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