static inline const fletcher_4_ops_t *
fletcher_4_impl_get(void)
{
fletcher_4_ops_t *ops = NULL;
const uint32_t impl = IMPL_READ(fletcher_4_impl_chosen);
switch (impl) {
case IMPL_FASTEST:
ASSERT(fletcher_4_initialized);
ops = &fletcher_4_fastest_impl;
break;
#if !defined(_KERNEL)
case IMPL_CYCLE: {
ASSERT(fletcher_4_initialized);
ASSERT3U(fletcher_4_supp_impls_cnt, >, 0);
static uint32_t cycle_count = 0;
uint32_t idx = (++cycle_count) % fletcher_4_supp_impls_cnt;
ops = fletcher_4_supp_impls[idx];
}
break;
#endif
default:
ASSERT3U(fletcher_4_supp_impls_cnt, >, 0);
ASSERT3U(impl, <, fletcher_4_supp_impls_cnt);
ops = fletcher_4_supp_impls[impl];
break;
}
ASSERT3P(ops, !=, NULL);
return (ops);
}
static void
fletcher_4_benchmark_impl(boolean_t native, char *data, uint64_t data_size)
{
struct fletcher_4_kstat *fastest_stat =
&fletcher_4_stat_data[fletcher_4_supp_impls_cnt];
hrtime_t start;
uint64_t run_bw, run_time_ns, best_run = 0;
zio_cksum_t zc;
uint32_t i, l, sel_save = IMPL_READ(fletcher_4_impl_chosen);
fletcher_checksum_func_t *fletcher_4_test = native ?
fletcher_4_native : fletcher_4_byteswap;
for (i = 0; i < fletcher_4_supp_impls_cnt; i++) {
struct fletcher_4_kstat *stat = &fletcher_4_stat_data[i];
uint64_t run_count = 0;
/* temporary set an implementation */
fletcher_4_impl_chosen = i;
kpreempt_disable();
start = gethrtime();
do {
for (l = 0; l < 32; l++, run_count++)
fletcher_4_test(data, data_size, NULL, &zc);
run_time_ns = gethrtime() - start;
} while (run_time_ns < FLETCHER_4_BENCH_NS);
kpreempt_enable();
run_bw = data_size * run_count * NANOSEC;
run_bw /= run_time_ns; /* B/s */
if (native)
stat->native = run_bw;
else
stat->byteswap = run_bw;
if (run_bw > best_run) {
best_run = run_bw;
if (native) {
fastest_stat->native = i;
FLETCHER_4_FASTEST_FN_COPY(native,
fletcher_4_supp_impls[i]);
} else {
fastest_stat->byteswap = i;
FLETCHER_4_FASTEST_FN_COPY(byteswap,
fletcher_4_supp_impls[i]);
}
}
}
/* restore original selection */
atomic_swap_32(&fletcher_4_impl_chosen, sel_save);
}
int
fletcher_4_impl_set(const char *val)
{
int err = -EINVAL;
uint32_t impl = IMPL_READ(fletcher_4_impl_chosen);
size_t i, val_len;
val_len = strlen(val);
while ((val_len > 0) && !!isspace(val[val_len-1])) /* trim '\n' */
val_len--;
/* check mandatory implementations */
for (i = 0; i < ARRAY_SIZE(fletcher_4_impl_selectors); i++) {
const char *name = fletcher_4_impl_selectors[i].fis_name;
if (val_len == strlen(name) &&
strncmp(val, name, val_len) == 0) {
impl = fletcher_4_impl_selectors[i].fis_sel;
err = 0;
break;
}
}
if (err != 0 && fletcher_4_initialized) {
/* check all supported implementations */
for (i = 0; i < fletcher_4_supp_impls_cnt; i++) {
const char *name = fletcher_4_supp_impls[i]->name;
if (val_len == strlen(name) &&
strncmp(val, name, val_len) == 0) {
impl = i;
err = 0;
break;
}
}
}
if (err == 0) {
atomic_swap_32(&fletcher_4_impl_chosen, impl);
membar_producer();
}
return (err);
}
static int
fletcher_4_param_get(char *buffer, struct kernel_param *unused)
{
const uint32_t impl = IMPL_READ(fletcher_4_impl_chosen);
char *fmt;
int i, cnt = 0;
/* list fastest */
fmt = (impl == IMPL_FASTEST) ? "[%s] " : "%s ";
cnt += sprintf(buffer + cnt, fmt, "fastest");
/* list all supported implementations */
for (i = 0; i < fletcher_4_supp_impls_cnt; i++) {
fmt = (i == impl) ? "[%s] " : "%s ";
cnt += sprintf(buffer + cnt, fmt,
fletcher_4_supp_impls[i]->name);
}
return (cnt);
}
