blob: aa00234e4576f17098e14ea2ace651b6b771a61c [file] [log] [blame]
Jens Axboe2866c822006-10-09 15:57:48 +02001/*
Jens Axboeda751ca2007-03-14 10:59:33 +01002 * splice engine
3 *
4 * IO engine that transfers data by doing splices to/from pipes and
5 * the files.
Jens Axboe2866c822006-10-09 15:57:48 +02006 *
7 */
8#include <stdio.h>
9#include <stdlib.h>
10#include <unistd.h>
11#include <errno.h>
12#include <assert.h>
13#include <sys/poll.h>
Jens Axboe81887d52007-09-07 20:36:08 +020014#include <sys/mman.h>
Jens Axboe5f350952006-11-07 15:20:59 +010015
16#include "../fio.h"
Jens Axboe2866c822006-10-09 15:57:48 +020017
Jens Axboe34cfcda2006-11-03 14:00:45 +010018#ifdef FIO_HAVE_SPLICE
19
Jens Axboe2866c822006-10-09 15:57:48 +020020struct spliceio_data {
Jens Axboe2866c822006-10-09 15:57:48 +020021 int pipe[2];
Jens Axboef24254e2007-06-13 19:04:16 +020022 int vmsplice_to_user;
Jens Axboe8b850242007-10-16 09:42:40 +020023 int vmsplice_to_user_map;
Jens Axboe2866c822006-10-09 15:57:48 +020024};
25
Jens Axboe2866c822006-10-09 15:57:48 +020026/*
Jens Axboef24254e2007-06-13 19:04:16 +020027 * vmsplice didn't use to support splicing to user space, this is the old
28 * variant of getting that job done. Doesn't make a lot of sense, but it
29 * uses splices to move data from the source into a pipe.
Jens Axboe2866c822006-10-09 15:57:48 +020030 */
Jens Axboef24254e2007-06-13 19:04:16 +020031static int fio_splice_read_old(struct thread_data *td, struct io_u *io_u)
Jens Axboe2866c822006-10-09 15:57:48 +020032{
33 struct spliceio_data *sd = td->io_ops->data;
Jens Axboe53cdc682006-10-18 11:50:58 +020034 struct fio_file *f = io_u->file;
Jens Axboe2866c822006-10-09 15:57:48 +020035 int ret, ret2, buflen;
36 off_t offset;
37 void *p;
38
39 offset = io_u->offset;
Jens Axboecec6b552007-02-06 20:15:38 +010040 buflen = io_u->xfer_buflen;
41 p = io_u->xfer_buf;
Jens Axboe2866c822006-10-09 15:57:48 +020042 while (buflen) {
43 int this_len = buflen;
44
45 if (this_len > SPLICE_DEF_SIZE)
46 this_len = SPLICE_DEF_SIZE;
47
Jens Axboe53cdc682006-10-18 11:50:58 +020048 ret = splice(f->fd, &offset, sd->pipe[1], NULL, this_len, SPLICE_F_MORE);
Jens Axboe2866c822006-10-09 15:57:48 +020049 if (ret < 0) {
50 if (errno == ENODATA || errno == EAGAIN)
51 continue;
52
Jens Axboeb4ba5f32007-04-02 10:36:59 +020053 return -errno;
Jens Axboe2866c822006-10-09 15:57:48 +020054 }
55
56 buflen -= ret;
57
58 while (ret) {
59 ret2 = read(sd->pipe[0], p, ret);
60 if (ret2 < 0)
Jens Axboeb4ba5f32007-04-02 10:36:59 +020061 return -errno;
Jens Axboe2866c822006-10-09 15:57:48 +020062
63 ret -= ret2;
64 p += ret2;
65 }
66 }
67
Jens Axboecec6b552007-02-06 20:15:38 +010068 return io_u->xfer_buflen;
Jens Axboe2866c822006-10-09 15:57:48 +020069}
70
71/*
Jens Axboef24254e2007-06-13 19:04:16 +020072 * We can now vmsplice into userspace, so do the transfer by splicing into
73 * a pipe and vmsplicing that into userspace.
74 */
75static int fio_splice_read(struct thread_data *td, struct io_u *io_u)
76{
77 struct spliceio_data *sd = td->io_ops->data;
78 struct fio_file *f = io_u->file;
79 struct iovec iov;
Jens Axboe8b850242007-10-16 09:42:40 +020080 int ret , buflen, mmap_len;
Jens Axboef24254e2007-06-13 19:04:16 +020081 off_t offset;
Jens Axboe81887d52007-09-07 20:36:08 +020082 void *p, *map;
Jens Axboef24254e2007-06-13 19:04:16 +020083
Jens Axboe8b850242007-10-16 09:42:40 +020084 ret = 0;
Jens Axboef24254e2007-06-13 19:04:16 +020085 offset = io_u->offset;
Jens Axboe81887d52007-09-07 20:36:08 +020086 mmap_len = buflen = io_u->xfer_buflen;
87
Jens Axboe8b850242007-10-16 09:42:40 +020088 if (sd->vmsplice_to_user_map) {
89 map = mmap(io_u->xfer_buf, buflen, PROT_READ, MAP_PRIVATE|OS_MAP_ANON, 0, 0);
90 if (map == MAP_FAILED) {
91 td_verror(td, errno, "mmap io_u");
92 return -1;
93 }
94
95 p = map;
96 } else {
97 map = NULL;
98 p = io_u->xfer_buf;
Jens Axboe81887d52007-09-07 20:36:08 +020099 }
100
Jens Axboef24254e2007-06-13 19:04:16 +0200101 while (buflen) {
102 int this_len = buflen;
Jens Axboe81887d52007-09-07 20:36:08 +0200103 int flags = 0;
Jens Axboef24254e2007-06-13 19:04:16 +0200104
Jens Axboe81887d52007-09-07 20:36:08 +0200105 if (this_len > SPLICE_DEF_SIZE) {
Jens Axboef24254e2007-06-13 19:04:16 +0200106 this_len = SPLICE_DEF_SIZE;
Jens Axboe81887d52007-09-07 20:36:08 +0200107 flags = SPLICE_F_MORE;
108 }
Jens Axboef24254e2007-06-13 19:04:16 +0200109
Jens Axboe81887d52007-09-07 20:36:08 +0200110 ret = splice(f->fd, &offset, sd->pipe[1], NULL, this_len,flags);
Jens Axboef24254e2007-06-13 19:04:16 +0200111 if (ret < 0) {
112 if (errno == ENODATA || errno == EAGAIN)
113 continue;
114
Jens Axboe81887d52007-09-07 20:36:08 +0200115 td_verror(td, errno, "splice-from-fd");
116 break;
Jens Axboef24254e2007-06-13 19:04:16 +0200117 }
118
119 buflen -= ret;
120 iov.iov_base = p;
121 iov.iov_len = ret;
Jens Axboef24254e2007-06-13 19:04:16 +0200122
123 while (iov.iov_len) {
124 ret = vmsplice(sd->pipe[0], &iov, 1, SPLICE_F_MOVE);
Jens Axboe81887d52007-09-07 20:36:08 +0200125 if (ret < 0) {
Jens Axboecd988012009-04-20 12:39:17 +0200126 if (errno == EFAULT &&
127 sd->vmsplice_to_user_map) {
Jens Axboe8b850242007-10-16 09:42:40 +0200128 sd->vmsplice_to_user_map = 0;
129 munmap(map, mmap_len);
Jens Axboecd988012009-04-20 12:39:17 +0200130 map = NULL;
131 p = io_u->xfer_buf;
132 iov.iov_base = p;
133 continue;
Jens Axboe8b850242007-10-16 09:42:40 +0200134 }
Jens Axboe3a6d2672007-10-16 09:44:47 +0200135 if (errno == EBADF) {
136 ret = -EBADF;
137 break;
138 }
Jens Axboe81887d52007-09-07 20:36:08 +0200139 td_verror(td, errno, "vmsplice");
140 break;
141 } else if (!ret) {
142 td_verror(td, ENODATA, "vmsplice");
143 ret = -1;
144 break;
145 }
Jens Axboef24254e2007-06-13 19:04:16 +0200146
147 iov.iov_len -= ret;
148 iov.iov_base += ret;
Jens Axboecd988012009-04-20 12:39:17 +0200149 p += ret;
Jens Axboef24254e2007-06-13 19:04:16 +0200150 }
Jens Axboe81887d52007-09-07 20:36:08 +0200151 if (ret < 0)
152 break;
Jens Axboef24254e2007-06-13 19:04:16 +0200153 }
154
Jens Axboe8b850242007-10-16 09:42:40 +0200155 if (sd->vmsplice_to_user_map && munmap(map, mmap_len) < 0) {
Jens Axboe81887d52007-09-07 20:36:08 +0200156 td_verror(td, errno, "munnap io_u");
157 return -1;
158 }
159 if (ret < 0)
160 return ret;
161
Jens Axboef24254e2007-06-13 19:04:16 +0200162 return io_u->xfer_buflen;
163}
164
Jens Axboef24254e2007-06-13 19:04:16 +0200165/*
Jens Axboe2866c822006-10-09 15:57:48 +0200166 * For splice writing, we can vmsplice our data buffer directly into a
167 * pipe and then splice that to a file.
168 */
169static int fio_splice_write(struct thread_data *td, struct io_u *io_u)
170{
171 struct spliceio_data *sd = td->io_ops->data;
Jens Axboef24254e2007-06-13 19:04:16 +0200172 struct iovec iov = {
173 .iov_base = io_u->xfer_buf,
174 .iov_len = io_u->xfer_buflen,
Jens Axboe2866c822006-10-09 15:57:48 +0200175 };
176 struct pollfd pfd = { .fd = sd->pipe[1], .events = POLLOUT, };
Jens Axboe53cdc682006-10-18 11:50:58 +0200177 struct fio_file *f = io_u->file;
Jens Axboe2866c822006-10-09 15:57:48 +0200178 off_t off = io_u->offset;
179 int ret, ret2;
180
Jens Axboef24254e2007-06-13 19:04:16 +0200181 while (iov.iov_len) {
Jens Axboe2866c822006-10-09 15:57:48 +0200182 if (poll(&pfd, 1, -1) < 0)
183 return errno;
184
Jens Axboef24254e2007-06-13 19:04:16 +0200185 ret = vmsplice(sd->pipe[1], &iov, 1, SPLICE_F_NONBLOCK);
Jens Axboe2866c822006-10-09 15:57:48 +0200186 if (ret < 0)
Jens Axboeb4ba5f32007-04-02 10:36:59 +0200187 return -errno;
Jens Axboe2866c822006-10-09 15:57:48 +0200188
Jens Axboef24254e2007-06-13 19:04:16 +0200189 iov.iov_len -= ret;
190 iov.iov_base += ret;
Jens Axboe2866c822006-10-09 15:57:48 +0200191
192 while (ret) {
Jens Axboe53cdc682006-10-18 11:50:58 +0200193 ret2 = splice(sd->pipe[0], NULL, f->fd, &off, ret, 0);
Jens Axboe2866c822006-10-09 15:57:48 +0200194 if (ret2 < 0)
Jens Axboeb4ba5f32007-04-02 10:36:59 +0200195 return -errno;
Jens Axboe2866c822006-10-09 15:57:48 +0200196
197 ret -= ret2;
198 }
199 }
200
Jens Axboecec6b552007-02-06 20:15:38 +0100201 return io_u->xfer_buflen;
Jens Axboe2866c822006-10-09 15:57:48 +0200202}
203
204static int fio_spliceio_queue(struct thread_data *td, struct io_u *io_u)
205{
Jens Axboef24254e2007-06-13 19:04:16 +0200206 struct spliceio_data *sd = td->io_ops->data;
Jens Axboef3f552b2008-06-13 10:27:05 +0200207 int uninitialized_var(ret);
Jens Axboe2866c822006-10-09 15:57:48 +0200208
Jens Axboe7101d9c2007-09-12 13:12:39 +0200209 fio_ro_check(td, io_u);
210
Jens Axboef24254e2007-06-13 19:04:16 +0200211 if (io_u->ddir == DDIR_READ) {
Jens Axboe43688572007-08-14 13:43:05 +0200212 if (sd->vmsplice_to_user) {
Jens Axboef24254e2007-06-13 19:04:16 +0200213 ret = fio_splice_read(td, io_u);
Jens Axboe43688572007-08-14 13:43:05 +0200214 /*
215 * This kernel doesn't support vmsplice to user
216 * space. Reset the vmsplice_to_user flag, so that
217 * we retry below and don't hit this path again.
218 */
219 if (ret == -EBADF)
220 sd->vmsplice_to_user = 0;
221 }
222 if (!sd->vmsplice_to_user)
Jens Axboef24254e2007-06-13 19:04:16 +0200223 ret = fio_splice_read_old(td, io_u);
224 } else if (io_u->ddir == DDIR_WRITE)
Jens Axboe2866c822006-10-09 15:57:48 +0200225 ret = fio_splice_write(td, io_u);
Jens Axboea5f30272010-07-19 16:19:55 -0600226 else if (io_u->ddir == DDIR_TRIM)
227 ret = do_io_u_trim(td, io_u);
Jens Axboe87dc1ab2006-10-24 14:41:26 +0200228 else
Jens Axboef0115312010-03-09 21:47:15 +0100229 ret = do_io_u_sync(td, io_u);
Jens Axboe2866c822006-10-09 15:57:48 +0200230
Jens Axboecec6b552007-02-06 20:15:38 +0100231 if (ret != (int) io_u->xfer_buflen) {
Jens Axboe22819ec2007-02-18 07:47:14 +0100232 if (ret >= 0) {
Jens Axboecec6b552007-02-06 20:15:38 +0100233 io_u->resid = io_u->xfer_buflen - ret;
234 io_u->error = 0;
Jens Axboe36167d82007-02-18 05:41:31 +0100235 return FIO_Q_COMPLETED;
Jens Axboe2866c822006-10-09 15:57:48 +0200236 } else
237 io_u->error = errno;
238 }
239
Jens Axboe0aa417a2008-06-09 19:25:49 +0200240 if (io_u->error) {
Jens Axboee1161c32007-02-22 19:36:48 +0100241 td_verror(td, io_u->error, "xfer");
Jens Axboe0aa417a2008-06-09 19:25:49 +0200242 if (io_u->error == EINVAL)
243 log_err("fio: looks like splice doesn't work on this"
244 " file system\n");
245 }
Jens Axboe2866c822006-10-09 15:57:48 +0200246
Jens Axboe36167d82007-02-18 05:41:31 +0100247 return FIO_Q_COMPLETED;
Jens Axboe2866c822006-10-09 15:57:48 +0200248}
249
250static void fio_spliceio_cleanup(struct thread_data *td)
251{
252 struct spliceio_data *sd = td->io_ops->data;
253
254 if (sd) {
255 close(sd->pipe[0]);
256 close(sd->pipe[1]);
257 free(sd);
Jens Axboe2866c822006-10-09 15:57:48 +0200258 }
259}
260
261static int fio_spliceio_init(struct thread_data *td)
262{
263 struct spliceio_data *sd = malloc(sizeof(*sd));
264
Jens Axboe2866c822006-10-09 15:57:48 +0200265 if (pipe(sd->pipe) < 0) {
Jens Axboee1161c32007-02-22 19:36:48 +0100266 td_verror(td, errno, "pipe");
Jens Axboe2866c822006-10-09 15:57:48 +0200267 free(sd);
268 return 1;
269 }
270
Jens Axboef24254e2007-06-13 19:04:16 +0200271 /*
Jens Axboe43688572007-08-14 13:43:05 +0200272 * Assume this work, we'll reset this if it doesn't
Jens Axboef24254e2007-06-13 19:04:16 +0200273 */
Jens Axboe43688572007-08-14 13:43:05 +0200274 sd->vmsplice_to_user = 1;
Jens Axboef24254e2007-06-13 19:04:16 +0200275
Jens Axboe81887d52007-09-07 20:36:08 +0200276 /*
Jens Axboe8b850242007-10-16 09:42:40 +0200277 * Works with "real" vmsplice to user, eg mapping pages directly.
278 * Reset if we fail.
279 */
280 sd->vmsplice_to_user_map = 1;
281
282 /*
Jens Axboe81887d52007-09-07 20:36:08 +0200283 * And if vmsplice_to_user works, we definitely need aligned
284 * buffers. Just set ->odirect to force that.
285 */
286 if (td_read(td))
Jens Axboe76a31792011-05-13 21:19:25 +0200287 td->o.mem_align = 1;
Jens Axboe81887d52007-09-07 20:36:08 +0200288
Jens Axboe2866c822006-10-09 15:57:48 +0200289 td->io_ops->data = sd;
290 return 0;
291}
292
Jens Axboe5f350952006-11-07 15:20:59 +0100293static struct ioengine_ops ioengine = {
Jens Axboe2866c822006-10-09 15:57:48 +0200294 .name = "splice",
295 .version = FIO_IOOPS_VERSION,
296 .init = fio_spliceio_init,
297 .queue = fio_spliceio_queue,
Jens Axboe2866c822006-10-09 15:57:48 +0200298 .cleanup = fio_spliceio_cleanup,
Jens Axboeb5af8292007-03-08 12:43:13 +0100299 .open_file = generic_open_file,
300 .close_file = generic_close_file,
Jens Axboedf9c26b2009-03-05 10:13:58 +0100301 .get_file_size = generic_get_file_size,
Jens Axboe9c0d2242009-07-01 12:26:28 +0200302 .flags = FIO_SYNCIO | FIO_PIPEIO,
Jens Axboe2866c822006-10-09 15:57:48 +0200303};
Jens Axboe34cfcda2006-11-03 14:00:45 +0100304
305#else /* FIO_HAVE_SPLICE */
306
307/*
308 * When we have a proper configure system in place, we simply wont build
309 * and install this io engine. For now install a crippled version that
310 * just complains and fails to load.
311 */
312static int fio_spliceio_init(struct thread_data fio_unused *td)
313{
Jens Axboea3edaf72010-09-26 10:53:40 +0900314 log_err("fio: splice not available\n");
Jens Axboe34cfcda2006-11-03 14:00:45 +0100315 return 1;
316}
317
Jens Axboe5f350952006-11-07 15:20:59 +0100318static struct ioengine_ops ioengine = {
Jens Axboe34cfcda2006-11-03 14:00:45 +0100319 .name = "splice",
320 .version = FIO_IOOPS_VERSION,
321 .init = fio_spliceio_init,
322};
323
324#endif
Jens Axboe5f350952006-11-07 15:20:59 +0100325
326static void fio_init fio_spliceio_register(void)
327{
328 register_ioengine(&ioengine);
329}
330
331static void fio_exit fio_spliceio_unregister(void)
332{
333 unregister_ioengine(&ioengine);
334}