#include "../fio.h"
#include "../optgroup.h"
+#ifdef CONFIG_RBD_BLKIN
+#include <zipkin_c.h>
+#endif
+
+#ifdef CONFIG_RBD_POLL
+/* add for poll */
+#include <poll.h>
+#include <sys/eventfd.h>
+#endif
struct fio_rbd_iou {
struct io_u *io_u;
rbd_completion_t completion;
int io_seen;
int io_complete;
+#ifdef CONFIG_RBD_BLKIN
+ struct blkin_trace_info info;
+#endif
};
struct rbd_data {
rbd_image_t image;
struct io_u **aio_events;
struct io_u **sort_events;
+ int fd; /* add for poll */
+ bool connected;
};
struct rbd_options {
{
struct rbd_data *rbd;
- if (td->io_ops->data)
+ if (td->io_ops_data)
return 0;
rbd = calloc(1, sizeof(struct rbd_data));
if (!rbd)
goto failed;
+ rbd->connected = false;
+
+ /* add for poll, init fd: -1 */
+ rbd->fd = -1;
+
rbd->aio_events = calloc(td->o.iodepth, sizeof(struct io_u *));
if (!rbd->aio_events)
goto failed;
return 0;
failed:
- if (rbd)
+ if (rbd) {
+ if (rbd->aio_events)
+ free(rbd->aio_events);
+ if (rbd->sort_events)
+ free(rbd->sort_events);
free(rbd);
+ }
return 1;
}
+#ifdef CONFIG_RBD_POLL
+static bool _fio_rbd_setup_poll(struct rbd_data *rbd)
+{
+ int r;
+
+ /* add for rbd poll */
+ rbd->fd = eventfd(0, EFD_NONBLOCK);
+ if (rbd->fd < 0) {
+ log_err("eventfd failed.\n");
+ return false;
+ }
+
+ r = rbd_set_image_notification(rbd->image, rbd->fd, EVENT_TYPE_EVENTFD);
+ if (r < 0) {
+ log_err("rbd_set_image_notification failed.\n");
+ close(rbd->fd);
+ rbd->fd = -1;
+ return false;
+ }
+
+ return true;
+}
+#else
+static bool _fio_rbd_setup_poll(struct rbd_data *rbd)
+{
+ return true;
+}
+#endif
+
static int _fio_rbd_connect(struct thread_data *td)
{
- struct rbd_data *rbd = td->io_ops->data;
+ struct rbd_data *rbd = td->io_ops_data;
struct rbd_options *o = td->eo;
int r;
char *client_name = NULL;
/*
- * If we specify cluser name, the rados_creat2
+ * If we specify cluser name, the rados_create2
* will not assume 'client.'. name is considered
* as a full type.id namestr
*/
- if (!index(o->client_name, '.')) {
- client_name = calloc(1, strlen("client.") +
- strlen(o->client_name) + 1);
- strcat(client_name, "client.");
- o->client_name = strcat(client_name, o->client_name);
+ if (o->client_name) {
+ if (!index(o->client_name, '.')) {
+ client_name = calloc(1, strlen("client.") +
+ strlen(o->client_name) + 1);
+ strcat(client_name, "client.");
+ strcat(client_name, o->client_name);
+ } else {
+ client_name = o->client_name;
+ }
}
+
r = rados_create2(&rbd->cluster, o->cluster_name,
- o->client_name, 0);
+ client_name, 0);
+
+ if (client_name && !index(o->client_name, '.'))
+ free(client_name);
} else
r = rados_create(&rbd->cluster, o->client_name);
log_err("rbd_open failed.\n");
goto failed_open;
}
+
+ if (!_fio_rbd_setup_poll(rbd))
+ goto failed_poll;
+
return 0;
+failed_poll:
+ rbd_close(rbd->image);
+ rbd->image = NULL;
failed_open:
rados_ioctx_destroy(rbd->io_ctx);
rbd->io_ctx = NULL;
if (!rbd)
return;
+ /* close eventfd */
+ if (rbd->fd != -1) {
+ close(rbd->fd);
+ rbd->fd = -1;
+ }
+
/* shutdown everything */
if (rbd->image) {
rbd_close(rbd->image);
static struct io_u *fio_rbd_event(struct thread_data *td, int event)
{
- struct rbd_data *rbd = td->io_ops->data;
+ struct rbd_data *rbd = td->io_ops_data;
return rbd->aio_events[event];
}
static int rbd_iter_events(struct thread_data *td, unsigned int *events,
unsigned int min_evts, int wait)
{
- struct rbd_data *rbd = td->io_ops->data;
+ struct rbd_data *rbd = td->io_ops_data;
unsigned int this_events = 0;
struct io_u *io_u;
- int i, sidx;
+ int i, sidx = 0;
+
+#ifdef CONFIG_RBD_POLL
+ int ret = 0;
+ int event_num = 0;
+ struct fio_rbd_iou *fri = NULL;
+ rbd_completion_t comps[min_evts];
+
+ struct pollfd pfd;
+ pfd.fd = rbd->fd;
+ pfd.events = POLLIN;
+
+ ret = poll(&pfd, 1, -1);
+ if (ret <= 0)
+ return 0;
+
+ assert(pfd.revents & POLLIN);
+
+ event_num = rbd_poll_io_events(rbd->image, comps, min_evts);
- sidx = 0;
+ for (i = 0; i < event_num; i++) {
+ fri = rbd_aio_get_arg(comps[i]);
+ io_u = fri->io_u;
+#else
io_u_qiter(&td->io_u_all, io_u, i) {
+#endif
if (!(io_u->flags & IO_U_F_FLIGHT))
continue;
if (rbd_io_u_seen(io_u))
static int fio_rbd_queue(struct thread_data *td, struct io_u *io_u)
{
- struct rbd_data *rbd = td->io_ops->data;
+ struct rbd_data *rbd = td->io_ops_data;
struct fio_rbd_iou *fri = io_u->engine_data;
int r = -1;
}
if (io_u->ddir == DDIR_WRITE) {
+#ifdef CONFIG_RBD_BLKIN
+ blkin_init_trace_info(&fri->info);
+ r = rbd_aio_write_traced(rbd->image, io_u->offset, io_u->xfer_buflen,
+ io_u->xfer_buf, fri->completion, &fri->info);
+#else
r = rbd_aio_write(rbd->image, io_u->offset, io_u->xfer_buflen,
io_u->xfer_buf, fri->completion);
+#endif
if (r < 0) {
log_err("rbd_aio_write failed.\n");
goto failed_comp;
}
} else if (io_u->ddir == DDIR_READ) {
+#ifdef CONFIG_RBD_BLKIN
+ blkin_init_trace_info(&fri->info);
+ r = rbd_aio_read_traced(rbd->image, io_u->offset, io_u->xfer_buflen,
+ io_u->xfer_buf, fri->completion, &fri->info);
+#else
r = rbd_aio_read(rbd->image, io_u->offset, io_u->xfer_buflen,
io_u->xfer_buf, fri->completion);
+#endif
if (r < 0) {
log_err("rbd_aio_read failed.\n");
static int fio_rbd_init(struct thread_data *td)
{
int r;
+ struct rbd_data *rbd = td->io_ops_data;
+
+ if (rbd->connected)
+ return 0;
r = _fio_rbd_connect(td);
if (r) {
static void fio_rbd_cleanup(struct thread_data *td)
{
- struct rbd_data *rbd = td->io_ops->data;
+ struct rbd_data *rbd = td->io_ops_data;
if (rbd) {
_fio_rbd_disconnect(rbd);
log_err("fio_setup_rbd_data failed.\n");
goto cleanup;
}
- td->io_ops->data = rbd;
+ td->io_ops_data = rbd;
/* librbd does not allow us to run first in the main thread and later
* in a fork child. It needs to be the same process context all the
log_err("fio_rbd_connect failed.\n");
goto cleanup;
}
+ rbd->connected = true;
/* get size of the RADOS block device */
r = rbd_stat(rbd->image, &info, sizeof(info));
if (r < 0) {
log_err("rbd_status failed.\n");
goto disconnect;
+ } else if (info.size == 0) {
+ log_err("image size should be larger than zero.\n");
+ r = -EINVAL;
+ goto disconnect;
}
+
dprint(FD_IO, "rbd-engine: image size: %lu\n", info.size);
/* taken from "net" engine. Pretend we deal with files,
/* disconnect, then we were only connected to determine
* the size of the RBD.
*/
- _fio_rbd_disconnect(rbd);
return 0;
disconnect:
static int fio_rbd_invalidate(struct thread_data *td, struct fio_file *f)
{
#if defined(CONFIG_RBD_INVAL)
- struct rbd_data *rbd = td->io_ops->data;
+ struct rbd_data *rbd = td->io_ops_data;
return rbd_invalidate_cache(rbd->image);
#else