1 #ifndef JEMALLOC_INTERNAL_TSD_H 2 #define JEMALLOC_INTERNAL_TSD_H 3 4 #include "jemalloc/internal/arena_types.h" 5 #include "jemalloc/internal/assert.h" 6 #include "jemalloc/internal/jemalloc_internal_externs.h" 7 #include "jemalloc/internal/prof_types.h" 8 #include "jemalloc/internal/ql.h" 9 #include "jemalloc/internal/rtree_tsd.h" 10 #include "jemalloc/internal/tcache_types.h" 11 #include "jemalloc/internal/tcache_structs.h" 12 #include "jemalloc/internal/util.h" 13 #include "jemalloc/internal/witness.h" 14 15 /* 16 * Thread-Specific-Data layout 17 * --- data accessed on tcache fast path: state, rtree_ctx, stats, prof --- 18 * s: state 19 * e: tcache_enabled 20 * m: thread_allocated (config_stats) 21 * f: thread_deallocated (config_stats) 22 * p: prof_tdata (config_prof) 23 * c: rtree_ctx (rtree cache accessed on deallocation) 24 * t: tcache 25 * --- data not accessed on tcache fast path: arena-related fields --- 26 * d: arenas_tdata_bypass 27 * r: reentrancy_level 28 * x: narenas_tdata 29 * i: iarena 30 * a: arena 31 * o: arenas_tdata 32 * Loading TSD data is on the critical path of basically all malloc operations. 33 * In particular, tcache and rtree_ctx rely on hot CPU cache to be effective. 34 * Use a compact layout to reduce cache footprint. 35 * +--- 64-bit and 64B cacheline; 1B each letter; First byte on the left. ---+ 36 * |---------------------------- 1st cacheline ----------------------------| 37 * | sedrxxxx mmmmmmmm ffffffff pppppppp [c * 32 ........ ........ .......] | 38 * |---------------------------- 2nd cacheline ----------------------------| 39 * | [c * 64 ........ ........ ........ ........ ........ ........ .......] | 40 * |---------------------------- 3nd cacheline ----------------------------| 41 * | [c * 32 ........ ........ .......] iiiiiiii aaaaaaaa oooooooo [t...... | 42 * +-------------------------------------------------------------------------+ 43 * Note: the entire tcache is embedded into TSD and spans multiple cachelines. 44 * 45 * The last 3 members (i, a and o) before tcache isn't really needed on tcache 46 * fast path. However we have a number of unused tcache bins and witnesses 47 * (never touched unless config_debug) at the end of tcache, so we place them 48 * there to avoid breaking the cachelines and possibly paging in an extra page. 49 */ 50 #ifdef JEMALLOC_JET 51 typedef void (*test_callback_t)(int *); 52 # define MALLOC_TSD_TEST_DATA_INIT 0x72b65c10 53 # define MALLOC_TEST_TSD \ 54 O(test_data, int, int) \ 55 O(test_callback, test_callback_t, int) 56 # define MALLOC_TEST_TSD_INITIALIZER , MALLOC_TSD_TEST_DATA_INIT, NULL 57 #else 58 # define MALLOC_TEST_TSD 59 # define MALLOC_TEST_TSD_INITIALIZER 60 #endif 61 62 /* O(name, type, nullable type */ 63 #define MALLOC_TSD \ 64 O(tcache_enabled, bool, bool) \ 65 O(arenas_tdata_bypass, bool, bool) \ 66 O(reentrancy_level, int8_t, int8_t) \ 67 O(narenas_tdata, uint32_t, uint32_t) \ 68 O(offset_state, uint64_t, uint64_t) \ 69 O(thread_allocated, uint64_t, uint64_t) \ 70 O(thread_deallocated, uint64_t, uint64_t) \ 71 O(prof_tdata, prof_tdata_t *, prof_tdata_t *) \ 72 O(rtree_ctx, rtree_ctx_t, rtree_ctx_t) \ 73 O(iarena, arena_t *, arena_t *) \ 74 O(arena, arena_t *, arena_t *) \ 75 O(arenas_tdata, arena_tdata_t *, arena_tdata_t *)\ 76 O(tcache, tcache_t, tcache_t) \ 77 O(witness_tsd, witness_tsd_t, witness_tsdn_t) \ 78 MALLOC_TEST_TSD 79 80 #define TSD_INITIALIZER { \ 81 tsd_state_uninitialized, \ 82 TCACHE_ENABLED_ZERO_INITIALIZER, \ 83 false, \ 84 0, \ 85 0, \ 86 0, \ 87 0, \ 88 0, \ 89 NULL, \ 90 RTREE_CTX_ZERO_INITIALIZER, \ 91 NULL, \ 92 NULL, \ 93 NULL, \ 94 TCACHE_ZERO_INITIALIZER, \ 95 WITNESS_TSD_INITIALIZER \ 96 MALLOC_TEST_TSD_INITIALIZER \ 97 } 98 99 enum { 100 tsd_state_nominal = 0, /* Common case --> jnz. */ 101 tsd_state_nominal_slow = 1, /* Initialized but on slow path. */ 102 /* the above 2 nominal states should be lower values. */ 103 tsd_state_nominal_max = 1, /* used for comparison only. */ 104 tsd_state_minimal_initialized = 2, 105 tsd_state_purgatory = 3, 106 tsd_state_reincarnated = 4, 107 tsd_state_uninitialized = 5 108 }; 109 110 /* Manually limit tsd_state_t to a single byte. */ 111 typedef uint8_t tsd_state_t; 112 113 /* The actual tsd. */ 114 struct tsd_s { 115 /* 116 * The contents should be treated as totally opaque outside the tsd 117 * module. Access any thread-local state through the getters and 118 * setters below. 119 */ 120 tsd_state_t state; 121 #define O(n, t, nt) \ 122 t use_a_getter_or_setter_instead_##n; 123 MALLOC_TSD 124 #undef O 125 }; 126 127 /* 128 * Wrapper around tsd_t that makes it possible to avoid implicit conversion 129 * between tsd_t and tsdn_t, where tsdn_t is "nullable" and has to be 130 * explicitly converted to tsd_t, which is non-nullable. 131 */ 132 struct tsdn_s { 133 tsd_t tsd; 134 }; 135 #define TSDN_NULL ((tsdn_t *)0) 136 JEMALLOC_ALWAYS_INLINE tsdn_t * 137 tsd_tsdn(tsd_t *tsd) { 138 return (tsdn_t *)tsd; 139 } 140 141 JEMALLOC_ALWAYS_INLINE bool 142 tsdn_null(const tsdn_t *tsdn) { 143 return tsdn == NULL; 144 } 145 146 JEMALLOC_ALWAYS_INLINE tsd_t * 147 tsdn_tsd(tsdn_t *tsdn) { 148 assert(!tsdn_null(tsdn)); 149 150 return &tsdn->tsd; 151 } 152 153 void *malloc_tsd_malloc(size_t size); 154 void malloc_tsd_dalloc(void *wrapper); 155 void malloc_tsd_cleanup_register(bool (*f)(void)); 156 tsd_t *malloc_tsd_boot0(void); 157 void malloc_tsd_boot1(void); 158 void tsd_cleanup(void *arg); 159 tsd_t *tsd_fetch_slow(tsd_t *tsd, bool internal); 160 void tsd_slow_update(tsd_t *tsd); 161 162 /* 163 * We put the platform-specific data declarations and inlines into their own 164 * header files to avoid cluttering this file. They define tsd_boot0, 165 * tsd_boot1, tsd_boot, tsd_booted_get, tsd_get_allocates, tsd_get, and tsd_set. 166 */ 167 #ifdef JEMALLOC_MALLOC_THREAD_CLEANUP 168 #include "jemalloc/internal/tsd_malloc_thread_cleanup.h" 169 #elif (defined(JEMALLOC_TLS)) 170 #include "jemalloc/internal/tsd_tls.h" 171 #elif (defined(_WIN32)) 172 #include "jemalloc/internal/tsd_win.h" 173 #else 174 #include "jemalloc/internal/tsd_generic.h" 175 #endif 176 177 /* 178 * tsd_foop_get_unsafe(tsd) returns a pointer to the thread-local instance of 179 * foo. This omits some safety checks, and so can be used during tsd 180 * initialization and cleanup. 181 */ 182 #define O(n, t, nt) \ 183 JEMALLOC_ALWAYS_INLINE t * \ 184 tsd_##n##p_get_unsafe(tsd_t *tsd) { \ 185 return &tsd->use_a_getter_or_setter_instead_##n; \ 186 } 187 MALLOC_TSD 188 #undef O 189 190 /* tsd_foop_get(tsd) returns a pointer to the thread-local instance of foo. */ 191 #define O(n, t, nt) \ 192 JEMALLOC_ALWAYS_INLINE t * \ 193 tsd_##n##p_get(tsd_t *tsd) { \ 194 assert(tsd->state == tsd_state_nominal || \ 195 tsd->state == tsd_state_nominal_slow || \ 196 tsd->state == tsd_state_reincarnated || \ 197 tsd->state == tsd_state_minimal_initialized); \ 198 return tsd_##n##p_get_unsafe(tsd); \ 199 } 200 MALLOC_TSD 201 #undef O 202 203 /* 204 * tsdn_foop_get(tsdn) returns either the thread-local instance of foo (if tsdn 205 * isn't NULL), or NULL (if tsdn is NULL), cast to the nullable pointer type. 206 */ 207 #define O(n, t, nt) \ 208 JEMALLOC_ALWAYS_INLINE nt * \ 209 tsdn_##n##p_get(tsdn_t *tsdn) { \ 210 if (tsdn_null(tsdn)) { \ 211 return NULL; \ 212 } \ 213 tsd_t *tsd = tsdn_tsd(tsdn); \ 214 return (nt *)tsd_##n##p_get(tsd); \ 215 } 216 MALLOC_TSD 217 #undef O 218 219 /* tsd_foo_get(tsd) returns the value of the thread-local instance of foo. */ 220 #define O(n, t, nt) \ 221 JEMALLOC_ALWAYS_INLINE t \ 222 tsd_##n##_get(tsd_t *tsd) { \ 223 return *tsd_##n##p_get(tsd); \ 224 } 225 MALLOC_TSD 226 #undef O 227 228 /* tsd_foo_set(tsd, val) updates the thread-local instance of foo to be val. */ 229 #define O(n, t, nt) \ 230 JEMALLOC_ALWAYS_INLINE void \ 231 tsd_##n##_set(tsd_t *tsd, t val) { \ 232 assert(tsd->state != tsd_state_reincarnated && \ 233 tsd->state != tsd_state_minimal_initialized); \ 234 *tsd_##n##p_get(tsd) = val; \ 235 } 236 MALLOC_TSD 237 #undef O 238 239 JEMALLOC_ALWAYS_INLINE void 240 tsd_assert_fast(tsd_t *tsd) { 241 assert(!malloc_slow && tsd_tcache_enabled_get(tsd) && 242 tsd_reentrancy_level_get(tsd) == 0); 243 } 244 245 JEMALLOC_ALWAYS_INLINE bool 246 tsd_fast(tsd_t *tsd) { 247 bool fast = (tsd->state == tsd_state_nominal); 248 if (fast) { 249 tsd_assert_fast(tsd); 250 } 251 252 return fast; 253 } 254 255 JEMALLOC_ALWAYS_INLINE tsd_t * 256 tsd_fetch_impl(bool init, bool minimal) { 257 tsd_t *tsd = tsd_get(init); 258 259 if (!init && tsd_get_allocates() && tsd == NULL) { 260 return NULL; 261 } 262 assert(tsd != NULL); 263 264 if (unlikely(tsd->state != tsd_state_nominal)) { 265 return tsd_fetch_slow(tsd, minimal); 266 } 267 assert(tsd_fast(tsd)); 268 tsd_assert_fast(tsd); 269 270 return tsd; 271 } 272 273 /* Get a minimal TSD that requires no cleanup. See comments in free(). */ 274 JEMALLOC_ALWAYS_INLINE tsd_t * 275 tsd_fetch_min(void) { 276 return tsd_fetch_impl(true, true); 277 } 278 279 /* For internal background threads use only. */ 280 JEMALLOC_ALWAYS_INLINE tsd_t * 281 tsd_internal_fetch(void) { 282 tsd_t *tsd = tsd_fetch_min(); 283 /* Use reincarnated state to prevent full initialization. */ 284 tsd->state = tsd_state_reincarnated; 285 286 return tsd; 287 } 288 289 JEMALLOC_ALWAYS_INLINE tsd_t * 290 tsd_fetch(void) { 291 return tsd_fetch_impl(true, false); 292 } 293 294 static inline bool 295 tsd_nominal(tsd_t *tsd) { 296 return (tsd->state <= tsd_state_nominal_max); 297 } 298 299 JEMALLOC_ALWAYS_INLINE tsdn_t * 300 tsdn_fetch(void) { 301 if (!tsd_booted_get()) { 302 return NULL; 303 } 304 305 return tsd_tsdn(tsd_fetch_impl(false, false)); 306 } 307 308 JEMALLOC_ALWAYS_INLINE rtree_ctx_t * 309 tsd_rtree_ctx(tsd_t *tsd) { 310 return tsd_rtree_ctxp_get(tsd); 311 } 312 313 JEMALLOC_ALWAYS_INLINE rtree_ctx_t * 314 tsdn_rtree_ctx(tsdn_t *tsdn, rtree_ctx_t *fallback) { 315 /* 316 * If tsd cannot be accessed, initialize the fallback rtree_ctx and 317 * return a pointer to it. 318 */ 319 if (unlikely(tsdn_null(tsdn))) { 320 rtree_ctx_data_init(fallback); 321 return fallback; 322 } 323 return tsd_rtree_ctx(tsdn_tsd(tsdn)); 324 } 325 326 #endif /* JEMALLOC_INTERNAL_TSD_H */ 327