xref: /linux/fs/btrfs/qgroup.h (revision 58b3bafff8257c6946df5d6aeb215b8ac839ed2a)
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 "ulist.h"
12 #include "delayed-ref.h"
13 
14 /*
15  * Btrfs qgroup overview
16  *
17  * Btrfs qgroup splits into 3 main part:
18  * 1) Reserve
19  *    Reserve metadata/data space for incoming operations
20  *    Affect how qgroup limit works
21  *
22  * 2) Trace
23  *    Tell btrfs qgroup to trace dirty extents.
24  *
25  *    Dirty extents including:
26  *    - Newly allocated extents
27  *    - Extents going to be deleted (in this trans)
28  *    - Extents whose owner is going to be modified
29  *
30  *    This is the main part affects whether qgroup numbers will stay
31  *    consistent.
32  *    Btrfs qgroup can trace clean extents and won't cause any problem,
33  *    but it will consume extra CPU time, it should be avoided if possible.
34  *
35  * 3) Account
36  *    Btrfs qgroup will updates its numbers, based on dirty extents traced
37  *    in previous step.
38  *
39  *    Normally at qgroup rescan and transaction commit time.
40  */
41 
42 /*
43  * Special performance optimization for balance.
44  *
45  * For balance, we need to swap subtree of subvolume and reloc trees.
46  * In theory, we need to trace all subtree blocks of both subvolume and reloc
47  * trees, since their owner has changed during such swap.
48  *
49  * However since balance has ensured that both subtrees are containing the
50  * same contents and have the same tree structures, such swap won't cause
51  * qgroup number change.
52  *
53  * But there is a race window between subtree swap and transaction commit,
54  * during that window, if we increase/decrease tree level or merge/split tree
55  * blocks, we still need to trace the original subtrees.
56  *
57  * So for balance, we use a delayed subtree tracing, whose workflow is:
58  *
59  * 1) Record the subtree root block get swapped.
60  *
61  *    During subtree swap:
62  *    O = Old tree blocks
63  *    N = New tree blocks
64  *          reloc tree                     subvolume tree X
65  *             Root                               Root
66  *            /    \                             /    \
67  *          NA     OB                          OA      OB
68  *        /  |     |  \                      /  |      |  \
69  *      NC  ND     OE  OF                   OC  OD     OE  OF
70  *
71  *   In this case, NA and OA are going to be swapped, record (NA, OA) into
72  *   subvolume tree X.
73  *
74  * 2) After subtree swap.
75  *          reloc tree                     subvolume tree X
76  *             Root                               Root
77  *            /    \                             /    \
78  *          OA     OB                          NA      OB
79  *        /  |     |  \                      /  |      |  \
80  *      OC  OD     OE  OF                   NC  ND     OE  OF
81  *
82  * 3a) COW happens for OB
83  *     If we are going to COW tree block OB, we check OB's bytenr against
84  *     tree X's swapped_blocks structure.
85  *     If it doesn't fit any, nothing will happen.
86  *
87  * 3b) COW happens for NA
88  *     Check NA's bytenr against tree X's swapped_blocks, and get a hit.
89  *     Then we do subtree scan on both subtrees OA and NA.
90  *     Resulting 6 tree blocks to be scanned (OA, OC, OD, NA, NC, ND).
91  *
92  *     Then no matter what we do to subvolume tree X, qgroup numbers will
93  *     still be correct.
94  *     Then NA's record gets removed from X's swapped_blocks.
95  *
96  * 4)  Transaction commit
97  *     Any record in X's swapped_blocks gets removed, since there is no
98  *     modification to the swapped subtrees, no need to trigger heavy qgroup
99  *     subtree rescan for them.
100  */
101 
102 /*
103  * Record a dirty extent, and info qgroup to update quota on it
104  * TODO: Use kmem cache to alloc it.
105  */
106 struct btrfs_qgroup_extent_record {
107 	struct rb_node node;
108 	u64 bytenr;
109 	u64 num_bytes;
110 
111 	/*
112 	 * For qgroup reserved data space freeing.
113 	 *
114 	 * @data_rsv_refroot and @data_rsv will be recorded after
115 	 * BTRFS_ADD_DELAYED_EXTENT is called.
116 	 * And will be used to free reserved qgroup space at
117 	 * transaction commit time.
118 	 */
119 	u32 data_rsv;		/* reserved data space needs to be freed */
120 	u64 data_rsv_refroot;	/* which root the reserved data belongs to */
121 	struct ulist *old_roots;
122 };
123 
124 struct btrfs_qgroup_swapped_block {
125 	struct rb_node node;
126 
127 	int level;
128 	bool trace_leaf;
129 
130 	/* bytenr/generation of the tree block in subvolume tree after swap */
131 	u64 subvol_bytenr;
132 	u64 subvol_generation;
133 
134 	/* bytenr/generation of the tree block in reloc tree after swap */
135 	u64 reloc_bytenr;
136 	u64 reloc_generation;
137 
138 	u64 last_snapshot;
139 	struct btrfs_key first_key;
140 };
141 
142 /*
143  * Qgroup reservation types:
144  *
145  * DATA:
146  *	space reserved for data
147  *
148  * META_PERTRANS:
149  * 	Space reserved for metadata (per-transaction)
150  * 	Due to the fact that qgroup data is only updated at transaction commit
151  * 	time, reserved space for metadata must be kept until transaction
152  * 	commits.
153  * 	Any metadata reserved that are used in btrfs_start_transaction() should
154  * 	be of this type.
155  *
156  * META_PREALLOC:
157  *	There are cases where metadata space is reserved before starting
158  *	transaction, and then btrfs_join_transaction() to get a trans handle.
159  *	Any metadata reserved for such usage should be of this type.
160  *	And after join_transaction() part (or all) of such reservation should
161  *	be converted into META_PERTRANS.
162  */
163 enum btrfs_qgroup_rsv_type {
164 	BTRFS_QGROUP_RSV_DATA,
165 	BTRFS_QGROUP_RSV_META_PERTRANS,
166 	BTRFS_QGROUP_RSV_META_PREALLOC,
167 	BTRFS_QGROUP_RSV_LAST,
168 };
169 
170 /*
171  * Represents how many bytes we have reserved for this qgroup.
172  *
173  * Each type should have different reservation behavior.
174  * E.g, data follows its io_tree flag modification, while
175  * *currently* meta is just reserve-and-clear during transaction.
176  *
177  * TODO: Add new type for reservation which can survive transaction commit.
178  * Current metadata reservation behavior is not suitable for such case.
179  */
180 struct btrfs_qgroup_rsv {
181 	u64 values[BTRFS_QGROUP_RSV_LAST];
182 };
183 
184 /*
185  * one struct for each qgroup, organized in fs_info->qgroup_tree.
186  */
187 struct btrfs_qgroup {
188 	u64 qgroupid;
189 
190 	/*
191 	 * state
192 	 */
193 	u64 rfer;	/* referenced */
194 	u64 rfer_cmpr;	/* referenced compressed */
195 	u64 excl;	/* exclusive */
196 	u64 excl_cmpr;	/* exclusive compressed */
197 
198 	/*
199 	 * limits
200 	 */
201 	u64 lim_flags;	/* which limits are set */
202 	u64 max_rfer;
203 	u64 max_excl;
204 	u64 rsv_rfer;
205 	u64 rsv_excl;
206 
207 	/*
208 	 * reservation tracking
209 	 */
210 	struct btrfs_qgroup_rsv rsv;
211 
212 	/*
213 	 * lists
214 	 */
215 	struct list_head groups;  /* groups this group is member of */
216 	struct list_head members; /* groups that are members of this group */
217 	struct list_head dirty;   /* dirty groups */
218 	struct rb_node node;	  /* tree of qgroups */
219 
220 	/*
221 	 * temp variables for accounting operations
222 	 * Refer to qgroup_shared_accounting() for details.
223 	 */
224 	u64 old_refcnt;
225 	u64 new_refcnt;
226 };
227 
228 /*
229  * For qgroup event trace points only
230  */
231 #define QGROUP_RESERVE		(1<<0)
232 #define QGROUP_RELEASE		(1<<1)
233 #define QGROUP_FREE		(1<<2)
234 
235 int btrfs_quota_enable(struct btrfs_fs_info *fs_info);
236 int btrfs_quota_disable(struct btrfs_fs_info *fs_info);
237 int btrfs_qgroup_rescan(struct btrfs_fs_info *fs_info);
238 void btrfs_qgroup_rescan_resume(struct btrfs_fs_info *fs_info);
239 int btrfs_qgroup_wait_for_completion(struct btrfs_fs_info *fs_info,
240 				     bool interruptible);
241 int btrfs_add_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
242 			      u64 dst);
243 int btrfs_del_qgroup_relation(struct btrfs_trans_handle *trans, u64 src,
244 			      u64 dst);
245 int btrfs_create_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid);
246 int btrfs_remove_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid);
247 int btrfs_limit_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid,
248 		       struct btrfs_qgroup_limit *limit);
249 int btrfs_read_qgroup_config(struct btrfs_fs_info *fs_info);
250 void btrfs_free_qgroup_config(struct btrfs_fs_info *fs_info);
251 struct btrfs_delayed_extent_op;
252 
253 /*
254  * Inform qgroup to trace one dirty extent, its info is recorded in @record.
255  * So qgroup can account it at transaction committing time.
256  *
257  * No lock version, caller must acquire delayed ref lock and allocated memory,
258  * then call btrfs_qgroup_trace_extent_post() after exiting lock context.
259  *
260  * Return 0 for success insert
261  * Return >0 for existing record, caller can free @record safely.
262  * Error is not possible
263  */
264 int btrfs_qgroup_trace_extent_nolock(
265 		struct btrfs_fs_info *fs_info,
266 		struct btrfs_delayed_ref_root *delayed_refs,
267 		struct btrfs_qgroup_extent_record *record);
268 
269 /*
270  * Post handler after qgroup_trace_extent_nolock().
271  *
272  * NOTE: Current qgroup does the expensive backref walk at transaction
273  * committing time with TRANS_STATE_COMMIT_DOING, this blocks incoming
274  * new transaction.
275  * This is designed to allow btrfs_find_all_roots() to get correct new_roots
276  * result.
277  *
278  * However for old_roots there is no need to do backref walk at that time,
279  * since we search commit roots to walk backref and result will always be
280  * correct.
281  *
282  * Due to the nature of no lock version, we can't do backref there.
283  * So we must call btrfs_qgroup_trace_extent_post() after exiting
284  * spinlock context.
285  *
286  * TODO: If we can fix and prove btrfs_find_all_roots() can get correct result
287  * using current root, then we can move all expensive backref walk out of
288  * transaction committing, but not now as qgroup accounting will be wrong again.
289  */
290 int btrfs_qgroup_trace_extent_post(struct btrfs_fs_info *fs_info,
291 				   struct btrfs_qgroup_extent_record *qrecord);
292 
293 /*
294  * Inform qgroup to trace one dirty extent, specified by @bytenr and
295  * @num_bytes.
296  * So qgroup can account it at commit trans time.
297  *
298  * Better encapsulated version, with memory allocation and backref walk for
299  * commit roots.
300  * So this can sleep.
301  *
302  * Return 0 if the operation is done.
303  * Return <0 for error, like memory allocation failure or invalid parameter
304  * (NULL trans)
305  */
306 int btrfs_qgroup_trace_extent(struct btrfs_trans_handle *trans, u64 bytenr,
307 			      u64 num_bytes, gfp_t gfp_flag);
308 
309 /*
310  * Inform qgroup to trace all leaf items of data
311  *
312  * Return 0 for success
313  * Return <0 for error(ENOMEM)
314  */
315 int btrfs_qgroup_trace_leaf_items(struct btrfs_trans_handle *trans,
316 				  struct extent_buffer *eb);
317 /*
318  * Inform qgroup to trace a whole subtree, including all its child tree
319  * blocks and data.
320  * The root tree block is specified by @root_eb.
321  *
322  * Normally used by relocation(tree block swap) and subvolume deletion.
323  *
324  * Return 0 for success
325  * Return <0 for error(ENOMEM or tree search error)
326  */
327 int btrfs_qgroup_trace_subtree(struct btrfs_trans_handle *trans,
328 			       struct extent_buffer *root_eb,
329 			       u64 root_gen, int root_level);
330 int btrfs_qgroup_account_extent(struct btrfs_trans_handle *trans, u64 bytenr,
331 				u64 num_bytes, struct ulist *old_roots,
332 				struct ulist *new_roots);
333 int btrfs_qgroup_account_extents(struct btrfs_trans_handle *trans);
334 int btrfs_run_qgroups(struct btrfs_trans_handle *trans);
335 int btrfs_qgroup_inherit(struct btrfs_trans_handle *trans, u64 srcid,
336 			 u64 objectid, struct btrfs_qgroup_inherit *inherit);
337 void btrfs_qgroup_free_refroot(struct btrfs_fs_info *fs_info,
338 			       u64 ref_root, u64 num_bytes,
339 			       enum btrfs_qgroup_rsv_type type);
340 
341 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
342 int btrfs_verify_qgroup_counts(struct btrfs_fs_info *fs_info, u64 qgroupid,
343 			       u64 rfer, u64 excl);
344 #endif
345 
346 /* New io_tree based accurate qgroup reserve API */
347 int btrfs_qgroup_reserve_data(struct inode *inode,
348 			struct extent_changeset **reserved, u64 start, u64 len);
349 int btrfs_qgroup_release_data(struct inode *inode, u64 start, u64 len);
350 int btrfs_qgroup_free_data(struct inode *inode,
351 			struct extent_changeset *reserved, u64 start, u64 len);
352 
353 int __btrfs_qgroup_reserve_meta(struct btrfs_root *root, int num_bytes,
354 				enum btrfs_qgroup_rsv_type type, bool enforce);
355 /* Reserve metadata space for pertrans and prealloc type */
356 static inline int btrfs_qgroup_reserve_meta_pertrans(struct btrfs_root *root,
357 				int num_bytes, bool enforce)
358 {
359 	return __btrfs_qgroup_reserve_meta(root, num_bytes,
360 			BTRFS_QGROUP_RSV_META_PERTRANS, enforce);
361 }
362 static inline int btrfs_qgroup_reserve_meta_prealloc(struct btrfs_root *root,
363 				int num_bytes, bool enforce)
364 {
365 	return __btrfs_qgroup_reserve_meta(root, num_bytes,
366 			BTRFS_QGROUP_RSV_META_PREALLOC, enforce);
367 }
368 
369 void __btrfs_qgroup_free_meta(struct btrfs_root *root, int num_bytes,
370 			     enum btrfs_qgroup_rsv_type type);
371 
372 /* Free per-transaction meta reservation for error handling */
373 static inline void btrfs_qgroup_free_meta_pertrans(struct btrfs_root *root,
374 						   int num_bytes)
375 {
376 	__btrfs_qgroup_free_meta(root, num_bytes,
377 			BTRFS_QGROUP_RSV_META_PERTRANS);
378 }
379 
380 /* Pre-allocated meta reservation can be freed at need */
381 static inline void btrfs_qgroup_free_meta_prealloc(struct btrfs_root *root,
382 						   int num_bytes)
383 {
384 	__btrfs_qgroup_free_meta(root, num_bytes,
385 			BTRFS_QGROUP_RSV_META_PREALLOC);
386 }
387 
388 /*
389  * Per-transaction meta reservation should be all freed at transaction commit
390  * time
391  */
392 void btrfs_qgroup_free_meta_all_pertrans(struct btrfs_root *root);
393 
394 /*
395  * Convert @num_bytes of META_PREALLOCATED reservation to META_PERTRANS.
396  *
397  * This is called when preallocated meta reservation needs to be used.
398  * Normally after btrfs_join_transaction() call.
399  */
400 void btrfs_qgroup_convert_reserved_meta(struct btrfs_root *root, int num_bytes);
401 
402 void btrfs_qgroup_check_reserved_leak(struct inode *inode);
403 
404 /* btrfs_qgroup_swapped_blocks related functions */
405 void btrfs_qgroup_init_swapped_blocks(
406 	struct btrfs_qgroup_swapped_blocks *swapped_blocks);
407 
408 void btrfs_qgroup_clean_swapped_blocks(struct btrfs_root *root);
409 int btrfs_qgroup_add_swapped_blocks(struct btrfs_trans_handle *trans,
410 		struct btrfs_root *subvol_root,
411 		struct btrfs_block_group *bg,
412 		struct extent_buffer *subvol_parent, int subvol_slot,
413 		struct extent_buffer *reloc_parent, int reloc_slot,
414 		u64 last_snapshot);
415 int btrfs_qgroup_trace_subtree_after_cow(struct btrfs_trans_handle *trans,
416 		struct btrfs_root *root, struct extent_buffer *eb);
417 
418 #endif
419