fs/btrfs/qgroup.c

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