static int aio_event_thread(void *data)
{
struct aio_threadinfo *tinfo = data;
struct aio_output *output = tinfo->output;
struct aio_threadinfo *other = &output->tinfo[2];
int err = -ENOMEM;
MARS_DBG("event thread has started.\n");
//set_user_nice(current, -20);
use_fake_mm();
if (!current->mm)
goto err;
err = aio_start_thread(output, &output->tinfo[2], aio_sync_thread, 'y');
if (unlikely(err < 0))
goto err;
while (!brick_thread_should_stop() || atomic_read(&tinfo->queued_sum) > 0) {
mm_segment_t oldfs;
int count;
int i;
struct timespec timeout = {
.tv_sec = 1,
};
struct io_event events[MARS_MAX_AIO_READ];
oldfs = get_fs();
set_fs(get_ds());
/* TODO: don't timeout upon termination.
* Probably we should submit a dummy request.
*/
count = sys_io_getevents(output->ctxp, 1, MARS_MAX_AIO_READ, events, &timeout);
set_fs(oldfs);
if (likely(count > 0)) {
atomic_sub(count, &output->submit_count);
}
for (i = 0; i < count; i++) {
struct aio_mref_aspect *mref_a = (void*)events[i].data;
struct mref_object *mref;
int err = events[i].res;
if (!mref_a) {
continue; // this was a dummy request
}
mref = mref_a->object;
MARS_IO("AIO done %p pos = %lld len = %d rw = %d\n", mref, mref->ref_pos, mref->ref_len, mref->ref_rw);
mapfree_set(output->mf, mref->ref_pos, mref->ref_pos + mref->ref_len);
if (output->brick->o_fdsync
&& err >= 0
&& mref->ref_rw != READ
&& !mref->ref_skip_sync
&& !mref_a->resubmit++) {
// workaround for non-implemented AIO FSYNC operation
if (output->mf &&
output->mf->mf_filp &&
output->mf->mf_filp->f_op &&
!output->mf->mf_filp->f_op->aio_fsync) {
mars_trace(mref, "aio_fsync");
_enqueue(other, mref_a, mref->ref_prio, true);
continue;
}
err = aio_submit(output, mref_a, true);
if (likely(err >= 0))
continue;
}
_complete(output, mref_a, err);
}
}
err = 0;
err:
MARS_DBG("event thread has stopped, err = %d\n", err);
aio_stop_thread(output, 2, false);
unuse_fake_mm();
tinfo->terminated = true;
wake_up_interruptible_all(&tinfo->terminate_event);
return err;
}
#if 1
/* This should go to fs/open.c (as long as vfs_submit() is not implemented)
*/
#include <linux/fdtable.h>
void fd_uninstall(unsigned int fd)
{
struct files_struct *files = current->files;
struct fdtable *fdt;
MARS_DBG("fd = %d\n", fd);
if (unlikely(fd < 0)) {
MARS_ERR("bad fd = %d\n", fd);
return;
}
spin_lock(&files->file_lock);
fdt = files_fdtable(files);
rcu_assign_pointer(fdt->fd[fd], NULL);
spin_unlock(&files->file_lock);
}
EXPORT_SYMBOL(fd_uninstall);
#endif
static
atomic_t ioctx_count = ATOMIC_INIT(0);
static
void _destroy_ioctx(struct aio_output *output)
{
if (unlikely(!output))
goto done;
aio_stop_thread(output, 1, true);
use_fake_mm();
if (likely(output->ctxp)) {
mm_segment_t oldfs;
int err;
MARS_DBG("ioctx count = %d destroying %p\n", atomic_read(&ioctx_count), (void*)output->ctxp);
oldfs = get_fs();
set_fs(get_ds());
err = sys_io_destroy(output->ctxp);
set_fs(oldfs);
atomic_dec(&ioctx_count);
MARS_DBG("ioctx count = %d status = %d\n", atomic_read(&ioctx_count), err);
output->ctxp = 0;
}
if (likely(output->fd >= 0)) {
MARS_DBG("destroying fd %d\n", output->fd);
fd_uninstall(output->fd);
put_unused_fd(output->fd);
output->fd = -1;
}
done:
if (likely(current->mm)) {
unuse_fake_mm();
}
}
static
int _create_ioctx(struct aio_output *output)
{
struct file *file;
mm_segment_t oldfs;
int err = -EINVAL;
CHECK_PTR_NULL(output, done);
CHECK_PTR_NULL(output->mf, done);
file = output->mf->mf_filp;
CHECK_PTR_NULL(file, done);
/* TODO: this is provisionary. We only need it for sys_io_submit()
* which uses userspace concepts like file handles.
* This should be accompanied by a future kernelsapce vfs_submit() or
* do_submit() which currently does not exist :(
*/
err = get_unused_fd();
MARS_DBG("file %p '%s' new fd = %d\n", file, output->mf->mf_name, err);
if (unlikely(err < 0)) {
MARS_ERR("cannot get fd, err=%d\n", err);
goto done;
}
output->fd = err;
fd_install(err, file);
use_fake_mm();
err = -ENOMEM;
if (unlikely(!current->mm)) {
MARS_ERR("cannot fake mm\n");
goto done;
}
MARS_DBG("ioctx count = %d old = %p\n", atomic_read(&ioctx_count), (void*)output->ctxp);
output->ctxp = 0;
oldfs = get_fs();
set_fs(get_ds());
err = sys_io_setup(MARS_MAX_AIO, &output->ctxp);
set_fs(oldfs);
if (likely(output->ctxp))
atomic_inc(&ioctx_count);
MARS_DBG("ioctx count = %d new = %p status = %d\n", atomic_read(&ioctx_count), (void*)output->ctxp, err);
if (unlikely(err < 0)) {
MARS_ERR("io_setup failed, err=%d\n", err);
goto done;
}
err = aio_start_thread(output, &output->tinfo[1], aio_event_thread, 'e');
if (unlikely(err < 0)) {
MARS_ERR("could not start event thread\n");
goto done;
}
done:
if (likely(current->mm)) {
unuse_fake_mm();
}
return err;
}
static int aio_submit_thread(void *data)
{
struct aio_threadinfo *tinfo = data;
struct aio_output *output = tinfo->output;
struct file *file;
int err = -EINVAL;
MARS_DBG("submit thread has started.\n");
file = output->mf->mf_filp;
use_fake_mm();
while (!brick_thread_should_stop() || atomic_read(&output->read_count) + atomic_read(&output->write_count) + atomic_read(&tinfo->queued_sum) > 0) {
struct aio_mref_aspect *mref_a;
struct mref_object *mref;
int sleeptime;
int status;
wait_event_interruptible_timeout(
tinfo->event,
atomic_read(&tinfo->queued_sum) > 0,
HZ / 4);
mref_a = _dequeue(tinfo);
if (!mref_a) {
continue;
}
mref = mref_a->object;
status = -EINVAL;
CHECK_PTR(mref, error);
mapfree_set(output->mf, mref->ref_pos, -1);
if (mref->ref_rw) {
insert_dirty(output, mref_a);
}
// check for reads exactly at EOF (special case)
if (mref->ref_pos == mref->ref_total_size &&
!mref->ref_rw &&
mref->ref_timeout > 0) {
loff_t total_size = i_size_read(file->f_mapping->host);
loff_t len = total_size - mref->ref_pos;
if (len > 0) {
mref->ref_total_size = total_size;
mref->ref_len = len;
} else {
if (!mref_a->start_jiffies) {
mref_a->start_jiffies = jiffies;
}
if ((long long)jiffies - mref_a->start_jiffies <= mref->ref_timeout) {
if (atomic_read(&tinfo->queued_sum) <= 0) {
atomic_inc(&output->total_msleep_count);
brick_msleep(1000 * 4 / HZ);
}
_enqueue(tinfo, mref_a, MARS_PRIO_LOW, true);
continue;
}
MARS_DBG("ENODATA %lld\n", len);
_complete(output, mref_a, -ENODATA);
continue;
}
}
sleeptime = 1;
for (;;) {
status = aio_submit(output, mref_a, false);
if (likely(status != -EAGAIN)) {
break;
}
atomic_inc(&output->total_delay_count);
brick_msleep(sleeptime);
if (sleeptime < 100) {
sleeptime++;
}
}
error:
if (unlikely(status < 0)) {
MARS_IO("submit_count = %d status = %d\n", atomic_read(&output->submit_count), status);
_complete_mref(output, mref, status);
}
}
MARS_DBG("submit thread has stopped, status = %d.\n", err);
if (likely(current->mm)) {
unuse_fake_mm();
}
tinfo->terminated = true;
wake_up_interruptible_all(&tinfo->terminate_event);
return err;
}
static int aio_get_info(struct aio_output *output, struct mars_info *info)
{
struct file *file;
loff_t min;
loff_t max;
if (unlikely(!output ||
!output->mf ||
!(file = output->mf->mf_filp) ||
!file->f_mapping ||
!file->f_mapping->host))
return -EINVAL;
info->tf_align = 1;
info->tf_min_size = 1;
/* Workaround for races in the page cache.
*
* It appears that concurrent reads and writes seem to
* result in inconsistent reads in some very rare cases, due to
* races. Sometimes, the inode claims that the file has been already
* appended by a write operation, but the data has not actually hit
* the page cache, such that a concurrent read gets NULL blocks.
*/
min = i_size_read(file->f_mapping->host);
max = 0;
if (!output->brick->is_static_device) {
get_dirty(output, &min, &max);
}
info->current_size = min;
MARS_DBG("determined file size = %lld\n", info->current_size);
return 0;
}
//////////////// informational / statistics ///////////////
static noinline
char *aio_statistics(struct aio_brick *brick, int verbose)
{
struct aio_output *output = brick->outputs[0];
char *res = brick_string_alloc(4096);
char *sync = NULL;
int pos = 0;
if (!res)
return NULL;
pos += report_timing(&timings[0], res + pos, 4096 - pos);
pos += report_timing(&timings[1], res + pos, 4096 - pos);
pos += report_timing(&timings[2], res + pos, 4096 - pos);
snprintf(res + pos, 4096 - pos,
"total "
"reads = %d "
"writes = %d "
"allocs = %d "
"submits = %d "
"again = %d "
"delays = %d "
"msleeps = %d "
"fdsyncs = %d "
"fdsync_waits = %d "
"map_free = %d | "
"flying reads = %d "
"writes = %d "
"allocs = %d "
"submits = %d "
"q0 = %d "
"q1 = %d "
"q2 = %d "
"| total "
"q0 = %d "
"q1 = %d "
"q2 = %d "
"%s\n",
atomic_read(&output->total_read_count),
atomic_read(&output->total_write_count),
atomic_read(&output->total_alloc_count),
atomic_read(&output->total_submit_count),
atomic_read(&output->total_again_count),
atomic_read(&output->total_delay_count),
atomic_read(&output->total_msleep_count),
atomic_read(&output->total_fdsync_count),
atomic_read(&output->total_fdsync_wait_count),
atomic_read(&output->total_mapfree_count),
atomic_read(&output->read_count),
atomic_read(&output->write_count),
atomic_read(&output->alloc_count),
atomic_read(&output->submit_count),
atomic_read(&output->tinfo[0].queued_sum),
atomic_read(&output->tinfo[1].queued_sum),
atomic_read(&output->tinfo[2].queued_sum),
atomic_read(&output->tinfo[0].total_enqueue_count),
atomic_read(&output->tinfo[1].total_enqueue_count),
atomic_read(&output->tinfo[2].total_enqueue_count),
sync ? sync : "");
if (sync)
brick_string_free(sync);
return res;
}
static noinline
void aio_reset_statistics(struct aio_brick *brick)
{
struct aio_output *output = brick->outputs[0];
int i;
atomic_set(&output->total_read_count, 0);
atomic_set(&output->total_write_count, 0);
atomic_set(&output->total_alloc_count, 0);
atomic_set(&output->total_submit_count, 0);
atomic_set(&output->total_again_count, 0);
atomic_set(&output->total_delay_count, 0);
atomic_set(&output->total_msleep_count, 0);
atomic_set(&output->total_fdsync_count, 0);
atomic_set(&output->total_fdsync_wait_count, 0);
atomic_set(&output->total_mapfree_count, 0);
for (i = 0; i < 3; i++) {
struct aio_threadinfo *tinfo = &output->tinfo[i];
atomic_set(&tinfo->total_enqueue_count, 0);
}
}
//////////////// object / aspect constructors / destructors ///////////////
static int aio_mref_aspect_init_fn(struct generic_aspect *_ini)
{
struct aio_mref_aspect *ini = (void*)_ini;
INIT_LIST_HEAD(&ini->io_head);
INIT_LIST_HEAD(&ini->dirty_head);
return 0;
}
static void aio_mref_aspect_exit_fn(struct generic_aspect *_ini)
{
struct aio_mref_aspect *ini = (void*)_ini;
CHECK_HEAD_EMPTY(&ini->dirty_head);
CHECK_HEAD_EMPTY(&ini->io_head);
}
MARS_MAKE_STATICS(aio);
////////////////////// brick constructors / destructors ////////////////////
static int aio_brick_construct(struct aio_brick *brick)
{
return 0;
}
static int aio_switch(struct aio_brick *brick)
{
static int index;
struct aio_output *output = brick->outputs[0];
const char *path = output->brick->brick_path;
int flags = O_RDWR | O_LARGEFILE;
int status = 0;
MARS_DBG("power.button = %d\n", brick->power.button);
if (!brick->power.button)
goto cleanup;
if (brick->power.led_on || output->mf)
goto done;
mars_power_led_off((void*)brick, false);
if (brick->o_creat) {
flags |= O_CREAT;
MARS_DBG("using O_CREAT on %s\n", path);
}
if (brick->o_direct) {
flags |= O_DIRECT;
MARS_DBG("using O_DIRECT on %s\n", path);
}
output->mf = mapfree_get(path, flags);
if (unlikely(!output->mf)) {
MARS_ERR("could not open file = '%s' flags = %d\n", path, flags);
status = -ENOENT;
goto err;
}
output->index = ++index;
status = _create_ioctx(output);
if (unlikely(status < 0)) {
MARS_ERR("could not create ioctx, status = %d\n", status);
goto err;
}
status = aio_start_thread(output, &output->tinfo[0], aio_submit_thread, 's');
if (unlikely(status < 0)) {
MARS_ERR("could not start theads, status = %d\n", status);
goto err;
}
MARS_DBG("opened file '%s'\n", path);
mars_power_led_on((void*)brick, true);
done:
return 0;
err:
MARS_ERR("status = %d\n", status);
cleanup:
if (brick->power.led_off) {
goto done;
}
mars_power_led_on((void*)brick, false);
aio_stop_thread(output, 0, false);
_destroy_ioctx(output);
mars_power_led_off((void*)brick,
(output->tinfo[0].thread == NULL &&
output->tinfo[1].thread == NULL &&
output->tinfo[2].thread == NULL));
MARS_DBG("switch off led_off = %d status = %d\n", brick->power.led_off, status);
if (brick->power.led_off) {
if (output->mf) {
MARS_DBG("closing file = '%s'\n", output->mf->mf_name);
mapfree_put(output->mf);
output->mf = NULL;
}
}
return status;
}
static int aio_output_construct(struct aio_output *output)
{
INIT_LIST_HEAD(&output->dirty_anchor);
spin_lock_init(&output->dirty_lock);
init_waitqueue_head(&output->fdsync_event);
output->fd = -1;
return 0;
}
/*
* stress_aio_linux
* stress asynchronous I/O using the linux specific aio ABI
*/
int stress_aio_linux(
uint64_t *const counter,
const uint32_t instance,
const uint64_t max_ops,
const char *name)
{
int fd, rc = EXIT_FAILURE;
char filename[PATH_MAX];
const pid_t pid = getpid();
aio_context_t ctx = 0;
if (!set_aio_linux_requests) {
if (opt_flags & OPT_FLAGS_MAXIMIZE)
opt_aio_linux_requests = MAX_AIO_REQUESTS;
if (opt_flags & OPT_FLAGS_MINIMIZE)
opt_aio_linux_requests = MIN_AIO_REQUESTS;
}
if (sys_io_setup(opt_aio_linux_requests, &ctx) < 0) {
pr_failed_err(name, "io_setup");
return EXIT_FAILURE;
}
if (stress_temp_dir_mk(name, pid, instance) < 0) {
return EXIT_FAILURE;
}
(void)stress_temp_filename(filename, sizeof(filename),
name, pid, instance, mwc32());
(void)umask(0077);
if ((fd = open(filename, O_CREAT | O_RDWR, S_IRUSR | S_IWUSR)) < 0) {
pr_failed_err(name, "open");
goto finish;
}
(void)unlink(filename);
do {
struct iocb cb[opt_aio_linux_requests];
struct iocb *cbs[opt_aio_linux_requests];
struct io_event events[opt_aio_linux_requests];
uint8_t buffers[opt_aio_linux_requests][BUFFER_SZ];
int ret, i;
long n;
for (i = 0; i < opt_aio_linux_requests; i++)
aio_linux_fill_buffer(i, buffers[i], BUFFER_SZ);
memset(cb, 0, sizeof(cb));
for (i = 0; i < opt_aio_linux_requests; i++) {
cb[i].aio_fildes = fd;
cb[i].aio_lio_opcode = IOCB_CMD_PWRITE;
cb[i].aio_buf = (long)buffers[i];
cb[i].aio_offset = mwc16() * BUFFER_SZ;
cb[i].aio_nbytes = BUFFER_SZ;
cbs[i] = &cb[i];
}
ret = sys_io_submit(ctx, opt_aio_linux_requests, cbs);
if (ret < 0) {
if (errno == EAGAIN)
continue;
pr_failed_err(name, "io_submit");
break;
}
n = opt_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 = sys_io_getevents(ctx, 1, n, events, timeout_ptr);
if (ret < 0) {
if ((errno == EINTR) && (opt_do_run))
continue;
pr_failed_err(name, "io_getevents");
break;
} else {
n -= ret;
}
} while ((n > 0) && opt_do_run);
(*counter)++;
} while (opt_do_run && (!max_ops || *counter < max_ops));
rc = EXIT_SUCCESS;
(void)close(fd);
finish:
(void)sys_io_destroy(ctx);
(void)stress_temp_dir_rm(name, pid, instance);
return rc;
}
