// Allocates a new block and updates core id => block mapping. 'old_block'
// points to the block that the caller thinks is attached to
// 'core_id'. 'old_block' may be NULL. Returns true if:
// - allocated a new block OR
// - 'core_id' => 'old_block' mapping changed (another thread allocated a
// block before lock was acquired).
static bool cl_allocate_core_local_block(uint32_t core_id,
cl_block* old_block) {
// Now that we have the lock, check if core-local mapping has changed.
cl_core_local_block* core_local_block = &g_log.core_local_blocks[core_id];
cl_block* block = cl_core_local_block_get_block(core_local_block);
if ((block != NULL) && (block != old_block)) {
return true;
}
if (block != NULL) {
cl_core_local_block_set_block(core_local_block, NULL);
cl_block_list_insert_at_tail(&g_log.dirty_block_list, block);
}
block = cl_allocate_block();
if (block == NULL) {
return false;
}
cl_core_local_block_set_block(core_local_block, block);
cl_block_enable_access(block);
return true;
}
/* Allocates a new block and updates core id => block mapping. 'old_block'
points to the block that the caller thinks is attached to
'core_id'. 'old_block' may be NULL. Returns non-zero if:
- allocated a new block OR
- 'core_id' => 'old_block' mapping changed (another thread allocated a
block before lock was acquired). */
static int cl_allocate_core_local_block(gpr_int32 core_id,
cl_block *old_block) {
/* Now that we have the lock, check if core-local mapping has changed. */
cl_core_local_block *core_local_block = &g_log.core_local_blocks[core_id];
cl_block *block = cl_core_local_block_get_block(core_local_block);
if ((block != NULL) && (block != old_block)) {
return 1;
}
if (block != NULL) {
cl_core_local_block_set_block(core_local_block, NULL);
cl_block_list_insert_at_tail(&g_log.dirty_block_list, block);
}
block = cl_allocate_block();
if (block == NULL) {
gpr_atm_rel_store(&g_log.is_full, 1);
return 0;
}
cl_core_local_block_set_block(core_local_block, block);
cl_block_enable_access(block);
return 1;
}
// External functions: primary stats_log interface
void census_log_initialize(size_t size_in_mb, int discard_old_records) {
// Check cacheline alignment.
GPR_ASSERT(sizeof(cl_block) % GPR_CACHELINE_SIZE == 0);
GPR_ASSERT(sizeof(cl_core_local_block) % GPR_CACHELINE_SIZE == 0);
GPR_ASSERT(!g_log.initialized);
g_log.discard_old_records = discard_old_records;
g_log.num_cores = gpr_cpu_num_cores();
// Ensure that we will not get any overflow in calaculating num_blocks
GPR_ASSERT(CL_LOG_2_MB >= CENSUS_LOG_2_MAX_RECORD_SIZE);
GPR_ASSERT(size_in_mb < 1000);
// Ensure at least 2x as many blocks as there are cores.
g_log.num_blocks =
(uint32_t)GPR_MAX(2 * g_log.num_cores, (size_in_mb << CL_LOG_2_MB) >>
CENSUS_LOG_2_MAX_RECORD_SIZE);
gpr_mu_init(&g_log.lock);
g_log.read_iterator_state = 0;
g_log.block_being_read = NULL;
g_log.core_local_blocks = (cl_core_local_block*)gpr_malloc_aligned(
g_log.num_cores * sizeof(cl_core_local_block), GPR_CACHELINE_SIZE_LOG);
memset(g_log.core_local_blocks, 0,
g_log.num_cores * sizeof(cl_core_local_block));
g_log.blocks = (cl_block*)gpr_malloc_aligned(
g_log.num_blocks * sizeof(cl_block), GPR_CACHELINE_SIZE_LOG);
memset(g_log.blocks, 0, g_log.num_blocks * sizeof(cl_block));
g_log.buffer = gpr_malloc(g_log.num_blocks * CENSUS_LOG_MAX_RECORD_SIZE);
memset(g_log.buffer, 0, g_log.num_blocks * CENSUS_LOG_MAX_RECORD_SIZE);
cl_block_list_initialize(&g_log.free_block_list);
cl_block_list_initialize(&g_log.dirty_block_list);
for (uint32_t i = 0; i < g_log.num_blocks; ++i) {
cl_block* block = g_log.blocks + i;
cl_block_initialize(block, g_log.buffer + (CENSUS_LOG_MAX_RECORD_SIZE * i));
cl_block_try_disable_access(block, 1 /* discard data */);
cl_block_list_insert_at_tail(&g_log.free_block_list, block);
}
gpr_atm_rel_store(&g_log.out_of_space_count, 0);
g_log.initialized = 1;
}
/* External functions: primary stats_log interface */
void census_log_initialize(size_t size_in_mb, int discard_old_records) {
gpr_int32 ix;
/* Check cacheline alignment. */
GPR_ASSERT(sizeof(cl_block) % GPR_CACHELINE_SIZE == 0);
GPR_ASSERT(sizeof(cl_core_local_block) % GPR_CACHELINE_SIZE == 0);
GPR_ASSERT(!g_log.initialized);
g_log.discard_old_records = discard_old_records;
g_log.num_cores = gpr_cpu_num_cores();
/* Ensure at least as many blocks as there are cores. */
g_log.num_blocks = GPR_MAX(
g_log.num_cores, (size_in_mb << 20) >> CENSUS_LOG_2_MAX_RECORD_SIZE);
gpr_mu_init(&g_log.lock);
g_log.read_iterator_state = 0;
g_log.block_being_read = NULL;
gpr_atm_rel_store(&g_log.is_full, 0);
g_log.core_local_blocks = (cl_core_local_block *)gpr_malloc_aligned(
g_log.num_cores * sizeof(cl_core_local_block), GPR_CACHELINE_SIZE_LOG);
memset(g_log.core_local_blocks, 0,
g_log.num_cores * sizeof(cl_core_local_block));
g_log.blocks = (cl_block *)gpr_malloc_aligned(
g_log.num_blocks * sizeof(cl_block), GPR_CACHELINE_SIZE_LOG);
memset(g_log.blocks, 0, g_log.num_blocks * sizeof(cl_block));
g_log.buffer = gpr_malloc(g_log.num_blocks * CENSUS_LOG_MAX_RECORD_SIZE);
memset(g_log.buffer, 0, g_log.num_blocks * CENSUS_LOG_MAX_RECORD_SIZE);
cl_block_list_initialize(&g_log.free_block_list);
cl_block_list_initialize(&g_log.dirty_block_list);
for (ix = 0; ix < g_log.num_blocks; ++ix) {
cl_block *block = g_log.blocks + ix;
cl_block_initialize(block,
g_log.buffer + (CENSUS_LOG_MAX_RECORD_SIZE * ix));
cl_block_try_disable_access(block, 1 /* discard data */);
cl_block_list_insert_at_tail(&g_log.free_block_list, block);
}
gpr_atm_rel_store(&g_log.out_of_space_count, 0);
g_log.initialized = 1;
}
