xref: /openbmc/qemu/block/qed-l2-cache.c (revision 2df1eb27)
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 "qemu/memalign.h"
55 #include "trace.h"
56 #include "qed.h"
57 
58 /* Each L2 holds 2GB so this let's us fully cache a 100GB disk */
59 #define MAX_L2_CACHE_SIZE 50
60 
61 /**
62  * Initialize the L2 cache
63  */
64 void qed_init_l2_cache(L2TableCache *l2_cache)
65 {
66     QTAILQ_INIT(&l2_cache->entries);
67     l2_cache->n_entries = 0;
68 }
69 
70 /**
71  * Free the L2 cache
72  */
73 void qed_free_l2_cache(L2TableCache *l2_cache)
74 {
75     CachedL2Table *entry, *next_entry;
76 
77     QTAILQ_FOREACH_SAFE(entry, &l2_cache->entries, node, next_entry) {
78         qemu_vfree(entry->table);
79         g_free(entry);
80     }
81 }
82 
83 /**
84  * Allocate an uninitialized entry from the cache
85  *
86  * The returned entry has a reference count of 1 and is owned by the caller.
87  * The caller must allocate the actual table field for this entry and it must
88  * be freeable using qemu_vfree().
89  */
90 CachedL2Table *qed_alloc_l2_cache_entry(L2TableCache *l2_cache)
91 {
92     CachedL2Table *entry;
93 
94     entry = g_malloc0(sizeof(*entry));
95     entry->ref++;
96 
97     trace_qed_alloc_l2_cache_entry(l2_cache, entry);
98 
99     return entry;
100 }
101 
102 /**
103  * Decrease an entry's reference count and free if necessary when the reference
104  * count drops to zero.
105  *
106  * Called with table_lock held.
107  */
108 void qed_unref_l2_cache_entry(CachedL2Table *entry)
109 {
110     if (!entry) {
111         return;
112     }
113 
114     entry->ref--;
115     trace_qed_unref_l2_cache_entry(entry, entry->ref);
116     if (entry->ref == 0) {
117         qemu_vfree(entry->table);
118         g_free(entry);
119     }
120 }
121 
122 /**
123  * Find an entry in the L2 cache.  This may return NULL and it's up to the
124  * caller to satisfy the cache miss.
125  *
126  * For a cached entry, this function increases the reference count and returns
127  * the entry.
128  *
129  * Called with table_lock held.
130  */
131 CachedL2Table *qed_find_l2_cache_entry(L2TableCache *l2_cache, uint64_t offset)
132 {
133     CachedL2Table *entry;
134 
135     QTAILQ_FOREACH(entry, &l2_cache->entries, node) {
136         if (entry->offset == offset) {
137             trace_qed_find_l2_cache_entry(l2_cache, entry, offset, entry->ref);
138             entry->ref++;
139             return entry;
140         }
141     }
142     return NULL;
143 }
144 
145 /**
146  * Commit an L2 cache entry into the cache.  This is meant to be used as part of
147  * the process to satisfy a cache miss.  A caller would allocate an entry which
148  * is not actually in the L2 cache and then once the entry was valid and
149  * present on disk, the entry can be committed into the cache.
150  *
151  * Since the cache is write-through, it's important that this function is not
152  * called until the entry is present on disk and the L1 has been updated to
153  * point to the entry.
154  *
155  * N.B. This function steals a reference to the l2_table from the caller so the
156  * caller must obtain a new reference by issuing a call to
157  * qed_find_l2_cache_entry().
158  *
159  * Called with table_lock held.
160  */
161 void qed_commit_l2_cache_entry(L2TableCache *l2_cache, CachedL2Table *l2_table)
162 {
163     CachedL2Table *entry;
164 
165     entry = qed_find_l2_cache_entry(l2_cache, l2_table->offset);
166     if (entry) {
167         qed_unref_l2_cache_entry(entry);
168         qed_unref_l2_cache_entry(l2_table);
169         return;
170     }
171 
172     /* Evict an unused cache entry so we have space.  If all entries are in use
173      * we can grow the cache temporarily and we try to shrink back down later.
174      */
175     if (l2_cache->n_entries >= MAX_L2_CACHE_SIZE) {
176         CachedL2Table *next;
177         QTAILQ_FOREACH_SAFE(entry, &l2_cache->entries, node, next) {
178             if (entry->ref > 1) {
179                 continue;
180             }
181 
182             QTAILQ_REMOVE(&l2_cache->entries, entry, node);
183             l2_cache->n_entries--;
184             qed_unref_l2_cache_entry(entry);
185 
186             /* Stop evicting when we've shrunk back to max size */
187             if (l2_cache->n_entries < MAX_L2_CACHE_SIZE) {
188                 break;
189             }
190         }
191     }
192 
193     l2_cache->n_entries++;
194     QTAILQ_INSERT_TAIL(&l2_cache->entries, l2_table, node);
195 }
196