1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License, Version 1.0 only 6 * (the "License"). You may not use this file except in compliance 7 * with the License. 8 * 9 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 10 * or http://www.opensolaris.org/os/licensing. 11 * See the License for the specific language governing permissions 12 * and limitations under the License. 13 * 14 * When distributing Covered Code, include this CDDL HEADER in each 15 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 16 * If applicable, add the following below this CDDL HEADER, with the 17 * fields enclosed by brackets "[]" replaced with your own identifying 18 * information: Portions Copyright [yyyy] [name of copyright owner] 19 * 20 * CDDL HEADER END 21 */ 22 /* 23 * Copyright 2004 Sun Microsystems, Inc. All rights reserved. 24 * Copyright 2012 Joyent, Inc. All rights reserved. 25 * Use is subject to license terms. 26 */ 27 28 #ifndef _UMEM_IMPL_H 29 #define _UMEM_IMPL_H 30 31 #include <umem.h> 32 33 #include <sys/sysmacros.h> 34 #include <sys/time.h> 35 #include <sys/vmem.h> 36 #include <thread.h> 37 38 #ifdef __cplusplus 39 extern "C" { 40 #endif 41 42 /* 43 * umem memory allocator: implementation-private data structures 44 */ 45 46 /* 47 * Internal flags for umem_cache_create 48 */ 49 #define UMC_QCACHE 0x00100000 50 #define UMC_INTERNAL 0x80000000 51 52 /* 53 * Cache flags 54 */ 55 #define UMF_AUDIT 0x00000001 /* transaction auditing */ 56 #define UMF_DEADBEEF 0x00000002 /* deadbeef checking */ 57 #define UMF_REDZONE 0x00000004 /* redzone checking */ 58 #define UMF_CONTENTS 0x00000008 /* freed-buffer content logging */ 59 #define UMF_CHECKSIGNAL 0x00000010 /* abort when in signal context */ 60 #define UMF_NOMAGAZINE 0x00000020 /* disable per-cpu magazines */ 61 #define UMF_FIREWALL 0x00000040 /* put all bufs before unmapped pages */ 62 #define UMF_LITE 0x00000100 /* lightweight debugging */ 63 64 #define UMF_HASH 0x00000200 /* cache has hash table */ 65 #define UMF_RANDOMIZE 0x00000400 /* randomize other umem_flags */ 66 67 #define UMF_BUFTAG (UMF_DEADBEEF | UMF_REDZONE) 68 #define UMF_TOUCH (UMF_BUFTAG | UMF_LITE | UMF_CONTENTS) 69 #define UMF_RANDOM (UMF_TOUCH | UMF_AUDIT | UMF_NOMAGAZINE) 70 #define UMF_DEBUG (UMF_RANDOM | UMF_FIREWALL) 71 72 #define UMEM_STACK_DEPTH umem_stack_depth 73 74 #define UMEM_FREE_PATTERN 0xdeadbeefdeadbeefULL 75 #define UMEM_UNINITIALIZED_PATTERN 0xbaddcafebaddcafeULL 76 #define UMEM_REDZONE_PATTERN 0xfeedfacefeedfaceULL 77 #define UMEM_REDZONE_BYTE 0xbb 78 79 #define UMEM_FATAL_FLAGS (UMEM_NOFAIL) 80 #define UMEM_SLEEP_FLAGS (0) 81 82 /* 83 * Redzone size encodings for umem_alloc() / umem_free(). We encode the 84 * allocation size, rather than storing it directly, so that umem_free() 85 * can distinguish frees of the wrong size from redzone violations. 86 */ 87 #define UMEM_SIZE_ENCODE(x) (251 * (x) + 1) 88 #define UMEM_SIZE_DECODE(x) ((x) / 251) 89 #define UMEM_SIZE_VALID(x) ((x) % 251 == 1) 90 91 /* 92 * The bufctl (buffer control) structure keeps some minimal information 93 * about each buffer: its address, its slab, and its current linkage, 94 * which is either on the slab's freelist (if the buffer is free), or 95 * on the cache's buf-to-bufctl hash table (if the buffer is allocated). 96 * In the case of non-hashed, or "raw", caches (the common case), only 97 * the freelist linkage is necessary: the buffer address is at a fixed 98 * offset from the bufctl address, and the slab is at the end of the page. 99 * 100 * NOTE: bc_next must be the first field; raw buffers have linkage only. 101 */ 102 typedef struct umem_bufctl { 103 struct umem_bufctl *bc_next; /* next bufctl struct */ 104 void *bc_addr; /* address of buffer */ 105 struct umem_slab *bc_slab; /* controlling slab */ 106 } umem_bufctl_t; 107 108 /* 109 * The UMF_AUDIT version of the bufctl structure. The beginning of this 110 * structure must be identical to the normal bufctl structure so that 111 * pointers are interchangeable. 112 */ 113 114 #define UMEM_BUFCTL_AUDIT_SIZE_DEPTH(frames) \ 115 ((size_t)(&((umem_bufctl_audit_t *)0)->bc_stack[frames])) 116 117 /* 118 * umem_bufctl_audits must be allocated from a UMC_NOHASH cache, so we 119 * require that 2 of them, plus 2 buftags, plus a umem_slab_t, all fit on 120 * a single page. 121 * 122 * For ILP32, this is about 1000 frames. 123 * For LP64, this is about 490 frames. 124 */ 125 126 #define UMEM_BUFCTL_AUDIT_ALIGN 32 127 128 #define UMEM_BUFCTL_AUDIT_MAX_SIZE \ 129 (P2ALIGN((PAGESIZE - sizeof (umem_slab_t))/2 - \ 130 sizeof (umem_buftag_t), UMEM_BUFCTL_AUDIT_ALIGN)) 131 132 #define UMEM_MAX_STACK_DEPTH \ 133 ((UMEM_BUFCTL_AUDIT_MAX_SIZE - \ 134 UMEM_BUFCTL_AUDIT_SIZE_DEPTH(0)) / sizeof (uintptr_t)) 135 136 typedef struct umem_bufctl_audit { 137 struct umem_bufctl *bc_next; /* next bufctl struct */ 138 void *bc_addr; /* address of buffer */ 139 struct umem_slab *bc_slab; /* controlling slab */ 140 umem_cache_t *bc_cache; /* controlling cache */ 141 hrtime_t bc_timestamp; /* transaction time */ 142 thread_t bc_thread; /* thread doing transaction */ 143 struct umem_bufctl *bc_lastlog; /* last log entry */ 144 void *bc_contents; /* contents at last free */ 145 int bc_depth; /* stack depth */ 146 uintptr_t bc_stack[1]; /* pc stack */ 147 } umem_bufctl_audit_t; 148 149 #define UMEM_LOCAL_BUFCTL_AUDIT(bcpp) \ 150 *(bcpp) = (umem_bufctl_audit_t *) \ 151 alloca(UMEM_BUFCTL_AUDIT_SIZE) 152 153 #define UMEM_BUFCTL_AUDIT_SIZE \ 154 UMEM_BUFCTL_AUDIT_SIZE_DEPTH(UMEM_STACK_DEPTH) 155 156 /* 157 * A umem_buftag structure is appended to each buffer whenever any of the 158 * UMF_BUFTAG flags (UMF_DEADBEEF, UMF_REDZONE, UMF_VERIFY) are set. 159 */ 160 typedef struct umem_buftag { 161 uint64_t bt_redzone; /* 64-bit redzone pattern */ 162 umem_bufctl_t *bt_bufctl; /* bufctl */ 163 intptr_t bt_bxstat; /* bufctl ^ (alloc/free) */ 164 } umem_buftag_t; 165 166 #define UMEM_BUFTAG(cp, buf) \ 167 ((umem_buftag_t *)((char *)(buf) + (cp)->cache_buftag)) 168 169 #define UMEM_BUFCTL(cp, buf) \ 170 ((umem_bufctl_t *)((char *)(buf) + (cp)->cache_bufctl)) 171 172 #define UMEM_BUF(cp, bcp) \ 173 ((void *)((char *)(bcp) - (cp)->cache_bufctl)) 174 175 #define UMEM_SLAB(cp, buf) \ 176 ((umem_slab_t *)P2END((uintptr_t)(buf), (cp)->cache_slabsize) - 1) 177 178 #define UMEM_CPU_CACHE(cp, cpu) \ 179 (umem_cpu_cache_t *)((char *)cp + cpu->cpu_cache_offset) 180 181 #define UMEM_MAGAZINE_VALID(cp, mp) \ 182 (((umem_slab_t *)P2END((uintptr_t)(mp), PAGESIZE) - 1)->slab_cache == \ 183 (cp)->cache_magtype->mt_cache) 184 185 #define UMEM_SLAB_MEMBER(sp, buf) \ 186 ((size_t)(buf) - (size_t)(sp)->slab_base < \ 187 (sp)->slab_cache->cache_slabsize) 188 189 #define UMEM_BUFTAG_ALLOC 0xa110c8edUL 190 #define UMEM_BUFTAG_FREE 0xf4eef4eeUL 191 192 typedef struct umem_slab { 193 struct umem_cache *slab_cache; /* controlling cache */ 194 void *slab_base; /* base of allocated memory */ 195 struct umem_slab *slab_next; /* next slab on freelist */ 196 struct umem_slab *slab_prev; /* prev slab on freelist */ 197 struct umem_bufctl *slab_head; /* first free buffer */ 198 long slab_refcnt; /* outstanding allocations */ 199 long slab_chunks; /* chunks (bufs) in this slab */ 200 } umem_slab_t; 201 202 #define UMEM_HASH_INITIAL 64 203 204 #define UMEM_HASH(cp, buf) \ 205 ((cp)->cache_hash_table + \ 206 (((uintptr_t)(buf) >> (cp)->cache_hash_shift) & (cp)->cache_hash_mask)) 207 208 typedef struct umem_magazine { 209 void *mag_next; 210 void *mag_round[1]; /* one or more rounds */ 211 } umem_magazine_t; 212 213 /* 214 * The magazine types for fast per-cpu allocation 215 */ 216 typedef struct umem_magtype { 217 int mt_magsize; /* magazine size (number of rounds) */ 218 int mt_align; /* magazine alignment */ 219 size_t mt_minbuf; /* all smaller buffers qualify */ 220 size_t mt_maxbuf; /* no larger buffers qualify */ 221 umem_cache_t *mt_cache; /* magazine cache */ 222 } umem_magtype_t; 223 224 #define UMEM_CPU_CACHE_SIZE 64 /* must be power of 2 */ 225 #define UMEM_CPU_PAD (UMEM_CPU_CACHE_SIZE - sizeof (mutex_t) - \ 226 2 * sizeof (uint_t) - 2 * sizeof (void *) - 4 * sizeof (int)) 227 #define UMEM_CACHE_SIZE(ncpus) \ 228 ((size_t)(&((umem_cache_t *)0)->cache_cpu[ncpus])) 229 230 typedef struct umem_cpu_cache { 231 mutex_t cc_lock; /* protects this cpu's local cache */ 232 uint_t cc_alloc; /* allocations from this cpu */ 233 uint_t cc_free; /* frees to this cpu */ 234 umem_magazine_t *cc_loaded; /* the currently loaded magazine */ 235 umem_magazine_t *cc_ploaded; /* the previously loaded magazine */ 236 int cc_rounds; /* number of objects in loaded mag */ 237 int cc_prounds; /* number of objects in previous mag */ 238 int cc_magsize; /* number of rounds in a full mag */ 239 int cc_flags; /* CPU-local copy of cache_flags */ 240 #ifndef _LP64 241 char cc_pad[UMEM_CPU_PAD]; /* for nice alignment (32-bit) */ 242 #endif 243 } umem_cpu_cache_t; 244 245 /* 246 * The magazine lists used in the depot. 247 */ 248 typedef struct umem_maglist { 249 umem_magazine_t *ml_list; /* magazine list */ 250 long ml_total; /* number of magazines */ 251 long ml_min; /* min since last update */ 252 long ml_reaplimit; /* max reapable magazines */ 253 uint64_t ml_alloc; /* allocations from this list */ 254 } umem_maglist_t; 255 256 #define UMEM_CACHE_NAMELEN 31 257 258 struct umem_cache { 259 /* 260 * Statistics 261 */ 262 uint64_t cache_slab_create; /* slab creates */ 263 uint64_t cache_slab_destroy; /* slab destroys */ 264 uint64_t cache_slab_alloc; /* slab layer allocations */ 265 uint64_t cache_slab_free; /* slab layer frees */ 266 uint64_t cache_alloc_fail; /* total failed allocations */ 267 uint64_t cache_buftotal; /* total buffers */ 268 uint64_t cache_bufmax; /* max buffers ever */ 269 uint64_t cache_rescale; /* # of hash table rescales */ 270 uint64_t cache_lookup_depth; /* hash lookup depth */ 271 uint64_t cache_depot_contention; /* mutex contention count */ 272 uint64_t cache_depot_contention_prev; /* previous snapshot */ 273 274 /* 275 * Cache properties 276 */ 277 char cache_name[UMEM_CACHE_NAMELEN + 1]; 278 size_t cache_bufsize; /* object size */ 279 size_t cache_align; /* object alignment */ 280 umem_constructor_t *cache_constructor; 281 umem_destructor_t *cache_destructor; 282 umem_reclaim_t *cache_reclaim; 283 void *cache_private; /* opaque arg to callbacks */ 284 vmem_t *cache_arena; /* vmem source for slabs */ 285 int cache_cflags; /* cache creation flags */ 286 int cache_flags; /* various cache state info */ 287 int cache_uflags; /* UMU_* flags */ 288 uint32_t cache_mtbf; /* induced alloc failure rate */ 289 umem_cache_t *cache_next; /* forward cache linkage */ 290 umem_cache_t *cache_prev; /* backward cache linkage */ 291 umem_cache_t *cache_unext; /* next in update list */ 292 umem_cache_t *cache_uprev; /* prev in update list */ 293 uint32_t cache_cpu_mask; /* mask for cpu offset */ 294 295 /* 296 * Slab layer 297 */ 298 mutex_t cache_lock; /* protects slab layer */ 299 size_t cache_chunksize; /* buf + alignment [+ debug] */ 300 size_t cache_slabsize; /* size of a slab */ 301 size_t cache_bufctl; /* buf-to-bufctl distance */ 302 size_t cache_buftag; /* buf-to-buftag distance */ 303 size_t cache_verify; /* bytes to verify */ 304 size_t cache_contents; /* bytes of saved content */ 305 size_t cache_color; /* next slab color */ 306 size_t cache_mincolor; /* maximum slab color */ 307 size_t cache_maxcolor; /* maximum slab color */ 308 size_t cache_hash_shift; /* get to interesting bits */ 309 size_t cache_hash_mask; /* hash table mask */ 310 umem_slab_t *cache_freelist; /* slab free list */ 311 umem_slab_t cache_nullslab; /* end of freelist marker */ 312 umem_cache_t *cache_bufctl_cache; /* source of bufctls */ 313 umem_bufctl_t **cache_hash_table; /* hash table base */ 314 /* 315 * Depot layer 316 */ 317 mutex_t cache_depot_lock; /* protects depot */ 318 umem_magtype_t *cache_magtype; /* magazine type */ 319 umem_maglist_t cache_full; /* full magazines */ 320 umem_maglist_t cache_empty; /* empty magazines */ 321 322 /* 323 * Per-CPU layer 324 */ 325 umem_cpu_cache_t cache_cpu[1]; /* cache_cpu_mask + 1 entries */ 326 }; 327 328 typedef struct umem_cpu_log_header { 329 mutex_t clh_lock; 330 char *clh_current; 331 size_t clh_avail; 332 int clh_chunk; 333 int clh_hits; 334 char clh_pad[64 - sizeof (mutex_t) - sizeof (char *) - 335 sizeof (size_t) - 2 * sizeof (int)]; 336 } umem_cpu_log_header_t; 337 338 typedef struct umem_log_header { 339 mutex_t lh_lock; 340 char *lh_base; 341 int *lh_free; 342 size_t lh_chunksize; 343 int lh_nchunks; 344 int lh_head; 345 int lh_tail; 346 int lh_hits; 347 umem_cpu_log_header_t lh_cpu[1]; /* actually umem_max_ncpus */ 348 } umem_log_header_t; 349 350 typedef struct umem_cpu { 351 uint32_t cpu_cache_offset; 352 uint32_t cpu_number; 353 } umem_cpu_t; 354 355 #define UMEM_MAXBUF 131072 356 357 #define UMEM_ALIGN 8 /* min guaranteed alignment */ 358 #define UMEM_ALIGN_SHIFT 3 /* log2(UMEM_ALIGN) */ 359 #define UMEM_VOID_FRACTION 8 /* never waste more than 1/8 of slab */ 360 361 /* 362 * For 64 bits, buffers >= 16 bytes must be 16-byte aligned 363 */ 364 #ifdef _LP64 365 #define UMEM_SECOND_ALIGN 16 366 #else 367 #define UMEM_SECOND_ALIGN UMEM_ALIGN 368 #endif 369 370 #define MALLOC_MAGIC 0x3a10c000 /* 8-byte tag */ 371 #define MEMALIGN_MAGIC 0x3e3a1000 372 373 #ifdef _LP64 374 #define MALLOC_SECOND_MAGIC 0x16ba7000 /* 8-byte tag, 16-aligned */ 375 #define MALLOC_OVERSIZE_MAGIC 0x06e47000 /* 16-byte tag, _LP64 */ 376 #endif 377 378 #define UMEM_MALLOC_ENCODE(type, sz) (uint32_t)((type) - (sz)) 379 #define UMEM_MALLOC_DECODE(stat, sz) (uint32_t)((stat) + (sz)) 380 #define UMEM_FREE_PATTERN_32 (uint32_t)(UMEM_FREE_PATTERN) 381 382 #define UMU_MAGAZINE_RESIZE 0x00000001 383 #define UMU_HASH_RESCALE 0x00000002 384 #define UMU_REAP 0x00000004 385 #define UMU_NOTIFY 0x08000000 386 #define UMU_ACTIVE 0x80000000 387 388 #define UMEM_READY_INIT_FAILED -1 389 #define UMEM_READY_STARTUP 1 390 #define UMEM_READY_INITING 2 391 #define UMEM_READY 3 392 393 #ifdef UMEM_STANDALONE 394 extern void umem_startup(caddr_t, size_t, size_t, caddr_t, caddr_t); 395 extern int umem_add(caddr_t, size_t); 396 #endif 397 398 #ifdef __cplusplus 399 } 400 #endif 401 402 #endif /* _UMEM_IMPL_H */ 403