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 <fcntl.h>
18#include <assert.h>
19
20#include "fio.h"
21#include "diskutil.h"
22
23static FLIST_HEAD(engine_list);
24
25static int check_engine_ops(struct ioengine_ops *ops)
26{
27	if (ops->version != FIO_IOOPS_VERSION) {
28		log_err("bad ioops version %d (want %d)\n", ops->version,
29							FIO_IOOPS_VERSION);
30		return 1;
31	}
32
33	if (!ops->queue) {
34		log_err("%s: no queue handler\n", ops->name);
35		return 1;
36	}
37
38	/*
39	 * sync engines only need a ->queue()
40	 */
41	if (ops->flags & FIO_SYNCIO)
42		return 0;
43
44	if (!ops->event) {
45		log_err("%s: no event handler\n", ops->name);
46		return 1;
47	}
48	if (!ops->getevents) {
49		log_err("%s: no getevents handler\n", ops->name);
50		return 1;
51	}
52	if (!ops->queue) {
53		log_err("%s: no queue handler\n", ops->name);
54		return 1;
55	}
56
57	return 0;
58}
59
60void unregister_ioengine(struct ioengine_ops *ops)
61{
62	dprint(FD_IO, "ioengine %s unregistered\n", ops->name);
63	flist_del(&ops->list);
64	INIT_FLIST_HEAD(&ops->list);
65}
66
67void register_ioengine(struct ioengine_ops *ops)
68{
69	dprint(FD_IO, "ioengine %s registered\n", ops->name);
70	INIT_FLIST_HEAD(&ops->list);
71	flist_add_tail(&ops->list, &engine_list);
72}
73
74static struct ioengine_ops *find_ioengine(const char *name)
75{
76	struct ioengine_ops *ops;
77	struct flist_head *entry;
78
79	flist_for_each(entry, &engine_list) {
80		ops = flist_entry(entry, struct ioengine_ops, list);
81		if (!strcmp(name, ops->name))
82			return ops;
83	}
84
85	return NULL;
86}
87
88static struct ioengine_ops *dlopen_ioengine(struct thread_data *td,
89					    const char *engine_lib)
90{
91	struct ioengine_ops *ops;
92	void *dlhandle;
93
94	dprint(FD_IO, "dload engine %s\n", engine_lib);
95
96	dlerror();
97	dlhandle = dlopen(engine_lib, RTLD_LAZY);
98	if (!dlhandle) {
99		td_vmsg(td, -1, dlerror(), "dlopen");
100		return NULL;
101	}
102
103	/*
104	 * Unlike the included modules, external engines should have a
105	 * non-static ioengine structure that we can reference.
106	 */
107	ops = dlsym(dlhandle, engine_lib);
108	if (!ops)
109		ops = dlsym(dlhandle, "ioengine");
110
111	/*
112	 * For some external engines (like C++ ones) it is not that trivial
113	 * to provide a non-static ionengine structure that we can reference.
114	 * Instead we call a method which allocates the required ioengine
115	 * structure.
116	 */
117	if (!ops) {
118		get_ioengine_t get_ioengine = dlsym(dlhandle, "get_ioengine");
119
120		if (get_ioengine)
121			get_ioengine(&ops);
122	}
123
124	if (!ops) {
125		td_vmsg(td, -1, dlerror(), "dlsym");
126		dlclose(dlhandle);
127		return NULL;
128	}
129
130	ops->dlhandle = dlhandle;
131	return ops;
132}
133
134struct ioengine_ops *load_ioengine(struct thread_data *td, const char *name)
135{
136	struct ioengine_ops *ops, *ret;
137	char engine[16];
138
139	dprint(FD_IO, "load ioengine %s\n", name);
140
141	strncpy(engine, name, sizeof(engine) - 1);
142
143	/*
144	 * linux libaio has alias names, so convert to what we want
145	 */
146	if (!strncmp(engine, "linuxaio", 8) || !strncmp(engine, "aio", 3))
147		strcpy(engine, "libaio");
148
149	ops = find_ioengine(engine);
150	if (!ops)
151		ops = dlopen_ioengine(td, name);
152
153	if (!ops) {
154		log_err("fio: engine %s not loadable\n", name);
155		return NULL;
156	}
157
158	/*
159	 * Check that the required methods are there.
160	 */
161	if (check_engine_ops(ops))
162		return NULL;
163
164	ret = malloc(sizeof(*ret));
165	memcpy(ret, ops, sizeof(*ret));
166	ret->data = NULL;
167
168	return ret;
169}
170
171/*
172 * For cleaning up an ioengine which never made it to init().
173 */
174void free_ioengine(struct thread_data *td)
175{
176	dprint(FD_IO, "free ioengine %s\n", td->io_ops->name);
177
178	if (td->eo && td->io_ops->options) {
179		options_free(td->io_ops->options, td->eo);
180		free(td->eo);
181		td->eo = NULL;
182	}
183
184	if (td->io_ops->dlhandle)
185		dlclose(td->io_ops->dlhandle);
186
187	free(td->io_ops);
188	td->io_ops = NULL;
189}
190
191void close_ioengine(struct thread_data *td)
192{
193	dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
194
195	if (td->io_ops->cleanup) {
196		td->io_ops->cleanup(td);
197		td->io_ops->data = NULL;
198	}
199
200	free_ioengine(td);
201}
202
203int td_io_prep(struct thread_data *td, struct io_u *io_u)
204{
205	dprint_io_u(io_u, "prep");
206	fio_ro_check(td, io_u);
207
208	lock_file(td, io_u->file, io_u->ddir);
209
210	if (td->io_ops->prep) {
211		int ret = td->io_ops->prep(td, io_u);
212
213		dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
214		if (ret)
215			unlock_file(td, io_u->file);
216		return ret;
217	}
218
219	return 0;
220}
221
222int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
223		    struct timespec *t)
224{
225	int r = 0;
226
227	/*
228	 * For ioengine=rdma one side operation RDMA_WRITE or RDMA_READ,
229	 * server side gets a message from the client
230	 * side that the task is finished, and
231	 * td->done is set to 1 after td_io_commit(). In this case,
232	 * there is no need to reap complete event in server side.
233	 */
234	if (td->done)
235		return 0;
236
237	if (min > 0 && td->io_ops->commit) {
238		r = td->io_ops->commit(td);
239		if (r < 0)
240			goto out;
241	}
242	if (max > td->cur_depth)
243		max = td->cur_depth;
244	if (min > max)
245		max = min;
246
247	r = 0;
248	if (max && td->io_ops->getevents)
249		r = td->io_ops->getevents(td, min, max, t);
250out:
251	if (r >= 0) {
252		/*
253		 * Reflect that our submitted requests were retrieved with
254		 * whatever OS async calls are in the underlying engine.
255		 */
256		td->io_u_in_flight -= r;
257		io_u_mark_complete(td, r);
258	} else
259		td_verror(td, r, "get_events");
260
261	dprint(FD_IO, "getevents: %d\n", r);
262	return r;
263}
264
265int td_io_queue(struct thread_data *td, struct io_u *io_u)
266{
267	int ret;
268
269	dprint_io_u(io_u, "queue");
270	fio_ro_check(td, io_u);
271
272	assert((io_u->flags & IO_U_F_FLIGHT) == 0);
273	io_u->flags |= IO_U_F_FLIGHT;
274
275	assert(fio_file_open(io_u->file));
276
277	/*
278	 * If using a write iolog, store this entry.
279	 */
280	log_io_u(td, io_u);
281
282	io_u->error = 0;
283	io_u->resid = 0;
284
285	if (td->io_ops->flags & FIO_SYNCIO) {
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	if (ddir_rw(acct_ddir(io_u)))
298		td->io_issues[acct_ddir(io_u)]++;
299
300	ret = td->io_ops->queue(td, io_u);
301
302	unlock_file(td, io_u->file);
303
304	/*
305	 * If an error was seen and the io engine didn't propagate it
306	 * back to 'td', do so.
307	 */
308	if (io_u->error && !td->error)
309		td_verror(td, io_u->error, "td_io_queue");
310
311	/*
312	 * Add warning for O_DIRECT so that users have an easier time
313	 * spotting potentially bad alignment. If this triggers for the first
314	 * IO, then it's likely an alignment problem or because the host fs
315	 * does not support O_DIRECT
316	 */
317	if (io_u->error == EINVAL && td->io_issues[io_u->ddir & 1] == 1 &&
318	    td->o.odirect) {
319
320		log_info("fio: first direct IO errored. File system may not "
321			 "support direct IO, or iomem_align= is bad.\n");
322	}
323
324	if (!td->io_ops->commit || ddir_trim(io_u->ddir)) {
325		io_u_mark_submit(td, 1);
326		io_u_mark_complete(td, 1);
327	}
328
329	if (ret == FIO_Q_COMPLETED) {
330		if (ddir_rw(io_u->ddir)) {
331			io_u_mark_depth(td, 1);
332			td->ts.total_io_u[io_u->ddir]++;
333		}
334	} else if (ret == FIO_Q_QUEUED) {
335		int r;
336
337		if (ddir_rw(io_u->ddir)) {
338			td->io_u_queued++;
339			td->ts.total_io_u[io_u->ddir]++;
340		}
341
342		if (td->io_u_queued >= td->o.iodepth_batch) {
343			r = td_io_commit(td);
344			if (r < 0)
345				return r;
346		}
347	}
348
349	if ((td->io_ops->flags & FIO_SYNCIO) == 0) {
350		if (fio_fill_issue_time(td))
351			fio_gettime(&io_u->issue_time, NULL);
352
353		/*
354		 * only used for iolog
355		 */
356		if (td->o.read_iolog_file)
357			memcpy(&td->last_issue, &io_u->issue_time,
358					sizeof(struct timeval));
359	}
360
361	return ret;
362}
363
364int td_io_init(struct thread_data *td)
365{
366	int ret = 0;
367
368	if (td->io_ops->init) {
369		ret = td->io_ops->init(td);
370		if (ret && td->o.iodepth > 1) {
371			log_err("fio: io engine init failed. Perhaps try"
372				" reducing io depth?\n");
373		}
374		if (!td->error)
375			td->error = ret;
376	}
377
378	if (!ret && (td->io_ops->flags & FIO_NOIO))
379		td->flags |= TD_F_NOIO;
380
381	return ret;
382}
383
384int td_io_commit(struct thread_data *td)
385{
386	int ret;
387
388	dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
389
390	if (!td->cur_depth || !td->io_u_queued)
391		return 0;
392
393	io_u_mark_depth(td, td->io_u_queued);
394
395	if (td->io_ops->commit) {
396		ret = td->io_ops->commit(td);
397		if (ret)
398			td_verror(td, -ret, "io commit");
399	}
400
401	/*
402	 * Reflect that events were submitted as async IO requests.
403	 */
404	td->io_u_in_flight += td->io_u_queued;
405	td->io_u_queued = 0;
406
407	return 0;
408}
409
410int td_io_open_file(struct thread_data *td, struct fio_file *f)
411{
412	assert(!fio_file_open(f));
413	assert(f->fd == -1);
414
415	if (td->io_ops->open_file(td, f)) {
416		if (td->error == EINVAL && td->o.odirect)
417			log_err("fio: destination does not support O_DIRECT\n");
418		if (td->error == EMFILE) {
419			log_err("fio: try reducing/setting openfiles (failed"
420				" at %u of %u)\n", td->nr_open_files,
421							td->o.nr_files);
422		}
423
424		assert(f->fd == -1);
425		assert(!fio_file_open(f));
426		return 1;
427	}
428
429	fio_file_reset(td, f);
430	fio_file_set_open(f);
431	fio_file_clear_closing(f);
432	disk_util_inc(f->du);
433
434	td->nr_open_files++;
435	get_file(f);
436
437	if (f->filetype == FIO_TYPE_PIPE) {
438		if (td_random(td)) {
439			log_err("fio: can't seek on pipes (no random io)\n");
440			goto err;
441		}
442	}
443
444	if (td->io_ops->flags & FIO_DISKLESSIO)
445		goto done;
446
447	if (td->o.invalidate_cache && file_invalidate_cache(td, f))
448		goto err;
449
450	if (td->o.fadvise_hint &&
451	    (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_FILE)) {
452		int flags;
453
454		if (td_random(td))
455			flags = POSIX_FADV_RANDOM;
456		else
457			flags = POSIX_FADV_SEQUENTIAL;
458
459		if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
460			td_verror(td, errno, "fadvise");
461			goto err;
462		}
463	}
464
465#ifdef FIO_OS_DIRECTIO
466	/*
467	 * Some OS's have a distinct call to mark the file non-buffered,
468	 * instead of using O_DIRECT (Solaris)
469	 */
470	if (td->o.odirect) {
471		int ret = fio_set_odirect(f->fd);
472
473		if (ret) {
474			td_verror(td, ret, "fio_set_odirect");
475			log_err("fio: the file system does not seem to support direct IO\n");
476			goto err;
477		}
478	}
479#endif
480
481done:
482	log_file(td, f, FIO_LOG_OPEN_FILE);
483	return 0;
484err:
485	disk_util_dec(f->du);
486	if (td->io_ops->close_file)
487		td->io_ops->close_file(td, f);
488	return 1;
489}
490
491int td_io_close_file(struct thread_data *td, struct fio_file *f)
492{
493	if (!fio_file_closing(f))
494		log_file(td, f, FIO_LOG_CLOSE_FILE);
495
496	/*
497	 * mark as closing, do real close when last io on it has completed
498	 */
499	fio_file_set_closing(f);
500
501	disk_util_dec(f->du);
502
503	if (td->o.file_lock_mode != FILE_LOCK_NONE)
504		unlock_file_all(td, f);
505
506	return put_file(td, f);
507}
508
509int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
510{
511	if (!td->io_ops->get_file_size)
512		return 0;
513
514	return td->io_ops->get_file_size(td, f);
515}
516
517static int do_sync_file_range(struct thread_data *td, struct fio_file *f)
518{
519	off64_t offset, nbytes;
520
521	offset = f->first_write;
522	nbytes = f->last_write - f->first_write;
523
524	if (!nbytes)
525		return 0;
526
527	return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
528}
529
530int do_io_u_sync(struct thread_data *td, struct io_u *io_u)
531{
532	int ret;
533
534	if (io_u->ddir == DDIR_SYNC) {
535		ret = fsync(io_u->file->fd);
536	} else if (io_u->ddir == DDIR_DATASYNC) {
537#ifdef CONFIG_FDATASYNC
538		ret = fdatasync(io_u->file->fd);
539#else
540		ret = io_u->xfer_buflen;
541		io_u->error = EINVAL;
542#endif
543	} else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
544		ret = do_sync_file_range(td, io_u->file);
545	else {
546		ret = io_u->xfer_buflen;
547		io_u->error = EINVAL;
548	}
549
550	if (ret < 0)
551		io_u->error = errno;
552
553	return ret;
554}
555
556int do_io_u_trim(struct thread_data *td, struct io_u *io_u)
557{
558#ifndef FIO_HAVE_TRIM
559	io_u->error = EINVAL;
560	return 0;
561#else
562	struct fio_file *f = io_u->file;
563	int ret;
564
565	ret = os_trim(f->fd, io_u->offset, io_u->xfer_buflen);
566	if (!ret)
567		return io_u->xfer_buflen;
568
569	io_u->error = ret;
570	return 0;
571#endif
572}
573
574int fio_show_ioengine_help(const char *engine)
575{
576	struct flist_head *entry;
577	struct thread_data td;
578	char *sep;
579	int ret = 1;
580
581	if (!engine || !*engine) {
582		log_info("Available IO engines:\n");
583		flist_for_each(entry, &engine_list) {
584			td.io_ops = flist_entry(entry, struct ioengine_ops,
585						list);
586			log_info("\t%s\n", td.io_ops->name);
587		}
588		return 0;
589	}
590	sep = strchr(engine, ',');
591	if (sep) {
592		*sep = 0;
593		sep++;
594	}
595
596	memset(&td, 0, sizeof(td));
597
598	td.io_ops = load_ioengine(&td, engine);
599	if (!td.io_ops) {
600		log_info("IO engine %s not found\n", engine);
601		return 1;
602	}
603
604	if (td.io_ops->options)
605		ret = show_cmd_help(td.io_ops->options, sep);
606	else
607		log_info("IO engine %s has no options\n", td.io_ops->name);
608
609	free_ioengine(&td);
610
611	return ret;
612}
613