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