xref: /linux/fs/btrfs/qgroup.h (revision d6a5c562214f26e442c8ec3ff1e28e16675d1bcf)
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 #define BTRFS_QGROUP_RUNTIME_FLAG_CANCEL_RESCAN		(1UL << 3)
104 #define BTRFS_QGROUP_RUNTIME_FLAG_NO_ACCOUNTING		(1UL << 4)
105 
106 /*
107  * Record a dirty extent, and info qgroup to update quota on it
108  * TODO: Use kmem cache to alloc it.
109  */
110 struct btrfs_qgroup_extent_record {
111 	struct rb_node node;
112 	u64 bytenr;
113 	u64 num_bytes;
114 
115 	/*
116 	 * For qgroup reserved data space freeing.
117 	 *
118 	 * @data_rsv_refroot and @data_rsv will be recorded after
119 	 * BTRFS_ADD_DELAYED_EXTENT is called.
120 	 * And will be used to free reserved qgroup space at
121 	 * transaction commit time.
122 	 */
123 	u32 data_rsv;		/* reserved data space needs to be freed */
124 	u64 data_rsv_refroot;	/* which root the reserved data belongs to */
125 	struct ulist *old_roots;
126 };
127 
128 struct btrfs_qgroup_swapped_block {
129 	struct rb_node node;
130 
131 	int level;
132 	bool trace_leaf;
133 
134 	/* bytenr/generation of the tree block in subvolume tree after swap */
135 	u64 subvol_bytenr;
136 	u64 subvol_generation;
137 
138 	/* bytenr/generation of the tree block in reloc tree after swap */
139 	u64 reloc_bytenr;
140 	u64 reloc_generation;
141 
142 	u64 last_snapshot;
143 	struct btrfs_key first_key;
144 };
145 
146 /*
147  * Qgroup reservation types:
148  *
149  * DATA:
150  *	space reserved for data
151  *
152  * META_PERTRANS:
153  * 	Space reserved for metadata (per-transaction)
154  * 	Due to the fact that qgroup data is only updated at transaction commit
155  * 	time, reserved space for metadata must be kept until transaction
156  * 	commits.
157  * 	Any metadata reserved that are used in btrfs_start_transaction() should
158  * 	be of this type.
159  *
160  * META_PREALLOC:
161  *	There are cases where metadata space is reserved before starting
162  *	transaction, and then btrfs_join_transaction() to get a trans handle.
163  *	Any metadata reserved for such usage should be of this type.
164  *	And after join_transaction() part (or all) of such reservation should
165  *	be converted into META_PERTRANS.
166  */
167 enum btrfs_qgroup_rsv_type {
168 	BTRFS_QGROUP_RSV_DATA,
169 	BTRFS_QGROUP_RSV_META_PERTRANS,
170 	BTRFS_QGROUP_RSV_META_PREALLOC,
171 	BTRFS_QGROUP_RSV_LAST,
172 };
173 
174 /*
175  * Represents how many bytes we have reserved for this qgroup.
176  *
177  * Each type should have different reservation behavior.
178  * E.g, data follows its io_tree flag modification, while
179  * *currently* meta is just reserve-and-clear during transaction.
180  *
181  * TODO: Add new type for reservation which can survive transaction commit.
182  * Current metadata reservation behavior is not suitable for such case.
183  */
184 struct btrfs_qgroup_rsv {
185 	u64 values[BTRFS_QGROUP_RSV_LAST];
186 };
187 
188 /*
189  * one struct for each qgroup, organized in fs_info->qgroup_tree.
190  */
191 struct btrfs_qgroup {
192 	u64 qgroupid;
193 
194 	/*
195 	 * state
196 	 */
197 	u64 rfer;	/* referenced */
198 	u64 rfer_cmpr;	/* referenced compressed */
199 	u64 excl;	/* exclusive */
200 	u64 excl_cmpr;	/* exclusive compressed */
201 
202 	/*
203 	 * limits
204 	 */
205 	u64 lim_flags;	/* which limits are set */
206 	u64 max_rfer;
207 	u64 max_excl;
208 	u64 rsv_rfer;
209 	u64 rsv_excl;
210 
211 	/*
212 	 * reservation tracking
213 	 */
214 	struct btrfs_qgroup_rsv rsv;
215 
216 	/*
217 	 * lists
218 	 */
219 	struct list_head groups;  /* groups this group is member of */
220 	struct list_head members; /* groups that are members of this group */
221 	struct list_head dirty;   /* dirty groups */
222 	struct rb_node node;	  /* tree of qgroups */
223 
224 	/*
225 	 * temp variables for accounting operations
226 	 * Refer to qgroup_shared_accounting() for details.
227 	 */
228 	u64 old_refcnt;
229 	u64 new_refcnt;
230 
231 	/*
232 	 * Sysfs kobjectid
233 	 */
234 	struct kobject kobj;
235 };
236 
237 static inline u64 btrfs_qgroup_subvolid(u64 qgroupid)
238 {
239 	return (qgroupid & ((1ULL << BTRFS_QGROUP_LEVEL_SHIFT) - 1));
240 }
241 
242 /*
243  * For qgroup event trace points only
244  */
245 #define QGROUP_RESERVE		(1<<0)
246 #define QGROUP_RELEASE		(1<<1)
247 #define QGROUP_FREE		(1<<2)
248 
249 int btrfs_quota_enable(struct btrfs_fs_info *fs_info);
250 int btrfs_quota_disable(struct btrfs_fs_info *fs_info);
251 int btrfs_qgroup_rescan(struct btrfs_fs_info *fs_info);
252 void btrfs_qgroup_rescan_resume(struct btrfs_fs_info *fs_info);
253 int btrfs_qgroup_wait_for_completion(struct btrfs_fs_info *fs_info,
254 				     bool interruptible);
255 int btrfs_add_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
256 			      u64 dst);
257 int btrfs_del_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
258 			      u64 dst);
259 int btrfs_create_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid);
260 int btrfs_remove_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid);
261 int btrfs_limit_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid,
262 		       struct btrfs_qgroup_limit *limit);
263 int btrfs_read_qgroup_config(struct btrfs_fs_info *fs_info);
264 void btrfs_free_qgroup_config(struct btrfs_fs_info *fs_info);
265 struct btrfs_delayed_extent_op;
266 
267 /*
268  * Inform qgroup to trace one dirty extent, its info is recorded in @record.
269  * So qgroup can account it at transaction committing time.
270  *
271  * No lock version, caller must acquire delayed ref lock and allocated memory,
272  * then call btrfs_qgroup_trace_extent_post() after exiting lock context.
273  *
274  * Return 0 for success insert
275  * Return >0 for existing record, caller can free @record safely.
276  * Error is not possible
277  */
278 int btrfs_qgroup_trace_extent_nolock(
279 		struct btrfs_fs_info *fs_info,
280 		struct btrfs_delayed_ref_root *delayed_refs,
281 		struct btrfs_qgroup_extent_record *record);
282 
283 /*
284  * Post handler after qgroup_trace_extent_nolock().
285  *
286  * NOTE: Current qgroup does the expensive backref walk at transaction
287  * committing time with TRANS_STATE_COMMIT_DOING, this blocks incoming
288  * new transaction.
289  * This is designed to allow btrfs_find_all_roots() to get correct new_roots
290  * result.
291  *
292  * However for old_roots there is no need to do backref walk at that time,
293  * since we search commit roots to walk backref and result will always be
294  * correct.
295  *
296  * Due to the nature of no lock version, we can't do backref there.
297  * So we must call btrfs_qgroup_trace_extent_post() after exiting
298  * spinlock context.
299  *
300  * TODO: If we can fix and prove btrfs_find_all_roots() can get correct result
301  * using current root, then we can move all expensive backref walk out of
302  * transaction committing, but not now as qgroup accounting will be wrong again.
303  */
304 int btrfs_qgroup_trace_extent_post(struct btrfs_trans_handle *trans,
305 				   struct btrfs_qgroup_extent_record *qrecord);
306 
307 /*
308  * Inform qgroup to trace one dirty extent, specified by @bytenr and
309  * @num_bytes.
310  * So qgroup can account it at commit trans time.
311  *
312  * Better encapsulated version, with memory allocation and backref walk for
313  * commit roots.
314  * So this can sleep.
315  *
316  * Return 0 if the operation is done.
317  * Return <0 for error, like memory allocation failure or invalid parameter
318  * (NULL trans)
319  */
320 int btrfs_qgroup_trace_extent(struct btrfs_trans_handle *trans, u64 bytenr,
321 			      u64 num_bytes, gfp_t gfp_flag);
322 
323 /*
324  * Inform qgroup to trace all leaf items of data
325  *
326  * Return 0 for success
327  * Return <0 for error(ENOMEM)
328  */
329 int btrfs_qgroup_trace_leaf_items(struct btrfs_trans_handle *trans,
330 				  struct extent_buffer *eb);
331 /*
332  * Inform qgroup to trace a whole subtree, including all its child tree
333  * blocks and data.
334  * The root tree block is specified by @root_eb.
335  *
336  * Normally used by relocation(tree block swap) and subvolume deletion.
337  *
338  * Return 0 for success
339  * Return <0 for error(ENOMEM or tree search error)
340  */
341 int btrfs_qgroup_trace_subtree(struct btrfs_trans_handle *trans,
342 			       struct extent_buffer *root_eb,
343 			       u64 root_gen, int root_level);
344 int btrfs_qgroup_account_extent(struct btrfs_trans_handle *trans, u64 bytenr,
345 				u64 num_bytes, struct ulist *old_roots,
346 				struct ulist *new_roots);
347 int btrfs_qgroup_account_extents(struct btrfs_trans_handle *trans);
348 int btrfs_run_qgroups(struct btrfs_trans_handle *trans);
349 int btrfs_qgroup_inherit(struct btrfs_trans_handle *trans, u64 srcid,
350 			 u64 objectid, struct btrfs_qgroup_inherit *inherit);
351 void btrfs_qgroup_free_refroot(struct btrfs_fs_info *fs_info,
352 			       u64 ref_root, u64 num_bytes,
353 			       enum btrfs_qgroup_rsv_type type);
354 
355 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
356 int btrfs_verify_qgroup_counts(struct btrfs_fs_info *fs_info, u64 qgroupid,
357 			       u64 rfer, u64 excl);
358 #endif
359 
360 /* New io_tree based accurate qgroup reserve API */
361 int btrfs_qgroup_reserve_data(struct btrfs_inode *inode,
362 			struct extent_changeset **reserved, u64 start, u64 len);
363 int btrfs_qgroup_release_data(struct btrfs_inode *inode, u64 start, u64 len);
364 int btrfs_qgroup_free_data(struct btrfs_inode *inode,
365 			   struct extent_changeset *reserved, u64 start,
366 			   u64 len);
367 int btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
368 			      enum btrfs_qgroup_rsv_type type, bool enforce);
369 int __btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
370 				enum btrfs_qgroup_rsv_type type, bool enforce,
371 				bool noflush);
372 /* Reserve metadata space for pertrans and prealloc type */
373 static inline int btrfs_qgroup_reserve_meta_pertrans(struct btrfs_root *root,
374 				int num_bytes, bool enforce)
375 {
376 	return __btrfs_qgroup_reserve_meta(root, num_bytes,
377 					   BTRFS_QGROUP_RSV_META_PERTRANS,
378 					   enforce, false);
379 }
380 static inline int btrfs_qgroup_reserve_meta_prealloc(struct btrfs_root *root,
381 						     int num_bytes, bool enforce,
382 						     bool noflush)
383 {
384 	return __btrfs_qgroup_reserve_meta(root, num_bytes,
385 					   BTRFS_QGROUP_RSV_META_PREALLOC,
386 					   enforce, noflush);
387 }
388 
389 void __btrfs_qgroup_free_meta(struct btrfs_root *root, int num_bytes,
390 			     enum btrfs_qgroup_rsv_type type);
391 
392 /* Free per-transaction meta reservation for error handling */
393 static inline void btrfs_qgroup_free_meta_pertrans(struct btrfs_root *root,
394 						   int num_bytes)
395 {
396 	__btrfs_qgroup_free_meta(root, num_bytes,
397 			BTRFS_QGROUP_RSV_META_PERTRANS);
398 }
399 
400 /* Pre-allocated meta reservation can be freed at need */
401 static inline void btrfs_qgroup_free_meta_prealloc(struct btrfs_root *root,
402 						   int num_bytes)
403 {
404 	__btrfs_qgroup_free_meta(root, num_bytes,
405 			BTRFS_QGROUP_RSV_META_PREALLOC);
406 }
407 
408 /*
409  * Per-transaction meta reservation should be all freed at transaction commit
410  * time
411  */
412 void btrfs_qgroup_free_meta_all_pertrans(struct btrfs_root *root);
413 
414 /*
415  * Convert @num_bytes of META_PREALLOCATED reservation to META_PERTRANS.
416  *
417  * This is called when preallocated meta reservation needs to be used.
418  * Normally after btrfs_join_transaction() call.
419  */
420 void btrfs_qgroup_convert_reserved_meta(struct btrfs_root *root, int num_bytes);
421 
422 void btrfs_qgroup_check_reserved_leak(struct btrfs_inode *inode);
423 
424 /* btrfs_qgroup_swapped_blocks related functions */
425 void btrfs_qgroup_init_swapped_blocks(
426 	struct btrfs_qgroup_swapped_blocks *swapped_blocks);
427 
428 void btrfs_qgroup_clean_swapped_blocks(struct btrfs_root *root);
429 int btrfs_qgroup_add_swapped_blocks(struct btrfs_trans_handle *trans,
430 		struct btrfs_root *subvol_root,
431 		struct btrfs_block_group *bg,
432 		struct extent_buffer *subvol_parent, int subvol_slot,
433 		struct extent_buffer *reloc_parent, int reloc_slot,
434 		u64 last_snapshot);
435 int btrfs_qgroup_trace_subtree_after_cow(struct btrfs_trans_handle *trans,
436 		struct btrfs_root *root, struct extent_buffer *eb);
437 void btrfs_qgroup_destroy_extent_records(struct btrfs_transaction *trans);
438 bool btrfs_check_quota_leak(struct btrfs_fs_info *fs_info);
439 
440 #endif
441