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