ioengines.c revision cb849a795d2bfd99efdd66f5a3e475c9fb3edd86 
1/*
2 * The io parts of the fio tool, includes workers for sync and mmap'ed
3 * io, as well as both posix and linux libaio support.
4 *
5 * sync io is implemented on top of aio.
6 *
7 * This is not really specific to fio, if the get_io_u/put_io_u and
8 * structures was pulled into this as well it would be a perfectly
9 * generic io engine that could be used for other projects.
10 *
11 */
12#include <stdio.h>
13#include <stdlib.h>
14#include <unistd.h>
15#include <string.h>
16#include <dlfcn.h>
17#include <assert.h>
18
19#include "fio.h"
20#include "diskutil.h"
21
22static FLIST_HEAD(engine_list);
23
24static int check_engine_ops(struct ioengine_ops *ops)
25{
26	if (ops->version != FIO_IOOPS_VERSION) {
27		log_err("bad ioops version %d (want %d)\n", ops->version,
28							FIO_IOOPS_VERSION);
29		return 1;
30	}
31
32	if (!ops->queue) {
33		log_err("%s: no queue handler\n", ops->name);
34		return 1;
35	}
36
37	/*
38	 * sync engines only need a ->queue()
39	 */
40	if (ops->flags & FIO_SYNCIO)
41		return 0;
42
43	if (!ops->event) {
44		log_err("%s: no event handler\n", ops->name);
45		return 1;
46	}
47	if (!ops->getevents) {
48		log_err("%s: no getevents handler\n", ops->name);
49		return 1;
50	}
51	if (!ops->queue) {
52		log_err("%s: no queue handler\n", ops->name);
53		return 1;
54	}
55
56	return 0;
57}
58
59void unregister_ioengine(struct ioengine_ops *ops)
60{
61	dprint(FD_IO, "ioengine %s unregistered\n", ops->name);
62	flist_del(&ops->list);
63	INIT_FLIST_HEAD(&ops->list);
64}
65
66void register_ioengine(struct ioengine_ops *ops)
67{
68	dprint(FD_IO, "ioengine %s registered\n", ops->name);
69	INIT_FLIST_HEAD(&ops->list);
70	flist_add_tail(&ops->list, &engine_list);
71}
72
73static struct ioengine_ops *find_ioengine(const char *name)
74{
75	struct ioengine_ops *ops;
76	struct flist_head *entry;
77
78	flist_for_each(entry, &engine_list) {
79		ops = flist_entry(entry, struct ioengine_ops, list);
80		if (!strcmp(name, ops->name))
81			return ops;
82	}
83
84	return NULL;
85}
86
87static struct ioengine_ops *dlopen_ioengine(struct thread_data *td,
88					    const char *engine_lib)
89{
90	struct ioengine_ops *ops;
91	void *dlhandle;
92
93	dprint(FD_IO, "dload engine %s\n", engine_lib);
94
95	dlerror();
96	dlhandle = dlopen(engine_lib, RTLD_LAZY);
97	if (!dlhandle) {
98		td_vmsg(td, -1, dlerror(), "dlopen");
99		return NULL;
100	}
101
102	/*
103	 * Unlike the included modules, external engines should have a
104	 * non-static ioengine structure that we can reference.
105	 */
106	ops = dlsym(dlhandle, "ioengine");
107	if (!ops) {
108		td_vmsg(td, -1, dlerror(), "dlsym");
109		dlclose(dlhandle);
110		return NULL;
111	}
112
113	ops->dlhandle = dlhandle;
114	return ops;
115}
116
117struct ioengine_ops *load_ioengine(struct thread_data *td, const char *name)
118{
119	struct ioengine_ops *ops, *ret;
120	char engine[16];
121
122	dprint(FD_IO, "load ioengine %s\n", name);
123
124	strncpy(engine, name, sizeof(engine) - 1);
125
126	/*
127	 * linux libaio has alias names, so convert to what we want
128	 */
129	if (!strncmp(engine, "linuxaio", 8) || !strncmp(engine, "aio", 3))
130		strcpy(engine, "libaio");
131
132	ops = find_ioengine(engine);
133	if (!ops)
134		ops = dlopen_ioengine(td, name);
135
136	if (!ops) {
137		log_err("fio: engine %s not loadable\n", name);
138		return NULL;
139	}
140
141	/*
142	 * Check that the required methods are there.
143	 */
144	if (check_engine_ops(ops))
145		return NULL;
146
147	ret = malloc(sizeof(*ret));
148	memcpy(ret, ops, sizeof(*ret));
149	ret->data = NULL;
150
151	return ret;
152}
153
154void close_ioengine(struct thread_data *td)
155{
156	dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
157
158	if (td->io_ops->cleanup) {
159		td->io_ops->cleanup(td);
160		td->io_ops->data = NULL;
161	}
162
163	if (td->io_ops->dlhandle)
164		dlclose(td->io_ops->dlhandle);
165
166	free(td->io_ops);
167	td->io_ops = NULL;
168}
169
170int td_io_prep(struct thread_data *td, struct io_u *io_u)
171{
172	dprint_io_u(io_u, "prep");
173	fio_ro_check(td, io_u);
174
175	lock_file(td, io_u->file, io_u->ddir);
176
177	if (td->io_ops->prep) {
178		int ret = td->io_ops->prep(td, io_u);
179
180		dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
181		if (ret)
182			unlock_file(td, io_u->file);
183		return ret;
184	}
185
186	return 0;
187}
188
189int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
190		    struct timespec *t)
191{
192	int r = 0;
193
194	if (min > 0 && td->io_ops->commit) {
195		r = td->io_ops->commit(td);
196		if (r < 0)
197			goto out;
198	}
199	if (max > td->cur_depth)
200		max = td->cur_depth;
201	if (min > max)
202		max = min;
203
204	r = 0;
205	if (max && td->io_ops->getevents)
206		r = td->io_ops->getevents(td, min, max, t);
207out:
208	if (r >= 0)
209		io_u_mark_complete(td, r);
210	dprint(FD_IO, "getevents: %d\n", r);
211	return r;
212}
213
214int td_io_queue(struct thread_data *td, struct io_u *io_u)
215{
216	int ret;
217
218	dprint_io_u(io_u, "queue");
219	fio_ro_check(td, io_u);
220
221	assert((io_u->flags & IO_U_F_FLIGHT) == 0);
222	io_u->flags |= IO_U_F_FLIGHT;
223
224	assert(fio_file_open(io_u->file));
225
226	io_u->error = 0;
227	io_u->resid = 0;
228
229	if (td->io_ops->flags & FIO_SYNCIO) {
230		if (fio_fill_issue_time(td))
231			fio_gettime(&io_u->issue_time, NULL);
232
233		/*
234		 * only used for iolog
235		 */
236		if (td->o.read_iolog_file)
237			memcpy(&td->last_issue, &io_u->issue_time,
238					sizeof(struct timeval));
239	}
240
241	if (!ddir_sync(io_u->ddir))
242		td->io_issues[io_u->ddir]++;
243
244	ret = td->io_ops->queue(td, io_u);
245
246	unlock_file(td, io_u->file);
247
248	/*
249	 * Add warning for O_DIRECT so that users have an easier time
250	 * spotting potentially bad alignment. If this triggers for the first
251	 * IO, then it's likely an alignment problem or because the host fs
252	 * does not support O_DIRECT
253	 */
254	if (io_u->error == EINVAL && td->io_issues[io_u->ddir] == 1 &&
255	    td->o.odirect) {
256		log_info("fio: first direct IO errored. File system may not "
257			 "support direct IO, or iomem_align= is bad.\n");
258	}
259
260	if (!td->io_ops->commit) {
261		io_u_mark_submit(td, 1);
262		io_u_mark_complete(td, 1);
263	}
264
265	if (ret == FIO_Q_COMPLETED) {
266		if (!ddir_sync(io_u->ddir)) {
267			io_u_mark_depth(td, 1);
268			td->ts.total_io_u[io_u->ddir]++;
269		}
270	} else if (ret == FIO_Q_QUEUED) {
271		int r;
272
273		if (!ddir_sync(io_u->ddir)) {
274			td->io_u_queued++;
275			td->ts.total_io_u[io_u->ddir]++;
276		}
277
278		if (td->io_u_queued >= td->o.iodepth_batch) {
279			r = td_io_commit(td);
280			if (r < 0)
281				return r;
282		}
283	}
284
285	if ((td->io_ops->flags & FIO_SYNCIO) == 0) {
286		if (fio_fill_issue_time(td))
287			fio_gettime(&io_u->issue_time, NULL);
288
289		/*
290		 * only used for iolog
291		 */
292		if (td->o.read_iolog_file)
293			memcpy(&td->last_issue, &io_u->issue_time,
294					sizeof(struct timeval));
295	}
296
297	return ret;
298}
299
300int td_io_init(struct thread_data *td)
301{
302	int ret = 0;
303
304	if (td->io_ops->init) {
305		ret = td->io_ops->init(td);
306		if (ret && td->o.iodepth > 1) {
307			log_err("fio: io engine init failed. Perhaps try"
308				" reducing io depth?\n");
309		}
310	}
311
312	return ret;
313}
314
315int td_io_commit(struct thread_data *td)
316{
317	dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
318
319	if (!td->cur_depth || !td->io_u_queued)
320		return 0;
321
322	io_u_mark_depth(td, td->io_u_queued);
323	td->io_u_queued = 0;
324
325	if (td->io_ops->commit)
326		return td->io_ops->commit(td);
327
328	return 0;
329}
330
331int td_io_open_file(struct thread_data *td, struct fio_file *f)
332{
333	assert(!fio_file_open(f));
334	assert(f->fd == -1);
335
336	if (td->io_ops->open_file(td, f)) {
337		if (td->error == EINVAL && td->o.odirect)
338			log_err("fio: destination does not support O_DIRECT\n");
339		if (td->error == EMFILE) {
340			log_err("fio: try reducing/setting openfiles (failed"
341				" at %u of %u)\n", td->nr_open_files,
342							td->o.nr_files);
343		}
344
345		assert(f->fd == -1);
346		assert(!fio_file_open(f));
347		return 1;
348	}
349
350	fio_file_reset(f);
351	fio_file_set_open(f);
352	fio_file_clear_closing(f);
353	disk_util_inc(f->du);
354
355	td->nr_open_files++;
356	get_file(f);
357
358	if (f->filetype == FIO_TYPE_PIPE) {
359		if (td_random(td)) {
360			log_err("fio: can't seek on pipes (no random io)\n");
361			goto err;
362		}
363	}
364
365	if (td->io_ops->flags & FIO_DISKLESSIO)
366		goto done;
367
368	if (td->o.invalidate_cache && file_invalidate_cache(td, f))
369		goto err;
370
371	if (td->o.fadvise_hint &&
372	    (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_FILE)) {
373		int flags;
374
375		if (td_random(td))
376			flags = POSIX_FADV_RANDOM;
377		else
378			flags = POSIX_FADV_SEQUENTIAL;
379
380		if (fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
381			td_verror(td, errno, "fadvise");
382			goto err;
383		}
384	}
385
386#ifdef FIO_OS_DIRECTIO
387	/*
388	 * Some OS's have a distinct call to mark the file non-buffered,
389	 * instead of using O_DIRECT (Solaris)
390	 */
391	if (td->o.odirect) {
392		int ret = fio_set_odirect(f->fd);
393
394		if (ret) {
395			td_verror(td, ret, "fio_set_odirect");
396			goto err;
397		}
398	}
399#endif
400
401done:
402	log_file(td, f, FIO_LOG_OPEN_FILE);
403	return 0;
404err:
405	disk_util_dec(f->du);
406	if (td->io_ops->close_file)
407		td->io_ops->close_file(td, f);
408	return 1;
409}
410
411int td_io_close_file(struct thread_data *td, struct fio_file *f)
412{
413	if (!fio_file_closing(f))
414		log_file(td, f, FIO_LOG_CLOSE_FILE);
415
416	/*
417	 * mark as closing, do real close when last io on it has completed
418	 */
419	fio_file_set_closing(f);
420
421	disk_util_dec(f->du);
422	unlock_file_all(td, f);
423
424	return put_file(td, f);
425}
426
427int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
428{
429	if (!td->io_ops->get_file_size)
430		return 0;
431
432	return td->io_ops->get_file_size(td, f);
433}
434
435static int do_sync_file_range(struct thread_data *td, struct fio_file *f)
436{
437	off64_t offset, nbytes;
438
439	offset = f->first_write;
440	nbytes = f->last_write - f->first_write;
441
442	if (!nbytes)
443		return 0;
444
445	return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
446}
447
448int do_io_u_sync(struct thread_data *td, struct io_u *io_u)
449{
450	int ret;
451
452	if (io_u->ddir == DDIR_SYNC) {
453		ret = fsync(io_u->file->fd);
454	} else if (io_u->ddir == DDIR_DATASYNC) {
455#ifdef FIO_HAVE_FDATASYNC
456		ret = fdatasync(io_u->file->fd);
457#else
458		ret = io_u->xfer_buflen;
459		io_u->error = EINVAL;
460#endif
461	} else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
462		ret = do_sync_file_range(td, io_u->file);
463	else {
464		ret = io_u->xfer_buflen;
465		io_u->error = EINVAL;
466	}
467
468	if (ret < 0)
469		io_u->error = errno;
470
471	return ret;
472}
473