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