fs/btrfs/qgroup.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2011 STRATO.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/writeback.h>
   9 #include <linux/blkdev.h>
  10 #include <linux/rbtree.h>
  11 #include <linux/slab.h>
  12 #include <linux/workqueue.h>
  13 #include <linux/btrfs.h>
  14 #include <linux/sched/mm.h>
  15
  16 #include "ctree.h"
  17 #include "transaction.h"
  18 #include "disk-io.h"
  19 #include "locking.h"
  20 #include "ulist.h"
  21 #include "backref.h"
  22 #include "extent_io.h"
  23 #include "qgroup.h"
  24 #include "block-group.h"
  25 #include "sysfs.h"
  26 #include "tree-mod-log.h"
  27
  28 /* TODO XXX FIXME
  29  *  - subvol delete -> delete when ref goes to 0? delete limits also?
  30  *  - reorganize keys
  31  *  - compressed
  32  *  - sync
  33  *  - copy also limits on subvol creation
  34  *  - limit
  35  *  - caches for ulists
  36  *  - performance benchmarks
  37  *  - check all ioctl parameters
  38  */
  39
  40 /*
  41  * Helpers to access qgroup reservation
  42  *
  43  * Callers should ensure the lock context and type are valid
  44  */
  45
  46 static u64 qgroup_rsv_total(const struct btrfs_qgroup *qgroup)
  47 {
  48         u64 ret = 0;
  49         int i;
  50
  51         for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++)
  52                 ret += qgroup->rsv.values[i];
  53
  54         return ret;
  55 }
  56
  57 #ifdef CONFIG_BTRFS_DEBUG
  58 static const char *qgroup_rsv_type_str(enum btrfs_qgroup_rsv_type type)
  59 {
  60         if (type == BTRFS_QGROUP_RSV_DATA)
  61                 return "data";
  62         if (type == BTRFS_QGROUP_RSV_META_PERTRANS)
  63                 return "meta_pertrans";
  64         if (type == BTRFS_QGROUP_RSV_META_PREALLOC)
  65                 return "meta_prealloc";
  66         return NULL;
  67 }
  68 #endif
  69
  70 static void qgroup_rsv_add(struct btrfs_fs_info *fs_info,
  71                            struct btrfs_qgroup *qgroup, u64 num_bytes,
  72                            enum btrfs_qgroup_rsv_type type)
  73 {
  74         trace_qgroup_update_reserve(fs_info, qgroup, num_bytes, type);
  75         qgroup->rsv.values[type] += num_bytes;
  76 }
  77
  78 static void qgroup_rsv_release(struct btrfs_fs_info *fs_info,
  79                                struct btrfs_qgroup *qgroup, u64 num_bytes,
  80                                enum btrfs_qgroup_rsv_type type)
  81 {
  82         trace_qgroup_update_reserve(fs_info, qgroup, -(s64)num_bytes, type);
  83         if (qgroup->rsv.values[type] >= num_bytes) {
  84                 qgroup->rsv.values[type] -= num_bytes;
  85                 return;
  86         }
  87 #ifdef CONFIG_BTRFS_DEBUG
  88         WARN_RATELIMIT(1,
  89                 "qgroup %llu %s reserved space underflow, have %llu to free %llu",
  90                 qgroup->qgroupid, qgroup_rsv_type_str(type),
  91                 qgroup->rsv.values[type], num_bytes);
  92 #endif
  93         qgroup->rsv.values[type] = 0;
  94 }
  95
  96 static void qgroup_rsv_add_by_qgroup(struct btrfs_fs_info *fs_info,
  97                                      struct btrfs_qgroup *dest,
  98                                      struct btrfs_qgroup *src)
  99 {
 100         int i;
 101
 102         for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++)
 103                 qgroup_rsv_add(fs_info, dest, src->rsv.values[i], i);
 104 }
 105
 106 static void qgroup_rsv_release_by_qgroup(struct btrfs_fs_info *fs_info,
 107                                          struct btrfs_qgroup *dest,
 108                                           struct btrfs_qgroup *src)
 109 {
 110         int i;
 111
 112         for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++)
 113                 qgroup_rsv_release(fs_info, dest, src->rsv.values[i], i);
 114 }
 115
 116 static void btrfs_qgroup_update_old_refcnt(struct btrfs_qgroup *qg, u64 seq,
 117                                            int mod)
 118 {
 119         if (qg->old_refcnt < seq)
 120                 qg->old_refcnt = seq;
 121         qg->old_refcnt += mod;
 122 }
 123
 124 static void btrfs_qgroup_update_new_refcnt(struct btrfs_qgroup *qg, u64 seq,
 125                                            int mod)
 126 {
 127         if (qg->new_refcnt < seq)
 128                 qg->new_refcnt = seq;
 129         qg->new_refcnt += mod;
 130 }
 131
 132 static inline u64 btrfs_qgroup_get_old_refcnt(struct btrfs_qgroup *qg, u64 seq)
 133 {
 134         if (qg->old_refcnt < seq)
 135                 return 0;
 136         return qg->old_refcnt - seq;
 137 }
 138
 139 static inline u64 btrfs_qgroup_get_new_refcnt(struct btrfs_qgroup *qg, u64 seq)
 140 {
 141         if (qg->new_refcnt < seq)
 142                 return 0;
 143         return qg->new_refcnt - seq;
 144 }
 145
 146 /*
 147  * glue structure to represent the relations between qgroups.
 148  */
 149 struct btrfs_qgroup_list {
 150         struct list_head next_group;
 151         struct list_head next_member;
 152         struct btrfs_qgroup *group;
 153         struct btrfs_qgroup *member;
 154 };
 155
 156 static inline u64 qgroup_to_aux(struct btrfs_qgroup *qg)
 157 {
 158         return (u64)(uintptr_t)qg;
 159 }
 160
 161 static inline struct btrfs_qgroup* unode_aux_to_qgroup(struct ulist_node *n)
 162 {
 163         return (struct btrfs_qgroup *)(uintptr_t)n->aux;
 164 }
 165
 166 static int
 167 qgroup_rescan_init(struct btrfs_fs_info *fs_info, u64 progress_objectid,
 168                    int init_flags);
 169 static void qgroup_rescan_zero_tracking(struct btrfs_fs_info *fs_info);
 170
 171 /* must be called with qgroup_ioctl_lock held */
 172 static struct btrfs_qgroup *find_qgroup_rb(struct btrfs_fs_info *fs_info,
 173                                            u64 qgroupid)
 174 {
 175         struct rb_node *n = fs_info->qgroup_tree.rb_node;
 176         struct btrfs_qgroup *qgroup;
 177
 178         while (n) {
 179                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
 180                 if (qgroup->qgroupid < qgroupid)
 181                         n = n->rb_left;
 182                 else if (qgroup->qgroupid > qgroupid)
 183                         n = n->rb_right;
 184                 else
 185                         return qgroup;
 186         }
 187         return NULL;
 188 }
 189
 190 /* must be called with qgroup_lock held */
 191 static struct btrfs_qgroup *add_qgroup_rb(struct btrfs_fs_info *fs_info,
 192                                           u64 qgroupid)
 193 {
 194         struct rb_node **p = &fs_info->qgroup_tree.rb_node;
 195         struct rb_node *parent = NULL;
 196         struct btrfs_qgroup *qgroup;
 197
 198         while (*p) {
 199                 parent = *p;
 200                 qgroup = rb_entry(parent, struct btrfs_qgroup, node);
 201
 202                 if (qgroup->qgroupid < qgroupid)
 203                         p = &(*p)->rb_left;
 204                 else if (qgroup->qgroupid > qgroupid)
 205                         p = &(*p)->rb_right;
 206                 else
 207                         return qgroup;
 208         }
 209
 210         qgroup = kzalloc(sizeof(*qgroup), GFP_ATOMIC);
 211         if (!qgroup)
 212                 return ERR_PTR(-ENOMEM);
 213
 214         qgroup->qgroupid = qgroupid;
 215         INIT_LIST_HEAD(&qgroup->groups);
 216         INIT_LIST_HEAD(&qgroup->members);
 217         INIT_LIST_HEAD(&qgroup->dirty);
 218
 219         rb_link_node(&qgroup->node, parent, p);
 220         rb_insert_color(&qgroup->node, &fs_info->qgroup_tree);
 221
 222         return qgroup;
 223 }
 224
 225 static void __del_qgroup_rb(struct btrfs_fs_info *fs_info,
 226                             struct btrfs_qgroup *qgroup)
 227 {
 228         struct btrfs_qgroup_list *list;
 229
 230         list_del(&qgroup->dirty);
 231         while (!list_empty(&qgroup->groups)) {
 232                 list = list_first_entry(&qgroup->groups,
 233                                         struct btrfs_qgroup_list, next_group);
 234                 list_del(&list->next_group);
 235                 list_del(&list->next_member);
 236                 kfree(list);
 237         }
 238
 239         while (!list_empty(&qgroup->members)) {
 240                 list = list_first_entry(&qgroup->members,
 241                                         struct btrfs_qgroup_list, next_member);
 242                 list_del(&list->next_group);
 243                 list_del(&list->next_member);
 244                 kfree(list);
 245         }
 246 }
 247
 248 /* must be called with qgroup_lock held */
 249 static int del_qgroup_rb(struct btrfs_fs_info *fs_info, u64 qgroupid)
 250 {
 251         struct btrfs_qgroup *qgroup = find_qgroup_rb(fs_info, qgroupid);
 252
 253         if (!qgroup)
 254                 return -ENOENT;
 255
 256         rb_erase(&qgroup->node, &fs_info->qgroup_tree);
 257         __del_qgroup_rb(fs_info, qgroup);
 258         return 0;
 259 }
 260
 261 /* must be called with qgroup_lock held */
 262 static int add_relation_rb(struct btrfs_fs_info *fs_info,
 263                            u64 memberid, u64 parentid)
 264 {
 265         struct btrfs_qgroup *member;
 266         struct btrfs_qgroup *parent;
 267         struct btrfs_qgroup_list *list;
 268
 269         member = find_qgroup_rb(fs_info, memberid);
 270         parent = find_qgroup_rb(fs_info, parentid);
 271         if (!member || !parent)
 272                 return -ENOENT;
 273
 274         list = kzalloc(sizeof(*list), GFP_ATOMIC);
 275         if (!list)
 276                 return -ENOMEM;
 277
 278         list->group = parent;
 279         list->member = member;
 280         list_add_tail(&list->next_group, &member->groups);
 281         list_add_tail(&list->next_member, &parent->members);
 282
 283         return 0;
 284 }
 285
 286 /* must be called with qgroup_lock held */
 287 static int del_relation_rb(struct btrfs_fs_info *fs_info,
 288                            u64 memberid, u64 parentid)
 289 {
 290         struct btrfs_qgroup *member;
 291         struct btrfs_qgroup *parent;
 292         struct btrfs_qgroup_list *list;
 293
 294         member = find_qgroup_rb(fs_info, memberid);
 295         parent = find_qgroup_rb(fs_info, parentid);
 296         if (!member || !parent)
 297                 return -ENOENT;
 298
 299         list_for_each_entry(list, &member->groups, next_group) {
 300                 if (list->group == parent) {
 301                         list_del(&list->next_group);
 302                         list_del(&list->next_member);
 303                         kfree(list);
 304                         return 0;
 305                 }
 306         }
 307         return -ENOENT;
 308 }
 309
 310 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 311 int btrfs_verify_qgroup_counts(struct btrfs_fs_info *fs_info, u64 qgroupid,
 312                                u64 rfer, u64 excl)
 313 {
 314         struct btrfs_qgroup *qgroup;
 315
 316         qgroup = find_qgroup_rb(fs_info, qgroupid);
 317         if (!qgroup)
 318                 return -EINVAL;
 319         if (qgroup->rfer != rfer || qgroup->excl != excl)
 320                 return -EINVAL;
 321         return 0;
 322 }
 323 #endif
 324
 325 /*
 326  * The full config is read in one go, only called from open_ctree()
 327  * It doesn't use any locking, as at this point we're still single-threaded
 328  */
 329 int btrfs_read_qgroup_config(struct btrfs_fs_info *fs_info)
 330 {
 331         struct btrfs_key key;
 332         struct btrfs_key found_key;
 333         struct btrfs_root *quota_root = fs_info->quota_root;
 334         struct btrfs_path *path = NULL;
 335         struct extent_buffer *l;
 336         int slot;
 337         int ret = 0;
 338         u64 flags = 0;
 339         u64 rescan_progress = 0;
 340
 341         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
 342                 return 0;
 343
 344         fs_info->qgroup_ulist = ulist_alloc(GFP_KERNEL);
 345         if (!fs_info->qgroup_ulist) {
 346                 ret = -ENOMEM;
 347                 goto out;
 348         }
 349
 350         path = btrfs_alloc_path();
 351         if (!path) {
 352                 ret = -ENOMEM;
 353                 goto out;
 354         }
 355
 356         ret = btrfs_sysfs_add_qgroups(fs_info);
 357         if (ret < 0)
 358                 goto out;
 359         /* default this to quota off, in case no status key is found */
 360         fs_info->qgroup_flags = 0;
 361
 362         /*
 363          * pass 1: read status, all qgroup infos and limits
 364          */
 365         key.objectid = 0;
 366         key.type = 0;
 367         key.offset = 0;
 368         ret = btrfs_search_slot_for_read(quota_root, &key, path, 1, 1);
 369         if (ret)
 370                 goto out;
 371
 372         while (1) {
 373                 struct btrfs_qgroup *qgroup;
 374
 375                 slot = path->slots[0];
 376                 l = path->nodes[0];
 377                 btrfs_item_key_to_cpu(l, &found_key, slot);
 378
 379                 if (found_key.type == BTRFS_QGROUP_STATUS_KEY) {
 380                         struct btrfs_qgroup_status_item *ptr;
 381
 382                         ptr = btrfs_item_ptr(l, slot,
 383                                              struct btrfs_qgroup_status_item);
 384
 385                         if (btrfs_qgroup_status_version(l, ptr) !=
 386                             BTRFS_QGROUP_STATUS_VERSION) {
 387                                 btrfs_err(fs_info,
 388                                  "old qgroup version, quota disabled");
 389                                 goto out;
 390                         }
 391                         if (btrfs_qgroup_status_generation(l, ptr) !=
 392                             fs_info->generation) {
 393                                 flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
 394                                 btrfs_err(fs_info,
 395                                         "qgroup generation mismatch, marked as inconsistent");
 396                         }
 397                         fs_info->qgroup_flags = btrfs_qgroup_status_flags(l,
 398                                                                           ptr);
 399                         rescan_progress = btrfs_qgroup_status_rescan(l, ptr);
 400                         goto next1;
 401                 }
 402
 403                 if (found_key.type != BTRFS_QGROUP_INFO_KEY &&
 404                     found_key.type != BTRFS_QGROUP_LIMIT_KEY)
 405                         goto next1;
 406
 407                 qgroup = find_qgroup_rb(fs_info, found_key.offset);
 408                 if ((qgroup && found_key.type == BTRFS_QGROUP_INFO_KEY) ||
 409                     (!qgroup && found_key.type == BTRFS_QGROUP_LIMIT_KEY)) {
 410                         btrfs_err(fs_info, "inconsistent qgroup config");
 411                         flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
 412                 }
 413                 if (!qgroup) {
 414                         qgroup = add_qgroup_rb(fs_info, found_key.offset);
 415                         if (IS_ERR(qgroup)) {
 416                                 ret = PTR_ERR(qgroup);
 417                                 goto out;
 418                         }
 419                 }
 420                 ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
 421                 if (ret < 0)
 422                         goto out;
 423
 424                 switch (found_key.type) {
 425                 case BTRFS_QGROUP_INFO_KEY: {
 426                         struct btrfs_qgroup_info_item *ptr;
 427
 428                         ptr = btrfs_item_ptr(l, slot,
 429                                              struct btrfs_qgroup_info_item);
 430                         qgroup->rfer = btrfs_qgroup_info_rfer(l, ptr);
 431                         qgroup->rfer_cmpr = btrfs_qgroup_info_rfer_cmpr(l, ptr);
 432                         qgroup->excl = btrfs_qgroup_info_excl(l, ptr);
 433                         qgroup->excl_cmpr = btrfs_qgroup_info_excl_cmpr(l, ptr);
 434                         /* generation currently unused */
 435                         break;
 436                 }
 437                 case BTRFS_QGROUP_LIMIT_KEY: {
 438                         struct btrfs_qgroup_limit_item *ptr;
 439
 440                         ptr = btrfs_item_ptr(l, slot,
 441                                              struct btrfs_qgroup_limit_item);
 442                         qgroup->lim_flags = btrfs_qgroup_limit_flags(l, ptr);
 443                         qgroup->max_rfer = btrfs_qgroup_limit_max_rfer(l, ptr);
 444                         qgroup->max_excl = btrfs_qgroup_limit_max_excl(l, ptr);
 445                         qgroup->rsv_rfer = btrfs_qgroup_limit_rsv_rfer(l, ptr);
 446                         qgroup->rsv_excl = btrfs_qgroup_limit_rsv_excl(l, ptr);
 447                         break;
 448                 }
 449                 }
 450 next1:
 451                 ret = btrfs_next_item(quota_root, path);
 452                 if (ret < 0)
 453                         goto out;
 454                 if (ret)
 455                         break;
 456         }
 457         btrfs_release_path(path);
 458
 459         /*
 460          * pass 2: read all qgroup relations
 461          */
 462         key.objectid = 0;
 463         key.type = BTRFS_QGROUP_RELATION_KEY;
 464         key.offset = 0;
 465         ret = btrfs_search_slot_for_read(quota_root, &key, path, 1, 0);
 466         if (ret)
 467                 goto out;
 468         while (1) {
 469                 slot = path->slots[0];
 470                 l = path->nodes[0];
 471                 btrfs_item_key_to_cpu(l, &found_key, slot);
 472
 473                 if (found_key.type != BTRFS_QGROUP_RELATION_KEY)
 474                         goto next2;
 475
 476                 if (found_key.objectid > found_key.offset) {
 477                         /* parent <- member, not needed to build config */
 478                         /* FIXME should we omit the key completely? */
 479                         goto next2;
 480                 }
 481
 482                 ret = add_relation_rb(fs_info, found_key.objectid,
 483                                       found_key.offset);
 484                 if (ret == -ENOENT) {
 485                         btrfs_warn(fs_info,
 486                                 "orphan qgroup relation 0x%llx->0x%llx",
 487                                 found_key.objectid, found_key.offset);
 488                         ret = 0;        /* ignore the error */
 489                 }
 490                 if (ret)
 491                         goto out;
 492 next2:
 493                 ret = btrfs_next_item(quota_root, path);
 494                 if (ret < 0)
 495                         goto out;
 496                 if (ret)
 497                         break;
 498         }
 499 out:
 500         btrfs_free_path(path);
 501         fs_info->qgroup_flags |= flags;
 502         if (!(fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_ON))
 503                 clear_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
 504         else if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN &&
 505                  ret >= 0)
 506                 ret = qgroup_rescan_init(fs_info, rescan_progress, 0);
 507
 508         if (ret < 0) {
 509                 ulist_free(fs_info->qgroup_ulist);
 510                 fs_info->qgroup_ulist = NULL;
 511                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
 512                 btrfs_sysfs_del_qgroups(fs_info);
 513         }
 514
 515         return ret < 0 ? ret : 0;
 516 }
 517
 518 /*
 519  * Called in close_ctree() when quota is still enabled.  This verifies we don't
 520  * leak some reserved space.
 521  *
 522  * Return false if no reserved space is left.
 523  * Return true if some reserved space is leaked.
 524  */
 525 bool btrfs_check_quota_leak(struct btrfs_fs_info *fs_info)
 526 {
 527         struct rb_node *node;
 528         bool ret = false;
 529
 530         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
 531                 return ret;
 532         /*
 533          * Since we're unmounting, there is no race and no need to grab qgroup
 534          * lock.  And here we don't go post-order to provide a more user
 535          * friendly sorted result.
 536          */
 537         for (node = rb_first(&fs_info->qgroup_tree); node; node = rb_next(node)) {
 538                 struct btrfs_qgroup *qgroup;
 539                 int i;
 540
 541                 qgroup = rb_entry(node, struct btrfs_qgroup, node);
 542                 for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++) {
 543                         if (qgroup->rsv.values[i]) {
 544                                 ret = true;
 545                                 btrfs_warn(fs_info,
 546                 "qgroup %hu/%llu has unreleased space, type %d rsv %llu",
 547                                    btrfs_qgroup_level(qgroup->qgroupid),
 548                                    btrfs_qgroup_subvolid(qgroup->qgroupid),
 549                                    i, qgroup->rsv.values[i]);
 550                         }
 551                 }
 552         }
 553         return ret;
 554 }
 555
 556 /*
 557  * This is called from close_ctree() or open_ctree() or btrfs_quota_disable(),
 558  * first two are in single-threaded paths.And for the third one, we have set
 559  * quota_root to be null with qgroup_lock held before, so it is safe to clean
 560  * up the in-memory structures without qgroup_lock held.
 561  */
 562 void btrfs_free_qgroup_config(struct btrfs_fs_info *fs_info)
 563 {
 564         struct rb_node *n;
 565         struct btrfs_qgroup *qgroup;
 566
 567         while ((n = rb_first(&fs_info->qgroup_tree))) {
 568                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
 569                 rb_erase(n, &fs_info->qgroup_tree);
 570                 __del_qgroup_rb(fs_info, qgroup);
 571                 btrfs_sysfs_del_one_qgroup(fs_info, qgroup);
 572                 kfree(qgroup);
 573         }
 574         /*
 575          * We call btrfs_free_qgroup_config() when unmounting
 576          * filesystem and disabling quota, so we set qgroup_ulist
 577          * to be null here to avoid double free.
 578          */
 579         ulist_free(fs_info->qgroup_ulist);
 580         fs_info->qgroup_ulist = NULL;
 581         btrfs_sysfs_del_qgroups(fs_info);
 582 }
 583
 584 static int add_qgroup_relation_item(struct btrfs_trans_handle *trans, u64 src,
 585                                     u64 dst)
 586 {
 587         int ret;
 588         struct btrfs_root *quota_root = trans->fs_info->quota_root;
 589         struct btrfs_path *path;
 590         struct btrfs_key key;
 591
 592         path = btrfs_alloc_path();
 593         if (!path)
 594                 return -ENOMEM;
 595
 596         key.objectid = src;
 597         key.type = BTRFS_QGROUP_RELATION_KEY;
 598         key.offset = dst;
 599
 600         ret = btrfs_insert_empty_item(trans, quota_root, path, &key, 0);
 601
 602         btrfs_mark_buffer_dirty(path->nodes[0]);
 603
 604         btrfs_free_path(path);
 605         return ret;
 606 }
 607
 608 static int del_qgroup_relation_item(struct btrfs_trans_handle *trans, u64 src,
 609                                     u64 dst)
 610 {
 611         int ret;
 612         struct btrfs_root *quota_root = trans->fs_info->quota_root;
 613         struct btrfs_path *path;
 614         struct btrfs_key key;
 615
 616         path = btrfs_alloc_path();
 617         if (!path)
 618                 return -ENOMEM;
 619
 620         key.objectid = src;
 621         key.type = BTRFS_QGROUP_RELATION_KEY;
 622         key.offset = dst;
 623
 624         ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
 625         if (ret < 0)
 626                 goto out;
 627
 628         if (ret > 0) {
 629                 ret = -ENOENT;
 630                 goto out;
 631         }
 632
 633         ret = btrfs_del_item(trans, quota_root, path);
 634 out:
 635         btrfs_free_path(path);
 636         return ret;
 637 }
 638
 639 static int add_qgroup_item(struct btrfs_trans_handle *trans,
 640                            struct btrfs_root *quota_root, u64 qgroupid)
 641 {
 642         int ret;
 643         struct btrfs_path *path;
 644         struct btrfs_qgroup_info_item *qgroup_info;
 645         struct btrfs_qgroup_limit_item *qgroup_limit;
 646         struct extent_buffer *leaf;
 647         struct btrfs_key key;
 648
 649         if (btrfs_is_testing(quota_root->fs_info))
 650                 return 0;
 651
 652         path = btrfs_alloc_path();
 653         if (!path)
 654                 return -ENOMEM;
 655
 656         key.objectid = 0;
 657         key.type = BTRFS_QGROUP_INFO_KEY;
 658         key.offset = qgroupid;
 659
 660         /*
 661          * Avoid a transaction abort by catching -EEXIST here. In that
 662          * case, we proceed by re-initializing the existing structure
 663          * on disk.
 664          */
 665
 666         ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
 667                                       sizeof(*qgroup_info));
 668         if (ret && ret != -EEXIST)
 669                 goto out;
 670
 671         leaf = path->nodes[0];
 672         qgroup_info = btrfs_item_ptr(leaf, path->slots[0],
 673                                  struct btrfs_qgroup_info_item);
 674         btrfs_set_qgroup_info_generation(leaf, qgroup_info, trans->transid);
 675         btrfs_set_qgroup_info_rfer(leaf, qgroup_info, 0);
 676         btrfs_set_qgroup_info_rfer_cmpr(leaf, qgroup_info, 0);
 677         btrfs_set_qgroup_info_excl(leaf, qgroup_info, 0);
 678         btrfs_set_qgroup_info_excl_cmpr(leaf, qgroup_info, 0);
 679
 680         btrfs_mark_buffer_dirty(leaf);
 681
 682         btrfs_release_path(path);
 683
 684         key.type = BTRFS_QGROUP_LIMIT_KEY;
 685         ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
 686                                       sizeof(*qgroup_limit));
 687         if (ret && ret != -EEXIST)
 688                 goto out;
 689
 690         leaf = path->nodes[0];
 691         qgroup_limit = btrfs_item_ptr(leaf, path->slots[0],
 692                                   struct btrfs_qgroup_limit_item);
 693         btrfs_set_qgroup_limit_flags(leaf, qgroup_limit, 0);
 694         btrfs_set_qgroup_limit_max_rfer(leaf, qgroup_limit, 0);
 695         btrfs_set_qgroup_limit_max_excl(leaf, qgroup_limit, 0);
 696         btrfs_set_qgroup_limit_rsv_rfer(leaf, qgroup_limit, 0);
 697         btrfs_set_qgroup_limit_rsv_excl(leaf, qgroup_limit, 0);
 698
 699         btrfs_mark_buffer_dirty(leaf);
 700
 701         ret = 0;
 702 out:
 703         btrfs_free_path(path);
 704         return ret;
 705 }
 706
 707 static int del_qgroup_item(struct btrfs_trans_handle *trans, u64 qgroupid)
 708 {
 709         int ret;
 710         struct btrfs_root *quota_root = trans->fs_info->quota_root;
 711         struct btrfs_path *path;
 712         struct btrfs_key key;
 713
 714         path = btrfs_alloc_path();
 715         if (!path)
 716                 return -ENOMEM;
 717
 718         key.objectid = 0;
 719         key.type = BTRFS_QGROUP_INFO_KEY;
 720         key.offset = qgroupid;
 721         ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
 722         if (ret < 0)
 723                 goto out;
 724
 725         if (ret > 0) {
 726                 ret = -ENOENT;
 727                 goto out;
 728         }
 729
 730         ret = btrfs_del_item(trans, quota_root, path);
 731         if (ret)
 732                 goto out;
 733
 734         btrfs_release_path(path);
 735
 736         key.type = BTRFS_QGROUP_LIMIT_KEY;
 737         ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
 738         if (ret < 0)
 739                 goto out;
 740
 741         if (ret > 0) {
 742                 ret = -ENOENT;
 743                 goto out;
 744         }
 745
 746         ret = btrfs_del_item(trans, quota_root, path);
 747
 748 out:
 749         btrfs_free_path(path);
 750         return ret;
 751 }
 752
 753 static int update_qgroup_limit_item(struct btrfs_trans_handle *trans,
 754                                     struct btrfs_qgroup *qgroup)
 755 {
 756         struct btrfs_root *quota_root = trans->fs_info->quota_root;
 757         struct btrfs_path *path;
 758         struct btrfs_key key;
 759         struct extent_buffer *l;
 760         struct btrfs_qgroup_limit_item *qgroup_limit;
 761         int ret;
 762         int slot;
 763
 764         key.objectid = 0;
 765         key.type = BTRFS_QGROUP_LIMIT_KEY;
 766         key.offset = qgroup->qgroupid;
 767
 768         path = btrfs_alloc_path();
 769         if (!path)
 770                 return -ENOMEM;
 771
 772         ret = btrfs_search_slot(trans, quota_root, &key, path, 0, 1);
 773         if (ret > 0)
 774                 ret = -ENOENT;
 775
 776         if (ret)
 777                 goto out;
 778
 779         l = path->nodes[0];
 780         slot = path->slots[0];
 781         qgroup_limit = btrfs_item_ptr(l, slot, struct btrfs_qgroup_limit_item);
 782         btrfs_set_qgroup_limit_flags(l, qgroup_limit, qgroup->lim_flags);
 783         btrfs_set_qgroup_limit_max_rfer(l, qgroup_limit, qgroup->max_rfer);
 784         btrfs_set_qgroup_limit_max_excl(l, qgroup_limit, qgroup->max_excl);
 785         btrfs_set_qgroup_limit_rsv_rfer(l, qgroup_limit, qgroup->rsv_rfer);
 786         btrfs_set_qgroup_limit_rsv_excl(l, qgroup_limit, qgroup->rsv_excl);
 787
 788         btrfs_mark_buffer_dirty(l);
 789
 790 out:
 791         btrfs_free_path(path);
 792         return ret;
 793 }
 794
 795 static int update_qgroup_info_item(struct btrfs_trans_handle *trans,
 796                                    struct btrfs_qgroup *qgroup)
 797 {
 798         struct btrfs_fs_info *fs_info = trans->fs_info;
 799         struct btrfs_root *quota_root = fs_info->quota_root;
 800         struct btrfs_path *path;
 801         struct btrfs_key key;
 802         struct extent_buffer *l;
 803         struct btrfs_qgroup_info_item *qgroup_info;
 804         int ret;
 805         int slot;
 806
 807         if (btrfs_is_testing(fs_info))
 808                 return 0;
 809
 810         key.objectid = 0;
 811         key.type = BTRFS_QGROUP_INFO_KEY;
 812         key.offset = qgroup->qgroupid;
 813
 814         path = btrfs_alloc_path();
 815         if (!path)
 816                 return -ENOMEM;
 817
 818         ret = btrfs_search_slot(trans, quota_root, &key, path, 0, 1);
 819         if (ret > 0)
 820                 ret = -ENOENT;
 821
 822         if (ret)
 823                 goto out;
 824
 825         l = path->nodes[0];
 826         slot = path->slots[0];
 827         qgroup_info = btrfs_item_ptr(l, slot, struct btrfs_qgroup_info_item);
 828         btrfs_set_qgroup_info_generation(l, qgroup_info, trans->transid);
 829         btrfs_set_qgroup_info_rfer(l, qgroup_info, qgroup->rfer);
 830         btrfs_set_qgroup_info_rfer_cmpr(l, qgroup_info, qgroup->rfer_cmpr);
 831         btrfs_set_qgroup_info_excl(l, qgroup_info, qgroup->excl);
 832         btrfs_set_qgroup_info_excl_cmpr(l, qgroup_info, qgroup->excl_cmpr);
 833
 834         btrfs_mark_buffer_dirty(l);
 835
 836 out:
 837         btrfs_free_path(path);
 838         return ret;
 839 }
 840
 841 static int update_qgroup_status_item(struct btrfs_trans_handle *trans)
 842 {
 843         struct btrfs_fs_info *fs_info = trans->fs_info;
 844         struct btrfs_root *quota_root = fs_info->quota_root;
 845         struct btrfs_path *path;
 846         struct btrfs_key key;
 847         struct extent_buffer *l;
 848         struct btrfs_qgroup_status_item *ptr;
 849         int ret;
 850         int slot;
 851
 852         key.objectid = 0;
 853         key.type = BTRFS_QGROUP_STATUS_KEY;
 854         key.offset = 0;
 855
 856         path = btrfs_alloc_path();
 857         if (!path)
 858                 return -ENOMEM;
 859
 860         ret = btrfs_search_slot(trans, quota_root, &key, path, 0, 1);
 861         if (ret > 0)
 862                 ret = -ENOENT;
 863
 864         if (ret)
 865                 goto out;
 866
 867         l = path->nodes[0];
 868         slot = path->slots[0];
 869         ptr = btrfs_item_ptr(l, slot, struct btrfs_qgroup_status_item);
 870         btrfs_set_qgroup_status_flags(l, ptr, fs_info->qgroup_flags);
 871         btrfs_set_qgroup_status_generation(l, ptr, trans->transid);
 872         btrfs_set_qgroup_status_rescan(l, ptr,
 873                                 fs_info->qgroup_rescan_progress.objectid);
 874
 875         btrfs_mark_buffer_dirty(l);
 876
 877 out:
 878         btrfs_free_path(path);
 879         return ret;
 880 }
 881
 882 /*
 883  * called with qgroup_lock held
 884  */
 885 static int btrfs_clean_quota_tree(struct btrfs_trans_handle *trans,
 886                                   struct btrfs_root *root)
 887 {
 888         struct btrfs_path *path;
 889         struct btrfs_key key;
 890         struct extent_buffer *leaf = NULL;
 891         int ret;
 892         int nr = 0;
 893
 894         path = btrfs_alloc_path();
 895         if (!path)
 896                 return -ENOMEM;
 897
 898         key.objectid = 0;
 899         key.offset = 0;
 900         key.type = 0;
 901
 902         while (1) {
 903                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 904                 if (ret < 0)
 905                         goto out;
 906                 leaf = path->nodes[0];
 907                 nr = btrfs_header_nritems(leaf);
 908                 if (!nr)
 909                         break;
 910                 /*
 911                  * delete the leaf one by one
 912                  * since the whole tree is going
 913                  * to be deleted.
 914                  */
 915                 path->slots[0] = 0;
 916                 ret = btrfs_del_items(trans, root, path, 0, nr);
 917                 if (ret)
 918                         goto out;
 919
 920                 btrfs_release_path(path);
 921         }
 922         ret = 0;
 923 out:
 924         btrfs_free_path(path);
 925         return ret;
 926 }
 927
 928 int btrfs_quota_enable(struct btrfs_fs_info *fs_info)
 929 {
 930         struct btrfs_root *quota_root;
 931         struct btrfs_root *tree_root = fs_info->tree_root;
 932         struct btrfs_path *path = NULL;
 933         struct btrfs_qgroup_status_item *ptr;
 934         struct extent_buffer *leaf;
 935         struct btrfs_key key;
 936         struct btrfs_key found_key;
 937         struct btrfs_qgroup *qgroup = NULL;
 938         struct btrfs_trans_handle *trans = NULL;
 939         struct ulist *ulist = NULL;
 940         int ret = 0;
 941         int slot;
 942
 943         /*
 944          * We need to have subvol_sem write locked, to prevent races between
 945          * concurrent tasks trying to enable quotas, because we will unlock
 946          * and relock qgroup_ioctl_lock before setting fs_info->quota_root
 947          * and before setting BTRFS_FS_QUOTA_ENABLED.
 948          */
 949         lockdep_assert_held_write(&fs_info->subvol_sem);
 950
 951         mutex_lock(&fs_info->qgroup_ioctl_lock);
 952         if (fs_info->quota_root)
 953                 goto out;
 954
 955         ulist = ulist_alloc(GFP_KERNEL);
 956         if (!ulist) {
 957                 ret = -ENOMEM;
 958                 goto out;
 959         }
 960
 961         ret = btrfs_sysfs_add_qgroups(fs_info);
 962         if (ret < 0)
 963                 goto out;
 964
 965         /*
 966          * Unlock qgroup_ioctl_lock before starting the transaction. This is to
 967          * avoid lock acquisition inversion problems (reported by lockdep) between
 968          * qgroup_ioctl_lock and the vfs freeze semaphores, acquired when we
 969          * start a transaction.
 970          * After we started the transaction lock qgroup_ioctl_lock again and
 971          * check if someone else created the quota root in the meanwhile. If so,
 972          * just return success and release the transaction handle.
 973          *
 974          * Also we don't need to worry about someone else calling
 975          * btrfs_sysfs_add_qgroups() after we unlock and getting an error because
 976          * that function returns 0 (success) when the sysfs entries already exist.
 977          */
 978         mutex_unlock(&fs_info->qgroup_ioctl_lock);
 979
 980         /*
 981          * 1 for quota root item
 982          * 1 for BTRFS_QGROUP_STATUS item
 983          *
 984          * Yet we also need 2*n items for a QGROUP_INFO/QGROUP_LIMIT items
 985          * per subvolume. However those are not currently reserved since it
 986          * would be a lot of overkill.
 987          */
 988         trans = btrfs_start_transaction(tree_root, 2);
 989
 990         mutex_lock(&fs_info->qgroup_ioctl_lock);
 991         if (IS_ERR(trans)) {
 992                 ret = PTR_ERR(trans);
 993                 trans = NULL;
 994                 goto out;
 995         }
 996
 997         if (fs_info->quota_root)
 998                 goto out;
 999
1000         fs_info->qgroup_ulist = ulist;
1001         ulist = NULL;
1002
1003         /*
1004          * initially create the quota tree
1005          */
1006         quota_root = btrfs_create_tree(trans, BTRFS_QUOTA_TREE_OBJECTID);
1007         if (IS_ERR(quota_root)) {
1008                 ret =  PTR_ERR(quota_root);
1009                 btrfs_abort_transaction(trans, ret);
1010                 goto out;
1011         }
1012
1013         path = btrfs_alloc_path();
1014         if (!path) {
1015                 ret = -ENOMEM;
1016                 btrfs_abort_transaction(trans, ret);
1017                 goto out_free_root;
1018         }
1019
1020         key.objectid = 0;
1021         key.type = BTRFS_QGROUP_STATUS_KEY;
1022         key.offset = 0;
1023
1024         ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
1025                                       sizeof(*ptr));
1026         if (ret) {
1027                 btrfs_abort_transaction(trans, ret);
1028                 goto out_free_path;
1029         }
1030
1031         leaf = path->nodes[0];
1032         ptr = btrfs_item_ptr(leaf, path->slots[0],
1033                                  struct btrfs_qgroup_status_item);
1034         btrfs_set_qgroup_status_generation(leaf, ptr, trans->transid);
1035         btrfs_set_qgroup_status_version(leaf, ptr, BTRFS_QGROUP_STATUS_VERSION);
1036         fs_info->qgroup_flags = BTRFS_QGROUP_STATUS_FLAG_ON |
1037                                 BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1038         btrfs_set_qgroup_status_flags(leaf, ptr, fs_info->qgroup_flags);
1039         btrfs_set_qgroup_status_rescan(leaf, ptr, 0);
1040
1041         btrfs_mark_buffer_dirty(leaf);
1042
1043         key.objectid = 0;
1044         key.type = BTRFS_ROOT_REF_KEY;
1045         key.offset = 0;
1046
1047         btrfs_release_path(path);
1048         ret = btrfs_search_slot_for_read(tree_root, &key, path, 1, 0);
1049         if (ret > 0)
1050                 goto out_add_root;
1051         if (ret < 0) {
1052                 btrfs_abort_transaction(trans, ret);
1053                 goto out_free_path;
1054         }
1055
1056         while (1) {
1057                 slot = path->slots[0];
1058                 leaf = path->nodes[0];
1059                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1060
1061                 if (found_key.type == BTRFS_ROOT_REF_KEY) {
1062
1063                         /* Release locks on tree_root before we access quota_root */
1064                         btrfs_release_path(path);
1065
1066                         ret = add_qgroup_item(trans, quota_root,
1067                                               found_key.offset);
1068                         if (ret) {
1069                                 btrfs_abort_transaction(trans, ret);
1070                                 goto out_free_path;
1071                         }
1072
1073                         qgroup = add_qgroup_rb(fs_info, found_key.offset);
1074                         if (IS_ERR(qgroup)) {
1075                                 ret = PTR_ERR(qgroup);
1076                                 btrfs_abort_transaction(trans, ret);
1077                                 goto out_free_path;
1078                         }
1079                         ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1080                         if (ret < 0) {
1081                                 btrfs_abort_transaction(trans, ret);
1082                                 goto out_free_path;
1083                         }
1084                         ret = btrfs_search_slot_for_read(tree_root, &found_key,
1085                                                          path, 1, 0);
1086                         if (ret < 0) {
1087                                 btrfs_abort_transaction(trans, ret);
1088                                 goto out_free_path;
1089                         }
1090                         if (ret > 0) {
1091                                 /*
1092                                  * Shouldn't happen, but in case it does we
1093                                  * don't need to do the btrfs_next_item, just
1094                                  * continue.
1095                                  */
1096                                 continue;
1097                         }
1098                 }
1099                 ret = btrfs_next_item(tree_root, path);
1100                 if (ret < 0) {
1101                         btrfs_abort_transaction(trans, ret);
1102                         goto out_free_path;
1103                 }
1104                 if (ret)
1105                         break;
1106         }
1107
1108 out_add_root:
1109         btrfs_release_path(path);
1110         ret = add_qgroup_item(trans, quota_root, BTRFS_FS_TREE_OBJECTID);
1111         if (ret) {
1112                 btrfs_abort_transaction(trans, ret);
1113                 goto out_free_path;
1114         }
1115
1116         qgroup = add_qgroup_rb(fs_info, BTRFS_FS_TREE_OBJECTID);
1117         if (IS_ERR(qgroup)) {
1118                 ret = PTR_ERR(qgroup);
1119                 btrfs_abort_transaction(trans, ret);
1120                 goto out_free_path;
1121         }
1122         ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1123         if (ret < 0) {
1124                 btrfs_abort_transaction(trans, ret);
1125                 goto out_free_path;
1126         }
1127
1128         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1129         /*
1130          * Commit the transaction while not holding qgroup_ioctl_lock, to avoid
1131          * a deadlock with tasks concurrently doing other qgroup operations, such
1132          * adding/removing qgroups or adding/deleting qgroup relations for example,
1133          * because all qgroup operations first start or join a transaction and then
1134          * lock the qgroup_ioctl_lock mutex.
1135          * We are safe from a concurrent task trying to enable quotas, by calling
1136          * this function, since we are serialized by fs_info->subvol_sem.
1137          */
1138         ret = btrfs_commit_transaction(trans);
1139         trans = NULL;
1140         mutex_lock(&fs_info->qgroup_ioctl_lock);
1141         if (ret)
1142                 goto out_free_path;
1143
1144         /*
1145          * Set quota enabled flag after committing the transaction, to avoid
1146          * deadlocks on fs_info->qgroup_ioctl_lock with concurrent snapshot
1147          * creation.
1148          */
1149         spin_lock(&fs_info->qgroup_lock);
1150         fs_info->quota_root = quota_root;
1151         set_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1152         spin_unlock(&fs_info->qgroup_lock);
1153
1154         ret = qgroup_rescan_init(fs_info, 0, 1);
1155         if (!ret) {
1156                 qgroup_rescan_zero_tracking(fs_info);
1157                 fs_info->qgroup_rescan_running = true;
1158                 btrfs_queue_work(fs_info->qgroup_rescan_workers,
1159                                  &fs_info->qgroup_rescan_work);
1160         }
1161
1162 out_free_path:
1163         btrfs_free_path(path);
1164 out_free_root:
1165         if (ret)
1166                 btrfs_put_root(quota_root);
1167 out:
1168         if (ret) {
1169                 ulist_free(fs_info->qgroup_ulist);
1170                 fs_info->qgroup_ulist = NULL;
1171                 btrfs_sysfs_del_qgroups(fs_info);
1172         }
1173         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1174         if (ret && trans)
1175                 btrfs_end_transaction(trans);
1176         else if (trans)
1177                 ret = btrfs_end_transaction(trans);
1178         ulist_free(ulist);
1179         return ret;
1180 }
1181
1182 int btrfs_quota_disable(struct btrfs_fs_info *fs_info)
1183 {
1184         struct btrfs_root *quota_root;
1185         struct btrfs_trans_handle *trans = NULL;
1186         int ret = 0;
1187
1188         /*
1189          * We need to have subvol_sem write locked, to prevent races between
1190          * concurrent tasks trying to disable quotas, because we will unlock
1191          * and relock qgroup_ioctl_lock across BTRFS_FS_QUOTA_ENABLED changes.
1192          */
1193         lockdep_assert_held_write(&fs_info->subvol_sem);
1194
1195         mutex_lock(&fs_info->qgroup_ioctl_lock);
1196         if (!fs_info->quota_root)
1197                 goto out;
1198
1199         /*
1200          * Request qgroup rescan worker to complete and wait for it. This wait
1201          * must be done before transaction start for quota disable since it may
1202          * deadlock with transaction by the qgroup rescan worker.
1203          */
1204         clear_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1205         btrfs_qgroup_wait_for_completion(fs_info, false);
1206         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1207
1208         /*
1209          * 1 For the root item
1210          *
1211          * We should also reserve enough items for the quota tree deletion in
1212          * btrfs_clean_quota_tree but this is not done.
1213          *
1214          * Also, we must always start a transaction without holding the mutex
1215          * qgroup_ioctl_lock, see btrfs_quota_enable().
1216          */
1217         trans = btrfs_start_transaction(fs_info->tree_root, 1);
1218
1219         mutex_lock(&fs_info->qgroup_ioctl_lock);
1220         if (IS_ERR(trans)) {
1221                 ret = PTR_ERR(trans);
1222                 trans = NULL;
1223                 set_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1224                 goto out;
1225         }
1226
1227         if (!fs_info->quota_root)
1228                 goto out;
1229
1230         spin_lock(&fs_info->qgroup_lock);
1231         quota_root = fs_info->quota_root;
1232         fs_info->quota_root = NULL;
1233         fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_ON;
1234         spin_unlock(&fs_info->qgroup_lock);
1235
1236         btrfs_free_qgroup_config(fs_info);
1237
1238         ret = btrfs_clean_quota_tree(trans, quota_root);
1239         if (ret) {
1240                 btrfs_abort_transaction(trans, ret);
1241                 goto out;
1242         }
1243
1244         ret = btrfs_del_root(trans, &quota_root->root_key);
1245         if (ret) {
1246                 btrfs_abort_transaction(trans, ret);
1247                 goto out;
1248         }
1249
1250         list_del(&quota_root->dirty_list);
1251
1252         btrfs_tree_lock(quota_root->node);
1253         btrfs_clean_tree_block(quota_root->node);
1254         btrfs_tree_unlock(quota_root->node);
1255         btrfs_free_tree_block(trans, btrfs_root_id(quota_root),
1256                               quota_root->node, 0, 1);
1257
1258         btrfs_put_root(quota_root);
1259
1260 out:
1261         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1262         if (ret && trans)
1263                 btrfs_end_transaction(trans);
1264         else if (trans)
1265                 ret = btrfs_end_transaction(trans);
1266
1267         return ret;
1268 }
1269
1270 static void qgroup_dirty(struct btrfs_fs_info *fs_info,
1271                          struct btrfs_qgroup *qgroup)
1272 {
1273         if (list_empty(&qgroup->dirty))
1274                 list_add(&qgroup->dirty, &fs_info->dirty_qgroups);
1275 }
1276
1277 /*
1278  * The easy accounting, we're updating qgroup relationship whose child qgroup
1279  * only has exclusive extents.
1280  *
1281  * In this case, all exclusive extents will also be exclusive for parent, so
1282  * excl/rfer just get added/removed.
1283  *
1284  * So is qgroup reservation space, which should also be added/removed to
1285  * parent.
1286  * Or when child tries to release reservation space, parent will underflow its
1287  * reservation (for relationship adding case).
1288  *
1289  * Caller should hold fs_info->qgroup_lock.
1290  */
1291 static int __qgroup_excl_accounting(struct btrfs_fs_info *fs_info,
1292                                     struct ulist *tmp, u64 ref_root,
1293                                     struct btrfs_qgroup *src, int sign)
1294 {
1295         struct btrfs_qgroup *qgroup;
1296         struct btrfs_qgroup_list *glist;
1297         struct ulist_node *unode;
1298         struct ulist_iterator uiter;
1299         u64 num_bytes = src->excl;
1300         int ret = 0;
1301
1302         qgroup = find_qgroup_rb(fs_info, ref_root);
1303         if (!qgroup)
1304                 goto out;
1305
1306         qgroup->rfer += sign * num_bytes;
1307         qgroup->rfer_cmpr += sign * num_bytes;
1308
1309         WARN_ON(sign < 0 && qgroup->excl < num_bytes);
1310         qgroup->excl += sign * num_bytes;
1311         qgroup->excl_cmpr += sign * num_bytes;
1312
1313         if (sign > 0)
1314                 qgroup_rsv_add_by_qgroup(fs_info, qgroup, src);
1315         else
1316                 qgroup_rsv_release_by_qgroup(fs_info, qgroup, src);
1317
1318         qgroup_dirty(fs_info, qgroup);
1319
1320         /* Get all of the parent groups that contain this qgroup */
1321         list_for_each_entry(glist, &qgroup->groups, next_group) {
1322                 ret = ulist_add(tmp, glist->group->qgroupid,
1323                                 qgroup_to_aux(glist->group), GFP_ATOMIC);
1324                 if (ret < 0)
1325                         goto out;
1326         }
1327
1328         /* Iterate all of the parents and adjust their reference counts */
1329         ULIST_ITER_INIT(&uiter);
1330         while ((unode = ulist_next(tmp, &uiter))) {
1331                 qgroup = unode_aux_to_qgroup(unode);
1332                 qgroup->rfer += sign * num_bytes;
1333                 qgroup->rfer_cmpr += sign * num_bytes;
1334                 WARN_ON(sign < 0 && qgroup->excl < num_bytes);
1335                 qgroup->excl += sign * num_bytes;
1336                 if (sign > 0)
1337                         qgroup_rsv_add_by_qgroup(fs_info, qgroup, src);
1338                 else
1339                         qgroup_rsv_release_by_qgroup(fs_info, qgroup, src);
1340                 qgroup->excl_cmpr += sign * num_bytes;
1341                 qgroup_dirty(fs_info, qgroup);
1342
1343                 /* Add any parents of the parents */
1344                 list_for_each_entry(glist, &qgroup->groups, next_group) {
1345                         ret = ulist_add(tmp, glist->group->qgroupid,
1346                                         qgroup_to_aux(glist->group), GFP_ATOMIC);
1347                         if (ret < 0)
1348                                 goto out;
1349                 }
1350         }
1351         ret = 0;
1352 out:
1353         return ret;
1354 }
1355
1356
1357 /*
1358  * Quick path for updating qgroup with only excl refs.
1359  *
1360  * In that case, just update all parent will be enough.
1361  * Or we needs to do a full rescan.
1362  * Caller should also hold fs_info->qgroup_lock.
1363  *
1364  * Return 0 for quick update, return >0 for need to full rescan
1365  * and mark INCONSISTENT flag.
1366  * Return < 0 for other error.
1367  */
1368 static int quick_update_accounting(struct btrfs_fs_info *fs_info,
1369                                    struct ulist *tmp, u64 src, u64 dst,
1370                                    int sign)
1371 {
1372         struct btrfs_qgroup *qgroup;
1373         int ret = 1;
1374         int err = 0;
1375
1376         qgroup = find_qgroup_rb(fs_info, src);
1377         if (!qgroup)
1378                 goto out;
1379         if (qgroup->excl == qgroup->rfer) {
1380                 ret = 0;
1381                 err = __qgroup_excl_accounting(fs_info, tmp, dst,
1382                                                qgroup, sign);
1383                 if (err < 0) {
1384                         ret = err;
1385                         goto out;
1386                 }
1387         }
1388 out:
1389         if (ret)
1390                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1391         return ret;
1392 }
1393
1394 int btrfs_add_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1395                               u64 dst)
1396 {
1397         struct btrfs_fs_info *fs_info = trans->fs_info;
1398         struct btrfs_qgroup *parent;
1399         struct btrfs_qgroup *member;
1400         struct btrfs_qgroup_list *list;
1401         struct ulist *tmp;
1402         unsigned int nofs_flag;
1403         int ret = 0;
1404
1405         /* Check the level of src and dst first */
1406         if (btrfs_qgroup_level(src) >= btrfs_qgroup_level(dst))
1407                 return -EINVAL;
1408
1409         /* We hold a transaction handle open, must do a NOFS allocation. */
1410         nofs_flag = memalloc_nofs_save();
1411         tmp = ulist_alloc(GFP_KERNEL);
1412         memalloc_nofs_restore(nofs_flag);
1413         if (!tmp)
1414                 return -ENOMEM;
1415
1416         mutex_lock(&fs_info->qgroup_ioctl_lock);
1417         if (!fs_info->quota_root) {
1418                 ret = -ENOTCONN;
1419                 goto out;
1420         }
1421         member = find_qgroup_rb(fs_info, src);
1422         parent = find_qgroup_rb(fs_info, dst);
1423         if (!member || !parent) {
1424                 ret = -EINVAL;
1425                 goto out;
1426         }
1427
1428         /* check if such qgroup relation exist firstly */
1429         list_for_each_entry(list, &member->groups, next_group) {
1430                 if (list->group == parent) {
1431                         ret = -EEXIST;
1432                         goto out;
1433                 }
1434         }
1435
1436         ret = add_qgroup_relation_item(trans, src, dst);
1437         if (ret)
1438                 goto out;
1439
1440         ret = add_qgroup_relation_item(trans, dst, src);
1441         if (ret) {
1442                 del_qgroup_relation_item(trans, src, dst);
1443                 goto out;
1444         }
1445
1446         spin_lock(&fs_info->qgroup_lock);
1447         ret = add_relation_rb(fs_info, src, dst);
1448         if (ret < 0) {
1449                 spin_unlock(&fs_info->qgroup_lock);
1450                 goto out;
1451         }
1452         ret = quick_update_accounting(fs_info, tmp, src, dst, 1);
1453         spin_unlock(&fs_info->qgroup_lock);
1454 out:
1455         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1456         ulist_free(tmp);
1457         return ret;
1458 }
1459
1460 static int __del_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1461                                  u64 dst)
1462 {
1463         struct btrfs_fs_info *fs_info = trans->fs_info;
1464         struct btrfs_qgroup *parent;
1465         struct btrfs_qgroup *member;
1466         struct btrfs_qgroup_list *list;
1467         struct ulist *tmp;
1468         bool found = false;
1469         unsigned int nofs_flag;
1470         int ret = 0;
1471         int ret2;
1472
1473         /* We hold a transaction handle open, must do a NOFS allocation. */
1474         nofs_flag = memalloc_nofs_save();
1475         tmp = ulist_alloc(GFP_KERNEL);
1476         memalloc_nofs_restore(nofs_flag);
1477         if (!tmp)
1478                 return -ENOMEM;
1479
1480         if (!fs_info->quota_root) {
1481                 ret = -ENOTCONN;
1482                 goto out;
1483         }
1484
1485         member = find_qgroup_rb(fs_info, src);
1486         parent = find_qgroup_rb(fs_info, dst);
1487         /*
1488          * The parent/member pair doesn't exist, then try to delete the dead
1489          * relation items only.
1490          */
1491         if (!member || !parent)
1492                 goto delete_item;
1493
1494         /* check if such qgroup relation exist firstly */
1495         list_for_each_entry(list, &member->groups, next_group) {
1496                 if (list->group == parent) {
1497                         found = true;
1498                         break;
1499                 }
1500         }
1501
1502 delete_item:
1503         ret = del_qgroup_relation_item(trans, src, dst);
1504         if (ret < 0 && ret != -ENOENT)
1505                 goto out;
1506         ret2 = del_qgroup_relation_item(trans, dst, src);
1507         if (ret2 < 0 && ret2 != -ENOENT)
1508                 goto out;
1509
1510         /* At least one deletion succeeded, return 0 */
1511         if (!ret || !ret2)
1512                 ret = 0;
1513
1514         if (found) {
1515                 spin_lock(&fs_info->qgroup_lock);
1516                 del_relation_rb(fs_info, src, dst);
1517                 ret = quick_update_accounting(fs_info, tmp, src, dst, -1);
1518                 spin_unlock(&fs_info->qgroup_lock);
1519         }
1520 out:
1521         ulist_free(tmp);
1522         return ret;
1523 }
1524
1525 int btrfs_del_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1526                               u64 dst)
1527 {
1528         struct btrfs_fs_info *fs_info = trans->fs_info;
1529         int ret = 0;
1530
1531         mutex_lock(&fs_info->qgroup_ioctl_lock);
1532         ret = __del_qgroup_relation(trans, src, dst);
1533         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1534
1535         return ret;
1536 }
1537
1538 int btrfs_create_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid)
1539 {
1540         struct btrfs_fs_info *fs_info = trans->fs_info;
1541         struct btrfs_root *quota_root;
1542         struct btrfs_qgroup *qgroup;
1543         int ret = 0;
1544
1545         mutex_lock(&fs_info->qgroup_ioctl_lock);
1546         if (!fs_info->quota_root) {
1547                 ret = -ENOTCONN;
1548                 goto out;
1549         }
1550         quota_root = fs_info->quota_root;
1551         qgroup = find_qgroup_rb(fs_info, qgroupid);
1552         if (qgroup) {
1553                 ret = -EEXIST;
1554                 goto out;
1555         }
1556
1557         ret = add_qgroup_item(trans, quota_root, qgroupid);
1558         if (ret)
1559                 goto out;
1560
1561         spin_lock(&fs_info->qgroup_lock);
1562         qgroup = add_qgroup_rb(fs_info, qgroupid);
1563         spin_unlock(&fs_info->qgroup_lock);
1564
1565         if (IS_ERR(qgroup)) {
1566                 ret = PTR_ERR(qgroup);
1567                 goto out;
1568         }
1569         ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1570 out:
1571         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1572         return ret;
1573 }
1574
1575 int btrfs_remove_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid)
1576 {
1577         struct btrfs_fs_info *fs_info = trans->fs_info;
1578         struct btrfs_qgroup *qgroup;
1579         struct btrfs_qgroup_list *list;
1580         int ret = 0;
1581
1582         mutex_lock(&fs_info->qgroup_ioctl_lock);
1583         if (!fs_info->quota_root) {
1584                 ret = -ENOTCONN;
1585                 goto out;
1586         }
1587
1588         qgroup = find_qgroup_rb(fs_info, qgroupid);
1589         if (!qgroup) {
1590                 ret = -ENOENT;
1591                 goto out;
1592         }
1593
1594         /* Check if there are no children of this qgroup */
1595         if (!list_empty(&qgroup->members)) {
1596                 ret = -EBUSY;
1597                 goto out;
1598         }
1599
1600         ret = del_qgroup_item(trans, qgroupid);
1601         if (ret && ret != -ENOENT)
1602                 goto out;
1603
1604         while (!list_empty(&qgroup->groups)) {
1605                 list = list_first_entry(&qgroup->groups,
1606                                         struct btrfs_qgroup_list, next_group);
1607                 ret = __del_qgroup_relation(trans, qgroupid,
1608                                             list->group->qgroupid);
1609                 if (ret)
1610                         goto out;
1611         }
1612
1613         spin_lock(&fs_info->qgroup_lock);
1614         del_qgroup_rb(fs_info, qgroupid);
1615         spin_unlock(&fs_info->qgroup_lock);
1616
1617         /*
1618          * Remove the qgroup from sysfs now without holding the qgroup_lock
1619          * spinlock, since the sysfs_remove_group() function needs to take
1620          * the mutex kernfs_mutex through kernfs_remove_by_name_ns().
1621          */
1622         btrfs_sysfs_del_one_qgroup(fs_info, qgroup);
1623         kfree(qgroup);
1624 out:
1625         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1626         return ret;
1627 }
1628
1629 int btrfs_limit_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid,
1630                        struct btrfs_qgroup_limit *limit)
1631 {
1632         struct btrfs_fs_info *fs_info = trans->fs_info;
1633         struct btrfs_qgroup *qgroup;
1634         int ret = 0;
1635         /* Sometimes we would want to clear the limit on this qgroup.
1636          * To meet this requirement, we treat the -1 as a special value
1637          * which tell kernel to clear the limit on this qgroup.
1638          */
1639         const u64 CLEAR_VALUE = -1;
1640
1641         mutex_lock(&fs_info->qgroup_ioctl_lock);
1642         if (!fs_info->quota_root) {
1643                 ret = -ENOTCONN;
1644                 goto out;
1645         }
1646
1647         qgroup = find_qgroup_rb(fs_info, qgroupid);
1648         if (!qgroup) {
1649                 ret = -ENOENT;
1650                 goto out;
1651         }
1652
1653         spin_lock(&fs_info->qgroup_lock);
1654         if (limit->flags & BTRFS_QGROUP_LIMIT_MAX_RFER) {
1655                 if (limit->max_rfer == CLEAR_VALUE) {
1656                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_MAX_RFER;
1657                         limit->flags &= ~BTRFS_QGROUP_LIMIT_MAX_RFER;
1658                         qgroup->max_rfer = 0;
1659                 } else {
1660                         qgroup->max_rfer = limit->max_rfer;
1661                 }
1662         }
1663         if (limit->flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) {
1664                 if (limit->max_excl == CLEAR_VALUE) {
1665                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_MAX_EXCL;
1666                         limit->flags &= ~BTRFS_QGROUP_LIMIT_MAX_EXCL;
1667                         qgroup->max_excl = 0;
1668                 } else {
1669                         qgroup->max_excl = limit->max_excl;
1670                 }
1671         }
1672         if (limit->flags & BTRFS_QGROUP_LIMIT_RSV_RFER) {
1673                 if (limit->rsv_rfer == CLEAR_VALUE) {
1674                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_RSV_RFER;
1675                         limit->flags &= ~BTRFS_QGROUP_LIMIT_RSV_RFER;
1676                         qgroup->rsv_rfer = 0;
1677                 } else {
1678                         qgroup->rsv_rfer = limit->rsv_rfer;
1679                 }
1680         }
1681         if (limit->flags & BTRFS_QGROUP_LIMIT_RSV_EXCL) {
1682                 if (limit->rsv_excl == CLEAR_VALUE) {
1683                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_RSV_EXCL;
1684                         limit->flags &= ~BTRFS_QGROUP_LIMIT_RSV_EXCL;
1685                         qgroup->rsv_excl = 0;
1686                 } else {
1687                         qgroup->rsv_excl = limit->rsv_excl;
1688                 }
1689         }
1690         qgroup->lim_flags |= limit->flags;
1691
1692         spin_unlock(&fs_info->qgroup_lock);
1693
1694         ret = update_qgroup_limit_item(trans, qgroup);
1695         if (ret) {
1696                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1697                 btrfs_info(fs_info, "unable to update quota limit for %llu",
1698                        qgroupid);
1699         }
1700
1701 out:
1702         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1703         return ret;
1704 }
1705
1706 int btrfs_qgroup_trace_extent_nolock(struct btrfs_fs_info *fs_info,
1707                                 struct btrfs_delayed_ref_root *delayed_refs,
1708                                 struct btrfs_qgroup_extent_record *record)
1709 {
1710         struct rb_node **p = &delayed_refs->dirty_extent_root.rb_node;
1711         struct rb_node *parent_node = NULL;
1712         struct btrfs_qgroup_extent_record *entry;
1713         u64 bytenr = record->bytenr;
1714
1715         lockdep_assert_held(&delayed_refs->lock);
1716         trace_btrfs_qgroup_trace_extent(fs_info, record);
1717
1718         while (*p) {
1719                 parent_node = *p;
1720                 entry = rb_entry(parent_node, struct btrfs_qgroup_extent_record,
1721                                  node);
1722                 if (bytenr < entry->bytenr) {
1723                         p = &(*p)->rb_left;
1724                 } else if (bytenr > entry->bytenr) {
1725                         p = &(*p)->rb_right;
1726                 } else {
1727                         if (record->data_rsv && !entry->data_rsv) {
1728                                 entry->data_rsv = record->data_rsv;
1729                                 entry->data_rsv_refroot =
1730                                         record->data_rsv_refroot;
1731                         }
1732                         return 1;
1733                 }
1734         }
1735
1736         rb_link_node(&record->node, parent_node, p);
1737         rb_insert_color(&record->node, &delayed_refs->dirty_extent_root);
1738         return 0;
1739 }
1740
1741 int btrfs_qgroup_trace_extent_post(struct btrfs_trans_handle *trans,
1742                                    struct btrfs_qgroup_extent_record *qrecord)
1743 {
1744         struct ulist *old_root;
1745         u64 bytenr = qrecord->bytenr;
1746         int ret;
1747
1748         /*
1749          * We are always called in a context where we are already holding a
1750          * transaction handle. Often we are called when adding a data delayed
1751          * reference from btrfs_truncate_inode_items() (truncating or unlinking),
1752          * in which case we will be holding a write lock on extent buffer from a
1753          * subvolume tree. In this case we can't allow btrfs_find_all_roots() to
1754          * acquire fs_info->commit_root_sem, because that is a higher level lock
1755          * that must be acquired before locking any extent buffers.
1756          *
1757          * So we want btrfs_find_all_roots() to not acquire the commit_root_sem
1758          * but we can't pass it a non-NULL transaction handle, because otherwise
1759          * it would not use commit roots and would lock extent buffers, causing
1760          * a deadlock if it ends up trying to read lock the same extent buffer
1761          * that was previously write locked at btrfs_truncate_inode_items().
1762          *
1763          * So pass a NULL transaction handle to btrfs_find_all_roots() and
1764          * explicitly tell it to not acquire the commit_root_sem - if we are
1765          * holding a transaction handle we don't need its protection.
1766          */
1767         ASSERT(trans != NULL);
1768
1769         ret = btrfs_find_all_roots(NULL, trans->fs_info, bytenr, 0, &old_root,
1770                                    true);
1771         if (ret < 0) {
1772                 trans->fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1773                 btrfs_warn(trans->fs_info,
1774 "error accounting new delayed refs extent (err code: %d), quota inconsistent",
1775                         ret);
1776                 return 0;
1777         }
1778
1779         /*
1780          * Here we don't need to get the lock of
1781          * trans->transaction->delayed_refs, since inserted qrecord won't
1782          * be deleted, only qrecord->node may be modified (new qrecord insert)
1783          *
1784          * So modifying qrecord->old_roots is safe here
1785          */
1786         qrecord->old_roots = old_root;
1787         return 0;
1788 }
1789
1790 int btrfs_qgroup_trace_extent(struct btrfs_trans_handle *trans, u64 bytenr,
1791                               u64 num_bytes, gfp_t gfp_flag)
1792 {
1793         struct btrfs_fs_info *fs_info = trans->fs_info;
1794         struct btrfs_qgroup_extent_record *record;
1795         struct btrfs_delayed_ref_root *delayed_refs;
1796         int ret;
1797
1798         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)
1799             || bytenr == 0 || num_bytes == 0)
1800                 return 0;
1801         record = kzalloc(sizeof(*record), gfp_flag);
1802         if (!record)
1803                 return -ENOMEM;
1804
1805         delayed_refs = &trans->transaction->delayed_refs;
1806         record->bytenr = bytenr;
1807         record->num_bytes = num_bytes;
1808         record->old_roots = NULL;
1809
1810         spin_lock(&delayed_refs->lock);
1811         ret = btrfs_qgroup_trace_extent_nolock(fs_info, delayed_refs, record);
1812         spin_unlock(&delayed_refs->lock);
1813         if (ret > 0) {
1814                 kfree(record);
1815                 return 0;
1816         }
1817         return btrfs_qgroup_trace_extent_post(trans, record);
1818 }
1819
1820 int btrfs_qgroup_trace_leaf_items(struct btrfs_trans_handle *trans,
1821                                   struct extent_buffer *eb)
1822 {
1823         struct btrfs_fs_info *fs_info = trans->fs_info;
1824         int nr = btrfs_header_nritems(eb);
1825         int i, extent_type, ret;
1826         struct btrfs_key key;
1827         struct btrfs_file_extent_item *fi;
1828         u64 bytenr, num_bytes;
1829
1830         /* We can be called directly from walk_up_proc() */
1831         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
1832                 return 0;
1833
1834         for (i = 0; i < nr; i++) {
1835                 btrfs_item_key_to_cpu(eb, &key, i);
1836
1837                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1838                         continue;
1839
1840                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
1841                 /* filter out non qgroup-accountable extents  */
1842                 extent_type = btrfs_file_extent_type(eb, fi);
1843
1844                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
1845                         continue;
1846
1847                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
1848                 if (!bytenr)
1849                         continue;
1850
1851                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
1852
1853                 ret = btrfs_qgroup_trace_extent(trans, bytenr, num_bytes,
1854                                                 GFP_NOFS);
1855                 if (ret)
1856                         return ret;
1857         }
1858         cond_resched();
1859         return 0;
1860 }
1861
1862 /*
1863  * Walk up the tree from the bottom, freeing leaves and any interior
1864  * nodes which have had all slots visited. If a node (leaf or
1865  * interior) is freed, the node above it will have it's slot
1866  * incremented. The root node will never be freed.
1867  *
1868  * At the end of this function, we should have a path which has all
1869  * slots incremented to the next position for a search. If we need to
1870  * read a new node it will be NULL and the node above it will have the
1871  * correct slot selected for a later read.
1872  *
1873  * If we increment the root nodes slot counter past the number of
1874  * elements, 1 is returned to signal completion of the search.
1875  */
1876 static int adjust_slots_upwards(struct btrfs_path *path, int root_level)
1877 {
1878         int level = 0;
1879         int nr, slot;
1880         struct extent_buffer *eb;
1881
1882         if (root_level == 0)
1883                 return 1;
1884
1885         while (level <= root_level) {
1886                 eb = path->nodes[level];
1887                 nr = btrfs_header_nritems(eb);
1888                 path->slots[level]++;
1889                 slot = path->slots[level];
1890                 if (slot >= nr || level == 0) {
1891                         /*
1892                          * Don't free the root -  we will detect this
1893                          * condition after our loop and return a
1894                          * positive value for caller to stop walking the tree.
1895                          */
1896                         if (level != root_level) {
1897                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
1898                                 path->locks[level] = 0;
1899
1900                                 free_extent_buffer(eb);
1901                                 path->nodes[level] = NULL;
1902                                 path->slots[level] = 0;
1903                         }
1904                 } else {
1905                         /*
1906                          * We have a valid slot to walk back down
1907                          * from. Stop here so caller can process these
1908                          * new nodes.
1909                          */
1910                         break;
1911                 }
1912
1913                 level++;
1914         }
1915
1916         eb = path->nodes[root_level];
1917         if (path->slots[root_level] >= btrfs_header_nritems(eb))
1918                 return 1;
1919
1920         return 0;
1921 }
1922
1923 /*
1924  * Helper function to trace a subtree tree block swap.
1925  *
1926  * The swap will happen in highest tree block, but there may be a lot of
1927  * tree blocks involved.
1928  *
1929  * For example:
1930  *  OO = Old tree blocks
1931  *  NN = New tree blocks allocated during balance
1932  *
1933  *           File tree (257)                  Reloc tree for 257
1934  * L2              OO                                NN
1935  *               /    \                            /    \
1936  * L1          OO      OO (a)                    OO      NN (a)
1937  *            / \     / \                       / \     / \
1938  * L0       OO   OO OO   OO                   OO   OO NN   NN
1939  *                  (b)  (c)                          (b)  (c)
1940  *
1941  * When calling qgroup_trace_extent_swap(), we will pass:
1942  * @src_eb = OO(a)
1943  * @dst_path = [ nodes[1] = NN(a), nodes[0] = NN(c) ]
1944  * @dst_level = 0
1945  * @root_level = 1
1946  *
1947  * In that case, qgroup_trace_extent_swap() will search from OO(a) to
1948  * reach OO(c), then mark both OO(c) and NN(c) as qgroup dirty.
1949  *
1950  * The main work of qgroup_trace_extent_swap() can be split into 3 parts:
1951  *
1952  * 1) Tree search from @src_eb
1953  *    It should acts as a simplified btrfs_search_slot().
1954  *    The key for search can be extracted from @dst_path->nodes[dst_level]
1955  *    (first key).
1956  *
1957  * 2) Mark the final tree blocks in @src_path and @dst_path qgroup dirty
1958  *    NOTE: In above case, OO(a) and NN(a) won't be marked qgroup dirty.
1959  *    They should be marked during previous (@dst_level = 1) iteration.
1960  *
1961  * 3) Mark file extents in leaves dirty
1962  *    We don't have good way to pick out new file extents only.
1963  *    So we still follow the old method by scanning all file extents in
1964  *    the leave.
1965  *
1966  * This function can free us from keeping two paths, thus later we only need
1967  * to care about how to iterate all new tree blocks in reloc tree.
1968  */
1969 static int qgroup_trace_extent_swap(struct btrfs_trans_handle* trans,
1970                                     struct extent_buffer *src_eb,
1971                                     struct btrfs_path *dst_path,
1972                                     int dst_level, int root_level,
1973                                     bool trace_leaf)
1974 {
1975         struct btrfs_key key;
1976         struct btrfs_path *src_path;
1977         struct btrfs_fs_info *fs_info = trans->fs_info;
1978         u32 nodesize = fs_info->nodesize;
1979         int cur_level = root_level;
1980         int ret;
1981
1982         BUG_ON(dst_level > root_level);
1983         /* Level mismatch */
1984         if (btrfs_header_level(src_eb) != root_level)
1985                 return -EINVAL;
1986
1987         src_path = btrfs_alloc_path();
1988         if (!src_path) {
1989                 ret = -ENOMEM;
1990                 goto out;
1991         }
1992
1993         if (dst_level)
1994                 btrfs_node_key_to_cpu(dst_path->nodes[dst_level], &key, 0);
1995         else
1996                 btrfs_item_key_to_cpu(dst_path->nodes[dst_level], &key, 0);
1997
1998         /* For src_path */
1999         atomic_inc(&src_eb->refs);
2000         src_path->nodes[root_level] = src_eb;
2001         src_path->slots[root_level] = dst_path->slots[root_level];
2002         src_path->locks[root_level] = 0;
2003
2004         /* A simplified version of btrfs_search_slot() */
2005         while (cur_level >= dst_level) {
2006                 struct btrfs_key src_key;
2007                 struct btrfs_key dst_key;
2008
2009                 if (src_path->nodes[cur_level] == NULL) {
2010                         struct extent_buffer *eb;
2011                         int parent_slot;
2012
2013                         eb = src_path->nodes[cur_level + 1];
2014                         parent_slot = src_path->slots[cur_level + 1];
2015
2016                         eb = btrfs_read_node_slot(eb, parent_slot);
2017                         if (IS_ERR(eb)) {
2018                                 ret = PTR_ERR(eb);
2019                                 goto out;
2020                         }
2021
2022                         src_path->nodes[cur_level] = eb;
2023
2024                         btrfs_tree_read_lock(eb);
2025                         src_path->locks[cur_level] = BTRFS_READ_LOCK;
2026                 }
2027
2028                 src_path->slots[cur_level] = dst_path->slots[cur_level];
2029                 if (cur_level) {
2030                         btrfs_node_key_to_cpu(dst_path->nodes[cur_level],
2031                                         &dst_key, dst_path->slots[cur_level]);
2032                         btrfs_node_key_to_cpu(src_path->nodes[cur_level],
2033                                         &src_key, src_path->slots[cur_level]);
2034                 } else {
2035                         btrfs_item_key_to_cpu(dst_path->nodes[cur_level],
2036                                         &dst_key, dst_path->slots[cur_level]);
2037                         btrfs_item_key_to_cpu(src_path->nodes[cur_level],
2038                                         &src_key, src_path->slots[cur_level]);
2039                 }
2040                 /* Content mismatch, something went wrong */
2041                 if (btrfs_comp_cpu_keys(&dst_key, &src_key)) {
2042                         ret = -ENOENT;
2043                         goto out;
2044                 }
2045                 cur_level--;
2046         }
2047
2048         /*
2049          * Now both @dst_path and @src_path have been populated, record the tree
2050          * blocks for qgroup accounting.
2051          */
2052         ret = btrfs_qgroup_trace_extent(trans, src_path->nodes[dst_level]->start,
2053                         nodesize, GFP_NOFS);
2054         if (ret < 0)
2055                 goto out;
2056         ret = btrfs_qgroup_trace_extent(trans,
2057                         dst_path->nodes[dst_level]->start,
2058                         nodesize, GFP_NOFS);
2059         if (ret < 0)
2060                 goto out;
2061
2062         /* Record leaf file extents */
2063         if (dst_level == 0 && trace_leaf) {
2064                 ret = btrfs_qgroup_trace_leaf_items(trans, src_path->nodes[0]);
2065                 if (ret < 0)
2066                         goto out;
2067                 ret = btrfs_qgroup_trace_leaf_items(trans, dst_path->nodes[0]);
2068         }
2069 out:
2070         btrfs_free_path(src_path);
2071         return ret;
2072 }
2073
2074 /*
2075  * Helper function to do recursive generation-aware depth-first search, to
2076  * locate all new tree blocks in a subtree of reloc tree.
2077  *
2078  * E.g. (OO = Old tree blocks, NN = New tree blocks, whose gen == last_snapshot)
2079  *         reloc tree
2080  * L2         NN (a)
2081  *          /    \
2082  * L1    OO        NN (b)
2083  *      /  \      /  \
2084  * L0  OO  OO    OO  NN
2085  *               (c) (d)
2086  * If we pass:
2087  * @dst_path = [ nodes[1] = NN(b), nodes[0] = NULL ],
2088  * @cur_level = 1
2089  * @root_level = 1
2090  *
2091  * We will iterate through tree blocks NN(b), NN(d) and info qgroup to trace
2092  * above tree blocks along with their counter parts in file tree.
2093  * While during search, old tree blocks OO(c) will be skipped as tree block swap
2094  * won't affect OO(c).
2095  */
2096 static int qgroup_trace_new_subtree_blocks(struct btrfs_trans_handle* trans,
2097                                            struct extent_buffer *src_eb,
2098                                            struct btrfs_path *dst_path,
2099                                            int cur_level, int root_level,
2100                                            u64 last_snapshot, bool trace_leaf)
2101 {
2102         struct btrfs_fs_info *fs_info = trans->fs_info;
2103         struct extent_buffer *eb;
2104         bool need_cleanup = false;
2105         int ret = 0;
2106         int i;
2107
2108         /* Level sanity check */
2109         if (cur_level < 0 || cur_level >= BTRFS_MAX_LEVEL - 1 ||
2110             root_level < 0 || root_level >= BTRFS_MAX_LEVEL - 1 ||
2111             root_level < cur_level) {
2112                 btrfs_err_rl(fs_info,
2113                         "%s: bad levels, cur_level=%d root_level=%d",
2114                         __func__, cur_level, root_level);
2115                 return -EUCLEAN;
2116         }
2117
2118         /* Read the tree block if needed */
2119         if (dst_path->nodes[cur_level] == NULL) {
2120                 int parent_slot;
2121                 u64 child_gen;
2122
2123                 /*
2124                  * dst_path->nodes[root_level] must be initialized before
2125                  * calling this function.
2126                  */
2127                 if (cur_level == root_level) {
2128                         btrfs_err_rl(fs_info,
2129         "%s: dst_path->nodes[%d] not initialized, root_level=%d cur_level=%d",
2130                                 __func__, root_level, root_level, cur_level);
2131                         return -EUCLEAN;
2132                 }
2133
2134                 /*
2135                  * We need to get child blockptr/gen from parent before we can
2136                  * read it.
2137                   */
2138                 eb = dst_path->nodes[cur_level + 1];
2139                 parent_slot = dst_path->slots[cur_level + 1];
2140                 child_gen = btrfs_node_ptr_generation(eb, parent_slot);
2141
2142                 /* This node is old, no need to trace */
2143                 if (child_gen < last_snapshot)
2144                         goto out;
2145
2146                 eb = btrfs_read_node_slot(eb, parent_slot);
2147                 if (IS_ERR(eb)) {
2148                         ret = PTR_ERR(eb);
2149                         goto out;
2150                 }
2151
2152                 dst_path->nodes[cur_level] = eb;
2153                 dst_path->slots[cur_level] = 0;
2154
2155                 btrfs_tree_read_lock(eb);
2156                 dst_path->locks[cur_level] = BTRFS_READ_LOCK;
2157                 need_cleanup = true;
2158         }
2159
2160         /* Now record this tree block and its counter part for qgroups */
2161         ret = qgroup_trace_extent_swap(trans, src_eb, dst_path, cur_level,
2162                                        root_level, trace_leaf);
2163         if (ret < 0)
2164                 goto cleanup;
2165
2166         eb = dst_path->nodes[cur_level];
2167
2168         if (cur_level > 0) {
2169                 /* Iterate all child tree blocks */
2170                 for (i = 0; i < btrfs_header_nritems(eb); i++) {
2171                         /* Skip old tree blocks as they won't be swapped */
2172                         if (btrfs_node_ptr_generation(eb, i) < last_snapshot)
2173                                 continue;
2174                         dst_path->slots[cur_level] = i;
2175
2176                         /* Recursive call (at most 7 times) */
2177                         ret = qgroup_trace_new_subtree_blocks(trans, src_eb,
2178                                         dst_path, cur_level - 1, root_level,
2179                                         last_snapshot, trace_leaf);
2180                         if (ret < 0)
2181                                 goto cleanup;
2182                 }
2183         }
2184
2185 cleanup:
2186         if (need_cleanup) {
2187                 /* Clean up */
2188                 btrfs_tree_unlock_rw(dst_path->nodes[cur_level],
2189                                      dst_path->locks[cur_level]);
2190                 free_extent_buffer(dst_path->nodes[cur_level]);
2191                 dst_path->nodes[cur_level] = NULL;
2192                 dst_path->slots[cur_level] = 0;
2193                 dst_path->locks[cur_level] = 0;
2194         }
2195 out:
2196         return ret;
2197 }
2198
2199 static int qgroup_trace_subtree_swap(struct btrfs_trans_handle *trans,
2200                                 struct extent_buffer *src_eb,
2201                                 struct extent_buffer *dst_eb,
2202                                 u64 last_snapshot, bool trace_leaf)
2203 {
2204         struct btrfs_fs_info *fs_info = trans->fs_info;
2205         struct btrfs_path *dst_path = NULL;
2206         int level;
2207         int ret;
2208
2209         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2210                 return 0;
2211
2212         /* Wrong parameter order */
2213         if (btrfs_header_generation(src_eb) > btrfs_header_generation(dst_eb)) {
2214                 btrfs_err_rl(fs_info,
2215                 "%s: bad parameter order, src_gen=%llu dst_gen=%llu", __func__,
2216                              btrfs_header_generation(src_eb),
2217                              btrfs_header_generation(dst_eb));
2218                 return -EUCLEAN;
2219         }
2220
2221         if (!extent_buffer_uptodate(src_eb) || !extent_buffer_uptodate(dst_eb)) {
2222                 ret = -EIO;
2223                 goto out;
2224         }
2225
2226         level = btrfs_header_level(dst_eb);
2227         dst_path = btrfs_alloc_path();
2228         if (!dst_path) {
2229                 ret = -ENOMEM;
2230                 goto out;
2231         }
2232         /* For dst_path */
2233         atomic_inc(&dst_eb->refs);
2234         dst_path->nodes[level] = dst_eb;
2235         dst_path->slots[level] = 0;
2236         dst_path->locks[level] = 0;
2237
2238         /* Do the generation aware breadth-first search */
2239         ret = qgroup_trace_new_subtree_blocks(trans, src_eb, dst_path, level,
2240                                               level, last_snapshot, trace_leaf);
2241         if (ret < 0)
2242                 goto out;
2243         ret = 0;
2244
2245 out:
2246         btrfs_free_path(dst_path);
2247         if (ret < 0)
2248                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2249         return ret;
2250 }
2251
2252 int btrfs_qgroup_trace_subtree(struct btrfs_trans_handle *trans,
2253                                struct extent_buffer *root_eb,
2254                                u64 root_gen, int root_level)
2255 {
2256         struct btrfs_fs_info *fs_info = trans->fs_info;
2257         int ret = 0;
2258         int level;
2259         struct extent_buffer *eb = root_eb;
2260         struct btrfs_path *path = NULL;
2261
2262         BUG_ON(root_level < 0 || root_level >= BTRFS_MAX_LEVEL);
2263         BUG_ON(root_eb == NULL);
2264
2265         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2266                 return 0;
2267
2268         if (!extent_buffer_uptodate(root_eb)) {
2269                 ret = btrfs_read_buffer(root_eb, root_gen, root_level, NULL);
2270                 if (ret)
2271                         goto out;
2272         }
2273
2274         if (root_level == 0) {
2275                 ret = btrfs_qgroup_trace_leaf_items(trans, root_eb);
2276                 goto out;
2277         }
2278
2279         path = btrfs_alloc_path();
2280         if (!path)
2281                 return -ENOMEM;
2282
2283         /*
2284          * Walk down the tree.  Missing extent blocks are filled in as
2285          * we go. Metadata is accounted every time we read a new
2286          * extent block.
2287          *
2288          * When we reach a leaf, we account for file extent items in it,
2289          * walk back up the tree (adjusting slot pointers as we go)
2290          * and restart the search process.
2291          */
2292         atomic_inc(&root_eb->refs);     /* For path */
2293         path->nodes[root_level] = root_eb;
2294         path->slots[root_level] = 0;
2295         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
2296 walk_down:
2297         level = root_level;
2298         while (level >= 0) {
2299                 if (path->nodes[level] == NULL) {
2300                         int parent_slot;
2301                         u64 child_bytenr;
2302
2303                         /*
2304                          * We need to get child blockptr from parent before we
2305                          * can read it.
2306                           */
2307                         eb = path->nodes[level + 1];
2308                         parent_slot = path->slots[level + 1];
2309                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
2310
2311                         eb = btrfs_read_node_slot(eb, parent_slot);
2312                         if (IS_ERR(eb)) {
2313                                 ret = PTR_ERR(eb);
2314                                 goto out;
2315                         }
2316
2317                         path->nodes[level] = eb;
2318                         path->slots[level] = 0;
2319
2320                         btrfs_tree_read_lock(eb);
2321                         path->locks[level] = BTRFS_READ_LOCK;
2322
2323                         ret = btrfs_qgroup_trace_extent(trans, child_bytenr,
2324                                                         fs_info->nodesize,
2325                                                         GFP_NOFS);
2326                         if (ret)
2327                                 goto out;
2328                 }
2329
2330                 if (level == 0) {
2331                         ret = btrfs_qgroup_trace_leaf_items(trans,
2332                                                             path->nodes[level]);
2333                         if (ret)
2334                                 goto out;
2335
2336                         /* Nonzero return here means we completed our search */
2337                         ret = adjust_slots_upwards(path, root_level);
2338                         if (ret)
2339                                 break;
2340
2341                         /* Restart search with new slots */
2342                         goto walk_down;
2343                 }
2344
2345                 level--;
2346         }
2347
2348         ret = 0;
2349 out:
2350         btrfs_free_path(path);
2351
2352         return ret;
2353 }
2354
2355 #define UPDATE_NEW      0
2356 #define UPDATE_OLD      1
2357 /*
2358  * Walk all of the roots that points to the bytenr and adjust their refcnts.
2359  */
2360 static int qgroup_update_refcnt(struct btrfs_fs_info *fs_info,
2361                                 struct ulist *roots, struct ulist *tmp,
2362                                 struct ulist *qgroups, u64 seq, int update_old)
2363 {
2364         struct ulist_node *unode;
2365         struct ulist_iterator uiter;
2366         struct ulist_node *tmp_unode;
2367         struct ulist_iterator tmp_uiter;
2368         struct btrfs_qgroup *qg;
2369         int ret = 0;
2370
2371         if (!roots)
2372                 return 0;
2373         ULIST_ITER_INIT(&uiter);
2374         while ((unode = ulist_next(roots, &uiter))) {
2375                 qg = find_qgroup_rb(fs_info, unode->val);
2376                 if (!qg)
2377                         continue;
2378
2379                 ulist_reinit(tmp);
2380                 ret = ulist_add(qgroups, qg->qgroupid, qgroup_to_aux(qg),
2381                                 GFP_ATOMIC);
2382                 if (ret < 0)
2383                         return ret;
2384                 ret = ulist_add(tmp, qg->qgroupid, qgroup_to_aux(qg), GFP_ATOMIC);
2385                 if (ret < 0)
2386                         return ret;
2387                 ULIST_ITER_INIT(&tmp_uiter);
2388                 while ((tmp_unode = ulist_next(tmp, &tmp_uiter))) {
2389                         struct btrfs_qgroup_list *glist;
2390
2391                         qg = unode_aux_to_qgroup(tmp_unode);
2392                         if (update_old)
2393                                 btrfs_qgroup_update_old_refcnt(qg, seq, 1);
2394                         else
2395                                 btrfs_qgroup_update_new_refcnt(qg, seq, 1);
2396                         list_for_each_entry(glist, &qg->groups, next_group) {
2397                                 ret = ulist_add(qgroups, glist->group->qgroupid,
2398                                                 qgroup_to_aux(glist->group),
2399                                                 GFP_ATOMIC);
2400                                 if (ret < 0)
2401                                         return ret;
2402                                 ret = ulist_add(tmp, glist->group->qgroupid,
2403                                                 qgroup_to_aux(glist->group),
2404                                                 GFP_ATOMIC);
2405                                 if (ret < 0)
2406                                         return ret;
2407                         }
2408                 }
2409         }
2410         return 0;
2411 }
2412
2413 /*
2414  * Update qgroup rfer/excl counters.
2415  * Rfer update is easy, codes can explain themselves.
2416  *
2417  * Excl update is tricky, the update is split into 2 parts.
2418  * Part 1: Possible exclusive <-> sharing detect:
2419  *      |       A       |       !A      |
2420  *  -------------------------------------
2421  *  B   |       *       |       -       |
2422  *  -------------------------------------
2423  *  !B  |       +       |       **      |
2424  *  -------------------------------------
2425  *
2426  * Conditions:
2427  * A:   cur_old_roots < nr_old_roots    (not exclusive before)
2428  * !A:  cur_old_roots == nr_old_roots   (possible exclusive before)
2429  * B:   cur_new_roots < nr_new_roots    (not exclusive now)
2430  * !B:  cur_new_roots == nr_new_roots   (possible exclusive now)
2431  *
2432  * Results:
2433  * +: Possible sharing -> exclusive     -: Possible exclusive -> sharing
2434  * *: Definitely not changed.           **: Possible unchanged.
2435  *
2436  * For !A and !B condition, the exception is cur_old/new_roots == 0 case.
2437  *
2438  * To make the logic clear, we first use condition A and B to split
2439  * combination into 4 results.
2440  *
2441  * Then, for result "+" and "-", check old/new_roots == 0 case, as in them
2442  * only on variant maybe 0.
2443  *
2444  * Lastly, check result **, since there are 2 variants maybe 0, split them
2445  * again(2x2).
2446  * But this time we don't need to consider other things, the codes and logic
2447  * is easy to understand now.
2448  */
2449 static int qgroup_update_counters(struct btrfs_fs_info *fs_info,
2450                                   struct ulist *qgroups,
2451                                   u64 nr_old_roots,
2452                                   u64 nr_new_roots,
2453                                   u64 num_bytes, u64 seq)
2454 {
2455         struct ulist_node *unode;
2456         struct ulist_iterator uiter;
2457         struct btrfs_qgroup *qg;
2458         u64 cur_new_count, cur_old_count;
2459
2460         ULIST_ITER_INIT(&uiter);
2461         while ((unode = ulist_next(qgroups, &uiter))) {
2462                 bool dirty = false;
2463
2464                 qg = unode_aux_to_qgroup(unode);
2465                 cur_old_count = btrfs_qgroup_get_old_refcnt(qg, seq);
2466                 cur_new_count = btrfs_qgroup_get_new_refcnt(qg, seq);
2467
2468                 trace_qgroup_update_counters(fs_info, qg, cur_old_count,
2469                                              cur_new_count);
2470
2471                 /* Rfer update part */
2472                 if (cur_old_count == 0 && cur_new_count > 0) {
2473                         qg->rfer += num_bytes;
2474                         qg->rfer_cmpr += num_bytes;
2475                         dirty = true;
2476                 }
2477                 if (cur_old_count > 0 && cur_new_count == 0) {
2478                         qg->rfer -= num_bytes;
2479                         qg->rfer_cmpr -= num_bytes;
2480                         dirty = true;
2481                 }
2482
2483                 /* Excl update part */
2484                 /* Exclusive/none -> shared case */
2485                 if (cur_old_count == nr_old_roots &&
2486                     cur_new_count < nr_new_roots) {
2487                         /* Exclusive -> shared */
2488                         if (cur_old_count != 0) {
2489                                 qg->excl -= num_bytes;
2490                                 qg->excl_cmpr -= num_bytes;
2491                                 dirty = true;
2492                         }
2493                 }
2494
2495                 /* Shared -> exclusive/none case */
2496                 if (cur_old_count < nr_old_roots &&
2497                     cur_new_count == nr_new_roots) {
2498                         /* Shared->exclusive */
2499                         if (cur_new_count != 0) {
2500                                 qg->excl += num_bytes;
2501                                 qg->excl_cmpr += num_bytes;
2502                                 dirty = true;
2503                         }
2504                 }
2505
2506                 /* Exclusive/none -> exclusive/none case */
2507                 if (cur_old_count == nr_old_roots &&
2508                     cur_new_count == nr_new_roots) {
2509                         if (cur_old_count == 0) {
2510                                 /* None -> exclusive/none */
2511
2512                                 if (cur_new_count != 0) {
2513                                         /* None -> exclusive */
2514                                         qg->excl += num_bytes;
2515                                         qg->excl_cmpr += num_bytes;
2516                                         dirty = true;
2517                                 }
2518                                 /* None -> none, nothing changed */
2519                         } else {
2520                                 /* Exclusive -> exclusive/none */
2521
2522                                 if (cur_new_count == 0) {
2523                                         /* Exclusive -> none */
2524                                         qg->excl -= num_bytes;
2525                                         qg->excl_cmpr -= num_bytes;
2526                                         dirty = true;
2527                                 }
2528                                 /* Exclusive -> exclusive, nothing changed */
2529                         }
2530                 }
2531
2532                 if (dirty)
2533                         qgroup_dirty(fs_info, qg);
2534         }
2535         return 0;
2536 }
2537
2538 /*
2539  * Check if the @roots potentially is a list of fs tree roots
2540  *
2541  * Return 0 for definitely not a fs/subvol tree roots ulist
2542  * Return 1 for possible fs/subvol tree roots in the list (considering an empty
2543  *          one as well)
2544  */
2545 static int maybe_fs_roots(struct ulist *roots)
2546 {
2547         struct ulist_node *unode;
2548         struct ulist_iterator uiter;
2549
2550         /* Empty one, still possible for fs roots */
2551         if (!roots || roots->nnodes == 0)
2552                 return 1;
2553
2554         ULIST_ITER_INIT(&uiter);
2555         unode = ulist_next(roots, &uiter);
2556         if (!unode)
2557                 return 1;
2558
2559         /*
2560          * If it contains fs tree roots, then it must belong to fs/subvol
2561          * trees.
2562          * If it contains a non-fs tree, it won't be shared with fs/subvol trees.
2563          */
2564         return is_fstree(unode->val);
2565 }
2566
2567 int btrfs_qgroup_account_extent(struct btrfs_trans_handle *trans, u64 bytenr,
2568                                 u64 num_bytes, struct ulist *old_roots,
2569                                 struct ulist *new_roots)
2570 {
2571         struct btrfs_fs_info *fs_info = trans->fs_info;
2572         struct ulist *qgroups = NULL;
2573         struct ulist *tmp = NULL;
2574         u64 seq;
2575         u64 nr_new_roots = 0;
2576         u64 nr_old_roots = 0;
2577         int ret = 0;
2578
2579         /*
2580          * If quotas get disabled meanwhile, the resources need to be freed and
2581          * we can't just exit here.
2582          */
2583         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2584                 goto out_free;
2585
2586         if (new_roots) {
2587                 if (!maybe_fs_roots(new_roots))
2588                         goto out_free;
2589                 nr_new_roots = new_roots->nnodes;
2590         }
2591         if (old_roots) {
2592                 if (!maybe_fs_roots(old_roots))
2593                         goto out_free;
2594                 nr_old_roots = old_roots->nnodes;
2595         }
2596
2597         /* Quick exit, either not fs tree roots, or won't affect any qgroup */
2598         if (nr_old_roots == 0 && nr_new_roots == 0)
2599                 goto out_free;
2600
2601         BUG_ON(!fs_info->quota_root);
2602
2603         trace_btrfs_qgroup_account_extent(fs_info, trans->transid, bytenr,
2604                                         num_bytes, nr_old_roots, nr_new_roots);
2605
2606         qgroups = ulist_alloc(GFP_NOFS);
2607         if (!qgroups) {
2608                 ret = -ENOMEM;
2609                 goto out_free;
2610         }
2611         tmp = ulist_alloc(GFP_NOFS);
2612         if (!tmp) {
2613                 ret = -ENOMEM;
2614                 goto out_free;
2615         }
2616
2617         mutex_lock(&fs_info->qgroup_rescan_lock);
2618         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
2619                 if (fs_info->qgroup_rescan_progress.objectid <= bytenr) {
2620                         mutex_unlock(&fs_info->qgroup_rescan_lock);
2621                         ret = 0;
2622                         goto out_free;
2623                 }
2624         }
2625         mutex_unlock(&fs_info->qgroup_rescan_lock);
2626
2627         spin_lock(&fs_info->qgroup_lock);
2628         seq = fs_info->qgroup_seq;
2629
2630         /* Update old refcnts using old_roots */
2631         ret = qgroup_update_refcnt(fs_info, old_roots, tmp, qgroups, seq,
2632                                    UPDATE_OLD);
2633         if (ret < 0)
2634                 goto out;
2635
2636         /* Update new refcnts using new_roots */
2637         ret = qgroup_update_refcnt(fs_info, new_roots, tmp, qgroups, seq,
2638                                    UPDATE_NEW);
2639         if (ret < 0)
2640                 goto out;
2641
2642         qgroup_update_counters(fs_info, qgroups, nr_old_roots, nr_new_roots,
2643                                num_bytes, seq);
2644
2645         /*
2646          * Bump qgroup_seq to avoid seq overlap
2647          */
2648         fs_info->qgroup_seq += max(nr_old_roots, nr_new_roots) + 1;
2649 out:
2650         spin_unlock(&fs_info->qgroup_lock);
2651 out_free:
2652         ulist_free(tmp);
2653         ulist_free(qgroups);
2654         ulist_free(old_roots);
2655         ulist_free(new_roots);
2656         return ret;
2657 }
2658
2659 int btrfs_qgroup_account_extents(struct btrfs_trans_handle *trans)
2660 {
2661         struct btrfs_fs_info *fs_info = trans->fs_info;
2662         struct btrfs_qgroup_extent_record *record;
2663         struct btrfs_delayed_ref_root *delayed_refs;
2664         struct ulist *new_roots = NULL;
2665         struct rb_node *node;
2666         u64 num_dirty_extents = 0;
2667         u64 qgroup_to_skip;
2668         int ret = 0;
2669
2670         delayed_refs = &trans->transaction->delayed_refs;
2671         qgroup_to_skip = delayed_refs->qgroup_to_skip;
2672         while ((node = rb_first(&delayed_refs->dirty_extent_root))) {
2673                 record = rb_entry(node, struct btrfs_qgroup_extent_record,
2674                                   node);
2675
2676                 num_dirty_extents++;
2677                 trace_btrfs_qgroup_account_extents(fs_info, record);
2678
2679                 if (!ret) {
2680                         /*
2681                          * Old roots should be searched when inserting qgroup
2682                          * extent record
2683                          */
2684                         if (WARN_ON(!record->old_roots)) {
2685                                 /* Search commit root to find old_roots */
2686                                 ret = btrfs_find_all_roots(NULL, fs_info,
2687                                                 record->bytenr, 0,
2688                                                 &record->old_roots, false);
2689                                 if (ret < 0)
2690                                         goto cleanup;
2691                         }
2692
2693                         /* Free the reserved data space */
2694                         btrfs_qgroup_free_refroot(fs_info,
2695                                         record->data_rsv_refroot,
2696                                         record->data_rsv,
2697                                         BTRFS_QGROUP_RSV_DATA);
2698                         /*
2699                          * Use BTRFS_SEQ_LAST as time_seq to do special search,
2700                          * which doesn't lock tree or delayed_refs and search
2701                          * current root. It's safe inside commit_transaction().
2702                          */
2703                         ret = btrfs_find_all_roots(trans, fs_info,
2704                            record->bytenr, BTRFS_SEQ_LAST, &new_roots, false);
2705                         if (ret < 0)
2706                                 goto cleanup;
2707                         if (qgroup_to_skip) {
2708                                 ulist_del(new_roots, qgroup_to_skip, 0);
2709                                 ulist_del(record->old_roots, qgroup_to_skip,
2710                                           0);
2711                         }
2712                         ret = btrfs_qgroup_account_extent(trans, record->bytenr,
2713                                                           record->num_bytes,
2714                                                           record->old_roots,
2715                                                           new_roots);
2716                         record->old_roots = NULL;
2717                         new_roots = NULL;
2718                 }
2719 cleanup:
2720                 ulist_free(record->old_roots);
2721                 ulist_free(new_roots);
2722                 new_roots = NULL;
2723                 rb_erase(node, &delayed_refs->dirty_extent_root);
2724                 kfree(record);
2725
2726         }
2727         trace_qgroup_num_dirty_extents(fs_info, trans->transid,
2728                                        num_dirty_extents);
2729         return ret;
2730 }
2731
2732 /*
2733  * called from commit_transaction. Writes all changed qgroups to disk.
2734  */
2735 int btrfs_run_qgroups(struct btrfs_trans_handle *trans)
2736 {
2737         struct btrfs_fs_info *fs_info = trans->fs_info;
2738         int ret = 0;
2739
2740         if (!fs_info->quota_root)
2741                 return ret;
2742
2743         spin_lock(&fs_info->qgroup_lock);
2744         while (!list_empty(&fs_info->dirty_qgroups)) {
2745                 struct btrfs_qgroup *qgroup;
2746                 qgroup = list_first_entry(&fs_info->dirty_qgroups,
2747                                           struct btrfs_qgroup, dirty);
2748                 list_del_init(&qgroup->dirty);
2749                 spin_unlock(&fs_info->qgroup_lock);
2750                 ret = update_qgroup_info_item(trans, qgroup);
2751                 if (ret)
2752                         fs_info->qgroup_flags |=
2753                                         BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2754                 ret = update_qgroup_limit_item(trans, qgroup);
2755                 if (ret)
2756                         fs_info->qgroup_flags |=
2757                                         BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2758                 spin_lock(&fs_info->qgroup_lock);
2759         }
2760         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2761                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_ON;
2762         else
2763                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_ON;
2764         spin_unlock(&fs_info->qgroup_lock);
2765
2766         ret = update_qgroup_status_item(trans);
2767         if (ret)
2768                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2769
2770         return ret;
2771 }
2772
2773 /*
2774  * Copy the accounting information between qgroups. This is necessary
2775  * when a snapshot or a subvolume is created. Throwing an error will
2776  * cause a transaction abort so we take extra care here to only error
2777  * when a readonly fs is a reasonable outcome.
2778  */
2779 int btrfs_qgroup_inherit(struct btrfs_trans_handle *trans, u64 srcid,
2780                          u64 objectid, struct btrfs_qgroup_inherit *inherit)
2781 {
2782         int ret = 0;
2783         int i;
2784         u64 *i_qgroups;
2785         bool committing = false;
2786         struct btrfs_fs_info *fs_info = trans->fs_info;
2787         struct btrfs_root *quota_root;
2788         struct btrfs_qgroup *srcgroup;
2789         struct btrfs_qgroup *dstgroup;
2790         bool need_rescan = false;
2791         u32 level_size = 0;
2792         u64 nums;
2793
2794         /*
2795          * There are only two callers of this function.
2796          *
2797          * One in create_subvol() in the ioctl context, which needs to hold
2798          * the qgroup_ioctl_lock.
2799          *
2800          * The other one in create_pending_snapshot() where no other qgroup
2801          * code can modify the fs as they all need to either start a new trans
2802          * or hold a trans handler, thus we don't need to hold
2803          * qgroup_ioctl_lock.
2804          * This would avoid long and complex lock chain and make lockdep happy.
2805          */
2806         spin_lock(&fs_info->trans_lock);
2807         if (trans->transaction->state == TRANS_STATE_COMMIT_DOING)
2808                 committing = true;
2809         spin_unlock(&fs_info->trans_lock);
2810
2811         if (!committing)
2812                 mutex_lock(&fs_info->qgroup_ioctl_lock);
2813         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2814                 goto out;
2815
2816         quota_root = fs_info->quota_root;
2817         if (!quota_root) {
2818                 ret = -EINVAL;
2819                 goto out;
2820         }
2821
2822         if (inherit) {
2823                 i_qgroups = (u64 *)(inherit + 1);
2824                 nums = inherit->num_qgroups + 2 * inherit->num_ref_copies +
2825                        2 * inherit->num_excl_copies;
2826                 for (i = 0; i < nums; ++i) {
2827                         srcgroup = find_qgroup_rb(fs_info, *i_qgroups);
2828
2829                         /*
2830                          * Zero out invalid groups so we can ignore
2831                          * them later.
2832                          */
2833                         if (!srcgroup ||
2834                             ((srcgroup->qgroupid >> 48) <= (objectid >> 48)))
2835                                 *i_qgroups = 0ULL;
2836
2837                         ++i_qgroups;
2838                 }
2839         }
2840
2841         /*
2842          * create a tracking group for the subvol itself
2843          */
2844         ret = add_qgroup_item(trans, quota_root, objectid);
2845         if (ret)
2846                 goto out;
2847
2848         /*
2849          * add qgroup to all inherited groups
2850          */
2851         if (inherit) {
2852                 i_qgroups = (u64 *)(inherit + 1);
2853                 for (i = 0; i < inherit->num_qgroups; ++i, ++i_qgroups) {
2854                         if (*i_qgroups == 0)
2855                                 continue;
2856                         ret = add_qgroup_relation_item(trans, objectid,
2857                                                        *i_qgroups);
2858                         if (ret && ret != -EEXIST)
2859                                 goto out;
2860                         ret = add_qgroup_relation_item(trans, *i_qgroups,
2861                                                        objectid);
2862                         if (ret && ret != -EEXIST)
2863                                 goto out;
2864                 }
2865                 ret = 0;
2866         }
2867
2868
2869         spin_lock(&fs_info->qgroup_lock);
2870
2871         dstgroup = add_qgroup_rb(fs_info, objectid);
2872         if (IS_ERR(dstgroup)) {
2873                 ret = PTR_ERR(dstgroup);
2874                 goto unlock;
2875         }
2876
2877         if (inherit && inherit->flags & BTRFS_QGROUP_INHERIT_SET_LIMITS) {
2878                 dstgroup->lim_flags = inherit->lim.flags;
2879                 dstgroup->max_rfer = inherit->lim.max_rfer;
2880                 dstgroup->max_excl = inherit->lim.max_excl;
2881                 dstgroup->rsv_rfer = inherit->lim.rsv_rfer;
2882                 dstgroup->rsv_excl = inherit->lim.rsv_excl;
2883
2884                 ret = update_qgroup_limit_item(trans, dstgroup);
2885                 if (ret) {
2886                         fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2887                         btrfs_info(fs_info,
2888                                    "unable to update quota limit for %llu",
2889                                    dstgroup->qgroupid);
2890                         goto unlock;
2891                 }
2892         }
2893
2894         if (srcid) {
2895                 srcgroup = find_qgroup_rb(fs_info, srcid);
2896                 if (!srcgroup)
2897                         goto unlock;
2898
2899                 /*
2900                  * We call inherit after we clone the root in order to make sure
2901                  * our counts don't go crazy, so at this point the only
2902                  * difference between the two roots should be the root node.
2903                  */
2904                 level_size = fs_info->nodesize;
2905                 dstgroup->rfer = srcgroup->rfer;
2906                 dstgroup->rfer_cmpr = srcgroup->rfer_cmpr;
2907                 dstgroup->excl = level_size;
2908                 dstgroup->excl_cmpr = level_size;
2909                 srcgroup->excl = level_size;
2910                 srcgroup->excl_cmpr = level_size;
2911
2912                 /* inherit the limit info */
2913                 dstgroup->lim_flags = srcgroup->lim_flags;
2914                 dstgroup->max_rfer = srcgroup->max_rfer;
2915                 dstgroup->max_excl = srcgroup->max_excl;
2916                 dstgroup->rsv_rfer = srcgroup->rsv_rfer;
2917                 dstgroup->rsv_excl = srcgroup->rsv_excl;
2918
2919                 qgroup_dirty(fs_info, dstgroup);
2920                 qgroup_dirty(fs_info, srcgroup);
2921         }
2922
2923         if (!inherit)
2924                 goto unlock;
2925
2926         i_qgroups = (u64 *)(inherit + 1);
2927         for (i = 0; i < inherit->num_qgroups; ++i) {
2928                 if (*i_qgroups) {
2929                         ret = add_relation_rb(fs_info, objectid, *i_qgroups);
2930                         if (ret)
2931                                 goto unlock;
2932                 }
2933                 ++i_qgroups;
2934
2935                 /*
2936                  * If we're doing a snapshot, and adding the snapshot to a new
2937                  * qgroup, the numbers are guaranteed to be incorrect.
2938                  */
2939                 if (srcid)
2940                         need_rescan = true;
2941         }
2942
2943         for (i = 0; i <  inherit->num_ref_copies; ++i, i_qgroups += 2) {
2944                 struct btrfs_qgroup *src;
2945                 struct btrfs_qgroup *dst;
2946
2947                 if (!i_qgroups[0] || !i_qgroups[1])
2948                         continue;
2949
2950                 src = find_qgroup_rb(fs_info, i_qgroups[0]);
2951                 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
2952
2953                 if (!src || !dst) {
2954                         ret = -EINVAL;
2955                         goto unlock;
2956                 }
2957
2958                 dst->rfer = src->rfer - level_size;
2959                 dst->rfer_cmpr = src->rfer_cmpr - level_size;
2960
2961                 /* Manually tweaking numbers certainly needs a rescan */
2962                 need_rescan = true;
2963         }
2964         for (i = 0; i <  inherit->num_excl_copies; ++i, i_qgroups += 2) {
2965                 struct btrfs_qgroup *src;
2966                 struct btrfs_qgroup *dst;
2967
2968                 if (!i_qgroups[0] || !i_qgroups[1])
2969                         continue;
2970
2971                 src = find_qgroup_rb(fs_info, i_qgroups[0]);
2972                 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
2973
2974                 if (!src || !dst) {
2975                         ret = -EINVAL;
2976                         goto unlock;
2977                 }
2978
2979                 dst->excl = src->excl + level_size;
2980                 dst->excl_cmpr = src->excl_cmpr + level_size;
2981                 need_rescan = true;
2982         }
2983
2984 unlock:
2985         spin_unlock(&fs_info->qgroup_lock);
2986         if (!ret)
2987                 ret = btrfs_sysfs_add_one_qgroup(fs_info, dstgroup);
2988 out:
2989         if (!committing)
2990                 mutex_unlock(&fs_info->qgroup_ioctl_lock);
2991         if (need_rescan)
2992                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2993         return ret;
2994 }
2995
2996 static bool qgroup_check_limits(const struct btrfs_qgroup *qg, u64 num_bytes)
2997 {
2998         if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_RFER) &&
2999             qgroup_rsv_total(qg) + (s64)qg->rfer + num_bytes > qg->max_rfer)
3000                 return false;
3001
3002         if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) &&
3003             qgroup_rsv_total(qg) + (s64)qg->excl + num_bytes > qg->max_excl)
3004                 return false;
3005
3006         return true;
3007 }
3008
3009 static int qgroup_reserve(struct btrfs_root *root, u64 num_bytes, bool enforce,
3010                           enum btrfs_qgroup_rsv_type type)
3011 {
3012         struct btrfs_qgroup *qgroup;
3013         struct btrfs_fs_info *fs_info = root->fs_info;
3014         u64 ref_root = root->root_key.objectid;
3015         int ret = 0;
3016         struct ulist_node *unode;
3017         struct ulist_iterator uiter;
3018
3019         if (!is_fstree(ref_root))
3020                 return 0;
3021
3022         if (num_bytes == 0)
3023                 return 0;
3024
3025         if (test_bit(BTRFS_FS_QUOTA_OVERRIDE, &fs_info->flags) &&
3026             capable(CAP_SYS_RESOURCE))
3027                 enforce = false;
3028
3029         spin_lock(&fs_info->qgroup_lock);
3030         if (!fs_info->quota_root)
3031                 goto out;
3032
3033         qgroup = find_qgroup_rb(fs_info, ref_root);
3034         if (!qgroup)
3035                 goto out;
3036
3037         /*
3038          * in a first step, we check all affected qgroups if any limits would
3039          * be exceeded
3040          */
3041         ulist_reinit(fs_info->qgroup_ulist);
3042         ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
3043                         qgroup_to_aux(qgroup), GFP_ATOMIC);
3044         if (ret < 0)
3045                 goto out;
3046         ULIST_ITER_INIT(&uiter);
3047         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
3048                 struct btrfs_qgroup *qg;
3049                 struct btrfs_qgroup_list *glist;
3050
3051                 qg = unode_aux_to_qgroup(unode);
3052
3053                 if (enforce && !qgroup_check_limits(qg, num_bytes)) {
3054                         ret = -EDQUOT;
3055                         goto out;
3056                 }
3057
3058                 list_for_each_entry(glist, &qg->groups, next_group) {
3059                         ret = ulist_add(fs_info->qgroup_ulist,
3060                                         glist->group->qgroupid,
3061                                         qgroup_to_aux(glist->group), GFP_ATOMIC);
3062                         if (ret < 0)
3063                                 goto out;
3064                 }
3065         }
3066         ret = 0;
3067         /*
3068          * no limits exceeded, now record the reservation into all qgroups
3069          */
3070         ULIST_ITER_INIT(&uiter);
3071         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
3072                 struct btrfs_qgroup *qg;
3073
3074                 qg = unode_aux_to_qgroup(unode);
3075
3076                 qgroup_rsv_add(fs_info, qg, num_bytes, type);
3077         }
3078
3079 out:
3080         spin_unlock(&fs_info->qgroup_lock);
3081         return ret;
3082 }
3083
3084 /*
3085  * Free @num_bytes of reserved space with @type for qgroup.  (Normally level 0
3086  * qgroup).
3087  *
3088  * Will handle all higher level qgroup too.
3089  *
3090  * NOTE: If @num_bytes is (u64)-1, this means to free all bytes of this qgroup.
3091  * This special case is only used for META_PERTRANS type.
3092  */
3093 void btrfs_qgroup_free_refroot(struct btrfs_fs_info *fs_info,
3094                                u64 ref_root, u64 num_bytes,
3095                                enum btrfs_qgroup_rsv_type type)
3096 {
3097         struct btrfs_qgroup *qgroup;
3098         struct ulist_node *unode;
3099         struct ulist_iterator uiter;
3100         int ret = 0;
3101
3102         if (!is_fstree(ref_root))
3103                 return;
3104
3105         if (num_bytes == 0)
3106                 return;
3107
3108         if (num_bytes == (u64)-1 && type != BTRFS_QGROUP_RSV_META_PERTRANS) {
3109                 WARN(1, "%s: Invalid type to free", __func__);
3110                 return;
3111         }
3112         spin_lock(&fs_info->qgroup_lock);
3113
3114         if (!fs_info->quota_root)
3115                 goto out;
3116
3117         qgroup = find_qgroup_rb(fs_info, ref_root);
3118         if (!qgroup)
3119                 goto out;
3120
3121         if (num_bytes == (u64)-1)
3122                 /*
3123                  * We're freeing all pertrans rsv, get reserved value from
3124                  * level 0 qgroup as real num_bytes to free.
3125                  */
3126                 num_bytes = qgroup->rsv.values[type];
3127
3128         ulist_reinit(fs_info->qgroup_ulist);
3129         ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
3130                         qgroup_to_aux(qgroup), GFP_ATOMIC);
3131         if (ret < 0)
3132                 goto out;
3133         ULIST_ITER_INIT(&uiter);
3134         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
3135                 struct btrfs_qgroup *qg;
3136                 struct btrfs_qgroup_list *glist;
3137
3138                 qg = unode_aux_to_qgroup(unode);
3139
3140                 qgroup_rsv_release(fs_info, qg, num_bytes, type);
3141
3142                 list_for_each_entry(glist, &qg->groups, next_group) {
3143                         ret = ulist_add(fs_info->qgroup_ulist,
3144                                         glist->group->qgroupid,
3145                                         qgroup_to_aux(glist->group), GFP_ATOMIC);
3146                         if (ret < 0)
3147                                 goto out;
3148                 }
3149         }
3150
3151 out:
3152         spin_unlock(&fs_info->qgroup_lock);
3153 }
3154
3155 /*
3156  * Check if the leaf is the last leaf. Which means all node pointers
3157  * are at their last position.
3158  */
3159 static bool is_last_leaf(struct btrfs_path *path)
3160 {
3161         int i;
3162
3163         for (i = 1; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
3164                 if (path->slots[i] != btrfs_header_nritems(path->nodes[i]) - 1)
3165                         return false;
3166         }
3167         return true;
3168 }
3169
3170 /*
3171  * returns < 0 on error, 0 when more leafs are to be scanned.
3172  * returns 1 when done.
3173  */
3174 static int qgroup_rescan_leaf(struct btrfs_trans_handle *trans,
3175                               struct btrfs_path *path)
3176 {
3177         struct btrfs_fs_info *fs_info = trans->fs_info;
3178         struct btrfs_root *extent_root;
3179         struct btrfs_key found;
3180         struct extent_buffer *scratch_leaf = NULL;
3181         struct ulist *roots = NULL;
3182         u64 num_bytes;
3183         bool done;
3184         int slot;
3185         int ret;
3186
3187         mutex_lock(&fs_info->qgroup_rescan_lock);
3188         extent_root = btrfs_extent_root(fs_info,
3189                                 fs_info->qgroup_rescan_progress.objectid);
3190         ret = btrfs_search_slot_for_read(extent_root,
3191                                          &fs_info->qgroup_rescan_progress,
3192                                          path, 1, 0);
3193
3194         btrfs_debug(fs_info,
3195                 "current progress key (%llu %u %llu), search_slot ret %d",
3196                 fs_info->qgroup_rescan_progress.objectid,
3197                 fs_info->qgroup_rescan_progress.type,
3198                 fs_info->qgroup_rescan_progress.offset, ret);
3199
3200         if (ret) {
3201                 /*
3202                  * The rescan is about to end, we will not be scanning any
3203                  * further blocks. We cannot unset the RESCAN flag here, because
3204                  * we want to commit the transaction if everything went well.
3205                  * To make the live accounting work in this phase, we set our
3206                  * scan progress pointer such that every real extent objectid
3207                  * will be smaller.
3208                  */
3209                 fs_info->qgroup_rescan_progress.objectid = (u64)-1;
3210                 btrfs_release_path(path);
3211                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3212                 return ret;
3213         }
3214         done = is_last_leaf(path);
3215
3216         btrfs_item_key_to_cpu(path->nodes[0], &found,
3217                               btrfs_header_nritems(path->nodes[0]) - 1);
3218         fs_info->qgroup_rescan_progress.objectid = found.objectid + 1;
3219
3220         scratch_leaf = btrfs_clone_extent_buffer(path->nodes[0]);
3221         if (!scratch_leaf) {
3222                 ret = -ENOMEM;
3223                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3224                 goto out;
3225         }
3226         slot = path->slots[0];
3227         btrfs_release_path(path);
3228         mutex_unlock(&fs_info->qgroup_rescan_lock);
3229
3230         for (; slot < btrfs_header_nritems(scratch_leaf); ++slot) {
3231                 btrfs_item_key_to_cpu(scratch_leaf, &found, slot);
3232                 if (found.type != BTRFS_EXTENT_ITEM_KEY &&
3233                     found.type != BTRFS_METADATA_ITEM_KEY)
3234                         continue;
3235                 if (found.type == BTRFS_METADATA_ITEM_KEY)
3236                         num_bytes = fs_info->nodesize;
3237                 else
3238                         num_bytes = found.offset;
3239
3240                 ret = btrfs_find_all_roots(NULL, fs_info, found.objectid, 0,
3241                                            &roots, false);
3242                 if (ret < 0)
3243                         goto out;
3244                 /* For rescan, just pass old_roots as NULL */
3245                 ret = btrfs_qgroup_account_extent(trans, found.objectid,
3246                                                   num_bytes, NULL, roots);
3247                 if (ret < 0)
3248                         goto out;
3249         }
3250 out:
3251         if (scratch_leaf)
3252                 free_extent_buffer(scratch_leaf);
3253
3254         if (done && !ret) {
3255                 ret = 1;
3256                 fs_info->qgroup_rescan_progress.objectid = (u64)-1;
3257         }
3258         return ret;
3259 }
3260
3261 static bool rescan_should_stop(struct btrfs_fs_info *fs_info)
3262 {
3263         return btrfs_fs_closing(fs_info) ||
3264                 test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state);
3265 }
3266
3267 static void btrfs_qgroup_rescan_worker(struct btrfs_work *work)
3268 {
3269         struct btrfs_fs_info *fs_info = container_of(work, struct btrfs_fs_info,
3270                                                      qgroup_rescan_work);
3271         struct btrfs_path *path;
3272         struct btrfs_trans_handle *trans = NULL;
3273         int err = -ENOMEM;
3274         int ret = 0;
3275         bool stopped = false;
3276
3277         path = btrfs_alloc_path();
3278         if (!path)
3279                 goto out;
3280         /*
3281          * Rescan should only search for commit root, and any later difference
3282          * should be recorded by qgroup
3283          */
3284         path->search_commit_root = 1;
3285         path->skip_locking = 1;
3286
3287         err = 0;
3288         while (!err && !(stopped = rescan_should_stop(fs_info))) {
3289                 trans = btrfs_start_transaction(fs_info->fs_root, 0);
3290                 if (IS_ERR(trans)) {
3291                         err = PTR_ERR(trans);
3292                         break;
3293                 }
3294                 if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
3295                         err = -EINTR;
3296                 } else {
3297                         err = qgroup_rescan_leaf(trans, path);
3298                 }
3299                 if (err > 0)
3300                         btrfs_commit_transaction(trans);
3301                 else
3302                         btrfs_end_transaction(trans);
3303         }
3304
3305 out:
3306         btrfs_free_path(path);
3307
3308         mutex_lock(&fs_info->qgroup_rescan_lock);
3309         if (err > 0 &&
3310             fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT) {
3311                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
3312         } else if (err < 0) {
3313                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
3314         }
3315         mutex_unlock(&fs_info->qgroup_rescan_lock);
3316
3317         /*
3318          * only update status, since the previous part has already updated the
3319          * qgroup info.
3320          */
3321         trans = btrfs_start_transaction(fs_info->quota_root, 1);
3322         if (IS_ERR(trans)) {
3323                 err = PTR_ERR(trans);
3324                 trans = NULL;
3325                 btrfs_err(fs_info,
3326                           "fail to start transaction for status update: %d",
3327                           err);
3328         }
3329
3330         mutex_lock(&fs_info->qgroup_rescan_lock);
3331         if (!stopped)
3332                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3333         if (trans) {
3334                 ret = update_qgroup_status_item(trans);
3335                 if (ret < 0) {
3336                         err = ret;
3337                         btrfs_err(fs_info, "fail to update qgroup status: %d",
3338                                   err);
3339                 }
3340         }
3341         fs_info->qgroup_rescan_running = false;
3342         complete_all(&fs_info->qgroup_rescan_completion);
3343         mutex_unlock(&fs_info->qgroup_rescan_lock);
3344
3345         if (!trans)
3346                 return;
3347
3348         btrfs_end_transaction(trans);
3349
3350         if (stopped) {
3351                 btrfs_info(fs_info, "qgroup scan paused");
3352         } else if (err >= 0) {
3353                 btrfs_info(fs_info, "qgroup scan completed%s",
3354                         err > 0 ? " (inconsistency flag cleared)" : "");
3355         } else {
3356                 btrfs_err(fs_info, "qgroup scan failed with %d", err);
3357         }
3358 }
3359
3360 /*
3361  * Checks that (a) no rescan is running and (b) quota is enabled. Allocates all
3362  * memory required for the rescan context.
3363  */
3364 static int
3365 qgroup_rescan_init(struct btrfs_fs_info *fs_info, u64 progress_objectid,
3366                    int init_flags)
3367 {
3368         int ret = 0;
3369
3370         if (!init_flags) {
3371                 /* we're resuming qgroup rescan at mount time */
3372                 if (!(fs_info->qgroup_flags &
3373                       BTRFS_QGROUP_STATUS_FLAG_RESCAN)) {
3374                         btrfs_warn(fs_info,
3375                         "qgroup rescan init failed, qgroup rescan is not queued");
3376                         ret = -EINVAL;
3377                 } else if (!(fs_info->qgroup_flags &
3378                              BTRFS_QGROUP_STATUS_FLAG_ON)) {
3379                         btrfs_warn(fs_info,
3380                         "qgroup rescan init failed, qgroup is not enabled");
3381                         ret = -EINVAL;
3382                 }
3383
3384                 if (ret)
3385                         return ret;
3386         }
3387
3388         mutex_lock(&fs_info->qgroup_rescan_lock);
3389
3390         if (init_flags) {
3391                 if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
3392                         btrfs_warn(fs_info,
3393                                    "qgroup rescan is already in progress");
3394                         ret = -EINPROGRESS;
3395                 } else if (!(fs_info->qgroup_flags &
3396                              BTRFS_QGROUP_STATUS_FLAG_ON)) {
3397                         btrfs_warn(fs_info,
3398                         "qgroup rescan init failed, qgroup is not enabled");
3399                         ret = -EINVAL;
3400                 } else if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
3401                         /* Quota disable is in progress */
3402                         ret = -EBUSY;
3403                 }
3404
3405                 if (ret) {
3406                         mutex_unlock(&fs_info->qgroup_rescan_lock);
3407                         return ret;
3408                 }
3409                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3410         }
3411
3412         memset(&fs_info->qgroup_rescan_progress, 0,
3413                 sizeof(fs_info->qgroup_rescan_progress));
3414         fs_info->qgroup_rescan_progress.objectid = progress_objectid;
3415         init_completion(&fs_info->qgroup_rescan_completion);
3416         mutex_unlock(&fs_info->qgroup_rescan_lock);
3417
3418         btrfs_init_work(&fs_info->qgroup_rescan_work,
3419                         btrfs_qgroup_rescan_worker, NULL, NULL);
3420         return 0;
3421 }
3422
3423 static void
3424 qgroup_rescan_zero_tracking(struct btrfs_fs_info *fs_info)
3425 {
3426         struct rb_node *n;
3427         struct btrfs_qgroup *qgroup;
3428
3429         spin_lock(&fs_info->qgroup_lock);
3430         /* clear all current qgroup tracking information */
3431         for (n = rb_first(&fs_info->qgroup_tree); n; n = rb_next(n)) {
3432                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
3433                 qgroup->rfer = 0;
3434                 qgroup->rfer_cmpr = 0;
3435                 qgroup->excl = 0;
3436                 qgroup->excl_cmpr = 0;
3437                 qgroup_dirty(fs_info, qgroup);
3438         }
3439         spin_unlock(&fs_info->qgroup_lock);
3440 }
3441
3442 int
3443 btrfs_qgroup_rescan(struct btrfs_fs_info *fs_info)
3444 {
3445         int ret = 0;
3446         struct btrfs_trans_handle *trans;
3447
3448         ret = qgroup_rescan_init(fs_info, 0, 1);
3449         if (ret)
3450                 return ret;
3451
3452         /*
3453          * We have set the rescan_progress to 0, which means no more
3454          * delayed refs will be accounted by btrfs_qgroup_account_ref.
3455          * However, btrfs_qgroup_account_ref may be right after its call
3456          * to btrfs_find_all_roots, in which case it would still do the
3457          * accounting.
3458          * To solve this, we're committing the transaction, which will
3459          * ensure we run all delayed refs and only after that, we are
3460          * going to clear all tracking information for a clean start.
3461          */
3462
3463         trans = btrfs_join_transaction(fs_info->fs_root);
3464         if (IS_ERR(trans)) {
3465                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3466                 return PTR_ERR(trans);
3467         }
3468         ret = btrfs_commit_transaction(trans);
3469         if (ret) {
3470                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3471                 return ret;
3472         }
3473
3474         qgroup_rescan_zero_tracking(fs_info);
3475
3476         mutex_lock(&fs_info->qgroup_rescan_lock);
3477         fs_info->qgroup_rescan_running = true;
3478         btrfs_queue_work(fs_info->qgroup_rescan_workers,
3479                          &fs_info->qgroup_rescan_work);
3480         mutex_unlock(&fs_info->qgroup_rescan_lock);
3481
3482         return 0;
3483 }
3484
3485 int btrfs_qgroup_wait_for_completion(struct btrfs_fs_info *fs_info,
3486                                      bool interruptible)
3487 {
3488         int running;
3489         int ret = 0;
3490
3491         mutex_lock(&fs_info->qgroup_rescan_lock);
3492         running = fs_info->qgroup_rescan_running;
3493         mutex_unlock(&fs_info->qgroup_rescan_lock);
3494
3495         if (!running)
3496                 return 0;
3497
3498         if (interruptible)
3499                 ret = wait_for_completion_interruptible(
3500                                         &fs_info->qgroup_rescan_completion);
3501         else
3502                 wait_for_completion(&fs_info->qgroup_rescan_completion);
3503
3504         return ret;
3505 }
3506
3507 /*
3508  * this is only called from open_ctree where we're still single threaded, thus
3509  * locking is omitted here.
3510  */
3511 void
3512 btrfs_qgroup_rescan_resume(struct btrfs_fs_info *fs_info)
3513 {
3514         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
3515                 mutex_lock(&fs_info->qgroup_rescan_lock);
3516                 fs_info->qgroup_rescan_running = true;
3517                 btrfs_queue_work(fs_info->qgroup_rescan_workers,
3518                                  &fs_info->qgroup_rescan_work);
3519                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3520         }
3521 }
3522
3523 #define rbtree_iterate_from_safe(node, next, start)                             \
3524        for (node = start; node && ({ next = rb_next(node); 1;}); node = next)
3525
3526 static int qgroup_unreserve_range(struct btrfs_inode *inode,
3527                                   struct extent_changeset *reserved, u64 start,
3528                                   u64 len)
3529 {
3530         struct rb_node *node;
3531         struct rb_node *next;
3532         struct ulist_node *entry;
3533         int ret = 0;
3534
3535         node = reserved->range_changed.root.rb_node;
3536         if (!node)
3537                 return 0;
3538         while (node) {
3539                 entry = rb_entry(node, struct ulist_node, rb_node);
3540                 if (entry->val < start)
3541                         node = node->rb_right;
3542                 else
3543                         node = node->rb_left;
3544         }
3545
3546         if (entry->val > start && rb_prev(&entry->rb_node))
3547                 entry = rb_entry(rb_prev(&entry->rb_node), struct ulist_node,
3548                                  rb_node);
3549
3550         rbtree_iterate_from_safe(node, next, &entry->rb_node) {
3551                 u64 entry_start;
3552                 u64 entry_end;
3553                 u64 entry_len;
3554                 int clear_ret;
3555
3556                 entry = rb_entry(node, struct ulist_node, rb_node);
3557                 entry_start = entry->val;
3558                 entry_end = entry->aux;
3559                 entry_len = entry_end - entry_start + 1;
3560
3561                 if (entry_start >= start + len)
3562                         break;
3563                 if (entry_start + entry_len <= start)
3564                         continue;
3565                 /*
3566                  * Now the entry is in [start, start + len), revert the
3567                  * EXTENT_QGROUP_RESERVED bit.
3568                  */
3569                 clear_ret = clear_extent_bits(&inode->io_tree, entry_start,
3570                                               entry_end, EXTENT_QGROUP_RESERVED);
3571                 if (!ret && clear_ret < 0)
3572                         ret = clear_ret;
3573
3574                 ulist_del(&reserved->range_changed, entry->val, entry->aux);
3575                 if (likely(reserved->bytes_changed >= entry_len)) {
3576                         reserved->bytes_changed -= entry_len;
3577                 } else {
3578                         WARN_ON(1);
3579                         reserved->bytes_changed = 0;
3580                 }
3581         }
3582
3583         return ret;
3584 }
3585
3586 /*
3587  * Try to free some space for qgroup.
3588  *
3589  * For qgroup, there are only 3 ways to free qgroup space:
3590  * - Flush nodatacow write
3591  *   Any nodatacow write will free its reserved data space at run_delalloc_range().
3592  *   In theory, we should only flush nodatacow inodes, but it's not yet
3593  *   possible, so we need to flush the whole root.
3594  *
3595  * - Wait for ordered extents
3596  *   When ordered extents are finished, their reserved metadata is finally
3597  *   converted to per_trans status, which can be freed by later commit
3598  *   transaction.
3599  *
3600  * - Commit transaction
3601  *   This would free the meta_per_trans space.
3602  *   In theory this shouldn't provide much space, but any more qgroup space
3603  *   is needed.
3604  */
3605 static int try_flush_qgroup(struct btrfs_root *root)
3606 {
3607         struct btrfs_trans_handle *trans;
3608         int ret;
3609
3610         /* Can't hold an open transaction or we run the risk of deadlocking. */
3611         ASSERT(current->journal_info == NULL);
3612         if (WARN_ON(current->journal_info))
3613                 return 0;
3614
3615         /*
3616          * We don't want to run flush again and again, so if there is a running
3617          * one, we won't try to start a new flush, but exit directly.
3618          */
3619         if (test_and_set_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state)) {
3620                 wait_event(root->qgroup_flush_wait,
3621                         !test_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state));
3622                 return 0;
3623         }
3624
3625         ret = btrfs_start_delalloc_snapshot(root, true);
3626         if (ret < 0)
3627                 goto out;
3628         btrfs_wait_ordered_extents(root, U64_MAX, 0, (u64)-1);
3629
3630         trans = btrfs_join_transaction(root);
3631         if (IS_ERR(trans)) {
3632                 ret = PTR_ERR(trans);
3633                 goto out;
3634         }
3635
3636         ret = btrfs_commit_transaction(trans);
3637 out:
3638         clear_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state);
3639         wake_up(&root->qgroup_flush_wait);
3640         return ret;
3641 }
3642
3643 static int qgroup_reserve_data(struct btrfs_inode *inode,
3644                         struct extent_changeset **reserved_ret, u64 start,
3645                         u64 len)
3646 {
3647         struct btrfs_root *root = inode->root;
3648         struct extent_changeset *reserved;
3649         bool new_reserved = false;
3650         u64 orig_reserved;
3651         u64 to_reserve;
3652         int ret;
3653
3654         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &root->fs_info->flags) ||
3655             !is_fstree(root->root_key.objectid) || len == 0)
3656                 return 0;
3657
3658         /* @reserved parameter is mandatory for qgroup */
3659         if (WARN_ON(!reserved_ret))
3660                 return -EINVAL;
3661         if (!*reserved_ret) {
3662                 new_reserved = true;
3663                 *reserved_ret = extent_changeset_alloc();
3664                 if (!*reserved_ret)
3665                         return -ENOMEM;
3666         }
3667         reserved = *reserved_ret;
3668         /* Record already reserved space */
3669         orig_reserved = reserved->bytes_changed;
3670         ret = set_record_extent_bits(&inode->io_tree, start,
3671                         start + len -1, EXTENT_QGROUP_RESERVED, reserved);
3672
3673         /* Newly reserved space */
3674         to_reserve = reserved->bytes_changed - orig_reserved;
3675         trace_btrfs_qgroup_reserve_data(&inode->vfs_inode, start, len,
3676                                         to_reserve, QGROUP_RESERVE);
3677         if (ret < 0)
3678                 goto out;
3679         ret = qgroup_reserve(root, to_reserve, true, BTRFS_QGROUP_RSV_DATA);
3680         if (ret < 0)
3681                 goto cleanup;
3682
3683         return ret;
3684
3685 cleanup:
3686         qgroup_unreserve_range(inode, reserved, start, len);
3687 out:
3688         if (new_reserved) {
3689                 extent_changeset_free(reserved);
3690                 *reserved_ret = NULL;
3691         }
3692         return ret;
3693 }
3694
3695 /*
3696  * Reserve qgroup space for range [start, start + len).
3697  *
3698  * This function will either reserve space from related qgroups or do nothing
3699  * if the range is already reserved.
3700  *
3701  * Return 0 for successful reservation
3702  * Return <0 for error (including -EQUOT)
3703  *
3704  * NOTE: This function may sleep for memory allocation, dirty page flushing and
3705  *       commit transaction. So caller should not hold any dirty page locked.
3706  */
3707 int btrfs_qgroup_reserve_data(struct btrfs_inode *inode,
3708                         struct extent_changeset **reserved_ret, u64 start,
3709                         u64 len)
3710 {
3711         int ret;
3712
3713         ret = qgroup_reserve_data(inode, reserved_ret, start, len);
3714         if (ret <= 0 && ret != -EDQUOT)
3715                 return ret;
3716
3717         ret = try_flush_qgroup(inode->root);
3718         if (ret < 0)
3719                 return ret;
3720         return qgroup_reserve_data(inode, reserved_ret, start, len);
3721 }
3722
3723 /* Free ranges specified by @reserved, normally in error path */
3724 static int qgroup_free_reserved_data(struct btrfs_inode *inode,
3725                         struct extent_changeset *reserved, u64 start, u64 len)
3726 {
3727         struct btrfs_root *root = inode->root;
3728         struct ulist_node *unode;
3729         struct ulist_iterator uiter;
3730         struct extent_changeset changeset;
3731         int freed = 0;
3732         int ret;
3733
3734         extent_changeset_init(&changeset);
3735         len = round_up(start + len, root->fs_info->sectorsize);
3736         start = round_down(start, root->fs_info->sectorsize);
3737
3738         ULIST_ITER_INIT(&uiter);
3739         while ((unode = ulist_next(&reserved->range_changed, &uiter))) {
3740                 u64 range_start = unode->val;
3741                 /* unode->aux is the inclusive end */
3742                 u64 range_len = unode->aux - range_start + 1;
3743                 u64 free_start;
3744                 u64 free_len;
3745
3746                 extent_changeset_release(&changeset);
3747
3748                 /* Only free range in range [start, start + len) */
3749                 if (range_start >= start + len ||
3750                     range_start + range_len <= start)
3751                         continue;
3752                 free_start = max(range_start, start);
3753                 free_len = min(start + len, range_start + range_len) -
3754                            free_start;
3755                 /*
3756                  * TODO: To also modify reserved->ranges_reserved to reflect
3757                  * the modification.
3758                  *
3759                  * However as long as we free qgroup reserved according to
3760                  * EXTENT_QGROUP_RESERVED, we won't double free.
3761                  * So not need to rush.
3762                  */
3763                 ret = clear_record_extent_bits(&inode->io_tree, free_start,
3764                                 free_start + free_len - 1,
3765                                 EXTENT_QGROUP_RESERVED, &changeset);
3766                 if (ret < 0)
3767                         goto out;
3768                 freed += changeset.bytes_changed;
3769         }
3770         btrfs_qgroup_free_refroot(root->fs_info, root->root_key.objectid, freed,
3771                                   BTRFS_QGROUP_RSV_DATA);
3772         ret = freed;
3773 out:
3774         extent_changeset_release(&changeset);
3775         return ret;
3776 }
3777
3778 static int __btrfs_qgroup_release_data(struct btrfs_inode *inode,
3779                         struct extent_changeset *reserved, u64 start, u64 len,
3780                         int free)
3781 {
3782         struct extent_changeset changeset;
3783         int trace_op = QGROUP_RELEASE;
3784         int ret;
3785
3786         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &inode->root->fs_info->flags))
3787                 return 0;
3788
3789         /* In release case, we shouldn't have @reserved */
3790         WARN_ON(!free && reserved);
3791         if (free && reserved)
3792                 return qgroup_free_reserved_data(inode, reserved, start, len);
3793         extent_changeset_init(&changeset);
3794         ret = clear_record_extent_bits(&inode->io_tree, start, start + len -1,
3795                                        EXTENT_QGROUP_RESERVED, &changeset);
3796         if (ret < 0)
3797                 goto out;
3798
3799         if (free)
3800                 trace_op = QGROUP_FREE;
3801         trace_btrfs_qgroup_release_data(&inode->vfs_inode, start, len,
3802                                         changeset.bytes_changed, trace_op);
3803         if (free)
3804                 btrfs_qgroup_free_refroot(inode->root->fs_info,
3805                                 inode->root->root_key.objectid,
3806                                 changeset.bytes_changed, BTRFS_QGROUP_RSV_DATA);
3807         ret = changeset.bytes_changed;
3808 out:
3809         extent_changeset_release(&changeset);
3810         return ret;
3811 }
3812
3813 /*
3814  * Free a reserved space range from io_tree and related qgroups
3815  *
3816  * Should be called when a range of pages get invalidated before reaching disk.
3817  * Or for error cleanup case.
3818  * if @reserved is given, only reserved range in [@start, @start + @len) will
3819  * be freed.
3820  *
3821  * For data written to disk, use btrfs_qgroup_release_data().
3822  *
3823  * NOTE: This function may sleep for memory allocation.
3824  */
3825 int btrfs_qgroup_free_data(struct btrfs_inode *inode,
3826                         struct extent_changeset *reserved, u64 start, u64 len)
3827 {
3828         return __btrfs_qgroup_release_data(inode, reserved, start, len, 1);
3829 }
3830
3831 /*
3832  * Release a reserved space range from io_tree only.
3833  *
3834  * Should be called when a range of pages get written to disk and corresponding
3835  * FILE_EXTENT is inserted into corresponding root.
3836  *
3837  * Since new qgroup accounting framework will only update qgroup numbers at
3838  * commit_transaction() time, its reserved space shouldn't be freed from
3839  * related qgroups.
3840  *
3841  * But we should release the range from io_tree, to allow further write to be
3842  * COWed.
3843  *
3844  * NOTE: This function may sleep for memory allocation.
3845  */
3846 int btrfs_qgroup_release_data(struct btrfs_inode *inode, u64 start, u64 len)
3847 {
3848         return __btrfs_qgroup_release_data(inode, NULL, start, len, 0);
3849 }
3850
3851 static void add_root_meta_rsv(struct btrfs_root *root, int num_bytes,
3852                               enum btrfs_qgroup_rsv_type type)
3853 {
3854         if (type != BTRFS_QGROUP_RSV_META_PREALLOC &&
3855             type != BTRFS_QGROUP_RSV_META_PERTRANS)
3856                 return;
3857         if (num_bytes == 0)
3858                 return;
3859
3860         spin_lock(&root->qgroup_meta_rsv_lock);
3861         if (type == BTRFS_QGROUP_RSV_META_PREALLOC)
3862                 root->qgroup_meta_rsv_prealloc += num_bytes;
3863         else
3864                 root->qgroup_meta_rsv_pertrans += num_bytes;
3865         spin_unlock(&root->qgroup_meta_rsv_lock);
3866 }
3867
3868 static int sub_root_meta_rsv(struct btrfs_root *root, int num_bytes,
3869                              enum btrfs_qgroup_rsv_type type)
3870 {
3871         if (type != BTRFS_QGROUP_RSV_META_PREALLOC &&
3872             type != BTRFS_QGROUP_RSV_META_PERTRANS)
3873                 return 0;
3874         if (num_bytes == 0)
3875                 return 0;
3876
3877         spin_lock(&root->qgroup_meta_rsv_lock);
3878         if (type == BTRFS_QGROUP_RSV_META_PREALLOC) {
3879                 num_bytes = min_t(u64, root->qgroup_meta_rsv_prealloc,
3880                                   num_bytes);
3881                 root->qgroup_meta_rsv_prealloc -= num_bytes;
3882         } else {
3883                 num_bytes = min_t(u64, root->qgroup_meta_rsv_pertrans,
3884                                   num_bytes);
3885                 root->qgroup_meta_rsv_pertrans -= num_bytes;
3886         }
3887         spin_unlock(&root->qgroup_meta_rsv_lock);
3888         return num_bytes;
3889 }
3890
3891 int btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
3892                               enum btrfs_qgroup_rsv_type type, bool enforce)
3893 {
3894         struct btrfs_fs_info *fs_info = root->fs_info;
3895         int ret;
3896
3897         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
3898             !is_fstree(root->root_key.objectid) || num_bytes == 0)
3899                 return 0;
3900
3901         BUG_ON(num_bytes != round_down(num_bytes, fs_info->nodesize));
3902         trace_qgroup_meta_reserve(root, (s64)num_bytes, type);
3903         ret = qgroup_reserve(root, num_bytes, enforce, type);
3904         if (ret < 0)
3905                 return ret;
3906         /*
3907          * Record what we have reserved into root.
3908          *
3909          * To avoid quota disabled->enabled underflow.
3910          * In that case, we may try to free space we haven't reserved
3911          * (since quota was disabled), so record what we reserved into root.
3912          * And ensure later release won't underflow this number.
3913          */
3914         add_root_meta_rsv(root, num_bytes, type);
3915         return ret;
3916 }
3917
3918 int __btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
3919                                 enum btrfs_qgroup_rsv_type type, bool enforce)
3920 {
3921         int ret;
3922
3923         ret = btrfs_qgroup_reserve_meta(root, num_bytes, type, enforce);
3924         if (ret <= 0 && ret != -EDQUOT)
3925                 return ret;
3926
3927         ret = try_flush_qgroup(root);
3928         if (ret < 0)
3929                 return ret;
3930         return btrfs_qgroup_reserve_meta(root, num_bytes, type, enforce);
3931 }
3932
3933 void btrfs_qgroup_free_meta_all_pertrans(struct btrfs_root *root)
3934 {
3935         struct btrfs_fs_info *fs_info = root->fs_info;
3936
3937         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
3938             !is_fstree(root->root_key.objectid))
3939                 return;
3940
3941         /* TODO: Update trace point to handle such free */
3942         trace_qgroup_meta_free_all_pertrans(root);
3943         /* Special value -1 means to free all reserved space */
3944         btrfs_qgroup_free_refroot(fs_info, root->root_key.objectid, (u64)-1,
3945                                   BTRFS_QGROUP_RSV_META_PERTRANS);
3946 }
3947
3948 void __btrfs_qgroup_free_meta(struct btrfs_root *root, int num_bytes,
3949                               enum btrfs_qgroup_rsv_type type)
3950 {
3951         struct btrfs_fs_info *fs_info = root->fs_info;
3952
3953         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
3954             !is_fstree(root->root_key.objectid))
3955                 return;
3956
3957         /*
3958          * reservation for META_PREALLOC can happen before quota is enabled,
3959          * which can lead to underflow.
3960          * Here ensure we will only free what we really have reserved.
3961          */
3962         num_bytes = sub_root_meta_rsv(root, num_bytes, type);
3963         BUG_ON(num_bytes != round_down(num_bytes, fs_info->nodesize));
3964         trace_qgroup_meta_reserve(root, -(s64)num_bytes, type);
3965         btrfs_qgroup_free_refroot(fs_info, root->root_key.objectid,
3966                                   num_bytes, type);
3967 }
3968
3969 static void qgroup_convert_meta(struct btrfs_fs_info *fs_info, u64 ref_root,
3970                                 int num_bytes)
3971 {
3972         struct btrfs_qgroup *qgroup;
3973         struct ulist_node *unode;
3974         struct ulist_iterator uiter;
3975         int ret = 0;
3976
3977         if (num_bytes == 0)
3978                 return;
3979         if (!fs_info->quota_root)
3980                 return;
3981
3982         spin_lock(&fs_info->qgroup_lock);
3983         qgroup = find_qgroup_rb(fs_info, ref_root);
3984         if (!qgroup)
3985                 goto out;
3986         ulist_reinit(fs_info->qgroup_ulist);
3987         ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
3988                        qgroup_to_aux(qgroup), GFP_ATOMIC);
3989         if (ret < 0)
3990                 goto out;
3991         ULIST_ITER_INIT(&uiter);
3992         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
3993                 struct btrfs_qgroup *qg;
3994                 struct btrfs_qgroup_list *glist;
3995
3996                 qg = unode_aux_to_qgroup(unode);
3997
3998                 qgroup_rsv_release(fs_info, qg, num_bytes,
3999                                 BTRFS_QGROUP_RSV_META_PREALLOC);
4000                 qgroup_rsv_add(fs_info, qg, num_bytes,
4001                                 BTRFS_QGROUP_RSV_META_PERTRANS);
4002                 list_for_each_entry(glist, &qg->groups, next_group) {
4003                         ret = ulist_add(fs_info->qgroup_ulist,
4004                                         glist->group->qgroupid,
4005                                         qgroup_to_aux(glist->group), GFP_ATOMIC);
4006                         if (ret < 0)
4007                                 goto out;
4008                 }
4009         }
4010 out:
4011         spin_unlock(&fs_info->qgroup_lock);
4012 }
4013
4014 void btrfs_qgroup_convert_reserved_meta(struct btrfs_root *root, int num_bytes)
4015 {
4016         struct btrfs_fs_info *fs_info = root->fs_info;
4017
4018         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
4019             !is_fstree(root->root_key.objectid))
4020                 return;
4021         /* Same as btrfs_qgroup_free_meta_prealloc() */
4022         num_bytes = sub_root_meta_rsv(root, num_bytes,
4023                                       BTRFS_QGROUP_RSV_META_PREALLOC);
4024         trace_qgroup_meta_convert(root, num_bytes);
4025         qgroup_convert_meta(fs_info, root->root_key.objectid, num_bytes);
4026 }
4027
4028 /*
4029  * Check qgroup reserved space leaking, normally at destroy inode
4030  * time
4031  */
4032 void btrfs_qgroup_check_reserved_leak(struct btrfs_inode *inode)
4033 {
4034         struct extent_changeset changeset;
4035         struct ulist_node *unode;
4036         struct ulist_iterator iter;
4037         int ret;
4038
4039         extent_changeset_init(&changeset);
4040         ret = clear_record_extent_bits(&inode->io_tree, 0, (u64)-1,
4041                         EXTENT_QGROUP_RESERVED, &changeset);
4042
4043         WARN_ON(ret < 0);
4044         if (WARN_ON(changeset.bytes_changed)) {
4045                 ULIST_ITER_INIT(&iter);
4046                 while ((unode = ulist_next(&changeset.range_changed, &iter))) {
4047                         btrfs_warn(inode->root->fs_info,
4048                 "leaking qgroup reserved space, ino: %llu, start: %llu, end: %llu",
4049                                 btrfs_ino(inode), unode->val, unode->aux);
4050                 }
4051                 btrfs_qgroup_free_refroot(inode->root->fs_info,
4052                                 inode->root->root_key.objectid,
4053                                 changeset.bytes_changed, BTRFS_QGROUP_RSV_DATA);
4054
4055         }
4056         extent_changeset_release(&changeset);
4057 }
4058
4059 void btrfs_qgroup_init_swapped_blocks(
4060         struct btrfs_qgroup_swapped_blocks *swapped_blocks)
4061 {
4062         int i;
4063
4064         spin_lock_init(&swapped_blocks->lock);
4065         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
4066                 swapped_blocks->blocks[i] = RB_ROOT;
4067         swapped_blocks->swapped = false;
4068 }
4069
4070 /*
4071  * Delete all swapped blocks record of @root.
4072  * Every record here means we skipped a full subtree scan for qgroup.
4073  *
4074  * Gets called when committing one transaction.
4075  */
4076 void btrfs_qgroup_clean_swapped_blocks(struct btrfs_root *root)
4077 {
4078         struct btrfs_qgroup_swapped_blocks *swapped_blocks;
4079         int i;
4080
4081         swapped_blocks = &root->swapped_blocks;
4082
4083         spin_lock(&swapped_blocks->lock);
4084         if (!swapped_blocks->swapped)
4085                 goto out;
4086         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
4087                 struct rb_root *cur_root = &swapped_blocks->blocks[i];
4088                 struct btrfs_qgroup_swapped_block *entry;
4089                 struct btrfs_qgroup_swapped_block *next;
4090
4091                 rbtree_postorder_for_each_entry_safe(entry, next, cur_root,
4092                                                      node)
4093                         kfree(entry);
4094                 swapped_blocks->blocks[i] = RB_ROOT;
4095         }
4096         swapped_blocks->swapped = false;
4097 out:
4098         spin_unlock(&swapped_blocks->lock);
4099 }
4100
4101 /*
4102  * Add subtree roots record into @subvol_root.
4103  *
4104  * @subvol_root:        tree root of the subvolume tree get swapped
4105  * @bg:                 block group under balance
4106  * @subvol_parent/slot: pointer to the subtree root in subvolume tree
4107  * @reloc_parent/slot:  pointer to the subtree root in reloc tree
4108  *                      BOTH POINTERS ARE BEFORE TREE SWAP
4109  * @last_snapshot:      last snapshot generation of the subvolume tree
4110  */
4111 int btrfs_qgroup_add_swapped_blocks(struct btrfs_trans_handle *trans,
4112                 struct btrfs_root *subvol_root,
4113                 struct btrfs_block_group *bg,
4114                 struct extent_buffer *subvol_parent, int subvol_slot,
4115                 struct extent_buffer *reloc_parent, int reloc_slot,
4116                 u64 last_snapshot)
4117 {
4118         struct btrfs_fs_info *fs_info = subvol_root->fs_info;
4119         struct btrfs_qgroup_swapped_blocks *blocks = &subvol_root->swapped_blocks;
4120         struct btrfs_qgroup_swapped_block *block;
4121         struct rb_node **cur;
4122         struct rb_node *parent = NULL;
4123         int level = btrfs_header_level(subvol_parent) - 1;
4124         int ret = 0;
4125
4126         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
4127                 return 0;
4128
4129         if (btrfs_node_ptr_generation(subvol_parent, subvol_slot) >
4130             btrfs_node_ptr_generation(reloc_parent, reloc_slot)) {
4131                 btrfs_err_rl(fs_info,
4132                 "%s: bad parameter order, subvol_gen=%llu reloc_gen=%llu",
4133                         __func__,
4134                         btrfs_node_ptr_generation(subvol_parent, subvol_slot),
4135                         btrfs_node_ptr_generation(reloc_parent, reloc_slot));
4136                 return -EUCLEAN;
4137         }
4138
4139         block = kmalloc(sizeof(*block), GFP_NOFS);
4140         if (!block) {
4141                 ret = -ENOMEM;
4142                 goto out;
4143         }
4144
4145         /*
4146          * @reloc_parent/slot is still before swap, while @block is going to
4147          * record the bytenr after swap, so we do the swap here.
4148          */
4149         block->subvol_bytenr = btrfs_node_blockptr(reloc_parent, reloc_slot);
4150         block->subvol_generation = btrfs_node_ptr_generation(reloc_parent,
4151                                                              reloc_slot);
4152         block->reloc_bytenr = btrfs_node_blockptr(subvol_parent, subvol_slot);
4153         block->reloc_generation = btrfs_node_ptr_generation(subvol_parent,
4154                                                             subvol_slot);
4155         block->last_snapshot = last_snapshot;
4156         block->level = level;
4157
4158         /*
4159          * If we have bg == NULL, we're called from btrfs_recover_relocation(),
4160          * no one else can modify tree blocks thus we qgroup will not change
4161          * no matter the value of trace_leaf.
4162          */
4163         if (bg && bg->flags & BTRFS_BLOCK_GROUP_DATA)
4164                 block->trace_leaf = true;
4165         else
4166                 block->trace_leaf = false;
4167         btrfs_node_key_to_cpu(reloc_parent, &block->first_key, reloc_slot);
4168
4169         /* Insert @block into @blocks */
4170         spin_lock(&blocks->lock);
4171         cur = &blocks->blocks[level].rb_node;
4172         while (*cur) {
4173                 struct btrfs_qgroup_swapped_block *entry;
4174
4175                 parent = *cur;
4176                 entry = rb_entry(parent, struct btrfs_qgroup_swapped_block,
4177                                  node);
4178
4179                 if (entry->subvol_bytenr < block->subvol_bytenr) {
4180                         cur = &(*cur)->rb_left;
4181                 } else if (entry->subvol_bytenr > block->subvol_bytenr) {
4182                         cur = &(*cur)->rb_right;
4183                 } else {
4184                         if (entry->subvol_generation !=
4185                                         block->subvol_generation ||
4186                             entry->reloc_bytenr != block->reloc_bytenr ||
4187                             entry->reloc_generation !=
4188                                         block->reloc_generation) {
4189                                 /*
4190                                  * Duplicated but mismatch entry found.
4191                                  * Shouldn't happen.
4192                                  *
4193                                  * Marking qgroup inconsistent should be enough
4194                                  * for end users.
4195                                  */
4196                                 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
4197                                 ret = -EEXIST;
4198                         }
4199                         kfree(block);
4200                         goto out_unlock;
4201                 }
4202         }
4203         rb_link_node(&block->node, parent, cur);
4204         rb_insert_color(&block->node, &blocks->blocks[level]);
4205         blocks->swapped = true;
4206 out_unlock:
4207         spin_unlock(&blocks->lock);
4208 out:
4209         if (ret < 0)
4210                 fs_info->qgroup_flags |=
4211                         BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
4212         return ret;
4213 }
4214
4215 /*
4216  * Check if the tree block is a subtree root, and if so do the needed
4217  * delayed subtree trace for qgroup.
4218  *
4219  * This is called during btrfs_cow_block().
4220  */
4221 int btrfs_qgroup_trace_subtree_after_cow(struct btrfs_trans_handle *trans,
4222                                          struct btrfs_root *root,
4223                                          struct extent_buffer *subvol_eb)
4224 {
4225         struct btrfs_fs_info *fs_info = root->fs_info;
4226         struct btrfs_qgroup_swapped_blocks *blocks = &root->swapped_blocks;
4227         struct btrfs_qgroup_swapped_block *block;
4228         struct extent_buffer *reloc_eb = NULL;
4229         struct rb_node *node;
4230         bool found = false;
4231         bool swapped = false;
4232         int level = btrfs_header_level(subvol_eb);
4233         int ret = 0;
4234         int i;
4235
4236         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
4237                 return 0;
4238         if (!is_fstree(root->root_key.objectid) || !root->reloc_root)
4239                 return 0;
4240
4241         spin_lock(&blocks->lock);
4242         if (!blocks->swapped) {
4243                 spin_unlock(&blocks->lock);
4244                 return 0;
4245         }
4246         node = blocks->blocks[level].rb_node;
4247
4248         while (node) {
4249                 block = rb_entry(node, struct btrfs_qgroup_swapped_block, node);
4250                 if (block->subvol_bytenr < subvol_eb->start) {
4251                         node = node->rb_left;
4252                 } else if (block->subvol_bytenr > subvol_eb->start) {
4253                         node = node->rb_right;
4254                 } else {
4255                         found = true;
4256                         break;
4257                 }
4258         }
4259         if (!found) {
4260                 spin_unlock(&blocks->lock);
4261                 goto out;
4262         }
4263         /* Found one, remove it from @blocks first and update blocks->swapped */
4264         rb_erase(&block->node, &blocks->blocks[level]);
4265         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
4266                 if (RB_EMPTY_ROOT(&blocks->blocks[i])) {
4267                         swapped = true;
4268                         break;
4269                 }
4270         }
4271         blocks->swapped = swapped;
4272         spin_unlock(&blocks->lock);
4273
4274         /* Read out reloc subtree root */
4275         reloc_eb = read_tree_block(fs_info, block->reloc_bytenr, 0,
4276                                    block->reloc_generation, block->level,
4277                                    &block->first_key);
4278         if (IS_ERR(reloc_eb)) {
4279                 ret = PTR_ERR(reloc_eb);
4280                 reloc_eb = NULL;
4281                 goto free_out;
4282         }
4283         if (!extent_buffer_uptodate(reloc_eb)) {
4284                 ret = -EIO;
4285                 goto free_out;
4286         }
4287
4288         ret = qgroup_trace_subtree_swap(trans, reloc_eb, subvol_eb,
4289                         block->last_snapshot, block->trace_leaf);
4290 free_out:
4291         kfree(block);
4292         free_extent_buffer(reloc_eb);
4293 out:
4294         if (ret < 0) {
4295                 btrfs_err_rl(fs_info,
4296                              "failed to account subtree at bytenr %llu: %d",
4297                              subvol_eb->start, ret);
4298                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
4299         }
4300         return ret;
4301 }
4302
4303 void btrfs_qgroup_destroy_extent_records(struct btrfs_transaction *trans)
4304 {
4305         struct btrfs_qgroup_extent_record *entry;
4306         struct btrfs_qgroup_extent_record *next;
4307         struct rb_root *root;
4308
4309         root = &trans->delayed_refs.dirty_extent_root;
4310         rbtree_postorder_for_each_entry_safe(entry, next, root, node) {
4311                 ulist_free(entry->old_roots);
4312                 kfree(entry);
4313         }
4314 }