int
readn(int fd, char *buf, int bufsize, ioTunnel *en)
{
int nleft, nread;
nleft = bufsize;
while (nleft > 0) {
#ifdef WIN32
nread = recv(fd, buf, nleft, 0);
#else
nread = en == NULL ? system_read(fd, buf, nleft) : en->eRead(fd, buf, nleft);
#endif /* WIN32 */
if (nread < 0)
return (nread);
else if (nread == 0)
break;
nleft -= nread;
buf += nread;
}
return (bufsize - nleft);
}
ssize_t intercept_read(gint fd, gpointer buf, gint n) {
return system_read(fd, buf, n);
}
/*
* stress_aiol
* stress asynchronous I/O using the linux specific aio ABI
*/
static int stress_aiol(const args_t *args)
{
int fd, ret, rc = EXIT_FAILURE;
char filename[PATH_MAX];
char buf[64];
io_context_t ctx = 0;
uint64_t aio_linux_requests = DEFAULT_AIO_LINUX_REQUESTS;
uint8_t *buffer;
uint64_t aio_max_nr = DEFAULT_AIO_MAX_NR;
if (!get_setting("aiol-requests", &aio_linux_requests)) {
if (g_opt_flags & OPT_FLAGS_MAXIMIZE)
aio_linux_requests = MAX_AIO_REQUESTS;
if (g_opt_flags & OPT_FLAGS_MINIMIZE)
aio_linux_requests = MIN_AIO_REQUESTS;
}
if ((aio_linux_requests < MIN_AIO_REQUESTS) ||
(aio_linux_requests > MAX_AIO_REQUESTS)) {
pr_err("%s: iol_requests out of range", args->name);
return EXIT_FAILURE;
}
ret = system_read("/proc/sys/fs/aio-max-nr", buf, sizeof(buf));
if (ret > 0) {
if (sscanf(buf, "%" SCNu64, &aio_max_nr) != 1) {
/* Guess max */
aio_max_nr = DEFAULT_AIO_MAX_NR;
}
} else {
/* Guess max */
aio_max_nr = DEFAULT_AIO_MAX_NR;
}
aio_max_nr /= (args->num_instances == 0) ? 1 : args->num_instances;
if (aio_max_nr < 1)
aio_max_nr = 1;
if (aio_linux_requests > aio_max_nr) {
aio_linux_requests = aio_max_nr;
if (args->instance == 0)
pr_inf("%s: Limiting AIO requests to "
"%" PRIu64 " per stressor (avoids running out of resources)\n",
args->name, aio_linux_requests);
}
ret = posix_memalign((void **)&buffer, 4096,
aio_linux_requests * BUFFER_SZ);
if (ret) {
pr_inf("%s: Out of memory allocating buffers, errno=%d (%s)",
args->name, errno, strerror(errno));
return EXIT_NO_RESOURCE;
}
ret = io_setup(aio_linux_requests, &ctx);
if (ret < 0) {
/*
* The libaio interface returns -errno in the
* return value, so set errno accordingly
*/
errno = -ret;
if ((errno == EAGAIN) || (errno == EACCES)) {
pr_err("%s: io_setup failed, ran out of "
"available events, consider increasing "
"/proc/sys/fs/aio-max-nr, errno=%d (%s)\n",
args->name, errno, strerror(errno));
rc = EXIT_NO_RESOURCE;
goto free_buffer;
} else if (errno == ENOMEM) {
pr_err("%s: io_setup failed, ran out of "
"memory, errno=%d (%s)\n",
args->name, errno, strerror(errno));
rc = EXIT_NO_RESOURCE;
goto free_buffer;
} else if (errno == ENOSYS) {
pr_err("%s: io_setup failed, no io_setup "
"system call with this kernel, "
"errno=%d (%s)\n",
args->name, errno, strerror(errno));
rc = EXIT_NO_RESOURCE;
goto free_buffer;
} else {
pr_fail_err("io_setup");
rc = EXIT_FAILURE;
goto free_buffer;
}
}
ret = stress_temp_dir_mk_args(args);
if (ret < 0) {
rc = exit_status(-ret);
goto free_buffer;
}
(void)stress_temp_filename_args(args,
filename, sizeof(filename), mwc32());
if ((fd = open(filename, O_CREAT | O_RDWR | O_DIRECT, S_IRUSR | S_IWUSR)) < 0) {
rc = exit_status(errno);
pr_fail_err("open");
goto finish;
}
(void)unlink(filename);
do {
struct iocb cb[aio_linux_requests];
struct iocb *cbs[aio_linux_requests];
struct io_event events[aio_linux_requests];
uint8_t *buffers[aio_linux_requests];
uint8_t *bufptr = buffer;
uint64_t i;
long n;
for (i = 0; i < aio_linux_requests; i++, bufptr += BUFFER_SZ) {
buffers[i] = bufptr;
aio_linux_fill_buffer(i, buffers[i], BUFFER_SZ);
}
(void)memset(cb, 0, sizeof(cb));
for (i = 0; i < aio_linux_requests; i++) {
cb[i].aio_fildes = fd;
cb[i].aio_lio_opcode = IO_CMD_PWRITE;
cb[i].u.c.buf = buffers[i];
cb[i].u.c.offset = mwc16() * BUFFER_SZ;
cb[i].u.c.nbytes = BUFFER_SZ;
cbs[i] = &cb[i];
}
ret = io_submit(ctx, (long)aio_linux_requests, cbs);
if (ret < 0) {
errno = -ret;
if (errno == EAGAIN)
continue;
pr_fail_err("io_submit");
break;
}
n = aio_linux_requests;
do {
struct timespec timeout, *timeout_ptr;
if (clock_gettime(CLOCK_REALTIME, &timeout) < 0) {
timeout_ptr = NULL;
} else {
timeout.tv_nsec += 1000000;
if (timeout.tv_nsec > 1000000000) {
timeout.tv_nsec -= 1000000000;
timeout.tv_sec++;
}
timeout_ptr = &timeout;
}
ret = io_getevents(ctx, 1, n, events, timeout_ptr);
if (ret < 0) {
errno = -ret;
if (errno == EINTR) {
if (g_keep_stressing_flag)
continue;
else
break;
}
pr_fail_err("io_getevents");
break;
} else {
n -= ret;
}
} while ((n > 0) && g_keep_stressing_flag);
inc_counter(args);
} while (keep_stressing());
rc = EXIT_SUCCESS;
(void)close(fd);
finish:
(void)io_destroy(ctx);
(void)stress_temp_dir_rm_args(args);
free_buffer:
free(buffer);
return rc;
}
void perf_stat_dump(FILE *yaml, proc_info_t *procs_head, const double duration)
{
bool no_perf_stats = true;
proc_info_t *pi;
#if defined(HAVE_LOCALE_H)
(void)setlocale(LC_ALL, "");
#endif
pr_yaml(yaml, "perfstats:\n");
for (pi = procs_head; pi; pi = pi->next) {
int p;
uint64_t counter_totals[STRESS_PERF_MAX];
uint64_t total_cpu_cycles = 0;
uint64_t total_cache_refs = 0;
uint64_t total_branches = 0;
bool got_data = false;
char *munged;
(void)memset(counter_totals, 0, sizeof(counter_totals));
/* Sum totals across all instances of the stressor */
for (p = 0; p < STRESS_PERF_MAX && perf_info[p].label; p++) {
int32_t j;
stress_perf_t *sp = &pi->stats[0]->sp;
if (!perf_stat_succeeded(sp))
continue;
for (j = 0; j < pi->started_procs; j++) {
const uint64_t counter = sp->perf_stat[p].counter;
if (counter == STRESS_PERF_INVALID) {
counter_totals[p] = STRESS_PERF_INVALID;
break;
}
counter_totals[p] += counter;
got_data |= (counter > 0);
}
if (perf_info[p].type == PERF_TYPE_HARDWARE) {
unsigned long config = perf_info[p].config;
if (config == PERF_COUNT_HW_CPU_CYCLES)
total_cpu_cycles = counter_totals[p];
else if (config == PERF_COUNT_HW_CACHE_REFERENCES)
total_cache_refs = counter_totals[p];
else if (config == PERF_COUNT_HW_BRANCH_INSTRUCTIONS)
total_branches = counter_totals[p];
}
}
if (!got_data)
continue;
munged = stress_munge_underscore(pi->stressor->name);
pr_inf("%s:\n", munged);
pr_yaml(yaml, " - stressor: %s\n", munged);
pr_yaml(yaml, " duration: %f\n", duration);
for (p = 0; p < STRESS_PERF_MAX && perf_info[p].label; p++) {
const char *l = perf_info[p].label;
uint64_t ct = counter_totals[p];
if (l && (ct != STRESS_PERF_INVALID)) {
char extra[32];
char yaml_label[128];
*extra = '\0';
no_perf_stats = false;
if (perf_info[p].type == PERF_TYPE_HARDWARE) {
unsigned long config = perf_info[p].config;
if ((config == PERF_COUNT_HW_INSTRUCTIONS) &&
(total_cpu_cycles > 0))
(void)snprintf(extra, sizeof(extra),
" (%.3f instr. per cycle)",
(double)ct / (double)total_cpu_cycles);
else if ((config == PERF_COUNT_HW_CACHE_MISSES) &&
(total_cache_refs > 0))
(void)snprintf(extra, sizeof(extra),
" (%5.2f%%)",
100.0 * (double)ct / (double)total_cache_refs);
else if ((config == PERF_COUNT_HW_BRANCH_MISSES) &&
(total_branches > 0))
(void)snprintf(extra, sizeof(extra),
" (%5.2f%%)",
100.0 * (double)ct / (double)total_branches);
}
pr_inf("%'26" PRIu64 " %-24s %s%s\n",
ct, l, perf_stat_scale(ct, duration),
extra);
perf_yaml_label(yaml_label, l, sizeof(yaml_label));
pr_yaml(yaml, " %s_total: %" PRIu64
"\n", yaml_label, ct);
pr_yaml(yaml, " %s_per_second: %f\n",
yaml_label, (double)ct / duration);
}
}
pr_yaml(yaml, "\n");
}
if (no_perf_stats) {
if (geteuid() != 0) {
char buffer[64];
int ret;
bool paranoid = false;
int level = 0;
static char *path = "/proc/sys/kernel/perf_event_paranoid";
ret = system_read(path, buffer, sizeof(buffer) - 1);
if (ret > 0) {
if (sscanf(buffer, "%5d", &level) == 1)
paranoid = true;
}
if (paranoid & (level > 1)) {
pr_inf("Cannot read perf counters, "
"do not have CAP_SYS_ADMIN capability "
"or %s is set too high (%d)\n",
path, level);
}
} else {
pr_inf("perf counters are not available "
"on this device\n");
}
}
}