->nodes[root_level] = src_eb;
2072         src_path->slots[root_level] = dst_path->slots[root_level];
2073         src_path->locks[root_level] = 0;
2074
2075         /* A simplified version of btrfs_search_slot() */
2076         while (cur_level >= dst_level) {
2077                 struct btrfs_key src_key;
2078                 struct btrfs_key dst_key;
2079
2080                 if (src_path->nodes[cur_level] == NULL) {
2081                         struct extent_buffer *eb;
2082                         int parent_slot;
2083
2084                         eb = src_path->nodes[cur_level + 1];
2085                         parent_slot = src_path->slots[cur_level + 1];
2086
2087                         eb = btrfs_read_node_slot(eb, parent_slot);
2088                         if (IS_ERR(eb)) {
2089                                 ret = PTR_ERR(eb);
2090                                 goto out;
2091                         }
2092
2093                         src_path->nodes[cur_level] = eb;
2094
2095                         btrfs_tree_read_lock(eb);
2096                         src_path->locks[cur_level] = BTRFS_READ_LOCK;
2097                 }
2098
2099                 src_path->slots[cur_level] = dst_path->slots[cur_level];
2100                 if (cur_level) {
2101                         btrfs_node_key_to_cpu(dst_path->nodes[cur_level],
2102                                         &dst_key, dst_path->slots[cur_level]);
2103                         btrfs_node_key_to_cpu(src_path->nodes[cur_level],
2104                                         &src_key, src_path->slots[cur_level]);
2105                 } else {
2106                         btrfs_item_key_to_cpu(dst_path->nodes[cur_level],
2107                                         &dst_key, dst_path->slots[cur_level]);
2108                         btrfs_item_key_to_cpu(src_path->nodes[cur_level],
2109                                         &src_key, src_path->slots[cur_level]);
2110                 }
2111                 /* Content mismatch, something went wrong */
2112                 if (btrfs_comp_cpu_keys(&dst_key, &src_key)) {
2113                         ret = -ENOENT;
2114                         goto out;
2115                 }
2116                 cur_level--;
2117         }
2118
2119         /*
2120          * Now both @dst_path and @src_path have been populated, record the tree
2121          * blocks for qgroup accounting.
2122          */
2123         ret = btrfs_qgroup_trace_extent(trans, src_path->nodes[dst_level]->start,
2124                                         nodesize);
2125         if (ret < 0)
2126                 goto out;
2127         ret = btrfs_qgroup_trace_extent(trans, dst_path->nodes[dst_level]->start,
2128                                         nodesize);
2129         if (ret < 0)
2130                 goto out;
2131
2132         /* Record leaf file extents */
2133         if (dst_level == 0 && trace_leaf) {
2134                 ret = btrfs_qgroup_trace_leaf_items(trans, src_path->nodes[0]);
2135                 if (ret < 0)
2136                         goto out;
2137                 ret = btrfs_qgroup_trace_leaf_items(trans, dst_path->nodes[0]);
2138         }
2139 out:
2140         btrfs_free_path(src_path);
2141         return ret;
2142 }
2143
2144 /*
2145  * Helper function to do recursive generation-aware depth-first search, to
2146  * locate all new tree blocks in a subtree of reloc tree.
2147  *
2148  * E.g. (OO = Old tree blocks, NN = New tree blocks, whose gen == last_snapshot)
2149  *         reloc tree
2150  * L2         NN (a)
2151  *          /    \
2152  * L1    OO        NN (b)
2153  *      /  \      /  \
2154  * L0  OO  OO    OO  NN
2155  *               (c) (d)
2156  * If we pass:
2157  * @dst_path = [ nodes[1] = NN(b), nodes[0] = NULL ],
2158  * @cur_level = 1
2159  * @root_level = 1
2160  *
2161  * We will iterate through tree blocks NN(b), NN(d) and info qgroup to trace
2162  * above tree blocks along with their counter parts in file tree.
2163  * While during search, old tree blocks OO(c) will be skipped as tree block swap
2164  * won't affect OO(c).
2165  */
2166 static int qgroup_trace_new_subtree_blocks(struct btrfs_trans_handle* trans,
2167                                            struct extent_buffer *src_eb,
2168                                            struct btrfs_path *dst_path,
2169                                            int cur_level, int root_level,
2170                                            u64 last_snapshot, bool trace_leaf)
2171 {
2172         struct btrfs_fs_info *fs_info = trans->fs_info;
2173         struct extent_buffer *eb;
2174         bool need_cleanup = false;
2175         int ret = 0;
2176         int i;
2177
2178         /* Level sanity check */
2179         if (cur_level < 0 || cur_level >= BTRFS_MAX_LEVEL - 1 ||
2180             root_level < 0 || root_level >= BTRFS_MAX_LEVEL - 1 ||
2181             root_level < cur_level) {
2182                 btrfs_err_rl(fs_info,
2183                         "%s: bad levels, cur_level=%d root_level=%d",
2184                         __func__, cur_level, root_level);
2185                 return -EUCLEAN;
2186         }
2187
2188         /* Read the tree block if needed */
2189         if (dst_path->nodes[cur_level] == NULL) {
2190                 int parent_slot;
2191                 u64 child_gen;
2192
2193                 /*
2194                  * dst_path->nodes[root_level] must be initialized before
2195                  * calling this function.
2196                  */
2197                 if (cur_level == root_level) {
2198                         btrfs_err_rl(fs_info,
2199         "%s: dst_path->nodes[%d] not initialized, root_level=%d cur_level=%d",
2200                                 __func__, root_level, root_level, cur_level);
2201                         return -EUCLEAN;
2202                 }
2203
2204                 /*
2205                  * We need to get child blockptr/gen from parent before we can
2206                  * read it.
2207                   */
2208                 eb = dst_path->nodes[cur_level + 1];
2209                 parent_slot = dst_path->slots[cur_level + 1];
2210                 child_gen = btrfs_node_ptr_generation(eb, parent_slot);
2211
2212                 /* This node is old, no need to trace */
2213                 if (child_gen < last_snapshot)
2214                         goto out;
2215
2216                 eb = btrfs_read_node_slot(eb, parent_slot);
2217                 if (IS_ERR(eb)) {
2218                         ret = PTR_ERR(eb);
2219                         goto out;
2220                 }
2221
2222                 dst_path->nodes[cur_level] = eb;
2223                 dst_path->slots[cur_level] = 0;
2224
2225                 btrfs_tree_read_lock(eb);
2226                 dst_path->locks[cur_level] = BTRFS_READ_LOCK;
2227                 need_cleanup = true;
2228         }
2229
2230         /* Now record this tree block and its counter part for qgroups */
2231         ret = qgroup_trace_extent_swap(trans, src_eb, dst_path, cur_level,
2232                                        root_level, trace_leaf);
2233         if (ret < 0)
2234                 goto cleanup;
2235
2236         eb = dst_path->nodes[cur_level];
2237
2238         if (cur_level > 0) {
2239                 /* Iterate all child tree blocks */
2240                 for (i = 0; i < btrfs_header_nritems(eb); i++) {
2241                         /* Skip old tree blocks as they won't be swapped */
2242                         if (btrfs_node_ptr_generation(eb, i) < last_snapshot)
2243                                 continue;
2244                         dst_path->slots[cur_level] = i;
2245
2246                         /* Recursive call (at most 7 times) */
2247                         ret = qgroup_trace_new_subtree_blocks(trans, src_eb,
2248                                         dst_path, cur_level - 1, root_level,
2249                                         last_snapshot, trace_leaf);
2250                         if (ret < 0)
2251                                 goto cleanup;
2252                 }
2253         }
2254
2255 cleanup:
2256         if (need_cleanup) {
2257                 /* Clean up */
2258                 btrfs_tree_unlock_rw(dst_path->nodes[cur_level],
2259                                      dst_path->locks[cur_level]);
2260                 free_extent_buffer(dst_path->nodes[cur_level]);
2261                 dst_path->nodes[cur_level] = NULL;
2262                 dst_path->slots[cur_level] = 0;
2263                 dst_path->locks[cur_level] = 0;
2264         }
2265 out:
2266         return ret;
2267 }
2268
2269 static int qgroup_trace_subtree_swap(struct btrfs_trans_handle *trans,
2270                                 struct extent_buffer *src_eb,
2271                                 struct extent_buffer *dst_eb,
2272                                 u64 last_snapshot, bool trace_leaf)
2273 {
2274         struct btrfs_fs_info *fs_info = trans->fs_info;
2275         struct btrfs_path *dst_path = NULL;
2276         int level;
2277         int ret;
2278
2279         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2280                 return 0;
2281
2282         /* Wrong parameter order */
2283         if (btrfs_header_generation(src_eb) > btrfs_header_generation(dst_eb)) {
2284                 btrfs_err_rl(fs_info,
2285                 "%s: bad parameter order, src_gen=%llu dst_gen=%llu", __func__,
2286                              btrfs_header_generation(src_eb),
2287                              btrfs_header_generation(dst_eb));
2288                 return -EUCLEAN;
2289         }
2290
2291         if (!extent_buffer_uptodate(src_eb) || !extent_buffer_uptodate(dst_eb)) {
2292                 ret = -EIO;
2293                 goto out;
2294         }
2295
2296         level = btrfs_header_level(dst_eb);
2297         dst_path = btrfs_alloc_path();
2298         if (!dst_path) {
2299                 ret = -ENOMEM;
2300                 goto out;
2301         }
2302         /* For dst_path */
2303         atomic_inc(&dst_eb->refs);
2304         dst_path->nodes[level] = dst_eb;
2305         dst_path->slots[level] = 0;
2306         dst_path->locks[level] = 0;
2307
2308         /* Do the generation aware breadth-first search */
2309         ret = qgroup_trace_new_subtree_blocks(trans, src_eb, dst_path, level,
2310                                               level, last_snapshot, trace_leaf);
2311         if (ret < 0)
2312                 goto out;
2313         ret = 0;
2314
2315 out:
2316         btrfs_free_path(dst_path);
2317         if (ret < 0)
2318                 qgroup_mark_inconsistent(fs_info);
2319         return ret;
2320 }
2321
2322 int btrfs_qgroup_trace_subtree(struct btrfs_trans_handle *trans,
2323                                struct extent_buffer *root_eb,
2324                                u64 root_gen, int root_level)
2325 {
2326         struct btrfs_fs_info *fs_info = trans->fs_info;
2327         int ret = 0;
2328         int level;
2329         u8 drop_subptree_thres;
2330         struct extent_buffer *eb = root_eb;
2331         struct btrfs_path *path = NULL;
2332
2333         BUG_ON(root_level < 0 || root_level >= BTRFS_MAX_LEVEL);
2334         BUG_ON(root_eb == NULL);
2335
2336         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2337                 return 0;
2338
2339         spin_lock(&fs_info->qgroup_lock);
2340         drop_subptree_thres = fs_info->qgroup_drop_subtree_thres;
2341         spin_unlock(&fs_info->qgroup_lock);
2342
2343         /*
2344          * This function only gets called for snapshot drop, if we hit a high
2345          * node here, it means we are going to change ownership for quite a lot
2346          * of extents, which will greatly slow down btrfs_commit_transaction().
2347          *
2348          * So here if we find a high tree here, we just skip the accounting and
2349          * mark qgroup inconsistent.
2350          */
2351         if (root_level >= drop_subptree_thres) {
2352                 qgroup_mark_inconsistent(fs_info);
2353                 return 0;
2354         }
2355
2356         if (!extent_buffer_uptodate(root_eb)) {
2357                 struct btrfs_tree_parent_check check = {
2358                         .has_first_key = false,
2359                         .transid = root_gen,
2360                         .level = root_level
2361                 };
2362
2363                 ret = btrfs_read_extent_buffer(root_eb, &check);
2364                 if (ret)
2365                         goto out;
2366         }
2367
2368         if (root_level == 0) {
2369                 ret = btrfs_qgroup_trace_leaf_items(trans, root_eb);
2370                 goto out;
2371         }
2372
2373         path = btrfs_alloc_path();
2374         if (!path)
2375                 return -ENOMEM;
2376
2377         /*
2378          * Walk down the tree.  Missing extent blocks are filled in as
2379          * we go. Metadata is accounted every time we read a new
2380          * extent block.
2381          *
2382          * When we reach a leaf, we account for file extent items in it,
2383          * walk back up the tree (adjusting slot pointers as we go)
2384          * and restart the search process.
2385          */
2386         atomic_inc(&root_eb->refs);     /* For path */
2387         path->nodes[root_level] = root_eb;
2388         path->slots[root_level] = 0;
2389         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
2390 walk_down:
2391         level = root_level;
2392         while (level >= 0) {
2393                 if (path->nodes[level] == NULL) {
2394                         int parent_slot;
2395                         u64 child_bytenr;
2396
2397                         /*
2398                          * We need to get child blockptr from parent before we
2399                          * can read it.
2400                           */
2401                         eb = path->nodes[level + 1];
2402                         parent_slot = path->slots[level + 1];
2403                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
2404
2405                         eb = btrfs_read_node_slot(eb, parent_slot);
2406                         if (IS_ERR(eb)) {
2407                                 ret = PTR_ERR(eb);
2408                                 goto out;
2409                         }
2410
2411                         path->nodes[level] = eb;
2412                         path->slots[level] = 0;
2413
2414                         btrfs_tree_read_lock(eb);
2415                         path->locks[level] = BTRFS_READ_LOCK;
2416
2417                         ret = btrfs_qgroup_trace_extent(trans, child_bytenr,
2418                                                         fs_info->nodesize);
2419                         if (ret)
2420                                 goto out;
2421                 }
2422
2423                 if (level == 0) {
2424                         ret = btrfs_qgroup_trace_leaf_items(trans,
2425                                                             path->nodes[level]);
2426                         if (ret)
2427                                 goto out;
2428
2429                         /* Nonzero return here means we completed our search */
2430                         ret = adjust_slots_upwards(path, root_level);
2431                         if (ret)
2432                                 break;
2433
2434                         /* Restart search with new slots */
2435                         goto walk_down;
2436                 }
2437
2438                 level--;
2439         }
2440
2441         ret = 0;
2442 out:
2443         btrfs_free_path(path);
2444
2445         return ret;
2446 }
2447
2448 #define UPDATE_NEW      0
2449 #define UPDATE_OLD      1
2450 /*
2451  * Walk all of the roots that points to the bytenr and adjust their refcnts.
2452  */
2453 static int qgroup_update_refcnt(struct btrfs_fs_info *fs_info,
2454                                 struct ulist *roots, struct ulist *tmp,
2455                                 struct ulist *qgroups, u64 seq, int update_old)
2456 {
2457         struct ulist_node *unode;
2458         struct ulist_iterator uiter;
2459         struct ulist_node *tmp_unode;
2460         struct ulist_iterator tmp_uiter;
2461         struct btrfs_qgroup *qg;
2462         int ret = 0;
2463
2464         if (!roots)
2465                 return 0;
2466         ULIST_ITER_INIT(&uiter);
2467         while ((unode = ulist_next(roots, &uiter))) {
2468                 qg = find_qgroup_rb(fs_info, unode->val);
2469                 if (!qg)
2470                         continue;
2471
2472                 ulist_reinit(tmp);
2473                 ret = ulist_add(qgroups, qg->qgroupid, qgroup_to_aux(qg),
2474                                 GFP_ATOMIC);
2475                 if (ret < 0)
2476                         return ret;
2477                 ret = ulist_add(tmp, qg->qgroupid, qgroup_to_aux(qg), GFP_ATOMIC);
2478                 if (ret < 0)
2479                         return ret;
2480                 ULIST_ITER_INIT(&tmp_uiter);
2481                 while ((tmp_unode = ulist_next(tmp, &tmp_uiter))) {
2482                         struct btrfs_qgroup_list *glist;
2483
2484                         qg = unode_aux_to_qgroup(tmp_unode);
2485                         if (update_old)
2486                                 btrfs_qgroup_update_old_refcnt(qg, seq, 1);
2487                         else
2488                                 btrfs_qgroup_update_new_refcnt(qg, seq, 1);
2489                         list_for_each_entry(glist, &qg->groups, next_group) {
2490                                 ret = ulist_add(qgroups, glist->group->qgroupid,
2491                                                 qgroup_to_aux(glist->group),
2492                                                 GFP_ATOMIC);
2493                                 if (ret < 0)
2494                                         return ret;
2495                                 ret = ulist_add(tmp, glist->group->qgroupid,
2496                                                 qgroup_to_aux(glist->group),
2497                                                 GFP_ATOMIC);
2498                                 if (ret < 0)
2499                                         return ret;
2500                         }
2501                 }
2502         }
2503         return 0;
2504 }
2505
2506 /*
2507  * Update qgroup rfer/excl counters.
2508  * Rfer update is easy, codes can explain themselves.
2509  *
2510  * Excl update is tricky, the update is split into 2 parts.
2511  * Part 1: Possible exclusive <-> sharing detect:
2512  *      |       A       |       !A      |
2513  *  -------------------------------------
2514  *  B   |       *       |       -       |
2515  *  -------------------------------------
2516  *  !B  |       +       |       **      |
2517  *  -------------------------------------
2518  *
2519  * Conditions:
2520  * A:   cur_old_roots < nr_old_roots    (not exclusive before)
2521  * !A:  cur_old_roots == nr_old_roots   (possible exclusive before)
2522  * B:   cur_new_roots < nr_new_roots    (not exclusive now)
2523  * !B:  cur_new_roots == nr_new_roots   (possible exclusive now)
2524  *
2525  * Results:
2526  * +: Possible sharing -> exclusive     -: Possible exclusive -> sharing
2527  * *: Definitely not changed.           **: Possible unchanged.
2528  *
2529  * For !A and !B condition, the exception is cur_old/new_roots == 0 case.
2530  *
2531  * To make the logic clear, we first use condition A and B to split
2532  * combination into 4 results.
2533  *
2534  * Then, for result "+" and "-", check old/new_roots == 0 case, as in them
2535  * only on variant maybe 0.
2536  *
2537  * Lastly, check result **, since there are 2 variants maybe 0, split them
2538  * again(2x2).
2539  * But this time we don't need to consider other things, the codes and logic
2540  * is easy to understand now.
2541  */
2542 static int qgroup_update_counters(struct btrfs_fs_info *fs_info,
2543                                   struct ulist *qgroups,
2544                                   u64 nr_old_roots,
2545                                   u64 nr_new_roots,
2546                                   u64 num_bytes, u64 seq)
2547 {
2548         struct ulist_node *unode;
2549         struct ulist_iterator uiter;
2550         struct btrfs_qgroup *qg;
2551         u64 cur_new_count, cur_old_count;
2552
2553         ULIST_ITER_INIT(&uiter);
2554         while ((unode = ulist_next(qgroups, &uiter))) {
2555                 bool dirty = false;
2556
2557                 qg = unode_aux_to_qgroup(unode);
2558                 cur_old_count = btrfs_qgroup_get_old_refcnt(qg, seq);
2559                 cur_new_count = btrfs_qgroup_get_new_refcnt(qg, seq);
2560
2561                 trace_qgroup_update_counters(fs_info, qg, cur_old_count,
2562                                              cur_new_count);
2563
2564                 /* Rfer update part */
2565                 if (cur_old_count == 0 && cur_new_count > 0) {
2566                         qg->rfer += num_bytes;
2567                         qg->rfer_cmpr += num_bytes;
2568                         dirty = true;
2569                 }
2570                 if (cur_old_count > 0 && cur_new_count == 0) {
2571                         qg->rfer -= num_bytes;
2572                         qg->rfer_cmpr -= num_bytes;
2573                         dirty = true;
2574                 }
2575
2576                 /* Excl update part */
2577                 /* Exclusive/none -> shared case */
2578                 if (cur_old_count == nr_old_roots &&
2579                     cur_new_count < nr_new_roots) {
2580                         /* Exclusive -> shared */
2581                         if (cur_old_count != 0) {
2582                                 qg->excl -= num_bytes;
2583                                 qg->excl_cmpr -= num_bytes;
2584                                 dirty = true;
2585                         }
2586                 }
2587
2588                 /* Shared -> exclusive/none case */
2589                 if (cur_old_count < nr_old_roots &&
2590                     cur_new_count == nr_new_roots) {
2591                         /* Shared->exclusive */
2592                         if (cur_new_count != 0) {
2593                                 qg->excl += num_bytes;
2594                                 qg->excl_cmpr += num_bytes;
2595                                 dirty = true;
2596                         }
2597                 }
2598
2599                 /* Exclusive/none -> exclusive/none case */
2600                 if (cur_old_count == nr_old_roots &&
2601                     cur_new_count == nr_new_roots) {
2602                         if (cur_old_count == 0) {
2603                                 /* None -> exclusive/none */
2604
2605                                 if (cur_new_count != 0) {
2606                                         /* None -> exclusive */
2607                                         qg->excl += num_bytes;
2608                                         qg->excl_cmpr += num_bytes;
2609                                         dirty = true;
2610                                 }
2611                                 /* None -> none, nothing changed */
2612                         } else {
2613                                 /* Exclusive -> exclusive/none */
2614
2615                                 if (cur_new_count == 0) {
2616                                         /* Exclusive -> none */
2617                                         qg->excl -= num_bytes;
2618                                         qg->excl_cmpr -= num_bytes;
2619                                         dirty = true;
2620                                 }
2621                                 /* Exclusive -> exclusive, nothing changed */
2622                         }
2623                 }
2624
2625                 if (dirty)
2626                         qgroup_dirty(fs_info, qg);
2627         }
2628         return 0;
2629 }
2630
2631 /*
2632  * Check if the @roots potentially is a list of fs tree roots
2633  *
2634  * Return 0 for definitely not a fs/subvol tree roots ulist
2635  * Return 1 for possible fs/subvol tree roots in the list (considering an empty
2636  *          one as well)
2637  */
2638 static int maybe_fs_roots(struct ulist *roots)
2639 {
2640         struct ulist_node *unode;
2641         struct ulist_iterator uiter;
2642
2643         /* Empty one, still possible for fs roots */
2644         if (!roots || roots->nnodes == 0)
2645                 return 1;
2646
2647         ULIST_ITER_INIT(&uiter);
2648         unode = ulist_next(roots, &uiter);
2649         if (!unode)
2650                 return 1;
2651
2652         /*
2653          * If it contains fs tree roots, then it must belong to fs/subvol
2654          * trees.
2655          * If it contains a non-fs tree, it won't be shared with fs/subvol trees.
2656          */
2657         return is_fstree(unode->val);
2658 }
2659
2660 int btrfs_qgroup_account_extent(struct btrfs_trans_handle *trans, u64 bytenr,
2661                                 u64 num_bytes, struct ulist *old_roots,
2662                                 struct ulist *new_roots)
2663 {
2664         struct btrfs_fs_info *fs_info = trans->fs_info;
2665         struct ulist *qgroups = NULL;
2666         struct ulist *tmp = NULL;
2667         u64 seq;
2668         u64 nr_new_roots = 0;
2669         u64 nr_old_roots = 0;
2670         int ret = 0;
2671
2672         /*
2673          * If quotas get disabled meanwhile, the resources need to be freed and
2674          * we can't just exit here.
2675          */
2676         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
2677             fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING)
2678                 goto out_free;
2679
2680         if (new_roots) {
2681                 if (!maybe_fs_roots(new_roots))
2682                         goto out_free;
2683                 nr_new_roots = new_roots->nnodes;
2684         }
2685         if (old_roots) {
2686                 if (!maybe_fs_roots(old_roots))
2687                         goto out_free;
2688                 nr_old_roots = old_roots->nnodes;
2689         }
2690
2691         /* Quick exit, either not fs tree roots, or won't affect any qgroup */
2692         if (nr_old_roots == 0 && nr_new_roots == 0)
2693                 goto out_free;
2694
2695         BUG_ON(!fs_info->quota_root);
2696
2697         trace_btrfs_qgroup_account_extent(fs_info, trans->transid, bytenr,
2698                                         num_bytes, nr_old_roots, nr_new_roots);
2699
2700         qgroups = ulist_alloc(GFP_NOFS);
2701         if (!qgroups) {
2702                 ret = -ENOMEM;
2703                 goto out_free;
2704         }
2705         tmp = ulist_alloc(GFP_NOFS);
2706         if (!tmp) {
2707                 ret = -ENOMEM;
2708                 goto out_free;
2709         }
2710
2711         mutex_lock(&fs_info->qgroup_rescan_lock);
2712         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
2713                 if (fs_info->qgroup_rescan_progress.objectid <= bytenr) {
2714                         mutex_unlock(&fs_info->qgroup_rescan_lock);
2715                         ret = 0;
2716                         goto out_free;
2717                 }
2718         }
2719         mutex_unlock(&fs_info->qgroup_rescan_lock);
2720
2721         spin_lock(&fs_info->qgroup_lock);
2722         seq = fs_info->qgroup_seq;
2723
2724         /* Update old refcnts using old_roots */
2725         ret = qgroup_update_refcnt(fs_info, old_roots, tmp, qgroups, seq,
2726                                    UPDATE_OLD);
2727         if (ret < 0)
2728                 goto out;
2729
2730         /* Update new refcnts using new_roots */
2731         ret = qgroup_update_refcnt(fs_info, new_roots, tmp, qgroups, seq,
2732                                    UPDATE_NEW);
2733         if (ret < 0)
2734                 goto out;
2735
2736         qgroup_update_counters(fs_info, qgroups, nr_old_roots, nr_new_roots,
2737                                num_bytes, seq);
2738
2739         /*
2740          * Bump qgroup_seq to avoid seq overlap
2741          */
2742         fs_info->qgroup_seq += max(nr_old_roots, nr_new_roots) + 1;
2743 out:
2744         spin_unlock(&fs_info->qgroup_lock);
2745 out_free:
2746         ulist_free(tmp);
2747         ulist_free(qgroups);
2748         ulist_free(old_roots);
2749         ulist_free(new_roots);
2750         return ret;
2751 }
2752
2753 int btrfs_qgroup_account_extents(struct btrfs_trans_handle *trans)
2754 {
2755         struct btrfs_fs_info *fs_info = trans->fs_info;
2756         struct btrfs_qgroup_extent_record *record;
2757         struct btrfs_delayed_ref_root *delayed_refs;
2758         struct ulist *new_roots = NULL;
2759         struct rb_node *node;
2760         u64 num_dirty_extents = 0;
2761         u64 qgroup_to_skip;
2762         int ret = 0;
2763
2764         delayed_refs = &trans->transaction->delayed_refs;
2765         qgroup_to_skip = delayed_refs->qgroup_to_skip;
2766         while ((node = rb_first(&delayed_refs->dirty_extent_root))) {
2767                 record = rb_entry(node, struct btrfs_qgroup_extent_record,
2768                                   node);
2769
2770                 num_dirty_extents++;
2771                 trace_btrfs_qgroup_account_extents(fs_info, record);
2772
2773                 if (!ret && !(fs_info->qgroup_flags &
2774                               BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING)) {
2775                         struct btrfs_backref_walk_ctx ctx = { 0 };
2776
2777                         ctx.bytenr = record->bytenr;
2778                         ctx.fs_info = fs_info;
2779
2780                         /*
2781                          * Old roots should be searched when inserting qgroup
2782                          * extent record.
2783                          *
2784                          * But for INCONSISTENT (NO_ACCOUNTING) -> rescan case,
2785                          * we may have some record inserted during
2786                          * NO_ACCOUNTING (thus no old_roots populated), but
2787                          * later we start rescan, which clears NO_ACCOUNTING,
2788                          * leaving some inserted records without old_roots
2789                          * populated.
2790                          *
2791                          * Those cases are rare and should not cause too much
2792                          * time spent during commit_transaction().
2793                          */
2794                         if (!record->old_roots) {
2795                                 /* Search commit root to find old_roots */
2796                                 ret = btrfs_find_all_roots(&ctx, false);
2797                                 if (ret < 0)
2798                                         goto cleanup;
2799                                 record->old_roots = ctx.roots;
2800                                 ctx.roots = NULL;
2801                         }
2802
2803                         /* Free the reserved data space */
2804                         btrfs_qgroup_free_refroot(fs_info,
2805                                         record->data_rsv_refroot,
2806                                         record->data_rsv,
2807                                         BTRFS_QGROUP_RSV_DATA);
2808                         /*
2809                          * Use BTRFS_SEQ_LAST as time_seq to do special search,
2810                          * which doesn't lock tree or delayed_refs and search
2811                          * current root. It's safe inside commit_transaction().
2812                          */
2813                         ctx.trans = trans;
2814                         ctx.time_seq = BTRFS_SEQ_LAST;
2815                         ret = btrfs_find_all_roots(&ctx, false);
2816                         if (ret < 0)
2817                                 goto cleanup;
2818                         new_roots = ctx.roots;
2819                         if (qgroup_to_skip) {
2820                                 ulist_del(new_roots, qgroup_to_skip, 0);
2821                                 ulist_del(record->old_roots, qgroup_to_skip,
2822                                           0);
2823                         }
2824                         ret = btrfs_qgroup_account_extent(trans, record->bytenr,
2825                                                           record->num_bytes,
2826                                                           record->old_roots,
2827                                                           new_roots);
2828                         record->old_roots = NULL;
2829                         new_roots = NULL;
2830                 }
2831 cleanup:
2832                 ulist_free(record->old_roots);
2833                 ulist_free(new_roots);
2834                 new_roots = NULL;
2835                 rb_erase(node, &delayed_refs->dirty_extent_root);
2836                 kfree(record);
2837
2838         }
2839         trace_qgroup_num_dirty_extents(fs_info, trans->transid,
2840                                        num_dirty_extents);
2841         return ret;
2842 }
2843
2844 /*
2845  * Writes all changed qgroups to disk.
2846  * Called by the transaction commit path and the qgroup assign ioctl.
2847  */
2848 int btrfs_run_qgroups(struct btrfs_trans_handle *trans)
2849 {
2850         struct btrfs_fs_info *fs_info = trans->fs_info;
2851         int ret = 0;
2852
2853         /*
2854          * In case we are called from the qgroup assign ioctl, assert that we
2855          * are holding the qgroup_ioctl_lock, otherwise we can race with a quota
2856          * disable operation (ioctl) and access a freed quota root.
2857          */
2858         if (trans->transaction->state != TRANS_STATE_COMMIT_DOING)
2859                 lockdep_assert_held(&fs_info->qgroup_ioctl_lock);
2860
2861         if (!fs_info->quota_root)
2862                 return ret;
2863
2864         spin_lock(&fs_info->qgroup_lock);
2865         while (!list_empty(&fs_info->dirty_qgroups)) {
2866                 struct btrfs_qgroup *qgroup;
2867                 qgroup = list_first_entry(&fs_info->dirty_qgroups,
2868                                           struct btrfs_qgroup, dirty);
2869                 list_del_init(&qgroup->dirty);
2870                 spin_unlock(&fs_info->qgroup_lock);
2871                 ret = update_qgroup_info_item(trans, qgroup);
2872                 if (ret)
2873                         qgroup_mark_inconsistent(fs_info);
2874                 ret = update_qgroup_limit_item(trans, qgroup);
2875                 if (ret)
2876                         qgroup_mark_inconsistent(fs_info);
2877                 spin_lock(&fs_info->qgroup_lock);
2878         }
2879         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2880                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_ON;
2881         else
2882                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_ON;
2883         spin_unlock(&fs_info->qgroup_lock);
2884
2885         ret = update_qgroup_status_item(trans);
2886         if (ret)
2887                 qgroup_mark_inconsistent(fs_info);
2888
2889         return ret;
2890 }
2891
2892 /*
2893  * Copy the accounting information between qgroups. This is necessary
2894  * when a snapshot or a subvolume is created. Throwing an error will
2895  * cause a transaction abort so we take extra care here to only error
2896  * when a readonly fs is a reasonable outcome.
2897  */
2898 int btrfs_qgroup_inherit(struct btrfs_trans_handle *trans, u64 srcid,
2899                          u64 objectid, struct btrfs_qgroup_inherit *inherit)
2900 {
2901         int ret = 0;
2902         int i;
2903         u64 *i_qgroups;
2904         bool committing = false;
2905         struct btrfs_fs_info *fs_info = trans->fs_info;
2906         struct btrfs_root *quota_root;
2907         struct btrfs_qgroup *srcgroup;
2908         struct btrfs_qgroup *dstgroup;
2909         bool need_rescan = false;
2910         u32 level_size = 0;
2911         u64 nums;
2912
2913         /*
2914          * There are only two callers of this function.
2915          *
2916          * One in create_subvol() in the ioctl context, which needs to hold
2917          * the qgroup_ioctl_lock.
2918          *
2919          * The other one in create_pending_snapshot() where no other qgroup
2920          * code can modify the fs as they all need to either start a new trans
2921          * or hold a trans handler, thus we don't need to hold
2922          * qgroup_ioctl_lock.
2923          * This would avoid long and complex lock chain and make lockdep happy.
2924          */
2925         spin_lock(&fs_info->trans_lock);
2926         if (trans->transaction->state == TRANS_STATE_COMMIT_DOING)
2927                 committing = true;
2928         spin_unlock(&fs_info->trans_lock);
2929
2930         if (!committing)
2931                 mutex_lock(&fs_info->qgroup_ioctl_lock);
2932         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2933                 goto out;
2934
2935         quota_root = fs_info->quota_root;
2936         if (!quota_root) {
2937                 ret = -EINVAL;
2938                 goto out;
2939         }
2940
2941         if (inherit) {
2942                 i_qgroups = (u64 *)(inherit + 1);
2943                 nums = inherit->num_qgroups + 2 * inherit->num_ref_copies +
2944                        2 * inherit->num_excl_copies;
2945                 for (i = 0; i < nums; ++i) {
2946                         srcgroup = find_qgroup_rb(fs_info, *i_qgroups);
2947
2948                         /*
2949                          * Zero out invalid groups so we can ignore
2950                          * them later.
2951                          */
2952                         if (!srcgroup ||
2953                             ((srcgroup->qgroupid >> 48) <= (objectid >> 48)))
2954                                 *i_qgroups = 0ULL;
2955
2956                         ++i_qgroups;
2957                 }
2958         }
2959
2960         /*
2961          * create a tracking group for the subvol itself
2962          */
2963         ret = add_qgroup_item(trans, quota_root, objectid);
2964         if (ret)
2965                 goto out;
2966
2967         /*
2968          * add qgroup to all inherited groups
2969          */
2970         if (inherit) {
2971                 i_qgroups = (u64 *)(inherit + 1);
2972                 for (i = 0; i < inherit->num_qgroups; ++i, ++i_qgroups) {
2973                         if (*i_qgroups == 0)
2974                                 continue;
2975                         ret = add_qgroup_relation_item(trans, objectid,
2976                                                        *i_qgroups);
2977                         if (ret && ret != -EEXIST)
2978                                 goto out;
2979                         ret = add_qgroup_relation_item(trans, *i_qgroups,
2980                                                        objectid);
2981                         if (ret && ret != -EEXIST)
2982                                 goto out;
2983                 }
2984                 ret = 0;
2985         }
2986
2987
2988         spin_lock(&fs_info->qgroup_lock);
2989
2990         dstgroup = add_qgroup_rb(fs_info, objectid);
2991         if (IS_ERR(dstgroup)) {
2992                 ret = PTR_ERR(dstgroup);
2993                 goto unlock;
2994         }
2995
2996         if (inherit && inherit->flags & BTRFS_QGROUP_INHERIT_SET_LIMITS) {
2997                 dstgroup->lim_flags = inherit->lim.flags;
2998                 dstgroup->max_rfer = inherit->lim.max_rfer;
2999                 dstgroup->max_excl = inherit->lim.max_excl;
3000                 dstgroup->rsv_rfer = inherit->lim.rsv_rfer;
3001                 dstgroup->rsv_excl = inherit->lim.rsv_excl;
3002
3003                 qgroup_dirty(fs_info, dstgroup);
3004         }
3005
3006         if (srcid) {
3007                 srcgroup = find_qgroup_rb(fs_info, srcid);
3008                 if (!srcgroup)
3009                         goto unlock;
3010
3011                 /*
3012                  * We call inherit after we clone the root in order to make sure
3013                  * our counts don't go crazy, so at this point the only
3014                  * difference between the two roots should be the root node.
3015                  */
3016                 level_size = fs_info->nodesize;
3017                 dstgroup->rfer = srcgroup->rfer;
3018                 dstgroup->rfer_cmpr = srcgroup->rfer_cmpr;
3019                 dstgroup->excl = level_size;
3020                 dstgroup->excl_cmpr = level_size;
3021                 srcgroup->excl = level_size;
3022                 srcgroup->excl_cmpr = level_size;
3023
3024                 /* inherit the limit info */
3025                 dstgroup->lim_flags = srcgroup->lim_flags;
3026                 dstgroup->max_rfer = srcgroup->max_rfer;
3027                 dstgroup->max_excl = srcgroup->max_excl;
3028                 dstgroup->rsv_rfer = srcgroup->rsv_rfer;
3029                 dstgroup->rsv_excl = srcgroup->rsv_excl;
3030
3031                 qgroup_dirty(fs_info, dstgroup);
3032                 qgroup_dirty(fs_info, srcgroup);
3033         }
3034
3035         if (!inherit)
3036                 goto unlock;
3037
3038         i_qgroups = (u64 *)(inherit + 1);
3039         for (i = 0; i < inherit->num_qgroups; ++i) {
3040                 if (*i_qgroups) {
3041                         ret = add_relation_rb(fs_info, objectid, *i_qgroups);
3042                         if (ret)
3043                                 goto unlock;
3044                 }
3045                 ++i_qgroups;
3046
3047                 /*
3048                  * If we're doing a snapshot, and adding the snapshot to a new
3049                  * qgroup, the numbers are guaranteed to be incorrect.
3050                  */
3051                 if (srcid)
3052                         need_rescan = true;
3053         }
3054
3055         for (i = 0; i <  inherit->num_ref_copies; ++i, i_qgroups += 2) {
3056                 struct btrfs_qgroup *src;
3057                 struct btrfs_qgroup *dst;
3058
3059                 if (!i_qgroups[0] || !i_qgroups[1])
3060                         continue;
3061
3062                 src = find_qgroup_rb(fs_info, i_qgroups[0]);
3063                 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
3064
3065                 if (!src || !dst) {
3066                         ret = -EINVAL;
3067                         goto unlock;
3068                 }
3069
3070                 dst->rfer = src->rfer - level_size;
3071                 dst->rfer_cmpr = src->rfer_cmpr - level_size;
3072
3073                 /* Manually tweaking numbers certainly needs a rescan */
3074                 need_rescan = true;
3075         }
3076         for (i = 0; i <  inherit->num_excl_copies; ++i, i_qgroups += 2) {
3077                 struct btrfs_qgroup *src;
3078                 struct btrfs_qgroup *dst;
3079
3080                 if (!i_qgroups[0] || !i_qgroups[1])
3081                         continue;
3082
3083                 src = find_qgroup_rb(fs_info, i_qgroups[0]);
3084                 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
3085
3086                 if (!src || !dst) {
3087                         ret = -EINVAL;
3088                         goto unlock;
3089                 }
3090
3091                 dst->excl = src->excl + level_size;
3092                 dst->excl_cmpr = src->excl_cmpr + level_size;
3093                 need_rescan = true;
3094         }
3095
3096 unlock:
3097         spin_unlock(&fs_info->qgroup_lock);
3098         if (!ret)
3099                 ret = btrfs_sysfs_add_one_qgroup(fs_info, dstgroup);
3100 out:
3101         if (!committing)
3102                 mutex_unlock(&fs_info->qgroup_ioctl_lock);
3103         if (need_rescan)
3104                 qgroup_mark_inconsistent(fs_info);
3105         return ret;
3106 }
3107
3108 static bool qgroup_check_limits(const struct btrfs_qgroup *qg, u64 num_bytes)
3109 {
3110         if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_RFER) &&
3111             qgroup_rsv_total(qg) + (s64)qg->rfer + num_bytes > qg->max_rfer)
3112                 return false;
3113
3114         if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) &&
3115             qgroup_rsv_total(qg) + (s64)qg->excl + num_bytes > qg->max_excl)
3116                 return false;
3117
3118         return true;
3119 }
3120
3121 static int qgroup_reserve(struct btrfs_root *root, u64 num_bytes, bool enforce,
3122                           enum btrfs_qgroup_rsv_type type)
3123 {
3124         struct btrfs_qgroup *qgroup;
3125         struct btrfs_fs_info *fs_info = root->fs_info;
3126         u64 ref_root = root->root_key.objectid;
3127         int ret = 0;
3128         struct ulist_node *unode;
3129         struct ulist_iterator uiter;
3130
3131         if (!is_fstree(ref_root))
3132                 return 0;
3133
3134         if (num_bytes == 0)
3135                 return 0;
3136
3137         if (test_bit(BTRFS_FS_QUOTA_OVERRIDE, &fs_info->flags) &&
3138             capable(CAP_SYS_RESOURCE))
3139                 enforce = false;
3140
3141         spin_lock(&fs_info->qgroup_lock);
3142         if (!fs_info->quota_root)
3143                 goto out;
3144
3145         qgroup = find_qgroup_rb(fs_info, ref_root);
3146         if (!qgroup)
3147                 goto out;
3148
3149         /*
3150          * in a first step, we check all affected qgroups if any limits would
3151          * be exceeded
3152          */
3153         ulist_reinit(fs_info->qgroup_ulist);
3154         ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
3155                         qgroup_to_aux(qgroup), GFP_ATOMIC);
3156         if (ret < 0)
3157                 goto out;
3158         ULIST_ITER_INIT(&uiter);
3159         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
3160                 struct btrfs_qgroup *qg;
3161                 struct btrfs_qgroup_list *glist;
3162
3163                 qg = unode_aux_to_qgroup(unode);
3164
3165                 if (enforce && !qgroup_check_limits(qg, num_bytes)) {
3166                         ret = -EDQUOT;
3167                         goto out;
3168                 }
3169
3170                 list_for_each_entry(glist, &qg->groups, next_group) {
3171                         ret = ulist_add(fs_info->qgroup_ulist,
3172                                         glist->group->qgroupid,
3173                                         qgroup_to_aux(glist->group), GFP_ATOMIC);
3174                         if (ret < 0)
3175                                 goto out;
3176                 }
3177         }
3178         ret = 0;
3179         /*
3180          * no limits exceeded, now record the reservation into all qgroups
3181          */
3182         ULIST_ITER_INIT(&uiter);
3183         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
3184                 struct btrfs_qgroup *qg;
3185
3186                 qg = unode_aux_to_qgroup(unode);
3187
3188                 qgroup_rsv_add(fs_info, qg, num_bytes, type);
3189         }
3190
3191 out:
3192         spin_unlock(&fs_info->qgroup_lock);
3193         return ret;
3194 }
3195
3196 /*
3197  * Free @num_bytes of reserved space with @type for qgroup.  (Normally level 0
3198  * qgroup).
3199  *
3200  * Will handle all higher level qgroup too.
3201  *
3202  * NOTE: If @num_bytes is (u64)-1, this means to free all bytes of this qgroup.
3203  * This special case is only used for META_PERTRANS type.
3204  */
3205 void btrfs_qgroup_free_refroot(struct btrfs_fs_info *fs_info,
3206                                u64 ref_root, u64 num_bytes,
3207                                enum btrfs_qgroup_rsv_type type)
3208 {
3209         struct btrfs_qgroup *qgroup;
3210         struct ulist_node *unode;
3211         struct ulist_iterator uiter;
3212         int ret = 0;
3213
3214         if (!is_fstree(ref_root))
3215                 return;
3216
3217         if (num_bytes == 0)
3218                 return;
3219
3220         if (num_bytes == (u64)-1 && type != BTRFS_QGROUP_RSV_META_PERTRANS) {
3221                 WARN(1, "%s: Invalid type to free", __func__);
3222                 return;
3223         }
3224         spin_lock(&fs_info->qgroup_lock);
3225
3226         if (!fs_info->quota_root)
3227                 goto out;
3228
3229         qgroup = find_qgroup_rb(fs_info, ref_root);
3230         if (!qgroup)
3231                 goto out;
3232
3233         if (num_bytes == (u64)-1)
3234                 /*
3235                  * We're freeing all pertrans rsv, get reserved value from
3236                  * level 0 qgroup as real num_bytes to free.
3237                  */
3238                 num_bytes = qgroup->rsv.values[type];
3239
3240         ulist_reinit(fs_info->qgroup_ulist);
3241         ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
3242                         qgroup_to_aux(qgroup), GFP_ATOMIC);
3243         if (ret < 0)
3244                 goto out;
3245         ULIST_ITER_INIT(&uiter);
3246         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
3247                 struct btrfs_qgroup *qg;
3248                 struct btrfs_qgroup_list *glist;
3249
3250                 qg = unode_aux_to_qgroup(unode);
3251
3252                 qgroup_rsv_release(fs_info, qg, num_bytes, type);
3253
3254                 list_for_each_entry(glist, &qg->groups, next_group) {
3255                         ret = ulist_add(fs_info->qgroup_ulist,
3256                                         glist->group->qgroupid,
3257                                         qgroup_to_aux(glist->group), GFP_ATOMIC);
3258                         if (ret < 0)
3259                                 goto out;
3260                 }
3261         }
3262
3263 out:
3264         spin_unlock(&fs_info->qgroup_lock);
3265 }
3266
3267 /*
3268  * Check if the leaf is the last leaf. Which means all node pointers
3269  * are at their last position.
3270  */
3271 static bool is_last_leaf(struct btrfs_path *path)
3272 {
3273         int i;
3274
3275         for (i = 1; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
3276                 if (path->slots[i] != btrfs_header_nritems(path->nodes[i]) - 1)
3277                         return false;
3278         }
3279         return true;
3280 }
3281
3282 /*
3283  * returns < 0 on error, 0 when more leafs are to be scanned.
3284  * returns 1 when done.
3285  */
3286 static int qgroup_rescan_leaf(struct btrfs_trans_handle *trans,
3287                               struct btrfs_path *path)
3288 {
3289         struct btrfs_fs_info *fs_info = trans->fs_info;
3290         struct btrfs_root *extent_root;
3291         struct btrfs_key found;
3292         struct extent_buffer *scratch_leaf = NULL;
3293         u64 num_bytes;
3294         bool done;
3295         int slot;
3296         int ret;
3297
3298         mutex_lock(&fs_info->qgroup_rescan_lock);
3299         extent_root = btrfs_extent_root(fs_info,
3300                                 fs_info->qgroup_rescan_progress.objectid);
3301         ret = btrfs_search_slot_for_read(extent_root,
3302                                          &fs_info->qgroup_rescan_progress,
3303                                          path, 1, 0);
3304
3305         btrfs_debug(fs_info,
3306                 "current progress key (%llu %u %llu), search_slot ret %d",
3307                 fs_info->qgroup_rescan_progress.objectid,
3308                 fs_info->qgroup_rescan_progress.type,
3309                 fs_info->qgroup_rescan_progress.offset, ret);
3310
3311         if (ret) {
3312                 /*
3313                  * The rescan is about to end, we will not be scanning any
3314                  * further blocks. We cannot unset the RESCAN flag here, because
3315                  * we want to commit the transaction if everything went well.
3316                  * To make the live accounting work in this phase, we set our
3317                  * scan progress pointer such that every real extent objectid
3318                  * will be smaller.
3319                  */
3320                 fs_info->qgroup_rescan_progress.objectid = (u64)-1;
3321                 btrfs_release_path(path);
3322                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3323                 return ret;
3324         }
3325         done = is_last_leaf(path);
3326
3327         btrfs_item_key_to_cpu(path->nodes[0], &found,
3328                               btrfs_header_nritems(path->nodes[0]) - 1);
3329         fs_info->qgroup_rescan_progress.objectid = found.objectid + 1;
3330
3331         scratch_leaf = btrfs_clone_extent_buffer(path->nodes[0]);
3332         if (!scratch_leaf) {
3333                 ret = -ENOMEM;
3334                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3335                 goto out;
3336         }
3337         slot = path->slots[0];
3338         btrfs_release_path(path);
3339         mutex_unlock(&fs_info->qgroup_rescan_lock);
3340
3341         for (; slot < btrfs_header_nritems(scratch_leaf); ++slot) {
3342                 struct btrfs_backref_walk_ctx ctx = { 0 };
3343
3344                 btrfs_item_key_to_cpu(scratch_leaf, &found, slot);
3345                 if (found.type != BTRFS_EXTENT_ITEM_KEY &&
3346                     found.type != BTRFS_METADATA_ITEM_KEY)
3347                         continue;
3348                 if (found.type == BTRFS_METADATA_ITEM_KEY)
3349                         num_bytes = fs_info->nodesize;
3350                 else
3351                         num_bytes = found.offset;
3352
3353                 ctx.bytenr = found.objectid;
3354                 ctx.fs_info = fs_info;
3355
3356                 ret = btrfs_find_all_roots(&ctx, false);
3357                 if (ret < 0)
3358                         goto out;
3359                 /* For rescan, just pass old_roots as NULL */
3360                 ret = btrfs_qgroup_account_extent(trans, found.objectid,
3361                                                   num_bytes, NULL, ctx.roots);
3362                 if (ret < 0)
3363                         goto out;
3364         }
3365 out:
3366         if (scratch_leaf)
3367                 free_extent_buffer(scratch_leaf);
3368
3369         if (done && !ret) {
3370                 ret = 1;
3371                 fs_info->qgroup_rescan_progress.objectid = (u64)-1;
3372         }
3373         return ret;
3374 }
3375
3376 static bool rescan_should_stop(struct btrfs_fs_info *fs_info)
3377 {
3378         return btrfs_fs_closing(fs_info) ||
3379                 test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state) ||
3380                 !test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
3381                           fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN;
3382 }
3383
3384 static void btrfs_qgroup_rescan_worker(struct btrfs_work *work)
3385 {
3386         struct btrfs_fs_info *fs_info = container_of(work, struct btrfs_fs_info,
3387                                                      qgroup_rescan_work);
3388         struct btrfs_path *path;
3389         struct btrfs_trans_handle *trans = NULL;
3390         int err = -ENOMEM;
3391         int ret = 0;
3392         bool stopped = false;
3393         bool did_leaf_rescans = false;
3394
3395         path = btrfs_alloc_path();
3396         if (!path)
3397                 goto out;
3398         /*
3399          * Rescan should only search for commit root, and any later difference
3400          * should be recorded by qgroup
3401          */
3402         path->search_commit_root = 1;
3403         path->skip_locking = 1;
3404
3405         err = 0;
3406         while (!err && !(stopped = rescan_should_stop(fs_info))) {
3407                 trans = btrfs_start_transaction(fs_info->fs_root, 0);
3408                 if (IS_ERR(trans)) {
3409                         err = PTR_ERR(trans);
3410                         break;
3411                 }
3412
3413                 err = qgroup_rescan_leaf(trans, path);
3414                 did_leaf_rescans = true;
3415
3416                 if (err > 0)
3417                         btrfs_commit_transaction(trans);
3418                 else
3419                         btrfs_end_transaction(trans);
3420         }
3421
3422 out:
3423         btrfs_free_path(path);
3424
3425         mutex_lock(&fs_info->qgroup_rescan_lock);
3426         if (err > 0 &&
3427             fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT) {
3428                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
3429         } else if (err < 0 || stopped) {
3430                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
3431         }
3432         mutex_unlock(&fs_info->qgroup_rescan_lock);
3433
3434         /*
3435          * Only update status, since the previous part has already updated the
3436          * qgroup info, and only if we did any actual work. This also prevents
3437          * race with a concurrent quota disable, which has already set
3438          * fs_info->quota_root to NULL and cleared BTRFS_FS_QUOTA_ENABLED at
3439          * btrfs_quota_disable().
3440          */
3441         if (did_leaf_rescans) {
3442                 trans = btrfs_start_transaction(fs_info->quota_root, 1);
3443                 if (IS_ERR(trans)) {
3444                         err = PTR_ERR(trans);
3445                         trans = NULL;
3446                         btrfs_err(fs_info,
3447                                   "fail to start transaction for status update: %d",
3448                                   err);
3449                 }
3450         } else {
3451                 trans = NULL;
3452         }
3453
3454         mutex_lock(&fs_info->qgroup_rescan_lock);
3455         if (!stopped ||
3456             fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN)
3457                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3458         if (trans) {
3459                 ret = update_qgroup_status_item(trans);
3460                 if (ret < 0) {
3461                         err = ret;
3462                         btrfs_err(fs_info, "fail to update qgroup status: %d",
3463                                   err);
3464                 }
3465         }
3466         fs_info->qgroup_rescan_running = false;
3467         fs_info->qgroup_flags &= ~BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN;
3468         complete_all(&fs_info->qgroup_rescan_completion);
3469         mutex_unlock(&fs_info->qgroup_rescan_lock);
3470
3471         if (!trans)
3472                 return;
3473
3474         btrfs_end_transaction(trans);
3475
3476         if (stopped) {
3477                 btrfs_info(fs_info, "qgroup scan paused");
3478         } else if (fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN) {
3479                 btrfs_info(fs_info, "qgroup scan cancelled");
3480         } else if (err >= 0) {
3481                 btrfs_info(fs_info, "qgroup scan completed%s",
3482                         err > 0 ? " (inconsistency flag cleared)" : "");
3483         } else {
3484                 btrfs_err(fs_info, "qgroup scan failed with %d", err);
3485         }
3486 }
3487
3488 /*
3489  * Checks that (a) no rescan is running and (b) quota is enabled. Allocates all
3490  * memory required for the rescan context.
3491  */
3492 static int
3493 qgroup_rescan_init(struct btrfs_fs_info *fs_info, u64 progress_objectid,
3494                    int init_flags)
3495 {
3496         int ret = 0;
3497
3498         if (!init_flags) {
3499                 /* we're resuming qgroup rescan at mount time */
3500                 if (!(fs_info->qgroup_flags &
3501                       BTRFS_QGROUP_STATUS_FLAG_RESCAN)) {
3502                         btrfs_warn(fs_info,
3503                         "qgroup rescan init failed, qgroup rescan is not queued");
3504                         ret = -EINVAL;
3505                 } else if (!(fs_info->qgroup_flags &
3506                              BTRFS_QGROUP_STATUS_FLAG_ON)) {
3507                         btrfs_warn(fs_info,
3508                         "qgroup rescan init failed, qgroup is not enabled");
3509                         ret = -EINVAL;
3510                 }
3511
3512                 if (ret)
3513                         return ret;
3514         }
3515
3516         mutex_lock(&fs_info->qgroup_rescan_lock);
3517
3518         if (init_flags) {
3519                 if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
3520                         btrfs_warn(fs_info,
3521                                    "qgroup rescan is already in progress");
3522                         ret = -EINPROGRESS;
3523                 } else if (!(fs_info->qgroup_flags &
3524                              BTRFS_QGROUP_STATUS_FLAG_ON)) {
3525                         btrfs_warn(fs_info,
3526                         "qgroup rescan init failed, qgroup is not enabled");
3527                         ret = -EINVAL;
3528                 } else if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
3529                         /* Quota disable is in progress */
3530                         ret = -EBUSY;
3531                 }
3532
3533                 if (ret) {
3534                         mutex_unlock(&fs_info->qgroup_rescan_lock);
3535                         return ret;
3536                 }
3537                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3538         }
3539
3540         memset(&fs_info->qgroup_rescan_progress, 0,
3541                 sizeof(fs_info->qgroup_rescan_progress));
3542         fs_info->qgroup_flags &= ~(BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN |
3543                                    BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING);
3544         fs_info->qgroup_rescan_progress.objectid = progress_objectid;
3545         init_completion(&fs_info->qgroup_rescan_completion);
3546         mutex_unlock(&fs_info->qgroup_rescan_lock);
3547
3548         btrfs_init_work(&fs_info->qgroup_rescan_work,
3549                         btrfs_qgroup_rescan_worker, NULL, NULL);
3550         return 0;
3551 }
3552
3553 static void
3554 qgroup_rescan_zero_tracking(struct btrfs_fs_info *fs_info)
3555 {
3556         struct rb_node *n;
3557         struct btrfs_qgroup *qgroup;
3558
3559         spin_lock(&fs_info->qgroup_lock);
3560         /* clear all current qgroup tracking information */
3561         for (n = rb_first(&fs_info->qgroup_tree); n; n = rb_next(n)) {
3562                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
3563                 qgroup->rfer = 0;
3564                 qgroup->rfer_cmpr = 0;
3565                 qgroup->excl = 0;
3566                 qgroup->excl_cmpr = 0;
3567                 qgroup_dirty(fs_info, qgroup);
3568         }
3569         spin_unlock(&fs_info->qgroup_lock);
3570 }
3571
3572 int
3573 btrfs_qgroup_rescan(struct btrfs_fs_info *fs_info)
3574 {
3575         int ret = 0;
3576         struct btrfs_trans_handle *trans;
3577
3578         ret = qgroup_rescan_init(fs_info, 0, 1);
3579         if (ret)
3580                 return ret;
3581
3582         /*
3583          * We have set the rescan_progress to 0, which means no more
3584          * delayed refs will be accounted by btrfs_qgroup_account_ref.
3585          * However, btrfs_qgroup_account_ref may be right after its call
3586          * to btrfs_find_all_roots, in which case it would still do the
3587          * accounting.
3588          * To solve this, we're committing the transaction, which will
3589          * ensure we run all delayed refs and only after that, we are
3590          * going to clear all tracking information for a clean start.
3591          */
3592
3593         trans = btrfs_attach_transaction_barrier(fs_info->fs_root);
3594         if (IS_ERR(trans) && trans != ERR_PTR(-ENOENT)) {
3595                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3596                 return PTR_ERR(trans);
3597         } else if (trans != ERR_PTR(-ENOENT)) {
3598                 ret = btrfs_commit_transaction(trans);
3599                 if (ret) {
3600                         fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3601                         return ret;
3602                 }
3603         }
3604
3605         qgroup_rescan_zero_tracking(fs_info);
3606
3607         mutex_lock(&fs_info->qgroup_rescan_lock);
3608         fs_info->qgroup_rescan_running = true;
3609         btrfs_queue_work(fs_info->qgroup_rescan_workers,
3610                          &fs_info->qgroup_rescan_work);
3611         mutex_unlock(&fs_info->qgroup_rescan_lock);
3612
3613         return 0;
3614 }
3615
3616 int btrfs_qgroup_wait_for_completion(struct btrfs_fs_info *fs_info,
3617                                      bool interruptible)
3618 {
3619         int running;
3620         int ret = 0;
3621
3622         mutex_lock(&fs_info->qgroup_rescan_lock);
3623         running = fs_info->qgroup_rescan_running;
3624         mutex_unlock(&fs_info->qgroup_rescan_lock);
3625
3626         if (!running)
3627                 return 0;
3628
3629         if (interruptible)
3630                 ret = wait_for_completion_interruptible(
3631                                         &fs_info->qgroup_rescan_completion);
3632         else
3633                 wait_for_completion(&fs_info->qgroup_rescan_completion);
3634
3635         return ret;
3636 }
3637
3638 /*
3639  * this is only called from open_ctree where we're still single threaded, thus
3640  * locking is omitted here.
3641  */
3642 void
3643 btrfs_qgroup_rescan_resume(struct btrfs_fs_info *fs_info)
3644 {
3645         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
3646                 mutex_lock(&fs_info->qgroup_rescan_lock);
3647                 fs_info->qgroup_rescan_running = true;
3648                 btrfs_queue_work(fs_info->qgroup_rescan_workers,
3649                                  &fs_info->qgroup_rescan_work);
3650                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3651         }
3652 }
3653
3654 #define rbtree_iterate_from_safe(node, next, start)                             \
3655        for (node = start; node && ({ next = rb_next(node); 1;}); node = next)
3656
3657 static int qgroup_unreserve_range(struct btrfs_inode *inode,
3658                                   struct extent_changeset *reserved, u64 start,
3659                                   u64 len)
3660 {
3661         struct rb_node *node;
3662         struct rb_node *next;
3663         struct ulist_node *entry;
3664         int ret = 0;
3665
3666         node = reserved->range_changed.root.rb_node;
3667         if (!node)
3668                 return 0;
3669         while (node) {
3670                 entry = rb_entry(node, struct ulist_node, rb_node);
3671                 if (entry->val < start)
3672                         node = node->rb_right;
3673                 else
3674                         node = node->rb_left;
3675         }
3676
3677         if (entry->val > start && rb_prev(&entry->rb_node))
3678                 entry = rb_entry(rb_prev(&entry->rb_node), struct ulist_node,
3679                                  rb_node);
3680
3681         rbtree_iterate_from_safe(node, next, &entry->rb_node) {
3682                 u64 entry_start;
3683                 u64 entry_end;
3684                 u64 entry_len;
3685                 int clear_ret;
3686
3687                 entry = rb_entry(node, struct ulist_node, rb_node);
3688                 entry_start = entry->val;
3689                 entry_end = entry->aux;
3690                 entry_len = entry_end - entry_start + 1;
3691
3692                 if (entry_start >= start + len)
3693                         break;
3694                 if (entry_start + entry_len <= start)
3695                         continue;
3696                 /*
3697                  * Now the entry is in [start, start + len), revert the
3698                  * EXTENT_QGROUP_RESERVED bit.
3699                  */
3700                 clear_ret = clear_extent_bits(&inode->io_tree, entry_start,
3701                                               entry_end, EXTENT_QGROUP_RESERVED);
3702                 if (!ret && clear_ret < 0)
3703                         ret = clear_ret;
3704
3705                 ulist_del(&reserved->range_changed, entry->val, entry->aux);
3706                 if (likely(reserved->bytes_changed >= entry_len)) {
3707                         reserved->bytes_changed -= entry_len;
3708                 } else {
3709                         WARN_ON(1);
3710                         reserved->bytes_changed = 0;
3711                 }
3712         }
3713
3714         return ret;
3715 }
3716
3717 /*
3718  * Try to free some space for qgroup.
3719  *
3720  * For qgroup, there are only 3 ways to free qgroup space:
3721  * - Flush nodatacow write
3722  *   Any nodatacow write will free its reserved data space at run_delalloc_range().
3723  *   In theory, we should only flush nodatacow inodes, but it's not yet
3724  *   possible, so we need to flush the whole root.
3725  *
3726  * - Wait for ordered extents
3727  *   When ordered extents are finished, their reserved metadata is finally
3728  *   converted to per_trans status, which can be freed by later commit
3729  *   transaction.
3730  *
3731  * - Commit transaction
3732  *   This would free the meta_per_trans space.
3733  *   In theory this shouldn't provide much space, but any more qgroup space
3734  *   is needed.
3735  */
3736 static int try_flush_qgroup(struct btrfs_root *root)
3737 {
3738         struct btrfs_trans_handle *trans;
3739         int ret;
3740
3741         /* Can't hold an open transaction or we run the risk of deadlocking. */
3742         ASSERT(current->journal_info == NULL);
3743         if (WARN_ON(current->journal_info))
3744                 return 0;
3745
3746         /*
3747          * We don't want to run flush again and again, so if there is a running
3748          * one, we won't try to start a new flush, but exit directly.
3749          */
3750         if (test_and_set_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state)) {
3751                 wait_event(root->qgroup_flush_wait,
3752                         !test_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state));
3753                 return 0;
3754         }
3755
3756         ret = btrfs_start_delalloc_snapshot(root, true);
3757         if (ret < 0)
3758                 goto out;
3759         btrfs_wait_ordered_extents(root, U64_MAX, 0, (u64)-1);
3760
3761         trans = btrfs_attach_transaction_barrier(root);
3762         if (IS_ERR(trans)) {
3763                 ret = PTR_ERR(trans);
3764                 if (ret == -ENOENT)
3765                         ret = 0;
3766                 goto out;
3767         }
3768
3769         ret = btrfs_commit_transaction(trans);
3770 out:
3771         clear_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state);
3772         wake_up(&root->qgroup_flush_wait);
3773         return ret;
3774 }
3775
3776 static int qgroup_reserve_data(struct btrfs_inode *inode,
3777                         struct extent_changeset **reserved_ret, u64 start,
3778                         u64 len)
3779 {
3780         struct btrfs_root *root = inode->root;
3781         struct extent_changeset *reserved;
3782         bool new_reserved = false;
3783         u64 orig_reserved;
3784         u64 to_reserve;
3785         int ret;
3786
3787         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &root->fs_info->flags) ||
3788             !is_fstree(root->root_key.objectid) || len == 0)
3789                 return 0;
3790
3791         /* @reserved parameter is mandatory for qgroup */
3792         if (WARN_ON(!reserved_ret))
3793                 return -EINVAL;
3794         if (!*reserved_ret) {
3795                 new_reserved = true;
3796                 *reserved_ret = extent_changeset_alloc();
3797                 if (!*reserved_ret)
3798                         return -ENOMEM;
3799         }
3800         reserved = *reserved_ret;
3801         /* Record already reserved space */
3802         orig_reserved = reserved->bytes_changed;
3803         ret = set_record_extent_bits(&inode->io_tree, start,
3804                         start + len -1, EXTENT_QGROUP_RESERVED, reserved);
3805
3806         /* Newly reserved space */
3807         to_reserve = reserved->bytes_changed - orig_reserved;
3808         trace_btrfs_qgroup_reserve_data(&inode->vfs_inode, start, len,
3809                                         to_reserve, QGROUP_RESERVE);
3810         if (ret < 0)
3811                 goto out;
3812         ret = qgroup_reserve(root, to_reserve, true, BTRFS_QGROUP_RSV_DATA);
3813         if (ret < 0)
3814                 goto cleanup;
3815
3816         return ret;
3817
3818 cleanup:
3819         qgroup_unreserve_range(inode, reserved, start, len);
3820 out:
3821         if (new_reserved) {
3822                 extent_changeset_free(reserved);
3823                 *reserved_ret = NULL;
3824         }
3825         return ret;
3826 }
3827
3828 /*
3829  * Reserve qgroup space for range [start, start + len).
3830  *
3831  * This function will either reserve space from related qgroups or do nothing
3832  * if the range is already reserved.
3833  *
3834  * Return 0 for successful reservation
3835  * Return <0 for error (including -EQUOT)
3836  *
3837  * NOTE: This function may sleep for memory allocation, dirty page flushing and
3838  *       commit transaction. So caller should not hold any dirty page locked.
3839  */
3840 int btrfs_qgroup_reserve_data(struct btrfs_inode *inode,
3841                         struct extent_changeset **reserved_ret, u64 start,
3842                         u64 len)
3843 {
3844         int ret;
3845
3846         ret = qgroup_reserve_data(inode, reserved_ret, start, len);
3847         if (ret <= 0 && ret != -EDQUOT)
3848                 return ret;
3849
3850         ret = try_flush_qgroup(inode->root);
3851         if (ret < 0)
3852                 return ret;
3853         return qgroup_reserve_data(inode, reserved_ret, start, len);
3854 }
3855
3856 /* Free ranges specified by @reserved, normally in error path */
3857 static int qgroup_free_reserved_data(struct btrfs_inode *inode,
3858                         struct extent_changeset *reserved, u64 start, u64 len)
3859 {
3860         struct btrfs_root *root = inode->root;
3861         struct ulist_node *unode;
3862         struct ulist_iterator uiter;
3863         struct extent_changeset changeset;
3864         int freed = 0;
3865         int ret;
3866
3867         extent_changeset_init(&changeset);
3868         len = round_up(start + len, root->fs_info->sectorsize);
3869         start = round_down(start, root->fs_info->sectorsize);
3870
3871         ULIST_ITER_INIT(&uiter);
3872         while ((unode = ulist_next(&reserved->range_changed, &uiter))) {
3873                 u64 range_start = unode->val;
3874                 /* unode->aux is the inclusive end */
3875                 u64 range_len = unode->aux - range_start + 1;
3876                 u64 free_start;
3877                 u64 free_len;
3878
3879                 extent_changeset_release(&changeset);
3880
3881                 /* Only free range in range [start, start + len) */
3882                 if (range_start >= start + len ||
3883                     range_start + range_len <= start)
3884                         continue;
3885                 free_start = max(range_start, start);
3886                 free_len = min(start + len, range_start + range_len) -
3887                            free_start;
3888                 /*
3889                  * TODO: To also modify reserved->ranges_reserved to reflect
3890                  * the modification.
3891                  *
3892                  * However as long as we free qgroup reserved according to
3893                  * EXTENT_QGROUP_RESERVED, we won't double free.
3894                  * So not need to rush.
3895                  */
3896                 ret = clear_record_extent_bits(&inode->io_tree, free_start,
3897                                 free_start + free_len - 1,
3898                                 EXTENT_QGROUP_RESERVED, &changeset);
3899                 if (ret < 0)
3900                         goto out;
3901                 freed += changeset.bytes_changed;
3902         }
3903         btrfs_qgroup_free_refroot(root->fs_info, root->root_key.objectid, freed,
3904                                   BTRFS_QGROUP_RSV_DATA);
3905         ret = freed;
3906 out:
3907         extent_changeset_release(&changeset);
3908         return ret;
3909 }
3910
3911 static int __btrfs_qgroup_release_data(struct btrfs_inode *inode,
3912                         struct extent_changeset *reserved, u64 start, u64 len,
3913                         int free)
3914 {
3915         struct extent_changeset changeset;
3916         int trace_op = QGROUP_RELEASE;
3917         int ret;
3918
3919         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &inode->root->fs_info->flags))
3920                 return 0;
3921
3922         /* In release case, we shouldn't have @reserved */
3923         WARN_ON(!free && reserved);
3924         if (free && reserved)
3925                 return qgroup_free_reserved_data(inode, reserved, start, len);
3926         extent_changeset_init(&changeset);
3927         ret = clear_record_extent_bits(&inode->io_tree, start, start + len -1,
3928                                        EXTENT_QGROUP_RESERVED, &changeset);
3929         if (ret < 0)
3930                 goto out;
3931
3932         if (free)
3933                 trace_op = QGROUP_FREE;
3934         trace_btrfs_qgroup_release_data(&inode->vfs_inode, start, len,
3935                                         changeset.bytes_changed, trace_op);
3936         if (free)
3937                 btrfs_qgroup_free_refroot(inode->root->fs_info,
3938                                 inode->root->root_key.objectid,
3939                                 changeset.bytes_changed, BTRFS_QGROUP_RSV_DATA);
3940         ret = changeset.bytes_changed;
3941 out:
3942         extent_changeset_release(&changeset);
3943         return ret;
3944 }
3945
3946 /*
3947  * Free a reserved space range from io_tree and related qgroups
3948  *
3949  * Should be called when a range of pages get invalidated before reaching disk.
3950  * Or for error cleanup case.
3951  * if @reserved is given, only reserved range in [@start, @start + @len) will
3952  * be freed.
3953  *
3954  * For data written to disk, use btrfs_qgroup_release_data().
3955  *
3956  * NOTE: This function may sleep for memory allocation.
3957  */
3958 int btrfs_qgroup_free_data(struct btrfs_inode *inode,
3959                         struct extent_changeset *reserved, u64 start, u64 len)
3960 {
3961         return __btrfs_qgroup_release_data(inode, reserved, start, len, 1);
3962 }
3963
3964 /*
3965  * Release a reserved space range from io_tree only.
3966  *
3967  * Should be called when a range of pages get written to disk and corresponding
3968  * FILE_EXTENT is inserted into corresponding root.
3969  *
3970  * Since new qgroup accounting framework will only update qgroup numbers at
3971  * commit_transaction() time, its reserved space shouldn't be freed from
3972  * related qgroups.
3973  *
3974  * But we should release the range from io_tree, to allow further write to be
3975  * COWed.
3976  *
3977  * NOTE: This function may sleep for memory allocation.
3978  */
3979 int btrfs_qgroup_release_data(struct btrfs_inode *inode, u64 start, u64 len)
3980 {
3981         return __btrfs_qgroup_release_data(inode, NULL, start, len, 0);
3982 }
3983
3984 static void add_root_meta_rsv(struct btrfs_root *root, int num_bytes,
3985                               enum btrfs_qgroup_rsv_type type)
3986 {
3987         if (type != BTRFS_QGROUP_RSV_META_PREALLOC &&
3988             type != BTRFS_QGROUP_RSV_META_PERTRANS)
3989                 return;
3990         if (num_bytes == 0)
3991                 return;
3992
3993         spin_lock(&root->qgroup_meta_rsv_lock);
3994         if (type == BTRFS_QGROUP_RSV_META_PREALLOC)
3995                 root->qgroup_meta_rsv_prealloc += num_bytes;
3996         else
3997                 root->qgroup_meta_rsv_pertrans += num_bytes;
3998         spin_unlock(&root->qgroup_meta_rsv_lock);
3999 }
4000
4001 static int sub_root_meta_rsv(struct btrfs_root *root, int num_bytes,
4002                              enum btrfs_qgroup_rsv_type type)
4003 {
4004         if (type != BTRFS_QGROUP_RSV_META_PREALLOC &&
4005             type != BTRFS_QGROUP_RSV_META_PERTRANS)
4006                 return 0;
4007         if (num_bytes == 0)
4008                 return 0;
4009
4010         spin_lock(&root->qgroup_meta_rsv_lock);
4011         if (type == BTRFS_QGROUP_RSV_META_PREALLOC) {
4012                 num_bytes = min_t(u64, root->qgroup_meta_rsv_prealloc,
4013                                   num_bytes);
4014                 root->qgroup_meta_rsv_prealloc -= num_bytes;
4015         } else {
4016                 num_bytes = min_t(u64, root->qgroup_meta_rsv_pertrans,
4017                                   num_bytes);
4018                 root->qgroup_meta_rsv_pertrans -= num_bytes;
4019         }
4020         spin_unlock(&root->qgroup_meta_rsv_lock);
4021         return num_bytes;
4022 }
4023
4024 int btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
4025                               enum btrfs_qgroup_rsv_type type, bool enforce)
4026 {
4027         struct btrfs_fs_info *fs_info = root->fs_info;
4028         int ret;
4029
4030         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
4031             !is_fstree(root->root_key.objectid) || num_bytes == 0)
4032                 return 0;
4033
4034         BUG_ON(num_bytes != round_down(num_bytes, fs_info->nodesize));
4035         trace_qgroup_meta_reserve(root, (s64)num_bytes, type);
4036         ret = qgroup_reserve(root, num_bytes, enforce, type);
4037         if (ret < 0)
4038                 return ret;
4039         /*
4040          * Record what we have reserved into root.
4041          *
4042          * To avoid quota disabled->enabled underflow.
4043          * In that case, we may try to free space we haven't reserved
4044          * (since quota was disabled), so record what we reserved into root.
4045          * And ensure later release won't underflow this number.
4046          */
4047         add_root_meta_rsv(root, num_bytes, type);
4048         return ret;
4049 }
4050
4051 int __btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
4052                                 enum btrfs_qgroup_rsv_type type, bool enforce,
4053                                 bool noflush)
4054 {
4055         int ret;
4056
4057         ret = btrfs_qgroup_reserve_meta(root, num_bytes, type, enforce);
4058         if ((ret <= 0 && ret != -EDQUOT) || noflush)
4059                 return ret;
4060
4061         ret = try_flush_qgroup(root);
4062         if (ret < 0)
4063                 return ret;
4064         return btrfs_qgroup_reserve_meta(root, num_bytes, type, enforce);
4065 }
4066
4067 void btrfs_qgroup_free_meta_all_pertrans(struct btrfs_root *root)
4068 {
4069         struct btrfs_fs_info *fs_info = root->fs_info;
4070
4071         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
4072             !is_fstree(root->root_key.objectid))
4073                 return;
4074
4075         /* TODO: Update trace point to handle such free */
4076         trace_qgroup_meta_free_all_pertrans(root);
4077         /* Special value -1 means to free all reserved space */
4078         btrfs_qgroup_free_refroot(fs_info, root->root_key.objectid, (u64)-1,
4079                                   BTRFS_QGROUP_RSV_META_PERTRANS);
4080 }
4081
4082 void __btrfs_qgroup_free_meta(struct btrfs_root *root, int num_bytes,
4083                               enum btrfs_qgroup_rsv_type type)
4084 {
4085         struct btrfs_fs_info *fs_info = root->fs_info;
4086
4087         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
4088             !is_fstree(root->root_key.objectid))
4089                 return;
4090
4091         /*
4092          * reservation for META_PREALLOC can happen before quota is enabled,
4093          * which can lead to underflow.
4094          * Here ensure we will only free what we really have reserved.
4095          */
4096         num_bytes = sub_root_meta_rsv(root, num_bytes, type);
4097         BUG_ON(num_bytes != round_down(num_bytes, fs_info->nodesize));
4098         trace_qgroup_meta_reserve(root, -(s64)num_bytes, type);
4099         btrfs_qgroup_free_refroot(fs_info, root->root_key.objectid,
4100                                   num_bytes, type);
4101 }
4102
4103 static void qgroup_convert_meta(struct btrfs_fs_info *fs_info, u64 ref_root,
4104                                 int num_bytes)
4105 {
4106         struct btrfs_qgroup *qgroup;
4107         struct ulist_node *unode;
4108         struct ulist_iterator uiter;
4109         int ret = 0;
4110
4111         if (num_bytes == 0)
4112                 return;
4113         if (!fs_info->quota_root)
4114                 return;
4115
4116         spin_lock(&fs_info->qgroup_lock);
4117         qgroup = find_qgroup_rb(fs_info, ref_root);
4118         if (!qgroup)
4119                 goto out;
4120         ulist_reinit(fs_info->qgroup_ulist);
4121         ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
4122                        qgroup_to_aux(qgroup), GFP_ATOMIC);
4123         if (ret < 0)
4124                 goto out;
4125         ULIST_ITER_INIT(&uiter);
4126         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
4127                 struct btrfs_qgroup *qg;
4128                 struct btrfs_qgroup_list *glist;
4129
4130                 qg = unode_aux_to_qgroup(unode);
4131
4132                 qgroup_rsv_release(fs_info, qg, num_bytes,
4133                                 BTRFS_QGROUP_RSV_META_PREALLOC);
4134                 qgroup_rsv_add(fs_info, qg, num_bytes,
4135                                 BTRFS_QGROUP_RSV_META_PERTRANS);
4136                 list_for_each_entry(glist, &qg->groups, next_group) {
4137                         ret = ulist_add(fs_info->qgroup_ulist,
4138                                         glist->group->qgroupid,
4139                                         qgroup_to_aux(glist->group), GFP_ATOMIC);
4140                         if (ret < 0)
4141                                 goto out;
4142                 }
4143         }
4144 out:
4145         spin_unlock(&fs_info->qgroup_lock);
4146 }
4147
4148 void btrfs_qgroup_convert_reserved_meta(struct btrfs_root *root, int num_bytes)
4149 {
4150         struct btrfs_fs_info *fs_info = root->fs_info;
4151
4152         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
4153             !is_fstree(root->root_key.objectid))
4154                 return;
4155         /* Same as btrfs_qgroup_free_meta_prealloc() */
4156         num_bytes = sub_root_meta_rsv(root, num_bytes,
4157                                       BTRFS_QGROUP_RSV_META_PREALLOC);
4158         trace_qgroup_meta_convert(root, num_bytes);
4159         qgroup_convert_meta(fs_info, root->root_key.objectid, num_bytes);
4160 }
4161
4162 /*
4163  * Check qgroup reserved space leaking, normally at destroy inode
4164  * time
4165  */
4166 void btrfs_qgroup_check_reserved_leak(struct btrfs_inode *inode)
4167 {
4168         struct extent_changeset changeset;
4169         struct ulist_node *unode;
4170         struct ulist_iterator iter;
4171         int ret;
4172
4173         extent_changeset_init(&changeset);
4174         ret = clear_record_extent_bits(&inode->io_tree, 0, (u64)-1,
4175                         EXTENT_QGROUP_RESERVED, &changeset);
4176
4177         WARN_ON(ret < 0);
4178         if (WARN_ON(changeset.bytes_changed)) {
4179                 ULIST_ITER_INIT(&iter);
4180                 while ((unode = ulist_next(&changeset.range_changed, &iter))) {
4181                         btrfs_warn(inode->root->fs_info,
4182                 "leaking qgroup reserved space, ino: %llu, start: %llu, end: %llu",
4183                                 btrfs_ino(inode), unode->val, unode->aux);
4184                 }
4185                 btrfs_qgroup_free_refroot(inode->root->fs_info,
4186                                 inode->root->root_key.objectid,
4187                                 changeset.bytes_changed, BTRFS_QGROUP_RSV_DATA);
4188
4189         }
4190         extent_changeset_release(&changeset);
4191 }
4192
4193 void btrfs_qgroup_init_swapped_blocks(
4194         struct btrfs_qgroup_swapped_blocks *swapped_blocks)
4195 {
4196         int i;
4197
4198         spin_lock_init(&swapped_blocks->lock);
4199         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
4200                 swapped_blocks->blocks[i] = RB_ROOT;
4201         swapped_blocks->swapped = false;
4202 }
4203
4204 /*
4205  * Delete all swapped blocks record of @root.
4206  * Every record here means we skipped a full subtree scan for qgroup.
4207  *
4208  * Gets called when committing one transaction.
4209  */
4210 void btrfs_qgroup_clean_swapped_blocks(struct btrfs_root *root)
4211 {
4212         struct btrfs_qgroup_swapped_blocks *swapped_blocks;
4213         int i;
4214
4215         swapped_blocks = &root->swapped_blocks;
4216
4217         spin_lock(&swapped_blocks->lock);
4218         if (!swapped_blocks->swapped)
4219                 goto out;
4220         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
4221                 struct rb_root *cur_root = &swapped_blocks->blocks[i];
4222                 struct btrfs_qgroup_swapped_block *entry;
4223                 struct btrfs_qgroup_swapped_block *next;
4224
4225                 rbtree_postorder_for_each_entry_safe(entry, next, cur_root,
4226                                                      node)
4227                         kfree(entry);
4228                 swapped_blocks->blocks[i] = RB_ROOT;
4229         }
4230         swapped_blocks->swapped = false;
4231 out:
4232         spin_unlock(&swapped_blocks->lock);
4233 }
4234
4235 /*
4236  * Add subtree roots record into @subvol_root.
4237  *
4238  * @subvol_root:        tree root of the subvolume tree get swapped
4239  * @bg:                 block group under balance
4240  * @subvol_parent/slot: pointer to the subtree root in subvolume tree
4241  * @reloc_parent/slot:  pointer to the subtree root in reloc tree
4242  *                      BOTH POINTERS ARE BEFORE TREE SWAP
4243  * @last_snapshot:      last snapshot generation of the subvolume tree
4244  */
4245 int btrfs_qgroup_add_swapped_blocks(struct btrfs_trans_handle *trans,
4246                 struct btrfs_root *subvol_root,
4247                 struct btrfs_block_group *bg,
4248                 struct extent_buffer *subvol_parent, int subvol_slot,
4249                 struct extent_buffer *reloc_parent, int reloc_slot,
4250                 u64 last_snapshot)
4251 {
4252         struct btrfs_fs_info *fs_info = subvol_root->fs_info;
4253         struct btrfs_qgroup_swapped_blocks *blocks = &subvol_root->swapped_blocks;
4254         struct btrfs_qgroup_swapped_block *block;
4255         struct rb_node **cur;
4256         struct rb_node *parent = NULL;
4257         int level = btrfs_header_level(subvol_parent) - 1;
4258         int ret = 0;
4259
4260         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
4261                 return 0;
4262
4263         if (btrfs_node_ptr_generation(subvol_parent, subvol_slot) >
4264             btrfs_node_ptr_generation(reloc_parent, reloc_slot)) {
4265                 btrfs_err_rl(fs_info,
4266                 "%s: bad parameter order, subvol_gen=%llu reloc_gen=%llu",
4267                         __func__,
4268                         btrfs_node_ptr_generation(subvol_parent, subvol_slot),
4269                         btrfs_node_ptr_generation(reloc_parent, reloc_slot));
4270                 return -EUCLEAN;
4271         }
4272
4273         block = kmalloc(sizeof(*block), GFP_NOFS);
4274         if (!block) {
4275                 ret = -ENOMEM;
4276                 goto out;
4277         }
4278
4279         /*
4280          * @reloc_parent/slot is still before swap, while @block is going to
4281          * record the bytenr after swap, so we do the swap here.
4282          */
4283         block->subvol_bytenr = btrfs_node_blockptr(reloc_parent, reloc_slot);
4284         block->subvol_generation = btrfs_node_ptr_generation(reloc_parent,
4285                                                              reloc_slot);
4286         block->reloc_bytenr = btrfs_node_blockptr(subvol_parent, subvol_slot);
4287         block->reloc_generation = btrfs_node_ptr_generation(subvol_parent,
4288                                                             subvol_slot);
4289         block->last_snapshot = last_snapshot;
4290         block->level = level;
4291
4292         /*
4293          * If we have bg == NULL, we're called from btrfs_recover_relocation(),
4294          * no one else can modify tree blocks thus we qgroup will not change
4295          * no matter the value of trace_leaf.
4296          */
4297         if (bg && bg->flags & BTRFS_BLOCK_GROUP_DATA)
4298                 block->trace_leaf = true;
4299         else
4300                 block->trace_leaf = false;
4301         btrfs_node_key_to_cpu(reloc_parent, &block->first_key, reloc_slot);
4302
4303         /* Insert @block into @blocks */
4304         spin_lock(&blocks->lock);
4305         cur = &blocks->blocks[level].rb_node;
4306         while (*cur) {
4307                 struct btrfs_qgroup_swapped_block *entry;
4308
4309                 parent = *cur;
4310                 entry = rb_entry(parent, struct btrfs_qgroup_swapped_block,
4311                                  node);
4312
4313                 if (entry->subvol_bytenr < block->subvol_bytenr) {
4314                         cur = &(*cur)->rb_left;
4315                 } else if (entry->subvol_bytenr > block->subvol_bytenr) {
4316                         cur = &(*cur)->rb_right;
4317                 } else {
4318                         if (entry->subvol_generation !=
4319                                         block->subvol_generation ||
4320                             entry->reloc_bytenr != block->reloc_bytenr ||
4321                             entry->reloc_generation !=
4322                                         block->reloc_generation) {
4323                                 /*
4324                                  * Duplicated but mismatch entry found.
4325                                  * Shouldn't happen.
4326                                  *
4327                                  * Marking qgroup inconsistent should be enough
4328                                  * for end users.
4329                                  */
4330                                 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
4331                                 ret = -EEXIST;
4332                         }
4333                         kfree(block);
4334                         goto out_unlock;
4335                 }
4336         }
4337         rb_link_node(&block->node, parent, cur);
4338         rb_insert_color(&block->node, &blocks->blocks[level]);
4339         blocks->swapped = true;
4340 out_unlock:
4341         spin_unlock(&blocks->lock);
4342 out:
4343         if (ret < 0)
4344                 qgroup_mark_inconsistent(fs_info);
4345         return ret;
4346 }
4347
4348 /*
4349  * Check if the tree block is a subtree root, and if so do the needed
4350  * delayed subtree trace for qgroup.
4351  *
4352  * This is called during btrfs_cow_block().
4353  */
4354 int btrfs_qgroup_trace_subtree_after_cow(struct btrfs_trans_handle *trans,
4355                                          struct btrfs_root *root,
4356                                          struct extent_buffer *subvol_eb)
4357 {
4358         struct btrfs_fs_info *fs_info = root->fs_info;
4359         struct btrfs_tree_parent_check check = { 0 };
4360         struct btrfs_qgroup_swapped_blocks *blocks = &root->swapped_blocks;
4361         struct btrfs_qgroup_swapped_block *block;
4362         struct extent_buffer *reloc_eb = NULL;
4363         struct rb_node *node;
4364         bool found = false;
4365         bool swapped = false;
4366         int level = btrfs_header_level(subvol_eb);
4367         int ret = 0;
4368         int i;
4369
4370         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
4371                 return 0;
4372         if (!is_fstree(root->root_key.objectid) || !root->reloc_root)
4373                 return 0;
4374
4375         spin_lock(&blocks->lock);
4376         if (!blocks->swapped) {
4377                 spin_unlock(&blocks->lock);
4378                 return 0;
4379         }
4380         node = blocks->blocks[level].rb_node;
4381
4382         while (node) {
4383                 block = rb_entry(node, struct btrfs_qgroup_swapped_block, node);
4384                 if (block->subvol_bytenr < subvol_eb->start) {
4385                         node = node->rb_left;
4386                 } else if (block->subvol_bytenr > subvol_eb->start) {
4387                         node = node->rb_right;
4388                 } else {
4389                         found = true;
4390                         break;
4391                 }
4392         }
4393         if (!found) {
4394                 spin_unlock(&blocks->lock);
4395                 goto out;
4396         }
4397         /* Found one, remove it from @blocks first and update blocks->swapped */
4398         rb_erase(&block->node, &blocks->blocks[level]);
4399         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
4400                 if (RB_EMPTY_ROOT(&blocks->blocks[i])) {
4401                         swapped = true;
4402                         break;
4403                 }
4404         }
4405         blocks->swapped = swapped;
4406         spin_unlock(&blocks->lock);
4407
4408         check.level = block->level;
4409         check.transid = block->reloc_generation;
4410         check.has_first_key = true;
4411         memcpy(&check.first_key, &block->first_key, sizeof(check.first_key));
4412
4413         /* Read out reloc subtree root */
4414         reloc_eb = read_tree_block(fs_info, block->reloc_bytenr, &check);
4415         if (IS_ERR(reloc_eb)) {
4416                 ret = PTR_ERR(reloc_eb);
4417                 reloc_eb = NULL;
4418                 goto free_out;
4419         }
4420         if (!extent_buffer_uptodate(reloc_eb)) {
4421                 ret = -EIO;
4422                 goto free_out;
4423         }
4424
4425         ret = qgroup_trace_subtree_swap(trans, reloc_eb, subvol_eb,
4426                         block->last_snapshot, block->trace_leaf);
4427 free_out:
4428         kfree(block);
4429         free_extent_buffer(reloc_eb);
4430 out:
4431         if (ret < 0) {
4432                 btrfs_err_rl(fs_info,
4433                              "failed to account subtree at bytenr %llu: %d",
4434                              subvol_eb->start, ret);
4435                 qgroup_mark_inconsistent(fs_info);
4436         }
4437         return ret;
4438 }
4439
4440 void btrfs_qgroup_destroy_extent_records(struct btrfs_transaction *trans)
4441 {
4442         struct btrfs_qgroup_extent_record *entry;
4443         struct btrfs_qgroup_extent_record *next;
4444         struct rb_root *root;
4445
4446         root = &trans->delayed_refs.dirty_extent_root;
4447         rbtree_postorder_for_each_entry_safe(entry, next, root, node) {
4448                 ulist_free(entry->old_roots);
4449                 kfree(entry);
4450         }
4451         *root = RB_ROOT;
4452 }