[fio.git] / engines / dev-dax.c

/*
 * device DAX engine
 *
 * IO engine that reads/writes from files by doing memcpy to/from
 * a memory mapped region of DAX enabled device.
 *
 * Copyright (C) 2016 Intel Corp
 *
 * This program is free software; you can redistribute it and/or
 * modify it under the terms of the GNU General Public License,
 * version 2 as published by the Free Software Foundation..
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
 * GNU General Public License for more details.
 *
 */

/*
 * device dax engine
 * IO engine that access a DAX device directly for read and write data
 *
 * To use:
 *   ioengine=dev-dax
 *
 *   Other relevant settings:
 *     iodepth=1
 *     direct=0	   REQUIRED
 *     filename=/dev/daxN.N
 *     bs=2m
 *
 *     direct should be left to 0. Using dev-dax implies that memory access
 *     is direct. However, dev-dax does not support O_DIRECT flag by design
 *     since it is not necessary.
 *
 *     bs should adhere to the device dax alignment at minimally.
 *
 * libpmem.so
 *   By default, the dev-dax engine will let the system find the libpmem.so
 *   that it uses. You can use an alternative libpmem by setting the
 *   FIO_PMEM_LIB environment variable to the full path to the desired
 *   libpmem.so.
 */

#include <stdio.h>
#include <limits.h>
#include <stdlib.h>
#include <unistd.h>
#include <errno.h>
#include <sys/mman.h>
#include <sys/stat.h>
#include <sys/sysmacros.h>
#include <dlfcn.h>
#include <libgen.h>

#include "../fio.h"
#include "../verify.h"

/*
 * Limits us to 1GiB of mapped files in total to model after
 * mmap engine behavior
 */
#define MMAP_TOTAL_SZ	(1 * 1024 * 1024 * 1024UL)

struct fio_devdax_data {
	void *devdax_ptr;
	size_t devdax_sz;
	off_t devdax_off;
};

static void * (*pmem_memcpy_persist)(void *dest, const void *src, size_t len);

static int fio_devdax_file(struct thread_data *td, struct fio_file *f,
			   size_t length, off_t off)
{
	struct fio_devdax_data *fdd = FILE_ENG_DATA(f);
	int flags = 0;

	if (td_rw(td))
		flags = PROT_READ | PROT_WRITE;
	else if (td_write(td)) {
		flags = PROT_WRITE;

		if (td->o.verify != VERIFY_NONE)
			flags |= PROT_READ;
	} else
		flags = PROT_READ;

	fdd->devdax_ptr = mmap(NULL, length, flags, MAP_SHARED, f->fd, off);
	if (fdd->devdax_ptr == MAP_FAILED) {
		fdd->devdax_ptr = NULL;
		td_verror(td, errno, "mmap");
	}

	if (td->error && fdd->devdax_ptr)
		munmap(fdd->devdax_ptr, length);

	return td->error;
}

/*
 * Just mmap an appropriate portion, we cannot mmap the full extent
 */
static int fio_devdax_prep_limited(struct thread_data *td, struct io_u *io_u)
{
	struct fio_file *f = io_u->file;
	struct fio_devdax_data *fdd = FILE_ENG_DATA(f);

	if (io_u->buflen > f->real_file_size) {
		log_err("fio: bs too big for dev-dax engine\n");
		return EIO;
	}

	fdd->devdax_sz = min(MMAP_TOTAL_SZ, f->real_file_size);
	if (fdd->devdax_sz > f->io_size)
		fdd->devdax_sz = f->io_size;

	fdd->devdax_off = io_u->offset;

	return fio_devdax_file(td, f, fdd->devdax_sz, fdd->devdax_off);
}

/*
 * Attempt to mmap the entire file
 */
static int fio_devdax_prep_full(struct thread_data *td, struct io_u *io_u)
{
	struct fio_file *f = io_u->file;
	struct fio_devdax_data *fdd = FILE_ENG_DATA(f);
	int ret;

	if (fio_file_partial_mmap(f))
		return EINVAL;

	if (io_u->offset != (size_t) io_u->offset ||
	    f->io_size != (size_t) f->io_size) {
		fio_file_set_partial_mmap(f);
		return EINVAL;
	}

	fdd->devdax_sz = f->io_size;
	fdd->devdax_off = 0;

	ret = fio_devdax_file(td, f, fdd->devdax_sz, fdd->devdax_off);
	if (ret)
		fio_file_set_partial_mmap(f);

	return ret;
}

static int fio_devdax_prep(struct thread_data *td, struct io_u *io_u)
{
	struct fio_file *f = io_u->file;
	struct fio_devdax_data *fdd = FILE_ENG_DATA(f);
	int ret;

	/*
	 * It fits within existing mapping, use it
	 */
	if (io_u->offset >= fdd->devdax_off &&
	    io_u->offset + io_u->buflen < fdd->devdax_off + fdd->devdax_sz)
		goto done;

	/*
	 * unmap any existing mapping
	 */
	if (fdd->devdax_ptr) {
		if (munmap(fdd->devdax_ptr, fdd->devdax_sz) < 0)
			return errno;
		fdd->devdax_ptr = NULL;
	}

	if (fio_devdax_prep_full(td, io_u)) {
		td_clear_error(td);
		ret = fio_devdax_prep_limited(td, io_u);
		if (ret)
			return ret;
	}

done:
	io_u->mmap_data = fdd->devdax_ptr + io_u->offset - fdd->devdax_off -
				f->file_offset;
	return 0;
}

static int fio_devdax_queue(struct thread_data *td, struct io_u *io_u)
{
	fio_ro_check(td, io_u);
	io_u->error = 0;

	switch (io_u->ddir) {
	case DDIR_READ:
		memcpy(io_u->xfer_buf, io_u->mmap_data, io_u->xfer_buflen);
		break;
	case DDIR_WRITE:
		pmem_memcpy_persist(io_u->mmap_data, io_u->xfer_buf,
				    io_u->xfer_buflen);
		break;
	case DDIR_SYNC:
	case DDIR_DATASYNC:
	case DDIR_SYNC_FILE_RANGE:
		break;
	default:
		io_u->error = EINVAL;
		break;
	}

	return FIO_Q_COMPLETED;
}

static int fio_devdax_init(struct thread_data *td)
{
	struct thread_options *o = &td->o;
	const char *path;
	void *dl;

	if ((o->rw_min_bs & page_mask) &&
	    (o->fsync_blocks || o->fdatasync_blocks)) {
		log_err("fio: mmap options dictate a minimum block size of "
			"%llu bytes\n", (unsigned long long) page_size);
		return 1;
	}

	path = getenv("FIO_PMEM_LIB");
	if (!path)
		path = "libpmem.so";

	dl = dlopen(path, RTLD_NOW | RTLD_NODELETE);
	if (!dl) {
		log_err("fio: unable to open libpmem: %s\n", dlerror());
		return 1;
	}

	pmem_memcpy_persist = dlsym(dl, "pmem_memcpy_persist");
	if (!pmem_memcpy_persist) {
		log_err("fio: unable to load libpmem: %s\n", dlerror());
		return 1;
	}

	return 0;
}

static int fio_devdax_open_file(struct thread_data *td, struct fio_file *f)
{
	struct fio_devdax_data *fdd;
	int ret;

	ret = generic_open_file(td, f);
	if (ret)
		return ret;

	fdd = calloc(1, sizeof(*fdd));
	if (!fdd) {
		int fio_unused __ret;
		__ret = generic_close_file(td, f);
		return 1;
	}

	FILE_SET_ENG_DATA(f, fdd);

	return 0;
}

static int fio_devdax_close_file(struct thread_data *td, struct fio_file *f)
{
	struct fio_devdax_data *fdd = FILE_ENG_DATA(f);

	FILE_SET_ENG_DATA(f, NULL);
	free(fdd);
	fio_file_clear_partial_mmap(f);

	return generic_close_file(td, f);
}

static int
fio_devdax_get_file_size(struct thread_data *td, struct fio_file *f)
{
	char spath[PATH_MAX];
	char npath[PATH_MAX];
	char *rpath;
	FILE *sfile;
	uint64_t size;
	struct stat st;
	int rc;

	if (fio_file_size_known(f))
		return 0;

	if (f->filetype != FIO_TYPE_CHAR)
		return -EINVAL;

	rc = stat(f->file_name, &st);
	if (rc < 0) {
		log_err("%s: failed to stat file %s: %d\n",
			td->o.name, f->file_name, errno);
		return -errno;
	}

	snprintf(spath, PATH_MAX, "/sys/dev/char/%d:%d/subsystem",
		 major(st.st_rdev), minor(st.st_rdev));

	rpath = realpath(spath, npath);
	if (!rpath) {
		log_err("%s: realpath on %s failed: %d\n",
			td->o.name, spath, errno);
		return -errno;
	}

	/* check if DAX device */
	if (strcmp("/sys/class/dax", rpath)) {
		log_err("%s: %s not a DAX device!\n",
			td->o.name, f->file_name);
	}

	snprintf(spath, PATH_MAX, "/sys/dev/char/%d:%d/size",
		 major(st.st_rdev), minor(st.st_rdev));

	sfile = fopen(spath, "r");
	if (!sfile) {
		log_err("%s: fopen on %s failed: %d\n",
			td->o.name, spath, errno);
		return 1;
	}

	rc = fscanf(sfile, "%lu", &size);
	if (rc < 0) {
		log_err("%s: fscanf on %s failed: %d\n",
			td->o.name, spath, errno);
		return 1;
	}

	f->real_file_size = size;

	fclose(sfile);

	if (f->file_offset > f->real_file_size) {
		log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
					(unsigned long long) f->file_offset,
					(unsigned long long) f->real_file_size);
		return 1;
	}

	fio_file_set_size_known(f);
	return 0;
}

static struct ioengine_ops ioengine = {
	.name		= "dev-dax",
	.version	= FIO_IOOPS_VERSION,
	.init		= fio_devdax_init,
	.prep		= fio_devdax_prep,
	.queue		= fio_devdax_queue,
	.open_file	= fio_devdax_open_file,
	.close_file	= fio_devdax_close_file,
	.get_file_size	= fio_devdax_get_file_size,
	.flags		= FIO_SYNCIO | FIO_DISKLESSIO | FIO_NOEXTEND | FIO_NODISKUTIL,
};

static void fio_init fio_devdax_register(void)
{
	register_ioengine(&ioengine);
}

static void fio_exit fio_devdax_unregister(void)
{
	unregister_ioengine(&ioengine);
}
Commit	Line	Data
cbb15e42 DJ	1	/*
	2	* device DAX engine
	3	*
	4	* IO engine that reads/writes from files by doing memcpy to/from
	5	* a memory mapped region of DAX enabled device.
	6	*
	7	* Copyright (C) 2016 Intel Corp
	8	*
	9	* This program is free software; you can redistribute it and/or
	10	* modify it under the terms of the GNU General Public License,
	11	* version 2 as published by the Free Software Foundation..
	12	*
	13	* This program is distributed in the hope that it will be useful,
	14	* but WITHOUT ANY WARRANTY; without even the implied warranty of
	15	* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
	16	* GNU General Public License for more details.
	17	*
	18	*/
	19
	20	/*
	21	* device dax engine
	22	* IO engine that access a DAX device directly for read and write data
	23	*
	24	* To use:
	25	* ioengine=dev-dax
	26	*
	27	* Other relevant settings:
	28	* iodepth=1
	29	* direct=0 REQUIRED
	30	* filename=/dev/daxN.N
	31	* bs=2m
	32	*
	33	* direct should be left to 0. Using dev-dax implies that memory access
	34	* is direct. However, dev-dax does not support O_DIRECT flag by design
	35	* since it is not necessary.
	36	*
	37	* bs should adhere to the device dax alignment at minimally.
	38	*
	39	* libpmem.so
	40	* By default, the dev-dax engine will let the system find the libpmem.so
	41	* that it uses. You can use an alternative libpmem by setting the
	42	* FIO_PMEM_LIB environment variable to the full path to the desired
	43	* libpmem.so.
	44	*/
	45
	46	#include <stdio.h>
	47	#include <limits.h>
	48	#include <stdlib.h>
	49	#include <unistd.h>
	50	#include <errno.h>
	51	#include <sys/mman.h>
	52	#include <sys/stat.h>
	53	#include <sys/sysmacros.h>
	54	#include <dlfcn.h>
	55	#include <libgen.h>
	56
	57	#include "../fio.h"
	58	#include "../verify.h"
	59
	60	/*
420b104a	61	* Limits us to 1GiB of mapped files in total to model after
cbb15e42 DJ	62	* mmap engine behavior
	63	*/
	64	#define MMAP_TOTAL_SZ (1 * 1024 * 1024 * 1024UL)
	65
	66	struct fio_devdax_data {
	67	void *devdax_ptr;
	68	size_t devdax_sz;
	69	off_t devdax_off;
	70	};
	71
	72	static void * (pmem_memcpy_persist)(void dest, const void *src, size_t len);
	73
	74	static int fio_devdax_file(struct thread_data td, struct fio_file f,
	75	size_t length, off_t off)
	76	{
	77	struct fio_devdax_data *fdd = FILE_ENG_DATA(f);
	78	int flags = 0;
	79
	80	if (td_rw(td))
	81	flags = PROT_READ \| PROT_WRITE;
	82	else if (td_write(td)) {
	83	flags = PROT_WRITE;
	84
	85	if (td->o.verify != VERIFY_NONE)
	86	flags \|= PROT_READ;
	87	} else
	88	flags = PROT_READ;
	89
	90	fdd->devdax_ptr = mmap(NULL, length, flags, MAP_SHARED, f->fd, off);
	91	if (fdd->devdax_ptr == MAP_FAILED) {
	92	fdd->devdax_ptr = NULL;
	93	td_verror(td, errno, "mmap");
	94	}
	95
	96	if (td->error && fdd->devdax_ptr)
	97	munmap(fdd->devdax_ptr, length);
	98
	99	return td->error;
	100	}
	101
	102	/*
	103	* Just mmap an appropriate portion, we cannot mmap the full extent
	104	*/
	105	static int fio_devdax_prep_limited(struct thread_data td, struct io_u io_u)
	106	{
	107	struct fio_file *f = io_u->file;
	108	struct fio_devdax_data *fdd = FILE_ENG_DATA(f);
	109
	110	if (io_u->buflen > f->real_file_size) {
	111	log_err("fio: bs too big for dev-dax engine\n");
	112	return EIO;
	113	}
	114
	115	fdd->devdax_sz = min(MMAP_TOTAL_SZ, f->real_file_size);
	116	if (fdd->devdax_sz > f->io_size)
	117	fdd->devdax_sz = f->io_size;
	118
	119	fdd->devdax_off = io_u->offset;
	120
	121	return fio_devdax_file(td, f, fdd->devdax_sz, fdd->devdax_off);
	122	}
	123
	124	/*
	125	* Attempt to mmap the entire file
126	*/
127	static int fio_devdax_prep_full(struct thread_data td, struct io_u io_u)
128	{
129	struct fio_file *f = io_u->file;
130	struct fio_devdax_data *fdd = FILE_ENG_DATA(f);
131	int ret;
132
133	if (fio_file_partial_mmap(f))
134	return EINVAL;
135
136	if (io_u->offset != (size_t) io_u->offset \|\|
137	f->io_size != (size_t) f->io_size) {
138	fio_file_set_partial_mmap(f);
139	return EINVAL;
140	}
141
142	fdd->devdax_sz = f->io_size;
143	fdd->devdax_off = 0;
144
145	ret = fio_devdax_file(td, f, fdd->devdax_sz, fdd->devdax_off);
146	if (ret)
147	fio_file_set_partial_mmap(f);
148
149	return ret;
150	}
151
152	static int fio_devdax_prep(struct thread_data td, struct io_u io_u)
153	{
154	struct fio_file *f = io_u->file;
155	struct fio_devdax_data *fdd = FILE_ENG_DATA(f);
156	int ret;
157
158	/*
159	* It fits within existing mapping, use it
160	*/
161	if (io_u->offset >= fdd->devdax_off &&
162	io_u->offset + io_u->buflen < fdd->devdax_off + fdd->devdax_sz)
163	goto done;
164
165	/*
166	* unmap any existing mapping
167	*/
168	if (fdd->devdax_ptr) {
169	if (munmap(fdd->devdax_ptr, fdd->devdax_sz) < 0)
170	return errno;
171	fdd->devdax_ptr = NULL;
172	}
173
174	if (fio_devdax_prep_full(td, io_u)) {
175	td_clear_error(td);
176	ret = fio_devdax_prep_limited(td, io_u);
177	if (ret)
178	return ret;
179	}
180
181	done:
182	io_u->mmap_data = fdd->devdax_ptr + io_u->offset - fdd->devdax_off -
183	f->file_offset;
184	return 0;
185	}
186
187	static int fio_devdax_queue(struct thread_data td, struct io_u io_u)
188	{
189	fio_ro_check(td, io_u);
190	io_u->error = 0;
191
192	switch (io_u->ddir) {
193	case DDIR_READ:
194	memcpy(io_u->xfer_buf, io_u->mmap_data, io_u->xfer_buflen);
195	break;
196	case DDIR_WRITE:
197	pmem_memcpy_persist(io_u->mmap_data, io_u->xfer_buf,
198	io_u->xfer_buflen);
199	break;
200	case DDIR_SYNC:
201	case DDIR_DATASYNC:
202	case DDIR_SYNC_FILE_RANGE:
203	break;
204	default:
205	io_u->error = EINVAL;
206	break;
207	}
208
209	return FIO_Q_COMPLETED;
210	}
211
212	static int fio_devdax_init(struct thread_data *td)
213	{
214	struct thread_options *o = &td->o;
215	const char *path;
216	void *dl;
217
218	if ((o->rw_min_bs & page_mask) &&
219	(o->fsync_blocks \|\| o->fdatasync_blocks)) {
220	log_err("fio: mmap options dictate a minimum block size of "
221	"%llu bytes\n", (unsigned long long) page_size);
222	return 1;
223	}
224
225	path = getenv("FIO_PMEM_LIB");
226	if (!path)
227	path = "libpmem.so";
228
229	dl = dlopen(path, RTLD_NOW \| RTLD_NODELETE);
230	if (!dl) {
231	log_err("fio: unable to open libpmem: %s\n", dlerror());
232	return 1;
233	}
234
235	pmem_memcpy_persist = dlsym(dl, "pmem_memcpy_persist");
236	if (!pmem_memcpy_persist) {
237	log_err("fio: unable to load libpmem: %s\n", dlerror());
238	return 1;
239	}
240
241	return 0;
242	}
243
244	static int fio_devdax_open_file(struct thread_data td, struct fio_file f)
245	{
246	struct fio_devdax_data *fdd;
247	int ret;
248
249	ret = generic_open_file(td, f);
250	if (ret)
251	return ret;
252
253	fdd = calloc(1, sizeof(*fdd));
254	if (!fdd) {
255	int fio_unused __ret;
256	__ret = generic_close_file(td, f);
257	return 1;
258	}
259
260	FILE_SET_ENG_DATA(f, fdd);
261
262	return 0;
263	}
264
265	static int fio_devdax_close_file(struct thread_data td, struct fio_file f)
266	{
267	struct fio_devdax_data *fdd = FILE_ENG_DATA(f);
268
269	FILE_SET_ENG_DATA(f, NULL);
270	free(fdd);
271	fio_file_clear_partial_mmap(f);
272
273	return generic_close_file(td, f);
274	}
275
276	static int
277	fio_devdax_get_file_size(struct thread_data td, struct fio_file f)
278	{
279	char spath[PATH_MAX];
280	char npath[PATH_MAX];
281	char *rpath;
282	FILE *sfile;
283	uint64_t size;
284	struct stat st;
285	int rc;
286
287	if (fio_file_size_known(f))
288	return 0;
289
290	if (f->filetype != FIO_TYPE_CHAR)
291	return -EINVAL;
292
293	rc = stat(f->file_name, &st);
294	if (rc < 0) {
295	log_err("%s: failed to stat file %s: %d\n",
296	td->o.name, f->file_name, errno);
297	return -errno;
298	}
299
300	snprintf(spath, PATH_MAX, "/sys/dev/char/%d:%d/subsystem",
301	major(st.st_rdev), minor(st.st_rdev));
302
303	rpath = realpath(spath, npath);
304	if (!rpath) {
305	log_err("%s: realpath on %s failed: %d\n",
306	td->o.name, spath, errno);
307	return -errno;
308	}
309
310	/* check if DAX device */
311	if (strcmp("/sys/class/dax", rpath)) {
312	log_err("%s: %s not a DAX device!\n",
313	td->o.name, f->file_name);
314	}
315
316	snprintf(spath, PATH_MAX, "/sys/dev/char/%d:%d/size",
317	major(st.st_rdev), minor(st.st_rdev));
318
319	sfile = fopen(spath, "r");
320	if (!sfile) {
321	log_err("%s: fopen on %s failed: %d\n",
322	td->o.name, spath, errno);
323	return 1;
324	}
325
326	rc = fscanf(sfile, "%lu", &size);
327	if (rc < 0) {
328	log_err("%s: fscanf on %s failed: %d\n",
329	td->o.name, spath, errno);
330	return 1;
331	}
332
333	f->real_file_size = size;
334
335	fclose(sfile);
336
337	if (f->file_offset > f->real_file_size) {
338	log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
339	(unsigned long long) f->file_offset,
340	(unsigned long long) f->real_file_size);
341	return 1;
342	}
343
344	fio_file_set_size_known(f);
345	return 0;
346	}
347
348	static struct ioengine_ops ioengine = {
349	.name = "dev-dax",
350	.version = FIO_IOOPS_VERSION,
351	.init = fio_devdax_init,
352	.prep = fio_devdax_prep,
353	.queue = fio_devdax_queue,
354	.open_file = fio_devdax_open_file,
355	.close_file = fio_devdax_close_file,
356	.get_file_size = fio_devdax_get_file_size,
357	.flags = FIO_SYNCIO \| FIO_DISKLESSIO \| FIO_NOEXTEND \| FIO_NODISKUTIL,
358	};
359
360	static void fio_init fio_devdax_register(void)
361	{
362	register_ioengine(&ioengine);
363	}
364
365	static void fio_exit fio_devdax_unregister(void)
366	{
367	unregister_ioengine(&ioengine);
368	}