1/*
2 * QEMU Enhanced Disk Format L2 Cache
3 *
4 * Copyright IBM, Corp. 2010
5 *
6 * Authors:
7 * Anthony Liguori <aliguori@us.ibm.com>
8 *
9 * This work is licensed under the terms of the GNU LGPL, version 2 or later.
10 * See the COPYING.LIB file in the top-level directory.
11 *
12 */
13
14/*
15 * L2 table cache usage is as follows:
16 *
17 * An open image has one L2 table cache that is used to avoid accessing the
18 * image file for recently referenced L2 tables.
19 *
20 * Cluster offset lookup translates the logical offset within the block device
21 * to a cluster offset within the image file. This is done by indexing into
22 * the L1 and L2 tables which store cluster offsets. It is here where the L2
23 * table cache serves up recently referenced L2 tables.
24 *
25 * If there is a cache miss, that L2 table is read from the image file and
26 * committed to the cache. Subsequent accesses to that L2 table will be served
27 * from the cache until the table is evicted from the cache.
28 *
29 * L2 tables are also committed to the cache when new L2 tables are allocated
30 * in the image file. Since the L2 table cache is write-through, the new L2
31 * table is first written out to the image file and then committed to the
32 * cache.
33 *
34 * Multiple I/O requests may be using an L2 table cache entry at any given
35 * time. That means an entry may be in use across several requests and
36 * reference counting is needed to free the entry at the correct time. In
37 * particular, an entry evicted from the cache will only be freed once all
38 * references are dropped.
39 *
40 * An in-flight I/O request will hold a reference to a L2 table cache entry for
41 * the period during which it needs to access the L2 table. This includes
42 * cluster offset lookup, L2 table allocation, and L2 table update when a new
43 * data cluster has been allocated.
44 *
45 * An interesting case occurs when two requests need to access an L2 table that
46 * is not in the cache. Since the operation to read the table from the image
47 * file takes some time to complete, both requests may see a cache miss and
48 * start reading the L2 table from the image file. The first to finish will
49 * commit its L2 table into the cache. When the second tries to commit its
50 * table will be deleted in favor of the existing cache entry.
51 */
52
53#include "qemu/osdep.h"
54#include "trace.h"
55#include "qed.h"
56
57/* Each L2 holds 2GB so this let's us fully cache a 100GB disk */
58#define MAX_L2_CACHE_SIZE 50
59
60/**
61 * Initialize the L2 cache
62 */
63void qed_init_l2_cache(L2TableCache *l2_cache)
64{
65 QTAILQ_INIT(&l2_cache->entries);
66 l2_cache->n_entries = 0;
67}
68
69/**
70 * Free the L2 cache
71 */
72void qed_free_l2_cache(L2TableCache *l2_cache)
73{
74 CachedL2Table *entry, *next_entry;
75
76 QTAILQ_FOREACH_SAFE(entry, &l2_cache->entries, node, next_entry) {
77 qemu_vfree(entry->table);
78 g_free(entry);
79 }
80}
81
82/**
83 * Allocate an uninitialized entry from the cache
84 *
85 * The returned entry has a reference count of 1 and is owned by the caller.
86 * The caller must allocate the actual table field for this entry and it must
87 * be freeable using qemu_vfree().
88 */
89CachedL2Table *qed_alloc_l2_cache_entry(L2TableCache *l2_cache)
90{
91 CachedL2Table *entry;
92
93 entry = g_malloc0(sizeof(*entry));
94 entry->ref++;
95
96 trace_qed_alloc_l2_cache_entry(l2_cache, entry);
97
98 return entry;
99}
100
101/**
102 * Decrease an entry's reference count and free if necessary when the reference
103 * count drops to zero.
104 *
105 * Called with table_lock held.
106 */
107void qed_unref_l2_cache_entry(CachedL2Table *entry)
108{
109 if (!entry) {
110 return;
111 }
112
113 entry->ref--;
114 trace_qed_unref_l2_cache_entry(entry, entry->ref);
115 if (entry->ref == 0) {
116 qemu_vfree(entry->table);
117 g_free(entry);
118 }
119}
120
121/**
122 * Find an entry in the L2 cache. This may return NULL and it's up to the
123 * caller to satisfy the cache miss.
124 *
125 * For a cached entry, this function increases the reference count and returns
126 * the entry.
127 *
128 * Called with table_lock held.
129 */
130CachedL2Table *qed_find_l2_cache_entry(L2TableCache *l2_cache, uint64_t offset)
131{
132 CachedL2Table *entry;
133
134 QTAILQ_FOREACH(entry, &l2_cache->entries, node) {
135 if (entry->offset == offset) {
136 trace_qed_find_l2_cache_entry(l2_cache, entry, offset, entry->ref);
137 entry->ref++;
138 return entry;
139 }
140 }
141 return NULL;
142}
143
144/**
145 * Commit an L2 cache entry into the cache. This is meant to be used as part of
146 * the process to satisfy a cache miss. A caller would allocate an entry which
147 * is not actually in the L2 cache and then once the entry was valid and
148 * present on disk, the entry can be committed into the cache.
149 *
150 * Since the cache is write-through, it's important that this function is not
151 * called until the entry is present on disk and the L1 has been updated to
152 * point to the entry.
153 *
154 * N.B. This function steals a reference to the l2_table from the caller so the
155 * caller must obtain a new reference by issuing a call to
156 * qed_find_l2_cache_entry().
157 *
158 * Called with table_lock held.
159 */
160void qed_commit_l2_cache_entry(L2TableCache *l2_cache, CachedL2Table *l2_table)
161{
162 CachedL2Table *entry;
163
164 entry = qed_find_l2_cache_entry(l2_cache, l2_table->offset);
165 if (entry) {
166 qed_unref_l2_cache_entry(entry);
167 qed_unref_l2_cache_entry(l2_table);
168 return;
169 }
170
171 /* Evict an unused cache entry so we have space. If all entries are in use
172 * we can grow the cache temporarily and we try to shrink back down later.
173 */
174 if (l2_cache->n_entries >= MAX_L2_CACHE_SIZE) {
175 CachedL2Table *next;
176 QTAILQ_FOREACH_SAFE(entry, &l2_cache->entries, node, next) {
177 if (entry->ref > 1) {
178 continue;
179 }
180
181 QTAILQ_REMOVE(&l2_cache->entries, entry, node);
182 l2_cache->n_entries--;
183 qed_unref_l2_cache_entry(entry);
184
185 /* Stop evicting when we've shrunk back to max size */
186 if (l2_cache->n_entries < MAX_L2_CACHE_SIZE) {
187 break;
188 }
189 }
190 }
191
192 l2_cache->n_entries++;
193 QTAILQ_INSERT_TAIL(&l2_cache->entries, l2_table, node);
194}
195