#include <unistd.h>
#include <string.h>
#include <dlfcn.h>
+#include <fcntl.h>
+#include <assert.h>
#include "fio.h"
-#include "os.h"
+#include "diskutil.h"
+
+static FLIST_HEAD(engine_list);
static int 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;
+ }
+
+ if (!ops->queue) {
+ log_err("%s: no queue handler\n", ops->name);
+ return 1;
+ }
+
/*
- * cpu thread doesn't need to provide anything
+ * sync engines only need a ->queue()
*/
- if (ops->flags & FIO_CPUIO)
+ if (ops->flags & FIO_SYNCIO)
return 0;
if (!ops->event) {
- log_err("%s: no event handler)\n", ops->name);
+ log_err("%s: no event handler\n", ops->name);
return 1;
}
if (!ops->getevents) {
- log_err("%s: no getevents handler)\n", ops->name);
+ 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;
- }
-
+
return 0;
}
-struct ioengine_ops *load_ioengine(struct thread_data *td, char *name)
+void unregister_ioengine(struct ioengine_ops *ops)
{
- char engine[16], engine_lib[256];
- struct ioengine_ops *ops, *ret;
- void *dlhandle;
+ dprint(FD_IO, "ioengine %s unregistered\n", ops->name);
+ flist_del(&ops->list);
+ INIT_FLIST_HEAD(&ops->list);
+}
- strcpy(engine, name);
+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);
+}
- /*
- * linux libaio has alias names, so convert to what we want
- */
- if (!strncmp(engine, "linuxaio", 8) || !strncmp(engine, "aio", 3))
- strcpy(engine, "libaio");
+static struct ioengine_ops *find_ioengine(const char *name)
+{
+ struct ioengine_ops *ops;
+ struct flist_head *entry;
+
+ flist_for_each(entry, &engine_list) {
+ ops = flist_entry(entry, struct ioengine_ops, list);
+ if (!strcmp(name, ops->name))
+ return ops;
+ }
+
+ return NULL;
+}
+
+static struct ioengine_ops *dlopen_ioengine(struct thread_data *td,
+ const char *engine_lib)
+{
+ struct ioengine_ops *ops;
+ void *dlhandle;
+
+ dprint(FD_IO, "dload engine %s\n", engine_lib);
- sprintf(engine_lib, "%s/lib/fio/fio-engine-%s.o", fio_inst_prefix, engine);
dlerror();
dlhandle = dlopen(engine_lib, RTLD_LAZY);
if (!dlhandle) {
- td_vmsg(td, -1, dlerror());
+ td_vmsg(td, -1, dlerror(), "dlopen");
return NULL;
}
- ops = dlsym(dlhandle, "ioengine");
+ /*
+ * Unlike the included modules, external engines should have a
+ * non-static ioengine structure that we can reference.
+ */
+ 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());
+ td_vmsg(td, -1, dlerror(), "dlsym");
dlclose(dlhandle);
return NULL;
}
- if (ops->version != FIO_IOOPS_VERSION) {
- log_err("bad ioops version %d (want %d)\n", ops->version, FIO_IOOPS_VERSION);
- dlclose(dlhandle);
+ ops->dlhandle = dlhandle;
+ return ops;
+}
+
+struct ioengine_ops *load_ioengine(struct thread_data *td, const char *name)
+{
+ struct ioengine_ops *ops, *ret;
+ char engine[16];
+
+ dprint(FD_IO, "load ioengine %s\n", name);
+
+ 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))
+ strcpy(engine, "libaio");
+
+ ops = find_ioengine(engine);
+ if (!ops)
+ ops = dlopen_ioengine(td, name);
+
+ if (!ops) {
+ log_err("fio: engine %s not loadable\n", name);
return NULL;
}
/*
* Check that the required methods are there.
*/
- if (check_engine_ops(ops)) {
- dlclose(dlhandle);
+ if (check_engine_ops(ops))
return NULL;
- }
ret = malloc(sizeof(*ret));
memcpy(ret, ops, sizeof(*ret));
ret->data = NULL;
- ret->dlhandle = dlhandle;
return ret;
}
-void close_ioengine(struct thread_data *td)
+/*
+ * For cleaning up an ioengine which never made it to init().
+ */
+void free_ioengine(struct thread_data *td)
{
- if (td->io_ops->cleanup)
- td->io_ops->cleanup(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);
- dlclose(td->io_ops->dlhandle);
free(td->io_ops);
td->io_ops = NULL;
}
+void close_ioengine(struct thread_data *td)
+{
+ dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
+
+ if (td->io_ops->cleanup) {
+ td->io_ops->cleanup(td);
+ td->io_ops->data = NULL;
+ }
+
+ free_ioengine(td);
+}
+
int td_io_prep(struct thread_data *td, struct io_u *io_u)
{
- if (td->io_ops->prep && td->io_ops->prep(td, io_u))
- return 1;
+ dprint_io_u(io_u, "prep");
+ fio_ro_check(td, io_u);
+
+ lock_file(td, io_u->file, io_u->ddir);
+
+ if (td->io_ops->prep) {
+ int ret = td->io_ops->prep(td, io_u);
+
+ dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
+ if (ret)
+ unlock_file(td, io_u->file);
+ return ret;
+ }
return 0;
}
-int td_io_sync(struct thread_data *td, struct fio_file *f)
+int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
+ 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)
+ goto out;
+ }
+ if (max > td->cur_depth)
+ max = td->cur_depth;
+ if (min > max)
+ max = min;
+
+ r = 0;
+ if (max && td->io_ops->getevents)
+ r = td->io_ops->getevents(td, min, max, t);
+out:
+ 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");
+
+ dprint(FD_IO, "getevents: %d\n", r);
+ return r;
+}
+
+int td_io_queue(struct thread_data *td, struct io_u *io_u)
{
- struct io_u *io_u = __get_io_u(td);
- struct io_completion_data icd;
+ const enum fio_ddir ddir = acct_ddir(io_u);
+ unsigned long buflen = io_u->xfer_buflen;
int ret;
- if (!io_u)
- return 1;
+ dprint_io_u(io_u, "queue");
+ fio_ro_check(td, io_u);
- io_u->ddir = DDIR_SYNC;
- io_u->file = f;
+ assert((io_u->flags & IO_U_F_FLIGHT) == 0);
+ io_u_set(io_u, IO_U_F_FLIGHT);
- if (td_io_prep(td, io_u)) {
- put_io_u(td, io_u);
- return 1;
+ 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 (fio_fill_issue_time(td))
+ fio_gettime(&io_u->issue_time, NULL);
+
+ /*
+ * only used for iolog
+ */
+ if (td->o.read_iolog_file)
+ memcpy(&td->last_issue, &io_u->issue_time,
+ sizeof(struct timeval));
}
- ret = td_io_queue(td, io_u);
- if (ret) {
- put_io_u(td, io_u);
- td_verror(td, ret);
- return 1;
+ if (ddir_rw(ddir)) {
+ td->io_issues[ddir]++;
+ td->io_issue_bytes[ddir] += buflen;
+ td->rate_io_issue_bytes[ddir] += buflen;
}
- ret = td_io_getevents(td, 1, td->cur_depth, NULL);
- if (ret < 0) {
- td_verror(td, -ret);
- return 1;
+ ret = td->io_ops->queue(td, io_u);
+
+ 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;
}
- icd.nr = ret;
- ios_completed(td, &icd);
- if (icd.error) {
- td_verror(td, icd.error);
+ /*
+ * 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] == 1 &&
+ td->o.odirect) {
+
+ log_info("fio: first direct IO errored. File system may not "
+ "support direct IO, or iomem_align= is bad. Try "
+ "setting direct=0.\n");
+ }
+
+ if (!td->io_ops->commit || io_u->ddir == DDIR_TRIM) {
+ io_u_mark_submit(td, 1);
+ io_u_mark_complete(td, 1);
+ }
+
+ if (ret == FIO_Q_COMPLETED) {
+ if (ddir_rw(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_rw(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_ops->flags & FIO_SYNCIO) == 0) {
+ if (fio_fill_issue_time(td))
+ fio_gettime(&io_u->issue_time, NULL);
+
+ /*
+ * only used for iolog
+ */
+ if (td->o.read_iolog_file)
+ memcpy(&td->last_issue, &io_u->issue_time,
+ sizeof(struct timeval));
+ }
+
+ return ret;
+}
+
+int td_io_init(struct thread_data *td)
+{
+ int ret = 0;
+
+ 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 (!td->error)
+ td->error = ret;
+ }
+
+ if (!ret && (td->io_ops->flags & FIO_NOIO))
+ td->flags |= TD_F_NOIO;
+
+ return ret;
+}
+
+int 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;
+
+ io_u_mark_depth(td, td->io_u_queued);
+
+ if (td->io_ops->commit) {
+ ret = td->io_ops->commit(td);
+ if (ret)
+ td_verror(td, -ret, "io commit");
+ }
+
+ /*
+ * Reflect that events were submitted as async IO requests.
+ */
+ td->io_u_in_flight += td->io_u_queued;
+ td->io_u_queued = 0;
+
+ return 0;
+}
+
+int td_io_open_file(struct thread_data *td, struct fio_file *f)
+{
+ assert(!fio_file_open(f));
+ assert(f->fd == -1);
+
+ if (td->io_ops->open_file(td, f)) {
+ if (td->error == EINVAL && td->o.odirect)
+ log_err("fio: destination does not support O_DIRECT\n");
+ if (td->error == EMFILE) {
+ log_err("fio: try reducing/setting openfiles (failed"
+ " at %u of %u)\n", td->nr_open_files,
+ td->o.nr_files);
+ }
+
+ assert(f->fd == -1);
+ assert(!fio_file_open(f));
return 1;
}
+ fio_file_reset(td, f);
+ fio_file_set_open(f);
+ fio_file_clear_closing(f);
+ disk_util_inc(f->du);
+
+ td->nr_open_files++;
+ get_file(f);
+
+ if (f->filetype == FIO_TYPE_PIPE) {
+ if (td_random(td)) {
+ log_err("fio: can't seek on pipes (no random io)\n");
+ goto err;
+ }
+ }
+
+ if (td->io_ops->flags & 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)) {
+ int flags;
+
+ if (td_random(td))
+ flags = POSIX_FADV_RANDOM;
+ else
+ flags = POSIX_FADV_SEQUENTIAL;
+
+ if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
+ td_verror(td, errno, "fadvise");
+ goto err;
+ }
+ }
+#ifdef FIO_HAVE_STREAMID
+ if (td->o.fadvise_stream &&
+ (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_FILE)) {
+ off_t stream = td->o.fadvise_stream;
+
+ if (posix_fadvise(f->fd, stream, f->io_size, POSIX_FADV_STREAMID) < 0) {
+ td_verror(td, errno, "fadvise streamid");
+ goto err;
+ }
+ }
+#endif
+
+#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);
+
+ if (ret) {
+ td_verror(td, ret, "fio_set_odirect");
+ log_err("fio: the file system does not seem to support direct IO\n");
+ goto err;
+ }
+ }
+#endif
+
+done:
+ log_file(td, f, FIO_LOG_OPEN_FILE);
return 0;
+err:
+ disk_util_dec(f->du);
+ if (td->io_ops->close_file)
+ td->io_ops->close_file(td, f);
+ return 1;
}
-int td_io_getevents(struct thread_data *td, int min, int max,
- struct timespec *t)
+int td_io_close_file(struct thread_data *td, struct fio_file *f)
{
- return td->io_ops->getevents(td, min, max, t);
+ if (!fio_file_closing(f))
+ log_file(td, f, FIO_LOG_CLOSE_FILE);
+
+ /*
+ * mark as closing, do real close when last io on it has completed
+ */
+ fio_file_set_closing(f);
+
+ disk_util_dec(f->du);
+
+ if (td->o.file_lock_mode != FILE_LOCK_NONE)
+ unlock_file_all(td, f);
+
+ return put_file(td, f);
}
-int td_io_queue(struct thread_data *td, struct io_u *io_u)
+int td_io_unlink_file(struct thread_data *td, struct fio_file *f)
{
- gettimeofday(&io_u->issue_time, NULL);
+ if (td->io_ops->unlink_file)
+ return td->io_ops->unlink_file(td, f);
+ else
+ return unlink(f->file_name);
+}
+
+int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
+{
+ if (!td->io_ops->get_file_size)
+ return 0;
- return td->io_ops->queue(td, io_u);
+ return td->io_ops->get_file_size(td, f);
}
-int td_io_init(struct thread_data *td)
+static int do_sync_file_range(const struct thread_data *td,
+ struct fio_file *f)
{
- if (td->io_ops->init)
- return td->io_ops->init(td);
+ off64_t offset, nbytes;
+
+ offset = f->first_write;
+ nbytes = f->last_write - f->first_write;
+
+ if (!nbytes)
+ return 0;
+
+ return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
+}
+
+int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
+{
+ int ret;
+ if (io_u->ddir == DDIR_SYNC) {
+ ret = fsync(io_u->file->fd);
+ } else if (io_u->ddir == DDIR_DATASYNC) {
+#ifdef CONFIG_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 (ret < 0)
+ io_u->error = errno;
+
+ return ret;
+}
+
+int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
+{
+#ifndef FIO_HAVE_TRIM
+ io_u->error = EINVAL;
return 0;
+#else
+ struct fio_file *f = io_u->file;
+ int ret;
+
+ ret = os_trim(f->fd, io_u->offset, io_u->xfer_buflen);
+ if (!ret)
+ return io_u->xfer_buflen;
+
+ io_u->error = ret;
+ return 0;
+#endif
+}
+
+int fio_show_ioengine_help(const char *engine)
+{
+ struct flist_head *entry;
+ struct thread_data td;
+ char *sep;
+ int ret = 1;
+
+ if (!engine || !*engine) {
+ log_info("Available IO engines:\n");
+ flist_for_each(entry, &engine_list) {
+ td.io_ops = flist_entry(entry, struct ioengine_ops,
+ list);
+ log_info("\t%s\n", td.io_ops->name);
+ }
+ return 0;
+ }
+ sep = strchr(engine, ',');
+ if (sep) {
+ *sep = 0;
+ sep++;
+ }
+
+ memset(&td, 0, sizeof(td));
+
+ td.io_ops = load_ioengine(&td, engine);
+ if (!td.io_ops) {
+ log_info("IO engine %s not found\n", engine);
+ return 1;
+ }
+
+ if (td.io_ops->options)
+ ret = show_cmd_help(td.io_ops->options, sep);
+ else
+ log_info("IO engine %s has no options\n", td.io_ops->name);
+
+ free_ioengine(&td);
+
+ return ret;
}