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         } else {
1161                 /*
1162                  * We have set both BTRFS_FS_QUOTA_ENABLED and
1163                  * BTRFS_QGROUP_STATUS_FLAG_ON, so we can only fail with
1164                  * -EINPROGRESS. That can happen because someone started the
1165                  * rescan worker by calling quota rescan ioctl before we
1166                  * attempted to initialize the rescan worker. Failure due to
1167                  * quotas disabled in the meanwhile is not possible, because
1168                  * we are holding a write lock on fs_info->subvol_sem, which
1169                  * is also acquired when disabling quotas.
1170                  * Ignore such error, and any other error would need to undo
1171                  * everything we did in the transaction we just committed.
1172                  */
1173                 ASSERT(ret == -EINPROGRESS);
1174                 ret = 0;
1175         }
1176
1177 out_free_path:
1178         btrfs_free_path(path);
1179 out_free_root:
1180         if (ret)
1181                 btrfs_put_root(quota_root);
1182 out:
1183         if (ret) {
1184                 ulist_free(fs_info->qgroup_ulist);
1185                 fs_info->qgroup_ulist = NULL;
1186                 btrfs_sysfs_del_qgroups(fs_info);
1187         }
1188         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1189         if (ret && trans)
1190                 btrfs_end_transaction(trans);
1191         else if (trans)
1192                 ret = btrfs_end_transaction(trans);
1193         ulist_free(ulist);
1194         return ret;
1195 }
1196
1197 int btrfs_quota_disable(struct btrfs_fs_info *fs_info)
1198 {
1199         struct btrfs_root *quota_root;
1200         struct btrfs_trans_handle *trans = NULL;
1201         int ret = 0;
1202
1203         /*
1204          * We need to have subvol_sem write locked, to prevent races between
1205          * concurrent tasks trying to disable quotas, because we will unlock
1206          * and relock qgroup_ioctl_lock across BTRFS_FS_QUOTA_ENABLED changes.
1207          */
1208         lockdep_assert_held_write(&fs_info->subvol_sem);
1209
1210         mutex_lock(&fs_info->qgroup_ioctl_lock);
1211         if (!fs_info->quota_root)
1212                 goto out;
1213
1214         /*
1215          * Unlock the qgroup_ioctl_lock mutex before waiting for the rescan worker to
1216          * complete. Otherwise we can deadlock because btrfs_remove_qgroup() needs
1217          * to lock that mutex while holding a transaction handle and the rescan
1218          * worker needs to commit a transaction.
1219          */
1220         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1221
1222         /*
1223          * Request qgroup rescan worker to complete and wait for it. This wait
1224          * must be done before transaction start for quota disable since it may
1225          * deadlock with transaction by the qgroup rescan worker.
1226          */
1227         clear_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1228         btrfs_qgroup_wait_for_completion(fs_info, false);
1229
1230         /*
1231          * 1 For the root item
1232          *
1233          * We should also reserve enough items for the quota tree deletion in
1234          * btrfs_clean_quota_tree but this is not done.
1235          *
1236          * Also, we must always start a transaction without holding the mutex
1237          * qgroup_ioctl_lock, see btrfs_quota_enable().
1238          */
1239         trans = btrfs_start_transaction(fs_info->tree_root, 1);
1240
1241         mutex_lock(&fs_info->qgroup_ioctl_lock);
1242         if (IS_ERR(trans)) {
1243                 ret = PTR_ERR(trans);
1244                 trans = NULL;
1245                 set_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1246                 goto out;
1247         }
1248
1249         if (!fs_info->quota_root)
1250                 goto out;
1251
1252         spin_lock(&fs_info->qgroup_lock);
1253         quota_root = fs_info->quota_root;
1254         fs_info->quota_root = NULL;
1255         fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_ON;
1256         spin_unlock(&fs_info->qgroup_lock);
1257
1258         btrfs_free_qgroup_config(fs_info);
1259
1260         ret = btrfs_clean_quota_tree(trans, quota_root);
1261         if (ret) {
1262                 btrfs_abort_transaction(trans, ret);
1263                 goto out;
1264         }
1265
1266         ret = btrfs_del_root(trans, &quota_root->root_key);
1267         if (ret) {
1268                 btrfs_abort_transaction(trans, ret);
1269                 goto out;
1270         }
1271
1272         list_del(&quota_root->dirty_list);
1273
1274         btrfs_tree_lock(quota_root->node);
1275         btrfs_clean_tree_block(quota_root->node);
1276         btrfs_tree_unlock(quota_root->node);
1277         btrfs_free_tree_block(trans, btrfs_root_id(quota_root),
1278                               quota_root->node, 0, 1);
1279
1280         btrfs_put_root(quota_root);
1281
1282 out:
1283         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1284         if (ret && trans)
1285                 btrfs_end_transaction(trans);
1286         else if (trans)
1287                 ret = btrfs_end_transaction(trans);
1288
1289         return ret;
1290 }
1291
1292 static void qgroup_dirty(struct btrfs_fs_info *fs_info,
1293                          struct btrfs_qgroup *qgroup)
1294 {
1295         if (list_empty(&qgroup->dirty))
1296                 list_add(&qgroup->dirty, &fs_info->dirty_qgroups);
1297 }
1298
1299 /*
1300  * The easy accounting, we're updating qgroup relationship whose child qgroup
1301  * only has exclusive extents.
1302  *
1303  * In this case, all exclusive extents will also be exclusive for parent, so
1304  * excl/rfer just get added/removed.
1305  *
1306  * So is qgroup reservation space, which should also be added/removed to
1307  * parent.
1308  * Or when child tries to release reservation space, parent will underflow its
1309  * reservation (for relationship adding case).
1310  *
1311  * Caller should hold fs_info->qgroup_lock.
1312  */
1313 static int __qgroup_excl_accounting(struct btrfs_fs_info *fs_info,
1314                                     struct ulist *tmp, u64 ref_root,
1315                                     struct btrfs_qgroup *src, int sign)
1316 {
1317         struct btrfs_qgroup *qgroup;
1318         struct btrfs_qgroup_list *glist;
1319         struct ulist_node *unode;
1320         struct ulist_iterator uiter;
1321         u64 num_bytes = src->excl;
1322         int ret = 0;
1323
1324         qgroup = find_qgroup_rb(fs_info, ref_root);
1325         if (!qgroup)
1326                 goto out;
1327
1328         qgroup->rfer += sign * num_bytes;
1329         qgroup->rfer_cmpr += sign * num_bytes;
1330
1331         WARN_ON(sign < 0 && qgroup->excl < num_bytes);
1332         qgroup->excl += sign * num_bytes;
1333         qgroup->excl_cmpr += sign * num_bytes;
1334
1335         if (sign > 0)
1336                 qgroup_rsv_add_by_qgroup(fs_info, qgroup, src);
1337         else
1338                 qgroup_rsv_release_by_qgroup(fs_info, qgroup, src);
1339
1340         qgroup_dirty(fs_info, qgroup);
1341
1342         /* Get all of the parent groups that contain this qgroup */
1343         list_for_each_entry(glist, &qgroup->groups, next_group) {
1344                 ret = ulist_add(tmp, glist->group->qgroupid,
1345                                 qgroup_to_aux(glist->group), GFP_ATOMIC);
1346                 if (ret < 0)
1347                         goto out;
1348         }
1349
1350         /* Iterate all of the parents and adjust their reference counts */
1351         ULIST_ITER_INIT(&uiter);
1352         while ((unode = ulist_next(tmp, &uiter))) {
1353                 qgroup = unode_aux_to_qgroup(unode);
1354                 qgroup->rfer += sign * num_bytes;
1355                 qgroup->rfer_cmpr += sign * num_bytes;
1356                 WARN_ON(sign < 0 && qgroup->excl < num_bytes);
1357                 qgroup->excl += sign * num_bytes;
1358                 if (sign > 0)
1359                         qgroup_rsv_add_by_qgroup(fs_info, qgroup, src);
1360                 else
1361                         qgroup_rsv_release_by_qgroup(fs_info, qgroup, src);
1362                 qgroup->excl_cmpr += sign * num_bytes;
1363                 qgroup_dirty(fs_info, qgroup);
1364
1365                 /* Add any parents of the parents */
1366                 list_for_each_entry(glist, &qgroup->groups, next_group) {
1367                         ret = ulist_add(tmp, glist->group->qgroupid,
1368                                         qgroup_to_aux(glist->group), GFP_ATOMIC);
1369                         if (ret < 0)
1370                                 goto out;
1371                 }
1372         }
1373         ret = 0;
1374 out:
1375         return ret;
1376 }
1377
1378
1379 /*
1380  * Quick path for updating qgroup with only excl refs.
1381  *
1382  * In that case, just update all parent will be enough.
1383  * Or we needs to do a full rescan.
1384  * Caller should also hold fs_info->qgroup_lock.
1385  *
1386  * Return 0 for quick update, return >0 for need to full rescan
1387  * and mark INCONSISTENT flag.
1388  * Return < 0 for other error.
1389  */
1390 static int quick_update_accounting(struct btrfs_fs_info *fs_info,
1391                                    struct ulist *tmp, u64 src, u64 dst,
1392                                    int sign)
1393 {
1394         struct btrfs_qgroup *qgroup;
1395         int ret = 1;
1396         int err = 0;
1397
1398         qgroup = find_qgroup_rb(fs_info, src);
1399         if (!qgroup)
1400                 goto out;
1401         if (qgroup->excl == qgroup->rfer) {
1402                 ret = 0;
1403                 err = __qgroup_excl_accounting(fs_info, tmp, dst,
1404                                                qgroup, sign);
1405                 if (err < 0) {
1406                         ret = err;
1407                         goto out;
1408                 }
1409         }
1410 out:
1411         if (ret)
1412                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1413         return ret;
1414 }
1415
1416 int btrfs_add_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1417                               u64 dst)
1418 {
1419         struct btrfs_fs_info *fs_info = trans->fs_info;
1420         struct btrfs_qgroup *parent;
1421         struct btrfs_qgroup *member;
1422         struct btrfs_qgroup_list *list;
1423         struct ulist *tmp;
1424         unsigned int nofs_flag;
1425         int ret = 0;
1426
1427         /* Check the level of src and dst first */
1428         if (btrfs_qgroup_level(src) >= btrfs_qgroup_level(dst))
1429                 return -EINVAL;
1430
1431         /* We hold a transaction handle open, must do a NOFS allocation. */
1432         nofs_flag = memalloc_nofs_save();
1433         tmp = ulist_alloc(GFP_KERNEL);
1434         memalloc_nofs_restore(nofs_flag);
1435         if (!tmp)
1436                 return -ENOMEM;
1437
1438         mutex_lock(&fs_info->qgroup_ioctl_lock);
1439         if (!fs_info->quota_root) {
1440                 ret = -ENOTCONN;
1441                 goto out;
1442         }
1443         member = find_qgroup_rb(fs_info, src);
1444         parent = find_qgroup_rb(fs_info, dst);
1445         if (!member || !parent) {
1446                 ret = -EINVAL;
1447                 goto out;
1448         }
1449
1450         /* check if such qgroup relation exist firstly */
1451         list_for_each_entry(list, &member->groups, next_group) {
1452                 if (list->group == parent) {
1453                         ret = -EEXIST;
1454                         goto out;
1455                 }
1456         }
1457
1458         ret = add_qgroup_relation_item(trans, src, dst);
1459         if (ret)
1460                 goto out;
1461
1462         ret = add_qgroup_relation_item(trans, dst, src);
1463         if (ret) {
1464                 del_qgroup_relation_item(trans, src, dst);
1465                 goto out;
1466         }
1467
1468         spin_lock(&fs_info->qgroup_lock);
1469         ret = add_relation_rb(fs_info, src, dst);
1470         if (ret < 0) {
1471                 spin_unlock(&fs_info->qgroup_lock);
1472                 goto out;
1473         }
1474         ret = quick_update_accounting(fs_info, tmp, src, dst, 1);
1475         spin_unlock(&fs_info->qgroup_lock);
1476 out:
1477         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1478         ulist_free(tmp);
1479         return ret;
1480 }
1481
1482 static int __del_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1483                                  u64 dst)
1484 {
1485         struct btrfs_fs_info *fs_info = trans->fs_info;
1486         struct btrfs_qgroup *parent;
1487         struct btrfs_qgroup *member;
1488         struct btrfs_qgroup_list *list;
1489         struct ulist *tmp;
1490         bool found = false;
1491         unsigned int nofs_flag;
1492         int ret = 0;
1493         int ret2;
1494
1495         /* We hold a transaction handle open, must do a NOFS allocation. */
1496         nofs_flag = memalloc_nofs_save();
1497         tmp = ulist_alloc(GFP_KERNEL);
1498         memalloc_nofs_restore(nofs_flag);
1499         if (!tmp)
1500                 return -ENOMEM;
1501
1502         if (!fs_info->quota_root) {
1503                 ret = -ENOTCONN;
1504                 goto out;
1505         }
1506
1507         member = find_qgroup_rb(fs_info, src);
1508         parent = find_qgroup_rb(fs_info, dst);
1509         /*
1510          * The parent/member pair doesn't exist, then try to delete the dead
1511          * relation items only.
1512          */
1513         if (!member || !parent)
1514                 goto delete_item;
1515
1516         /* check if such qgroup relation exist firstly */
1517         list_for_each_entry(list, &member->groups, next_group) {
1518                 if (list->group == parent) {
1519                         found = true;
1520                         break;
1521                 }
1522         }
1523
1524 delete_item:
1525         ret = del_qgroup_relation_item(trans, src, dst);
1526         if (ret < 0 && ret != -ENOENT)
1527                 goto out;
1528         ret2 = del_qgroup_relation_item(trans, dst, src);
1529         if (ret2 < 0 && ret2 != -ENOENT)
1530                 goto out;
1531
1532         /* At least one deletion succeeded, return 0 */
1533         if (!ret || !ret2)
1534                 ret = 0;
1535
1536         if (found) {
1537                 spin_lock(&fs_info->qgroup_lock);
1538                 del_relation_rb(fs_info, src, dst);
1539                 ret = quick_update_accounting(fs_info, tmp, src, dst, -1);
1540                 spin_unlock(&fs_info->qgroup_lock);
1541         }
1542 out:
1543         ulist_free(tmp);
1544         return ret;
1545 }
1546
1547 int btrfs_del_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1548                               u64 dst)
1549 {
1550         struct btrfs_fs_info *fs_info = trans->fs_info;
1551         int ret = 0;
1552
1553         mutex_lock(&fs_info->qgroup_ioctl_lock);
1554         ret = __del_qgroup_relation(trans, src, dst);
1555         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1556
1557         return ret;
1558 }
1559
1560 int btrfs_create_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid)
1561 {
1562         struct btrfs_fs_info *fs_info = trans->fs_info;
1563         struct btrfs_root *quota_root;
1564         struct btrfs_qgroup *qgroup;
1565         int ret = 0;
1566
1567         mutex_lock(&fs_info->qgroup_ioctl_lock);
1568         if (!fs_info->quota_root) {
1569                 ret = -ENOTCONN;
1570                 goto out;
1571         }
1572         quota_root = fs_info->quota_root;
1573         qgroup = find_qgroup_rb(fs_info, qgroupid);
1574         if (qgroup) {
1575                 ret = -EEXIST;
1576                 goto out;
1577         }
1578
1579         ret = add_qgroup_item(trans, quota_root, qgroupid);
1580         if (ret)
1581                 goto out;
1582
1583         spin_lock(&fs_info->qgroup_lock);
1584         qgroup = add_qgroup_rb(fs_info, qgroupid);
1585         spin_unlock(&fs_info->qgroup_lock);
1586
1587         if (IS_ERR(qgroup)) {
1588                 ret = PTR_ERR(qgroup);
1589                 goto out;
1590         }
1591         ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1592 out:
1593         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1594         return ret;
1595 }
1596
1597 int btrfs_remove_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid)
1598 {
1599         struct btrfs_fs_info *fs_info = trans->fs_info;
1600         struct btrfs_qgroup *qgroup;
1601         struct btrfs_qgroup_list *list;
1602         int ret = 0;
1603
1604         mutex_lock(&fs_info->qgroup_ioctl_lock);
1605         if (!fs_info->quota_root) {
1606                 ret = -ENOTCONN;
1607                 goto out;
1608         }
1609
1610         qgroup = find_qgroup_rb(fs_info, qgroupid);
1611         if (!qgroup) {
1612                 ret = -ENOENT;
1613                 goto out;
1614         }
1615
1616         /* Check if there are no children of this qgroup */
1617         if (!list_empty(&qgroup->members)) {
1618                 ret = -EBUSY;
1619                 goto out;
1620         }
1621
1622         ret = del_qgroup_item(trans, qgroupid);
1623         if (ret && ret != -ENOENT)
1624                 goto out;
1625
1626         while (!list_empty(&qgroup->groups)) {
1627                 list = list_first_entry(&qgroup->groups,
1628                                         struct btrfs_qgroup_list, next_group);
1629                 ret = __del_qgroup_relation(trans, qgroupid,
1630                                             list->group->qgroupid);
1631                 if (ret)
1632                         goto out;
1633         }
1634
1635         spin_lock(&fs_info->qgroup_lock);
1636         del_qgroup_rb(fs_info, qgroupid);
1637         spin_unlock(&fs_info->qgroup_lock);
1638
1639         /*
1640          * Remove the qgroup from sysfs now without holding the qgroup_lock
1641          * spinlock, since the sysfs_remove_group() function needs to take
1642          * the mutex kernfs_mutex through kernfs_remove_by_name_ns().
1643          */
1644         btrfs_sysfs_del_one_qgroup(fs_info, qgroup);
1645         kfree(qgroup);
1646 out:
1647         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1648         return ret;
1649 }
1650
1651 int btrfs_limit_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid,
1652                        struct btrfs_qgroup_limit *limit)
1653 {
1654         struct btrfs_fs_info *fs_info = trans->fs_info;
1655         struct btrfs_qgroup *qgroup;
1656         int ret = 0;
1657         /* Sometimes we would want to clear the limit on this qgroup.
1658          * To meet this requirement, we treat the -1 as a special value
1659          * which tell kernel to clear the limit on this qgroup.
1660          */
1661         const u64 CLEAR_VALUE = -1;
1662
1663         mutex_lock(&fs_info->qgroup_ioctl_lock);
1664         if (!fs_info->quota_root) {
1665                 ret = -ENOTCONN;
1666                 goto out;
1667         }
1668
1669         qgroup = find_qgroup_rb(fs_info, qgroupid);
1670         if (!qgroup) {
1671                 ret = -ENOENT;
1672                 goto out;
1673         }
1674
1675         spin_lock(&fs_info->qgroup_lock);
1676         if (limit->flags & BTRFS_QGROUP_LIMIT_MAX_RFER) {
1677                 if (limit->max_rfer == CLEAR_VALUE) {
1678                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_MAX_RFER;
1679                         limit->flags &= ~BTRFS_QGROUP_LIMIT_MAX_RFER;
1680                         qgroup->max_rfer = 0;
1681                 } else {
1682                         qgroup->max_rfer = limit->max_rfer;
1683                 }
1684         }
1685         if (limit->flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) {
1686                 if (limit->max_excl == CLEAR_VALUE) {
1687                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_MAX_EXCL;
1688                         limit->flags &= ~BTRFS_QGROUP_LIMIT_MAX_EXCL;
1689                         qgroup->max_excl = 0;
1690                 } else {
1691                         qgroup->max_excl = limit->max_excl;
1692                 }
1693         }
1694         if (limit->flags & BTRFS_QGROUP_LIMIT_RSV_RFER) {
1695                 if (limit->rsv_rfer == CLEAR_VALUE) {
1696                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_RSV_RFER;
1697                         limit->flags &= ~BTRFS_QGROUP_LIMIT_RSV_RFER;
1698                         qgroup->rsv_rfer = 0;
1699                 } else {
1700                         qgroup->rsv_rfer = limit->rsv_rfer;
1701                 }
1702         }
1703         if (limit->flags & BTRFS_QGROUP_LIMIT_RSV_EXCL) {
1704                 if (limit->rsv_excl == CLEAR_VALUE) {
1705                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_RSV_EXCL;
1706                         limit->flags &= ~BTRFS_QGROUP_LIMIT_RSV_EXCL;
1707                         qgroup->rsv_excl = 0;
1708                 } else {
1709                         qgroup->rsv_excl = limit->rsv_excl;
1710                 }
1711         }
1712         qgroup->lim_flags |= limit->flags;
1713
1714         spin_unlock(&fs_info->qgroup_lock);
1715
1716         ret = update_qgroup_limit_item(trans, qgroup);
1717         if (ret) {
1718                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1719                 btrfs_info(fs_info, "unable to update quota limit for %llu",
1720                        qgroupid);
1721         }
1722
1723 out:
1724         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1725         return ret;
1726 }
1727
1728 int btrfs_qgroup_trace_extent_nolock(struct btrfs_fs_info *fs_info,
1729                                 struct btrfs_delayed_ref_root *delayed_refs,
1730                                 struct btrfs_qgroup_extent_record *record)
1731 {
1732         struct rb_node **p = &delayed_refs->dirty_extent_root.rb_node;
1733         struct rb_node *parent_node = NULL;
1734         struct btrfs_qgroup_extent_record *entry;
1735         u64 bytenr = record->bytenr;
1736
1737         lockdep_assert_held(&delayed_refs->lock);
1738         trace_btrfs_qgroup_trace_extent(fs_info, record);
1739
1740         while (*p) {
1741                 parent_node = *p;
1742                 entry = rb_entry(parent_node, struct btrfs_qgroup_extent_record,
1743                                  node);
1744                 if (bytenr < entry->bytenr) {
1745                         p = &(*p)->rb_left;
1746                 } else if (bytenr > entry->bytenr) {
1747                         p = &(*p)->rb_right;
1748                 } else {
1749                         if (record->data_rsv && !entry->data_rsv) {
1750                                 entry->data_rsv = record->data_rsv;
1751                                 entry->data_rsv_refroot =
1752                                         record->data_rsv_refroot;
1753                         }
1754                         return 1;
1755                 }
1756         }
1757
1758         rb_link_node(&record->node, parent_node, p);
1759         rb_insert_color(&record->node, &delayed_refs->dirty_extent_root);
1760         return 0;
1761 }
1762
1763 int btrfs_qgroup_trace_extent_post(struct btrfs_trans_handle *trans,
1764                                    struct btrfs_qgroup_extent_record *qrecord)
1765 {
1766         struct ulist *old_root;
1767         u64 bytenr = qrecord->bytenr;
1768         int ret;
1769
1770         /*
1771          * We are always called in a context where we are already holding a
1772          * transaction handle. Often we are called when adding a data delayed
1773          * reference from btrfs_truncate_inode_items() (truncating or unlinking),
1774          * in which case we will be holding a write lock on extent buffer from a
1775          * subvolume tree. In this case we can't allow btrfs_find_all_roots() to
1776          * acquire fs_info->commit_root_sem, because that is a higher level lock
1777          * that must be acquired before locking any extent buffers.
1778          *
1779          * So we want btrfs_find_all_roots() to not acquire the commit_root_sem
1780          * but we can't pass it a non-NULL transaction handle, because otherwise
1781          * it would not use commit roots and would lock extent buffers, causing
1782          * a deadlock if it ends up trying to read lock the same extent buffer
1783          * that was previously write locked at btrfs_truncate_inode_items().
1784          *
1785          * So pass a NULL transaction handle to btrfs_find_all_roots() and
1786          * explicitly tell it to not acquire the commit_root_sem - if we are
1787          * holding a transaction handle we don't need its protection.
1788          */
1789         ASSERT(trans != NULL);
1790
1791         ret = btrfs_find_all_roots(NULL, trans->fs_info, bytenr, 0, &old_root,
1792                                    true);
1793         if (ret < 0) {
1794                 trans->fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1795                 btrfs_warn(trans->fs_info,
1796 "error accounting new delayed refs extent (err code: %d), quota inconsistent",
1797                         ret);
1798                 return 0;
1799         }
1800
1801         /*
1802          * Here we don't need to get the lock of
1803          * trans->transaction->delayed_refs, since inserted qrecord won't
1804          * be deleted, only qrecord->node may be modified (new qrecord insert)
1805          *
1806          * So modifying qrecord->old_roots is safe here
1807          */
1808         qrecord->old_roots = old_root;
1809         return 0;
1810 }
1811
1812 int btrfs_qgroup_trace_extent(struct btrfs_trans_handle *trans, u64 bytenr,
1813                               u64 num_bytes, gfp_t gfp_flag)
1814 {
1815         struct btrfs_fs_info *fs_info = trans->fs_info;
1816         struct btrfs_qgroup_extent_record *record;
1817         struct btrfs_delayed_ref_root *delayed_refs;
1818         int ret;
1819
1820         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)
1821             || bytenr == 0 || num_bytes == 0)
1822                 return 0;
1823         record = kzalloc(sizeof(*record), gfp_flag);
1824         if (!record)
1825                 return -ENOMEM;
1826
1827         delayed_refs = &trans->transaction->delayed_refs;
1828         record->bytenr = bytenr;
1829         record->num_bytes = num_bytes;
1830         record->old_roots = NULL;
1831
1832         spin_lock(&delayed_refs->lock);
1833         ret = btrfs_qgroup_trace_extent_nolock(fs_info, delayed_refs, record);
1834         spin_unlock(&delayed_refs->lock);
1835         if (ret > 0) {
1836                 kfree(record);
1837                 return 0;
1838         }
1839         return btrfs_qgroup_trace_extent_post(trans, record);
1840 }
1841
1842 int btrfs_qgroup_trace_leaf_items(struct btrfs_trans_handle *trans,
1843                                   struct extent_buffer *eb)
1844 {
1845         struct btrfs_fs_info *fs_info = trans->fs_info;
1846         int nr = btrfs_header_nritems(eb);
1847         int i, extent_type, ret;
1848         struct btrfs_key key;
1849         struct btrfs_file_extent_item *fi;
1850         u64 bytenr, num_bytes;
1851
1852         /* We can be called directly from walk_up_proc() */
1853         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
1854                 return 0;
1855
1856         for (i = 0; i < nr; i++) {
1857                 btrfs_item_key_to_cpu(eb, &key, i);
1858
1859                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1860                         continue;
1861
1862                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
1863                 /* filter out non qgroup-accountable extents  */
1864                 extent_type = btrfs_file_extent_type(eb, fi);
1865
1866                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
1867                         continue;
1868
1869                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
1870                 if (!bytenr)
1871                         continue;
1872
1873                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
1874
1875                 ret = btrfs_qgroup_trace_extent(trans, bytenr, num_bytes,
1876                                                 GFP_NOFS);
1877                 if (ret)
1878                         return ret;
1879         }
1880         cond_resched();
1881         return 0;
1882 }
1883
1884 /*
1885  * Walk up the tree from the bottom, freeing leaves and any interior
1886  * nodes which have had all slots visited. If a node (leaf or
1887  * interior) is freed, the node above it will have it's slot
1888  * incremented. The root node will never be freed.
1889  *
1890  * At the end of this function, we should have a path which has all
1891  * slots incremented to the next position for a search. If we need to
1892  * read a new node it will be NULL and the node above it will have the
1893  * correct slot selected for a later read.
1894  *
1895  * If we increment the root nodes slot counter past the number of
1896  * elements, 1 is returned to signal completion of the search.
1897  */
1898 static int adjust_slots_upwards(struct btrfs_path *path, int root_level)
1899 {
1900         int level = 0;
1901         int nr, slot;
1902         struct extent_buffer *eb;
1903
1904         if (root_level == 0)
1905                 return 1;
1906
1907         while (level <= root_level) {
1908                 eb = path->nodes[level];
1909                 nr = btrfs_header_nritems(eb);
1910                 path->slots[level]++;
1911                 slot = path->slots[level];
1912                 if (slot >= nr || level == 0) {
1913                         /*
1914                          * Don't free the root -  we will detect this
1915                          * condition after our loop and return a
1916                          * positive value for caller to stop walking the tree.
1917                          */
1918                         if (level != root_level) {
1919                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
1920                                 path->locks[level] = 0;
1921
1922                                 free_extent_buffer(eb);
1923                                 path->nodes[level] = NULL;
1924                                 path->slots[level] = 0;
1925                         }
1926                 } else {
1927                         /*
1928                          * We have a valid slot to walk back down
1929                          * from. Stop here so caller can process these
1930                          * new nodes.
1931                          */
1932                         break;
1933                 }
1934
1935                 level++;
1936         }
1937
1938         eb = path->nodes[root_level];
1939         if (path->slots[root_level] >= btrfs_header_nritems(eb))
1940                 return 1;
1941
1942         return 0;
1943 }
1944
1945 /*
1946  * Helper function to trace a subtree tree block swap.
1947  *
1948  * The swap will happen in highest tree block, but there may be a lot of
1949  * tree blocks involved.
1950  *
1951  * For example:
1952  *  OO = Old tree blocks
1953  *  NN = New tree blocks allocated during balance
1954  *
1955  *           File tree (257)                  Reloc tree for 257
1956  * L2              OO                                NN
1957  *               /    \                            /    \
1958  * L1          OO      OO (a)                    OO      NN (a)
1959  *            / \     / \                       / \     / \
1960  * L0       OO   OO OO   OO                   OO   OO NN   NN
1961  *                  (b)  (c)                          (b)  (c)
1962  *
1963  * When calling qgroup_trace_extent_swap(), we will pass:
1964  * @src_eb = OO(a)
1965  * @dst_path = [ nodes[1] = NN(a), nodes[0] = NN(c) ]
1966  * @dst_level = 0
1967  * @root_level = 1
1968  *
1969  * In that case, qgroup_trace_extent_swap() will search from OO(a) to
1970  * reach OO(c), then mark both OO(c) and NN(c) as qgroup dirty.
1971  *
1972  * The main work of qgroup_trace_extent_swap() can be split into 3 parts:
1973  *
1974  * 1) Tree search from @src_eb
1975  *    It should acts as a simplified btrfs_search_slot().
1976  *    The key for search can be extracted from @dst_path->nodes[dst_level]
1977  *    (first key).
1978  *
1979  * 2) Mark the final tree blocks in @src_path and @dst_path qgroup dirty
1980  *    NOTE: In above case, OO(a) and NN(a) won't be marked qgroup dirty.
1981  *    They should be marked during previous (@dst_level = 1) iteration.
1982  *
1983  * 3) Mark file extents in leaves dirty
1984  *    We don't have good way to pick out new file extents only.
1985  *    So we still follow the old method by scanning all file extents in
1986  *    the leave.
1987  *
1988  * This function can free us from keeping two paths, thus later we only need
1989  * to care about how to iterate all new tree blocks in reloc tree.
1990  */
1991 static int qgroup_trace_extent_swap(struct btrfs_trans_handle* trans,
1992                                     struct extent_buffer *src_eb,
1993                                     struct btrfs_path *dst_path,
1994                                     int dst_level, int root_level,
1995                                     bool trace_leaf)
1996 {
1997         struct btrfs_key key;
1998         struct btrfs_path *src_path;
1999         struct btrfs_fs_info *fs_info = trans->fs_info;
2000         u32 nodesize = fs_info->nodesize;
2001         int cur_level = root_level;
2002         int ret;
2003
2004         BUG_ON(dst_level > root_level);
2005         /* Level mismatch */
2006         if (btrfs_header_level(src_eb) != root_level)
2007                 return -EINVAL;
2008
2009         src_path = btrfs_alloc_path();
2010         if (!src_path) {
2011                 ret = -ENOMEM;
2012                 goto out;
2013         }
2014
2015         if (dst_level)
2016                 btrfs_node_key_to_cpu(dst_path->nodes[dst_level], &key, 0);
2017         else
2018                 btrfs_item_key_to_cpu(dst_path->nodes[dst_level], &key, 0);
2019
2020         /* For src_path */
2021         atomic_inc(&src_eb->refs);
2022         src_path->nodes[root_level] = src_eb;
2023         src_path->slots[root_level] = dst_path->slots[root_level];
2024         src_path->locks[root_level] = 0;
2025
2026         /* A simplified version of btrfs_search_slot() */
2027         while (cur_level >= dst_level) {
2028                 struct btrfs_key src_key;
2029                 struct btrfs_key dst_key;
2030
2031                 if (src_path->nodes[cur_level] == NULL) {
2032                         struct extent_buffer *eb;
2033                         int parent_slot;
2034
2035                         eb = src_path->nodes[cur_level + 1];
2036                         parent_slot = src_path->slots[cur_level + 1];
2037
2038                         eb = btrfs_read_node_slot(eb, parent_slot);
2039                         if (IS_ERR(eb)) {
2040                                 ret = PTR_ERR(eb);
2041                                 goto out;
2042                         }
2043
2044                         src_path->nodes[cur_level] = eb;
2045
2046                         btrfs_tree_read_lock(eb);
2047                         src_path->locks[cur_level] = BTRFS_READ_LOCK;
2048                 }
2049
2050                 src_path->slots[cur_level] = dst_path->slots[cur_level];
2051                 if (cur_level) {
2052                         btrfs_node_key_to_cpu(dst_path->nodes[cur_level],
2053                                         &dst_key, dst_path->slots[cur_level]);
2054                         btrfs_node_key_to_cpu(src_path->nodes[cur_level],
2055                                         &src_key, src_path->slots[cur_level]);
2056                 } else {
2057                         btrfs_item_key_to_cpu(dst_path->nodes[cur_level],
2058                                         &dst_key, dst_path->slots[cur_level]);
2059                         btrfs_item_key_to_cpu(src_path->nodes[cur_level],
2060                                         &src_key, src_path->slots[cur_level]);
2061                 }
2062                 /* Content mismatch, something went wrong */
2063                 if (btrfs_comp_cpu_keys(&dst_key, &src_key)) {
2064                         ret = -ENOENT;
2065                         goto out;
2066                 }
2067                 cur_level--;
2068         }
2069
2070         /*
2071          * Now both @dst_path and @src_path have been populated, record the tree
2072          * blocks for qgroup accounting.
2073          */
2074         ret = btrfs_qgroup_trace_extent(trans, src_path->nodes[dst_level]->start,
2075                         nodesize, GFP_NOFS);
2076         if (ret < 0)
2077                 goto out;
2078         ret = btrfs_qgroup_trace_extent(trans,
2079                         dst_path->nodes[dst_level]->start,
2080                         nodesize, GFP_NOFS);
2081         if (ret < 0)
2082                 goto out;
2083
2084         /* Record leaf file extents */
2085         if (dst_level == 0 && trace_leaf) {
2086                 ret = btrfs_qgroup_trace_leaf_items(trans, src_path->nodes[0]);
2087                 if (ret < 0)
2088                         goto out;
2089                 ret = btrfs_qgroup_trace_leaf_items(trans, dst_path->nodes[0]);
2090         }
2091 out:
2092         btrfs_free_path(src_path);
2093         return ret;
2094 }
2095
2096 /*
2097  * Helper function to do recursive generation-aware depth-first search, to
2098  * locate all new tree blocks in a subtree of reloc tree.
2099  *
2100  * E.g. (OO = Old tree blocks, NN = New tree blocks, whose gen == last_snapshot)
2101  *         reloc tree
2102  * L2         NN (a)
2103  *          /    \
2104  * L1    OO        NN (b)
2105  *      /  \      /  \
2106  * L0  OO  OO    OO  NN
2107  *               (c) (d)
2108  * If we pass:
2109  * @dst_path = [ nodes[1] = NN(b), nodes[0] = NULL ],
2110  * @cur_level = 1
2111  * @root_level = 1
2112  *
2113  * We will iterate through tree blocks NN(b), NN(d) and info qgroup to trace
2114  * above tree blocks along with their counter parts in file tree.
2115  * While during search, old tree blocks OO(c) will be skipped as tree block swap
2116  * won't affect OO(c).
2117  */
2118 static int qgroup_trace_new_subtree_blocks(struct btrfs_trans_handle* trans,
2119                                            struct extent_buffer *src_eb,
2120                                            struct btrfs_path *dst_path,
2121                                            int cur_level, int root_level,
2122                                            u64 last_snapshot, bool trace_leaf)
2123 {
2124         struct btrfs_fs_info *fs_info = trans->fs_info;
2125         struct extent_buffer *eb;
2126         bool need_cleanup = false;
2127         int ret = 0;
2128         int i;
2129
2130         /* Level sanity check */
2131         if (cur_level < 0 || cur_level >= BTRFS_MAX_LEVEL - 1 ||
2132             root_level < 0 || root_level >= BTRFS_MAX_LEVEL - 1 ||
2133             root_level < cur_level) {
2134                 btrfs_err_rl(fs_info,
2135                         "%s: bad levels, cur_level=%d root_level=%d",
2136                         __func__, cur_level, root_level);
2137                 return -EUCLEAN;
2138         }
2139
2140         /* Read the tree block if needed */
2141         if (dst_path->nodes[cur_level] == NULL) {
2142                 int parent_slot;
2143                 u64 child_gen;
2144
2145                 /*
2146                  * dst_path->nodes[root_level] must be initialized before
2147                  * calling this function.
2148                  */
2149                 if (cur_level == root_level) {
2150                         btrfs_err_rl(fs_info,
2151         "%s: dst_path->nodes[%d] not initialized, root_level=%d cur_level=%d",
2152                                 __func__, root_level, root_level, cur_level);
2153                         return -EUCLEAN;
2154                 }
2155
2156                 /*
2157                  * We need to get child blockptr/gen from parent before we can
2158                  * read it.
2159                   */
2160                 eb = dst_path->nodes[cur_level + 1];
2161                 parent_slot = dst_path->slots[cur_level + 1];
2162                 child_gen = btrfs_node_ptr_generation(eb, parent_slot);
2163
2164                 /* This node is old, no need to trace */
2165                 if (child_gen < last_snapshot)
2166                         goto out;
2167
2168                 eb = btrfs_read_node_slot(eb, parent_slot);
2169                 if (IS_ERR(eb)) {
2170                         ret = PTR_ERR(eb);
2171                         goto out;
2172                 }
2173
2174                 dst_path->nodes[cur_level] = eb;
2175                 dst_path->slots[cur_level] = 0;
2176
2177                 btrfs_tree_read_lock(eb);
2178                 dst_path->locks[cur_level] = BTRFS_READ_LOCK;
2179                 need_cleanup = true;
2180         }
2181
2182         /* Now record this tree block and its counter part for qgroups */
2183         ret = qgroup_trace_extent_swap(trans, src_eb, dst_path, cur_level,
2184                                        root_level, trace_leaf);
2185         if (ret < 0)
2186                 goto cleanup;
2187
2188         eb = dst_path->nodes[cur_level];
2189
2190         if (cur_level > 0) {
2191                 /* Iterate all child tree blocks */
2192                 for (i = 0; i < btrfs_header_nritems(eb); i++) {
2193                         /* Skip old tree blocks as they won't be swapped */
2194                         if (btrfs_node_ptr_generation(eb, i) < last_snapshot)
2195                                 continue;
2196                         dst_path->slots[cur_level] = i;
2197
2198                         /* Recursive call (at most 7 times) */
2199                         ret = qgroup_trace_new_subtree_blocks(trans, src_eb,
2200                                         dst_path, cur_level - 1, root_level,
2201                                         last_snapshot, trace_leaf);
2202                         if (ret < 0)
2203                                 goto cleanup;
2204                 }
2205         }
2206
2207 cleanup:
2208         if (need_cleanup) {
2209                 /* Clean up */
2210                 btrfs_tree_unlock_rw(dst_path->nodes[cur_level],
2211                                      dst_path->locks[cur_level]);
2212                 free_extent_buffer(dst_path->nodes[cur_level]);
2213                 dst_path->nodes[cur_level] = NULL;
2214                 dst_path->slots[cur_level] = 0;
2215                 dst_path->locks[cur_level] = 0;
2216         }
2217 out:
2218         return ret;
2219 }
2220
2221 static int qgroup_trace_subtree_swap(struct btrfs_trans_handle *trans,
2222                                 struct extent_buffer *src_eb,
2223                                 struct extent_buffer *dst_eb,
2224                                 u64 last_snapshot, bool trace_leaf)
2225 {
2226         struct btrfs_fs_info *fs_info = trans->fs_info;
2227         struct btrfs_path *dst_path = NULL;
2228         int level;
2229         int ret;
2230
2231         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2232                 return 0;
2233
2234         /* Wrong parameter order */
2235         if (btrfs_header_generation(src_eb) > btrfs_header_generation(dst_eb)) {
2236                 btrfs_err_rl(fs_info,
2237                 "%s: bad parameter order, src_gen=%llu dst_gen=%llu", __func__,
2238                              btrfs_header_generation(src_eb),
2239                              btrfs_header_generation(dst_eb));
2240                 return -EUCLEAN;
2241         }
2242
2243         if (!extent_buffer_uptodate(src_eb) || !extent_buffer_uptodate(dst_eb)) {
2244                 ret = -EIO;
2245                 goto out;
2246         }
2247
2248         level = btrfs_header_level(dst_eb);
2249         dst_path = btrfs_alloc_path();
2250         if (!dst_path) {
2251                 ret = -ENOMEM;
2252                 goto out;
2253         }
2254         /* For dst_path */
2255         atomic_inc(&dst_eb->refs);
2256         dst_path->nodes[level] = dst_eb;
2257         dst_path->slots[level] = 0;
2258         dst_path->locks[level] = 0;
2259
2260         /* Do the generation aware breadth-first search */
2261         ret = qgroup_trace_new_subtree_blocks(trans, src_eb, dst_path, level,
2262                                               level, last_snapshot, trace_leaf);
2263         if (ret < 0)
2264                 goto out;
2265         ret = 0;
2266
2267 out:
2268         btrfs_free_path(dst_path);
2269         if (ret < 0)
2270                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2271         return ret;
2272 }
2273
2274 int btrfs_qgroup_trace_subtree(struct btrfs_trans_handle *trans,
2275                                struct extent_buffer *root_eb,
2276                                u64 root_gen, int root_level)
2277 {
2278         struct btrfs_fs_info *fs_info = trans->fs_info;
2279         int ret = 0;
2280         int level;
2281         struct extent_buffer *eb = root_eb;
2282         struct btrfs_path *path = NULL;
2283
2284         BUG_ON(root_level < 0 || root_level >= BTRFS_MAX_LEVEL);
2285         BUG_ON(root_eb == NULL);
2286
2287         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2288                 return 0;
2289
2290         if (!extent_buffer_uptodate(root_eb)) {
2291                 ret = btrfs_read_buffer(root_eb, root_gen, root_level, NULL);
2292                 if (ret)
2293                         goto out;
2294         }
2295
2296         if (root_level == 0) {
2297                 ret = btrfs_qgroup_trace_leaf_items(trans, root_eb);
2298                 goto out;
2299         }
2300
2301         path = btrfs_alloc_path();
2302         if (!path)
2303                 return -ENOMEM;
2304
2305         /*
2306          * Walk down the tree.  Missing extent blocks are filled in as
2307          * we go. Metadata is accounted every time we read a new
2308          * extent block.
2309          *
2310          * When we reach a leaf, we account for file extent items in it,
2311          * walk back up the tree (adjusting slot pointers as we go)
2312          * and restart the search process.
2313          */
2314         atomic_inc(&root_eb->refs);     /* For path */
2315         path->nodes[root_level] = root_eb;
2316         path->slots[root_level] = 0;
2317         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
2318 walk_down:
2319         level = root_level;
2320         while (level >= 0) {
2321                 if (path->nodes[level] == NULL) {
2322                         int parent_slot;
2323                         u64 child_bytenr;
2324
2325                         /*
2326                          * We need to get child blockptr from parent before we
2327                          * can read it.
2328                           */
2329                         eb = path->nodes[level + 1];
2330                         parent_slot = path->slots[level + 1];
2331                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
2332
2333                         eb = btrfs_read_node_slot(eb, parent_slot);
2334                         if (IS_ERR(eb)) {
2335                                 ret = PTR_ERR(eb);
2336                                 goto out;
2337                         }
2338
2339                         path->nodes[level] = eb;
2340                         path->slots[level] = 0;
2341
2342                         btrfs_tree_read_lock(eb);
2343                         path->locks[level] = BTRFS_READ_LOCK;
2344
2345                         ret = btrfs_qgroup_trace_extent(trans, child_bytenr,
2346                                                         fs_info->nodesize,
2347                                                         GFP_NOFS);
2348                         if (ret)
2349                                 goto out;
2350                 }
2351
2352                 if (level == 0) {
2353                         ret = btrfs_qgroup_trace_leaf_items(trans,
2354                                                             path->nodes[level]);
2355                         if (ret)
2356                                 goto out;
2357
2358                         /* Nonzero return here means we completed our search */
2359                         ret = adjust_slots_upwards(path, root_level);
2360                         if (ret)
2361                                 break;
2362
2363                         /* Restart search with new slots */
2364                         goto walk_down;
2365                 }
2366
2367                 level--;
2368         }
2369
2370         ret = 0;
2371 out:
2372         btrfs_free_path(path);
2373
2374         return ret;
2375 }
2376
2377 #define UPDATE_NEW      0
2378 #define UPDATE_OLD      1
2379 /*
2380  * Walk all of the roots that points to the bytenr and adjust their refcnts.
2381  */
2382 static int qgroup_update_refcnt(struct btrfs_fs_info *fs_info,
2383                                 struct ulist *roots, struct ulist *tmp,
2384                                 struct ulist *qgroups, u64 seq, int update_old)
2385 {
2386         struct ulist_node *unode;
2387         struct ulist_iterator uiter;
2388         struct ulist_node *tmp_unode;
2389         struct ulist_iterator tmp_uiter;
2390         struct btrfs_qgroup *qg;
2391         int ret = 0;
2392
2393         if (!roots)
2394                 return 0;
2395         ULIST_ITER_INIT(&uiter);
2396         while ((unode = ulist_next(roots, &uiter))) {
2397                 qg = find_qgroup_rb(fs_info, unode->val);
2398                 if (!qg)
2399                         continue;
2400
2401                 ulist_reinit(tmp);
2402                 ret = ulist_add(qgroups, qg->qgroupid, qgroup_to_aux(qg),
2403                                 GFP_ATOMIC);
2404                 if (ret < 0)
2405                         return ret;
2406                 ret = ulist_add(tmp, qg->qgroupid, qgroup_to_aux(qg), GFP_ATOMIC);
2407                 if (ret < 0)
2408                         return ret;
2409                 ULIST_ITER_INIT(&tmp_uiter);
2410                 while ((tmp_unode = ulist_next(tmp, &tmp_uiter))) {
2411                         struct btrfs_qgroup_list *glist;
2412
2413                         qg = unode_aux_to_qgroup(tmp_unode);
2414                         if (update_old)
2415                                 btrfs_qgroup_update_old_refcnt(qg, seq, 1);
2416                         else
2417                                 btrfs_qgroup_update_new_refcnt(qg, seq, 1);
2418                         list_for_each_entry(glist, &qg->groups, next_group) {
2419                                 ret = ulist_add(qgroups, glist->group->qgroupid,
2420                                                 qgroup_to_aux(glist->group),
2421                                                 GFP_ATOMIC);
2422                                 if (ret < 0)
2423                                         return ret;
2424                                 ret = ulist_add(tmp, glist->group->qgroupid,
2425                                                 qgroup_to_aux(glist->group),
2426                                                 GFP_ATOMIC);
2427                                 if (ret < 0)
2428                                         return ret;
2429                         }
2430                 }
2431         }
2432         return 0;
2433 }
2434
2435 /*
2436  * Update qgroup rfer/excl counters.
2437  * Rfer update is easy, codes can explain themselves.
2438  *
2439  * Excl update is tricky, the update is split into 2 parts.
2440  * Part 1: Possible exclusive <-> sharing detect:
2441  *      |       A       |       !A      |
2442  *  -------------------------------------
2443  *  B   |       *       |       -       |
2444  *  -------------------------------------
2445  *  !B  |       +       |       **      |
2446  *  -------------------------------------
2447  *
2448  * Conditions:
2449  * A:   cur_old_roots < nr_old_roots    (not exclusive before)
2450  * !A:  cur_old_roots == nr_old_roots   (possible exclusive before)
2451  * B:   cur_new_roots < nr_new_roots    (not exclusive now)
2452  * !B:  cur_new_roots == nr_new_roots   (possible exclusive now)
2453  *
2454  * Results:
2455  * +: Possible sharing -> exclusive     -: Possible exclusive -> sharing
2456  * *: Definitely not changed.           **: Possible unchanged.
2457  *
2458  * For !A and !B condition, the exception is cur_old/new_roots == 0 case.
2459  *
2460  * To make the logic clear, we first use condition A and B to split
2461  * combination into 4 results.
2462  *
2463  * Then, for result "+" and "-", check old/new_roots == 0 case, as in them
2464  * only on variant maybe 0.
2465  *
2466  * Lastly, check result **, since there are 2 variants maybe 0, split them
2467  * again(2x2).
2468  * But this time we don't need to consider other things, the codes and logic
2469  * is easy to understand now.
2470  */
2471 static int qgroup_update_counters(struct btrfs_fs_info *fs_info,
2472                                   struct ulist *qgroups,
2473                                   u64 nr_old_roots,
2474                                   u64 nr_new_roots,
2475                                   u64 num_bytes, u64 seq)
2476 {
2477         struct ulist_node *unode;
2478         struct ulist_iterator uiter;
2479         struct btrfs_qgroup *qg;
2480         u64 cur_new_count, cur_old_count;
2481
2482         ULIST_ITER_INIT(&uiter);
2483         while ((unode = ulist_next(qgroups, &uiter))) {
2484                 bool dirty = false;
2485
2486                 qg = unode_aux_to_qgroup(unode);
2487                 cur_old_count = btrfs_qgroup_get_old_refcnt(qg, seq);
2488                 cur_new_count = btrfs_qgroup_get_new_refcnt(qg, seq);
2489
2490                 trace_qgroup_update_counters(fs_info, qg, cur_old_count,
2491                                              cur_new_count);
2492
2493                 /* Rfer update part */
2494                 if (cur_old_count == 0 && cur_new_count > 0) {
2495                         qg->rfer += num_bytes;
2496                         qg->rfer_cmpr += num_bytes;
2497                         dirty = true;
2498                 }
2499                 if (cur_old_count > 0 && cur_new_count == 0) {
2500                         qg->rfer -= num_bytes;
2501                         qg->rfer_cmpr -= num_bytes;
2502                         dirty = true;
2503                 }
2504
2505                 /* Excl update part */
2506                 /* Exclusive/none -> shared case */
2507                 if (cur_old_count == nr_old_roots &&
2508                     cur_new_count < nr_new_roots) {
2509                         /* Exclusive -> shared */
2510                         if (cur_old_count != 0) {
2511                                 qg->excl -= num_bytes;
2512                                 qg->excl_cmpr -= num_bytes;
2513                                 dirty = true;
2514                         }
2515                 }
2516
2517                 /* Shared -> exclusive/none case */
2518                 if (cur_old_count < nr_old_roots &&
2519                     cur_new_count == nr_new_roots) {
2520                         /* Shared->exclusive */
2521                         if (cur_new_count != 0) {
2522                                 qg->excl += num_bytes;
2523                                 qg->excl_cmpr += num_bytes;
2524                                 dirty = true;
2525                         }
2526                 }
2527
2528                 /* Exclusive/none -> exclusive/none case */
2529                 if (cur_old_count == nr_old_roots &&
2530                     cur_new_count == nr_new_roots) {
2531                         if (cur_old_count == 0) {
2532                                 /* None -> exclusive/none */
2533
2534                                 if (cur_new_count != 0) {
2535                                         /* None -> exclusive */
2536                                         qg->excl += num_bytes;
2537                                         qg->excl_cmpr += num_bytes;
2538                                         dirty = true;
2539                                 }
2540                                 /* None -> none, nothing changed */
2541                         } else {
2542                                 /* Exclusive -> exclusive/none */
2543
2544                                 if (cur_new_count == 0) {
2545                                         /* Exclusive -> none */
2546                                         qg->excl -= num_bytes;
2547                                         qg->excl_cmpr -= num_bytes;
2548                                         dirty = true;
2549                                 }
2550                                 /* Exclusive -> exclusive, nothing changed */
2551                         }
2552                 }
2553
2554                 if (dirty)
2555                         qgroup_dirty(fs_info, qg);
2556         }
2557         return 0;
2558 }
2559
2560 /*
2561  * Check if the @roots potentially is a list of fs tree roots
2562  *
2563  * Return 0 for definitely not a fs/subvol tree roots ulist
2564  * Return 1 for possible fs/subvol tree roots in the list (considering an empty
2565  *          one as well)
2566  */
2567 static int maybe_fs_roots(struct ulist *roots)
2568 {
2569         struct ulist_node *unode;
2570         struct ulist_iterator uiter;
2571
2572         /* Empty one, still possible for fs roots */
2573         if (!roots || roots->nnodes == 0)
2574                 return 1;
2575
2576         ULIST_ITER_INIT(&uiter);
2577         unode = ulist_next(roots, &uiter);
2578         if (!unode)
2579                 return 1;
2580
2581         /*
2582          * If it contains fs tree roots, then it must belong to fs/subvol
2583          * trees.
2584          * If it contains a non-fs tree, it won't be shared with fs/subvol trees.
2585          */
2586         return is_fstree(unode->val);
2587 }
2588
2589 int btrfs_qgroup_account_extent(struct btrfs_trans_handle *trans, u64 bytenr,
2590                                 u64 num_bytes, struct ulist *old_roots,
2591                                 struct ulist *new_roots)
2592 {
2593         struct btrfs_fs_info *fs_info = trans->fs_info;
2594         struct ulist *qgroups = NULL;
2595         struct ulist *tmp = NULL;
2596         u64 seq;
2597         u64 nr_new_roots = 0;
2598         u64 nr_old_roots = 0;
2599         int ret = 0;
2600
2601         /*
2602          * If quotas get disabled meanwhile, the resources need to be freed and
2603          * we can't just exit here.
2604          */
2605         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2606                 goto out_free;
2607
2608         if (new_roots) {
2609                 if (!maybe_fs_roots(new_roots))
2610                         goto out_free;
2611                 nr_new_roots = new_roots->nnodes;
2612         }
2613         if (old_roots) {
2614                 if (!maybe_fs_roots(old_roots))
2615                         goto out_free;
2616                 nr_old_roots = old_roots->nnodes;
2617         }
2618
2619         /* Quick exit, either not fs tree roots, or won't affect any qgroup */
2620         if (nr_old_roots == 0 && nr_new_roots == 0)
2621                 goto out_free;
2622
2623         BUG_ON(!fs_info->quota_root);
2624
2625         trace_btrfs_qgroup_account_extent(fs_info, trans->transid, bytenr,
2626                                         num_bytes, nr_old_roots, nr_new_roots);
2627
2628         qgroups = ulist_alloc(GFP_NOFS);
2629         if (!qgroups) {
2630                 ret = -ENOMEM;
2631                 goto out_free;
2632         }
2633         tmp = ulist_alloc(GFP_NOFS);
2634         if (!tmp) {
2635                 ret = -ENOMEM;
2636                 goto out_free;
2637         }
2638
2639         mutex_lock(&fs_info->qgroup_rescan_lock);
2640         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
2641                 if (fs_info->qgroup_rescan_progress.objectid <= bytenr) {
2642                         mutex_unlock(&fs_info->qgroup_rescan_lock);
2643                         ret = 0;
2644                         goto out_free;
2645                 }
2646         }
2647         mutex_unlock(&fs_info->qgroup_rescan_lock);
2648
2649         spin_lock(&fs_info->qgroup_lock);
2650         seq = fs_info->qgroup_seq;
2651
2652         /* Update old refcnts using old_roots */
2653         ret = qgroup_update_refcnt(fs_info, old_roots, tmp, qgroups, seq,
2654                                    UPDATE_OLD);
2655         if (ret < 0)
2656                 goto out;
2657
2658         /* Update new refcnts using new_roots */
2659         ret = qgroup_update_refcnt(fs_info, new_roots, tmp, qgroups, seq,
2660                                    UPDATE_NEW);
2661         if (ret < 0)
2662                 goto out;
2663
2664         qgroup_update_counters(fs_info, qgroups, nr_old_roots, nr_new_roots,
2665                                num_bytes, seq);
2666
2667         /*
2668          * Bump qgroup_seq to avoid seq overlap
2669          */
2670         fs_info->qgroup_seq += max(nr_old_roots, nr_new_roots) + 1;
2671 out:
2672         spin_unlock(&fs_info->qgroup_lock);
2673 out_free:
2674         ulist_free(tmp);
2675         ulist_free(qgroups);
2676         ulist_free(old_roots);
2677         ulist_free(new_roots);
2678         return ret;
2679 }
2680
2681 int btrfs_qgroup_account_extents(struct btrfs_trans_handle *trans)
2682 {
2683         struct btrfs_fs_info *fs_info = trans->fs_info;
2684         struct btrfs_qgroup_extent_record *record;
2685         struct btrfs_delayed_ref_root *delayed_refs;
2686         struct ulist *new_roots = NULL;
2687         struct rb_node *node;
2688         u64 num_dirty_extents = 0;
2689         u64 qgroup_to_skip;
2690         int ret = 0;
2691
2692         delayed_refs = &trans->transaction->delayed_refs;
2693         qgroup_to_skip = delayed_refs->qgroup_to_skip;
2694         while ((node = rb_first(&delayed_refs->dirty_extent_root))) {
2695                 record = rb_entry(node, struct btrfs_qgroup_extent_record,
2696                                   node);
2697
2698                 num_dirty_extents++;
2699                 trace_btrfs_qgroup_account_extents(fs_info, record);
2700
2701                 if (!ret) {
2702                         /*
2703                          * Old roots should be searched when inserting qgroup
2704                          * extent record
2705                          */
2706                         if (WARN_ON(!record->old_roots)) {
2707                                 /* Search commit root to find old_roots */
2708                                 ret = btrfs_find_all_roots(NULL, fs_info,
2709                                                 record->bytenr, 0,
2710                                                 &record->old_roots, false);
2711                                 if (ret < 0)
2712                                         goto cleanup;
2713                         }
2714
2715                         /* Free the reserved data space */
2716                         btrfs_qgroup_free_refroot(fs_info,
2717                                         record->data_rsv_refroot,
2718                                         record->data_rsv,
2719                                         BTRFS_QGROUP_RSV_DATA);
2720                         /*
2721                          * Use BTRFS_SEQ_LAST as time_seq to do special search,
2722                          * which doesn't lock tree or delayed_refs and search
2723                          * current root. It's safe inside commit_transaction().
2724                          */
2725                         ret = btrfs_find_all_roots(trans, fs_info,
2726                            record->bytenr, BTRFS_SEQ_LAST, &new_roots, false);
2727                         if (ret < 0)
2728                                 goto cleanup;
2729                         if (qgroup_to_skip) {
2730                                 ulist_del(new_roots, qgroup_to_skip, 0);
2731                                 ulist_del(record->old_roots, qgroup_to_skip,
2732                                           0);
2733                         }
2734                         ret = btrfs_qgroup_account_extent(trans, record->bytenr,
2735                                                           record->num_bytes,
2736                                                           record->old_roots,
2737                                                           new_roots);
2738                         record->old_roots = NULL;
2739                         new_roots = NULL;
2740                 }
2741 cleanup:
2742                 ulist_free(record->old_roots);
2743                 ulist_free(new_roots);
2744                 new_roots = NULL;
2745                 rb_erase(node, &delayed_refs->dirty_extent_root);
2746                 kfree(record);
2747
2748         }
2749         trace_qgroup_num_dirty_extents(fs_info, trans->transid,
2750                                        num_dirty_extents);
2751         return ret;
2752 }
2753
2754 /*
2755  * Writes all changed qgroups to disk.
2756  * Called by the transaction commit path and the qgroup assign ioctl.
2757  */
2758 int btrfs_run_qgroups(struct btrfs_trans_handle *trans)
2759 {
2760         struct btrfs_fs_info *fs_info = trans->fs_info;
2761         int ret = 0;
2762
2763         /*
2764          * In case we are called from the qgroup assign ioctl, assert that we
2765          * are holding the qgroup_ioctl_lock, otherwise we can race with a quota
2766          * disable operation (ioctl) and access a freed quota root.
2767          */
2768         if (trans->transaction->state != TRANS_STATE_COMMIT_DOING)
2769                 lockdep_assert_held(&fs_info->qgroup_ioctl_lock);
2770
2771         if (!fs_info->quota_root)
2772                 return ret;
2773
2774         spin_lock(&fs_info->qgroup_lock);
2775         while (!list_empty(&fs_info->dirty_qgroups)) {
2776                 struct btrfs_qgroup *qgroup;
2777                 qgroup = list_first_entry(&fs_info->dirty_qgroups,
2778                                           struct btrfs_qgroup, dirty);
2779                 list_del_init(&qgroup->dirty);
2780                 spin_unlock(&fs_info->qgroup_lock);
2781                 ret = update_qgroup_info_item(trans, qgroup);
2782                 if (ret)
2783                         fs_info->qgroup_flags |=
2784                                         BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2785                 ret = update_qgroup_limit_item(trans, qgroup);
2786                 if (ret)
2787                         fs_info->qgroup_flags |=
2788                                         BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2789                 spin_lock(&fs_info->qgroup_lock);
2790         }
2791         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2792                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_ON;
2793         else
2794                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_ON;
2795         spin_unlock(&fs_info->qgroup_lock);
2796
2797         ret = update_qgroup_status_item(trans);
2798         if (ret)
2799                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2800
2801         return ret;
2802 }
2803
2804 /*
2805  * Copy the accounting information between qgroups. This is necessary
2806  * when a snapshot or a subvolume is created. Throwing an error will
2807  * cause a transaction abort so we take extra care here to only error
2808  * when a readonly fs is a reasonable outcome.
2809  */
2810 int btrfs_qgroup_inherit(struct btrfs_trans_handle *trans, u64 srcid,
2811                          u64 objectid, struct btrfs_qgroup_inherit *inherit)
2812 {
2813         int ret = 0;
2814         int i;
2815         u64 *i_qgroups;
2816         bool committing = false;
2817         struct btrfs_fs_info *fs_info = trans->fs_info;
2818         struct btrfs_root *quota_root;
2819         struct btrfs_qgroup *srcgroup;
2820         struct btrfs_qgroup *dstgroup;
2821         bool need_rescan = false;
2822         u32 level_size = 0;
2823         u64 nums;
2824
2825         /*
2826          * There are only two callers of this function.
2827          *
2828          * One in create_subvol() in the ioctl context, which needs to hold
2829          * the qgroup_ioctl_lock.
2830          *
2831          * The other one in create_pending_snapshot() where no other qgroup
2832          * code can modify the fs as they all need to either start a new trans
2833          * or hold a trans handler, thus we don't need to hold
2834          * qgroup_ioctl_lock.
2835          * This would avoid long and complex lock chain and make lockdep happy.
2836          */
2837         spin_lock(&fs_info->trans_lock);
2838         if (trans->transaction->state == TRANS_STATE_COMMIT_DOING)
2839                 committing = true;
2840         spin_unlock(&fs_info->trans_lock);
2841
2842         if (!committing)
2843                 mutex_lock(&fs_info->qgroup_ioctl_lock);
2844         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2845                 goto out;
2846
2847         quota_root = fs_info->quota_root;
2848         if (!quota_root) {
2849                 ret = -EINVAL;
2850                 goto out;
2851         }
2852
2853         if (inherit) {
2854                 i_qgroups = (u64 *)(inherit + 1);
2855                 nums = inherit->num_qgroups + 2 * inherit->num_ref_copies +
2856                        2 * inherit->num_excl_copies;
2857                 for (i = 0; i < nums; ++i) {
2858                         srcgroup = find_qgroup_rb(fs_info, *i_qgroups);
2859
2860                         /*
2861                          * Zero out invalid groups so we can ignore
2862                          * them later.
2863                          */
2864                         if (!srcgroup ||
2865                             ((srcgroup->qgroupid >> 48) <= (objectid >> 48)))
2866                                 *i_qgroups = 0ULL;
2867
2868                         ++i_qgroups;
2869                 }
2870         }
2871
2872         /*
2873          * create a tracking group for the subvol itself
2874          */
2875         ret = add_qgroup_item(trans, quota_root, objectid);
2876         if (ret)
2877                 goto out;
2878
2879         /*
2880          * add qgroup to all inherited groups
2881          */
2882         if (inherit) {
2883                 i_qgroups = (u64 *)(inherit + 1);
2884                 for (i = 0; i < inherit->num_qgroups; ++i, ++i_qgroups) {
2885                         if (*i_qgroups == 0)
2886                                 continue;
2887                         ret = add_qgroup_relation_item(trans, objectid,
2888                                                        *i_qgroups);
2889                         if (ret && ret != -EEXIST)
2890                                 goto out;
2891                         ret = add_qgroup_relation_item(trans, *i_qgroups,
2892                                                        objectid);
2893                         if (ret && ret != -EEXIST)
2894                                 goto out;
2895                 }
2896                 ret = 0;
2897         }
2898
2899
2900         spin_lock(&fs_info->qgroup_lock);
2901
2902         dstgroup = add_qgroup_rb(fs_info, objectid);
2903         if (IS_ERR(dstgroup)) {
2904                 ret = PTR_ERR(dstgroup);
2905                 goto unlock;
2906         }
2907
2908         if (inherit && inherit->flags & BTRFS_QGROUP_INHERIT_SET_LIMITS) {
2909                 dstgroup->lim_flags = inherit->lim.flags;
2910                 dstgroup->max_rfer = inherit->lim.max_rfer;
2911                 dstgroup->max_excl = inherit->lim.max_excl;
2912                 dstgroup->rsv_rfer = inherit->lim.rsv_rfer;
2913                 dstgroup->rsv_excl = inherit->lim.rsv_excl;
2914
2915                 qgroup_dirty(fs_info, dstgroup);
2916         }
2917
2918         if (srcid) {
2919                 srcgroup = find_qgroup_rb(fs_info, srcid);
2920                 if (!srcgroup)
2921                         goto unlock;
2922
2923                 /*
2924                  * We call inherit after we clone the root in order to make sure
2925                  * our counts don't go crazy, so at this point the only
2926                  * difference between the two roots should be the root node.
2927                  */
2928                 level_size = fs_info->nodesize;
2929                 dstgroup->rfer = srcgroup->rfer;
2930                 dstgroup->rfer_cmpr = srcgroup->rfer_cmpr;
2931                 dstgroup->excl = level_size;
2932                 dstgroup->excl_cmpr = level_size;
2933                 srcgroup->excl = level_size;
2934                 srcgroup->excl_cmpr = level_size;
2935
2936                 /* inherit the limit info */
2937                 dstgroup->lim_flags = srcgroup->lim_flags;
2938                 dstgroup->max_rfer = srcgroup->max_rfer;
2939                 dstgroup->max_excl = srcgroup->max_excl;
2940                 dstgroup->rsv_rfer = srcgroup->rsv_rfer;
2941                 dstgroup->rsv_excl = srcgroup->rsv_excl;
2942
2943                 qgroup_dirty(fs_info, dstgroup);
2944                 qgroup_dirty(fs_info, srcgroup);
2945         }
2946
2947         if (!inherit)
2948                 goto unlock;
2949
2950         i_qgroups = (u64 *)(inherit + 1);
2951         for (i = 0; i < inherit->num_qgroups; ++i) {
2952                 if (*i_qgroups) {
2953                         ret = add_relation_rb(fs_info, objectid, *i_qgroups);
2954                         if (ret)
2955                                 goto unlock;
2956                 }
2957                 ++i_qgroups;
2958
2959                 /*
2960                  * If we're doing a snapshot, and adding the snapshot to a new
2961                  * qgroup, the numbers are guaranteed to be incorrect.
2962                  */
2963                 if (srcid)
2964                         need_rescan = true;
2965         }
2966
2967         for (i = 0; i <  inherit->num_ref_copies; ++i, i_qgroups += 2) {
2968                 struct btrfs_qgroup *src;
2969                 struct btrfs_qgroup *dst;
2970
2971                 if (!i_qgroups[0] || !i_qgroups[1])
2972                         continue;
2973
2974                 src = find_qgroup_rb(fs_info, i_qgroups[0]);
2975                 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
2976
2977                 if (!src || !dst) {
2978                         ret = -EINVAL;
2979                         goto unlock;
2980                 }
2981
2982                 dst->rfer = src->rfer - level_size;
2983                 dst->rfer_cmpr = src->rfer_cmpr - level_size;
2984
2985                 /* Manually tweaking numbers certainly needs a rescan */
2986                 need_rescan = true;
2987         }
2988         for (i = 0; i <  inherit->num_excl_copies; ++i, i_qgroups += 2) {
2989                 struct btrfs_qgroup *src;
2990                 struct btrfs_qgroup *dst;
2991
2992                 if (!i_qgroups[0] || !i_qgroups[1])
2993                         continue;
2994
2995                 src = find_qgroup_rb(fs_info, i_qgroups[0]);
2996                 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
2997
2998                 if (!src || !dst) {
2999                         ret = -EINVAL;
3000                         goto unlock;
3001                 }
3002
3003                 dst->excl = src->excl + level_size;
3004                 dst->excl_cmpr = src->excl_cmpr + level_size;
3005                 need_rescan = true;
3006         }
3007
3008 unlock:
3009         spin_unlock(&fs_info->qgroup_lock);
3010         if (!ret)
3011                 ret = btrfs_sysfs_add_one_qgroup(fs_info, dstgroup);
3012 out:
3013         if (!committing)
3014                 mutex_unlock(&fs_info->qgroup_ioctl_lock);
3015         if (need_rescan)
3016                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
3017         return ret;
3018 }
3019
3020 static bool qgroup_check_limits(const struct btrfs_qgroup *qg, u64 num_bytes)
3021 {
3022         if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_RFER) &&
3023             qgroup_rsv_total(qg) + (s64)qg->rfer + num_bytes > qg->max_rfer)
3024                 return false;
3025
3026         if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) &&
3027             qgroup_rsv_total(qg) + (s64)qg->excl + num_bytes > qg->max_excl)
3028                 return false;
3029
3030         return true;
3031 }
3032
3033 static int qgroup_reserve(struct btrfs_root *root, u64 num_bytes, bool enforce,
3034                           enum btrfs_qgroup_rsv_type type)
3035 {
3036         struct btrfs_qgroup *qgroup;
3037         struct btrfs_fs_info *fs_info = root->fs_info;
3038         u64 ref_root = root->root_key.objectid;
3039         int ret = 0;
3040         struct ulist_node *unode;
3041         struct ulist_iterator uiter;
3042
3043         if (!is_fstree(ref_root))
3044                 return 0;
3045
3046         if (num_bytes == 0)
3047                 return 0;
3048
3049         if (test_bit(BTRFS_FS_QUOTA_OVERRIDE, &fs_info->flags) &&
3050             capable(CAP_SYS_RESOURCE))
3051                 enforce = false;
3052
3053         spin_lock(&fs_info->qgroup_lock);
3054         if (!fs_info->quota_root)
3055                 goto out;
3056
3057         qgroup = find_qgroup_rb(fs_info, ref_root);
3058         if (!qgroup)
3059                 goto out;
3060
3061         /*
3062          * in a first step, we check all affected qgroups if any limits would
3063          * be exceeded
3064          */
3065         ulist_reinit(fs_info->qgroup_ulist);
3066         ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
3067                         qgroup_to_aux(qgroup), GFP_ATOMIC);
3068         if (ret < 0)
3069                 goto out;
3070         ULIST_ITER_INIT(&uiter);
3071         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
3072                 struct btrfs_qgroup *qg;
3073                 struct btrfs_qgroup_list *glist;
3074
3075                 qg = unode_aux_to_qgroup(unode);
3076
3077                 if (enforce && !qgroup_check_limits(qg, num_bytes)) {
3078                         ret = -EDQUOT;
3079                         goto out;
3080                 }
3081
3082                 list_for_each_entry(glist, &qg->groups, next_group) {
3083                         ret = ulist_add(fs_info->qgroup_ulist,
3084                                         glist->group->qgroupid,
3085                                         qgroup_to_aux(glist->group), GFP_ATOMIC);
3086                         if (ret < 0)
3087                                 goto out;
3088                 }
3089         }
3090         ret = 0;
3091         /*
3092          * no limits exceeded, now record the reservation into all qgroups
3093          */
3094         ULIST_ITER_INIT(&uiter);
3095         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
3096                 struct btrfs_qgroup *qg;
3097
3098                 qg = unode_aux_to_qgroup(unode);
3099
3100                 qgroup_rsv_add(fs_info, qg, num_bytes, type);
3101         }
3102
3103 out:
3104         spin_unlock(&fs_info->qgroup_lock);
3105         return ret;
3106 }
3107
3108 /*
3109  * Free @num_bytes of reserved space with @type for qgroup.  (Normally level 0
3110  * qgroup).
3111  *
3112  * Will handle all higher level qgroup too.
3113  *
3114  * NOTE: If @num_bytes is (u64)-1, this means to free all bytes of this qgroup.
3115  * This special case is only used for META_PERTRANS type.
3116  */
3117 void btrfs_qgroup_free_refroot(struct btrfs_fs_info *fs_info,
3118                                u64 ref_root, u64 num_bytes,
3119                                enum btrfs_qgroup_rsv_type type)
3120 {
3121         struct btrfs_qgroup *qgroup;
3122         struct ulist_node *unode;
3123         struct ulist_iterator uiter;
3124         int ret = 0;
3125
3126         if (!is_fstree(ref_root))
3127                 return;
3128
3129         if (num_bytes == 0)
3130                 return;
3131
3132         if (num_bytes == (u64)-1 && type != BTRFS_QGROUP_RSV_META_PERTRANS) {
3133                 WARN(1, "%s: Invalid type to free", __func__);
3134                 return;
3135         }
3136         spin_lock(&fs_info->qgroup_lock);
3137
3138         if (!fs_info->quota_root)
3139                 goto out;
3140
3141         qgroup = find_qgroup_rb(fs_info, ref_root);
3142         if (!qgroup)
3143                 goto out;
3144
3145         if (num_bytes == (u64)-1)
3146                 /*
3147                  * We're freeing all pertrans rsv, get reserved value from
3148                  * level 0 qgroup as real num_bytes to free.
3149                  */
3150                 num_bytes = qgroup->rsv.values[type];
3151
3152         ulist_reinit(fs_info->qgroup_ulist);
3153         ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
3154                         qgroup_to_aux(qgroup), GFP_ATOMIC);
3155         if (ret < 0)
3156                 goto out;
3157         ULIST_ITER_INIT(&uiter);
3158         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
3159                 struct btrfs_qgroup *qg;
3160                 struct btrfs_qgroup_list *glist;
3161
3162                 qg = unode_aux_to_qgroup(unode);
3163
3164                 qgroup_rsv_release(fs_info, qg, num_bytes, type);
3165
3166                 list_for_each_entry(glist, &qg->groups, next_group) {
3167                         ret = ulist_add(fs_info->qgroup_ulist,
3168                                         glist->group->qgroupid,
3169                                         qgroup_to_aux(glist->group), GFP_ATOMIC);
3170                         if (ret < 0)
3171                                 goto out;
3172                 }
3173         }
3174
3175 out:
3176         spin_unlock(&fs_info->qgroup_lock);
3177 }
3178
3179 /*
3180  * Check if the leaf is the last leaf. Which means all node pointers
3181  * are at their last position.
3182  */
3183 static bool is_last_leaf(struct btrfs_path *path)
3184 {
3185         int i;
3186
3187         for (i = 1; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
3188                 if (path->slots[i] != btrfs_header_nritems(path->nodes[i]) - 1)
3189                         return false;
3190         }
3191         return true;
3192 }
3193
3194 /*
3195  * returns < 0 on error, 0 when more leafs are to be scanned.
3196  * returns 1 when done.
3197  */
3198 static int qgroup_rescan_leaf(struct btrfs_trans_handle *trans,
3199                               struct btrfs_path *path)
3200 {
3201         struct btrfs_fs_info *fs_info = trans->fs_info;
3202         struct btrfs_key found;
3203         struct extent_buffer *scratch_leaf = NULL;
3204         struct ulist *roots = NULL;
3205         u64 num_bytes;
3206         bool done;
3207         int slot;
3208         int ret;
3209
3210         mutex_lock(&fs_info->qgroup_rescan_lock);
3211         ret = btrfs_search_slot_for_read(fs_info->extent_root,
3212                                          &fs_info->qgroup_rescan_progress,
3213                                          path, 1, 0);
3214
3215         btrfs_debug(fs_info,
3216                 "current progress key (%llu %u %llu), search_slot ret %d",
3217                 fs_info->qgroup_rescan_progress.objectid,
3218                 fs_info->qgroup_rescan_progress.type,
3219                 fs_info->qgroup_rescan_progress.offset, ret);
3220
3221         if (ret) {
3222                 /*
3223                  * The rescan is about to end, we will not be scanning any
3224                  * further blocks. We cannot unset the RESCAN flag here, because
3225                  * we want to commit the transaction if everything went well.
3226                  * To make the live accounting work in this phase, we set our
3227                  * scan progress pointer such that every real extent objectid
3228                  * will be smaller.
3229                  */
3230                 fs_info->qgroup_rescan_progress.objectid = (u64)-1;
3231                 btrfs_release_path(path);
3232                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3233                 return ret;
3234         }
3235         done = is_last_leaf(path);
3236
3237         btrfs_item_key_to_cpu(path->nodes[0], &found,
3238                               btrfs_header_nritems(path->nodes[0]) - 1);
3239         fs_info->qgroup_rescan_progress.objectid = found.objectid + 1;
3240
3241         scratch_leaf = btrfs_clone_extent_buffer(path->nodes[0]);
3242         if (!scratch_leaf) {
3243                 ret = -ENOMEM;
3244                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3245                 goto out;
3246         }
3247         slot = path->slots[0];
3248         btrfs_release_path(path);
3249         mutex_unlock(&fs_info->qgroup_rescan_lock);
3250
3251         for (; slot < btrfs_header_nritems(scratch_leaf); ++slot) {
3252                 btrfs_item_key_to_cpu(scratch_leaf, &found, slot);
3253                 if (found.type != BTRFS_EXTENT_ITEM_KEY &&
3254                     found.type != BTRFS_METADATA_ITEM_KEY)
3255                         continue;
3256                 if (found.type == BTRFS_METADATA_ITEM_KEY)
3257                         num_bytes = fs_info->nodesize;
3258                 else
3259                         num_bytes = found.offset;
3260
3261                 ret = btrfs_find_all_roots(NULL, fs_info, found.objectid, 0,
3262                                            &roots, false);
3263                 if (ret < 0)
3264                         goto out;
3265                 /* For rescan, just pass old_roots as NULL */
3266                 ret = btrfs_qgroup_account_extent(trans, found.objectid,
3267                                                   num_bytes, NULL, roots);
3268                 if (ret < 0)
3269                         goto out;
3270         }
3271 out:
3272         if (scratch_leaf)
3273                 free_extent_buffer(scratch_leaf);
3274
3275         if (done && !ret) {
3276                 ret = 1;
3277                 fs_info->qgroup_rescan_progress.objectid = (u64)-1;
3278         }
3279         return ret;
3280 }
3281
3282 static bool rescan_should_stop(struct btrfs_fs_info *fs_info)
3283 {
3284         return btrfs_fs_closing(fs_info) ||
3285                 test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state) ||
3286                 !test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
3287 }
3288
3289 static void btrfs_qgroup_rescan_worker(struct btrfs_work *work)
3290 {
3291         struct btrfs_fs_info *fs_info = container_of(work, struct btrfs_fs_info,
3292                                                      qgroup_rescan_work);
3293         struct btrfs_path *path;
3294         struct btrfs_trans_handle *trans = NULL;
3295         int err = -ENOMEM;
3296         int ret = 0;
3297         bool stopped = false;
3298         bool did_leaf_rescans = false;
3299
3300         path = btrfs_alloc_path();
3301         if (!path)
3302                 goto out;
3303         /*
3304          * Rescan should only search for commit root, and any later difference
3305          * should be recorded by qgroup
3306          */
3307         path->search_commit_root = 1;
3308         path->skip_locking = 1;
3309
3310         err = 0;
3311         while (!err && !(stopped = rescan_should_stop(fs_info))) {
3312                 trans = btrfs_start_transaction(fs_info->fs_root, 0);
3313                 if (IS_ERR(trans)) {
3314                         err = PTR_ERR(trans);
3315                         break;
3316                 }
3317
3318                 err = qgroup_rescan_leaf(trans, path);
3319                 did_leaf_rescans = true;
3320
3321                 if (err > 0)
3322                         btrfs_commit_transaction(trans);
3323                 else
3324                         btrfs_end_transaction(trans);
3325         }
3326
3327 out:
3328         btrfs_free_path(path);
3329
3330         mutex_lock(&fs_info->qgroup_rescan_lock);
3331         if (err > 0 &&
3332             fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT) {
3333                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
3334         } else if (err < 0 || stopped) {
3335                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
3336         }
3337         mutex_unlock(&fs_info->qgroup_rescan_lock);
3338
3339         /*
3340          * Only update status, since the previous part has already updated the
3341          * qgroup info, and only if we did any actual work. This also prevents
3342          * race with a concurrent quota disable, which has already set
3343          * fs_info->quota_root to NULL and cleared BTRFS_FS_QUOTA_ENABLED at
3344          * btrfs_quota_disable().
3345          */
3346         if (did_leaf_rescans) {
3347                 trans = btrfs_start_transaction(fs_info->quota_root, 1);
3348                 if (IS_ERR(trans)) {
3349                         err = PTR_ERR(trans);
3350                         trans = NULL;
3351                         btrfs_err(fs_info,
3352                                   "fail to start transaction for status update: %d",
3353                                   err);
3354                 }
3355         } else {
3356                 trans = NULL;
3357         }
3358
3359         mutex_lock(&fs_info->qgroup_rescan_lock);
3360         if (!stopped)
3361                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3362         if (trans) {
3363                 ret = update_qgroup_status_item(trans);
3364                 if (ret < 0) {
3365                         err = ret;
3366                         btrfs_err(fs_info, "fail to update qgroup status: %d",
3367                                   err);
3368                 }
3369         }
3370         fs_info->qgroup_rescan_running = false;
3371         complete_all(&fs_info->qgroup_rescan_completion);
3372         mutex_unlock(&fs_info->qgroup_rescan_lock);
3373
3374         if (!trans)
3375                 return;
3376
3377         btrfs_end_transaction(trans);
3378
3379         if (stopped) {
3380                 btrfs_info(fs_info, "qgroup scan paused");
3381         } else if (err >= 0) {
3382                 btrfs_info(fs_info, "qgroup scan completed%s",
3383                         err > 0 ? " (inconsistency flag cleared)" : "");
3384         } else {
3385                 btrfs_err(fs_info, "qgroup scan failed with %d", err);
3386         }
3387 }
3388
3389 /*
3390  * Checks that (a) no rescan is running and (b) quota is enabled. Allocates all
3391  * memory required for the rescan context.
3392  */
3393 static int
3394 qgroup_rescan_init(struct btrfs_fs_info *fs_info, u64 progress_objectid,
3395                    int init_flags)
3396 {
3397         int ret = 0;
3398
3399         if (!init_flags) {
3400                 /* we're resuming qgroup rescan at mount time */
3401                 if (!(fs_info->qgroup_flags &
3402                       BTRFS_QGROUP_STATUS_FLAG_RESCAN)) {
3403                         btrfs_warn(fs_info,
3404                         "qgroup rescan init failed, qgroup rescan is not queued");
3405                         ret = -EINVAL;
3406                 } else if (!(fs_info->qgroup_flags &
3407                              BTRFS_QGROUP_STATUS_FLAG_ON)) {
3408                         btrfs_warn(fs_info,
3409                         "qgroup rescan init failed, qgroup is not enabled");
3410                         ret = -EINVAL;
3411                 }
3412
3413                 if (ret)
3414                         return ret;
3415         }
3416
3417         mutex_lock(&fs_info->qgroup_rescan_lock);
3418
3419         if (init_flags) {
3420                 if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
3421                         btrfs_warn(fs_info,
3422                                    "qgroup rescan is already in progress");
3423                         ret = -EINPROGRESS;
3424                 } else if (!(fs_info->qgroup_flags &
3425                              BTRFS_QGROUP_STATUS_FLAG_ON)) {
3426                         btrfs_warn(fs_info,
3427                         "qgroup rescan init failed, qgroup is not enabled");
3428                         ret = -EINVAL;
3429                 } else if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
3430                         /* Quota disable is in progress */
3431                         ret = -EBUSY;
3432                 }
3433
3434                 if (ret) {
3435                         mutex_unlock(&fs_info->qgroup_rescan_lock);
3436                         return ret;
3437                 }
3438                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3439         }
3440
3441         memset(&fs_info->qgroup_rescan_progress, 0,
3442                 sizeof(fs_info->qgroup_rescan_progress));
3443         fs_info->qgroup_rescan_progress.objectid = progress_objectid;
3444         init_completion(&fs_info->qgroup_rescan_completion);
3445         mutex_unlock(&fs_info->qgroup_rescan_lock);
3446
3447         btrfs_init_work(&fs_info->qgroup_rescan_work,
3448                         btrfs_qgroup_rescan_worker, NULL, NULL);
3449         return 0;
3450 }
3451
3452 static void
3453 qgroup_rescan_zero_tracking(struct btrfs_fs_info *fs_info)
3454 {
3455         struct rb_node *n;
3456         struct btrfs_qgroup *qgroup;
3457
3458         spin_lock(&fs_info->qgroup_lock);
3459         /* clear all current qgroup tracking information */
3460         for (n = rb_first(&fs_info->qgroup_tree); n; n = rb_next(n)) {
3461                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
3462                 qgroup->rfer = 0;
3463                 qgroup->rfer_cmpr = 0;
3464                 qgroup->excl = 0;
3465                 qgroup->excl_cmpr = 0;
3466                 qgroup_dirty(fs_info, qgroup);
3467         }
3468         spin_unlock(&fs_info->qgroup_lock);
3469 }
3470
3471 int
3472 btrfs_qgroup_rescan(struct btrfs_fs_info *fs_info)
3473 {
3474         int ret = 0;
3475         struct btrfs_trans_handle *trans;
3476
3477         ret = qgroup_rescan_init(fs_info, 0, 1);
3478         if (ret)
3479                 return ret;
3480
3481         /*
3482          * We have set the rescan_progress to 0, which means no more
3483          * delayed refs will be accounted by btrfs_qgroup_account_ref.
3484          * However, btrfs_qgroup_account_ref may be right after its call
3485          * to btrfs_find_all_roots, in which case it would still do the
3486          * accounting.
3487          * To solve this, we're committing the transaction, which will
3488          * ensure we run all delayed refs and only after that, we are
3489          * going to clear all tracking information for a clean start.
3490          */
3491
3492         trans = btrfs_join_transaction(fs_info->fs_root);
3493         if (IS_ERR(trans)) {
3494                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3495                 return PTR_ERR(trans);
3496         }
3497         ret = btrfs_commit_transaction(trans);
3498         if (ret) {
3499                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3500                 return ret;
3501         }
3502
3503         qgroup_rescan_zero_tracking(fs_info);
3504
3505         mutex_lock(&fs_info->qgroup_rescan_lock);
3506         fs_info->qgroup_rescan_running = true;
3507         btrfs_queue_work(fs_info->qgroup_rescan_workers,
3508                          &fs_info->qgroup_rescan_work);
3509         mutex_unlock(&fs_info->qgroup_rescan_lock);
3510
3511         return 0;
3512 }
3513
3514 int btrfs_qgroup_wait_for_completion(struct btrfs_fs_info *fs_info,
3515                                      bool interruptible)
3516 {
3517         int running;
3518         int ret = 0;
3519
3520         mutex_lock(&fs_info->qgroup_rescan_lock);
3521         running = fs_info->qgroup_rescan_running;
3522         mutex_unlock(&fs_info->qgroup_rescan_lock);
3523
3524         if (!running)
3525                 return 0;
3526
3527         if (interruptible)
3528                 ret = wait_for_completion_interruptible(
3529                                         &fs_info->qgroup_rescan_completion);
3530         else
3531                 wait_for_completion(&fs_info->qgroup_rescan_completion);
3532
3533         return ret;
3534 }
3535
3536 /*
3537  * this is only called from open_ctree where we're still single threaded, thus
3538  * locking is omitted here.
3539  */
3540 void
3541 btrfs_qgroup_rescan_resume(struct btrfs_fs_info *fs_info)
3542 {
3543         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
3544                 mutex_lock(&fs_info->qgroup_rescan_lock);
3545                 fs_info->qgroup_rescan_running = true;
3546                 btrfs_queue_work(fs_info->qgroup_rescan_workers,
3547                                  &fs_info->qgroup_rescan_work);
3548                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3549         }
3550 }
3551
3552 #define rbtree_iterate_from_safe(node, next, start)                             \
3553        for (node = start; node && ({ next = rb_next(node); 1;}); node = next)
3554
3555 static int qgroup_unreserve_range(struct btrfs_inode *inode,
3556                                   struct extent_changeset *reserved, u64 start,
3557                                   u64 len)
3558 {
3559         struct rb_node *node;
3560         struct rb_node *next;
3561         struct ulist_node *entry;
3562         int ret = 0;
3563
3564         node = reserved->range_changed.root.rb_node;
3565         if (!node)
3566                 return 0;
3567         while (node) {
3568                 entry = rb_entry(node, struct ulist_node, rb_node);
3569                 if (entry->val < start)
3570                         node = node->rb_right;
3571                 else
3572                         node = node->rb_left;
3573         }
3574
3575         if (entry->val > start && rb_prev(&entry->rb_node))
3576                 entry = rb_entry(rb_prev(&entry->rb_node), struct ulist_node,
3577                                  rb_node);
3578
3579         rbtree_iterate_from_safe(node, next, &entry->rb_node) {
3580                 u64 entry_start;
3581                 u64 entry_end;
3582                 u64 entry_len;
3583                 int clear_ret;
3584
3585                 entry = rb_entry(node, struct ulist_node, rb_node);
3586                 entry_start = entry->val;
3587                 entry_end = entry->aux;
3588                 entry_len = entry_end - entry_start + 1;
3589
3590                 if (entry_start >= start + len)
3591                         break;
3592                 if (entry_start + entry_len <= start)
3593                         continue;
3594                 /*
3595                  * Now the entry is in [start, start + len), revert the
3596                  * EXTENT_QGROUP_RESERVED bit.
3597                  */
3598                 clear_ret = clear_extent_bits(&inode->io_tree, entry_start,
3599                                               entry_end, EXTENT_QGROUP_RESERVED);
3600                 if (!ret && clear_ret < 0)
3601                         ret = clear_ret;
3602
3603                 ulist_del(&reserved->range_changed, entry->val, entry->aux);
3604                 if (likely(reserved->bytes_changed >= entry_len)) {
3605                         reserved->bytes_changed -= entry_len;
3606                 } else {
3607                         WARN_ON(1);
3608                         reserved->bytes_changed = 0;
3609                 }
3610         }
3611
3612         return ret;
3613 }
3614
3615 /*
3616  * Try to free some space for qgroup.
3617  *
3618  * For qgroup, there are only 3 ways to free qgroup space:
3619  * - Flush nodatacow write
3620  *   Any nodatacow write will free its reserved data space at run_delalloc_range().
3621  *   In theory, we should only flush nodatacow inodes, but it's not yet
3622  *   possible, so we need to flush the whole root.
3623  *
3624  * - Wait for ordered extents
3625  *   When ordered extents are finished, their reserved metadata is finally
3626  *   converted to per_trans status, which can be freed by later commit
3627  *   transaction.
3628  *
3629  * - Commit transaction
3630  *   This would free the meta_per_trans space.
3631  *   In theory this shouldn't provide much space, but any more qgroup space
3632  *   is needed.
3633  */
3634 static int try_flush_qgroup(struct btrfs_root *root)
3635 {
3636         struct btrfs_trans_handle *trans;
3637         int ret;
3638
3639         /* Can't hold an open transaction or we run the risk of deadlocking. */
3640         ASSERT(current->journal_info == NULL);
3641         if (WARN_ON(current->journal_info))
3642                 return 0;
3643
3644         /*
3645          * We don't want to run flush again and again, so if there is a running
3646          * one, we won't try to start a new flush, but exit directly.
3647          */
3648         if (test_and_set_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state)) {
3649                 wait_event(root->qgroup_flush_wait,
3650                         !test_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state));
3651                 return 0;
3652         }
3653
3654         ret = btrfs_start_delalloc_snapshot(root, true);
3655         if (ret < 0)
3656                 goto out;
3657         btrfs_wait_ordered_extents(root, U64_MAX, 0, (u64)-1);
3658
3659         trans = btrfs_join_transaction(root);
3660         if (IS_ERR(trans)) {
3661                 ret = PTR_ERR(trans);
3662                 goto out;
3663         }
3664
3665         ret = btrfs_commit_transaction(trans);
3666 out:
3667         clear_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state);
3668         wake_up(&root->qgroup_flush_wait);
3669         return ret;
3670 }
3671
3672 static int qgroup_reserve_data(struct btrfs_inode *inode,
3673                         struct extent_changeset **reserved_ret, u64 start,
3674                         u64 len)
3675 {
3676         struct btrfs_root *root = inode->root;
3677         struct extent_changeset *reserved;
3678         bool new_reserved = false;
3679         u64 orig_reserved;
3680         u64 to_reserve;
3681         int ret;
3682
3683         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &root->fs_info->flags) ||
3684             !is_fstree(root->root_key.objectid) || len == 0)
3685                 return 0;
3686
3687         /* @reserved parameter is mandatory for qgroup */
3688         if (WARN_ON(!reserved_ret))
3689                 return -EINVAL;
3690         if (!*reserved_ret) {
3691                 new_reserved = true;
3692                 *reserved_ret = extent_changeset_alloc();
3693                 if (!*reserved_ret)
3694                         return -ENOMEM;
3695         }
3696         reserved = *reserved_ret;
3697         /* Record already reserved space */
3698         orig_reserved = reserved->bytes_changed;
3699         ret = set_record_extent_bits(&inode->io_tree, start,
3700                         start + len -1, EXTENT_QGROUP_RESERVED, reserved);
3701
3702         /* Newly reserved space */
3703         to_reserve = reserved->bytes_changed - orig_reserved;
3704         trace_btrfs_qgroup_reserve_data(&inode->vfs_inode, start, len,
3705                                         to_reserve, QGROUP_RESERVE);
3706         if (ret < 0)
3707                 goto out;
3708         ret = qgroup_reserve(root, to_reserve, true, BTRFS_QGROUP_RSV_DATA);
3709         if (ret < 0)
3710                 goto cleanup;
3711
3712         return ret;
3713
3714 cleanup:
3715         qgroup_unreserve_range(inode, reserved, start, len);
3716 out:
3717         if (new_reserved) {
3718                 extent_changeset_free(reserved);
3719                 *reserved_ret = NULL;
3720         }
3721         return ret;
3722 }
3723
3724 /*
3725  * Reserve qgroup space for range [start, start + len).
3726  *
3727  * This function will either reserve space from related qgroups or do nothing
3728  * if the range is already reserved.
3729  *
3730  * Return 0 for successful reservation
3731  * Return <0 for error (including -EQUOT)
3732  *
3733  * NOTE: This function may sleep for memory allocation, dirty page flushing and
3734  *       commit transaction. So caller should not hold any dirty page locked.
3735  */
3736 int btrfs_qgroup_reserve_data(struct btrfs_inode *inode,
3737                         struct extent_changeset **reserved_ret, u64 start,
3738                         u64 len)
3739 {
3740         int ret;
3741
3742         ret = qgroup_reserve_data(inode, reserved_ret, start, len);
3743         if (ret <= 0 && ret != -EDQUOT)
3744                 return ret;
3745
3746         ret = try_flush_qgroup(inode->root);
3747         if (ret < 0)
3748                 return ret;
3749         return qgroup_reserve_data(inode, reserved_ret, start, len);
3750 }
3751
3752 /* Free ranges specified by @reserved, normally in error path */
3753 static int qgroup_free_reserved_data(struct btrfs_inode *inode,
3754                         struct extent_changeset *reserved, u64 start, u64 len)
3755 {
3756         struct btrfs_root *root = inode->root;
3757         struct ulist_node *unode;
3758         struct ulist_iterator uiter;
3759         struct extent_changeset changeset;
3760         int freed = 0;
3761         int ret;
3762
3763         extent_changeset_init(&changeset);
3764         len = round_up(start + len, root->fs_info->sectorsize);
3765         start = round_down(start, root->fs_info->sectorsize);
3766
3767         ULIST_ITER_INIT(&uiter);
3768         while ((unode = ulist_next(&reserved->range_changed, &uiter))) {
3769                 u64 range_start = unode->val;
3770                 /* unode->aux is the inclusive end */
3771                 u64 range_len = unode->aux - range_start + 1;
3772                 u64 free_start;
3773                 u64 free_len;
3774
3775                 extent_changeset_release(&changeset);
3776
3777                 /* Only free range in range [start, start + len) */
3778                 if (range_start >= start + len ||
3779                     range_start + range_len <= start)
3780                         continue;
3781                 free_start = max(range_start, start);
3782                 free_len = min(start + len, range_start + range_len) -
3783                            free_start;
3784                 /*
3785                  * TODO: To also modify reserved->ranges_reserved to reflect
3786                  * the modification.
3787                  *
3788                  * However as long as we free qgroup reserved according to
3789                  * EXTENT_QGROUP_RESERVED, we won't double free.
3790                  * So not need to rush.
3791                  */
3792                 ret = clear_record_extent_bits(&inode->io_tree, free_start,
3793                                 free_start + free_len - 1,
3794                                 EXTENT_QGROUP_RESERVED, &changeset);
3795                 if (ret < 0)
3796                         goto out;
3797                 freed += changeset.bytes_changed;
3798         }
3799         btrfs_qgroup_free_refroot(root->fs_info, root->root_key.objectid, freed,
3800                                   BTRFS_QGROUP_RSV_DATA);
3801         ret = freed;
3802 out:
3803         extent_changeset_release(&changeset);
3804         return ret;
3805 }
3806
3807 static int __btrfs_qgroup_release_data(struct btrfs_inode *inode,
3808                         struct extent_changeset *reserved, u64 start, u64 len,
3809                         int free)
3810 {
3811         struct extent_changeset changeset;
3812         int trace_op = QGROUP_RELEASE;
3813         int ret;
3814
3815         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &inode->root->fs_info->flags))
3816                 return 0;
3817
3818         /* In release case, we shouldn't have @reserved */
3819         WARN_ON(!free && reserved);
3820         if (free && reserved)
3821                 return qgroup_free_reserved_data(inode, reserved, start, len);
3822         extent_changeset_init(&changeset);
3823         ret = clear_record_extent_bits(&inode->io_tree, start, start + len -1,
3824                                        EXTENT_QGROUP_RESERVED, &changeset);
3825         if (ret < 0)
3826                 goto out;
3827
3828         if (free)
3829                 trace_op = QGROUP_FREE;
3830         trace_btrfs_qgroup_release_data(&inode->vfs_inode, start, len,
3831                                         changeset.bytes_changed, trace_op);
3832         if (free)
3833                 btrfs_qgroup_free_refroot(inode->root->fs_info,
3834                                 inode->root->root_key.objectid,
3835                                 changeset.bytes_changed, BTRFS_QGROUP_RSV_DATA);
3836         ret = changeset.bytes_changed;
3837 out:
3838         extent_changeset_release(&changeset);
3839         return ret;
3840 }
3841
3842 /*
3843  * Free a reserved space range from io_tree and related qgroups
3844  *
3845  * Should be called when a range of pages get invalidated before reaching disk.
3846  * Or for error cleanup case.
3847  * if @reserved is given, only reserved range in [@start, @start + @len) will
3848  * be freed.
3849  *
3850  * For data written to disk, use btrfs_qgroup_release_data().
3851  *
3852  * NOTE: This function may sleep for memory allocation.
3853  */
3854 int btrfs_qgroup_free_data(struct btrfs_inode *inode,
3855                         struct extent_changeset *reserved, u64 start, u64 len)
3856 {
3857         return __btrfs_qgroup_release_data(inode, reserved, start, len, 1);
3858 }
3859
3860 /*
3861  * Release a reserved space range from io_tree only.
3862  *
3863  * Should be called when a range of pages get written to disk and corresponding
3864  * FILE_EXTENT is inserted into corresponding root.
3865  *
3866  * Since new qgroup accounting framework will only update qgroup numbers at
3867  * commit_transaction() time, its reserved space shouldn't be freed from
3868  * related qgroups.
3869  *
3870  * But we should release the range from io_tree, to allow further write to be
3871  * COWed.
3872  *
3873  * NOTE: This function may sleep for memory allocation.
3874  */
3875 int btrfs_qgroup_release_data(struct btrfs_inode *inode, u64 start, u64 len)
3876 {
3877         return __btrfs_qgroup_release_data(inode, NULL, start, len, 0);
3878 }
3879
3880 static void add_root_meta_rsv(struct btrfs_root *root, int num_bytes,
3881                               enum btrfs_qgroup_rsv_type type)
3882 {
3883         if (type != BTRFS_QGROUP_RSV_META_PREALLOC &&
3884             type != BTRFS_QGROUP_RSV_META_PERTRANS)
3885                 return;
3886         if (num_bytes == 0)
3887                 return;
3888
3889         spin_lock(&root->qgroup_meta_rsv_lock);
3890         if (type == BTRFS_QGROUP_RSV_META_PREALLOC)
3891                 root->qgroup_meta_rsv_prealloc += num_bytes;
3892         else
3893                 root->qgroup_meta_rsv_pertrans += num_bytes;
3894         spin_unlock(&root->qgroup_meta_rsv_lock);
3895 }
3896
3897 static int sub_root_meta_rsv(struct btrfs_root *root, int num_bytes,
3898                              enum btrfs_qgroup_rsv_type type)
3899 {
3900         if (type != BTRFS_QGROUP_RSV_META_PREALLOC &&
3901             type != BTRFS_QGROUP_RSV_META_PERTRANS)
3902                 return 0;
3903         if (num_bytes == 0)
3904                 return 0;
3905
3906         spin_lock(&root->qgroup_meta_rsv_lock);
3907         if (type == BTRFS_QGROUP_RSV_META_PREALLOC) {
3908                 num_bytes = min_t(u64, root->qgroup_meta_rsv_prealloc,
3909                                   num_bytes);
3910                 root->qgroup_meta_rsv_prealloc -= num_bytes;
3911         } else {
3912                 num_bytes = min_t(u64, root->qgroup_meta_rsv_pertrans,
3913                                   num_bytes);
3914                 root->qgroup_meta_rsv_pertrans -= num_bytes;
3915         }
3916         spin_unlock(&root->qgroup_meta_rsv_lock);
3917         return num_bytes;
3918 }
3919
3920 int btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
3921                               enum btrfs_qgroup_rsv_type type, bool enforce)
3922 {
3923         struct btrfs_fs_info *fs_info = root->fs_info;
3924         int ret;
3925
3926         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
3927             !is_fstree(root->root_key.objectid) || num_bytes == 0)
3928                 return 0;
3929
3930         BUG_ON(num_bytes != round_down(num_bytes, fs_info->nodesize));
3931         trace_qgroup_meta_reserve(root, (s64)num_bytes, type);
3932         ret = qgroup_reserve(root, num_bytes, enforce, type);
3933         if (ret < 0)
3934                 return ret;
3935         /*
3936          * Record what we have reserved into root.
3937          *
3938          * To avoid quota disabled->enabled underflow.
3939          * In that case, we may try to free space we haven't reserved
3940          * (since quota was disabled), so record what we reserved into root.
3941          * And ensure later release won't underflow this number.
3942          */
3943         add_root_meta_rsv(root, num_bytes, type);
3944         return ret;
3945 }
3946
3947 int __btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
3948                                 enum btrfs_qgroup_rsv_type type, bool enforce)
3949 {
3950         int ret;
3951
3952         ret = btrfs_qgroup_reserve_meta(root, num_bytes, type, enforce);
3953         if (ret <= 0 && ret != -EDQUOT)
3954                 return ret;
3955
3956         ret = try_flush_qgroup(root);
3957         if (ret < 0)
3958                 return ret;
3959         return btrfs_qgroup_reserve_meta(root, num_bytes, type, enforce);
3960 }
3961
3962 void btrfs_qgroup_free_meta_all_pertrans(struct btrfs_root *root)
3963 {
3964         struct btrfs_fs_info *fs_info = root->fs_info;
3965
3966         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
3967             !is_fstree(root->root_key.objectid))
3968                 return;
3969
3970         /* TODO: Update trace point to handle such free */
3971         trace_qgroup_meta_free_all_pertrans(root);
3972         /* Special value -1 means to free all reserved space */
3973         btrfs_qgroup_free_refroot(fs_info, root->root_key.objectid, (u64)-1,
3974                                   BTRFS_QGROUP_RSV_META_PERTRANS);
3975 }
3976
3977 void __btrfs_qgroup_free_meta(struct btrfs_root *root, int num_bytes,
3978                               enum btrfs_qgroup_rsv_type type)
3979 {
3980         struct btrfs_fs_info *fs_info = root->fs_info;
3981
3982         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
3983             !is_fstree(root->root_key.objectid))
3984                 return;
3985
3986         /*
3987          * reservation for META_PREALLOC can happen before quota is enabled,
3988          * which can lead to underflow.
3989          * Here ensure we will only free what we really have reserved.
3990          */
3991         num_bytes = sub_root_meta_rsv(root, num_bytes, type);
3992         BUG_ON(num_bytes != round_down(num_bytes, fs_info->nodesize));
3993         trace_qgroup_meta_reserve(root, -(s64)num_bytes, type);
3994         btrfs_qgroup_free_refroot(fs_info, root->root_key.objectid,
3995                                   num_bytes, type);
3996 }
3997
3998 static void qgroup_convert_meta(struct btrfs_fs_info *fs_info, u64 ref_root,
3999                                 int num_bytes)
4000 {
4001         struct btrfs_qgroup *qgroup;
4002         struct ulist_node *unode;
4003         struct ulist_iterator uiter;
4004         int ret = 0;
4005
4006         if (num_bytes == 0)
4007                 return;
4008         if (!fs_info->quota_root)
4009                 return;
4010
4011         spin_lock(&fs_info->qgroup_lock);
4012         qgroup = find_qgroup_rb(fs_info, ref_root);
4013         if (!qgroup)
4014                 goto out;
4015         ulist_reinit(fs_info->qgroup_ulist);
4016         ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
4017                        qgroup_to_aux(qgroup), GFP_ATOMIC);
4018         if (ret < 0)
4019                 goto out;
4020         ULIST_ITER_INIT(&uiter);
4021         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
4022                 struct btrfs_qgroup *qg;
4023                 struct btrfs_qgroup_list *glist;
4024
4025                 qg = unode_aux_to_qgroup(unode);
4026
4027                 qgroup_rsv_release(fs_info, qg, num_bytes,
4028                                 BTRFS_QGROUP_RSV_META_PREALLOC);
4029                 qgroup_rsv_add(fs_info, qg, num_bytes,
4030                                 BTRFS_QGROUP_RSV_META_PERTRANS);
4031                 list_for_each_entry(glist, &qg->groups, next_group) {
4032                         ret = ulist_add(fs_info->qgroup_ulist,
4033                                         glist->group->qgroupid,
4034                                         qgroup_to_aux(glist->group), GFP_ATOMIC);
4035                         if (ret < 0)
4036                                 goto out;
4037                 }
4038         }
4039 out:
4040         spin_unlock(&fs_info->qgroup_lock);
4041 }
4042
4043 void btrfs_qgroup_convert_reserved_meta(struct btrfs_root *root, int num_bytes)
4044 {
4045         struct btrfs_fs_info *fs_info = root->fs_info;
4046
4047         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
4048             !is_fstree(root->root_key.objectid))
4049                 return;
4050         /* Same as btrfs_qgroup_free_meta_prealloc() */
4051         num_bytes = sub_root_meta_rsv(root, num_bytes,
4052                                       BTRFS_QGROUP_RSV_META_PREALLOC);
4053         trace_qgroup_meta_convert(root, num_bytes);
4054         qgroup_convert_meta(fs_info, root->root_key.objectid, num_bytes);
4055 }
4056
4057 /*
4058  * Check qgroup reserved space leaking, normally at destroy inode
4059  * time
4060  */
4061 void btrfs_qgroup_check_reserved_leak(struct btrfs_inode *inode)
4062 {
4063         struct extent_changeset changeset;
4064         struct ulist_node *unode;
4065         struct ulist_iterator iter;
4066         int ret;
4067
4068         extent_changeset_init(&changeset);
4069         ret = clear_record_extent_bits(&inode->io_tree, 0, (u64)-1,
4070                         EXTENT_QGROUP_RESERVED, &changeset);
4071
4072         WARN_ON(ret < 0);
4073         if (WARN_ON(changeset.bytes_changed)) {
4074                 ULIST_ITER_INIT(&iter);
4075                 while ((unode = ulist_next(&changeset.range_changed, &iter))) {
4076                         btrfs_warn(inode->root->fs_info,
4077                 "leaking qgroup reserved space, ino: %llu, start: %llu, end: %llu",
4078                                 btrfs_ino(inode), unode->val, unode->aux);
4079                 }
4080                 btrfs_qgroup_free_refroot(inode->root->fs_info,
4081                                 inode->root->root_key.objectid,
4082                                 changeset.bytes_changed, BTRFS_QGROUP_RSV_DATA);
4083
4084         }
4085         extent_changeset_release(&changeset);
4086 }
4087
4088 void btrfs_qgroup_init_swapped_blocks(
4089         struct btrfs_qgroup_swapped_blocks *swapped_blocks)
4090 {
4091         int i;
4092
4093         spin_lock_init(&swapped_blocks->lock);
4094         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
4095                 swapped_blocks->blocks[i] = RB_ROOT;
4096         swapped_blocks->swapped = false;
4097 }
4098
4099 /*
4100  * Delete all swapped blocks record of @root.
4101  * Every record here means we skipped a full subtree scan for qgroup.
4102  *
4103  * Gets called when committing one transaction.
4104  */
4105 void btrfs_qgroup_clean_swapped_blocks(struct btrfs_root *root)
4106 {
4107         struct btrfs_qgroup_swapped_blocks *swapped_blocks;
4108         int i;
4109
4110         swapped_blocks = &root->swapped_blocks;
4111
4112         spin_lock(&swapped_blocks->lock);
4113         if (!swapped_blocks->swapped)
4114                 goto out;
4115         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
4116                 struct rb_root *cur_root = &swapped_blocks->blocks[i];
4117                 struct btrfs_qgroup_swapped_block *entry;
4118                 struct btrfs_qgroup_swapped_block *next;
4119
4120                 rbtree_postorder_for_each_entry_safe(entry, next, cur_root,
4121                                                      node)
4122                         kfree(entry);
4123                 swapped_blocks->blocks[i] = RB_ROOT;
4124         }
4125         swapped_blocks->swapped = false;
4126 out:
4127         spin_unlock(&swapped_blocks->lock);
4128 }
4129
4130 /*
4131  * Add subtree roots record into @subvol_root.
4132  *
4133  * @subvol_root:        tree root of the subvolume tree get swapped
4134  * @bg:                 block group under balance
4135  * @subvol_parent/slot: pointer to the subtree root in subvolume tree
4136  * @reloc_parent/slot:  pointer to the subtree root in reloc tree
4137  *                      BOTH POINTERS ARE BEFORE TREE SWAP
4138  * @last_snapshot:      last snapshot generation of the subvolume tree
4139  */
4140 int btrfs_qgroup_add_swapped_blocks(struct btrfs_trans_handle *trans,
4141                 struct btrfs_root *subvol_root,
4142                 struct btrfs_block_group *bg,
4143                 struct extent_buffer *subvol_parent, int subvol_slot,
4144                 struct extent_buffer *reloc_parent, int reloc_slot,
4145                 u64 last_snapshot)
4146 {
4147         struct btrfs_fs_info *fs_info = subvol_root->fs_info;
4148         struct btrfs_qgroup_swapped_blocks *blocks = &subvol_root->swapped_blocks;
4149         struct btrfs_qgroup_swapped_block *block;
4150         struct rb_node **cur;
4151         struct rb_node *parent = NULL;
4152         int level = btrfs_header_level(subvol_parent) - 1;
4153         int ret = 0;
4154
4155         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
4156                 return 0;
4157
4158         if (btrfs_node_ptr_generation(subvol_parent, subvol_slot) >
4159             btrfs_node_ptr_generation(reloc_parent, reloc_slot)) {
4160                 btrfs_err_rl(fs_info,
4161                 "%s: bad parameter order, subvol_gen=%llu reloc_gen=%llu",
4162                         __func__,
4163                         btrfs_node_ptr_generation(subvol_parent, subvol_slot),
4164                         btrfs_node_ptr_generation(reloc_parent, reloc_slot));
4165                 return -EUCLEAN;
4166         }
4167
4168         block = kmalloc(sizeof(*block), GFP_NOFS);
4169         if (!block) {
4170                 ret = -ENOMEM;
4171                 goto out;
4172         }
4173
4174         /*
4175          * @reloc_parent/slot is still before swap, while @block is going to
4176          * record the bytenr after swap, so we do the swap here.
4177          */
4178         block->subvol_bytenr = btrfs_node_blockptr(reloc_parent, reloc_slot);
4179         block->subvol_generation = btrfs_node_ptr_generation(reloc_parent,
4180                                                              reloc_slot);
4181         block->reloc_bytenr = btrfs_node_blockptr(subvol_parent, subvol_slot);
4182         block->reloc_generation = btrfs_node_ptr_generation(subvol_parent,
4183                                                             subvol_slot);
4184         block->last_snapshot = last_snapshot;
4185         block->level = level;
4186
4187         /*
4188          * If we have bg == NULL, we're called from btrfs_recover_relocation(),
4189          * no one else can modify tree blocks thus we qgroup will not change
4190          * no matter the value of trace_leaf.
4191          */
4192         if (bg && bg->flags & BTRFS_BLOCK_GROUP_DATA)
4193                 block->trace_leaf = true;
4194         else
4195                 block->trace_leaf = false;
4196         btrfs_node_key_to_cpu(reloc_parent, &block->first_key, reloc_slot);
4197
4198         /* Insert @block into @blocks */
4199         spin_lock(&blocks->lock);
4200         cur = &blocks->blocks[level].rb_node;
4201         while (*cur) {
4202                 struct btrfs_qgroup_swapped_block *entry;
4203
4204                 parent = *cur;
4205                 entry = rb_entry(parent, struct btrfs_qgroup_swapped_block,
4206                                  node);
4207
4208                 if (entry->subvol_bytenr < block->subvol_bytenr) {
4209                         cur = &(*cur)->rb_left;
4210                 } else if (entry->subvol_bytenr > block->subvol_bytenr) {
4211                         cur = &(*cur)->rb_right;
4212                 } else {
4213                         if (entry->subvol_generation !=
4214                                         block->subvol_generation ||
4215                             entry->reloc_bytenr != block->reloc_bytenr ||
4216                             entry->reloc_generation !=
4217                                         block->reloc_generation) {
4218                                 /*
4219                                  * Duplicated but mismatch entry found.
4220                                  * Shouldn't happen.
4221                                  *
4222                                  * Marking qgroup inconsistent should be enough
4223                                  * for end users.
4224                                  */
4225                                 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
4226                                 ret = -EEXIST;
4227                         }
4228                         kfree(block);
4229                         goto out_unlock;
4230                 }
4231         }
4232         rb_link_node(&block->node, parent, cur);
4233         rb_insert_color(&block->node, &blocks->blocks[level]);
4234         blocks->swapped = true;
4235 out_unlock:
4236         spin_unlock(&blocks->lock);
4237 out:
4238         if (ret < 0)
4239                 fs_info->qgroup_flags |=
4240                         BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
4241         return ret;
4242 }
4243
4244 /*
4245  * Check if the tree block is a subtree root, and if so do the needed
4246  * delayed subtree trace for qgroup.
4247  *
4248  * This is called during btrfs_cow_block().
4249  */
4250 int btrfs_qgroup_trace_subtree_after_cow(struct btrfs_trans_handle *trans,
4251                                          struct btrfs_root *root,
4252                                          struct extent_buffer *subvol_eb)
4253 {
4254         struct btrfs_fs_info *fs_info = root->fs_info;
4255         struct btrfs_qgroup_swapped_blocks *blocks = &root->swapped_blocks;
4256         struct btrfs_qgroup_swapped_block *block;
4257         struct extent_buffer *reloc_eb = NULL;
4258         struct rb_node *node;
4259         bool found = false;
4260         bool swapped = false;
4261         int level = btrfs_header_level(subvol_eb);
4262         int ret = 0;
4263         int i;
4264
4265         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
4266                 return 0;
4267         if (!is_fstree(root->root_key.objectid) || !root->reloc_root)
4268                 return 0;
4269
4270         spin_lock(&blocks->lock);
4271         if (!blocks->swapped) {
4272                 spin_unlock(&blocks->lock);
4273                 return 0;
4274         }
4275         node = blocks->blocks[level].rb_node;
4276
4277         while (node) {
4278                 block = rb_entry(node, struct btrfs_qgroup_swapped_block, node);
4279                 if (block->subvol_bytenr < subvol_eb->start) {
4280                         node = node->rb_left;
4281                 } else if (block->subvol_bytenr > subvol_eb->start) {
4282                         node = node->rb_right;
4283                 } else {
4284                         found = true;
4285                         break;
4286                 }
4287         }
4288         if (!found) {
4289                 spin_unlock(&blocks->lock);
4290                 goto out;
4291         }
4292         /* Found one, remove it from @blocks first and update blocks->swapped */
4293         rb_erase(&block->node, &blocks->blocks[level]);
4294         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
4295                 if (RB_EMPTY_ROOT(&blocks->blocks[i])) {
4296                         swapped = true;
4297                         break;
4298                 }
4299         }
4300         blocks->swapped = swapped;
4301         spin_unlock(&blocks->lock);
4302
4303         /* Read out reloc subtree root */
4304         reloc_eb = read_tree_block(fs_info, block->reloc_bytenr, 0,
4305                                    block->reloc_generation, block->level,
4306                                    &block->first_key);
4307         if (IS_ERR(reloc_eb)) {
4308                 ret = PTR_ERR(reloc_eb);
4309                 reloc_eb = NULL;
4310                 goto free_out;
4311         }
4312         if (!extent_buffer_uptodate(reloc_eb)) {
4313                 ret = -EIO;
4314                 goto free_out;
4315         }
4316
4317         ret = qgroup_trace_subtree_swap(trans, reloc_eb, subvol_eb,
4318                         block->last_snapshot, block->trace_leaf);
4319 free_out:
4320         kfree(block);
4321         free_extent_buffer(reloc_eb);
4322 out:
4323         if (ret < 0) {
4324                 btrfs_err_rl(fs_info,
4325                              "failed to account subtree at bytenr %llu: %d",
4326                              subvol_eb->start, ret);
4327                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
4328         }
4329         return ret;
4330 }
4331
4332 void btrfs_qgroup_destroy_extent_records(struct btrfs_transaction *trans)
4333 {
4334         struct btrfs_qgroup_extent_record *entry;
4335         struct btrfs_qgroup_extent_record *next;
4336         struct rb_root *root;
4337
4338         root = &trans->delayed_refs.dirty_extent_root;
4339         rbtree_postorder_for_each_entry_safe(entry, next, root, node) {
4340                 ulist_free(entry->old_roots);
4341                 kfree(entry);
4342         }
4343 }