|
| 1 | +#include <stdbool.h> |
| 2 | +#include "sds.h" |
| 3 | +#include "server.h" |
| 4 | +#include "serverassert.h" |
| 5 | +#include "entry.h" |
| 6 | + |
| 7 | +#include <stdbool.h> |
| 8 | + |
| 9 | +/*----------------------------------------------------------------------------- |
| 10 | + * Hash Entry API |
| 11 | + *----------------------------------------------------------------------------*/ |
| 12 | + |
| 13 | +/* The entry pointer is the field sds. We encode the entry layout type |
| 14 | + * in the field SDS header. Field type SDS_TYPE_5 doesn't have any spare bits to |
| 15 | + * encode this so we use it only for the first layout type. |
| 16 | + * |
| 17 | + * Entry with embedded value, used for small sizes. The value is stored as |
| 18 | + * SDS_TYPE_8. The field can use any SDS type. |
| 19 | + * |
| 20 | + * +--------------+---------------+ |
| 21 | + * | field | value | |
| 22 | + * | hdr "foo" \0 | hdr8 "bar" \0 | |
| 23 | + * +------^-------+---------------+ |
| 24 | + * | |
| 25 | + * | |
| 26 | + * entry pointer = field sds |
| 27 | + * |
| 28 | + * Entry with value pointer, used for larger fields and values. The field is SDS |
| 29 | + * type 8 or higher. |
| 30 | + * |
| 31 | + * +-------+--------------+ |
| 32 | + * | value | field | |
| 33 | + * | ptr | hdr "foo" \0 | |
| 34 | + * +-------+------^-------+ |
| 35 | + * | |
| 36 | + * | |
| 37 | + * entry pointer = field sds |
| 38 | + */ |
| 39 | + |
| 40 | +/* The maximum allocation size we want to use for entries with embedded |
| 41 | + * values. */ |
| 42 | +#define EMBED_VALUE_MAX_ALLOC_SIZE 128 |
| 43 | + |
| 44 | +/* SDS aux flag. If set, it indicates that the entry has TTL metadata set. */ |
| 45 | +#define FIELD_SDS_AUX_BIT_ENTRY_HAS_TTL 0 |
| 46 | + |
| 47 | +/* SDS aux flag. If set, it indicates that the entry has an embedded value |
| 48 | + * pointer located in memory before the embedded field. If unset, the entry |
| 49 | + * instead has an embedded value located after the embedded field. */ |
| 50 | +#define FIELD_SDS_AUX_BIT_ENTRY_HAS_VALUE_PTR 2 |
| 51 | + |
| 52 | +bool entryHasValuePtr(const entry *entry) { |
| 53 | + return sdsGetAuxBit(entry, FIELD_SDS_AUX_BIT_ENTRY_HAS_VALUE_PTR); |
| 54 | +} |
| 55 | + |
| 56 | +bool entryHasExpiry(const entry *entry) { |
| 57 | + return sdsGetAuxBit(entry, FIELD_SDS_AUX_BIT_ENTRY_HAS_TTL); |
| 58 | +} |
| 59 | + |
| 60 | +/* Returns the location of a pointer to a separately allocated value. Only for |
| 61 | + * an entry without an embedded value. */ |
| 62 | +sds *entryGetValueRef(const entry *entry) { |
| 63 | + serverAssert(entryHasValuePtr(entry)); |
| 64 | + char *field_data = sdsAllocPtr(entry); |
| 65 | + field_data -= sizeof(sds *); |
| 66 | + return (sds *)field_data; |
| 67 | +} |
| 68 | + |
| 69 | +/* The entry pointer is the field sds, but that's an implementation detail. */ |
| 70 | +sds entryGetField(const entry *entry) { |
| 71 | + return (sds)entry; |
| 72 | +} |
| 73 | + |
| 74 | +sds entryGetValue(const entry *entry) { |
| 75 | + if (entryHasValuePtr(entry)) { |
| 76 | + return *entryGetValueRef(entry); |
| 77 | + } else { |
| 78 | + /* Skip field content, field null terminator and value sds8 hdr. */ |
| 79 | + size_t offset = sdslen(entry) + 1 + sdsHdrSize(SDS_TYPE_8); |
| 80 | + serverAssert((char *)entry + offset); |
| 81 | + |
| 82 | + return (char *)entry + offset; |
| 83 | + } |
| 84 | +} |
| 85 | + |
| 86 | +entry *entrySetValue(entry *e, sds value) { |
| 87 | + if (entryHasValuePtr(e)) { |
| 88 | + sds *value_ref = entryGetValueRef(e); |
| 89 | + sdsfree(*value_ref); |
| 90 | + *value_ref = value; |
| 91 | + return e; |
| 92 | + } else { |
| 93 | + entry *new_entry = entryUpdate(e, value, entryGetExpiry(e)); |
| 94 | + return new_entry; |
| 95 | + } |
| 96 | +} |
| 97 | + |
| 98 | +/* Returns the address of the entry allocation. */ |
| 99 | +void *entryAllocPtr(const entry *entry) { |
| 100 | + char *buf = sdsAllocPtr(entry); |
| 101 | + if (entryHasValuePtr(entry)) buf -= sizeof(sds *); |
| 102 | + if (entryHasExpiry(entry)) buf -= sizeof(long long); |
| 103 | + return buf; |
| 104 | +} |
| 105 | + |
| 106 | +/**************************************** Entry Expiry API *****************************************/ |
| 107 | +long long entryGetExpiry(const entry *entry) { |
| 108 | + long long expiry = EXPIRY_NONE; |
| 109 | + if (entryHasExpiry(entry)) { |
| 110 | + char *buf = sdsAllocPtr(entry); |
| 111 | + if (entryHasValuePtr(entry)) buf -= sizeof(sds *); |
| 112 | + buf -= sizeof(expiry); |
| 113 | + memcpy(&expiry, buf, sizeof(expiry)); |
| 114 | + } |
| 115 | + return expiry; |
| 116 | +} |
| 117 | + |
| 118 | +entry *entrySetExpiry(entry *e, long long expiry) { |
| 119 | + if (entryHasExpiry(e)) { |
| 120 | + char *buf = sdsAllocPtr(e); |
| 121 | + if (entryHasValuePtr(e)) buf -= sizeof(sds *); |
| 122 | + buf -= sizeof(expiry); |
| 123 | + memcpy(buf, &expiry, sizeof(expiry)); |
| 124 | + return e; |
| 125 | + } |
| 126 | + entry *new_entry = entryUpdate(e, NULL, expiry); |
| 127 | + return new_entry; |
| 128 | +} |
| 129 | + |
| 130 | +int entryIsExpired(entry *entry) { |
| 131 | + /* Don't expire anything while loading. It will be done later. */ |
| 132 | + if (server.loading) return 0; |
| 133 | + if (!timestampIsExpired(entryGetExpiry(entry))) return 0; |
| 134 | + if (server.primary_host == NULL && server.import_mode) { |
| 135 | + if (server.current_client && server.current_client->flag.import_source) return 0; |
| 136 | + } |
| 137 | + return 1; |
| 138 | +} |
| 139 | +/**************************************** Entry Expiry API - End *****************************************/ |
| 140 | + |
| 141 | +void freeentry(entry *entry) { |
| 142 | + if (entryHasValuePtr(entry)) { |
| 143 | + sdsfree(*entryGetValueRef(entry)); |
| 144 | + } |
| 145 | + zfree(entryAllocPtr(entry)); |
| 146 | +} |
| 147 | + |
| 148 | +/* takes ownership of value, does not take ownership of field */ |
| 149 | +entry *entryCreate(sds field, sds value, long long expiry) { |
| 150 | + /* In case simple sds just return the same field we got. */ |
| 151 | + if (!value && expiry == EXPIRY_NONE) |
| 152 | + return field; |
| 153 | + sds embedded_field_sds; |
| 154 | + size_t expiry_size = (expiry == EXPIRY_NONE) ? 0 : sizeof(long long); |
| 155 | + size_t field_len = sdslen(field); |
| 156 | + int field_sds_type = sdsReqType(field_len); |
| 157 | + if (field_sds_type == SDS_TYPE_5 && (expiry_size > 0)) { |
| 158 | + field_sds_type = SDS_TYPE_8; |
| 159 | + } |
| 160 | + size_t field_size = sdsReqSize(field_len, field_sds_type); |
| 161 | + size_t value_len = value ? sdslen(value) : 0; |
| 162 | + size_t value_size = value ? sdsReqSize(value_len, SDS_TYPE_8) : 0; |
| 163 | + size_t alloc_size = field_size + expiry_size; |
| 164 | + bool embed_value = false; |
| 165 | + if (value) { |
| 166 | + if (alloc_size + value_size <= EMBED_VALUE_MAX_ALLOC_SIZE) { |
| 167 | + /* Embed field and value. Value is fixed to SDS_TYPE_8. Unused |
| 168 | + * allocation space is recorded in the embedded value's SDS header. |
| 169 | + * |
| 170 | + * +------+--------------+---------------+ |
| 171 | + * | TTL | field | value | |
| 172 | + * | | hdr "foo" \0 | hdr8 "bar" \0 | |
| 173 | + * +------+--------------+---------------+ |
| 174 | + */ |
| 175 | + embed_value = true; |
| 176 | + alloc_size += value_size; |
| 177 | + } else { |
| 178 | + /* Embed field, but not value. Field must be >= SDS_TYPE_8 to encode to |
| 179 | + * indicate this type of entry. |
| 180 | + * |
| 181 | + * +------+-------+---------------+ |
| 182 | + * | TTL | value | field | |
| 183 | + * | | ptr | hdr8 "foo" \0 | |
| 184 | + * +------+-------+---------------+ |
| 185 | + */ |
| 186 | + embed_value = false; |
| 187 | + alloc_size += sizeof(sds *); |
| 188 | + if (field_sds_type == SDS_TYPE_5) { |
| 189 | + field_sds_type = SDS_TYPE_8; |
| 190 | + alloc_size -= field_size; |
| 191 | + field_size = sdsReqSize(field_len, field_sds_type); |
| 192 | + alloc_size += field_size; |
| 193 | + } |
| 194 | + } |
| 195 | + } |
| 196 | + /* allocate the buffer */ |
| 197 | + size_t buf_size; |
| 198 | + char *buf = zmalloc_usable(alloc_size, &buf_size); |
| 199 | + |
| 200 | + /* Set The expiry if exists */ |
| 201 | + if (expiry_size) { |
| 202 | + memcpy(buf, &expiry, expiry_size); |
| 203 | + buf += expiry_size; |
| 204 | + buf_size -= expiry_size; |
| 205 | + } |
| 206 | + if (value) { |
| 207 | + if (!embed_value) { |
| 208 | + *(sds *)buf = value; |
| 209 | + buf += sizeof(sds *); |
| 210 | + buf_size -= sizeof(sds *); |
| 211 | + } else { |
| 212 | + sdswrite(buf + field_size, buf_size - field_size, SDS_TYPE_8, value, value_len); |
| 213 | + sdsfree(value); |
| 214 | + buf_size -= value_size; |
| 215 | + } |
| 216 | + } |
| 217 | + /* Set the field data */ |
| 218 | + embedded_field_sds = sdswrite(buf, field_size, field_sds_type, field, field_len); |
| 219 | + |
| 220 | + /* Field sds aux bits are zero, which we use for this entry encoding. */ |
| 221 | + sdsSetAuxBit(embedded_field_sds, FIELD_SDS_AUX_BIT_ENTRY_HAS_VALUE_PTR, embed_value ? 0 : 1); |
| 222 | + sdsSetAuxBit(embedded_field_sds, FIELD_SDS_AUX_BIT_ENTRY_HAS_TTL, expiry_size > 0 ? 1 : 0); |
| 223 | + return (void *)embedded_field_sds; |
| 224 | +} |
| 225 | + |
| 226 | +/* Frees previous value, takes ownership of new value, returns entry (may be |
| 227 | + * reallocated). */ |
| 228 | +entry *entryUpdate(entry *e, sds value, long long expiry) { |
| 229 | + sds field = (sds)e; |
| 230 | + |
| 231 | + bool update_value = value ? true : false; |
| 232 | + long long ttl = entryGetExpiry(e); |
| 233 | + bool update_expiry = (expiry != ttl) ? true : false; |
| 234 | + if (!update_value && !update_expiry) |
| 235 | + return e; |
| 236 | + if (update_expiry) ttl = expiry; |
| 237 | + value = update_value ? value : entryGetValue(e); |
| 238 | + size_t expiry_size = ttl != EXPIRY_NONE ? sizeof(ttl) : 0; |
| 239 | + int field_sds_type = sdsReqType(sdslen(field)); |
| 240 | + if (field_sds_type == SDS_TYPE_5 && (expiry_size > 0)) { |
| 241 | + field_sds_type = SDS_TYPE_8; |
| 242 | + } |
| 243 | + size_t field_size = sdsHdrSize(field_sds_type) + sdsalloc(field) + 1; |
| 244 | + size_t value_len = value ? sdslen(value) : 0; |
| 245 | + size_t value_size = value ? sdsReqSize(value_len, SDS_TYPE_8) : 0; |
| 246 | + |
| 247 | + size_t required_size = field_size + value_size + expiry_size; |
| 248 | + size_t current_allocation_size = entryMemUsage(e); |
| 249 | + bool create_new_entry = (update_expiry && (entryGetExpiry(e) == EXPIRY_NONE || ttl == EXPIRY_NONE)) || |
| 250 | + !(update_value && !entryHasValuePtr(e) && |
| 251 | + required_size <= EMBED_VALUE_MAX_ALLOC_SIZE && |
| 252 | + required_size <= current_allocation_size && |
| 253 | + required_size >= current_allocation_size * 3 / 4); |
| 254 | + |
| 255 | + if (!create_new_entry) { |
| 256 | + /* In this case we are sure we do not have to allocate new entry, so expiry must already be set. */ |
| 257 | + if (update_expiry) { |
| 258 | + serverAssert(entryHasExpiry(e)); |
| 259 | + char *buf = sdsAllocPtr(e); |
| 260 | + if (entryHasValuePtr(e)) buf -= sizeof(sds *); |
| 261 | + buf -= sizeof(expiry); |
| 262 | + memcpy(buf, &expiry, sizeof(expiry)); |
| 263 | + } |
| 264 | + /* In this case we are sure we do not have to allocate new entry, so value must already be set or we have enough room to embed it. */ |
| 265 | + if (update_value) { |
| 266 | + if (entryHasValuePtr(e)) { |
| 267 | + sds *value_ref = entryGetValueRef(e); |
| 268 | + sdsfree(*value_ref); |
| 269 | + *value_ref = value; |
| 270 | + } else { |
| 271 | + /* Skip field content, field null terminator and value sds8 hdr. */ |
| 272 | + sds old_value = entryGetValue(e); |
| 273 | + sdswrite(sdsAllocPtr(old_value), sdsAllocSize(old_value), SDS_TYPE_8, value, sdslen(value)); |
| 274 | + sdsfree(value); |
| 275 | + } |
| 276 | + } |
| 277 | + return e; |
| 278 | + |
| 279 | + } else { |
| 280 | + /* Check if the value can be reused. */ |
| 281 | + int value_was_embedded = !entryHasValuePtr(e); |
| 282 | + /* In case the original entry value is embedded and we know the destination entry will be able to embed the value |
| 283 | + * We should duplicate the value. |
| 284 | + * TODO: We could basically optimize this case better by signaling entryCreate to take the value and avoid freeing it. */ |
| 285 | + if (value_was_embedded && required_size <= EMBED_VALUE_MAX_ALLOC_SIZE) |
| 286 | + value = sdsdup(value); |
| 287 | + /* if not we have to duplicate it, remove it from the original entry since we are going to delete it.*/ |
| 288 | + else if (!value_was_embedded) { |
| 289 | + sds *value_ref = entryGetValueRef(e); |
| 290 | + *value_ref = NULL; |
| 291 | + } |
| 292 | + } |
| 293 | + |
| 294 | + entry *new_entry = entryCreate(entryGetField(e), value, ttl); |
| 295 | + if (new_entry != e) |
| 296 | + freeentry(e); |
| 297 | + return new_entry; |
| 298 | +} |
| 299 | + |
| 300 | +/* Returns memory usage of a entry, including all allocations owned by |
| 301 | + * the entry. */ |
| 302 | +size_t entryMemUsage(entry *entry) { |
| 303 | + size_t mem = 0; |
| 304 | + |
| 305 | + if (entryHasValuePtr(entry)) { |
| 306 | + /* In case the value is not embedded we might not be able to sum all the allocation sizes since the field |
| 307 | + * header could be too small for holding the real allocation size. */ |
| 308 | + mem += zmalloc_usable_size(entryAllocPtr(entry)); |
| 309 | + } else { |
| 310 | + mem += sdsReqSize(sdslen(entry), sdsType(entry)); |
| 311 | + if (entryHasExpiry(entry)) mem += sizeof(long long); |
| 312 | + } |
| 313 | + mem += sdsAllocSize(entryGetValue(entry)); |
| 314 | + return mem; |
| 315 | +} |
| 316 | + |
| 317 | +/* Defragments a hashtable entry (field-value pair) if needed, using the |
| 318 | + * provided defrag functions. The defrag functions return NULL if the allocation |
| 319 | + * was not moved, otherwise they return a pointer to the new memory location. |
| 320 | + * A separate sds defrag function is needed because of the unique memory layout |
| 321 | + * of sds strings. |
| 322 | + * If the location of the entry changed we return the new location, |
| 323 | + * otherwise we return NULL. */ |
| 324 | +entry *entryDefrag(entry *entry, void *(*defragfn)(void *), sds (*sdsdefragfn)(sds)) { |
| 325 | + if (entryHasValuePtr(entry)) { |
| 326 | + sds *value_ref = entryGetValueRef(entry); |
| 327 | + sds new_value = sdsdefragfn(*value_ref); |
| 328 | + if (new_value) *value_ref = new_value; |
| 329 | + } |
| 330 | + char *allocation = entryAllocPtr(entry); |
| 331 | + char *new_allocation = defragfn(allocation); |
| 332 | + if (new_allocation != NULL) { |
| 333 | + /* Return the same offset into the new allocation as the entry's offset |
| 334 | + * in the old allocation. */ |
| 335 | + return new_allocation + ((char *)entry - allocation); |
| 336 | + } |
| 337 | + return NULL; |
| 338 | +} |
| 339 | + |
| 340 | +/* Used for releasing memory to OS to avoid unnecessary CoW. Called when we've |
| 341 | + * forked and memory won't be used again. See zmadvise_dontneed() */ |
| 342 | +void dismissEntry(entry *entry) { |
| 343 | + /* Only dismiss values memory since the field size usually is small. */ |
| 344 | + if (entryHasValuePtr(entry)) { |
| 345 | + dismissSds(*entryGetValueRef(entry)); |
| 346 | + } |
| 347 | +} |
0 commit comments