| 1 | #define USE_THE_REPOSITORY_VARIABLE |
| 2 | #define DISABLE_SIGN_COMPARE_WARNINGS |
| 3 | |
| 4 | #include "builtin.h" |
| 5 | #include "abspath.h" |
| 6 | #include "config.h" |
| 7 | #include "dir.h" |
| 8 | #include "environment.h" |
| 9 | #include "gettext.h" |
| 10 | #include "parse-options.h" |
| 11 | #include "fsmonitor-ll.h" |
| 12 | #include "fsmonitor-ipc.h" |
| 13 | #include "fsmonitor-settings.h" |
| 14 | #include "compat/fsmonitor/fsm-health.h" |
| 15 | #include "compat/fsmonitor/fsm-listen.h" |
| 16 | #include "fsmonitor--daemon.h" |
| 17 | |
| 18 | #include "simple-ipc.h" |
| 19 | #include "strmap.h" |
| 20 | #include "run-command.h" |
| 21 | #include "trace.h" |
| 22 | #include "trace2.h" |
| 23 | |
| 24 | static const char * const builtin_fsmonitor__daemon_usage[] = { |
| 25 | N_("git fsmonitor--daemon start [<options>]"), |
| 26 | N_("git fsmonitor--daemon run [<options>]"), |
| 27 | "git fsmonitor--daemon stop", |
| 28 | "git fsmonitor--daemon status", |
| 29 | NULL |
| 30 | }; |
| 31 | |
| 32 | #ifdef HAVE_FSMONITOR_DAEMON_BACKEND |
| 33 | /* |
| 34 | * Global state loaded from config. |
| 35 | */ |
| 36 | #define FSMONITOR__IPC_THREADS "fsmonitor.ipcthreads" |
| 37 | static int fsmonitor__ipc_threads = 8; |
| 38 | |
| 39 | #define FSMONITOR__START_TIMEOUT "fsmonitor.starttimeout" |
| 40 | static int fsmonitor__start_timeout_sec = 60; |
| 41 | |
| 42 | #define FSMONITOR__ANNOUNCE_STARTUP "fsmonitor.announcestartup" |
| 43 | static int fsmonitor__announce_startup = 0; |
| 44 | |
| 45 | static int fsmonitor_config(const char *var, const char *value, |
| 46 | const struct config_context *ctx, void *cb) |
| 47 | { |
| 48 | if (!strcmp(var, FSMONITOR__IPC_THREADS)) { |
| 49 | int i = git_config_int(var, value, ctx->kvi); |
| 50 | if (i < 1) |
| 51 | return error(_("value of '%s' out of range: %d"), |
| 52 | FSMONITOR__IPC_THREADS, i); |
| 53 | fsmonitor__ipc_threads = i; |
| 54 | return 0; |
| 55 | } |
| 56 | |
| 57 | if (!strcmp(var, FSMONITOR__START_TIMEOUT)) { |
| 58 | int i = git_config_int(var, value, ctx->kvi); |
| 59 | if (i < 0) |
| 60 | return error(_("value of '%s' out of range: %d"), |
| 61 | FSMONITOR__START_TIMEOUT, i); |
| 62 | fsmonitor__start_timeout_sec = i; |
| 63 | return 0; |
| 64 | } |
| 65 | |
| 66 | if (!strcmp(var, FSMONITOR__ANNOUNCE_STARTUP)) { |
| 67 | int is_bool; |
| 68 | int i = git_config_bool_or_int(var, value, ctx->kvi, &is_bool); |
| 69 | if (i < 0) |
| 70 | return error(_("value of '%s' not bool or int: %d"), |
| 71 | var, i); |
| 72 | fsmonitor__announce_startup = i; |
| 73 | return 0; |
| 74 | } |
| 75 | |
| 76 | return git_default_config(var, value, ctx, cb); |
| 77 | } |
| 78 | |
| 79 | /* |
| 80 | * Acting as a CLIENT. |
| 81 | * |
| 82 | * Send a "quit" command to the `git-fsmonitor--daemon` (if running) |
| 83 | * and wait for it to shutdown. |
| 84 | */ |
| 85 | static int do_as_client__send_stop(void) |
| 86 | { |
| 87 | struct strbuf answer = STRBUF_INIT; |
| 88 | int ret; |
| 89 | int max_wait_ms = 30000; |
| 90 | int elapsed_ms = 0; |
| 91 | |
| 92 | ret = fsmonitor_ipc__send_command("quit", &answer); |
| 93 | |
| 94 | /* The quit command does not return any response data. */ |
| 95 | strbuf_release(&answer); |
| 96 | |
| 97 | if (ret) |
| 98 | return ret; |
| 99 | |
| 100 | trace2_region_enter("fsm_client", "polling-for-daemon-exit", NULL); |
| 101 | while (fsmonitor_ipc__get_state() == IPC_STATE__LISTENING) { |
| 102 | if (elapsed_ms >= max_wait_ms) { |
| 103 | trace2_region_leave("fsm_client", |
| 104 | "polling-for-daemon-exit", NULL); |
| 105 | return error(_("daemon did not stop within %d seconds"), |
| 106 | max_wait_ms / 1000); |
| 107 | } |
| 108 | sleep_millisec(50); |
| 109 | elapsed_ms += 50; |
| 110 | } |
| 111 | trace2_region_leave("fsm_client", "polling-for-daemon-exit", NULL); |
| 112 | |
| 113 | return 0; |
| 114 | } |
| 115 | |
| 116 | static int do_as_client__status(void) |
| 117 | { |
| 118 | enum ipc_active_state state = fsmonitor_ipc__get_state(); |
| 119 | |
| 120 | switch (state) { |
| 121 | case IPC_STATE__LISTENING: |
| 122 | printf(_("fsmonitor-daemon is watching '%s'\n"), |
| 123 | the_repository->worktree); |
| 124 | return 0; |
| 125 | |
| 126 | default: |
| 127 | printf(_("fsmonitor-daemon is not watching '%s'\n"), |
| 128 | the_repository->worktree); |
| 129 | return 1; |
| 130 | } |
| 131 | } |
| 132 | |
| 133 | enum fsmonitor_cookie_item_result { |
| 134 | FCIR_ERROR = -1, /* could not create cookie file ? */ |
| 135 | FCIR_INIT, |
| 136 | FCIR_SEEN, |
| 137 | FCIR_ABORT, |
| 138 | }; |
| 139 | |
| 140 | struct fsmonitor_cookie_item { |
| 141 | struct hashmap_entry entry; |
| 142 | char *name; |
| 143 | enum fsmonitor_cookie_item_result result; |
| 144 | }; |
| 145 | |
| 146 | static int cookies_cmp(const void *data UNUSED, |
| 147 | const struct hashmap_entry *he1, |
| 148 | const struct hashmap_entry *he2, const void *keydata) |
| 149 | { |
| 150 | const struct fsmonitor_cookie_item *a = |
| 151 | container_of(he1, const struct fsmonitor_cookie_item, entry); |
| 152 | const struct fsmonitor_cookie_item *b = |
| 153 | container_of(he2, const struct fsmonitor_cookie_item, entry); |
| 154 | |
| 155 | return strcmp(a->name, keydata ? keydata : b->name); |
| 156 | } |
| 157 | |
| 158 | static enum fsmonitor_cookie_item_result with_lock__wait_for_cookie( |
| 159 | struct fsmonitor_daemon_state *state) |
| 160 | { |
| 161 | /* assert current thread holding state->main_lock */ |
| 162 | |
| 163 | int fd; |
| 164 | struct fsmonitor_cookie_item *cookie; |
| 165 | struct strbuf cookie_pathname = STRBUF_INIT; |
| 166 | struct strbuf cookie_filename = STRBUF_INIT; |
| 167 | enum fsmonitor_cookie_item_result result; |
| 168 | int my_cookie_seq; |
| 169 | |
| 170 | CALLOC_ARRAY(cookie, 1); |
| 171 | |
| 172 | my_cookie_seq = state->cookie_seq++; |
| 173 | |
| 174 | strbuf_addf(&cookie_filename, "%i-%i", getpid(), my_cookie_seq); |
| 175 | |
| 176 | strbuf_addbuf(&cookie_pathname, &state->path_cookie_prefix); |
| 177 | strbuf_addbuf(&cookie_pathname, &cookie_filename); |
| 178 | |
| 179 | cookie->name = strbuf_detach(&cookie_filename, NULL); |
| 180 | cookie->result = FCIR_INIT; |
| 181 | hashmap_entry_init(&cookie->entry, strhash(cookie->name)); |
| 182 | |
| 183 | hashmap_add(&state->cookies, &cookie->entry); |
| 184 | |
| 185 | trace_printf_key(&trace_fsmonitor, "cookie-wait: '%s' '%s'", |
| 186 | cookie->name, cookie_pathname.buf); |
| 187 | |
| 188 | /* |
| 189 | * Create the cookie file on disk and then wait for a notification |
| 190 | * that the listener thread has seen it. |
| 191 | */ |
| 192 | fd = open(cookie_pathname.buf, O_WRONLY | O_CREAT | O_EXCL, 0600); |
| 193 | if (fd < 0) { |
| 194 | error_errno(_("could not create fsmonitor cookie '%s'"), |
| 195 | cookie->name); |
| 196 | |
| 197 | cookie->result = FCIR_ERROR; |
| 198 | goto done; |
| 199 | } |
| 200 | |
| 201 | /* |
| 202 | * Technically, close() and unlink() can fail, but we don't |
| 203 | * care here. We only created the file to trigger a watch |
| 204 | * event from the FS to know that when we're up to date. |
| 205 | */ |
| 206 | close(fd); |
| 207 | unlink(cookie_pathname.buf); |
| 208 | |
| 209 | /* |
| 210 | * Wait for the listener thread to observe the cookie file. |
| 211 | * Time out after a short interval so that the client |
| 212 | * does not hang forever if the filesystem does not deliver |
| 213 | * events (e.g., on certain container/overlay filesystems |
| 214 | * where inotify watches succeed but events never arrive). |
| 215 | */ |
| 216 | { |
| 217 | struct timeval now; |
| 218 | struct timespec ts; |
| 219 | int err = 0; |
| 220 | |
| 221 | gettimeofday(&now, NULL); |
| 222 | ts.tv_sec = now.tv_sec + 1; |
| 223 | ts.tv_nsec = now.tv_usec * 1000; |
| 224 | |
| 225 | while (cookie->result == FCIR_INIT && !err) |
| 226 | err = pthread_cond_timedwait(&state->cookies_cond, |
| 227 | &state->main_lock, |
| 228 | &ts); |
| 229 | if (err == ETIMEDOUT && cookie->result == FCIR_INIT) { |
| 230 | trace_printf_key(&trace_fsmonitor, |
| 231 | "cookie_wait timed out"); |
| 232 | cookie->result = FCIR_ERROR; |
| 233 | } |
| 234 | } |
| 235 | |
| 236 | done: |
| 237 | hashmap_remove(&state->cookies, &cookie->entry, NULL); |
| 238 | |
| 239 | result = cookie->result; |
| 240 | |
| 241 | free(cookie->name); |
| 242 | free(cookie); |
| 243 | strbuf_release(&cookie_pathname); |
| 244 | |
| 245 | return result; |
| 246 | } |
| 247 | |
| 248 | /* |
| 249 | * Mark these cookies as _SEEN and wake up the corresponding client threads. |
| 250 | */ |
| 251 | static void with_lock__mark_cookies_seen(struct fsmonitor_daemon_state *state, |
| 252 | const struct string_list *cookie_names) |
| 253 | { |
| 254 | /* assert current thread holding state->main_lock */ |
| 255 | |
| 256 | int k; |
| 257 | int nr_seen = 0; |
| 258 | |
| 259 | for (k = 0; k < cookie_names->nr; k++) { |
| 260 | struct fsmonitor_cookie_item key; |
| 261 | struct fsmonitor_cookie_item *cookie; |
| 262 | |
| 263 | key.name = cookie_names->items[k].string; |
| 264 | hashmap_entry_init(&key.entry, strhash(key.name)); |
| 265 | |
| 266 | cookie = hashmap_get_entry(&state->cookies, &key, entry, NULL); |
| 267 | if (cookie) { |
| 268 | trace_printf_key(&trace_fsmonitor, "cookie-seen: '%s'", |
| 269 | cookie->name); |
| 270 | cookie->result = FCIR_SEEN; |
| 271 | nr_seen++; |
| 272 | } |
| 273 | } |
| 274 | |
| 275 | if (nr_seen) |
| 276 | pthread_cond_broadcast(&state->cookies_cond); |
| 277 | } |
| 278 | |
| 279 | /* |
| 280 | * Set _ABORT on all pending cookies and wake up all client threads. |
| 281 | */ |
| 282 | static void with_lock__abort_all_cookies(struct fsmonitor_daemon_state *state) |
| 283 | { |
| 284 | /* assert current thread holding state->main_lock */ |
| 285 | |
| 286 | struct hashmap_iter iter; |
| 287 | struct fsmonitor_cookie_item *cookie; |
| 288 | int nr_aborted = 0; |
| 289 | |
| 290 | hashmap_for_each_entry(&state->cookies, &iter, cookie, entry) { |
| 291 | trace_printf_key(&trace_fsmonitor, "cookie-abort: '%s'", |
| 292 | cookie->name); |
| 293 | cookie->result = FCIR_ABORT; |
| 294 | nr_aborted++; |
| 295 | } |
| 296 | |
| 297 | if (nr_aborted) |
| 298 | pthread_cond_broadcast(&state->cookies_cond); |
| 299 | } |
| 300 | |
| 301 | /* |
| 302 | * Requests to and from a FSMonitor Protocol V2 provider use an opaque |
| 303 | * "token" as a virtual timestamp. Clients can request a summary of all |
| 304 | * created/deleted/modified files relative to a token. In the response, |
| 305 | * clients receive a new token for the next (relative) request. |
| 306 | * |
| 307 | * |
| 308 | * Token Format |
| 309 | * ============ |
| 310 | * |
| 311 | * The contents of the token are private and provider-specific. |
| 312 | * |
| 313 | * For the built-in fsmonitor--daemon, we define a token as follows: |
| 314 | * |
| 315 | * "builtin" ":" <token_id> ":" <sequence_nr> |
| 316 | * |
| 317 | * The "builtin" prefix is used as a namespace to avoid conflicts |
| 318 | * with other providers (such as Watchman). |
| 319 | * |
| 320 | * The <token_id> is an arbitrary OPAQUE string, such as a GUID, |
| 321 | * UUID, or {timestamp,pid}. It is used to group all filesystem |
| 322 | * events that happened while the daemon was monitoring (and in-sync |
| 323 | * with the filesystem). |
| 324 | * |
| 325 | * Unlike FSMonitor Protocol V1, it is not defined as a timestamp |
| 326 | * and does not define less-than/greater-than relationships. |
| 327 | * (There are too many race conditions to rely on file system |
| 328 | * event timestamps.) |
| 329 | * |
| 330 | * The <sequence_nr> is a simple integer incremented whenever the |
| 331 | * daemon needs to make its state public. For example, if 1000 file |
| 332 | * system events come in, but no clients have requested the data, |
| 333 | * the daemon can continue to accumulate file changes in the same |
| 334 | * bin and does not need to advance the sequence number. However, |
| 335 | * as soon as a client does arrive, the daemon needs to start a new |
| 336 | * bin and increment the sequence number. |
| 337 | * |
| 338 | * The sequence number serves as the boundary between 2 sets |
| 339 | * of bins -- the older ones that the client has already seen |
| 340 | * and the newer ones that it hasn't. |
| 341 | * |
| 342 | * When a new <token_id> is created, the <sequence_nr> is reset to |
| 343 | * zero. |
| 344 | * |
| 345 | * |
| 346 | * About Token Ids |
| 347 | * =============== |
| 348 | * |
| 349 | * A new token_id is created: |
| 350 | * |
| 351 | * [1] each time the daemon is started. |
| 352 | * |
| 353 | * [2] any time that the daemon must re-sync with the filesystem |
| 354 | * (such as when the kernel drops or we miss events on a very |
| 355 | * active volume). |
| 356 | * |
| 357 | * [3] in response to a client "flush" command (for dropped event |
| 358 | * testing). |
| 359 | * |
| 360 | * When a new token_id is created, the daemon is free to discard all |
| 361 | * cached filesystem events associated with any previous token_ids. |
| 362 | * Events associated with a non-current token_id will never be sent |
| 363 | * to a client. A token_id change implicitly means that the daemon |
| 364 | * has gap in its event history. |
| 365 | * |
| 366 | * Therefore, clients that present a token with a stale (non-current) |
| 367 | * token_id will always be given a trivial response. |
| 368 | */ |
| 369 | struct fsmonitor_token_data { |
| 370 | struct strbuf token_id; |
| 371 | struct fsmonitor_batch *batch_head; |
| 372 | struct fsmonitor_batch *batch_tail; |
| 373 | uint64_t client_ref_count; |
| 374 | }; |
| 375 | |
| 376 | struct fsmonitor_batch { |
| 377 | struct fsmonitor_batch *next; |
| 378 | uint64_t batch_seq_nr; |
| 379 | const char **interned_paths; |
| 380 | size_t nr, alloc; |
| 381 | time_t pinned_time; |
| 382 | }; |
| 383 | |
| 384 | static struct fsmonitor_token_data *fsmonitor_new_token_data(void) |
| 385 | { |
| 386 | static int test_env_value = -1; |
| 387 | static uint64_t flush_count = 0; |
| 388 | struct fsmonitor_token_data *token; |
| 389 | struct fsmonitor_batch *batch; |
| 390 | |
| 391 | CALLOC_ARRAY(token, 1); |
| 392 | batch = fsmonitor_batch__new(); |
| 393 | |
| 394 | strbuf_init(&token->token_id, 0); |
| 395 | token->batch_head = batch; |
| 396 | token->batch_tail = batch; |
| 397 | token->client_ref_count = 0; |
| 398 | |
| 399 | if (test_env_value < 0) |
| 400 | test_env_value = git_env_bool("GIT_TEST_FSMONITOR_TOKEN", 0); |
| 401 | |
| 402 | if (!test_env_value) { |
| 403 | struct timeval tv; |
| 404 | struct tm tm; |
| 405 | time_t secs; |
| 406 | |
| 407 | gettimeofday(&tv, NULL); |
| 408 | secs = tv.tv_sec; |
| 409 | gmtime_r(&secs, &tm); |
| 410 | |
| 411 | strbuf_addf(&token->token_id, |
| 412 | "%"PRIu64".%d.%4d%02d%02dT%02d%02d%02d.%06ldZ", |
| 413 | flush_count++, |
| 414 | getpid(), |
| 415 | tm.tm_year + 1900, tm.tm_mon + 1, tm.tm_mday, |
| 416 | tm.tm_hour, tm.tm_min, tm.tm_sec, |
| 417 | (long)tv.tv_usec); |
| 418 | } else { |
| 419 | strbuf_addf(&token->token_id, "test_%08x", test_env_value++); |
| 420 | } |
| 421 | |
| 422 | /* |
| 423 | * We created a new <token_id> and are starting a new series |
| 424 | * of tokens with a zero <seq_nr>. |
| 425 | * |
| 426 | * Since clients cannot guess our new (non test) <token_id> |
| 427 | * they will always receive a trivial response (because of the |
| 428 | * mismatch on the <token_id>). The trivial response will |
| 429 | * tell them our new <token_id> so that subsequent requests |
| 430 | * will be relative to our new series. (And when sending that |
| 431 | * response, we pin the current head of the batch list.) |
| 432 | * |
| 433 | * Even if the client correctly guesses the <token_id>, their |
| 434 | * request of "builtin:<token_id>:0" asks for all changes MORE |
| 435 | * RECENT than batch/bin 0. |
| 436 | * |
| 437 | * This implies that it is a waste to accumulate paths in the |
| 438 | * initial batch/bin (because they will never be transmitted). |
| 439 | * |
| 440 | * So the daemon could be running for days and watching the |
| 441 | * file system, but doesn't need to actually accumulate any |
| 442 | * paths UNTIL we need to set a reference point for a later |
| 443 | * relative request. |
| 444 | * |
| 445 | * However, it is very useful for testing to always have a |
| 446 | * reference point set. Pin batch 0 to force early file system |
| 447 | * events to accumulate. |
| 448 | */ |
| 449 | if (test_env_value) |
| 450 | batch->pinned_time = time(NULL); |
| 451 | |
| 452 | return token; |
| 453 | } |
| 454 | |
| 455 | struct fsmonitor_batch *fsmonitor_batch__new(void) |
| 456 | { |
| 457 | struct fsmonitor_batch *batch; |
| 458 | |
| 459 | CALLOC_ARRAY(batch, 1); |
| 460 | |
| 461 | return batch; |
| 462 | } |
| 463 | |
| 464 | void fsmonitor_batch__free_list(struct fsmonitor_batch *batch) |
| 465 | { |
| 466 | while (batch) { |
| 467 | struct fsmonitor_batch *next = batch->next; |
| 468 | |
| 469 | /* |
| 470 | * The actual strings within the array of this batch |
| 471 | * are interned, so we don't own them. We only own |
| 472 | * the array. |
| 473 | */ |
| 474 | free(batch->interned_paths); |
| 475 | free(batch); |
| 476 | |
| 477 | batch = next; |
| 478 | } |
| 479 | } |
| 480 | |
| 481 | void fsmonitor_batch__add_path(struct fsmonitor_batch *batch, |
| 482 | const char *path) |
| 483 | { |
| 484 | const char *interned_path = strintern(path); |
| 485 | |
| 486 | trace_printf_key(&trace_fsmonitor, "event: %s", interned_path); |
| 487 | |
| 488 | ALLOC_GROW(batch->interned_paths, batch->nr + 1, batch->alloc); |
| 489 | batch->interned_paths[batch->nr++] = interned_path; |
| 490 | } |
| 491 | |
| 492 | static void fsmonitor_batch__combine(struct fsmonitor_batch *batch_dest, |
| 493 | const struct fsmonitor_batch *batch_src) |
| 494 | { |
| 495 | size_t k; |
| 496 | |
| 497 | ALLOC_GROW(batch_dest->interned_paths, |
| 498 | batch_dest->nr + batch_src->nr + 1, |
| 499 | batch_dest->alloc); |
| 500 | |
| 501 | for (k = 0; k < batch_src->nr; k++) |
| 502 | batch_dest->interned_paths[batch_dest->nr++] = |
| 503 | batch_src->interned_paths[k]; |
| 504 | } |
| 505 | |
| 506 | /* |
| 507 | * To keep the batch list from growing unbounded in response to filesystem |
| 508 | * activity, we try to truncate old batches from the end of the list as |
| 509 | * they become irrelevant. |
| 510 | * |
| 511 | * We assume that the .git/index will be updated with the most recent token |
| 512 | * any time the index is updated. And future commands will only ask for |
| 513 | * recent changes *since* that new token. So as tokens advance into the |
| 514 | * future, older batch items will never be requested/needed. So we can |
| 515 | * truncate them without loss of functionality. |
| 516 | * |
| 517 | * However, multiple commands may be talking to the daemon concurrently |
| 518 | * or perform a slow command, so a little "token skew" is possible. |
| 519 | * Therefore, we want this to be a little bit lazy and have a generous |
| 520 | * delay. |
| 521 | * |
| 522 | * The current reader thread walked backwards in time from `token->batch_head` |
| 523 | * back to `batch_marker` somewhere in the middle of the batch list. |
| 524 | * |
| 525 | * Let's walk backwards in time from that marker an arbitrary delay |
| 526 | * and truncate the list there. Note that these timestamps are completely |
| 527 | * artificial (based on when we pinned the batch item) and not on any |
| 528 | * filesystem activity. |
| 529 | * |
| 530 | * Return the obsolete portion of the list after we have removed it from |
| 531 | * the official list so that the caller can free it after leaving the lock. |
| 532 | */ |
| 533 | #define MY_TIME_DELAY_SECONDS (5 * 60) /* seconds */ |
| 534 | |
| 535 | static struct fsmonitor_batch *with_lock__truncate_old_batches( |
| 536 | struct fsmonitor_daemon_state *state, |
| 537 | const struct fsmonitor_batch *batch_marker) |
| 538 | { |
| 539 | /* assert current thread holding state->main_lock */ |
| 540 | |
| 541 | const struct fsmonitor_batch *batch; |
| 542 | struct fsmonitor_batch *remainder; |
| 543 | |
| 544 | if (!batch_marker) |
| 545 | return NULL; |
| 546 | |
| 547 | trace_printf_key(&trace_fsmonitor, "Truncate: mark (%"PRIu64",%"PRIu64")", |
| 548 | batch_marker->batch_seq_nr, |
| 549 | (uint64_t)batch_marker->pinned_time); |
| 550 | |
| 551 | for (batch = batch_marker; batch; batch = batch->next) { |
| 552 | time_t t; |
| 553 | |
| 554 | if (!batch->pinned_time) /* an overflow batch */ |
| 555 | continue; |
| 556 | |
| 557 | t = batch->pinned_time + MY_TIME_DELAY_SECONDS; |
| 558 | if (t > batch_marker->pinned_time) /* too close to marker */ |
| 559 | continue; |
| 560 | |
| 561 | goto truncate_past_here; |
| 562 | } |
| 563 | |
| 564 | return NULL; |
| 565 | |
| 566 | truncate_past_here: |
| 567 | state->current_token_data->batch_tail = (struct fsmonitor_batch *)batch; |
| 568 | |
| 569 | remainder = ((struct fsmonitor_batch *)batch)->next; |
| 570 | ((struct fsmonitor_batch *)batch)->next = NULL; |
| 571 | |
| 572 | return remainder; |
| 573 | } |
| 574 | |
| 575 | static void fsmonitor_free_token_data(struct fsmonitor_token_data *token) |
| 576 | { |
| 577 | if (!token) |
| 578 | return; |
| 579 | |
| 580 | assert(token->client_ref_count == 0); |
| 581 | |
| 582 | strbuf_release(&token->token_id); |
| 583 | |
| 584 | fsmonitor_batch__free_list(token->batch_head); |
| 585 | |
| 586 | free(token); |
| 587 | } |
| 588 | |
| 589 | /* |
| 590 | * Flush all of our cached data about the filesystem. Call this if we |
| 591 | * lose sync with the filesystem and miss some notification events. |
| 592 | * |
| 593 | * [1] If we are missing events, then we no longer have a complete |
| 594 | * history of the directory (relative to our current start token). |
| 595 | * We should create a new token and start fresh (as if we just |
| 596 | * booted up). |
| 597 | * |
| 598 | * [2] Some of those lost events may have been for cookie files. We |
| 599 | * should assume the worst and abort them rather letting them starve. |
| 600 | * |
| 601 | * If there are no concurrent threads reading the current token data |
| 602 | * series, we can free it now. Otherwise, let the last reader free |
| 603 | * it. |
| 604 | * |
| 605 | * Either way, the old token data series is no longer associated with |
| 606 | * our state data. |
| 607 | */ |
| 608 | static void with_lock__do_force_resync(struct fsmonitor_daemon_state *state) |
| 609 | { |
| 610 | /* assert current thread holding state->main_lock */ |
| 611 | |
| 612 | struct fsmonitor_token_data *free_me = NULL; |
| 613 | struct fsmonitor_token_data *new_one = NULL; |
| 614 | |
| 615 | new_one = fsmonitor_new_token_data(); |
| 616 | |
| 617 | if (state->current_token_data->client_ref_count == 0) |
| 618 | free_me = state->current_token_data; |
| 619 | state->current_token_data = new_one; |
| 620 | |
| 621 | fsmonitor_free_token_data(free_me); |
| 622 | |
| 623 | with_lock__abort_all_cookies(state); |
| 624 | } |
| 625 | |
| 626 | void fsmonitor_force_resync(struct fsmonitor_daemon_state *state) |
| 627 | { |
| 628 | pthread_mutex_lock(&state->main_lock); |
| 629 | with_lock__do_force_resync(state); |
| 630 | pthread_mutex_unlock(&state->main_lock); |
| 631 | } |
| 632 | |
| 633 | /* |
| 634 | * Format an opaque token string to send to the client. |
| 635 | */ |
| 636 | static void with_lock__format_response_token( |
| 637 | struct strbuf *response_token, |
| 638 | const struct strbuf *response_token_id, |
| 639 | const struct fsmonitor_batch *batch) |
| 640 | { |
| 641 | /* assert current thread holding state->main_lock */ |
| 642 | |
| 643 | strbuf_reset(response_token); |
| 644 | strbuf_addf(response_token, "builtin:%s:%"PRIu64, |
| 645 | response_token_id->buf, batch->batch_seq_nr); |
| 646 | } |
| 647 | |
| 648 | /* |
| 649 | * Parse an opaque token from the client. |
| 650 | * Returns -1 on error. |
| 651 | */ |
| 652 | static int fsmonitor_parse_client_token(const char *buf_token, |
| 653 | struct strbuf *requested_token_id, |
| 654 | uint64_t *seq_nr) |
| 655 | { |
| 656 | const char *p; |
| 657 | char *p_end; |
| 658 | |
| 659 | strbuf_reset(requested_token_id); |
| 660 | *seq_nr = 0; |
| 661 | |
| 662 | if (!skip_prefix(buf_token, "builtin:", &p)) |
| 663 | return -1; |
| 664 | |
| 665 | while (*p && *p != ':') |
| 666 | strbuf_addch(requested_token_id, *p++); |
| 667 | if (!*p++) |
| 668 | return -1; |
| 669 | |
| 670 | *seq_nr = (uint64_t)strtoumax(p, &p_end, 10); |
| 671 | if (*p_end) |
| 672 | return -1; |
| 673 | |
| 674 | return 0; |
| 675 | } |
| 676 | |
| 677 | static int do_handle_client(struct fsmonitor_daemon_state *state, |
| 678 | const char *command, |
| 679 | ipc_server_reply_cb *reply, |
| 680 | struct ipc_server_reply_data *reply_data) |
| 681 | { |
| 682 | struct fsmonitor_token_data *token_data = NULL; |
| 683 | struct strbuf response_token = STRBUF_INIT; |
| 684 | struct strbuf requested_token_id = STRBUF_INIT; |
| 685 | struct strbuf payload = STRBUF_INIT; |
| 686 | uint64_t requested_oldest_seq_nr = 0; |
| 687 | uint64_t total_response_len = 0; |
| 688 | const char *p; |
| 689 | const struct fsmonitor_batch *batch_head; |
| 690 | const struct fsmonitor_batch *batch; |
| 691 | struct fsmonitor_batch *remainder = NULL; |
| 692 | intmax_t count = 0, duplicates = 0; |
| 693 | struct strset shown = STRSET_INIT; |
| 694 | int do_trivial = 0; |
| 695 | int do_flush = 0; |
| 696 | int do_cookie = 0; |
| 697 | enum fsmonitor_cookie_item_result cookie_result; |
| 698 | |
| 699 | /* |
| 700 | * We expect `command` to be of the form: |
| 701 | * |
| 702 | * <command> := quit NUL |
| 703 | * | flush NUL |
| 704 | * | <V1-time-since-epoch-ns> NUL |
| 705 | * | <V2-opaque-fsmonitor-token> NUL |
| 706 | */ |
| 707 | |
| 708 | if (!strcmp(command, "quit")) { |
| 709 | /* |
| 710 | * A client has requested over the socket/pipe that the |
| 711 | * daemon shutdown. |
| 712 | * |
| 713 | * Tell the IPC thread pool to shutdown (which completes |
| 714 | * the await in the main thread (which can stop the |
| 715 | * fsmonitor listener thread)). |
| 716 | * |
| 717 | * There is no reply to the client. |
| 718 | */ |
| 719 | return SIMPLE_IPC_QUIT; |
| 720 | |
| 721 | } else if (!strcmp(command, "flush")) { |
| 722 | /* |
| 723 | * Flush all of our cached data and generate a new token |
| 724 | * just like if we lost sync with the filesystem. |
| 725 | * |
| 726 | * Then send a trivial response using the new token. |
| 727 | */ |
| 728 | do_flush = 1; |
| 729 | do_trivial = 1; |
| 730 | |
| 731 | } else if (!skip_prefix(command, "builtin:", &p)) { |
| 732 | /* assume V1 timestamp or garbage */ |
| 733 | |
| 734 | char *p_end; |
| 735 | |
| 736 | strtoumax(command, &p_end, 10); |
| 737 | trace_printf_key(&trace_fsmonitor, |
| 738 | ((*p_end) ? |
| 739 | "fsmonitor: invalid command line '%s'" : |
| 740 | "fsmonitor: unsupported V1 protocol '%s'"), |
| 741 | command); |
| 742 | do_trivial = 1; |
| 743 | do_cookie = 1; |
| 744 | |
| 745 | } else { |
| 746 | /* We have "builtin:*" */ |
| 747 | if (fsmonitor_parse_client_token(command, &requested_token_id, |
| 748 | &requested_oldest_seq_nr)) { |
| 749 | trace_printf_key(&trace_fsmonitor, |
| 750 | "fsmonitor: invalid V2 protocol token '%s'", |
| 751 | command); |
| 752 | do_trivial = 1; |
| 753 | do_cookie = 1; |
| 754 | |
| 755 | } else { |
| 756 | /* |
| 757 | * We have a V2 valid token: |
| 758 | * "builtin:<token_id>:<seq_nr>" |
| 759 | */ |
| 760 | do_cookie = 1; |
| 761 | } |
| 762 | } |
| 763 | |
| 764 | pthread_mutex_lock(&state->main_lock); |
| 765 | |
| 766 | if (!state->current_token_data) |
| 767 | BUG("fsmonitor state does not have a current token"); |
| 768 | |
| 769 | /* |
| 770 | * Write a cookie file inside the directory being watched in |
| 771 | * an effort to flush out existing filesystem events that we |
| 772 | * actually care about. Suspend this client thread until we |
| 773 | * see the filesystem events for this cookie file. |
| 774 | * |
| 775 | * Creating the cookie lets us guarantee that our FS listener |
| 776 | * thread has drained the kernel queue and we are caught up |
| 777 | * with the kernel. |
| 778 | * |
| 779 | * If we cannot create the cookie (or otherwise guarantee that |
| 780 | * we are caught up), we send a trivial response. We have to |
| 781 | * assume that there might be some very, very recent activity |
| 782 | * on the FS still in flight. |
| 783 | */ |
| 784 | if (do_cookie) { |
| 785 | cookie_result = with_lock__wait_for_cookie(state); |
| 786 | if (cookie_result != FCIR_SEEN) { |
| 787 | error(_("fsmonitor: cookie_result '%d' != SEEN"), |
| 788 | cookie_result); |
| 789 | do_trivial = 1; |
| 790 | } |
| 791 | } |
| 792 | |
| 793 | if (do_flush) |
| 794 | with_lock__do_force_resync(state); |
| 795 | |
| 796 | /* |
| 797 | * We mark the current head of the batch list as "pinned" so |
| 798 | * that the listener thread will treat this item as read-only |
| 799 | * (and prevent any more paths from being added to it) from |
| 800 | * now on. |
| 801 | */ |
| 802 | token_data = state->current_token_data; |
| 803 | batch_head = token_data->batch_head; |
| 804 | ((struct fsmonitor_batch *)batch_head)->pinned_time = time(NULL); |
| 805 | |
| 806 | /* |
| 807 | * FSMonitor Protocol V2 requires that we send a response header |
| 808 | * with a "new current token" and then all of the paths that changed |
| 809 | * since the "requested token". We send the seq_nr of the just-pinned |
| 810 | * head batch so that future requests from a client will be relative |
| 811 | * to it. |
| 812 | */ |
| 813 | with_lock__format_response_token(&response_token, |
| 814 | &token_data->token_id, batch_head); |
| 815 | |
| 816 | reply(reply_data, response_token.buf, response_token.len + 1); |
| 817 | total_response_len += response_token.len + 1; |
| 818 | |
| 819 | trace2_data_string("fsmonitor", the_repository, "response/token", |
| 820 | response_token.buf); |
| 821 | trace_printf_key(&trace_fsmonitor, "response token: %s", |
| 822 | response_token.buf); |
| 823 | |
| 824 | if (!do_trivial) { |
| 825 | if (strcmp(requested_token_id.buf, token_data->token_id.buf)) { |
| 826 | /* |
| 827 | * The client last spoke to a different daemon |
| 828 | * instance -OR- the daemon had to resync with |
| 829 | * the filesystem (and lost events), so reject. |
| 830 | */ |
| 831 | trace2_data_string("fsmonitor", the_repository, |
| 832 | "response/token", "different"); |
| 833 | do_trivial = 1; |
| 834 | |
| 835 | } else if (requested_oldest_seq_nr < |
| 836 | token_data->batch_tail->batch_seq_nr) { |
| 837 | /* |
| 838 | * The client wants older events than we have for |
| 839 | * this token_id. This means that the end of our |
| 840 | * batch list was truncated and we cannot give the |
| 841 | * client a complete snapshot relative to their |
| 842 | * request. |
| 843 | */ |
| 844 | trace_printf_key(&trace_fsmonitor, |
| 845 | "client requested truncated data"); |
| 846 | do_trivial = 1; |
| 847 | } |
| 848 | } |
| 849 | |
| 850 | if (do_trivial) { |
| 851 | pthread_mutex_unlock(&state->main_lock); |
| 852 | |
| 853 | reply(reply_data, "/", 2); |
| 854 | |
| 855 | trace2_data_intmax("fsmonitor", the_repository, |
| 856 | "response/trivial", 1); |
| 857 | |
| 858 | goto cleanup; |
| 859 | } |
| 860 | |
| 861 | /* |
| 862 | * We're going to hold onto a pointer to the current |
| 863 | * token-data while we walk the list of batches of files. |
| 864 | * During this time, we will NOT be under the lock. |
| 865 | * So we ref-count it. |
| 866 | * |
| 867 | * This allows the listener thread to continue prepending |
| 868 | * new batches of items to the token-data (which we'll ignore). |
| 869 | * |
| 870 | * AND it allows the listener thread to do a token-reset |
| 871 | * (and install a new `current_token_data`). |
| 872 | */ |
| 873 | token_data->client_ref_count++; |
| 874 | |
| 875 | pthread_mutex_unlock(&state->main_lock); |
| 876 | |
| 877 | /* |
| 878 | * The client request is relative to the token that they sent, |
| 879 | * so walk the batch list backwards from the current head back |
| 880 | * to the batch (sequence number) they named. |
| 881 | * |
| 882 | * We use a strset to de-dup the list of pathnames. |
| 883 | * |
| 884 | * NEEDSWORK: each batch contains a list of interned strings, |
| 885 | * so we only need to do pointer comparisons here to build the |
| 886 | * hash table. Currently, we're still comparing the string |
| 887 | * values. |
| 888 | */ |
| 889 | strset_init_with_options(&shown, NULL, 0); |
| 890 | for (batch = batch_head; |
| 891 | batch && batch->batch_seq_nr > requested_oldest_seq_nr; |
| 892 | batch = batch->next) { |
| 893 | size_t k; |
| 894 | |
| 895 | for (k = 0; k < batch->nr; k++) { |
| 896 | const char *s = batch->interned_paths[k]; |
| 897 | size_t s_len; |
| 898 | |
| 899 | if (!strset_add(&shown, s)) |
| 900 | duplicates++; |
| 901 | else { |
| 902 | trace_printf_key(&trace_fsmonitor, |
| 903 | "send[%"PRIuMAX"]: %s", |
| 904 | count, s); |
| 905 | |
| 906 | /* Each path gets written with a trailing NUL */ |
| 907 | s_len = strlen(s) + 1; |
| 908 | |
| 909 | if (payload.len + s_len >= |
| 910 | LARGE_PACKET_DATA_MAX) { |
| 911 | reply(reply_data, payload.buf, |
| 912 | payload.len); |
| 913 | total_response_len += payload.len; |
| 914 | strbuf_reset(&payload); |
| 915 | } |
| 916 | |
| 917 | strbuf_add(&payload, s, s_len); |
| 918 | count++; |
| 919 | } |
| 920 | } |
| 921 | } |
| 922 | |
| 923 | if (payload.len) { |
| 924 | reply(reply_data, payload.buf, payload.len); |
| 925 | total_response_len += payload.len; |
| 926 | } |
| 927 | |
| 928 | pthread_mutex_lock(&state->main_lock); |
| 929 | |
| 930 | if (token_data->client_ref_count > 0) |
| 931 | token_data->client_ref_count--; |
| 932 | |
| 933 | if (token_data->client_ref_count == 0) { |
| 934 | if (token_data != state->current_token_data) { |
| 935 | /* |
| 936 | * The listener thread did a token-reset while we were |
| 937 | * walking the batch list. Therefore, this token is |
| 938 | * stale and can be discarded completely. If we are |
| 939 | * the last reader thread using this token, we own |
| 940 | * that work. |
| 941 | */ |
| 942 | fsmonitor_free_token_data(token_data); |
| 943 | } else if (batch) { |
| 944 | /* |
| 945 | * We are holding the lock and are the only |
| 946 | * reader of the ref-counted portion of the |
| 947 | * list, so we get the honor of seeing if the |
| 948 | * list can be truncated to save memory. |
| 949 | * |
| 950 | * The main loop did not walk to the end of the |
| 951 | * list, so this batch is the first item in the |
| 952 | * batch-list that is older than the requested |
| 953 | * end-point sequence number. See if the tail |
| 954 | * end of the list is obsolete. |
| 955 | */ |
| 956 | remainder = with_lock__truncate_old_batches(state, |
| 957 | batch); |
| 958 | } |
| 959 | } |
| 960 | |
| 961 | pthread_mutex_unlock(&state->main_lock); |
| 962 | |
| 963 | if (remainder) |
| 964 | fsmonitor_batch__free_list(remainder); |
| 965 | |
| 966 | trace2_data_intmax("fsmonitor", the_repository, "response/length", total_response_len); |
| 967 | trace2_data_intmax("fsmonitor", the_repository, "response/count/files", count); |
| 968 | trace2_data_intmax("fsmonitor", the_repository, "response/count/duplicates", duplicates); |
| 969 | |
| 970 | cleanup: |
| 971 | strset_clear(&shown); |
| 972 | strbuf_release(&response_token); |
| 973 | strbuf_release(&requested_token_id); |
| 974 | strbuf_release(&payload); |
| 975 | |
| 976 | return 0; |
| 977 | } |
| 978 | |
| 979 | static ipc_server_application_cb handle_client; |
| 980 | |
| 981 | static int handle_client(void *data, |
| 982 | const char *command, size_t command_len, |
| 983 | ipc_server_reply_cb *reply, |
| 984 | struct ipc_server_reply_data *reply_data) |
| 985 | { |
| 986 | struct fsmonitor_daemon_state *state = data; |
| 987 | int result; |
| 988 | |
| 989 | /* |
| 990 | * The Simple IPC API now supports {char*, len} arguments, but |
| 991 | * FSMonitor always uses proper null-terminated strings, so |
| 992 | * we can ignore the command_len argument. (Trust, but verify.) |
| 993 | */ |
| 994 | if (command_len != strlen(command)) |
| 995 | BUG("FSMonitor assumes text messages"); |
| 996 | |
| 997 | trace_printf_key(&trace_fsmonitor, "requested token: %s", command); |
| 998 | |
| 999 | trace2_region_enter("fsmonitor", "handle_client", the_repository); |
| 1000 | trace2_data_string("fsmonitor", the_repository, "request", command); |
| 1001 | |
| 1002 | result = do_handle_client(state, command, reply, reply_data); |
| 1003 | |
| 1004 | trace2_region_leave("fsmonitor", "handle_client", the_repository); |
| 1005 | |
| 1006 | return result; |
| 1007 | } |
| 1008 | |
| 1009 | #define FSMONITOR_DIR "fsmonitor--daemon" |
| 1010 | #define FSMONITOR_COOKIE_DIR "cookies" |
| 1011 | #define FSMONITOR_COOKIE_PREFIX (FSMONITOR_DIR "/" FSMONITOR_COOKIE_DIR "/") |
| 1012 | |
| 1013 | enum fsmonitor_path_type fsmonitor_classify_path_workdir_relative( |
| 1014 | const char *rel) |
| 1015 | { |
| 1016 | if (fspathncmp(rel, ".git", 4)) |
| 1017 | return IS_WORKDIR_PATH; |
| 1018 | rel += 4; |
| 1019 | |
| 1020 | if (!*rel) |
| 1021 | return IS_DOT_GIT; |
| 1022 | if (*rel != '/') |
| 1023 | return IS_WORKDIR_PATH; /* e.g. .gitignore */ |
| 1024 | rel++; |
| 1025 | |
| 1026 | if (!fspathncmp(rel, FSMONITOR_COOKIE_PREFIX, |
| 1027 | strlen(FSMONITOR_COOKIE_PREFIX))) |
| 1028 | return IS_INSIDE_DOT_GIT_WITH_COOKIE_PREFIX; |
| 1029 | |
| 1030 | return IS_INSIDE_DOT_GIT; |
| 1031 | } |
| 1032 | |
| 1033 | enum fsmonitor_path_type fsmonitor_classify_path_gitdir_relative( |
| 1034 | const char *rel) |
| 1035 | { |
| 1036 | if (!fspathncmp(rel, FSMONITOR_COOKIE_PREFIX, |
| 1037 | strlen(FSMONITOR_COOKIE_PREFIX))) |
| 1038 | return IS_INSIDE_GITDIR_WITH_COOKIE_PREFIX; |
| 1039 | |
| 1040 | return IS_INSIDE_GITDIR; |
| 1041 | } |
| 1042 | |
| 1043 | static enum fsmonitor_path_type try_classify_workdir_abs_path( |
| 1044 | struct fsmonitor_daemon_state *state, |
| 1045 | const char *path) |
| 1046 | { |
| 1047 | const char *rel; |
| 1048 | |
| 1049 | if (fspathncmp(path, state->path_worktree_watch.buf, |
| 1050 | state->path_worktree_watch.len)) |
| 1051 | return IS_OUTSIDE_CONE; |
| 1052 | |
| 1053 | rel = path + state->path_worktree_watch.len; |
| 1054 | |
| 1055 | if (!*rel) |
| 1056 | return IS_WORKDIR_PATH; /* it is the root dir exactly */ |
| 1057 | if (*rel != '/') |
| 1058 | return IS_OUTSIDE_CONE; |
| 1059 | rel++; |
| 1060 | |
| 1061 | return fsmonitor_classify_path_workdir_relative(rel); |
| 1062 | } |
| 1063 | |
| 1064 | enum fsmonitor_path_type fsmonitor_classify_path_absolute( |
| 1065 | struct fsmonitor_daemon_state *state, |
| 1066 | const char *path) |
| 1067 | { |
| 1068 | const char *rel; |
| 1069 | enum fsmonitor_path_type t; |
| 1070 | |
| 1071 | t = try_classify_workdir_abs_path(state, path); |
| 1072 | if (state->nr_paths_watching == 1) |
| 1073 | return t; |
| 1074 | if (t != IS_OUTSIDE_CONE) |
| 1075 | return t; |
| 1076 | |
| 1077 | if (fspathncmp(path, state->path_gitdir_watch.buf, |
| 1078 | state->path_gitdir_watch.len)) |
| 1079 | return IS_OUTSIDE_CONE; |
| 1080 | |
| 1081 | rel = path + state->path_gitdir_watch.len; |
| 1082 | |
| 1083 | if (!*rel) |
| 1084 | return IS_GITDIR; /* it is the <gitdir> exactly */ |
| 1085 | if (*rel != '/') |
| 1086 | return IS_OUTSIDE_CONE; |
| 1087 | rel++; |
| 1088 | |
| 1089 | return fsmonitor_classify_path_gitdir_relative(rel); |
| 1090 | } |
| 1091 | |
| 1092 | /* |
| 1093 | * We try to combine small batches at the front of the batch-list to avoid |
| 1094 | * having a long list. This hopefully makes it a little easier when we want |
| 1095 | * to truncate and maintain the list. However, we don't want the paths array |
| 1096 | * to just keep growing and growing with realloc, so we insert an arbitrary |
| 1097 | * limit. |
| 1098 | */ |
| 1099 | #define MY_COMBINE_LIMIT (1024) |
| 1100 | |
| 1101 | void fsmonitor_publish(struct fsmonitor_daemon_state *state, |
| 1102 | struct fsmonitor_batch *batch, |
| 1103 | const struct string_list *cookie_names) |
| 1104 | { |
| 1105 | if (!batch && !cookie_names->nr) |
| 1106 | return; |
| 1107 | |
| 1108 | pthread_mutex_lock(&state->main_lock); |
| 1109 | |
| 1110 | if (batch) { |
| 1111 | struct fsmonitor_batch *head; |
| 1112 | |
| 1113 | head = state->current_token_data->batch_head; |
| 1114 | if (!head) { |
| 1115 | BUG("token does not have batch"); |
| 1116 | } else if (head->pinned_time) { |
| 1117 | /* |
| 1118 | * We cannot alter the current batch list |
| 1119 | * because: |
| 1120 | * |
| 1121 | * [a] it is being transmitted to at least one |
| 1122 | * client and the handle_client() thread has a |
| 1123 | * ref-count, but not a lock on the batch list |
| 1124 | * starting with this item. |
| 1125 | * |
| 1126 | * [b] it has been transmitted in the past to |
| 1127 | * at least one client such that future |
| 1128 | * requests are relative to this head batch. |
| 1129 | * |
| 1130 | * So, we can only prepend a new batch onto |
| 1131 | * the front of the list. |
| 1132 | */ |
| 1133 | batch->batch_seq_nr = head->batch_seq_nr + 1; |
| 1134 | batch->next = head; |
| 1135 | state->current_token_data->batch_head = batch; |
| 1136 | } else if (!head->batch_seq_nr) { |
| 1137 | /* |
| 1138 | * Batch 0 is unpinned. See the note in |
| 1139 | * `fsmonitor_new_token_data()` about why we |
| 1140 | * don't need to accumulate these paths. |
| 1141 | */ |
| 1142 | fsmonitor_batch__free_list(batch); |
| 1143 | } else if (head->nr + batch->nr > MY_COMBINE_LIMIT) { |
| 1144 | /* |
| 1145 | * The head batch in the list has never been |
| 1146 | * transmitted to a client, but folding the |
| 1147 | * contents of the new batch onto it would |
| 1148 | * exceed our arbitrary limit, so just prepend |
| 1149 | * the new batch onto the list. |
| 1150 | */ |
| 1151 | batch->batch_seq_nr = head->batch_seq_nr + 1; |
| 1152 | batch->next = head; |
| 1153 | state->current_token_data->batch_head = batch; |
| 1154 | } else { |
| 1155 | /* |
| 1156 | * We are free to add the paths in the given |
| 1157 | * batch onto the end of the current head batch. |
| 1158 | */ |
| 1159 | fsmonitor_batch__combine(head, batch); |
| 1160 | fsmonitor_batch__free_list(batch); |
| 1161 | } |
| 1162 | } |
| 1163 | |
| 1164 | if (cookie_names->nr) |
| 1165 | with_lock__mark_cookies_seen(state, cookie_names); |
| 1166 | |
| 1167 | pthread_mutex_unlock(&state->main_lock); |
| 1168 | } |
| 1169 | |
| 1170 | static void *fsm_health__thread_proc(void *_state) |
| 1171 | { |
| 1172 | struct fsmonitor_daemon_state *state = _state; |
| 1173 | |
| 1174 | trace2_thread_start("fsm-health"); |
| 1175 | |
| 1176 | fsm_health__loop(state); |
| 1177 | |
| 1178 | trace2_thread_exit(); |
| 1179 | return NULL; |
| 1180 | } |
| 1181 | |
| 1182 | static void *fsm_listen__thread_proc(void *_state) |
| 1183 | { |
| 1184 | struct fsmonitor_daemon_state *state = _state; |
| 1185 | |
| 1186 | trace2_thread_start("fsm-listen"); |
| 1187 | |
| 1188 | trace_printf_key(&trace_fsmonitor, "Watching: worktree '%s'", |
| 1189 | state->path_worktree_watch.buf); |
| 1190 | if (state->nr_paths_watching > 1) |
| 1191 | trace_printf_key(&trace_fsmonitor, "Watching: gitdir '%s'", |
| 1192 | state->path_gitdir_watch.buf); |
| 1193 | |
| 1194 | fsm_listen__loop(state); |
| 1195 | |
| 1196 | pthread_mutex_lock(&state->main_lock); |
| 1197 | if (state->current_token_data && |
| 1198 | state->current_token_data->client_ref_count == 0) |
| 1199 | fsmonitor_free_token_data(state->current_token_data); |
| 1200 | state->current_token_data = NULL; |
| 1201 | pthread_mutex_unlock(&state->main_lock); |
| 1202 | |
| 1203 | trace2_thread_exit(); |
| 1204 | return NULL; |
| 1205 | } |
| 1206 | |
| 1207 | static int fsmonitor_run_daemon_1(struct fsmonitor_daemon_state *state) |
| 1208 | { |
| 1209 | struct ipc_server_opts ipc_opts = { |
| 1210 | .nr_threads = fsmonitor__ipc_threads, |
| 1211 | |
| 1212 | /* |
| 1213 | * We know that there are no other active threads yet, |
| 1214 | * so we can let the IPC layer temporarily chdir() if |
| 1215 | * it needs to when creating the server side of the |
| 1216 | * Unix domain socket. |
| 1217 | */ |
| 1218 | .uds_disallow_chdir = 0 |
| 1219 | }; |
| 1220 | int health_started = 0; |
| 1221 | int listener_started = 0; |
| 1222 | int err = 0; |
| 1223 | |
| 1224 | /* |
| 1225 | * Start the IPC thread pool before the we've started the file |
| 1226 | * system event listener thread so that we have the IPC handle |
| 1227 | * before we need it. |
| 1228 | */ |
| 1229 | if (ipc_server_init_async(&state->ipc_server_data, |
| 1230 | state->path_ipc.buf, &ipc_opts, |
| 1231 | handle_client, state)) |
| 1232 | return error_errno( |
| 1233 | _("could not start IPC thread pool on '%s'"), |
| 1234 | state->path_ipc.buf); |
| 1235 | |
| 1236 | /* |
| 1237 | * Start the fsmonitor listener thread to collect filesystem |
| 1238 | * events. |
| 1239 | */ |
| 1240 | if (pthread_create(&state->listener_thread, NULL, |
| 1241 | fsm_listen__thread_proc, state)) { |
| 1242 | ipc_server_stop_async(state->ipc_server_data); |
| 1243 | err = error(_("could not start fsmonitor listener thread")); |
| 1244 | goto cleanup; |
| 1245 | } |
| 1246 | listener_started = 1; |
| 1247 | |
| 1248 | /* |
| 1249 | * Start the health thread to watch over our process. |
| 1250 | */ |
| 1251 | if (pthread_create(&state->health_thread, NULL, |
| 1252 | fsm_health__thread_proc, state)) { |
| 1253 | ipc_server_stop_async(state->ipc_server_data); |
| 1254 | err = error(_("could not start fsmonitor health thread")); |
| 1255 | goto cleanup; |
| 1256 | } |
| 1257 | health_started = 1; |
| 1258 | |
| 1259 | /* |
| 1260 | * The daemon is now fully functional in background threads. |
| 1261 | * Our primary thread should now just wait while the threads |
| 1262 | * do all the work. |
| 1263 | */ |
| 1264 | cleanup: |
| 1265 | /* |
| 1266 | * Wait for the IPC thread pool to shutdown (whether by client |
| 1267 | * request, from filesystem activity, or an error). |
| 1268 | */ |
| 1269 | ipc_server_await(state->ipc_server_data); |
| 1270 | |
| 1271 | /* |
| 1272 | * The fsmonitor listener thread may have received a shutdown |
| 1273 | * event from the IPC thread pool, but it doesn't hurt to tell |
| 1274 | * it again. And wait for it to shutdown. |
| 1275 | */ |
| 1276 | if (listener_started) { |
| 1277 | fsm_listen__stop_async(state); |
| 1278 | pthread_join(state->listener_thread, NULL); |
| 1279 | } |
| 1280 | |
| 1281 | if (health_started) { |
| 1282 | fsm_health__stop_async(state); |
| 1283 | pthread_join(state->health_thread, NULL); |
| 1284 | } |
| 1285 | |
| 1286 | if (err) |
| 1287 | return err; |
| 1288 | if (state->listen_error_code) |
| 1289 | return state->listen_error_code; |
| 1290 | if (state->health_error_code) |
| 1291 | return state->health_error_code; |
| 1292 | return 0; |
| 1293 | } |
| 1294 | |
| 1295 | static int fsmonitor_run_daemon(void) |
| 1296 | { |
| 1297 | struct fsmonitor_daemon_state state; |
| 1298 | const char *home; |
| 1299 | int err; |
| 1300 | |
| 1301 | memset(&state, 0, sizeof(state)); |
| 1302 | |
| 1303 | hashmap_init(&state.cookies, cookies_cmp, NULL, 0); |
| 1304 | pthread_mutex_init(&state.main_lock, NULL); |
| 1305 | pthread_cond_init(&state.cookies_cond, NULL); |
| 1306 | state.listen_error_code = 0; |
| 1307 | state.health_error_code = 0; |
| 1308 | state.current_token_data = fsmonitor_new_token_data(); |
| 1309 | |
| 1310 | /* Prepare to (recursively) watch the <worktree-root> directory. */ |
| 1311 | strbuf_init(&state.path_worktree_watch, 0); |
| 1312 | strbuf_addstr(&state.path_worktree_watch, |
| 1313 | absolute_path(repo_get_work_tree(the_repository))); |
| 1314 | state.nr_paths_watching = 1; |
| 1315 | |
| 1316 | strbuf_init(&state.alias.alias, 0); |
| 1317 | strbuf_init(&state.alias.points_to, 0); |
| 1318 | if ((err = fsmonitor__get_alias(state.path_worktree_watch.buf, &state.alias))) |
| 1319 | goto done; |
| 1320 | |
| 1321 | /* |
| 1322 | * We create and delete cookie files somewhere inside the .git |
| 1323 | * directory to help us keep sync with the file system. If |
| 1324 | * ".git" is not a directory, then <gitdir> is not inside the |
| 1325 | * cone of <worktree-root>, so set up a second watch to watch |
| 1326 | * the <gitdir> so that we get events for the cookie files. |
| 1327 | */ |
| 1328 | strbuf_init(&state.path_gitdir_watch, 0); |
| 1329 | strbuf_addbuf(&state.path_gitdir_watch, &state.path_worktree_watch); |
| 1330 | strbuf_addstr(&state.path_gitdir_watch, "/.git"); |
| 1331 | if (!is_directory(state.path_gitdir_watch.buf)) { |
| 1332 | strbuf_reset(&state.path_gitdir_watch); |
| 1333 | strbuf_addstr(&state.path_gitdir_watch, |
| 1334 | absolute_path(repo_get_git_dir(the_repository))); |
| 1335 | strbuf_strip_suffix(&state.path_gitdir_watch, "/."); |
| 1336 | state.nr_paths_watching = 2; |
| 1337 | } |
| 1338 | |
| 1339 | /* |
| 1340 | * We will write filesystem syncing cookie files into |
| 1341 | * <gitdir>/<fsmonitor-dir>/<cookie-dir>/<pid>-<seq>. |
| 1342 | * |
| 1343 | * The extra layers of subdirectories here keep us from |
| 1344 | * changing the mtime on ".git/" or ".git/foo/" when we create |
| 1345 | * or delete cookie files. |
| 1346 | * |
| 1347 | * There have been problems with some IDEs that do a |
| 1348 | * non-recursive watch of the ".git/" directory and run a |
| 1349 | * series of commands any time something happens. |
| 1350 | * |
| 1351 | * For example, if we place our cookie files directly in |
| 1352 | * ".git/" or ".git/foo/" then a `git status` (or similar |
| 1353 | * command) from the IDE will cause a cookie file to be |
| 1354 | * created in one of those dirs. This causes the mtime of |
| 1355 | * those dirs to change. This triggers the IDE's watch |
| 1356 | * notification. This triggers the IDE to run those commands |
| 1357 | * again. And the process repeats and the machine never goes |
| 1358 | * idle. |
| 1359 | * |
| 1360 | * Adding the extra layers of subdirectories prevents the |
| 1361 | * mtime of ".git/" and ".git/foo" from changing when a |
| 1362 | * cookie file is created. |
| 1363 | */ |
| 1364 | strbuf_init(&state.path_cookie_prefix, 0); |
| 1365 | strbuf_addbuf(&state.path_cookie_prefix, &state.path_gitdir_watch); |
| 1366 | |
| 1367 | strbuf_addch(&state.path_cookie_prefix, '/'); |
| 1368 | strbuf_addstr(&state.path_cookie_prefix, FSMONITOR_DIR); |
| 1369 | mkdir(state.path_cookie_prefix.buf, 0777); |
| 1370 | |
| 1371 | strbuf_addch(&state.path_cookie_prefix, '/'); |
| 1372 | strbuf_addstr(&state.path_cookie_prefix, FSMONITOR_COOKIE_DIR); |
| 1373 | mkdir(state.path_cookie_prefix.buf, 0777); |
| 1374 | |
| 1375 | strbuf_addch(&state.path_cookie_prefix, '/'); |
| 1376 | |
| 1377 | /* |
| 1378 | * We create a named-pipe or unix domain socket inside of the |
| 1379 | * ".git" directory. (Well, on Windows, we base our named |
| 1380 | * pipe in the NPFS on the absolute path of the git |
| 1381 | * directory.) |
| 1382 | */ |
| 1383 | strbuf_init(&state.path_ipc, 0); |
| 1384 | strbuf_addstr(&state.path_ipc, |
| 1385 | absolute_path(fsmonitor_ipc__get_path(the_repository))); |
| 1386 | |
| 1387 | /* |
| 1388 | * Confirm that we can create platform-specific resources for the |
| 1389 | * filesystem listener before we bother starting all the threads. |
| 1390 | */ |
| 1391 | if (fsm_listen__ctor(&state)) { |
| 1392 | err = error(_("could not initialize listener thread")); |
| 1393 | goto done; |
| 1394 | } |
| 1395 | |
| 1396 | if (fsm_health__ctor(&state)) { |
| 1397 | err = error(_("could not initialize health thread")); |
| 1398 | goto done; |
| 1399 | } |
| 1400 | |
| 1401 | /* |
| 1402 | * CD out of the worktree root directory. |
| 1403 | * |
| 1404 | * The common Git startup mechanism causes our CWD to be the |
| 1405 | * root of the worktree. On Windows, this causes our process |
| 1406 | * to hold a locked handle on the CWD. This prevents the |
| 1407 | * worktree from being moved or deleted while the daemon is |
| 1408 | * running. |
| 1409 | * |
| 1410 | * We assume that our FS and IPC listener threads have either |
| 1411 | * opened all of the handles that they need or will do |
| 1412 | * everything using absolute paths. |
| 1413 | */ |
| 1414 | home = getenv("HOME"); |
| 1415 | if (home && *home && chdir(home)) |
| 1416 | die_errno(_("could not cd home '%s'"), home); |
| 1417 | |
| 1418 | err = fsmonitor_run_daemon_1(&state); |
| 1419 | |
| 1420 | done: |
| 1421 | fsmonitor_free_token_data(state.current_token_data); |
| 1422 | state.current_token_data = NULL; |
| 1423 | pthread_cond_destroy(&state.cookies_cond); |
| 1424 | pthread_mutex_destroy(&state.main_lock); |
| 1425 | { |
| 1426 | struct hashmap_iter iter; |
| 1427 | struct fsmonitor_cookie_item *cookie; |
| 1428 | |
| 1429 | hashmap_for_each_entry(&state.cookies, &iter, cookie, entry) |
| 1430 | free(cookie->name); |
| 1431 | hashmap_clear_and_free(&state.cookies, |
| 1432 | struct fsmonitor_cookie_item, entry); |
| 1433 | } |
| 1434 | fsm_listen__dtor(&state); |
| 1435 | fsm_health__dtor(&state); |
| 1436 | |
| 1437 | ipc_server_free(state.ipc_server_data); |
| 1438 | |
| 1439 | strbuf_release(&state.path_worktree_watch); |
| 1440 | strbuf_release(&state.path_gitdir_watch); |
| 1441 | strbuf_release(&state.path_cookie_prefix); |
| 1442 | strbuf_release(&state.path_ipc); |
| 1443 | strbuf_release(&state.alias.alias); |
| 1444 | strbuf_release(&state.alias.points_to); |
| 1445 | |
| 1446 | return err; |
| 1447 | } |
| 1448 | |
| 1449 | static int try_to_run_foreground_daemon(int detach_console) |
| 1450 | { |
| 1451 | /* |
| 1452 | * Technically, we don't need to probe for an existing daemon |
| 1453 | * process, since we could just call `fsmonitor_run_daemon()` |
| 1454 | * and let it fail if the pipe/socket is busy. |
| 1455 | * |
| 1456 | * However, this method gives us a nicer error message for a |
| 1457 | * common error case. |
| 1458 | */ |
| 1459 | if (fsmonitor_ipc__get_state() == IPC_STATE__LISTENING) |
| 1460 | die(_("fsmonitor--daemon is already running '%s'"), |
| 1461 | the_repository->worktree); |
| 1462 | |
| 1463 | if (fsmonitor__announce_startup) { |
| 1464 | fprintf(stderr, _("running fsmonitor-daemon in '%s'\n"), |
| 1465 | the_repository->worktree); |
| 1466 | fflush(stderr); |
| 1467 | } |
| 1468 | |
| 1469 | if (detach_console) { |
| 1470 | #ifdef GIT_WINDOWS_NATIVE |
| 1471 | FreeConsole(); |
| 1472 | #else |
| 1473 | /* |
| 1474 | * Create a new session so that the daemon is detached |
| 1475 | * from the parent's process group. This prevents |
| 1476 | * shells with job control (e.g. bash with "set -m") |
| 1477 | * from waiting on the daemon when they wait for a |
| 1478 | * foreground command that implicitly spawned it. |
| 1479 | */ |
| 1480 | if (setsid() == -1) |
| 1481 | warning_errno(_("setsid failed")); |
| 1482 | #endif |
| 1483 | } |
| 1484 | |
| 1485 | return !!fsmonitor_run_daemon(); |
| 1486 | } |
| 1487 | |
| 1488 | static start_bg_wait_cb bg_wait_cb; |
| 1489 | |
| 1490 | static int bg_wait_cb(const struct child_process *cp UNUSED, |
| 1491 | void *cb_data UNUSED) |
| 1492 | { |
| 1493 | enum ipc_active_state s = fsmonitor_ipc__get_state(); |
| 1494 | |
| 1495 | switch (s) { |
| 1496 | case IPC_STATE__LISTENING: |
| 1497 | /* child is "ready" */ |
| 1498 | return 0; |
| 1499 | |
| 1500 | case IPC_STATE__NOT_LISTENING: |
| 1501 | case IPC_STATE__PATH_NOT_FOUND: |
| 1502 | /* give child more time */ |
| 1503 | return 1; |
| 1504 | |
| 1505 | default: |
| 1506 | case IPC_STATE__INVALID_PATH: |
| 1507 | case IPC_STATE__OTHER_ERROR: |
| 1508 | /* all the time in world won't help */ |
| 1509 | return -1; |
| 1510 | } |
| 1511 | } |
| 1512 | |
| 1513 | static int try_to_start_background_daemon(void) |
| 1514 | { |
| 1515 | struct child_process cp = CHILD_PROCESS_INIT; |
| 1516 | enum start_bg_result sbgr; |
| 1517 | |
| 1518 | /* |
| 1519 | * Before we try to create a background daemon process, see |
| 1520 | * if a daemon process is already listening. This makes it |
| 1521 | * easier for us to report an already-listening error to the |
| 1522 | * console, since our spawn/daemon can only report the success |
| 1523 | * of creating the background process (and not whether it |
| 1524 | * immediately exited). |
| 1525 | */ |
| 1526 | if (fsmonitor_ipc__get_state() == IPC_STATE__LISTENING) |
| 1527 | die(_("fsmonitor--daemon is already running '%s'"), |
| 1528 | the_repository->worktree); |
| 1529 | |
| 1530 | if (fsmonitor__announce_startup) { |
| 1531 | fprintf(stderr, _("starting fsmonitor-daemon in '%s'\n"), |
| 1532 | the_repository->worktree); |
| 1533 | fflush(stderr); |
| 1534 | } |
| 1535 | |
| 1536 | cp.git_cmd = 1; |
| 1537 | |
| 1538 | strvec_push(&cp.args, "fsmonitor--daemon"); |
| 1539 | strvec_push(&cp.args, "run"); |
| 1540 | strvec_push(&cp.args, "--detach"); |
| 1541 | strvec_pushf(&cp.args, "--ipc-threads=%d", fsmonitor__ipc_threads); |
| 1542 | |
| 1543 | cp.no_stdin = 1; |
| 1544 | cp.no_stdout = 1; |
| 1545 | cp.no_stderr = 1; |
| 1546 | cp.close_fd_above_stderr = 1; |
| 1547 | |
| 1548 | sbgr = start_bg_command(&cp, bg_wait_cb, NULL, |
| 1549 | fsmonitor__start_timeout_sec); |
| 1550 | |
| 1551 | switch (sbgr) { |
| 1552 | case SBGR_READY: |
| 1553 | return 0; |
| 1554 | |
| 1555 | default: |
| 1556 | case SBGR_ERROR: |
| 1557 | case SBGR_CB_ERROR: |
| 1558 | return error(_("daemon failed to start")); |
| 1559 | |
| 1560 | case SBGR_TIMEOUT: |
| 1561 | return error(_("daemon not online yet")); |
| 1562 | |
| 1563 | case SBGR_DIED: |
| 1564 | return error(_("daemon terminated")); |
| 1565 | } |
| 1566 | } |
| 1567 | |
| 1568 | int cmd_fsmonitor__daemon(int argc, |
| 1569 | const char **argv, |
| 1570 | const char *prefix, |
| 1571 | struct repository *repo UNUSED) |
| 1572 | { |
| 1573 | const char *subcmd; |
| 1574 | enum fsmonitor_reason reason; |
| 1575 | int detach_console = 0; |
| 1576 | |
| 1577 | struct option options[] = { |
| 1578 | OPT_BOOL(0, "detach", &detach_console, N_("detach from console")), |
| 1579 | OPT_INTEGER(0, "ipc-threads", |
| 1580 | &fsmonitor__ipc_threads, |
| 1581 | N_("use <n> ipc worker threads")), |
| 1582 | OPT_INTEGER(0, "start-timeout", |
| 1583 | &fsmonitor__start_timeout_sec, |
| 1584 | N_("max seconds to wait for background daemon startup")), |
| 1585 | |
| 1586 | OPT_END() |
| 1587 | }; |
| 1588 | |
| 1589 | repo_config(the_repository, fsmonitor_config, NULL); |
| 1590 | |
| 1591 | argc = parse_options(argc, argv, prefix, options, |
| 1592 | builtin_fsmonitor__daemon_usage, 0); |
| 1593 | if (argc != 1) |
| 1594 | usage_with_options(builtin_fsmonitor__daemon_usage, options); |
| 1595 | subcmd = argv[0]; |
| 1596 | |
| 1597 | if (fsmonitor__ipc_threads < 1) |
| 1598 | die(_("invalid 'ipc-threads' value (%d)"), |
| 1599 | fsmonitor__ipc_threads); |
| 1600 | |
| 1601 | prepare_repo_settings(the_repository); |
| 1602 | /* |
| 1603 | * If the repo is fsmonitor-compatible, explicitly set IPC-mode |
| 1604 | * (without bothering to load the `core.fsmonitor` config settings). |
| 1605 | * |
| 1606 | * If the repo is not compatible, the repo-settings will be set to |
| 1607 | * incompatible rather than IPC, so we can use one of the __get |
| 1608 | * routines to detect the discrepancy. |
| 1609 | */ |
| 1610 | fsm_settings__set_ipc(the_repository); |
| 1611 | |
| 1612 | reason = fsm_settings__get_reason(the_repository); |
| 1613 | if (reason > FSMONITOR_REASON_OK) |
| 1614 | die("%s", |
| 1615 | fsm_settings__get_incompatible_msg(the_repository, |
| 1616 | reason)); |
| 1617 | |
| 1618 | if (!strcmp(subcmd, "start")) |
| 1619 | return !!try_to_start_background_daemon(); |
| 1620 | |
| 1621 | if (!strcmp(subcmd, "run")) |
| 1622 | return !!try_to_run_foreground_daemon(detach_console); |
| 1623 | |
| 1624 | if (!strcmp(subcmd, "stop")) |
| 1625 | return !!do_as_client__send_stop(); |
| 1626 | |
| 1627 | if (!strcmp(subcmd, "status")) |
| 1628 | return !!do_as_client__status(); |
| 1629 | |
| 1630 | die(_("Unhandled subcommand '%s'"), subcmd); |
| 1631 | } |
| 1632 | |
| 1633 | #else |
| 1634 | int cmd_fsmonitor__daemon(int argc, const char **argv, const char *prefix UNUSED, struct repository *repo UNUSED) |
| 1635 | { |
| 1636 | struct option options[] = { |
| 1637 | OPT_END() |
| 1638 | }; |
| 1639 | |
| 1640 | show_usage_with_options_if_asked(argc, argv, |
| 1641 | builtin_fsmonitor__daemon_usage, options); |
| 1642 | |
| 1643 | die(_("fsmonitor--daemon not supported on this platform")); |
| 1644 | } |
| 1645 | #endif |