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 #include "fs.h"
  28 #include "accessors.h"
  29 #include "extent-tree.h"
  30 #include "root-tree.h"
  31 #include "tree-checker.h"
  32
  33 /*
  34  * Helpers to access qgroup reservation
  35  *
  36  * Callers should ensure the lock context and type are valid
  37  */
  38
  39 static u64 qgroup_rsv_total(const struct btrfs_qgroup *qgroup)
  40 {
  41         u64 ret = 0;
  42         int i;
  43
  44         for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++)
  45                 ret += qgroup->rsv.values[i];
  46
  47         return ret;
  48 }
  49
  50 #ifdef CONFIG_BTRFS_DEBUG
  51 static const char *qgroup_rsv_type_str(enum btrfs_qgroup_rsv_type type)
  52 {
  53         if (type == BTRFS_QGROUP_RSV_DATA)
  54                 return "data";
  55         if (type == BTRFS_QGROUP_RSV_META_PERTRANS)
  56                 return "meta_pertrans";
  57         if (type == BTRFS_QGROUP_RSV_META_PREALLOC)
  58                 return "meta_prealloc";
  59         return NULL;
  60 }
  61 #endif
  62
  63 static void qgroup_rsv_add(struct btrfs_fs_info *fs_info,
  64                            struct btrfs_qgroup *qgroup, u64 num_bytes,
  65                            enum btrfs_qgroup_rsv_type type)
  66 {
  67         trace_qgroup_update_reserve(fs_info, qgroup, num_bytes, type);
  68         qgroup->rsv.values[type] += num_bytes;
  69 }
  70
  71 static void qgroup_rsv_release(struct btrfs_fs_info *fs_info,
  72                                struct btrfs_qgroup *qgroup, u64 num_bytes,
  73                                enum btrfs_qgroup_rsv_type type)
  74 {
  75         trace_qgroup_update_reserve(fs_info, qgroup, -(s64)num_bytes, type);
  76         if (qgroup->rsv.values[type] >= num_bytes) {
  77                 qgroup->rsv.values[type] -= num_bytes;
  78                 return;
  79         }
  80 #ifdef CONFIG_BTRFS_DEBUG
  81         WARN_RATELIMIT(1,
  82                 "qgroup %llu %s reserved space underflow, have %llu to free %llu",
  83                 qgroup->qgroupid, qgroup_rsv_type_str(type),
  84                 qgroup->rsv.values[type], num_bytes);
  85 #endif
  86         qgroup->rsv.values[type] = 0;
  87 }
  88
  89 static void qgroup_rsv_add_by_qgroup(struct btrfs_fs_info *fs_info,
  90                                      struct btrfs_qgroup *dest,
  91                                      struct btrfs_qgroup *src)
  92 {
  93         int i;
  94
  95         for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++)
  96                 qgroup_rsv_add(fs_info, dest, src->rsv.values[i], i);
  97 }
  98
  99 static void qgroup_rsv_release_by_qgroup(struct btrfs_fs_info *fs_info,
 100                                          struct btrfs_qgroup *dest,
 101                                           struct btrfs_qgroup *src)
 102 {
 103         int i;
 104
 105         for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++)
 106                 qgroup_rsv_release(fs_info, dest, src->rsv.values[i], i);
 107 }
 108
 109 static void btrfs_qgroup_update_old_refcnt(struct btrfs_qgroup *qg, u64 seq,
 110                                            int mod)
 111 {
 112         if (qg->old_refcnt < seq)
 113                 qg->old_refcnt = seq;
 114         qg->old_refcnt += mod;
 115 }
 116
 117 static void btrfs_qgroup_update_new_refcnt(struct btrfs_qgroup *qg, u64 seq,
 118                                            int mod)
 119 {
 120         if (qg->new_refcnt < seq)
 121                 qg->new_refcnt = seq;
 122         qg->new_refcnt += mod;
 123 }
 124
 125 static inline u64 btrfs_qgroup_get_old_refcnt(struct btrfs_qgroup *qg, u64 seq)
 126 {
 127         if (qg->old_refcnt < seq)
 128                 return 0;
 129         return qg->old_refcnt - seq;
 130 }
 131
 132 static inline u64 btrfs_qgroup_get_new_refcnt(struct btrfs_qgroup *qg, u64 seq)
 133 {
 134         if (qg->new_refcnt < seq)
 135                 return 0;
 136         return qg->new_refcnt - seq;
 137 }
 138
 139 /*
 140  * glue structure to represent the relations between qgroups.
 141  */
 142 struct btrfs_qgroup_list {
 143         struct list_head next_group;
 144         struct list_head next_member;
 145         struct btrfs_qgroup *group;
 146         struct btrfs_qgroup *member;
 147 };
 148
 149 static inline u64 qgroup_to_aux(struct btrfs_qgroup *qg)
 150 {
 151         return (u64)(uintptr_t)qg;
 152 }
 153
 154 static inline struct btrfs_qgroup* unode_aux_to_qgroup(struct ulist_node *n)
 155 {
 156         return (struct btrfs_qgroup *)(uintptr_t)n->aux;
 157 }
 158
 159 static int
 160 qgroup_rescan_init(struct btrfs_fs_info *fs_info, u64 progress_objectid,
 161                    int init_flags);
 162 static void qgroup_rescan_zero_tracking(struct btrfs_fs_info *fs_info);
 163
 164 /* must be called with qgroup_ioctl_lock held */
 165 static struct btrfs_qgroup *find_qgroup_rb(struct btrfs_fs_info *fs_info,
 166                                            u64 qgroupid)
 167 {
 168         struct rb_node *n = fs_info->qgroup_tree.rb_node;
 169         struct btrfs_qgroup *qgroup;
 170
 171         while (n) {
 172                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
 173                 if (qgroup->qgroupid < qgroupid)
 174                         n = n->rb_left;
 175                 else if (qgroup->qgroupid > qgroupid)
 176                         n = n->rb_right;
 177                 else
 178                         return qgroup;
 179         }
 180         return NULL;
 181 }
 182
 183 /* must be called with qgroup_lock held */
 184 static struct btrfs_qgroup *add_qgroup_rb(struct btrfs_fs_info *fs_info,
 185                                           u64 qgroupid)
 186 {
 187         struct rb_node **p = &fs_info->qgroup_tree.rb_node;
 188         struct rb_node *parent = NULL;
 189         struct btrfs_qgroup *qgroup;
 190
 191         while (*p) {
 192                 parent = *p;
 193                 qgroup = rb_entry(parent, struct btrfs_qgroup, node);
 194
 195                 if (qgroup->qgroupid < qgroupid)
 196                         p = &(*p)->rb_left;
 197                 else if (qgroup->qgroupid > qgroupid)
 198                         p = &(*p)->rb_right;
 199                 else
 200                         return qgroup;
 201         }
 202
 203         qgroup = kzalloc(sizeof(*qgroup), GFP_ATOMIC);
 204         if (!qgroup)
 205                 return ERR_PTR(-ENOMEM);
 206
 207         qgroup->qgroupid = qgroupid;
 208         INIT_LIST_HEAD(&qgroup->groups);
 209         INIT_LIST_HEAD(&qgroup->members);
 210         INIT_LIST_HEAD(&qgroup->dirty);
 211         INIT_LIST_HEAD(&qgroup->iterator);
 212
 213         rb_link_node(&qgroup->node, parent, p);
 214         rb_insert_color(&qgroup->node, &fs_info->qgroup_tree);
 215
 216         return qgroup;
 217 }
 218
 219 static void __del_qgroup_rb(struct btrfs_fs_info *fs_info,
 220                             struct btrfs_qgroup *qgroup)
 221 {
 222         struct btrfs_qgroup_list *list;
 223
 224         list_del(&qgroup->dirty);
 225         while (!list_empty(&qgroup->groups)) {
 226                 list = list_first_entry(&qgroup->groups,
 227                                         struct btrfs_qgroup_list, next_group);
 228                 list_del(&list->next_group);
 229                 list_del(&list->next_member);
 230                 kfree(list);
 231         }
 232
 233         while (!list_empty(&qgroup->members)) {
 234                 list = list_first_entry(&qgroup->members,
 235                                         struct btrfs_qgroup_list, next_member);
 236                 list_del(&list->next_group);
 237                 list_del(&list->next_member);
 238                 kfree(list);
 239         }
 240 }
 241
 242 /* must be called with qgroup_lock held */
 243 static int del_qgroup_rb(struct btrfs_fs_info *fs_info, u64 qgroupid)
 244 {
 245         struct btrfs_qgroup *qgroup = find_qgroup_rb(fs_info, qgroupid);
 246
 247         if (!qgroup)
 248                 return -ENOENT;
 249
 250         rb_erase(&qgroup->node, &fs_info->qgroup_tree);
 251         __del_qgroup_rb(fs_info, qgroup);
 252         return 0;
 253 }
 254
 255 /*
 256  * Add relation specified by two qgroups.
 257  *
 258  * Must be called with qgroup_lock held.
 259  *
 260  * Return: 0        on success
 261  *         -ENOENT  if one of the qgroups is NULL
 262  *         <0       other errors
 263  */
 264 static int __add_relation_rb(struct btrfs_qgroup *member, struct btrfs_qgroup *parent)
 265 {
 266         struct btrfs_qgroup_list *list;
 267
 268         if (!member || !parent)
 269                 return -ENOENT;
 270
 271         list = kzalloc(sizeof(*list), GFP_ATOMIC);
 272         if (!list)
 273                 return -ENOMEM;
 274
 275         list->group = parent;
 276         list->member = member;
 277         list_add_tail(&list->next_group, &member->groups);
 278         list_add_tail(&list->next_member, &parent->members);
 279
 280         return 0;
 281 }
 282
 283 /*
 284  * Add relation specified by two qgroup ids.
 285  *
 286  * Must be called with qgroup_lock held.
 287  *
 288  * Return: 0        on success
 289  *         -ENOENT  if one of the ids does not exist
 290  *         <0       other errors
 291  */
 292 static int add_relation_rb(struct btrfs_fs_info *fs_info, u64 memberid, u64 parentid)
 293 {
 294         struct btrfs_qgroup *member;
 295         struct btrfs_qgroup *parent;
 296
 297         member = find_qgroup_rb(fs_info, memberid);
 298         parent = find_qgroup_rb(fs_info, parentid);
 299
 300         return __add_relation_rb(member, parent);
 301 }
 302
 303 /* Must be called with qgroup_lock held */
 304 static int del_relation_rb(struct btrfs_fs_info *fs_info,
 305                            u64 memberid, u64 parentid)
 306 {
 307         struct btrfs_qgroup *member;
 308         struct btrfs_qgroup *parent;
 309         struct btrfs_qgroup_list *list;
 310
 311         member = find_qgroup_rb(fs_info, memberid);
 312         parent = find_qgroup_rb(fs_info, parentid);
 313         if (!member || !parent)
 314                 return -ENOENT;
 315
 316         list_for_each_entry(list, &member->groups, next_group) {
 317                 if (list->group == parent) {
 318                         list_del(&list->next_group);
 319                         list_del(&list->next_member);
 320                         kfree(list);
 321                         return 0;
 322                 }
 323         }
 324         return -ENOENT;
 325 }
 326
 327 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 328 int btrfs_verify_qgroup_counts(struct btrfs_fs_info *fs_info, u64 qgroupid,
 329                                u64 rfer, u64 excl)
 330 {
 331         struct btrfs_qgroup *qgroup;
 332
 333         qgroup = find_qgroup_rb(fs_info, qgroupid);
 334         if (!qgroup)
 335                 return -EINVAL;
 336         if (qgroup->rfer != rfer || qgroup->excl != excl)
 337                 return -EINVAL;
 338         return 0;
 339 }
 340 #endif
 341
 342 static void qgroup_mark_inconsistent(struct btrfs_fs_info *fs_info)
 343 {
 344         fs_info->qgroup_flags |= (BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT |
 345                                   BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN |
 346                                   BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING);
 347 }
 348
 349 /*
 350  * The full config is read in one go, only called from open_ctree()
 351  * It doesn't use any locking, as at this point we're still single-threaded
 352  */
 353 int btrfs_read_qgroup_config(struct btrfs_fs_info *fs_info)
 354 {
 355         struct btrfs_key key;
 356         struct btrfs_key found_key;
 357         struct btrfs_root *quota_root = fs_info->quota_root;
 358         struct btrfs_path *path = NULL;
 359         struct extent_buffer *l;
 360         int slot;
 361         int ret = 0;
 362         u64 flags = 0;
 363         u64 rescan_progress = 0;
 364
 365         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
 366                 return 0;
 367
 368         fs_info->qgroup_ulist = ulist_alloc(GFP_KERNEL);
 369         if (!fs_info->qgroup_ulist) {
 370                 ret = -ENOMEM;
 371                 goto out;
 372         }
 373
 374         path = btrfs_alloc_path();
 375         if (!path) {
 376                 ret = -ENOMEM;
 377                 goto out;
 378         }
 379
 380         ret = btrfs_sysfs_add_qgroups(fs_info);
 381         if (ret < 0)
 382                 goto out;
 383         /* default this to quota off, in case no status key is found */
 384         fs_info->qgroup_flags = 0;
 385
 386         /*
 387          * pass 1: read status, all qgroup infos and limits
 388          */
 389         key.objectid = 0;
 390         key.type = 0;
 391         key.offset = 0;
 392         ret = btrfs_search_slot_for_read(quota_root, &key, path, 1, 1);
 393         if (ret)
 394                 goto out;
 395
 396         while (1) {
 397                 struct btrfs_qgroup *qgroup;
 398
 399                 slot = path->slots[0];
 400                 l = path->nodes[0];
 401                 btrfs_item_key_to_cpu(l, &found_key, slot);
 402
 403                 if (found_key.type == BTRFS_QGROUP_STATUS_KEY) {
 404                         struct btrfs_qgroup_status_item *ptr;
 405
 406                         ptr = btrfs_item_ptr(l, slot,
 407                                              struct btrfs_qgroup_status_item);
 408
 409                         if (btrfs_qgroup_status_version(l, ptr) !=
 410                             BTRFS_QGROUP_STATUS_VERSION) {
 411                                 btrfs_err(fs_info,
 412                                  "old qgroup version, quota disabled");
 413                                 goto out;
 414                         }
 415                         if (btrfs_qgroup_status_generation(l, ptr) !=
 416                             fs_info->generation) {
 417                                 qgroup_mark_inconsistent(fs_info);
 418                                 btrfs_err(fs_info,
 419                                         "qgroup generation mismatch, marked as inconsistent");
 420                         }
 421                         fs_info->qgroup_flags = btrfs_qgroup_status_flags(l,
 422                                                                           ptr);
 423                         rescan_progress = btrfs_qgroup_status_rescan(l, ptr);
 424                         goto next1;
 425                 }
 426
 427                 if (found_key.type != BTRFS_QGROUP_INFO_KEY &&
 428                     found_key.type != BTRFS_QGROUP_LIMIT_KEY)
 429                         goto next1;
 430
 431                 qgroup = find_qgroup_rb(fs_info, found_key.offset);
 432                 if ((qgroup && found_key.type == BTRFS_QGROUP_INFO_KEY) ||
 433                     (!qgroup && found_key.type == BTRFS_QGROUP_LIMIT_KEY)) {
 434                         btrfs_err(fs_info, "inconsistent qgroup config");
 435                         qgroup_mark_inconsistent(fs_info);
 436                 }
 437                 if (!qgroup) {
 438                         qgroup = add_qgroup_rb(fs_info, found_key.offset);
 439                         if (IS_ERR(qgroup)) {
 440                                 ret = PTR_ERR(qgroup);
 441                                 goto out;
 442                         }
 443                 }
 444                 ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
 445                 if (ret < 0)
 446                         goto out;
 447
 448                 switch (found_key.type) {
 449                 case BTRFS_QGROUP_INFO_KEY: {
 450                         struct btrfs_qgroup_info_item *ptr;
 451
 452                         ptr = btrfs_item_ptr(l, slot,
 453                                              struct btrfs_qgroup_info_item);
 454                         qgroup->rfer = btrfs_qgroup_info_rfer(l, ptr);
 455                         qgroup->rfer_cmpr = btrfs_qgroup_info_rfer_cmpr(l, ptr);
 456                         qgroup->excl = btrfs_qgroup_info_excl(l, ptr);
 457                         qgroup->excl_cmpr = btrfs_qgroup_info_excl_cmpr(l, ptr);
 458                         /* generation currently unused */
 459                         break;
 460                 }
 461                 case BTRFS_QGROUP_LIMIT_KEY: {
 462                         struct btrfs_qgroup_limit_item *ptr;
 463
 464                         ptr = btrfs_item_ptr(l, slot,
 465                                              struct btrfs_qgroup_limit_item);
 466                         qgroup->lim_flags = btrfs_qgroup_limit_flags(l, ptr);
 467                         qgroup->max_rfer = btrfs_qgroup_limit_max_rfer(l, ptr);
 468                         qgroup->max_excl = btrfs_qgroup_limit_max_excl(l, ptr);
 469                         qgroup->rsv_rfer = btrfs_qgroup_limit_rsv_rfer(l, ptr);
 470                         qgroup->rsv_excl = btrfs_qgroup_limit_rsv_excl(l, ptr);
 471                         break;
 472                 }
 473                 }
 474 next1:
 475                 ret = btrfs_next_item(quota_root, path);
 476                 if (ret < 0)
 477                         goto out;
 478                 if (ret)
 479                         break;
 480         }
 481         btrfs_release_path(path);
 482
 483         /*
 484          * pass 2: read all qgroup relations
 485          */
 486         key.objectid = 0;
 487         key.type = BTRFS_QGROUP_RELATION_KEY;
 488         key.offset = 0;
 489         ret = btrfs_search_slot_for_read(quota_root, &key, path, 1, 0);
 490         if (ret)
 491                 goto out;
 492         while (1) {
 493                 slot = path->slots[0];
 494                 l = path->nodes[0];
 495                 btrfs_item_key_to_cpu(l, &found_key, slot);
 496
 497                 if (found_key.type != BTRFS_QGROUP_RELATION_KEY)
 498                         goto next2;
 499
 500                 if (found_key.objectid > found_key.offset) {
 501                         /* parent <- member, not needed to build config */
 502                         /* FIXME should we omit the key completely? */
 503                         goto next2;
 504                 }
 505
 506                 ret = add_relation_rb(fs_info, found_key.objectid,
 507                                       found_key.offset);
 508                 if (ret == -ENOENT) {
 509                         btrfs_warn(fs_info,
 510                                 "orphan qgroup relation 0x%llx->0x%llx",
 511                                 found_key.objectid, found_key.offset);
 512                         ret = 0;        /* ignore the error */
 513                 }
 514                 if (ret)
 515                         goto out;
 516 next2:
 517                 ret = btrfs_next_item(quota_root, path);
 518                 if (ret < 0)
 519                         goto out;
 520                 if (ret)
 521                         break;
 522         }
 523 out:
 524         btrfs_free_path(path);
 525         fs_info->qgroup_flags |= flags;
 526         if (!(fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_ON))
 527                 clear_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
 528         else if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN &&
 529                  ret >= 0)
 530                 ret = qgroup_rescan_init(fs_info, rescan_progress, 0);
 531
 532         if (ret < 0) {
 533                 ulist_free(fs_info->qgroup_ulist);
 534                 fs_info->qgroup_ulist = NULL;
 535                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
 536                 btrfs_sysfs_del_qgroups(fs_info);
 537         }
 538
 539         return ret < 0 ? ret : 0;
 540 }
 541
 542 /*
 543  * Called in close_ctree() when quota is still enabled.  This verifies we don't
 544  * leak some reserved space.
 545  *
 546  * Return false if no reserved space is left.
 547  * Return true if some reserved space is leaked.
 548  */
 549 bool btrfs_check_quota_leak(struct btrfs_fs_info *fs_info)
 550 {
 551         struct rb_node *node;
 552         bool ret = false;
 553
 554         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
 555                 return ret;
 556         /*
 557          * Since we're unmounting, there is no race and no need to grab qgroup
 558          * lock.  And here we don't go post-order to provide a more user
 559          * friendly sorted result.
 560          */
 561         for (node = rb_first(&fs_info->qgroup_tree); node; node = rb_next(node)) {
 562                 struct btrfs_qgroup *qgroup;
 563                 int i;
 564
 565                 qgroup = rb_entry(node, struct btrfs_qgroup, node);
 566                 for (i = 0; i < BTRFS_QGROUP_RSV_LAST; i++) {
 567                         if (qgroup->rsv.values[i]) {
 568                                 ret = true;
 569                                 btrfs_warn(fs_info,
 570                 "qgroup %hu/%llu has unreleased space, type %d rsv %llu",
 571                                    btrfs_qgroup_level(qgroup->qgroupid),
 572                                    btrfs_qgroup_subvolid(qgroup->qgroupid),
 573                                    i, qgroup->rsv.values[i]);
 574                         }
 575                 }
 576         }
 577         return ret;
 578 }
 579
 580 /*
 581  * This is called from close_ctree() or open_ctree() or btrfs_quota_disable(),
 582  * first two are in single-threaded paths.And for the third one, we have set
 583  * quota_root to be null with qgroup_lock held before, so it is safe to clean
 584  * up the in-memory structures without qgroup_lock held.
 585  */
 586 void btrfs_free_qgroup_config(struct btrfs_fs_info *fs_info)
 587 {
 588         struct rb_node *n;
 589         struct btrfs_qgroup *qgroup;
 590
 591         while ((n = rb_first(&fs_info->qgroup_tree))) {
 592                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
 593                 rb_erase(n, &fs_info->qgroup_tree);
 594                 __del_qgroup_rb(fs_info, qgroup);
 595                 btrfs_sysfs_del_one_qgroup(fs_info, qgroup);
 596                 kfree(qgroup);
 597         }
 598         /*
 599          * We call btrfs_free_qgroup_config() when unmounting
 600          * filesystem and disabling quota, so we set qgroup_ulist
 601          * to be null here to avoid double free.
 602          */
 603         ulist_free(fs_info->qgroup_ulist);
 604         fs_info->qgroup_ulist = NULL;
 605         btrfs_sysfs_del_qgroups(fs_info);
 606 }
 607
 608 static int add_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_insert_empty_item(trans, quota_root, path, &key, 0);
 625
 626         btrfs_mark_buffer_dirty(trans, path->nodes[0]);
 627
 628         btrfs_free_path(path);
 629         return ret;
 630 }
 631
 632 static int del_qgroup_relation_item(struct btrfs_trans_handle *trans, u64 src,
 633                                     u64 dst)
 634 {
 635         int ret;
 636         struct btrfs_root *quota_root = trans->fs_info->quota_root;
 637         struct btrfs_path *path;
 638         struct btrfs_key key;
 639
 640         path = btrfs_alloc_path();
 641         if (!path)
 642                 return -ENOMEM;
 643
 644         key.objectid = src;
 645         key.type = BTRFS_QGROUP_RELATION_KEY;
 646         key.offset = dst;
 647
 648         ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
 649         if (ret < 0)
 650                 goto out;
 651
 652         if (ret > 0) {
 653                 ret = -ENOENT;
 654                 goto out;
 655         }
 656
 657         ret = btrfs_del_item(trans, quota_root, path);
 658 out:
 659         btrfs_free_path(path);
 660         return ret;
 661 }
 662
 663 static int add_qgroup_item(struct btrfs_trans_handle *trans,
 664                            struct btrfs_root *quota_root, u64 qgroupid)
 665 {
 666         int ret;
 667         struct btrfs_path *path;
 668         struct btrfs_qgroup_info_item *qgroup_info;
 669         struct btrfs_qgroup_limit_item *qgroup_limit;
 670         struct extent_buffer *leaf;
 671         struct btrfs_key key;
 672
 673         if (btrfs_is_testing(quota_root->fs_info))
 674                 return 0;
 675
 676         path = btrfs_alloc_path();
 677         if (!path)
 678                 return -ENOMEM;
 679
 680         key.objectid = 0;
 681         key.type = BTRFS_QGROUP_INFO_KEY;
 682         key.offset = qgroupid;
 683
 684         /*
 685          * Avoid a transaction abort by catching -EEXIST here. In that
 686          * case, we proceed by re-initializing the existing structure
 687          * on disk.
 688          */
 689
 690         ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
 691                                       sizeof(*qgroup_info));
 692         if (ret && ret != -EEXIST)
 693                 goto out;
 694
 695         leaf = path->nodes[0];
 696         qgroup_info = btrfs_item_ptr(leaf, path->slots[0],
 697                                  struct btrfs_qgroup_info_item);
 698         btrfs_set_qgroup_info_generation(leaf, qgroup_info, trans->transid);
 699         btrfs_set_qgroup_info_rfer(leaf, qgroup_info, 0);
 700         btrfs_set_qgroup_info_rfer_cmpr(leaf, qgroup_info, 0);
 701         btrfs_set_qgroup_info_excl(leaf, qgroup_info, 0);
 702         btrfs_set_qgroup_info_excl_cmpr(leaf, qgroup_info, 0);
 703
 704         btrfs_mark_buffer_dirty(trans, leaf);
 705
 706         btrfs_release_path(path);
 707
 708         key.type = BTRFS_QGROUP_LIMIT_KEY;
 709         ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
 710                                       sizeof(*qgroup_limit));
 711         if (ret && ret != -EEXIST)
 712                 goto out;
 713
 714         leaf = path->nodes[0];
 715         qgroup_limit = btrfs_item_ptr(leaf, path->slots[0],
 716                                   struct btrfs_qgroup_limit_item);
 717         btrfs_set_qgroup_limit_flags(leaf, qgroup_limit, 0);
 718         btrfs_set_qgroup_limit_max_rfer(leaf, qgroup_limit, 0);
 719         btrfs_set_qgroup_limit_max_excl(leaf, qgroup_limit, 0);
 720         btrfs_set_qgroup_limit_rsv_rfer(leaf, qgroup_limit, 0);
 721         btrfs_set_qgroup_limit_rsv_excl(leaf, qgroup_limit, 0);
 722
 723         btrfs_mark_buffer_dirty(trans, leaf);
 724
 725         ret = 0;
 726 out:
 727         btrfs_free_path(path);
 728         return ret;
 729 }
 730
 731 static int del_qgroup_item(struct btrfs_trans_handle *trans, u64 qgroupid)
 732 {
 733         int ret;
 734         struct btrfs_root *quota_root = trans->fs_info->quota_root;
 735         struct btrfs_path *path;
 736         struct btrfs_key key;
 737
 738         path = btrfs_alloc_path();
 739         if (!path)
 740                 return -ENOMEM;
 741
 742         key.objectid = 0;
 743         key.type = BTRFS_QGROUP_INFO_KEY;
 744         key.offset = qgroupid;
 745         ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
 746         if (ret < 0)
 747                 goto out;
 748
 749         if (ret > 0) {
 750                 ret = -ENOENT;
 751                 goto out;
 752         }
 753
 754         ret = btrfs_del_item(trans, quota_root, path);
 755         if (ret)
 756                 goto out;
 757
 758         btrfs_release_path(path);
 759
 760         key.type = BTRFS_QGROUP_LIMIT_KEY;
 761         ret = btrfs_search_slot(trans, quota_root, &key, path, -1, 1);
 762         if (ret < 0)
 763                 goto out;
 764
 765         if (ret > 0) {
 766                 ret = -ENOENT;
 767                 goto out;
 768         }
 769
 770         ret = btrfs_del_item(trans, quota_root, path);
 771
 772 out:
 773         btrfs_free_path(path);
 774         return ret;
 775 }
 776
 777 static int update_qgroup_limit_item(struct btrfs_trans_handle *trans,
 778                                     struct btrfs_qgroup *qgroup)
 779 {
 780         struct btrfs_root *quota_root = trans->fs_info->quota_root;
 781         struct btrfs_path *path;
 782         struct btrfs_key key;
 783         struct extent_buffer *l;
 784         struct btrfs_qgroup_limit_item *qgroup_limit;
 785         int ret;
 786         int slot;
 787
 788         key.objectid = 0;
 789         key.type = BTRFS_QGROUP_LIMIT_KEY;
 790         key.offset = qgroup->qgroupid;
 791
 792         path = btrfs_alloc_path();
 793         if (!path)
 794                 return -ENOMEM;
 795
 796         ret = btrfs_search_slot(trans, quota_root, &key, path, 0, 1);
 797         if (ret > 0)
 798                 ret = -ENOENT;
 799
 800         if (ret)
 801                 goto out;
 802
 803         l = path->nodes[0];
 804         slot = path->slots[0];
 805         qgroup_limit = btrfs_item_ptr(l, slot, struct btrfs_qgroup_limit_item);
 806         btrfs_set_qgroup_limit_flags(l, qgroup_limit, qgroup->lim_flags);
 807         btrfs_set_qgroup_limit_max_rfer(l, qgroup_limit, qgroup->max_rfer);
 808         btrfs_set_qgroup_limit_max_excl(l, qgroup_limit, qgroup->max_excl);
 809         btrfs_set_qgroup_limit_rsv_rfer(l, qgroup_limit, qgroup->rsv_rfer);
 810         btrfs_set_qgroup_limit_rsv_excl(l, qgroup_limit, qgroup->rsv_excl);
 811
 812         btrfs_mark_buffer_dirty(trans, l);
 813
 814 out:
 815         btrfs_free_path(path);
 816         return ret;
 817 }
 818
 819 static int update_qgroup_info_item(struct btrfs_trans_handle *trans,
 820                                    struct btrfs_qgroup *qgroup)
 821 {
 822         struct btrfs_fs_info *fs_info = trans->fs_info;
 823         struct btrfs_root *quota_root = fs_info->quota_root;
 824         struct btrfs_path *path;
 825         struct btrfs_key key;
 826         struct extent_buffer *l;
 827         struct btrfs_qgroup_info_item *qgroup_info;
 828         int ret;
 829         int slot;
 830
 831         if (btrfs_is_testing(fs_info))
 832                 return 0;
 833
 834         key.objectid = 0;
 835         key.type = BTRFS_QGROUP_INFO_KEY;
 836         key.offset = qgroup->qgroupid;
 837
 838         path = btrfs_alloc_path();
 839         if (!path)
 840                 return -ENOMEM;
 841
 842         ret = btrfs_search_slot(trans, quota_root, &key, path, 0, 1);
 843         if (ret > 0)
 844                 ret = -ENOENT;
 845
 846         if (ret)
 847                 goto out;
 848
 849         l = path->nodes[0];
 850         slot = path->slots[0];
 851         qgroup_info = btrfs_item_ptr(l, slot, struct btrfs_qgroup_info_item);
 852         btrfs_set_qgroup_info_generation(l, qgroup_info, trans->transid);
 853         btrfs_set_qgroup_info_rfer(l, qgroup_info, qgroup->rfer);
 854         btrfs_set_qgroup_info_rfer_cmpr(l, qgroup_info, qgroup->rfer_cmpr);
 855         btrfs_set_qgroup_info_excl(l, qgroup_info, qgroup->excl);
 856         btrfs_set_qgroup_info_excl_cmpr(l, qgroup_info, qgroup->excl_cmpr);
 857
 858         btrfs_mark_buffer_dirty(trans, l);
 859
 860 out:
 861         btrfs_free_path(path);
 862         return ret;
 863 }
 864
 865 static int update_qgroup_status_item(struct btrfs_trans_handle *trans)
 866 {
 867         struct btrfs_fs_info *fs_info = trans->fs_info;
 868         struct btrfs_root *quota_root = fs_info->quota_root;
 869         struct btrfs_path *path;
 870         struct btrfs_key key;
 871         struct extent_buffer *l;
 872         struct btrfs_qgroup_status_item *ptr;
 873         int ret;
 874         int slot;
 875
 876         key.objectid = 0;
 877         key.type = BTRFS_QGROUP_STATUS_KEY;
 878         key.offset = 0;
 879
 880         path = btrfs_alloc_path();
 881         if (!path)
 882                 return -ENOMEM;
 883
 884         ret = btrfs_search_slot(trans, quota_root, &key, path, 0, 1);
 885         if (ret > 0)
 886                 ret = -ENOENT;
 887
 888         if (ret)
 889                 goto out;
 890
 891         l = path->nodes[0];
 892         slot = path->slots[0];
 893         ptr = btrfs_item_ptr(l, slot, struct btrfs_qgroup_status_item);
 894         btrfs_set_qgroup_status_flags(l, ptr, fs_info->qgroup_flags &
 895                                       BTRFS_QGROUP_STATUS_FLAGS_MASK);
 896         btrfs_set_qgroup_status_generation(l, ptr, trans->transid);
 897         btrfs_set_qgroup_status_rescan(l, ptr,
 898                                 fs_info->qgroup_rescan_progress.objectid);
 899
 900         btrfs_mark_buffer_dirty(trans, l);
 901
 902 out:
 903         btrfs_free_path(path);
 904         return ret;
 905 }
 906
 907 /*
 908  * called with qgroup_lock held
 909  */
 910 static int btrfs_clean_quota_tree(struct btrfs_trans_handle *trans,
 911                                   struct btrfs_root *root)
 912 {
 913         struct btrfs_path *path;
 914         struct btrfs_key key;
 915         struct extent_buffer *leaf = NULL;
 916         int ret;
 917         int nr = 0;
 918
 919         path = btrfs_alloc_path();
 920         if (!path)
 921                 return -ENOMEM;
 922
 923         key.objectid = 0;
 924         key.offset = 0;
 925         key.type = 0;
 926
 927         while (1) {
 928                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 929                 if (ret < 0)
 930                         goto out;
 931                 leaf = path->nodes[0];
 932                 nr = btrfs_header_nritems(leaf);
 933                 if (!nr)
 934                         break;
 935                 /*
 936                  * delete the leaf one by one
 937                  * since the whole tree is going
 938                  * to be deleted.
 939                  */
 940                 path->slots[0] = 0;
 941                 ret = btrfs_del_items(trans, root, path, 0, nr);
 942                 if (ret)
 943                         goto out;
 944
 945                 btrfs_release_path(path);
 946         }
 947         ret = 0;
 948 out:
 949         btrfs_free_path(path);
 950         return ret;
 951 }
 952
 953 int btrfs_quota_enable(struct btrfs_fs_info *fs_info)
 954 {
 955         struct btrfs_root *quota_root;
 956         struct btrfs_root *tree_root = fs_info->tree_root;
 957         struct btrfs_path *path = NULL;
 958         struct btrfs_qgroup_status_item *ptr;
 959         struct extent_buffer *leaf;
 960         struct btrfs_key key;
 961         struct btrfs_key found_key;
 962         struct btrfs_qgroup *qgroup = NULL;
 963         struct btrfs_trans_handle *trans = NULL;
 964         struct ulist *ulist = NULL;
 965         int ret = 0;
 966         int slot;
 967
 968         /*
 969          * We need to have subvol_sem write locked, to prevent races between
 970          * concurrent tasks trying to enable quotas, because we will unlock
 971          * and relock qgroup_ioctl_lock before setting fs_info->quota_root
 972          * and before setting BTRFS_FS_QUOTA_ENABLED.
 973          */
 974         lockdep_assert_held_write(&fs_info->subvol_sem);
 975
 976         if (btrfs_fs_incompat(fs_info, EXTENT_TREE_V2)) {
 977                 btrfs_err(fs_info,
 978                           "qgroups are currently unsupported in extent tree v2");
 979                 return -EINVAL;
 980         }
 981
 982         mutex_lock(&fs_info->qgroup_ioctl_lock);
 983         if (fs_info->quota_root)
 984                 goto out;
 985
 986         ulist = ulist_alloc(GFP_KERNEL);
 987         if (!ulist) {
 988                 ret = -ENOMEM;
 989                 goto out;
 990         }
 991
 992         ret = btrfs_sysfs_add_qgroups(fs_info);
 993         if (ret < 0)
 994                 goto out;
 995
 996         /*
 997          * Unlock qgroup_ioctl_lock before starting the transaction. This is to
 998          * avoid lock acquisition inversion problems (reported by lockdep) between
 999          * qgroup_ioctl_lock and the vfs freeze semaphores, acquired when we
1000          * start a transaction.
1001          * After we started the transaction lock qgroup_ioctl_lock again and
1002          * check if someone else created the quota root in the meanwhile. If so,
1003          * just return success and release the transaction handle.
1004          *
1005          * Also we don't need to worry about someone else calling
1006          * btrfs_sysfs_add_qgroups() after we unlock and getting an error because
1007          * that function returns 0 (success) when the sysfs entries already exist.
1008          */
1009         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1010
1011         /*
1012          * 1 for quota root item
1013          * 1 for BTRFS_QGROUP_STATUS item
1014          *
1015          * Yet we also need 2*n items for a QGROUP_INFO/QGROUP_LIMIT items
1016          * per subvolume. However those are not currently reserved since it
1017          * would be a lot of overkill.
1018          */
1019         trans = btrfs_start_transaction(tree_root, 2);
1020
1021         mutex_lock(&fs_info->qgroup_ioctl_lock);
1022         if (IS_ERR(trans)) {
1023                 ret = PTR_ERR(trans);
1024                 trans = NULL;
1025                 goto out;
1026         }
1027
1028         if (fs_info->quota_root)
1029                 goto out;
1030
1031         fs_info->qgroup_ulist = ulist;
1032         ulist = NULL;
1033
1034         /*
1035          * initially create the quota tree
1036          */
1037         quota_root = btrfs_create_tree(trans, BTRFS_QUOTA_TREE_OBJECTID);
1038         if (IS_ERR(quota_root)) {
1039                 ret =  PTR_ERR(quota_root);
1040                 btrfs_abort_transaction(trans, ret);
1041                 goto out;
1042         }
1043
1044         path = btrfs_alloc_path();
1045         if (!path) {
1046                 ret = -ENOMEM;
1047                 btrfs_abort_transaction(trans, ret);
1048                 goto out_free_root;
1049         }
1050
1051         key.objectid = 0;
1052         key.type = BTRFS_QGROUP_STATUS_KEY;
1053         key.offset = 0;
1054
1055         ret = btrfs_insert_empty_item(trans, quota_root, path, &key,
1056                                       sizeof(*ptr));
1057         if (ret) {
1058                 btrfs_abort_transaction(trans, ret);
1059                 goto out_free_path;
1060         }
1061
1062         leaf = path->nodes[0];
1063         ptr = btrfs_item_ptr(leaf, path->slots[0],
1064                                  struct btrfs_qgroup_status_item);
1065         btrfs_set_qgroup_status_generation(leaf, ptr, trans->transid);
1066         btrfs_set_qgroup_status_version(leaf, ptr, BTRFS_QGROUP_STATUS_VERSION);
1067         fs_info->qgroup_flags = BTRFS_QGROUP_STATUS_FLAG_ON |
1068                                 BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1069         btrfs_set_qgroup_status_flags(leaf, ptr, fs_info->qgroup_flags &
1070                                       BTRFS_QGROUP_STATUS_FLAGS_MASK);
1071         btrfs_set_qgroup_status_rescan(leaf, ptr, 0);
1072
1073         btrfs_mark_buffer_dirty(trans, leaf);
1074
1075         key.objectid = 0;
1076         key.type = BTRFS_ROOT_REF_KEY;
1077         key.offset = 0;
1078
1079         btrfs_release_path(path);
1080         ret = btrfs_search_slot_for_read(tree_root, &key, path, 1, 0);
1081         if (ret > 0)
1082                 goto out_add_root;
1083         if (ret < 0) {
1084                 btrfs_abort_transaction(trans, ret);
1085                 goto out_free_path;
1086         }
1087
1088         while (1) {
1089                 slot = path->slots[0];
1090                 leaf = path->nodes[0];
1091                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
1092
1093                 if (found_key.type == BTRFS_ROOT_REF_KEY) {
1094
1095                         /* Release locks on tree_root before we access quota_root */
1096                         btrfs_release_path(path);
1097
1098                         ret = add_qgroup_item(trans, quota_root,
1099                                               found_key.offset);
1100                         if (ret) {
1101                                 btrfs_abort_transaction(trans, ret);
1102                                 goto out_free_path;
1103                         }
1104
1105                         qgroup = add_qgroup_rb(fs_info, found_key.offset);
1106                         if (IS_ERR(qgroup)) {
1107                                 ret = PTR_ERR(qgroup);
1108                                 btrfs_abort_transaction(trans, ret);
1109                                 goto out_free_path;
1110                         }
1111                         ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1112                         if (ret < 0) {
1113                                 btrfs_abort_transaction(trans, ret);
1114                                 goto out_free_path;
1115                         }
1116                         ret = btrfs_search_slot_for_read(tree_root, &found_key,
1117                                                          path, 1, 0);
1118                         if (ret < 0) {
1119                                 btrfs_abort_transaction(trans, ret);
1120                                 goto out_free_path;
1121                         }
1122                         if (ret > 0) {
1123                                 /*
1124                                  * Shouldn't happen, but in case it does we
1125                                  * don't need to do the btrfs_next_item, just
1126                                  * continue.
1127                                  */
1128                                 continue;
1129                         }
1130                 }
1131                 ret = btrfs_next_item(tree_root, path);
1132                 if (ret < 0) {
1133                         btrfs_abort_transaction(trans, ret);
1134                         goto out_free_path;
1135                 }
1136                 if (ret)
1137                         break;
1138         }
1139
1140 out_add_root:
1141         btrfs_release_path(path);
1142         ret = add_qgroup_item(trans, quota_root, BTRFS_FS_TREE_OBJECTID);
1143         if (ret) {
1144                 btrfs_abort_transaction(trans, ret);
1145                 goto out_free_path;
1146         }
1147
1148         qgroup = add_qgroup_rb(fs_info, BTRFS_FS_TREE_OBJECTID);
1149         if (IS_ERR(qgroup)) {
1150                 ret = PTR_ERR(qgroup);
1151                 btrfs_abort_transaction(trans, ret);
1152                 goto out_free_path;
1153         }
1154         ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1155         if (ret < 0) {
1156                 btrfs_abort_transaction(trans, ret);
1157                 goto out_free_path;
1158         }
1159
1160         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1161         /*
1162          * Commit the transaction while not holding qgroup_ioctl_lock, to avoid
1163          * a deadlock with tasks concurrently doing other qgroup operations, such
1164          * adding/removing qgroups or adding/deleting qgroup relations for example,
1165          * because all qgroup operations first start or join a transaction and then
1166          * lock the qgroup_ioctl_lock mutex.
1167          * We are safe from a concurrent task trying to enable quotas, by calling
1168          * this function, since we are serialized by fs_info->subvol_sem.
1169          */
1170         ret = btrfs_commit_transaction(trans);
1171         trans = NULL;
1172         mutex_lock(&fs_info->qgroup_ioctl_lock);
1173         if (ret)
1174                 goto out_free_path;
1175
1176         /*
1177          * Set quota enabled flag after committing the transaction, to avoid
1178          * deadlocks on fs_info->qgroup_ioctl_lock with concurrent snapshot
1179          * creation.
1180          */
1181         spin_lock(&fs_info->qgroup_lock);
1182         fs_info->quota_root = quota_root;
1183         set_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1184         spin_unlock(&fs_info->qgroup_lock);
1185
1186         ret = qgroup_rescan_init(fs_info, 0, 1);
1187         if (!ret) {
1188                 qgroup_rescan_zero_tracking(fs_info);
1189                 fs_info->qgroup_rescan_running = true;
1190                 btrfs_queue_work(fs_info->qgroup_rescan_workers,
1191                                  &fs_info->qgroup_rescan_work);
1192         } else {
1193                 /*
1194                  * We have set both BTRFS_FS_QUOTA_ENABLED and
1195                  * BTRFS_QGROUP_STATUS_FLAG_ON, so we can only fail with
1196                  * -EINPROGRESS. That can happen because someone started the
1197                  * rescan worker by calling quota rescan ioctl before we
1198                  * attempted to initialize the rescan worker. Failure due to
1199                  * quotas disabled in the meanwhile is not possible, because
1200                  * we are holding a write lock on fs_info->subvol_sem, which
1201                  * is also acquired when disabling quotas.
1202                  * Ignore such error, and any other error would need to undo
1203                  * everything we did in the transaction we just committed.
1204                  */
1205                 ASSERT(ret == -EINPROGRESS);
1206                 ret = 0;
1207         }
1208
1209 out_free_path:
1210         btrfs_free_path(path);
1211 out_free_root:
1212         if (ret)
1213                 btrfs_put_root(quota_root);
1214 out:
1215         if (ret) {
1216                 ulist_free(fs_info->qgroup_ulist);
1217                 fs_info->qgroup_ulist = NULL;
1218                 btrfs_sysfs_del_qgroups(fs_info);
1219         }
1220         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1221         if (ret && trans)
1222                 btrfs_end_transaction(trans);
1223         else if (trans)
1224                 ret = btrfs_end_transaction(trans);
1225         ulist_free(ulist);
1226         return ret;
1227 }
1228
1229 int btrfs_quota_disable(struct btrfs_fs_info *fs_info)
1230 {
1231         struct btrfs_root *quota_root;
1232         struct btrfs_trans_handle *trans = NULL;
1233         int ret = 0;
1234
1235         /*
1236          * We need to have subvol_sem write locked to prevent races with
1237          * snapshot creation.
1238          */
1239         lockdep_assert_held_write(&fs_info->subvol_sem);
1240
1241         /*
1242          * Lock the cleaner mutex to prevent races with concurrent relocation,
1243          * because relocation may be building backrefs for blocks of the quota
1244          * root while we are deleting the root. This is like dropping fs roots
1245          * of deleted snapshots/subvolumes, we need the same protection.
1246          *
1247          * This also prevents races between concurrent tasks trying to disable
1248          * quotas, because we will unlock and relock qgroup_ioctl_lock across
1249          * BTRFS_FS_QUOTA_ENABLED changes.
1250          */
1251         mutex_lock(&fs_info->cleaner_mutex);
1252
1253         mutex_lock(&fs_info->qgroup_ioctl_lock);
1254         if (!fs_info->quota_root)
1255                 goto out;
1256
1257         /*
1258          * Unlock the qgroup_ioctl_lock mutex before waiting for the rescan worker to
1259          * complete. Otherwise we can deadlock because btrfs_remove_qgroup() needs
1260          * to lock that mutex while holding a transaction handle and the rescan
1261          * worker needs to commit a transaction.
1262          */
1263         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1264
1265         /*
1266          * Request qgroup rescan worker to complete and wait for it. This wait
1267          * must be done before transaction start for quota disable since it may
1268          * deadlock with transaction by the qgroup rescan worker.
1269          */
1270         clear_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1271         btrfs_qgroup_wait_for_completion(fs_info, false);
1272
1273         /*
1274          * 1 For the root item
1275          *
1276          * We should also reserve enough items for the quota tree deletion in
1277          * btrfs_clean_quota_tree but this is not done.
1278          *
1279          * Also, we must always start a transaction without holding the mutex
1280          * qgroup_ioctl_lock, see btrfs_quota_enable().
1281          */
1282         trans = btrfs_start_transaction(fs_info->tree_root, 1);
1283
1284         mutex_lock(&fs_info->qgroup_ioctl_lock);
1285         if (IS_ERR(trans)) {
1286                 ret = PTR_ERR(trans);
1287                 trans = NULL;
1288                 set_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags);
1289                 goto out;
1290         }
1291
1292         if (!fs_info->quota_root)
1293                 goto out;
1294
1295         spin_lock(&fs_info->qgroup_lock);
1296         quota_root = fs_info->quota_root;
1297         fs_info->quota_root = NULL;
1298         fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_ON;
1299         fs_info->qgroup_drop_subtree_thres = BTRFS_MAX_LEVEL;
1300         spin_unlock(&fs_info->qgroup_lock);
1301
1302         btrfs_free_qgroup_config(fs_info);
1303
1304         ret = btrfs_clean_quota_tree(trans, quota_root);
1305         if (ret) {
1306                 btrfs_abort_transaction(trans, ret);
1307                 goto out;
1308         }
1309
1310         ret = btrfs_del_root(trans, &quota_root->root_key);
1311         if (ret) {
1312                 btrfs_abort_transaction(trans, ret);
1313                 goto out;
1314         }
1315
1316         spin_lock(&fs_info->trans_lock);
1317         list_del(&quota_root->dirty_list);
1318         spin_unlock(&fs_info->trans_lock);
1319
1320         btrfs_tree_lock(quota_root->node);
1321         btrfs_clear_buffer_dirty(trans, quota_root->node);
1322         btrfs_tree_unlock(quota_root->node);
1323         btrfs_free_tree_block(trans, btrfs_root_id(quota_root),
1324                               quota_root->node, 0, 1);
1325
1326         btrfs_put_root(quota_root);
1327
1328 out:
1329         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1330         if (ret && trans)
1331                 btrfs_end_transaction(trans);
1332         else if (trans)
1333                 ret = btrfs_end_transaction(trans);
1334         mutex_unlock(&fs_info->cleaner_mutex);
1335
1336         return ret;
1337 }
1338
1339 static void qgroup_dirty(struct btrfs_fs_info *fs_info,
1340                          struct btrfs_qgroup *qgroup)
1341 {
1342         if (list_empty(&qgroup->dirty))
1343                 list_add(&qgroup->dirty, &fs_info->dirty_qgroups);
1344 }
1345
1346 static void qgroup_iterator_add(struct list_head *head, struct btrfs_qgroup *qgroup)
1347 {
1348         if (!list_empty(&qgroup->iterator))
1349                 return;
1350
1351         list_add_tail(&qgroup->iterator, head);
1352 }
1353
1354 static void qgroup_iterator_clean(struct list_head *head)
1355 {
1356         while (!list_empty(head)) {
1357                 struct btrfs_qgroup *qgroup;
1358
1359                 qgroup = list_first_entry(head, struct btrfs_qgroup, iterator);
1360                 list_del_init(&qgroup->iterator);
1361         }
1362 }
1363
1364 /*
1365  * The easy accounting, we're updating qgroup relationship whose child qgroup
1366  * only has exclusive extents.
1367  *
1368  * In this case, all exclusive extents will also be exclusive for parent, so
1369  * excl/rfer just get added/removed.
1370  *
1371  * So is qgroup reservation space, which should also be added/removed to
1372  * parent.
1373  * Or when child tries to release reservation space, parent will underflow its
1374  * reservation (for relationship adding case).
1375  *
1376  * Caller should hold fs_info->qgroup_lock.
1377  */
1378 static int __qgroup_excl_accounting(struct btrfs_fs_info *fs_info,
1379                                     struct ulist *tmp, u64 ref_root,
1380                                     struct btrfs_qgroup *src, int sign)
1381 {
1382         struct btrfs_qgroup *qgroup;
1383         struct btrfs_qgroup_list *glist;
1384         struct ulist_node *unode;
1385         struct ulist_iterator uiter;
1386         u64 num_bytes = src->excl;
1387         int ret = 0;
1388
1389         qgroup = find_qgroup_rb(fs_info, ref_root);
1390         if (!qgroup)
1391                 goto out;
1392
1393         qgroup->rfer += sign * num_bytes;
1394         qgroup->rfer_cmpr += sign * num_bytes;
1395
1396         WARN_ON(sign < 0 && qgroup->excl < num_bytes);
1397         qgroup->excl += sign * num_bytes;
1398         qgroup->excl_cmpr += sign * num_bytes;
1399
1400         if (sign > 0)
1401                 qgroup_rsv_add_by_qgroup(fs_info, qgroup, src);
1402         else
1403                 qgroup_rsv_release_by_qgroup(fs_info, qgroup, src);
1404
1405         qgroup_dirty(fs_info, qgroup);
1406
1407         /* Get all of the parent groups that contain this qgroup */
1408         list_for_each_entry(glist, &qgroup->groups, next_group) {
1409                 ret = ulist_add(tmp, glist->group->qgroupid,
1410                                 qgroup_to_aux(glist->group), GFP_ATOMIC);
1411                 if (ret < 0)
1412                         goto out;
1413         }
1414
1415         /* Iterate all of the parents and adjust their reference counts */
1416         ULIST_ITER_INIT(&uiter);
1417         while ((unode = ulist_next(tmp, &uiter))) {
1418                 qgroup = unode_aux_to_qgroup(unode);
1419                 qgroup->rfer += sign * num_bytes;
1420                 qgroup->rfer_cmpr += sign * num_bytes;
1421                 WARN_ON(sign < 0 && qgroup->excl < num_bytes);
1422                 qgroup->excl += sign * num_bytes;
1423                 if (sign > 0)
1424                         qgroup_rsv_add_by_qgroup(fs_info, qgroup, src);
1425                 else
1426                         qgroup_rsv_release_by_qgroup(fs_info, qgroup, src);
1427                 qgroup->excl_cmpr += sign * num_bytes;
1428                 qgroup_dirty(fs_info, qgroup);
1429
1430                 /* Add any parents of the parents */
1431                 list_for_each_entry(glist, &qgroup->groups, next_group) {
1432                         ret = ulist_add(tmp, glist->group->qgroupid,
1433                                         qgroup_to_aux(glist->group), GFP_ATOMIC);
1434                         if (ret < 0)
1435                                 goto out;
1436                 }
1437         }
1438         ret = 0;
1439 out:
1440         return ret;
1441 }
1442
1443
1444 /*
1445  * Quick path for updating qgroup with only excl refs.
1446  *
1447  * In that case, just update all parent will be enough.
1448  * Or we needs to do a full rescan.
1449  * Caller should also hold fs_info->qgroup_lock.
1450  *
1451  * Return 0 for quick update, return >0 for need to full rescan
1452  * and mark INCONSISTENT flag.
1453  * Return < 0 for other error.
1454  */
1455 static int quick_update_accounting(struct btrfs_fs_info *fs_info,
1456                                    struct ulist *tmp, u64 src, u64 dst,
1457                                    int sign)
1458 {
1459         struct btrfs_qgroup *qgroup;
1460         int ret = 1;
1461         int err = 0;
1462
1463         qgroup = find_qgroup_rb(fs_info, src);
1464         if (!qgroup)
1465                 goto out;
1466         if (qgroup->excl == qgroup->rfer) {
1467                 ret = 0;
1468                 err = __qgroup_excl_accounting(fs_info, tmp, dst,
1469                                                qgroup, sign);
1470                 if (err < 0) {
1471                         ret = err;
1472                         goto out;
1473                 }
1474         }
1475 out:
1476         if (ret)
1477                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
1478         return ret;
1479 }
1480
1481 int btrfs_add_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1482                               u64 dst)
1483 {
1484         struct btrfs_fs_info *fs_info = trans->fs_info;
1485         struct btrfs_qgroup *parent;
1486         struct btrfs_qgroup *member;
1487         struct btrfs_qgroup_list *list;
1488         struct ulist *tmp;
1489         unsigned int nofs_flag;
1490         int ret = 0;
1491
1492         /* Check the level of src and dst first */
1493         if (btrfs_qgroup_level(src) >= btrfs_qgroup_level(dst))
1494                 return -EINVAL;
1495
1496         /* We hold a transaction handle open, must do a NOFS allocation. */
1497         nofs_flag = memalloc_nofs_save();
1498         tmp = ulist_alloc(GFP_KERNEL);
1499         memalloc_nofs_restore(nofs_flag);
1500         if (!tmp)
1501                 return -ENOMEM;
1502
1503         mutex_lock(&fs_info->qgroup_ioctl_lock);
1504         if (!fs_info->quota_root) {
1505                 ret = -ENOTCONN;
1506                 goto out;
1507         }
1508         member = find_qgroup_rb(fs_info, src);
1509         parent = find_qgroup_rb(fs_info, dst);
1510         if (!member || !parent) {
1511                 ret = -EINVAL;
1512                 goto out;
1513         }
1514
1515         /* check if such qgroup relation exist firstly */
1516         list_for_each_entry(list, &member->groups, next_group) {
1517                 if (list->group == parent) {
1518                         ret = -EEXIST;
1519                         goto out;
1520                 }
1521         }
1522
1523         ret = add_qgroup_relation_item(trans, src, dst);
1524         if (ret)
1525                 goto out;
1526
1527         ret = add_qgroup_relation_item(trans, dst, src);
1528         if (ret) {
1529                 del_qgroup_relation_item(trans, src, dst);
1530                 goto out;
1531         }
1532
1533         spin_lock(&fs_info->qgroup_lock);
1534         ret = __add_relation_rb(member, parent);
1535         if (ret < 0) {
1536                 spin_unlock(&fs_info->qgroup_lock);
1537                 goto out;
1538         }
1539         ret = quick_update_accounting(fs_info, tmp, src, dst, 1);
1540         spin_unlock(&fs_info->qgroup_lock);
1541 out:
1542         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1543         ulist_free(tmp);
1544         return ret;
1545 }
1546
1547 static int __del_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1548                                  u64 dst)
1549 {
1550         struct btrfs_fs_info *fs_info = trans->fs_info;
1551         struct btrfs_qgroup *parent;
1552         struct btrfs_qgroup *member;
1553         struct btrfs_qgroup_list *list;
1554         struct ulist *tmp;
1555         bool found = false;
1556         unsigned int nofs_flag;
1557         int ret = 0;
1558         int ret2;
1559
1560         /* We hold a transaction handle open, must do a NOFS allocation. */
1561         nofs_flag = memalloc_nofs_save();
1562         tmp = ulist_alloc(GFP_KERNEL);
1563         memalloc_nofs_restore(nofs_flag);
1564         if (!tmp)
1565                 return -ENOMEM;
1566
1567         if (!fs_info->quota_root) {
1568                 ret = -ENOTCONN;
1569                 goto out;
1570         }
1571
1572         member = find_qgroup_rb(fs_info, src);
1573         parent = find_qgroup_rb(fs_info, dst);
1574         /*
1575          * The parent/member pair doesn't exist, then try to delete the dead
1576          * relation items only.
1577          */
1578         if (!member || !parent)
1579                 goto delete_item;
1580
1581         /* check if such qgroup relation exist firstly */
1582         list_for_each_entry(list, &member->groups, next_group) {
1583                 if (list->group == parent) {
1584                         found = true;
1585                         break;
1586                 }
1587         }
1588
1589 delete_item:
1590         ret = del_qgroup_relation_item(trans, src, dst);
1591         if (ret < 0 && ret != -ENOENT)
1592                 goto out;
1593         ret2 = del_qgroup_relation_item(trans, dst, src);
1594         if (ret2 < 0 && ret2 != -ENOENT)
1595                 goto out;
1596
1597         /* At least one deletion succeeded, return 0 */
1598         if (!ret || !ret2)
1599                 ret = 0;
1600
1601         if (found) {
1602                 spin_lock(&fs_info->qgroup_lock);
1603                 del_relation_rb(fs_info, src, dst);
1604                 ret = quick_update_accounting(fs_info, tmp, src, dst, -1);
1605                 spin_unlock(&fs_info->qgroup_lock);
1606         }
1607 out:
1608         ulist_free(tmp);
1609         return ret;
1610 }
1611
1612 int btrfs_del_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
1613                               u64 dst)
1614 {
1615         struct btrfs_fs_info *fs_info = trans->fs_info;
1616         int ret = 0;
1617
1618         mutex_lock(&fs_info->qgroup_ioctl_lock);
1619         ret = __del_qgroup_relation(trans, src, dst);
1620         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1621
1622         return ret;
1623 }
1624
1625 int btrfs_create_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid)
1626 {
1627         struct btrfs_fs_info *fs_info = trans->fs_info;
1628         struct btrfs_root *quota_root;
1629         struct btrfs_qgroup *qgroup;
1630         int ret = 0;
1631
1632         mutex_lock(&fs_info->qgroup_ioctl_lock);
1633         if (!fs_info->quota_root) {
1634                 ret = -ENOTCONN;
1635                 goto out;
1636         }
1637         quota_root = fs_info->quota_root;
1638         qgroup = find_qgroup_rb(fs_info, qgroupid);
1639         if (qgroup) {
1640                 ret = -EEXIST;
1641                 goto out;
1642         }
1643
1644         ret = add_qgroup_item(trans, quota_root, qgroupid);
1645         if (ret)
1646                 goto out;
1647
1648         spin_lock(&fs_info->qgroup_lock);
1649         qgroup = add_qgroup_rb(fs_info, qgroupid);
1650         spin_unlock(&fs_info->qgroup_lock);
1651
1652         if (IS_ERR(qgroup)) {
1653                 ret = PTR_ERR(qgroup);
1654                 goto out;
1655         }
1656         ret = btrfs_sysfs_add_one_qgroup(fs_info, qgroup);
1657 out:
1658         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1659         return ret;
1660 }
1661
1662 int btrfs_remove_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid)
1663 {
1664         struct btrfs_fs_info *fs_info = trans->fs_info;
1665         struct btrfs_qgroup *qgroup;
1666         struct btrfs_qgroup_list *list;
1667         int ret = 0;
1668
1669         mutex_lock(&fs_info->qgroup_ioctl_lock);
1670         if (!fs_info->quota_root) {
1671                 ret = -ENOTCONN;
1672                 goto out;
1673         }
1674
1675         qgroup = find_qgroup_rb(fs_info, qgroupid);
1676         if (!qgroup) {
1677                 ret = -ENOENT;
1678                 goto out;
1679         }
1680
1681         /* Check if there are no children of this qgroup */
1682         if (!list_empty(&qgroup->members)) {
1683                 ret = -EBUSY;
1684                 goto out;
1685         }
1686
1687         ret = del_qgroup_item(trans, qgroupid);
1688         if (ret && ret != -ENOENT)
1689                 goto out;
1690
1691         while (!list_empty(&qgroup->groups)) {
1692                 list = list_first_entry(&qgroup->groups,
1693                                         struct btrfs_qgroup_list, next_group);
1694                 ret = __del_qgroup_relation(trans, qgroupid,
1695                                             list->group->qgroupid);
1696                 if (ret)
1697                         goto out;
1698         }
1699
1700         spin_lock(&fs_info->qgroup_lock);
1701         del_qgroup_rb(fs_info, qgroupid);
1702         spin_unlock(&fs_info->qgroup_lock);
1703
1704         /*
1705          * Remove the qgroup from sysfs now without holding the qgroup_lock
1706          * spinlock, since the sysfs_remove_group() function needs to take
1707          * the mutex kernfs_mutex through kernfs_remove_by_name_ns().
1708          */
1709         btrfs_sysfs_del_one_qgroup(fs_info, qgroup);
1710         kfree(qgroup);
1711 out:
1712         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1713         return ret;
1714 }
1715
1716 int btrfs_limit_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid,
1717                        struct btrfs_qgroup_limit *limit)
1718 {
1719         struct btrfs_fs_info *fs_info = trans->fs_info;
1720         struct btrfs_qgroup *qgroup;
1721         int ret = 0;
1722         /* Sometimes we would want to clear the limit on this qgroup.
1723          * To meet this requirement, we treat the -1 as a special value
1724          * which tell kernel to clear the limit on this qgroup.
1725          */
1726         const u64 CLEAR_VALUE = -1;
1727
1728         mutex_lock(&fs_info->qgroup_ioctl_lock);
1729         if (!fs_info->quota_root) {
1730                 ret = -ENOTCONN;
1731                 goto out;
1732         }
1733
1734         qgroup = find_qgroup_rb(fs_info, qgroupid);
1735         if (!qgroup) {
1736                 ret = -ENOENT;
1737                 goto out;
1738         }
1739
1740         spin_lock(&fs_info->qgroup_lock);
1741         if (limit->flags & BTRFS_QGROUP_LIMIT_MAX_RFER) {
1742                 if (limit->max_rfer == CLEAR_VALUE) {
1743                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_MAX_RFER;
1744                         limit->flags &= ~BTRFS_QGROUP_LIMIT_MAX_RFER;
1745                         qgroup->max_rfer = 0;
1746                 } else {
1747                         qgroup->max_rfer = limit->max_rfer;
1748                 }
1749         }
1750         if (limit->flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) {
1751                 if (limit->max_excl == CLEAR_VALUE) {
1752                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_MAX_EXCL;
1753                         limit->flags &= ~BTRFS_QGROUP_LIMIT_MAX_EXCL;
1754                         qgroup->max_excl = 0;
1755                 } else {
1756                         qgroup->max_excl = limit->max_excl;
1757                 }
1758         }
1759         if (limit->flags & BTRFS_QGROUP_LIMIT_RSV_RFER) {
1760                 if (limit->rsv_rfer == CLEAR_VALUE) {
1761                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_RSV_RFER;
1762                         limit->flags &= ~BTRFS_QGROUP_LIMIT_RSV_RFER;
1763                         qgroup->rsv_rfer = 0;
1764                 } else {
1765                         qgroup->rsv_rfer = limit->rsv_rfer;
1766                 }
1767         }
1768         if (limit->flags & BTRFS_QGROUP_LIMIT_RSV_EXCL) {
1769                 if (limit->rsv_excl == CLEAR_VALUE) {
1770                         qgroup->lim_flags &= ~BTRFS_QGROUP_LIMIT_RSV_EXCL;
1771                         limit->flags &= ~BTRFS_QGROUP_LIMIT_RSV_EXCL;
1772                         qgroup->rsv_excl = 0;
1773                 } else {
1774                         qgroup->rsv_excl = limit->rsv_excl;
1775                 }
1776         }
1777         qgroup->lim_flags |= limit->flags;
1778
1779         spin_unlock(&fs_info->qgroup_lock);
1780
1781         ret = update_qgroup_limit_item(trans, qgroup);
1782         if (ret) {
1783                 qgroup_mark_inconsistent(fs_info);
1784                 btrfs_info(fs_info, "unable to update quota limit for %llu",
1785                        qgroupid);
1786         }
1787
1788 out:
1789         mutex_unlock(&fs_info->qgroup_ioctl_lock);
1790         return ret;
1791 }
1792
1793 int btrfs_qgroup_trace_extent_nolock(struct btrfs_fs_info *fs_info,
1794                                 struct btrfs_delayed_ref_root *delayed_refs,
1795                                 struct btrfs_qgroup_extent_record *record)
1796 {
1797         struct rb_node **p = &delayed_refs->dirty_extent_root.rb_node;
1798         struct rb_node *parent_node = NULL;
1799         struct btrfs_qgroup_extent_record *entry;
1800         u64 bytenr = record->bytenr;
1801
1802         lockdep_assert_held(&delayed_refs->lock);
1803         trace_btrfs_qgroup_trace_extent(fs_info, record);
1804
1805         while (*p) {
1806                 parent_node = *p;
1807                 entry = rb_entry(parent_node, struct btrfs_qgroup_extent_record,
1808                                  node);
1809                 if (bytenr < entry->bytenr) {
1810                         p = &(*p)->rb_left;
1811                 } else if (bytenr > entry->bytenr) {
1812                         p = &(*p)->rb_right;
1813                 } else {
1814                         if (record->data_rsv && !entry->data_rsv) {
1815                                 entry->data_rsv = record->data_rsv;
1816                                 entry->data_rsv_refroot =
1817                                         record->data_rsv_refroot;
1818                         }
1819                         return 1;
1820                 }
1821         }
1822
1823         rb_link_node(&record->node, parent_node, p);
1824         rb_insert_color(&record->node, &delayed_refs->dirty_extent_root);
1825         return 0;
1826 }
1827
1828 int btrfs_qgroup_trace_extent_post(struct btrfs_trans_handle *trans,
1829                                    struct btrfs_qgroup_extent_record *qrecord)
1830 {
1831         struct btrfs_backref_walk_ctx ctx = { 0 };
1832         int ret;
1833
1834         /*
1835          * We are always called in a context where we are already holding a
1836          * transaction handle. Often we are called when adding a data delayed
1837          * reference from btrfs_truncate_inode_items() (truncating or unlinking),
1838          * in which case we will be holding a write lock on extent buffer from a
1839          * subvolume tree. In this case we can't allow btrfs_find_all_roots() to
1840          * acquire fs_info->commit_root_sem, because that is a higher level lock
1841          * that must be acquired before locking any extent buffers.
1842          *
1843          * So we want btrfs_find_all_roots() to not acquire the commit_root_sem
1844          * but we can't pass it a non-NULL transaction handle, because otherwise
1845          * it would not use commit roots and would lock extent buffers, causing
1846          * a deadlock if it ends up trying to read lock the same extent buffer
1847          * that was previously write locked at btrfs_truncate_inode_items().
1848          *
1849          * So pass a NULL transaction handle to btrfs_find_all_roots() and
1850          * explicitly tell it to not acquire the commit_root_sem - if we are
1851          * holding a transaction handle we don't need its protection.
1852          */
1853         ASSERT(trans != NULL);
1854
1855         if (trans->fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING)
1856                 return 0;
1857
1858         ctx.bytenr = qrecord->bytenr;
1859         ctx.fs_info = trans->fs_info;
1860
1861         ret = btrfs_find_all_roots(&ctx, true);
1862         if (ret < 0) {
1863                 qgroup_mark_inconsistent(trans->fs_info);
1864                 btrfs_warn(trans->fs_info,
1865 "error accounting new delayed refs extent (err code: %d), quota inconsistent",
1866                         ret);
1867                 return 0;
1868         }
1869
1870         /*
1871          * Here we don't need to get the lock of
1872          * trans->transaction->delayed_refs, since inserted qrecord won't
1873          * be deleted, only qrecord->node may be modified (new qrecord insert)
1874          *
1875          * So modifying qrecord->old_roots is safe here
1876          */
1877         qrecord->old_roots = ctx.roots;
1878         return 0;
1879 }
1880
1881 int btrfs_qgroup_trace_extent(struct btrfs_trans_handle *trans, u64 bytenr,
1882                               u64 num_bytes)
1883 {
1884         struct btrfs_fs_info *fs_info = trans->fs_info;
1885         struct btrfs_qgroup_extent_record *record;
1886         struct btrfs_delayed_ref_root *delayed_refs;
1887         int ret;
1888
1889         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)
1890             || bytenr == 0 || num_bytes == 0)
1891                 return 0;
1892         record = kzalloc(sizeof(*record), GFP_NOFS);
1893         if (!record)
1894                 return -ENOMEM;
1895
1896         delayed_refs = &trans->transaction->delayed_refs;
1897         record->bytenr = bytenr;
1898         record->num_bytes = num_bytes;
1899         record->old_roots = NULL;
1900
1901         spin_lock(&delayed_refs->lock);
1902         ret = btrfs_qgroup_trace_extent_nolock(fs_info, delayed_refs, record);
1903         spin_unlock(&delayed_refs->lock);
1904         if (ret > 0) {
1905                 kfree(record);
1906                 return 0;
1907         }
1908         return btrfs_qgroup_trace_extent_post(trans, record);
1909 }
1910
1911 int btrfs_qgroup_trace_leaf_items(struct btrfs_trans_handle *trans,
1912                                   struct extent_buffer *eb)
1913 {
1914         struct btrfs_fs_info *fs_info = trans->fs_info;
1915         int nr = btrfs_header_nritems(eb);
1916         int i, extent_type, ret;
1917         struct btrfs_key key;
1918         struct btrfs_file_extent_item *fi;
1919         u64 bytenr, num_bytes;
1920
1921         /* We can be called directly from walk_up_proc() */
1922         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
1923                 return 0;
1924
1925         for (i = 0; i < nr; i++) {
1926                 btrfs_item_key_to_cpu(eb, &key, i);
1927
1928                 if (key.type != BTRFS_EXTENT_DATA_KEY)
1929                         continue;
1930
1931                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
1932                 /* filter out non qgroup-accountable extents  */
1933                 extent_type = btrfs_file_extent_type(eb, fi);
1934
1935                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
1936                         continue;
1937
1938                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
1939                 if (!bytenr)
1940                         continue;
1941
1942                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
1943
1944                 ret = btrfs_qgroup_trace_extent(trans, bytenr, num_bytes);
1945                 if (ret)
1946                         return ret;
1947         }
1948         cond_resched();
1949         return 0;
1950 }
1951
1952 /*
1953  * Walk up the tree from the bottom, freeing leaves and any interior
1954  * nodes which have had all slots visited. If a node (leaf or
1955  * interior) is freed, the node above it will have it's slot
1956  * incremented. The root node will never be freed.
1957  *
1958  * At the end of this function, we should have a path which has all
1959  * slots incremented to the next position for a search. If we need to
1960  * read a new node it will be NULL and the node above it will have the
1961  * correct slot selected for a later read.
1962  *
1963  * If we increment the root nodes slot counter past the number of
1964  * elements, 1 is returned to signal completion of the search.
1965  */
1966 static int adjust_slots_upwards(struct btrfs_path *path, int root_level)
1967 {
1968         int level = 0;
1969         int nr, slot;
1970         struct extent_buffer *eb;
1971
1972         if (root_level == 0)
1973                 return 1;
1974
1975         while (level <= root_level) {
1976                 eb = path->nodes[level];
1977                 nr = btrfs_header_nritems(eb);
1978                 path->slots[level]++;
1979                 slot = path->slots[level];
1980                 if (slot >= nr || level == 0) {
1981                         /*
1982                          * Don't free the root -  we will detect this
1983                          * condition after our loop and return a
1984                          * positive value for caller to stop walking the tree.
1985                          */
1986                         if (level != root_level) {
1987                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
1988                                 path->locks[level] = 0;
1989
1990                                 free_extent_buffer(eb);
1991                                 path->nodes[level] = NULL;
1992                                 path->slots[level] = 0;
1993                         }
1994                 } else {
1995                         /*
1996                          * We have a valid slot to walk back down
1997                          * from. Stop here so caller can process these
1998                          * new nodes.
1999                          */
2000                         break;
2001                 }
2002
2003                 level++;
2004         }
2005
2006         eb = path->nodes[root_level];
2007         if (path->slots[root_level] >= btrfs_header_nritems(eb))
2008                 return 1;
2009
2010         return 0;
2011 }
2012
2013 /*
2014  * Helper function to trace a subtree tree block swap.
2015  *
2016  * The swap will happen in highest tree block, but there may be a lot of
2017  * tree blocks involved.
2018  *
2019  * For example:
2020  *  OO = Old tree blocks
2021  *  NN = New tree blocks allocated during balance
2022  *
2023  *           File tree (257)                  Reloc tree for 257
2024  * L2              OO                                NN
2025  *               /    \                            /    \
2026  * L1          OO      OO (a)                    OO      NN (a)
2027  *            / \     / \                       / \     / \
2028  * L0       OO   OO OO   OO                   OO   OO NN   NN
2029  *                  (b)  (c)                          (b)  (c)
2030  *
2031  * When calling qgroup_trace_extent_swap(), we will pass:
2032  * @src_eb = OO(a)
2033  * @dst_path = [ nodes[1] = NN(a), nodes[0] = NN(c) ]
2034  * @dst_level = 0
2035  * @root_level = 1
2036  *
2037  * In that case, qgroup_trace_extent_swap() will search from OO(a) to
2038  * reach OO(c), then mark both OO(c) and NN(c) as qgroup dirty.
2039  *
2040  * The main work of qgroup_trace_extent_swap() can be split into 3 parts:
2041  *
2042  * 1) Tree search from @src_eb
2043  *    It should acts as a simplified btrfs_search_slot().
2044  *    The key for search can be extracted from @dst_path->nodes[dst_level]
2045  *    (first key).
2046  *
2047  * 2) Mark the final tree blocks in @src_path and @dst_path qgroup dirty
2048  *    NOTE: In above case, OO(a) and NN(a) won't be marked qgroup dirty.
2049  *    They should be marked during previous (@dst_level = 1) iteration.
2050  *
2051  * 3) Mark file extents in leaves dirty
2052  *    We don't have good way to pick out new file extents only.
2053  *    So we still follow the old method by scanning all file extents in
2054  *    the leave.
2055  *
2056  * This function can free us from keeping two paths, thus later we only need
2057  * to care about how to iterate all new tree blocks in reloc tree.
2058  */
2059 static int qgroup_trace_extent_swap(struct btrfs_trans_handle* trans,
2060                                     struct extent_buffer *src_eb,
2061                                     struct btrfs_path *dst_path,
2062                                     int dst_level, int root_level,
2063                                     bool trace_leaf)
2064 {
2065         struct btrfs_key key;
2066         struct btrfs_path *src_path;
2067         struct btrfs_fs_info *fs_info = trans->fs_info;
2068         u32 nodesize = fs_info->nodesize;
2069         int cur_level = root_level;
2070         int ret;
2071
2072         BUG_ON(dst_level > root_level);
2073         /* Level mismatch */
2074         if (btrfs_header_level(src_eb) != root_level)
2075                 return -EINVAL;
2076
2077         src_path = btrfs_alloc_path();
2078         if (!src_path) {
2079                 ret = -ENOMEM;
2080                 goto out;
2081         }
2082
2083         if (dst_level)
2084                 btrfs_node_key_to_cpu(dst_path->nodes[dst_level], &key, 0);
2085         else
2086                 btrfs_item_key_to_cpu(dst_path->nodes[dst_level], &key, 0);
2087
2088         /* For src_path */
2089         atomic_inc(&src_eb->refs);
2090         src_path->nodes[root_level] = src_eb;
2091         src_path->slots[root_level] = dst_path->slots[root_level];
2092         src_path->locks[root_level] = 0;
2093
2094         /* A simplified version of btrfs_search_slot() */
2095         while (cur_level >= dst_level) {
2096                 struct btrfs_key src_key;
2097                 struct btrfs_key dst_key;
2098
2099                 if (src_path->nodes[cur_level] == NULL) {
2100                         struct extent_buffer *eb;
2101                         int parent_slot;
2102
2103                         eb = src_path->nodes[cur_level + 1];
2104                         parent_slot = src_path->slots[cur_level + 1];
2105
2106                         eb = btrfs_read_node_slot(eb, parent_slot);
2107                         if (IS_ERR(eb)) {
2108                                 ret = PTR_ERR(eb);
2109                                 goto out;
2110                         }
2111
2112                         src_path->nodes[cur_level] = eb;
2113
2114                         btrfs_tree_read_lock(eb);
2115                         src_path->locks[cur_level] = BTRFS_READ_LOCK;
2116                 }
2117
2118                 src_path->slots[cur_level] = dst_path->slots[cur_level];
2119                 if (cur_level) {
2120                         btrfs_node_key_to_cpu(dst_path->nodes[cur_level],
2121                                         &dst_key, dst_path->slots[cur_level]);
2122                         btrfs_node_key_to_cpu(src_path->nodes[cur_level],
2123                                         &src_key, src_path->slots[cur_level]);
2124                 } else {
2125                         btrfs_item_key_to_cpu(dst_path->nodes[cur_level],
2126                                         &dst_key, dst_path->slots[cur_level]);
2127                         btrfs_item_key_to_cpu(src_path->nodes[cur_level],
2128                                         &src_key, src_path->slots[cur_level]);
2129                 }
2130                 /* Content mismatch, something went wrong */
2131                 if (btrfs_comp_cpu_keys(&dst_key, &src_key)) {
2132                         ret = -ENOENT;
2133                         goto out;
2134                 }
2135                 cur_level--;
2136         }
2137
2138         /*
2139          * Now both @dst_path and @src_path have been populated, record the tree
2140          * blocks for qgroup accounting.
2141          */
2142         ret = btrfs_qgroup_trace_extent(trans, src_path->nodes[dst_level]->start,
2143                                         nodesize);
2144         if (ret < 0)
2145                 goto out;
2146         ret = btrfs_qgroup_trace_extent(trans, dst_path->nodes[dst_level]->start,
2147                                         nodesize);
2148         if (ret < 0)
2149                 goto out;
2150
2151         /* Record leaf file extents */
2152         if (dst_level == 0 && trace_leaf) {
2153                 ret = btrfs_qgroup_trace_leaf_items(trans, src_path->nodes[0]);
2154                 if (ret < 0)
2155                         goto out;
2156                 ret = btrfs_qgroup_trace_leaf_items(trans, dst_path->nodes[0]);
2157         }
2158 out:
2159         btrfs_free_path(src_path);
2160         return ret;
2161 }
2162
2163 /*
2164  * Helper function to do recursive generation-aware depth-first search, to
2165  * locate all new tree blocks in a subtree of reloc tree.
2166  *
2167  * E.g. (OO = Old tree blocks, NN = New tree blocks, whose gen == last_snapshot)
2168  *         reloc tree
2169  * L2         NN (a)
2170  *          /    \
2171  * L1    OO        NN (b)
2172  *      /  \      /  \
2173  * L0  OO  OO    OO  NN
2174  *               (c) (d)
2175  * If we pass:
2176  * @dst_path = [ nodes[1] = NN(b), nodes[0] = NULL ],
2177  * @cur_level = 1
2178  * @root_level = 1
2179  *
2180  * We will iterate through tree blocks NN(b), NN(d) and info qgroup to trace
2181  * above tree blocks along with their counter parts in file tree.
2182  * While during search, old tree blocks OO(c) will be skipped as tree block swap
2183  * won't affect OO(c).
2184  */
2185 static int qgroup_trace_new_subtree_blocks(struct btrfs_trans_handle* trans,
2186                                            struct extent_buffer *src_eb,
2187                                            struct btrfs_path *dst_path,
2188                                            int cur_level, int root_level,
2189                                            u64 last_snapshot, bool trace_leaf)
2190 {
2191         struct btrfs_fs_info *fs_info = trans->fs_info;
2192         struct extent_buffer *eb;
2193         bool need_cleanup = false;
2194         int ret = 0;
2195         int i;
2196
2197         /* Level sanity check */
2198         if (cur_level < 0 || cur_level >= BTRFS_MAX_LEVEL - 1 ||
2199             root_level < 0 || root_level >= BTRFS_MAX_LEVEL - 1 ||
2200             root_level < cur_level) {
2201                 btrfs_err_rl(fs_info,
2202                         "%s: bad levels, cur_level=%d root_level=%d",
2203                         __func__, cur_level, root_level);
2204                 return -EUCLEAN;
2205         }
2206
2207         /* Read the tree block if needed */
2208         if (dst_path->nodes[cur_level] == NULL) {
2209                 int parent_slot;
2210                 u64 child_gen;
2211
2212                 /*
2213                  * dst_path->nodes[root_level] must be initialized before
2214                  * calling this function.
2215                  */
2216                 if (cur_level == root_level) {
2217                         btrfs_err_rl(fs_info,
2218         "%s: dst_path->nodes[%d] not initialized, root_level=%d cur_level=%d",
2219                                 __func__, root_level, root_level, cur_level);
2220                         return -EUCLEAN;
2221                 }
2222
2223                 /*
2224                  * We need to get child blockptr/gen from parent before we can
2225                  * read it.
2226                   */
2227                 eb = dst_path->nodes[cur_level + 1];
2228                 parent_slot = dst_path->slots[cur_level + 1];
2229                 child_gen = btrfs_node_ptr_generation(eb, parent_slot);
2230
2231                 /* This node is old, no need to trace */
2232                 if (child_gen < last_snapshot)
2233                         goto out;
2234
2235                 eb = btrfs_read_node_slot(eb, parent_slot);
2236                 if (IS_ERR(eb)) {
2237                         ret = PTR_ERR(eb);
2238                         goto out;
2239                 }
2240
2241                 dst_path->nodes[cur_level] = eb;
2242                 dst_path->slots[cur_level] = 0;
2243
2244                 btrfs_tree_read_lock(eb);
2245                 dst_path->locks[cur_level] = BTRFS_READ_LOCK;
2246                 need_cleanup = true;
2247         }
2248
2249         /* Now record this tree block and its counter part for qgroups */
2250         ret = qgroup_trace_extent_swap(trans, src_eb, dst_path, cur_level,
2251                                        root_level, trace_leaf);
2252         if (ret < 0)
2253                 goto cleanup;
2254
2255         eb = dst_path->nodes[cur_level];
2256
2257         if (cur_level > 0) {
2258                 /* Iterate all child tree blocks */
2259                 for (i = 0; i < btrfs_header_nritems(eb); i++) {
2260                         /* Skip old tree blocks as they won't be swapped */
2261                         if (btrfs_node_ptr_generation(eb, i) < last_snapshot)
2262                                 continue;
2263                         dst_path->slots[cur_level] = i;
2264
2265                         /* Recursive call (at most 7 times) */
2266                         ret = qgroup_trace_new_subtree_blocks(trans, src_eb,
2267                                         dst_path, cur_level - 1, root_level,
2268                                         last_snapshot, trace_leaf);
2269                         if (ret < 0)
2270                                 goto cleanup;
2271                 }
2272         }
2273
2274 cleanup:
2275         if (need_cleanup) {
2276                 /* Clean up */
2277                 btrfs_tree_unlock_rw(dst_path->nodes[cur_level],
2278                                      dst_path->locks[cur_level]);
2279                 free_extent_buffer(dst_path->nodes[cur_level]);
2280                 dst_path->nodes[cur_level] = NULL;
2281                 dst_path->slots[cur_level] = 0;
2282                 dst_path->locks[cur_level] = 0;
2283         }
2284 out:
2285         return ret;
2286 }
2287
2288 static int qgroup_trace_subtree_swap(struct btrfs_trans_handle *trans,
2289                                 struct extent_buffer *src_eb,
2290                                 struct extent_buffer *dst_eb,
2291                                 u64 last_snapshot, bool trace_leaf)
2292 {
2293         struct btrfs_fs_info *fs_info = trans->fs_info;
2294         struct btrfs_path *dst_path = NULL;
2295         int level;
2296         int ret;
2297
2298         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2299                 return 0;
2300
2301         /* Wrong parameter order */
2302         if (btrfs_header_generation(src_eb) > btrfs_header_generation(dst_eb)) {
2303                 btrfs_err_rl(fs_info,
2304                 "%s: bad parameter order, src_gen=%llu dst_gen=%llu", __func__,
2305                              btrfs_header_generation(src_eb),
2306                              btrfs_header_generation(dst_eb));
2307                 return -EUCLEAN;
2308         }
2309
2310         if (!extent_buffer_uptodate(src_eb) || !extent_buffer_uptodate(dst_eb)) {
2311                 ret = -EIO;
2312                 goto out;
2313         }
2314
2315         level = btrfs_header_level(dst_eb);
2316         dst_path = btrfs_alloc_path();
2317         if (!dst_path) {
2318                 ret = -ENOMEM;
2319                 goto out;
2320         }
2321         /* For dst_path */
2322         atomic_inc(&dst_eb->refs);
2323         dst_path->nodes[level] = dst_eb;
2324         dst_path->slots[level] = 0;
2325         dst_path->locks[level] = 0;
2326
2327         /* Do the generation aware breadth-first search */
2328         ret = qgroup_trace_new_subtree_blocks(trans, src_eb, dst_path, level,
2329                                               level, last_snapshot, trace_leaf);
2330         if (ret < 0)
2331                 goto out;
2332         ret = 0;
2333
2334 out:
2335         btrfs_free_path(dst_path);
2336         if (ret < 0)
2337                 qgroup_mark_inconsistent(fs_info);
2338         return ret;
2339 }
2340
2341 int btrfs_qgroup_trace_subtree(struct btrfs_trans_handle *trans,
2342                                struct extent_buffer *root_eb,
2343                                u64 root_gen, int root_level)
2344 {
2345         struct btrfs_fs_info *fs_info = trans->fs_info;
2346         int ret = 0;
2347         int level;
2348         u8 drop_subptree_thres;
2349         struct extent_buffer *eb = root_eb;
2350         struct btrfs_path *path = NULL;
2351
2352         BUG_ON(root_level < 0 || root_level >= BTRFS_MAX_LEVEL);
2353         BUG_ON(root_eb == NULL);
2354
2355         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2356                 return 0;
2357
2358         spin_lock(&fs_info->qgroup_lock);
2359         drop_subptree_thres = fs_info->qgroup_drop_subtree_thres;
2360         spin_unlock(&fs_info->qgroup_lock);
2361
2362         /*
2363          * This function only gets called for snapshot drop, if we hit a high
2364          * node here, it means we are going to change ownership for quite a lot
2365          * of extents, which will greatly slow down btrfs_commit_transaction().
2366          *
2367          * So here if we find a high tree here, we just skip the accounting and
2368          * mark qgroup inconsistent.
2369          */
2370         if (root_level >= drop_subptree_thres) {
2371                 qgroup_mark_inconsistent(fs_info);
2372                 return 0;
2373         }
2374
2375         if (!extent_buffer_uptodate(root_eb)) {
2376                 struct btrfs_tree_parent_check check = {
2377                         .has_first_key = false,
2378                         .transid = root_gen,
2379                         .level = root_level
2380                 };
2381
2382                 ret = btrfs_read_extent_buffer(root_eb, &check);
2383                 if (ret)
2384                         goto out;
2385         }
2386
2387         if (root_level == 0) {
2388                 ret = btrfs_qgroup_trace_leaf_items(trans, root_eb);
2389                 goto out;
2390         }
2391
2392         path = btrfs_alloc_path();
2393         if (!path)
2394                 return -ENOMEM;
2395
2396         /*
2397          * Walk down the tree.  Missing extent blocks are filled in as
2398          * we go. Metadata is accounted every time we read a new
2399          * extent block.
2400          *
2401          * When we reach a leaf, we account for file extent items in it,
2402          * walk back up the tree (adjusting slot pointers as we go)
2403          * and restart the search process.
2404          */
2405         atomic_inc(&root_eb->refs);     /* For path */
2406         path->nodes[root_level] = root_eb;
2407         path->slots[root_level] = 0;
2408         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
2409 walk_down:
2410         level = root_level;
2411         while (level >= 0) {
2412                 if (path->nodes[level] == NULL) {
2413                         int parent_slot;
2414                         u64 child_bytenr;
2415
2416                         /*
2417                          * We need to get child blockptr from parent before we
2418                          * can read it.
2419                           */
2420                         eb = path->nodes[level + 1];
2421                         parent_slot = path->slots[level + 1];
2422                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
2423
2424                         eb = btrfs_read_node_slot(eb, parent_slot);
2425                         if (IS_ERR(eb)) {
2426                                 ret = PTR_ERR(eb);
2427                                 goto out;
2428                         }
2429
2430                         path->nodes[level] = eb;
2431                         path->slots[level] = 0;
2432
2433                         btrfs_tree_read_lock(eb);
2434                         path->locks[level] = BTRFS_READ_LOCK;
2435
2436                         ret = btrfs_qgroup_trace_extent(trans, child_bytenr,
2437                                                         fs_info->nodesize);
2438                         if (ret)
2439                                 goto out;
2440                 }
2441
2442                 if (level == 0) {
2443                         ret = btrfs_qgroup_trace_leaf_items(trans,
2444                                                             path->nodes[level]);
2445                         if (ret)
2446                                 goto out;
2447
2448                         /* Nonzero return here means we completed our search */
2449                         ret = adjust_slots_upwards(path, root_level);
2450                         if (ret)
2451                                 break;
2452
2453                         /* Restart search with new slots */
2454                         goto walk_down;
2455                 }
2456
2457                 level--;
2458         }
2459
2460         ret = 0;
2461 out:
2462         btrfs_free_path(path);
2463
2464         return ret;
2465 }
2466
2467 #define UPDATE_NEW      0
2468 #define UPDATE_OLD      1
2469 /*
2470  * Walk all of the roots that points to the bytenr and adjust their refcnts.
2471  */
2472 static int qgroup_update_refcnt(struct btrfs_fs_info *fs_info,
2473                                 struct ulist *roots, struct ulist *tmp,
2474                                 struct ulist *qgroups, u64 seq, int update_old)
2475 {
2476         struct ulist_node *unode;
2477         struct ulist_iterator uiter;
2478         struct ulist_node *tmp_unode;
2479         struct ulist_iterator tmp_uiter;
2480         struct btrfs_qgroup *qg;
2481         int ret = 0;
2482
2483         if (!roots)
2484                 return 0;
2485         ULIST_ITER_INIT(&uiter);
2486         while ((unode = ulist_next(roots, &uiter))) {
2487                 qg = find_qgroup_rb(fs_info, unode->val);
2488                 if (!qg)
2489                         continue;
2490
2491                 ulist_reinit(tmp);
2492                 ret = ulist_add(qgroups, qg->qgroupid, qgroup_to_aux(qg),
2493                                 GFP_ATOMIC);
2494                 if (ret < 0)
2495                         return ret;
2496                 ret = ulist_add(tmp, qg->qgroupid, qgroup_to_aux(qg), GFP_ATOMIC);
2497                 if (ret < 0)
2498                         return ret;
2499                 ULIST_ITER_INIT(&tmp_uiter);
2500                 while ((tmp_unode = ulist_next(tmp, &tmp_uiter))) {
2501                         struct btrfs_qgroup_list *glist;
2502
2503                         qg = unode_aux_to_qgroup(tmp_unode);
2504                         if (update_old)
2505                                 btrfs_qgroup_update_old_refcnt(qg, seq, 1);
2506                         else
2507                                 btrfs_qgroup_update_new_refcnt(qg, seq, 1);
2508                         list_for_each_entry(glist, &qg->groups, next_group) {
2509                                 ret = ulist_add(qgroups, glist->group->qgroupid,
2510                                                 qgroup_to_aux(glist->group),
2511                                                 GFP_ATOMIC);
2512                                 if (ret < 0)
2513                                         return ret;
2514                                 ret = ulist_add(tmp, glist->group->qgroupid,
2515                                                 qgroup_to_aux(glist->group),
2516                                                 GFP_ATOMIC);
2517                                 if (ret < 0)
2518                                         return ret;
2519                         }
2520                 }
2521         }
2522         return 0;
2523 }
2524
2525 /*
2526  * Update qgroup rfer/excl counters.
2527  * Rfer update is easy, codes can explain themselves.
2528  *
2529  * Excl update is tricky, the update is split into 2 parts.
2530  * Part 1: Possible exclusive <-> sharing detect:
2531  *      |       A       |       !A      |
2532  *  -------------------------------------
2533  *  B   |       *       |       -       |
2534  *  -------------------------------------
2535  *  !B  |       +       |       **      |
2536  *  -------------------------------------
2537  *
2538  * Conditions:
2539  * A:   cur_old_roots < nr_old_roots    (not exclusive before)
2540  * !A:  cur_old_roots == nr_old_roots   (possible exclusive before)
2541  * B:   cur_new_roots < nr_new_roots    (not exclusive now)
2542  * !B:  cur_new_roots == nr_new_roots   (possible exclusive now)
2543  *
2544  * Results:
2545  * +: Possible sharing -> exclusive     -: Possible exclusive -> sharing
2546  * *: Definitely not changed.           **: Possible unchanged.
2547  *
2548  * For !A and !B condition, the exception is cur_old/new_roots == 0 case.
2549  *
2550  * To make the logic clear, we first use condition A and B to split
2551  * combination into 4 results.
2552  *
2553  * Then, for result "+" and "-", check old/new_roots == 0 case, as in them
2554  * only on variant maybe 0.
2555  *
2556  * Lastly, check result **, since there are 2 variants maybe 0, split them
2557  * again(2x2).
2558  * But this time we don't need to consider other things, the codes and logic
2559  * is easy to understand now.
2560  */
2561 static int qgroup_update_counters(struct btrfs_fs_info *fs_info,
2562                                   struct ulist *qgroups,
2563                                   u64 nr_old_roots,
2564                                   u64 nr_new_roots,
2565                                   u64 num_bytes, u64 seq)
2566 {
2567         struct ulist_node *unode;
2568         struct ulist_iterator uiter;
2569         struct btrfs_qgroup *qg;
2570         u64 cur_new_count, cur_old_count;
2571
2572         ULIST_ITER_INIT(&uiter);
2573         while ((unode = ulist_next(qgroups, &uiter))) {
2574                 bool dirty = false;
2575
2576                 qg = unode_aux_to_qgroup(unode);
2577                 cur_old_count = btrfs_qgroup_get_old_refcnt(qg, seq);
2578                 cur_new_count = btrfs_qgroup_get_new_refcnt(qg, seq);
2579
2580                 trace_qgroup_update_counters(fs_info, qg, cur_old_count,
2581                                              cur_new_count);
2582
2583                 /* Rfer update part */
2584                 if (cur_old_count == 0 && cur_new_count > 0) {
2585                         qg->rfer += num_bytes;
2586                         qg->rfer_cmpr += num_bytes;
2587                         dirty = true;
2588                 }
2589                 if (cur_old_count > 0 && cur_new_count == 0) {
2590                         qg->rfer -= num_bytes;
2591                         qg->rfer_cmpr -= num_bytes;
2592                         dirty = true;
2593                 }
2594
2595                 /* Excl update part */
2596                 /* Exclusive/none -> shared case */
2597                 if (cur_old_count == nr_old_roots &&
2598                     cur_new_count < nr_new_roots) {
2599                         /* Exclusive -> shared */
2600                         if (cur_old_count != 0) {
2601                                 qg->excl -= num_bytes;
2602                                 qg->excl_cmpr -= num_bytes;
2603                                 dirty = true;
2604                         }
2605                 }
2606
2607                 /* Shared -> exclusive/none case */
2608                 if (cur_old_count < nr_old_roots &&
2609                     cur_new_count == nr_new_roots) {
2610                         /* Shared->exclusive */
2611                         if (cur_new_count != 0) {
2612                                 qg->excl += num_bytes;
2613                                 qg->excl_cmpr += num_bytes;
2614                                 dirty = true;
2615                         }
2616                 }
2617
2618                 /* Exclusive/none -> exclusive/none case */
2619                 if (cur_old_count == nr_old_roots &&
2620                     cur_new_count == nr_new_roots) {
2621                         if (cur_old_count == 0) {
2622                                 /* None -> exclusive/none */
2623
2624                                 if (cur_new_count != 0) {
2625                                         /* None -> exclusive */
2626                                         qg->excl += num_bytes;
2627                                         qg->excl_cmpr += num_bytes;
2628                                         dirty = true;
2629                                 }
2630                                 /* None -> none, nothing changed */
2631                         } else {
2632                                 /* Exclusive -> exclusive/none */
2633
2634                                 if (cur_new_count == 0) {
2635                                         /* Exclusive -> none */
2636                                         qg->excl -= num_bytes;
2637                                         qg->excl_cmpr -= num_bytes;
2638                                         dirty = true;
2639                                 }
2640                                 /* Exclusive -> exclusive, nothing changed */
2641                         }
2642                 }
2643
2644                 if (dirty)
2645                         qgroup_dirty(fs_info, qg);
2646         }
2647         return 0;
2648 }
2649
2650 /*
2651  * Check if the @roots potentially is a list of fs tree roots
2652  *
2653  * Return 0 for definitely not a fs/subvol tree roots ulist
2654  * Return 1 for possible fs/subvol tree roots in the list (considering an empty
2655  *          one as well)
2656  */
2657 static int maybe_fs_roots(struct ulist *roots)
2658 {
2659         struct ulist_node *unode;
2660         struct ulist_iterator uiter;
2661
2662         /* Empty one, still possible for fs roots */
2663         if (!roots || roots->nnodes == 0)
2664                 return 1;
2665
2666         ULIST_ITER_INIT(&uiter);
2667         unode = ulist_next(roots, &uiter);
2668         if (!unode)
2669                 return 1;
2670
2671         /*
2672          * If it contains fs tree roots, then it must belong to fs/subvol
2673          * trees.
2674          * If it contains a non-fs tree, it won't be shared with fs/subvol trees.
2675          */
2676         return is_fstree(unode->val);
2677 }
2678
2679 int btrfs_qgroup_account_extent(struct btrfs_trans_handle *trans, u64 bytenr,
2680                                 u64 num_bytes, struct ulist *old_roots,
2681                                 struct ulist *new_roots)
2682 {
2683         struct btrfs_fs_info *fs_info = trans->fs_info;
2684         struct ulist *qgroups = NULL;
2685         struct ulist *tmp = NULL;
2686         u64 seq;
2687         u64 nr_new_roots = 0;
2688         u64 nr_old_roots = 0;
2689         int ret = 0;
2690
2691         /*
2692          * If quotas get disabled meanwhile, the resources need to be freed and
2693          * we can't just exit here.
2694          */
2695         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
2696             fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING)
2697                 goto out_free;
2698
2699         if (new_roots) {
2700                 if (!maybe_fs_roots(new_roots))
2701                         goto out_free;
2702                 nr_new_roots = new_roots->nnodes;
2703         }
2704         if (old_roots) {
2705                 if (!maybe_fs_roots(old_roots))
2706                         goto out_free;
2707                 nr_old_roots = old_roots->nnodes;
2708         }
2709
2710         /* Quick exit, either not fs tree roots, or won't affect any qgroup */
2711         if (nr_old_roots == 0 && nr_new_roots == 0)
2712                 goto out_free;
2713
2714         BUG_ON(!fs_info->quota_root);
2715
2716         trace_btrfs_qgroup_account_extent(fs_info, trans->transid, bytenr,
2717                                         num_bytes, nr_old_roots, nr_new_roots);
2718
2719         qgroups = ulist_alloc(GFP_NOFS);
2720         if (!qgroups) {
2721                 ret = -ENOMEM;
2722                 goto out_free;
2723         }
2724         tmp = ulist_alloc(GFP_NOFS);
2725         if (!tmp) {
2726                 ret = -ENOMEM;
2727                 goto out_free;
2728         }
2729
2730         mutex_lock(&fs_info->qgroup_rescan_lock);
2731         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
2732                 if (fs_info->qgroup_rescan_progress.objectid <= bytenr) {
2733                         mutex_unlock(&fs_info->qgroup_rescan_lock);
2734                         ret = 0;
2735                         goto out_free;
2736                 }
2737         }
2738         mutex_unlock(&fs_info->qgroup_rescan_lock);
2739
2740         spin_lock(&fs_info->qgroup_lock);
2741         seq = fs_info->qgroup_seq;
2742
2743         /* Update old refcnts using old_roots */
2744         ret = qgroup_update_refcnt(fs_info, old_roots, tmp, qgroups, seq,
2745                                    UPDATE_OLD);
2746         if (ret < 0)
2747                 goto out;
2748
2749         /* Update new refcnts using new_roots */
2750         ret = qgroup_update_refcnt(fs_info, new_roots, tmp, qgroups, seq,
2751                                    UPDATE_NEW);
2752         if (ret < 0)
2753                 goto out;
2754
2755         qgroup_update_counters(fs_info, qgroups, nr_old_roots, nr_new_roots,
2756                                num_bytes, seq);
2757
2758         /*
2759          * Bump qgroup_seq to avoid seq overlap
2760          */
2761         fs_info->qgroup_seq += max(nr_old_roots, nr_new_roots) + 1;
2762 out:
2763         spin_unlock(&fs_info->qgroup_lock);
2764 out_free:
2765         ulist_free(tmp);
2766         ulist_free(qgroups);
2767         ulist_free(old_roots);
2768         ulist_free(new_roots);
2769         return ret;
2770 }
2771
2772 int btrfs_qgroup_account_extents(struct btrfs_trans_handle *trans)
2773 {
2774         struct btrfs_fs_info *fs_info = trans->fs_info;
2775         struct btrfs_qgroup_extent_record *record;
2776         struct btrfs_delayed_ref_root *delayed_refs;
2777         struct ulist *new_roots = NULL;
2778         struct rb_node *node;
2779         u64 num_dirty_extents = 0;
2780         u64 qgroup_to_skip;
2781         int ret = 0;
2782
2783         delayed_refs = &trans->transaction->delayed_refs;
2784         qgroup_to_skip = delayed_refs->qgroup_to_skip;
2785         while ((node = rb_first(&delayed_refs->dirty_extent_root))) {
2786                 record = rb_entry(node, struct btrfs_qgroup_extent_record,
2787                                   node);
2788
2789                 num_dirty_extents++;
2790                 trace_btrfs_qgroup_account_extents(fs_info, record);
2791
2792                 if (!ret && !(fs_info->qgroup_flags &
2793                               BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING)) {
2794                         struct btrfs_backref_walk_ctx ctx = { 0 };
2795
2796                         ctx.bytenr = record->bytenr;
2797                         ctx.fs_info = fs_info;
2798
2799                         /*
2800                          * Old roots should be searched when inserting qgroup
2801                          * extent record.
2802                          *
2803                          * But for INCONSISTENT (NO_ACCOUNTING) -> rescan case,
2804                          * we may have some record inserted during
2805                          * NO_ACCOUNTING (thus no old_roots populated), but
2806                          * later we start rescan, which clears NO_ACCOUNTING,
2807                          * leaving some inserted records without old_roots
2808                          * populated.
2809                          *
2810                          * Those cases are rare and should not cause too much
2811                          * time spent during commit_transaction().
2812                          */
2813                         if (!record->old_roots) {
2814                                 /* Search commit root to find old_roots */
2815                                 ret = btrfs_find_all_roots(&ctx, false);
2816                                 if (ret < 0)
2817                                         goto cleanup;
2818                                 record->old_roots = ctx.roots;
2819                                 ctx.roots = NULL;
2820                         }
2821
2822                         /* Free the reserved data space */
2823                         btrfs_qgroup_free_refroot(fs_info,
2824                                         record->data_rsv_refroot,
2825                                         record->data_rsv,
2826                                         BTRFS_QGROUP_RSV_DATA);
2827                         /*
2828                          * Use BTRFS_SEQ_LAST as time_seq to do special search,
2829                          * which doesn't lock tree or delayed_refs and search
2830                          * current root. It's safe inside commit_transaction().
2831                          */
2832                         ctx.trans = trans;
2833                         ctx.time_seq = BTRFS_SEQ_LAST;
2834                         ret = btrfs_find_all_roots(&ctx, false);
2835                         if (ret < 0)
2836                                 goto cleanup;
2837                         new_roots = ctx.roots;
2838                         if (qgroup_to_skip) {
2839                                 ulist_del(new_roots, qgroup_to_skip, 0);
2840                                 ulist_del(record->old_roots, qgroup_to_skip,
2841                                           0);
2842                         }
2843                         ret = btrfs_qgroup_account_extent(trans, record->bytenr,
2844                                                           record->num_bytes,
2845                                                           record->old_roots,
2846                                                           new_roots);
2847                         record->old_roots = NULL;
2848                         new_roots = NULL;
2849                 }
2850 cleanup:
2851                 ulist_free(record->old_roots);
2852                 ulist_free(new_roots);
2853                 new_roots = NULL;
2854                 rb_erase(node, &delayed_refs->dirty_extent_root);
2855                 kfree(record);
2856
2857         }
2858         trace_qgroup_num_dirty_extents(fs_info, trans->transid,
2859                                        num_dirty_extents);
2860         return ret;
2861 }
2862
2863 /*
2864  * Writes all changed qgroups to disk.
2865  * Called by the transaction commit path and the qgroup assign ioctl.
2866  */
2867 int btrfs_run_qgroups(struct btrfs_trans_handle *trans)
2868 {
2869         struct btrfs_fs_info *fs_info = trans->fs_info;
2870         int ret = 0;
2871
2872         /*
2873          * In case we are called from the qgroup assign ioctl, assert that we
2874          * are holding the qgroup_ioctl_lock, otherwise we can race with a quota
2875          * disable operation (ioctl) and access a freed quota root.
2876          */
2877         if (trans->transaction->state != TRANS_STATE_COMMIT_DOING)
2878                 lockdep_assert_held(&fs_info->qgroup_ioctl_lock);
2879
2880         if (!fs_info->quota_root)
2881                 return ret;
2882
2883         spin_lock(&fs_info->qgroup_lock);
2884         while (!list_empty(&fs_info->dirty_qgroups)) {
2885                 struct btrfs_qgroup *qgroup;
2886                 qgroup = list_first_entry(&fs_info->dirty_qgroups,
2887                                           struct btrfs_qgroup, dirty);
2888                 list_del_init(&qgroup->dirty);
2889                 spin_unlock(&fs_info->qgroup_lock);
2890                 ret = update_qgroup_info_item(trans, qgroup);
2891                 if (ret)
2892                         qgroup_mark_inconsistent(fs_info);
2893                 ret = update_qgroup_limit_item(trans, qgroup);
2894                 if (ret)
2895                         qgroup_mark_inconsistent(fs_info);
2896                 spin_lock(&fs_info->qgroup_lock);
2897         }
2898         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2899                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_ON;
2900         else
2901                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_ON;
2902         spin_unlock(&fs_info->qgroup_lock);
2903
2904         ret = update_qgroup_status_item(trans);
2905         if (ret)
2906                 qgroup_mark_inconsistent(fs_info);
2907
2908         return ret;
2909 }
2910
2911 /*
2912  * Copy the accounting information between qgroups. This is necessary
2913  * when a snapshot or a subvolume is created. Throwing an error will
2914  * cause a transaction abort so we take extra care here to only error
2915  * when a readonly fs is a reasonable outcome.
2916  */
2917 int btrfs_qgroup_inherit(struct btrfs_trans_handle *trans, u64 srcid,
2918                          u64 objectid, struct btrfs_qgroup_inherit *inherit)
2919 {
2920         int ret = 0;
2921         int i;
2922         u64 *i_qgroups;
2923         bool committing = false;
2924         struct btrfs_fs_info *fs_info = trans->fs_info;
2925         struct btrfs_root *quota_root;
2926         struct btrfs_qgroup *srcgroup;
2927         struct btrfs_qgroup *dstgroup;
2928         bool need_rescan = false;
2929         u32 level_size = 0;
2930         u64 nums;
2931
2932         /*
2933          * There are only two callers of this function.
2934          *
2935          * One in create_subvol() in the ioctl context, which needs to hold
2936          * the qgroup_ioctl_lock.
2937          *
2938          * The other one in create_pending_snapshot() where no other qgroup
2939          * code can modify the fs as they all need to either start a new trans
2940          * or hold a trans handler, thus we don't need to hold
2941          * qgroup_ioctl_lock.
2942          * This would avoid long and complex lock chain and make lockdep happy.
2943          */
2944         spin_lock(&fs_info->trans_lock);
2945         if (trans->transaction->state == TRANS_STATE_COMMIT_DOING)
2946                 committing = true;
2947         spin_unlock(&fs_info->trans_lock);
2948
2949         if (!committing)
2950                 mutex_lock(&fs_info->qgroup_ioctl_lock);
2951         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
2952                 goto out;
2953
2954         quota_root = fs_info->quota_root;
2955         if (!quota_root) {
2956                 ret = -EINVAL;
2957                 goto out;
2958         }
2959
2960         if (inherit) {
2961                 i_qgroups = (u64 *)(inherit + 1);
2962                 nums = inherit->num_qgroups + 2 * inherit->num_ref_copies +
2963                        2 * inherit->num_excl_copies;
2964                 for (i = 0; i < nums; ++i) {
2965                         srcgroup = find_qgroup_rb(fs_info, *i_qgroups);
2966
2967                         /*
2968                          * Zero out invalid groups so we can ignore
2969                          * them later.
2970                          */
2971                         if (!srcgroup ||
2972                             ((srcgroup->qgroupid >> 48) <= (objectid >> 48)))
2973                                 *i_qgroups = 0ULL;
2974
2975                         ++i_qgroups;
2976                 }
2977         }
2978
2979         /*
2980          * create a tracking group for the subvol itself
2981          */
2982         ret = add_qgroup_item(trans, quota_root, objectid);
2983         if (ret)
2984                 goto out;
2985
2986         /*
2987          * add qgroup to all inherited groups
2988          */
2989         if (inherit) {
2990                 i_qgroups = (u64 *)(inherit + 1);
2991                 for (i = 0; i < inherit->num_qgroups; ++i, ++i_qgroups) {
2992                         if (*i_qgroups == 0)
2993                                 continue;
2994                         ret = add_qgroup_relation_item(trans, objectid,
2995                                                        *i_qgroups);
2996                         if (ret && ret != -EEXIST)
2997                                 goto out;
2998                         ret = add_qgroup_relation_item(trans, *i_qgroups,
2999                                                        objectid);
3000                         if (ret && ret != -EEXIST)
3001                                 goto out;
3002                 }
3003                 ret = 0;
3004         }
3005
3006
3007         spin_lock(&fs_info->qgroup_lock);
3008
3009         dstgroup = add_qgroup_rb(fs_info, objectid);
3010         if (IS_ERR(dstgroup)) {
3011                 ret = PTR_ERR(dstgroup);
3012                 goto unlock;
3013         }
3014
3015         if (inherit && inherit->flags & BTRFS_QGROUP_INHERIT_SET_LIMITS) {
3016                 dstgroup->lim_flags = inherit->lim.flags;
3017                 dstgroup->max_rfer = inherit->lim.max_rfer;
3018                 dstgroup->max_excl = inherit->lim.max_excl;
3019                 dstgroup->rsv_rfer = inherit->lim.rsv_rfer;
3020                 dstgroup->rsv_excl = inherit->lim.rsv_excl;
3021
3022                 qgroup_dirty(fs_info, dstgroup);
3023         }
3024
3025         if (srcid) {
3026                 srcgroup = find_qgroup_rb(fs_info, srcid);
3027                 if (!srcgroup)
3028                         goto unlock;
3029
3030                 /*
3031                  * We call inherit after we clone the root in order to make sure
3032                  * our counts don't go crazy, so at this point the only
3033                  * difference between the two roots should be the root node.
3034                  */
3035                 level_size = fs_info->nodesize;
3036                 dstgroup->rfer = srcgroup->rfer;
3037                 dstgroup->rfer_cmpr = srcgroup->rfer_cmpr;
3038                 dstgroup->excl = level_size;
3039                 dstgroup->excl_cmpr = level_size;
3040                 srcgroup->excl = level_size;
3041                 srcgroup->excl_cmpr = level_size;
3042
3043                 /* inherit the limit info */
3044                 dstgroup->lim_flags = srcgroup->lim_flags;
3045                 dstgroup->max_rfer = srcgroup->max_rfer;
3046                 dstgroup->max_excl = srcgroup->max_excl;
3047                 dstgroup->rsv_rfer = srcgroup->rsv_rfer;
3048                 dstgroup->rsv_excl = srcgroup->rsv_excl;
3049
3050                 qgroup_dirty(fs_info, dstgroup);
3051                 qgroup_dirty(fs_info, srcgroup);
3052         }
3053
3054         if (!inherit)
3055                 goto unlock;
3056
3057         i_qgroups = (u64 *)(inherit + 1);
3058         for (i = 0; i < inherit->num_qgroups; ++i) {
3059                 if (*i_qgroups) {
3060                         ret = add_relation_rb(fs_info, objectid, *i_qgroups);
3061                         if (ret)
3062                                 goto unlock;
3063                 }
3064                 ++i_qgroups;
3065
3066                 /*
3067                  * If we're doing a snapshot, and adding the snapshot to a new
3068                  * qgroup, the numbers are guaranteed to be incorrect.
3069                  */
3070                 if (srcid)
3071                         need_rescan = true;
3072         }
3073
3074         for (i = 0; i <  inherit->num_ref_copies; ++i, i_qgroups += 2) {
3075                 struct btrfs_qgroup *src;
3076                 struct btrfs_qgroup *dst;
3077
3078                 if (!i_qgroups[0] || !i_qgroups[1])
3079                         continue;
3080
3081                 src = find_qgroup_rb(fs_info, i_qgroups[0]);
3082                 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
3083
3084                 if (!src || !dst) {
3085                         ret = -EINVAL;
3086                         goto unlock;
3087                 }
3088
3089                 dst->rfer = src->rfer - level_size;
3090                 dst->rfer_cmpr = src->rfer_cmpr - level_size;
3091
3092                 /* Manually tweaking numbers certainly needs a rescan */
3093                 need_rescan = true;
3094         }
3095         for (i = 0; i <  inherit->num_excl_copies; ++i, i_qgroups += 2) {
3096                 struct btrfs_qgroup *src;
3097                 struct btrfs_qgroup *dst;
3098
3099                 if (!i_qgroups[0] || !i_qgroups[1])
3100                         continue;
3101
3102                 src = find_qgroup_rb(fs_info, i_qgroups[0]);
3103                 dst = find_qgroup_rb(fs_info, i_qgroups[1]);
3104
3105                 if (!src || !dst) {
3106                         ret = -EINVAL;
3107                         goto unlock;
3108                 }
3109
3110                 dst->excl = src->excl + level_size;
3111                 dst->excl_cmpr = src->excl_cmpr + level_size;
3112                 need_rescan = true;
3113         }
3114
3115 unlock:
3116         spin_unlock(&fs_info->qgroup_lock);
3117         if (!ret)
3118                 ret = btrfs_sysfs_add_one_qgroup(fs_info, dstgroup);
3119 out:
3120         if (!committing)
3121                 mutex_unlock(&fs_info->qgroup_ioctl_lock);
3122         if (need_rescan)
3123                 qgroup_mark_inconsistent(fs_info);
3124         return ret;
3125 }
3126
3127 static bool qgroup_check_limits(const struct btrfs_qgroup *qg, u64 num_bytes)
3128 {
3129         if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_RFER) &&
3130             qgroup_rsv_total(qg) + (s64)qg->rfer + num_bytes > qg->max_rfer)
3131                 return false;
3132
3133         if ((qg->lim_flags & BTRFS_QGROUP_LIMIT_MAX_EXCL) &&
3134             qgroup_rsv_total(qg) + (s64)qg->excl + num_bytes > qg->max_excl)
3135                 return false;
3136
3137         return true;
3138 }
3139
3140 static int qgroup_reserve(struct btrfs_root *root, u64 num_bytes, bool enforce,
3141                           enum btrfs_qgroup_rsv_type type)
3142 {
3143         struct btrfs_qgroup *qgroup;
3144         struct btrfs_fs_info *fs_info = root->fs_info;
3145         u64 ref_root = root->root_key.objectid;
3146         int ret = 0;
3147         LIST_HEAD(qgroup_list);
3148
3149         if (!is_fstree(ref_root))
3150                 return 0;
3151
3152         if (num_bytes == 0)
3153                 return 0;
3154
3155         if (test_bit(BTRFS_FS_QUOTA_OVERRIDE, &fs_info->flags) &&
3156             capable(CAP_SYS_RESOURCE))
3157                 enforce = false;
3158
3159         spin_lock(&fs_info->qgroup_lock);
3160         if (!fs_info->quota_root)
3161                 goto out;
3162
3163         qgroup = find_qgroup_rb(fs_info, ref_root);
3164         if (!qgroup)
3165                 goto out;
3166
3167         qgroup_iterator_add(&qgroup_list, qgroup);
3168         list_for_each_entry(qgroup, &qgroup_list, iterator) {
3169                 struct btrfs_qgroup_list *glist;
3170
3171                 if (enforce && !qgroup_check_limits(qgroup, num_bytes)) {
3172                         ret = -EDQUOT;
3173                         goto out;
3174                 }
3175
3176                 list_for_each_entry(glist, &qgroup->groups, next_group)
3177                         qgroup_iterator_add(&qgroup_list, glist->group);
3178         }
3179
3180         ret = 0;
3181         /*
3182          * no limits exceeded, now record the reservation into all qgroups
3183          */
3184         list_for_each_entry(qgroup, &qgroup_list, iterator)
3185                 qgroup_rsv_add(fs_info, qgroup, num_bytes, type);
3186
3187 out:
3188         qgroup_iterator_clean(&qgroup_list);
3189         spin_unlock(&fs_info->qgroup_lock);
3190         return ret;
3191 }
3192
3193 /*
3194  * Free @num_bytes of reserved space with @type for qgroup.  (Normally level 0
3195  * qgroup).
3196  *
3197  * Will handle all higher level qgroup too.
3198  *
3199  * NOTE: If @num_bytes is (u64)-1, this means to free all bytes of this qgroup.
3200  * This special case is only used for META_PERTRANS type.
3201  */
3202 void btrfs_qgroup_free_refroot(struct btrfs_fs_info *fs_info,
3203                                u64 ref_root, u64 num_bytes,
3204                                enum btrfs_qgroup_rsv_type type)
3205 {
3206         struct btrfs_qgroup *qgroup;
3207         struct ulist_node *unode;
3208         struct ulist_iterator uiter;
3209         int ret = 0;
3210
3211         if (!is_fstree(ref_root))
3212                 return;
3213
3214         if (num_bytes == 0)
3215                 return;
3216
3217         if (num_bytes == (u64)-1 && type != BTRFS_QGROUP_RSV_META_PERTRANS) {
3218                 WARN(1, "%s: Invalid type to free", __func__);
3219                 return;
3220         }
3221         spin_lock(&fs_info->qgroup_lock);
3222
3223         if (!fs_info->quota_root)
3224                 goto out;
3225
3226         qgroup = find_qgroup_rb(fs_info, ref_root);
3227         if (!qgroup)
3228                 goto out;
3229
3230         if (num_bytes == (u64)-1)
3231                 /*
3232                  * We're freeing all pertrans rsv, get reserved value from
3233                  * level 0 qgroup as real num_bytes to free.
3234                  */
3235                 num_bytes = qgroup->rsv.values[type];
3236
3237         ulist_reinit(fs_info->qgroup_ulist);
3238         ret = ulist_add(fs_info->qgroup_ulist, qgroup->qgroupid,
3239                         qgroup_to_aux(qgroup), GFP_ATOMIC);
3240         if (ret < 0)
3241                 goto out;
3242         ULIST_ITER_INIT(&uiter);
3243         while ((unode = ulist_next(fs_info->qgroup_ulist, &uiter))) {
3244                 struct btrfs_qgroup *qg;
3245                 struct btrfs_qgroup_list *glist;
3246
3247                 qg = unode_aux_to_qgroup(unode);
3248
3249                 qgroup_rsv_release(fs_info, qg, num_bytes, type);
3250
3251                 list_for_each_entry(glist, &qg->groups, next_group) {
3252                         ret = ulist_add(fs_info->qgroup_ulist,
3253                                         glist->group->qgroupid,
3254                                         qgroup_to_aux(glist->group), GFP_ATOMIC);
3255                         if (ret < 0)
3256                                 goto out;
3257                 }
3258         }
3259
3260 out:
3261         spin_unlock(&fs_info->qgroup_lock);
3262 }
3263
3264 /*
3265  * Check if the leaf is the last leaf. Which means all node pointers
3266  * are at their last position.
3267  */
3268 static bool is_last_leaf(struct btrfs_path *path)
3269 {
3270         int i;
3271
3272         for (i = 1; i < BTRFS_MAX_LEVEL && path->nodes[i]; i++) {
3273                 if (path->slots[i] != btrfs_header_nritems(path->nodes[i]) - 1)
3274                         return false;
3275         }
3276         return true;
3277 }
3278
3279 /*
3280  * returns < 0 on error, 0 when more leafs are to be scanned.
3281  * returns 1 when done.
3282  */
3283 static int qgroup_rescan_leaf(struct btrfs_trans_handle *trans,
3284                               struct btrfs_path *path)
3285 {
3286         struct btrfs_fs_info *fs_info = trans->fs_info;
3287         struct btrfs_root *extent_root;
3288         struct btrfs_key found;
3289         struct extent_buffer *scratch_leaf = NULL;
3290         u64 num_bytes;
3291         bool done;
3292         int slot;
3293         int ret;
3294
3295         mutex_lock(&fs_info->qgroup_rescan_lock);
3296         extent_root = btrfs_extent_root(fs_info,
3297                                 fs_info->qgroup_rescan_progress.objectid);
3298         ret = btrfs_search_slot_for_read(extent_root,
3299                                          &fs_info->qgroup_rescan_progress,
3300                                          path, 1, 0);
3301
3302         btrfs_debug(fs_info,
3303                 "current progress key (%llu %u %llu), search_slot ret %d",
3304                 fs_info->qgroup_rescan_progress.objectid,
3305                 fs_info->qgroup_rescan_progress.type,
3306                 fs_info->qgroup_rescan_progress.offset, ret);
3307
3308         if (ret) {
3309                 /*
3310                  * The rescan is about to end, we will not be scanning any
3311                  * further blocks. We cannot unset the RESCAN flag here, because
3312                  * we want to commit the transaction if everything went well.
3313                  * To make the live accounting work in this phase, we set our
3314                  * scan progress pointer such that every real extent objectid
3315                  * will be smaller.
3316                  */
3317                 fs_info->qgroup_rescan_progress.objectid = (u64)-1;
3318                 btrfs_release_path(path);
3319                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3320                 return ret;
3321         }
3322         done = is_last_leaf(path);
3323
3324         btrfs_item_key_to_cpu(path->nodes[0], &found,
3325                               btrfs_header_nritems(path->nodes[0]) - 1);
3326         fs_info->qgroup_rescan_progress.objectid = found.objectid + 1;
3327
3328         scratch_leaf = btrfs_clone_extent_buffer(path->nodes[0]);
3329         if (!scratch_leaf) {
3330                 ret = -ENOMEM;
3331                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3332                 goto out;
3333         }
3334         slot = path->slots[0];
3335         btrfs_release_path(path);
3336         mutex_unlock(&fs_info->qgroup_rescan_lock);
3337
3338         for (; slot < btrfs_header_nritems(scratch_leaf); ++slot) {
3339                 struct btrfs_backref_walk_ctx ctx = { 0 };
3340
3341                 btrfs_item_key_to_cpu(scratch_leaf, &found, slot);
3342                 if (found.type != BTRFS_EXTENT_ITEM_KEY &&
3343                     found.type != BTRFS_METADATA_ITEM_KEY)
3344                         continue;
3345                 if (found.type == BTRFS_METADATA_ITEM_KEY)
3346                         num_bytes = fs_info->nodesize;
3347                 else
3348                         num_bytes = found.offset;
3349
3350                 ctx.bytenr = found.objectid;
3351                 ctx.fs_info = fs_info;
3352
3353                 ret = btrfs_find_all_roots(&ctx, false);
3354                 if (ret < 0)
3355                         goto out;
3356                 /* For rescan, just pass old_roots as NULL */
3357                 ret = btrfs_qgroup_account_extent(trans, found.objectid,
3358                                                   num_bytes, NULL, ctx.roots);
3359                 if (ret < 0)
3360                         goto out;
3361         }
3362 out:
3363         if (scratch_leaf)
3364                 free_extent_buffer(scratch_leaf);
3365
3366         if (done && !ret) {
3367                 ret = 1;
3368                 fs_info->qgroup_rescan_progress.objectid = (u64)-1;
3369         }
3370         return ret;
3371 }
3372
3373 static bool rescan_should_stop(struct btrfs_fs_info *fs_info)
3374 {
3375         return btrfs_fs_closing(fs_info) ||
3376                 test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state) ||
3377                 !test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
3378                           fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN;
3379 }
3380
3381 static void btrfs_qgroup_rescan_worker(struct btrfs_work *work)
3382 {
3383         struct btrfs_fs_info *fs_info = container_of(work, struct btrfs_fs_info,
3384                                                      qgroup_rescan_work);
3385         struct btrfs_path *path;
3386         struct btrfs_trans_handle *trans = NULL;
3387         int err = -ENOMEM;
3388         int ret = 0;
3389         bool stopped = false;
3390         bool did_leaf_rescans = false;
3391
3392         path = btrfs_alloc_path();
3393         if (!path)
3394                 goto out;
3395         /*
3396          * Rescan should only search for commit root, and any later difference
3397          * should be recorded by qgroup
3398          */
3399         path->search_commit_root = 1;
3400         path->skip_locking = 1;
3401
3402         err = 0;
3403         while (!err && !(stopped = rescan_should_stop(fs_info))) {
3404                 trans = btrfs_start_transaction(fs_info->fs_root, 0);
3405                 if (IS_ERR(trans)) {
3406                         err = PTR_ERR(trans);
3407                         break;
3408                 }
3409
3410                 err = qgroup_rescan_leaf(trans, path);
3411                 did_leaf_rescans = true;
3412
3413                 if (err > 0)
3414                         btrfs_commit_transaction(trans);
3415                 else
3416                         btrfs_end_transaction(trans);
3417         }
3418
3419 out:
3420         btrfs_free_path(path);
3421
3422         mutex_lock(&fs_info->qgroup_rescan_lock);
3423         if (err > 0 &&
3424             fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT) {
3425                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
3426         } else if (err < 0 || stopped) {
3427                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_INCONSISTENT;
3428         }
3429         mutex_unlock(&fs_info->qgroup_rescan_lock);
3430
3431         /*
3432          * Only update status, since the previous part has already updated the
3433          * qgroup info, and only if we did any actual work. This also prevents
3434          * race with a concurrent quota disable, which has already set
3435          * fs_info->quota_root to NULL and cleared BTRFS_FS_QUOTA_ENABLED at
3436          * btrfs_quota_disable().
3437          */
3438         if (did_leaf_rescans) {
3439                 trans = btrfs_start_transaction(fs_info->quota_root, 1);
3440                 if (IS_ERR(trans)) {
3441                         err = PTR_ERR(trans);
3442                         trans = NULL;
3443                         btrfs_err(fs_info,
3444                                   "fail to start transaction for status update: %d",
3445                                   err);
3446                 }
3447         } else {
3448                 trans = NULL;
3449         }
3450
3451         mutex_lock(&fs_info->qgroup_rescan_lock);
3452         if (!stopped ||
3453             fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN)
3454                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3455         if (trans) {
3456                 ret = update_qgroup_status_item(trans);
3457                 if (ret < 0) {
3458                         err = ret;
3459                         btrfs_err(fs_info, "fail to update qgroup status: %d",
3460                                   err);
3461                 }
3462         }
3463         fs_info->qgroup_rescan_running = false;
3464         fs_info->qgroup_flags &= ~BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN;
3465         complete_all(&fs_info->qgroup_rescan_completion);
3466         mutex_unlock(&fs_info->qgroup_rescan_lock);
3467
3468         if (!trans)
3469                 return;
3470
3471         btrfs_end_transaction(trans);
3472
3473         if (stopped) {
3474                 btrfs_info(fs_info, "qgroup scan paused");
3475         } else if (fs_info->qgroup_flags & BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN) {
3476                 btrfs_info(fs_info, "qgroup scan cancelled");
3477         } else if (err >= 0) {
3478                 btrfs_info(fs_info, "qgroup scan completed%s",
3479                         err > 0 ? " (inconsistency flag cleared)" : "");
3480         } else {
3481                 btrfs_err(fs_info, "qgroup scan failed with %d", err);
3482         }
3483 }
3484
3485 /*
3486  * Checks that (a) no rescan is running and (b) quota is enabled. Allocates all
3487  * memory required for the rescan context.
3488  */
3489 static int
3490 qgroup_rescan_init(struct btrfs_fs_info *fs_info, u64 progress_objectid,
3491                    int init_flags)
3492 {
3493         int ret = 0;
3494
3495         if (!init_flags) {
3496                 /* we're resuming qgroup rescan at mount time */
3497                 if (!(fs_info->qgroup_flags &
3498                       BTRFS_QGROUP_STATUS_FLAG_RESCAN)) {
3499                         btrfs_warn(fs_info,
3500                         "qgroup rescan init failed, qgroup rescan is not queued");
3501                         ret = -EINVAL;
3502                 } else if (!(fs_info->qgroup_flags &
3503                              BTRFS_QGROUP_STATUS_FLAG_ON)) {
3504                         btrfs_warn(fs_info,
3505                         "qgroup rescan init failed, qgroup is not enabled");
3506                         ret = -EINVAL;
3507                 }
3508
3509                 if (ret)
3510                         return ret;
3511         }
3512
3513         mutex_lock(&fs_info->qgroup_rescan_lock);
3514
3515         if (init_flags) {
3516                 if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
3517                         btrfs_warn(fs_info,
3518                                    "qgroup rescan is already in progress");
3519                         ret = -EINPROGRESS;
3520                 } else if (!(fs_info->qgroup_flags &
3521                              BTRFS_QGROUP_STATUS_FLAG_ON)) {
3522                         btrfs_warn(fs_info,
3523                         "qgroup rescan init failed, qgroup is not enabled");
3524                         ret = -EINVAL;
3525                 } else if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags)) {
3526                         /* Quota disable is in progress */
3527                         ret = -EBUSY;
3528                 }
3529
3530                 if (ret) {
3531                         mutex_unlock(&fs_info->qgroup_rescan_lock);
3532                         return ret;
3533                 }
3534                 fs_info->qgroup_flags |= BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3535         }
3536
3537         memset(&fs_info->qgroup_rescan_progress, 0,
3538                 sizeof(fs_info->qgroup_rescan_progress));
3539         fs_info->qgroup_flags &= ~(BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN |
3540                                    BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING);
3541         fs_info->qgroup_rescan_progress.objectid = progress_objectid;
3542         init_completion(&fs_info->qgroup_rescan_completion);
3543         mutex_unlock(&fs_info->qgroup_rescan_lock);
3544
3545         btrfs_init_work(&fs_info->qgroup_rescan_work,
3546                         btrfs_qgroup_rescan_worker, NULL, NULL);
3547         return 0;
3548 }
3549
3550 static void
3551 qgroup_rescan_zero_tracking(struct btrfs_fs_info *fs_info)
3552 {
3553         struct rb_node *n;
3554         struct btrfs_qgroup *qgroup;
3555
3556         spin_lock(&fs_info->qgroup_lock);
3557         /* clear all current qgroup tracking information */
3558         for (n = rb_first(&fs_info->qgroup_tree); n; n = rb_next(n)) {
3559                 qgroup = rb_entry(n, struct btrfs_qgroup, node);
3560                 qgroup->rfer = 0;
3561                 qgroup->rfer_cmpr = 0;
3562                 qgroup->excl = 0;
3563                 qgroup->excl_cmpr = 0;
3564                 qgroup_dirty(fs_info, qgroup);
3565         }
3566         spin_unlock(&fs_info->qgroup_lock);
3567 }
3568
3569 int
3570 btrfs_qgroup_rescan(struct btrfs_fs_info *fs_info)
3571 {
3572         int ret = 0;
3573         struct btrfs_trans_handle *trans;
3574
3575         ret = qgroup_rescan_init(fs_info, 0, 1);
3576         if (ret)
3577                 return ret;
3578
3579         /*
3580          * We have set the rescan_progress to 0, which means no more
3581          * delayed refs will be accounted by btrfs_qgroup_account_ref.
3582          * However, btrfs_qgroup_account_ref may be right after its call
3583          * to btrfs_find_all_roots, in which case it would still do the
3584          * accounting.
3585          * To solve this, we're committing the transaction, which will
3586          * ensure we run all delayed refs and only after that, we are
3587          * going to clear all tracking information for a clean start.
3588          */
3589
3590         trans = btrfs_attach_transaction_barrier(fs_info->fs_root);
3591         if (IS_ERR(trans) && trans != ERR_PTR(-ENOENT)) {
3592                 fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3593                 return PTR_ERR(trans);
3594         } else if (trans != ERR_PTR(-ENOENT)) {
3595                 ret = btrfs_commit_transaction(trans);
3596                 if (ret) {
3597                         fs_info->qgroup_flags &= ~BTRFS_QGROUP_STATUS_FLAG_RESCAN;
3598                         return ret;
3599                 }
3600         }
3601
3602         qgroup_rescan_zero_tracking(fs_info);
3603
3604         mutex_lock(&fs_info->qgroup_rescan_lock);
3605         fs_info->qgroup_rescan_running = true;
3606         btrfs_queue_work(fs_info->qgroup_rescan_workers,
3607                          &fs_info->qgroup_rescan_work);
3608         mutex_unlock(&fs_info->qgroup_rescan_lock);
3609
3610         return 0;
3611 }
3612
3613 int btrfs_qgroup_wait_for_completion(struct btrfs_fs_info *fs_info,
3614                                      bool interruptible)
3615 {
3616         int running;
3617         int ret = 0;
3618
3619         mutex_lock(&fs_info->qgroup_rescan_lock);
3620         running = fs_info->qgroup_rescan_running;
3621         mutex_unlock(&fs_info->qgroup_rescan_lock);
3622
3623         if (!running)
3624                 return 0;
3625
3626         if (interruptible)
3627                 ret = wait_for_completion_interruptible(
3628                                         &fs_info->qgroup_rescan_completion);
3629         else
3630                 wait_for_completion(&fs_info->qgroup_rescan_completion);
3631
3632         return ret;
3633 }
3634
3635 /*
3636  * this is only called from open_ctree where we're still single threaded, thus
3637  * locking is omitted here.
3638  */
3639 void
3640 btrfs_qgroup_rescan_resume(struct btrfs_fs_info *fs_info)
3641 {
3642         if (fs_info->qgroup_flags & BTRFS_QGROUP_STATUS_FLAG_RESCAN) {
3643                 mutex_lock(&fs_info->qgroup_rescan_lock);
3644                 fs_info->qgroup_rescan_running = true;
3645                 btrfs_queue_work(fs_info->qgroup_rescan_workers,
3646                                  &fs_info->qgroup_rescan_work);
3647                 mutex_unlock(&fs_info->qgroup_rescan_lock);
3648         }
3649 }
3650
3651 #define rbtree_iterate_from_safe(node, next, start)                             \
3652        for (node = start; node && ({ next = rb_next(node); 1;}); node = next)
3653
3654 static int qgroup_unreserve_range(struct btrfs_inode *inode,
3655                                   struct extent_changeset *reserved, u64 start,
3656                                   u64 len)
3657 {
3658         struct rb_node *node;
3659         struct rb_node *next;
3660         struct ulist_node *entry;
3661         int ret = 0;
3662
3663         node = reserved->range_changed.root.rb_node;
3664         if (!node)
3665                 return 0;
3666         while (node) {
3667                 entry = rb_entry(node, struct ulist_node, rb_node);
3668                 if (entry->val < start)
3669                         node = node->rb_right;
3670                 else
3671                         node = node->rb_left;
3672         }
3673
3674         if (entry->val > start && rb_prev(&entry->rb_node))
3675                 entry = rb_entry(rb_prev(&entry->rb_node), struct ulist_node,
3676                                  rb_node);
3677
3678         rbtree_iterate_from_safe(node, next, &entry->rb_node) {
3679                 u64 entry_start;
3680                 u64 entry_end;
3681                 u64 entry_len;
3682                 int clear_ret;
3683
3684                 entry = rb_entry(node, struct ulist_node, rb_node);
3685                 entry_start = entry->val;
3686                 entry_end = entry->aux;
3687                 entry_len = entry_end - entry_start + 1;
3688
3689                 if (entry_start >= start + len)
3690                         break;
3691                 if (entry_start + entry_len <= start)
3692                         continue;
3693                 /*
3694                  * Now the entry is in [start, start + len), revert the
3695                  * EXTENT_QGROUP_RESERVED bit.
3696                  */
3697                 clear_ret = clear_extent_bits(&inode->io_tree, entry_start,
3698                                               entry_end, EXTENT_QGROUP_RESERVED);
3699                 if (!ret && clear_ret < 0)
3700                         ret = clear_ret;
3701
3702                 ulist_del(&reserved->range_changed, entry->val, entry->aux);
3703                 if (likely(reserved->bytes_changed >= entry_len)) {
3704                         reserved->bytes_changed -= entry_len;
3705                 } else {
3706                         WARN_ON(1);
3707                         reserved->bytes_changed = 0;
3708                 }
3709         }
3710
3711         return ret;
3712 }
3713
3714 /*
3715  * Try to free some space for qgroup.
3716  *
3717  * For qgroup, there are only 3 ways to free qgroup space:
3718  * - Flush nodatacow write
3719  *   Any nodatacow write will free its reserved data space at run_delalloc_range().
3720  *   In theory, we should only flush nodatacow inodes, but it's not yet
3721  *   possible, so we need to flush the whole root.
3722  *
3723  * - Wait for ordered extents
3724  *   When ordered extents are finished, their reserved metadata is finally
3725  *   converted to per_trans status, which can be freed by later commit
3726  *   transaction.
3727  *
3728  * - Commit transaction
3729  *   This would free the meta_per_trans space.
3730  *   In theory this shouldn't provide much space, but any more qgroup space
3731  *   is needed.
3732  */
3733 static int try_flush_qgroup(struct btrfs_root *root)
3734 {
3735         struct btrfs_trans_handle *trans;
3736         int ret;
3737
3738         /* Can't hold an open transaction or we run the risk of deadlocking. */
3739         ASSERT(current->journal_info == NULL);
3740         if (WARN_ON(current->journal_info))
3741                 return 0;
3742
3743         /*
3744          * We don't want to run flush again and again, so if there is a running
3745          * one, we won't try to start a new flush, but exit directly.
3746          */
3747         if (test_and_set_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state)) {
3748                 wait_event(root->qgroup_flush_wait,
3749                         !test_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state));
3750                 return 0;
3751         }
3752
3753         ret = btrfs_start_delalloc_snapshot(root, true);
3754         if (ret < 0)
3755                 goto out;
3756         btrfs_wait_ordered_extents(root, U64_MAX, 0, (u64)-1);
3757
3758         trans = btrfs_attach_transaction_barrier(root);
3759         if (IS_ERR(trans)) {
3760                 ret = PTR_ERR(trans);
3761                 if (ret == -ENOENT)
3762                         ret = 0;
3763                 goto out;
3764         }
3765
3766         ret = btrfs_commit_transaction(trans);
3767 out:
3768         clear_bit(BTRFS_ROOT_QGROUP_FLUSHING, &root->state);
3769         wake_up(&root->qgroup_flush_wait);
3770         return ret;
3771 }
3772
3773 static int qgroup_reserve_data(struct btrfs_inode *inode,
3774                         struct extent_changeset **reserved_ret, u64 start,
3775                         u64 len)
3776 {
3777         struct btrfs_root *root = inode->root;
3778         struct extent_changeset *reserved;
3779         bool new_reserved = false;
3780         u64 orig_reserved;
3781         u64 to_reserve;
3782         int ret;
3783
3784         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &root->fs_info->flags) ||
3785             !is_fstree(root->root_key.objectid) || len == 0)
3786                 return 0;
3787
3788         /* @reserved parameter is mandatory for qgroup */
3789         if (WARN_ON(!reserved_ret))
3790                 return -EINVAL;
3791         if (!*reserved_ret) {
3792                 new_reserved = true;
3793                 *reserved_ret = extent_changeset_alloc();
3794                 if (!*reserved_ret)
3795                         return -ENOMEM;
3796         }
3797         reserved = *reserved_ret;
3798         /* Record already reserved space */
3799         orig_reserved = reserved->bytes_changed;
3800         ret = set_record_extent_bits(&inode->io_tree, start,
3801                         start + len -1, EXTENT_QGROUP_RESERVED, reserved);
3802
3803         /* Newly reserved space */
3804         to_reserve = reserved->bytes_changed - orig_reserved;
3805         trace_btrfs_qgroup_reserve_data(&inode->vfs_inode, start, len,
3806                                         to_reserve, QGROUP_RESERVE);
3807         if (ret < 0)
3808                 goto out;
3809         ret = qgroup_reserve(root, to_reserve, true, BTRFS_QGROUP_RSV_DATA);
3810         if (ret < 0)
3811                 goto cleanup;
3812
3813         return ret;
3814
3815 cleanup:
3816         qgroup_unreserve_range(inode, reserved, start, len);
3817 out:
3818         if (new_reserved) {
3819                 extent_changeset_free(reserved);
3820                 *reserved_ret = NULL;
3821         }
3822         return ret;
3823 }
3824
3825 /*
3826  * Reserve qgroup space for range [start, start + len).
3827  *
3828  * This function will either reserve space from related qgroups or do nothing
3829  * if the range is already reserved.
3830  *
3831  * Return 0 for successful reservation
3832  * Return <0 for error (including -EQUOT)
3833  *
3834  * NOTE: This function may sleep for memory allocation, dirty page flushing and
3835  *       commit transaction. So caller should not hold any dirty page locked.
3836  */
3837 int btrfs_qgroup_reserve_data(struct btrfs_inode *inode,
3838                         struct extent_changeset **reserved_ret, u64 start,
3839                         u64 len)
3840 {
3841         int ret;
3842
3843         ret = qgroup_reserve_data(inode, reserved_ret, start, len);
3844         if (ret <= 0 && ret != -EDQUOT)
3845                 return ret;
3846
3847         ret = try_flush_qgroup(inode->root);
3848         if (ret < 0)
3849                 return ret;
3850         return qgroup_reserve_data(inode, reserved_ret, start, len);
3851 }
3852
3853 /* Free ranges specified by @reserved, normally in error path */
3854 static int qgroup_free_reserved_data(struct btrfs_inode *inode,
3855                                      struct extent_changeset *reserved,
3856                                      u64 start, u64 len, u64 *freed_ret)
3857 {
3858         struct btrfs_root *root = inode->root;
3859         struct ulist_node *unode;
3860         struct ulist_iterator uiter;
3861         struct extent_changeset changeset;
3862         u64 freed = 0;
3863         int ret;
3864
3865         extent_changeset_init(&changeset);
3866         len = round_up(start + len, root->fs_info->sectorsize);
3867         start = round_down(start, root->fs_info->sectorsize);
3868
3869         ULIST_ITER_INIT(&uiter);
3870         while ((unode = ulist_next(&reserved->range_changed, &uiter))) {
3871                 u64 range_start = unode->val;
3872                 /* unode->aux is the inclusive end */
3873                 u64 range_len = unode->aux - range_start + 1;
3874                 u64 free_start;
3875                 u64 free_len;
3876
3877                 extent_changeset_release(&changeset);
3878
3879                 /* Only free range in range [start, start + len) */
3880                 if (range_start >= start + len ||
3881                     range_start + range_len <= start)
3882                         continue;
3883                 free_start = max(range_start, start);
3884                 free_len = min(start + len, range_start + range_len) -
3885                            free_start;
3886                 /*
3887                  * TODO: To also modify reserved->ranges_reserved to reflect
3888                  * the modification.
3889                  *
3890                  * However as long as we free qgroup reserved according to
3891                  * EXTENT_QGROUP_RESERVED, we won't double free.
3892                  * So not need to rush.
3893                  */
3894                 ret = clear_record_extent_bits(&inode->io_tree, free_start,
3895                                 free_start + free_len - 1,
3896                                 EXTENT_QGROUP_RESERVED, &changeset);
3897                 if (ret < 0)
3898                         goto out;
3899                 freed += changeset.bytes_changed;
3900         }
3901         btrfs_qgroup_free_refroot(root->fs_info, root->root_key.objectid, freed,
3902                                   BTRFS_QGROUP_RSV_DATA);
3903         if (freed_ret)
3904                 *freed_ret = freed;
3905         ret = 0;
3906 out:
3907         extent_changeset_release(&changeset);
3908         return ret;
3909 }
3910
3911 static int __btrfs_qgroup_release_data(struct btrfs_inode *inode,
3912                         struct extent_changeset *reserved, u64 start, u64 len,
3913                         u64 *released, int free)
3914 {
3915         struct extent_changeset changeset;
3916         int trace_op = QGROUP_RELEASE;
3917         int ret;
3918
3919         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &inode->root->fs_info->flags))
3920                 return 0;
3921
3922         /* In release case, we shouldn't have @reserved */
3923         WARN_ON(!free && reserved);
3924         if (free && reserved)
3925                 return qgroup_free_reserved_data(inode, reserved, start, len, released);
3926         extent_changeset_init(&changeset);
3927         ret = clear_record_extent_bits(&inode->io_tree, start, start + len -1,
3928                                        EXTENT_QGROUP_RESERVED, &changeset);
3929         if (ret < 0)
3930                 goto out;
3931
3932         if (free)
3933                 trace_op = QGROUP_FREE;
3934         trace_btrfs_qgroup_release_data(&inode->vfs_inode, start, len,
3935                                         changeset.bytes_changed, trace_op);
3936         if (free)
3937                 btrfs_qgroup_free_refroot(inode->root->fs_info,
3938                                 inode->root->root_key.objectid,
3939                                 changeset.bytes_changed, BTRFS_QGROUP_RSV_DATA);
3940         if (released)
3941                 *released = changeset.bytes_changed;
3942 out:
3943         extent_changeset_release(&changeset);
3944         return ret;
3945 }
3946
3947 /*
3948  * Free a reserved space range from io_tree and related qgroups
3949  *
3950  * Should be called when a range of pages get invalidated before reaching disk.
3951  * Or for error cleanup case.
3952  * if @reserved is given, only reserved range in [@start, @start + @len) will
3953  * be freed.
3954  *
3955  * For data written to disk, use btrfs_qgroup_release_data().
3956  *
3957  * NOTE: This function may sleep for memory allocation.
3958  */
3959 int btrfs_qgroup_free_data(struct btrfs_inode *inode,
3960                            struct extent_changeset *reserved,
3961                            u64 start, u64 len, u64 *freed)
3962 {
3963         return __btrfs_qgroup_release_data(inode, reserved, start, len, freed, 1);
3964 }
3965
3966 /*
3967  * Release a reserved space range from io_tree only.
3968  *
3969  * Should be called when a range of pages get written to disk and corresponding
3970  * FILE_EXTENT is inserted into corresponding root.
3971  *
3972  * Since new qgroup accounting framework will only update qgroup numbers at
3973  * commit_transaction() time, its reserved space shouldn't be freed from
3974  * related qgroups.
3975  *
3976  * But we should release the range from io_tree, to allow further write to be
3977  * COWed.
3978  *
3979  * NOTE: This function may sleep for memory allocation.
3980  */
3981 int btrfs_qgroup_release_data(struct btrfs_inode *inode, u64 start, u64 len, u64 *released)
3982 {
3983         return __btrfs_qgroup_release_data(inode, NULL, start, len, released, 0);
3984 }
3985
3986 static void add_root_meta_rsv(struct btrfs_root *root, int num_bytes,
3987                               enum btrfs_qgroup_rsv_type type)
3988 {
3989         if (type != BTRFS_QGROUP_RSV_META_PREALLOC &&
3990             type != BTRFS_QGROUP_RSV_META_PERTRANS)
3991                 return;
3992         if (num_bytes == 0)
3993                 return;
3994
3995         spin_lock(&root->qgroup_meta_rsv_lock);
3996         if (type == BTRFS_QGROUP_RSV_META_PREALLOC)
3997                 root->qgroup_meta_rsv_prealloc += num_bytes;
3998         else
3999                 root->qgroup_meta_rsv_pertrans += num_bytes;
4000         spin_unlock(&root->qgroup_meta_rsv_lock);
4001 }
4002
4003 static int sub_root_meta_rsv(struct btrfs_root *root, int num_bytes,
4004                              enum btrfs_qgroup_rsv_type type)
4005 {
4006         if (type != BTRFS_QGROUP_RSV_META_PREALLOC &&
4007             type != BTRFS_QGROUP_RSV_META_PERTRANS)
4008                 return 0;
4009         if (num_bytes == 0)
4010                 return 0;
4011
4012         spin_lock(&root->qgroup_meta_rsv_lock);
4013         if (type == BTRFS_QGROUP_RSV_META_PREALLOC) {
4014                 num_bytes = min_t(u64, root->qgroup_meta_rsv_prealloc,
4015                                   num_bytes);
4016                 root->qgroup_meta_rsv_prealloc -= num_bytes;
4017         } else {
4018                 num_bytes = min_t(u64, root->qgroup_meta_rsv_pertrans,
4019                                   num_bytes);
4020                 root->qgroup_meta_rsv_pertrans -= num_bytes;
4021         }
4022         spin_unlock(&root->qgroup_meta_rsv_lock);
4023         return num_bytes;
4024 }
4025
4026 int btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
4027                               enum btrfs_qgroup_rsv_type type, bool enforce)
4028 {
4029         struct btrfs_fs_info *fs_info = root->fs_info;
4030         int ret;
4031
4032         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
4033             !is_fstree(root->root_key.objectid) || num_bytes == 0)
4034                 return 0;
4035
4036         BUG_ON(num_bytes != round_down(num_bytes, fs_info->nodesize));
4037         trace_qgroup_meta_reserve(root, (s64)num_bytes, type);
4038         ret = qgroup_reserve(root, num_bytes, enforce, type);
4039         if (ret < 0)
4040                 return ret;
4041         /*
4042          * Record what we have reserved into root.
4043          *
4044          * To avoid quota disabled->enabled underflow.
4045          * In that case, we may try to free space we haven't reserved
4046          * (since quota was disabled), so record what we reserved into root.
4047          * And ensure later release won't underflow this number.
4048          */
4049         add_root_meta_rsv(root, num_bytes, type);
4050         return ret;
4051 }
4052
4053 int __btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
4054                                 enum btrfs_qgroup_rsv_type type, bool enforce,
4055                                 bool noflush)
4056 {
4057         int ret;
4058
4059         ret = btrfs_qgroup_reserve_meta(root, num_bytes, type, enforce);
4060         if ((ret <= 0 && ret != -EDQUOT) || noflush)
4061                 return ret;
4062
4063         ret = try_flush_qgroup(root);
4064         if (ret < 0)
4065                 return ret;
4066         return btrfs_qgroup_reserve_meta(root, num_bytes, type, enforce);
4067 }
4068
4069 void btrfs_qgroup_free_meta_all_pertrans(struct btrfs_root *root)
4070 {
4071         struct btrfs_fs_info *fs_info = root->fs_info;
4072
4073         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
4074             !is_fstree(root->root_key.objectid))
4075                 return;
4076
4077         /* TODO: Update trace point to handle such free */
4078         trace_qgroup_meta_free_all_pertrans(root);
4079         /* Special value -1 means to free all reserved space */
4080         btrfs_qgroup_free_refroot(fs_info, root->root_key.objectid, (u64)-1,
4081                                   BTRFS_QGROUP_RSV_META_PERTRANS);
4082 }
4083
4084 void __btrfs_qgroup_free_meta(struct btrfs_root *root, int num_bytes,
4085                               enum btrfs_qgroup_rsv_type type)
4086 {
4087         struct btrfs_fs_info *fs_info = root->fs_info;
4088
4089         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
4090             !is_fstree(root->root_key.objectid))
4091                 return;
4092
4093         /*
4094          * reservation for META_PREALLOC can happen before quota is enabled,
4095          * which can lead to underflow.
4096          * Here ensure we will only free what we really have reserved.
4097          */
4098         num_bytes = sub_root_meta_rsv(root, num_bytes, type);
4099         BUG_ON(num_bytes != round_down(num_bytes, fs_info->nodesize));
4100         trace_qgroup_meta_reserve(root, -(s64)num_bytes, type);
4101         btrfs_qgroup_free_refroot(fs_info, root->root_key.objectid,
4102                                   num_bytes, type);
4103 }
4104
4105 static void qgroup_convert_meta(struct btrfs_fs_info *fs_info, u64 ref_root,
4106                                 int num_bytes)
4107 {
4108         struct btrfs_qgroup *qgroup;
4109         LIST_HEAD(qgroup_list);
4110
4111         if (num_bytes == 0)
4112                 return;
4113         if (!fs_info->quota_root)
4114                 return;
4115
4116         spin_lock(&fs_info->qgroup_lock);
4117         qgroup = find_qgroup_rb(fs_info, ref_root);
4118         if (!qgroup)
4119                 goto out;
4120
4121         qgroup_iterator_add(&qgroup_list, qgroup);
4122         list_for_each_entry(qgroup, &qgroup_list, iterator) {
4123                 struct btrfs_qgroup_list *glist;
4124
4125                 qgroup_rsv_release(fs_info, qgroup, num_bytes,
4126                                 BTRFS_QGROUP_RSV_META_PREALLOC);
4127                 if (!sb_rdonly(fs_info->sb))
4128                         qgroup_rsv_add(fs_info, qgroup, num_bytes,
4129                                        BTRFS_QGROUP_RSV_META_PERTRANS);
4130
4131                 list_for_each_entry(glist, &qgroup->groups, next_group)
4132                         qgroup_iterator_add(&qgroup_list, glist->group);
4133         }
4134 out:
4135         qgroup_iterator_clean(&qgroup_list);
4136         spin_unlock(&fs_info->qgroup_lock);
4137 }
4138
4139 void btrfs_qgroup_convert_reserved_meta(struct btrfs_root *root, int num_bytes)
4140 {
4141         struct btrfs_fs_info *fs_info = root->fs_info;
4142
4143         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) ||
4144             !is_fstree(root->root_key.objectid))
4145                 return;
4146         /* Same as btrfs_qgroup_free_meta_prealloc() */
4147         num_bytes = sub_root_meta_rsv(root, num_bytes,
4148                                       BTRFS_QGROUP_RSV_META_PREALLOC);
4149         trace_qgroup_meta_convert(root, num_bytes);
4150         qgroup_convert_meta(fs_info, root->root_key.objectid, num_bytes);
4151 }
4152
4153 /*
4154  * Check qgroup reserved space leaking, normally at destroy inode
4155  * time
4156  */
4157 void btrfs_qgroup_check_reserved_leak(struct btrfs_inode *inode)
4158 {
4159         struct extent_changeset changeset;
4160         struct ulist_node *unode;
4161         struct ulist_iterator iter;
4162         int ret;
4163
4164         extent_changeset_init(&changeset);
4165         ret = clear_record_extent_bits(&inode->io_tree, 0, (u64)-1,
4166                         EXTENT_QGROUP_RESERVED, &changeset);
4167
4168         WARN_ON(ret < 0);
4169         if (WARN_ON(changeset.bytes_changed)) {
4170                 ULIST_ITER_INIT(&iter);
4171                 while ((unode = ulist_next(&changeset.range_changed, &iter))) {
4172                         btrfs_warn(inode->root->fs_info,
4173                 "leaking qgroup reserved space, ino: %llu, start: %llu, end: %llu",
4174                                 btrfs_ino(inode), unode->val, unode->aux);
4175                 }
4176                 btrfs_qgroup_free_refroot(inode->root->fs_info,
4177                                 inode->root->root_key.objectid,
4178                                 changeset.bytes_changed, BTRFS_QGROUP_RSV_DATA);
4179
4180         }
4181         extent_changeset_release(&changeset);
4182 }
4183
4184 void btrfs_qgroup_init_swapped_blocks(
4185         struct btrfs_qgroup_swapped_blocks *swapped_blocks)
4186 {
4187         int i;
4188
4189         spin_lock_init(&swapped_blocks->lock);
4190         for (i = 0; i < BTRFS_MAX_LEVEL; i++)
4191                 swapped_blocks->blocks[i] = RB_ROOT;
4192         swapped_blocks->swapped = false;
4193 }
4194
4195 /*
4196  * Delete all swapped blocks record of @root.
4197  * Every record here means we skipped a full subtree scan for qgroup.
4198  *
4199  * Gets called when committing one transaction.
4200  */
4201 void btrfs_qgroup_clean_swapped_blocks(struct btrfs_root *root)
4202 {
4203         struct btrfs_qgroup_swapped_blocks *swapped_blocks;
4204         int i;
4205
4206         swapped_blocks = &root->swapped_blocks;
4207
4208         spin_lock(&swapped_blocks->lock);
4209         if (!swapped_blocks->swapped)
4210                 goto out;
4211         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
4212                 struct rb_root *cur_root = &swapped_blocks->blocks[i];
4213                 struct btrfs_qgroup_swapped_block *entry;
4214                 struct btrfs_qgroup_swapped_block *next;
4215
4216                 rbtree_postorder_for_each_entry_safe(entry, next, cur_root,
4217                                                      node)
4218                         kfree(entry);
4219                 swapped_blocks->blocks[i] = RB_ROOT;
4220         }
4221         swapped_blocks->swapped = false;
4222 out:
4223         spin_unlock(&swapped_blocks->lock);
4224 }
4225
4226 /*
4227  * Add subtree roots record into @subvol_root.
4228  *
4229  * @subvol_root:        tree root of the subvolume tree get swapped
4230  * @bg:                 block group under balance
4231  * @subvol_parent/slot: pointer to the subtree root in subvolume tree
4232  * @reloc_parent/slot:  pointer to the subtree root in reloc tree
4233  *                      BOTH POINTERS ARE BEFORE TREE SWAP
4234  * @last_snapshot:      last snapshot generation of the subvolume tree
4235  */
4236 int btrfs_qgroup_add_swapped_blocks(struct btrfs_trans_handle *trans,
4237                 struct btrfs_root *subvol_root,
4238                 struct btrfs_block_group *bg,
4239                 struct extent_buffer *subvol_parent, int subvol_slot,
4240                 struct extent_buffer *reloc_parent, int reloc_slot,
4241                 u64 last_snapshot)
4242 {
4243         struct btrfs_fs_info *fs_info = subvol_root->fs_info;
4244         struct btrfs_qgroup_swapped_blocks *blocks = &subvol_root->swapped_blocks;
4245         struct btrfs_qgroup_swapped_block *block;
4246         struct rb_node **cur;
4247         struct rb_node *parent = NULL;
4248         int level = btrfs_header_level(subvol_parent) - 1;
4249         int ret = 0;
4250
4251         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
4252                 return 0;
4253
4254         if (btrfs_node_ptr_generation(subvol_parent, subvol_slot) >
4255             btrfs_node_ptr_generation(reloc_parent, reloc_slot)) {
4256                 btrfs_err_rl(fs_info,
4257                 "%s: bad parameter order, subvol_gen=%llu reloc_gen=%llu",
4258                         __func__,
4259                         btrfs_node_ptr_generation(subvol_parent, subvol_slot),
4260                         btrfs_node_ptr_generation(reloc_parent, reloc_slot));
4261                 return -EUCLEAN;
4262         }
4263
4264         block = kmalloc(sizeof(*block), GFP_NOFS);
4265         if (!block) {
4266                 ret = -ENOMEM;
4267                 goto out;
4268         }
4269
4270         /*
4271          * @reloc_parent/slot is still before swap, while @block is going to
4272          * record the bytenr after swap, so we do the swap here.
4273          */
4274         block->subvol_bytenr = btrfs_node_blockptr(reloc_parent, reloc_slot);
4275         block->subvol_generation = btrfs_node_ptr_generation(reloc_parent,
4276                                                              reloc_slot);
4277         block->reloc_bytenr = btrfs_node_blockptr(subvol_parent, subvol_slot);
4278         block->reloc_generation = btrfs_node_ptr_generation(subvol_parent,
4279                                                             subvol_slot);
4280         block->last_snapshot = last_snapshot;
4281         block->level = level;
4282
4283         /*
4284          * If we have bg == NULL, we're called from btrfs_recover_relocation(),
4285          * no one else can modify tree blocks thus we qgroup will not change
4286          * no matter the value of trace_leaf.
4287          */
4288         if (bg && bg->flags & BTRFS_BLOCK_GROUP_DATA)
4289                 block->trace_leaf = true;
4290         else
4291                 block->trace_leaf = false;
4292         btrfs_node_key_to_cpu(reloc_parent, &block->first_key, reloc_slot);
4293
4294         /* Insert @block into @blocks */
4295         spin_lock(&blocks->lock);
4296         cur = &blocks->blocks[level].rb_node;
4297         while (*cur) {
4298                 struct btrfs_qgroup_swapped_block *entry;
4299
4300                 parent = *cur;
4301                 entry = rb_entry(parent, struct btrfs_qgroup_swapped_block,
4302                                  node);
4303
4304                 if (entry->subvol_bytenr < block->subvol_bytenr) {
4305                         cur = &(*cur)->rb_left;
4306                 } else if (entry->subvol_bytenr > block->subvol_bytenr) {
4307                         cur = &(*cur)->rb_right;
4308                 } else {
4309                         if (entry->subvol_generation !=
4310                                         block->subvol_generation ||
4311                             entry->reloc_bytenr != block->reloc_bytenr ||
4312                             entry->reloc_generation !=
4313                                         block->reloc_generation) {
4314                                 /*
4315                                  * Duplicated but mismatch entry found.
4316                                  * Shouldn't happen.
4317                                  *
4318                                  * Marking qgroup inconsistent should be enough
4319                                  * for end users.
4320                                  */
4321                                 WARN_ON(IS_ENABLED(CONFIG_BTRFS_DEBUG));
4322                                 ret = -EEXIST;
4323                         }
4324                         kfree(block);
4325                         goto out_unlock;
4326                 }
4327         }
4328         rb_link_node(&block->node, parent, cur);
4329         rb_insert_color(&block->node, &blocks->blocks[level]);
4330         blocks->swapped = true;
4331 out_unlock:
4332         spin_unlock(&blocks->lock);
4333 out:
4334         if (ret < 0)
4335                 qgroup_mark_inconsistent(fs_info);
4336         return ret;
4337 }
4338
4339 /*
4340  * Check if the tree block is a subtree root, and if so do the needed
4341  * delayed subtree trace for qgroup.
4342  *
4343  * This is called during btrfs_cow_block().
4344  */
4345 int btrfs_qgroup_trace_subtree_after_cow(struct btrfs_trans_handle *trans,
4346                                          struct btrfs_root *root,
4347                                          struct extent_buffer *subvol_eb)
4348 {
4349         struct btrfs_fs_info *fs_info = root->fs_info;
4350         struct btrfs_tree_parent_check check = { 0 };
4351         struct btrfs_qgroup_swapped_blocks *blocks = &root->swapped_blocks;
4352         struct btrfs_qgroup_swapped_block *block;
4353         struct extent_buffer *reloc_eb = NULL;
4354         struct rb_node *node;
4355         bool found = false;
4356         bool swapped = false;
4357         int level = btrfs_header_level(subvol_eb);
4358         int ret = 0;
4359         int i;
4360
4361         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags))
4362                 return 0;
4363         if (!is_fstree(root->root_key.objectid) || !root->reloc_root)
4364                 return 0;
4365
4366         spin_lock(&blocks->lock);
4367         if (!blocks->swapped) {
4368                 spin_unlock(&blocks->lock);
4369                 return 0;
4370         }
4371         node = blocks->blocks[level].rb_node;
4372
4373         while (node) {
4374                 block = rb_entry(node, struct btrfs_qgroup_swapped_block, node);
4375                 if (block->subvol_bytenr < subvol_eb->start) {
4376                         node = node->rb_left;
4377                 } else if (block->subvol_bytenr > subvol_eb->start) {
4378                         node = node->rb_right;
4379                 } else {
4380                         found = true;
4381                         break;
4382                 }
4383         }
4384         if (!found) {
4385                 spin_unlock(&blocks->lock);
4386                 goto out;
4387         }
4388         /* Found one, remove it from @blocks first and update blocks->swapped */
4389         rb_erase(&block->node, &blocks->blocks[level]);
4390         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
4391                 if (RB_EMPTY_ROOT(&blocks->blocks[i])) {
4392                         swapped = true;
4393                         break;
4394                 }
4395         }
4396         blocks->swapped = swapped;
4397         spin_unlock(&blocks->lock);
4398
4399         check.level = block->level;
4400         check.transid = block->reloc_generation;
4401         check.has_first_key = true;
4402         memcpy(&check.first_key, &block->first_key, sizeof(check.first_key));
4403
4404         /* Read out reloc subtree root */
4405         reloc_eb = read_tree_block(fs_info, block->reloc_bytenr, &check);
4406         if (IS_ERR(reloc_eb)) {
4407                 ret = PTR_ERR(reloc_eb);
4408                 reloc_eb = NULL;
4409                 goto free_out;
4410         }
4411         if (!extent_buffer_uptodate(reloc_eb)) {
4412                 ret = -EIO;
4413                 goto free_out;
4414         }
4415
4416         ret = qgroup_trace_subtree_swap(trans, reloc_eb, subvol_eb,
4417                         block->last_snapshot, block->trace_leaf);
4418 free_out:
4419         kfree(block);
4420         free_extent_buffer(reloc_eb);
4421 out:
4422         if (ret < 0) {
4423                 btrfs_err_rl(fs_info,
4424                              "failed to account subtree at bytenr %llu: %d",
4425                              subvol_eb->start, ret);
4426                 qgroup_mark_inconsistent(fs_info);
4427         }
4428         return ret;
4429 }
4430
4431 void btrfs_qgroup_destroy_extent_records(struct btrfs_transaction *trans)
4432 {
4433         struct btrfs_qgroup_extent_record *entry;
4434         struct btrfs_qgroup_extent_record *next;
4435         struct rb_root *root;
4436
4437         root = &trans->delayed_refs.dirty_extent_root;
4438         rbtree_postorder_for_each_entry_safe(entry, next, root, node) {
4439                 ulist_free(entry->old_roots);
4440                 kfree(entry);
4441         }
4442         *root = RB_ROOT;
4443 }