fs/btrfs/qgroup.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2011 STRATO.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/writeback.h>
   9 #include <linux/blkdev.h>
  10 #include <linux/rbtree.h>
  11 #include <linux/slab.h>
  12 #include <linux/workqueue.h>
  13 #include <linux/btrfs.h>
  14 #include <linux/sched/mm.h>
  15
  16 #include "ctree.h"
  17 #include "transaction.h"
  18 #include "disk-io.h"
  19 #include "locking.h"
  20 #include "ulist.h"
  21 #include "backref.h"
  22 #include "extent_io.h"
  23 #include "qgroup.h"
  24 #include "block-group.h"
  25 #include "sysfs.h"
  26 #include "tree-mod-log.h"
  27
  28 /* TODO XXX FIXME
  29  *  - subvol delete -> delete when ref goes to 0? delete limits also?
  30  *  - reorganize keys
  31  *  - compressed
  32  *  - sync
  33  *  - copy also limits on subvol creation
  34  *  - limit
  35  *  - caches for ulists
  36  *  - performance benchmarks
  37  *  - check all ioctl parameters
  38  */
  39
  40 /*
  41  * Helpers to access qgroup reservation
  42  *
  43  * Callers should ensure the lock context and type are valid
  44  */
  45
  46 static u64 qgroup_rsv_total(const struct btrfs_qgroup *qgroup)
  47 {
  48         u64 ret = 0;
  49         int i;
  50
  51         for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++)
  52                 ret += qgroup->rsv.values[i];
  53
  54         return ret;
  55 }
  56
  57 #ifdef CONFIG_BTRFS_DEBUG
  58 static const char *qgroup_rsv_type_str(enum btrfs_qgroup_rsv_type type)
  59 {
  60         if (type == BTRFS_QGROUP_RSV_DATA)
  61                 return "data";
  62         if (type == BTRFS_QGROUP_RSV_META_PERTRANS)
  63                 return "meta_pertrans";
  64         if (type == BTRFS_QGROUP_RSV_META_PREALLOC)
  65                 return "meta_prealloc";
  66         return NULL;
  67 }
  68 #endif
  69
  70 static void qgroup_rsv_add(struct btrfs_fs_info *fs_info,
  71                            struct btrfs_qgroup *qgroup, u64 num_bytes,
  72                            enum btrfs_qgroup_rsv_type type)
  73 {
  74         trace_qgroup_update_reserve(fs_info, qgroup, num_bytes, type);
  75         qgroup->rsv.values[type] += num_bytes;
  76 }
  77
  78 static void qgroup_rsv_release(struct btrfs_fs_info *fs_info,
  79                                struct btrfs_qgroup *qgroup, u64 num_bytes,
  80                                enum btrfs_qgroup_rsv_type type)
  81 {
  82         trace_qgroup_update_reserve(fs_info, qgroup, -(s64)num_bytes, type);
  83         if (qgroup->rsv.values[type] >= num_bytes) {
  84                 qgroup->rsv.values[type] -= num_bytes;
  85                 return;
  86         }
  87 #ifdef CONFIG_BTRFS_DEBUG
  88         WARN_RATELIMIT(1,
  89                 "qgroup %llu %s reserved space underflow, have %llu to free %llu",
  90                 qgroup->qgroupid, qgroup_rsv_type_str(type),
  91                 qgroup->rsv.values[type], num_bytes);
  92 #endif
  93         qgroup->rsv.values[type] = 0;
  94 }
  95
  96 static void qgroup_rsv_add_by_qgroup(struct btrfs_fs_info *fs_info,
  97                                      struct btrfs_qgroup *dest,
  98                                      struct btrfs_qgroup *src)
  99 {
 100         int i;
 101
 102         for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++)
 103                 qgroup_rsv_add(fs_info, dest, src->rsv.values[i], i);
 104 }
 105
 106 static void qgroup_rsv_release_by_qgroup(struct btrfs_fs_info *fs_info,
 107                                          struct btrfs_qgroup *dest,
 108                                           struct btrfs_qgroup *src)
 109 {
 110         int i;
 111
 112         for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++)
 113                 qgroup_rsv_release(fs_info, dest, src->rsv.values[i], i);
 114 }
 115
 116 static void btrfs_qgroup_update_old_refcnt(struct btrfs_qgroup *qg, u64 seq,
 117                                            int mod)
 118 {
 119         if (qg->old_refcnt < seq)
 120                 qg->old_refcnt = seq;
 121         qg->old_refcnt += mod;
 122 }
 123
 124 static void btrfs_qgroup_update_new_refcnt(struct btrfs_qgroup *qg, u64 seq,
 125                                            int mod)
 126 {
 127         if (qg->new_refcnt < seq)
 128                 qg->new_refcnt = seq;
 129         qg->new_refcnt += mod;
 130 }
 131
 132 static inline u64 btrfs_qgroup_get_old_refcnt(struct btrfs_qgroup *qg, u64 seq)
 133 {
 134         if (qg->old_refcnt < seq)
 135                 return 0;
 136         return qg->old_refcnt - seq;
 137 }
 138
 139 static inline u64 btrfs_qgroup_get_new_refcnt(struct btrfs_qgroup *qg, u64 seq)
 140 {
 141         if (qg->new_refcnt < seq)
 142                 return 0;
 143         return qg->new_refcnt - seq;
 144 }
 145
 146 /*
 147  * glue structure to represent the relations between qgroups.
 148  */
 149 struct btrfs_qgroup_list {
 150         struct list_head next_group;
 151         struct list_head next_member;
 152         struct btrfs_qgroup *group;
 153         struct btrfs_qgroup *member;
 154 };
 155
 156 static inline u64 qgroup_to_aux(struct btrfs_qgroup *qg)
 157 {
 158         return (u64)(uintptr_t)qg;
 159 }
 160
 161 static inline struct btrfs_qgroup* unode_aux_to_qgroup(struct ulist_node *n)
 162 {
 163         return (struct btrfs_qgroup *)(uintptr_t)n->aux;
 164 }
 165
 166 static int
 167 qgroup_rescan_init(struct btrfs_fs_info *fs_info, u64 progress_objectid,
 168                    int init_flags);
 169 static void qgroup_rescan_zero_tracking(struct btrfs_fs_info *fs_info);
 170
 171 /* must be called with qgroup_ioctl_lock held */
 172 static struct btrfs_qgroup *find_qgroup_rb(struct btrfs_fs_info *fs_info,
 173                                            u64 qgroupid)
 174 {
 175         struct rb_node *n = fs_info->qgroup_tree.rb_node;
 176         struct btrfs_qgroup *qgroup;
 177
 178         while (n) {
 179                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
 180                 if (qgroup->qgroupid < qgroupid)
 181                         n = n->rb_left;
 182                 else if (qgroup->qgroupid > qgroupid)
 183                         n = n->rb_right;
 184                 else
 185                         return qgroup;
 186         }
 187         return NULL;
 188 }
 189
 190 /* must be called with qgroup_lock held */
 191 static struct btrfs_qgroup *add_qgroup_rb(struct btrfs_fs_info *fs_info,
 192                                           u64 qgroupid)
 193 {
 194         struct rb_node **p = &fs_info->qgroup_tree.rb_node;
 195         struct rb_node *parent = NULL;
 196         struct btrfs_qgroup *qgroup;
 197
 198         while (*p) {
 199                 parent = *p;
 200                 qgroup = rb_entry(parent, struct btrfs_qgroup, node);
 201
 202                 if (qgroup->qgroupid < qgroupid)
 203                         p = &(*p)->rb_left;
 204                 else if (qgroup->qgroupid > qgroupid)
 205                         p = &(*p)->rb_right;
 206                 else
 207                         return qgroup;
 208         }
 209
 210         qgroup = kzalloc(sizeof(*qgroup), GFP_ATOMIC);
 211         if (!qgroup)
 212                 return ERR_PTR(-ENOMEM);
 213
 214         qgroup->qgroupid = qgroupid;
 215         INIT_LIST_HEAD(&qgroup->groups);
 216         INIT_LIST_HEAD(&qgroup->members);
 217         INIT_LIST_HEAD(&qgroup->dirty);
 218
 219         rb_link_node(&qgroup->node, parent, p);
 220         rb_insert_color(&qgroup->node, &fs_info->qgroup_tree);
 221
 222         return qgroup;
 223 }
 224
 225 static void __del_qgroup_rb(struct btrfs_fs_info *fs_info,
 226                             struct btrfs_qgroup *qgroup)
 227 {
 228         struct btrfs_qgroup_list *list;
 229
 230         list_del(&qgroup->dirty);
 231         while (!list_empty(&qgroup->groups)) {
 232                 list = list_first_entry(&qgroup->groups,
 233                                         struct btrfs_qgroup_list, next_group);
 234                 list_del(&list->next_group);
 235                 list_del(&list->next_member);
 236                 kfree(list);
 237         }
 238
 239         while (!list_empty(&qgroup->members)) {
 240                 list = list_first_entry(&qgroup->members,
 241                                         struct btrfs_qgroup_list, next_member);
 242                 list_del(&list->next_group);
 243                 list_del(&list->next_member);
 244                 kfree(list);
 245         }
 246 }
 247
 248 /* must be called with qgroup_lock held */
 249 static int del_qgroup_rb(struct btrfs_fs_info *fs_info, u64 qgroupid)
 250 {
 251         struct btrfs_qgroup *qgroup = find_qgroup_rb(fs_info, qgroupid);
 252
 253         if (!qgroup)
 254                 return -ENOENT;
 255
 256         rb_erase(&qgroup->node, &fs_info->qgroup_tree);
 257         __del_qgroup_rb(fs_info, qgroup);
 258         return 0;
 259 }
 260
 261 /* must be called with qgroup_lock held */
 262 static int add_relation_rb(struct btrfs_fs_info *fs_info,
 263                            u64 memberid, u64 parentid)
 264 {
 265         struct btrfs_qgroup *member;
 266         struct btrfs_qgroup *parent;
 267         struct btrfs_qgroup_list *list;
 268
 269         member = find_qgroup_rb(fs_info, memberid);
 270         parent = find_qgroup_rb(fs_info, parentid);
 271         if (!member || !parent)
 272                 return -ENOENT;
 273
 274         list = kzalloc(sizeof(*list), GFP_ATOMIC);
 275         if (!list)
 276                 return -ENOMEM;
 277
 278         list->group = parent;
 279         list->member = member;
 280         list_add_tail(&list->next_group, &member->groups);
 281         list_add_tail(&list->next_member, &parent->members);
 282
 283         return 0;
 284 }
 285
 286 /* must be called with qgroup_lock held */
 287 static int del_relation_rb(struct btrfs_fs_info *fs_info,
 288                            u64 memberid, u64 parentid)
 289 {
 290         struct btrfs_qgroup *member;
 291         struct btrfs_qgroup *parent;
 292         struct btrfs_qgroup_list *list;
 293
 294         member = find_qgroup_rb(fs_info, memberid);
 295         parent = find_qgroup_rb(fs_info, parentid);
 296         if (!member || !parent)
 297                 return -ENOENT;
 298
 299         list_for_each_entry(list, &member->groups, next_group) {
 300                 if (list->group == parent) {
 301                         list_del(&list->next_group);
 302                         list_del(&list->next_member);
 303                         kfree(list);
 304                         return 0;
 305                 }
 306         }
 307         return -ENOENT;
 308 }
 309
 310 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 311 int btrfs_verify_qgroup_counts(struct btrfs_fs_info *fs_info, u64 qgroupid,
 312                                u64 rfer, u64 excl)
 313 {
 314         struct btrfs_qgroup *qgroup;
 315
 316         qgroup = find_qgroup_rb(fs_info, qgroupid);
 317         if (!qgroup)
 318                 return -EINVAL;
 319         if (qgroup->rfer != rfer || qgroup->excl != excl)
 320                 return -EINVAL;
 321         return 0;
 322 }
 323 #endif
 324
 325 /*
 326  * The full config is read in one go, only called from open_ctree()
 327  * It doesn't use any locking, as at this point we're still single-threaded
 328  */
 329 int btrfs_read_qgroup_config(struct btrfs_fs_info *fs_info)
 330 {
 331         struct btrfs_key key;
 332         struct btrfs_key found_key;
 333         struct btrfs_root *quota_root = fs_info->quota_root;
 334         struct btrfs_path *path = NULL;
 335         struct extent_buffer *l;
 336         int slot;
 337         int ret = 0;
 338         u64 flags = 0;
 339         u64 rescan_progress = 0;
 340
 341         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
 342                 return 0;
 343
 344         fs_info->qgroup_ulist = ulist_alloc(GFP_KERNEL);
 345         if (!fs_info->qgroup_ulist) {
 346                 ret = -ENOMEM;
 347                 goto out;
 348         }
 349
 350         path = btrfs_alloc_path();
 351         if (!path) {
 352                 ret = -ENOMEM;
 353                 goto out;
 354         }
 355
 356         ret = btrfs_sysfs_add_qgroups(fs_info);
 357         if (ret < 0)
 358                 goto out;
 359         /* default this to quota off, in case no status key is found */
 360         fs_info->qgroup_flags = 0;
 361
 362         /*
 363          * pass 1: read status, all qgroup infos and limits
 364          */
 365         key.objectid = 0;
 366         key.type = 0;
 367         key.offset = 0;
 368         ret = btrfs_search_slot_for_read(quota_root, &key, path, 1, 1);
 369         if (ret)
 370                 goto out;
 371
 372         while (1) {
 373                 struct btrfs_qgroup *qgroup;
 374
 375                 slot = path->slots[0];
 376                 l = path->nodes[0];
 377                 btrfs_item_key_to_cpu(l, &found_key, slot);
 378
 379                 if (found_key.type == BTRFS_QGROUP_STATUS_KEY) {
 380                         struct btrfs_qgroup_status_item *ptr;
 381
 382                         ptr = btrfs_item_ptr(l, slot,
 383                                              struct btrfs_qgroup_status_item);
 384
 385                         if (btrfs_qgroup_status_version(l, ptr) !=
 386                             BTRFS_QGROUP_STATUS_VERSION) {
 387                                 btrfs_err(fs_info,
 388                                  "old qgroup version, quota disabled");
 389                                 goto out;
 390                         }
 391                         if (btrfs_qgroup_status_generation(l, ptr) !=
 392                             fs_info->generation) {
 393                                 flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
 394                                 btrfs_err(fs_info,
 395                                         "qgroup generation mismatch, marked as inconsistent");
 396                         }
 397                         fs_info->qgroup_flags = btrfs_qgroup_status_flags(l,
 398                                                                           ptr);
 399                         rescan_progress = btrfs_qgroup_status_rescan(l, ptr);
 400                         goto next1;
 401                 }
 402
 403                 if (found_key.type != BTRFS_QGROUP_INFO_KEY &&
 404                     found_key.type != BTRFS_QGROUP_LIMIT_KEY)
 405                         goto next1;
 406
 407                 qgroup = find_qgroup_rb(fs_info, found_key.offset);
 408                 if ((qgroup && found_key.type == BTRFS_QGROUP_INFO_KEY) ||
 409                     (!qgroup && found_key.type == BTRFS_QGROUP_LIMIT_KEY)) {
 410                         btrfs_err(fs_info, "inconsistent qgroup config");
 411                         flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
 412                 }
 413                 if (!qgroup) {
 414                         qgroup = add_qgroup_rb(fs_info, found_key.offset);
 415                         if (IS_ERR(qgroup)) {
 416                                 ret = PTR_ERR(qgroup);
 417                                 goto out;
 418                         }
 419                 }
 420                 ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
 421                 if (ret < 0)
 422                         goto out;
 423
 424                 switch (found_key.type) {
 425                 case BTRFS_QGROUP_INFO_KEY: {
 426                         struct btrfs_qgroup_info_item *ptr;
 427
 428                         ptr = btrfs_item_ptr(l, slot,
 429                                              struct btrfs_qgroup_info_item);
 430                         qgroup->rfer = btrfs_qgroup_info_rfer(l, ptr);
 431                         qgroup->rfer_cmpr = btrfs_qgroup_info_rfer_cmpr(l, ptr);
 432                         qgroup->excl = btrfs_qgroup_info_excl(l, ptr);
 433                         qgroup->excl_cmpr = btrfs_qgroup_info_excl_cmpr(l, ptr);
 434                         /* generation currently unused */
 435                         break;
 436                 }
 437                 case BTRFS_QGROUP_LIMIT_KEY: {
 438                         struct btrfs_qgroup_limit_item *ptr;
 439
 440                         ptr = btrfs_item_ptr(l, slot,
 441                                              struct btrfs_qgroup_limit_item);
 442                         qgroup->lim_flags = btrfs_qgroup_limit_flags(l, ptr);
 443                         qgroup->max_rfer = btrfs_qgroup_limit_max_rfer(l, ptr);
 444                         qgroup->max_excl = btrfs_qgroup_limit_max_excl(l, ptr);
 445                         qgroup->rsv_rfer = btrfs_qgroup_limit_rsv_rfer(l, ptr);
 446                         qgroup->rsv_excl = btrfs_qgroup_limit_rsv_excl(l, ptr);
 447                         break;
 448                 }
 449                 }
 450 next1:
 451                 ret = btrfs_next_item(quota_root, path);
 452                 if (ret < 0)
 453                         goto out;
 454                 if (ret)
 455                         break;
 456         }
 457         btrfs_release_path(path);
 458
 459         /*
 460          * pass 2: read all qgroup relations
 461          */
 462         key.objectid = 0;
 463         key.type = BTRFS_QGROUP_RELATION_KEY;
 464         key.offset = 0;
 465         ret = btrfs_search_slot_for_read(quota_root, &key, path, 1, 0);
 466         if (ret)
 467                 goto out;
 468         while (1) {
 469                 slot = path->slots[0];
 470                 l = path->nodes[0];
 471                 btrfs_item_key_to_cpu(l, &found_key, slot);
 472
 473                 if (found_key.type != BTRFS_QGROUP_RELATION_KEY)
 474                         goto next2;
 475
 476                 if (found_key.objectid > found_key.offset) {
 477                         /* parent <- member, not needed to build config */
 478                         /* FIXME should we omit the key completely? */
 479                         goto next2;
 480                 }
 481
 482                 ret = add_relation_rb(fs_info, found_key.objectid,
 483                                       found_key.offset);
 484                 if (ret == -ENOENT) {
 485                         btrfs_warn(fs_info,
 486                                 "orphan qgroup relation 0x%llx->0x%llx",
 487                                 found_key.objectid, found_key.offset);
 488                         ret = 0;        /* ignore the error */
 489                 }
 490                 if (ret)
 491                         goto out;
 492 next2:
 493                 ret = btrfs_next_item(quota_root, path);
 494                 if (ret < 0)
 495                         goto out;
 496                 if (ret)
 497                         break;
 498         }
 499 out:
 500         btrfs_free_path(path);
 501         fs_info->qgroup_flags |= flags;
 502         if (!(fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_ON))
 503                 clear_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
 504         else if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN &&
 505                  ret >= 0)
 506                 ret = qgroup_rescan_init(fs_info, rescan_progress, 0);
 507
 508         if (ret < 0) {
 509                 ulist_free(fs_info->qgroup_ulist);
 510                 fs_info->qgroup_ulist = NULL;
 511                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
 512                 btrfs_sysfs_del_qgroups(fs_info);
 513         }
 514
 515         return ret < 0 ? ret : 0;
 516 }
 517
 518 /*
 519  * Called in close_ctree() when quota is still enabled.  This verifies we don't
 520  * leak some reserved space.
 521  *
 522  * Return false if no reserved space is left.
 523  * Return true if some reserved space is leaked.
 524  */
 525 bool btrfs_check_quota_leak(struct btrfs_fs_info *fs_info)
 526 {
 527         struct rb_node *node;
 528         bool ret = false;
 529
 530         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
 531                 return ret;
 532         /*
 533          * Since we're unmounting, there is no race and no need to grab qgroup
 534          * lock.  And here we don't go post-order to provide a more user
 535          * friendly sorted result.
 536          */
 537         for (node = rb_first(&fs_info->qgroup_tree); node; node = rb_next(node)) {
 538                 struct btrfs_qgroup *qgroup;
 539                 int i;
 540
 541                 qgroup = rb_entry(node, struct btrfs_qgroup, node);
 542                 for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++) {
 543                         if (qgroup->rsv.values[i]) {
 544                                 ret = true;
 545                                 btrfs_warn(fs_info,
 546                 "qgroup %hu/%llu has unreleased space, type %d rsv %llu",
 547                                    btrfs_qgroup_level(qgroup->qgroupid),
 548                                    btrfs_qgroup_subvolid(qgroup->qgroupid),
 549                                    i, qgroup->rsv.values[i]);
 550                         }
 551                 }
 552         }
 553         return ret;
 554 }
 555
 556 /*
 557  * This is called from close_ctree() or open_ctree() or btrfs_quota_disable(),
 558  * first two are in single-threaded paths.And for the third one, we have set
 559  * quota_root to be null with qgroup_lock held before, so it is safe to clean
 560  * up the in-memory structures without qgroup_lock held.
 561  */
 562 void btrfs_free_qgroup_config(struct btrfs_fs_info *fs_info)
 563 {
 564         struct rb_node *n;
 565         struct btrfs_qgroup *qgroup;
 566
 567         while ((n = rb_first(&fs_info->qgroup_tree))) {
 568                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
 569                 rb_erase(n, &fs_info->qgroup_tree);
 570                 __del_qgroup_rb(fs_info, qgroup);
 571                 btrfs_sysfs_del_one_qgroup(fs_info, qgroup);
 572                 kfree(qgroup);
 573         }
 574         /*
 575          * We call btrfs_free_qgroup_config() when unmounting
 576          * filesystem and disabling quota, so we set qgroup_ulist
 577          * to be null here to avoid double free.
 578          */
 579         ulist_free(fs_info->qgroup_ulist);
 580         fs_info->qgroup_ulist = NULL;
 581         btrfs_sysfs_del_qgroups(fs_info);
 582 }
 583
 584 static int add_qgroup_relation_item(struct btrfs_trans_handle *trans, u64 src,
 585                                     u64 dst)
 586 {
 587         int ret;
 588         struct btrfs_root *quota_root = trans->fs_info->quota_root;
 589         struct btrfs_path *path;
 590         struct btrfs_key key;
 591
 592         path = btrfs_alloc_path();
 593         if (!path)
 594                 return -ENOMEM;
 595
 596         key.objectid = src;
 597         key.type = BTRFS_QGROUP_RELATION_KEY;
 598         key.offset = dst;
 599
 600         ret = btrfs_insert_empty_item(trans, quota_root, path, &key, 0);
 601
 602         btrfs_mark_buffer_dirty(path->nodes[0]);
 603
 604         btrfs_free_path(path);
 605         return ret;
 606 }
 607
 608 static int del_qgroup_relation_item(struct btrfs_trans_handle *trans, u64 src,
 609                                     u64 dst)
 610 {
 611         int ret;
 612         struct btrfs_root *quota_root = trans->fs_info->quota_root;
 613         struct btrfs_path *path;
 614         struct btrfs_key key;
 615
 616         path = btrfs_alloc_path();
 617         if (!path)
 618                 return -ENOMEM;
 619
 620         key.objectid = src;
 621         key.type = BTRFS_QGROUP_RELATION_KEY;
 622         key.offset = dst;
 623
 624         ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
 625         if (ret < 0)
 626                 goto out;
 627
 628         if (ret > 0) {
 629                 ret = -ENOENT;
 630                 goto out;
 631         }
 632
 633         ret = btrfs_del_item(trans, quota_root, path);
 634 out:
 635         btrfs_free_path(path);
 636         return ret;
 637 }
 638
 639 static int add_qgroup_item(struct btrfs_trans_handle *trans,
 640                            struct btrfs_root *quota_root, u64 qgroupid)
 641 {
 642         int ret;
 643         struct btrfs_path *path;
 644         struct btrfs_qgroup_info_item *qgroup_info;
 645         struct btrfs_qgroup_limit_item *qgroup_limit;
 646         struct extent_buffer *leaf;
 647         struct btrfs_key key;
 648
 649         if (btrfs_is_testing(quota_root->fs_info))
 650                 return 0;
 651
 652         path = btrfs_alloc_path();
 653         if (!path)
 654                 return -ENOMEM;
 655
 656         key.objectid = 0;
 657         key.type = BTRFS_QGROUP_INFO_KEY;
 658         key.offset = qgroupid;
 659
 660         /*
 661          * Avoid a transaction abort by catching -EEXIST here. In that
 662          * case, we proceed by re-initializing the existing structure
 663          * on disk.
 664          */
 665
 666         ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
 667                                       sizeof(*qgroup_info));
 668         if (ret && ret != -EEXIST)
 669                 goto out;
 670
 671         leaf = path->nodes[0];
 672         qgroup_info = btrfs_item_ptr(leaf, path->slots[0],
 673                                  struct btrfs_qgroup_info_item);
 674         btrfs_set_qgroup_info_generation(leaf, qgroup_info, trans->transid);
 675         btrfs_set_qgroup_info_rfer(leaf, qgroup_info, 0);
 676         btrfs_set_qgroup_info_rfer_cmpr(leaf, qgroup_info, 0);
 677         btrfs_set_qgroup_info_excl(leaf, qgroup_info, 0);
 678         btrfs_set_qgroup_info_excl_cmpr(leaf, qgroup_info, 0);
 679
 680         btrfs_mark_buffer_dirty(leaf);
 681
 682         btrfs_release_path(path);
 683
 684         key.type = BTRFS_QGROUP_LIMIT_KEY;
 685         ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
 686                                       sizeof(*qgroup_limit));
 687         if (ret && ret != -EEXIST)
 688                 goto out;
 689
 690         leaf = path->nodes[0];
 691         qgroup_limit = btrfs_item_ptr(leaf, path->slots[0],
 692                                   struct btrfs_qgroup_limit_item);
 693         btrfs_set_qgroup_limit_flags(leaf, qgroup_limit, 0);
 694         btrfs_set_qgroup_limit_max_rfer(leaf, qgroup_limit, 0);
 695         btrfs_set_qgroup_limit_max_excl(leaf, qgroup_limit, 0);
 696         btrfs_set_qgroup_limit_rsv_rfer(leaf, qgroup_limit, 0);
 697         btrfs_set_qgroup_limit_rsv_excl(leaf, qgroup_limit, 0);
 698
 699         btrfs_mark_buffer_dirty(leaf);
 700
 701         ret = 0;
 702 out:
 703         btrfs_free_path(path);
 704         return ret;
 705 }
 706
 707 static int del_qgroup_item(struct btrfs_trans_handle *trans, u64 qgroupid)
 708 {
 709         int ret;
 710         struct btrfs_root *quota_root = trans->fs_info->quota_root;
 711         struct btrfs_path *path;
 712         struct btrfs_key key;
 713
 714         path = btrfs_alloc_path();
 715         if (!path)
 716                 return -ENOMEM;
 717
 718         key.objectid = 0;
 719         key.type = BTRFS_QGROUP_INFO_KEY;
 720         key.offset = qgroupid;
 721         ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
 722         if (ret < 0)
 723                 goto out;
 724
 725         if (ret > 0) {
 726                 ret = -ENOENT;
 727                 goto out;
 728         }
 729
 730         ret = btrfs_del_item(trans, quota_root, path);
 731         if (ret)
 732                 goto out;
 733
 734         btrfs_release_path(path);
 735
 736         key.type = BTRFS_QGROUP_LIMIT_KEY;
 737         ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
 738         if (ret < 0)
 739                 goto out;
 740
 741         if (ret > 0) {
 742                 ret = -ENOENT;
 743                 goto out;
 744         }
 745
 746         ret = btrfs_del_item(trans, quota_root, path);
 747
 748 out:
 749         btrfs_free_path(path);
 750         return ret;
 751 }
 752
 753 static int update_qgroup_limit_item(struct btrfs_trans_handle *trans,
 754                                     struct btrfs_qgroup *qgroup)
 755 {
 756         struct btrfs_root *quota_root = trans->fs_info->quota_root;
 757         struct btrfs_path *path;
 758         struct btrfs_key key;
 759         struct extent_buffer *l;
 760         struct btrfs_qgroup_limit_item *qgroup_limit;
 761         int ret;
 762         int slot;
 763
 764         key.objectid = 0;
 765         key.type = BTRFS_QGROUP_LIMIT_KEY;
 766         key.offset = qgroup->qgroupid;
 767
 768         path = btrfs_alloc_path();
 769         if (!path)
 770                 return -ENOMEM;
 771
 772         ret = btrfs_search_slot(trans, quota_root, &key, path, 0, 1);
 773         if (ret > 0)
 774                 ret = -ENOENT;
 775
 776         if (ret)
 777                 goto out;
 778
 779         l = path->nodes[0];
 780         slot = path->slots[0];
 781         qgroup_limit = btrfs_item_ptr(l, slot, struct btrfs_qgroup_limit_item);
 782         btrfs_set_qgroup_limit_flags(l, qgroup_limit, qgroup->lim_flags);
 783         btrfs_set_qgroup_limit_max_rfer(l, qgroup_limit, qgroup->max_rfer);
 784         btrfs_set_qgroup_limit_max_excl(l, qgroup_limit, qgroup->max_excl);
 785         btrfs_set_qgroup_limit_rsv_rfer(l, qgroup_limit, qgroup->rsv_rfer);
 786         btrfs_set_qgroup_limit_rsv_excl(l, qgroup_limit, qgroup->rsv_excl);
 787
 788         btrfs_mark_buffer_dirty(l);
 789
 790 out:
 791         btrfs_free_path(path);
 792         return ret;
 793 }
 794
 795 static int update_qgroup_info_item(struct btrfs_trans_handle *trans,
 796                                    struct btrfs_qgroup *qgroup)
 797 {
 798         struct btrfs_fs_info *fs_info = trans->fs_info;
 799         struct btrfs_root *quota_root = fs_info->quota_root;
 800         struct btrfs_path *path;
 801         struct btrfs_key key;
 802         struct extent_buffer *l;
 803         struct btrfs_qgroup_info_item *qgroup_info;
 804         int ret;
 805         int slot;
 806
 807         if (btrfs_is_testing(fs_info))
 808                 return 0;
 809
 810         key.objectid = 0;
 811         key.type = BTRFS_QGROUP_INFO_KEY;
 812         key.offset = qgroup->qgroupid;
 813
 814         path = btrfs_alloc_path();
 815         if (!path)
 816                 return -ENOMEM;
 817
 818         ret = btrfs_search_slot(trans, quota_root, &key, path, 0, 1);
 819         if (ret > 0)
 820                 ret = -ENOENT;
 821
 822         if (ret)
 823                 goto out;
 824
 825         l = path->nodes[0];
 826         slot = path->slots[0];
 827         qgroup_info = btrfs_item_ptr(l, slot, struct btrfs_qgroup_info_item);
 828         btrfs_set_qgroup_info_generation(l, qgroup_info, trans->transid);
 829         btrfs_set_qgroup_info_rfer(l, qgroup_info, qgroup->rfer);
 830         btrfs_set_qgroup_info_rfer_cmpr(l, qgroup_info, qgroup->rfer_cmpr);
 831         btrfs_set_qgroup_info_excl(l, qgroup_info, qgroup->excl);
 832         btrfs_set_qgroup_info_excl_cmpr(l, qgroup_info, qgroup->excl_cmpr);
 833
 834         btrfs_mark_buffer_dirty(l);
 835
 836 out:
 837         btrfs_free_path(path);
 838         return ret;
 839 }
 840
 841 static int update_qgroup_status_item(struct btrfs_trans_handle *trans)
 842 {
 843         struct btrfs_fs_info *fs_info = trans->fs_info;
 844         struct btrfs_root *quota_root = fs_info->quota_root;
 845         struct btrfs_path *path;
 846         struct btrfs_key key;
 847         struct extent_buffer *l;
 848         struct btrfs_qgroup_status_item *ptr;
 849         int ret;
 850         int slot;
 851
 852         key.objectid = 0;
 853         key.type = BTRFS_QGROUP_STATUS_KEY;
 854         key.offset = 0;
 855
 856         path = btrfs_alloc_path();
 857         if (!path)
 858                 return -ENOMEM;
 859
 860         ret = btrfs_search_slot(trans, quota_root, &key, path, 0, 1);
 861         if (ret > 0)
 862                 ret = -ENOENT;
 863
 864         if (ret)
 865                 goto out;
 866
 867         l = path->nodes[0];
 868         slot = path->slots[0];
 869         ptr = btrfs_item_ptr(l, slot, struct btrfs_qgroup_status_item);
 870         btrfs_set_qgroup_status_flags(l, ptr, fs_info->qgroup_flags);
 871         btrfs_set_qgroup_status_generation(l, ptr, trans->transid);
 872         btrfs_set_qgroup_status_rescan(l, ptr,
 873                                 fs_info->qgroup_rescan_progress.objectid);
 874
 875         btrfs_mark_buffer_dirty(l);
 876
 877 out:
 878         btrfs_free_path(path);
 879         return ret;
 880 }
 881
 882 /*
 883  * called with qgroup_lock held
 884  */
 885 static int btrfs_clean_quota_tree(struct btrfs_trans_handle *trans,
 886                                   struct btrfs_root *root)
 887 {
 888         struct btrfs_path *path;
 889         struct btrfs_key key;
 890         struct extent_buffer *leaf = NULL;
 891         int ret;
 892         int nr = 0;
 893
 894         path = btrfs_alloc_path();
 895         if (!path)
 896                 return -ENOMEM;
 897
 898         key.objectid = 0;
 899         key.offset = 0;
 900         key.type = 0;
 901
 902         while (1) {
 903                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 904                 if (ret < 0)
 905                         goto out;
 906                 leaf = path->nodes[0];
 907                 nr = btrfs_header_nritems(leaf);
 908                 if (!nr)
 909                         break;
 910                 /*
 911                  * delete the leaf one by one
 912                  * since the whole tree is going
 913                  * to be deleted.
 914                  */
 915                 path->slots[0] = 0;
 916                 ret = btrfs_del_items(trans, root, path, 0, nr);
 917                 if (ret)
 918                         goto out;
 919
 920                 btrfs_release_path(path);
 921         }
 922         ret = 0;
 923 out:
 924         btrfs_free_path(path);
 925         return ret;
 926 }
 927
 928 int btrfs_quota_enable(struct btrfs_fs_info *fs_info)
 929 {
 930         struct btrfs_root *quota_root;
 931         struct btrfs_root *tree_root = fs_info->tree_root;
 932         struct btrfs_path *path = NULL;
 933         struct btrfs_qgroup_status_item *ptr;
 934         struct extent_buffer *leaf;
 935         struct btrfs_key key;
 936         struct btrfs_key found_key;
 937         struct btrfs_qgroup *qgroup = NULL;
 938         struct btrfs_trans_handle *trans = NULL;
 939         struct ulist *ulist = NULL;
 940         int ret = 0;
 941         int slot;
 942
 943         /*
 944          * We need to have subvol_sem write locked, to prevent races between
 945          * concurrent tasks trying to enable quotas, because we will unlock
 946          * and relock qgroup_ioctl_lock before setting fs_info->quota_root
 947          * and before setting BTRFS_FS_QUOTA_ENABLED.
 948          */
 949         lockdep_assert_held_write(&fs_info->subvol_sem);
 950
 951         mutex_lock(&fs_info->qgroup_ioctl_lock);
 952         if (fs_info->quota_root)
 953                 goto out;
 954
 955         ulist = ulist_alloc(GFP_KERNEL);
 956         if (!ulist) {
 957                 ret = -ENOMEM;
 958                 goto out;
 959         }
 960
 961         ret = btrfs_sysfs_add_qgroups(fs_info);
 962         if (ret < 0)
 963                 goto out;
 964
 965         /*
 966          * Unlock qgroup_ioctl_lock before starting the transaction. This is to
 967          * avoid lock acquisition inversion problems (reported by lockdep) between
 968          * qgroup_ioctl_lock and the vfs freeze semaphores, acquired when we
 969          * start a transaction.
 970          * After we started the transaction lock qgroup_ioctl_lock again and
 971          * check if someone else created the quota root in the meanwhile. If so,
 972          * just return success and release the transaction handle.
 973          *
 974          * Also we don't need to worry about someone else calling
 975          * btrfs_sysfs_add_qgroups() after we unlock and getting an error because
 976          * that function returns 0 (success) when the sysfs entries already exist.
 977          */
 978         mutex_unlock(&fs_info->qgroup_ioctl_lock);
 979
 980         /*
 981          * 1 for quota root item
 982          * 1 for BTRFS_QGROUP_STATUS item
 983          *
 984          * Yet we also need 2*n items for a QGROUP_INFO/QGROUP_LIMIT items
 985          * per subvolume. However those are not currently reserved since it
 986          * would be a lot of overkill.
 987          */
 988         trans = btrfs_start_transaction(tree_root, 2);
 989
 990         mutex_lock(&fs_info->qgroup_ioctl_lock);
 991         if (IS_ERR(trans)) {
 992                 ret = PTR_ERR(trans);
 993                 trans = NULL;
 994                 goto out;
 995         }
 996
 997         if (fs_info->quota_root)
 998                 goto out;
 999
1000         fs_info->qgroup_ulist = ulist;
1001         ulist = NULL;
1002
1003         /*
1004          * initially create the quota tree
1005          */
1006         quota_root = btrfs_create_tree(trans, BTRFS_QUOTA_TREE_OBJECTID);
1007         if (IS_ERR(quota_root)) {
1008                 ret =  PTR_ERR(quota_root);
1009                 btrfs_abort_transaction(trans, ret);
1010                 goto out;
1011         }
1012
1013         path = btrfs_alloc_path();
1014         if (!path) {
1015                 ret = -ENOMEM;
1016                 btrfs_abort_transaction(trans, ret);
1017                 goto out_free_root;
1018         }
1019
1020         key.objectid = 0;
1021         key.type = BTRFS_QGROUP_STATUS_KEY;
1022         key.offset = 0;
1023
1024         ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
1025                                       sizeof(*ptr));
1026         if (ret) {
1027                 btrfs_abort_transaction(trans, ret);
1028                 goto out_free_path;
1029         }
1030
1031         leaf = path->nodes[0];
1032         ptr = btrfs_item_ptr(leaf, path->slots[0],
1033                                  struct btrfs_qgroup_status_item);
1034         btrfs_set_qgroup_status_generation(leaf, ptr, trans->transid);
1035         btrfs_set_qgroup_status_version(leaf, ptr, BTRFS_QGROUP_STATUS_VERSION);
1036         fs_info->qgroup_flags = BTRFS_QGROUP_STATUS_FLAG_ON |
1037                                 BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1038         btrfs_set_qgroup_status_flags(leaf, ptr, fs_info->qgroup_flags);
1039         btrfs_set_qgroup_status_rescan(leaf, ptr, 0);
1040
1041         btrfs_mark_buffer_dirty(leaf);
1042
1043         key.objectid = 0;
1044         key.type = BTRFS_ROOT_REF_KEY;
1045         key.offset = 0;
1046
1047         btrfs_release_path(path);
1048         ret = btrfs_search_slot_for_read(tree_root, &key, path, 1, 0);
1049         if (ret > 0)
1050                 goto out_add_root;
1051         if (ret < 0) {
1052                 btrfs_abort_transaction(trans, ret);
1053                 goto out_free_path;
1054         }
1055
1056         while (1) {
1057                 slot = path->slots[0];
1058                 leaf = path->nodes[0];
1059                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1060
1061                 if (found_key.type == BTRFS_ROOT_REF_KEY) {
1062
1063                         /* Release locks on tree_root before we access quota_root */
1064                         btrfs_release_path(path);
1065
1066                         ret = add_qgroup_item(trans, quota_root,
1067                                               found_key.offset);
1068                         if (ret) {
1069                                 btrfs_abort_transaction(trans, ret);
1070                                 goto out_free_path;
1071                         }
1072
1073                         qgroup = add_qgroup_rb(fs_info, found_key.offset);
1074                         if (IS_ERR(qgroup)) {
1075                                 ret = PTR_ERR(qgroup);
1076                                 btrfs_abort_transaction(trans, ret);
1077                                 goto out_free_path;
1078                         }
1079                         ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1080                         if (ret < 0) {
1081                                 btrfs_abort_transaction(trans, ret);
1082                                 goto out_free_path;
1083                         }
1084                         ret = btrfs_search_slot_for_read(tree_root, &found_key,
1085                                                          path, 1, 0);
1086                         if (ret < 0) {
1087                                 btrfs_abort_transaction(trans, ret);
1088                                 goto out_free_path;
1089                         }
1090                         if (ret > 0) {
1091                                 /*
1092                                  * Shouldn't happen, but in case it does we
1093                                  * don't need to do the btrfs_next_item, just
1094                                  * continue.
1095                                  */
1096                                 continue;
1097                         }
1098                 }
1099                 ret = btrfs_next_item(tree_root, path);
1100                 if (ret < 0) {
1101                         btrfs_abort_transaction(trans, ret);
1102                         goto out_free_path;
1103                 }
1104                 if (ret)
1105                         break;
1106         }
1107
1108 out_add_root:
1109         btrfs_release_path(path);
1110         ret = add_qgroup_item(trans, quota_root, BTRFS_FS_TREE_OBJECTID);
1111         if (ret) {
1112                 btrfs_abort_transaction(trans, ret);
1113                 goto out_free_path;
1114         }
1115
1116         qgroup = add_qgroup_rb(fs_info, BTRFS_FS_TREE_OBJECTID);
1117         if (IS_ERR(qgroup)) {
1118                 ret = PTR_ERR(qgroup);
1119                 btrfs_abort_transaction(trans, ret);
1120                 goto out_free_path;
1121         }
1122         ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1123         if (ret < 0) {
1124                 btrfs_abort_transaction(trans, ret);
1125                 goto out_free_path;
1126         }
1127
1128         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1129         /*
1130          * Commit the transaction while not holding qgroup_ioctl_lock, to avoid
1131          * a deadlock with tasks concurrently doing other qgroup operations, such
1132          * adding/removing qgroups or adding/deleting qgroup relations for example,
1133          * because all qgroup operations first start or join a transaction and then
1134          * lock the qgroup_ioctl_lock mutex.
1135          * We are safe from a concurrent task trying to enable quotas, by calling
1136          * this function, since we are serialized by fs_info->subvol_sem.
1137          */
1138         ret = btrfs_commit_transaction(trans);
1139         trans = NULL;
1140         mutex_lock(&fs_info->qgroup_ioctl_lock);
1141         if (ret)
1142                 goto out_free_path;
1143
1144         /*
1145          * Set quota enabled flag after committing the transaction, to avoid
1146          * deadlocks on fs_info->qgroup_ioctl_lock with concurrent snapshot
1147          * creation.
1148          */
1149         spin_lock(&fs_info->qgroup_lock);
1150         fs_info->quota_root = quota_root;
1151         set_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1152         spin_unlock(&fs_info->qgroup_lock);
1153
1154         ret = qgroup_rescan_init(fs_info, 0, 1);
1155         if (!ret) {
1156                 qgroup_rescan_zero_tracking(fs_info);
1157                 fs_info->qgroup_rescan_running = true;
1158                 btrfs_queue_work(fs_info->qgroup_rescan_workers,
1159                                  &fs_info->qgroup_rescan_work);
1160         }
1161
1162 out_free_path:
1163         btrfs_free_path(path);
1164 out_free_root:
1165         if (ret)
1166                 btrfs_put_root(quota_root);
1167 out:
1168         if (ret) {
1169                 ulist_free(fs_info->qgroup_ulist);
1170                 fs_info->qgroup_ulist = NULL;
1171                 btrfs_sysfs_del_qgroups(fs_info);
1172         }
1173         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1174         if (ret && trans)
1175                 btrfs_end_transaction(trans);
1176         else if (trans)
1177                 ret = btrfs_end_transaction(trans);
1178         ulist_free(ulist);
1179         return ret;
1180 }
1181
1182 int btrfs_quota_disable(struct btrfs_fs_info *fs_info)
1183 {
1184         struct btrfs_root *quota_root;
1185         struct btrfs_trans_handle *trans = NULL;
1186         int ret = 0;
1187
1188         /*
1189          * We need to have subvol_sem write locked, to prevent races between
1190          * concurrent tasks trying to disable quotas, because we will unlock
1191          * and relock qgroup_ioctl_lock across BTRFS_FS_QUOTA_ENABLED changes.
1192          */
1193         lockdep_assert_held_write(&fs_info->subvol_sem);
1194
1195         mutex_lock(&fs_info->qgroup_ioctl_lock);
1196         if (!fs_info->quota_root)
1197                 goto out;
1198
1199         /*
1200          * Unlock the qgroup_ioctl_lock mutex before waiting for the rescan worker to
1201          * complete. Otherwise we can deadlock because btrfs_remove_qgroup() needs
1202          * to lock that mutex while holding a transaction handle and the rescan
1203          * worker needs to commit a transaction.
1204          */
1205         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1206
1207         /*
1208          * Request qgroup rescan worker to complete and wait for it. This wait
1209          * must be done before transaction start for quota disable since it may
1210          * deadlock with transaction by the qgroup rescan worker.
1211          */
1212         clear_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1213         btrfs_qgroup_wait_for_completion(fs_info, false);
1214
1215         /*
1216          * 1 For the root item
1217          *
1218          * We should also reserve enough items for the quota tree deletion in
1219          * btrfs_clean_quota_tree but this is not done.
1220          *
1221          * Also, we must always start a transaction without holding the mutex
1222          * qgroup_ioctl_lock, see btrfs_quota_enable().
1223          */
1224         trans = btrfs_start_transaction(fs_info->tree_root, 1);
1225
1226         mutex_lock(&fs_info->qgroup_ioctl_lock);
1227         if (IS_ERR(trans)) {
1228                 ret = PTR_ERR(trans);
1229                 trans = NULL;
1230                 set_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1231                 goto out;
1232         }
1233
1234         if (!fs_info->quota_root)
1235                 goto out;
1236
1237         spin_lock(&fs_info->qgroup_lock);
1238         quota_root = fs_info->quota_root;
1239         fs_info->quota_root = NULL;
1240         fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_ON;
1241         spin_unlock(&fs_info->qgroup_lock);
1242
1243         btrfs_free_qgroup_config(fs_info);
1244
1245         ret = btrfs_clean_quota_tree(trans, quota_root);
1246         if (ret) {
1247                 btrfs_abort_transaction(trans, ret);
1248                 goto out;
1249         }
1250
1251         ret = btrfs_del_root(trans, &quota_root->root_key);
1252         if (ret) {
1253                 btrfs_abort_transaction(trans, ret);
1254                 goto out;
1255         }
1256
1257         list_del(&quota_root->dirty_list);
1258
1259         btrfs_tree_lock(quota_root->node);
1260         btrfs_clean_tree_block(quota_root->node);
1261         btrfs_tree_unlock(quota_root->node);
1262         btrfs_free_tree_block(trans, quota_root, quota_root->node, 0, 1);
1263
1264         btrfs_put_root(quota_root);
1265
1266 out:
1267         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1268         if (ret && trans)
1269                 btrfs_end_transaction(trans);
1270         else if (trans)
1271                 ret = btrfs_end_transaction(trans);
1272
1273         return ret;
1274 }
1275
1276 static void qgroup_dirty(struct btrfs_fs_info *fs_info,
1277                          struct btrfs_qgroup *qgroup)
1278 {
1279         if (list_empty(&qgroup->dirty))
1280                 list_add(&qgroup->dirty, &fs_info->dirty_qgroups);
1281 }
1282
1283 /*
1284  * The easy accounting, we're updating qgroup relationship whose child qgroup
1285  * only has exclusive extents.
1286  *
1287  * In this case, all exclusive extents will also be exclusive for parent, so
1288  * excl/rfer just get added/removed.
1289  *
1290  * So is qgroup reservation space, which should also be added/removed to
1291  * parent.
1292  * Or when child tries to release reservation space, parent will underflow its
1293  * reservation (for relationship adding case).
1294  *
1295  * Caller should hold fs_info->qgroup_lock.
1296  */
1297 static int __qgroup_excl_accounting(struct btrfs_fs_info *fs_info,
1298                                     struct ulist *tmp, u64 ref_root,
1299                                     struct btrfs_qgroup *src, int sign)
1300 {
1301         struct btrfs_qgroup *qgroup;
1302         struct btrfs_qgroup_list *glist;
1303         struct ulist_node *unode;
1304         struct ulist_iterator uiter;
1305         u64 num_bytes = src->excl;
1306         int ret = 0;
1307
1308         qgroup = find_qgroup_rb(fs_info, ref_root);
1309         if (!qgroup)
1310                 goto out;
1311
1312         qgroup->rfer += sign * num_bytes;
1313         qgroup->rfer_cmpr += sign * num_bytes;
1314
1315         WARN_ON(sign < 0 && qgroup->excl < num_bytes);
1316         qgroup->excl += sign * num_bytes;
1317         qgroup->excl_cmpr += sign * num_bytes;
1318
1319         if (sign > 0)
1320                 qgroup_rsv_add_by_qgroup(fs_info, qgroup, src);
1321         else
1322                 qgroup_rsv_release_by_qgroup(fs_info, qgroup, src);
1323
1324         qgroup_dirty(fs_info, qgroup);
1325
1326         /* Get all of the parent groups that contain this qgroup */
1327         list_for_each_entry(glist, &qgroup->groups, next_group) {
1328                 ret = ulist_add(tmp, glist->group->qgroupid,
1329                                 qgroup_to_aux(glist->group), GFP_ATOMIC);
1330                 if (ret < 0)
1331                         goto out;
1332         }
1333
1334         /* Iterate all of the parents and adjust their reference counts */
1335         ULIST_ITER_INIT(&uiter);
1336         while ((unode = ulist_next(tmp, &uiter))) {
1337                 qgroup = unode_aux_to_qgroup(unode);
1338                 qgroup->rfer += sign * num_bytes;
1339                 qgroup->rfer_cmpr += sign * num_bytes;
1340                 WARN_ON(sign < 0 && qgroup->excl < num_bytes);
1341                 qgroup->excl += sign * num_bytes;
1342                 if (sign > 0)
1343                         qgroup_rsv_add_by_qgroup(fs_info, qgroup, src);
1344                 else
1345                         qgroup_rsv_release_by_qgroup(fs_info, qgroup, src);
1346                 qgroup->excl_cmpr += sign * num_bytes;
1347                 qgroup_dirty(fs_info, qgroup);
1348
1349                 /* Add any parents of the parents */
1350                 list_for_each_entry(glist, &qgroup->groups, next_group) {
1351                         ret = ulist_add(tmp, glist->group->qgroupid,
1352                                         qgroup_to_aux(glist->group), GFP_ATOMIC);
1353                         if (ret < 0)
1354                                 goto out;
1355                 }
1356         }
1357         ret = 0;
1358 out:
1359         return ret;
1360 }
1361
1362
1363 /*
1364  * Quick path for updating qgroup with only excl refs.
1365  *
1366  * In that case, just update all parent will be enough.
1367  * Or we needs to do a full rescan.
1368  * Caller should also hold fs_info->qgroup_lock.
1369  *
1370  * Return 0 for quick update, return >0 for need to full rescan
1371  * and mark INCONSISTENT flag.
1372  * Return < 0 for other error.
1373  */
1374 static int quick_update_accounting(struct btrfs_fs_info *fs_info,
1375                                    struct ulist *tmp, u64 src, u64 dst,
1376                                    int sign)
1377 {
1378         struct btrfs_qgroup *qgroup;
1379         int ret = 1;
1380         int err = 0;
1381
1382         qgroup = find_qgroup_rb(fs_info, src);
1383         if (!qgroup)
1384                 goto out;
1385         if (qgroup->excl == qgroup->rfer) {
1386                 ret = 0;
1387                 err = __qgroup_excl_accounting(fs_info, tmp, dst,
1388                                                qgroup, sign);
1389                 if (err < 0) {
1390                         ret = err;
1391                         goto out;
1392                 }
1393         }
1394 out:
1395         if (ret)
1396                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1397         return ret;
1398 }
1399
1400 int btrfs_add_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1401                               u64 dst)
1402 {
1403         struct btrfs_fs_info *fs_info = trans->fs_info;
1404         struct btrfs_qgroup *parent;
1405         struct btrfs_qgroup *member;
1406         struct btrfs_qgroup_list *list;
1407         struct ulist *tmp;
1408         unsigned int nofs_flag;
1409         int ret = 0;
1410
1411         /* Check the level of src and dst first */
1412         if (btrfs_qgroup_level(src) >= btrfs_qgroup_level(dst))
1413                 return -EINVAL;
1414
1415         /* We hold a transaction handle open, must do a NOFS allocation. */
1416         nofs_flag = memalloc_nofs_save();
1417         tmp = ulist_alloc(GFP_KERNEL);
1418         memalloc_nofs_restore(nofs_flag);
1419         if (!tmp)
1420                 return -ENOMEM;
1421
1422         mutex_lock(&fs_info->qgroup_ioctl_lock);
1423         if (!fs_info->quota_root) {
1424                 ret = -ENOTCONN;
1425                 goto out;
1426         }
1427         member = find_qgroup_rb(fs_info, src);
1428         parent = find_qgroup_rb(fs_info, dst);
1429         if (!member || !parent) {
1430                 ret = -EINVAL;
1431                 goto out;
1432         }
1433
1434         /* check if such qgroup relation exist firstly */
1435         list_for_each_entry(list, &member->groups, next_group) {
1436                 if (list->group == parent) {
1437                         ret = -EEXIST;
1438                         goto out;
1439                 }
1440         }
1441
1442         ret = add_qgroup_relation_item(trans, src, dst);
1443         if (ret)
1444                 goto out;
1445
1446         ret = add_qgroup_relation_item(trans, dst, src);
1447         if (ret) {
1448                 del_qgroup_relation_item(trans, src, dst);
1449                 goto out;
1450         }
1451
1452         spin_lock(&fs_info->qgroup_lock);
1453         ret = add_relation_rb(fs_info, src, dst);
1454         if (ret < 0) {
1455                 spin_unlock(&fs_info->qgroup_lock);
1456                 goto out;
1457         }
1458         ret = quick_update_accounting(fs_info, tmp, src, dst, 1);
1459         spin_unlock(&fs_info->qgroup_lock);
1460 out:
1461         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1462         ulist_free(tmp);
1463         return ret;
1464 }
1465
1466 static int __del_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1467                                  u64 dst)
1468 {
1469         struct btrfs_fs_info *fs_info = trans->fs_info;
1470         struct btrfs_qgroup *parent;
1471         struct btrfs_qgroup *member;
1472         struct btrfs_qgroup_list *list;
1473         struct ulist *tmp;
1474         bool found = false;
1475         unsigned int nofs_flag;
1476         int ret = 0;
1477         int ret2;
1478
1479         /* We hold a transaction handle open, must do a NOFS allocation. */
1480         nofs_flag = memalloc_nofs_save();
1481         tmp = ulist_alloc(GFP_KERNEL);
1482         memalloc_nofs_restore(nofs_flag);
1483         if (!tmp)
1484                 return -ENOMEM;
1485
1486         if (!fs_info->quota_root) {
1487                 ret = -ENOTCONN;
1488                 goto out;
1489         }
1490
1491         member = find_qgroup_rb(fs_info, src);
1492         parent = find_qgroup_rb(fs_info, dst);
1493         /*
1494          * The parent/member pair doesn't exist, then try to delete the dead
1495          * relation items only.
1496          */
1497         if (!member || !parent)
1498                 goto delete_item;
1499
1500         /* check if such qgroup relation exist firstly */
1501         list_for_each_entry(list, &member->groups, next_group) {
1502                 if (list->group == parent) {
1503                         found = true;
1504                         break;
1505                 }
1506         }
1507
1508 delete_item:
1509         ret = del_qgroup_relation_item(trans, src, dst);
1510         if (ret < 0 && ret != -ENOENT)
1511                 goto out;
1512         ret2 = del_qgroup_relation_item(trans, dst, src);
1513         if (ret2 < 0 && ret2 != -ENOENT)
1514                 goto out;
1515
1516         /* At least one deletion succeeded, return 0 */
1517         if (!ret || !ret2)
1518                 ret = 0;
1519
1520         if (found) {
1521                 spin_lock(&fs_info->qgroup_lock);
1522                 del_relation_rb(fs_info, src, dst);
1523                 ret = quick_update_accounting(fs_info, tmp, src, dst, -1);
1524                 spin_unlock(&fs_info->qgroup_lock);
1525         }
1526 out:
1527         ulist_free(tmp);
1528         return ret;
1529 }
1530
1531 int btrfs_del_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1532                               u64 dst)
1533 {
1534         struct btrfs_fs_info *fs_info = trans->fs_info;
1535         int ret = 0;
1536
1537         mutex_lock(&fs_info->qgroup_ioctl_lock);
1538         ret = __del_qgroup_relation(trans, src, dst);
1539         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1540
1541         return ret;
1542 }
1543
1544 int btrfs_create_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid)
1545 {
1546         struct btrfs_fs_info *fs_info = trans->fs_info;
1547         struct btrfs_root *quota_root;
1548         struct btrfs_qgroup *qgroup;
1549         int ret = 0;
1550
1551         mutex_lock(&fs_info->qgroup_ioctl_lock);
1552         if (!fs_info->quota_root) {
1553                 ret = -ENOTCONN;
1554                 goto out;
1555         }
1556         quota_root = fs_info->quota_root;
1557         qgroup = find_qgroup_rb(fs_info, qgroupid);
1558         if (qgroup) {
1559                 ret = -EEXIST;
1560                 goto out;
1561         }
1562
1563         ret = add_qgroup_item(trans, quota_root, qgroupid);
1564         if (ret)
1565                 goto out;
1566
1567         spin_lock(&fs_info->qgroup_lock);
1568         qgroup = add_qgroup_rb(fs_info, qgroupid);
1569         spin_unlock(&fs_info->qgroup_lock);
1570
1571         if (IS_ERR(qgroup)) {
1572                 ret = PTR_ERR(qgroup);
1573                 goto out;
1574         }
1575         ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1576 out:
1577         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1578         return ret;
1579 }
1580
1581 int btrfs_remove_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid)
1582 {
1583         struct btrfs_fs_info *fs_info = trans->fs_info;
1584         struct btrfs_qgroup *qgroup;
1585         struct btrfs_qgroup_list *list;
1586         int ret = 0;
1587
1588         mutex_lock(&fs_info->qgroup_ioctl_lock);
1589         if (!fs_info->quota_root) {
1590                 ret = -ENOTCONN;
1591                 goto out;
1592         }
1593
1594         qgroup = find_qgroup_rb(fs_info, qgroupid);
1595         if (!qgroup) {
1596                 ret = -ENOENT;
1597                 goto out;
1598         }
1599
1600         /* Check if there are no children of this qgroup */
1601         if (!list_empty(&qgroup->members)) {
1602                 ret = -EBUSY;
1603                 goto out;
1604         }
1605
1606         ret = del_qgroup_item(trans, qgroupid);
1607         if (ret && ret != -ENOENT)
1608                 goto out;
1609
1610         while (!list_empty(&qgroup->groups)) {
1611                 list = list_first_entry(&qgroup->groups,
1612                                         struct btrfs_qgroup_list, next_group);
1613                 ret = __del_qgroup_relation(trans, qgroupid,
1614                                             list->group->qgroupid);
1615                 if (ret)
1616                         goto out;
1617         }
1618
1619         spin_lock(&fs_info->qgroup_lock);
1620         del_qgroup_rb(fs_info, qgroupid);
1621         spin_unlock(&fs_info->qgroup_lock);
1622
1623         /*
1624          * Remove the qgroup from sysfs now without holding the qgroup_lock
1625          * spinlock, since the sysfs_remove_group() function needs to take
1626          * the mutex kernfs_mutex through kernfs_remove_by_name_ns().
1627          */
1628         btrfs_sysfs_del_one_qgroup(fs_info, qgroup);
1629         kfree(qgroup);
1630 out:
1631         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1632         return ret;
1633 }
1634
1635 int btrfs_limit_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid,
1636                        struct btrfs_qgroup_limit *limit)
1637 {
1638         struct btrfs_fs_info *fs_info = trans->fs_info;
1639         struct btrfs_qgroup *qgroup;
1640         int ret = 0;
1641         /* Sometimes we would want to clear the limit on this qgroup.
1642          * To meet this requirement, we treat the -1 as a special value
1643          * which tell kernel to clear the limit on this qgroup.
1644          */
1645         const u64 CLEAR_VALUE = -1;
1646
1647         mutex_lock(&fs_info->qgroup_ioctl_lock);
1648         if (!fs_info->quota_root) {
1649                 ret = -ENOTCONN;
1650                 goto out;
1651         }
1652
1653         qgroup = find_qgroup_rb(fs_info, qgroupid);
1654         if (!qgroup) {
1655                 ret = -ENOENT;
1656                 goto out;
1657         }
1658
1659         spin_lock(&fs_info->qgroup_lock);
1660         if (limit->flags & BTRFS_QGROUP_LIMIT_MAX_RFER) {
1661                 if (limit->max_rfer == CLEAR_VALUE) {
1662                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_MAX_RFER;
1663                         limit->flags &= ~BTRFS_QGROUP_LIMIT_MAX_RFER;
1664                         qgroup->max_rfer = 0;
1665                 } else {
1666                         qgroup->max_rfer = limit->max_rfer;
1667                 }
1668         }
1669         if (limit->flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) {
1670                 if (limit->max_excl == CLEAR_VALUE) {
1671                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_MAX_EXCL;
1672                         limit->flags &= ~BTRFS_QGROUP_LIMIT_MAX_EXCL;
1673                         qgroup->max_excl = 0;
1674                 } else {
1675                         qgroup->max_excl = limit->max_excl;
1676                 }
1677         }
1678         if (limit->flags & BTRFS_QGROUP_LIMIT_RSV_RFER) {
1679                 if (limit->rsv_rfer == CLEAR_VALUE) {
1680                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_RSV_RFER;
1681                         limit->flags &= ~BTRFS_QGROUP_LIMIT_RSV_RFER;
1682                         qgroup->rsv_rfer = 0;
1683                 } else {
1684                         qgroup->rsv_rfer = limit->rsv_rfer;
1685                 }
1686         }
1687         if (limit->flags & BTRFS_QGROUP_LIMIT_RSV_EXCL) {
1688                 if (limit->rsv_excl == CLEAR_VALUE) {
1689                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_RSV_EXCL;
1690                         limit->flags &= ~BTRFS_QGROUP_LIMIT_RSV_EXCL;
1691                         qgroup->rsv_excl = 0;
1692                 } else {
1693                         qgroup->rsv_excl = limit->rsv_excl;
1694                 }
1695         }
1696         qgroup->lim_flags |= limit->flags;
1697
1698         spin_unlock(&fs_info->qgroup_lock);
1699
1700         ret = update_qgroup_limit_item(trans, qgroup);
1701         if (ret) {
1702                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1703                 btrfs_info(fs_info, "unable to update quota limit for %llu",
1704                        qgroupid);
1705         }
1706
1707 out:
1708         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1709         return ret;
1710 }
1711
1712 int btrfs_qgroup_trace_extent_nolock(struct btrfs_fs_info *fs_info,
1713                                 struct btrfs_delayed_ref_root *delayed_refs,
1714                                 struct btrfs_qgroup_extent_record *record)
1715 {
1716         struct rb_node **p = &delayed_refs->dirty_extent_root.rb_node;
1717         struct rb_node *parent_node = NULL;
1718         struct btrfs_qgroup_extent_record *entry;
1719         u64 bytenr = record->bytenr;
1720
1721         lockdep_assert_held(&delayed_refs->lock);
1722         trace_btrfs_qgroup_trace_extent(fs_info, record);
1723
1724         while (*p) {
1725                 parent_node = *p;
1726                 entry = rb_entry(parent_node, struct btrfs_qgroup_extent_record,
1727                                  node);
1728                 if (bytenr < entry->bytenr) {
1729                         p = &(*p)->rb_left;
1730                 } else if (bytenr > entry->bytenr) {
1731                         p = &(*p)->rb_right;
1732                 } else {
1733                         if (record->data_rsv && !entry->data_rsv) {
1734                                 entry->data_rsv = record->data_rsv;
1735                                 entry->data_rsv_refroot =
1736                                         record->data_rsv_refroot;
1737                         }
1738                         return 1;
1739                 }
1740         }
1741
1742         rb_link_node(&record->node, parent_node, p);
1743         rb_insert_color(&record->node, &delayed_refs->dirty_extent_root);
1744         return 0;
1745 }
1746
1747 int btrfs_qgroup_trace_extent_post(struct btrfs_trans_handle *trans,
1748                                    struct btrfs_qgroup_extent_record *qrecord)
1749 {
1750         struct ulist *old_root;
1751         u64 bytenr = qrecord->bytenr;
1752         int ret;
1753
1754         /*
1755          * We are always called in a context where we are already holding a
1756          * transaction handle. Often we are called when adding a data delayed
1757          * reference from btrfs_truncate_inode_items() (truncating or unlinking),
1758          * in which case we will be holding a write lock on extent buffer from a
1759          * subvolume tree. In this case we can't allow btrfs_find_all_roots() to
1760          * acquire fs_info->commit_root_sem, because that is a higher level lock
1761          * that must be acquired before locking any extent buffers.
1762          *
1763          * So we want btrfs_find_all_roots() to not acquire the commit_root_sem
1764          * but we can't pass it a non-NULL transaction handle, because otherwise
1765          * it would not use commit roots and would lock extent buffers, causing
1766          * a deadlock if it ends up trying to read lock the same extent buffer
1767          * that was previously write locked at btrfs_truncate_inode_items().
1768          *
1769          * So pass a NULL transaction handle to btrfs_find_all_roots() and
1770          * explicitly tell it to not acquire the commit_root_sem - if we are
1771          * holding a transaction handle we don't need its protection.
1772          */
1773         ASSERT(trans != NULL);
1774
1775         ret = btrfs_find_all_roots(NULL, trans->fs_info, bytenr, 0, &old_root,
1776                                    true);
1777         if (ret < 0) {
1778                 trans->fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1779                 btrfs_warn(trans->fs_info,
1780 "error accounting new delayed refs extent (err code: %d), quota inconsistent",
1781                         ret);
1782                 return 0;
1783         }
1784
1785         /*
1786          * Here we don't need to get the lock of
1787          * trans->transaction->delayed_refs, since inserted qrecord won't
1788          * be deleted, only qrecord->node may be modified (new qrecord insert)
1789          *
1790          * So modifying qrecord->old_roots is safe here
1791          */
1792         qrecord->old_roots = old_root;
1793         return 0;
1794 }
1795
1796 int btrfs_qgroup_trace_extent(struct btrfs_trans_handle *trans, u64 bytenr,
1797                               u64 num_bytes, gfp_t gfp_flag)
1798 {
1799         struct btrfs_fs_info *fs_info = trans->fs_info;
1800         struct btrfs_qgroup_extent_record *record;
1801         struct btrfs_delayed_ref_root *delayed_refs;
1802         int ret;
1803
1804         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)
1805             || bytenr == 0 || num_bytes == 0)
1806                 return 0;
1807         record = kzalloc(sizeof(*record), gfp_flag);
1808         if (!record)
1809                 return -ENOMEM;
1810
1811         delayed_refs = &trans->transaction->delayed_refs;
1812         record->bytenr = bytenr;
1813         record->num_bytes = num_bytes;
1814         record->old_roots = NULL;
1815
1816         spin_lock(&delayed_refs->lock);
1817         ret = btrfs_qgroup_trace_extent_nolock(fs_info, delayed_refs, record);
1818         spin_unlock(&delayed_refs->lock);
1819         if (ret > 0) {
1820                 kfree(record);
1821                 return 0;
1822         }
1823         return btrfs_qgroup_trace_extent_post(trans, record);
1824 }
1825
1826 int btrfs_qgroup_trace_leaf_items(struct btrfs_trans_handle *trans,
1827                                   struct extent_buffer *eb)
1828 {
1829         struct btrfs_fs_info *fs_info = trans->fs_info;
1830         int nr = btrfs_header_nritems(eb);
1831         int i, extent_type, ret;
1832         struct btrfs_key key;
1833         struct btrfs_file_extent_item *fi;
1834         u64 bytenr, num_bytes;
1835
1836         /* We can be called directly from walk_up_proc() */
1837         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
1838                 return 0;
1839
1840         for (i = 0; i < nr; i++) {
1841                 btrfs_item_key_to_cpu(eb, &key, i);
1842
1843                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1844                         continue;
1845
1846                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
1847                 /* filter out non qgroup-accountable extents  */
1848                 extent_type = btrfs_file_extent_type(eb, fi);
1849
1850                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
1851                         continue;
1852
1853                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
1854                 if (!bytenr)
1855                         continue;
1856
1857                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
1858
1859                 ret = btrfs_qgroup_trace_extent(trans, bytenr, num_bytes,
1860                                                 GFP_NOFS);
1861                 if (ret)
1862                         return ret;
1863         }
1864         cond_resched();
1865         return 0;
1866 }
1867
1868 /*
1869  * Walk up the tree from the bottom, freeing leaves and any interior
1870  * nodes which have had all slots visited. If a node (leaf or
1871  * interior) is freed, the node above it will have it's slot
1872  * incremented. The root node will never be freed.
1873  *
1874  * At the end of this function, we should have a path which has all
1875  * slots incremented to the next position for a search. If we need to
1876  * read a new node it will be NULL and the node above it will have the
1877  * correct slot selected for a later read.
1878  *
1879  * If we increment the root nodes slot counter past the number of
1880  * elements, 1 is returned to signal completion of the search.
1881  */
1882 static int adjust_slots_upwards(struct btrfs_path *path, int root_level)
1883 {
1884         int level = 0;
1885         int nr, slot;
1886         struct extent_buffer *eb;
1887
1888         if (root_level == 0)
1889                 return 1;
1890
1891         while (level <= root_level) {
1892                 eb = path->nodes[level];
1893                 nr = btrfs_header_nritems(eb);
1894                 path->slots[level]++;
1895                 slot = path->slots[level];
1896                 if (slot >= nr || level == 0) {
1897                         /*
1898                          * Don't free the root -  we will detect this
1899                          * condition after our loop and return a
1900                          * positive value for caller to stop walking the tree.
1901                          */
1902                         if (level != root_level) {
1903                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
1904                                 path->locks[level] = 0;
1905
1906                                 free_extent_buffer(eb);
1907                                 path->nodes[level] = NULL;
1908                                 path->slots[level] = 0;
1909                         }
1910                 } else {
1911                         /*
1912                          * We have a valid slot to walk back down
1913                          * from. Stop here so caller can process these
1914                          * new nodes.
1915                          */
1916                         break;
1917                 }
1918
1919                 level++;
1920         }
1921
1922         eb = path->nodes[root_level];
1923         if (path->slots[root_level] >= btrfs_header_nritems(eb))
1924                 return 1;
1925
1926         return 0;
1927 }
1928
1929 /*
1930  * Helper function to trace a subtree tree block swap.
1931  *
1932  * The swap will happen in highest tree block, but there may be a lot of
1933  * tree blocks involved.
1934  *
1935  * For example:
1936  *  OO = Old tree blocks
1937  *  NN = New tree blocks allocated during balance
1938  *
1939  *           File tree (257)                  Reloc tree for 257
1940  * L2              OO                                NN
1941  *               /    \                            /    \
1942  * L1          OO      OO (a)                    OO      NN (a)
1943  *            / \     / \                       / \     / \
1944  * L0       OO   OO OO   OO                   OO   OO NN   NN
1945  *                  (b)  (c)                          (b)  (c)
1946  *
1947  * When calling qgroup_trace_extent_swap(), we will pass:
1948  * @src_eb = OO(a)
1949  * @dst_path = [ nodes[1] = NN(a), nodes[0] = NN(c) ]
1950  * @dst_level = 0
1951  * @root_level = 1
1952  *
1953  * In that case, qgroup_trace_extent_swap() will search from OO(a) to
1954  * reach OO(c), then mark both OO(c) and NN(c) as qgroup dirty.
1955  *
1956  * The main work of qgroup_trace_extent_swap() can be split into 3 parts:
1957  *
1958  * 1) Tree search from @src_eb
1959  *    It should acts as a simplified btrfs_search_slot().
1960  *    The key for search can be extracted from @dst_path->nodes[dst_level]
1961  *    (first key).
1962  *
1963  * 2) Mark the final tree blocks in @src_path and @dst_path qgroup dirty
1964  *    NOTE: In above case, OO(a) and NN(a) won't be marked qgroup dirty.
1965  *    They should be marked during previous (@dst_level = 1) iteration.
1966  *
1967  * 3) Mark file extents in leaves dirty
1968  *    We don't have good way to pick out new file extents only.
1969  *    So we still follow the old method by scanning all file extents in
1970  *    the leave.
1971  *
1972  * This function can free us from keeping two paths, thus later we only need
1973  * to care about how to iterate all new tree blocks in reloc tree.
1974  */
1975 static int qgroup_trace_extent_swap(struct btrfs_trans_handle* trans,
1976                                     struct extent_buffer *src_eb,
1977                                     struct btrfs_path *dst_path,
1978                                     int dst_level, int root_level,
1979                                     bool trace_leaf)
1980 {
1981         struct btrfs_key key;
1982         struct btrfs_path *src_path;
1983         struct btrfs_fs_info *fs_info = trans->fs_info;
1984         u32 nodesize = fs_info->nodesize;
1985         int cur_level = root_level;
1986         int ret;
1987
1988         BUG_ON(dst_level > root_level);
1989         /* Level mismatch */
1990         if (btrfs_header_level(src_eb) != root_level)
1991                 return -EINVAL;
1992
1993         src_path = btrfs_alloc_path();
1994         if (!src_path) {
1995                 ret = -ENOMEM;
1996                 goto out;
1997         }
1998
1999         if (dst_level)
2000                 btrfs_node_key_to_cpu(dst_path->nodes[dst_level], &key, 0);
2001         else
2002                 btrfs_item_key_to_cpu(dst_path->nodes[dst_level], &key, 0);
2003
2004         /* For src_path */
2005         atomic_inc(&src_eb->refs);
2006         src_path->nodes[root_level] = src_eb;
2007         src_path->slots[root_level] = dst_path->slots[root_level];
2008         src_path->locks[root_level] = 0;
2009
2010         /* A simplified version of btrfs_search_slot() */
2011         while (cur_level >= dst_level) {
2012                 struct btrfs_key src_key;
2013                 struct btrfs_key dst_key;
2014
2015                 if (src_path->nodes[cur_level] == NULL) {
2016                         struct extent_buffer *eb;
2017                         int parent_slot;
2018
2019                         eb = src_path->nodes[cur_level + 1];
2020                         parent_slot = src_path->slots[cur_level + 1];
2021
2022                         eb = btrfs_read_node_slot(eb, parent_slot);
2023                         if (IS_ERR(eb)) {
2024                                 ret = PTR_ERR(eb);
2025                                 goto out;
2026                         }
2027
2028                         src_path->nodes[cur_level] = eb;
2029
2030                         btrfs_tree_read_lock(eb);
2031                         src_path->locks[cur_level] = BTRFS_READ_LOCK;
2032                 }
2033
2034                 src_path->slots[cur_level] = dst_path->slots[cur_level];
2035                 if (cur_level) {
2036                         btrfs_node_key_to_cpu(dst_path->nodes[cur_level],
2037                                         &dst_key, dst_path->slots[cur_level]);
2038                         btrfs_node_key_to_cpu(src_path->nodes[cur_level],
2039                                         &src_key, src_path->slots[cur_level]);
2040                 } else {
2041                         btrfs_item_key_to_cpu(dst_path->nodes[cur_level],
2042                                         &dst_key, dst_path->slots[cur_level]);
2043                         btrfs_item_key_to_cpu(src_path->nodes[cur_level],
2044                                         &src_key, src_path->slots[cur_level]);
2045                 }
2046                 /* Content mismatch, something went wrong */
2047                 if (btrfs_comp_cpu_keys(&dst_key, &src_key)) {
2048                         ret = -ENOENT;
2049                         goto out;
2050                 }
2051                 cur_level--;
2052         }
2053
2054         /*
2055          * Now both @dst_path and @src_path have been populated, record the tree
2056          * blocks for qgroup accounting.
2057          */
2058         ret = btrfs_qgroup_trace_extent(trans, src_path->nodes[dst_level]->start,
2059                         nodesize, GFP_NOFS);
2060         if (ret < 0)
2061                 goto out;
2062         ret = btrfs_qgroup_trace_extent(trans,
2063                         dst_path->nodes[dst_level]->start,
2064                         nodesize, GFP_NOFS);
2065         if (ret < 0)
2066                 goto out;
2067
2068         /* Record leaf file extents */
2069         if (dst_level == 0 && trace_leaf) {
2070                 ret = btrfs_qgroup_trace_leaf_items(trans, src_path->nodes[0]);
2071                 if (ret < 0)
2072                         goto out;
2073                 ret = btrfs_qgroup_trace_leaf_items(trans, dst_path->nodes[0]);
2074         }
2075 out:
2076         btrfs_free_path(src_path);
2077         return ret;
2078 }
2079
2080 /*
2081  * Helper function to do recursive generation-aware depth-first search, to
2082  * locate all new tree blocks in a subtree of reloc tree.
2083  *
2084  * E.g. (OO = Old tree blocks, NN = New tree blocks, whose gen == last_snapshot)
2085  *         reloc tree
2086  * L2         NN (a)
2087  *          /    \
2088  * L1    OO        NN (b)
2089  *      /  \      /  \
2090  * L0  OO  OO    OO  NN
2091  *               (c) (d)
2092  * If we pass:
2093  * @dst_path = [ nodes[1] = NN(b), nodes[0] = NULL ],
2094  * @cur_level = 1
2095  * @root_level = 1
2096  *
2097  * We will iterate through tree blocks NN(b), NN(d) and info qgroup to trace
2098  * above tree blocks along with their counter parts in file tree.
2099  * While during search, old tree blocks OO(c) will be skipped as tree block swap
2100  * won't affect OO(c).
2101  */
2102 static int qgroup_trace_new_subtree_blocks(struct btrfs_trans_handle* trans,
2103                                            struct extent_buffer *src_eb,
2104                                            struct btrfs_path *dst_path,
2105                                            int cur_level, int root_level,
2106                                            u64 last_snapshot, bool trace_leaf)
2107 {
2108         struct btrfs_fs_info *fs_info = trans->fs_info;
2109         struct extent_buffer *eb;
2110         bool need_cleanup = false;
2111         int ret = 0;
2112         int i;
2113
2114         /* Level sanity check */
2115         if (cur_level < 0 || cur_level >= BTRFS_MAX_LEVEL - 1 ||
2116             root_level < 0 || root_level >= BTRFS_MAX_LEVEL - 1 ||
2117             root_level < cur_level) {
2118                 btrfs_err_rl(fs_info,
2119                         "%s: bad levels, cur_level=%d root_level=%d",
2120                         __func__, cur_level, root_level);
2121                 return -EUCLEAN;
2122         }
2123
2124         /* Read the tree block if needed */
2125         if (dst_path->nodes[cur_level] == NULL) {
2126                 int parent_slot;
2127                 u64 child_gen;
2128
2129                 /*
2130                  * dst_path->nodes[root_level] must be initialized before
2131                  * calling this function.
2132                  */
2133                 if (cur_level == root_level) {
2134                         btrfs_err_rl(fs_info,
2135         "%s: dst_path->nodes[%d] not initialized, root_level=%d cur_level=%d",
2136                                 __func__, root_level, root_level, cur_level);
2137                         return -EUCLEAN;
2138                 }
2139
2140                 /*
2141                  * We need to get child blockptr/gen from parent before we can
2142                  * read it.
2143                   */
2144                 eb = dst_path->nodes[cur_level + 1];
2145                 parent_slot = dst_path->slots[cur_level + 1];
2146                 child_gen = btrfs_node_ptr_generation(eb, parent_slot);
2147
2148                 /* This node is old, no need to trace */
2149                 if (child_gen < last_snapshot)
2150                         goto out;
2151
2152                 eb = btrfs_read_node_slot(eb, parent_slot);
2153                 if (IS_ERR(eb)) {
2154                         ret = PTR_ERR(eb);
2155                         goto out;
2156                 }
2157
2158                 dst_path->nodes[cur_level] = eb;
2159                 dst_path->slots[cur_level] = 0;
2160
2161                 btrfs_tree_read_lock(eb);
2162                 dst_path->locks[cur_level] = BTRFS_READ_LOCK;
2163                 need_cleanup = true;
2164         }
2165
2166         /* Now record this tree block and its counter part for qgroups */
2167         ret = qgroup_trace_extent_swap(trans, src_eb, dst_path, cur_level,
2168                                        root_level, trace_leaf);
2169         if (ret < 0)
2170                 goto cleanup;
2171
2172         eb = dst_path->nodes[cur_level];
2173
2174         if (cur_level > 0) {
2175                 /* Iterate all child tree blocks */
2176                 for (i = 0; i < btrfs_header_nritems(eb); i++) {
2177                         /* Skip old tree blocks as they won't be swapped */
2178                         if (btrfs_node_ptr_generation(eb, i) < last_snapshot)
2179                                 continue;
2180                         dst_path->slots[cur_level] = i;
2181
2182                         /* Recursive call (at most 7 times) */
2183                         ret = qgroup_trace_new_subtree_blocks(trans, src_eb,
2184                                         dst_path, cur_level - 1, root_level,
2185                                         last_snapshot, trace_leaf);
2186                         if (ret < 0)
2187                                 goto cleanup;
2188                 }
2189         }
2190
2191 cleanup:
2192         if (need_cleanup) {
2193                 /* Clean up */
2194                 btrfs_tree_unlock_rw(dst_path->nodes[cur_level],
2195                                      dst_path->locks[cur_level]);
2196                 free_extent_buffer(dst_path->nodes[cur_level]);
2197                 dst_path->nodes[cur_level] = NULL;
2198                 dst_path->slots[cur_level] = 0;
2199                 dst_path->locks[cur_level] = 0;
2200         }
2201 out:
2202         return ret;
2203 }
2204
2205 static int qgroup_trace_subtree_swap(struct btrfs_trans_handle *trans,
2206                                 struct extent_buffer *src_eb,
2207                                 struct extent_buffer *dst_eb,
2208                                 u64 last_snapshot, bool trace_leaf)
2209 {
2210         struct btrfs_fs_info *fs_info = trans->fs_info;
2211         struct btrfs_path *dst_path = NULL;
2212         int level;
2213         int ret;
2214
2215         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2216                 return 0;
2217
2218         /* Wrong parameter order */
2219         if (btrfs_header_generation(src_eb) > btrfs_header_generation(dst_eb)) {
2220                 btrfs_err_rl(fs_info,
2221                 "%s: bad parameter order, src_gen=%llu dst_gen=%llu", __func__,
2222                              btrfs_header_generation(src_eb),
2223                              btrfs_header_generation(dst_eb));
2224                 return -EUCLEAN;
2225         }
2226
2227         if (!extent_buffer_uptodate(src_eb) || !extent_buffer_uptodate(dst_eb)) {
2228                 ret = -EIO;
2229                 goto out;
2230         }
2231
2232         level = btrfs_header_level(dst_eb);
2233         dst_path = btrfs_alloc_path();
2234         if (!dst_path) {
2235                 ret = -ENOMEM;
2236                 goto out;
2237         }
2238         /* For dst_path */
2239         atomic_inc(&dst_eb->refs);
2240         dst_path->nodes[level] = dst_eb;
2241         dst_path->slots[level] = 0;
2242         dst_path->locks[level] = 0;
2243
2244         /* Do the generation aware breadth-first search */
2245         ret = qgroup_trace_new_subtree_blocks(trans, src_eb, dst_path, level,
2246                                               level, last_snapshot, trace_leaf);
2247         if (ret < 0)
2248                 goto out;
2249         ret = 0;
2250
2251 out:
2252         btrfs_free_path(dst_path);
2253         if (ret < 0)
2254                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2255         return ret;
2256 }
2257
2258 int btrfs_qgroup_trace_subtree(struct btrfs_trans_handle *trans,
2259                                struct extent_buffer *root_eb,
2260                                u64 root_gen, int root_level)
2261 {
2262         struct btrfs_fs_info *fs_info = trans->fs_info;
2263         int ret = 0;
2264         int level;
2265         struct extent_buffer *eb = root_eb;
2266         struct btrfs_path *path = NULL;
2267
2268         BUG_ON(root_level < 0 || root_level >= BTRFS_MAX_LEVEL);
2269         BUG_ON(root_eb == NULL);
2270
2271         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2272                 return 0;
2273
2274         if (!extent_buffer_uptodate(root_eb)) {
2275                 ret = btrfs_read_buffer(root_eb, root_gen, root_level, NULL);
2276                 if (ret)
2277                         goto out;
2278         }
2279
2280         if (root_level == 0) {
2281                 ret = btrfs_qgroup_trace_leaf_items(trans, root_eb);
2282                 goto out;
2283         }
2284
2285         path = btrfs_alloc_path();
2286         if (!path)
2287                 return -ENOMEM;
2288
2289         /*
2290          * Walk down the tree.  Missing extent blocks are filled in as
2291          * we go. Metadata is accounted every time we read a new
2292          * extent block.
2293          *
2294          * When we reach a leaf, we account for file extent items in it,
2295          * walk back up the tree (adjusting slot pointers as we go)
2296          * and restart the search process.
2297          */
2298         atomic_inc(&root_eb->refs);     /* For path */
2299         path->nodes[root_level] = root_eb;
2300         path->slots[root_level] = 0;
2301         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
2302 walk_down:
2303         level = root_level;
2304         while (level >= 0) {
2305                 if (path->nodes[level] == NULL) {
2306                         int parent_slot;
2307                         u64 child_bytenr;
2308
2309                         /*
2310                          * We need to get child blockptr from parent before we
2311                          * can read it.
2312                           */
2313                         eb = path->nodes[level + 1];
2314                         parent_slot = path->slots[level + 1];
2315                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
2316
2317                         eb = btrfs_read_node_slot(eb, parent_slot);
2318                         if (IS_ERR(eb)) {
2319                                 ret = PTR_ERR(eb);
2320                                 goto out;
2321                         }
2322
2323                         path->nodes[level] = eb;
2324                         path->slots[level] = 0;
2325
2326                         btrfs_tree_read_lock(eb);
2327                         path->locks[level] = BTRFS_READ_LOCK;
2328
2329                         ret = btrfs_qgroup_trace_extent(trans, child_bytenr,
2330                                                         fs_info->nodesize,
2331                                                         GFP_NOFS);
2332                         if (ret)
2333                                 goto out;
2334                 }
2335
2336                 if (level == 0) {
2337                         ret = btrfs_qgroup_trace_leaf_items(trans,
2338                                                             path->nodes[level]);
2339                         if (ret)
2340                                 goto out;
2341
2342                         /* Nonzero return here means we completed our search */
2343                         ret = adjust_slots_upwards(path, root_level);
2344                         if (ret)
2345                                 break;
2346
2347                         /* Restart search with new slots */
2348                         goto walk_down;
2349                 }
2350
2351                 level--;
2352         }
2353
2354         ret = 0;
2355 out:
2356         btrfs_free_path(path);
2357
2358         return ret;
2359 }
2360
2361 #define UPDATE_NEW      0
2362 #define UPDATE_OLD      1
2363 /*
2364  * Walk all of the roots that points to the bytenr and adjust their refcnts.
2365  */
2366 static int qgroup_update_refcnt(struct btrfs_fs_info *fs_info,
2367                                 struct ulist *roots, struct ulist *tmp,
2368                                 struct ulist *qgroups, u64 seq, int update_old)
2369 {
2370         struct ulist_node *unode;
2371         struct ulist_iterator uiter;
2372         struct ulist_node *tmp_unode;
2373         struct ulist_iterator tmp_uiter;
2374         struct btrfs_qgroup *qg;
2375         int ret = 0;
2376
2377         if (!roots)
2378                 return 0;
2379         ULIST_ITER_INIT(&uiter);
2380         while ((unode = ulist_next(roots, &uiter))) {
2381                 qg = find_qgroup_rb(fs_info, unode->val);
2382                 if (!qg)
2383                         continue;
2384
2385                 ulist_reinit(tmp);
2386                 ret = ulist_add(qgroups, qg->qgroupid, qgroup_to_aux(qg),
2387                                 GFP_ATOMIC);
2388                 if (ret < 0)
2389                         return ret;
2390                 ret = ulist_add(tmp, qg->qgroupid, qgroup_to_aux(qg), GFP_ATOMIC);
2391                 if (ret < 0)
2392                         return ret;
2393                 ULIST_ITER_INIT(&tmp_uiter);
2394                 while ((tmp_unode = ulist_next(tmp, &tmp_uiter))) {
2395                         struct btrfs_qgroup_list *glist;
2396
2397                         qg = unode_aux_to_qgroup(tmp_unode);
2398                         if (update_old)
2399                                 btrfs_qgroup_update_old_refcnt(qg, seq, 1);
2400                         else
2401                                 btrfs_qgroup_update_new_refcnt(qg, seq, 1);
2402                         list_for_each_entry(glist, &qg->groups, next_group) {
2403                                 ret = ulist_add(qgroups, glist->group->qgroupid,
2404                                                 qgroup_to_aux(glist->group),
2405                                                 GFP_ATOMIC);
2406                                 if (ret < 0)
2407                                         return ret;
2408                                 ret = ulist_add(tmp, glist->group->qgroupid,
2409                                                 qgroup_to_aux(glist->group),
2410                                                 GFP_ATOMIC);
2411                                 if (ret < 0)
2412                                         return ret;
2413                         }
2414                 }
2415         }
2416         return 0;
2417 }
2418
2419 /*
2420  * Update qgroup rfer/excl counters.
2421  * Rfer update is easy, codes can explain themselves.
2422  *
2423  * Excl update is tricky, the update is split into 2 parts.
2424  * Part 1: Possible exclusive <-> sharing detect:
2425  *      |       A       |       !A      |
2426  *  -------------------------------------
2427  *  B   |       *       |       -       |
2428  *  -------------------------------------
2429  *  !B  |       +       |       **      |
2430  *  -------------------------------------
2431  *
2432  * Conditions:
2433  * A:   cur_old_roots < nr_old_roots    (not exclusive before)
2434  * !A:  cur_old_roots == nr_old_roots   (possible exclusive before)
2435  * B:   cur_new_roots < nr_new_roots    (not exclusive now)
2436  * !B:  cur_new_roots == nr_new_roots   (possible exclusive now)
2437  *
2438  * Results:
2439  * +: Possible sharing -> exclusive     -: Possible exclusive -> sharing
2440  * *: Definitely not changed.           **: Possible unchanged.
2441  *
2442  * For !A and !B condition, the exception is cur_old/new_roots == 0 case.
2443  *
2444  * To make the logic clear, we first use condition A and B to split
2445  * combination into 4 results.
2446  *
2447  * Then, for result "+" and "-", check old/new_roots == 0 case, as in them
2448  * only on variant maybe 0.
2449  *
2450  * Lastly, check result **, since there are 2 variants maybe 0, split them
2451  * again(2x2).
2452  * But this time we don't need to consider other things, the codes and logic
2453  * is easy to understand now.
2454  */
2455 static int qgroup_update_counters(struct btrfs_fs_info *fs_info,
2456                                   struct ulist *qgroups,
2457                                   u64 nr_old_roots,
2458                                   u64 nr_new_roots,
2459                                   u64 num_bytes, u64 seq)
2460 {
2461         struct ulist_node *unode;
2462         struct ulist_iterator uiter;
2463         struct btrfs_qgroup *qg;
2464         u64 cur_new_count, cur_old_count;
2465
2466         ULIST_ITER_INIT(&uiter);
2467         while ((unode = ulist_next(qgroups, &uiter))) {
2468                 bool dirty = false;
2469
2470                 qg = unode_aux_to_qgroup(unode);
2471                 cur_old_count = btrfs_qgroup_get_old_refcnt(qg, seq);
2472                 cur_new_count = btrfs_qgroup_get_new_refcnt(qg, seq);
2473
2474                 trace_qgroup_update_counters(fs_info, qg, cur_old_count,
2475                                              cur_new_count);
2476
2477                 /* Rfer update part */
2478                 if (cur_old_count == 0 && cur_new_count > 0) {
2479                         qg->rfer += num_bytes;
2480                         qg->rfer_cmpr += num_bytes;
2481                         dirty = true;
2482                 }
2483                 if (cur_old_count > 0 && cur_new_count == 0) {
2484                         qg->rfer -= num_bytes;
2485                         qg->rfer_cmpr -= num_bytes;
2486                         dirty = true;
2487                 }
2488
2489                 /* Excl update part */
2490                 /* Exclusive/none -> shared case */
2491                 if (cur_old_count == nr_old_roots &&
2492                     cur_new_count < nr_new_roots) {
2493                         /* Exclusive -> shared */
2494                         if (cur_old_count != 0) {
2495                                 qg->excl -= num_bytes;
2496                                 qg->excl_cmpr -= num_bytes;
2497                                 dirty = true;
2498                         }
2499                 }
2500
2501                 /* Shared -> exclusive/none case */
2502                 if (cur_old_count < nr_old_roots &&
2503                     cur_new_count == nr_new_roots) {
2504                         /* Shared->exclusive */
2505                         if (cur_new_count != 0) {
2506                                 qg->excl += num_bytes;
2507                                 qg->excl_cmpr += num_bytes;
2508                                 dirty = true;
2509                         }
2510                 }
2511
2512                 /* Exclusive/none -> exclusive/none case */
2513                 if (cur_old_count == nr_old_roots &&
2514                     cur_new_count == nr_new_roots) {
2515                         if (cur_old_count == 0) {
2516                                 /* None -> exclusive/none */
2517
2518                                 if (cur_new_count != 0) {
2519                                         /* None -> exclusive */
2520                                         qg->excl += num_bytes;
2521                                         qg->excl_cmpr += num_bytes;
2522                                         dirty = true;
2523                                 }
2524                                 /* None -> none, nothing changed */
2525                         } else {
2526                                 /* Exclusive -> exclusive/none */
2527
2528                                 if (cur_new_count == 0) {
2529                                         /* Exclusive -> none */
2530                                         qg->excl -= num_bytes;
2531                                         qg->excl_cmpr -= num_bytes;
2532                                         dirty = true;
2533                                 }
2534                                 /* Exclusive -> exclusive, nothing changed */
2535                         }
2536                 }
2537
2538                 if (dirty)
2539                         qgroup_dirty(fs_info, qg);
2540         }
2541         return 0;
2542 }
2543
2544 /*
2545  * Check if the @roots potentially is a list of fs tree roots
2546  *
2547  * Return 0 for definitely not a fs/subvol tree roots ulist
2548  * Return 1 for possible fs/subvol tree roots in the list (considering an empty
2549  *          one as well)
2550  */
2551 static int maybe_fs_roots(struct ulist *roots)
2552 {
2553         struct ulist_node *unode;
2554         struct ulist_iterator uiter;
2555
2556         /* Empty one, still possible for fs roots */
2557         if (!roots || roots->nnodes == 0)
2558                 return 1;
2559
2560         ULIST_ITER_INIT(&uiter);
2561         unode = ulist_next(roots, &uiter);
2562         if (!unode)
2563                 return 1;
2564
2565         /*
2566          * If it contains fs tree roots, then it must belong to fs/subvol
2567          * trees.
2568          * If it contains a non-fs tree, it won't be shared with fs/subvol trees.
2569          */
2570         return is_fstree(unode->val);
2571 }
2572
2573 int btrfs_qgroup_account_extent(struct btrfs_trans_handle *trans, u64 bytenr,
2574                                 u64 num_bytes, struct ulist *old_roots,
2575                                 struct ulist *new_roots)
2576 {
2577         struct btrfs_fs_info *fs_info = trans->fs_info;
2578         struct ulist *qgroups = NULL;
2579         struct ulist *tmp = NULL;
2580         u64 seq;
2581         u64 nr_new_roots = 0;
2582         u64 nr_old_roots = 0;
2583         int ret = 0;
2584
2585         /*
2586          * If quotas get disabled meanwhile, the resources need to be freed and
2587          * we can't just exit here.
2588          */
2589         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2590                 goto out_free;
2591
2592         if (new_roots) {
2593                 if (!maybe_fs_roots(new_roots))
2594                         goto out_free;
2595                 nr_new_roots = new_roots->nnodes;
2596         }
2597         if (old_roots) {
2598                 if (!maybe_fs_roots(old_roots))
2599                         goto out_free;
2600                 nr_old_roots = old_roots->nnodes;
2601         }
2602
2603         /* Quick exit, either not fs tree roots, or won't affect any qgroup */
2604         if (nr_old_roots == 0 && nr_new_roots == 0)
2605                 goto out_free;
2606
2607         BUG_ON(!fs_info->quota_root);
2608
2609         trace_btrfs_qgroup_account_extent(fs_info, trans->transid, bytenr,
2610                                         num_bytes, nr_old_roots, nr_new_roots);
2611
2612         qgroups = ulist_alloc(GFP_NOFS);
2613         if (!qgroups) {
2614                 ret = -ENOMEM;
2615                 goto out_free;
2616         }
2617         tmp = ulist_alloc(GFP_NOFS);
2618         if (!tmp) {
2619                 ret = -ENOMEM;
2620                 goto out_free;
2621         }
2622
2623         mutex_lock(&fs_info->qgroup_rescan_lock);
2624         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
2625                 if (fs_info->qgroup_rescan_progress.objectid <= bytenr) {
2626                         mutex_unlock(&fs_info->qgroup_rescan_lock);
2627                         ret = 0;
2628                         goto out_free;
2629                 }
2630         }
2631         mutex_unlock(&fs_info->qgroup_rescan_lock);
2632
2633         spin_lock(&fs_info->qgroup_lock);
2634         seq = fs_info->qgroup_seq;
2635
2636         /* Update old refcnts using old_roots */
2637         ret = qgroup_update_refcnt(fs_info, old_roots, tmp, qgroups, seq,
2638                                    UPDATE_OLD);
2639         if (ret < 0)
2640                 goto out;
2641
2642         /* Update new refcnts using new_roots */
2643         ret = qgroup_update_refcnt(fs_info, new_roots, tmp, qgroups, seq,
2644                                    UPDATE_NEW);
2645         if (ret < 0)
2646                 goto out;
2647
2648         qgroup_update_counters(fs_info, qgroups, nr_old_roots, nr_new_roots,
2649                                num_bytes, seq);
2650
2651         /*
2652          * Bump qgroup_seq to avoid seq overlap
2653          */
2654         fs_info->qgroup_seq += max(nr_old_roots, nr_new_roots) + 1;
2655 out:
2656         spin_unlock(&fs_info->qgroup_lock);
2657 out_free:
2658         ulist_free(tmp);
2659         ulist_free(qgroups);
2660         ulist_free(old_roots);
2661         ulist_free(new_roots);
2662         return ret;
2663 }
2664
2665 int btrfs_qgroup_account_extents(struct btrfs_trans_handle *trans)
2666 {
2667         struct btrfs_fs_info *fs_info = trans->fs_info;
2668         struct btrfs_qgroup_extent_record *record;
2669         struct btrfs_delayed_ref_root *delayed_refs;
2670         struct ulist *new_roots = NULL;
2671         struct rb_node *node;
2672         u64 num_dirty_extents = 0;
2673         u64 qgroup_to_skip;
2674         int ret = 0;
2675
2676         delayed_refs = &trans->transaction->delayed_refs;
2677         qgroup_to_skip = delayed_refs->qgroup_to_skip;
2678         while ((node = rb_first(&delayed_refs->dirty_extent_root))) {
2679                 record = rb_entry(node, struct btrfs_qgroup_extent_record,
2680                                   node);
2681
2682                 num_dirty_extents++;
2683                 trace_btrfs_qgroup_account_extents(fs_info, record);
2684
2685                 if (!ret) {
2686                         /*
2687                          * Old roots should be searched when inserting qgroup
2688                          * extent record
2689                          */
2690                         if (WARN_ON(!record->old_roots)) {
2691                                 /* Search commit root to find old_roots */
2692                                 ret = btrfs_find_all_roots(NULL, fs_info,
2693                                                 record->bytenr, 0,
2694                                                 &record->old_roots, false);
2695                                 if (ret < 0)
2696                                         goto cleanup;
2697                         }
2698
2699                         /* Free the reserved data space */
2700                         btrfs_qgroup_free_refroot(fs_info,
2701                                         record->data_rsv_refroot,
2702                                         record->data_rsv,
2703                                         BTRFS_QGROUP_RSV_DATA);
2704                         /*
2705                          * Use BTRFS_SEQ_LAST as time_seq to do special search,
2706                          * which doesn't lock tree or delayed_refs and search
2707                          * current root. It's safe inside commit_transaction().
2708                          */
2709                         ret = btrfs_find_all_roots(trans, fs_info,
2710                            record->bytenr, BTRFS_SEQ_LAST, &new_roots, false);
2711                         if (ret < 0)
2712                                 goto cleanup;
2713                         if (qgroup_to_skip) {
2714                                 ulist_del(new_roots, qgroup_to_skip, 0);
2715                                 ulist_del(record->old_roots, qgroup_to_skip,
2716                                           0);
2717                         }
2718                         ret = btrfs_qgroup_account_extent(trans, record->bytenr,
2719                                                           record->num_bytes,
2720                                                           record->old_roots,
2721                                                           new_roots);
2722                         record->old_roots = NULL;
2723                         new_roots = NULL;
2724                 }
2725 cleanup:
2726                 ulist_free(record->old_roots);
2727                 ulist_free(new_roots);
2728                 new_roots = NULL;
2729                 rb_erase(node, &delayed_refs->dirty_extent_root);
2730                 kfree(record);
2731
2732         }
2733         trace_qgroup_num_dirty_extents(fs_info, trans->transid,
2734                                        num_dirty_extents);
2735         return ret;
2736 }
2737
2738 /*
2739  * called from commit_transaction. Writes all changed qgroups to disk.
2740  */
2741 int btrfs_run_qgroups(struct btrfs_trans_handle *trans)
2742 {
2743         struct btrfs_fs_info *fs_info = trans->fs_info;
2744         int ret = 0;
2745
2746         if (!fs_info->quota_root)
2747                 return ret;
2748
2749         spin_lock(&fs_info->qgroup_lock);
2750         while (!list_empty(&fs_info->dirty_qgroups)) {
2751                 struct btrfs_qgroup *qgroup;
2752                 qgroup = list_first_entry(&fs_info->dirty_qgroups,
2753                                           struct btrfs_qgroup, dirty);
2754                 list_del_init(&qgroup->dirty);
2755                 spin_unlock(&fs_info->qgroup_lock);
2756                 ret = update_qgroup_info_item(trans, qgroup);
2757                 if (ret)
2758                         fs_info->qgroup_flags |=
2759                                         BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2760                 ret = update_qgroup_limit_item(trans, qgroup);
2761                 if (ret)
2762                         fs_info->qgroup_flags |=
2763                                         BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2764                 spin_lock(&fs_info->qgroup_lock);
2765         }
2766         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2767                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_ON;
2768         else
2769                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_ON;
2770         spin_unlock(&fs_info->qgroup_lock);
2771
2772         ret = update_qgroup_status_item(trans);
2773         if (ret)
2774                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2775
2776         return ret;
2777 }
2778
2779 /*
2780  * Copy the accounting information between qgroups. This is necessary
2781  * when a snapshot or a subvolume is created. Throwing an error will
2782  * cause a transaction abort so we take extra care here to only error
2783  * when a readonly fs is a reasonable outcome.
2784  */
2785 int btrfs_qgroup_inherit(struct btrfs_trans_handle *trans, u64 srcid,
2786                          u64 objectid, struct btrfs_qgroup_inherit *inherit)
2787 {
2788         int ret = 0;
2789         int i;
2790         u64 *i_qgroups;
2791         bool committing = false;
2792         struct btrfs_fs_info *fs_info = trans->fs_info;
2793         struct btrfs_root *quota_root;
2794         struct btrfs_qgroup *srcgroup;
2795         struct btrfs_qgroup *dstgroup;
2796         bool need_rescan = false;
2797         u32 level_size = 0;
2798         u64 nums;
2799
2800         /*
2801          * There are only two callers of this function.
2802          *
2803          * One in create_subvol() in the ioctl context, which needs to hold
2804          * the qgroup_ioctl_lock.
2805          *
2806          * The other one in create_pending_snapshot() where no other qgroup
2807          * code can modify the fs as they all need to either start a new trans
2808          * or hold a trans handler, thus we don't need to hold
2809          * qgroup_ioctl_lock.
2810          * This would avoid long and complex lock chain and make lockdep happy.
2811          */
2812         spin_lock(&fs_info->trans_lock);
2813         if (trans->transaction->state == TRANS_STATE_COMMIT_DOING)
2814                 committing = true;
2815         spin_unlock(&fs_info->trans_lock);
2816
2817         if (!committing)
2818                 mutex_lock(&fs_info->qgroup_ioctl_lock);
2819         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2820                 goto out;
2821
2822         quota_root = fs_info->quota_root;
2823         if (!quota_root) {
2824                 ret = -EINVAL;
2825                 goto out;
2826         }
2827
2828         if (inherit) {
2829                 i_qgroups = (u64 *)(inherit + 1);
2830                 nums = inherit->num_qgroups + 2 * inherit->num_ref_copies +
2831                        2 * inherit->num_excl_copies;
2832                 for (i = 0; i < nums; ++i) {
2833                         srcgroup = find_qgroup_rb(fs_info, *i_qgroups);
2834
2835                         /*
2836                          * Zero out invalid groups so we can ignore
2837                          * them later.
2838                          */
2839                         if (!srcgroup ||
2840                             ((srcgroup->qgroupid >> 48) <= (objectid >> 48)))
2841                                 *i_qgroups = 0ULL;
2842
2843                         ++i_qgroups;
2844                 }
2845         }
2846
2847         /*
2848          * create a tracking group for the subvol itself
2849          */
2850         ret = add_qgroup_item(trans, quota_root, objectid);
2851         if (ret)
2852                 goto out;
2853
2854         /*
2855          * add qgroup to all inherited groups
2856          */
2857         if (inherit) {
2858                 i_qgroups = (u64 *)(inherit + 1);
2859                 for (i = 0; i < inherit->num_qgroups; ++i, ++i_qgroups) {
2860                         if (*i_qgroups == 0)
2861                                 continue;
2862                         ret = add_qgroup_relation_item(trans, objectid,
2863                                                        *i_qgroups);
2864                         if (ret && ret != -EEXIST)
2865                                 goto out;
2866                         ret = add_qgroup_relation_item(trans, *i_qgroups,
2867                                                        objectid);
2868                         if (ret && ret != -EEXIST)
2869                                 goto out;
2870                 }
2871                 ret = 0;
2872         }
2873
2874
2875         spin_lock(&fs_info->qgroup_lock);
2876
2877         dstgroup = add_qgroup_rb(fs_info, objectid);
2878         if (IS_ERR(dstgroup)) {
2879                 ret = PTR_ERR(dstgroup);
2880                 goto unlock;
2881         }
2882
2883         if (inherit && inherit->flags & BTRFS_QGROUP_INHERIT_SET_LIMITS) {
2884                 dstgroup->lim_flags = inherit->lim.flags;
2885                 dstgroup->max_rfer = inherit->lim.max_rfer;
2886                 dstgroup->max_excl = inherit->lim.max_excl;
2887                 dstgroup->rsv_rfer = inherit->lim.rsv_rfer;
2888                 dstgroup->rsv_excl = inherit->lim.rsv_excl;
2889
2890                 ret = update_qgroup_limit_item(trans, dstgroup);
2891                 if (ret) {
2892                         fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2893                         btrfs_info(fs_info,
2894                                    "unable to update quota limit for %llu",
2895                                    dstgroup->qgroupid);
2896                         goto unlock;
2897                 }
2898         }
2899
2900         if (srcid) {
2901                 srcgroup = find_qgroup_rb(fs_info, srcid);
2902                 if (!srcgroup)
2903                         goto unlock;
2904
2905                 /*
2906                  * We call inherit after we clone the root in order to make sure
2907                  * our counts don't go crazy, so at this point the only
2908                  * difference between the two roots should be the root node.
2909                  */
2910                 level_size = fs_info->nodesize;
2911                 dstgroup->rfer = srcgroup->rfer;
2912                 dstgroup->rfer_cmpr = srcgroup->rfer_cmpr;
2913                 dstgroup->excl = level_size;
2914                 dstgroup->excl_cmpr = level_size;
2915                 srcgroup->excl = level_size;
2916                 srcgroup->excl_cmpr = level_size;
2917
2918                 /* inherit the limit info */
2919                 dstgroup->lim_flags = srcgroup->lim_flags;
2920                 dstgroup->max_rfer = srcgroup->max_rfer;
2921                 dstgroup->max_excl = srcgroup->max_excl;
2922                 dstgroup->rsv_rfer = srcgroup->rsv_rfer;
2923                 dstgroup->rsv_excl = srcgroup->rsv_excl;
2924
2925                 qgroup_dirty(fs_info, dstgroup);
2926                 qgroup_dirty(fs_info, srcgroup);
2927         }
2928
2929         if (!inherit)
2930                 goto unlock;
2931
2932         i_qgroups = (u64 *)(inherit + 1);
2933         for (i = 0; i < inherit->num_qgroups; ++i) {
2934                 if (*i_qgroups) {
2935                         ret = add_relation_rb(fs_info, objectid, *i_qgroups);
2936                         if (ret)
2937                                 goto unlock;
2938                 }
2939                 ++i_qgroups;
2940
2941                 /*
2942                  * If we're doing a snapshot, and adding the snapshot to a new
2943                  * qgroup, the numbers are guaranteed to be incorrect.
2944                  */
2945                 if (srcid)
2946                         need_rescan = true;
2947         }
2948
2949         for (i = 0; i <  inherit->num_ref_copies; ++i, i_qgroups += 2) {
2950                 struct btrfs_qgroup *src;
2951                 struct btrfs_qgroup *dst;
2952
2953                 if (!i_qgroups[0] || !i_qgroups[1])
2954                         continue;
2955
2956                 src = find_qgroup_rb(fs_info, i_qgroups[0]);
2957                 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
2958
2959                 if (!src || !dst) {
2960                         ret = -EINVAL;
2961                         goto unlock;
2962                 }
2963
2964                 dst->rfer = src->rfer - level_size;
2965                 dst->rfer_cmpr = src->rfer_cmpr - level_size;
2966
2967                 /* Manually tweaking numbers certainly needs a rescan */
2968                 need_rescan = true;
2969         }
2970         for (i = 0; i <  inherit->num_excl_copies; ++i, i_qgroups += 2) {
2971                 struct btrfs_qgroup *src;
2972                 struct btrfs_qgroup *dst;
2973
2974                 if (!i_qgroups[0] || !i_qgroups[1])
2975                         continue;
2976
2977                 src = find_qgroup_rb(fs_info, i_qgroups[0]);
2978                 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
2979
2980                 if (!src || !dst) {
2981                         ret = -EINVAL;
2982                         goto unlock;
2983                 }
2984
2985                 dst->excl = src->excl + level_size;
2986                 dst->excl_cmpr = src->excl_cmpr + level_size;
2987                 need_rescan = true;
2988         }
2989
2990 unlock:
2991         spin_unlock(&fs_info->qgroup_lock);
2992         if (!ret)
2993                 ret = btrfs_sysfs_add_one_qgroup(fs_info, dstgroup);
2994 out:
2995         if (!committing)
2996                 mutex_unlock(&fs_info->qgroup_ioctl_lock);
2997         if (need_rescan)
2998                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
2999         return ret;
3000 }
3001
3002 static bool qgroup_check_limits(const struct btrfs_qgroup *qg, u64 num_bytes)
3003 {
3004         if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_RFER) &&
3005             qgroup_rsv_total(qg) + (s64)qg->rfer + num_bytes > qg->max_rfer)
3006                 return false;
3007
3008         if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) &&
3009             qgroup_rsv_total(qg) + (s64)qg->excl + num_bytes > qg->max_excl)
3010                 return false;
3011
3012         return true;
3013 }
3014
3015 static int qgroup_reserve(struct btrfs_root *root, u64 num_bytes, bool enforce,
3016                           enum btrfs_qgroup_rsv_type type)
3017 {
3018         struct btrfs_qgroup *qgroup;
3019         struct btrfs_fs_info *fs_info = root->fs_info;
3020         u64 ref_root = root->root_key.objectid;
3021         int ret = 0;
3022         struct ulist_node *unode;
3023         struct ulist_iterator uiter;
3024
3025         if (!is_fstree(ref_root))
3026                 return 0;
3027
3028         if (num_bytes == 0)
3029                 return 0;
3030
3031         if (test_bit(BTRFS_FS_QUOTA_OVERRIDE, &fs_info->flags) &&
3032             capable(CAP_SYS_RESOURCE))
3033                 enforce = false;
3034
3035         spin_lock(&fs_info->qgroup_lock);
3036         if (!fs_info->quota_root)
3037                 goto out;
3038
3039         qgroup = find_qgroup_rb(fs_info, ref_root);
3040         if (!qgroup)
3041                 goto out;
3042
3043         /*
3044          * in a first step, we check all affected qgroups if any limits would
3045          * be exceeded
3046          */
3047         ulist_reinit(fs_info->qgroup_ulist);
3048         ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
3049                         qgroup_to_aux(qgroup), GFP_ATOMIC);
3050         if (ret < 0)
3051                 goto out;
3052         ULIST_ITER_INIT(&uiter);
3053         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
3054                 struct btrfs_qgroup *qg;
3055                 struct btrfs_qgroup_list *glist;
3056
3057                 qg = unode_aux_to_qgroup(unode);
3058
3059                 if (enforce && !qgroup_check_limits(qg, num_bytes)) {
3060                         ret = -EDQUOT;
3061                         goto out;
3062                 }
3063
3064                 list_for_each_entry(glist, &qg->groups, next_group) {
3065                         ret = ulist_add(fs_info->qgroup_ulist,
3066                                         glist->group->qgroupid,
3067                                         qgroup_to_aux(glist->group), GFP_ATOMIC);
3068                         if (ret < 0)
3069                                 goto out;
3070                 }
3071         }
3072         ret = 0;
3073         /*
3074          * no limits exceeded, now record the reservation into all qgroups
3075          */
3076         ULIST_ITER_INIT(&uiter);
3077         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
3078                 struct btrfs_qgroup *qg;
3079
3080                 qg = unode_aux_to_qgroup(unode);
3081
3082                 qgroup_rsv_add(fs_info, qg, num_bytes, type);
3083         }
3084
3085 out:
3086         spin_unlock(&fs_info->qgroup_lock);
3087         return ret;
3088 }
3089
3090 /*
3091  * Free @num_bytes of reserved space with @type for qgroup.  (Normally level 0
3092  * qgroup).
3093  *
3094  * Will handle all higher level qgroup too.
3095  *
3096  * NOTE: If @num_bytes is (u64)-1, this means to free all bytes of this qgroup.
3097  * This special case is only used for META_PERTRANS type.
3098  */
3099 void btrfs_qgroup_free_refroot(struct btrfs_fs_info *fs_info,
3100                                u64 ref_root, u64 num_bytes,
3101                                enum btrfs_qgroup_rsv_type type)
3102 {
3103         struct btrfs_qgroup *qgroup;
3104         struct ulist_node *unode;
3105         struct ulist_iterator uiter;
3106         int ret = 0;
3107
3108         if (!is_fstree(ref_root))
3109                 return;
3110
3111         if (num_bytes == 0)
3112                 return;
3113
3114         if (num_bytes == (u64)-1 && type != BTRFS_QGROUP_RSV_META_PERTRANS) {
3115                 WARN(1, "%s: Invalid type to free", __func__);
3116                 return;
3117         }
3118         spin_lock(&fs_info->qgroup_lock);
3119
3120         if (!fs_info->quota_root)
3121                 goto out;
3122
3123         qgroup = find_qgroup_rb(fs_info, ref_root);
3124         if (!qgroup)
3125                 goto out;
3126
3127         if (num_bytes == (u64)-1)
3128                 /*
3129                  * We're freeing all pertrans rsv, get reserved value from
3130                  * level 0 qgroup as real num_bytes to free.
3131                  */
3132                 num_bytes = qgroup->rsv.values[type];
3133
3134         ulist_reinit(fs_info->qgroup_ulist);
3135         ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
3136                         qgroup_to_aux(qgroup), GFP_ATOMIC);
3137         if (ret < 0)
3138                 goto out;
3139         ULIST_ITER_INIT(&uiter);
3140         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
3141                 struct btrfs_qgroup *qg;
3142                 struct btrfs_qgroup_list *glist;
3143
3144                 qg = unode_aux_to_qgroup(unode);
3145
3146                 qgroup_rsv_release(fs_info, qg, num_bytes, type);
3147
3148                 list_for_each_entry(glist, &qg->groups, next_group) {
3149                         ret = ulist_add(fs_info->qgroup_ulist,
3150                                         glist->group->qgroupid,
3151                                         qgroup_to_aux(glist->group), GFP_ATOMIC);
3152                         if (ret < 0)
3153                                 goto out;
3154                 }
3155         }
3156
3157 out:
3158         spin_unlock(&fs_info->qgroup_lock);
3159 }
3160
3161 /*
3162  * Check if the leaf is the last leaf. Which means all node pointers
3163  * are at their last position.
3164  */
3165 static bool is_last_leaf(struct btrfs_path *path)
3166 {
3167         int i;
3168
3169         for (i = 1; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
3170                 if (path->slots[i] != btrfs_header_nritems(path->nodes[i]) - 1)
3171                         return false;
3172         }
3173         return true;
3174 }
3175
3176 /*
3177  * returns < 0 on error, 0 when more leafs are to be scanned.
3178  * returns 1 when done.
3179  */
3180 static int qgroup_rescan_leaf(struct btrfs_trans_handle *trans,
3181                               struct btrfs_path *path)
3182 {
3183         struct btrfs_fs_info *fs_info = trans->fs_info;
3184         struct btrfs_key found;
3185         struct extent_buffer *scratch_leaf = NULL;
3186         struct ulist *roots = NULL;
3187         u64 num_bytes;
3188         bool done;
3189         int slot;
3190         int ret;
3191
3192         mutex_lock(&fs_info->qgroup_rescan_lock);
3193         ret = btrfs_search_slot_for_read(fs_info->extent_root,
3194                                          &fs_info->qgroup_rescan_progress,
3195                                          path, 1, 0);
3196
3197         btrfs_debug(fs_info,
3198                 "current progress key (%llu %u %llu), search_slot ret %d",
3199                 fs_info->qgroup_rescan_progress.objectid,
3200                 fs_info->qgroup_rescan_progress.type,
3201                 fs_info->qgroup_rescan_progress.offset, ret);
3202
3203         if (ret) {
3204                 /*
3205                  * The rescan is about to end, we will not be scanning any
3206                  * further blocks. We cannot unset the RESCAN flag here, because
3207                  * we want to commit the transaction if everything went well.
3208                  * To make the live accounting work in this phase, we set our
3209                  * scan progress pointer such that every real extent objectid
3210                  * will be smaller.
3211                  */
3212                 fs_info->qgroup_rescan_progress.objectid = (u64)-1;
3213                 btrfs_release_path(path);
3214                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3215                 return ret;
3216         }
3217         done = is_last_leaf(path);
3218
3219         btrfs_item_key_to_cpu(path->nodes[0], &found,
3220                               btrfs_header_nritems(path->nodes[0]) - 1);
3221         fs_info->qgroup_rescan_progress.objectid = found.objectid + 1;
3222
3223         scratch_leaf = btrfs_clone_extent_buffer(path->nodes[0]);
3224         if (!scratch_leaf) {
3225                 ret = -ENOMEM;
3226                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3227                 goto out;
3228         }
3229         slot = path->slots[0];
3230         btrfs_release_path(path);
3231         mutex_unlock(&fs_info->qgroup_rescan_lock);
3232
3233         for (; slot < btrfs_header_nritems(scratch_leaf); ++slot) {
3234                 btrfs_item_key_to_cpu(scratch_leaf, &found, slot);
3235                 if (found.type != BTRFS_EXTENT_ITEM_KEY &&
3236                     found.type != BTRFS_METADATA_ITEM_KEY)
3237                         continue;
3238                 if (found.type == BTRFS_METADATA_ITEM_KEY)
3239                         num_bytes = fs_info->nodesize;
3240                 else
3241                         num_bytes = found.offset;
3242
3243                 ret = btrfs_find_all_roots(NULL, fs_info, found.objectid, 0,
3244                                            &roots, false);
3245                 if (ret < 0)
3246                         goto out;
3247                 /* For rescan, just pass old_roots as NULL */
3248                 ret = btrfs_qgroup_account_extent(trans, found.objectid,
3249                                                   num_bytes, NULL, roots);
3250                 if (ret < 0)
3251                         goto out;
3252         }
3253 out:
3254         if (scratch_leaf)
3255                 free_extent_buffer(scratch_leaf);
3256
3257         if (done && !ret) {
3258                 ret = 1;
3259                 fs_info->qgroup_rescan_progress.objectid = (u64)-1;
3260         }
3261         return ret;
3262 }
3263
3264 static bool rescan_should_stop(struct btrfs_fs_info *fs_info)
3265 {
3266         return btrfs_fs_closing(fs_info) ||
3267                 test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state);
3268 }
3269
3270 static void btrfs_qgroup_rescan_worker(struct btrfs_work *work)
3271 {
3272         struct btrfs_fs_info *fs_info = container_of(work, struct btrfs_fs_info,
3273                                                      qgroup_rescan_work);
3274         struct btrfs_path *path;
3275         struct btrfs_trans_handle *trans = NULL;
3276         int err = -ENOMEM;
3277         int ret = 0;
3278         bool stopped = false;
3279
3280         path = btrfs_alloc_path();
3281         if (!path)
3282                 goto out;
3283         /*
3284          * Rescan should only search for commit root, and any later difference
3285          * should be recorded by qgroup
3286          */
3287         path->search_commit_root = 1;
3288         path->skip_locking = 1;
3289
3290         err = 0;
3291         while (!err && !(stopped = rescan_should_stop(fs_info))) {
3292                 trans = btrfs_start_transaction(fs_info->fs_root, 0);
3293                 if (IS_ERR(trans)) {
3294                         err = PTR_ERR(trans);
3295                         break;
3296                 }
3297                 if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
3298                         err = -EINTR;
3299                 } else {
3300                         err = qgroup_rescan_leaf(trans, path);
3301                 }
3302                 if (err > 0)
3303                         btrfs_commit_transaction(trans);
3304                 else
3305                         btrfs_end_transaction(trans);
3306         }
3307
3308 out:
3309         btrfs_free_path(path);
3310
3311         mutex_lock(&fs_info->qgroup_rescan_lock);
3312         if (err > 0 &&
3313             fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT) {
3314                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
3315         } else if (err < 0) {
3316                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
3317         }
3318         mutex_unlock(&fs_info->qgroup_rescan_lock);
3319
3320         /*
3321          * only update status, since the previous part has already updated the
3322          * qgroup info.
3323          */
3324         trans = btrfs_start_transaction(fs_info->quota_root, 1);
3325         if (IS_ERR(trans)) {
3326                 err = PTR_ERR(trans);
3327                 trans = NULL;
3328                 btrfs_err(fs_info,
3329                           "fail to start transaction for status update: %d",
3330                           err);
3331         }
3332
3333         mutex_lock(&fs_info->qgroup_rescan_lock);
3334         if (!stopped)
3335                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3336         if (trans) {
3337                 ret = update_qgroup_status_item(trans);
3338                 if (ret < 0) {
3339                         err = ret;
3340                         btrfs_err(fs_info, "fail to update qgroup status: %d",
3341                                   err);
3342                 }
3343         }
3344         fs_info->qgroup_rescan_running = false;
3345         complete_all(&fs_info->qgroup_rescan_completion);
3346         mutex_unlock(&fs_info->qgroup_rescan_lock);
3347
3348         if (!trans)
3349                 return;
3350
3351         btrfs_end_transaction(trans);
3352
3353         if (stopped) {
3354                 btrfs_info(fs_info, "qgroup scan paused");
3355         } else if (err >= 0) {
3356                 btrfs_info(fs_info, "qgroup scan completed%s",
3357                         err > 0 ? " (inconsistency flag cleared)" : "");
3358         } else {
3359                 btrfs_err(fs_info, "qgroup scan failed with %d", err);
3360         }
3361 }
3362
3363 /*
3364  * Checks that (a) no rescan is running and (b) quota is enabled. Allocates all
3365  * memory required for the rescan context.
3366  */
3367 static int
3368 qgroup_rescan_init(struct btrfs_fs_info *fs_info, u64 progress_objectid,
3369                    int init_flags)
3370 {
3371         int ret = 0;
3372
3373         if (!init_flags) {
3374                 /* we're resuming qgroup rescan at mount time */
3375                 if (!(fs_info->qgroup_flags &
3376                       BTRFS_QGROUP_STATUS_FLAG_RESCAN)) {
3377                         btrfs_warn(fs_info,
3378                         "qgroup rescan init failed, qgroup rescan is not queued");
3379                         ret = -EINVAL;
3380                 } else if (!(fs_info->qgroup_flags &
3381                              BTRFS_QGROUP_STATUS_FLAG_ON)) {
3382                         btrfs_warn(fs_info,
3383                         "qgroup rescan init failed, qgroup is not enabled");
3384                         ret = -EINVAL;
3385                 }
3386
3387                 if (ret)
3388                         return ret;
3389         }
3390
3391         mutex_lock(&fs_info->qgroup_rescan_lock);
3392
3393         if (init_flags) {
3394                 if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
3395                         btrfs_warn(fs_info,
3396                                    "qgroup rescan is already in progress");
3397                         ret = -EINPROGRESS;
3398                 } else if (!(fs_info->qgroup_flags &
3399                              BTRFS_QGROUP_STATUS_FLAG_ON)) {
3400                         btrfs_warn(fs_info,
3401                         "qgroup rescan init failed, qgroup is not enabled");
3402                         ret = -EINVAL;
3403                 } else if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
3404                         /* Quota disable is in progress */
3405                         ret = -EBUSY;
3406                 }
3407
3408                 if (ret) {
3409                         mutex_unlock(&fs_info->qgroup_rescan_lock);
3410                         return ret;
3411                 }
3412                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3413         }
3414
3415         memset(&fs_info->qgroup_rescan_progress, 0,
3416                 sizeof(fs_info->qgroup_rescan_progress));
3417         fs_info->qgroup_rescan_progress.objectid = progress_objectid;
3418         init_completion(&fs_info->qgroup_rescan_completion);
3419         mutex_unlock(&fs_info->qgroup_rescan_lock);
3420
3421         btrfs_init_work(&fs_info->qgroup_rescan_work,
3422                         btrfs_qgroup_rescan_worker, NULL, NULL);
3423         return 0;
3424 }
3425
3426 static void
3427 qgroup_rescan_zero_tracking(struct btrfs_fs_info *fs_info)
3428 {
3429         struct rb_node *n;
3430         struct btrfs_qgroup *qgroup;
3431
3432         spin_lock(&fs_info->qgroup_lock);
3433         /* clear all current qgroup tracking information */
3434         for (n = rb_first(&fs_info->qgroup_tree); n; n = rb_next(n)) {
3435                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
3436                 qgroup->rfer = 0;
3437                 qgroup->rfer_cmpr = 0;
3438                 qgroup->excl = 0;
3439                 qgroup->excl_cmpr = 0;
3440                 qgroup_dirty(fs_info, qgroup);
3441         }
3442         spin_unlock(&fs_info->qgroup_lock);
3443 }
3444
3445 int
3446 btrfs_qgroup_rescan(struct btrfs_fs_info *fs_info)
3447 {
3448         int ret = 0;
3449         struct btrfs_trans_handle *trans;
3450
3451         ret = qgroup_rescan_init(fs_info, 0, 1);
3452         if (ret)
3453                 return ret;
3454
3455         /*
3456          * We have set the rescan_progress to 0, which means no more
3457          * delayed refs will be accounted by btrfs_qgroup_account_ref.
3458          * However, btrfs_qgroup_account_ref may be right after its call
3459          * to btrfs_find_all_roots, in which case it would still do the
3460          * accounting.
3461          * To solve this, we're committing the transaction, which will
3462          * ensure we run all delayed refs and only after that, we are
3463          * going to clear all tracking information for a clean start.
3464          */
3465
3466         trans = btrfs_join_transaction(fs_info->fs_root);
3467         if (IS_ERR(trans)) {
3468                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3469                 return PTR_ERR(trans);
3470         }
3471         ret = btrfs_commit_transaction(trans);
3472         if (ret) {
3473                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3474                 return ret;
3475         }
3476
3477         qgroup_rescan_zero_tracking(fs_info);
3478
3479         mutex_lock(&fs_info->qgroup_rescan_lock);
3480         fs_info->qgroup_rescan_running = true;
3481         btrfs_queue_work(fs_info->qgroup_rescan_workers,
3482                          &fs_info->qgroup_rescan_work);
3483         mutex_unlock(&fs_info->qgroup_rescan_lock);
3484
3485         return 0;
3486 }
3487
3488 int btrfs_qgroup_wait_for_completion(struct btrfs_fs_info *fs_info,
3489                                      bool interruptible)
3490 {
3491         int running;
3492         int ret = 0;
3493
3494         mutex_lock(&fs_info->qgroup_rescan_lock);
3495         running = fs_info->qgroup_rescan_running;
3496         mutex_unlock(&fs_info->qgroup_rescan_lock);
3497
3498         if (!running)
3499                 return 0;
3500
3501         if (interruptible)
3502                 ret = wait_for_completion_interruptible(
3503                                         &fs_info->qgroup_rescan_completion);
3504         else
3505                 wait_for_completion(&fs_info->qgroup_rescan_completion);
3506
3507         return ret;
3508 }
3509
3510 /*
3511  * this is only called from open_ctree where we're still single threaded, thus
3512  * locking is omitted here.
3513  */
3514 void
3515 btrfs_qgroup_rescan_resume(struct btrfs_fs_info *fs_info)
3516 {
3517         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
3518                 mutex_lock(&fs_info->qgroup_rescan_lock);
3519                 fs_info->qgroup_rescan_running = true;
3520                 btrfs_queue_work(fs_info->qgroup_rescan_workers,
3521                                  &fs_info->qgroup_rescan_work);
3522                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3523         }
3524 }
3525
3526 #define rbtree_iterate_from_safe(node, next, start)                             \
3527        for (node = start; node && ({ next = rb_next(node); 1;}); node = next)
3528
3529 static int qgroup_unreserve_range(struct btrfs_inode *inode,
3530                                   struct extent_changeset *reserved, u64 start,
3531                                   u64 len)
3532 {
3533         struct rb_node *node;
3534         struct rb_node *next;
3535         struct ulist_node *entry;
3536         int ret = 0;
3537
3538         node = reserved->range_changed.root.rb_node;
3539         if (!node)
3540                 return 0;
3541         while (node) {
3542                 entry = rb_entry(node, struct ulist_node, rb_node);
3543                 if (entry->val < start)
3544                         node = node->rb_right;
3545                 else
3546                         node = node->rb_left;
3547         }
3548
3549         if (entry->val > start && rb_prev(&entry->rb_node))
3550                 entry = rb_entry(rb_prev(&entry->rb_node), struct ulist_node,
3551                                  rb_node);
3552
3553         rbtree_iterate_from_safe(node, next, &entry->rb_node) {
3554                 u64 entry_start;
3555                 u64 entry_end;
3556                 u64 entry_len;
3557                 int clear_ret;
3558
3559                 entry = rb_entry(node, struct ulist_node, rb_node);
3560                 entry_start = entry->val;
3561                 entry_end = entry->aux;
3562                 entry_len = entry_end - entry_start + 1;
3563
3564                 if (entry_start >= start + len)
3565                         break;
3566                 if (entry_start + entry_len <= start)
3567                         continue;
3568                 /*
3569                  * Now the entry is in [start, start + len), revert the
3570                  * EXTENT_QGROUP_RESERVED bit.
3571                  */
3572                 clear_ret = clear_extent_bits(&inode->io_tree, entry_start,
3573                                               entry_end, EXTENT_QGROUP_RESERVED);
3574                 if (!ret && clear_ret < 0)
3575                         ret = clear_ret;
3576
3577                 ulist_del(&reserved->range_changed, entry->val, entry->aux);
3578                 if (likely(reserved->bytes_changed >= entry_len)) {
3579                         reserved->bytes_changed -= entry_len;
3580                 } else {
3581                         WARN_ON(1);
3582                         reserved->bytes_changed = 0;
3583                 }
3584         }
3585
3586         return ret;
3587 }
3588
3589 /*
3590  * Try to free some space for qgroup.
3591  *
3592  * For qgroup, there are only 3 ways to free qgroup space:
3593  * - Flush nodatacow write
3594  *   Any nodatacow write will free its reserved data space at run_delalloc_range().
3595  *   In theory, we should only flush nodatacow inodes, but it's not yet
3596  *   possible, so we need to flush the whole root.
3597  *
3598  * - Wait for ordered extents
3599  *   When ordered extents are finished, their reserved metadata is finally
3600  *   converted to per_trans status, which can be freed by later commit
3601  *   transaction.
3602  *
3603  * - Commit transaction
3604  *   This would free the meta_per_trans space.
3605  *   In theory this shouldn't provide much space, but any more qgroup space
3606  *   is needed.
3607  */
3608 static int try_flush_qgroup(struct btrfs_root *root)
3609 {
3610         struct btrfs_trans_handle *trans;
3611         int ret;
3612
3613         /* Can't hold an open transaction or we run the risk of deadlocking. */
3614         ASSERT(current->journal_info == NULL);
3615         if (WARN_ON(current->journal_info))
3616                 return 0;
3617
3618         /*
3619          * We don't want to run flush again and again, so if there is a running
3620          * one, we won't try to start a new flush, but exit directly.
3621          */
3622         if (test_and_set_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state)) {
3623                 wait_event(root->qgroup_flush_wait,
3624                         !test_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state));
3625                 return 0;
3626         }
3627
3628         ret = btrfs_start_delalloc_snapshot(root, true);
3629         if (ret < 0)
3630                 goto out;
3631         btrfs_wait_ordered_extents(root, U64_MAX, 0, (u64)-1);
3632
3633         trans = btrfs_join_transaction(root);
3634         if (IS_ERR(trans)) {
3635                 ret = PTR_ERR(trans);
3636                 goto out;
3637         }
3638
3639         ret = btrfs_commit_transaction(trans);
3640 out:
3641         clear_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state);
3642         wake_up(&root->qgroup_flush_wait);
3643         return ret;
3644 }
3645
3646 static int qgroup_reserve_data(struct btrfs_inode *inode,
3647                         struct extent_changeset **reserved_ret, u64 start,
3648                         u64 len)
3649 {
3650         struct btrfs_root *root = inode->root;
3651         struct extent_changeset *reserved;
3652         bool new_reserved = false;
3653         u64 orig_reserved;
3654         u64 to_reserve;
3655         int ret;
3656
3657         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &root->fs_info->flags) ||
3658             !is_fstree(root->root_key.objectid) || len == 0)
3659                 return 0;
3660
3661         /* @reserved parameter is mandatory for qgroup */
3662         if (WARN_ON(!reserved_ret))
3663                 return -EINVAL;
3664         if (!*reserved_ret) {
3665                 new_reserved = true;
3666                 *reserved_ret = extent_changeset_alloc();
3667                 if (!*reserved_ret)
3668                         return -ENOMEM;
3669         }
3670         reserved = *reserved_ret;
3671         /* Record already reserved space */
3672         orig_reserved = reserved->bytes_changed;
3673         ret = set_record_extent_bits(&inode->io_tree, start,
3674                         start + len -1, EXTENT_QGROUP_RESERVED, reserved);
3675
3676         /* Newly reserved space */
3677         to_reserve = reserved->bytes_changed - orig_reserved;
3678         trace_btrfs_qgroup_reserve_data(&inode->vfs_inode, start, len,
3679                                         to_reserve, QGROUP_RESERVE);
3680         if (ret < 0)
3681                 goto out;
3682         ret = qgroup_reserve(root, to_reserve, true, BTRFS_QGROUP_RSV_DATA);
3683         if (ret < 0)
3684                 goto cleanup;
3685
3686         return ret;
3687
3688 cleanup:
3689         qgroup_unreserve_range(inode, reserved, start, len);
3690 out:
3691         if (new_reserved) {
3692                 extent_changeset_free(reserved);
3693                 *reserved_ret = NULL;
3694         }
3695         return ret;
3696 }
3697
3698 /*
3699  * Reserve qgroup space for range [start, start + len).
3700  *
3701  * This function will either reserve space from related qgroups or do nothing
3702  * if the range is already reserved.
3703  *
3704  * Return 0 for successful reservation
3705  * Return <0 for error (including -EQUOT)
3706  *
3707  * NOTE: This function may sleep for memory allocation, dirty page flushing and
3708  *       commit transaction. So caller should not hold any dirty page locked.
3709  */
3710 int btrfs_qgroup_reserve_data(struct btrfs_inode *inode,
3711                         struct extent_changeset **reserved_ret, u64 start,
3712                         u64 len)
3713 {
3714         int ret;
3715
3716         ret = qgroup_reserve_data(inode, reserved_ret, start, len);
3717         if (ret <= 0 && ret != -EDQUOT)
3718                 return ret;
3719
3720         ret = try_flush_qgroup(inode->root);
3721         if (ret < 0)
3722                 return ret;
3723         return qgroup_reserve_data(inode, reserved_ret, start, len);
3724 }
3725
3726 /* Free ranges specified by @reserved, normally in error path */
3727 static int qgroup_free_reserved_data(struct btrfs_inode *inode,
3728                         struct extent_changeset *reserved, u64 start, u64 len)
3729 {
3730         struct btrfs_root *root = inode->root;
3731         struct ulist_node *unode;
3732         struct ulist_iterator uiter;
3733         struct extent_changeset changeset;
3734         int freed = 0;
3735         int ret;
3736
3737         extent_changeset_init(&changeset);
3738         len = round_up(start + len, root->fs_info->sectorsize);
3739         start = round_down(start, root->fs_info->sectorsize);
3740
3741         ULIST_ITER_INIT(&uiter);
3742         while ((unode = ulist_next(&reserved->range_changed, &uiter))) {
3743                 u64 range_start = unode->val;
3744                 /* unode->aux is the inclusive end */
3745                 u64 range_len = unode->aux - range_start + 1;
3746                 u64 free_start;
3747                 u64 free_len;
3748
3749                 extent_changeset_release(&changeset);
3750
3751                 /* Only free range in range [start, start + len) */
3752                 if (range_start >= start + len ||
3753                     range_start + range_len <= start)
3754                         continue;
3755                 free_start = max(range_start, start);
3756                 free_len = min(start + len, range_start + range_len) -
3757                            free_start;
3758                 /*
3759                  * TODO: To also modify reserved->ranges_reserved to reflect
3760                  * the modification.
3761                  *
3762                  * However as long as we free qgroup reserved according to
3763                  * EXTENT_QGROUP_RESERVED, we won't double free.
3764                  * So not need to rush.
3765                  */
3766                 ret = clear_record_extent_bits(&inode->io_tree, free_start,
3767                                 free_start + free_len - 1,
3768                                 EXTENT_QGROUP_RESERVED, &changeset);
3769                 if (ret < 0)
3770                         goto out;
3771                 freed += changeset.bytes_changed;
3772         }
3773         btrfs_qgroup_free_refroot(root->fs_info, root->root_key.objectid, freed,
3774                                   BTRFS_QGROUP_RSV_DATA);
3775         ret = freed;
3776 out:
3777         extent_changeset_release(&changeset);
3778         return ret;
3779 }
3780
3781 static int __btrfs_qgroup_release_data(struct btrfs_inode *inode,
3782                         struct extent_changeset *reserved, u64 start, u64 len,
3783                         int free)
3784 {
3785         struct extent_changeset changeset;
3786         int trace_op = QGROUP_RELEASE;
3787         int ret;
3788
3789         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &inode->root->fs_info->flags))
3790                 return 0;
3791
3792         /* In release case, we shouldn't have @reserved */
3793         WARN_ON(!free && reserved);
3794         if (free && reserved)
3795                 return qgroup_free_reserved_data(inode, reserved, start, len);
3796         extent_changeset_init(&changeset);
3797         ret = clear_record_extent_bits(&inode->io_tree, start, start + len -1,
3798                                        EXTENT_QGROUP_RESERVED, &changeset);
3799         if (ret < 0)
3800                 goto out;
3801
3802         if (free)
3803                 trace_op = QGROUP_FREE;
3804         trace_btrfs_qgroup_release_data(&inode->vfs_inode, start, len,
3805                                         changeset.bytes_changed, trace_op);
3806         if (free)
3807                 btrfs_qgroup_free_refroot(inode->root->fs_info,
3808                                 inode->root->root_key.objectid,
3809                                 changeset.bytes_changed, BTRFS_QGROUP_RSV_DATA);
3810         ret = changeset.bytes_changed;
3811 out:
3812         extent_changeset_release(&changeset);
3813         return ret;
3814 }
3815
3816 /*
3817  * Free a reserved space range from io_tree and related qgroups
3818  *
3819  * Should be called when a range of pages get invalidated before reaching disk.
3820  * Or for error cleanup case.
3821  * if @reserved is given, only reserved range in [@start, @start + @len) will
3822  * be freed.
3823  *
3824  * For data written to disk, use btrfs_qgroup_release_data().
3825  *
3826  * NOTE: This function may sleep for memory allocation.
3827  */
3828 int btrfs_qgroup_free_data(struct btrfs_inode *inode,
3829                         struct extent_changeset *reserved, u64 start, u64 len)
3830 {
3831         return __btrfs_qgroup_release_data(inode, reserved, start, len, 1);
3832 }
3833
3834 /*
3835  * Release a reserved space range from io_tree only.
3836  *
3837  * Should be called when a range of pages get written to disk and corresponding
3838  * FILE_EXTENT is inserted into corresponding root.
3839  *
3840  * Since new qgroup accounting framework will only update qgroup numbers at
3841  * commit_transaction() time, its reserved space shouldn't be freed from
3842  * related qgroups.
3843  *
3844  * But we should release the range from io_tree, to allow further write to be
3845  * COWed.
3846  *
3847  * NOTE: This function may sleep for memory allocation.
3848  */
3849 int btrfs_qgroup_release_data(struct btrfs_inode *inode, u64 start, u64 len)
3850 {
3851         return __btrfs_qgroup_release_data(inode, NULL, start, len, 0);
3852 }
3853
3854 static void add_root_meta_rsv(struct btrfs_root *root, int num_bytes,
3855                               enum btrfs_qgroup_rsv_type type)
3856 {
3857         if (type != BTRFS_QGROUP_RSV_META_PREALLOC &&
3858             type != BTRFS_QGROUP_RSV_META_PERTRANS)
3859                 return;
3860         if (num_bytes == 0)
3861                 return;
3862
3863         spin_lock(&root->qgroup_meta_rsv_lock);
3864         if (type == BTRFS_QGROUP_RSV_META_PREALLOC)
3865                 root->qgroup_meta_rsv_prealloc += num_bytes;
3866         else
3867                 root->qgroup_meta_rsv_pertrans += num_bytes;
3868         spin_unlock(&root->qgroup_meta_rsv_lock);
3869 }
3870
3871 static int sub_root_meta_rsv(struct btrfs_root *root, int num_bytes,
3872                              enum btrfs_qgroup_rsv_type type)
3873 {
3874         if (type != BTRFS_QGROUP_RSV_META_PREALLOC &&
3875             type != BTRFS_QGROUP_RSV_META_PERTRANS)
3876                 return 0;
3877         if (num_bytes == 0)
3878                 return 0;
3879
3880         spin_lock(&root->qgroup_meta_rsv_lock);
3881         if (type == BTRFS_QGROUP_RSV_META_PREALLOC) {
3882                 num_bytes = min_t(u64, root->qgroup_meta_rsv_prealloc,
3883                                   num_bytes);
3884                 root->qgroup_meta_rsv_prealloc -= num_bytes;
3885         } else {
3886                 num_bytes = min_t(u64, root->qgroup_meta_rsv_pertrans,
3887                                   num_bytes);
3888                 root->qgroup_meta_rsv_pertrans -= num_bytes;
3889         }
3890         spin_unlock(&root->qgroup_meta_rsv_lock);
3891         return num_bytes;
3892 }
3893
3894 int btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
3895                               enum btrfs_qgroup_rsv_type type, bool enforce)
3896 {
3897         struct btrfs_fs_info *fs_info = root->fs_info;
3898         int ret;
3899
3900         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
3901             !is_fstree(root->root_key.objectid) || num_bytes == 0)
3902                 return 0;
3903
3904         BUG_ON(num_bytes != round_down(num_bytes, fs_info->nodesize));
3905         trace_qgroup_meta_reserve(root, (s64)num_bytes, type);
3906         ret = qgroup_reserve(root, num_bytes, enforce, type);
3907         if (ret < 0)
3908                 return ret;
3909         /*
3910          * Record what we have reserved into root.
3911          *
3912          * To avoid quota disabled->enabled underflow.
3913          * In that case, we may try to free space we haven't reserved
3914          * (since quota was disabled), so record what we reserved into root.
3915          * And ensure later release won't underflow this number.
3916          */
3917         add_root_meta_rsv(root, num_bytes, type);
3918         return ret;
3919 }
3920
3921 int __btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
3922                                 enum btrfs_qgroup_rsv_type type, bool enforce)
3923 {
3924         int ret;
3925
3926         ret = btrfs_qgroup_reserve_meta(root, num_bytes, type, enforce);
3927         if (ret <= 0 && ret != -EDQUOT)
3928                 return ret;
3929
3930         ret = try_flush_qgroup(root);
3931         if (ret < 0)
3932                 return ret;
3933         return btrfs_qgroup_reserve_meta(root, num_bytes, type, enforce);
3934 }
3935
3936 void btrfs_qgroup_free_meta_all_pertrans(struct btrfs_root *root)
3937 {
3938         struct btrfs_fs_info *fs_info = root->fs_info;
3939
3940         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
3941             !is_fstree(root->root_key.objectid))
3942                 return;
3943
3944         /* TODO: Update trace point to handle such free */
3945         trace_qgroup_meta_free_all_pertrans(root);
3946         /* Special value -1 means to free all reserved space */
3947         btrfs_qgroup_free_refroot(fs_info, root->root_key.objectid, (u64)-1,
3948                                   BTRFS_QGROUP_RSV_META_PERTRANS);
3949 }
3950
3951 void __btrfs_qgroup_free_meta(struct btrfs_root *root, int num_bytes,
3952                               enum btrfs_qgroup_rsv_type type)
3953 {
3954         struct btrfs_fs_info *fs_info = root->fs_info;
3955
3956         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
3957             !is_fstree(root->root_key.objectid))
3958                 return;
3959
3960         /*
3961          * reservation for META_PREALLOC can happen before quota is enabled,
3962          * which can lead to underflow.
3963          * Here ensure we will only free what we really have reserved.
3964          */
3965         num_bytes = sub_root_meta_rsv(root, num_bytes, type);
3966         BUG_ON(num_bytes != round_down(num_bytes, fs_info->nodesize));
3967         trace_qgroup_meta_reserve(root, -(s64)num_bytes, type);
3968         btrfs_qgroup_free_refroot(fs_info, root->root_key.objectid,
3969                                   num_bytes, type);
3970 }
3971
3972 static void qgroup_convert_meta(struct btrfs_fs_info *fs_info, u64 ref_root,
3973                                 int num_bytes)
3974 {
3975         struct btrfs_qgroup *qgroup;
3976         struct ulist_node *unode;
3977         struct ulist_iterator uiter;
3978         int ret = 0;
3979
3980         if (num_bytes == 0)
3981                 return;
3982         if (!fs_info->quota_root)
3983                 return;
3984
3985         spin_lock(&fs_info->qgroup_lock);
3986         qgroup = find_qgroup_rb(fs_info, ref_root);
3987         if (!qgroup)
3988                 goto out;
3989         ulist_reinit(fs_info->qgroup_ulist);
3990         ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
3991                        qgroup_to_aux(qgroup), GFP_ATOMIC);
3992         if (ret < 0)
3993                 goto out;
3994         ULIST_ITER_INIT(&uiter);
3995         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
3996                 struct btrfs_qgroup *qg;
3997                 struct btrfs_qgroup_list *glist;
3998
3999                 qg = unode_aux_to_qgroup(unode);
4000
4001                 qgroup_rsv_release(fs_info, qg, num_bytes,
4002                                 BTRFS_QGROUP_RSV_META_PREALLOC);
4003                 qgroup_rsv_add(fs_info, qg, num_bytes,
4004                                 BTRFS_QGROUP_RSV_META_PERTRANS);
4005                 list_for_each_entry(glist, &qg->groups, next_group) {
4006                         ret = ulist_add(fs_info->qgroup_ulist,
4007                                         glist->group->qgroupid,
4008                                         qgroup_to_aux(glist->group), GFP_ATOMIC);
4009                         if (ret < 0)
4010                                 goto out;
4011                 }
4012         }
4013 out:
4014         spin_unlock(&fs_info->qgroup_lock);
4015 }
4016
4017 void btrfs_qgroup_convert_reserved_meta(struct btrfs_root *root, int num_bytes)
4018 {
4019         struct btrfs_fs_info *fs_info = root->fs_info;
4020
4021         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
4022             !is_fstree(root->root_key.objectid))
4023                 return;
4024         /* Same as btrfs_qgroup_free_meta_prealloc() */
4025         num_bytes = sub_root_meta_rsv(root, num_bytes,
4026                                       BTRFS_QGROUP_RSV_META_PREALLOC);
4027         trace_qgroup_meta_convert(root, num_bytes);
4028         qgroup_convert_meta(fs_info, root->root_key.objectid, num_bytes);
4029 }
4030
4031 /*
4032  * Check qgroup reserved space leaking, normally at destroy inode
4033  * time
4034  */
4035 void btrfs_qgroup_check_reserved_leak(struct btrfs_inode *inode)
4036 {
4037         struct extent_changeset changeset;
4038         struct ulist_node *unode;
4039         struct ulist_iterator iter;
4040         int ret;
4041
4042         extent_changeset_init(&changeset);
4043         ret = clear_record_extent_bits(&inode->io_tree, 0, (u64)-1,
4044                         EXTENT_QGROUP_RESERVED, &changeset);
4045
4046         WARN_ON(ret < 0);
4047         if (WARN_ON(changeset.bytes_changed)) {
4048                 ULIST_ITER_INIT(&iter);
4049                 while ((unode = ulist_next(&changeset.range_changed, &iter))) {
4050                         btrfs_warn(inode->root->fs_info,
4051                 "leaking qgroup reserved space, ino: %llu, start: %llu, end: %llu",
4052                                 btrfs_ino(inode), unode->val, unode->aux);
4053                 }
4054                 btrfs_qgroup_free_refroot(inode->root->fs_info,
4055                                 inode->root->root_key.objectid,
4056                                 changeset.bytes_changed, BTRFS_QGROUP_RSV_DATA);
4057
4058         }
4059         extent_changeset_release(&changeset);
4060 }
4061
4062 void btrfs_qgroup_init_swapped_blocks(
4063         struct btrfs_qgroup_swapped_blocks *swapped_blocks)
4064 {
4065         int i;
4066
4067         spin_lock_init(&swapped_blocks->lock);
4068         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
4069                 swapped_blocks->blocks[i] = RB_ROOT;
4070         swapped_blocks->swapped = false;
4071 }
4072
4073 /*
4074  * Delete all swapped blocks record of @root.
4075  * Every record here means we skipped a full subtree scan for qgroup.
4076  *
4077  * Gets called when committing one transaction.
4078  */
4079 void btrfs_qgroup_clean_swapped_blocks(struct btrfs_root *root)
4080 {
4081         struct btrfs_qgroup_swapped_blocks *swapped_blocks;
4082         int i;
4083
4084         swapped_blocks = &root->swapped_blocks;
4085
4086         spin_lock(&swapped_blocks->lock);
4087         if (!swapped_blocks->swapped)
4088                 goto out;
4089         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
4090                 struct rb_root *cur_root = &swapped_blocks->blocks[i];
4091                 struct btrfs_qgroup_swapped_block *entry;
4092                 struct btrfs_qgroup_swapped_block *next;
4093
4094                 rbtree_postorder_for_each_entry_safe(entry, next, cur_root,
4095                                                      node)
4096                         kfree(entry);
4097                 swapped_blocks->blocks[i] = RB_ROOT;
4098         }
4099         swapped_blocks->swapped = false;
4100 out:
4101         spin_unlock(&swapped_blocks->lock);
4102 }
4103
4104 /*
4105  * Add subtree roots record into @subvol_root.
4106  *
4107  * @subvol_root:        tree root of the subvolume tree get swapped
4108  * @bg:                 block group under balance
4109  * @subvol_parent/slot: pointer to the subtree root in subvolume tree
4110  * @reloc_parent/slot:  pointer to the subtree root in reloc tree
4111  *                      BOTH POINTERS ARE BEFORE TREE SWAP
4112  * @last_snapshot:      last snapshot generation of the subvolume tree
4113  */
4114 int btrfs_qgroup_add_swapped_blocks(struct btrfs_trans_handle *trans,
4115                 struct btrfs_root *subvol_root,
4116                 struct btrfs_block_group *bg,
4117                 struct extent_buffer *subvol_parent, int subvol_slot,
4118                 struct extent_buffer *reloc_parent, int reloc_slot,
4119                 u64 last_snapshot)
4120 {
4121         struct btrfs_fs_info *fs_info = subvol_root->fs_info;
4122         struct btrfs_qgroup_swapped_blocks *blocks = &subvol_root->swapped_blocks;
4123         struct btrfs_qgroup_swapped_block *block;
4124         struct rb_node **cur;
4125         struct rb_node *parent = NULL;
4126         int level = btrfs_header_level(subvol_parent) - 1;
4127         int ret = 0;
4128
4129         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
4130                 return 0;
4131
4132         if (btrfs_node_ptr_generation(subvol_parent, subvol_slot) >
4133             btrfs_node_ptr_generation(reloc_parent, reloc_slot)) {
4134                 btrfs_err_rl(fs_info,
4135                 "%s: bad parameter order, subvol_gen=%llu reloc_gen=%llu",
4136                         __func__,
4137                         btrfs_node_ptr_generation(subvol_parent, subvol_slot),
4138                         btrfs_node_ptr_generation(reloc_parent, reloc_slot));
4139                 return -EUCLEAN;
4140         }
4141
4142         block = kmalloc(sizeof(*block), GFP_NOFS);
4143         if (!block) {
4144                 ret = -ENOMEM;
4145                 goto out;
4146         }
4147
4148         /*
4149          * @reloc_parent/slot is still before swap, while @block is going to
4150          * record the bytenr after swap, so we do the swap here.
4151          */
4152         block->subvol_bytenr = btrfs_node_blockptr(reloc_parent, reloc_slot);
4153         block->subvol_generation = btrfs_node_ptr_generation(reloc_parent,
4154                                                              reloc_slot);
4155         block->reloc_bytenr = btrfs_node_blockptr(subvol_parent, subvol_slot);
4156         block->reloc_generation = btrfs_node_ptr_generation(subvol_parent,
4157                                                             subvol_slot);
4158         block->last_snapshot = last_snapshot;
4159         block->level = level;
4160
4161         /*
4162          * If we have bg == NULL, we're called from btrfs_recover_relocation(),
4163          * no one else can modify tree blocks thus we qgroup will not change
4164          * no matter the value of trace_leaf.
4165          */
4166         if (bg && bg->flags & BTRFS_BLOCK_GROUP_DATA)
4167                 block->trace_leaf = true;
4168         else
4169                 block->trace_leaf = false;
4170         btrfs_node_key_to_cpu(reloc_parent, &block->first_key, reloc_slot);
4171
4172         /* Insert @block into @blocks */
4173         spin_lock(&blocks->lock);
4174         cur = &blocks->blocks[level].rb_node;
4175         while (*cur) {
4176                 struct btrfs_qgroup_swapped_block *entry;
4177
4178                 parent = *cur;
4179                 entry = rb_entry(parent, struct btrfs_qgroup_swapped_block,
4180                                  node);
4181
4182                 if (entry->subvol_bytenr < block->subvol_bytenr) {
4183                         cur = &(*cur)->rb_left;
4184                 } else if (entry->subvol_bytenr > block->subvol_bytenr) {
4185                         cur = &(*cur)->rb_right;
4186                 } else {
4187                         if (entry->subvol_generation !=
4188                                         block->subvol_generation ||
4189                             entry->reloc_bytenr != block->reloc_bytenr ||
4190                             entry->reloc_generation !=
4191                                         block->reloc_generation) {
4192                                 /*
4193                                  * Duplicated but mismatch entry found.
4194                                  * Shouldn't happen.
4195                                  *
4196                                  * Marking qgroup inconsistent should be enough
4197                                  * for end users.
4198                                  */
4199                                 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
4200                                 ret = -EEXIST;
4201                         }
4202                         kfree(block);
4203                         goto out_unlock;
4204                 }
4205         }
4206         rb_link_node(&block->node, parent, cur);
4207         rb_insert_color(&block->node, &blocks->blocks[level]);
4208         blocks->swapped = true;
4209 out_unlock:
4210         spin_unlock(&blocks->lock);
4211 out:
4212         if (ret < 0)
4213                 fs_info->qgroup_flags |=
4214                         BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
4215         return ret;
4216 }
4217
4218 /*
4219  * Check if the tree block is a subtree root, and if so do the needed
4220  * delayed subtree trace for qgroup.
4221  *
4222  * This is called during btrfs_cow_block().
4223  */
4224 int btrfs_qgroup_trace_subtree_after_cow(struct btrfs_trans_handle *trans,
4225                                          struct btrfs_root *root,
4226                                          struct extent_buffer *subvol_eb)
4227 {
4228         struct btrfs_fs_info *fs_info = root->fs_info;
4229         struct btrfs_qgroup_swapped_blocks *blocks = &root->swapped_blocks;
4230         struct btrfs_qgroup_swapped_block *block;
4231         struct extent_buffer *reloc_eb = NULL;
4232         struct rb_node *node;
4233         bool found = false;
4234         bool swapped = false;
4235         int level = btrfs_header_level(subvol_eb);
4236         int ret = 0;
4237         int i;
4238
4239         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
4240                 return 0;
4241         if (!is_fstree(root->root_key.objectid) || !root->reloc_root)
4242                 return 0;
4243
4244         spin_lock(&blocks->lock);
4245         if (!blocks->swapped) {
4246                 spin_unlock(&blocks->lock);
4247                 return 0;
4248         }
4249         node = blocks->blocks[level].rb_node;
4250
4251         while (node) {
4252                 block = rb_entry(node, struct btrfs_qgroup_swapped_block, node);
4253                 if (block->subvol_bytenr < subvol_eb->start) {
4254                         node = node->rb_left;
4255                 } else if (block->subvol_bytenr > subvol_eb->start) {
4256                         node = node->rb_right;
4257                 } else {
4258                         found = true;
4259                         break;
4260                 }
4261         }
4262         if (!found) {
4263                 spin_unlock(&blocks->lock);
4264                 goto out;
4265         }
4266         /* Found one, remove it from @blocks first and update blocks->swapped */
4267         rb_erase(&block->node, &blocks->blocks[level]);
4268         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
4269                 if (RB_EMPTY_ROOT(&blocks->blocks[i])) {
4270                         swapped = true;
4271                         break;
4272                 }
4273         }
4274         blocks->swapped = swapped;
4275         spin_unlock(&blocks->lock);
4276
4277         /* Read out reloc subtree root */
4278         reloc_eb = read_tree_block(fs_info, block->reloc_bytenr, 0,
4279                                    block->reloc_generation, block->level,
4280                                    &block->first_key);
4281         if (IS_ERR(reloc_eb)) {
4282                 ret = PTR_ERR(reloc_eb);
4283                 reloc_eb = NULL;
4284                 goto free_out;
4285         }
4286         if (!extent_buffer_uptodate(reloc_eb)) {
4287                 ret = -EIO;
4288                 goto free_out;
4289         }
4290
4291         ret = qgroup_trace_subtree_swap(trans, reloc_eb, subvol_eb,
4292                         block->last_snapshot, block->trace_leaf);
4293 free_out:
4294         kfree(block);
4295         free_extent_buffer(reloc_eb);
4296 out:
4297         if (ret < 0) {
4298                 btrfs_err_rl(fs_info,
4299                              "failed to account subtree at bytenr %llu: %d",
4300                              subvol_eb->start, ret);
4301                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
4302         }
4303         return ret;
4304 }
4305
4306 void btrfs_qgroup_destroy_extent_records(struct btrfs_transaction *trans)
4307 {
4308         struct btrfs_qgroup_extent_record *entry;
4309         struct btrfs_qgroup_extent_record *next;
4310         struct rb_root *root;
4311
4312         root = &trans->delayed_refs.dirty_extent_root;
4313         rbtree_postorder_for_each_entry_safe(entry, next, root, node) {
4314                 ulist_free(entry->old_roots);
4315                 kfree(entry);
4316         }
4317 }