erts_sspa_data_t *
erts_sspa_create(size_t blk_sz, int pa_size, int nthreads, const char* name)
{
erts_sspa_data_t *data;
size_t tot_size;
size_t chunk_mem_size;
char *p;
char *chunk_start;
int cix;
int no_blocks = pa_size;
int no_blocks_per_chunk;
size_t aligned_blk_sz;
#if !defined(ERTS_STRUCTURE_ALIGNED_ALLOC)
/* Force 64-bit alignment... */
aligned_blk_sz = ((blk_sz - 1) / 8) * 8 + 8;
#else
/* Alignment of structure is enough... */
aligned_blk_sz = blk_sz;
#endif
if (!name) { /* schedulers only variant */
ASSERT(!nthreads);
nthreads = erts_no_schedulers;
}
else {
ASSERT(nthreads > 0);
}
if (nthreads == 1)
no_blocks_per_chunk = no_blocks;
else {
int extra = (no_blocks - 1)/4 + 1;
if (extra == 0)
extra = 1;
no_blocks_per_chunk = no_blocks;
no_blocks_per_chunk += extra * nthreads;
no_blocks_per_chunk /= nthreads;
}
no_blocks = no_blocks_per_chunk * nthreads;
chunk_mem_size = ERTS_ALC_CACHE_LINE_ALIGN_SIZE(sizeof(erts_sspa_chunk_header_t));
chunk_mem_size += aligned_blk_sz * no_blocks_per_chunk;
chunk_mem_size = ERTS_ALC_CACHE_LINE_ALIGN_SIZE(chunk_mem_size);
tot_size = ERTS_ALC_CACHE_LINE_ALIGN_SIZE(sizeof(erts_sspa_data_t));
tot_size += chunk_mem_size * nthreads;
p = erts_alloc_permanent_cache_aligned(ERTS_ALC_T_PRE_ALLOC_DATA, tot_size);
data = (erts_sspa_data_t *) p;
p += ERTS_ALC_CACHE_LINE_ALIGN_SIZE(sizeof(erts_sspa_data_t));
chunk_start = p;
data->chunks_mem_size = chunk_mem_size;
data->start = chunk_start;
data->end = chunk_start + chunk_mem_size * nthreads;
data->nthreads = nthreads;
if (name) { /* thread variant */
erts_tsd_key_create(&data->tsd_key, (char*)name);
erts_atomic_init_nob(&data->id_generator, 0);
}
/* Initialize all chunks */
for (cix = 0; cix < nthreads; cix++) {
erts_sspa_chunk_t *chnk = erts_sspa_cix2chunk(data, cix);
erts_sspa_chunk_header_t *chdr = &chnk->aligned.header;
erts_sspa_blk_t *blk;
int i;
erts_atomic_init_nob(&chdr->tail.data.last, (erts_aint_t) &chdr->tail.data.marker);
erts_atomic_init_nob(&chdr->tail.data.marker.next_atmc, ERTS_AINT_NULL);
erts_atomic_init_nob(&chdr->tail.data.um_refc[0], 0);
erts_atomic_init_nob(&chdr->tail.data.um_refc[1], 0);
erts_atomic32_init_nob(&chdr->tail.data.um_refc_ix, 0);
chdr->head.no_thr_progress_check = 0;
chdr->head.used_marker = 1;
chdr->head.first = &chdr->tail.data.marker;
chdr->head.unref_end = &chdr->tail.data.marker;
chdr->head.next.thr_progress = erts_thr_progress_current();
chdr->head.next.thr_progress_reached = 1;
chdr->head.next.um_refc_ix = 1;
chdr->head.next.unref_end = &chdr->tail.data.marker;
p = &chnk->data[0];
chdr->local.first = (erts_sspa_blk_t *) p;
blk = (erts_sspa_blk_t *) p;
for (i = 0; i < no_blocks_per_chunk; i++) {
blk = (erts_sspa_blk_t *) p;
p += aligned_blk_sz;
blk->next_ptr = (erts_sspa_blk_t *) p;
}
blk->next_ptr = NULL;
chdr->local.last = blk;
chdr->local.cnt = no_blocks_per_chunk;
chdr->local.lim = no_blocks_per_chunk / 3;
ERTS_SSPA_DBG_CHK_LCL(chdr);
}
return data;
}
erts_sspa_data_t *
erts_sspa_create(size_t blk_sz, int pa_size)
{
erts_sspa_data_t *data;
size_t tot_size;
size_t chunk_mem_size;
char *p;
char *chunk_start;
int cix;
int no_blocks = pa_size;
int no_blocks_per_chunk;
if (erts_no_schedulers == 1)
no_blocks_per_chunk = no_blocks;
else {
int extra = (no_blocks - 1)/4 + 1;
if (extra == 0)
extra = 1;
no_blocks_per_chunk = no_blocks;
no_blocks_per_chunk += extra*erts_no_schedulers;
no_blocks_per_chunk /= erts_no_schedulers;
}
no_blocks = no_blocks_per_chunk * erts_no_schedulers;
chunk_mem_size = ERTS_ALC_CACHE_LINE_ALIGN_SIZE(sizeof(erts_sspa_chunk_header_t));
chunk_mem_size += blk_sz * no_blocks_per_chunk;
chunk_mem_size = ERTS_ALC_CACHE_LINE_ALIGN_SIZE(chunk_mem_size);
tot_size = ERTS_ALC_CACHE_LINE_ALIGN_SIZE(sizeof(erts_sspa_data_t));
tot_size += chunk_mem_size*erts_no_schedulers;
p = erts_alloc_permanent_cache_aligned(ERTS_ALC_T_PRE_ALLOC_DATA, tot_size);
data = (erts_sspa_data_t *) p;
p += ERTS_ALC_CACHE_LINE_ALIGN_SIZE(sizeof(erts_sspa_data_t));
chunk_start = p;
data->chunks_mem_size = chunk_mem_size;
data->start = chunk_start;
data->end = chunk_start + chunk_mem_size*erts_no_schedulers;
/* Initialize all chunks */
for (cix = 0; cix < erts_no_schedulers; cix++) {
erts_sspa_chunk_t *chnk = erts_sspa_cix2chunk(data, cix);
erts_sspa_chunk_header_t *chdr = &chnk->aligned.header;
erts_sspa_blk_t *blk;
int i;
erts_atomic_init_nob(&chdr->tail.data.last, (erts_aint_t) &chdr->tail.data.marker);
erts_atomic_init_nob(&chdr->tail.data.marker.next_atmc, ERTS_AINT_NULL);
erts_atomic_init_nob(&chdr->tail.data.um_refc[0], 0);
erts_atomic_init_nob(&chdr->tail.data.um_refc[1], 0);
erts_atomic32_init_nob(&chdr->tail.data.um_refc_ix, 0);
chdr->head.no_thr_progress_check = 0;
chdr->head.used_marker = 1;
chdr->head.first = &chdr->tail.data.marker;
chdr->head.unref_end = &chdr->tail.data.marker;
chdr->head.next.thr_progress = erts_thr_progress_current();
chdr->head.next.thr_progress_reached = 1;
chdr->head.next.um_refc_ix = 1;
chdr->head.next.unref_end = &chdr->tail.data.marker;
p = &chnk->data[0];
chdr->local.first = (erts_sspa_blk_t *) p;
blk = (erts_sspa_blk_t *) p;
for (i = 0; i < no_blocks_per_chunk; i++) {
blk = (erts_sspa_blk_t *) p;
p += blk_sz;
blk->next_ptr = (erts_sspa_blk_t *) p;
}
blk->next_ptr = NULL;
chdr->local.last = blk;
chdr->local.cnt = no_blocks_per_chunk;
chdr->local.lim = no_blocks_per_chunk / 3;
ERTS_SSPA_DBG_CHK_LCL(chdr);
}
return data;
}
