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