| 1 | /* |
| 2 | * binject engine |
| 3 | * |
| 4 | * IO engine that uses the Linux binject interface to directly inject |
| 5 | * bio's to block devices. |
| 6 | * |
| 7 | */ |
| 8 | #include <stdio.h> |
| 9 | #include <stdlib.h> |
| 10 | #include <unistd.h> |
| 11 | #include <errno.h> |
| 12 | #include <assert.h> |
| 13 | #include <string.h> |
| 14 | #include <sys/poll.h> |
| 15 | |
| 16 | #include "../fio.h" |
| 17 | |
| 18 | #ifdef FIO_HAVE_BINJECT |
| 19 | |
| 20 | struct binject_data { |
| 21 | struct b_user_cmd *cmds; |
| 22 | struct io_u **events; |
| 23 | struct pollfd *pfds; |
| 24 | int *fd_flags; |
| 25 | }; |
| 26 | |
| 27 | static void binject_buc_init(struct binject_data *bd, struct io_u *io_u) |
| 28 | { |
| 29 | struct b_user_cmd *buc = &io_u->buc; |
| 30 | |
| 31 | memset(buc, 0, sizeof(*buc)); |
| 32 | binject_buc_set_magic(buc); |
| 33 | |
| 34 | buc->buf = (unsigned long) io_u->xfer_buf; |
| 35 | buc->len = io_u->xfer_buflen; |
| 36 | buc->offset = io_u->offset; |
| 37 | buc->usr_ptr = (unsigned long) io_u; |
| 38 | |
| 39 | buc->flags = B_FLAG_NOIDLE | B_FLAG_UNPLUG; |
| 40 | assert(buc->buf); |
| 41 | } |
| 42 | |
| 43 | static int pollin_events(struct pollfd *pfds, int fds) |
| 44 | { |
| 45 | int i; |
| 46 | |
| 47 | for (i = 0; i < fds; i++) |
| 48 | if (pfds[i].revents & POLLIN) |
| 49 | return 1; |
| 50 | |
| 51 | return 0; |
| 52 | } |
| 53 | |
| 54 | static int fio_binject_getevents(struct thread_data *td, unsigned int min, |
| 55 | unsigned int max, struct timespec fio_unused *t) |
| 56 | { |
| 57 | struct binject_data *bd = td->io_ops->data; |
| 58 | int left = max, ret, r = 0, ev_index = 0; |
| 59 | void *buf = bd->cmds; |
| 60 | unsigned int i, events; |
| 61 | struct fio_file *f; |
| 62 | |
| 63 | /* |
| 64 | * Fill in the file descriptors |
| 65 | */ |
| 66 | for_each_file(td, f, i) { |
| 67 | /* |
| 68 | * don't block for min events == 0 |
| 69 | */ |
| 70 | if (!min) { |
| 71 | bd->fd_flags[i] = fcntl(f->fd, F_GETFL); |
| 72 | fcntl(f->fd, F_SETFL, bd->fd_flags[i] | O_NONBLOCK); |
| 73 | } |
| 74 | bd->pfds[i].fd = f->fd; |
| 75 | bd->pfds[i].events = POLLIN; |
| 76 | } |
| 77 | |
| 78 | while (left) { |
| 79 | void *p; |
| 80 | |
| 81 | do { |
| 82 | if (!min) |
| 83 | break; |
| 84 | |
| 85 | ret = poll(bd->pfds, td->o.nr_files, -1); |
| 86 | if (ret < 0) { |
| 87 | if (!r) |
| 88 | r = -errno; |
| 89 | td_verror(td, errno, "poll"); |
| 90 | break; |
| 91 | } else if (!ret) |
| 92 | continue; |
| 93 | |
| 94 | if (pollin_events(bd->pfds, td->o.nr_files)) |
| 95 | break; |
| 96 | } while (1); |
| 97 | |
| 98 | if (r < 0) |
| 99 | break; |
| 100 | |
| 101 | re_read: |
| 102 | p = buf; |
| 103 | events = 0; |
| 104 | for_each_file(td, f, i) { |
| 105 | ret = read(f->fd, p, left * sizeof(struct b_user_cmd)); |
| 106 | if (ret < 0) { |
| 107 | if (errno == EAGAIN) |
| 108 | continue; |
| 109 | r = -errno; |
| 110 | td_verror(td, errno, "read"); |
| 111 | break; |
| 112 | } else if (ret) { |
| 113 | p += ret; |
| 114 | events += ret / sizeof(struct b_user_cmd); |
| 115 | } |
| 116 | } |
| 117 | |
| 118 | if (r < 0) |
| 119 | break; |
| 120 | if (!events) { |
| 121 | usleep(1000); |
| 122 | goto re_read; |
| 123 | } |
| 124 | |
| 125 | left -= events; |
| 126 | r += events; |
| 127 | |
| 128 | for (i = 0; i < events; i++) { |
| 129 | struct b_user_cmd *buc = (struct b_user_cmd *) buf + i; |
| 130 | |
| 131 | bd->events[ev_index] = (struct io_u *) buc->usr_ptr; |
| 132 | ev_index++; |
| 133 | } |
| 134 | } |
| 135 | |
| 136 | if (!min) { |
| 137 | for_each_file(td, f, i) |
| 138 | fcntl(f->fd, F_SETFL, bd->fd_flags[i]); |
| 139 | } |
| 140 | |
| 141 | if (r > 0) |
| 142 | assert(ev_index == r); |
| 143 | |
| 144 | return r; |
| 145 | } |
| 146 | |
| 147 | static int fio_binject_doio(struct thread_data *td, struct io_u *io_u) |
| 148 | { |
| 149 | struct b_user_cmd *buc = &io_u->buc; |
| 150 | struct fio_file *f = io_u->file; |
| 151 | int ret; |
| 152 | |
| 153 | ret = write(f->fd, buc, sizeof(*buc)); |
| 154 | if (ret < 0) |
| 155 | return ret; |
| 156 | |
| 157 | return FIO_Q_QUEUED; |
| 158 | } |
| 159 | |
| 160 | static int fio_binject_prep(struct thread_data *td, struct io_u *io_u) |
| 161 | { |
| 162 | struct binject_data *bd = td->io_ops->data; |
| 163 | unsigned int bs = io_u->file->file_data; |
| 164 | struct b_user_cmd *buc = &io_u->buc; |
| 165 | |
| 166 | if (io_u->xfer_buflen & (bs - 1)) { |
| 167 | log_err("read/write not sector aligned\n"); |
| 168 | return EINVAL; |
| 169 | } |
| 170 | |
| 171 | if (io_u->ddir == DDIR_READ) { |
| 172 | binject_buc_init(bd, io_u); |
| 173 | buc->type = B_TYPE_READ; |
| 174 | } else if (io_u->ddir == DDIR_WRITE) { |
| 175 | binject_buc_init(bd, io_u); |
| 176 | buc->type = B_TYPE_WRITE; |
| 177 | } else if (io_u->ddir == DDIR_TRIM) { |
| 178 | binject_buc_init(bd, io_u); |
| 179 | buc->type = B_TYPE_DISCARD; |
| 180 | } else { |
| 181 | assert(0); |
| 182 | } |
| 183 | |
| 184 | return 0; |
| 185 | } |
| 186 | |
| 187 | static int fio_binject_queue(struct thread_data *td, struct io_u *io_u) |
| 188 | { |
| 189 | int ret; |
| 190 | |
| 191 | fio_ro_check(td, io_u); |
| 192 | |
| 193 | ret = fio_binject_doio(td, io_u); |
| 194 | |
| 195 | if (ret < 0) |
| 196 | io_u->error = errno; |
| 197 | |
| 198 | if (io_u->error) { |
| 199 | td_verror(td, io_u->error, "xfer"); |
| 200 | return FIO_Q_COMPLETED; |
| 201 | } |
| 202 | |
| 203 | return ret; |
| 204 | } |
| 205 | |
| 206 | static struct io_u *fio_binject_event(struct thread_data *td, int event) |
| 207 | { |
| 208 | struct binject_data *bd = td->io_ops->data; |
| 209 | |
| 210 | return bd->events[event]; |
| 211 | } |
| 212 | |
| 213 | static int fio_binject_open_file(struct thread_data *td, struct fio_file *f) |
| 214 | { |
| 215 | unsigned int bs; |
| 216 | int ret; |
| 217 | |
| 218 | ret = generic_open_file(td, f); |
| 219 | if (ret) |
| 220 | return 1; |
| 221 | |
| 222 | if (f->filetype != FIO_TYPE_BD) { |
| 223 | log_err("fio: binject only works with block devices\n"); |
| 224 | return 1; |
| 225 | } |
| 226 | if (ioctl(f->fd, BLKSSZGET, &bs) < 0) { |
| 227 | td_verror(td, errno, "BLKSSZGET"); |
| 228 | return 1; |
| 229 | } |
| 230 | |
| 231 | f->file_data = bs; |
| 232 | return 0; |
| 233 | } |
| 234 | |
| 235 | static void fio_binject_cleanup(struct thread_data *td) |
| 236 | { |
| 237 | struct binject_data *bd = td->io_ops->data; |
| 238 | |
| 239 | if (bd) { |
| 240 | free(bd->events); |
| 241 | free(bd->cmds); |
| 242 | free(bd->fd_flags); |
| 243 | free(bd->pfds); |
| 244 | free(bd); |
| 245 | } |
| 246 | } |
| 247 | |
| 248 | static int fio_binject_init(struct thread_data *td) |
| 249 | { |
| 250 | struct binject_data *bd; |
| 251 | |
| 252 | bd = malloc(sizeof(*bd)); |
| 253 | memset(bd, 0, sizeof(*bd)); |
| 254 | |
| 255 | bd->cmds = malloc(td->o.iodepth * sizeof(struct b_user_cmd)); |
| 256 | memset(bd->cmds, 0, td->o.iodepth * sizeof(struct b_user_cmd)); |
| 257 | |
| 258 | bd->events = malloc(td->o.iodepth * sizeof(struct io_u *)); |
| 259 | memset(bd->events, 0, td->o.iodepth * sizeof(struct io_u *)); |
| 260 | |
| 261 | bd->pfds = malloc(sizeof(struct pollfd) * td->o.nr_files); |
| 262 | memset(bd->pfds, 0, sizeof(struct pollfd) * td->o.nr_files); |
| 263 | |
| 264 | bd->fd_flags = malloc(sizeof(int) * td->o.nr_files); |
| 265 | memset(bd->fd_flags, 0, sizeof(int) * td->o.nr_files); |
| 266 | |
| 267 | td->io_ops->data = bd; |
| 268 | return 0; |
| 269 | } |
| 270 | |
| 271 | static struct ioengine_ops ioengine = { |
| 272 | .name = "binject", |
| 273 | .version = FIO_IOOPS_VERSION, |
| 274 | .init = fio_binject_init, |
| 275 | .prep = fio_binject_prep, |
| 276 | .queue = fio_binject_queue, |
| 277 | .getevents = fio_binject_getevents, |
| 278 | .event = fio_binject_event, |
| 279 | .cleanup = fio_binject_cleanup, |
| 280 | .open_file = fio_binject_open_file, |
| 281 | .close_file = generic_close_file, |
| 282 | .get_file_size = generic_get_file_size, |
| 283 | .flags = FIO_RAWIO, |
| 284 | }; |
| 285 | |
| 286 | #else /* FIO_HAVE_BINJECT */ |
| 287 | |
| 288 | /* |
| 289 | * When we have a proper configure system in place, we simply wont build |
| 290 | * and install this io engine. For now install a crippled version that |
| 291 | * just complains and fails to load. |
| 292 | */ |
| 293 | static int fio_binject_init(struct thread_data fio_unused *td) |
| 294 | { |
| 295 | fprintf(stderr, "fio: ioengine binject not available\n"); |
| 296 | return 1; |
| 297 | } |
| 298 | |
| 299 | static struct ioengine_ops ioengine = { |
| 300 | .name = "binject", |
| 301 | .version = FIO_IOOPS_VERSION, |
| 302 | .init = fio_binject_init, |
| 303 | }; |
| 304 | |
| 305 | #endif |
| 306 | |
| 307 | static void fio_init fio_binject_register(void) |
| 308 | { |
| 309 | register_ioengine(&ioengine); |
| 310 | } |
| 311 | |
| 312 | static void fio_exit fio_binject_unregister(void) |
| 313 | { |
| 314 | unregister_ioengine(&ioengine); |
| 315 | } |