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 "trace.h" 54 #include "qed.h" 55 56 /* Each L2 holds 2GB so this let's us fully cache a 100GB disk */ 57 #define MAX_L2_CACHE_SIZE 50 58 59 /** 60 * Initialize the L2 cache 61 */ 62 void qed_init_l2_cache(L2TableCache *l2_cache) 63 { 64 QTAILQ_INIT(&l2_cache->entries); 65 l2_cache->n_entries = 0; 66 } 67 68 /** 69 * Free the L2 cache 70 */ 71 void qed_free_l2_cache(L2TableCache *l2_cache) 72 { 73 CachedL2Table *entry, *next_entry; 74 75 QTAILQ_FOREACH_SAFE(entry, &l2_cache->entries, node, next_entry) { 76 qemu_vfree(entry->table); 77 g_free(entry); 78 } 79 } 80 81 /** 82 * Allocate an uninitialized entry from the cache 83 * 84 * The returned entry has a reference count of 1 and is owned by the caller. 85 * The caller must allocate the actual table field for this entry and it must 86 * be freeable using qemu_vfree(). 87 */ 88 CachedL2Table *qed_alloc_l2_cache_entry(L2TableCache *l2_cache) 89 { 90 CachedL2Table *entry; 91 92 entry = g_malloc0(sizeof(*entry)); 93 entry->ref++; 94 95 trace_qed_alloc_l2_cache_entry(l2_cache, entry); 96 97 return entry; 98 } 99 100 /** 101 * Decrease an entry's reference count and free if necessary when the reference 102 * count drops to zero. 103 */ 104 void qed_unref_l2_cache_entry(CachedL2Table *entry) 105 { 106 if (!entry) { 107 return; 108 } 109 110 entry->ref--; 111 trace_qed_unref_l2_cache_entry(entry, entry->ref); 112 if (entry->ref == 0) { 113 qemu_vfree(entry->table); 114 g_free(entry); 115 } 116 } 117 118 /** 119 * Find an entry in the L2 cache. This may return NULL and it's up to the 120 * caller to satisfy the cache miss. 121 * 122 * For a cached entry, this function increases the reference count and returns 123 * the entry. 124 */ 125 CachedL2Table *qed_find_l2_cache_entry(L2TableCache *l2_cache, uint64_t offset) 126 { 127 CachedL2Table *entry; 128 129 QTAILQ_FOREACH(entry, &l2_cache->entries, node) { 130 if (entry->offset == offset) { 131 trace_qed_find_l2_cache_entry(l2_cache, entry, offset, entry->ref); 132 entry->ref++; 133 return entry; 134 } 135 } 136 return NULL; 137 } 138 139 /** 140 * Commit an L2 cache entry into the cache. This is meant to be used as part of 141 * the process to satisfy a cache miss. A caller would allocate an entry which 142 * is not actually in the L2 cache and then once the entry was valid and 143 * present on disk, the entry can be committed into the cache. 144 * 145 * Since the cache is write-through, it's important that this function is not 146 * called until the entry is present on disk and the L1 has been updated to 147 * point to the entry. 148 * 149 * N.B. This function steals a reference to the l2_table from the caller so the 150 * caller must obtain a new reference by issuing a call to 151 * qed_find_l2_cache_entry(). 152 */ 153 void qed_commit_l2_cache_entry(L2TableCache *l2_cache, CachedL2Table *l2_table) 154 { 155 CachedL2Table *entry; 156 157 entry = qed_find_l2_cache_entry(l2_cache, l2_table->offset); 158 if (entry) { 159 qed_unref_l2_cache_entry(entry); 160 qed_unref_l2_cache_entry(l2_table); 161 return; 162 } 163 164 /* Evict an unused cache entry so we have space. If all entries are in use 165 * we can grow the cache temporarily and we try to shrink back down later. 166 */ 167 if (l2_cache->n_entries >= MAX_L2_CACHE_SIZE) { 168 CachedL2Table *next; 169 QTAILQ_FOREACH_SAFE(entry, &l2_cache->entries, node, next) { 170 if (entry->ref > 1) { 171 continue; 172 } 173 174 QTAILQ_REMOVE(&l2_cache->entries, entry, node); 175 l2_cache->n_entries--; 176 qed_unref_l2_cache_entry(entry); 177 178 /* Stop evicting when we've shrunk back to max size */ 179 if (l2_cache->n_entries < MAX_L2_CACHE_SIZE) { 180 break; 181 } 182 } 183 } 184 185 l2_cache->n_entries++; 186 QTAILQ_INSERT_TAIL(&l2_cache->entries, l2_table, node); 187 } 188