xref: /linux/drivers/md/dm-cache-policy.h (revision ab520be8cd5d56867fc95cfbc34b90880faf1f9d)
1 /*
2  * Copyright (C) 2012 Red Hat. All rights reserved.
3  *
4  * This file is released under the GPL.
5  */
6 
7 #ifndef DM_CACHE_POLICY_H
8 #define DM_CACHE_POLICY_H
9 
10 #include "dm-cache-block-types.h"
11 
12 #include <linux/device-mapper.h>
13 
14 /*----------------------------------------------------------------*/
15 
16 /* FIXME: make it clear which methods are optional.  Get debug policy to
17  * double check this at start.
18  */
19 
20 /*
21  * The cache policy makes the important decisions about which blocks get to
22  * live on the faster cache device.
23  *
24  * When the core target has to remap a bio it calls the 'map' method of the
25  * policy.  This returns an instruction telling the core target what to do.
26  *
27  * POLICY_HIT:
28  *   That block is in the cache.  Remap to the cache and carry on.
29  *
30  * POLICY_MISS:
31  *   This block is on the origin device.  Remap and carry on.
32  *
33  * POLICY_NEW:
34  *   This block is currently on the origin device, but the policy wants to
35  *   move it.  The core should:
36  *
37  *   - hold any further io to this origin block
38  *   - copy the origin to the given cache block
39  *   - release all the held blocks
40  *   - remap the original block to the cache
41  *
42  * POLICY_REPLACE:
43  *   This block is currently on the origin device.  The policy wants to
44  *   move it to the cache, with the added complication that the destination
45  *   cache block needs a writeback first.  The core should:
46  *
47  *   - hold any further io to this origin block
48  *   - hold any further io to the origin block that's being written back
49  *   - writeback
50  *   - copy new block to cache
51  *   - release held blocks
52  *   - remap bio to cache and reissue.
53  *
54  * Should the core run into trouble while processing a POLICY_NEW or
55  * POLICY_REPLACE instruction it will roll back the policies mapping using
56  * remove_mapping() or force_mapping().  These methods must not fail.  This
57  * approach avoids having transactional semantics in the policy (ie, the
58  * core informing the policy when a migration is complete), and hence makes
59  * it easier to write new policies.
60  *
61  * In general policy methods should never block, except in the case of the
62  * map function when can_migrate is set.  So be careful to implement using
63  * bounded, preallocated memory.
64  */
65 enum policy_operation {
66 	POLICY_HIT,
67 	POLICY_MISS,
68 	POLICY_NEW,
69 	POLICY_REPLACE
70 };
71 
72 /*
73  * When issuing a POLICY_REPLACE the policy needs to make a callback to
74  * lock the block being demoted.  This doesn't need to occur during a
75  * writeback operation since the block remains in the cache.
76  */
77 struct policy_locker;
78 typedef int (*policy_lock_fn)(struct policy_locker *l, dm_oblock_t oblock);
79 
80 struct policy_locker {
81 	policy_lock_fn fn;
82 };
83 
84 /*
85  * This is the instruction passed back to the core target.
86  */
87 struct policy_result {
88 	enum policy_operation op;
89 	dm_oblock_t old_oblock;	/* POLICY_REPLACE */
90 	dm_cblock_t cblock;	/* POLICY_HIT, POLICY_NEW, POLICY_REPLACE */
91 };
92 
93 /*
94  * The cache policy object.  Just a bunch of methods.  It is envisaged that
95  * this structure will be embedded in a bigger, policy specific structure
96  * (ie. use container_of()).
97  */
98 struct dm_cache_policy {
99 
100 	/*
101 	 * FIXME: make it clear which methods are optional, and which may
102 	 * block.
103 	 */
104 
105 	/*
106 	 * Destroys this object.
107 	 */
108 	void (*destroy)(struct dm_cache_policy *p);
109 
110 	/*
111 	 * See large comment above.
112 	 *
113 	 * oblock      - the origin block we're interested in.
114 	 *
115 	 * can_block - indicates whether the current thread is allowed to
116 	 *             block.  -EWOULDBLOCK returned if it can't and would.
117 	 *
118 	 * can_migrate - gives permission for POLICY_NEW or POLICY_REPLACE
119 	 *               instructions.  If denied and the policy would have
120 	 *               returned one of these instructions it should
121 	 *               return -EWOULDBLOCK.
122 	 *
123 	 * discarded_oblock - indicates whether the whole origin block is
124 	 *               in a discarded state (FIXME: better to tell the
125 	 *               policy about this sooner, so it can recycle that
126 	 *               cache block if it wants.)
127 	 * bio         - the bio that triggered this call.
128 	 * result      - gets filled in with the instruction.
129 	 *
130 	 * May only return 0, or -EWOULDBLOCK (if !can_migrate)
131 	 */
132 	int (*map)(struct dm_cache_policy *p, dm_oblock_t oblock,
133 		   bool can_block, bool can_migrate, bool discarded_oblock,
134 		   struct bio *bio, struct policy_locker *locker,
135 		   struct policy_result *result);
136 
137 	/*
138 	 * Sometimes we want to see if a block is in the cache, without
139 	 * triggering any update of stats.  (ie. it's not a real hit).
140 	 *
141 	 * Must not block.
142 	 *
143 	 * Returns 0 if in cache, -ENOENT if not, < 0 for other errors
144 	 * (-EWOULDBLOCK would be typical).
145 	 */
146 	int (*lookup)(struct dm_cache_policy *p, dm_oblock_t oblock, dm_cblock_t *cblock);
147 
148 	void (*set_dirty)(struct dm_cache_policy *p, dm_oblock_t oblock);
149 	void (*clear_dirty)(struct dm_cache_policy *p, dm_oblock_t oblock);
150 
151 	/*
152 	 * Called when a cache target is first created.  Used to load a
153 	 * mapping from the metadata device into the policy.
154 	 */
155 	int (*load_mapping)(struct dm_cache_policy *p, dm_oblock_t oblock,
156 			    dm_cblock_t cblock, uint32_t hint, bool hint_valid);
157 
158 	/*
159 	 * Gets the hint for a given cblock.  Called in a single threaded
160 	 * context.  So no locking required.
161 	 */
162 	uint32_t (*get_hint)(struct dm_cache_policy *p, dm_cblock_t cblock);
163 
164 	/*
165 	 * Override functions used on the error paths of the core target.
166 	 * They must succeed.
167 	 */
168 	void (*remove_mapping)(struct dm_cache_policy *p, dm_oblock_t oblock);
169 	void (*force_mapping)(struct dm_cache_policy *p, dm_oblock_t current_oblock,
170 			      dm_oblock_t new_oblock);
171 
172 	/*
173 	 * This is called via the invalidate_cblocks message.  It is
174 	 * possible the particular cblock has already been removed due to a
175 	 * write io in passthrough mode.  In which case this should return
176 	 * -ENODATA.
177 	 */
178 	int (*remove_cblock)(struct dm_cache_policy *p, dm_cblock_t cblock);
179 
180 	/*
181 	 * Provide a dirty block to be written back by the core target.  If
182 	 * critical_only is set then the policy should only provide work if
183 	 * it urgently needs it.
184 	 *
185 	 * Returns:
186 	 *
187 	 * 0 and @cblock,@oblock: block to write back provided
188 	 *
189 	 * -ENODATA: no dirty blocks available
190 	 */
191 	int (*writeback_work)(struct dm_cache_policy *p, dm_oblock_t *oblock, dm_cblock_t *cblock,
192 			      bool critical_only);
193 
194 	/*
195 	 * How full is the cache?
196 	 */
197 	dm_cblock_t (*residency)(struct dm_cache_policy *p);
198 
199 	/*
200 	 * Because of where we sit in the block layer, we can be asked to
201 	 * map a lot of little bios that are all in the same block (no
202 	 * queue merging has occurred).  To stop the policy being fooled by
203 	 * these, the core target sends regular tick() calls to the policy.
204 	 * The policy should only count an entry as hit once per tick.
205 	 */
206 	void (*tick)(struct dm_cache_policy *p, bool can_block);
207 
208 	/*
209 	 * Configuration.
210 	 */
211 	int (*emit_config_values)(struct dm_cache_policy *p, char *result,
212 				  unsigned maxlen, ssize_t *sz_ptr);
213 	int (*set_config_value)(struct dm_cache_policy *p,
214 				const char *key, const char *value);
215 
216 	/*
217 	 * Book keeping ptr for the policy register, not for general use.
218 	 */
219 	void *private;
220 };
221 
222 /*----------------------------------------------------------------*/
223 
224 /*
225  * We maintain a little register of the different policy types.
226  */
227 #define CACHE_POLICY_NAME_SIZE 16
228 #define CACHE_POLICY_VERSION_SIZE 3
229 
230 struct dm_cache_policy_type {
231 	/* For use by the register code only. */
232 	struct list_head list;
233 
234 	/*
235 	 * Policy writers should fill in these fields.  The name field is
236 	 * what gets passed on the target line to select your policy.
237 	 */
238 	char name[CACHE_POLICY_NAME_SIZE];
239 	unsigned version[CACHE_POLICY_VERSION_SIZE];
240 
241 	/*
242 	 * For use by an alias dm_cache_policy_type to point to the
243 	 * real dm_cache_policy_type.
244 	 */
245 	struct dm_cache_policy_type *real;
246 
247 	/*
248 	 * Policies may store a hint for each each cache block.
249 	 * Currently the size of this hint must be 0 or 4 bytes but we
250 	 * expect to relax this in future.
251 	 */
252 	size_t hint_size;
253 
254 	struct module *owner;
255 	struct dm_cache_policy *(*create)(dm_cblock_t cache_size,
256 					  sector_t origin_size,
257 					  sector_t block_size);
258 };
259 
260 int dm_cache_policy_register(struct dm_cache_policy_type *type);
261 void dm_cache_policy_unregister(struct dm_cache_policy_type *type);
262 
263 /*----------------------------------------------------------------*/
264 
265 #endif	/* DM_CACHE_POLICY_H */
266