1 /* SPDX-License-Identifier: GPL-2.0 */
2 /*
3 * Copyright (C) 2014 Facebook. All rights reserved.
4 */
5
6 #ifndef BTRFS_QGROUP_H
7 #define BTRFS_QGROUP_H
8
9 #include <linux/spinlock.h>
10 #include <linux/rbtree.h>
11 #include <linux/kobject.h>
12 #include "ulist.h"
13 #include "delayed-ref.h"
14 #include "misc.h"
15
16 /*
17 * Btrfs qgroup overview
18 *
19 * Btrfs qgroup splits into 3 main part:
20 * 1) Reserve
21 * Reserve metadata/data space for incoming operations
22 * Affect how qgroup limit works
23 *
24 * 2) Trace
25 * Tell btrfs qgroup to trace dirty extents.
26 *
27 * Dirty extents including:
28 * - Newly allocated extents
29 * - Extents going to be deleted (in this trans)
30 * - Extents whose owner is going to be modified
31 *
32 * This is the main part affects whether qgroup numbers will stay
33 * consistent.
34 * Btrfs qgroup can trace clean extents and won't cause any problem,
35 * but it will consume extra CPU time, it should be avoided if possible.
36 *
37 * 3) Account
38 * Btrfs qgroup will updates its numbers, based on dirty extents traced
39 * in previous step.
40 *
41 * Normally at qgroup rescan and transaction commit time.
42 */
43
44 /*
45 * Special performance optimization for balance.
46 *
47 * For balance, we need to swap subtree of subvolume and reloc trees.
48 * In theory, we need to trace all subtree blocks of both subvolume and reloc
49 * trees, since their owner has changed during such swap.
50 *
51 * However since balance has ensured that both subtrees are containing the
52 * same contents and have the same tree structures, such swap won't cause
53 * qgroup number change.
54 *
55 * But there is a race window between subtree swap and transaction commit,
56 * during that window, if we increase/decrease tree level or merge/split tree
57 * blocks, we still need to trace the original subtrees.
58 *
59 * So for balance, we use a delayed subtree tracing, whose workflow is:
60 *
61 * 1) Record the subtree root block get swapped.
62 *
63 * During subtree swap:
64 * O = Old tree blocks
65 * N = New tree blocks
66 * reloc tree subvolume tree X
67 * Root Root
68 * / \ / \
69 * NA OB OA OB
70 * / | | \ / | | \
71 * NC ND OE OF OC OD OE OF
72 *
73 * In this case, NA and OA are going to be swapped, record (NA, OA) into
74 * subvolume tree X.
75 *
76 * 2) After subtree swap.
77 * reloc tree subvolume tree X
78 * Root Root
79 * / \ / \
80 * OA OB NA OB
81 * / | | \ / | | \
82 * OC OD OE OF NC ND OE OF
83 *
84 * 3a) COW happens for OB
85 * If we are going to COW tree block OB, we check OB's bytenr against
86 * tree X's swapped_blocks structure.
87 * If it doesn't fit any, nothing will happen.
88 *
89 * 3b) COW happens for NA
90 * Check NA's bytenr against tree X's swapped_blocks, and get a hit.
91 * Then we do subtree scan on both subtrees OA and NA.
92 * Resulting 6 tree blocks to be scanned (OA, OC, OD, NA, NC, ND).
93 *
94 * Then no matter what we do to subvolume tree X, qgroup numbers will
95 * still be correct.
96 * Then NA's record gets removed from X's swapped_blocks.
97 *
98 * 4) Transaction commit
99 * Any record in X's swapped_blocks gets removed, since there is no
100 * modification to the swapped subtrees, no need to trigger heavy qgroup
101 * subtree rescan for them.
102 */
103
104 #define BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN (1UL << 3)
105 #define BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING (1UL << 4)
106
107 /*
108 * Record a dirty extent, and info qgroup to update quota on it
109 * TODO: Use kmem cache to alloc it.
110 */
111 struct btrfs_qgroup_extent_record {
112 struct rb_node node;
113 u64 bytenr;
114 u64 num_bytes;
115
116 /*
117 * For qgroup reserved data space freeing.
118 *
119 * @data_rsv_refroot and @data_rsv will be recorded after
120 * BTRFS_ADD_DELAYED_EXTENT is called.
121 * And will be used to free reserved qgroup space at
122 * transaction commit time.
123 */
124 u32 data_rsv; /* reserved data space needs to be freed */
125 u64 data_rsv_refroot; /* which root the reserved data belongs to */
126 struct ulist *old_roots;
127 };
128
129 struct btrfs_qgroup_swapped_block {
130 struct rb_node node;
131
132 int level;
133 bool trace_leaf;
134
135 /* bytenr/generation of the tree block in subvolume tree after swap */
136 u64 subvol_bytenr;
137 u64 subvol_generation;
138
139 /* bytenr/generation of the tree block in reloc tree after swap */
140 u64 reloc_bytenr;
141 u64 reloc_generation;
142
143 u64 last_snapshot;
144 struct btrfs_key first_key;
145 };
146
147 /*
148 * Qgroup reservation types:
149 *
150 * DATA:
151 * space reserved for data
152 *
153 * META_PERTRANS:
154 * Space reserved for metadata (per-transaction)
155 * Due to the fact that qgroup data is only updated at transaction commit
156 * time, reserved space for metadata must be kept until transaction
157 * commits.
158 * Any metadata reserved that are used in btrfs_start_transaction() should
159 * be of this type.
160 *
161 * META_PREALLOC:
162 * There are cases where metadata space is reserved before starting
163 * transaction, and then btrfs_join_transaction() to get a trans handle.
164 * Any metadata reserved for such usage should be of this type.
165 * And after join_transaction() part (or all) of such reservation should
166 * be converted into META_PERTRANS.
167 */
168 enum btrfs_qgroup_rsv_type {
169 BTRFS_QGROUP_RSV_DATA,
170 BTRFS_QGROUP_RSV_META_PERTRANS,
171 BTRFS_QGROUP_RSV_META_PREALLOC,
172 BTRFS_QGROUP_RSV_LAST,
173 };
174
175 /*
176 * Represents how many bytes we have reserved for this qgroup.
177 *
178 * Each type should have different reservation behavior.
179 * E.g, data follows its io_tree flag modification, while
180 * *currently* meta is just reserve-and-clear during transaction.
181 *
182 * TODO: Add new type for reservation which can survive transaction commit.
183 * Current metadata reservation behavior is not suitable for such case.
184 */
185 struct btrfs_qgroup_rsv {
186 u64 values[BTRFS_QGROUP_RSV_LAST];
187 };
188
189 /*
190 * one struct for each qgroup, organized in fs_info->qgroup_tree.
191 */
192 struct btrfs_qgroup {
193 u64 qgroupid;
194
195 /*
196 * state
197 */
198 u64 rfer; /* referenced */
199 u64 rfer_cmpr; /* referenced compressed */
200 u64 excl; /* exclusive */
201 u64 excl_cmpr; /* exclusive compressed */
202
203 /*
204 * limits
205 */
206 u64 lim_flags; /* which limits are set */
207 u64 max_rfer;
208 u64 max_excl;
209 u64 rsv_rfer;
210 u64 rsv_excl;
211
212 /*
213 * reservation tracking
214 */
215 struct btrfs_qgroup_rsv rsv;
216
217 /*
218 * lists
219 */
220 struct list_head groups; /* groups this group is member of */
221 struct list_head members; /* groups that are members of this group */
222 struct list_head dirty; /* dirty groups */
223
224 /*
225 * For qgroup iteration usage.
226 *
227 * The iteration list should always be empty until qgroup_iterator_add()
228 * is called. And should be reset to empty after the iteration is
229 * finished.
230 */
231 struct list_head iterator;
232 struct rb_node node; /* tree of qgroups */
233
234 /*
235 * temp variables for accounting operations
236 * Refer to qgroup_shared_accounting() for details.
237 */
238 u64 old_refcnt;
239 u64 new_refcnt;
240
241 /*
242 * Sysfs kobjectid
243 */
244 struct kobject kobj;
245 };
246
btrfs_qgroup_subvolid(u64 qgroupid)247 static inline u64 btrfs_qgroup_subvolid(u64 qgroupid)
248 {
249 return (qgroupid & ((1ULL << BTRFS_QGROUP_LEVEL_SHIFT) - 1));
250 }
251
252 /*
253 * For qgroup event trace points only
254 */
255 enum {
256 ENUM_BIT(QGROUP_RESERVE),
257 ENUM_BIT(QGROUP_RELEASE),
258 ENUM_BIT(QGROUP_FREE),
259 };
260
261 int btrfs_quota_enable(struct btrfs_fs_info *fs_info);
262 int btrfs_quota_disable(struct btrfs_fs_info *fs_info);
263 int btrfs_qgroup_rescan(struct btrfs_fs_info *fs_info);
264 void btrfs_qgroup_rescan_resume(struct btrfs_fs_info *fs_info);
265 int btrfs_qgroup_wait_for_completion(struct btrfs_fs_info *fs_info,
266 bool interruptible);
267 int btrfs_add_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
268 u64 dst);
269 int btrfs_del_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
270 u64 dst);
271 int btrfs_create_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid);
272 int btrfs_remove_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid);
273 int btrfs_limit_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid,
274 struct btrfs_qgroup_limit *limit);
275 int btrfs_read_qgroup_config(struct btrfs_fs_info *fs_info);
276 void btrfs_free_qgroup_config(struct btrfs_fs_info *fs_info);
277 struct btrfs_delayed_extent_op;
278
279 /*
280 * Inform qgroup to trace one dirty extent, its info is recorded in @record.
281 * So qgroup can account it at transaction committing time.
282 *
283 * No lock version, caller must acquire delayed ref lock and allocated memory,
284 * then call btrfs_qgroup_trace_extent_post() after exiting lock context.
285 *
286 * Return 0 for success insert
287 * Return >0 for existing record, caller can free @record safely.
288 * Error is not possible
289 */
290 int btrfs_qgroup_trace_extent_nolock(
291 struct btrfs_fs_info *fs_info,
292 struct btrfs_delayed_ref_root *delayed_refs,
293 struct btrfs_qgroup_extent_record *record);
294
295 /*
296 * Post handler after qgroup_trace_extent_nolock().
297 *
298 * NOTE: Current qgroup does the expensive backref walk at transaction
299 * committing time with TRANS_STATE_COMMIT_DOING, this blocks incoming
300 * new transaction.
301 * This is designed to allow btrfs_find_all_roots() to get correct new_roots
302 * result.
303 *
304 * However for old_roots there is no need to do backref walk at that time,
305 * since we search commit roots to walk backref and result will always be
306 * correct.
307 *
308 * Due to the nature of no lock version, we can't do backref there.
309 * So we must call btrfs_qgroup_trace_extent_post() after exiting
310 * spinlock context.
311 *
312 * TODO: If we can fix and prove btrfs_find_all_roots() can get correct result
313 * using current root, then we can move all expensive backref walk out of
314 * transaction committing, but not now as qgroup accounting will be wrong again.
315 */
316 int btrfs_qgroup_trace_extent_post(struct btrfs_trans_handle *trans,
317 struct btrfs_qgroup_extent_record *qrecord);
318
319 /*
320 * Inform qgroup to trace one dirty extent, specified by @bytenr and
321 * @num_bytes.
322 * So qgroup can account it at commit trans time.
323 *
324 * Better encapsulated version, with memory allocation and backref walk for
325 * commit roots.
326 * So this can sleep.
327 *
328 * Return 0 if the operation is done.
329 * Return <0 for error, like memory allocation failure or invalid parameter
330 * (NULL trans)
331 */
332 int btrfs_qgroup_trace_extent(struct btrfs_trans_handle *trans, u64 bytenr,
333 u64 num_bytes);
334
335 /*
336 * Inform qgroup to trace all leaf items of data
337 *
338 * Return 0 for success
339 * Return <0 for error(ENOMEM)
340 */
341 int btrfs_qgroup_trace_leaf_items(struct btrfs_trans_handle *trans,
342 struct extent_buffer *eb);
343 /*
344 * Inform qgroup to trace a whole subtree, including all its child tree
345 * blocks and data.
346 * The root tree block is specified by @root_eb.
347 *
348 * Normally used by relocation(tree block swap) and subvolume deletion.
349 *
350 * Return 0 for success
351 * Return <0 for error(ENOMEM or tree search error)
352 */
353 int btrfs_qgroup_trace_subtree(struct btrfs_trans_handle *trans,
354 struct extent_buffer *root_eb,
355 u64 root_gen, int root_level);
356 int btrfs_qgroup_account_extent(struct btrfs_trans_handle *trans, u64 bytenr,
357 u64 num_bytes, struct ulist *old_roots,
358 struct ulist *new_roots);
359 int btrfs_qgroup_account_extents(struct btrfs_trans_handle *trans);
360 int btrfs_run_qgroups(struct btrfs_trans_handle *trans);
361 int btrfs_qgroup_inherit(struct btrfs_trans_handle *trans, u64 srcid,
362 u64 objectid, struct btrfs_qgroup_inherit *inherit);
363 void btrfs_qgroup_free_refroot(struct btrfs_fs_info *fs_info,
364 u64 ref_root, u64 num_bytes,
365 enum btrfs_qgroup_rsv_type type);
366
367 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
368 int btrfs_verify_qgroup_counts(struct btrfs_fs_info *fs_info, u64 qgroupid,
369 u64 rfer, u64 excl);
370 #endif
371
372 /* New io_tree based accurate qgroup reserve API */
373 int btrfs_qgroup_reserve_data(struct btrfs_inode *inode,
374 struct extent_changeset **reserved, u64 start, u64 len);
375 int btrfs_qgroup_release_data(struct btrfs_inode *inode, u64 start, u64 len, u64 *released);
376 int btrfs_qgroup_free_data(struct btrfs_inode *inode,
377 struct extent_changeset *reserved, u64 start,
378 u64 len, u64 *freed);
379 int btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
380 enum btrfs_qgroup_rsv_type type, bool enforce);
381 int __btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
382 enum btrfs_qgroup_rsv_type type, bool enforce,
383 bool noflush);
384 /* Reserve metadata space for pertrans and prealloc type */
btrfs_qgroup_reserve_meta_pertrans(struct btrfs_root * root,int num_bytes,bool enforce)385 static inline int btrfs_qgroup_reserve_meta_pertrans(struct btrfs_root *root,
386 int num_bytes, bool enforce)
387 {
388 return __btrfs_qgroup_reserve_meta(root, num_bytes,
389 BTRFS_QGROUP_RSV_META_PERTRANS,
390 enforce, false);
391 }
btrfs_qgroup_reserve_meta_prealloc(struct btrfs_root * root,int num_bytes,bool enforce,bool noflush)392 static inline int btrfs_qgroup_reserve_meta_prealloc(struct btrfs_root *root,
393 int num_bytes, bool enforce,
394 bool noflush)
395 {
396 return __btrfs_qgroup_reserve_meta(root, num_bytes,
397 BTRFS_QGROUP_RSV_META_PREALLOC,
398 enforce, noflush);
399 }
400
401 void __btrfs_qgroup_free_meta(struct btrfs_root *root, int num_bytes,
402 enum btrfs_qgroup_rsv_type type);
403
404 /* Free per-transaction meta reservation for error handling */
btrfs_qgroup_free_meta_pertrans(struct btrfs_root * root,int num_bytes)405 static inline void btrfs_qgroup_free_meta_pertrans(struct btrfs_root *root,
406 int num_bytes)
407 {
408 __btrfs_qgroup_free_meta(root, num_bytes,
409 BTRFS_QGROUP_RSV_META_PERTRANS);
410 }
411
412 /* Pre-allocated meta reservation can be freed at need */
btrfs_qgroup_free_meta_prealloc(struct btrfs_root * root,int num_bytes)413 static inline void btrfs_qgroup_free_meta_prealloc(struct btrfs_root *root,
414 int num_bytes)
415 {
416 __btrfs_qgroup_free_meta(root, num_bytes,
417 BTRFS_QGROUP_RSV_META_PREALLOC);
418 }
419
420 /*
421 * Per-transaction meta reservation should be all freed at transaction commit
422 * time
423 */
424 void btrfs_qgroup_free_meta_all_pertrans(struct btrfs_root *root);
425
426 /*
427 * Convert @num_bytes of META_PREALLOCATED reservation to META_PERTRANS.
428 *
429 * This is called when preallocated meta reservation needs to be used.
430 * Normally after btrfs_join_transaction() call.
431 */
432 void btrfs_qgroup_convert_reserved_meta(struct btrfs_root *root, int num_bytes);
433
434 void btrfs_qgroup_check_reserved_leak(struct btrfs_inode *inode);
435
436 /* btrfs_qgroup_swapped_blocks related functions */
437 void btrfs_qgroup_init_swapped_blocks(
438 struct btrfs_qgroup_swapped_blocks *swapped_blocks);
439
440 void btrfs_qgroup_clean_swapped_blocks(struct btrfs_root *root);
441 int btrfs_qgroup_add_swapped_blocks(struct btrfs_trans_handle *trans,
442 struct btrfs_root *subvol_root,
443 struct btrfs_block_group *bg,
444 struct extent_buffer *subvol_parent, int subvol_slot,
445 struct extent_buffer *reloc_parent, int reloc_slot,
446 u64 last_snapshot);
447 int btrfs_qgroup_trace_subtree_after_cow(struct btrfs_trans_handle *trans,
448 struct btrfs_root *root, struct extent_buffer *eb);
449 void btrfs_qgroup_destroy_extent_records(struct btrfs_transaction *trans);
450 bool btrfs_check_quota_leak(struct btrfs_fs_info *fs_info);
451
452 #endif
453