1/* SPDX-License-Identifier: GPL-2.0-or-later */
2/*
3 lru_cache.c
4
5 This file is part of DRBD by Philipp Reisner and Lars Ellenberg.
6
7 Copyright (C) 2003-2008, LINBIT Information Technologies GmbH.
8 Copyright (C) 2003-2008, Philipp Reisner <philipp.reisner@linbit.com>.
9 Copyright (C) 2003-2008, Lars Ellenberg <lars.ellenberg@linbit.com>.
10
11
12 */
13
14#ifndef LRU_CACHE_H
15#define LRU_CACHE_H
16
17#include <linux/list.h>
18#include <linux/slab.h>
19#include <linux/bitops.h>
20#include <linux/string.h> /* for memset */
21#include <linux/seq_file.h>
22
23/*
24This header file (and its .c file; kernel-doc of functions see there)
25 define a helper framework to easily keep track of index:label associations,
26 and changes to an "active set" of objects, as well as pending transactions,
27 to persistently record those changes.
28
29 We use an LRU policy if it is necessary to "cool down" a region currently in
30 the active set before we can "heat" a previously unused region.
31
32 Because of this later property, it is called "lru_cache".
33 As it actually Tracks Objects in an Active SeT, we could also call it
34 toast (incidentally that is what may happen to the data on the
35 backend storage upon next resync, if we don't get it right).
36
37What for?
38
39We replicate IO (more or less synchronously) to local and remote disk.
40
41For crash recovery after replication node failure,
42 we need to resync all regions that have been target of in-flight WRITE IO
43 (in use, or "hot", regions), as we don't know whether or not those WRITEs
44 have made it to stable storage.
45
46 To avoid a "full resync", we need to persistently track these regions.
47
48 This is known as "write intent log", and can be implemented as on-disk
49 (coarse or fine grained) bitmap, or other meta data.
50
51 To avoid the overhead of frequent extra writes to this meta data area,
52 usually the condition is softened to regions that _may_ have been target of
53 in-flight WRITE IO, e.g. by only lazily clearing the on-disk write-intent
54 bitmap, trading frequency of meta data transactions against amount of
55 (possibly unnecessary) resync traffic.
56
57 If we set a hard limit on the area that may be "hot" at any given time, we
58 limit the amount of resync traffic needed for crash recovery.
59
60For recovery after replication link failure,
61 we need to resync all blocks that have been changed on the other replica
62 in the mean time, or, if both replica have been changed independently [*],
63 all blocks that have been changed on either replica in the mean time.
64 [*] usually as a result of a cluster split-brain and insufficient protection.
65 but there are valid use cases to do this on purpose.
66
67 Tracking those blocks can be implemented as "dirty bitmap".
68 Having it fine-grained reduces the amount of resync traffic.
69 It should also be persistent, to allow for reboots (or crashes)
70 while the replication link is down.
71
72There are various possible implementations for persistently storing
73write intent log information, three of which are mentioned here.
74
75"Chunk dirtying"
76 The on-disk "dirty bitmap" may be re-used as "write-intent" bitmap as well.
77 To reduce the frequency of bitmap updates for write-intent log purposes,
78 one could dirty "chunks" (of some size) at a time of the (fine grained)
79 on-disk bitmap, while keeping the in-memory "dirty" bitmap as clean as
80 possible, flushing it to disk again when a previously "hot" (and on-disk
81 dirtied as full chunk) area "cools down" again (no IO in flight anymore,
82 and none expected in the near future either).
83
84"Explicit (coarse) write intent bitmap"
85 An other implementation could chose a (probably coarse) explicit bitmap,
86 for write-intent log purposes, additionally to the fine grained dirty bitmap.
87
88"Activity log"
89 Yet an other implementation may keep track of the hot regions, by starting
90 with an empty set, and writing down a journal of region numbers that have
91 become "hot", or have "cooled down" again.
92
93 To be able to use a ring buffer for this journal of changes to the active
94 set, we not only record the actual changes to that set, but also record the
95 not changing members of the set in a round robin fashion. To do so, we use a
96 fixed (but configurable) number of slots which we can identify by index, and
97 associate region numbers (labels) with these indices.
98 For each transaction recording a change to the active set, we record the
99 change itself (index: -old_label, +new_label), and which index is associated
100 with which label (index: current_label) within a certain sliding window that
101 is moved further over the available indices with each such transaction.
102
103 Thus, for crash recovery, if the ringbuffer is sufficiently large, we can
104 accurately reconstruct the active set.
105
106 Sufficiently large depends only on maximum number of active objects, and the
107 size of the sliding window recording "index: current_label" associations within
108 each transaction.
109
110 This is what we call the "activity log".
111
112 Currently we need one activity log transaction per single label change, which
113 does not give much benefit over the "dirty chunks of bitmap" approach, other
114 than potentially less seeks.
115
116 We plan to change the transaction format to support multiple changes per
117 transaction, which then would reduce several (disjoint, "random") updates to
118 the bitmap into one transaction to the activity log ring buffer.
119*/
120
121/* this defines an element in a tracked set
122 * .colision is for hash table lookup.
123 * When we process a new IO request, we know its sector, thus can deduce the
124 * region number (label) easily. To do the label -> object lookup without a
125 * full list walk, we use a simple hash table.
126 *
127 * .list is on one of three lists:
128 * in_use: currently in use (refcnt > 0, lc_number != LC_FREE)
129 * lru: unused but ready to be reused or recycled
130 * (lc_refcnt == 0, lc_number != LC_FREE),
131 * free: unused but ready to be recycled
132 * (lc_refcnt == 0, lc_number == LC_FREE),
133 *
134 * an element is said to be "in the active set",
135 * if either on "in_use" or "lru", i.e. lc_number != LC_FREE.
136 *
137 * DRBD currently (May 2009) only uses 61 elements on the resync lru_cache
138 * (total memory usage 2 pages), and up to 3833 elements on the act_log
139 * lru_cache, totalling ~215 kB for 64bit architecture, ~53 pages.
140 *
141 * We usually do not actually free these objects again, but only "recycle"
142 * them, as the change "index: -old_label, +LC_FREE" would need a transaction
143 * as well. Which also means that using a kmem_cache to allocate the objects
144 * from wastes some resources.
145 * But it avoids high order page allocations in kmalloc.
146 */
147struct lc_element {
148 struct hlist_node colision;
149 struct list_head list; /* LRU list or free list */
150 unsigned refcnt;
151 /* back "pointer" into lc_cache->element[index],
152 * for paranoia, and for "lc_element_to_index" */
153 unsigned lc_index;
154 /* if we want to track a larger set of objects,
155 * it needs to become an architecture independent u64 */
156 unsigned lc_number;
157 /* special label when on free list */
158#define LC_FREE (~0U)
159
160 /* for pending changes */
161 unsigned lc_new_number;
162};
163
164struct lru_cache {
165 /* the least recently used item is kept at lru->prev */
166 struct list_head lru;
167 struct list_head free;
168 struct list_head in_use;
169 struct list_head to_be_changed;
170
171 /* the pre-created kmem cache to allocate the objects from */
172 struct kmem_cache *lc_cache;
173
174 /* size of tracked objects, used to memset(,0,) them in lc_reset */
175 size_t element_size;
176 /* offset of struct lc_element member in the tracked object */
177 size_t element_off;
178
179 /* number of elements (indices) */
180 unsigned int nr_elements;
181 /* Arbitrary limit on maximum tracked objects. Practical limit is much
182 * lower due to allocation failures, probably. For typical use cases,
183 * nr_elements should be a few thousand at most.
184 * This also limits the maximum value of lc_element.lc_index, allowing the
185 * 8 high bits of .lc_index to be overloaded with flags in the future. */
186#define LC_MAX_ACTIVE (1<<24)
187
188 /* allow to accumulate a few (index:label) changes,
189 * but no more than max_pending_changes */
190 unsigned int max_pending_changes;
191 /* number of elements currently on to_be_changed list */
192 unsigned int pending_changes;
193
194 /* statistics */
195 unsigned used; /* number of elements currently on in_use list */
196 unsigned long hits, misses, starving, locked, changed;
197
198 /* see below: flag-bits for lru_cache */
199 unsigned long flags;
200
201
202 const char *name;
203
204 /* nr_elements there */
205 struct hlist_head *lc_slot;
206 struct lc_element **lc_element;
207};
208
209
210/* flag-bits for lru_cache */
211enum {
212 /* debugging aid, to catch concurrent access early.
213 * user needs to guarantee exclusive access by proper locking! */
214 __LC_PARANOIA,
215
216 /* annotate that the set is "dirty", possibly accumulating further
217 * changes, until a transaction is finally triggered */
218 __LC_DIRTY,
219
220 /* Locked, no further changes allowed.
221 * Also used to serialize changing transactions. */
222 __LC_LOCKED,
223
224 /* if we need to change the set, but currently there is no free nor
225 * unused element available, we are "starving", and must not give out
226 * further references, to guarantee that eventually some refcnt will
227 * drop to zero and we will be able to make progress again, changing
228 * the set, writing the transaction.
229 * if the statistics say we are frequently starving,
230 * nr_elements is too small. */
231 __LC_STARVING,
232};
233#define LC_PARANOIA (1<<__LC_PARANOIA)
234#define LC_DIRTY (1<<__LC_DIRTY)
235#define LC_LOCKED (1<<__LC_LOCKED)
236#define LC_STARVING (1<<__LC_STARVING)
237
238extern struct lru_cache *lc_create(const char *name, struct kmem_cache *cache,
239 unsigned max_pending_changes,
240 unsigned e_count, size_t e_size, size_t e_off);
241extern void lc_reset(struct lru_cache *lc);
242extern void lc_destroy(struct lru_cache *lc);
243extern void lc_del(struct lru_cache *lc, struct lc_element *element);
244
245extern struct lc_element *lc_get_cumulative(struct lru_cache *lc, unsigned int enr);
246extern struct lc_element *lc_try_get(struct lru_cache *lc, unsigned int enr);
247extern struct lc_element *lc_find(struct lru_cache *lc, unsigned int enr);
248extern struct lc_element *lc_get(struct lru_cache *lc, unsigned int enr);
249extern unsigned int lc_put(struct lru_cache *lc, struct lc_element *e);
250extern void lc_committed(struct lru_cache *lc);
251
252struct seq_file;
253extern void lc_seq_printf_stats(struct seq_file *seq, struct lru_cache *lc);
254
255extern void lc_seq_dump_details(struct seq_file *seq, struct lru_cache *lc, char *utext,
256 void (*detail) (struct seq_file *, struct lc_element *));
257
258/**
259 * lc_try_lock_for_transaction - can be used to stop lc_get() from changing the tracked set
260 * @lc: the lru cache to operate on
261 *
262 * Allows (expects) the set to be "dirty". Note that the reference counts and
263 * order on the active and lru lists may still change. Used to serialize
264 * changing transactions. Returns true if we acquired the lock.
265 */
266static inline int lc_try_lock_for_transaction(struct lru_cache *lc)
267{
268 return !test_and_set_bit(nr: __LC_LOCKED, addr: &lc->flags);
269}
270
271/**
272 * lc_try_lock - variant to stop lc_get() from changing the tracked set
273 * @lc: the lru cache to operate on
274 *
275 * Note that the reference counts and order on the active and lru lists may
276 * still change. Only works on a "clean" set. Returns true if we acquired the
277 * lock, which means there are no pending changes, and any further attempt to
278 * change the set will not succeed until the next lc_unlock().
279 */
280extern int lc_try_lock(struct lru_cache *lc);
281
282/**
283 * lc_unlock - unlock @lc, allow lc_get() to change the set again
284 * @lc: the lru cache to operate on
285 */
286static inline void lc_unlock(struct lru_cache *lc)
287{
288 clear_bit(nr: __LC_DIRTY, addr: &lc->flags);
289 clear_bit_unlock(nr: __LC_LOCKED, addr: &lc->flags);
290}
291
292extern bool lc_is_used(struct lru_cache *lc, unsigned int enr);
293
294#define lc_entry(ptr, type, member) \
295 container_of(ptr, type, member)
296
297extern struct lc_element *lc_element_by_index(struct lru_cache *lc, unsigned i);
298
299#endif
300

source code of linux/include/linux/lru_cache.h