| 1 | #include <stdlib.h> |
| 2 | #include <poll.h> |
| 3 | #include <nfsc/libnfs.h> |
| 4 | #include <nfsc/libnfs-raw.h> |
| 5 | #include <nfsc/libnfs-raw-mount.h> |
| 6 | |
| 7 | #include "../fio.h" |
| 8 | #include "../optgroup.h" |
| 9 | |
| 10 | enum nfs_op_type { |
| 11 | NFS_READ_WRITE = 0, |
| 12 | NFS_STAT_MKDIR_RMDIR, |
| 13 | NFS_STAT_TOUCH_RM, |
| 14 | }; |
| 15 | |
| 16 | struct fio_libnfs_options { |
| 17 | struct nfs_context *context; |
| 18 | char *nfs_url; |
| 19 | unsigned int queue_depth; /* nfs_callback needs this info, but doesn't have fio td structure to pull it from */ |
| 20 | /* the following implement a circular queue of outstanding IOs */ |
| 21 | int outstanding_events; /* IOs issued to libnfs, that have not returned yet */ |
| 22 | int prev_requested_event_index; /* event last returned via fio_libnfs_event */ |
| 23 | int next_buffered_event; /* round robin-pointer within events[] */ |
| 24 | int buffered_event_count; /* IOs completed by libnfs, waiting for FIO */ |
| 25 | int free_event_buffer_index; /* next free buffer */ |
| 26 | struct io_u**events; |
| 27 | }; |
| 28 | |
| 29 | struct nfs_data { |
| 30 | struct nfsfh *nfsfh; |
| 31 | struct fio_libnfs_options *options; |
| 32 | }; |
| 33 | |
| 34 | static struct fio_option options[] = { |
| 35 | { |
| 36 | .name = "nfs_url", |
| 37 | .lname = "nfs_url", |
| 38 | .type = FIO_OPT_STR_STORE, |
| 39 | .help = "URL in libnfs format, eg nfs://<server|ipv4|ipv6>/path[?arg=val[&arg=val]*]", |
| 40 | .off1 = offsetof(struct fio_libnfs_options, nfs_url), |
| 41 | .category = FIO_OPT_C_ENGINE, |
| 42 | .group = __FIO_OPT_G_NFS, |
| 43 | }, |
| 44 | { |
| 45 | .name = NULL, |
| 46 | }, |
| 47 | }; |
| 48 | |
| 49 | static struct io_u *fio_libnfs_event(struct thread_data *td, int event) |
| 50 | { |
| 51 | struct fio_libnfs_options *o = td->eo; |
| 52 | struct io_u *io_u = o->events[o->next_buffered_event]; |
| 53 | assert(o->events[o->next_buffered_event]); |
| 54 | o->events[o->next_buffered_event] = NULL; |
| 55 | o->next_buffered_event = (o->next_buffered_event + 1) % td->o.iodepth; |
| 56 | /* validate our state machine */ |
| 57 | assert(o->buffered_event_count); |
| 58 | o->buffered_event_count--; |
| 59 | assert(io_u); |
| 60 | /* assert that fio_libnfs_event is being called in sequential fashion */ |
| 61 | assert(event == 0 || o->prev_requested_event_index + 1 == event); |
| 62 | if (o->buffered_event_count == 0) { |
| 63 | o->prev_requested_event_index = -1; |
| 64 | } else { |
| 65 | o->prev_requested_event_index = event; |
| 66 | } |
| 67 | return io_u; |
| 68 | } |
| 69 | |
| 70 | static int nfs_event_loop(struct thread_data *td, bool flush) { |
| 71 | struct fio_libnfs_options *o = td->eo; |
| 72 | struct pollfd pfds[1]; /* nfs:0 */ |
| 73 | /* we already have stuff queued for fio, no need to waste cpu on poll() */ |
| 74 | if (o->buffered_event_count) |
| 75 | return o->buffered_event_count; |
| 76 | /* fio core logic seems to stop calling this event-loop if we ever return with 0 events */ |
| 77 | #define SHOULD_WAIT() (o->outstanding_events == td->o.iodepth || (flush && o->outstanding_events)) |
| 78 | |
| 79 | do { |
| 80 | int timeout = SHOULD_WAIT() ? -1 : 0; |
| 81 | int ret = 0; |
| 82 | pfds[0].fd = nfs_get_fd(o->context); |
| 83 | pfds[0].events = nfs_which_events(o->context); |
| 84 | ret = poll(&pfds[0], 1, timeout); |
| 85 | if (ret < 0) { |
| 86 | if (errno == EINTR || errno == EAGAIN) { |
| 87 | continue; |
| 88 | } |
| 89 | log_err("nfs: failed to poll events: %s.\n", |
| 90 | strerror(errno)); |
| 91 | break; |
| 92 | } |
| 93 | |
| 94 | ret = nfs_service(o->context, pfds[0].revents); |
| 95 | if (ret < 0) { |
| 96 | log_err("nfs: socket is in an unrecoverable error state.\n"); |
| 97 | break; |
| 98 | } |
| 99 | } while (SHOULD_WAIT()); |
| 100 | return o->buffered_event_count; |
| 101 | #undef SHOULD_WAIT |
| 102 | } |
| 103 | |
| 104 | static int fio_libnfs_getevents(struct thread_data *td, unsigned int min, |
| 105 | unsigned int max, const struct timespec *t) |
| 106 | { |
| 107 | return nfs_event_loop(td, false); |
| 108 | } |
| 109 | |
| 110 | static void nfs_callback(int res, struct nfs_context *nfs, void *data, |
| 111 | void *private_data) |
| 112 | { |
| 113 | struct io_u *io_u = private_data; |
| 114 | struct nfs_data *nfs_data = io_u->file->engine_data; |
| 115 | struct fio_libnfs_options *o = nfs_data->options; |
| 116 | if (res < 0) { |
| 117 | log_err("Failed NFS operation(code:%d): %s\n", res, nfs_get_error(o->context)); |
| 118 | io_u->error = -res; |
| 119 | /* res is used for read math below, don't wanna pass negative there */ |
| 120 | res = 0; |
| 121 | } else if (io_u->ddir == DDIR_READ) { |
| 122 | memcpy(io_u->buf, data, res); |
| 123 | if (res == 0) |
| 124 | log_err("Got NFS EOF, this is probably not expected\n"); |
| 125 | } |
| 126 | /* fio uses resid to track remaining data */ |
| 127 | io_u->resid = io_u->xfer_buflen - res; |
| 128 | |
| 129 | assert(!o->events[o->free_event_buffer_index]); |
| 130 | o->events[o->free_event_buffer_index] = io_u; |
| 131 | o->free_event_buffer_index = (o->free_event_buffer_index + 1) % o->queue_depth; |
| 132 | o->outstanding_events--; |
| 133 | o->buffered_event_count++; |
| 134 | } |
| 135 | |
| 136 | static int queue_write(struct fio_libnfs_options *o, struct io_u *io_u) { |
| 137 | struct nfs_data *nfs_data = io_u->engine_data; |
| 138 | return nfs_pwrite_async(o->context, nfs_data->nfsfh, |
| 139 | io_u->offset, io_u->buflen, io_u->buf, nfs_callback, |
| 140 | io_u); |
| 141 | } |
| 142 | |
| 143 | static int queue_read(struct fio_libnfs_options *o, struct io_u *io_u) { |
| 144 | struct nfs_data *nfs_data = io_u->engine_data; |
| 145 | return nfs_pread_async(o->context, nfs_data->nfsfh, io_u->offset, io_u->buflen, nfs_callback, io_u); |
| 146 | } |
| 147 | |
| 148 | static enum fio_q_status fio_libnfs_queue(struct thread_data *td, |
| 149 | struct io_u *io_u) |
| 150 | { |
| 151 | struct nfs_data *nfs_data = io_u->file->engine_data; |
| 152 | struct fio_libnfs_options *o = nfs_data->options; |
| 153 | struct nfs_context *nfs = o->context; |
| 154 | int err; |
| 155 | enum fio_q_status ret = FIO_Q_QUEUED; |
| 156 | |
| 157 | io_u->engine_data = nfs_data; |
| 158 | switch(io_u->ddir) { |
| 159 | case DDIR_WRITE: |
| 160 | err = queue_write(o, io_u); |
| 161 | break; |
| 162 | case DDIR_READ: |
| 163 | err = queue_read(o, io_u); |
| 164 | break; |
| 165 | case DDIR_TRIM: |
| 166 | log_err("nfs: trim is not supported"); |
| 167 | err = -1; |
| 168 | break; |
| 169 | default: |
| 170 | log_err("nfs: unhandled io %d\n", io_u->ddir); |
| 171 | err = -1; |
| 172 | } |
| 173 | if (err) { |
| 174 | log_err("nfs: Failed to queue nfs op: %s\n", nfs_get_error(nfs)); |
| 175 | td->error = 1; |
| 176 | return FIO_Q_COMPLETED; |
| 177 | } |
| 178 | o->outstanding_events++; |
| 179 | return ret; |
| 180 | } |
| 181 | |
| 182 | /* |
| 183 | * Do a mount if one has not been done before |
| 184 | */ |
| 185 | static int do_mount(struct thread_data *td, const char *url) |
| 186 | { |
| 187 | size_t event_size = sizeof(struct io_u **) * td->o.iodepth; |
| 188 | struct fio_libnfs_options *options = td->eo; |
| 189 | struct nfs_url *nfs_url = NULL; |
| 190 | int ret = 0; |
| 191 | int path_len = 0; |
| 192 | char *mnt_dir = NULL; |
| 193 | |
| 194 | if (options->context) |
| 195 | return 0; |
| 196 | |
| 197 | options->context = nfs_init_context(); |
| 198 | if (options->context == NULL) { |
| 199 | log_err("nfs: failed to init nfs context\n"); |
| 200 | return -1; |
| 201 | } |
| 202 | |
| 203 | options->events = malloc(event_size); |
| 204 | memset(options->events, 0, event_size); |
| 205 | |
| 206 | options->prev_requested_event_index = -1; |
| 207 | options->queue_depth = td->o.iodepth; |
| 208 | |
| 209 | nfs_url = nfs_parse_url_full(options->context, url); |
| 210 | path_len = strlen(nfs_url->path); |
| 211 | mnt_dir = malloc(path_len + strlen(nfs_url->file) + 1); |
| 212 | strcpy(mnt_dir, nfs_url->path); |
| 213 | strcpy(mnt_dir + strlen(nfs_url->path), nfs_url->file); |
| 214 | ret = nfs_mount(options->context, nfs_url->server, mnt_dir); |
| 215 | free(mnt_dir); |
| 216 | nfs_destroy_url(nfs_url); |
| 217 | return ret; |
| 218 | } |
| 219 | |
| 220 | static int fio_libnfs_setup(struct thread_data *td) |
| 221 | { |
| 222 | /* Using threads with libnfs causes fio to hang on exit, lower performance */ |
| 223 | td->o.use_thread = 0; |
| 224 | return 0; |
| 225 | } |
| 226 | |
| 227 | static void fio_libnfs_cleanup(struct thread_data *td) |
| 228 | { |
| 229 | struct fio_libnfs_options *o = td->eo; |
| 230 | nfs_umount(o->context); |
| 231 | nfs_destroy_context(o->context); |
| 232 | free(o->events); |
| 233 | } |
| 234 | |
| 235 | static int fio_libnfs_open(struct thread_data *td, struct fio_file *f) |
| 236 | { |
| 237 | int ret; |
| 238 | struct fio_libnfs_options *options = td->eo; |
| 239 | struct nfs_data *nfs_data = NULL; |
| 240 | int flags = 0; |
| 241 | |
| 242 | if (!options->nfs_url) { |
| 243 | log_err("nfs: nfs_url is a required parameter\n"); |
| 244 | return -1; |
| 245 | } |
| 246 | |
| 247 | ret = do_mount(td, options->nfs_url); |
| 248 | |
| 249 | if (ret != 0) { |
| 250 | log_err("nfs: Failed to mount %s with code %d: %s\n", options->nfs_url, ret, nfs_get_error(options->context)); |
| 251 | return ret; |
| 252 | } |
| 253 | nfs_data = malloc(sizeof(struct nfs_data)); |
| 254 | memset(nfs_data, 0, sizeof(struct nfs_data)); |
| 255 | nfs_data->options = options; |
| 256 | |
| 257 | if (td->o.td_ddir == TD_DDIR_WRITE) { |
| 258 | flags |= O_CREAT | O_RDWR; |
| 259 | } else { |
| 260 | flags |= O_RDWR; |
| 261 | } |
| 262 | ret = nfs_open(options->context, f->file_name, flags, &nfs_data->nfsfh); |
| 263 | |
| 264 | if (ret != 0) |
| 265 | log_err("Failed to open %s: %s\n", f->file_name, nfs_get_error(options->context)); |
| 266 | f->engine_data = nfs_data; |
| 267 | return ret; |
| 268 | } |
| 269 | |
| 270 | static int fio_libnfs_close(struct thread_data *td, struct fio_file *f) |
| 271 | { |
| 272 | struct nfs_data *nfs_data = f->engine_data; |
| 273 | struct fio_libnfs_options *o = nfs_data->options; |
| 274 | int ret = 0; |
| 275 | if (nfs_data->nfsfh) |
| 276 | ret = nfs_close(o->context, nfs_data->nfsfh); |
| 277 | free(nfs_data); |
| 278 | f->engine_data = NULL; |
| 279 | return ret; |
| 280 | } |
| 281 | |
| 282 | /* |
| 283 | * Hook for writing out outstanding data. |
| 284 | */ |
| 285 | static int fio_libnfs_commit(struct thread_data *td) { |
| 286 | nfs_event_loop(td, true); |
| 287 | return 0; |
| 288 | } |
| 289 | |
| 290 | struct ioengine_ops ioengine = { |
| 291 | .name = "nfs", |
| 292 | .version = FIO_IOOPS_VERSION, |
| 293 | .setup = fio_libnfs_setup, |
| 294 | .queue = fio_libnfs_queue, |
| 295 | .getevents = fio_libnfs_getevents, |
| 296 | .event = fio_libnfs_event, |
| 297 | .cleanup = fio_libnfs_cleanup, |
| 298 | .open_file = fio_libnfs_open, |
| 299 | .close_file = fio_libnfs_close, |
| 300 | .commit = fio_libnfs_commit, |
| 301 | .flags = FIO_DISKLESSIO | FIO_NOEXTEND | FIO_NODISKUTIL, |
| 302 | .options = options, |
| 303 | .option_struct_size = sizeof(struct fio_libnfs_options), |
| 304 | }; |
| 305 | |
| 306 | static void fio_init fio_nfs_register(void) |
| 307 | { |
| 308 | register_ioengine(&ioengine); |
| 309 | } |
| 310 | |
| 311 | static void fio_exit fio_nfs_unregister(void) |
| 312 | { |
| 313 | unregister_ioengine(&ioengine); |
| 314 | } |