fs/btrfs/qgroup.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2011 STRATO.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/writeback.h>
   9 #include <linux/blkdev.h>
  10 #include <linux/rbtree.h>
  11 #include <linux/slab.h>
  12 #include <linux/workqueue.h>
  13 #include <linux/btrfs.h>
  14 #include <linux/sched/mm.h>
  15
  16 #include "ctree.h"
  17 #include "transaction.h"
  18 #include "disk-io.h"
  19 #include "locking.h"
  20 #include "ulist.h"
  21 #include "backref.h"
  22 #include "extent_io.h"
  23 #include "qgroup.h"
  24 #include "block-group.h"
  25 #include "sysfs.h"
  26 #include "tree-mod-log.h"
  27
  28 /* TODO XXX FIXME
  29  *  - subvol delete -> delete when ref goes to 0? delete limits also?
  30  *  - reorganize keys
  31  *  - compressed
  32  *  - sync
  33  *  - copy also limits on subvol creation
  34  *  - limit
  35  *  - caches for ulists
  36  *  - performance benchmarks
  37  *  - check all ioctl parameters
  38  */
  39
  40 /*
  41  * Helpers to access qgroup reservation
  42  *
  43  * Callers should ensure the lock context and type are valid
  44  */
  45
  46 static u64 qgroup_rsv_total(const struct btrfs_qgroup *qgroup)
  47 {
  48         u64 ret = 0;
  49         int i;
  50
  51         for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++)
  52                 ret += qgroup->rsv.values[i];
  53
  54         return ret;
  55 }
  56
  57 #ifdef CONFIG_BTRFS_DEBUG
  58 static const char *qgroup_rsv_type_str(enum btrfs_qgroup_rsv_type type)
  59 {
  60         if (type == BTRFS_QGROUP_RSV_DATA)
  61                 return "data";
  62         if (type == BTRFS_QGROUP_RSV_META_PERTRANS)
  63                 return "meta_pertrans";
  64         if (type == BTRFS_QGROUP_RSV_META_PREALLOC)
  65                 return "meta_prealloc";
  66         return NULL;
  67 }
  68 #endif
  69
  70 static void qgroup_rsv_add(struct btrfs_fs_info *fs_info,
  71                            struct btrfs_qgroup *qgroup, u64 num_bytes,
  72                            enum btrfs_qgroup_rsv_type type)
  73 {
  74         trace_qgroup_update_reserve(fs_info, qgroup, num_bytes, type);
  75         qgroup->rsv.values[type] += num_bytes;
  76 }
  77
  78 static void qgroup_rsv_release(struct btrfs_fs_info *fs_info,
  79                                struct btrfs_qgroup *qgroup, u64 num_bytes,
  80                                enum btrfs_qgroup_rsv_type type)
  81 {
  82         trace_qgroup_update_reserve(fs_info, qgroup, -(s64)num_bytes, type);
  83         if (qgroup->rsv.values[type] >= num_bytes) {
  84                 qgroup->rsv.values[type] -= num_bytes;
  85                 return;
  86         }
  87 #ifdef CONFIG_BTRFS_DEBUG
  88         WARN_RATELIMIT(1,
  89                 "qgroup %llu %s reserved space underflow, have %llu to free %llu",
  90                 qgroup->qgroupid, qgroup_rsv_type_str(type),
  91                 qgroup->rsv.values[type], num_bytes);
  92 #endif
  93         qgroup->rsv.values[type] = 0;
  94 }
  95
  96 static void qgroup_rsv_add_by_qgroup(struct btrfs_fs_info *fs_info,
  97                                      struct btrfs_qgroup *dest,
  98                                      struct btrfs_qgroup *src)
  99 {
 100         int i;
 101
 102         for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++)
 103                 qgroup_rsv_add(fs_info, dest, src->rsv.values[i], i);
 104 }
 105
 106 static void qgroup_rsv_release_by_qgroup(struct btrfs_fs_info *fs_info,
 107                                          struct btrfs_qgroup *dest,
 108                                           struct btrfs_qgroup *src)
 109 {
 110         int i;
 111
 112         for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++)
 113                 qgroup_rsv_release(fs_info, dest, src->rsv.values[i], i);
 114 }
 115
 116 static void btrfs_qgroup_update_old_refcnt(struct btrfs_qgroup *qg, u64 seq,
 117                                            int mod)
 118 {
 119         if (qg->old_refcnt < seq)
 120                 qg->old_refcnt = seq;
 121         qg->old_refcnt += mod;
 122 }
 123
 124 static void btrfs_qgroup_update_new_refcnt(struct btrfs_qgroup *qg, u64 seq,
 125                                            int mod)
 126 {
 127         if (qg->new_refcnt < seq)
 128                 qg->new_refcnt = seq;
 129         qg->new_refcnt += mod;
 130 }
 131
 132 static inline u64 btrfs_qgroup_get_old_refcnt(struct btrfs_qgroup *qg, u64 seq)
 133 {
 134         if (qg->old_refcnt < seq)
 135                 return 0;
 136         return qg->old_refcnt - seq;
 137 }
 138
 139 static inline u64 btrfs_qgroup_get_new_refcnt(struct btrfs_qgroup *qg, u64 seq)
 140 {
 141         if (qg->new_refcnt < seq)
 142                 return 0;
 143         return qg->new_refcnt - seq;
 144 }
 145
 146 /*
 147  * glue structure to represent the relations between qgroups.
 148  */
 149 struct btrfs_qgroup_list {
 150         struct list_head next_group;
 151         struct list_head next_member;
 152         struct btrfs_qgroup *group;
 153         struct btrfs_qgroup *member;
 154 };
 155
 156 static inline u64 qgroup_to_aux(struct btrfs_qgroup *qg)
 157 {
 158         return (u64)(uintptr_t)qg;
 159 }
 160
 161 static inline struct btrfs_qgroup* unode_aux_to_qgroup(struct ulist_node *n)
 162 {
 163         return (struct btrfs_qgroup *)(uintptr_t)n->aux;
 164 }
 165
 166 static int
 167 qgroup_rescan_init(struct btrfs_fs_info *fs_info, u64 progress_objectid,
 168                    int init_flags);
 169 static void qgroup_rescan_zero_tracking(struct btrfs_fs_info *fs_info);
 170
 171 /* must be called with qgroup_ioctl_lock held */
 172 static struct btrfs_qgroup *find_qgroup_rb(struct btrfs_fs_info *fs_info,
 173                                            u64 qgroupid)
 174 {
 175         struct rb_node *n = fs_info->qgroup_tree.rb_node;
 176         struct btrfs_qgroup *qgroup;
 177
 178         while (n) {
 179                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
 180                 if (qgroup->qgroupid < qgroupid)
 181                         n = n->rb_left;
 182                 else if (qgroup->qgroupid > qgroupid)
 183                         n = n->rb_right;
 184                 else
 185                         return qgroup;
 186         }
 187         return NULL;
 188 }
 189
 190 /* must be called with qgroup_lock held */
 191 static struct btrfs_qgroup *add_qgroup_rb(struct btrfs_fs_info *fs_info,
 192                                           u64 qgroupid)
 193 {
 194         struct rb_node **p = &fs_info->qgroup_tree.rb_node;
 195         struct rb_node *parent = NULL;
 196         struct btrfs_qgroup *qgroup;
 197
 198         while (*p) {
 199                 parent = *p;
 200                 qgroup = rb_entry(parent, struct btrfs_qgroup, node);
 201
 202                 if (qgroup->qgroupid < qgroupid)
 203                         p = &(*p)->rb_left;
 204                 else if (qgroup->qgroupid > qgroupid)
 205                         p = &(*p)->rb_right;
 206                 else
 207                         return qgroup;
 208         }
 209
 210         qgroup = kzalloc(sizeof(*qgroup), GFP_ATOMIC);
 211         if (!qgroup)
 212                 return ERR_PTR(-ENOMEM);
 213
 214         qgroup->qgroupid = qgroupid;
 215         INIT_LIST_HEAD(&qgroup->groups);
 216         INIT_LIST_HEAD(&qgroup->members);
 217         INIT_LIST_HEAD(&qgroup->dirty);
 218
 219         rb_link_node(&qgroup->node, parent, p);
 220         rb_insert_color(&qgroup->node, &fs_info->qgroup_tree);
 221
 222         return qgroup;
 223 }
 224
 225 static void __del_qgroup_rb(struct btrfs_fs_info *fs_info,
 226                             struct btrfs_qgroup *qgroup)
 227 {
 228         struct btrfs_qgroup_list *list;
 229
 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 }
 247
 248 /* must be called with qgroup_lock held */
 249 static int del_qgroup_rb(struct btrfs_fs_info *fs_info, u64 qgroupid)
 250 {
 251         struct btrfs_qgroup *qgroup = find_qgroup_rb(fs_info, qgroupid);
 252
 253         if (!qgroup)
 254                 return -ENOENT;
 255
 256         rb_erase(&qgroup->node, &fs_info->qgroup_tree);
 257         __del_qgroup_rb(fs_info, qgroup);
 258         return 0;
 259 }
 260
 261 /* must be called with qgroup_lock held */
 262 static int add_relation_rb(struct btrfs_fs_info *fs_info,
 263                            u64 memberid, u64 parentid)
 264 {
 265         struct btrfs_qgroup *member;
 266         struct btrfs_qgroup *parent;
 267         struct btrfs_qgroup_list *list;
 268
 269         member = find_qgroup_rb(fs_info, memberid);
 270         parent = find_qgroup_rb(fs_info, parentid);
 271         if (!member || !parent)
 272                 return -ENOENT;
 273
 274         list = kzalloc(sizeof(*list), GFP_ATOMIC);
 275         if (!list)
 276                 return -ENOMEM;
 277
 278         list->group = parent;
 279         list->member = member;
 280         list_add_tail(&list->next_group, &member->groups);
 281         list_add_tail(&list->next_member, &parent->members);
 282
 283         return 0;
 284 }
 285
 286 /* must be called with qgroup_lock held */
 287 static int del_relation_rb(struct btrfs_fs_info *fs_info,
 288                            u64 memberid, u64 parentid)
 289 {
 290         struct btrfs_qgroup *member;
 291         struct btrfs_qgroup *parent;
 292         struct btrfs_qgroup_list *list;
 293
 294         member = find_qgroup_rb(fs_info, memberid);
 295         parent = find_qgroup_rb(fs_info, parentid);
 296         if (!member || !parent)
 297                 return -ENOENT;
 298
 299         list_for_each_entry(list, &member->groups, next_group) {
 300                 if (list->group == parent) {
 301                         list_del(&list->next_group);
 302                         list_del(&list->next_member);
 303                         kfree(list);
 304                         return 0;
 305                 }
 306         }
 307         return -ENOENT;
 308 }
 309
 310 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 311 int btrfs_verify_qgroup_counts(struct btrfs_fs_info *fs_info, u64 qgroupid,
 312                                u64 rfer, u64 excl)
 313 {
 314         struct btrfs_qgroup *qgroup;
 315
 316         qgroup = find_qgroup_rb(fs_info, qgroupid);
 317         if (!qgroup)
 318                 return -EINVAL;
 319         if (qgroup->rfer != rfer || qgroup->excl != excl)
 320                 return -EINVAL;
 321         return 0;
 322 }
 323 #endif
 324
 325 /*
 326  * The full config is read in one go, only called from open_ctree()
 327  * It doesn't use any locking, as at this point we're still single-threaded
 328  */
 329 int btrfs_read_qgroup_config(struct btrfs_fs_info *fs_info)
 330 {
 331         struct btrfs_key key;
 332         struct btrfs_key found_key;
 333         struct btrfs_root *quota_root = fs_info->quota_root;
 334         struct btrfs_path *path = NULL;
 335         struct extent_buffer *l;
 336         int slot;
 337         int ret = 0;
 338         u64 flags = 0;
 339         u64 rescan_progress = 0;
 340
 341         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
 342                 return 0;
 343
 344         fs_info->qgroup_ulist = ulist_alloc(GFP_KERNEL);
 345         if (!fs_info->qgroup_ulist) {
 346                 ret = -ENOMEM;
 347                 goto out;
 348         }
 349
 350         path = btrfs_alloc_path();
 351         if (!path) {
 352                 ret = -ENOMEM;
 353                 goto out;
 354         }
 355
 356         ret = btrfs_sysfs_add_qgroups(fs_info);
 357         if (ret < 0)
 358                 goto out;
 359         /* default this to quota off, in case no status key is found */
 360         fs_info->qgroup_flags = 0;
 361
 362         /*
 363          * pass 1: read status, all qgroup infos and limits
 364          */
 365         key.objectid = 0;
 366         key.type = 0;
 367         key.offset = 0;
 368         ret = btrfs_search_slot_for_read(quota_root, &key, path, 1, 1);
 369         if (ret)
 370                 goto out;
 371
 372         while (1) {
 373                 struct btrfs_qgroup *qgroup;
 374
 375                 slot = path->slots[0];
 376                 l = path->nodes[0];
 377                 btrfs_item_key_to_cpu(l, &found_key, slot);
 378
 379                 if (found_key.type == BTRFS_QGROUP_STATUS_KEY) {
 380                         struct btrfs_qgroup_status_item *ptr;
 381
 382                         ptr = btrfs_item_ptr(l, slot,
 383                                              struct btrfs_qgroup_status_item);
 384
 385                         if (btrfs_qgroup_status_version(l, ptr) !=
 386                             BTRFS_QGROUP_STATUS_VERSION) {
 387                                 btrfs_err(fs_info,
 388                                  "old qgroup version, quota disabled");
 389                                 goto out;
 390                         }
 391                         if (btrfs_qgroup_status_generation(l, ptr) !=
 392                             fs_info->generation) {
 393                                 flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
 394                                 btrfs_err(fs_info,
 395                                         "qgroup generation mismatch, marked as inconsistent");
 396                         }
 397                         fs_info->qgroup_flags = btrfs_qgroup_status_flags(l,
 398                                                                           ptr);
 399                         rescan_progress = btrfs_qgroup_status_rescan(l, ptr);
 400                         goto next1;
 401                 }
 402
 403                 if (found_key.type != BTRFS_QGROUP_INFO_KEY &&
 404                     found_key.type != BTRFS_QGROUP_LIMIT_KEY)
 405                         goto next1;
 406
 407                 qgroup = find_qgroup_rb(fs_info, found_key.offset);
 408                 if ((qgroup && found_key.type == BTRFS_QGROUP_INFO_KEY) ||
 409                     (!qgroup && found_key.type == BTRFS_QGROUP_LIMIT_KEY)) {
 410                         btrfs_err(fs_info, "inconsistent qgroup config");
 411                         flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
 412                 }
 413                 if (!qgroup) {
 414                         qgroup = add_qgroup_rb(fs_info, found_key.offset);
 415                         if (IS_ERR(qgroup)) {
 416                                 ret = PTR_ERR(qgroup);
 417                                 goto out;
 418                         }
 419                 }
 420                 ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
 421                 if (ret < 0)
 422                         goto out;
 423
 424                 switch (found_key.type) {
 425                 case BTRFS_QGROUP_INFO_KEY: {
 426                         struct btrfs_qgroup_info_item *ptr;
 427
 428                         ptr = btrfs_item_ptr(l, slot,
 429                                              struct btrfs_qgroup_info_item);
 430                         qgroup->rfer = btrfs_qgroup_info_rfer(l, ptr);
 431                         qgroup->rfer_cmpr = btrfs_qgroup_info_rfer_cmpr(l, ptr);
 432                         qgroup->excl = btrfs_qgroup_info_excl(l, ptr);
 433                         qgroup->excl_cmpr = btrfs_qgroup_info_excl_cmpr(l, ptr);
 434                         /* generation currently unused */
 435                         break;
 436                 }
 437                 case BTRFS_QGROUP_LIMIT_KEY: {
 438                         struct btrfs_qgroup_limit_item *ptr;
 439
 440                         ptr = btrfs_item_ptr(l, slot,
 441                                              struct btrfs_qgroup_limit_item);
 442                         qgroup->lim_flags = btrfs_qgroup_limit_flags(l, ptr);
 443                         qgroup->max_rfer = btrfs_qgroup_limit_max_rfer(l, ptr);
 444                         qgroup->max_excl = btrfs_qgroup_limit_max_excl(l, ptr);
 445                         qgroup->rsv_rfer = btrfs_qgroup_limit_rsv_rfer(l, ptr);
 446                         qgroup->rsv_excl = btrfs_qgroup_limit_rsv_excl(l, ptr);
 447                         break;
 448                 }
 449                 }
 450 next1:
 451                 ret = btrfs_next_item(quota_root, path);
 452                 if (ret < 0)
 453                         goto out;
 454                 if (ret)
 455                         break;
 456         }
 457         btrfs_release_path(path);
 458
 459         /*
 460          * pass 2: read all qgroup relations
 461          */
 462         key.objectid = 0;
 463         key.type = BTRFS_QGROUP_RELATION_KEY;
 464         key.offset = 0;
 465         ret = btrfs_search_slot_for_read(quota_root, &key, path, 1, 0);
 466         if (ret)
 467                 goto out;
 468         while (1) {
 469                 slot = path->slots[0];
 470                 l = path->nodes[0];
 471                 btrfs_item_key_to_cpu(l, &found_key, slot);
 472
 473                 if (found_key.type != BTRFS_QGROUP_RELATION_KEY)
 474                         goto next2;
 475
 476                 if (found_key.objectid > found_key.offset) {
 477                         /* parent <- member, not needed to build config */
 478                         /* FIXME should we omit the key completely? */
 479                         goto next2;
 480                 }
 481
 482                 ret = add_relation_rb(fs_info, found_key.objectid,
 483                                       found_key.offset);
 484                 if (ret == -ENOENT) {
 485                         btrfs_warn(fs_info,
 486                                 "orphan qgroup relation 0x%llx->0x%llx",
 487                                 found_key.objectid, found_key.offset);
 488                         ret = 0;        /* ignore the error */
 489                 }
 490                 if (ret)
 491                         goto out;
 492 next2:
 493                 ret = btrfs_next_item(quota_root, path);
 494                 if (ret < 0)
 495                         goto out;
 496                 if (ret)
 497                         break;
 498         }
 499 out:
 500         btrfs_free_path(path);
 501         fs_info->qgroup_flags |= flags;
 502         if (!(fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_ON))
 503                 clear_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
 504         else if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN &&
 505                  ret >= 0)
 506                 ret = qgroup_rescan_init(fs_info, rescan_progress, 0);
 507
 508         if (ret < 0) {
 509                 ulist_free(fs_info->qgroup_ulist);
 510                 fs_info->qgroup_ulist = NULL;
 511                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
 512                 btrfs_sysfs_del_qgroups(fs_info);
 513         }
 514
 515         return ret < 0 ? ret : 0;
 516 }
 517
 518 /*
 519  * Called in close_ctree() when quota is still enabled.  This verifies we don't
 520  * leak some reserved space.
 521  *
 522  * Return false if no reserved space is left.
 523  * Return true if some reserved space is leaked.
 524  */
 525 bool btrfs_check_quota_leak(struct btrfs_fs_info *fs_info)
 526 {
 527         struct rb_node *node;
 528         bool ret = false;
 529
 530         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
 531                 return ret;
 532         /*
 533          * Since we're unmounting, there is no race and no need to grab qgroup
 534          * lock.  And here we don't go post-order to provide a more user
 535          * friendly sorted result.
 536          */
 537         for (node = rb_first(&fs_info->qgroup_tree); node; node = rb_next(node)) {
 538                 struct btrfs_qgroup *qgroup;
 539                 int i;
 540
 541                 qgroup = rb_entry(node, struct btrfs_qgroup, node);
 542                 for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++) {
 543                         if (qgroup->rsv.values[i]) {
 544                                 ret = true;
 545                                 btrfs_warn(fs_info,
 546                 "qgroup %hu/%llu has unreleased space, type %d rsv %llu",
 547                                    btrfs_qgroup_level(qgroup->qgroupid),
 548                                    btrfs_qgroup_subvolid(qgroup->qgroupid),
 549                                    i, qgroup->rsv.values[i]);
 550                         }
 551                 }
 552         }
 553         return ret;
 554 }
 555
 556 /*
 557  * This is called from close_ctree() or open_ctree() or btrfs_quota_disable(),
 558  * first two are in single-threaded paths.And for the third one, we have set
 559  * quota_root to be null with qgroup_lock held before, so it is safe to clean
 560  * up the in-memory structures without qgroup_lock held.
 561  */
 562 void btrfs_free_qgroup_config(struct btrfs_fs_info *fs_info)
 563 {
 564         struct rb_node *n;
 565         struct btrfs_qgroup *qgroup;
 566
 567         while ((n = rb_first(&fs_info->qgroup_tree))) {
 568                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
 569                 rb_erase(n, &fs_info->qgroup_tree);
 570                 __del_qgroup_rb(fs_info, qgroup);
 571                 btrfs_sysfs_del_one_qgroup(fs_info, qgroup);
 572                 kfree(qgroup);
 573         }
 574         /*
 575          * We call btrfs_free_qgroup_config() when unmounting
 576          * filesystem and disabling quota, so we set qgroup_ulist
 577          * to be null here to avoid double free.
 578          */
 579         ulist_free(fs_info->qgroup_ulist);
 580         fs_info->qgroup_ulist = NULL;
 581         btrfs_sysfs_del_qgroups(fs_info);
 582 }
 583
 584 static int add_qgroup_relation_item(struct btrfs_trans_handle *trans, u64 src,
 585                                     u64 dst)
 586 {
 587         int ret;
 588         struct btrfs_root *quota_root = trans->fs_info->quota_root;
 589         struct btrfs_path *path;
 590         struct btrfs_key key;
 591
 592         path = btrfs_alloc_path();
 593         if (!path)
 594                 return -ENOMEM;
 595
 596         key.objectid = src;
 597         key.type = BTRFS_QGROUP_RELATION_KEY;
 598         key.offset = dst;
 599
 600         ret = btrfs_insert_empty_item(trans, quota_root, path, &key, 0);
 601
 602         btrfs_mark_buffer_dirty(path->nodes[0]);
 603
 604         btrfs_free_path(path);
 605         return ret;
 606 }
 607
 608 static int del_qgroup_relation_item(struct btrfs_trans_handle *trans, u64 src,
 609                                     u64 dst)
 610 {
 611         int ret;
 612         struct btrfs_root *quota_root = trans->fs_info->quota_root;
 613         struct btrfs_path *path;
 614         struct btrfs_key key;
 615
 616         path = btrfs_alloc_path();
 617         if (!path)
 618                 return -ENOMEM;
 619
 620         key.objectid = src;
 621         key.type = BTRFS_QGROUP_RELATION_KEY;
 622         key.offset = dst;
 623
 624         ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
 625         if (ret < 0)
 626                 goto out;
 627
 628         if (ret > 0) {
 629                 ret = -ENOENT;
 630                 goto out;
 631         }
 632
 633         ret = btrfs_del_item(trans, quota_root, path);
 634 out:
 635         btrfs_free_path(path);
 636         return ret;
 637 }
 638
 639 static int add_qgroup_item(struct btrfs_trans_handle *trans,
 640                            struct btrfs_root *quota_root, u64 qgroupid)
 641 {
 642         int ret;
 643         struct btrfs_path *path;
 644         struct btrfs_qgroup_info_item *qgroup_info;
 645         struct btrfs_qgroup_limit_item *qgroup_limit;
 646         struct extent_buffer *leaf;
 647         struct btrfs_key key;
 648
 649         if (btrfs_is_testing(quota_root->fs_info))
 650                 return 0;
 651
 652         path = btrfs_alloc_path();
 653         if (!path)
 654                 return -ENOMEM;
 655
 656         key.objectid = 0;
 657         key.type = BTRFS_QGROUP_INFO_KEY;
 658         key.offset = qgroupid;
 659
 660         /*
 661          * Avoid a transaction abort by catching -EEXIST here. In that
 662          * case, we proceed by re-initializing the existing structure
 663          * on disk.
 664          */
 665
 666         ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
 667                                       sizeof(*qgroup_info));
 668         if (ret && ret != -EEXIST)
 669                 goto out;
 670
 671         leaf = path->nodes[0];
 672         qgroup_info = btrfs_item_ptr(leaf, path->slots[0],
 673                                  struct btrfs_qgroup_info_item);
 674         btrfs_set_qgroup_info_generation(leaf, qgroup_info, trans->transid);
 675         btrfs_set_qgroup_info_rfer(leaf, qgroup_info, 0);
 676         btrfs_set_qgroup_info_rfer_cmpr(leaf, qgroup_info, 0);
 677         btrfs_set_qgroup_info_excl(leaf, qgroup_info, 0);
 678         btrfs_set_qgroup_info_excl_cmpr(leaf, qgroup_info, 0);
 679
 680         btrfs_mark_buffer_dirty(leaf);
 681
 682         btrfs_release_path(path);
 683
 684         key.type = BTRFS_QGROUP_LIMIT_KEY;
 685         ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
 686                                       sizeof(*qgroup_limit));
 687         if (ret && ret != -EEXIST)
 688                 goto out;
 689
 690         leaf = path->nodes[0];
 691         qgroup_limit = btrfs_item_ptr(leaf, path->slots[0],
 692                                   struct btrfs_qgroup_limit_item);
 693         btrfs_set_qgroup_limit_flags(leaf, qgroup_limit, 0);
 694         btrfs_set_qgroup_limit_max_rfer(leaf, qgroup_limit, 0);
 695         btrfs_set_qgroup_limit_max_excl(leaf, qgroup_limit, 0);
 696         btrfs_set_qgroup_limit_rsv_rfer(leaf, qgroup_limit, 0);
 697         btrfs_set_qgroup_limit_rsv_excl(leaf, qgroup_limit, 0);
 698
 699         btrfs_mark_buffer_dirty(leaf);
 700
 701         ret = 0;
 702 out:
 703         btrfs_free_path(path);
 704         return ret;
 705 }
 706
 707 static int del_qgroup_item(struct btrfs_trans_handle *trans, u64 qgroupid)
 708 {
 709         int ret;
 710         struct btrfs_root *quota_root = trans->fs_info->quota_root;
 711         struct btrfs_path *path;
 712         struct btrfs_key key;
 713
 714         path = btrfs_alloc_path();
 715         if (!path)
 716                 return -ENOMEM;
 717
 718         key.objectid = 0;
 719         key.type = BTRFS_QGROUP_INFO_KEY;
 720         key.offset = qgroupid;
 721         ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
 722         if (ret < 0)
 723                 goto out;
 724
 725         if (ret > 0) {
 726                 ret = -ENOENT;
 727                 goto out;
 728         }
 729
 730         ret = btrfs_del_item(trans, quota_root, path);
 731         if (ret)
 732                 goto out;
 733
 734         btrfs_release_path(path);
 735
 736         key.type = BTRFS_QGROUP_LIMIT_KEY;
 737         ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
 738         if (ret < 0)
 739                 goto out;
 740
 741         if (ret > 0) {
 742                 ret = -ENOENT;
 743                 goto out;
 744         }
 745
 746         ret = btrfs_del_item(trans, quota_root, path);
 747
 748 out:
 749         btrfs_free_path(path);
 750         return ret;
 751 }
 752
 753 static int update_qgroup_limit_item(struct btrfs_trans_handle *trans,
 754                                     struct btrfs_qgroup *qgroup)
 755 {
 756         struct btrfs_root *quota_root = trans->fs_info->quota_root;
 757         struct btrfs_path *path;
 758         struct btrfs_key key;
 759         struct extent_buffer *l;
 760         struct btrfs_qgroup_limit_item *qgroup_limit;
 761         int ret;
 762         int slot;
 763
 764         key.objectid = 0;
 765         key.type = BTRFS_QGROUP_LIMIT_KEY;
 766         key.offset = qgroup->qgroupid;
 767
 768         path = btrfs_alloc_path();
 769         if (!path)
 770                 return -ENOMEM;
 771
 772         ret = btrfs_search_slot(trans, quota_root, &key, path, 0, 1);
 773         if (ret > 0)
 774                 ret = -ENOENT;
 775
 776         if (ret)
 777                 goto out;
 778
 779         l = path->nodes[0];
 780         slot = path->slots[0];
 781         qgroup_limit = btrfs_item_ptr(l, slot, struct btrfs_qgroup_limit_item);
 782         btrfs_set_qgroup_limit_flags(l, qgroup_limit, qgroup->lim_flags);
 783         btrfs_set_qgroup_limit_max_rfer(l, qgroup_limit, qgroup->max_rfer);
 784         btrfs_set_qgroup_limit_max_excl(l, qgroup_limit, qgroup->max_excl);
 785         btrfs_set_qgroup_limit_rsv_rfer(l, qgroup_limit, qgroup->rsv_rfer);
 786         btrfs_set_qgroup_limit_rsv_excl(l, qgroup_limit, qgroup->rsv_excl);
 787
 788         btrfs_mark_buffer_dirty(l);
 789
 790 out:
 791         btrfs_free_path(path);
 792         return ret;
 793 }
 794
 795 static int update_qgroup_info_item(struct btrfs_trans_handle *trans,
 796                                    struct btrfs_qgroup *qgroup)
 797 {
 798         struct btrfs_fs_info *fs_info = trans->fs_info;
 799         struct btrfs_root *quota_root = fs_info->quota_root;
 800         struct btrfs_path *path;
 801         struct btrfs_key key;
 802         struct extent_buffer *l;
 803         struct btrfs_qgroup_info_item *qgroup_info;
 804         int ret;
 805         int slot;
 806
 807         if (btrfs_is_testing(fs_info))
 808                 return 0;
 809
 810         key.objectid = 0;
 811         key.type = BTRFS_QGROUP_INFO_KEY;
 812         key.offset = qgroup->qgroupid;
 813
 814         path = btrfs_alloc_path();
 815         if (!path)
 816                 return -ENOMEM;
 817
 818         ret = btrfs_search_slot(trans, quota_root, &key, path, 0, 1);
 819         if (ret > 0)
 820                 ret = -ENOENT;
 821
 822         if (ret)
 823                 goto out;
 824
 825         l = path->nodes[0];
 826         slot = path->slots[0];
 827         qgroup_info = btrfs_item_ptr(l, slot, struct btrfs_qgroup_info_item);
 828         btrfs_set_qgroup_info_generation(l, qgroup_info, trans->transid);
 829         btrfs_set_qgroup_info_rfer(l, qgroup_info, qgroup->rfer);
 830         btrfs_set_qgroup_info_rfer_cmpr(l, qgroup_info, qgroup->rfer_cmpr);
 831         btrfs_set_qgroup_info_excl(l, qgroup_info, qgroup->excl);
 832         btrfs_set_qgroup_info_excl_cmpr(l, qgroup_info, qgroup->excl_cmpr);
 833
 834         btrfs_mark_buffer_dirty(l);
 835
 836 out:
 837         btrfs_free_path(path);
 838         return ret;
 839 }
 840
 841 static int update_qgroup_status_item(struct btrfs_trans_handle *trans)
 842 {
 843         struct btrfs_fs_info *fs_info = trans->fs_info;
 844         struct btrfs_root *quota_root = fs_info->quota_root;
 845         struct btrfs_path *path;
 846         struct btrfs_key key;
 847         struct extent_buffer *l;
 848         struct btrfs_qgroup_status_item *ptr;
 849         int ret;
 850         int slot;
 851
 852         key.objectid = 0;
 853         key.type = BTRFS_QGROUP_STATUS_KEY;
 854         key.offset = 0;
 855
 856         path = btrfs_alloc_path();
 857         if (!path)
 858                 return -ENOMEM;
 859
 860         ret = btrfs_search_slot(trans, quota_root, &key, path, 0, 1);
 861         if (ret > 0)
 862                 ret = -ENOENT;
 863
 864         if (ret)
 865                 goto out;
 866
 867         l = path->nodes[0];
 868         slot = path->slots[0];
 869         ptr = btrfs_item_ptr(l, slot, struct btrfs_qgroup_status_item);
 870         btrfs_set_qgroup_status_flags(l, ptr, fs_info->qgroup_flags);
 871         btrfs_set_qgroup_status_generation(l, ptr, trans->transid);
 872         btrfs_set_qgroup_status_rescan(l, ptr,
 873                                 fs_info->qgroup_rescan_progress.objectid);
 874
 875         btrfs_mark_buffer_dirty(l);
 876
 877 out:
 878         btrfs_free_path(path);
 879         return ret;
 880 }
 881
 882 /*
 883  * called with qgroup_lock held
 884  */
 885 static int btrfs_clean_quota_tree(struct btrfs_trans_handle *trans,
 886                                   struct btrfs_root *root)
 887 {
 888         struct btrfs_path *path;
 889         struct btrfs_key key;
 890         struct extent_buffer *leaf = NULL;
 891         int ret;
 892         int nr = 0;
 893
 894         path = btrfs_alloc_path();
 895         if (!path)
 896                 return -ENOMEM;
 897
 898         key.objectid = 0;
 899         key.offset = 0;
 900         key.type = 0;
 901
 902         while (1) {
 903                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 904                 if (ret < 0)
 905                         goto out;
 906                 leaf = path->nodes[0];
 907                 nr = btrfs_header_nritems(leaf);
 908                 if (!nr)
 909                         break;
 910                 /*
 911                  * delete the leaf one by one
 912                  * since the whole tree is going
 913                  * to be deleted.
 914                  */
 915                 path->slots[0] = 0;
 916                 ret = btrfs_del_items(trans, root, path, 0, nr);
 917                 if (ret)
 918                         goto out;
 919
 920                 btrfs_release_path(path);
 921         }
 922         ret = 0;
 923 out:
 924         btrfs_free_path(path);
 925         return ret;
 926 }
 927
 928 int btrfs_quota_enable(struct btrfs_fs_info *fs_info)
 929 {
 930         struct btrfs_root *quota_root;
 931         struct btrfs_root *tree_root = fs_info->tree_root;
 932         struct btrfs_path *path = NULL;
 933         struct btrfs_qgroup_status_item *ptr;
 934         struct extent_buffer *leaf;
 935         struct btrfs_key key;
 936         struct btrfs_key found_key;
 937         struct btrfs_qgroup *qgroup = NULL;
 938         struct btrfs_trans_handle *trans = NULL;
 939         struct ulist *ulist = NULL;
 940         int ret = 0;
 941         int slot;
 942
 943         /*
 944          * We need to have subvol_sem write locked, to prevent races between
 945          * concurrent tasks trying to enable quotas, because we will unlock
 946          * and relock qgroup_ioctl_lock before setting fs_info->quota_root
 947          * and before setting BTRFS_FS_QUOTA_ENABLED.
 948          */
 949         lockdep_assert_held_write(&fs_info->subvol_sem);
 950
 951         mutex_lock(&fs_info->qgroup_ioctl_lock);
 952         if (fs_info->quota_root)
 953                 goto out;
 954
 955         ulist = ulist_alloc(GFP_KERNEL);
 956         if (!ulist) {
 957                 ret = -ENOMEM;
 958                 goto out;
 959         }
 960
 961         ret = btrfs_sysfs_add_qgroups(fs_info);
 962         if (ret < 0)
 963                 goto out;
 964
 965         /*
 966          * Unlock qgroup_ioctl_lock before starting the transaction. This is to
 967          * avoid lock acquisition inversion problems (reported by lockdep) between
 968          * qgroup_ioctl_lock and the vfs freeze semaphores, acquired when we
 969          * start a transaction.
 970          * After we started the transaction lock qgroup_ioctl_lock again and
 971          * check if someone else created the quota root in the meanwhile. If so,
 972          * just return success and release the transaction handle.
 973          *
 974          * Also we don't need to worry about someone else calling
 975          * btrfs_sysfs_add_qgroups() after we unlock and getting an error because
 976          * that function returns 0 (success) when the sysfs entries already exist.
 977          */
 978         mutex_unlock(&fs_info->qgroup_ioctl_lock);
 979
 980         /*
 981          * 1 for quota root item
 982          * 1 for BTRFS_QGROUP_STATUS item
 983          *
 984          * Yet we also need 2*n items for a QGROUP_INFO/QGROUP_LIMIT items
 985          * per subvolume. However those are not currently reserved since it
 986          * would be a lot of overkill.
 987          */
 988         trans = btrfs_start_transaction(tree_root, 2);
 989
 990         mutex_lock(&fs_info->qgroup_ioctl_lock);
 991         if (IS_ERR(trans)) {
 992                 ret = PTR_ERR(trans);
 993                 trans = NULL;
 994                 goto out;
 995         }
 996
 997         if (fs_info->quota_root)
 998                 goto out;
 999
1000         fs_info->qgroup_ulist = ulist;
1001         ulist = NULL;
1002
1003         /*
1004          * initially create the quota tree
1005          */
1006         quota_root = btrfs_create_tree(trans, BTRFS_QUOTA_TREE_OBJECTID);
1007         if (IS_ERR(quota_root)) {
1008                 ret =  PTR_ERR(quota_root);
1009                 btrfs_abort_transaction(trans, ret);
1010                 goto out;
1011         }
1012
1013         path = btrfs_alloc_path();
1014         if (!path) {
1015                 ret = -ENOMEM;
1016                 btrfs_abort_transaction(trans, ret);
1017                 goto out_free_root;
1018         }
1019
1020         key.objectid = 0;
1021         key.type = BTRFS_QGROUP_STATUS_KEY;
1022         key.offset = 0;
1023
1024         ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
1025                                       sizeof(*ptr));
1026         if (ret) {
1027                 btrfs_abort_transaction(trans, ret);
1028                 goto out_free_path;
1029         }
1030
1031         leaf = path->nodes[0];
1032         ptr = btrfs_item_ptr(leaf, path->slots[0],
1033                                  struct btrfs_qgroup_status_item);
1034         btrfs_set_qgroup_status_generation(leaf, ptr, trans->transid);
1035         btrfs_set_qgroup_status_version(leaf, ptr, BTRFS_QGROUP_STATUS_VERSION);
1036         fs_info->qgroup_flags = BTRFS_QGROUP_STATUS_FLAG_ON |
1037                                 BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1038         btrfs_set_qgroup_status_flags(leaf, ptr, fs_info->qgroup_flags);
1039         btrfs_set_qgroup_status_rescan(leaf, ptr, 0);
1040
1041         btrfs_mark_buffer_dirty(leaf);
1042
1043         key.objectid = 0;
1044         key.type = BTRFS_ROOT_REF_KEY;
1045         key.offset = 0;
1046
1047         btrfs_release_path(path);
1048         ret = btrfs_search_slot_for_read(tree_root, &key, path, 1, 0);
1049         if (ret > 0)
1050                 goto out_add_root;
1051         if (ret < 0) {
1052                 btrfs_abort_transaction(trans, ret);
1053                 goto out_free_path;
1054         }
1055
1056         while (1) {
1057                 slot = path->slots[0];
1058                 leaf = path->nodes[0];
1059                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1060
1061                 if (found_key.type == BTRFS_ROOT_REF_KEY) {
1062
1063                         /* Release locks on tree_root before we access quota_root */
1064                         btrfs_release_path(path);
1065
1066                         ret = add_qgroup_item(trans, quota_root,
1067                                               found_key.offset);
1068                         if (ret) {
1069                                 btrfs_abort_transaction(trans, ret);
1070                                 goto out_free_path;
1071                         }
1072
1073                         qgroup = add_qgroup_rb(fs_info, found_key.offset);
1074                         if (IS_ERR(qgroup)) {
1075                                 ret = PTR_ERR(qgroup);
1076                                 btrfs_abort_transaction(trans, ret);
1077                                 goto out_free_path;
1078                         }
1079                         ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1080                         if (ret < 0) {
1081                                 btrfs_abort_transaction(trans, ret);
1082                                 goto out_free_path;
1083                         }
1084                         ret = btrfs_search_slot_for_read(tree_root, &found_key,
1085                                                          path, 1, 0);
1086                         if (ret < 0) {
1087                                 btrfs_abort_transaction(trans, ret);
1088                                 goto out_free_path;
1089                         }
1090                         if (ret > 0) {
1091                                 /*
1092                                  * Shouldn't happen, but in case it does we
1093                                  * don't need to do the btrfs_next_item, just
1094                                  * continue.
1095                                  */
1096                                 continue;
1097                         }
1098                 }
1099                 ret = btrfs_next_item(tree_root, path);
1100                 if (ret < 0) {
1101                         btrfs_abort_transaction(trans, ret);
1102                         goto out_free_path;
1103                 }
1104                 if (ret)
1105                         break;
1106         }
1107
1108 out_add_root:
1109         btrfs_release_path(path);
1110         ret = add_qgroup_item(trans, quota_root, BTRFS_FS_TREE_OBJECTID);
1111         if (ret) {
1112                 btrfs_abort_transaction(trans, ret);
1113                 goto out_free_path;
1114         }
1115
1116         qgroup = add_qgroup_rb(fs_info, BTRFS_FS_TREE_OBJECTID);
1117         if (IS_ERR(qgroup)) {
1118                 ret = PTR_ERR(qgroup);
1119                 btrfs_abort_transaction(trans, ret);
1120                 goto out_free_path;
1121         }
1122         ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1123         if (ret < 0) {
1124                 btrfs_abort_transaction(trans, ret);
1125                 goto out_free_path;
1126         }
1127
1128         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1129         /*
1130          * Commit the transaction while not holding qgroup_ioctl_lock, to avoid
1131          * a deadlock with tasks concurrently doing other qgroup operations, such
1132          * adding/removing qgroups or adding/deleting qgroup relations for example,
1133          * because all qgroup operations first start or join a transaction and then
1134          * lock the qgroup_ioctl_lock mutex.
1135          * We are safe from a concurrent task trying to enable quotas, by calling
1136          * this function, since we are serialized by fs_info->subvol_sem.
1137          */
1138         ret = btrfs_commit_transaction(trans);
1139         trans = NULL;
1140         mutex_lock(&fs_info->qgroup_ioctl_lock);
1141         if (ret)
1142                 goto out_free_path;
1143
1144         /*
1145          * Set quota enabled flag after committing the transaction, to avoid
1146          * deadlocks on fs_info->qgroup_ioctl_lock with concurrent snapshot
1147          * creation.
1148          */
1149         spin_lock(&fs_info->qgroup_lock);
1150         fs_info->quota_root = quota_root;
1151         set_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1152         spin_unlock(&fs_info->qgroup_lock);
1153
1154         ret = qgroup_rescan_init(fs_info, 0, 1);
1155         if (!ret) {
1156                 qgroup_rescan_zero_tracking(fs_info);
1157                 fs_info->qgroup_rescan_running = true;
1158                 btrfs_queue_work(fs_info->qgroup_rescan_workers,
1159                                  &fs_info->qgroup_rescan_work);
1160         }
1161
1162 out_free_path:
1163         btrfs_free_path(path);
1164 out_free_root:
1165         if (ret)
1166                 btrfs_put_root(quota_root);
1167 out:
1168         if (ret) {
1169                 ulist_free(fs_info->qgroup_ulist);
1170                 fs_info->qgroup_ulist = NULL;
1171                 btrfs_sysfs_del_qgroups(fs_info);
1172         }
1173         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1174         if (ret && trans)
1175                 btrfs_end_transaction(trans);
1176         else if (trans)
1177                 ret = btrfs_end_transaction(trans);
1178         ulist_free(ulist);
1179         return ret;
1180 }
1181
1182 int btrfs_quota_disable(struct btrfs_fs_info *fs_info)
1183 {
1184         struct btrfs_root *quota_root;
1185         struct btrfs_trans_handle *trans = NULL;
1186         int ret = 0;
1187
1188         /*
1189          * We need to have subvol_sem write locked, to prevent races between
1190          * concurrent tasks trying to disable quotas, because we will unlock
1191          * and relock qgroup_ioctl_lock across BTRFS_FS_QUOTA_ENABLED changes.
1192          */
1193         lockdep_assert_held_write(&fs_info->subvol_sem);
1194
1195         mutex_lock(&fs_info->qgroup_ioctl_lock);
1196         if (!fs_info->quota_root)
1197                 goto out;
1198
1199         /*
1200          * Unlock the qgroup_ioctl_lock mutex before waiting for the rescan worker to
1201          * complete. Otherwise we can deadlock because btrfs_remove_qgroup() needs
1202          * to lock that mutex while holding a transaction handle and the rescan
1203          * worker needs to commit a transaction.
1204          */
1205         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1206
1207         /*
1208          * Request qgroup rescan worker to complete and wait for it. This wait
1209          * must be done before transaction start for quota disable since it may
1210          * deadlock with transaction by the qgroup rescan worker.
1211          */
1212         clear_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1213         btrfs_qgroup_wait_for_completion(fs_info, false);
1214
1215         /*
1216          * 1 For the root item
1217          *
1218          * We should also reserve enough items for the quota tree deletion in
1219          * btrfs_clean_quota_tree but this is not done.
1220          *
1221          * Also, we must always start a transaction without holding the mutex
1222          * qgroup_ioctl_lock, see btrfs_quota_enable().
1223          */
1224         trans = btrfs_start_transaction(fs_info->tree_root, 1);
1225
1226         mutex_lock(&fs_info->qgroup_ioctl_lock);
1227         if (IS_ERR(trans)) {
1228                 ret = PTR_ERR(trans);
1229                 trans = NULL;
1230                 set_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1231                 goto out;
1232         }
1233
1234         if (!fs_info->quota_root)
1235                 goto out;
1236
1237         spin_lock(&fs_info->qgroup_lock);
1238         quota_root = fs_info->quota_root;
1239         fs_info->quota_root = NULL;
1240         fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_ON;
1241         spin_unlock(&fs_info->qgroup_lock);
1242
1243         btrfs_free_qgroup_config(fs_info);
1244
1245         ret = btrfs_clean_quota_tree(trans, quota_root);
1246         if (ret) {
1247                 btrfs_abort_transaction(trans, ret);
1248                 goto out;
1249         }
1250
1251         ret = btrfs_del_root(trans, &quota_root->root_key);
1252         if (ret) {
1253                 btrfs_abort_transaction(trans, ret);
1254                 goto out;
1255         }
1256
1257         list_del(&quota_root->dirty_list);
1258
1259         btrfs_tree_lock(quota_root->node);
1260         btrfs_clean_tree_block(quota_root->node);
1261         btrfs_tree_unlock(quota_root->node);
1262         btrfs_free_tree_block(trans, btrfs_root_id(quota_root),
1263                               quota_root->node, 0, 1);
1264
1265         btrfs_put_root(quota_root);
1266
1267 out:
1268         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1269         if (ret && trans)
1270                 btrfs_end_transaction(trans);
1271         else if (trans)
1272                 ret = btrfs_end_transaction(trans);
1273
1274         return ret;
1275 }
1276
1277 static void qgroup_dirty(struct btrfs_fs_info *fs_info,
1278                          struct btrfs_qgroup *qgroup)
1279 {
1280         if (list_empty(&qgroup->dirty))
1281                 list_add(&qgroup->dirty, &fs_info->dirty_qgroups);
1282 }
1283
1284 /*
1285  * The easy accounting, we're updating qgroup relationship whose child qgroup
1286  * only has exclusive extents.
1287  *
1288  * In this case, all exclusive extents will also be exclusive for parent, so
1289  * excl/rfer just get added/removed.
1290  *
1291  * So is qgroup reservation space, which should also be added/removed to
1292  * parent.
1293  * Or when child tries to release reservation space, parent will underflow its
1294  * reservation (for relationship adding case).
1295  *
1296  * Caller should hold fs_info->qgroup_lock.
1297  */
1298 static int __qgroup_excl_accounting(struct btrfs_fs_info *fs_info,
1299                                     struct ulist *tmp, u64 ref_root,
1300                                     struct btrfs_qgroup *src, int sign)
1301 {
1302         struct btrfs_qgroup *qgroup;
1303         struct btrfs_qgroup_list *glist;
1304         struct ulist_node *unode;
1305         struct ulist_iterator uiter;
1306         u64 num_bytes = src->excl;
1307         int ret = 0;
1308
1309         qgroup = find_qgroup_rb(fs_info, ref_root);
1310         if (!qgroup)
1311                 goto out;
1312
1313         qgroup->rfer += sign * num_bytes;
1314         qgroup->rfer_cmpr += sign * num_bytes;
1315
1316         WARN_ON(sign < 0 && qgroup->excl < num_bytes);
1317         qgroup->excl += sign * num_bytes;
1318         qgroup->excl_cmpr += sign * num_bytes;
1319
1320         if (sign > 0)
1321                 qgroup_rsv_add_by_qgroup(fs_info, qgroup, src);
1322         else
1323                 qgroup_rsv_release_by_qgroup(fs_info, qgroup, src);
1324
1325         qgroup_dirty(fs_info, qgroup);
1326
1327         /* Get all of the parent groups that contain this qgroup */
1328         list_for_each_entry(glist, &qgroup->groups, next_group) {
1329                 ret = ulist_add(tmp, glist->group->qgroupid,
1330                                 qgroup_to_aux(glist->group), GFP_ATOMIC);
1331                 if (ret < 0)
1332                         goto out;
1333         }
1334
1335         /* Iterate all of the parents and adjust their reference counts */
1336         ULIST_ITER_INIT(&uiter);
1337         while ((unode = ulist_next(tmp, &uiter))) {
1338                 qgroup = unode_aux_to_qgroup(unode);
1339                 qgroup->rfer += sign * num_bytes;
1340                 qgroup->rfer_cmpr += sign * num_bytes;
1341                 WARN_ON(sign < 0 && qgroup->excl < num_bytes);
1342                 qgroup->excl += sign * num_bytes;
1343                 if (sign > 0)
1344                         qgroup_rsv_add_by_qgroup(fs_info, qgroup, src);
1345                 else
1346                         qgroup_rsv_release_by_qgroup(fs_info, qgroup, src);
1347                 qgroup->excl_cmpr += sign * num_bytes;
1348                 qgroup_dirty(fs_info, qgroup);
1349
1350                 /* Add any parents of the parents */
1351                 list_for_each_entry(glist, &qgroup->groups, next_group) {
1352                         ret = ulist_add(tmp, glist->group->qgroupid,
1353                                         qgroup_to_aux(glist->group), GFP_ATOMIC);
1354                         if (ret < 0)
1355                                 goto out;
1356                 }
1357         }
1358         ret = 0;
1359 out:
1360         return ret;
1361 }
1362
1363
1364 /*
1365  * Quick path for updating qgroup with only excl refs.
1366  *
1367  * In that case, just update all parent will be enough.
1368  * Or we needs to do a full rescan.
1369  * Caller should also hold fs_info->qgroup_lock.
1370  *
1371  * Return 0 for quick update, return >0 for need to full rescan
1372  * and mark INCONSISTENT flag.
1373  * Return < 0 for other error.
1374  */
1375 static int quick_update_accounting(struct btrfs_fs_info *fs_info,
1376                                    struct ulist *tmp, u64 src, u64 dst,
1377                                    int sign)
1378 {
1379         struct btrfs_qgroup *qgroup;
1380         int ret = 1;
1381         int err = 0;
1382
1383         qgroup = find_qgroup_rb(fs_info, src);
1384         if (!qgroup)
1385                 goto out;
1386         if (qgroup->excl == qgroup->rfer) {
1387                 ret = 0;
1388                 err = __qgroup_excl_accounting(fs_info, tmp, dst,
1389                                                qgroup, sign);
1390                 if (err < 0) {
1391                         ret = err;
1392                         goto out;
1393                 }
1394         }
1395 out:
1396         if (ret)
1397                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1398         return ret;
1399 }
1400
1401 int btrfs_add_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1402                               u64 dst)
1403 {
1404         struct btrfs_fs_info *fs_info = trans->fs_info;
1405         struct btrfs_qgroup *parent;
1406         struct btrfs_qgroup *member;
1407         struct btrfs_qgroup_list *list;
1408         struct ulist *tmp;
1409         unsigned int nofs_flag;
1410         int ret = 0;
1411
1412         /* Check the level of src and dst first */
1413         if (btrfs_qgroup_level(src) >= btrfs_qgroup_level(dst))
1414                 return -EINVAL;
1415
1416         /* We hold a transaction handle open, must do a NOFS allocation. */
1417         nofs_flag = memalloc_nofs_save();
1418         tmp = ulist_alloc(GFP_KERNEL);
1419         memalloc_nofs_restore(nofs_flag);
1420         if (!tmp)
1421                 return -ENOMEM;
1422
1423         mutex_lock(&fs_info->qgroup_ioctl_lock);
1424         if (!fs_info->quota_root) {
1425                 ret = -ENOTCONN;
1426                 goto out;
1427         }
1428         member = find_qgroup_rb(fs_info, src);
1429         parent = find_qgroup_rb(fs_info, dst);
1430         if (!member || !parent) {
1431                 ret = -EINVAL;
1432                 goto out;
1433         }
1434
1435         /* check if such qgroup relation exist firstly */
1436         list_for_each_entry(list, &member->groups, next_group) {
1437                 if (list->group == parent) {
1438                         ret = -EEXIST;
1439                         goto out;
1440                 }
1441         }
1442
1443         ret = add_qgroup_relation_item(trans, src, dst);
1444         if (ret)
1445                 goto out;
1446
1447         ret = add_qgroup_relation_item(trans, dst, src);
1448         if (ret) {
1449                 del_qgroup_relation_item(trans, src, dst);
1450                 goto out;
1451         }
1452
1453         spin_lock(&fs_info->qgroup_lock);
1454         ret = add_relation_rb(fs_info, src, dst);
1455         if (ret < 0) {
1456                 spin_unlock(&fs_info->qgroup_lock);
1457                 goto out;
1458         }
1459         ret = quick_update_accounting(fs_info, tmp, src, dst, 1);
1460         spin_unlock(&fs_info->qgroup_lock);
1461 out:
1462         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1463         ulist_free(tmp);
1464         return ret;
1465 }
1466
1467 static int __del_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1468                                  u64 dst)
1469 {
1470         struct btrfs_fs_info *fs_info = trans->fs_info;
1471         struct btrfs_qgroup *parent;
1472         struct btrfs_qgroup *member;
1473         struct btrfs_qgroup_list *list;
1474         struct ulist *tmp;
1475         bool found = false;
1476         unsigned int nofs_flag;
1477         int ret = 0;
1478         int ret2;
1479
1480         /* We hold a transaction handle open, must do a NOFS allocation. */
1481         nofs_flag = memalloc_nofs_save();
1482         tmp = ulist_alloc(GFP_KERNEL);
1483         memalloc_nofs_restore(nofs_flag);
1484         if (!tmp)
1485                 return -ENOMEM;
1486
1487         if (!fs_info->quota_root) {
1488                 ret = -ENOTCONN;
1489                 goto out;
1490         }
1491
1492         member = find_qgroup_rb(fs_info, src);
1493         parent = find_qgroup_rb(fs_info, dst);
1494         /*
1495          * The parent/member pair doesn't exist, then try to delete the dead
1496          * relation items only.
1497          */
1498         if (!member || !parent)
1499                 goto delete_item;
1500
1501         /* check if such qgroup relation exist firstly */
1502         list_for_each_entry(list, &member->groups, next_group) {
1503                 if (list->group == parent) {
1504                         found = true;
1505                         break;
1506                 }
1507         }
1508
1509 delete_item:
1510         ret = del_qgroup_relation_item(trans, src, dst);
1511         if (ret < 0 && ret != -ENOENT)
1512                 goto out;
1513         ret2 = del_qgroup_relation_item(trans, dst, src);
1514         if (ret2 < 0 && ret2 != -ENOENT)
1515                 goto out;
1516
1517         /* At least one deletion succeeded, return 0 */
1518         if (!ret || !ret2)
1519                 ret = 0;
1520
1521         if (found) {
1522                 spin_lock(&fs_info->qgroup_lock);
1523                 del_relation_rb(fs_info, src, dst);
1524                 ret = quick_update_accounting(fs_info, tmp, src, dst, -1);
1525                 spin_unlock(&fs_info->qgroup_lock);
1526         }
1527 out:
1528         ulist_free(tmp);
1529         return ret;
1530 }
1531
1532 int btrfs_del_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1533                               u64 dst)
1534 {
1535         struct btrfs_fs_info *fs_info = trans->fs_info;
1536         int ret = 0;
1537
1538         mutex_lock(&fs_info->qgroup_ioctl_lock);
1539         ret = __del_qgroup_relation(trans, src, dst);
1540         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1541
1542         return ret;
1543 }
1544
1545 int btrfs_create_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid)
1546 {
1547         struct btrfs_fs_info *fs_info = trans->fs_info;
1548         struct btrfs_root *quota_root;
1549         struct btrfs_qgroup *qgroup;
1550         int ret = 0;
1551
1552         mutex_lock(&fs_info->qgroup_ioctl_lock);
1553         if (!fs_info->quota_root) {
1554                 ret = -ENOTCONN;
1555                 goto out;
1556         }
1557         quota_root = fs_info->quota_root;
1558         qgroup = find_qgroup_rb(fs_info, qgroupid);
1559         if (qgroup) {
1560                 ret = -EEXIST;
1561                 goto out;
1562         }
1563
1564         ret = add_qgroup_item(trans, quota_root, qgroupid);
1565         if (ret)
1566                 goto out;
1567
1568         spin_lock(&fs_info->qgroup_lock);
1569         qgroup = add_qgroup_rb(fs_info, qgroupid);
1570         spin_unlock(&fs_info->qgroup_lock);
1571
1572         if (IS_ERR(qgroup)) {
1573                 ret = PTR_ERR(qgroup);
1574                 goto out;
1575         }
1576         ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1577 out:
1578         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1579         return ret;
1580 }
1581
1582 int btrfs_remove_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid)
1583 {
1584         struct btrfs_fs_info *fs_info = trans->fs_info;
1585         struct btrfs_qgroup *qgroup;
1586         struct btrfs_qgroup_list *list;
1587         int ret = 0;
1588
1589         mutex_lock(&fs_info->qgroup_ioctl_lock);
1590         if (!fs_info->quota_root) {
1591                 ret = -ENOTCONN;
1592                 goto out;
1593         }
1594
1595         qgroup = find_qgroup_rb(fs_info, qgroupid);
1596         if (!qgroup) {
1597                 ret = -ENOENT;
1598                 goto out;
1599         }
1600
1601         /* Check if there are no children of this qgroup */
1602         if (!list_empty(&qgroup->members)) {
1603                 ret = -EBUSY;
1604                 goto out;
1605         }
1606
1607         ret = del_qgroup_item(trans, qgroupid);
1608         if (ret && ret != -ENOENT)
1609                 goto out;
1610
1611         while (!list_empty(&qgroup->groups)) {
1612                 list = list_first_entry(&qgroup->groups,
1613                                         struct btrfs_qgroup_list, next_group);
1614                 ret = __del_qgroup_relation(trans, qgroupid,
1615                                             list->group->qgroupid);
1616                 if (ret)
1617                         goto out;
1618         }
1619
1620         spin_lock(&fs_info->qgroup_lock);
1621         del_qgroup_rb(fs_info, qgroupid);
1622         spin_unlock(&fs_info->qgroup_lock);
1623
1624         /*
1625          * Remove the qgroup from sysfs now without holding the qgroup_lock
1626          * spinlock, since the sysfs_remove_group() function needs to take
1627          * the mutex kernfs_mutex through kernfs_remove_by_name_ns().
1628          */
1629         btrfs_sysfs_del_one_qgroup(fs_info, qgroup);
1630         kfree(qgroup);
1631 out:
1632         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1633         return ret;
1634 }
1635
1636 int btrfs_limit_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid,
1637                        struct btrfs_qgroup_limit *limit)
1638 {
1639         struct btrfs_fs_info *fs_info = trans->fs_info;
1640         struct btrfs_qgroup *qgroup;
1641         int ret = 0;
1642         /* Sometimes we would want to clear the limit on this qgroup.
1643          * To meet this requirement, we treat the -1 as a special value
1644          * which tell kernel to clear the limit on this qgroup.
1645          */
1646         const u64 CLEAR_VALUE = -1;
1647
1648         mutex_lock(&fs_info->qgroup_ioctl_lock);
1649         if (!fs_info->quota_root) {
1650                 ret = -ENOTCONN;
1651                 goto out;
1652         }
1653
1654         qgroup = find_qgroup_rb(fs_info, qgroupid);
1655         if (!qgroup) {
1656                 ret = -ENOENT;
1657                 goto out;
1658         }
1659
1660         spin_lock(&fs_info->qgroup_lock);
1661         if (limit->flags & BTRFS_QGROUP_LIMIT_MAX_RFER) {
1662                 if (limit->max_rfer == CLEAR_VALUE) {
1663                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_MAX_RFER;
1664                         limit->flags &= ~BTRFS_QGROUP_LIMIT_MAX_RFER;
1665                         qgroup->max_rfer = 0;
1666                 } else {
1667                         qgroup->max_rfer = limit->max_rfer;
1668                 }
1669         }
1670         if (limit->flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) {
1671                 if (limit->max_excl == CLEAR_VALUE) {
1672                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_MAX_EXCL;
1673                         limit->flags &= ~BTRFS_QGROUP_LIMIT_MAX_EXCL;
1674                         qgroup->max_excl = 0;
1675                 } else {
1676                         qgroup->max_excl = limit->max_excl;
1677                 }
1678         }
1679         if (limit->flags & BTRFS_QGROUP_LIMIT_RSV_RFER) {
1680                 if (limit->rsv_rfer == CLEAR_VALUE) {
1681                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_RSV_RFER;
1682                         limit->flags &= ~BTRFS_QGROUP_LIMIT_RSV_RFER;
1683                         qgroup->rsv_rfer = 0;
1684                 } else {
1685                         qgroup->rsv_rfer = limit->rsv_rfer;
1686                 }
1687         }
1688         if (limit->flags & BTRFS_QGROUP_LIMIT_RSV_EXCL) {
1689                 if (limit->rsv_excl == CLEAR_VALUE) {
1690                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_RSV_EXCL;
1691                         limit->flags &= ~BTRFS_QGROUP_LIMIT_RSV_EXCL;
1692                         qgroup->rsv_excl = 0;
1693                 } else {
1694                         qgroup->rsv_excl = limit->rsv_excl;
1695                 }
1696         }
1697         qgroup->lim_flags |= limit->flags;
1698
1699         spin_unlock(&fs_info->qgroup_lock);
1700
1701         ret = update_qgroup_limit_item(trans, qgroup);
1702         if (ret) {
1703                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1704                 btrfs_info(fs_info, "unable to update quota limit for %llu",
1705                        qgroupid);
1706         }
1707
1708 out:
1709         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1710         return ret;
1711 }
1712
1713 int btrfs_qgroup_trace_extent_nolock(struct btrfs_fs_info *fs_info,
1714                                 struct btrfs_delayed_ref_root *delayed_refs,
1715                                 struct btrfs_qgroup_extent_record *record)
1716 {
1717         struct rb_node **p = &delayed_refs->dirty_extent_root.rb_node;
1718         struct rb_node *parent_node = NULL;
1719         struct btrfs_qgroup_extent_record *entry;
1720         u64 bytenr = record->bytenr;
1721
1722         lockdep_assert_held(&delayed_refs->lock);
1723         trace_btrfs_qgroup_trace_extent(fs_info, record);
1724
1725         while (*p) {
1726                 parent_node = *p;
1727                 entry = rb_entry(parent_node, struct btrfs_qgroup_extent_record,
1728                                  node);
1729                 if (bytenr < entry->bytenr) {
1730                         p = &(*p)->rb_left;
1731                 } else if (bytenr > entry->bytenr) {
1732                         p = &(*p)->rb_right;
1733                 } else {
1734                         if (record->data_rsv && !entry->data_rsv) {
1735                                 entry->data_rsv = record->data_rsv;
1736                                 entry->data_rsv_refroot =
1737                                         record->data_rsv_refroot;
1738                         }
1739                         return 1;
1740                 }
1741         }
1742
1743         rb_link_node(&record->node, parent_node, p);
1744         rb_insert_color(&record->node, &delayed_refs->dirty_extent_root);
1745         return 0;
1746 }
1747
1748 int btrfs_qgroup_trace_extent_post(struct btrfs_trans_handle *trans,
1749                                    struct btrfs_qgroup_extent_record *qrecord)
1750 {
1751         struct ulist *old_root;
1752         u64 bytenr = qrecord->bytenr;
1753         int ret;
1754
1755         /*
1756          * We are always called in a context where we are already holding a
1757          * transaction handle. Often we are called when adding a data delayed
1758          * reference from btrfs_truncate_inode_items() (truncating or unlinking),
1759          * in which case we will be holding a write lock on extent buffer from a
1760          * subvolume tree. In this case we can't allow btrfs_find_all_roots() to
1761          * acquire fs_info->commit_root_sem, because that is a higher level lock
1762          * that must be acquired before locking any extent buffers.
1763          *
1764          * So we want btrfs_find_all_roots() to not acquire the commit_root_sem
1765          * but we can't pass it a non-NULL transaction handle, because otherwise
1766          * it would not use commit roots and would lock extent buffers, causing
1767          * a deadlock if it ends up trying to read lock the same extent buffer
1768          * that was previously write locked at btrfs_truncate_inode_items().
1769          *
1770          * So pass a NULL transaction handle to btrfs_find_all_roots() and
1771          * explicitly tell it to not acquire the commit_root_sem - if we are
1772          * holding a transaction handle we don't need its protection.
1773          */
1774         ASSERT(trans != NULL);
1775
1776         ret = btrfs_find_all_roots(NULL, trans->fs_info, bytenr, 0, &old_root,
1777                                    true);
1778         if (ret < 0) {
1779                 trans->fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1780                 btrfs_warn(trans->fs_info,
1781 "error accounting new delayed refs extent (err code: %d), quota inconsistent",
1782                         ret);
1783                 return 0;
1784         }
1785
1786         /*
1787          * Here we don't need to get the lock of
1788          * trans->transaction->delayed_refs, since inserted qrecord won't
1789          * be deleted, only qrecord->node may be modified (new qrecord insert)
1790          *
1791          * So modifying qrecord->old_roots is safe here
1792          */
1793         qrecord->old_roots = old_root;
1794         return 0;
1795 }
1796
1797 int btrfs_qgroup_trace_extent(struct btrfs_trans_handle *trans, u64 bytenr,
1798                               u64 num_bytes, gfp_t gfp_flag)
1799 {
1800         struct btrfs_fs_info *fs_info = trans->fs_info;
1801         struct btrfs_qgroup_extent_record *record;
1802         struct btrfs_delayed_ref_root *delayed_refs;
1803         int ret;
1804
1805         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)
1806             || bytenr == 0 || num_bytes == 0)
1807                 return 0;
1808         record = kzalloc(sizeof(*record), gfp_flag);
1809         if (!record)
1810                 return -ENOMEM;
1811
1812         delayed_refs = &trans->transaction->delayed_refs;
1813         record->bytenr = bytenr;
1814         record->num_bytes = num_bytes;
1815         record->old_roots = NULL;
1816
1817         spin_lock(&delayed_refs->lock);
1818         ret = btrfs_qgroup_trace_extent_nolock(fs_info, delayed_refs, record);
1819         spin_unlock(&delayed_refs->lock);
1820         if (ret > 0) {
1821                 kfree(record);
1822                 return 0;
1823         }
1824         return btrfs_qgroup_trace_extent_post(trans, record);
1825 }
1826
1827 int btrfs_qgroup_trace_leaf_items(struct btrfs_trans_handle *trans,
1828                                   struct extent_buffer *eb)
1829 {
1830         struct btrfs_fs_info *fs_info = trans->fs_info;
1831         int nr = btrfs_header_nritems(eb);
1832         int i, extent_type, ret;
1833         struct btrfs_key key;
1834         struct btrfs_file_extent_item *fi;
1835         u64 bytenr, num_bytes;
1836
1837         /* We can be called directly from walk_up_proc() */
1838         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
1839                 return 0;
1840
1841         for (i = 0; i < nr; i++) {
1842                 btrfs_item_key_to_cpu(eb, &key, i);
1843
1844                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1845                         continue;
1846
1847                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
1848                 /* filter out non qgroup-accountable extents  */
1849                 extent_type = btrfs_file_extent_type(eb, fi);
1850
1851                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
1852                         continue;
1853
1854                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
1855                 if (!bytenr)
1856                         continue;
1857
1858                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
1859
1860                 ret = btrfs_qgroup_trace_extent(trans, bytenr, num_bytes,
1861                                                 GFP_NOFS);
1862                 if (ret)
1863                         return ret;
1864         }
1865         cond_resched();
1866         return 0;
1867 }
1868
1869 /*
1870  * Walk up the tree from the bottom, freeing leaves and any interior
1871  * nodes which have had all slots visited. If a node (leaf or
1872  * interior) is freed, the node above it will have it's slot
1873  * incremented. The root node will never be freed.
1874  *
1875  * At the end of this function, we should have a path which has all
1876  * slots incremented to the next position for a search. If we need to
1877  * read a new node it will be NULL and the node above it will have the
1878  * correct slot selected for a later read.
1879  *
1880  * If we increment the root nodes slot counter past the number of
1881  * elements, 1 is returned to signal completion of the search.
1882  */
1883 static int adjust_slots_upwards(struct btrfs_path *path, int root_level)
1884 {
1885         int level = 0;
1886         int nr, slot;
1887         struct extent_buffer *eb;
1888
1889         if (root_level == 0)
1890                 return 1;
1891
1892         while (level <= root_level) {
1893                 eb = path->nodes[level];
1894                 nr = btrfs_header_nritems(eb);
1895                 path->slots[level]++;
1896                 slot = path->slots[level];
1897                 if (slot >= nr || level == 0) {
1898                         /*
1899                          * Don't free the root -  we will detect this
1900                          * condition after our loop and return a
1901                          * positive value for caller to stop walking the tree.
1902                          */
1903                         if (level != root_level) {
1904                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
1905                                 path->locks[level] = 0;
1906
1907                                 free_extent_buffer(eb);
1908                                 path->nodes[level] = NULL;
1909                                 path->slots[level] = 0;
1910                         }
1911                 } else {
1912                         /*
1913                          * We have a valid slot to walk back down
1914                          * from. Stop here so caller can process these
1915                          * new nodes.
1916                          */
1917                         break;
1918                 }
1919
1920                 level++;
1921         }
1922
1923         eb = path->nodes[root_level];
1924         if (path->slots[root_level] >= btrfs_header_nritems(eb))
1925                 return 1;
1926
1927         return 0;
1928 }
1929
1930 /*
1931  * Helper function to trace a subtree tree block swap.
1932  *
1933  * The swap will happen in highest tree block, but there may be a lot of
1934  * tree blocks involved.
1935  *
1936  * For example:
1937  *  OO = Old tree blocks
1938  *  NN = New tree blocks allocated during balance
1939  *
1940  *           File tree (257)                  Reloc tree for 257
1941  * L2              OO                                NN
1942  *               /    \                            /    \
1943  * L1          OO      OO (a)                    OO      NN (a)
1944  *            / \     / \                       / \     / \
1945  * L0       OO   OO OO   OO                   OO   OO NN   NN
1946  *                  (b)  (c)                          (b)  (c)
1947  *
1948  * When calling qgroup_trace_extent_swap(), we will pass:
1949  * @src_eb = OO(a)
1950  * @dst_path = [ nodes[1] = NN(a), nodes[0] = NN(c) ]
1951  * @dst_level = 0
1952  * @root_level = 1
1953  *
1954  * In that case, qgroup_trace_extent_swap() will search from OO(a) to
1955  * reach OO(c), then mark both OO(c) and NN(c) as qgroup dirty.
1956  *
1957  * The main work of qgroup_trace_extent_swap() can be split into 3 parts:
1958  *
1959  * 1) Tree search from @src_eb
1960  *    It should acts as a simplified btrfs_search_slot().
1961  *    The key for search can be extracted from @dst_path->nodes[dst_level]
1962  *    (first key).
1963  *
1964  * 2) Mark the final tree blocks in @src_path and @dst_path qgroup dirty
1965  *    NOTE: In above case, OO(a) and NN(a) won't be marked qgroup dirty.
1966  *    They should be marked during previous (@dst_level = 1) iteration.
1967  *
1968  * 3) Mark file extents in leaves dirty
1969  *    We don't have good way to pick out new file extents only.
1970  *    So we still follow the old method by scanning all file extents in
1971  *    the leave.
1972  *
1973  * This function can free us from keeping two paths, thus later we only need
1974  * to care about how to iterate all new tree blocks in reloc tree.
1975  */
1976 static int qgroup_trace_extent_swap(struct btrfs_trans_handle* trans,
1977                                     struct extent_buffer *src_eb,
1978                                     struct btrfs_path *dst_path,
1979                                     int dst_level, int root_level,
1980                                     bool trace_leaf)
1981 {
1982         struct btrfs_key key;
1983         struct btrfs_path *src_path;
1984         struct btrfs_fs_info *fs_info = trans->fs_info;
1985         u32 nodesize = fs_info->nodesize;
1986         int cur_level = root_level;
1987         int ret;
1988
1989         BUG_ON(dst_level > root_level);
1990         /* Level mismatch */
1991         if (btrfs_header_level(src_eb) != root_level)
1992                 return -EINVAL;
1993
1994         src_path = btrfs_alloc_path();
1995         if (!src_path) {
1996                 ret = -ENOMEM;
1997                 goto out;
1998         }
1999
2000         if (dst_level)
2001                 btrfs_node_key_to_cpu(dst_path->nodes[dst_level], &key, 0);
2002         else
2003                 btrfs_item_key_to_cpu(dst_path->nodes[dst_level], &key, 0);
2004
2005         /* For src_path */
2006         atomic_inc(&src_eb->refs);
2007         src_path->nodes[root_level] = src_eb;
2008         src_path->slots[root_level] = dst_path->slots[root_level];
2009         src_path->locks[root_level] = 0;
2010
2011         /* A simplified version of btrfs_search_slot() */
2012         while (cur_level >= dst_level) {
2013                 struct btrfs_key src_key;
2014                 struct btrfs_key dst_key;
2015
2016                 if (src_path->nodes[cur_level] == NULL) {
2017                         struct extent_buffer *eb;
2018                         int parent_slot;
2019
2020                         eb = src_path->nodes[cur_level + 1];
2021                         parent_slot = src_path->slots[cur_level + 1];
2022
2023                         eb = btrfs_read_node_slot(eb, parent_slot);
2024                         if (IS_ERR(eb)) {
2025                                 ret = PTR_ERR(eb);
2026                                 goto out;
2027                         }
2028
2029                         src_path->nodes[cur_level] = eb;
2030
2031                         btrfs_tree_read_lock(eb);
2032                         src_path->locks[cur_level] = BTRFS_READ_LOCK;
2033                 }
2034
2035                 src_path->slots[cur_level] = dst_path->slots[cur_level];
2036                 if (cur_level) {
2037                         btrfs_node_key_to_cpu(dst_path->nodes[cur_level],
2038                                         &dst_key, dst_path->slots[cur_level]);
2039                         btrfs_node_key_to_cpu(src_path->nodes[cur_level],
2040                                         &src_key, src_path->slots[cur_level]);
2041                 } else {
2042                         btrfs_item_key_to_cpu(dst_path->nodes[cur_level],
2043                                         &dst_key, dst_path->slots[cur_level]);
2044                         btrfs_item_key_to_cpu(src_path->nodes[cur_level],
2045                                         &src_key, src_path->slots[cur_level]);
2046                 }
2047                 /* Content mismatch, something went wrong */
2048                 if (btrfs_comp_cpu_keys(&dst_key, &src_key)) {
2049                         ret = -ENOENT;
2050                         goto out;
2051                 }
2052                 cur_level--;
2053         }
2054
2055         /*
2056          * Now both @dst_path and @src_path have been populated, record the tree
2057          * blocks for qgroup accounting.
2058          */
2059         ret = btrfs_qgroup_trace_extent(trans, src_path->nodes[dst_level]->start,
2060                         nodesize, GFP_NOFS);
2061         if (ret < 0)
2062                 goto out;
2063         ret = btrfs_qgroup_trace_extent(trans,
2064                         dst_path->nodes[dst_level]->start,
2065                         nodesize, GFP_NOFS);
2066         if (ret < 0)
2067                 goto out;
2068
2069         /* Record leaf file extents */
2070         if (dst_level == 0 && trace_leaf) {
2071                 ret = btrfs_qgroup_trace_leaf_items(trans, src_path->nodes[0]);
2072                 if (ret < 0)
2073                         goto out;
2074                 ret = btrfs_qgroup_trace_leaf_items(trans, dst_path->nodes[0]);
2075         }
2076 out:
2077         btrfs_free_path(src_path);
2078         return ret;
2079 }
2080
2081 /*
2082  * Helper function to do recursive generation-aware depth-first search, to
2083  * locate all new tree blocks in a subtree of reloc tree.
2084  *
2085  * E.g. (OO = Old tree blocks, NN = New tree blocks, whose gen == last_snapshot)
2086  *         reloc tree
2087  * L2         NN (a)
2088  *          /    \
2089  * L1    OO        NN (b)
2090  *      /  \      /  \
2091  * L0  OO  OO    OO  NN
2092  *               (c) (d)
2093  * If we pass:
2094  * @dst_path = [ nodes[1] = NN(b), nodes[0] = NULL ],
2095  * @cur_level = 1
2096  * @root_level = 1
2097  *
2098  * We will iterate through tree blocks NN(b), NN(d) and info qgroup to trace
2099  * above tree blocks along with their counter parts in file tree.
2100  * While during search, old tree blocks OO(c) will be skipped as tree block swap
2101  * won't affect OO(c).
2102  */
2103 static int qgroup_trace_new_subtree_blocks(struct btrfs_trans_handle* trans,
2104                                            struct extent_buffer *src_eb,
2105                                            struct btrfs_path *dst_path,
2106                                            int cur_level, int root_level,
2107                                            u64 last_snapshot, bool trace_leaf)
2108 {
2109         struct btrfs_fs_info *fs_info = trans->fs_info;
2110         struct extent_buffer *eb;
2111         bool need_cleanup = false;
2112         int ret = 0;
2113         int i;
2114
2115         /* Level sanity check */
2116         if (cur_level < 0 || cur_level >= BTRFS_MAX_LEVEL - 1 ||
2117             root_level < 0 || root_level >= BTRFS_MAX_LEVEL - 1 ||
2118             root_level < cur_level) {
2119                 btrfs_err_rl(fs_info,
2120                         "%s: bad levels, cur_level=%d root_level=%d",
2121                         __func__, cur_level, root_level);
2122                 return -EUCLEAN;
2123         }
2124
2125         /* Read the tree block if needed */
2126         if (dst_path->nodes[cur_level] == NULL) {
2127                 int parent_slot;
2128                 u64 child_gen;
2129
2130                 /*
2131                  * dst_path->nodes[root_level] must be initialized before
2132                  * calling this function.
2133                  */
2134                 if (cur_level == root_level) {
2135                         btrfs_err_rl(fs_info,
2136         "%s: dst_path->nodes[%d] not initialized, root_level=%d cur_level=%d",
2137                                 __func__, root_level, root_level, cur_level);
2138                         return -EUCLEAN;
2139                 }
2140
2141                 /*
2142                  * We need to get child blockptr/gen from parent before we can
2143                  * read it.
2144                   */
2145                 eb = dst_path->nodes[cur_level + 1];
2146                 parent_slot = dst_path->slots[cur_level + 1];
2147                 child_gen = btrfs_node_ptr_generation(eb, parent_slot);
2148
2149                 /* This node is old, no need to trace */
2150                 if (child_gen < last_snapshot)
2151                         goto out;
2152
2153                 eb = btrfs_read_node_slot(eb, parent_slot);
2154                 if (IS_ERR(eb)) {
2155                         ret = PTR_ERR(eb);
2156                         goto out;
2157                 }
2158
2159                 dst_path->nodes[cur_level] = eb;
2160                 dst_path->slots[cur_level] = 0;
2161
2162                 btrfs_tree_read_lock(eb);
2163                 dst_path->locks[cur_level] = BTRFS_READ_LOCK;
2164                 need_cleanup = true;
2165         }
2166
2167         /* Now record this tree block and its counter part for qgroups */
2168         ret = qgroup_trace_extent_swap(trans, src_eb, dst_path, cur_level,
2169                                        root_level, trace_leaf);
2170         if (ret < 0)
2171                 goto cleanup;
2172
2173         eb = dst_path->nodes[cur_level];
2174
2175         if (cur_level > 0) {
2176                 /* Iterate all child tree blocks */
2177                 for (i = 0; i < btrfs_header_nritems(eb); i++) {
2178                         /* Skip old tree blocks as they won't be swapped */
2179                         if (btrfs_node_ptr_generation(eb, i) < last_snapshot)
2180                                 continue;
2181                         dst_path->slots[cur_level] = i;
2182
2183                         /* Recursive call (at most 7 times) */
2184                         ret = qgroup_trace_new_subtree_blocks(trans, src_eb,
2185                                         dst_path, cur_level - 1, root_level,
2186                                         last_snapshot, trace_leaf);
2187                         if (ret < 0)
2188                                 goto cleanup;
2189                 }
2190         }
2191
2192 cleanup:
2193         if (need_cleanup) {
2194                 /* Clean up */
2195                 btrfs_tree_unlock_rw(dst_path->nodes[cur_level],
2196                                      dst_path->locks[cur_level]);
2197                 free_extent_buffer(dst_path->nodes[cur_level]);
2198                 dst_path->nodes[cur_level] = NULL;
2199                 dst_path->slots[cur_level] = 0;
2200                 dst_path->locks[cur_level] = 0;
2201         }
2202 out:
2203         return ret;
2204 }
2205
2206 static int qgroup_trace_subtree_swap(struct btrfs_trans_handle *trans,
2207                                 struct extent_buffer *src_eb,
2208                                 struct extent_buffer *dst_eb,
2209                                 u64 last_snapshot, bool trace_leaf)
2210 {
2211         struct btrfs_fs_info *fs_info = trans->fs_info;
2212         struct btrfs_path *dst_path = NULL;
2213         int level;
2214         int ret;
2215
2216         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2217                 return 0;
2218
2219         /* Wrong parameter order */
2220         if (btrfs_header_generation(src_eb) > btrfs_header_generation(dst_eb)) {
2221                 btrfs_err_rl(fs_info,
2222                 "%s: bad parameter order, src_gen=%llu dst_gen=%llu", __func__,
2223                              btrfs_header_generation(src_eb),
2224                              btrfs_header_generation(dst_eb));
2225                 return -EUCLEAN;
2226         }
2227
2228         if (!extent_buffer_uptodate(src_eb) || !extent_buffer_uptodate(dst_eb)) {
2229                 ret = -EIO;
2230                 goto out;
2231         }
2232
2233         level = btrfs_header_level(dst_eb);
2234         dst_path = btrfs_alloc_path();
2235         if (!dst_path) {
2236                 ret = -ENOMEM;
2237                 goto out;
2238         }
2239         /* For dst_path */
2240         atomic_inc(&dst_eb->refs);
2241         dst_path->nodes[level] = dst_eb;
2242         dst_path->slots[level] = 0;
2243         dst_path->locks[level] = 0;
2244
2245         /* Do the generation aware breadth-first search */
2246         ret = qgroup_trace_new_subtree_blocks(trans, src_eb, dst_path, level,
2247                                               level, last_snapshot, trace_leaf);
2248         if (ret < 0)
2249                 goto out;
2250         ret = 0;
2251
2252 out:
2253         btrfs_free_path(dst_path);
2254         if (ret < 0)
2255                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2256         return ret;
2257 }
2258
2259 int btrfs_qgroup_trace_subtree(struct btrfs_trans_handle *trans,
2260                                struct extent_buffer *root_eb,
2261                                u64 root_gen, int root_level)
2262 {
2263         struct btrfs_fs_info *fs_info = trans->fs_info;
2264         int ret = 0;
2265         int level;
2266         struct extent_buffer *eb = root_eb;
2267         struct btrfs_path *path = NULL;
2268
2269         BUG_ON(root_level < 0 || root_level >= BTRFS_MAX_LEVEL);
2270         BUG_ON(root_eb == NULL);
2271
2272         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2273                 return 0;
2274
2275         if (!extent_buffer_uptodate(root_eb)) {
2276                 ret = btrfs_read_buffer(root_eb, root_gen, root_level, NULL);
2277                 if (ret)
2278                         goto out;
2279         }
2280
2281         if (root_level == 0) {
2282                 ret = btrfs_qgroup_trace_leaf_items(trans, root_eb);
2283                 goto out;
2284         }
2285
2286         path = btrfs_alloc_path();
2287         if (!path)
2288                 return -ENOMEM;
2289
2290         /*
2291          * Walk down the tree.  Missing extent blocks are filled in as
2292          * we go. Metadata is accounted every time we read a new
2293          * extent block.
2294          *
2295          * When we reach a leaf, we account for file extent items in it,
2296          * walk back up the tree (adjusting slot pointers as we go)
2297          * and restart the search process.
2298          */
2299         atomic_inc(&root_eb->refs);     /* For path */
2300         path->nodes[root_level] = root_eb;
2301         path->slots[root_level] = 0;
2302         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
2303 walk_down:
2304         level = root_level;
2305         while (level >= 0) {
2306                 if (path->nodes[level] == NULL) {
2307                         int parent_slot;
2308                         u64 child_bytenr;
2309
2310                         /*
2311                          * We need to get child blockptr from parent before we
2312                          * can read it.
2313                           */
2314                         eb = path->nodes[level + 1];
2315                         parent_slot = path->slots[level + 1];
2316                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
2317
2318                         eb = btrfs_read_node_slot(eb, parent_slot);
2319                         if (IS_ERR(eb)) {
2320                                 ret = PTR_ERR(eb);
2321                                 goto out;
2322                         }
2323
2324                         path->nodes[level] = eb;
2325                         path->slots[level] = 0;
2326
2327                         btrfs_tree_read_lock(eb);
2328                         path->locks[level] = BTRFS_READ_LOCK;
2329
2330                         ret = btrfs_qgroup_trace_extent(trans, child_bytenr,
2331                                                         fs_info->nodesize,
2332                                                         GFP_NOFS);
2333                         if (ret)
2334                                 goto out;
2335                 }
2336
2337                 if (level == 0) {
2338                         ret = btrfs_qgroup_trace_leaf_items(trans,
2339                                                             path->nodes[level]);
2340                         if (ret)
2341                                 goto out;
2342
2343                         /* Nonzero return here means we completed our search */
2344                         ret = adjust_slots_upwards(path, root_level);
2345                         if (ret)
2346                                 break;
2347
2348                         /* Restart search with new slots */
2349                         goto walk_down;
2350                 }
2351
2352                 level--;
2353         }
2354
2355         ret = 0;
2356 out:
2357         btrfs_free_path(path);
2358
2359         return ret;
2360 }
2361
2362 #define UPDATE_NEW      0
2363 #define UPDATE_OLD      1
2364 /*
2365  * Walk all of the roots that points to the bytenr and adjust their refcnts.
2366  */
2367 static int qgroup_update_refcnt(struct btrfs_fs_info *fs_info,
2368                                 struct ulist *roots, struct ulist *tmp,
2369                                 struct ulist *qgroups, u64 seq, int update_old)
2370 {
2371         struct ulist_node *unode;
2372         struct ulist_iterator uiter;
2373         struct ulist_node *tmp_unode;
2374         struct ulist_iterator tmp_uiter;
2375         struct btrfs_qgroup *qg;
2376         int ret = 0;
2377
2378         if (!roots)
2379                 return 0;
2380         ULIST_ITER_INIT(&uiter);
2381         while ((unode = ulist_next(roots, &uiter))) {
2382                 qg = find_qgroup_rb(fs_info, unode->val);
2383                 if (!qg)
2384                         continue;
2385
2386                 ulist_reinit(tmp);
2387                 ret = ulist_add(qgroups, qg->qgroupid, qgroup_to_aux(qg),
2388                                 GFP_ATOMIC);
2389                 if (ret < 0)
2390                         return ret;
2391                 ret = ulist_add(tmp, qg->qgroupid, qgroup_to_aux(qg), GFP_ATOMIC);
2392                 if (ret < 0)
2393                         return ret;
2394                 ULIST_ITER_INIT(&tmp_uiter);
2395                 while ((tmp_unode = ulist_next(tmp, &tmp_uiter))) {
2396                         struct btrfs_qgroup_list *glist;
2397
2398                         qg = unode_aux_to_qgroup(tmp_unode);
2399                         if (update_old)
2400                                 btrfs_qgroup_update_old_refcnt(qg, seq, 1);
2401                         else
2402                                 btrfs_qgroup_update_new_refcnt(qg, seq, 1);
2403                         list_for_each_entry(glist, &qg->groups, next_group) {
2404                                 ret = ulist_add(qgroups, glist->group->qgroupid,
2405                                                 qgroup_to_aux(glist->group),
2406                                                 GFP_ATOMIC);
2407                                 if (ret < 0)
2408                                         return ret;
2409                                 ret = ulist_add(tmp, glist->group->qgroupid,
2410                                                 qgroup_to_aux(glist->group),
2411                                                 GFP_ATOMIC);
2412                                 if (ret < 0)
2413                                         return ret;
2414                         }
2415                 }
2416         }
2417         return 0;
2418 }
2419
2420 /*
2421  * Update qgroup rfer/excl counters.
2422  * Rfer update is easy, codes can explain themselves.
2423  *
2424  * Excl update is tricky, the update is split into 2 parts.
2425  * Part 1: Possible exclusive <-> sharing detect:
2426  *      |       A       |       !A      |
2427  *  -------------------------------------
2428  *  B   |       *       |       -       |
2429  *  -------------------------------------
2430  *  !B  |       +       |       **      |
2431  *  -------------------------------------
2432  *
2433  * Conditions:
2434  * A:   cur_old_roots < nr_old_roots    (not exclusive before)
2435  * !A:  cur_old_roots == nr_old_roots   (possible exclusive before)
2436  * B:   cur_new_roots < nr_new_roots    (not exclusive now)
2437  * !B:  cur_new_roots == nr_new_roots   (possible exclusive now)
2438  *
2439  * Results:
2440  * +: Possible sharing -> exclusive     -: Possible exclusive -> sharing
2441  * *: Definitely not changed.           **: Possible unchanged.
2442  *
2443  * For !A and !B condition, the exception is cur_old/new_roots == 0 case.
2444  *
2445  * To make the logic clear, we first use condition A and B to split
2446  * combination into 4 results.
2447  *
2448  * Then, for result "+" and "-", check old/new_roots == 0 case, as in them
2449  * only on variant maybe 0.
2450  *
2451  * Lastly, check result **, since there are 2 variants maybe 0, split them
2452  * again(2x2).
2453  * But this time we don't need to consider other things, the codes and logic
2454  * is easy to understand now.
2455  */
2456 static int qgroup_update_counters(struct btrfs_fs_info *fs_info,
2457                                   struct ulist *qgroups,
2458                                   u64 nr_old_roots,
2459                                   u64 nr_new_roots,
2460                                   u64 num_bytes, u64 seq)
2461 {
2462         struct ulist_node *unode;
2463         struct ulist_iterator uiter;
2464         struct btrfs_qgroup *qg;
2465         u64 cur_new_count, cur_old_count;
2466
2467         ULIST_ITER_INIT(&uiter);
2468         while ((unode = ulist_next(qgroups, &uiter))) {
2469                 bool dirty = false;
2470
2471                 qg = unode_aux_to_qgroup(unode);
2472                 cur_old_count = btrfs_qgroup_get_old_refcnt(qg, seq);
2473                 cur_new_count = btrfs_qgroup_get_new_refcnt(qg, seq);
2474
2475                 trace_qgroup_update_counters(fs_info, qg, cur_old_count,
2476                                              cur_new_count);
2477
2478                 /* Rfer update part */
2479                 if (cur_old_count == 0 && cur_new_count > 0) {
2480                         qg->rfer += num_bytes;
2481                         qg->rfer_cmpr += num_bytes;
2482                         dirty = true;
2483                 }
2484                 if (cur_old_count > 0 && cur_new_count == 0) {
2485                         qg->rfer -= num_bytes;
2486                         qg->rfer_cmpr -= num_bytes;
2487                         dirty = true;
2488                 }
2489
2490                 /* Excl update part */
2491                 /* Exclusive/none -> shared case */
2492                 if (cur_old_count == nr_old_roots &&
2493                     cur_new_count < nr_new_roots) {
2494                         /* Exclusive -> shared */
2495                         if (cur_old_count != 0) {
2496                                 qg->excl -= num_bytes;
2497                                 qg->excl_cmpr -= num_bytes;
2498                                 dirty = true;
2499                         }
2500                 }
2501
2502                 /* Shared -> exclusive/none case */
2503                 if (cur_old_count < nr_old_roots &&
2504                     cur_new_count == nr_new_roots) {
2505                         /* Shared->exclusive */
2506                         if (cur_new_count != 0) {
2507                                 qg->excl += num_bytes;
2508                                 qg->excl_cmpr += num_bytes;
2509                                 dirty = true;
2510                         }
2511                 }
2512
2513                 /* Exclusive/none -> exclusive/none case */
2514                 if (cur_old_count == nr_old_roots &&
2515                     cur_new_count == nr_new_roots) {
2516                         if (cur_old_count == 0) {
2517                                 /* None -> exclusive/none */
2518
2519                                 if (cur_new_count != 0) {
2520                                         /* None -> exclusive */
2521                                         qg->excl += num_bytes;
2522                                         qg->excl_cmpr += num_bytes;
2523                                         dirty = true;
2524                                 }
2525                                 /* None -> none, nothing changed */
2526                         } else {
2527                                 /* Exclusive -> exclusive/none */
2528
2529                                 if (cur_new_count == 0) {
2530                                         /* Exclusive -> none */
2531                                         qg->excl -= num_bytes;
2532                                         qg->excl_cmpr -= num_bytes;
2533                                         dirty = true;
2534                                 }
2535                                 /* Exclusive -> exclusive, nothing changed */
2536                         }
2537                 }
2538
2539                 if (dirty)
2540                         qgroup_dirty(fs_info, qg);
2541         }
2542         return 0;
2543 }
2544
2545 /*
2546  * Check if the @roots potentially is a list of fs tree roots
2547  *
2548  * Return 0 for definitely not a fs/subvol tree roots ulist
2549  * Return 1 for possible fs/subvol tree roots in the list (considering an empty
2550  *          one as well)
2551  */
2552 static int maybe_fs_roots(struct ulist *roots)
2553 {
2554         struct ulist_node *unode;
2555         struct ulist_iterator uiter;
2556
2557         /* Empty one, still possible for fs roots */
2558         if (!roots || roots->nnodes == 0)
2559                 return 1;
2560
2561         ULIST_ITER_INIT(&uiter);
2562         unode = ulist_next(roots, &uiter);
2563         if (!unode)
2564                 return 1;
2565
2566         /*
2567          * If it contains fs tree roots, then it must belong to fs/subvol
2568          * trees.
2569          * If it contains a non-fs tree, it won't be shared with fs/subvol trees.
2570          */
2571         return is_fstree(unode->val);
2572 }
2573
2574 int btrfs_qgroup_account_extent(struct btrfs_trans_handle *trans, u64 bytenr,
2575                                 u64 num_bytes, struct ulist *old_roots,
2576                                 struct ulist *new_roots)
2577 {
2578         struct btrfs_fs_info *fs_info = trans->fs_info;
2579         struct ulist *qgroups = NULL;
2580         struct ulist *tmp = NULL;
2581         u64 seq;
2582         u64 nr_new_roots = 0;
2583         u64 nr_old_roots = 0;
2584         int ret = 0;
2585
2586         /*
2587          * If quotas get disabled meanwhile, the resources need to be freed and
2588          * we can't just exit here.
2589          */
2590         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2591                 goto out_free;
2592
2593         if (new_roots) {
2594                 if (!maybe_fs_roots(new_roots))
2595                         goto out_free;
2596                 nr_new_roots = new_roots->nnodes;
2597         }
2598         if (old_roots) {
2599                 if (!maybe_fs_roots(old_roots))
2600                         goto out_free;
2601                 nr_old_roots = old_roots->nnodes;
2602         }
2603
2604         /* Quick exit, either not fs tree roots, or won't affect any qgroup */
2605         if (nr_old_roots == 0 && nr_new_roots == 0)
2606                 goto out_free;
2607
2608         BUG_ON(!fs_info->quota_root);
2609
2610         trace_btrfs_qgroup_account_extent(fs_info, trans->transid, bytenr,
2611                                         num_bytes, nr_old_roots, nr_new_roots);
2612
2613         qgroups = ulist_alloc(GFP_NOFS);
2614         if (!qgroups) {
2615                 ret = -ENOMEM;
2616                 goto out_free;
2617         }
2618         tmp = ulist_alloc(GFP_NOFS);
2619         if (!tmp) {
2620                 ret = -ENOMEM;
2621                 goto out_free;
2622         }
2623
2624         mutex_lock(&fs_info->qgroup_rescan_lock);
2625         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
2626                 if (fs_info->qgroup_rescan_progress.objectid <= bytenr) {
2627                         mutex_unlock(&fs_info->qgroup_rescan_lock);
2628                         ret = 0;
2629                         goto out_free;
2630                 }
2631         }
2632         mutex_unlock(&fs_info->qgroup_rescan_lock);
2633
2634         spin_lock(&fs_info->qgroup_lock);
2635         seq = fs_info->qgroup_seq;
2636
2637         /* Update old refcnts using old_roots */
2638         ret = qgroup_update_refcnt(fs_info, old_roots, tmp, qgroups, seq,
2639                                    UPDATE_OLD);
2640         if (ret < 0)
2641                 goto out;
2642
2643         /* Update new refcnts using new_roots */
2644         ret = qgroup_update_refcnt(fs_info, new_roots, tmp, qgroups, seq,
2645                                    UPDATE_NEW);
2646         if (ret < 0)
2647                 goto out;
2648
2649         qgroup_update_counters(fs_info, qgroups, nr_old_roots, nr_new_roots,
2650                                num_bytes, seq);
2651
2652         /*
2653          * Bump qgroup_seq to avoid seq overlap
2654          */
2655         fs_info->qgroup_seq += max(nr_old_roots, nr_new_roots) + 1;
2656 out:
2657         spin_unlock(&fs_info->qgroup_lock);
2658 out_free:
2659         ulist_free(tmp);
2660         ulist_free(qgroups);
2661         ulist_free(old_roots);
2662         ulist_free(new_roots);
2663         return ret;
2664 }
2665
2666 int btrfs_qgroup_account_extents(struct btrfs_trans_handle *trans)
2667 {
2668         struct btrfs_fs_info *fs_info = trans->fs_info;
2669         struct btrfs_qgroup_extent_record *record;
2670         struct btrfs_delayed_ref_root *delayed_refs;
2671         struct ulist *new_roots = NULL;
2672         struct rb_node *node;
2673         u64 num_dirty_extents = 0;
2674         u64 qgroup_to_skip;
2675         int ret = 0;
2676
2677         delayed_refs = &trans->transaction->delayed_refs;
2678         qgroup_to_skip = delayed_refs->qgroup_to_skip;
2679         while ((node = rb_first(&delayed_refs->dirty_extent_root))) {
2680                 record = rb_entry(node, struct btrfs_qgroup_extent_record,
2681                                   node);
2682
2683                 num_dirty_extents++;
2684                 trace_btrfs_qgroup_account_extents(fs_info, record);
2685
2686                 if (!ret) {
2687                         /*
2688                          * Old roots should be searched when inserting qgroup
2689                          * extent record
2690                          */
2691                         if (WARN_ON(!record->old_roots)) {
2692                                 /* Search commit root to find old_roots */
2693                                 ret = btrfs_find_all_roots(NULL, fs_info,
2694                                                 record->bytenr, 0,
2695                                                 &record->old_roots, false);
2696                                 if (ret < 0)
2697                                         goto cleanup;
2698                         }
2699
2700                         /* Free the reserved data space */
2701                         btrfs_qgroup_free_refroot(fs_info,
2702                                         record->data_rsv_refroot,
2703                                         record->data_rsv,
2704                                         BTRFS_QGROUP_RSV_DATA);
2705                         /*
2706                          * Use BTRFS_SEQ_LAST as time_seq to do special search,
2707                          * which doesn't lock tree or delayed_refs and search
2708                          * current root. It's safe inside commit_transaction().
2709                          */
2710                         ret = btrfs_find_all_roots(trans, fs_info,
2711                            record->bytenr, BTRFS_SEQ_LAST, &new_roots, false);
2712                         if (ret < 0)
2713                                 goto cleanup;
2714                         if (qgroup_to_skip) {
2715                                 ulist_del(new_roots, qgroup_to_skip, 0);
2716                                 ulist_del(record->old_roots, qgroup_to_skip,
2717                                           0);
2718                         }
2719                         ret = btrfs_qgroup_account_extent(trans, record->bytenr,
2720                                                           record->num_bytes,
2721                                                           record->old_roots,
2722                                                           new_roots);
2723                         record->old_roots = NULL;
2724                         new_roots = NULL;
2725                 }
2726 cleanup:
2727                 ulist_free(record->old_roots);
2728                 ulist_free(new_roots);
2729                 new_roots = NULL;
2730                 rb_erase(node, &delayed_refs->dirty_extent_root);
2731                 kfree(record);
2732
2733         }
2734         trace_qgroup_num_dirty_extents(fs_info, trans->transid,
2735                                        num_dirty_extents);
2736         return ret;
2737 }
2738
2739 /*
2740  * called from commit_transaction. Writes all changed qgroups to disk.
2741  */
2742 int btrfs_run_qgroups(struct btrfs_trans_handle *trans)
2743 {
2744         struct btrfs_fs_info *fs_info = trans->fs_info;
2745         int ret = 0;
2746
2747         if (!fs_info->quota_root)
2748                 return ret;
2749
2750         spin_lock(&fs_info->qgroup_lock);
2751         while (!list_empty(&fs_info->dirty_qgroups)) {
2752                 struct btrfs_qgroup *qgroup;
2753                 qgroup = list_first_entry(&fs_info->dirty_qgroups,
2754                                           struct btrfs_qgroup, dirty);
2755                 list_del_init(&qgroup->dirty);
2756                 spin_unlock(&fs_info->qgroup_lock);
2757                 ret = update_qgroup_info_item(trans, qgroup);
2758                 if (ret)
2759                         fs_info->qgroup_flags |=
2760                                         BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2761                 ret = update_qgroup_limit_item(trans, qgroup);
2762                 if (ret)
2763                         fs_info->qgroup_flags |=
2764                                         BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2765                 spin_lock(&fs_info->qgroup_lock);
2766         }
2767         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2768                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_ON;
2769         else
2770                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_ON;
2771         spin_unlock(&fs_info->qgroup_lock);
2772
2773         ret = update_qgroup_status_item(trans);
2774         if (ret)
2775                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2776
2777         return ret;
2778 }
2779
2780 /*
2781  * Copy the accounting information between qgroups. This is necessary
2782  * when a snapshot or a subvolume is created. Throwing an error will
2783  * cause a transaction abort so we take extra care here to only error
2784  * when a readonly fs is a reasonable outcome.
2785  */
2786 int btrfs_qgroup_inherit(struct btrfs_trans_handle *trans, u64 srcid,
2787                          u64 objectid, struct btrfs_qgroup_inherit *inherit)
2788 {
2789         int ret = 0;
2790         int i;
2791         u64 *i_qgroups;
2792         bool committing = false;
2793         struct btrfs_fs_info *fs_info = trans->fs_info;
2794         struct btrfs_root *quota_root;
2795         struct btrfs_qgroup *srcgroup;
2796         struct btrfs_qgroup *dstgroup;
2797         bool need_rescan = false;
2798         u32 level_size = 0;
2799         u64 nums;
2800
2801         /*
2802          * There are only two callers of this function.
2803          *
2804          * One in create_subvol() in the ioctl context, which needs to hold
2805          * the qgroup_ioctl_lock.
2806          *
2807          * The other one in create_pending_snapshot() where no other qgroup
2808          * code can modify the fs as they all need to either start a new trans
2809          * or hold a trans handler, thus we don't need to hold
2810          * qgroup_ioctl_lock.
2811          * This would avoid long and complex lock chain and make lockdep happy.
2812          */
2813         spin_lock(&fs_info->trans_lock);
2814         if (trans->transaction->state == TRANS_STATE_COMMIT_DOING)
2815                 committing = true;
2816         spin_unlock(&fs_info->trans_lock);
2817
2818         if (!committing)
2819                 mutex_lock(&fs_info->qgroup_ioctl_lock);
2820         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2821                 goto out;
2822
2823         quota_root = fs_info->quota_root;
2824         if (!quota_root) {
2825                 ret = -EINVAL;
2826                 goto out;
2827         }
2828
2829         if (inherit) {
2830                 i_qgroups = (u64 *)(inherit + 1);
2831                 nums = inherit->num_qgroups + 2 * inherit->num_ref_copies +
2832                        2 * inherit->num_excl_copies;
2833                 for (i = 0; i < nums; ++i) {
2834                         srcgroup = find_qgroup_rb(fs_info, *i_qgroups);
2835
2836                         /*
2837                          * Zero out invalid groups so we can ignore
2838                          * them later.
2839                          */
2840                         if (!srcgroup ||
2841                             ((srcgroup->qgroupid >> 48) <= (objectid >> 48)))
2842                                 *i_qgroups = 0ULL;
2843
2844                         ++i_qgroups;
2845                 }
2846         }
2847
2848         /*
2849          * create a tracking group for the subvol itself
2850          */
2851         ret = add_qgroup_item(trans, quota_root, objectid);
2852         if (ret)
2853                 goto out;
2854
2855         /*
2856          * add qgroup to all inherited groups
2857          */
2858         if (inherit) {
2859                 i_qgroups = (u64 *)(inherit + 1);
2860                 for (i = 0; i < inherit->num_qgroups; ++i, ++i_qgroups) {
2861                         if (*i_qgroups == 0)
2862                                 continue;
2863                         ret = add_qgroup_relation_item(trans, objectid,
2864                                                        *i_qgroups);
2865                         if (ret && ret != -EEXIST)
2866                                 goto out;
2867                         ret = add_qgroup_relation_item(trans, *i_qgroups,
2868                                                        objectid);
2869                         if (ret && ret != -EEXIST)
2870                                 goto out;
2871                 }
2872                 ret = 0;
2873         }
2874
2875
2876         spin_lock(&fs_info->qgroup_lock);
2877
2878         dstgroup = add_qgroup_rb(fs_info, objectid);
2879         if (IS_ERR(dstgroup)) {
2880                 ret = PTR_ERR(dstgroup);
2881                 goto unlock;
2882         }
2883
2884         if (inherit && inherit->flags & BTRFS_QGROUP_INHERIT_SET_LIMITS) {
2885                 dstgroup->lim_flags = inherit->lim.flags;
2886                 dstgroup->max_rfer = inherit->lim.max_rfer;
2887                 dstgroup->max_excl = inherit->lim.max_excl;
2888                 dstgroup->rsv_rfer = inherit->lim.rsv_rfer;
2889                 dstgroup->rsv_excl = inherit->lim.rsv_excl;
2890
2891                 ret = update_qgroup_limit_item(trans, dstgroup);
2892                 if (ret) {
2893                         fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2894                         btrfs_info(fs_info,
2895                                    "unable to update quota limit for %llu",
2896                                    dstgroup->qgroupid);
2897                         goto unlock;
2898                 }
2899         }
2900
2901         if (srcid) {
2902                 srcgroup = find_qgroup_rb(fs_info, srcid);
2903                 if (!srcgroup)
2904                         goto unlock;
2905
2906                 /*
2907                  * We call inherit after we clone the root in order to make sure
2908                  * our counts don't go crazy, so at this point the only
2909                  * difference between the two roots should be the root node.
2910                  */
2911                 level_size = fs_info->nodesize;
2912                 dstgroup->rfer = srcgroup->rfer;
2913                 dstgroup->rfer_cmpr = srcgroup->rfer_cmpr;
2914                 dstgroup->excl = level_size;
2915                 dstgroup->excl_cmpr = level_size;
2916                 srcgroup->excl = level_size;
2917                 srcgroup->excl_cmpr = level_size;
2918
2919                 /* inherit the limit info */
2920                 dstgroup->lim_flags = srcgroup->lim_flags;
2921                 dstgroup->max_rfer = srcgroup->max_rfer;
2922                 dstgroup->max_excl = srcgroup->max_excl;
2923                 dstgroup->rsv_rfer = srcgroup->rsv_rfer;
2924                 dstgroup->rsv_excl = srcgroup->rsv_excl;
2925
2926                 qgroup_dirty(fs_info, dstgroup);
2927                 qgroup_dirty(fs_info, srcgroup);
2928         }
2929
2930         if (!inherit)
2931                 goto unlock;
2932
2933         i_qgroups = (u64 *)(inherit + 1);
2934         for (i = 0; i < inherit->num_qgroups; ++i) {
2935                 if (*i_qgroups) {
2936                         ret = add_relation_rb(fs_info, objectid, *i_qgroups);
2937                         if (ret)
2938                                 goto unlock;
2939                 }
2940                 ++i_qgroups;
2941
2942                 /*
2943                  * If we're doing a snapshot, and adding the snapshot to a new
2944                  * qgroup, the numbers are guaranteed to be incorrect.
2945                  */
2946                 if (srcid)
2947                         need_rescan = true;
2948         }
2949
2950         for (i = 0; i <  inherit->num_ref_copies; ++i, i_qgroups += 2) {
2951                 struct btrfs_qgroup *src;
2952                 struct btrfs_qgroup *dst;
2953
2954                 if (!i_qgroups[0] || !i_qgroups[1])
2955                         continue;
2956
2957                 src = find_qgroup_rb(fs_info, i_qgroups[0]);
2958                 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
2959
2960                 if (!src || !dst) {
2961                         ret = -EINVAL;
2962                         goto unlock;
2963                 }
2964
2965                 dst->rfer = src->rfer - level_size;
2966                 dst->rfer_cmpr = src->rfer_cmpr - level_size;
2967
2968                 /* Manually tweaking numbers certainly needs a rescan */
2969                 need_rescan = true;
2970         }
2971         for (i = 0; i <  inherit->num_excl_copies; ++i, i_qgroups += 2) {
2972                 struct btrfs_qgroup *src;
2973                 struct btrfs_qgroup *dst;
2974
2975                 if (!i_qgroups[0] || !i_qgroups[1])
2976                         continue;
2977
2978                 src = find_qgroup_rb(fs_info, i_qgroups[0]);
2979                 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
2980
2981                 if (!src || !dst) {
2982                         ret = -EINVAL;
2983                         goto unlock;
2984                 }
2985
2986                 dst->excl = src->excl + level_size;
2987                 dst->excl_cmpr = src->excl_cmpr + level_size;
2988                 need_rescan = true;
2989         }
2990
2991 unlock:
2992         spin_unlock(&fs_info->qgroup_lock);
2993         if (!ret)
2994                 ret = btrfs_sysfs_add_one_qgroup(fs_info, dstgroup);
2995 out:
2996         if (!committing)
2997                 mutex_unlock(&fs_info->qgroup_ioctl_lock);
2998         if (need_rescan)
2999                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
3000         return ret;
3001 }
3002
3003 static bool qgroup_check_limits(const struct btrfs_qgroup *qg, u64 num_bytes)
3004 {
3005         if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_RFER) &&
3006             qgroup_rsv_total(qg) + (s64)qg->rfer + num_bytes > qg->max_rfer)
3007                 return false;
3008
3009         if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) &&
3010             qgroup_rsv_total(qg) + (s64)qg->excl + num_bytes > qg->max_excl)
3011                 return false;
3012
3013         return true;
3014 }
3015
3016 static int qgroup_reserve(struct btrfs_root *root, u64 num_bytes, bool enforce,
3017                           enum btrfs_qgroup_rsv_type type)
3018 {
3019         struct btrfs_qgroup *qgroup;
3020         struct btrfs_fs_info *fs_info = root->fs_info;
3021         u64 ref_root = root->root_key.objectid;
3022         int ret = 0;
3023         struct ulist_node *unode;
3024         struct ulist_iterator uiter;
3025
3026         if (!is_fstree(ref_root))
3027                 return 0;
3028
3029         if (num_bytes == 0)
3030                 return 0;
3031
3032         if (test_bit(BTRFS_FS_QUOTA_OVERRIDE, &fs_info->flags) &&
3033             capable(CAP_SYS_RESOURCE))
3034                 enforce = false;
3035
3036         spin_lock(&fs_info->qgroup_lock);
3037         if (!fs_info->quota_root)
3038                 goto out;
3039
3040         qgroup = find_qgroup_rb(fs_info, ref_root);
3041         if (!qgroup)
3042                 goto out;
3043
3044         /*
3045          * in a first step, we check all affected qgroups if any limits would
3046          * be exceeded
3047          */
3048         ulist_reinit(fs_info->qgroup_ulist);
3049         ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
3050                         qgroup_to_aux(qgroup), GFP_ATOMIC);
3051         if (ret < 0)
3052                 goto out;
3053         ULIST_ITER_INIT(&uiter);
3054         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
3055                 struct btrfs_qgroup *qg;
3056                 struct btrfs_qgroup_list *glist;
3057
3058                 qg = unode_aux_to_qgroup(unode);
3059
3060                 if (enforce && !qgroup_check_limits(qg, num_bytes)) {
3061                         ret = -EDQUOT;
3062                         goto out;
3063                 }
3064
3065                 list_for_each_entry(glist, &qg->groups, next_group) {
3066                         ret = ulist_add(fs_info->qgroup_ulist,
3067                                         glist->group->qgroupid,
3068                                         qgroup_to_aux(glist->group), GFP_ATOMIC);
3069                         if (ret < 0)
3070                                 goto out;
3071                 }
3072         }
3073         ret = 0;
3074         /*
3075          * no limits exceeded, now record the reservation into all qgroups
3076          */
3077         ULIST_ITER_INIT(&uiter);
3078         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
3079                 struct btrfs_qgroup *qg;
3080
3081                 qg = unode_aux_to_qgroup(unode);
3082
3083                 qgroup_rsv_add(fs_info, qg, num_bytes, type);
3084         }
3085
3086 out:
3087         spin_unlock(&fs_info->qgroup_lock);
3088         return ret;
3089 }
3090
3091 /*
3092  * Free @num_bytes of reserved space with @type for qgroup.  (Normally level 0
3093  * qgroup).
3094  *
3095  * Will handle all higher level qgroup too.
3096  *
3097  * NOTE: If @num_bytes is (u64)-1, this means to free all bytes of this qgroup.
3098  * This special case is only used for META_PERTRANS type.
3099  */
3100 void btrfs_qgroup_free_refroot(struct btrfs_fs_info *fs_info,
3101                                u64 ref_root, u64 num_bytes,
3102                                enum btrfs_qgroup_rsv_type type)
3103 {
3104         struct btrfs_qgroup *qgroup;
3105         struct ulist_node *unode;
3106         struct ulist_iterator uiter;
3107         int ret = 0;
3108
3109         if (!is_fstree(ref_root))
3110                 return;
3111
3112         if (num_bytes == 0)
3113                 return;
3114
3115         if (num_bytes == (u64)-1 && type != BTRFS_QGROUP_RSV_META_PERTRANS) {
3116                 WARN(1, "%s: Invalid type to free", __func__);
3117                 return;
3118         }
3119         spin_lock(&fs_info->qgroup_lock);
3120
3121         if (!fs_info->quota_root)
3122                 goto out;
3123
3124         qgroup = find_qgroup_rb(fs_info, ref_root);
3125         if (!qgroup)
3126                 goto out;
3127
3128         if (num_bytes == (u64)-1)
3129                 /*
3130                  * We're freeing all pertrans rsv, get reserved value from
3131                  * level 0 qgroup as real num_bytes to free.
3132                  */
3133                 num_bytes = qgroup->rsv.values[type];
3134
3135         ulist_reinit(fs_info->qgroup_ulist);
3136         ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
3137                         qgroup_to_aux(qgroup), GFP_ATOMIC);
3138         if (ret < 0)
3139                 goto out;
3140         ULIST_ITER_INIT(&uiter);
3141         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
3142                 struct btrfs_qgroup *qg;
3143                 struct btrfs_qgroup_list *glist;
3144
3145                 qg = unode_aux_to_qgroup(unode);
3146
3147                 qgroup_rsv_release(fs_info, qg, num_bytes, type);
3148
3149                 list_for_each_entry(glist, &qg->groups, next_group) {
3150                         ret = ulist_add(fs_info->qgroup_ulist,
3151                                         glist->group->qgroupid,
3152                                         qgroup_to_aux(glist->group), GFP_ATOMIC);
3153                         if (ret < 0)
3154                                 goto out;
3155                 }
3156         }
3157
3158 out:
3159         spin_unlock(&fs_info->qgroup_lock);
3160 }
3161
3162 /*
3163  * Check if the leaf is the last leaf. Which means all node pointers
3164  * are at their last position.
3165  */
3166 static bool is_last_leaf(struct btrfs_path *path)
3167 {
3168         int i;
3169
3170         for (i = 1; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
3171                 if (path->slots[i] != btrfs_header_nritems(path->nodes[i]) - 1)
3172                         return false;
3173         }
3174         return true;
3175 }
3176
3177 /*
3178  * returns < 0 on error, 0 when more leafs are to be scanned.
3179  * returns 1 when done.
3180  */
3181 static int qgroup_rescan_leaf(struct btrfs_trans_handle *trans,
3182                               struct btrfs_path *path)
3183 {
3184         struct btrfs_fs_info *fs_info = trans->fs_info;
3185         struct btrfs_root *extent_root;
3186         struct btrfs_key found;
3187         struct extent_buffer *scratch_leaf = NULL;
3188         struct ulist *roots = NULL;
3189         u64 num_bytes;
3190         bool done;
3191         int slot;
3192         int ret;
3193
3194         mutex_lock(&fs_info->qgroup_rescan_lock);
3195         extent_root = btrfs_extent_root(fs_info,
3196                                 fs_info->qgroup_rescan_progress.objectid);
3197         ret = btrfs_search_slot_for_read(extent_root,
3198                                          &fs_info->qgroup_rescan_progress,
3199                                          path, 1, 0);
3200
3201         btrfs_debug(fs_info,
3202                 "current progress key (%llu %u %llu), search_slot ret %d",
3203                 fs_info->qgroup_rescan_progress.objectid,
3204                 fs_info->qgroup_rescan_progress.type,
3205                 fs_info->qgroup_rescan_progress.offset, ret);
3206
3207         if (ret) {
3208                 /*
3209                  * The rescan is about to end, we will not be scanning any
3210                  * further blocks. We cannot unset the RESCAN flag here, because
3211                  * we want to commit the transaction if everything went well.
3212                  * To make the live accounting work in this phase, we set our
3213                  * scan progress pointer such that every real extent objectid
3214                  * will be smaller.
3215                  */
3216                 fs_info->qgroup_rescan_progress.objectid = (u64)-1;
3217                 btrfs_release_path(path);
3218                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3219                 return ret;
3220         }
3221         done = is_last_leaf(path);
3222
3223         btrfs_item_key_to_cpu(path->nodes[0], &found,
3224                               btrfs_header_nritems(path->nodes[0]) - 1);
3225         fs_info->qgroup_rescan_progress.objectid = found.objectid + 1;
3226
3227         scratch_leaf = btrfs_clone_extent_buffer(path->nodes[0]);
3228         if (!scratch_leaf) {
3229                 ret = -ENOMEM;
3230                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3231                 goto out;
3232         }
3233         slot = path->slots[0];
3234         btrfs_release_path(path);
3235         mutex_unlock(&fs_info->qgroup_rescan_lock);
3236
3237         for (; slot < btrfs_header_nritems(scratch_leaf); ++slot) {
3238                 btrfs_item_key_to_cpu(scratch_leaf, &found, slot);
3239                 if (found.type != BTRFS_EXTENT_ITEM_KEY &&
3240                     found.type != BTRFS_METADATA_ITEM_KEY)
3241                         continue;
3242                 if (found.type == BTRFS_METADATA_ITEM_KEY)
3243                         num_bytes = fs_info->nodesize;
3244                 else
3245                         num_bytes = found.offset;
3246
3247                 ret = btrfs_find_all_roots(NULL, fs_info, found.objectid, 0,
3248                                            &roots, false);
3249                 if (ret < 0)
3250                         goto out;
3251                 /* For rescan, just pass old_roots as NULL */
3252                 ret = btrfs_qgroup_account_extent(trans, found.objectid,
3253                                                   num_bytes, NULL, roots);
3254                 if (ret < 0)
3255                         goto out;
3256         }
3257 out:
3258         if (scratch_leaf)
3259                 free_extent_buffer(scratch_leaf);
3260
3261         if (done && !ret) {
3262                 ret = 1;
3263                 fs_info->qgroup_rescan_progress.objectid = (u64)-1;
3264         }
3265         return ret;
3266 }
3267
3268 static bool rescan_should_stop(struct btrfs_fs_info *fs_info)
3269 {
3270         return btrfs_fs_closing(fs_info) ||
3271                 test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state);
3272 }
3273
3274 static void btrfs_qgroup_rescan_worker(struct btrfs_work *work)
3275 {
3276         struct btrfs_fs_info *fs_info = container_of(work, struct btrfs_fs_info,
3277                                                      qgroup_rescan_work);
3278         struct btrfs_path *path;
3279         struct btrfs_trans_handle *trans = NULL;
3280         int err = -ENOMEM;
3281         int ret = 0;
3282         bool stopped = false;
3283
3284         path = btrfs_alloc_path();
3285         if (!path)
3286                 goto out;
3287         /*
3288          * Rescan should only search for commit root, and any later difference
3289          * should be recorded by qgroup
3290          */
3291         path->search_commit_root = 1;
3292         path->skip_locking = 1;
3293
3294         err = 0;
3295         while (!err && !(stopped = rescan_should_stop(fs_info))) {
3296                 trans = btrfs_start_transaction(fs_info->fs_root, 0);
3297                 if (IS_ERR(trans)) {
3298                         err = PTR_ERR(trans);
3299                         break;
3300                 }
3301                 if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
3302                         err = -EINTR;
3303                 } else {
3304                         err = qgroup_rescan_leaf(trans, path);
3305                 }
3306                 if (err > 0)
3307                         btrfs_commit_transaction(trans);
3308                 else
3309                         btrfs_end_transaction(trans);
3310         }
3311
3312 out:
3313         btrfs_free_path(path);
3314
3315         mutex_lock(&fs_info->qgroup_rescan_lock);
3316         if (err > 0 &&
3317             fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT) {
3318                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
3319         } else if (err < 0) {
3320                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
3321         }
3322         mutex_unlock(&fs_info->qgroup_rescan_lock);
3323
3324         /*
3325          * only update status, since the previous part has already updated the
3326          * qgroup info.
3327          */
3328         trans = btrfs_start_transaction(fs_info->quota_root, 1);
3329         if (IS_ERR(trans)) {
3330                 err = PTR_ERR(trans);
3331                 trans = NULL;
3332                 btrfs_err(fs_info,
3333                           "fail to start transaction for status update: %d",
3334                           err);
3335         }
3336
3337         mutex_lock(&fs_info->qgroup_rescan_lock);
3338         if (!stopped)
3339                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3340         if (trans) {
3341                 ret = update_qgroup_status_item(trans);
3342                 if (ret < 0) {
3343                         err = ret;
3344                         btrfs_err(fs_info, "fail to update qgroup status: %d",
3345                                   err);
3346                 }
3347         }
3348         fs_info->qgroup_rescan_running = false;
3349         complete_all(&fs_info->qgroup_rescan_completion);
3350         mutex_unlock(&fs_info->qgroup_rescan_lock);
3351
3352         if (!trans)
3353                 return;
3354
3355         btrfs_end_transaction(trans);
3356
3357         if (stopped) {
3358                 btrfs_info(fs_info, "qgroup scan paused");
3359         } else if (err >= 0) {
3360                 btrfs_info(fs_info, "qgroup scan completed%s",
3361                         err > 0 ? " (inconsistency flag cleared)" : "");
3362         } else {
3363                 btrfs_err(fs_info, "qgroup scan failed with %d", err);
3364         }
3365 }
3366
3367 /*
3368  * Checks that (a) no rescan is running and (b) quota is enabled. Allocates all
3369  * memory required for the rescan context.
3370  */
3371 static int
3372 qgroup_rescan_init(struct btrfs_fs_info *fs_info, u64 progress_objectid,
3373                    int init_flags)
3374 {
3375         int ret = 0;
3376
3377         if (!init_flags) {
3378                 /* we're resuming qgroup rescan at mount time */
3379                 if (!(fs_info->qgroup_flags &
3380                       BTRFS_QGROUP_STATUS_FLAG_RESCAN)) {
3381                         btrfs_warn(fs_info,
3382                         "qgroup rescan init failed, qgroup rescan is not queued");
3383                         ret = -EINVAL;
3384                 } else if (!(fs_info->qgroup_flags &
3385                              BTRFS_QGROUP_STATUS_FLAG_ON)) {
3386                         btrfs_warn(fs_info,
3387                         "qgroup rescan init failed, qgroup is not enabled");
3388                         ret = -EINVAL;
3389                 }
3390
3391                 if (ret)
3392                         return ret;
3393         }
3394
3395         mutex_lock(&fs_info->qgroup_rescan_lock);
3396
3397         if (init_flags) {
3398                 if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
3399                         btrfs_warn(fs_info,
3400                                    "qgroup rescan is already in progress");
3401                         ret = -EINPROGRESS;
3402                 } else if (!(fs_info->qgroup_flags &
3403                              BTRFS_QGROUP_STATUS_FLAG_ON)) {
3404                         btrfs_warn(fs_info,
3405                         "qgroup rescan init failed, qgroup is not enabled");
3406                         ret = -EINVAL;
3407                 } else if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
3408                         /* Quota disable is in progress */
3409                         ret = -EBUSY;
3410                 }
3411
3412                 if (ret) {
3413                         mutex_unlock(&fs_info->qgroup_rescan_lock);
3414                         return ret;
3415                 }
3416                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3417         }
3418
3419         memset(&fs_info->qgroup_rescan_progress, 0,
3420                 sizeof(fs_info->qgroup_rescan_progress));
3421         fs_info->qgroup_rescan_progress.objectid = progress_objectid;
3422         init_completion(&fs_info->qgroup_rescan_completion);
3423         mutex_unlock(&fs_info->qgroup_rescan_lock);
3424
3425         btrfs_init_work(&fs_info->qgroup_rescan_work,
3426                         btrfs_qgroup_rescan_worker, NULL, NULL);
3427         return 0;
3428 }
3429
3430 static void
3431 qgroup_rescan_zero_tracking(struct btrfs_fs_info *fs_info)
3432 {
3433         struct rb_node *n;
3434         struct btrfs_qgroup *qgroup;
3435
3436         spin_lock(&fs_info->qgroup_lock);
3437         /* clear all current qgroup tracking information */
3438         for (n = rb_first(&fs_info->qgroup_tree); n; n = rb_next(n)) {
3439                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
3440                 qgroup->rfer = 0;
3441                 qgroup->rfer_cmpr = 0;
3442                 qgroup->excl = 0;
3443                 qgroup->excl_cmpr = 0;
3444                 qgroup_dirty(fs_info, qgroup);
3445         }
3446         spin_unlock(&fs_info->qgroup_lock);
3447 }
3448
3449 int
3450 btrfs_qgroup_rescan(struct btrfs_fs_info *fs_info)
3451 {
3452         int ret = 0;
3453         struct btrfs_trans_handle *trans;
3454
3455         ret = qgroup_rescan_init(fs_info, 0, 1);
3456         if (ret)
3457                 return ret;
3458
3459         /*
3460          * We have set the rescan_progress to 0, which means no more
3461          * delayed refs will be accounted by btrfs_qgroup_account_ref.
3462          * However, btrfs_qgroup_account_ref may be right after its call
3463          * to btrfs_find_all_roots, in which case it would still do the
3464          * accounting.
3465          * To solve this, we're committing the transaction, which will
3466          * ensure we run all delayed refs and only after that, we are
3467          * going to clear all tracking information for a clean start.
3468          */
3469
3470         trans = btrfs_join_transaction(fs_info->fs_root);
3471         if (IS_ERR(trans)) {
3472                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3473                 return PTR_ERR(trans);
3474         }
3475         ret = btrfs_commit_transaction(trans);
3476         if (ret) {
3477                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3478                 return ret;
3479         }
3480
3481         qgroup_rescan_zero_tracking(fs_info);
3482
3483         mutex_lock(&fs_info->qgroup_rescan_lock);
3484         fs_info->qgroup_rescan_running = true;
3485         btrfs_queue_work(fs_info->qgroup_rescan_workers,
3486                          &fs_info->qgroup_rescan_work);
3487         mutex_unlock(&fs_info->qgroup_rescan_lock);
3488
3489         return 0;
3490 }
3491
3492 int btrfs_qgroup_wait_for_completion(struct btrfs_fs_info *fs_info,
3493                                      bool interruptible)
3494 {
3495         int running;
3496         int ret = 0;
3497
3498         mutex_lock(&fs_info->qgroup_rescan_lock);
3499         running = fs_info->qgroup_rescan_running;
3500         mutex_unlock(&fs_info->qgroup_rescan_lock);
3501
3502         if (!running)
3503                 return 0;
3504
3505         if (interruptible)
3506                 ret = wait_for_completion_interruptible(
3507                                         &fs_info->qgroup_rescan_completion);
3508         else
3509                 wait_for_completion(&fs_info->qgroup_rescan_completion);
3510
3511         return ret;
3512 }
3513
3514 /*
3515  * this is only called from open_ctree where we're still single threaded, thus
3516  * locking is omitted here.
3517  */
3518 void
3519 btrfs_qgroup_rescan_resume(struct btrfs_fs_info *fs_info)
3520 {
3521         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
3522                 mutex_lock(&fs_info->qgroup_rescan_lock);
3523                 fs_info->qgroup_rescan_running = true;
3524                 btrfs_queue_work(fs_info->qgroup_rescan_workers,
3525                                  &fs_info->qgroup_rescan_work);
3526                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3527         }
3528 }
3529
3530 #define rbtree_iterate_from_safe(node, next, start)                             \
3531        for (node = start; node && ({ next = rb_next(node); 1;}); node = next)
3532
3533 static int qgroup_unreserve_range(struct btrfs_inode *inode,
3534                                   struct extent_changeset *reserved, u64 start,
3535                                   u64 len)
3536 {
3537         struct rb_node *node;
3538         struct rb_node *next;
3539         struct ulist_node *entry;
3540         int ret = 0;
3541
3542         node = reserved->range_changed.root.rb_node;
3543         if (!node)
3544                 return 0;
3545         while (node) {
3546                 entry = rb_entry(node, struct ulist_node, rb_node);
3547                 if (entry->val < start)
3548                         node = node->rb_right;
3549                 else
3550                         node = node->rb_left;
3551         }
3552
3553         if (entry->val > start && rb_prev(&entry->rb_node))
3554                 entry = rb_entry(rb_prev(&entry->rb_node), struct ulist_node,
3555                                  rb_node);
3556
3557         rbtree_iterate_from_safe(node, next, &entry->rb_node) {
3558                 u64 entry_start;
3559                 u64 entry_end;
3560                 u64 entry_len;
3561                 int clear_ret;
3562
3563                 entry = rb_entry(node, struct ulist_node, rb_node);
3564                 entry_start = entry->val;
3565                 entry_end = entry->aux;
3566                 entry_len = entry_end - entry_start + 1;
3567
3568                 if (entry_start >= start + len)
3569                         break;
3570                 if (entry_start + entry_len <= start)
3571                         continue;
3572                 /*
3573                  * Now the entry is in [start, start + len), revert the
3574                  * EXTENT_QGROUP_RESERVED bit.
3575                  */
3576                 clear_ret = clear_extent_bits(&inode->io_tree, entry_start,
3577                                               entry_end, EXTENT_QGROUP_RESERVED);
3578                 if (!ret && clear_ret < 0)
3579                         ret = clear_ret;
3580
3581                 ulist_del(&reserved->range_changed, entry->val, entry->aux);
3582                 if (likely(reserved->bytes_changed >= entry_len)) {
3583                         reserved->bytes_changed -= entry_len;
3584                 } else {
3585                         WARN_ON(1);
3586                         reserved->bytes_changed = 0;
3587                 }
3588         }
3589
3590         return ret;
3591 }
3592
3593 /*
3594  * Try to free some space for qgroup.
3595  *
3596  * For qgroup, there are only 3 ways to free qgroup space:
3597  * - Flush nodatacow write
3598  *   Any nodatacow write will free its reserved data space at run_delalloc_range().
3599  *   In theory, we should only flush nodatacow inodes, but it's not yet
3600  *   possible, so we need to flush the whole root.
3601  *
3602  * - Wait for ordered extents
3603  *   When ordered extents are finished, their reserved metadata is finally
3604  *   converted to per_trans status, which can be freed by later commit
3605  *   transaction.
3606  *
3607  * - Commit transaction
3608  *   This would free the meta_per_trans space.
3609  *   In theory this shouldn't provide much space, but any more qgroup space
3610  *   is needed.
3611  */
3612 static int try_flush_qgroup(struct btrfs_root *root)
3613 {
3614         struct btrfs_trans_handle *trans;
3615         int ret;
3616
3617         /* Can't hold an open transaction or we run the risk of deadlocking. */
3618         ASSERT(current->journal_info == NULL);
3619         if (WARN_ON(current->journal_info))
3620                 return 0;
3621
3622         /*
3623          * We don't want to run flush again and again, so if there is a running
3624          * one, we won't try to start a new flush, but exit directly.
3625          */
3626         if (test_and_set_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state)) {
3627                 wait_event(root->qgroup_flush_wait,
3628                         !test_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state));
3629                 return 0;
3630         }
3631
3632         ret = btrfs_start_delalloc_snapshot(root, true);
3633         if (ret < 0)
3634                 goto out;
3635         btrfs_wait_ordered_extents(root, U64_MAX, 0, (u64)-1);
3636
3637         trans = btrfs_join_transaction(root);
3638         if (IS_ERR(trans)) {
3639                 ret = PTR_ERR(trans);
3640                 goto out;
3641         }
3642
3643         ret = btrfs_commit_transaction(trans);
3644 out:
3645         clear_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state);
3646         wake_up(&root->qgroup_flush_wait);
3647         return ret;
3648 }
3649
3650 static int qgroup_reserve_data(struct btrfs_inode *inode,
3651                         struct extent_changeset **reserved_ret, u64 start,
3652                         u64 len)
3653 {
3654         struct btrfs_root *root = inode->root;
3655         struct extent_changeset *reserved;
3656         bool new_reserved = false;
3657         u64 orig_reserved;
3658         u64 to_reserve;
3659         int ret;
3660
3661         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &root->fs_info->flags) ||
3662             !is_fstree(root->root_key.objectid) || len == 0)
3663                 return 0;
3664
3665         /* @reserved parameter is mandatory for qgroup */
3666         if (WARN_ON(!reserved_ret))
3667                 return -EINVAL;
3668         if (!*reserved_ret) {
3669                 new_reserved = true;
3670                 *reserved_ret = extent_changeset_alloc();
3671                 if (!*reserved_ret)
3672                         return -ENOMEM;
3673         }
3674         reserved = *reserved_ret;
3675         /* Record already reserved space */
3676         orig_reserved = reserved->bytes_changed;
3677         ret = set_record_extent_bits(&inode->io_tree, start,
3678                         start + len -1, EXTENT_QGROUP_RESERVED, reserved);
3679
3680         /* Newly reserved space */
3681         to_reserve = reserved->bytes_changed - orig_reserved;
3682         trace_btrfs_qgroup_reserve_data(&inode->vfs_inode, start, len,
3683                                         to_reserve, QGROUP_RESERVE);
3684         if (ret < 0)
3685                 goto out;
3686         ret = qgroup_reserve(root, to_reserve, true, BTRFS_QGROUP_RSV_DATA);
3687         if (ret < 0)
3688                 goto cleanup;
3689
3690         return ret;
3691
3692 cleanup:
3693         qgroup_unreserve_range(inode, reserved, start, len);
3694 out:
3695         if (new_reserved) {
3696                 extent_changeset_free(reserved);
3697                 *reserved_ret = NULL;
3698         }
3699         return ret;
3700 }
3701
3702 /*
3703  * Reserve qgroup space for range [start, start + len).
3704  *
3705  * This function will either reserve space from related qgroups or do nothing
3706  * if the range is already reserved.
3707  *
3708  * Return 0 for successful reservation
3709  * Return <0 for error (including -EQUOT)
3710  *
3711  * NOTE: This function may sleep for memory allocation, dirty page flushing and
3712  *       commit transaction. So caller should not hold any dirty page locked.
3713  */
3714 int btrfs_qgroup_reserve_data(struct btrfs_inode *inode,
3715                         struct extent_changeset **reserved_ret, u64 start,
3716                         u64 len)
3717 {
3718         int ret;
3719
3720         ret = qgroup_reserve_data(inode, reserved_ret, start, len);
3721         if (ret <= 0 && ret != -EDQUOT)
3722                 return ret;
3723
3724         ret = try_flush_qgroup(inode->root);
3725         if (ret < 0)
3726                 return ret;
3727         return qgroup_reserve_data(inode, reserved_ret, start, len);
3728 }
3729
3730 /* Free ranges specified by @reserved, normally in error path */
3731 static int qgroup_free_reserved_data(struct btrfs_inode *inode,
3732                         struct extent_changeset *reserved, u64 start, u64 len)
3733 {
3734         struct btrfs_root *root = inode->root;
3735         struct ulist_node *unode;
3736         struct ulist_iterator uiter;
3737         struct extent_changeset changeset;
3738         int freed = 0;
3739         int ret;
3740
3741         extent_changeset_init(&changeset);
3742         len = round_up(start + len, root->fs_info->sectorsize);
3743         start = round_down(start, root->fs_info->sectorsize);
3744
3745         ULIST_ITER_INIT(&uiter);
3746         while ((unode = ulist_next(&reserved->range_changed, &uiter))) {
3747                 u64 range_start = unode->val;
3748                 /* unode->aux is the inclusive end */
3749                 u64 range_len = unode->aux - range_start + 1;
3750                 u64 free_start;
3751                 u64 free_len;
3752
3753                 extent_changeset_release(&changeset);
3754
3755                 /* Only free range in range [start, start + len) */
3756                 if (range_start >= start + len ||
3757                     range_start + range_len <= start)
3758                         continue;
3759                 free_start = max(range_start, start);
3760                 free_len = min(start + len, range_start + range_len) -
3761                            free_start;
3762                 /*
3763                  * TODO: To also modify reserved->ranges_reserved to reflect
3764                  * the modification.
3765                  *
3766                  * However as long as we free qgroup reserved according to
3767                  * EXTENT_QGROUP_RESERVED, we won't double free.
3768                  * So not need to rush.
3769                  */
3770                 ret = clear_record_extent_bits(&inode->io_tree, free_start,
3771                                 free_start + free_len - 1,
3772                                 EXTENT_QGROUP_RESERVED, &changeset);
3773                 if (ret < 0)
3774                         goto out;
3775                 freed += changeset.bytes_changed;
3776         }
3777         btrfs_qgroup_free_refroot(root->fs_info, root->root_key.objectid, freed,
3778                                   BTRFS_QGROUP_RSV_DATA);
3779         ret = freed;
3780 out:
3781         extent_changeset_release(&changeset);
3782         return ret;
3783 }
3784
3785 static int __btrfs_qgroup_release_data(struct btrfs_inode *inode,
3786                         struct extent_changeset *reserved, u64 start, u64 len,
3787                         int free)
3788 {
3789         struct extent_changeset changeset;
3790         int trace_op = QGROUP_RELEASE;
3791         int ret;
3792
3793         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &inode->root->fs_info->flags))
3794                 return 0;
3795
3796         /* In release case, we shouldn't have @reserved */
3797         WARN_ON(!free && reserved);
3798         if (free && reserved)
3799                 return qgroup_free_reserved_data(inode, reserved, start, len);
3800         extent_changeset_init(&changeset);
3801         ret = clear_record_extent_bits(&inode->io_tree, start, start + len -1,
3802                                        EXTENT_QGROUP_RESERVED, &changeset);
3803         if (ret < 0)
3804                 goto out;
3805
3806         if (free)
3807                 trace_op = QGROUP_FREE;
3808         trace_btrfs_qgroup_release_data(&inode->vfs_inode, start, len,
3809                                         changeset.bytes_changed, trace_op);
3810         if (free)
3811                 btrfs_qgroup_free_refroot(inode->root->fs_info,
3812                                 inode->root->root_key.objectid,
3813                                 changeset.bytes_changed, BTRFS_QGROUP_RSV_DATA);
3814         ret = changeset.bytes_changed;
3815 out:
3816         extent_changeset_release(&changeset);
3817         return ret;
3818 }
3819
3820 /*
3821  * Free a reserved space range from io_tree and related qgroups
3822  *
3823  * Should be called when a range of pages get invalidated before reaching disk.
3824  * Or for error cleanup case.
3825  * if @reserved is given, only reserved range in [@start, @start + @len) will
3826  * be freed.
3827  *
3828  * For data written to disk, use btrfs_qgroup_release_data().
3829  *
3830  * NOTE: This function may sleep for memory allocation.
3831  */
3832 int btrfs_qgroup_free_data(struct btrfs_inode *inode,
3833                         struct extent_changeset *reserved, u64 start, u64 len)
3834 {
3835         return __btrfs_qgroup_release_data(inode, reserved, start, len, 1);
3836 }
3837
3838 /*
3839  * Release a reserved space range from io_tree only.
3840  *
3841  * Should be called when a range of pages get written to disk and corresponding
3842  * FILE_EXTENT is inserted into corresponding root.
3843  *
3844  * Since new qgroup accounting framework will only update qgroup numbers at
3845  * commit_transaction() time, its reserved space shouldn't be freed from
3846  * related qgroups.
3847  *
3848  * But we should release the range from io_tree, to allow further write to be
3849  * COWed.
3850  *
3851  * NOTE: This function may sleep for memory allocation.
3852  */
3853 int btrfs_qgroup_release_data(struct btrfs_inode *inode, u64 start, u64 len)
3854 {
3855         return __btrfs_qgroup_release_data(inode, NULL, start, len, 0);
3856 }
3857
3858 static void add_root_meta_rsv(struct btrfs_root *root, int num_bytes,
3859                               enum btrfs_qgroup_rsv_type type)
3860 {
3861         if (type != BTRFS_QGROUP_RSV_META_PREALLOC &&
3862             type != BTRFS_QGROUP_RSV_META_PERTRANS)
3863                 return;
3864         if (num_bytes == 0)
3865                 return;
3866
3867         spin_lock(&root->qgroup_meta_rsv_lock);
3868         if (type == BTRFS_QGROUP_RSV_META_PREALLOC)
3869                 root->qgroup_meta_rsv_prealloc += num_bytes;
3870         else
3871                 root->qgroup_meta_rsv_pertrans += num_bytes;
3872         spin_unlock(&root->qgroup_meta_rsv_lock);
3873 }
3874
3875 static int sub_root_meta_rsv(struct btrfs_root *root, int num_bytes,
3876                              enum btrfs_qgroup_rsv_type type)
3877 {
3878         if (type != BTRFS_QGROUP_RSV_META_PREALLOC &&
3879             type != BTRFS_QGROUP_RSV_META_PERTRANS)
3880                 return 0;
3881         if (num_bytes == 0)
3882                 return 0;
3883
3884         spin_lock(&root->qgroup_meta_rsv_lock);
3885         if (type == BTRFS_QGROUP_RSV_META_PREALLOC) {
3886                 num_bytes = min_t(u64, root->qgroup_meta_rsv_prealloc,
3887                                   num_bytes);
3888                 root->qgroup_meta_rsv_prealloc -= num_bytes;
3889         } else {
3890                 num_bytes = min_t(u64, root->qgroup_meta_rsv_pertrans,
3891                                   num_bytes);
3892                 root->qgroup_meta_rsv_pertrans -= num_bytes;
3893         }
3894         spin_unlock(&root->qgroup_meta_rsv_lock);
3895         return num_bytes;
3896 }
3897
3898 int btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
3899                               enum btrfs_qgroup_rsv_type type, bool enforce)
3900 {
3901         struct btrfs_fs_info *fs_info = root->fs_info;
3902         int ret;
3903
3904         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
3905             !is_fstree(root->root_key.objectid) || num_bytes == 0)
3906                 return 0;
3907
3908         BUG_ON(num_bytes != round_down(num_bytes, fs_info->nodesize));
3909         trace_qgroup_meta_reserve(root, (s64)num_bytes, type);
3910         ret = qgroup_reserve(root, num_bytes, enforce, type);
3911         if (ret < 0)
3912                 return ret;
3913         /*
3914          * Record what we have reserved into root.
3915          *
3916          * To avoid quota disabled->enabled underflow.
3917          * In that case, we may try to free space we haven't reserved
3918          * (since quota was disabled), so record what we reserved into root.
3919          * And ensure later release won't underflow this number.
3920          */
3921         add_root_meta_rsv(root, num_bytes, type);
3922         return ret;
3923 }
3924
3925 int __btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
3926                                 enum btrfs_qgroup_rsv_type type, bool enforce)
3927 {
3928         int ret;
3929
3930         ret = btrfs_qgroup_reserve_meta(root, num_bytes, type, enforce);
3931         if (ret <= 0 && ret != -EDQUOT)
3932                 return ret;
3933
3934         ret = try_flush_qgroup(root);
3935         if (ret < 0)
3936                 return ret;
3937         return btrfs_qgroup_reserve_meta(root, num_bytes, type, enforce);
3938 }
3939
3940 void btrfs_qgroup_free_meta_all_pertrans(struct btrfs_root *root)
3941 {
3942         struct btrfs_fs_info *fs_info = root->fs_info;
3943
3944         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
3945             !is_fstree(root->root_key.objectid))
3946                 return;
3947
3948         /* TODO: Update trace point to handle such free */
3949         trace_qgroup_meta_free_all_pertrans(root);
3950         /* Special value -1 means to free all reserved space */
3951         btrfs_qgroup_free_refroot(fs_info, root->root_key.objectid, (u64)-1,
3952                                   BTRFS_QGROUP_RSV_META_PERTRANS);
3953 }
3954
3955 void __btrfs_qgroup_free_meta(struct btrfs_root *root, int num_bytes,
3956                               enum btrfs_qgroup_rsv_type type)
3957 {
3958         struct btrfs_fs_info *fs_info = root->fs_info;
3959
3960         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
3961             !is_fstree(root->root_key.objectid))
3962                 return;
3963
3964         /*
3965          * reservation for META_PREALLOC can happen before quota is enabled,
3966          * which can lead to underflow.
3967          * Here ensure we will only free what we really have reserved.
3968          */
3969         num_bytes = sub_root_meta_rsv(root, num_bytes, type);
3970         BUG_ON(num_bytes != round_down(num_bytes, fs_info->nodesize));
3971         trace_qgroup_meta_reserve(root, -(s64)num_bytes, type);
3972         btrfs_qgroup_free_refroot(fs_info, root->root_key.objectid,
3973                                   num_bytes, type);
3974 }
3975
3976 static void qgroup_convert_meta(struct btrfs_fs_info *fs_info, u64 ref_root,
3977                                 int num_bytes)
3978 {
3979         struct btrfs_qgroup *qgroup;
3980         struct ulist_node *unode;
3981         struct ulist_iterator uiter;
3982         int ret = 0;
3983
3984         if (num_bytes == 0)
3985                 return;
3986         if (!fs_info->quota_root)
3987                 return;
3988
3989         spin_lock(&fs_info->qgroup_lock);
3990         qgroup = find_qgroup_rb(fs_info, ref_root);
3991         if (!qgroup)
3992                 goto out;
3993         ulist_reinit(fs_info->qgroup_ulist);
3994         ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
3995                        qgroup_to_aux(qgroup), GFP_ATOMIC);
3996         if (ret < 0)
3997                 goto out;
3998         ULIST_ITER_INIT(&uiter);
3999         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
4000                 struct btrfs_qgroup *qg;
4001                 struct btrfs_qgroup_list *glist;
4002
4003                 qg = unode_aux_to_qgroup(unode);
4004
4005                 qgroup_rsv_release(fs_info, qg, num_bytes,
4006                                 BTRFS_QGROUP_RSV_META_PREALLOC);
4007                 qgroup_rsv_add(fs_info, qg, num_bytes,
4008                                 BTRFS_QGROUP_RSV_META_PERTRANS);
4009                 list_for_each_entry(glist, &qg->groups, next_group) {
4010                         ret = ulist_add(fs_info->qgroup_ulist,
4011                                         glist->group->qgroupid,
4012                                         qgroup_to_aux(glist->group), GFP_ATOMIC);
4013                         if (ret < 0)
4014                                 goto out;
4015                 }
4016         }
4017 out:
4018         spin_unlock(&fs_info->qgroup_lock);
4019 }
4020
4021 void btrfs_qgroup_convert_reserved_meta(struct btrfs_root *root, int num_bytes)
4022 {
4023         struct btrfs_fs_info *fs_info = root->fs_info;
4024
4025         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
4026             !is_fstree(root->root_key.objectid))
4027                 return;
4028         /* Same as btrfs_qgroup_free_meta_prealloc() */
4029         num_bytes = sub_root_meta_rsv(root, num_bytes,
4030                                       BTRFS_QGROUP_RSV_META_PREALLOC);
4031         trace_qgroup_meta_convert(root, num_bytes);
4032         qgroup_convert_meta(fs_info, root->root_key.objectid, num_bytes);
4033 }
4034
4035 /*
4036  * Check qgroup reserved space leaking, normally at destroy inode
4037  * time
4038  */
4039 void btrfs_qgroup_check_reserved_leak(struct btrfs_inode *inode)
4040 {
4041         struct extent_changeset changeset;
4042         struct ulist_node *unode;
4043         struct ulist_iterator iter;
4044         int ret;
4045
4046         extent_changeset_init(&changeset);
4047         ret = clear_record_extent_bits(&inode->io_tree, 0, (u64)-1,
4048                         EXTENT_QGROUP_RESERVED, &changeset);
4049
4050         WARN_ON(ret < 0);
4051         if (WARN_ON(changeset.bytes_changed)) {
4052                 ULIST_ITER_INIT(&iter);
4053                 while ((unode = ulist_next(&changeset.range_changed, &iter))) {
4054                         btrfs_warn(inode->root->fs_info,
4055                 "leaking qgroup reserved space, ino: %llu, start: %llu, end: %llu",
4056                                 btrfs_ino(inode), unode->val, unode->aux);
4057                 }
4058                 btrfs_qgroup_free_refroot(inode->root->fs_info,
4059                                 inode->root->root_key.objectid,
4060                                 changeset.bytes_changed, BTRFS_QGROUP_RSV_DATA);
4061
4062         }
4063         extent_changeset_release(&changeset);
4064 }
4065
4066 void btrfs_qgroup_init_swapped_blocks(
4067         struct btrfs_qgroup_swapped_blocks *swapped_blocks)
4068 {
4069         int i;
4070
4071         spin_lock_init(&swapped_blocks->lock);
4072         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
4073                 swapped_blocks->blocks[i] = RB_ROOT;
4074         swapped_blocks->swapped = false;
4075 }
4076
4077 /*
4078  * Delete all swapped blocks record of @root.
4079  * Every record here means we skipped a full subtree scan for qgroup.
4080  *
4081  * Gets called when committing one transaction.
4082  */
4083 void btrfs_qgroup_clean_swapped_blocks(struct btrfs_root *root)
4084 {
4085         struct btrfs_qgroup_swapped_blocks *swapped_blocks;
4086         int i;
4087
4088         swapped_blocks = &root->swapped_blocks;
4089
4090         spin_lock(&swapped_blocks->lock);
4091         if (!swapped_blocks->swapped)
4092                 goto out;
4093         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
4094                 struct rb_root *cur_root = &swapped_blocks->blocks[i];
4095                 struct btrfs_qgroup_swapped_block *entry;
4096                 struct btrfs_qgroup_swapped_block *next;
4097
4098                 rbtree_postorder_for_each_entry_safe(entry, next, cur_root,
4099                                                      node)
4100                         kfree(entry);
4101                 swapped_blocks->blocks[i] = RB_ROOT;
4102         }
4103         swapped_blocks->swapped = false;
4104 out:
4105         spin_unlock(&swapped_blocks->lock);
4106 }
4107
4108 /*
4109  * Add subtree roots record into @subvol_root.
4110  *
4111  * @subvol_root:        tree root of the subvolume tree get swapped
4112  * @bg:                 block group under balance
4113  * @subvol_parent/slot: pointer to the subtree root in subvolume tree
4114  * @reloc_parent/slot:  pointer to the subtree root in reloc tree
4115  *                      BOTH POINTERS ARE BEFORE TREE SWAP
4116  * @last_snapshot:      last snapshot generation of the subvolume tree
4117  */
4118 int btrfs_qgroup_add_swapped_blocks(struct btrfs_trans_handle *trans,
4119                 struct btrfs_root *subvol_root,
4120                 struct btrfs_block_group *bg,
4121                 struct extent_buffer *subvol_parent, int subvol_slot,
4122                 struct extent_buffer *reloc_parent, int reloc_slot,
4123                 u64 last_snapshot)
4124 {
4125         struct btrfs_fs_info *fs_info = subvol_root->fs_info;
4126         struct btrfs_qgroup_swapped_blocks *blocks = &subvol_root->swapped_blocks;
4127         struct btrfs_qgroup_swapped_block *block;
4128         struct rb_node **cur;
4129         struct rb_node *parent = NULL;
4130         int level = btrfs_header_level(subvol_parent) - 1;
4131         int ret = 0;
4132
4133         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
4134                 return 0;
4135
4136         if (btrfs_node_ptr_generation(subvol_parent, subvol_slot) >
4137             btrfs_node_ptr_generation(reloc_parent, reloc_slot)) {
4138                 btrfs_err_rl(fs_info,
4139                 "%s: bad parameter order, subvol_gen=%llu reloc_gen=%llu",
4140                         __func__,
4141                         btrfs_node_ptr_generation(subvol_parent, subvol_slot),
4142                         btrfs_node_ptr_generation(reloc_parent, reloc_slot));
4143                 return -EUCLEAN;
4144         }
4145
4146         block = kmalloc(sizeof(*block), GFP_NOFS);
4147         if (!block) {
4148                 ret = -ENOMEM;
4149                 goto out;
4150         }
4151
4152         /*
4153          * @reloc_parent/slot is still before swap, while @block is going to
4154          * record the bytenr after swap, so we do the swap here.
4155          */
4156         block->subvol_bytenr = btrfs_node_blockptr(reloc_parent, reloc_slot);
4157         block->subvol_generation = btrfs_node_ptr_generation(reloc_parent,
4158                                                              reloc_slot);
4159         block->reloc_bytenr = btrfs_node_blockptr(subvol_parent, subvol_slot);
4160         block->reloc_generation = btrfs_node_ptr_generation(subvol_parent,
4161                                                             subvol_slot);
4162         block->last_snapshot = last_snapshot;
4163         block->level = level;
4164
4165         /*
4166          * If we have bg == NULL, we're called from btrfs_recover_relocation(),
4167          * no one else can modify tree blocks thus we qgroup will not change
4168          * no matter the value of trace_leaf.
4169          */
4170         if (bg && bg->flags & BTRFS_BLOCK_GROUP_DATA)
4171                 block->trace_leaf = true;
4172         else
4173                 block->trace_leaf = false;
4174         btrfs_node_key_to_cpu(reloc_parent, &block->first_key, reloc_slot);
4175
4176         /* Insert @block into @blocks */
4177         spin_lock(&blocks->lock);
4178         cur = &blocks->blocks[level].rb_node;
4179         while (*cur) {
4180                 struct btrfs_qgroup_swapped_block *entry;
4181
4182                 parent = *cur;
4183                 entry = rb_entry(parent, struct btrfs_qgroup_swapped_block,
4184                                  node);
4185
4186                 if (entry->subvol_bytenr < block->subvol_bytenr) {
4187                         cur = &(*cur)->rb_left;
4188                 } else if (entry->subvol_bytenr > block->subvol_bytenr) {
4189                         cur = &(*cur)->rb_right;
4190                 } else {
4191                         if (entry->subvol_generation !=
4192                                         block->subvol_generation ||
4193                             entry->reloc_bytenr != block->reloc_bytenr ||
4194                             entry->reloc_generation !=
4195                                         block->reloc_generation) {
4196                                 /*
4197                                  * Duplicated but mismatch entry found.
4198                                  * Shouldn't happen.
4199                                  *
4200                                  * Marking qgroup inconsistent should be enough
4201                                  * for end users.
4202                                  */
4203                                 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
4204                                 ret = -EEXIST;
4205                         }
4206                         kfree(block);
4207                         goto out_unlock;
4208                 }
4209         }
4210         rb_link_node(&block->node, parent, cur);
4211         rb_insert_color(&block->node, &blocks->blocks[level]);
4212         blocks->swapped = true;
4213 out_unlock:
4214         spin_unlock(&blocks->lock);
4215 out:
4216         if (ret < 0)
4217                 fs_info->qgroup_flags |=
4218                         BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
4219         return ret;
4220 }
4221
4222 /*
4223  * Check if the tree block is a subtree root, and if so do the needed
4224  * delayed subtree trace for qgroup.
4225  *
4226  * This is called during btrfs_cow_block().
4227  */
4228 int btrfs_qgroup_trace_subtree_after_cow(struct btrfs_trans_handle *trans,
4229                                          struct btrfs_root *root,
4230                                          struct extent_buffer *subvol_eb)
4231 {
4232         struct btrfs_fs_info *fs_info = root->fs_info;
4233         struct btrfs_qgroup_swapped_blocks *blocks = &root->swapped_blocks;
4234         struct btrfs_qgroup_swapped_block *block;
4235         struct extent_buffer *reloc_eb = NULL;
4236         struct rb_node *node;
4237         bool found = false;
4238         bool swapped = false;
4239         int level = btrfs_header_level(subvol_eb);
4240         int ret = 0;
4241         int i;
4242
4243         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
4244                 return 0;
4245         if (!is_fstree(root->root_key.objectid) || !root->reloc_root)
4246                 return 0;
4247
4248         spin_lock(&blocks->lock);
4249         if (!blocks->swapped) {
4250                 spin_unlock(&blocks->lock);
4251                 return 0;
4252         }
4253         node = blocks->blocks[level].rb_node;
4254
4255         while (node) {
4256                 block = rb_entry(node, struct btrfs_qgroup_swapped_block, node);
4257                 if (block->subvol_bytenr < subvol_eb->start) {
4258                         node = node->rb_left;
4259                 } else if (block->subvol_bytenr > subvol_eb->start) {
4260                         node = node->rb_right;
4261                 } else {
4262                         found = true;
4263                         break;
4264                 }
4265         }
4266         if (!found) {
4267                 spin_unlock(&blocks->lock);
4268                 goto out;
4269         }
4270         /* Found one, remove it from @blocks first and update blocks->swapped */
4271         rb_erase(&block->node, &blocks->blocks[level]);
4272         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
4273                 if (RB_EMPTY_ROOT(&blocks->blocks[i])) {
4274                         swapped = true;
4275                         break;
4276                 }
4277         }
4278         blocks->swapped = swapped;
4279         spin_unlock(&blocks->lock);
4280
4281         /* Read out reloc subtree root */
4282         reloc_eb = read_tree_block(fs_info, block->reloc_bytenr, 0,
4283                                    block->reloc_generation, block->level,
4284                                    &block->first_key);
4285         if (IS_ERR(reloc_eb)) {
4286                 ret = PTR_ERR(reloc_eb);
4287                 reloc_eb = NULL;
4288                 goto free_out;
4289         }
4290         if (!extent_buffer_uptodate(reloc_eb)) {
4291                 ret = -EIO;
4292                 goto free_out;
4293         }
4294
4295         ret = qgroup_trace_subtree_swap(trans, reloc_eb, subvol_eb,
4296                         block->last_snapshot, block->trace_leaf);
4297 free_out:
4298         kfree(block);
4299         free_extent_buffer(reloc_eb);
4300 out:
4301         if (ret < 0) {
4302                 btrfs_err_rl(fs_info,
4303                              "failed to account subtree at bytenr %llu: %d",
4304                              subvol_eb->start, ret);
4305                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
4306         }
4307         return ret;
4308 }
4309
4310 void btrfs_qgroup_destroy_extent_records(struct btrfs_transaction *trans)
4311 {
4312         struct btrfs_qgroup_extent_record *entry;
4313         struct btrfs_qgroup_extent_record *next;
4314         struct rb_root *root;
4315
4316         root = &trans->delayed_refs.dirty_extent_root;
4317         rbtree_postorder_for_each_entry_safe(entry, next, root, node) {
4318                 ulist_free(entry->old_roots);
4319                 kfree(entry);
4320         }
4321 }