Commit | Line | Data |
---|---|---|
2866c82d | 1 | /* |
da751ca9 JA |
2 | * splice engine |
3 | * | |
4 | * IO engine that transfers data by doing splices to/from pipes and | |
5 | * the files. | |
2866c82d JA |
6 | * |
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> | |
5f350952 JA |
14 | |
15 | #include "../fio.h" | |
2866c82d | 16 | |
34cfcdaf JA |
17 | #ifdef FIO_HAVE_SPLICE |
18 | ||
2866c82d | 19 | struct spliceio_data { |
2866c82d | 20 | int pipe[2]; |
f24254e1 | 21 | int vmsplice_to_user; |
2866c82d JA |
22 | }; |
23 | ||
2866c82d | 24 | /* |
f24254e1 JA |
25 | * vmsplice didn't use to support splicing to user space, this is the old |
26 | * variant of getting that job done. Doesn't make a lot of sense, but it | |
27 | * uses splices to move data from the source into a pipe. | |
2866c82d | 28 | */ |
f24254e1 | 29 | static int fio_splice_read_old(struct thread_data *td, struct io_u *io_u) |
2866c82d JA |
30 | { |
31 | struct spliceio_data *sd = td->io_ops->data; | |
53cdc686 | 32 | struct fio_file *f = io_u->file; |
2866c82d JA |
33 | int ret, ret2, buflen; |
34 | off_t offset; | |
35 | void *p; | |
36 | ||
37 | offset = io_u->offset; | |
cec6b55d JA |
38 | buflen = io_u->xfer_buflen; |
39 | p = io_u->xfer_buf; | |
2866c82d JA |
40 | while (buflen) { |
41 | int this_len = buflen; | |
42 | ||
43 | if (this_len > SPLICE_DEF_SIZE) | |
44 | this_len = SPLICE_DEF_SIZE; | |
45 | ||
53cdc686 | 46 | ret = splice(f->fd, &offset, sd->pipe[1], NULL, this_len, SPLICE_F_MORE); |
2866c82d JA |
47 | if (ret < 0) { |
48 | if (errno == ENODATA || errno == EAGAIN) | |
49 | continue; | |
50 | ||
b4ba5f30 | 51 | return -errno; |
2866c82d JA |
52 | } |
53 | ||
54 | buflen -= ret; | |
55 | ||
56 | while (ret) { | |
57 | ret2 = read(sd->pipe[0], p, ret); | |
58 | if (ret2 < 0) | |
b4ba5f30 | 59 | return -errno; |
2866c82d JA |
60 | |
61 | ret -= ret2; | |
62 | p += ret2; | |
63 | } | |
64 | } | |
65 | ||
cec6b55d | 66 | return io_u->xfer_buflen; |
2866c82d JA |
67 | } |
68 | ||
a26fc7c5 JA |
69 | static void splice_unmap_io_u(struct thread_data *td, struct io_u *io_u) |
70 | { | |
71 | struct spliceio_data *sd = td->io_ops->data; | |
72 | struct iovec iov = { | |
73 | .iov_base = io_u->xfer_buf, | |
74 | .iov_len = io_u->xfer_buflen, | |
75 | }; | |
76 | ||
77 | vmsplice(sd->pipe[0], &iov, 1, SPLICE_F_UNMAP); | |
78 | } | |
79 | ||
f24254e1 JA |
80 | /* |
81 | * We can now vmsplice into userspace, so do the transfer by splicing into | |
82 | * a pipe and vmsplicing that into userspace. | |
83 | */ | |
84 | static int fio_splice_read(struct thread_data *td, struct io_u *io_u) | |
85 | { | |
86 | struct spliceio_data *sd = td->io_ops->data; | |
87 | struct fio_file *f = io_u->file; | |
88 | struct iovec iov; | |
89 | int ret, buflen; | |
90 | off_t offset; | |
91 | void *p; | |
92 | ||
93 | offset = io_u->offset; | |
94 | buflen = io_u->xfer_buflen; | |
95 | p = io_u->xfer_buf; | |
bb446c11 | 96 | io_u->xfer_buf = NULL; |
f24254e1 JA |
97 | while (buflen) { |
98 | int this_len = buflen; | |
99 | ||
100 | if (this_len > SPLICE_DEF_SIZE) | |
101 | this_len = SPLICE_DEF_SIZE; | |
102 | ||
103 | ret = splice(f->fd, &offset, sd->pipe[1], NULL, this_len, SPLICE_F_MORE); | |
104 | if (ret < 0) { | |
105 | if (errno == ENODATA || errno == EAGAIN) | |
106 | continue; | |
107 | ||
108 | return -errno; | |
109 | } | |
110 | ||
111 | buflen -= ret; | |
112 | iov.iov_base = p; | |
113 | iov.iov_len = ret; | |
114 | p += ret; | |
115 | ||
116 | while (iov.iov_len) { | |
117 | ret = vmsplice(sd->pipe[0], &iov, 1, SPLICE_F_MOVE); | |
118 | if (ret < 0) | |
119 | return -errno; | |
120 | else if (!ret) | |
121 | return -ENODATA; | |
122 | ||
bb446c11 JA |
123 | if (!io_u->xfer_buf) |
124 | io_u->xfer_buf = iov.iov_base; | |
f24254e1 JA |
125 | iov.iov_len -= ret; |
126 | iov.iov_base += ret; | |
127 | } | |
128 | } | |
129 | ||
bb446c11 | 130 | io_u->unmap = splice_unmap_io_u; |
f24254e1 JA |
131 | return io_u->xfer_buflen; |
132 | } | |
133 | ||
2866c82d JA |
134 | /* |
135 | * For splice writing, we can vmsplice our data buffer directly into a | |
136 | * pipe and then splice that to a file. | |
137 | */ | |
138 | static int fio_splice_write(struct thread_data *td, struct io_u *io_u) | |
139 | { | |
140 | struct spliceio_data *sd = td->io_ops->data; | |
f24254e1 JA |
141 | struct iovec iov = { |
142 | .iov_base = io_u->xfer_buf, | |
143 | .iov_len = io_u->xfer_buflen, | |
2866c82d JA |
144 | }; |
145 | struct pollfd pfd = { .fd = sd->pipe[1], .events = POLLOUT, }; | |
53cdc686 | 146 | struct fio_file *f = io_u->file; |
2866c82d JA |
147 | off_t off = io_u->offset; |
148 | int ret, ret2; | |
149 | ||
f24254e1 | 150 | while (iov.iov_len) { |
2866c82d JA |
151 | if (poll(&pfd, 1, -1) < 0) |
152 | return errno; | |
153 | ||
f24254e1 | 154 | ret = vmsplice(sd->pipe[1], &iov, 1, SPLICE_F_NONBLOCK); |
2866c82d | 155 | if (ret < 0) |
b4ba5f30 | 156 | return -errno; |
2866c82d | 157 | |
f24254e1 JA |
158 | iov.iov_len -= ret; |
159 | iov.iov_base += ret; | |
2866c82d JA |
160 | |
161 | while (ret) { | |
53cdc686 | 162 | ret2 = splice(sd->pipe[0], NULL, f->fd, &off, ret, 0); |
2866c82d | 163 | if (ret2 < 0) |
b4ba5f30 | 164 | return -errno; |
2866c82d JA |
165 | |
166 | ret -= ret2; | |
167 | } | |
168 | } | |
169 | ||
cec6b55d | 170 | return io_u->xfer_buflen; |
2866c82d JA |
171 | } |
172 | ||
173 | static int fio_spliceio_queue(struct thread_data *td, struct io_u *io_u) | |
174 | { | |
f24254e1 | 175 | struct spliceio_data *sd = td->io_ops->data; |
cec6b55d | 176 | int ret; |
2866c82d | 177 | |
f24254e1 | 178 | if (io_u->ddir == DDIR_READ) { |
43688575 | 179 | if (sd->vmsplice_to_user) { |
f24254e1 | 180 | ret = fio_splice_read(td, io_u); |
43688575 JA |
181 | /* |
182 | * This kernel doesn't support vmsplice to user | |
183 | * space. Reset the vmsplice_to_user flag, so that | |
184 | * we retry below and don't hit this path again. | |
185 | */ | |
186 | if (ret == -EBADF) | |
187 | sd->vmsplice_to_user = 0; | |
188 | } | |
189 | if (!sd->vmsplice_to_user) | |
f24254e1 JA |
190 | ret = fio_splice_read_old(td, io_u); |
191 | } else if (io_u->ddir == DDIR_WRITE) | |
2866c82d | 192 | ret = fio_splice_write(td, io_u); |
87dc1ab1 JA |
193 | else |
194 | ret = fsync(io_u->file->fd); | |
2866c82d | 195 | |
cec6b55d | 196 | if (ret != (int) io_u->xfer_buflen) { |
22819ec2 | 197 | if (ret >= 0) { |
cec6b55d JA |
198 | io_u->resid = io_u->xfer_buflen - ret; |
199 | io_u->error = 0; | |
36167d82 | 200 | return FIO_Q_COMPLETED; |
2866c82d JA |
201 | } else |
202 | io_u->error = errno; | |
203 | } | |
204 | ||
36167d82 | 205 | if (io_u->error) |
e1161c32 | 206 | td_verror(td, io_u->error, "xfer"); |
2866c82d | 207 | |
36167d82 | 208 | return FIO_Q_COMPLETED; |
2866c82d JA |
209 | } |
210 | ||
211 | static void fio_spliceio_cleanup(struct thread_data *td) | |
212 | { | |
213 | struct spliceio_data *sd = td->io_ops->data; | |
214 | ||
215 | if (sd) { | |
216 | close(sd->pipe[0]); | |
217 | close(sd->pipe[1]); | |
218 | free(sd); | |
219 | td->io_ops->data = NULL; | |
220 | } | |
221 | } | |
222 | ||
223 | static int fio_spliceio_init(struct thread_data *td) | |
224 | { | |
225 | struct spliceio_data *sd = malloc(sizeof(*sd)); | |
226 | ||
2866c82d | 227 | if (pipe(sd->pipe) < 0) { |
e1161c32 | 228 | td_verror(td, errno, "pipe"); |
2866c82d JA |
229 | free(sd); |
230 | return 1; | |
231 | } | |
232 | ||
f24254e1 | 233 | /* |
43688575 | 234 | * Assume this work, we'll reset this if it doesn't |
f24254e1 | 235 | */ |
43688575 | 236 | sd->vmsplice_to_user = 1; |
f24254e1 | 237 | |
2866c82d JA |
238 | td->io_ops->data = sd; |
239 | return 0; | |
240 | } | |
241 | ||
5f350952 | 242 | static struct ioengine_ops ioengine = { |
2866c82d JA |
243 | .name = "splice", |
244 | .version = FIO_IOOPS_VERSION, | |
245 | .init = fio_spliceio_init, | |
246 | .queue = fio_spliceio_queue, | |
2866c82d | 247 | .cleanup = fio_spliceio_cleanup, |
b5af8293 JA |
248 | .open_file = generic_open_file, |
249 | .close_file = generic_close_file, | |
2866c82d JA |
250 | .flags = FIO_SYNCIO, |
251 | }; | |
34cfcdaf JA |
252 | |
253 | #else /* FIO_HAVE_SPLICE */ | |
254 | ||
255 | /* | |
256 | * When we have a proper configure system in place, we simply wont build | |
257 | * and install this io engine. For now install a crippled version that | |
258 | * just complains and fails to load. | |
259 | */ | |
260 | static int fio_spliceio_init(struct thread_data fio_unused *td) | |
261 | { | |
262 | fprintf(stderr, "fio: splice not available\n"); | |
263 | return 1; | |
264 | } | |
265 | ||
5f350952 | 266 | static struct ioengine_ops ioengine = { |
34cfcdaf JA |
267 | .name = "splice", |
268 | .version = FIO_IOOPS_VERSION, | |
269 | .init = fio_spliceio_init, | |
270 | }; | |
271 | ||
272 | #endif | |
5f350952 JA |
273 | |
274 | static void fio_init fio_spliceio_register(void) | |
275 | { | |
276 | register_ioengine(&ioengine); | |
277 | } | |
278 | ||
279 | static void fio_exit fio_spliceio_unregister(void) | |
280 | { | |
281 | unregister_ioengine(&ioengine); | |
282 | } |