fs/btrfs/qgroup.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2011 STRATO.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/writeback.h>
   9 #include <linux/blkdev.h>
  10 #include <linux/rbtree.h>
  11 #include <linux/slab.h>
  12 #include <linux/workqueue.h>
  13 #include <linux/btrfs.h>
  14 #include <linux/sched/mm.h>
  15
  16 #include "ctree.h"
  17 #include "transaction.h"
  18 #include "disk-io.h"
  19 #include "locking.h"
  20 #include "ulist.h"
  21 #include "backref.h"
  22 #include "extent_io.h"
  23 #include "qgroup.h"
  24 #include "block-group.h"
  25 #include "sysfs.h"
  26
  27 /* TODO XXX FIXME
  28  *  - subvol delete -> delete when ref goes to 0? delete limits also?
  29  *  - reorganize keys
  30  *  - compressed
  31  *  - sync
  32  *  - copy also limits on subvol creation
  33  *  - limit
  34  *  - caches for ulists
  35  *  - performance benchmarks
  36  *  - check all ioctl parameters
  37  */
  38
  39 /*
  40  * Helpers to access qgroup reservation
  41  *
  42  * Callers should ensure the lock context and type are valid
  43  */
  44
  45 static u64 qgroup_rsv_total(const struct btrfs_qgroup *qgroup)
  46 {
  47         u64 ret = 0;
  48         int i;
  49
  50         for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++)
  51                 ret += qgroup->rsv.values[i];
  52
  53         return ret;
  54 }
  55
  56 #ifdef CONFIG_BTRFS_DEBUG
  57 static const char *qgroup_rsv_type_str(enum btrfs_qgroup_rsv_type type)
  58 {
  59         if (type == BTRFS_QGROUP_RSV_DATA)
  60                 return "data";
  61         if (type == BTRFS_QGROUP_RSV_META_PERTRANS)
  62                 return "meta_pertrans";
  63         if (type == BTRFS_QGROUP_RSV_META_PREALLOC)
  64                 return "meta_prealloc";
  65         return NULL;
  66 }
  67 #endif
  68
  69 static void qgroup_rsv_add(struct btrfs_fs_info *fs_info,
  70                            struct btrfs_qgroup *qgroup, u64 num_bytes,
  71                            enum btrfs_qgroup_rsv_type type)
  72 {
  73         trace_qgroup_update_reserve(fs_info, qgroup, num_bytes, type);
  74         qgroup->rsv.values[type] += num_bytes;
  75 }
  76
  77 static void qgroup_rsv_release(struct btrfs_fs_info *fs_info,
  78                                struct btrfs_qgroup *qgroup, u64 num_bytes,
  79                                enum btrfs_qgroup_rsv_type type)
  80 {
  81         trace_qgroup_update_reserve(fs_info, qgroup, -(s64)num_bytes, type);
  82         if (qgroup->rsv.values[type] >= num_bytes) {
  83                 qgroup->rsv.values[type] -= num_bytes;
  84                 return;
  85         }
  86 #ifdef CONFIG_BTRFS_DEBUG
  87         WARN_RATELIMIT(1,
  88                 "qgroup %llu %s reserved space underflow, have %llu to free %llu",
  89                 qgroup->qgroupid, qgroup_rsv_type_str(type),
  90                 qgroup->rsv.values[type], num_bytes);
  91 #endif
  92         qgroup->rsv.values[type] = 0;
  93 }
  94
  95 static void qgroup_rsv_add_by_qgroup(struct btrfs_fs_info *fs_info,
  96                                      struct btrfs_qgroup *dest,
  97                                      struct btrfs_qgroup *src)
  98 {
  99         int i;
 100
 101         for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++)
 102                 qgroup_rsv_add(fs_info, dest, src->rsv.values[i], i);
 103 }
 104
 105 static void qgroup_rsv_release_by_qgroup(struct btrfs_fs_info *fs_info,
 106                                          struct btrfs_qgroup *dest,
 107                                           struct btrfs_qgroup *src)
 108 {
 109         int i;
 110
 111         for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++)
 112                 qgroup_rsv_release(fs_info, dest, src->rsv.values[i], i);
 113 }
 114
 115 static void btrfs_qgroup_update_old_refcnt(struct btrfs_qgroup *qg, u64 seq,
 116                                            int mod)
 117 {
 118         if (qg->old_refcnt < seq)
 119                 qg->old_refcnt = seq;
 120         qg->old_refcnt += mod;
 121 }
 122
 123 static void btrfs_qgroup_update_new_refcnt(struct btrfs_qgroup *qg, u64 seq,
 124                                            int mod)
 125 {
 126         if (qg->new_refcnt < seq)
 127                 qg->new_refcnt = seq;
 128         qg->new_refcnt += mod;
 129 }
 130
 131 static inline u64 btrfs_qgroup_get_old_refcnt(struct btrfs_qgroup *qg, u64 seq)
 132 {
 133         if (qg->old_refcnt < seq)
 134                 return 0;
 135         return qg->old_refcnt - seq;
 136 }
 137
 138 static inline u64 btrfs_qgroup_get_new_refcnt(struct btrfs_qgroup *qg, u64 seq)
 139 {
 140         if (qg->new_refcnt < seq)
 141                 return 0;
 142         return qg->new_refcnt - seq;
 143 }
 144
 145 /*
 146  * glue structure to represent the relations between qgroups.
 147  */
 148 struct btrfs_qgroup_list {
 149         struct list_head next_group;
 150         struct list_head next_member;
 151         struct btrfs_qgroup *group;
 152         struct btrfs_qgroup *member;
 153 };
 154
 155 static inline u64 qgroup_to_aux(struct btrfs_qgroup *qg)
 156 {
 157         return (u64)(uintptr_t)qg;
 158 }
 159
 160 static inline struct btrfs_qgroup* unode_aux_to_qgroup(struct ulist_node *n)
 161 {
 162         return (struct btrfs_qgroup *)(uintptr_t)n->aux;
 163 }
 164
 165 static int
 166 qgroup_rescan_init(struct btrfs_fs_info *fs_info, u64 progress_objectid,
 167                    int init_flags);
 168 static void qgroup_rescan_zero_tracking(struct btrfs_fs_info *fs_info);
 169
 170 /* must be called with qgroup_ioctl_lock held */
 171 static struct btrfs_qgroup *find_qgroup_rb(struct btrfs_fs_info *fs_info,
 172                                            u64 qgroupid)
 173 {
 174         struct rb_node *n = fs_info->qgroup_tree.rb_node;
 175         struct btrfs_qgroup *qgroup;
 176
 177         while (n) {
 178                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
 179                 if (qgroup->qgroupid < qgroupid)
 180                         n = n->rb_left;
 181                 else if (qgroup->qgroupid > qgroupid)
 182                         n = n->rb_right;
 183                 else
 184                         return qgroup;
 185         }
 186         return NULL;
 187 }
 188
 189 /* must be called with qgroup_lock held */
 190 static struct btrfs_qgroup *add_qgroup_rb(struct btrfs_fs_info *fs_info,
 191                                           u64 qgroupid)
 192 {
 193         struct rb_node **p = &fs_info->qgroup_tree.rb_node;
 194         struct rb_node *parent = NULL;
 195         struct btrfs_qgroup *qgroup;
 196
 197         while (*p) {
 198                 parent = *p;
 199                 qgroup = rb_entry(parent, struct btrfs_qgroup, node);
 200
 201                 if (qgroup->qgroupid < qgroupid)
 202                         p = &(*p)->rb_left;
 203                 else if (qgroup->qgroupid > qgroupid)
 204                         p = &(*p)->rb_right;
 205                 else
 206                         return qgroup;
 207         }
 208
 209         qgroup = kzalloc(sizeof(*qgroup), GFP_ATOMIC);
 210         if (!qgroup)
 211                 return ERR_PTR(-ENOMEM);
 212
 213         qgroup->qgroupid = qgroupid;
 214         INIT_LIST_HEAD(&qgroup->groups);
 215         INIT_LIST_HEAD(&qgroup->members);
 216         INIT_LIST_HEAD(&qgroup->dirty);
 217
 218         rb_link_node(&qgroup->node, parent, p);
 219         rb_insert_color(&qgroup->node, &fs_info->qgroup_tree);
 220
 221         return qgroup;
 222 }
 223
 224 static void __del_qgroup_rb(struct btrfs_fs_info *fs_info,
 225                             struct btrfs_qgroup *qgroup)
 226 {
 227         struct btrfs_qgroup_list *list;
 228
 229         btrfs_sysfs_del_one_qgroup(fs_info, qgroup);
 230         list_del(&qgroup->dirty);
 231         while (!list_empty(&qgroup->groups)) {
 232                 list = list_first_entry(&qgroup->groups,
 233                                         struct btrfs_qgroup_list, next_group);
 234                 list_del(&list->next_group);
 235                 list_del(&list->next_member);
 236                 kfree(list);
 237         }
 238
 239         while (!list_empty(&qgroup->members)) {
 240                 list = list_first_entry(&qgroup->members,
 241                                         struct btrfs_qgroup_list, next_member);
 242                 list_del(&list->next_group);
 243                 list_del(&list->next_member);
 244                 kfree(list);
 245         }
 246         kfree(qgroup);
 247 }
 248
 249 /* must be called with qgroup_lock held */
 250 static int del_qgroup_rb(struct btrfs_fs_info *fs_info, u64 qgroupid)
 251 {
 252         struct btrfs_qgroup *qgroup = find_qgroup_rb(fs_info, qgroupid);
 253
 254         if (!qgroup)
 255                 return -ENOENT;
 256
 257         rb_erase(&qgroup->node, &fs_info->qgroup_tree);
 258         __del_qgroup_rb(fs_info, qgroup);
 259         return 0;
 260 }
 261
 262 /* must be called with qgroup_lock held */
 263 static int add_relation_rb(struct btrfs_fs_info *fs_info,
 264                            u64 memberid, u64 parentid)
 265 {
 266         struct btrfs_qgroup *member;
 267         struct btrfs_qgroup *parent;
 268         struct btrfs_qgroup_list *list;
 269
 270         member = find_qgroup_rb(fs_info, memberid);
 271         parent = find_qgroup_rb(fs_info, parentid);
 272         if (!member || !parent)
 273                 return -ENOENT;
 274
 275         list = kzalloc(sizeof(*list), GFP_ATOMIC);
 276         if (!list)
 277                 return -ENOMEM;
 278
 279         list->group = parent;
 280         list->member = member;
 281         list_add_tail(&list->next_group, &member->groups);
 282         list_add_tail(&list->next_member, &parent->members);
 283
 284         return 0;
 285 }
 286
 287 /* must be called with qgroup_lock held */
 288 static int del_relation_rb(struct btrfs_fs_info *fs_info,
 289                            u64 memberid, u64 parentid)
 290 {
 291         struct btrfs_qgroup *member;
 292         struct btrfs_qgroup *parent;
 293         struct btrfs_qgroup_list *list;
 294
 295         member = find_qgroup_rb(fs_info, memberid);
 296         parent = find_qgroup_rb(fs_info, parentid);
 297         if (!member || !parent)
 298                 return -ENOENT;
 299
 300         list_for_each_entry(list, &member->groups, next_group) {
 301                 if (list->group == parent) {
 302                         list_del(&list->next_group);
 303                         list_del(&list->next_member);
 304                         kfree(list);
 305                         return 0;
 306                 }
 307         }
 308         return -ENOENT;
 309 }
 310
 311 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 312 int btrfs_verify_qgroup_counts(struct btrfs_fs_info *fs_info, u64 qgroupid,
 313                                u64 rfer, u64 excl)
 314 {
 315         struct btrfs_qgroup *qgroup;
 316
 317         qgroup = find_qgroup_rb(fs_info, qgroupid);
 318         if (!qgroup)
 319                 return -EINVAL;
 320         if (qgroup->rfer != rfer || qgroup->excl != excl)
 321                 return -EINVAL;
 322         return 0;
 323 }
 324 #endif
 325
 326 /*
 327  * The full config is read in one go, only called from open_ctree()
 328  * It doesn't use any locking, as at this point we're still single-threaded
 329  */
 330 int btrfs_read_qgroup_config(struct btrfs_fs_info *fs_info)
 331 {
 332         struct btrfs_key key;
 333         struct btrfs_key found_key;
 334         struct btrfs_root *quota_root = fs_info->quota_root;
 335         struct btrfs_path *path = NULL;
 336         struct extent_buffer *l;
 337         int slot;
 338         int ret = 0;
 339         u64 flags = 0;
 340         u64 rescan_progress = 0;
 341
 342         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
 343                 return 0;
 344
 345         fs_info->qgroup_ulist = ulist_alloc(GFP_KERNEL);
 346         if (!fs_info->qgroup_ulist) {
 347                 ret = -ENOMEM;
 348                 goto out;
 349         }
 350
 351         path = btrfs_alloc_path();
 352         if (!path) {
 353                 ret = -ENOMEM;
 354                 goto out;
 355         }
 356
 357         ret = btrfs_sysfs_add_qgroups(fs_info);
 358         if (ret < 0)
 359                 goto out;
 360         /* default this to quota off, in case no status key is found */
 361         fs_info->qgroup_flags = 0;
 362
 363         /*
 364          * pass 1: read status, all qgroup infos and limits
 365          */
 366         key.objectid = 0;
 367         key.type = 0;
 368         key.offset = 0;
 369         ret = btrfs_search_slot_for_read(quota_root, &key, path, 1, 1);
 370         if (ret)
 371                 goto out;
 372
 373         while (1) {
 374                 struct btrfs_qgroup *qgroup;
 375
 376                 slot = path->slots[0];
 377                 l = path->nodes[0];
 378                 btrfs_item_key_to_cpu(l, &found_key, slot);
 379
 380                 if (found_key.type == BTRFS_QGROUP_STATUS_KEY) {
 381                         struct btrfs_qgroup_status_item *ptr;
 382
 383                         ptr = btrfs_item_ptr(l, slot,
 384                                              struct btrfs_qgroup_status_item);
 385
 386                         if (btrfs_qgroup_status_version(l, ptr) !=
 387                             BTRFS_QGROUP_STATUS_VERSION) {
 388                                 btrfs_err(fs_info,
 389                                  "old qgroup version, quota disabled");
 390                                 goto out;
 391                         }
 392                         if (btrfs_qgroup_status_generation(l, ptr) !=
 393                             fs_info->generation) {
 394                                 flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
 395                                 btrfs_err(fs_info,
 396                                         "qgroup generation mismatch, marked as inconsistent");
 397                         }
 398                         fs_info->qgroup_flags = btrfs_qgroup_status_flags(l,
 399                                                                           ptr);
 400                         rescan_progress = btrfs_qgroup_status_rescan(l, ptr);
 401                         goto next1;
 402                 }
 403
 404                 if (found_key.type != BTRFS_QGROUP_INFO_KEY &&
 405                     found_key.type != BTRFS_QGROUP_LIMIT_KEY)
 406                         goto next1;
 407
 408                 qgroup = find_qgroup_rb(fs_info, found_key.offset);
 409                 if ((qgroup && found_key.type == BTRFS_QGROUP_INFO_KEY) ||
 410                     (!qgroup && found_key.type == BTRFS_QGROUP_LIMIT_KEY)) {
 411                         btrfs_err(fs_info, "inconsistent qgroup config");
 412                         flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
 413                 }
 414                 if (!qgroup) {
 415                         qgroup = add_qgroup_rb(fs_info, found_key.offset);
 416                         if (IS_ERR(qgroup)) {
 417                                 ret = PTR_ERR(qgroup);
 418                                 goto out;
 419                         }
 420                 }
 421                 ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
 422                 if (ret < 0)
 423                         goto out;
 424
 425                 switch (found_key.type) {
 426                 case BTRFS_QGROUP_INFO_KEY: {
 427                         struct btrfs_qgroup_info_item *ptr;
 428
 429                         ptr = btrfs_item_ptr(l, slot,
 430                                              struct btrfs_qgroup_info_item);
 431                         qgroup->rfer = btrfs_qgroup_info_rfer(l, ptr);
 432                         qgroup->rfer_cmpr = btrfs_qgroup_info_rfer_cmpr(l, ptr);
 433                         qgroup->excl = btrfs_qgroup_info_excl(l, ptr);
 434                         qgroup->excl_cmpr = btrfs_qgroup_info_excl_cmpr(l, ptr);
 435                         /* generation currently unused */
 436                         break;
 437                 }
 438                 case BTRFS_QGROUP_LIMIT_KEY: {
 439                         struct btrfs_qgroup_limit_item *ptr;
 440
 441                         ptr = btrfs_item_ptr(l, slot,
 442                                              struct btrfs_qgroup_limit_item);
 443                         qgroup->lim_flags = btrfs_qgroup_limit_flags(l, ptr);
 444                         qgroup->max_rfer = btrfs_qgroup_limit_max_rfer(l, ptr);
 445                         qgroup->max_excl = btrfs_qgroup_limit_max_excl(l, ptr);
 446                         qgroup->rsv_rfer = btrfs_qgroup_limit_rsv_rfer(l, ptr);
 447                         qgroup->rsv_excl = btrfs_qgroup_limit_rsv_excl(l, ptr);
 448                         break;
 449                 }
 450                 }
 451 next1:
 452                 ret = btrfs_next_item(quota_root, path);
 453                 if (ret < 0)
 454                         goto out;
 455                 if (ret)
 456                         break;
 457         }
 458         btrfs_release_path(path);
 459
 460         /*
 461          * pass 2: read all qgroup relations
 462          */
 463         key.objectid = 0;
 464         key.type = BTRFS_QGROUP_RELATION_KEY;
 465         key.offset = 0;
 466         ret = btrfs_search_slot_for_read(quota_root, &key, path, 1, 0);
 467         if (ret)
 468                 goto out;
 469         while (1) {
 470                 slot = path->slots[0];
 471                 l = path->nodes[0];
 472                 btrfs_item_key_to_cpu(l, &found_key, slot);
 473
 474                 if (found_key.type != BTRFS_QGROUP_RELATION_KEY)
 475                         goto next2;
 476
 477                 if (found_key.objectid > found_key.offset) {
 478                         /* parent <- member, not needed to build config */
 479                         /* FIXME should we omit the key completely? */
 480                         goto next2;
 481                 }
 482
 483                 ret = add_relation_rb(fs_info, found_key.objectid,
 484                                       found_key.offset);
 485                 if (ret == -ENOENT) {
 486                         btrfs_warn(fs_info,
 487                                 "orphan qgroup relation 0x%llx->0x%llx",
 488                                 found_key.objectid, found_key.offset);
 489                         ret = 0;        /* ignore the error */
 490                 }
 491                 if (ret)
 492                         goto out;
 493 next2:
 494                 ret = btrfs_next_item(quota_root, path);
 495                 if (ret < 0)
 496                         goto out;
 497                 if (ret)
 498                         break;
 499         }
 500 out:
 501         btrfs_free_path(path);
 502         fs_info->qgroup_flags |= flags;
 503         if (!(fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_ON))
 504                 clear_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
 505         else if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN &&
 506                  ret >= 0)
 507                 ret = qgroup_rescan_init(fs_info, rescan_progress, 0);
 508
 509         if (ret < 0) {
 510                 ulist_free(fs_info->qgroup_ulist);
 511                 fs_info->qgroup_ulist = NULL;
 512                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
 513                 btrfs_sysfs_del_qgroups(fs_info);
 514         }
 515
 516         return ret < 0 ? ret : 0;
 517 }
 518
 519 /*
 520  * Called in close_ctree() when quota is still enabled.  This verifies we don't
 521  * leak some reserved space.
 522  *
 523  * Return false if no reserved space is left.
 524  * Return true if some reserved space is leaked.
 525  */
 526 bool btrfs_check_quota_leak(struct btrfs_fs_info *fs_info)
 527 {
 528         struct rb_node *node;
 529         bool ret = false;
 530
 531         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
 532                 return ret;
 533         /*
 534          * Since we're unmounting, there is no race and no need to grab qgroup
 535          * lock.  And here we don't go post-order to provide a more user
 536          * friendly sorted result.
 537          */
 538         for (node = rb_first(&fs_info->qgroup_tree); node; node = rb_next(node)) {
 539                 struct btrfs_qgroup *qgroup;
 540                 int i;
 541
 542                 qgroup = rb_entry(node, struct btrfs_qgroup, node);
 543                 for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++) {
 544                         if (qgroup->rsv.values[i]) {
 545                                 ret = true;
 546                                 btrfs_warn(fs_info,
 547                 "qgroup %hu/%llu has unreleased space, type %d rsv %llu",
 548                                    btrfs_qgroup_level(qgroup->qgroupid),
 549                                    btrfs_qgroup_subvolid(qgroup->qgroupid),
 550                                    i, qgroup->rsv.values[i]);
 551                         }
 552                 }
 553         }
 554         return ret;
 555 }
 556
 557 /*
 558  * This is called from close_ctree() or open_ctree() or btrfs_quota_disable(),
 559  * first two are in single-threaded paths.And for the third one, we have set
 560  * quota_root to be null with qgroup_lock held before, so it is safe to clean
 561  * up the in-memory structures without qgroup_lock held.
 562  */
 563 void btrfs_free_qgroup_config(struct btrfs_fs_info *fs_info)
 564 {
 565         struct rb_node *n;
 566         struct btrfs_qgroup *qgroup;
 567
 568         while ((n = rb_first(&fs_info->qgroup_tree))) {
 569                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
 570                 rb_erase(n, &fs_info->qgroup_tree);
 571                 __del_qgroup_rb(fs_info, qgroup);
 572         }
 573         /*
 574          * We call btrfs_free_qgroup_config() when unmounting
 575          * filesystem and disabling quota, so we set qgroup_ulist
 576          * to be null here to avoid double free.
 577          */
 578         ulist_free(fs_info->qgroup_ulist);
 579         fs_info->qgroup_ulist = NULL;
 580         btrfs_sysfs_del_qgroups(fs_info);
 581 }
 582
 583 static int add_qgroup_relation_item(struct btrfs_trans_handle *trans, u64 src,
 584                                     u64 dst)
 585 {
 586         int ret;
 587         struct btrfs_root *quota_root = trans->fs_info->quota_root;
 588         struct btrfs_path *path;
 589         struct btrfs_key key;
 590
 591         path = btrfs_alloc_path();
 592         if (!path)
 593                 return -ENOMEM;
 594
 595         key.objectid = src;
 596         key.type = BTRFS_QGROUP_RELATION_KEY;
 597         key.offset = dst;
 598
 599         ret = btrfs_insert_empty_item(trans, quota_root, path, &key, 0);
 600
 601         btrfs_mark_buffer_dirty(path->nodes[0]);
 602
 603         btrfs_free_path(path);
 604         return ret;
 605 }
 606
 607 static int del_qgroup_relation_item(struct btrfs_trans_handle *trans, u64 src,
 608                                     u64 dst)
 609 {
 610         int ret;
 611         struct btrfs_root *quota_root = trans->fs_info->quota_root;
 612         struct btrfs_path *path;
 613         struct btrfs_key key;
 614
 615         path = btrfs_alloc_path();
 616         if (!path)
 617                 return -ENOMEM;
 618
 619         key.objectid = src;
 620         key.type = BTRFS_QGROUP_RELATION_KEY;
 621         key.offset = dst;
 622
 623         ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
 624         if (ret < 0)
 625                 goto out;
 626
 627         if (ret > 0) {
 628                 ret = -ENOENT;
 629                 goto out;
 630         }
 631
 632         ret = btrfs_del_item(trans, quota_root, path);
 633 out:
 634         btrfs_free_path(path);
 635         return ret;
 636 }
 637
 638 static int add_qgroup_item(struct btrfs_trans_handle *trans,
 639                            struct btrfs_root *quota_root, u64 qgroupid)
 640 {
 641         int ret;
 642         struct btrfs_path *path;
 643         struct btrfs_qgroup_info_item *qgroup_info;
 644         struct btrfs_qgroup_limit_item *qgroup_limit;
 645         struct extent_buffer *leaf;
 646         struct btrfs_key key;
 647
 648         if (btrfs_is_testing(quota_root->fs_info))
 649                 return 0;
 650
 651         path = btrfs_alloc_path();
 652         if (!path)
 653                 return -ENOMEM;
 654
 655         key.objectid = 0;
 656         key.type = BTRFS_QGROUP_INFO_KEY;
 657         key.offset = qgroupid;
 658
 659         /*
 660          * Avoid a transaction abort by catching -EEXIST here. In that
 661          * case, we proceed by re-initializing the existing structure
 662          * on disk.
 663          */
 664
 665         ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
 666                                       sizeof(*qgroup_info));
 667         if (ret && ret != -EEXIST)
 668                 goto out;
 669
 670         leaf = path->nodes[0];
 671         qgroup_info = btrfs_item_ptr(leaf, path->slots[0],
 672                                  struct btrfs_qgroup_info_item);
 673         btrfs_set_qgroup_info_generation(leaf, qgroup_info, trans->transid);
 674         btrfs_set_qgroup_info_rfer(leaf, qgroup_info, 0);
 675         btrfs_set_qgroup_info_rfer_cmpr(leaf, qgroup_info, 0);
 676         btrfs_set_qgroup_info_excl(leaf, qgroup_info, 0);
 677         btrfs_set_qgroup_info_excl_cmpr(leaf, qgroup_info, 0);
 678
 679         btrfs_mark_buffer_dirty(leaf);
 680
 681         btrfs_release_path(path);
 682
 683         key.type = BTRFS_QGROUP_LIMIT_KEY;
 684         ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
 685                                       sizeof(*qgroup_limit));
 686         if (ret && ret != -EEXIST)
 687                 goto out;
 688
 689         leaf = path->nodes[0];
 690         qgroup_limit = btrfs_item_ptr(leaf, path->slots[0],
 691                                   struct btrfs_qgroup_limit_item);
 692         btrfs_set_qgroup_limit_flags(leaf, qgroup_limit, 0);
 693         btrfs_set_qgroup_limit_max_rfer(leaf, qgroup_limit, 0);
 694         btrfs_set_qgroup_limit_max_excl(leaf, qgroup_limit, 0);
 695         btrfs_set_qgroup_limit_rsv_rfer(leaf, qgroup_limit, 0);
 696         btrfs_set_qgroup_limit_rsv_excl(leaf, qgroup_limit, 0);
 697
 698         btrfs_mark_buffer_dirty(leaf);
 699
 700         ret = 0;
 701 out:
 702         btrfs_free_path(path);
 703         return ret;
 704 }
 705
 706 static int del_qgroup_item(struct btrfs_trans_handle *trans, u64 qgroupid)
 707 {
 708         int ret;
 709         struct btrfs_root *quota_root = trans->fs_info->quota_root;
 710         struct btrfs_path *path;
 711         struct btrfs_key key;
 712
 713         path = btrfs_alloc_path();
 714         if (!path)
 715                 return -ENOMEM;
 716
 717         key.objectid = 0;
 718         key.type = BTRFS_QGROUP_INFO_KEY;
 719         key.offset = qgroupid;
 720         ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
 721         if (ret < 0)
 722                 goto out;
 723
 724         if (ret > 0) {
 725                 ret = -ENOENT;
 726                 goto out;
 727         }
 728
 729         ret = btrfs_del_item(trans, quota_root, path);
 730         if (ret)
 731                 goto out;
 732
 733         btrfs_release_path(path);
 734
 735         key.type = BTRFS_QGROUP_LIMIT_KEY;
 736         ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
 737         if (ret < 0)
 738                 goto out;
 739
 740         if (ret > 0) {
 741                 ret = -ENOENT;
 742                 goto out;
 743         }
 744
 745         ret = btrfs_del_item(trans, quota_root, path);
 746
 747 out:
 748         btrfs_free_path(path);
 749         return ret;
 750 }
 751
 752 static int update_qgroup_limit_item(struct btrfs_trans_handle *trans,
 753                                     struct btrfs_qgroup *qgroup)
 754 {
 755         struct btrfs_root *quota_root = trans->fs_info->quota_root;
 756         struct btrfs_path *path;
 757         struct btrfs_key key;
 758         struct extent_buffer *l;
 759         struct btrfs_qgroup_limit_item *qgroup_limit;
 760         int ret;
 761         int slot;
 762
 763         key.objectid = 0;
 764         key.type = BTRFS_QGROUP_LIMIT_KEY;
 765         key.offset = qgroup->qgroupid;
 766
 767         path = btrfs_alloc_path();
 768         if (!path)
 769                 return -ENOMEM;
 770
 771         ret = btrfs_search_slot(trans, quota_root, &key, path, 0, 1);
 772         if (ret > 0)
 773                 ret = -ENOENT;
 774
 775         if (ret)
 776                 goto out;
 777
 778         l = path->nodes[0];
 779         slot = path->slots[0];
 780         qgroup_limit = btrfs_item_ptr(l, slot, struct btrfs_qgroup_limit_item);
 781         btrfs_set_qgroup_limit_flags(l, qgroup_limit, qgroup->lim_flags);
 782         btrfs_set_qgroup_limit_max_rfer(l, qgroup_limit, qgroup->max_rfer);
 783         btrfs_set_qgroup_limit_max_excl(l, qgroup_limit, qgroup->max_excl);
 784         btrfs_set_qgroup_limit_rsv_rfer(l, qgroup_limit, qgroup->rsv_rfer);
 785         btrfs_set_qgroup_limit_rsv_excl(l, qgroup_limit, qgroup->rsv_excl);
 786
 787         btrfs_mark_buffer_dirty(l);
 788
 789 out:
 790         btrfs_free_path(path);
 791         return ret;
 792 }
 793
 794 static int update_qgroup_info_item(struct btrfs_trans_handle *trans,
 795                                    struct btrfs_qgroup *qgroup)
 796 {
 797         struct btrfs_fs_info *fs_info = trans->fs_info;
 798         struct btrfs_root *quota_root = fs_info->quota_root;
 799         struct btrfs_path *path;
 800         struct btrfs_key key;
 801         struct extent_buffer *l;
 802         struct btrfs_qgroup_info_item *qgroup_info;
 803         int ret;
 804         int slot;
 805
 806         if (btrfs_is_testing(fs_info))
 807                 return 0;
 808
 809         key.objectid = 0;
 810         key.type = BTRFS_QGROUP_INFO_KEY;
 811         key.offset = qgroup->qgroupid;
 812
 813         path = btrfs_alloc_path();
 814         if (!path)
 815                 return -ENOMEM;
 816
 817         ret = btrfs_search_slot(trans, quota_root, &key, path, 0, 1);
 818         if (ret > 0)
 819                 ret = -ENOENT;
 820
 821         if (ret)
 822                 goto out;
 823
 824         l = path->nodes[0];
 825         slot = path->slots[0];
 826         qgroup_info = btrfs_item_ptr(l, slot, struct btrfs_qgroup_info_item);
 827         btrfs_set_qgroup_info_generation(l, qgroup_info, trans->transid);
 828         btrfs_set_qgroup_info_rfer(l, qgroup_info, qgroup->rfer);
 829         btrfs_set_qgroup_info_rfer_cmpr(l, qgroup_info, qgroup->rfer_cmpr);
 830         btrfs_set_qgroup_info_excl(l, qgroup_info, qgroup->excl);
 831         btrfs_set_qgroup_info_excl_cmpr(l, qgroup_info, qgroup->excl_cmpr);
 832
 833         btrfs_mark_buffer_dirty(l);
 834
 835 out:
 836         btrfs_free_path(path);
 837         return ret;
 838 }
 839
 840 static int update_qgroup_status_item(struct btrfs_trans_handle *trans)
 841 {
 842         struct btrfs_fs_info *fs_info = trans->fs_info;
 843         struct btrfs_root *quota_root = fs_info->quota_root;
 844         struct btrfs_path *path;
 845         struct btrfs_key key;
 846         struct extent_buffer *l;
 847         struct btrfs_qgroup_status_item *ptr;
 848         int ret;
 849         int slot;
 850
 851         key.objectid = 0;
 852         key.type = BTRFS_QGROUP_STATUS_KEY;
 853         key.offset = 0;
 854
 855         path = btrfs_alloc_path();
 856         if (!path)
 857                 return -ENOMEM;
 858
 859         ret = btrfs_search_slot(trans, quota_root, &key, path, 0, 1);
 860         if (ret > 0)
 861                 ret = -ENOENT;
 862
 863         if (ret)
 864                 goto out;
 865
 866         l = path->nodes[0];
 867         slot = path->slots[0];
 868         ptr = btrfs_item_ptr(l, slot, struct btrfs_qgroup_status_item);
 869         btrfs_set_qgroup_status_flags(l, ptr, fs_info->qgroup_flags);
 870         btrfs_set_qgroup_status_generation(l, ptr, trans->transid);
 871         btrfs_set_qgroup_status_rescan(l, ptr,
 872                                 fs_info->qgroup_rescan_progress.objectid);
 873
 874         btrfs_mark_buffer_dirty(l);
 875
 876 out:
 877         btrfs_free_path(path);
 878         return ret;
 879 }
 880
 881 /*
 882  * called with qgroup_lock held
 883  */
 884 static int btrfs_clean_quota_tree(struct btrfs_trans_handle *trans,
 885                                   struct btrfs_root *root)
 886 {
 887         struct btrfs_path *path;
 888         struct btrfs_key key;
 889         struct extent_buffer *leaf = NULL;
 890         int ret;
 891         int nr = 0;
 892
 893         path = btrfs_alloc_path();
 894         if (!path)
 895                 return -ENOMEM;
 896
 897         key.objectid = 0;
 898         key.offset = 0;
 899         key.type = 0;
 900
 901         while (1) {
 902                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 903                 if (ret < 0)
 904                         goto out;
 905                 leaf = path->nodes[0];
 906                 nr = btrfs_header_nritems(leaf);
 907                 if (!nr)
 908                         break;
 909                 /*
 910                  * delete the leaf one by one
 911                  * since the whole tree is going
 912                  * to be deleted.
 913                  */
 914                 path->slots[0] = 0;
 915                 ret = btrfs_del_items(trans, root, path, 0, nr);
 916                 if (ret)
 917                         goto out;
 918
 919                 btrfs_release_path(path);
 920         }
 921         ret = 0;
 922 out:
 923         btrfs_free_path(path);
 924         return ret;
 925 }
 926
 927 int btrfs_quota_enable(struct btrfs_fs_info *fs_info)
 928 {
 929         struct btrfs_root *quota_root;
 930         struct btrfs_root *tree_root = fs_info->tree_root;
 931         struct btrfs_path *path = NULL;
 932         struct btrfs_qgroup_status_item *ptr;
 933         struct extent_buffer *leaf;
 934         struct btrfs_key key;
 935         struct btrfs_key found_key;
 936         struct btrfs_qgroup *qgroup = NULL;
 937         struct btrfs_trans_handle *trans = NULL;
 938         struct ulist *ulist = NULL;
 939         int ret = 0;
 940         int slot;
 941
 942         mutex_lock(&fs_info->qgroup_ioctl_lock);
 943         if (fs_info->quota_root)
 944                 goto out;
 945
 946         ulist = ulist_alloc(GFP_KERNEL);
 947         if (!ulist) {
 948                 ret = -ENOMEM;
 949                 goto out;
 950         }
 951
 952         ret = btrfs_sysfs_add_qgroups(fs_info);
 953         if (ret < 0)
 954                 goto out;
 955
 956         /*
 957          * Unlock qgroup_ioctl_lock before starting the transaction. This is to
 958          * avoid lock acquisition inversion problems (reported by lockdep) between
 959          * qgroup_ioctl_lock and the vfs freeze semaphores, acquired when we
 960          * start a transaction.
 961          * After we started the transaction lock qgroup_ioctl_lock again and
 962          * check if someone else created the quota root in the meanwhile. If so,
 963          * just return success and release the transaction handle.
 964          *
 965          * Also we don't need to worry about someone else calling
 966          * btrfs_sysfs_add_qgroups() after we unlock and getting an error because
 967          * that function returns 0 (success) when the sysfs entries already exist.
 968          */
 969         mutex_unlock(&fs_info->qgroup_ioctl_lock);
 970
 971         /*
 972          * 1 for quota root item
 973          * 1 for BTRFS_QGROUP_STATUS item
 974          *
 975          * Yet we also need 2*n items for a QGROUP_INFO/QGROUP_LIMIT items
 976          * per subvolume. However those are not currently reserved since it
 977          * would be a lot of overkill.
 978          */
 979         trans = btrfs_start_transaction(tree_root, 2);
 980
 981         mutex_lock(&fs_info->qgroup_ioctl_lock);
 982         if (IS_ERR(trans)) {
 983                 ret = PTR_ERR(trans);
 984                 trans = NULL;
 985                 goto out;
 986         }
 987
 988         if (fs_info->quota_root)
 989                 goto out;
 990
 991         fs_info->qgroup_ulist = ulist;
 992         ulist = NULL;
 993
 994         /*
 995          * initially create the quota tree
 996          */
 997         quota_root = btrfs_create_tree(trans, BTRFS_QUOTA_TREE_OBJECTID);
 998         if (IS_ERR(quota_root)) {
 999                 ret =  PTR_ERR(quota_root);
1000                 btrfs_abort_transaction(trans, ret);
1001                 goto out;
1002         }
1003
1004         path = btrfs_alloc_path();
1005         if (!path) {
1006                 ret = -ENOMEM;
1007                 btrfs_abort_transaction(trans, ret);
1008                 goto out_free_root;
1009         }
1010
1011         key.objectid = 0;
1012         key.type = BTRFS_QGROUP_STATUS_KEY;
1013         key.offset = 0;
1014
1015         ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
1016                                       sizeof(*ptr));
1017         if (ret) {
1018                 btrfs_abort_transaction(trans, ret);
1019                 goto out_free_path;
1020         }
1021
1022         leaf = path->nodes[0];
1023         ptr = btrfs_item_ptr(leaf, path->slots[0],
1024                                  struct btrfs_qgroup_status_item);
1025         btrfs_set_qgroup_status_generation(leaf, ptr, trans->transid);
1026         btrfs_set_qgroup_status_version(leaf, ptr, BTRFS_QGROUP_STATUS_VERSION);
1027         fs_info->qgroup_flags = BTRFS_QGROUP_STATUS_FLAG_ON |
1028                                 BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1029         btrfs_set_qgroup_status_flags(leaf, ptr, fs_info->qgroup_flags);
1030         btrfs_set_qgroup_status_rescan(leaf, ptr, 0);
1031
1032         btrfs_mark_buffer_dirty(leaf);
1033
1034         key.objectid = 0;
1035         key.type = BTRFS_ROOT_REF_KEY;
1036         key.offset = 0;
1037
1038         btrfs_release_path(path);
1039         ret = btrfs_search_slot_for_read(tree_root, &key, path, 1, 0);
1040         if (ret > 0)
1041                 goto out_add_root;
1042         if (ret < 0) {
1043                 btrfs_abort_transaction(trans, ret);
1044                 goto out_free_path;
1045         }
1046
1047         while (1) {
1048                 slot = path->slots[0];
1049                 leaf = path->nodes[0];
1050                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1051
1052                 if (found_key.type == BTRFS_ROOT_REF_KEY) {
1053
1054                         /* Release locks on tree_root before we access quota_root */
1055                         btrfs_release_path(path);
1056
1057                         ret = add_qgroup_item(trans, quota_root,
1058                                               found_key.offset);
1059                         if (ret) {
1060                                 btrfs_abort_transaction(trans, ret);
1061                                 goto out_free_path;
1062                         }
1063
1064                         qgroup = add_qgroup_rb(fs_info, found_key.offset);
1065                         if (IS_ERR(qgroup)) {
1066                                 ret = PTR_ERR(qgroup);
1067                                 btrfs_abort_transaction(trans, ret);
1068                                 goto out_free_path;
1069                         }
1070                         ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1071                         if (ret < 0) {
1072                                 btrfs_abort_transaction(trans, ret);
1073                                 goto out_free_path;
1074                         }
1075                         ret = btrfs_search_slot_for_read(tree_root, &found_key,
1076                                                          path, 1, 0);
1077                         if (ret < 0) {
1078                                 btrfs_abort_transaction(trans, ret);
1079                                 goto out_free_path;
1080                         }
1081                         if (ret > 0) {
1082                                 /*
1083                                  * Shouldn't happen, but in case it does we
1084                                  * don't need to do the btrfs_next_item, just
1085                                  * continue.
1086                                  */
1087                                 continue;
1088                         }
1089                 }
1090                 ret = btrfs_next_item(tree_root, path);
1091                 if (ret < 0) {
1092                         btrfs_abort_transaction(trans, ret);
1093                         goto out_free_path;
1094                 }
1095                 if (ret)
1096                         break;
1097         }
1098
1099 out_add_root:
1100         btrfs_release_path(path);
1101         ret = add_qgroup_item(trans, quota_root, BTRFS_FS_TREE_OBJECTID);
1102         if (ret) {
1103                 btrfs_abort_transaction(trans, ret);
1104                 goto out_free_path;
1105         }
1106
1107         qgroup = add_qgroup_rb(fs_info, BTRFS_FS_TREE_OBJECTID);
1108         if (IS_ERR(qgroup)) {
1109                 ret = PTR_ERR(qgroup);
1110                 btrfs_abort_transaction(trans, ret);
1111                 goto out_free_path;
1112         }
1113         ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1114         if (ret < 0) {
1115                 btrfs_abort_transaction(trans, ret);
1116                 goto out_free_path;
1117         }
1118
1119         ret = btrfs_commit_transaction(trans);
1120         trans = NULL;
1121         if (ret)
1122                 goto out_free_path;
1123
1124         /*
1125          * Set quota enabled flag after committing the transaction, to avoid
1126          * deadlocks on fs_info->qgroup_ioctl_lock with concurrent snapshot
1127          * creation.
1128          */
1129         spin_lock(&fs_info->qgroup_lock);
1130         fs_info->quota_root = quota_root;
1131         set_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1132         spin_unlock(&fs_info->qgroup_lock);
1133
1134         ret = qgroup_rescan_init(fs_info, 0, 1);
1135         if (!ret) {
1136                 qgroup_rescan_zero_tracking(fs_info);
1137                 fs_info->qgroup_rescan_running = true;
1138                 btrfs_queue_work(fs_info->qgroup_rescan_workers,
1139                                  &fs_info->qgroup_rescan_work);
1140         }
1141
1142 out_free_path:
1143         btrfs_free_path(path);
1144 out_free_root:
1145         if (ret)
1146                 btrfs_put_root(quota_root);
1147 out:
1148         if (ret) {
1149                 ulist_free(fs_info->qgroup_ulist);
1150                 fs_info->qgroup_ulist = NULL;
1151                 btrfs_sysfs_del_qgroups(fs_info);
1152         }
1153         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1154         if (ret && trans)
1155                 btrfs_end_transaction(trans);
1156         else if (trans)
1157                 ret = btrfs_end_transaction(trans);
1158         ulist_free(ulist);
1159         return ret;
1160 }
1161
1162 int btrfs_quota_disable(struct btrfs_fs_info *fs_info)
1163 {
1164         struct btrfs_root *quota_root;
1165         struct btrfs_trans_handle *trans = NULL;
1166         int ret = 0;
1167
1168         mutex_lock(&fs_info->qgroup_ioctl_lock);
1169         if (!fs_info->quota_root)
1170                 goto out;
1171         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1172
1173         /*
1174          * 1 For the root item
1175          *
1176          * We should also reserve enough items for the quota tree deletion in
1177          * btrfs_clean_quota_tree but this is not done.
1178          *
1179          * Also, we must always start a transaction without holding the mutex
1180          * qgroup_ioctl_lock, see btrfs_quota_enable().
1181          */
1182         trans = btrfs_start_transaction(fs_info->tree_root, 1);
1183
1184         mutex_lock(&fs_info->qgroup_ioctl_lock);
1185         if (IS_ERR(trans)) {
1186                 ret = PTR_ERR(trans);
1187                 trans = NULL;
1188                 goto out;
1189         }
1190
1191         if (!fs_info->quota_root)
1192                 goto out;
1193
1194         clear_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1195         btrfs_qgroup_wait_for_completion(fs_info, false);
1196         spin_lock(&fs_info->qgroup_lock);
1197         quota_root = fs_info->quota_root;
1198         fs_info->quota_root = NULL;
1199         fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_ON;
1200         spin_unlock(&fs_info->qgroup_lock);
1201
1202         btrfs_free_qgroup_config(fs_info);
1203
1204         ret = btrfs_clean_quota_tree(trans, quota_root);
1205         if (ret) {
1206                 btrfs_abort_transaction(trans, ret);
1207                 goto out;
1208         }
1209
1210         ret = btrfs_del_root(trans, &quota_root->root_key);
1211         if (ret) {
1212                 btrfs_abort_transaction(trans, ret);
1213                 goto out;
1214         }
1215
1216         list_del(&quota_root->dirty_list);
1217
1218         btrfs_tree_lock(quota_root->node);
1219         btrfs_clean_tree_block(quota_root->node);
1220         btrfs_tree_unlock(quota_root->node);
1221         btrfs_free_tree_block(trans, quota_root, quota_root->node, 0, 1);
1222
1223         btrfs_put_root(quota_root);
1224
1225 out:
1226         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1227         if (ret && trans)
1228                 btrfs_end_transaction(trans);
1229         else if (trans)
1230                 ret = btrfs_end_transaction(trans);
1231
1232         return ret;
1233 }
1234
1235 static void qgroup_dirty(struct btrfs_fs_info *fs_info,
1236                          struct btrfs_qgroup *qgroup)
1237 {
1238         if (list_empty(&qgroup->dirty))
1239                 list_add(&qgroup->dirty, &fs_info->dirty_qgroups);
1240 }
1241
1242 /*
1243  * The easy accounting, we're updating qgroup relationship whose child qgroup
1244  * only has exclusive extents.
1245  *
1246  * In this case, all exclusive extents will also be exclusive for parent, so
1247  * excl/rfer just get added/removed.
1248  *
1249  * So is qgroup reservation space, which should also be added/removed to
1250  * parent.
1251  * Or when child tries to release reservation space, parent will underflow its
1252  * reservation (for relationship adding case).
1253  *
1254  * Caller should hold fs_info->qgroup_lock.
1255  */
1256 static int __qgroup_excl_accounting(struct btrfs_fs_info *fs_info,
1257                                     struct ulist *tmp, u64 ref_root,
1258                                     struct btrfs_qgroup *src, int sign)
1259 {
1260         struct btrfs_qgroup *qgroup;
1261         struct btrfs_qgroup_list *glist;
1262         struct ulist_node *unode;
1263         struct ulist_iterator uiter;
1264         u64 num_bytes = src->excl;
1265         int ret = 0;
1266
1267         qgroup = find_qgroup_rb(fs_info, ref_root);
1268         if (!qgroup)
1269                 goto out;
1270
1271         qgroup->rfer += sign * num_bytes;
1272         qgroup->rfer_cmpr += sign * num_bytes;
1273
1274         WARN_ON(sign < 0 && qgroup->excl < num_bytes);
1275         qgroup->excl += sign * num_bytes;
1276         qgroup->excl_cmpr += sign * num_bytes;
1277
1278         if (sign > 0)
1279                 qgroup_rsv_add_by_qgroup(fs_info, qgroup, src);
1280         else
1281                 qgroup_rsv_release_by_qgroup(fs_info, qgroup, src);
1282
1283         qgroup_dirty(fs_info, qgroup);
1284
1285         /* Get all of the parent groups that contain this qgroup */
1286         list_for_each_entry(glist, &qgroup->groups, next_group) {
1287                 ret = ulist_add(tmp, glist->group->qgroupid,
1288                                 qgroup_to_aux(glist->group), GFP_ATOMIC);
1289                 if (ret < 0)
1290                         goto out;
1291         }
1292
1293         /* Iterate all of the parents and adjust their reference counts */
1294         ULIST_ITER_INIT(&uiter);
1295         while ((unode = ulist_next(tmp, &uiter))) {
1296                 qgroup = unode_aux_to_qgroup(unode);
1297                 qgroup->rfer += sign * num_bytes;
1298                 qgroup->rfer_cmpr += sign * num_bytes;
1299                 WARN_ON(sign < 0 && qgroup->excl < num_bytes);
1300                 qgroup->excl += sign * num_bytes;
1301                 if (sign > 0)
1302                         qgroup_rsv_add_by_qgroup(fs_info, qgroup, src);
1303                 else
1304                         qgroup_rsv_release_by_qgroup(fs_info, qgroup, src);
1305                 qgroup->excl_cmpr += sign * num_bytes;
1306                 qgroup_dirty(fs_info, qgroup);
1307
1308                 /* Add any parents of the parents */
1309                 list_for_each_entry(glist, &qgroup->groups, next_group) {
1310                         ret = ulist_add(tmp, glist->group->qgroupid,
1311                                         qgroup_to_aux(glist->group), GFP_ATOMIC);
1312                         if (ret < 0)
1313                                 goto out;
1314                 }
1315         }
1316         ret = 0;
1317 out:
1318         return ret;
1319 }
1320
1321
1322 /*
1323  * Quick path for updating qgroup with only excl refs.
1324  *
1325  * In that case, just update all parent will be enough.
1326  * Or we needs to do a full rescan.
1327  * Caller should also hold fs_info->qgroup_lock.
1328  *
1329  * Return 0 for quick update, return >0 for need to full rescan
1330  * and mark INCONSISTENT flag.
1331  * Return < 0 for other error.
1332  */
1333 static int quick_update_accounting(struct btrfs_fs_info *fs_info,
1334                                    struct ulist *tmp, u64 src, u64 dst,
1335                                    int sign)
1336 {
1337         struct btrfs_qgroup *qgroup;
1338         int ret = 1;
1339         int err = 0;
1340
1341         qgroup = find_qgroup_rb(fs_info, src);
1342         if (!qgroup)
1343                 goto out;
1344         if (qgroup->excl == qgroup->rfer) {
1345                 ret = 0;
1346                 err = __qgroup_excl_accounting(fs_info, tmp, dst,
1347                                                qgroup, sign);
1348                 if (err < 0) {
1349                         ret = err;
1350                         goto out;
1351                 }
1352         }
1353 out:
1354         if (ret)
1355                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1356         return ret;
1357 }
1358
1359 int btrfs_add_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1360                               u64 dst)
1361 {
1362         struct btrfs_fs_info *fs_info = trans->fs_info;
1363         struct btrfs_qgroup *parent;
1364         struct btrfs_qgroup *member;
1365         struct btrfs_qgroup_list *list;
1366         struct ulist *tmp;
1367         unsigned int nofs_flag;
1368         int ret = 0;
1369
1370         /* Check the level of src and dst first */
1371         if (btrfs_qgroup_level(src) >= btrfs_qgroup_level(dst))
1372                 return -EINVAL;
1373
1374         /* We hold a transaction handle open, must do a NOFS allocation. */
1375         nofs_flag = memalloc_nofs_save();
1376         tmp = ulist_alloc(GFP_KERNEL);
1377         memalloc_nofs_restore(nofs_flag);
1378         if (!tmp)
1379                 return -ENOMEM;
1380
1381         mutex_lock(&fs_info->qgroup_ioctl_lock);
1382         if (!fs_info->quota_root) {
1383                 ret = -ENOTCONN;
1384                 goto out;
1385         }
1386         member = find_qgroup_rb(fs_info, src);
1387         parent = find_qgroup_rb(fs_info, dst);
1388         if (!member || !parent) {
1389                 ret = -EINVAL;
1390                 goto out;
1391         }
1392
1393         /* check if such qgroup relation exist firstly */
1394         list_for_each_entry(list, &member->groups, next_group) {
1395                 if (list->group == parent) {
1396                         ret = -EEXIST;
1397                         goto out;
1398                 }
1399         }
1400
1401         ret = add_qgroup_relation_item(trans, src, dst);
1402         if (ret)
1403                 goto out;
1404
1405         ret = add_qgroup_relation_item(trans, dst, src);
1406         if (ret) {
1407                 del_qgroup_relation_item(trans, src, dst);
1408                 goto out;
1409         }
1410
1411         spin_lock(&fs_info->qgroup_lock);
1412         ret = add_relation_rb(fs_info, src, dst);
1413         if (ret < 0) {
1414                 spin_unlock(&fs_info->qgroup_lock);
1415                 goto out;
1416         }
1417         ret = quick_update_accounting(fs_info, tmp, src, dst, 1);
1418         spin_unlock(&fs_info->qgroup_lock);
1419 out:
1420         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1421         ulist_free(tmp);
1422         return ret;
1423 }
1424
1425 static int __del_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1426                                  u64 dst)
1427 {
1428         struct btrfs_fs_info *fs_info = trans->fs_info;
1429         struct btrfs_qgroup *parent;
1430         struct btrfs_qgroup *member;
1431         struct btrfs_qgroup_list *list;
1432         struct ulist *tmp;
1433         bool found = false;
1434         unsigned int nofs_flag;
1435         int ret = 0;
1436         int ret2;
1437
1438         /* We hold a transaction handle open, must do a NOFS allocation. */
1439         nofs_flag = memalloc_nofs_save();
1440         tmp = ulist_alloc(GFP_KERNEL);
1441         memalloc_nofs_restore(nofs_flag);
1442         if (!tmp)
1443                 return -ENOMEM;
1444
1445         if (!fs_info->quota_root) {
1446                 ret = -ENOTCONN;
1447                 goto out;
1448         }
1449
1450         member = find_qgroup_rb(fs_info, src);
1451         parent = find_qgroup_rb(fs_info, dst);
1452         /*
1453          * The parent/member pair doesn't exist, then try to delete the dead
1454          * relation items only.
1455          */
1456         if (!member || !parent)
1457                 goto delete_item;
1458
1459         /* check if such qgroup relation exist firstly */
1460         list_for_each_entry(list, &member->groups, next_group) {
1461                 if (list->group == parent) {
1462                         found = true;
1463                         break;
1464                 }
1465         }
1466
1467 delete_item:
1468         ret = del_qgroup_relation_item(trans, src, dst);
1469         if (ret < 0 && ret != -ENOENT)
1470                 goto out;
1471         ret2 = del_qgroup_relation_item(trans, dst, src);
1472         if (ret2 < 0 && ret2 != -ENOENT)
1473                 goto out;
1474
1475         /* At least one deletion succeeded, return 0 */
1476         if (!ret || !ret2)
1477                 ret = 0;
1478
1479         if (found) {
1480                 spin_lock(&fs_info->qgroup_lock);
1481                 del_relation_rb(fs_info, src, dst);
1482                 ret = quick_update_accounting(fs_info, tmp, src, dst, -1);
1483                 spin_unlock(&fs_info->qgroup_lock);
1484         }
1485 out:
1486         ulist_free(tmp);
1487         return ret;
1488 }
1489
1490 int btrfs_del_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1491                               u64 dst)
1492 {
1493         struct btrfs_fs_info *fs_info = trans->fs_info;
1494         int ret = 0;
1495
1496         mutex_lock(&fs_info->qgroup_ioctl_lock);
1497         ret = __del_qgroup_relation(trans, src, dst);
1498         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1499
1500         return ret;
1501 }
1502
1503 int btrfs_create_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid)
1504 {
1505         struct btrfs_fs_info *fs_info = trans->fs_info;
1506         struct btrfs_root *quota_root;
1507         struct btrfs_qgroup *qgroup;
1508         int ret = 0;
1509
1510         mutex_lock(&fs_info->qgroup_ioctl_lock);
1511         if (!fs_info->quota_root) {
1512                 ret = -ENOTCONN;
1513                 goto out;
1514         }
1515         quota_root = fs_info->quota_root;
1516         qgroup = find_qgroup_rb(fs_info, qgroupid);
1517         if (qgroup) {
1518                 ret = -EEXIST;
1519                 goto out;
1520         }
1521
1522         ret = add_qgroup_item(trans, quota_root, qgroupid);
1523         if (ret)
1524                 goto out;
1525
1526         spin_lock(&fs_info->qgroup_lock);
1527         qgroup = add_qgroup_rb(fs_info, qgroupid);
1528         spin_unlock(&fs_info->qgroup_lock);
1529
1530         if (IS_ERR(qgroup)) {
1531                 ret = PTR_ERR(qgroup);
1532                 goto out;
1533         }
1534         ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1535 out:
1536         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1537         return ret;
1538 }
1539
1540 int btrfs_remove_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid)
1541 {
1542         struct btrfs_fs_info *fs_info = trans->fs_info;
1543         struct btrfs_qgroup *qgroup;
1544         struct btrfs_qgroup_list *list;
1545         int ret = 0;
1546
1547         mutex_lock(&fs_info->qgroup_ioctl_lock);
1548         if (!fs_info->quota_root) {
1549                 ret = -ENOTCONN;
1550                 goto out;
1551         }
1552
1553         qgroup = find_qgroup_rb(fs_info, qgroupid);
1554         if (!qgroup) {
1555                 ret = -ENOENT;
1556                 goto out;
1557         }
1558
1559         /* Check if there are no children of this qgroup */
1560         if (!list_empty(&qgroup->members)) {
1561                 ret = -EBUSY;
1562                 goto out;
1563         }
1564
1565         ret = del_qgroup_item(trans, qgroupid);
1566         if (ret && ret != -ENOENT)
1567                 goto out;
1568
1569         while (!list_empty(&qgroup->groups)) {
1570                 list = list_first_entry(&qgroup->groups,
1571                                         struct btrfs_qgroup_list, next_group);
1572                 ret = __del_qgroup_relation(trans, qgroupid,
1573                                             list->group->qgroupid);
1574                 if (ret)
1575                         goto out;
1576         }
1577
1578         spin_lock(&fs_info->qgroup_lock);
1579         del_qgroup_rb(fs_info, qgroupid);
1580         spin_unlock(&fs_info->qgroup_lock);
1581 out:
1582         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1583         return ret;
1584 }
1585
1586 int btrfs_limit_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid,
1587                        struct btrfs_qgroup_limit *limit)
1588 {
1589         struct btrfs_fs_info *fs_info = trans->fs_info;
1590         struct btrfs_qgroup *qgroup;
1591         int ret = 0;
1592         /* Sometimes we would want to clear the limit on this qgroup.
1593          * To meet this requirement, we treat the -1 as a special value
1594          * which tell kernel to clear the limit on this qgroup.
1595          */
1596         const u64 CLEAR_VALUE = -1;
1597
1598         mutex_lock(&fs_info->qgroup_ioctl_lock);
1599         if (!fs_info->quota_root) {
1600                 ret = -ENOTCONN;
1601                 goto out;
1602         }
1603
1604         qgroup = find_qgroup_rb(fs_info, qgroupid);
1605         if (!qgroup) {
1606                 ret = -ENOENT;
1607                 goto out;
1608         }
1609
1610         spin_lock(&fs_info->qgroup_lock);
1611         if (limit->flags & BTRFS_QGROUP_LIMIT_MAX_RFER) {
1612                 if (limit->max_rfer == CLEAR_VALUE) {
1613                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_MAX_RFER;
1614                         limit->flags &= ~BTRFS_QGROUP_LIMIT_MAX_RFER;
1615                         qgroup->max_rfer = 0;
1616                 } else {
1617                         qgroup->max_rfer = limit->max_rfer;
1618                 }
1619         }
1620         if (limit->flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) {
1621                 if (limit->max_excl == CLEAR_VALUE) {
1622                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_MAX_EXCL;
1623                         limit->flags &= ~BTRFS_QGROUP_LIMIT_MAX_EXCL;
1624                         qgroup->max_excl = 0;
1625                 } else {
1626                         qgroup->max_excl = limit->max_excl;
1627                 }
1628         }
1629         if (limit->flags & BTRFS_QGROUP_LIMIT_RSV_RFER) {
1630                 if (limit->rsv_rfer == CLEAR_VALUE) {
1631                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_RSV_RFER;
1632                         limit->flags &= ~BTRFS_QGROUP_LIMIT_RSV_RFER;
1633                         qgroup->rsv_rfer = 0;
1634                 } else {
1635                         qgroup->rsv_rfer = limit->rsv_rfer;
1636                 }
1637         }
1638         if (limit->flags & BTRFS_QGROUP_LIMIT_RSV_EXCL) {
1639                 if (limit->rsv_excl == CLEAR_VALUE) {
1640                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_RSV_EXCL;
1641                         limit->flags &= ~BTRFS_QGROUP_LIMIT_RSV_EXCL;
1642                         qgroup->rsv_excl = 0;
1643                 } else {
1644                         qgroup->rsv_excl = limit->rsv_excl;
1645                 }
1646         }
1647         qgroup->lim_flags |= limit->flags;
1648
1649         spin_unlock(&fs_info->qgroup_lock);
1650
1651         ret = update_qgroup_limit_item(trans, qgroup);
1652         if (ret) {
1653                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1654                 btrfs_info(fs_info, "unable to update quota limit for %llu",
1655                        qgroupid);
1656         }
1657
1658 out:
1659         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1660         return ret;
1661 }
1662
1663 int btrfs_qgroup_trace_extent_nolock(struct btrfs_fs_info *fs_info,
1664                                 struct btrfs_delayed_ref_root *delayed_refs,
1665                                 struct btrfs_qgroup_extent_record *record)
1666 {
1667         struct rb_node **p = &delayed_refs->dirty_extent_root.rb_node;
1668         struct rb_node *parent_node = NULL;
1669         struct btrfs_qgroup_extent_record *entry;
1670         u64 bytenr = record->bytenr;
1671
1672         lockdep_assert_held(&delayed_refs->lock);
1673         trace_btrfs_qgroup_trace_extent(fs_info, record);
1674
1675         while (*p) {
1676                 parent_node = *p;
1677                 entry = rb_entry(parent_node, struct btrfs_qgroup_extent_record,
1678                                  node);
1679                 if (bytenr < entry->bytenr) {
1680                         p = &(*p)->rb_left;
1681                 } else if (bytenr > entry->bytenr) {
1682                         p = &(*p)->rb_right;
1683                 } else {
1684                         if (record->data_rsv && !entry->data_rsv) {
1685                                 entry->data_rsv = record->data_rsv;
1686                                 entry->data_rsv_refroot =
1687                                         record->data_rsv_refroot;
1688                         }
1689                         return 1;
1690                 }
1691         }
1692
1693         rb_link_node(&record->node, parent_node, p);
1694         rb_insert_color(&record->node, &delayed_refs->dirty_extent_root);
1695         return 0;
1696 }
1697
1698 int btrfs_qgroup_trace_extent_post(struct btrfs_fs_info *fs_info,
1699                                    struct btrfs_qgroup_extent_record *qrecord)
1700 {
1701         struct ulist *old_root;
1702         u64 bytenr = qrecord->bytenr;
1703         int ret;
1704
1705         ret = btrfs_find_all_roots(NULL, fs_info, bytenr, 0, &old_root, false);
1706         if (ret < 0) {
1707                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1708                 btrfs_warn(fs_info,
1709 "error accounting new delayed refs extent (err code: %d), quota inconsistent",
1710                         ret);
1711                 return 0;
1712         }
1713
1714         /*
1715          * Here we don't need to get the lock of
1716          * trans->transaction->delayed_refs, since inserted qrecord won't
1717          * be deleted, only qrecord->node may be modified (new qrecord insert)
1718          *
1719          * So modifying qrecord->old_roots is safe here
1720          */
1721         qrecord->old_roots = old_root;
1722         return 0;
1723 }
1724
1725 int btrfs_qgroup_trace_extent(struct btrfs_trans_handle *trans, u64 bytenr,
1726                               u64 num_bytes, gfp_t gfp_flag)
1727 {
1728         struct btrfs_fs_info *fs_info = trans->fs_info;
1729         struct btrfs_qgroup_extent_record *record;
1730         struct btrfs_delayed_ref_root *delayed_refs;
1731         int ret;
1732
1733         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)
1734             || bytenr == 0 || num_bytes == 0)
1735                 return 0;
1736         record = kzalloc(sizeof(*record), gfp_flag);
1737         if (!record)
1738                 return -ENOMEM;
1739
1740         delayed_refs = &trans->transaction->delayed_refs;
1741         record->bytenr = bytenr;
1742         record->num_bytes = num_bytes;
1743         record->old_roots = NULL;
1744
1745         spin_lock(&delayed_refs->lock);
1746         ret = btrfs_qgroup_trace_extent_nolock(fs_info, delayed_refs, record);
1747         spin_unlock(&delayed_refs->lock);
1748         if (ret > 0) {
1749                 kfree(record);
1750                 return 0;
1751         }
1752         return btrfs_qgroup_trace_extent_post(fs_info, record);
1753 }
1754
1755 int btrfs_qgroup_trace_leaf_items(struct btrfs_trans_handle *trans,
1756                                   struct extent_buffer *eb)
1757 {
1758         struct btrfs_fs_info *fs_info = trans->fs_info;
1759         int nr = btrfs_header_nritems(eb);
1760         int i, extent_type, ret;
1761         struct btrfs_key key;
1762         struct btrfs_file_extent_item *fi;
1763         u64 bytenr, num_bytes;
1764
1765         /* We can be called directly from walk_up_proc() */
1766         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
1767                 return 0;
1768
1769         for (i = 0; i < nr; i++) {
1770                 btrfs_item_key_to_cpu(eb, &key, i);
1771
1772                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1773                         continue;
1774
1775                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
1776                 /* filter out non qgroup-accountable extents  */
1777                 extent_type = btrfs_file_extent_type(eb, fi);
1778
1779                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
1780                         continue;
1781
1782                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
1783                 if (!bytenr)
1784                         continue;
1785
1786                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
1787
1788                 ret = btrfs_qgroup_trace_extent(trans, bytenr, num_bytes,
1789                                                 GFP_NOFS);
1790                 if (ret)
1791                         return ret;
1792         }
1793         cond_resched();
1794         return 0;
1795 }
1796
1797 /*
1798  * Walk up the tree from the bottom, freeing leaves and any interior
1799  * nodes which have had all slots visited. If a node (leaf or
1800  * interior) is freed, the node above it will have it's slot
1801  * incremented. The root node will never be freed.
1802  *
1803  * At the end of this function, we should have a path which has all
1804  * slots incremented to the next position for a search. If we need to
1805  * read a new node it will be NULL and the node above it will have the
1806  * correct slot selected for a later read.
1807  *
1808  * If we increment the root nodes slot counter past the number of
1809  * elements, 1 is returned to signal completion of the search.
1810  */
1811 static int adjust_slots_upwards(struct btrfs_path *path, int root_level)
1812 {
1813         int level = 0;
1814         int nr, slot;
1815         struct extent_buffer *eb;
1816
1817         if (root_level == 0)
1818                 return 1;
1819
1820         while (level <= root_level) {
1821                 eb = path->nodes[level];
1822                 nr = btrfs_header_nritems(eb);
1823                 path->slots[level]++;
1824                 slot = path->slots[level];
1825                 if (slot >= nr || level == 0) {
1826                         /*
1827                          * Don't free the root -  we will detect this
1828                          * condition after our loop and return a
1829                          * positive value for caller to stop walking the tree.
1830                          */
1831                         if (level != root_level) {
1832                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
1833                                 path->locks[level] = 0;
1834
1835                                 free_extent_buffer(eb);
1836                                 path->nodes[level] = NULL;
1837                                 path->slots[level] = 0;
1838                         }
1839                 } else {
1840                         /*
1841                          * We have a valid slot to walk back down
1842                          * from. Stop here so caller can process these
1843                          * new nodes.
1844                          */
1845                         break;
1846                 }
1847
1848                 level++;
1849         }
1850
1851         eb = path->nodes[root_level];
1852         if (path->slots[root_level] >= btrfs_header_nritems(eb))
1853                 return 1;
1854
1855         return 0;
1856 }
1857
1858 /*
1859  * Helper function to trace a subtree tree block swap.
1860  *
1861  * The swap will happen in highest tree block, but there may be a lot of
1862  * tree blocks involved.
1863  *
1864  * For example:
1865  *  OO = Old tree blocks
1866  *  NN = New tree blocks allocated during balance
1867  *
1868  *           File tree (257)                  Reloc tree for 257
1869  * L2              OO                                NN
1870  *               /    \                            /    \
1871  * L1          OO      OO (a)                    OO      NN (a)
1872  *            / \     / \                       / \     / \
1873  * L0       OO   OO OO   OO                   OO   OO NN   NN
1874  *                  (b)  (c)                          (b)  (c)
1875  *
1876  * When calling qgroup_trace_extent_swap(), we will pass:
1877  * @src_eb = OO(a)
1878  * @dst_path = [ nodes[1] = NN(a), nodes[0] = NN(c) ]
1879  * @dst_level = 0
1880  * @root_level = 1
1881  *
1882  * In that case, qgroup_trace_extent_swap() will search from OO(a) to
1883  * reach OO(c), then mark both OO(c) and NN(c) as qgroup dirty.
1884  *
1885  * The main work of qgroup_trace_extent_swap() can be split into 3 parts:
1886  *
1887  * 1) Tree search from @src_eb
1888  *    It should acts as a simplified btrfs_search_slot().
1889  *    The key for search can be extracted from @dst_path->nodes[dst_level]
1890  *    (first key).
1891  *
1892  * 2) Mark the final tree blocks in @src_path and @dst_path qgroup dirty
1893  *    NOTE: In above case, OO(a) and NN(a) won't be marked qgroup dirty.
1894  *    They should be marked during previous (@dst_level = 1) iteration.
1895  *
1896  * 3) Mark file extents in leaves dirty
1897  *    We don't have good way to pick out new file extents only.
1898  *    So we still follow the old method by scanning all file extents in
1899  *    the leave.
1900  *
1901  * This function can free us from keeping two paths, thus later we only need
1902  * to care about how to iterate all new tree blocks in reloc tree.
1903  */
1904 static int qgroup_trace_extent_swap(struct btrfs_trans_handle* trans,
1905                                     struct extent_buffer *src_eb,
1906                                     struct btrfs_path *dst_path,
1907                                     int dst_level, int root_level,
1908                                     bool trace_leaf)
1909 {
1910         struct btrfs_key key;
1911         struct btrfs_path *src_path;
1912         struct btrfs_fs_info *fs_info = trans->fs_info;
1913         u32 nodesize = fs_info->nodesize;
1914         int cur_level = root_level;
1915         int ret;
1916
1917         BUG_ON(dst_level > root_level);
1918         /* Level mismatch */
1919         if (btrfs_header_level(src_eb) != root_level)
1920                 return -EINVAL;
1921
1922         src_path = btrfs_alloc_path();
1923         if (!src_path) {
1924                 ret = -ENOMEM;
1925                 goto out;
1926         }
1927
1928         if (dst_level)
1929                 btrfs_node_key_to_cpu(dst_path->nodes[dst_level], &key, 0);
1930         else
1931                 btrfs_item_key_to_cpu(dst_path->nodes[dst_level], &key, 0);
1932
1933         /* For src_path */
1934         atomic_inc(&src_eb->refs);
1935         src_path->nodes[root_level] = src_eb;
1936         src_path->slots[root_level] = dst_path->slots[root_level];
1937         src_path->locks[root_level] = 0;
1938
1939         /* A simplified version of btrfs_search_slot() */
1940         while (cur_level >= dst_level) {
1941                 struct btrfs_key src_key;
1942                 struct btrfs_key dst_key;
1943
1944                 if (src_path->nodes[cur_level] == NULL) {
1945                         struct extent_buffer *eb;
1946                         int parent_slot;
1947
1948                         eb = src_path->nodes[cur_level + 1];
1949                         parent_slot = src_path->slots[cur_level + 1];
1950
1951                         eb = btrfs_read_node_slot(eb, parent_slot);
1952                         if (IS_ERR(eb)) {
1953                                 ret = PTR_ERR(eb);
1954                                 goto out;
1955                         }
1956
1957                         src_path->nodes[cur_level] = eb;
1958
1959                         btrfs_tree_read_lock(eb);
1960                         src_path->locks[cur_level] = BTRFS_READ_LOCK;
1961                 }
1962
1963                 src_path->slots[cur_level] = dst_path->slots[cur_level];
1964                 if (cur_level) {
1965                         btrfs_node_key_to_cpu(dst_path->nodes[cur_level],
1966                                         &dst_key, dst_path->slots[cur_level]);
1967                         btrfs_node_key_to_cpu(src_path->nodes[cur_level],
1968                                         &src_key, src_path->slots[cur_level]);
1969                 } else {
1970                         btrfs_item_key_to_cpu(dst_path->nodes[cur_level],
1971                                         &dst_key, dst_path->slots[cur_level]);
1972                         btrfs_item_key_to_cpu(src_path->nodes[cur_level],
1973                                         &src_key, src_path->slots[cur_level]);
1974                 }
1975                 /* Content mismatch, something went wrong */
1976                 if (btrfs_comp_cpu_keys(&dst_key, &src_key)) {
1977                         ret = -ENOENT;
1978                         goto out;
1979                 }
1980                 cur_level--;
1981         }
1982
1983         /*
1984          * Now both @dst_path and @src_path have been populated, record the tree
1985          * blocks for qgroup accounting.
1986          */
1987         ret = btrfs_qgroup_trace_extent(trans, src_path->nodes[dst_level]->start,
1988                         nodesize, GFP_NOFS);
1989         if (ret < 0)
1990                 goto out;
1991         ret = btrfs_qgroup_trace_extent(trans,
1992                         dst_path->nodes[dst_level]->start,
1993                         nodesize, GFP_NOFS);
1994         if (ret < 0)
1995                 goto out;
1996
1997         /* Record leaf file extents */
1998         if (dst_level == 0 && trace_leaf) {
1999                 ret = btrfs_qgroup_trace_leaf_items(trans, src_path->nodes[0]);
2000                 if (ret < 0)
2001                         goto out;
2002                 ret = btrfs_qgroup_trace_leaf_items(trans, dst_path->nodes[0]);
2003         }
2004 out:
2005         btrfs_free_path(src_path);
2006         return ret;
2007 }
2008
2009 /*
2010  * Helper function to do recursive generation-aware depth-first search, to
2011  * locate all new tree blocks in a subtree of reloc tree.
2012  *
2013  * E.g. (OO = Old tree blocks, NN = New tree blocks, whose gen == last_snapshot)
2014  *         reloc tree
2015  * L2         NN (a)
2016  *          /    \
2017  * L1    OO        NN (b)
2018  *      /  \      /  \
2019  * L0  OO  OO    OO  NN
2020  *               (c) (d)
2021  * If we pass:
2022  * @dst_path = [ nodes[1] = NN(b), nodes[0] = NULL ],
2023  * @cur_level = 1
2024  * @root_level = 1
2025  *
2026  * We will iterate through tree blocks NN(b), NN(d) and info qgroup to trace
2027  * above tree blocks along with their counter parts in file tree.
2028  * While during search, old tree blocks OO(c) will be skipped as tree block swap
2029  * won't affect OO(c).
2030  */
2031 static int qgroup_trace_new_subtree_blocks(struct btrfs_trans_handle* trans,
2032                                            struct extent_buffer *src_eb,
2033                                            struct btrfs_path *dst_path,
2034                                            int cur_level, int root_level,
2035                                            u64 last_snapshot, bool trace_leaf)
2036 {
2037         struct btrfs_fs_info *fs_info = trans->fs_info;
2038         struct extent_buffer *eb;
2039         bool need_cleanup = false;
2040         int ret = 0;
2041         int i;
2042
2043         /* Level sanity check */
2044         if (cur_level < 0 || cur_level >= BTRFS_MAX_LEVEL - 1 ||
2045             root_level < 0 || root_level >= BTRFS_MAX_LEVEL - 1 ||
2046             root_level < cur_level) {
2047                 btrfs_err_rl(fs_info,
2048                         "%s: bad levels, cur_level=%d root_level=%d",
2049                         __func__, cur_level, root_level);
2050                 return -EUCLEAN;
2051         }
2052
2053         /* Read the tree block if needed */
2054         if (dst_path->nodes[cur_level] == NULL) {
2055                 int parent_slot;
2056                 u64 child_gen;
2057
2058                 /*
2059                  * dst_path->nodes[root_level] must be initialized before
2060                  * calling this function.
2061                  */
2062                 if (cur_level == root_level) {
2063                         btrfs_err_rl(fs_info,
2064         "%s: dst_path->nodes[%d] not initialized, root_level=%d cur_level=%d",
2065                                 __func__, root_level, root_level, cur_level);
2066                         return -EUCLEAN;
2067                 }
2068
2069                 /*
2070                  * We need to get child blockptr/gen from parent before we can
2071                  * read it.
2072                   */
2073                 eb = dst_path->nodes[cur_level + 1];
2074                 parent_slot = dst_path->slots[cur_level + 1];
2075                 child_gen = btrfs_node_ptr_generation(eb, parent_slot);
2076
2077                 /* This node is old, no need to trace */
2078                 if (child_gen < last_snapshot)
2079                         goto out;
2080
2081                 eb = btrfs_read_node_slot(eb, parent_slot);
2082                 if (IS_ERR(eb)) {
2083                         ret = PTR_ERR(eb);
2084                         goto out;
2085                 }
2086
2087                 dst_path->nodes[cur_level] = eb;
2088                 dst_path->slots[cur_level] = 0;
2089
2090                 btrfs_tree_read_lock(eb);
2091                 dst_path->locks[cur_level] = BTRFS_READ_LOCK;
2092                 need_cleanup = true;
2093         }
2094
2095         /* Now record this tree block and its counter part for qgroups */
2096         ret = qgroup_trace_extent_swap(trans, src_eb, dst_path, cur_level,
2097                                        root_level, trace_leaf);
2098         if (ret < 0)
2099                 goto cleanup;
2100
2101         eb = dst_path->nodes[cur_level];
2102
2103         if (cur_level > 0) {
2104                 /* Iterate all child tree blocks */
2105                 for (i = 0; i < btrfs_header_nritems(eb); i++) {
2106                         /* Skip old tree blocks as they won't be swapped */
2107                         if (btrfs_node_ptr_generation(eb, i) < last_snapshot)
2108                                 continue;
2109                         dst_path->slots[cur_level] = i;
2110
2111                         /* Recursive call (at most 7 times) */
2112                         ret = qgroup_trace_new_subtree_blocks(trans, src_eb,
2113                                         dst_path, cur_level - 1, root_level,
2114                                         last_snapshot, trace_leaf);
2115                         if (ret < 0)
2116                                 goto cleanup;
2117                 }
2118         }
2119
2120 cleanup:
2121         if (need_cleanup) {
2122                 /* Clean up */
2123                 btrfs_tree_unlock_rw(dst_path->nodes[cur_level],
2124                                      dst_path->locks[cur_level]);
2125                 free_extent_buffer(dst_path->nodes[cur_level]);
2126                 dst_path->nodes[cur_level] = NULL;
2127                 dst_path->slots[cur_level] = 0;
2128                 dst_path->locks[cur_level] = 0;
2129         }
2130 out:
2131         return ret;
2132 }
2133
2134 static int qgroup_trace_subtree_swap(struct btrfs_trans_handle *trans,
2135                                 struct extent_buffer *src_eb,
2136                                 struct extent_buffer *dst_eb,
2137                                 u64 last_snapshot, bool trace_leaf)
2138 {
2139         struct btrfs_fs_info *fs_info = trans->fs_info;
2140         struct btrfs_path *dst_path = NULL;
2141         int level;
2142         int ret;
2143
2144         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2145                 return 0;
2146
2147         /* Wrong parameter order */
2148         if (btrfs_header_generation(src_eb) > btrfs_header_generation(dst_eb)) {
2149                 btrfs_err_rl(fs_info,
2150                 "%s: bad parameter order, src_gen=%llu dst_gen=%llu", __func__,
2151                              btrfs_header_generation(src_eb),
2152                              btrfs_header_generation(dst_eb));
2153                 return -EUCLEAN;
2154         }
2155
2156         if (!extent_buffer_uptodate(src_eb) || !extent_buffer_uptodate(dst_eb)) {
2157                 ret = -EIO;
2158                 goto out;
2159         }
2160
2161         level = btrfs_header_level(dst_eb);
2162         dst_path = btrfs_alloc_path();
2163         if (!dst_path) {
2164                 ret = -ENOMEM;
2165                 goto out;
2166         }
2167         /* For dst_path */
2168         atomic_inc(&dst_eb->refs);
2169         dst_path->nodes[level] = dst_eb;
2170         dst_path->slots[level] = 0;
2171         dst_path->locks[level] = 0;
2172
2173         /* Do the generation aware breadth-first search */
2174         ret = qgroup_trace_new_subtree_blocks(trans, src_eb, dst_path, level,
2175                                               level, last_snapshot, trace_leaf);
2176         if (ret < 0)
2177                 goto out;
2178         ret = 0;
2179
2180 out:
2181         btrfs_free_path(dst_path);
2182         if (ret < 0)
2183                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2184         return ret;
2185 }
2186
2187 int btrfs_qgroup_trace_subtree(struct btrfs_trans_handle *trans,
2188                                struct extent_buffer *root_eb,
2189                                u64 root_gen, int root_level)
2190 {
2191         struct btrfs_fs_info *fs_info = trans->fs_info;
2192         int ret = 0;
2193         int level;
2194         struct extent_buffer *eb = root_eb;
2195         struct btrfs_path *path = NULL;
2196
2197         BUG_ON(root_level < 0 || root_level >= BTRFS_MAX_LEVEL);
2198         BUG_ON(root_eb == NULL);
2199
2200         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2201                 return 0;
2202
2203         if (!extent_buffer_uptodate(root_eb)) {
2204                 ret = btrfs_read_buffer(root_eb, root_gen, root_level, NULL);
2205                 if (ret)
2206                         goto out;
2207         }
2208
2209         if (root_level == 0) {
2210                 ret = btrfs_qgroup_trace_leaf_items(trans, root_eb);
2211                 goto out;
2212         }
2213
2214         path = btrfs_alloc_path();
2215         if (!path)
2216                 return -ENOMEM;
2217
2218         /*
2219          * Walk down the tree.  Missing extent blocks are filled in as
2220          * we go. Metadata is accounted every time we read a new
2221          * extent block.
2222          *
2223          * When we reach a leaf, we account for file extent items in it,
2224          * walk back up the tree (adjusting slot pointers as we go)
2225          * and restart the search process.
2226          */
2227         atomic_inc(&root_eb->refs);     /* For path */
2228         path->nodes[root_level] = root_eb;
2229         path->slots[root_level] = 0;
2230         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
2231 walk_down:
2232         level = root_level;
2233         while (level >= 0) {
2234                 if (path->nodes[level] == NULL) {
2235                         int parent_slot;
2236                         u64 child_bytenr;
2237
2238                         /*
2239                          * We need to get child blockptr from parent before we
2240                          * can read it.
2241                           */
2242                         eb = path->nodes[level + 1];
2243                         parent_slot = path->slots[level + 1];
2244                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
2245
2246                         eb = btrfs_read_node_slot(eb, parent_slot);
2247                         if (IS_ERR(eb)) {
2248                                 ret = PTR_ERR(eb);
2249                                 goto out;
2250                         }
2251
2252                         path->nodes[level] = eb;
2253                         path->slots[level] = 0;
2254
2255                         btrfs_tree_read_lock(eb);
2256                         path->locks[level] = BTRFS_READ_LOCK;
2257
2258                         ret = btrfs_qgroup_trace_extent(trans, child_bytenr,
2259                                                         fs_info->nodesize,
2260                                                         GFP_NOFS);
2261                         if (ret)
2262                                 goto out;
2263                 }
2264
2265                 if (level == 0) {
2266                         ret = btrfs_qgroup_trace_leaf_items(trans,
2267                                                             path->nodes[level]);
2268                         if (ret)
2269                                 goto out;
2270
2271                         /* Nonzero return here means we completed our search */
2272                         ret = adjust_slots_upwards(path, root_level);
2273                         if (ret)
2274                                 break;
2275
2276                         /* Restart search with new slots */
2277                         goto walk_down;
2278                 }
2279
2280                 level--;
2281         }
2282
2283         ret = 0;
2284 out:
2285         btrfs_free_path(path);
2286
2287         return ret;
2288 }
2289
2290 #define UPDATE_NEW      0
2291 #define UPDATE_OLD      1
2292 /*
2293  * Walk all of the roots that points to the bytenr and adjust their refcnts.
2294  */
2295 static int qgroup_update_refcnt(struct btrfs_fs_info *fs_info,
2296                                 struct ulist *roots, struct ulist *tmp,
2297                                 struct ulist *qgroups, u64 seq, int update_old)
2298 {
2299         struct ulist_node *unode;
2300         struct ulist_iterator uiter;
2301         struct ulist_node *tmp_unode;
2302         struct ulist_iterator tmp_uiter;
2303         struct btrfs_qgroup *qg;
2304         int ret = 0;
2305
2306         if (!roots)
2307                 return 0;
2308         ULIST_ITER_INIT(&uiter);
2309         while ((unode = ulist_next(roots, &uiter))) {
2310                 qg = find_qgroup_rb(fs_info, unode->val);
2311                 if (!qg)
2312                         continue;
2313
2314                 ulist_reinit(tmp);
2315                 ret = ulist_add(qgroups, qg->qgroupid, qgroup_to_aux(qg),
2316                                 GFP_ATOMIC);
2317                 if (ret < 0)
2318                         return ret;
2319                 ret = ulist_add(tmp, qg->qgroupid, qgroup_to_aux(qg), GFP_ATOMIC);
2320                 if (ret < 0)
2321                         return ret;
2322                 ULIST_ITER_INIT(&tmp_uiter);
2323                 while ((tmp_unode = ulist_next(tmp, &tmp_uiter))) {
2324                         struct btrfs_qgroup_list *glist;
2325
2326                         qg = unode_aux_to_qgroup(tmp_unode);
2327                         if (update_old)
2328                                 btrfs_qgroup_update_old_refcnt(qg, seq, 1);
2329                         else
2330                                 btrfs_qgroup_update_new_refcnt(qg, seq, 1);
2331                         list_for_each_entry(glist, &qg->groups, next_group) {
2332                                 ret = ulist_add(qgroups, glist->group->qgroupid,
2333                                                 qgroup_to_aux(glist->group),
2334                                                 GFP_ATOMIC);
2335                                 if (ret < 0)
2336                                         return ret;
2337                                 ret = ulist_add(tmp, glist->group->qgroupid,
2338                                                 qgroup_to_aux(glist->group),
2339                                                 GFP_ATOMIC);
2340                                 if (ret < 0)
2341                                         return ret;
2342                         }
2343                 }
2344         }
2345         return 0;
2346 }
2347
2348 /*
2349  * Update qgroup rfer/excl counters.
2350  * Rfer update is easy, codes can explain themselves.
2351  *
2352  * Excl update is tricky, the update is split into 2 parts.
2353  * Part 1: Possible exclusive <-> sharing detect:
2354  *      |       A       |       !A      |
2355  *  -------------------------------------
2356  *  B   |       *       |       -       |
2357  *  -------------------------------------
2358  *  !B  |       +       |       **      |
2359  *  -------------------------------------
2360  *
2361  * Conditions:
2362  * A:   cur_old_roots < nr_old_roots    (not exclusive before)
2363  * !A:  cur_old_roots == nr_old_roots   (possible exclusive before)
2364  * B:   cur_new_roots < nr_new_roots    (not exclusive now)
2365  * !B:  cur_new_roots == nr_new_roots   (possible exclusive now)
2366  *
2367  * Results:
2368  * +: Possible sharing -> exclusive     -: Possible exclusive -> sharing
2369  * *: Definitely not changed.           **: Possible unchanged.
2370  *
2371  * For !A and !B condition, the exception is cur_old/new_roots == 0 case.
2372  *
2373  * To make the logic clear, we first use condition A and B to split
2374  * combination into 4 results.
2375  *
2376  * Then, for result "+" and "-", check old/new_roots == 0 case, as in them
2377  * only on variant maybe 0.
2378  *
2379  * Lastly, check result **, since there are 2 variants maybe 0, split them
2380  * again(2x2).
2381  * But this time we don't need to consider other things, the codes and logic
2382  * is easy to understand now.
2383  */
2384 static int qgroup_update_counters(struct btrfs_fs_info *fs_info,
2385                                   struct ulist *qgroups,
2386                                   u64 nr_old_roots,
2387                                   u64 nr_new_roots,
2388                                   u64 num_bytes, u64 seq)
2389 {
2390         struct ulist_node *unode;
2391         struct ulist_iterator uiter;
2392         struct btrfs_qgroup *qg;
2393         u64 cur_new_count, cur_old_count;
2394
2395         ULIST_ITER_INIT(&uiter);
2396         while ((unode = ulist_next(qgroups, &uiter))) {
2397                 bool dirty = false;
2398
2399                 qg = unode_aux_to_qgroup(unode);
2400                 cur_old_count = btrfs_qgroup_get_old_refcnt(qg, seq);
2401                 cur_new_count = btrfs_qgroup_get_new_refcnt(qg, seq);
2402
2403                 trace_qgroup_update_counters(fs_info, qg, cur_old_count,
2404                                              cur_new_count);
2405
2406                 /* Rfer update part */
2407                 if (cur_old_count == 0 && cur_new_count > 0) {
2408                         qg->rfer += num_bytes;
2409                         qg->rfer_cmpr += num_bytes;
2410                         dirty = true;
2411                 }
2412                 if (cur_old_count > 0 && cur_new_count == 0) {
2413                         qg->rfer -= num_bytes;
2414                         qg->rfer_cmpr -= num_bytes;
2415                         dirty = true;
2416                 }
2417
2418                 /* Excl update part */
2419                 /* Exclusive/none -> shared case */
2420                 if (cur_old_count == nr_old_roots &&
2421                     cur_new_count < nr_new_roots) {
2422                         /* Exclusive -> shared */
2423                         if (cur_old_count != 0) {
2424                                 qg->excl -= num_bytes;
2425                                 qg->excl_cmpr -= num_bytes;
2426                                 dirty = true;
2427                         }
2428                 }
2429
2430                 /* Shared -> exclusive/none case */
2431                 if (cur_old_count < nr_old_roots &&
2432                     cur_new_count == nr_new_roots) {
2433                         /* Shared->exclusive */
2434                         if (cur_new_count != 0) {
2435                                 qg->excl += num_bytes;
2436                                 qg->excl_cmpr += num_bytes;
2437                                 dirty = true;
2438                         }
2439                 }
2440
2441                 /* Exclusive/none -> exclusive/none case */
2442                 if (cur_old_count == nr_old_roots &&
2443                     cur_new_count == nr_new_roots) {
2444                         if (cur_old_count == 0) {
2445                                 /* None -> exclusive/none */
2446
2447                                 if (cur_new_count != 0) {
2448                                         /* None -> exclusive */
2449                                         qg->excl += num_bytes;
2450                                         qg->excl_cmpr += num_bytes;
2451                                         dirty = true;
2452                                 }
2453                                 /* None -> none, nothing changed */
2454                         } else {
2455                                 /* Exclusive -> exclusive/none */
2456
2457                                 if (cur_new_count == 0) {
2458                                         /* Exclusive -> none */
2459                                         qg->excl -= num_bytes;
2460                                         qg->excl_cmpr -= num_bytes;
2461                                         dirty = true;
2462                                 }
2463                                 /* Exclusive -> exclusive, nothing changed */
2464                         }
2465                 }
2466
2467                 if (dirty)
2468                         qgroup_dirty(fs_info, qg);
2469         }
2470         return 0;
2471 }
2472
2473 /*
2474  * Check if the @roots potentially is a list of fs tree roots
2475  *
2476  * Return 0 for definitely not a fs/subvol tree roots ulist
2477  * Return 1 for possible fs/subvol tree roots in the list (considering an empty
2478  *          one as well)
2479  */
2480 static int maybe_fs_roots(struct ulist *roots)
2481 {
2482         struct ulist_node *unode;
2483         struct ulist_iterator uiter;
2484
2485         /* Empty one, still possible for fs roots */
2486         if (!roots || roots->nnodes == 0)
2487                 return 1;
2488
2489         ULIST_ITER_INIT(&uiter);
2490         unode = ulist_next(roots, &uiter);
2491         if (!unode)
2492                 return 1;
2493
2494         /*
2495          * If it contains fs tree roots, then it must belong to fs/subvol
2496          * trees.
2497          * If it contains a non-fs tree, it won't be shared with fs/subvol trees.
2498          */
2499         return is_fstree(unode->val);
2500 }
2501
2502 int btrfs_qgroup_account_extent(struct btrfs_trans_handle *trans, u64 bytenr,
2503                                 u64 num_bytes, struct ulist *old_roots,
2504                                 struct ulist *new_roots)
2505 {
2506         struct btrfs_fs_info *fs_info = trans->fs_info;
2507         struct ulist *qgroups = NULL;
2508         struct ulist *tmp = NULL;
2509         u64 seq;
2510         u64 nr_new_roots = 0;
2511         u64 nr_old_roots = 0;
2512         int ret = 0;
2513
2514         /*
2515          * If quotas get disabled meanwhile, the resouces need to be freed and
2516          * we can't just exit here.
2517          */
2518         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2519                 goto out_free;
2520
2521         if (new_roots) {
2522                 if (!maybe_fs_roots(new_roots))
2523                         goto out_free;
2524                 nr_new_roots = new_roots->nnodes;
2525         }
2526         if (old_roots) {
2527                 if (!maybe_fs_roots(old_roots))
2528                         goto out_free;
2529                 nr_old_roots = old_roots->nnodes;
2530         }
2531
2532         /* Quick exit, either not fs tree roots, or won't affect any qgroup */
2533         if (nr_old_roots == 0 && nr_new_roots == 0)
2534                 goto out_free;
2535
2536         BUG_ON(!fs_info->quota_root);
2537
2538         trace_btrfs_qgroup_account_extent(fs_info, trans->transid, bytenr,
2539                                         num_bytes, nr_old_roots, nr_new_roots);
2540
2541         qgroups = ulist_alloc(GFP_NOFS);
2542         if (!qgroups) {
2543                 ret = -ENOMEM;
2544                 goto out_free;
2545         }
2546         tmp = ulist_alloc(GFP_NOFS);
2547         if (!tmp) {
2548                 ret = -ENOMEM;
2549                 goto out_free;
2550         }
2551
2552         mutex_lock(&fs_info->qgroup_rescan_lock);
2553         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
2554                 if (fs_info->qgroup_rescan_progress.objectid <= bytenr) {
2555                         mutex_unlock(&fs_info->qgroup_rescan_lock);
2556                         ret = 0;
2557                         goto out_free;
2558                 }
2559         }
2560         mutex_unlock(&fs_info->qgroup_rescan_lock);
2561
2562         spin_lock(&fs_info->qgroup_lock);
2563         seq = fs_info->qgroup_seq;
2564
2565         /* Update old refcnts using old_roots */
2566         ret = qgroup_update_refcnt(fs_info, old_roots, tmp, qgroups, seq,
2567                                    UPDATE_OLD);
2568         if (ret < 0)
2569                 goto out;
2570
2571         /* Update new refcnts using new_roots */
2572         ret = qgroup_update_refcnt(fs_info, new_roots, tmp, qgroups, seq,
2573                                    UPDATE_NEW);
2574         if (ret < 0)
2575                 goto out;
2576
2577         qgroup_update_counters(fs_info, qgroups, nr_old_roots, nr_new_roots,
2578                                num_bytes, seq);
2579
2580         /*
2581          * Bump qgroup_seq to avoid seq overlap
2582          */
2583         fs_info->qgroup_seq += max(nr_old_roots, nr_new_roots) + 1;
2584 out:
2585         spin_unlock(&fs_info->qgroup_lock);
2586 out_free:
2587         ulist_free(tmp);
2588         ulist_free(qgroups);
2589         ulist_free(old_roots);
2590         ulist_free(new_roots);
2591         return ret;
2592 }
2593
2594 int btrfs_qgroup_account_extents(struct btrfs_trans_handle *trans)
2595 {
2596         struct btrfs_fs_info *fs_info = trans->fs_info;
2597         struct btrfs_qgroup_extent_record *record;
2598         struct btrfs_delayed_ref_root *delayed_refs;
2599         struct ulist *new_roots = NULL;
2600         struct rb_node *node;
2601         u64 num_dirty_extents = 0;
2602         u64 qgroup_to_skip;
2603         int ret = 0;
2604
2605         delayed_refs = &trans->transaction->delayed_refs;
2606         qgroup_to_skip = delayed_refs->qgroup_to_skip;
2607         while ((node = rb_first(&delayed_refs->dirty_extent_root))) {
2608                 record = rb_entry(node, struct btrfs_qgroup_extent_record,
2609                                   node);
2610
2611                 num_dirty_extents++;
2612                 trace_btrfs_qgroup_account_extents(fs_info, record);
2613
2614                 if (!ret) {
2615                         /*
2616                          * Old roots should be searched when inserting qgroup
2617                          * extent record
2618                          */
2619                         if (WARN_ON(!record->old_roots)) {
2620                                 /* Search commit root to find old_roots */
2621                                 ret = btrfs_find_all_roots(NULL, fs_info,
2622                                                 record->bytenr, 0,
2623                                                 &record->old_roots, false);
2624                                 if (ret < 0)
2625                                         goto cleanup;
2626                         }
2627
2628                         /* Free the reserved data space */
2629                         btrfs_qgroup_free_refroot(fs_info,
2630                                         record->data_rsv_refroot,
2631                                         record->data_rsv,
2632                                         BTRFS_QGROUP_RSV_DATA);
2633                         /*
2634                          * Use SEQ_LAST as time_seq to do special search, which
2635                          * doesn't lock tree or delayed_refs and search current
2636                          * root. It's safe inside commit_transaction().
2637                          */
2638                         ret = btrfs_find_all_roots(trans, fs_info,
2639                                 record->bytenr, SEQ_LAST, &new_roots, false);
2640                         if (ret < 0)
2641                                 goto cleanup;
2642                         if (qgroup_to_skip) {
2643                                 ulist_del(new_roots, qgroup_to_skip, 0);
2644                                 ulist_del(record->old_roots, qgroup_to_skip,
2645                                           0);
2646                         }
2647                         ret = btrfs_qgroup_account_extent(trans, record->bytenr,
2648                                                           record->num_bytes,
2649                                                           record->old_roots,
2650                                                           new_roots);
2651                         record->old_roots = NULL;
2652                         new_roots = NULL;
2653                 }
2654 cleanup:
2655                 ulist_free(record->old_roots);
2656                 ulist_free(new_roots);
2657                 new_roots = NULL;
2658                 rb_erase(node, &delayed_refs->dirty_extent_root);
2659                 kfree(record);
2660
2661         }
2662         trace_qgroup_num_dirty_extents(fs_info, trans->transid,
2663                                        num_dirty_extents);
2664         return ret;
2665 }
2666
2667 /*
2668  * called from commit_transaction. Writes all changed qgroups to disk.
2669  */
2670 int btrfs_run_qgroups(struct btrfs_trans_handle *trans)
2671 {
2672         struct btrfs_fs_info *fs_info = trans->fs_info;
2673         int ret = 0;
2674
2675         if (!fs_info->quota_root)
2676                 return ret;
2677
2678         spin_lock(&fs_info->qgroup_lock);
2679         while (!list_empty(&fs_info->dirty_qgroups)) {
2680                 struct btrfs_qgroup *qgroup;
2681                 qgroup = list_first_entry(&fs_info->dirty_qgroups,
2682                                           struct btrfs_qgroup, dirty);
2683                 list_del_init(&qgroup->dirty);
2684                 spin_unlock(&fs_info->qgroup_lock);
2685                 ret = update_qgroup_info_item(trans, qgroup);
2686                 if (ret)
2687                         fs_info->qgroup_flags |=
2688                                         BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2689                 ret = update_qgroup_limit_item(trans, qgroup);
2690                 if (ret)
2691                         fs_info->qgroup_flags |=
2692                                         BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2693                 spin_lock(&fs_info->qgroup_lock);
2694         }
2695         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2696                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_ON;
2697         else
2698                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_ON;
2699         spin_unlock(&fs_info->qgroup_lock);
2700
2701         ret = update_qgroup_status_item(trans);
2702         if (ret)
2703                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2704
2705         return ret;
2706 }
2707
2708 /*
2709  * Copy the accounting information between qgroups. This is necessary
2710  * when a snapshot or a subvolume is created. Throwing an error will
2711  * cause a transaction abort so we take extra care here to only error
2712  * when a readonly fs is a reasonable outcome.
2713  */
2714 int btrfs_qgroup_inherit(struct btrfs_trans_handle *trans, u64 srcid,
2715                          u64 objectid, struct btrfs_qgroup_inherit *inherit)
2716 {
2717         int ret = 0;
2718         int i;
2719         u64 *i_qgroups;
2720         bool committing = false;
2721         struct btrfs_fs_info *fs_info = trans->fs_info;
2722         struct btrfs_root *quota_root;
2723         struct btrfs_qgroup *srcgroup;
2724         struct btrfs_qgroup *dstgroup;
2725         bool need_rescan = false;
2726         u32 level_size = 0;
2727         u64 nums;
2728
2729         /*
2730          * There are only two callers of this function.
2731          *
2732          * One in create_subvol() in the ioctl context, which needs to hold
2733          * the qgroup_ioctl_lock.
2734          *
2735          * The other one in create_pending_snapshot() where no other qgroup
2736          * code can modify the fs as they all need to either start a new trans
2737          * or hold a trans handler, thus we don't need to hold
2738          * qgroup_ioctl_lock.
2739          * This would avoid long and complex lock chain and make lockdep happy.
2740          */
2741         spin_lock(&fs_info->trans_lock);
2742         if (trans->transaction->state == TRANS_STATE_COMMIT_DOING)
2743                 committing = true;
2744         spin_unlock(&fs_info->trans_lock);
2745
2746         if (!committing)
2747                 mutex_lock(&fs_info->qgroup_ioctl_lock);
2748         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2749                 goto out;
2750
2751         quota_root = fs_info->quota_root;
2752         if (!quota_root) {
2753                 ret = -EINVAL;
2754                 goto out;
2755         }
2756
2757         if (inherit) {
2758                 i_qgroups = (u64 *)(inherit + 1);
2759                 nums = inherit->num_qgroups + 2 * inherit->num_ref_copies +
2760                        2 * inherit->num_excl_copies;
2761                 for (i = 0; i < nums; ++i) {
2762                         srcgroup = find_qgroup_rb(fs_info, *i_qgroups);
2763
2764                         /*
2765                          * Zero out invalid groups so we can ignore
2766                          * them later.
2767                          */
2768                         if (!srcgroup ||
2769                             ((srcgroup->qgroupid >> 48) <= (objectid >> 48)))
2770                                 *i_qgroups = 0ULL;
2771
2772                         ++i_qgroups;
2773                 }
2774         }
2775
2776         /*
2777          * create a tracking group for the subvol itself
2778          */
2779         ret = add_qgroup_item(trans, quota_root, objectid);
2780         if (ret)
2781                 goto out;
2782
2783         /*
2784          * add qgroup to all inherited groups
2785          */
2786         if (inherit) {
2787                 i_qgroups = (u64 *)(inherit + 1);
2788                 for (i = 0; i < inherit->num_qgroups; ++i, ++i_qgroups) {
2789                         if (*i_qgroups == 0)
2790                                 continue;
2791                         ret = add_qgroup_relation_item(trans, objectid,
2792                                                        *i_qgroups);
2793                         if (ret && ret != -EEXIST)
2794                                 goto out;
2795                         ret = add_qgroup_relation_item(trans, *i_qgroups,
2796                                                        objectid);
2797                         if (ret && ret != -EEXIST)
2798                                 goto out;
2799                 }
2800                 ret = 0;
2801         }
2802
2803
2804         spin_lock(&fs_info->qgroup_lock);
2805
2806         dstgroup = add_qgroup_rb(fs_info, objectid);
2807         if (IS_ERR(dstgroup)) {
2808                 ret = PTR_ERR(dstgroup);
2809                 goto unlock;
2810         }
2811
2812         if (inherit && inherit->flags & BTRFS_QGROUP_INHERIT_SET_LIMITS) {
2813                 dstgroup->lim_flags = inherit->lim.flags;
2814                 dstgroup->max_rfer = inherit->lim.max_rfer;
2815                 dstgroup->max_excl = inherit->lim.max_excl;
2816                 dstgroup->rsv_rfer = inherit->lim.rsv_rfer;
2817                 dstgroup->rsv_excl = inherit->lim.rsv_excl;
2818
2819                 ret = update_qgroup_limit_item(trans, dstgroup);
2820                 if (ret) {
2821                         fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2822                         btrfs_info(fs_info,
2823                                    "unable to update quota limit for %llu",
2824                                    dstgroup->qgroupid);
2825                         goto unlock;
2826                 }
2827         }
2828
2829         if (srcid) {
2830                 srcgroup = find_qgroup_rb(fs_info, srcid);
2831                 if (!srcgroup)
2832                         goto unlock;
2833
2834                 /*
2835                  * We call inherit after we clone the root in order to make sure
2836                  * our counts don't go crazy, so at this point the only
2837                  * difference between the two roots should be the root node.
2838                  */
2839                 level_size = fs_info->nodesize;
2840                 dstgroup->rfer = srcgroup->rfer;
2841                 dstgroup->rfer_cmpr = srcgroup->rfer_cmpr;
2842                 dstgroup->excl = level_size;
2843                 dstgroup->excl_cmpr = level_size;
2844                 srcgroup->excl = level_size;
2845                 srcgroup->excl_cmpr = level_size;
2846
2847                 /* inherit the limit info */
2848                 dstgroup->lim_flags = srcgroup->lim_flags;
2849                 dstgroup->max_rfer = srcgroup->max_rfer;
2850                 dstgroup->max_excl = srcgroup->max_excl;
2851                 dstgroup->rsv_rfer = srcgroup->rsv_rfer;
2852                 dstgroup->rsv_excl = srcgroup->rsv_excl;
2853
2854                 qgroup_dirty(fs_info, dstgroup);
2855                 qgroup_dirty(fs_info, srcgroup);
2856         }
2857
2858         if (!inherit)
2859                 goto unlock;
2860
2861         i_qgroups = (u64 *)(inherit + 1);
2862         for (i = 0; i < inherit->num_qgroups; ++i) {
2863                 if (*i_qgroups) {
2864                         ret = add_relation_rb(fs_info, objectid, *i_qgroups);
2865                         if (ret)
2866                                 goto unlock;
2867                 }
2868                 ++i_qgroups;
2869
2870                 /*
2871                  * If we're doing a snapshot, and adding the snapshot to a new
2872                  * qgroup, the numbers are guaranteed to be incorrect.
2873                  */
2874                 if (srcid)
2875                         need_rescan = true;
2876         }
2877
2878         for (i = 0; i <  inherit->num_ref_copies; ++i, i_qgroups += 2) {
2879                 struct btrfs_qgroup *src;
2880                 struct btrfs_qgroup *dst;
2881
2882                 if (!i_qgroups[0] || !i_qgroups[1])
2883                         continue;
2884
2885                 src = find_qgroup_rb(fs_info, i_qgroups[0]);
2886                 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
2887
2888                 if (!src || !dst) {
2889                         ret = -EINVAL;
2890                         goto unlock;
2891                 }
2892
2893                 dst->rfer = src->rfer - level_size;
2894                 dst->rfer_cmpr = src->rfer_cmpr - level_size;
2895
2896                 /* Manually tweaking numbers certainly needs a rescan */
2897                 need_rescan = true;
2898         }
2899         for (i = 0; i <  inherit->num_excl_copies; ++i, i_qgroups += 2) {
2900                 struct btrfs_qgroup *src;
2901                 struct btrfs_qgroup *dst;
2902
2903                 if (!i_qgroups[0] || !i_qgroups[1])
2904                         continue;
2905
2906                 src = find_qgroup_rb(fs_info, i_qgroups[0]);
2907                 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
2908
2909                 if (!src || !dst) {
2910                         ret = -EINVAL;
2911                         goto unlock;
2912                 }
2913
2914                 dst->excl = src->excl + level_size;
2915                 dst->excl_cmpr = src->excl_cmpr + level_size;
2916                 need_rescan = true;
2917         }
2918
2919 unlock:
2920         spin_unlock(&fs_info->qgroup_lock);
2921         if (!ret)
2922                 ret = btrfs_sysfs_add_one_qgroup(fs_info, dstgroup);
2923 out:
2924         if (!committing)
2925                 mutex_unlock(&fs_info->qgroup_ioctl_lock);
2926         if (need_rescan)
2927                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2928         return ret;
2929 }
2930
2931 static bool qgroup_check_limits(const struct btrfs_qgroup *qg, u64 num_bytes)
2932 {
2933         if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_RFER) &&
2934             qgroup_rsv_total(qg) + (s64)qg->rfer + num_bytes > qg->max_rfer)
2935                 return false;
2936
2937         if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) &&
2938             qgroup_rsv_total(qg) + (s64)qg->excl + num_bytes > qg->max_excl)
2939                 return false;
2940
2941         return true;
2942 }
2943
2944 static int qgroup_reserve(struct btrfs_root *root, u64 num_bytes, bool enforce,
2945                           enum btrfs_qgroup_rsv_type type)
2946 {
2947         struct btrfs_qgroup *qgroup;
2948         struct btrfs_fs_info *fs_info = root->fs_info;
2949         u64 ref_root = root->root_key.objectid;
2950         int ret = 0;
2951         struct ulist_node *unode;
2952         struct ulist_iterator uiter;
2953
2954         if (!is_fstree(ref_root))
2955                 return 0;
2956
2957         if (num_bytes == 0)
2958                 return 0;
2959
2960         if (test_bit(BTRFS_FS_QUOTA_OVERRIDE, &fs_info->flags) &&
2961             capable(CAP_SYS_RESOURCE))
2962                 enforce = false;
2963
2964         spin_lock(&fs_info->qgroup_lock);
2965         if (!fs_info->quota_root)
2966                 goto out;
2967
2968         qgroup = find_qgroup_rb(fs_info, ref_root);
2969         if (!qgroup)
2970                 goto out;
2971
2972         /*
2973          * in a first step, we check all affected qgroups if any limits would
2974          * be exceeded
2975          */
2976         ulist_reinit(fs_info->qgroup_ulist);
2977         ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
2978                         qgroup_to_aux(qgroup), GFP_ATOMIC);
2979         if (ret < 0)
2980                 goto out;
2981         ULIST_ITER_INIT(&uiter);
2982         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
2983                 struct btrfs_qgroup *qg;
2984                 struct btrfs_qgroup_list *glist;
2985
2986                 qg = unode_aux_to_qgroup(unode);
2987
2988                 if (enforce && !qgroup_check_limits(qg, num_bytes)) {
2989                         ret = -EDQUOT;
2990                         goto out;
2991                 }
2992
2993                 list_for_each_entry(glist, &qg->groups, next_group) {
2994                         ret = ulist_add(fs_info->qgroup_ulist,
2995                                         glist->group->qgroupid,
2996                                         qgroup_to_aux(glist->group), GFP_ATOMIC);
2997                         if (ret < 0)
2998                                 goto out;
2999                 }
3000         }
3001         ret = 0;
3002         /*
3003          * no limits exceeded, now record the reservation into all qgroups
3004          */
3005         ULIST_ITER_INIT(&uiter);
3006         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
3007                 struct btrfs_qgroup *qg;
3008
3009                 qg = unode_aux_to_qgroup(unode);
3010
3011                 qgroup_rsv_add(fs_info, qg, num_bytes, type);
3012         }
3013
3014 out:
3015         spin_unlock(&fs_info->qgroup_lock);
3016         return ret;
3017 }
3018
3019 /*
3020  * Free @num_bytes of reserved space with @type for qgroup.  (Normally level 0
3021  * qgroup).
3022  *
3023  * Will handle all higher level qgroup too.
3024  *
3025  * NOTE: If @num_bytes is (u64)-1, this means to free all bytes of this qgroup.
3026  * This special case is only used for META_PERTRANS type.
3027  */
3028 void btrfs_qgroup_free_refroot(struct btrfs_fs_info *fs_info,
3029                                u64 ref_root, u64 num_bytes,
3030                                enum btrfs_qgroup_rsv_type type)
3031 {
3032         struct btrfs_qgroup *qgroup;
3033         struct ulist_node *unode;
3034         struct ulist_iterator uiter;
3035         int ret = 0;
3036
3037         if (!is_fstree(ref_root))
3038                 return;
3039
3040         if (num_bytes == 0)
3041                 return;
3042
3043         if (num_bytes == (u64)-1 && type != BTRFS_QGROUP_RSV_META_PERTRANS) {
3044                 WARN(1, "%s: Invalid type to free", __func__);
3045                 return;
3046         }
3047         spin_lock(&fs_info->qgroup_lock);
3048
3049         if (!fs_info->quota_root)
3050                 goto out;
3051
3052         qgroup = find_qgroup_rb(fs_info, ref_root);
3053         if (!qgroup)
3054                 goto out;
3055
3056         if (num_bytes == (u64)-1)
3057                 /*
3058                  * We're freeing all pertrans rsv, get reserved value from
3059                  * level 0 qgroup as real num_bytes to free.
3060                  */
3061                 num_bytes = qgroup->rsv.values[type];
3062
3063         ulist_reinit(fs_info->qgroup_ulist);
3064         ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
3065                         qgroup_to_aux(qgroup), GFP_ATOMIC);
3066         if (ret < 0)
3067                 goto out;
3068         ULIST_ITER_INIT(&uiter);
3069         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
3070                 struct btrfs_qgroup *qg;
3071                 struct btrfs_qgroup_list *glist;
3072
3073                 qg = unode_aux_to_qgroup(unode);
3074
3075                 qgroup_rsv_release(fs_info, qg, num_bytes, type);
3076
3077                 list_for_each_entry(glist, &qg->groups, next_group) {
3078                         ret = ulist_add(fs_info->qgroup_ulist,
3079                                         glist->group->qgroupid,
3080                                         qgroup_to_aux(glist->group), GFP_ATOMIC);
3081                         if (ret < 0)
3082                                 goto out;
3083                 }
3084         }
3085
3086 out:
3087         spin_unlock(&fs_info->qgroup_lock);
3088 }
3089
3090 /*
3091  * Check if the leaf is the last leaf. Which means all node pointers
3092  * are at their last position.
3093  */
3094 static bool is_last_leaf(struct btrfs_path *path)
3095 {
3096         int i;
3097
3098         for (i = 1; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
3099                 if (path->slots[i] != btrfs_header_nritems(path->nodes[i]) - 1)
3100                         return false;
3101         }
3102         return true;
3103 }
3104
3105 /*
3106  * returns < 0 on error, 0 when more leafs are to be scanned.
3107  * returns 1 when done.
3108  */
3109 static int qgroup_rescan_leaf(struct btrfs_trans_handle *trans,
3110                               struct btrfs_path *path)
3111 {
3112         struct btrfs_fs_info *fs_info = trans->fs_info;
3113         struct btrfs_key found;
3114         struct extent_buffer *scratch_leaf = NULL;
3115         struct ulist *roots = NULL;
3116         u64 num_bytes;
3117         bool done;
3118         int slot;
3119         int ret;
3120
3121         mutex_lock(&fs_info->qgroup_rescan_lock);
3122         ret = btrfs_search_slot_for_read(fs_info->extent_root,
3123                                          &fs_info->qgroup_rescan_progress,
3124                                          path, 1, 0);
3125
3126         btrfs_debug(fs_info,
3127                 "current progress key (%llu %u %llu), search_slot ret %d",
3128                 fs_info->qgroup_rescan_progress.objectid,
3129                 fs_info->qgroup_rescan_progress.type,
3130                 fs_info->qgroup_rescan_progress.offset, ret);
3131
3132         if (ret) {
3133                 /*
3134                  * The rescan is about to end, we will not be scanning any
3135                  * further blocks. We cannot unset the RESCAN flag here, because
3136                  * we want to commit the transaction if everything went well.
3137                  * To make the live accounting work in this phase, we set our
3138                  * scan progress pointer such that every real extent objectid
3139                  * will be smaller.
3140                  */
3141                 fs_info->qgroup_rescan_progress.objectid = (u64)-1;
3142                 btrfs_release_path(path);
3143                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3144                 return ret;
3145         }
3146         done = is_last_leaf(path);
3147
3148         btrfs_item_key_to_cpu(path->nodes[0], &found,
3149                               btrfs_header_nritems(path->nodes[0]) - 1);
3150         fs_info->qgroup_rescan_progress.objectid = found.objectid + 1;
3151
3152         scratch_leaf = btrfs_clone_extent_buffer(path->nodes[0]);
3153         if (!scratch_leaf) {
3154                 ret = -ENOMEM;
3155                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3156                 goto out;
3157         }
3158         slot = path->slots[0];
3159         btrfs_release_path(path);
3160         mutex_unlock(&fs_info->qgroup_rescan_lock);
3161
3162         for (; slot < btrfs_header_nritems(scratch_leaf); ++slot) {
3163                 btrfs_item_key_to_cpu(scratch_leaf, &found, slot);
3164                 if (found.type != BTRFS_EXTENT_ITEM_KEY &&
3165                     found.type != BTRFS_METADATA_ITEM_KEY)
3166                         continue;
3167                 if (found.type == BTRFS_METADATA_ITEM_KEY)
3168                         num_bytes = fs_info->nodesize;
3169                 else
3170                         num_bytes = found.offset;
3171
3172                 ret = btrfs_find_all_roots(NULL, fs_info, found.objectid, 0,
3173                                            &roots, false);
3174                 if (ret < 0)
3175                         goto out;
3176                 /* For rescan, just pass old_roots as NULL */
3177                 ret = btrfs_qgroup_account_extent(trans, found.objectid,
3178                                                   num_bytes, NULL, roots);
3179                 if (ret < 0)
3180                         goto out;
3181         }
3182 out:
3183         if (scratch_leaf)
3184                 free_extent_buffer(scratch_leaf);
3185
3186         if (done && !ret) {
3187                 ret = 1;
3188                 fs_info->qgroup_rescan_progress.objectid = (u64)-1;
3189         }
3190         return ret;
3191 }
3192
3193 static void btrfs_qgroup_rescan_worker(struct btrfs_work *work)
3194 {
3195         struct btrfs_fs_info *fs_info = container_of(work, struct btrfs_fs_info,
3196                                                      qgroup_rescan_work);
3197         struct btrfs_path *path;
3198         struct btrfs_trans_handle *trans = NULL;
3199         int err = -ENOMEM;
3200         int ret = 0;
3201
3202         path = btrfs_alloc_path();
3203         if (!path)
3204                 goto out;
3205         /*
3206          * Rescan should only search for commit root, and any later difference
3207          * should be recorded by qgroup
3208          */
3209         path->search_commit_root = 1;
3210         path->skip_locking = 1;
3211
3212         err = 0;
3213         while (!err && !btrfs_fs_closing(fs_info)) {
3214                 trans = btrfs_start_transaction(fs_info->fs_root, 0);
3215                 if (IS_ERR(trans)) {
3216                         err = PTR_ERR(trans);
3217                         break;
3218                 }
3219                 if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
3220                         err = -EINTR;
3221                 } else {
3222                         err = qgroup_rescan_leaf(trans, path);
3223                 }
3224                 if (err > 0)
3225                         btrfs_commit_transaction(trans);
3226                 else
3227                         btrfs_end_transaction(trans);
3228         }
3229
3230 out:
3231         btrfs_free_path(path);
3232
3233         mutex_lock(&fs_info->qgroup_rescan_lock);
3234         if (err > 0 &&
3235             fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT) {
3236                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
3237         } else if (err < 0) {
3238                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
3239         }
3240         mutex_unlock(&fs_info->qgroup_rescan_lock);
3241
3242         /*
3243          * only update status, since the previous part has already updated the
3244          * qgroup info.
3245          */
3246         trans = btrfs_start_transaction(fs_info->quota_root, 1);
3247         if (IS_ERR(trans)) {
3248                 err = PTR_ERR(trans);
3249                 trans = NULL;
3250                 btrfs_err(fs_info,
3251                           "fail to start transaction for status update: %d",
3252                           err);
3253         }
3254
3255         mutex_lock(&fs_info->qgroup_rescan_lock);
3256         if (!btrfs_fs_closing(fs_info))
3257                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3258         if (trans) {
3259                 ret = update_qgroup_status_item(trans);
3260                 if (ret < 0) {
3261                         err = ret;
3262                         btrfs_err(fs_info, "fail to update qgroup status: %d",
3263                                   err);
3264                 }
3265         }
3266         fs_info->qgroup_rescan_running = false;
3267         complete_all(&fs_info->qgroup_rescan_completion);
3268         mutex_unlock(&fs_info->qgroup_rescan_lock);
3269
3270         if (!trans)
3271                 return;
3272
3273         btrfs_end_transaction(trans);
3274
3275         if (btrfs_fs_closing(fs_info)) {
3276                 btrfs_info(fs_info, "qgroup scan paused");
3277         } else if (err >= 0) {
3278                 btrfs_info(fs_info, "qgroup scan completed%s",
3279                         err > 0 ? " (inconsistency flag cleared)" : "");
3280         } else {
3281                 btrfs_err(fs_info, "qgroup scan failed with %d", err);
3282         }
3283 }
3284
3285 /*
3286  * Checks that (a) no rescan is running and (b) quota is enabled. Allocates all
3287  * memory required for the rescan context.
3288  */
3289 static int
3290 qgroup_rescan_init(struct btrfs_fs_info *fs_info, u64 progress_objectid,
3291                    int init_flags)
3292 {
3293         int ret = 0;
3294
3295         if (!init_flags) {
3296                 /* we're resuming qgroup rescan at mount time */
3297                 if (!(fs_info->qgroup_flags &
3298                       BTRFS_QGROUP_STATUS_FLAG_RESCAN)) {
3299                         btrfs_warn(fs_info,
3300                         "qgroup rescan init failed, qgroup rescan is not queued");
3301                         ret = -EINVAL;
3302                 } else if (!(fs_info->qgroup_flags &
3303                              BTRFS_QGROUP_STATUS_FLAG_ON)) {
3304                         btrfs_warn(fs_info,
3305                         "qgroup rescan init failed, qgroup is not enabled");
3306                         ret = -EINVAL;
3307                 }
3308
3309                 if (ret)
3310                         return ret;
3311         }
3312
3313         mutex_lock(&fs_info->qgroup_rescan_lock);
3314
3315         if (init_flags) {
3316                 if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
3317                         btrfs_warn(fs_info,
3318                                    "qgroup rescan is already in progress");
3319                         ret = -EINPROGRESS;
3320                 } else if (!(fs_info->qgroup_flags &
3321                              BTRFS_QGROUP_STATUS_FLAG_ON)) {
3322                         btrfs_warn(fs_info,
3323                         "qgroup rescan init failed, qgroup is not enabled");
3324                         ret = -EINVAL;
3325                 }
3326
3327                 if (ret) {
3328                         mutex_unlock(&fs_info->qgroup_rescan_lock);
3329                         return ret;
3330                 }
3331                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3332         }
3333
3334         memset(&fs_info->qgroup_rescan_progress, 0,
3335                 sizeof(fs_info->qgroup_rescan_progress));
3336         fs_info->qgroup_rescan_progress.objectid = progress_objectid;
3337         init_completion(&fs_info->qgroup_rescan_completion);
3338         mutex_unlock(&fs_info->qgroup_rescan_lock);
3339
3340         btrfs_init_work(&fs_info->qgroup_rescan_work,
3341                         btrfs_qgroup_rescan_worker, NULL, NULL);
3342         return 0;
3343 }
3344
3345 static void
3346 qgroup_rescan_zero_tracking(struct btrfs_fs_info *fs_info)
3347 {
3348         struct rb_node *n;
3349         struct btrfs_qgroup *qgroup;
3350
3351         spin_lock(&fs_info->qgroup_lock);
3352         /* clear all current qgroup tracking information */
3353         for (n = rb_first(&fs_info->qgroup_tree); n; n = rb_next(n)) {
3354                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
3355                 qgroup->rfer = 0;
3356                 qgroup->rfer_cmpr = 0;
3357                 qgroup->excl = 0;
3358                 qgroup->excl_cmpr = 0;
3359                 qgroup_dirty(fs_info, qgroup);
3360         }
3361         spin_unlock(&fs_info->qgroup_lock);
3362 }
3363
3364 int
3365 btrfs_qgroup_rescan(struct btrfs_fs_info *fs_info)
3366 {
3367         int ret = 0;
3368         struct btrfs_trans_handle *trans;
3369
3370         ret = qgroup_rescan_init(fs_info, 0, 1);
3371         if (ret)
3372                 return ret;
3373
3374         /*
3375          * We have set the rescan_progress to 0, which means no more
3376          * delayed refs will be accounted by btrfs_qgroup_account_ref.
3377          * However, btrfs_qgroup_account_ref may be right after its call
3378          * to btrfs_find_all_roots, in which case it would still do the
3379          * accounting.
3380          * To solve this, we're committing the transaction, which will
3381          * ensure we run all delayed refs and only after that, we are
3382          * going to clear all tracking information for a clean start.
3383          */
3384
3385         trans = btrfs_join_transaction(fs_info->fs_root);
3386         if (IS_ERR(trans)) {
3387                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3388                 return PTR_ERR(trans);
3389         }
3390         ret = btrfs_commit_transaction(trans);
3391         if (ret) {
3392                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3393                 return ret;
3394         }
3395
3396         qgroup_rescan_zero_tracking(fs_info);
3397
3398         mutex_lock(&fs_info->qgroup_rescan_lock);
3399         fs_info->qgroup_rescan_running = true;
3400         btrfs_queue_work(fs_info->qgroup_rescan_workers,
3401                          &fs_info->qgroup_rescan_work);
3402         mutex_unlock(&fs_info->qgroup_rescan_lock);
3403
3404         return 0;
3405 }
3406
3407 int btrfs_qgroup_wait_for_completion(struct btrfs_fs_info *fs_info,
3408                                      bool interruptible)
3409 {
3410         int running;
3411         int ret = 0;
3412
3413         mutex_lock(&fs_info->qgroup_rescan_lock);
3414         running = fs_info->qgroup_rescan_running;
3415         mutex_unlock(&fs_info->qgroup_rescan_lock);
3416
3417         if (!running)
3418                 return 0;
3419
3420         if (interruptible)
3421                 ret = wait_for_completion_interruptible(
3422                                         &fs_info->qgroup_rescan_completion);
3423         else
3424                 wait_for_completion(&fs_info->qgroup_rescan_completion);
3425
3426         return ret;
3427 }
3428
3429 /*
3430  * this is only called from open_ctree where we're still single threaded, thus
3431  * locking is omitted here.
3432  */
3433 void
3434 btrfs_qgroup_rescan_resume(struct btrfs_fs_info *fs_info)
3435 {
3436         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
3437                 mutex_lock(&fs_info->qgroup_rescan_lock);
3438                 fs_info->qgroup_rescan_running = true;
3439                 btrfs_queue_work(fs_info->qgroup_rescan_workers,
3440                                  &fs_info->qgroup_rescan_work);
3441                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3442         }
3443 }
3444
3445 #define rbtree_iterate_from_safe(node, next, start)                             \
3446        for (node = start; node && ({ next = rb_next(node); 1;}); node = next)
3447
3448 static int qgroup_unreserve_range(struct btrfs_inode *inode,
3449                                   struct extent_changeset *reserved, u64 start,
3450                                   u64 len)
3451 {
3452         struct rb_node *node;
3453         struct rb_node *next;
3454         struct ulist_node *entry;
3455         int ret = 0;
3456
3457         node = reserved->range_changed.root.rb_node;
3458         if (!node)
3459                 return 0;
3460         while (node) {
3461                 entry = rb_entry(node, struct ulist_node, rb_node);
3462                 if (entry->val < start)
3463                         node = node->rb_right;
3464                 else
3465                         node = node->rb_left;
3466         }
3467
3468         if (entry->val > start && rb_prev(&entry->rb_node))
3469                 entry = rb_entry(rb_prev(&entry->rb_node), struct ulist_node,
3470                                  rb_node);
3471
3472         rbtree_iterate_from_safe(node, next, &entry->rb_node) {
3473                 u64 entry_start;
3474                 u64 entry_end;
3475                 u64 entry_len;
3476                 int clear_ret;
3477
3478                 entry = rb_entry(node, struct ulist_node, rb_node);
3479                 entry_start = entry->val;
3480                 entry_end = entry->aux;
3481                 entry_len = entry_end - entry_start + 1;
3482
3483                 if (entry_start >= start + len)
3484                         break;
3485                 if (entry_start + entry_len <= start)
3486                         continue;
3487                 /*
3488                  * Now the entry is in [start, start + len), revert the
3489                  * EXTENT_QGROUP_RESERVED bit.
3490                  */
3491                 clear_ret = clear_extent_bits(&inode->io_tree, entry_start,
3492                                               entry_end, EXTENT_QGROUP_RESERVED);
3493                 if (!ret && clear_ret < 0)
3494                         ret = clear_ret;
3495
3496                 ulist_del(&reserved->range_changed, entry->val, entry->aux);
3497                 if (likely(reserved->bytes_changed >= entry_len)) {
3498                         reserved->bytes_changed -= entry_len;
3499                 } else {
3500                         WARN_ON(1);
3501                         reserved->bytes_changed = 0;
3502                 }
3503         }
3504
3505         return ret;
3506 }
3507
3508 /*
3509  * Try to free some space for qgroup.
3510  *
3511  * For qgroup, there are only 3 ways to free qgroup space:
3512  * - Flush nodatacow write
3513  *   Any nodatacow write will free its reserved data space at run_delalloc_range().
3514  *   In theory, we should only flush nodatacow inodes, but it's not yet
3515  *   possible, so we need to flush the whole root.
3516  *
3517  * - Wait for ordered extents
3518  *   When ordered extents are finished, their reserved metadata is finally
3519  *   converted to per_trans status, which can be freed by later commit
3520  *   transaction.
3521  *
3522  * - Commit transaction
3523  *   This would free the meta_per_trans space.
3524  *   In theory this shouldn't provide much space, but any more qgroup space
3525  *   is needed.
3526  */
3527 static int try_flush_qgroup(struct btrfs_root *root)
3528 {
3529         struct btrfs_trans_handle *trans;
3530         int ret;
3531         bool can_commit = true;
3532
3533         /*
3534          * We don't want to run flush again and again, so if there is a running
3535          * one, we won't try to start a new flush, but exit directly.
3536          */
3537         if (test_and_set_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state)) {
3538                 wait_event(root->qgroup_flush_wait,
3539                         !test_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state));
3540                 return 0;
3541         }
3542
3543         /*
3544          * If current process holds a transaction, we shouldn't flush, as we
3545          * assume all space reservation happens before a transaction handle is
3546          * held.
3547          *
3548          * But there are cases like btrfs_delayed_item_reserve_metadata() where
3549          * we try to reserve space with one transction handle already held.
3550          * In that case we can't commit transaction, but at least try to end it
3551          * and hope the started data writes can free some space.
3552          */
3553         if (current->journal_info &&
3554             current->journal_info != BTRFS_SEND_TRANS_STUB)
3555                 can_commit = false;
3556
3557         ret = btrfs_start_delalloc_snapshot(root);
3558         if (ret < 0)
3559                 goto out;
3560         btrfs_wait_ordered_extents(root, U64_MAX, 0, (u64)-1);
3561
3562         trans = btrfs_join_transaction(root);
3563         if (IS_ERR(trans)) {
3564                 ret = PTR_ERR(trans);
3565                 goto out;
3566         }
3567
3568         if (can_commit)
3569                 ret = btrfs_commit_transaction(trans);
3570         else
3571                 ret = btrfs_end_transaction(trans);
3572 out:
3573         clear_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state);
3574         wake_up(&root->qgroup_flush_wait);
3575         return ret;
3576 }
3577
3578 static int qgroup_reserve_data(struct btrfs_inode *inode,
3579                         struct extent_changeset **reserved_ret, u64 start,
3580                         u64 len)
3581 {
3582         struct btrfs_root *root = inode->root;
3583         struct extent_changeset *reserved;
3584         bool new_reserved = false;
3585         u64 orig_reserved;
3586         u64 to_reserve;
3587         int ret;
3588
3589         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &root->fs_info->flags) ||
3590             !is_fstree(root->root_key.objectid) || len == 0)
3591                 return 0;
3592
3593         /* @reserved parameter is mandatory for qgroup */
3594         if (WARN_ON(!reserved_ret))
3595                 return -EINVAL;
3596         if (!*reserved_ret) {
3597                 new_reserved = true;
3598                 *reserved_ret = extent_changeset_alloc();
3599                 if (!*reserved_ret)
3600                         return -ENOMEM;
3601         }
3602         reserved = *reserved_ret;
3603         /* Record already reserved space */
3604         orig_reserved = reserved->bytes_changed;
3605         ret = set_record_extent_bits(&inode->io_tree, start,
3606                         start + len -1, EXTENT_QGROUP_RESERVED, reserved);
3607
3608         /* Newly reserved space */
3609         to_reserve = reserved->bytes_changed - orig_reserved;
3610         trace_btrfs_qgroup_reserve_data(&inode->vfs_inode, start, len,
3611                                         to_reserve, QGROUP_RESERVE);
3612         if (ret < 0)
3613                 goto out;
3614         ret = qgroup_reserve(root, to_reserve, true, BTRFS_QGROUP_RSV_DATA);
3615         if (ret < 0)
3616                 goto cleanup;
3617
3618         return ret;
3619
3620 cleanup:
3621         qgroup_unreserve_range(inode, reserved, start, len);
3622 out:
3623         if (new_reserved) {
3624                 extent_changeset_release(reserved);
3625                 kfree(reserved);
3626                 *reserved_ret = NULL;
3627         }
3628         return ret;
3629 }
3630
3631 /*
3632  * Reserve qgroup space for range [start, start + len).
3633  *
3634  * This function will either reserve space from related qgroups or do nothing
3635  * if the range is already reserved.
3636  *
3637  * Return 0 for successful reservation
3638  * Return <0 for error (including -EQUOT)
3639  *
3640  * NOTE: This function may sleep for memory allocation, dirty page flushing and
3641  *       commit transaction. So caller should not hold any dirty page locked.
3642  */
3643 int btrfs_qgroup_reserve_data(struct btrfs_inode *inode,
3644                         struct extent_changeset **reserved_ret, u64 start,
3645                         u64 len)
3646 {
3647         int ret;
3648
3649         ret = qgroup_reserve_data(inode, reserved_ret, start, len);
3650         if (ret <= 0 && ret != -EDQUOT)
3651                 return ret;
3652
3653         ret = try_flush_qgroup(inode->root);
3654         if (ret < 0)
3655                 return ret;
3656         return qgroup_reserve_data(inode, reserved_ret, start, len);
3657 }
3658
3659 /* Free ranges specified by @reserved, normally in error path */
3660 static int qgroup_free_reserved_data(struct btrfs_inode *inode,
3661                         struct extent_changeset *reserved, u64 start, u64 len)
3662 {
3663         struct btrfs_root *root = inode->root;
3664         struct ulist_node *unode;
3665         struct ulist_iterator uiter;
3666         struct extent_changeset changeset;
3667         int freed = 0;
3668         int ret;
3669
3670         extent_changeset_init(&changeset);
3671         len = round_up(start + len, root->fs_info->sectorsize);
3672         start = round_down(start, root->fs_info->sectorsize);
3673
3674         ULIST_ITER_INIT(&uiter);
3675         while ((unode = ulist_next(&reserved->range_changed, &uiter))) {
3676                 u64 range_start = unode->val;
3677                 /* unode->aux is the inclusive end */
3678                 u64 range_len = unode->aux - range_start + 1;
3679                 u64 free_start;
3680                 u64 free_len;
3681
3682                 extent_changeset_release(&changeset);
3683
3684                 /* Only free range in range [start, start + len) */
3685                 if (range_start >= start + len ||
3686                     range_start + range_len <= start)
3687                         continue;
3688                 free_start = max(range_start, start);
3689                 free_len = min(start + len, range_start + range_len) -
3690                            free_start;
3691                 /*
3692                  * TODO: To also modify reserved->ranges_reserved to reflect
3693                  * the modification.
3694                  *
3695                  * However as long as we free qgroup reserved according to
3696                  * EXTENT_QGROUP_RESERVED, we won't double free.
3697                  * So not need to rush.
3698                  */
3699                 ret = clear_record_extent_bits(&inode->io_tree, free_start,
3700                                 free_start + free_len - 1,
3701                                 EXTENT_QGROUP_RESERVED, &changeset);
3702                 if (ret < 0)
3703                         goto out;
3704                 freed += changeset.bytes_changed;
3705         }
3706         btrfs_qgroup_free_refroot(root->fs_info, root->root_key.objectid, freed,
3707                                   BTRFS_QGROUP_RSV_DATA);
3708         ret = freed;
3709 out:
3710         extent_changeset_release(&changeset);
3711         return ret;
3712 }
3713
3714 static int __btrfs_qgroup_release_data(struct btrfs_inode *inode,
3715                         struct extent_changeset *reserved, u64 start, u64 len,
3716                         int free)
3717 {
3718         struct extent_changeset changeset;
3719         int trace_op = QGROUP_RELEASE;
3720         int ret;
3721
3722         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &inode->root->fs_info->flags))
3723                 return 0;
3724
3725         /* In release case, we shouldn't have @reserved */
3726         WARN_ON(!free && reserved);
3727         if (free && reserved)
3728                 return qgroup_free_reserved_data(inode, reserved, start, len);
3729         extent_changeset_init(&changeset);
3730         ret = clear_record_extent_bits(&inode->io_tree, start, start + len -1,
3731                                        EXTENT_QGROUP_RESERVED, &changeset);
3732         if (ret < 0)
3733                 goto out;
3734
3735         if (free)
3736                 trace_op = QGROUP_FREE;
3737         trace_btrfs_qgroup_release_data(&inode->vfs_inode, start, len,
3738                                         changeset.bytes_changed, trace_op);
3739         if (free)
3740                 btrfs_qgroup_free_refroot(inode->root->fs_info,
3741                                 inode->root->root_key.objectid,
3742                                 changeset.bytes_changed, BTRFS_QGROUP_RSV_DATA);
3743         ret = changeset.bytes_changed;
3744 out:
3745         extent_changeset_release(&changeset);
3746         return ret;
3747 }
3748
3749 /*
3750  * Free a reserved space range from io_tree and related qgroups
3751  *
3752  * Should be called when a range of pages get invalidated before reaching disk.
3753  * Or for error cleanup case.
3754  * if @reserved is given, only reserved range in [@start, @start + @len) will
3755  * be freed.
3756  *
3757  * For data written to disk, use btrfs_qgroup_release_data().
3758  *
3759  * NOTE: This function may sleep for memory allocation.
3760  */
3761 int btrfs_qgroup_free_data(struct btrfs_inode *inode,
3762                         struct extent_changeset *reserved, u64 start, u64 len)
3763 {
3764         return __btrfs_qgroup_release_data(inode, reserved, start, len, 1);
3765 }
3766
3767 /*
3768  * Release a reserved space range from io_tree only.
3769  *
3770  * Should be called when a range of pages get written to disk and corresponding
3771  * FILE_EXTENT is inserted into corresponding root.
3772  *
3773  * Since new qgroup accounting framework will only update qgroup numbers at
3774  * commit_transaction() time, its reserved space shouldn't be freed from
3775  * related qgroups.
3776  *
3777  * But we should release the range from io_tree, to allow further write to be
3778  * COWed.
3779  *
3780  * NOTE: This function may sleep for memory allocation.
3781  */
3782 int btrfs_qgroup_release_data(struct btrfs_inode *inode, u64 start, u64 len)
3783 {
3784         return __btrfs_qgroup_release_data(inode, NULL, start, len, 0);
3785 }
3786
3787 static void add_root_meta_rsv(struct btrfs_root *root, int num_bytes,
3788                               enum btrfs_qgroup_rsv_type type)
3789 {
3790         if (type != BTRFS_QGROUP_RSV_META_PREALLOC &&
3791             type != BTRFS_QGROUP_RSV_META_PERTRANS)
3792                 return;
3793         if (num_bytes == 0)
3794                 return;
3795
3796         spin_lock(&root->qgroup_meta_rsv_lock);
3797         if (type == BTRFS_QGROUP_RSV_META_PREALLOC)
3798                 root->qgroup_meta_rsv_prealloc += num_bytes;
3799         else
3800                 root->qgroup_meta_rsv_pertrans += num_bytes;
3801         spin_unlock(&root->qgroup_meta_rsv_lock);
3802 }
3803
3804 static int sub_root_meta_rsv(struct btrfs_root *root, int num_bytes,
3805                              enum btrfs_qgroup_rsv_type type)
3806 {
3807         if (type != BTRFS_QGROUP_RSV_META_PREALLOC &&
3808             type != BTRFS_QGROUP_RSV_META_PERTRANS)
3809                 return 0;
3810         if (num_bytes == 0)
3811                 return 0;
3812
3813         spin_lock(&root->qgroup_meta_rsv_lock);
3814         if (type == BTRFS_QGROUP_RSV_META_PREALLOC) {
3815                 num_bytes = min_t(u64, root->qgroup_meta_rsv_prealloc,
3816                                   num_bytes);
3817                 root->qgroup_meta_rsv_prealloc -= num_bytes;
3818         } else {
3819                 num_bytes = min_t(u64, root->qgroup_meta_rsv_pertrans,
3820                                   num_bytes);
3821                 root->qgroup_meta_rsv_pertrans -= num_bytes;
3822         }
3823         spin_unlock(&root->qgroup_meta_rsv_lock);
3824         return num_bytes;
3825 }
3826
3827 static int qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
3828                                 enum btrfs_qgroup_rsv_type type, bool enforce)
3829 {
3830         struct btrfs_fs_info *fs_info = root->fs_info;
3831         int ret;
3832
3833         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
3834             !is_fstree(root->root_key.objectid) || num_bytes == 0)
3835                 return 0;
3836
3837         BUG_ON(num_bytes != round_down(num_bytes, fs_info->nodesize));
3838         trace_qgroup_meta_reserve(root, (s64)num_bytes, type);
3839         ret = qgroup_reserve(root, num_bytes, enforce, type);
3840         if (ret < 0)
3841                 return ret;
3842         /*
3843          * Record what we have reserved into root.
3844          *
3845          * To avoid quota disabled->enabled underflow.
3846          * In that case, we may try to free space we haven't reserved
3847          * (since quota was disabled), so record what we reserved into root.
3848          * And ensure later release won't underflow this number.
3849          */
3850         add_root_meta_rsv(root, num_bytes, type);
3851         return ret;
3852 }
3853
3854 int __btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
3855                                 enum btrfs_qgroup_rsv_type type, bool enforce)
3856 {
3857         int ret;
3858
3859         ret = qgroup_reserve_meta(root, num_bytes, type, enforce);
3860         if (ret <= 0 && ret != -EDQUOT)
3861                 return ret;
3862
3863         ret = try_flush_qgroup(root);
3864         if (ret < 0)
3865                 return ret;
3866         return qgroup_reserve_meta(root, num_bytes, type, enforce);
3867 }
3868
3869 void btrfs_qgroup_free_meta_all_pertrans(struct btrfs_root *root)
3870 {
3871         struct btrfs_fs_info *fs_info = root->fs_info;
3872
3873         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
3874             !is_fstree(root->root_key.objectid))
3875                 return;
3876
3877         /* TODO: Update trace point to handle such free */
3878         trace_qgroup_meta_free_all_pertrans(root);
3879         /* Special value -1 means to free all reserved space */
3880         btrfs_qgroup_free_refroot(fs_info, root->root_key.objectid, (u64)-1,
3881                                   BTRFS_QGROUP_RSV_META_PERTRANS);
3882 }
3883
3884 void __btrfs_qgroup_free_meta(struct btrfs_root *root, int num_bytes,
3885                               enum btrfs_qgroup_rsv_type type)
3886 {
3887         struct btrfs_fs_info *fs_info = root->fs_info;
3888
3889         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
3890             !is_fstree(root->root_key.objectid))
3891                 return;
3892
3893         /*
3894          * reservation for META_PREALLOC can happen before quota is enabled,
3895          * which can lead to underflow.
3896          * Here ensure we will only free what we really have reserved.
3897          */
3898         num_bytes = sub_root_meta_rsv(root, num_bytes, type);
3899         BUG_ON(num_bytes != round_down(num_bytes, fs_info->nodesize));
3900         trace_qgroup_meta_reserve(root, -(s64)num_bytes, type);
3901         btrfs_qgroup_free_refroot(fs_info, root->root_key.objectid,
3902                                   num_bytes, type);
3903 }
3904
3905 static void qgroup_convert_meta(struct btrfs_fs_info *fs_info, u64 ref_root,
3906                                 int num_bytes)
3907 {
3908         struct btrfs_qgroup *qgroup;
3909         struct ulist_node *unode;
3910         struct ulist_iterator uiter;
3911         int ret = 0;
3912
3913         if (num_bytes == 0)
3914                 return;
3915         if (!fs_info->quota_root)
3916                 return;
3917
3918         spin_lock(&fs_info->qgroup_lock);
3919         qgroup = find_qgroup_rb(fs_info, ref_root);
3920         if (!qgroup)
3921                 goto out;
3922         ulist_reinit(fs_info->qgroup_ulist);
3923         ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
3924                        qgroup_to_aux(qgroup), GFP_ATOMIC);
3925         if (ret < 0)
3926                 goto out;
3927         ULIST_ITER_INIT(&uiter);
3928         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
3929                 struct btrfs_qgroup *qg;
3930                 struct btrfs_qgroup_list *glist;
3931
3932                 qg = unode_aux_to_qgroup(unode);
3933
3934                 qgroup_rsv_release(fs_info, qg, num_bytes,
3935                                 BTRFS_QGROUP_RSV_META_PREALLOC);
3936                 qgroup_rsv_add(fs_info, qg, num_bytes,
3937                                 BTRFS_QGROUP_RSV_META_PERTRANS);
3938                 list_for_each_entry(glist, &qg->groups, next_group) {
3939                         ret = ulist_add(fs_info->qgroup_ulist,
3940                                         glist->group->qgroupid,
3941                                         qgroup_to_aux(glist->group), GFP_ATOMIC);
3942                         if (ret < 0)
3943                                 goto out;
3944                 }
3945         }
3946 out:
3947         spin_unlock(&fs_info->qgroup_lock);
3948 }
3949
3950 void btrfs_qgroup_convert_reserved_meta(struct btrfs_root *root, int num_bytes)
3951 {
3952         struct btrfs_fs_info *fs_info = root->fs_info;
3953
3954         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
3955             !is_fstree(root->root_key.objectid))
3956                 return;
3957         /* Same as btrfs_qgroup_free_meta_prealloc() */
3958         num_bytes = sub_root_meta_rsv(root, num_bytes,
3959                                       BTRFS_QGROUP_RSV_META_PREALLOC);
3960         trace_qgroup_meta_convert(root, num_bytes);
3961         qgroup_convert_meta(fs_info, root->root_key.objectid, num_bytes);
3962 }
3963
3964 /*
3965  * Check qgroup reserved space leaking, normally at destroy inode
3966  * time
3967  */
3968 void btrfs_qgroup_check_reserved_leak(struct btrfs_inode *inode)
3969 {
3970         struct extent_changeset changeset;
3971         struct ulist_node *unode;
3972         struct ulist_iterator iter;
3973         int ret;
3974
3975         extent_changeset_init(&changeset);
3976         ret = clear_record_extent_bits(&inode->io_tree, 0, (u64)-1,
3977                         EXTENT_QGROUP_RESERVED, &changeset);
3978
3979         WARN_ON(ret < 0);
3980         if (WARN_ON(changeset.bytes_changed)) {
3981                 ULIST_ITER_INIT(&iter);
3982                 while ((unode = ulist_next(&changeset.range_changed, &iter))) {
3983                         btrfs_warn(inode->root->fs_info,
3984                 "leaking qgroup reserved space, ino: %llu, start: %llu, end: %llu",
3985                                 btrfs_ino(inode), unode->val, unode->aux);
3986                 }
3987                 btrfs_qgroup_free_refroot(inode->root->fs_info,
3988                                 inode->root->root_key.objectid,
3989                                 changeset.bytes_changed, BTRFS_QGROUP_RSV_DATA);
3990
3991         }
3992         extent_changeset_release(&changeset);
3993 }
3994
3995 void btrfs_qgroup_init_swapped_blocks(
3996         struct btrfs_qgroup_swapped_blocks *swapped_blocks)
3997 {
3998         int i;
3999
4000         spin_lock_init(&swapped_blocks->lock);
4001         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
4002                 swapped_blocks->blocks[i] = RB_ROOT;
4003         swapped_blocks->swapped = false;
4004 }
4005
4006 /*
4007  * Delete all swapped blocks record of @root.
4008  * Every record here means we skipped a full subtree scan for qgroup.
4009  *
4010  * Gets called when committing one transaction.
4011  */
4012 void btrfs_qgroup_clean_swapped_blocks(struct btrfs_root *root)
4013 {
4014         struct btrfs_qgroup_swapped_blocks *swapped_blocks;
4015         int i;
4016
4017         swapped_blocks = &root->swapped_blocks;
4018
4019         spin_lock(&swapped_blocks->lock);
4020         if (!swapped_blocks->swapped)
4021                 goto out;
4022         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
4023                 struct rb_root *cur_root = &swapped_blocks->blocks[i];
4024                 struct btrfs_qgroup_swapped_block *entry;
4025                 struct btrfs_qgroup_swapped_block *next;
4026
4027                 rbtree_postorder_for_each_entry_safe(entry, next, cur_root,
4028                                                      node)
4029                         kfree(entry);
4030                 swapped_blocks->blocks[i] = RB_ROOT;
4031         }
4032         swapped_blocks->swapped = false;
4033 out:
4034         spin_unlock(&swapped_blocks->lock);
4035 }
4036
4037 /*
4038  * Add subtree roots record into @subvol_root.
4039  *
4040  * @subvol_root:        tree root of the subvolume tree get swapped
4041  * @bg:                 block group under balance
4042  * @subvol_parent/slot: pointer to the subtree root in subvolume tree
4043  * @reloc_parent/slot:  pointer to the subtree root in reloc tree
4044  *                      BOTH POINTERS ARE BEFORE TREE SWAP
4045  * @last_snapshot:      last snapshot generation of the subvolume tree
4046  */
4047 int btrfs_qgroup_add_swapped_blocks(struct btrfs_trans_handle *trans,
4048                 struct btrfs_root *subvol_root,
4049                 struct btrfs_block_group *bg,
4050                 struct extent_buffer *subvol_parent, int subvol_slot,
4051                 struct extent_buffer *reloc_parent, int reloc_slot,
4052                 u64 last_snapshot)
4053 {
4054         struct btrfs_fs_info *fs_info = subvol_root->fs_info;
4055         struct btrfs_qgroup_swapped_blocks *blocks = &subvol_root->swapped_blocks;
4056         struct btrfs_qgroup_swapped_block *block;
4057         struct rb_node **cur;
4058         struct rb_node *parent = NULL;
4059         int level = btrfs_header_level(subvol_parent) - 1;
4060         int ret = 0;
4061
4062         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
4063                 return 0;
4064
4065         if (btrfs_node_ptr_generation(subvol_parent, subvol_slot) >
4066             btrfs_node_ptr_generation(reloc_parent, reloc_slot)) {
4067                 btrfs_err_rl(fs_info,
4068                 "%s: bad parameter order, subvol_gen=%llu reloc_gen=%llu",
4069                         __func__,
4070                         btrfs_node_ptr_generation(subvol_parent, subvol_slot),
4071                         btrfs_node_ptr_generation(reloc_parent, reloc_slot));
4072                 return -EUCLEAN;
4073         }
4074
4075         block = kmalloc(sizeof(*block), GFP_NOFS);
4076         if (!block) {
4077                 ret = -ENOMEM;
4078                 goto out;
4079         }
4080
4081         /*
4082          * @reloc_parent/slot is still before swap, while @block is going to
4083          * record the bytenr after swap, so we do the swap here.
4084          */
4085         block->subvol_bytenr = btrfs_node_blockptr(reloc_parent, reloc_slot);
4086         block->subvol_generation = btrfs_node_ptr_generation(reloc_parent,
4087                                                              reloc_slot);
4088         block->reloc_bytenr = btrfs_node_blockptr(subvol_parent, subvol_slot);
4089         block->reloc_generation = btrfs_node_ptr_generation(subvol_parent,
4090                                                             subvol_slot);
4091         block->last_snapshot = last_snapshot;
4092         block->level = level;
4093
4094         /*
4095          * If we have bg == NULL, we're called from btrfs_recover_relocation(),
4096          * no one else can modify tree blocks thus we qgroup will not change
4097          * no matter the value of trace_leaf.
4098          */
4099         if (bg && bg->flags & BTRFS_BLOCK_GROUP_DATA)
4100                 block->trace_leaf = true;
4101         else
4102                 block->trace_leaf = false;
4103         btrfs_node_key_to_cpu(reloc_parent, &block->first_key, reloc_slot);
4104
4105         /* Insert @block into @blocks */
4106         spin_lock(&blocks->lock);
4107         cur = &blocks->blocks[level].rb_node;
4108         while (*cur) {
4109                 struct btrfs_qgroup_swapped_block *entry;
4110
4111                 parent = *cur;
4112                 entry = rb_entry(parent, struct btrfs_qgroup_swapped_block,
4113                                  node);
4114
4115                 if (entry->subvol_bytenr < block->subvol_bytenr) {
4116                         cur = &(*cur)->rb_left;
4117                 } else if (entry->subvol_bytenr > block->subvol_bytenr) {
4118                         cur = &(*cur)->rb_right;
4119                 } else {
4120                         if (entry->subvol_generation !=
4121                                         block->subvol_generation ||
4122                             entry->reloc_bytenr != block->reloc_bytenr ||
4123                             entry->reloc_generation !=
4124                                         block->reloc_generation) {
4125                                 /*
4126                                  * Duplicated but mismatch entry found.
4127                                  * Shouldn't happen.
4128                                  *
4129                                  * Marking qgroup inconsistent should be enough
4130                                  * for end users.
4131                                  */
4132                                 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
4133                                 ret = -EEXIST;
4134                         }
4135                         kfree(block);
4136                         goto out_unlock;
4137                 }
4138         }
4139         rb_link_node(&block->node, parent, cur);
4140         rb_insert_color(&block->node, &blocks->blocks[level]);
4141         blocks->swapped = true;
4142 out_unlock:
4143         spin_unlock(&blocks->lock);
4144 out:
4145         if (ret < 0)
4146                 fs_info->qgroup_flags |=
4147                         BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
4148         return ret;
4149 }
4150
4151 /*
4152  * Check if the tree block is a subtree root, and if so do the needed
4153  * delayed subtree trace for qgroup.
4154  *
4155  * This is called during btrfs_cow_block().
4156  */
4157 int btrfs_qgroup_trace_subtree_after_cow(struct btrfs_trans_handle *trans,
4158                                          struct btrfs_root *root,
4159                                          struct extent_buffer *subvol_eb)
4160 {
4161         struct btrfs_fs_info *fs_info = root->fs_info;
4162         struct btrfs_qgroup_swapped_blocks *blocks = &root->swapped_blocks;
4163         struct btrfs_qgroup_swapped_block *block;
4164         struct extent_buffer *reloc_eb = NULL;
4165         struct rb_node *node;
4166         bool found = false;
4167         bool swapped = false;
4168         int level = btrfs_header_level(subvol_eb);
4169         int ret = 0;
4170         int i;
4171
4172         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
4173                 return 0;
4174         if (!is_fstree(root->root_key.objectid) || !root->reloc_root)
4175                 return 0;
4176
4177         spin_lock(&blocks->lock);
4178         if (!blocks->swapped) {
4179                 spin_unlock(&blocks->lock);
4180                 return 0;
4181         }
4182         node = blocks->blocks[level].rb_node;
4183
4184         while (node) {
4185                 block = rb_entry(node, struct btrfs_qgroup_swapped_block, node);
4186                 if (block->subvol_bytenr < subvol_eb->start) {
4187                         node = node->rb_left;
4188                 } else if (block->subvol_bytenr > subvol_eb->start) {
4189                         node = node->rb_right;
4190                 } else {
4191                         found = true;
4192                         break;
4193                 }
4194         }
4195         if (!found) {
4196                 spin_unlock(&blocks->lock);
4197                 goto out;
4198         }
4199         /* Found one, remove it from @blocks first and update blocks->swapped */
4200         rb_erase(&block->node, &blocks->blocks[level]);
4201         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
4202                 if (RB_EMPTY_ROOT(&blocks->blocks[i])) {
4203                         swapped = true;
4204                         break;
4205                 }
4206         }
4207         blocks->swapped = swapped;
4208         spin_unlock(&blocks->lock);
4209
4210         /* Read out reloc subtree root */
4211         reloc_eb = read_tree_block(fs_info, block->reloc_bytenr, 0,
4212                                    block->reloc_generation, block->level,
4213                                    &block->first_key);
4214         if (IS_ERR(reloc_eb)) {
4215                 ret = PTR_ERR(reloc_eb);
4216                 reloc_eb = NULL;
4217                 goto free_out;
4218         }
4219         if (!extent_buffer_uptodate(reloc_eb)) {
4220                 ret = -EIO;
4221                 goto free_out;
4222         }
4223
4224         ret = qgroup_trace_subtree_swap(trans, reloc_eb, subvol_eb,
4225                         block->last_snapshot, block->trace_leaf);
4226 free_out:
4227         kfree(block);
4228         free_extent_buffer(reloc_eb);
4229 out:
4230         if (ret < 0) {
4231                 btrfs_err_rl(fs_info,
4232                              "failed to account subtree at bytenr %llu: %d",
4233                              subvol_eb->start, ret);
4234                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
4235         }
4236         return ret;
4237 }
4238
4239 void btrfs_qgroup_destroy_extent_records(struct btrfs_transaction *trans)
4240 {
4241         struct btrfs_qgroup_extent_record *entry;
4242         struct btrfs_qgroup_extent_record *next;
4243         struct rb_root *root;
4244
4245         root = &trans->delayed_refs.dirty_extent_root;
4246         rbtree_postorder_for_each_entry_safe(entry, next, root, node) {
4247                 ulist_free(entry->old_roots);
4248                 kfree(entry);
4249         }
4250 }