* generic io engine that could be used for other projects.
*
*/
-#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <string.h>
#include <dlfcn.h>
+#include <fcntl.h>
#include <assert.h>
#include "fio.h"
#include "diskutil.h"
+#include "zbd.h"
static FLIST_HEAD(engine_list);
-static int check_engine_ops(struct ioengine_ops *ops)
+static bool check_engine_ops(struct ioengine_ops *ops)
{
if (ops->version != FIO_IOOPS_VERSION) {
log_err("bad ioops version %d (want %d)\n", ops->version,
FIO_IOOPS_VERSION);
- return 1;
+ return true;
}
if (!ops->queue) {
log_err("%s: no queue handler\n", ops->name);
- return 1;
+ return true;
}
/*
* sync engines only need a ->queue()
*/
if (ops->flags & FIO_SYNCIO)
- return 0;
+ return false;
- if (!ops->event) {
- log_err("%s: no event handler\n", ops->name);
- return 1;
- }
- if (!ops->getevents) {
- log_err("%s: no getevents handler\n", ops->name);
- return 1;
- }
- if (!ops->queue) {
- log_err("%s: no queue handler\n", ops->name);
- return 1;
+ if (!ops->event || !ops->getevents) {
+ log_err("%s: no event/getevents handler\n", ops->name);
+ return true;
}
- return 0;
+ return false;
}
void unregister_ioengine(struct ioengine_ops *ops)
{
dprint(FD_IO, "ioengine %s unregistered\n", ops->name);
- flist_del(&ops->list);
- INIT_FLIST_HEAD(&ops->list);
+ flist_del_init(&ops->list);
}
void register_ioengine(struct ioengine_ops *ops)
{
dprint(FD_IO, "ioengine %s registered\n", ops->name);
- INIT_FLIST_HEAD(&ops->list);
flist_add_tail(&ops->list, &engine_list);
}
* Unlike the included modules, external engines should have a
* non-static ioengine structure that we can reference.
*/
- ops = dlsym(dlhandle, "ioengine");
+ ops = dlsym(dlhandle, engine_lib);
+ if (!ops)
+ ops = dlsym(dlhandle, "ioengine");
+
+ /*
+ * For some external engines (like C++ ones) it is not that trivial
+ * to provide a non-static ionengine structure that we can reference.
+ * Instead we call a method which allocates the required ioengine
+ * structure.
+ */
+ if (!ops) {
+ get_ioengine_t get_ioengine = dlsym(dlhandle, "get_ioengine");
+
+ if (get_ioengine)
+ get_ioengine(&ops);
+ }
+
if (!ops) {
td_vmsg(td, -1, dlerror(), "dlsym");
dlclose(dlhandle);
return NULL;
}
- ops->dlhandle = dlhandle;
+ td->io_ops_dlhandle = dlhandle;
return ops;
}
-struct ioengine_ops *load_ioengine(struct thread_data *td, const char *name)
+static struct ioengine_ops *__load_ioengine(const char *name)
{
- struct ioengine_ops *ops, *ret;
- char engine[16];
-
- dprint(FD_IO, "load ioengine %s\n", name);
+ char engine[64];
+ engine[sizeof(engine) - 1] = '\0';
strncpy(engine, name, sizeof(engine) - 1);
/*
* linux libaio has alias names, so convert to what we want
*/
- if (!strncmp(engine, "linuxaio", 8) || !strncmp(engine, "aio", 3))
+ if (!strncmp(engine, "linuxaio", 8) || !strncmp(engine, "aio", 3)) {
+ dprint(FD_IO, "converting ioengine name: %s -> libaio\n", name);
strcpy(engine, "libaio");
+ }
+
+ dprint(FD_IO, "load ioengine %s\n", engine);
+ return find_ioengine(engine);
+}
- ops = find_ioengine(engine);
+struct ioengine_ops *load_ioengine(struct thread_data *td)
+{
+ struct ioengine_ops *ops = NULL;
+ const char *name;
+
+ /*
+ * Use ->ioengine_so_path if an external ioengine path is specified.
+ * In this case, ->ioengine is "external" which also means the prefix
+ * for external ioengines "external:" is properly used.
+ */
+ name = td->o.ioengine_so_path ?: td->o.ioengine;
+
+ /*
+ * Try to load ->ioengine first, and if failed try to dlopen(3) either
+ * ->ioengine or ->ioengine_so_path. This is redundant for an external
+ * ioengine with prefix, and also leaves the possibility of unexpected
+ * behavior (e.g. if the "external" ioengine exists), but we do this
+ * so as not to break job files not using the prefix.
+ */
+ ops = __load_ioengine(td->o.ioengine);
if (!ops)
ops = dlopen_ioengine(td, name);
+ /*
+ * If ops is NULL, we failed to load ->ioengine, and also failed to
+ * dlopen(3) either ->ioengine or ->ioengine_so_path as a path.
+ */
if (!ops) {
log_err("fio: engine %s not loadable\n", name);
return NULL;
if (check_engine_ops(ops))
return NULL;
- ret = malloc(sizeof(*ret));
- memcpy(ret, ops, sizeof(*ret));
- ret->data = NULL;
+ return ops;
+}
- return ret;
+/*
+ * For cleaning up an ioengine which never made it to init().
+ */
+void free_ioengine(struct thread_data *td)
+{
+ dprint(FD_IO, "free ioengine %s\n", td->io_ops->name);
+
+ if (td->eo && td->io_ops->options) {
+ options_free(td->io_ops->options, td->eo);
+ free(td->eo);
+ td->eo = NULL;
+ }
+
+ if (td->io_ops_dlhandle) {
+ dlclose(td->io_ops_dlhandle);
+ td->io_ops_dlhandle = NULL;
+ }
+
+ td->io_ops = NULL;
}
void close_ioengine(struct thread_data *td)
if (td->io_ops->cleanup) {
td->io_ops->cleanup(td);
- td->io_ops->data = NULL;
+ td->io_ops_data = NULL;
}
- if (td->io_ops->dlhandle)
- dlclose(td->io_ops->dlhandle);
-
- free(td->io_ops);
- td->io_ops = NULL;
+ free_ioengine(td);
}
int td_io_prep(struct thread_data *td, struct io_u *io_u)
if (td->io_ops->prep) {
int ret = td->io_ops->prep(td, io_u);
- dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
+ dprint(FD_IO, "prep: io_u %p: ret=%d\n", io_u, ret);
+
if (ret)
unlock_file(td, io_u->file);
return ret;
}
int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
- struct timespec *t)
+ const struct timespec *t)
{
int r = 0;
+ /*
+ * For ioengine=rdma one side operation RDMA_WRITE or RDMA_READ,
+ * server side gets a message from the client
+ * side that the task is finished, and
+ * td->done is set to 1 after td_io_commit(). In this case,
+ * there is no need to reap complete event in server side.
+ */
+ if (td->done)
+ return 0;
+
if (min > 0 && td->io_ops->commit) {
r = td->io_ops->commit(td);
if (r < 0)
if (max && td->io_ops->getevents)
r = td->io_ops->getevents(td, min, max, t);
out:
- if (r >= 0)
+ if (r >= 0) {
+ /*
+ * Reflect that our submitted requests were retrieved with
+ * whatever OS async calls are in the underlying engine.
+ */
+ td->io_u_in_flight -= r;
io_u_mark_complete(td, r);
- else
- td_verror(td, -r, "get_events");
+ } else
+ td_verror(td, r, "get_events");
dprint(FD_IO, "getevents: %d\n", r);
return r;
}
-int td_io_queue(struct thread_data *td, struct io_u *io_u)
+enum fio_q_status td_io_queue(struct thread_data *td, struct io_u *io_u)
{
- int ret;
+ const enum fio_ddir ddir = acct_ddir(io_u);
+ unsigned long long buflen = io_u->xfer_buflen;
+ enum fio_q_status ret;
dprint_io_u(io_u, "queue");
fio_ro_check(td, io_u);
assert((io_u->flags & IO_U_F_FLIGHT) == 0);
- io_u->flags |= IO_U_F_FLIGHT;
+ io_u_set(td, io_u, IO_U_F_FLIGHT);
assert(fio_file_open(io_u->file));
+ /*
+ * If using a write iolog, store this entry.
+ */
+ log_io_u(td, io_u);
+
io_u->error = 0;
io_u->resid = 0;
- if (td->io_ops->flags & FIO_SYNCIO) {
+ if (td_ioengine_flagged(td, FIO_SYNCIO)) {
if (fio_fill_issue_time(td))
fio_gettime(&io_u->issue_time, NULL);
*/
if (td->o.read_iolog_file)
memcpy(&td->last_issue, &io_u->issue_time,
- sizeof(struct timeval));
+ sizeof(io_u->issue_time));
}
- if (!ddir_sync(io_u->ddir))
- td->io_issues[io_u->ddir]++;
+ if (ddir_rw(ddir)) {
+ if (!(io_u->flags & IO_U_F_VER_LIST)) {
+ td->io_issues[ddir]++;
+ td->io_issue_bytes[ddir] += buflen;
+ }
+ td->rate_io_issue_bytes[ddir] += buflen;
+ }
ret = td->io_ops->queue(td, io_u);
+ if (ret != FIO_Q_BUSY && io_u->post_submit) {
+ io_u->post_submit(io_u, io_u->error == 0);
+ io_u->post_submit = NULL;
+ }
unlock_file(td, io_u->file);
+ if (ret == FIO_Q_BUSY && ddir_rw(ddir)) {
+ td->io_issues[ddir]--;
+ td->io_issue_bytes[ddir] -= buflen;
+ td->rate_io_issue_bytes[ddir] -= buflen;
+ io_u_clear(td, io_u, IO_U_F_FLIGHT);
+ }
+
+ /*
+ * If an error was seen and the io engine didn't propagate it
+ * back to 'td', do so.
+ */
+ if (io_u->error && !td->error)
+ td_verror(td, io_u->error, "td_io_queue");
+
/*
* Add warning for O_DIRECT so that users have an easier time
* spotting potentially bad alignment. If this triggers for the first
* IO, then it's likely an alignment problem or because the host fs
* does not support O_DIRECT
*/
- if (io_u->error == EINVAL && td->io_issues[io_u->ddir] == 1 &&
+ if (io_u->error == EINVAL && td->io_issues[io_u->ddir & 1] == 1 &&
td->o.odirect) {
+
log_info("fio: first direct IO errored. File system may not "
- "support direct IO, or iomem_align= is bad.\n");
+ "support direct IO, or iomem_align= is bad, or "
+ "invalid block size. Try setting direct=0.\n");
+ }
+
+ if (zbd_unaligned_write(io_u->error) &&
+ td->io_issues[io_u->ddir & 1] == 1 &&
+ td->o.zone_mode != ZONE_MODE_ZBD) {
+ log_info("fio: first I/O failed. If %s is a zoned block device, consider --zonemode=zbd\n",
+ io_u->file->file_name);
}
if (!td->io_ops->commit) {
}
if (ret == FIO_Q_COMPLETED) {
- if (!ddir_sync(io_u->ddir)) {
+ if (ddir_rw(io_u->ddir) || ddir_sync(io_u->ddir)) {
io_u_mark_depth(td, 1);
td->ts.total_io_u[io_u->ddir]++;
}
} else if (ret == FIO_Q_QUEUED) {
- int r;
+ td->io_u_queued++;
- if (!ddir_sync(io_u->ddir)) {
- td->io_u_queued++;
+ if (ddir_rw(io_u->ddir) || ddir_sync(io_u->ddir))
td->ts.total_io_u[io_u->ddir]++;
- }
- if (td->io_u_queued >= td->o.iodepth_batch) {
- r = td_io_commit(td);
- if (r < 0)
- return r;
- }
+ if (td->io_u_queued >= td->o.iodepth_batch)
+ td_io_commit(td);
}
- if ((td->io_ops->flags & FIO_SYNCIO) == 0) {
+ if (!td_ioengine_flagged(td, FIO_SYNCIO)) {
if (fio_fill_issue_time(td))
fio_gettime(&io_u->issue_time, NULL);
*/
if (td->o.read_iolog_file)
memcpy(&td->last_issue, &io_u->issue_time,
- sizeof(struct timeval));
+ sizeof(io_u->issue_time));
}
return ret;
if (td->io_ops->init) {
ret = td->io_ops->init(td);
- if (ret && td->o.iodepth > 1) {
- log_err("fio: io engine init failed. Perhaps try"
- " reducing io depth?\n");
- }
+ if (ret)
+ log_err("fio: io engine %s init failed.%s\n",
+ td->io_ops->name,
+ td->o.iodepth > 1 ?
+ " Perhaps try reducing io depth?" : "");
+ else
+ td->io_ops_init = 1;
+ if (!td->error)
+ td->error = ret;
}
return ret;
}
-int td_io_commit(struct thread_data *td)
+void td_io_commit(struct thread_data *td)
{
int ret;
dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
if (!td->cur_depth || !td->io_u_queued)
- return 0;
+ return;
io_u_mark_depth(td, td->io_u_queued);
- td->io_u_queued = 0;
if (td->io_ops->commit) {
ret = td->io_ops->commit(td);
td_verror(td, -ret, "io commit");
}
- return 0;
+ /*
+ * Reflect that events were submitted as async IO requests.
+ */
+ td->io_u_in_flight += td->io_u_queued;
+ td->io_u_queued = 0;
}
int td_io_open_file(struct thread_data *td, struct fio_file *f)
{
+ if (fio_file_closing(f)) {
+ /*
+ * Open translates to undo closing.
+ */
+ fio_file_clear_closing(f);
+ get_file(f);
+ return 0;
+ }
assert(!fio_file_open(f));
assert(f->fd == -1);
+ assert(td->io_ops->open_file);
if (td->io_ops->open_file(td, f)) {
if (td->error == EINVAL && td->o.odirect)
return 1;
}
- fio_file_reset(f);
+ fio_file_reset(td, f);
fio_file_set_open(f);
fio_file_clear_closing(f);
disk_util_inc(f->du);
}
}
- if (td->io_ops->flags & FIO_DISKLESSIO)
+ if (td_ioengine_flagged(td, FIO_DISKLESSIO))
goto done;
if (td->o.invalidate_cache && file_invalidate_cache(td, f))
goto err;
- if (td->o.fadvise_hint &&
- (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_FILE)) {
+ if (td->o.fadvise_hint != F_ADV_NONE &&
+ (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
int flags;
- if (td_random(td))
+ if (td->o.fadvise_hint == F_ADV_TYPE) {
+ if (td_random(td))
+ flags = POSIX_FADV_RANDOM;
+ else
+ flags = POSIX_FADV_SEQUENTIAL;
+ } else if (td->o.fadvise_hint == F_ADV_RANDOM)
flags = POSIX_FADV_RANDOM;
- else
+ else if (td->o.fadvise_hint == F_ADV_SEQUENTIAL)
flags = POSIX_FADV_SEQUENTIAL;
+ else {
+ log_err("fio: unknown fadvise type %d\n",
+ td->o.fadvise_hint);
+ flags = POSIX_FADV_NORMAL;
+ }
- if (fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
- td_verror(td, errno, "fadvise");
- goto err;
+ if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
+ if (!fio_did_warn(FIO_WARN_FADVISE))
+ log_err("fio: fadvise hint failed\n");
}
}
+#ifdef FIO_HAVE_WRITE_HINT
+ if (fio_option_is_set(&td->o, write_hint) &&
+ (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
+ uint64_t hint = td->o.write_hint;
+ int cmd;
-#ifdef FIO_OS_DIRECTIO
- /*
- * Some OS's have a distinct call to mark the file non-buffered,
- * instead of using O_DIRECT (Solaris)
- */
- if (td->o.odirect) {
- int ret = fio_set_odirect(f->fd);
+ /*
+ * For direct IO, we just need/want to set the hint on
+ * the file descriptor. For buffered IO, we need to set
+ * it on the inode.
+ */
+ if (td->o.odirect)
+ cmd = F_SET_FILE_RW_HINT;
+ else
+ cmd = F_SET_RW_HINT;
- if (ret) {
- td_verror(td, ret, "fio_set_odirect");
+ if (fcntl(f->fd, cmd, &hint) < 0) {
+ td_verror(td, errno, "fcntl write hint");
goto err;
}
}
#endif
+ if (td->o.odirect && !OS_O_DIRECT && fio_set_directio(td, f))
+ goto err;
+
done:
log_file(td, f, FIO_LOG_OPEN_FILE);
return 0;
*/
fio_file_set_closing(f);
- disk_util_dec(f->du);
- unlock_file_all(td, f);
-
return put_file(td, f);
}
+int td_io_unlink_file(struct thread_data *td, struct fio_file *f)
+{
+ if (td->io_ops->unlink_file)
+ return td->io_ops->unlink_file(td, f);
+ else {
+ int ret;
+
+ ret = unlink(f->file_name);
+ if (ret < 0)
+ return errno;
+
+ return 0;
+ }
+}
+
int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
{
if (!td->io_ops->get_file_size)
return td->io_ops->get_file_size(td, f);
}
-static int do_sync_file_range(struct thread_data *td, struct fio_file *f)
+int fio_show_ioengine_help(const char *engine)
{
- off64_t offset, nbytes;
-
- offset = f->first_write;
- nbytes = f->last_write - f->first_write;
-
- if (!nbytes)
+ struct flist_head *entry;
+ struct thread_data td;
+ struct ioengine_ops *io_ops;
+ char *sep;
+ int ret = 1;
+
+ if (!engine || !*engine) {
+ log_info("Available IO engines:\n");
+ flist_for_each(entry, &engine_list) {
+ io_ops = flist_entry(entry, struct ioengine_ops, list);
+ log_info("\t%s\n", io_ops->name);
+ }
return 0;
+ }
+ sep = strchr(engine, ',');
+ if (sep) {
+ *sep = 0;
+ sep++;
+ }
- return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
-}
-
-int do_io_u_sync(struct thread_data *td, struct io_u *io_u)
-{
- int ret;
+ memset(&td, 0, sizeof(struct thread_data));
+ td.o.ioengine = (char *)engine;
+ io_ops = load_ioengine(&td);
- if (io_u->ddir == DDIR_SYNC) {
- ret = fsync(io_u->file->fd);
- } else if (io_u->ddir == DDIR_DATASYNC) {
-#ifdef FIO_HAVE_FDATASYNC
- ret = fdatasync(io_u->file->fd);
-#else
- ret = io_u->xfer_buflen;
- io_u->error = EINVAL;
-#endif
- } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
- ret = do_sync_file_range(td, io_u->file);
- else {
- ret = io_u->xfer_buflen;
- io_u->error = EINVAL;
+ if (!io_ops) {
+ log_info("IO engine %s not found\n", engine);
+ return 1;
}
- if (ret < 0)
- io_u->error = errno;
+ if (io_ops->options)
+ ret = show_cmd_help(io_ops->options, sep);
+ else
+ log_info("IO engine %s has no options\n", io_ops->name);
+ free_ioengine(&td);
return ret;
}