Commit | Line | Data |
---|---|---|
2866c82d JA |
1 | /* |
2 | * splice io engine | |
3 | * | |
4 | */ | |
5 | #include <stdio.h> | |
6 | #include <stdlib.h> | |
7 | #include <unistd.h> | |
8 | #include <errno.h> | |
9 | #include <assert.h> | |
10 | #include <sys/poll.h> | |
11 | #include "fio.h" | |
12 | #include "os.h" | |
13 | ||
14 | struct spliceio_data { | |
15 | struct io_u *last_io_u; | |
16 | int pipe[2]; | |
17 | }; | |
18 | ||
7a16dd02 JA |
19 | static int fio_spliceio_sync(struct thread_data fio_unused *td, |
20 | struct fio_file *f) | |
2866c82d | 21 | { |
53cdc686 | 22 | return fsync(f->fd); |
2866c82d JA |
23 | } |
24 | ||
25 | static int fio_spliceio_getevents(struct thread_data *td, int fio_unused min, | |
26 | int max, struct timespec fio_unused *t) | |
27 | { | |
28 | assert(max <= 1); | |
29 | ||
30 | /* | |
31 | * we can only have one finished io_u for sync io, since the depth | |
32 | * is always 1 | |
33 | */ | |
34 | if (list_empty(&td->io_u_busylist)) | |
35 | return 0; | |
36 | ||
37 | return 1; | |
38 | } | |
39 | ||
40 | static struct io_u *fio_spliceio_event(struct thread_data *td, int event) | |
41 | { | |
42 | struct spliceio_data *sd = td->io_ops->data; | |
43 | ||
44 | assert(event == 0); | |
45 | ||
46 | return sd->last_io_u; | |
47 | } | |
48 | ||
49 | /* | |
50 | * For splice reading, we unfortunately cannot (yet) vmsplice the other way. | |
51 | * So just splice the data from the file into the pipe, and use regular | |
52 | * read to fill the buffer. Doesn't make a lot of sense, but... | |
53 | */ | |
54 | static int fio_splice_read(struct thread_data *td, struct io_u *io_u) | |
55 | { | |
56 | struct spliceio_data *sd = td->io_ops->data; | |
53cdc686 | 57 | struct fio_file *f = io_u->file; |
2866c82d JA |
58 | int ret, ret2, buflen; |
59 | off_t offset; | |
60 | void *p; | |
61 | ||
62 | offset = io_u->offset; | |
63 | buflen = io_u->buflen; | |
64 | p = io_u->buf; | |
65 | while (buflen) { | |
66 | int this_len = buflen; | |
67 | ||
68 | if (this_len > SPLICE_DEF_SIZE) | |
69 | this_len = SPLICE_DEF_SIZE; | |
70 | ||
53cdc686 | 71 | ret = splice(f->fd, &offset, sd->pipe[1], NULL, this_len, SPLICE_F_MORE); |
2866c82d JA |
72 | if (ret < 0) { |
73 | if (errno == ENODATA || errno == EAGAIN) | |
74 | continue; | |
75 | ||
76 | return errno; | |
77 | } | |
78 | ||
79 | buflen -= ret; | |
80 | ||
81 | while (ret) { | |
82 | ret2 = read(sd->pipe[0], p, ret); | |
83 | if (ret2 < 0) | |
84 | return errno; | |
85 | ||
86 | ret -= ret2; | |
87 | p += ret2; | |
88 | } | |
89 | } | |
90 | ||
91 | return io_u->buflen; | |
92 | } | |
93 | ||
94 | /* | |
95 | * For splice writing, we can vmsplice our data buffer directly into a | |
96 | * pipe and then splice that to a file. | |
97 | */ | |
98 | static int fio_splice_write(struct thread_data *td, struct io_u *io_u) | |
99 | { | |
100 | struct spliceio_data *sd = td->io_ops->data; | |
101 | struct iovec iov[1] = { | |
102 | { | |
103 | .iov_base = io_u->buf, | |
104 | .iov_len = io_u->buflen, | |
105 | } | |
106 | }; | |
107 | struct pollfd pfd = { .fd = sd->pipe[1], .events = POLLOUT, }; | |
53cdc686 | 108 | struct fio_file *f = io_u->file; |
2866c82d JA |
109 | off_t off = io_u->offset; |
110 | int ret, ret2; | |
111 | ||
112 | while (iov[0].iov_len) { | |
113 | if (poll(&pfd, 1, -1) < 0) | |
114 | return errno; | |
115 | ||
116 | ret = vmsplice(sd->pipe[1], iov, 1, SPLICE_F_NONBLOCK); | |
117 | if (ret < 0) | |
118 | return errno; | |
119 | ||
120 | iov[0].iov_len -= ret; | |
121 | iov[0].iov_base += ret; | |
122 | ||
123 | while (ret) { | |
53cdc686 | 124 | ret2 = splice(sd->pipe[0], NULL, f->fd, &off, ret, 0); |
2866c82d JA |
125 | if (ret2 < 0) |
126 | return errno; | |
127 | ||
128 | ret -= ret2; | |
129 | } | |
130 | } | |
131 | ||
132 | return io_u->buflen; | |
133 | } | |
134 | ||
135 | static int fio_spliceio_queue(struct thread_data *td, struct io_u *io_u) | |
136 | { | |
137 | struct spliceio_data *sd = td->io_ops->data; | |
138 | int ret; | |
139 | ||
140 | if (io_u->ddir == DDIR_READ) | |
141 | ret = fio_splice_read(td, io_u); | |
142 | else | |
143 | ret = fio_splice_write(td, io_u); | |
144 | ||
145 | if ((unsigned int) ret != io_u->buflen) { | |
146 | if (ret > 0) { | |
147 | io_u->resid = io_u->buflen - ret; | |
148 | io_u->error = ENODATA; | |
149 | } else | |
150 | io_u->error = errno; | |
151 | } | |
152 | ||
153 | if (!io_u->error) | |
154 | sd->last_io_u = io_u; | |
155 | ||
156 | return io_u->error; | |
157 | } | |
158 | ||
159 | static void fio_spliceio_cleanup(struct thread_data *td) | |
160 | { | |
161 | struct spliceio_data *sd = td->io_ops->data; | |
162 | ||
163 | if (sd) { | |
164 | close(sd->pipe[0]); | |
165 | close(sd->pipe[1]); | |
166 | free(sd); | |
167 | td->io_ops->data = NULL; | |
168 | } | |
169 | } | |
170 | ||
171 | static int fio_spliceio_init(struct thread_data *td) | |
172 | { | |
173 | struct spliceio_data *sd = malloc(sizeof(*sd)); | |
174 | ||
175 | sd->last_io_u = NULL; | |
176 | if (pipe(sd->pipe) < 0) { | |
177 | td_verror(td, errno); | |
178 | free(sd); | |
179 | return 1; | |
180 | } | |
181 | ||
182 | td->io_ops->data = sd; | |
183 | return 0; | |
184 | } | |
185 | ||
186 | struct ioengine_ops ioengine = { | |
187 | .name = "splice", | |
188 | .version = FIO_IOOPS_VERSION, | |
189 | .init = fio_spliceio_init, | |
190 | .queue = fio_spliceio_queue, | |
191 | .getevents = fio_spliceio_getevents, | |
192 | .event = fio_spliceio_event, | |
193 | .cleanup = fio_spliceio_cleanup, | |
194 | .sync = fio_spliceio_sync, | |
195 | .flags = FIO_SYNCIO, | |
196 | }; |