X-Git-Url: https://git.xonotic.org/?p=xonotic%2Fgmqcc.git;a=blobdiff_plain;f=stat.c;h=650bccdf040a0f83a324c70b94a545086d70aef4;hp=3b54d6e494f63f4cc23db2ca79afc75bdcd717fc;hb=edb38ce70e86acde9530af5861cfcf8f396b5e75;hpb=6db2e69f9a9411f4d5cb6923b1ead3706d7fe195 diff --git a/stat.c b/stat.c index 3b54d6e..650bccd 100644 --- a/stat.c +++ b/stat.c @@ -1,5 +1,5 @@ /* - * Copyright (C) 2012, 2013 + * Copyright (C) 2012, 2013, 2014 * Dale Weiler * Wolfgang Bumiller * @@ -27,6 +27,30 @@ #include "gmqcc.h" +typedef struct stat_mem_block_s stat_mem_block_t; + +#define IDENT_SIZE 4 +#define IDENT_VEC "vec" +#define IDENT_MEM "mem" +#define IDENT_VEC_TOP (sizeof(vector_t) + IDENT_SIZE) +#define IDENT_MEM_TOP (sizeof(stat_mem_block_t) + IDENT_SIZE) + +/* + * For the valgrind integration of our allocator. This allows us to have + * more `accurate` valgrind output for our allocator, and also secures the + * possible underflows (where one could obtain access to the redzone that + * represents info about that allocation). + */ +#ifndef NVALGRIND +# include +# include +#else +# define VALGRIND_MALLOCLIKE_BLOCK(PTR, ALLOC_SIZE, REDZONE_SIZE, ZEROED) +# define VALGRIND_FREELIKE_BLOCK(PTR, REDZONE_SIZE) +# define VALGRIND_MAKE_MEM_DEFINED(PTR, REDZONE_SIZE) +# define VALGRIND_MAKE_MEM_NOACCESS(PTR, REDZONE_SIZE) +#endif + /* * GMQCC performs tons of allocations, constructions, and crazyness * all around. When trying to optimizes systems, or just get fancy @@ -36,13 +60,14 @@ */ #define ST_SIZE 1024 -typedef struct stat_mem_block_s { +struct stat_mem_block_s { const char *file; size_t line; size_t size; + const char *expr; struct stat_mem_block_s *next; struct stat_mem_block_s *prev; -} stat_mem_block_t; +}; typedef struct { size_t key; @@ -55,6 +80,7 @@ static uint64_t stat_mem_allocated_total = 0; static uint64_t stat_mem_deallocated_total = 0; static uint64_t stat_mem_high = 0; static uint64_t stat_mem_peak = 0; +static uint64_t stat_mem_strdups = 0; static uint64_t stat_used_strdups = 0; static uint64_t stat_used_vectors = 0; static uint64_t stat_used_hashtables = 0; @@ -71,8 +97,8 @@ static stat_mem_block_t *stat_mem_block_root = NULL; */ static stat_size_table_t stat_size_new(void) { return (stat_size_table_t)memset( - mem_a(sizeof(stat_size_entry_t) * ST_SIZE), - 0, ST_SIZE * sizeof(stat_size_entry_t) + mem_a(sizeof(stat_size_entry_t*) * ST_SIZE), + 0, ST_SIZE * sizeof(stat_size_entry_t*) ); } @@ -102,21 +128,29 @@ static void stat_size_put(stat_size_table_t table, size_t key, size_t value) { * information as a header, returns the memory + 1 past it, can be * retrieved again with - 1. Where type is stat_mem_block_t*. */ -void *stat_mem_allocate(size_t size, size_t line, const char *file) { - stat_mem_block_t *info = (stat_mem_block_t*)malloc(sizeof(stat_mem_block_t) + size); - void *data = (void*)(info + 1); +void *stat_mem_allocate(size_t size, size_t line, const char *file, const char *expr) { + stat_mem_block_t *info = (stat_mem_block_t*)malloc(size + IDENT_MEM_TOP); + void *data = (void *)((char*)info + IDENT_MEM_TOP); - if(!info) + if(GMQCC_UNLIKELY(!info)) return NULL; info->line = line; info->size = size; info->file = file; + info->expr = expr; info->prev = NULL; info->next = stat_mem_block_root; - if (stat_mem_block_root) + /* Write identifier */ + memcpy(info + 1, IDENT_MEM, IDENT_SIZE); + + /* likely since it only happens once */ + if (GMQCC_LIKELY(stat_mem_block_root != NULL)) { + VALGRIND_MAKE_MEM_DEFINED(stat_mem_block_root, IDENT_MEM_TOP); stat_mem_block_root->prev = info; + VALGRIND_MAKE_MEM_NOACCESS(stat_mem_block_root, IDENT_MEM_TOP); + } stat_mem_block_root = info; stat_mem_allocated += size; @@ -126,69 +160,151 @@ void *stat_mem_allocate(size_t size, size_t line, const char *file) { if (stat_mem_high > stat_mem_peak) stat_mem_peak = stat_mem_high; + VALGRIND_MALLOCLIKE_BLOCK(data, size, IDENT_MEM_TOP, 0); return data; } -void stat_mem_deallocate(void *ptr) { - stat_mem_block_t *info = NULL; +void stat_mem_deallocate(void *ptr, size_t line, const char *file) { + stat_mem_block_t *info = NULL; + char *ident = (char *)ptr - IDENT_SIZE; + + if (GMQCC_UNLIKELY(!ptr)) + return; - if (!ptr) + /* Validate usage */ + VALGRIND_MAKE_MEM_DEFINED(ident, IDENT_SIZE); + if (!strcmp(ident, IDENT_VEC)) { + vector_t *vec = (vector_t*)((char *)ptr - IDENT_VEC_TOP); + stat_mem_block_t *block = (stat_mem_block_t*)((char *)vec - IDENT_MEM_TOP); + + VALGRIND_MAKE_MEM_DEFINED(block, sizeof(stat_mem_block_t)); + con_err("internal warning: invalid use of mem_d:\n"); + con_err("internal warning: vector (used elements: %u, allocated elements: %u)\n", + (unsigned)vec->used, + (unsigned)vec->allocated + ); + con_err("internal warning: vector was last (re)allocated with (size: %u (bytes), at location: %s:%u)\n", + (unsigned)block->size, + block->file, + (unsigned)block->line + ); + con_err("internal warning: released with wrong routine at %s:%u\n", file, (unsigned)line); + con_err("internal warning: forwarding to vec_free, please fix it\n"); + VALGRIND_MAKE_MEM_NOACCESS(block, sizeof(stat_mem_block_t)); + VALGRIND_MAKE_MEM_NOACCESS(ident, IDENT_SIZE); + vec_free(ptr); return; + } + VALGRIND_MAKE_MEM_NOACCESS(ident, IDENT_SIZE); + info = (stat_mem_block_t*)((char *)ptr - IDENT_MEM_TOP); - info = ((stat_mem_block_t*)ptr - 1); + /* + * we need access to the redzone that represents the info block + * so lets do that. + */ + VALGRIND_MAKE_MEM_DEFINED(info, IDENT_MEM_TOP); stat_mem_deallocated += info->size; stat_mem_high -= info->size; stat_mem_deallocated_total ++; - if (info->prev) info->prev->next = info->next; - if (info->next) info->next->prev = info->prev; + if (info->prev) { + /* just need access for a short period */ + VALGRIND_MAKE_MEM_DEFINED(info->prev, IDENT_MEM_TOP); + info->prev->next = info->next; + /* don't need access anymore */ + VALGRIND_MAKE_MEM_NOACCESS(info->prev, IDENT_MEM_TOP); + } + if (info->next) { + /* just need access for a short period */ + VALGRIND_MAKE_MEM_DEFINED(info->next, IDENT_MEM_TOP); + info->next->prev = info->prev; + /* don't need access anymore */ + VALGRIND_MAKE_MEM_NOACCESS(info->next, IDENT_MEM_TOP); + } /* move ahead */ if (info == stat_mem_block_root) stat_mem_block_root = info->next; free(info); + VALGRIND_MAKE_MEM_NOACCESS(info, IDENT_MEM_TOP); + VALGRIND_FREELIKE_BLOCK(ptr, IDENT_MEM_TOP); } -void *stat_mem_reallocate(void *ptr, size_t size, size_t line, const char *file) { +void *stat_mem_reallocate(void *ptr, size_t size, size_t line, const char *file, const char *expr) { stat_mem_block_t *oldinfo = NULL; stat_mem_block_t *newinfo; - if (!ptr) - return stat_mem_allocate(size, line, file); + if (GMQCC_UNLIKELY(!ptr)) + return stat_mem_allocate(size, line, file, expr); - /* stay consistent with glic */ - if (!size) { - stat_mem_deallocate(ptr); + /* stay consistent with glibc */ + if (GMQCC_UNLIKELY(!size)) { + stat_mem_deallocate(ptr, line, file); return NULL; } - oldinfo = ((stat_mem_block_t*)ptr - 1); - newinfo = ((stat_mem_block_t*)malloc(sizeof(stat_mem_block_t) + size)); + oldinfo = (stat_mem_block_t*)((char *)ptr - IDENT_MEM_TOP); + newinfo = (stat_mem_block_t*)malloc(size + IDENT_MEM_TOP); - if (!newinfo) { - stat_mem_deallocate(ptr); + if (GMQCC_UNLIKELY(!newinfo)) { + stat_mem_deallocate(ptr, line, file); return NULL; } - memcpy(newinfo+1, oldinfo+1, oldinfo->size); + VALGRIND_MALLOCLIKE_BLOCK((char *)newinfo + IDENT_MEM_TOP, size, IDENT_MEM_TOP, 0); + + /* we need access to the old info redzone */ + VALGRIND_MAKE_MEM_DEFINED(oldinfo, IDENT_MEM_TOP); - if (oldinfo->prev) oldinfo->prev->next = oldinfo->next; - if (oldinfo->next) oldinfo->next->prev = oldinfo->prev; + /* We need access to the new info redzone */ + VALGRIND_MAKE_MEM_DEFINED(newinfo, IDENT_MEM_TOP); + memcpy((char *)(newinfo + 1), IDENT_MEM, IDENT_SIZE); + memcpy((char *)newinfo + IDENT_MEM_TOP, (char *)oldinfo + IDENT_MEM_TOP, oldinfo->size); + VALGRIND_MAKE_MEM_NOACCESS(newinfo, IDENT_MEM_TOP); + + if (oldinfo->prev) { + /* just need access for a short period */ + VALGRIND_MAKE_MEM_DEFINED(oldinfo->prev, IDENT_MEM_TOP); + oldinfo->prev->next = oldinfo->next; + /* don't need access anymore */ + VALGRIND_MAKE_MEM_NOACCESS(oldinfo->prev, IDENT_MEM_TOP); + } + + if (oldinfo->next) { + /* just need access for a short period */ + VALGRIND_MAKE_MEM_DEFINED(oldinfo->next, IDENT_MEM_TOP); + oldinfo->next->prev = oldinfo->prev; + /* don't need access anymore */ + VALGRIND_MAKE_MEM_NOACCESS(oldinfo->next, IDENT_MEM_TOP); + } /* move ahead */ if (oldinfo == stat_mem_block_root) stat_mem_block_root = oldinfo->next; + /* we need access to the redzone for the newinfo block */ + VALGRIND_MAKE_MEM_DEFINED(newinfo, IDENT_MEM_TOP); + newinfo->line = line; newinfo->size = size; newinfo->file = file; + newinfo->expr = expr; newinfo->prev = NULL; newinfo->next = stat_mem_block_root; - if (stat_mem_block_root) + /* + * likely since the only time there is no root is when it's + * being initialized first. + */ + if (GMQCC_LIKELY(stat_mem_block_root != NULL)) { + /* we need access to the root */ + VALGRIND_MAKE_MEM_DEFINED(stat_mem_block_root, IDENT_MEM_TOP); stat_mem_block_root->prev = newinfo; + /* kill access */ + VALGRIND_MAKE_MEM_NOACCESS(stat_mem_block_root, IDENT_MEM_TOP); + } stat_mem_block_root = newinfo; stat_mem_allocated -= oldinfo->size; @@ -196,12 +312,19 @@ void *stat_mem_reallocate(void *ptr, size_t size, size_t line, const char *file) stat_mem_allocated += newinfo->size; stat_mem_high += newinfo->size; + /* + * we're finished with the redzones, lets kill the access + * to them. + */ + VALGRIND_MAKE_MEM_NOACCESS(newinfo, IDENT_MEM_TOP); + VALGRIND_MAKE_MEM_NOACCESS(oldinfo, IDENT_MEM_TOP); + if (stat_mem_high > stat_mem_peak) stat_mem_peak = stat_mem_high; free(oldinfo); - - return newinfo + 1; + VALGRIND_FREELIKE_BLOCK(ptr, IDENT_MEM_TOP); + return (char *)newinfo + IDENT_MEM_TOP; } /* @@ -217,12 +340,13 @@ char *stat_mem_strdup(const char *src, size_t line, const char *file, bool empty return NULL; len = strlen(src); - if (((!empty) ? len : true) && (ptr = (char*)stat_mem_allocate(len + 1, line, file))) { + if (((!empty) ? len : true) && (ptr = (char*)stat_mem_allocate(len + 1, line, file, "strdup"))) { memcpy(ptr, src, len); ptr[len] = '\0'; } stat_used_strdups ++; + stat_mem_strdups += len; return ptr; } @@ -230,17 +354,17 @@ char *stat_mem_strdup(const char *src, size_t line, const char *file, bool empty * The reallocate function for resizing vectors. */ void _util_vec_grow(void **a, size_t i, size_t s) { - vector_t *d = vec_meta(*a); + vector_t *d = (vector_t*)((char *)*a - IDENT_VEC_TOP); size_t m = 0; stat_size_entry_t *e = NULL; void *p = NULL; if (*a) { m = 2 * d->allocated + i; - p = mem_r(d, s * m + sizeof(vector_t)); + p = mem_r(d, s * m + IDENT_VEC_TOP); } else { m = i + 1; - p = mem_a(s * m + sizeof(vector_t)); + p = mem_a(s * m + IDENT_VEC_TOP); ((vector_t*)p)->used = 0; stat_used_vectors++; } @@ -255,8 +379,30 @@ void _util_vec_grow(void **a, size_t i, size_t s) { stat_type_vectors++; } - *a = (vector_t*)p + 1; - vec_meta(*a)->allocated = m; + d = (vector_t*)p; + d->allocated = m; + memcpy(d + 1, IDENT_VEC, IDENT_SIZE); + *a = (void *)((char *)d + IDENT_VEC_TOP); +} + +void _util_vec_delete(void *data, size_t line, const char *file) { + char *ident = (char *)data - IDENT_SIZE; + if (!strcmp(ident, IDENT_MEM)) { + stat_mem_block_t *block = (stat_mem_block_t*)((char *)data - IDENT_MEM_TOP); + VALGRIND_MAKE_MEM_DEFINED(block, sizeof(stat_mem_block_t)); + con_err("internal warning: invalid use of vec_free:\n"); + con_err("internal warning: memory block last allocated (size: %u (bytes), at %s:%u)\n", + (unsigned)block->size, + block->file, + (unsigned)block->line); + con_err("internal warning: released with with wrong routine at %s:%u\n", file, (unsigned)line); + con_err("internal warning: forwarding to mem_d, please fix it\n"); + VALGRIND_MAKE_MEM_NOACCESS(block, sizeof(stat_mem_block_t)); + mem_d(data); + return; + } + /* forward */ + stat_mem_deallocate((void*)(ident - sizeof(vector_t)), line, file); } /* @@ -270,139 +416,11 @@ typedef struct hash_node_t { struct hash_node_t *next; /* next node (linked list) */ } hash_node_t; -/* - * This is a patched version of the Murmur2 hashing function to use - * a proper pre-mix and post-mix setup. Infact this is Murmur3 for - * the most part just reinvented. - * - * Murmur 2 contains an inner loop such as: - * while (l >= 4) { - * u32 k = *(u32*)d; - * k *= m; - * k ^= k >> r; - * k *= m; - * - * h *= m; - * h ^= k; - * d += 4; - * l -= 4; - * } - * - * The two u32s that form the key are the same value x (pulled from data) - * this premix stage will perform the same results for both values. Unrolled - * this produces just: - * x *= m; - * x ^= x >> r; - * x *= m; - * - * h *= m; - * h ^= x; - * h *= m; - * h ^= x; - * - * This appears to be fine, except what happens when m == 1? well x - * cancels out entierly, leaving just: - * x ^= x >> r; - * h ^= x; - * h ^= x; - * - * So all keys hash to the same value, but how often does m == 1? - * well, it turns out testing x for all possible values yeilds only - * 172,013,942 unique results instead of 2^32. So nearly ~4.6 bits - * are cancelled out on average! - * - * This means we have a 14.5% (rounded) chance of colliding more, which - * results in another bucket/chain for the hashtable. - * - * We fix it buy upgrading the pre and post mix ssystems to align with murmur - * hash 3. - */ -#if 1 -#define GMQCC_ROTL32(X, R) (((X) << (R)) | ((X) >> (32 - (R)))) -GMQCC_INLINE size_t util_hthash(hash_table_t *ht, const char *key) { - const unsigned char *data = (const unsigned char *)key; - const size_t len = strlen(key); - const size_t block = len / 4; - const uint32_t mask1 = 0xCC9E2D51; - const uint32_t mask2 = 0x1B873593; - const uint32_t *blocks = (const uint32_t*)(data + block * 4); - const unsigned char *tail = (const unsigned char *)(data + block * 4); - - size_t i; - uint32_t k; - uint32_t h = 0x1EF0 ^ len; - - for (i = -block; i; i++) { - k = blocks[i]; - k *= mask1; - k = GMQCC_ROTL32(k, 15); - k *= mask2; - h ^= k; - h = GMQCC_ROTL32(h, 13); - h = h * 5 + 0xE6546B64; - } - - k = 0; - switch (len & 3) { - case 3: - k ^= tail[2] << 16; - case 2: - k ^= tail[1] << 8; - case 1: - k ^= tail[0]; - k *= mask1; - k = GMQCC_ROTL32(k, 15); - k *= mask2; - h ^= k; - } - - h ^= len; - h ^= h >> 16; - h *= 0x85EBCA6B; - h ^= h >> 13; - h *= 0xC2B2AE35; - h ^= h >> 16; - - return (size_t) (h % ht->size); -} -#undef GMQCC_ROTL32 -#else -/* We keep the old for reference */ -GMQCC_INLINE size_t util_hthash(hash_table_t *ht, const char *key) { - const uint32_t mix = 0x5BD1E995; - const uint32_t rot = 24; - size_t size = strlen(key); - uint32_t hash = 0x1EF0 /* LICRC TAB */ ^ size; - uint32_t alias = 0; - const unsigned char *data = (const unsigned char*)key; - - while (size >= 4) { - alias = (data[0] | (data[1] << 8) | (data[2] << 16) | (data[3] << 24)); - alias *= mix; - alias ^= alias >> rot; - alias *= mix; - - hash *= mix; - hash ^= alias; - - data += 4; - size -= 4; - } - - switch (size) { - case 3: hash ^= data[2] << 16; - case 2: hash ^= data[1] << 8; - case 1: hash ^= data[0]; - hash *= mix; - } - hash ^= hash >> 13; - hash *= mix; - hash ^= hash >> 15; - - return (size_t) (hash % ht->size); +size_t hash(const char *key); +size_t util_hthash(hash_table_t *ht, const char *key) { + return hash(key) % ht->size; } -#endif static hash_node_t *_util_htnewpair(const char *key, void *value) { hash_node_t *node; @@ -507,6 +525,7 @@ void *util_htget(hash_table_t *ht, const char *key) { return util_htgeth(ht, key, util_hthash(ht, key)); } +void *code_util_str_htgeth(hash_table_t *ht, const char *key, size_t bin); void *code_util_str_htgeth(hash_table_t *ht, const char *key, size_t bin) { hash_node_t *pair; size_t len, keylen; @@ -600,39 +619,60 @@ void util_htdel(hash_table_t *ht) { * The following functions below implement printing / dumping of statistical * information. */ -static void stat_dump_mem_contents(stat_mem_block_t *memory, uint16_t cols) { - uint32_t i, j; - for (i = 0; i < memory->size + ((memory->size % cols) ? (cols - memory->size % cols) : 0); i++) { - if (i % cols == 0) con_out(" 0x%06X: ", i); - if (i < memory->size) con_out("%02X " , 0xFF & ((unsigned char*)(memory + 1))[i]); - else con_out(" "); - - if ((uint16_t)(i % cols) == (cols - 1)) { - for (j = i - (cols - 1); j <= i; j++) { - con_out("%c", - (j >= memory->size) - ? ' ' - : (util_isprint(((unsigned char*)(memory + 1))[j])) - ? 0xFF & ((unsigned char*)(memory + 1)) [j] - : '.' - ); - } - con_out("\n"); +static void stat_dump_mem_contents(stat_mem_block_t *block, uint16_t cols) { + unsigned char *buffer = (unsigned char *)mem_a(cols); + unsigned char *memory = (unsigned char *)(block + 1); + size_t i; + + for (i = 0; i < block->size; i++) { + if (!(i % 16)) { + if (i != 0) + con_out(" %s\n", buffer); + con_out(" 0x%08X: ", i); } + + con_out(" %02X", memory[i]); + + buffer[i % cols] = ((memory[i] < 0x20) || (memory[i] > 0x7E)) + ? '.' + : memory[i]; + + buffer[(i % cols) + 1] = '\0'; + } + + while ((i % cols) != 0) { + con_out(" "); + i++; } + + con_out(" %s\n", buffer); + mem_d(buffer); } static void stat_dump_mem_leaks(void) { stat_mem_block_t *info; + /* we need access to the root for this */ + VALGRIND_MAKE_MEM_DEFINED(stat_mem_block_root, sizeof(stat_mem_block_t)); for (info = stat_mem_block_root; info; info = info->next) { - con_out("lost: %u (bytes) at %s:%u\n", + /* we need access to the block */ + VALGRIND_MAKE_MEM_DEFINED(info, sizeof(stat_mem_block_t)); + con_out("lost: %u (bytes) at %s:%u from expression `%s`\n", info->size, info->file, - info->line + info->line, + info->expr ); stat_dump_mem_contents(info, OPTS_OPTION_U16(OPTION_MEMDUMPCOLS)); + + /* + * we're finished with the access, the redzone should be marked + * inaccesible so that invalid read/writes that could 'step-into' + * those redzones will show up as invalid read/writes in valgrind. + */ + VALGRIND_MAKE_MEM_NOACCESS(info, sizeof(stat_mem_block_t)); } + VALGRIND_MAKE_MEM_NOACCESS(stat_mem_block_root, sizeof(stat_mem_block_t)); } static void stat_dump_mem_info(void) { @@ -679,12 +719,14 @@ void stat_info() { uint64_t mem = 0; con_out("Memory Statistics:\n\ - Total vectors allocated: %llu\n\ - Total string duplicates: %llu\n\ - Total hashtables allocated: %llu\n\ - Total unique vector sizes: %llu\n", + Total vectors allocated: %llu\n\ + Total string duplicates: %llu\n\ + Total string duplicate memory: %f (MB)\n\ + Total hashtables allocated: %llu\n\ + Total unique vector sizes: %llu\n", stat_used_vectors, stat_used_strdups, + (float)(stat_mem_strdups) / 1048576.0f, stat_used_hashtables, stat_type_vectors );