]>
Commit | Line | Data |
---|---|---|
16d9d617 | 1 | #include "builtin.h" |
0b027f6c | 2 | #include "abspath.h" |
36bf1958 | 3 | #include "alloc.h" |
16d9d617 | 4 | #include "config.h" |
32a8f510 | 5 | #include "environment.h" |
f394e093 | 6 | #include "gettext.h" |
16d9d617 | 7 | #include "parse-options.h" |
68d68646 | 8 | #include "fsmonitor-ll.h" |
16d9d617 | 9 | #include "fsmonitor-ipc.h" |
12fd27df | 10 | #include "fsmonitor-path-utils.h" |
68d68646 | 11 | #include "fsmonitor-settings.h" |
d0605550 | 12 | #include "compat/fsmonitor/fsm-health.h" |
9dcba0ba JH |
13 | #include "compat/fsmonitor/fsm-listen.h" |
14 | #include "fsmonitor--daemon.h" | |
16d9d617 JH |
15 | #include "simple-ipc.h" |
16 | #include "khash.h" | |
518a522f | 17 | #include "pkt-line.h" |
68d68646 | 18 | #include "trace.h" |
74ea5c95 | 19 | #include "trace2.h" |
16d9d617 JH |
20 | |
21 | static const char * const builtin_fsmonitor__daemon_usage[] = { | |
c284e27b | 22 | N_("git fsmonitor--daemon start [<options>]"), |
9dcba0ba | 23 | N_("git fsmonitor--daemon run [<options>]"), |
02cb8b9e AH |
24 | "git fsmonitor--daemon stop", |
25 | "git fsmonitor--daemon status", | |
16d9d617 JH |
26 | NULL |
27 | }; | |
28 | ||
29 | #ifdef HAVE_FSMONITOR_DAEMON_BACKEND | |
9dcba0ba JH |
30 | /* |
31 | * Global state loaded from config. | |
32 | */ | |
33 | #define FSMONITOR__IPC_THREADS "fsmonitor.ipcthreads" | |
34 | static int fsmonitor__ipc_threads = 8; | |
35 | ||
c284e27b JH |
36 | #define FSMONITOR__START_TIMEOUT "fsmonitor.starttimeout" |
37 | static int fsmonitor__start_timeout_sec = 60; | |
38 | ||
9dcba0ba JH |
39 | #define FSMONITOR__ANNOUNCE_STARTUP "fsmonitor.announcestartup" |
40 | static int fsmonitor__announce_startup = 0; | |
41 | ||
a4e7e317 GC |
42 | static int fsmonitor_config(const char *var, const char *value, |
43 | const struct config_context *ctx, void *cb) | |
9dcba0ba JH |
44 | { |
45 | if (!strcmp(var, FSMONITOR__IPC_THREADS)) { | |
8868b1eb | 46 | int i = git_config_int(var, value, ctx->kvi); |
9dcba0ba JH |
47 | if (i < 1) |
48 | return error(_("value of '%s' out of range: %d"), | |
49 | FSMONITOR__IPC_THREADS, i); | |
50 | fsmonitor__ipc_threads = i; | |
51 | return 0; | |
52 | } | |
53 | ||
c284e27b | 54 | if (!strcmp(var, FSMONITOR__START_TIMEOUT)) { |
8868b1eb | 55 | int i = git_config_int(var, value, ctx->kvi); |
c284e27b JH |
56 | if (i < 0) |
57 | return error(_("value of '%s' out of range: %d"), | |
58 | FSMONITOR__START_TIMEOUT, i); | |
59 | fsmonitor__start_timeout_sec = i; | |
60 | return 0; | |
61 | } | |
62 | ||
9dcba0ba JH |
63 | if (!strcmp(var, FSMONITOR__ANNOUNCE_STARTUP)) { |
64 | int is_bool; | |
8868b1eb | 65 | int i = git_config_bool_or_int(var, value, ctx->kvi, &is_bool); |
9dcba0ba JH |
66 | if (i < 0) |
67 | return error(_("value of '%s' not bool or int: %d"), | |
68 | var, i); | |
69 | fsmonitor__announce_startup = i; | |
70 | return 0; | |
71 | } | |
72 | ||
a4e7e317 | 73 | return git_default_config(var, value, ctx, cb); |
9dcba0ba JH |
74 | } |
75 | ||
abc9dbc0 JH |
76 | /* |
77 | * Acting as a CLIENT. | |
78 | * | |
79 | * Send a "quit" command to the `git-fsmonitor--daemon` (if running) | |
80 | * and wait for it to shutdown. | |
81 | */ | |
82 | static int do_as_client__send_stop(void) | |
83 | { | |
84 | struct strbuf answer = STRBUF_INIT; | |
85 | int ret; | |
86 | ||
87 | ret = fsmonitor_ipc__send_command("quit", &answer); | |
88 | ||
89 | /* The quit command does not return any response data. */ | |
90 | strbuf_release(&answer); | |
91 | ||
92 | if (ret) | |
93 | return ret; | |
94 | ||
95 | trace2_region_enter("fsm_client", "polling-for-daemon-exit", NULL); | |
96 | while (fsmonitor_ipc__get_state() == IPC_STATE__LISTENING) | |
97 | sleep_millisec(50); | |
98 | trace2_region_leave("fsm_client", "polling-for-daemon-exit", NULL); | |
99 | ||
100 | return 0; | |
101 | } | |
102 | ||
103 | static int do_as_client__status(void) | |
104 | { | |
105 | enum ipc_active_state state = fsmonitor_ipc__get_state(); | |
106 | ||
107 | switch (state) { | |
108 | case IPC_STATE__LISTENING: | |
109 | printf(_("fsmonitor-daemon is watching '%s'\n"), | |
110 | the_repository->worktree); | |
111 | return 0; | |
112 | ||
113 | default: | |
114 | printf(_("fsmonitor-daemon is not watching '%s'\n"), | |
115 | the_repository->worktree); | |
116 | return 1; | |
117 | } | |
118 | } | |
16d9d617 | 119 | |
b05880d3 JH |
120 | enum fsmonitor_cookie_item_result { |
121 | FCIR_ERROR = -1, /* could not create cookie file ? */ | |
122 | FCIR_INIT, | |
123 | FCIR_SEEN, | |
124 | FCIR_ABORT, | |
125 | }; | |
126 | ||
127 | struct fsmonitor_cookie_item { | |
128 | struct hashmap_entry entry; | |
129 | char *name; | |
130 | enum fsmonitor_cookie_item_result result; | |
131 | }; | |
132 | ||
133 | static int cookies_cmp(const void *data, const struct hashmap_entry *he1, | |
134 | const struct hashmap_entry *he2, const void *keydata) | |
135 | { | |
136 | const struct fsmonitor_cookie_item *a = | |
137 | container_of(he1, const struct fsmonitor_cookie_item, entry); | |
138 | const struct fsmonitor_cookie_item *b = | |
139 | container_of(he2, const struct fsmonitor_cookie_item, entry); | |
140 | ||
141 | return strcmp(a->name, keydata ? keydata : b->name); | |
142 | } | |
143 | ||
144 | static enum fsmonitor_cookie_item_result with_lock__wait_for_cookie( | |
145 | struct fsmonitor_daemon_state *state) | |
146 | { | |
147 | /* assert current thread holding state->main_lock */ | |
148 | ||
149 | int fd; | |
150 | struct fsmonitor_cookie_item *cookie; | |
151 | struct strbuf cookie_pathname = STRBUF_INIT; | |
152 | struct strbuf cookie_filename = STRBUF_INIT; | |
153 | enum fsmonitor_cookie_item_result result; | |
154 | int my_cookie_seq; | |
155 | ||
156 | CALLOC_ARRAY(cookie, 1); | |
157 | ||
158 | my_cookie_seq = state->cookie_seq++; | |
159 | ||
160 | strbuf_addf(&cookie_filename, "%i-%i", getpid(), my_cookie_seq); | |
161 | ||
162 | strbuf_addbuf(&cookie_pathname, &state->path_cookie_prefix); | |
163 | strbuf_addbuf(&cookie_pathname, &cookie_filename); | |
164 | ||
165 | cookie->name = strbuf_detach(&cookie_filename, NULL); | |
166 | cookie->result = FCIR_INIT; | |
167 | hashmap_entry_init(&cookie->entry, strhash(cookie->name)); | |
168 | ||
169 | hashmap_add(&state->cookies, &cookie->entry); | |
170 | ||
171 | trace_printf_key(&trace_fsmonitor, "cookie-wait: '%s' '%s'", | |
172 | cookie->name, cookie_pathname.buf); | |
173 | ||
174 | /* | |
175 | * Create the cookie file on disk and then wait for a notification | |
176 | * that the listener thread has seen it. | |
177 | */ | |
178 | fd = open(cookie_pathname.buf, O_WRONLY | O_CREAT | O_EXCL, 0600); | |
179 | if (fd < 0) { | |
180 | error_errno(_("could not create fsmonitor cookie '%s'"), | |
181 | cookie->name); | |
182 | ||
183 | cookie->result = FCIR_ERROR; | |
184 | goto done; | |
185 | } | |
186 | ||
187 | /* | |
188 | * Technically, close() and unlink() can fail, but we don't | |
189 | * care here. We only created the file to trigger a watch | |
190 | * event from the FS to know that when we're up to date. | |
191 | */ | |
192 | close(fd); | |
193 | unlink(cookie_pathname.buf); | |
194 | ||
195 | /* | |
196 | * Technically, this is an infinite wait (well, unless another | |
197 | * thread sends us an abort). I'd like to change this to | |
198 | * use `pthread_cond_timedwait()` and return an error/timeout | |
199 | * and let the caller do the trivial response thing, but we | |
200 | * don't have that routine in our thread-utils. | |
201 | * | |
202 | * After extensive beta testing I'm not really worried about | |
203 | * this. Also note that the above open() and unlink() calls | |
204 | * will cause at least two FS events on that path, so the odds | |
205 | * of getting stuck are pretty slim. | |
206 | */ | |
207 | while (cookie->result == FCIR_INIT) | |
208 | pthread_cond_wait(&state->cookies_cond, | |
209 | &state->main_lock); | |
210 | ||
211 | done: | |
212 | hashmap_remove(&state->cookies, &cookie->entry, NULL); | |
213 | ||
214 | result = cookie->result; | |
215 | ||
216 | free(cookie->name); | |
217 | free(cookie); | |
218 | strbuf_release(&cookie_pathname); | |
219 | ||
220 | return result; | |
221 | } | |
222 | ||
223 | /* | |
224 | * Mark these cookies as _SEEN and wake up the corresponding client threads. | |
225 | */ | |
226 | static void with_lock__mark_cookies_seen(struct fsmonitor_daemon_state *state, | |
227 | const struct string_list *cookie_names) | |
228 | { | |
229 | /* assert current thread holding state->main_lock */ | |
230 | ||
231 | int k; | |
232 | int nr_seen = 0; | |
233 | ||
234 | for (k = 0; k < cookie_names->nr; k++) { | |
235 | struct fsmonitor_cookie_item key; | |
236 | struct fsmonitor_cookie_item *cookie; | |
237 | ||
238 | key.name = cookie_names->items[k].string; | |
239 | hashmap_entry_init(&key.entry, strhash(key.name)); | |
240 | ||
241 | cookie = hashmap_get_entry(&state->cookies, &key, entry, NULL); | |
242 | if (cookie) { | |
243 | trace_printf_key(&trace_fsmonitor, "cookie-seen: '%s'", | |
244 | cookie->name); | |
245 | cookie->result = FCIR_SEEN; | |
246 | nr_seen++; | |
247 | } | |
248 | } | |
249 | ||
250 | if (nr_seen) | |
251 | pthread_cond_broadcast(&state->cookies_cond); | |
252 | } | |
253 | ||
254 | /* | |
255 | * Set _ABORT on all pending cookies and wake up all client threads. | |
256 | */ | |
257 | static void with_lock__abort_all_cookies(struct fsmonitor_daemon_state *state) | |
258 | { | |
259 | /* assert current thread holding state->main_lock */ | |
260 | ||
261 | struct hashmap_iter iter; | |
262 | struct fsmonitor_cookie_item *cookie; | |
263 | int nr_aborted = 0; | |
264 | ||
265 | hashmap_for_each_entry(&state->cookies, &iter, cookie, entry) { | |
266 | trace_printf_key(&trace_fsmonitor, "cookie-abort: '%s'", | |
267 | cookie->name); | |
268 | cookie->result = FCIR_ABORT; | |
269 | nr_aborted++; | |
270 | } | |
271 | ||
272 | if (nr_aborted) | |
273 | pthread_cond_broadcast(&state->cookies_cond); | |
274 | } | |
275 | ||
aeef767a JH |
276 | /* |
277 | * Requests to and from a FSMonitor Protocol V2 provider use an opaque | |
278 | * "token" as a virtual timestamp. Clients can request a summary of all | |
279 | * created/deleted/modified files relative to a token. In the response, | |
280 | * clients receive a new token for the next (relative) request. | |
281 | * | |
282 | * | |
283 | * Token Format | |
284 | * ============ | |
285 | * | |
286 | * The contents of the token are private and provider-specific. | |
287 | * | |
288 | * For the built-in fsmonitor--daemon, we define a token as follows: | |
289 | * | |
290 | * "builtin" ":" <token_id> ":" <sequence_nr> | |
291 | * | |
292 | * The "builtin" prefix is used as a namespace to avoid conflicts | |
293 | * with other providers (such as Watchman). | |
294 | * | |
295 | * The <token_id> is an arbitrary OPAQUE string, such as a GUID, | |
296 | * UUID, or {timestamp,pid}. It is used to group all filesystem | |
297 | * events that happened while the daemon was monitoring (and in-sync | |
298 | * with the filesystem). | |
299 | * | |
300 | * Unlike FSMonitor Protocol V1, it is not defined as a timestamp | |
301 | * and does not define less-than/greater-than relationships. | |
302 | * (There are too many race conditions to rely on file system | |
303 | * event timestamps.) | |
304 | * | |
305 | * The <sequence_nr> is a simple integer incremented whenever the | |
306 | * daemon needs to make its state public. For example, if 1000 file | |
307 | * system events come in, but no clients have requested the data, | |
308 | * the daemon can continue to accumulate file changes in the same | |
309 | * bin and does not need to advance the sequence number. However, | |
310 | * as soon as a client does arrive, the daemon needs to start a new | |
311 | * bin and increment the sequence number. | |
312 | * | |
313 | * The sequence number serves as the boundary between 2 sets | |
314 | * of bins -- the older ones that the client has already seen | |
315 | * and the newer ones that it hasn't. | |
316 | * | |
317 | * When a new <token_id> is created, the <sequence_nr> is reset to | |
318 | * zero. | |
319 | * | |
320 | * | |
321 | * About Token Ids | |
322 | * =============== | |
323 | * | |
324 | * A new token_id is created: | |
325 | * | |
326 | * [1] each time the daemon is started. | |
327 | * | |
328 | * [2] any time that the daemon must re-sync with the filesystem | |
329 | * (such as when the kernel drops or we miss events on a very | |
330 | * active volume). | |
331 | * | |
332 | * [3] in response to a client "flush" command (for dropped event | |
333 | * testing). | |
334 | * | |
335 | * When a new token_id is created, the daemon is free to discard all | |
336 | * cached filesystem events associated with any previous token_ids. | |
337 | * Events associated with a non-current token_id will never be sent | |
338 | * to a client. A token_id change implicitly means that the daemon | |
339 | * has gap in its event history. | |
340 | * | |
341 | * Therefore, clients that present a token with a stale (non-current) | |
342 | * token_id will always be given a trivial response. | |
343 | */ | |
344 | struct fsmonitor_token_data { | |
345 | struct strbuf token_id; | |
346 | struct fsmonitor_batch *batch_head; | |
347 | struct fsmonitor_batch *batch_tail; | |
348 | uint64_t client_ref_count; | |
349 | }; | |
350 | ||
bec486b9 JH |
351 | struct fsmonitor_batch { |
352 | struct fsmonitor_batch *next; | |
353 | uint64_t batch_seq_nr; | |
354 | const char **interned_paths; | |
355 | size_t nr, alloc; | |
356 | time_t pinned_time; | |
357 | }; | |
358 | ||
aeef767a JH |
359 | static struct fsmonitor_token_data *fsmonitor_new_token_data(void) |
360 | { | |
361 | static int test_env_value = -1; | |
362 | static uint64_t flush_count = 0; | |
363 | struct fsmonitor_token_data *token; | |
bec486b9 | 364 | struct fsmonitor_batch *batch; |
aeef767a JH |
365 | |
366 | CALLOC_ARRAY(token, 1); | |
bec486b9 | 367 | batch = fsmonitor_batch__new(); |
aeef767a JH |
368 | |
369 | strbuf_init(&token->token_id, 0); | |
bec486b9 JH |
370 | token->batch_head = batch; |
371 | token->batch_tail = batch; | |
aeef767a JH |
372 | token->client_ref_count = 0; |
373 | ||
374 | if (test_env_value < 0) | |
375 | test_env_value = git_env_bool("GIT_TEST_FSMONITOR_TOKEN", 0); | |
376 | ||
377 | if (!test_env_value) { | |
378 | struct timeval tv; | |
379 | struct tm tm; | |
380 | time_t secs; | |
381 | ||
382 | gettimeofday(&tv, NULL); | |
383 | secs = tv.tv_sec; | |
384 | gmtime_r(&secs, &tm); | |
385 | ||
386 | strbuf_addf(&token->token_id, | |
387 | "%"PRIu64".%d.%4d%02d%02dT%02d%02d%02d.%06ldZ", | |
388 | flush_count++, | |
389 | getpid(), | |
390 | tm.tm_year + 1900, tm.tm_mon + 1, tm.tm_mday, | |
391 | tm.tm_hour, tm.tm_min, tm.tm_sec, | |
392 | (long)tv.tv_usec); | |
393 | } else { | |
394 | strbuf_addf(&token->token_id, "test_%08x", test_env_value++); | |
395 | } | |
396 | ||
bec486b9 JH |
397 | /* |
398 | * We created a new <token_id> and are starting a new series | |
399 | * of tokens with a zero <seq_nr>. | |
400 | * | |
401 | * Since clients cannot guess our new (non test) <token_id> | |
402 | * they will always receive a trivial response (because of the | |
403 | * mismatch on the <token_id>). The trivial response will | |
404 | * tell them our new <token_id> so that subsequent requests | |
405 | * will be relative to our new series. (And when sending that | |
406 | * response, we pin the current head of the batch list.) | |
407 | * | |
408 | * Even if the client correctly guesses the <token_id>, their | |
409 | * request of "builtin:<token_id>:0" asks for all changes MORE | |
410 | * RECENT than batch/bin 0. | |
411 | * | |
412 | * This implies that it is a waste to accumulate paths in the | |
413 | * initial batch/bin (because they will never be transmitted). | |
414 | * | |
415 | * So the daemon could be running for days and watching the | |
416 | * file system, but doesn't need to actually accumulate any | |
417 | * paths UNTIL we need to set a reference point for a later | |
418 | * relative request. | |
419 | * | |
420 | * However, it is very useful for testing to always have a | |
421 | * reference point set. Pin batch 0 to force early file system | |
422 | * events to accumulate. | |
423 | */ | |
424 | if (test_env_value) | |
425 | batch->pinned_time = time(NULL); | |
426 | ||
aeef767a JH |
427 | return token; |
428 | } | |
429 | ||
bec486b9 JH |
430 | struct fsmonitor_batch *fsmonitor_batch__new(void) |
431 | { | |
432 | struct fsmonitor_batch *batch; | |
433 | ||
434 | CALLOC_ARRAY(batch, 1); | |
435 | ||
436 | return batch; | |
437 | } | |
438 | ||
439 | void fsmonitor_batch__free_list(struct fsmonitor_batch *batch) | |
440 | { | |
441 | while (batch) { | |
442 | struct fsmonitor_batch *next = batch->next; | |
443 | ||
444 | /* | |
445 | * The actual strings within the array of this batch | |
446 | * are interned, so we don't own them. We only own | |
447 | * the array. | |
448 | */ | |
449 | free(batch->interned_paths); | |
450 | free(batch); | |
451 | ||
452 | batch = next; | |
453 | } | |
454 | } | |
455 | ||
456 | void fsmonitor_batch__add_path(struct fsmonitor_batch *batch, | |
457 | const char *path) | |
458 | { | |
459 | const char *interned_path = strintern(path); | |
460 | ||
461 | trace_printf_key(&trace_fsmonitor, "event: %s", interned_path); | |
462 | ||
463 | ALLOC_GROW(batch->interned_paths, batch->nr + 1, batch->alloc); | |
464 | batch->interned_paths[batch->nr++] = interned_path; | |
465 | } | |
466 | ||
467 | static void fsmonitor_batch__combine(struct fsmonitor_batch *batch_dest, | |
468 | const struct fsmonitor_batch *batch_src) | |
469 | { | |
470 | size_t k; | |
471 | ||
472 | ALLOC_GROW(batch_dest->interned_paths, | |
473 | batch_dest->nr + batch_src->nr + 1, | |
474 | batch_dest->alloc); | |
475 | ||
476 | for (k = 0; k < batch_src->nr; k++) | |
477 | batch_dest->interned_paths[batch_dest->nr++] = | |
478 | batch_src->interned_paths[k]; | |
479 | } | |
480 | ||
50c725d6 JH |
481 | /* |
482 | * To keep the batch list from growing unbounded in response to filesystem | |
483 | * activity, we try to truncate old batches from the end of the list as | |
484 | * they become irrelevant. | |
485 | * | |
486 | * We assume that the .git/index will be updated with the most recent token | |
487 | * any time the index is updated. And future commands will only ask for | |
488 | * recent changes *since* that new token. So as tokens advance into the | |
489 | * future, older batch items will never be requested/needed. So we can | |
490 | * truncate them without loss of functionality. | |
491 | * | |
492 | * However, multiple commands may be talking to the daemon concurrently | |
493 | * or perform a slow command, so a little "token skew" is possible. | |
494 | * Therefore, we want this to be a little bit lazy and have a generous | |
495 | * delay. | |
496 | * | |
497 | * The current reader thread walked backwards in time from `token->batch_head` | |
498 | * back to `batch_marker` somewhere in the middle of the batch list. | |
499 | * | |
500 | * Let's walk backwards in time from that marker an arbitrary delay | |
501 | * and truncate the list there. Note that these timestamps are completely | |
502 | * artificial (based on when we pinned the batch item) and not on any | |
503 | * filesystem activity. | |
504 | * | |
505 | * Return the obsolete portion of the list after we have removed it from | |
506 | * the official list so that the caller can free it after leaving the lock. | |
507 | */ | |
508 | #define MY_TIME_DELAY_SECONDS (5 * 60) /* seconds */ | |
509 | ||
510 | static struct fsmonitor_batch *with_lock__truncate_old_batches( | |
511 | struct fsmonitor_daemon_state *state, | |
512 | const struct fsmonitor_batch *batch_marker) | |
513 | { | |
514 | /* assert current thread holding state->main_lock */ | |
515 | ||
516 | const struct fsmonitor_batch *batch; | |
517 | struct fsmonitor_batch *remainder; | |
518 | ||
519 | if (!batch_marker) | |
520 | return NULL; | |
521 | ||
522 | trace_printf_key(&trace_fsmonitor, "Truncate: mark (%"PRIu64",%"PRIu64")", | |
523 | batch_marker->batch_seq_nr, | |
524 | (uint64_t)batch_marker->pinned_time); | |
525 | ||
526 | for (batch = batch_marker; batch; batch = batch->next) { | |
527 | time_t t; | |
528 | ||
529 | if (!batch->pinned_time) /* an overflow batch */ | |
530 | continue; | |
531 | ||
532 | t = batch->pinned_time + MY_TIME_DELAY_SECONDS; | |
533 | if (t > batch_marker->pinned_time) /* too close to marker */ | |
534 | continue; | |
535 | ||
536 | goto truncate_past_here; | |
537 | } | |
538 | ||
539 | return NULL; | |
540 | ||
541 | truncate_past_here: | |
542 | state->current_token_data->batch_tail = (struct fsmonitor_batch *)batch; | |
543 | ||
544 | remainder = ((struct fsmonitor_batch *)batch)->next; | |
545 | ((struct fsmonitor_batch *)batch)->next = NULL; | |
546 | ||
547 | return remainder; | |
548 | } | |
549 | ||
bec486b9 JH |
550 | static void fsmonitor_free_token_data(struct fsmonitor_token_data *token) |
551 | { | |
552 | if (!token) | |
553 | return; | |
554 | ||
555 | assert(token->client_ref_count == 0); | |
556 | ||
557 | strbuf_release(&token->token_id); | |
558 | ||
559 | fsmonitor_batch__free_list(token->batch_head); | |
560 | ||
561 | free(token); | |
562 | } | |
563 | ||
564 | /* | |
565 | * Flush all of our cached data about the filesystem. Call this if we | |
566 | * lose sync with the filesystem and miss some notification events. | |
567 | * | |
568 | * [1] If we are missing events, then we no longer have a complete | |
569 | * history of the directory (relative to our current start token). | |
570 | * We should create a new token and start fresh (as if we just | |
571 | * booted up). | |
572 | * | |
b05880d3 JH |
573 | * [2] Some of those lost events may have been for cookie files. We |
574 | * should assume the worst and abort them rather letting them starve. | |
575 | * | |
bec486b9 JH |
576 | * If there are no concurrent threads reading the current token data |
577 | * series, we can free it now. Otherwise, let the last reader free | |
578 | * it. | |
579 | * | |
580 | * Either way, the old token data series is no longer associated with | |
581 | * our state data. | |
582 | */ | |
583 | static void with_lock__do_force_resync(struct fsmonitor_daemon_state *state) | |
584 | { | |
585 | /* assert current thread holding state->main_lock */ | |
586 | ||
587 | struct fsmonitor_token_data *free_me = NULL; | |
588 | struct fsmonitor_token_data *new_one = NULL; | |
589 | ||
590 | new_one = fsmonitor_new_token_data(); | |
591 | ||
592 | if (state->current_token_data->client_ref_count == 0) | |
593 | free_me = state->current_token_data; | |
594 | state->current_token_data = new_one; | |
595 | ||
596 | fsmonitor_free_token_data(free_me); | |
b05880d3 JH |
597 | |
598 | with_lock__abort_all_cookies(state); | |
bec486b9 JH |
599 | } |
600 | ||
601 | void fsmonitor_force_resync(struct fsmonitor_daemon_state *state) | |
602 | { | |
603 | pthread_mutex_lock(&state->main_lock); | |
604 | with_lock__do_force_resync(state); | |
605 | pthread_mutex_unlock(&state->main_lock); | |
606 | } | |
607 | ||
518a522f JH |
608 | /* |
609 | * Format an opaque token string to send to the client. | |
610 | */ | |
611 | static void with_lock__format_response_token( | |
612 | struct strbuf *response_token, | |
613 | const struct strbuf *response_token_id, | |
614 | const struct fsmonitor_batch *batch) | |
615 | { | |
616 | /* assert current thread holding state->main_lock */ | |
617 | ||
618 | strbuf_reset(response_token); | |
619 | strbuf_addf(response_token, "builtin:%s:%"PRIu64, | |
620 | response_token_id->buf, batch->batch_seq_nr); | |
621 | } | |
622 | ||
623 | /* | |
624 | * Parse an opaque token from the client. | |
625 | * Returns -1 on error. | |
626 | */ | |
627 | static int fsmonitor_parse_client_token(const char *buf_token, | |
628 | struct strbuf *requested_token_id, | |
629 | uint64_t *seq_nr) | |
630 | { | |
631 | const char *p; | |
632 | char *p_end; | |
633 | ||
634 | strbuf_reset(requested_token_id); | |
635 | *seq_nr = 0; | |
636 | ||
637 | if (!skip_prefix(buf_token, "builtin:", &p)) | |
638 | return -1; | |
639 | ||
640 | while (*p && *p != ':') | |
641 | strbuf_addch(requested_token_id, *p++); | |
642 | if (!*p++) | |
643 | return -1; | |
644 | ||
645 | *seq_nr = (uint64_t)strtoumax(p, &p_end, 10); | |
646 | if (*p_end) | |
647 | return -1; | |
648 | ||
649 | return 0; | |
650 | } | |
651 | ||
652 | KHASH_INIT(str, const char *, int, 0, kh_str_hash_func, kh_str_hash_equal) | |
653 | ||
654 | static int do_handle_client(struct fsmonitor_daemon_state *state, | |
655 | const char *command, | |
656 | ipc_server_reply_cb *reply, | |
657 | struct ipc_server_reply_data *reply_data) | |
658 | { | |
659 | struct fsmonitor_token_data *token_data = NULL; | |
660 | struct strbuf response_token = STRBUF_INIT; | |
661 | struct strbuf requested_token_id = STRBUF_INIT; | |
662 | struct strbuf payload = STRBUF_INIT; | |
663 | uint64_t requested_oldest_seq_nr = 0; | |
664 | uint64_t total_response_len = 0; | |
665 | const char *p; | |
666 | const struct fsmonitor_batch *batch_head; | |
667 | const struct fsmonitor_batch *batch; | |
50c725d6 | 668 | struct fsmonitor_batch *remainder = NULL; |
518a522f JH |
669 | intmax_t count = 0, duplicates = 0; |
670 | kh_str_t *shown; | |
671 | int hash_ret; | |
672 | int do_trivial = 0; | |
673 | int do_flush = 0; | |
b05880d3 JH |
674 | int do_cookie = 0; |
675 | enum fsmonitor_cookie_item_result cookie_result; | |
518a522f JH |
676 | |
677 | /* | |
678 | * We expect `command` to be of the form: | |
679 | * | |
680 | * <command> := quit NUL | |
681 | * | flush NUL | |
682 | * | <V1-time-since-epoch-ns> NUL | |
683 | * | <V2-opaque-fsmonitor-token> NUL | |
684 | */ | |
685 | ||
686 | if (!strcmp(command, "quit")) { | |
687 | /* | |
688 | * A client has requested over the socket/pipe that the | |
689 | * daemon shutdown. | |
690 | * | |
691 | * Tell the IPC thread pool to shutdown (which completes | |
692 | * the await in the main thread (which can stop the | |
693 | * fsmonitor listener thread)). | |
694 | * | |
695 | * There is no reply to the client. | |
696 | */ | |
697 | return SIMPLE_IPC_QUIT; | |
698 | ||
699 | } else if (!strcmp(command, "flush")) { | |
700 | /* | |
701 | * Flush all of our cached data and generate a new token | |
702 | * just like if we lost sync with the filesystem. | |
703 | * | |
704 | * Then send a trivial response using the new token. | |
705 | */ | |
706 | do_flush = 1; | |
707 | do_trivial = 1; | |
708 | ||
709 | } else if (!skip_prefix(command, "builtin:", &p)) { | |
710 | /* assume V1 timestamp or garbage */ | |
711 | ||
712 | char *p_end; | |
713 | ||
714 | strtoumax(command, &p_end, 10); | |
715 | trace_printf_key(&trace_fsmonitor, | |
716 | ((*p_end) ? | |
717 | "fsmonitor: invalid command line '%s'" : | |
718 | "fsmonitor: unsupported V1 protocol '%s'"), | |
719 | command); | |
720 | do_trivial = 1; | |
6692d454 | 721 | do_cookie = 1; |
518a522f JH |
722 | |
723 | } else { | |
724 | /* We have "builtin:*" */ | |
725 | if (fsmonitor_parse_client_token(command, &requested_token_id, | |
726 | &requested_oldest_seq_nr)) { | |
727 | trace_printf_key(&trace_fsmonitor, | |
728 | "fsmonitor: invalid V2 protocol token '%s'", | |
729 | command); | |
730 | do_trivial = 1; | |
6692d454 | 731 | do_cookie = 1; |
518a522f JH |
732 | |
733 | } else { | |
734 | /* | |
735 | * We have a V2 valid token: | |
736 | * "builtin:<token_id>:<seq_nr>" | |
737 | */ | |
b05880d3 | 738 | do_cookie = 1; |
518a522f JH |
739 | } |
740 | } | |
741 | ||
742 | pthread_mutex_lock(&state->main_lock); | |
743 | ||
744 | if (!state->current_token_data) | |
745 | BUG("fsmonitor state does not have a current token"); | |
746 | ||
b05880d3 JH |
747 | /* |
748 | * Write a cookie file inside the directory being watched in | |
749 | * an effort to flush out existing filesystem events that we | |
750 | * actually care about. Suspend this client thread until we | |
751 | * see the filesystem events for this cookie file. | |
752 | * | |
753 | * Creating the cookie lets us guarantee that our FS listener | |
754 | * thread has drained the kernel queue and we are caught up | |
755 | * with the kernel. | |
756 | * | |
757 | * If we cannot create the cookie (or otherwise guarantee that | |
758 | * we are caught up), we send a trivial response. We have to | |
759 | * assume that there might be some very, very recent activity | |
760 | * on the FS still in flight. | |
761 | */ | |
762 | if (do_cookie) { | |
763 | cookie_result = with_lock__wait_for_cookie(state); | |
764 | if (cookie_result != FCIR_SEEN) { | |
765 | error(_("fsmonitor: cookie_result '%d' != SEEN"), | |
766 | cookie_result); | |
767 | do_trivial = 1; | |
768 | } | |
769 | } | |
770 | ||
518a522f JH |
771 | if (do_flush) |
772 | with_lock__do_force_resync(state); | |
773 | ||
774 | /* | |
775 | * We mark the current head of the batch list as "pinned" so | |
776 | * that the listener thread will treat this item as read-only | |
777 | * (and prevent any more paths from being added to it) from | |
778 | * now on. | |
779 | */ | |
780 | token_data = state->current_token_data; | |
781 | batch_head = token_data->batch_head; | |
782 | ((struct fsmonitor_batch *)batch_head)->pinned_time = time(NULL); | |
783 | ||
784 | /* | |
785 | * FSMonitor Protocol V2 requires that we send a response header | |
786 | * with a "new current token" and then all of the paths that changed | |
787 | * since the "requested token". We send the seq_nr of the just-pinned | |
788 | * head batch so that future requests from a client will be relative | |
789 | * to it. | |
790 | */ | |
791 | with_lock__format_response_token(&response_token, | |
792 | &token_data->token_id, batch_head); | |
793 | ||
794 | reply(reply_data, response_token.buf, response_token.len + 1); | |
795 | total_response_len += response_token.len + 1; | |
796 | ||
797 | trace2_data_string("fsmonitor", the_repository, "response/token", | |
798 | response_token.buf); | |
799 | trace_printf_key(&trace_fsmonitor, "response token: %s", | |
800 | response_token.buf); | |
801 | ||
802 | if (!do_trivial) { | |
803 | if (strcmp(requested_token_id.buf, token_data->token_id.buf)) { | |
804 | /* | |
805 | * The client last spoke to a different daemon | |
806 | * instance -OR- the daemon had to resync with | |
807 | * the filesystem (and lost events), so reject. | |
808 | */ | |
809 | trace2_data_string("fsmonitor", the_repository, | |
810 | "response/token", "different"); | |
811 | do_trivial = 1; | |
812 | ||
813 | } else if (requested_oldest_seq_nr < | |
814 | token_data->batch_tail->batch_seq_nr) { | |
815 | /* | |
816 | * The client wants older events than we have for | |
817 | * this token_id. This means that the end of our | |
818 | * batch list was truncated and we cannot give the | |
819 | * client a complete snapshot relative to their | |
820 | * request. | |
821 | */ | |
822 | trace_printf_key(&trace_fsmonitor, | |
823 | "client requested truncated data"); | |
824 | do_trivial = 1; | |
825 | } | |
826 | } | |
827 | ||
828 | if (do_trivial) { | |
829 | pthread_mutex_unlock(&state->main_lock); | |
830 | ||
831 | reply(reply_data, "/", 2); | |
832 | ||
833 | trace2_data_intmax("fsmonitor", the_repository, | |
834 | "response/trivial", 1); | |
835 | ||
836 | goto cleanup; | |
837 | } | |
838 | ||
839 | /* | |
840 | * We're going to hold onto a pointer to the current | |
841 | * token-data while we walk the list of batches of files. | |
842 | * During this time, we will NOT be under the lock. | |
843 | * So we ref-count it. | |
844 | * | |
845 | * This allows the listener thread to continue prepending | |
846 | * new batches of items to the token-data (which we'll ignore). | |
847 | * | |
848 | * AND it allows the listener thread to do a token-reset | |
849 | * (and install a new `current_token_data`). | |
850 | */ | |
851 | token_data->client_ref_count++; | |
852 | ||
853 | pthread_mutex_unlock(&state->main_lock); | |
854 | ||
855 | /* | |
856 | * The client request is relative to the token that they sent, | |
857 | * so walk the batch list backwards from the current head back | |
858 | * to the batch (sequence number) they named. | |
859 | * | |
860 | * We use khash to de-dup the list of pathnames. | |
861 | * | |
862 | * NEEDSWORK: each batch contains a list of interned strings, | |
863 | * so we only need to do pointer comparisons here to build the | |
864 | * hash table. Currently, we're still comparing the string | |
865 | * values. | |
866 | */ | |
867 | shown = kh_init_str(); | |
868 | for (batch = batch_head; | |
869 | batch && batch->batch_seq_nr > requested_oldest_seq_nr; | |
870 | batch = batch->next) { | |
871 | size_t k; | |
872 | ||
873 | for (k = 0; k < batch->nr; k++) { | |
874 | const char *s = batch->interned_paths[k]; | |
875 | size_t s_len; | |
876 | ||
877 | if (kh_get_str(shown, s) != kh_end(shown)) | |
878 | duplicates++; | |
879 | else { | |
880 | kh_put_str(shown, s, &hash_ret); | |
881 | ||
882 | trace_printf_key(&trace_fsmonitor, | |
883 | "send[%"PRIuMAX"]: %s", | |
884 | count, s); | |
885 | ||
886 | /* Each path gets written with a trailing NUL */ | |
887 | s_len = strlen(s) + 1; | |
888 | ||
889 | if (payload.len + s_len >= | |
890 | LARGE_PACKET_DATA_MAX) { | |
891 | reply(reply_data, payload.buf, | |
892 | payload.len); | |
893 | total_response_len += payload.len; | |
894 | strbuf_reset(&payload); | |
895 | } | |
896 | ||
897 | strbuf_add(&payload, s, s_len); | |
898 | count++; | |
899 | } | |
900 | } | |
901 | } | |
902 | ||
903 | if (payload.len) { | |
904 | reply(reply_data, payload.buf, payload.len); | |
905 | total_response_len += payload.len; | |
906 | } | |
907 | ||
908 | kh_release_str(shown); | |
909 | ||
910 | pthread_mutex_lock(&state->main_lock); | |
911 | ||
912 | if (token_data->client_ref_count > 0) | |
913 | token_data->client_ref_count--; | |
914 | ||
915 | if (token_data->client_ref_count == 0) { | |
916 | if (token_data != state->current_token_data) { | |
917 | /* | |
918 | * The listener thread did a token-reset while we were | |
919 | * walking the batch list. Therefore, this token is | |
920 | * stale and can be discarded completely. If we are | |
921 | * the last reader thread using this token, we own | |
922 | * that work. | |
923 | */ | |
924 | fsmonitor_free_token_data(token_data); | |
50c725d6 JH |
925 | } else if (batch) { |
926 | /* | |
927 | * We are holding the lock and are the only | |
928 | * reader of the ref-counted portion of the | |
929 | * list, so we get the honor of seeing if the | |
930 | * list can be truncated to save memory. | |
931 | * | |
932 | * The main loop did not walk to the end of the | |
933 | * list, so this batch is the first item in the | |
934 | * batch-list that is older than the requested | |
935 | * end-point sequence number. See if the tail | |
936 | * end of the list is obsolete. | |
937 | */ | |
938 | remainder = with_lock__truncate_old_batches(state, | |
939 | batch); | |
518a522f JH |
940 | } |
941 | } | |
942 | ||
943 | pthread_mutex_unlock(&state->main_lock); | |
944 | ||
50c725d6 JH |
945 | if (remainder) |
946 | fsmonitor_batch__free_list(remainder); | |
947 | ||
518a522f JH |
948 | trace2_data_intmax("fsmonitor", the_repository, "response/length", total_response_len); |
949 | trace2_data_intmax("fsmonitor", the_repository, "response/count/files", count); | |
950 | trace2_data_intmax("fsmonitor", the_repository, "response/count/duplicates", duplicates); | |
951 | ||
952 | cleanup: | |
953 | strbuf_release(&response_token); | |
954 | strbuf_release(&requested_token_id); | |
955 | strbuf_release(&payload); | |
956 | ||
957 | return 0; | |
958 | } | |
959 | ||
9dcba0ba JH |
960 | static ipc_server_application_cb handle_client; |
961 | ||
962 | static int handle_client(void *data, | |
963 | const char *command, size_t command_len, | |
964 | ipc_server_reply_cb *reply, | |
965 | struct ipc_server_reply_data *reply_data) | |
966 | { | |
518a522f | 967 | struct fsmonitor_daemon_state *state = data; |
9dcba0ba JH |
968 | int result; |
969 | ||
970 | /* | |
971 | * The Simple IPC API now supports {char*, len} arguments, but | |
972 | * FSMonitor always uses proper null-terminated strings, so | |
973 | * we can ignore the command_len argument. (Trust, but verify.) | |
974 | */ | |
975 | if (command_len != strlen(command)) | |
976 | BUG("FSMonitor assumes text messages"); | |
977 | ||
518a522f JH |
978 | trace_printf_key(&trace_fsmonitor, "requested token: %s", command); |
979 | ||
9dcba0ba JH |
980 | trace2_region_enter("fsmonitor", "handle_client", the_repository); |
981 | trace2_data_string("fsmonitor", the_repository, "request", command); | |
982 | ||
518a522f | 983 | result = do_handle_client(state, command, reply, reply_data); |
9dcba0ba JH |
984 | |
985 | trace2_region_leave("fsmonitor", "handle_client", the_repository); | |
986 | ||
987 | return result; | |
988 | } | |
989 | ||
b05880d3 JH |
990 | #define FSMONITOR_DIR "fsmonitor--daemon" |
991 | #define FSMONITOR_COOKIE_DIR "cookies" | |
992 | #define FSMONITOR_COOKIE_PREFIX (FSMONITOR_DIR "/" FSMONITOR_COOKIE_DIR "/") | |
0ae7a1d9 JH |
993 | |
994 | enum fsmonitor_path_type fsmonitor_classify_path_workdir_relative( | |
995 | const char *rel) | |
996 | { | |
997 | if (fspathncmp(rel, ".git", 4)) | |
998 | return IS_WORKDIR_PATH; | |
999 | rel += 4; | |
1000 | ||
1001 | if (!*rel) | |
1002 | return IS_DOT_GIT; | |
1003 | if (*rel != '/') | |
1004 | return IS_WORKDIR_PATH; /* e.g. .gitignore */ | |
1005 | rel++; | |
1006 | ||
1007 | if (!fspathncmp(rel, FSMONITOR_COOKIE_PREFIX, | |
1008 | strlen(FSMONITOR_COOKIE_PREFIX))) | |
1009 | return IS_INSIDE_DOT_GIT_WITH_COOKIE_PREFIX; | |
1010 | ||
1011 | return IS_INSIDE_DOT_GIT; | |
1012 | } | |
1013 | ||
1014 | enum fsmonitor_path_type fsmonitor_classify_path_gitdir_relative( | |
1015 | const char *rel) | |
1016 | { | |
1017 | if (!fspathncmp(rel, FSMONITOR_COOKIE_PREFIX, | |
1018 | strlen(FSMONITOR_COOKIE_PREFIX))) | |
1019 | return IS_INSIDE_GITDIR_WITH_COOKIE_PREFIX; | |
1020 | ||
1021 | return IS_INSIDE_GITDIR; | |
1022 | } | |
1023 | ||
1024 | static enum fsmonitor_path_type try_classify_workdir_abs_path( | |
1025 | struct fsmonitor_daemon_state *state, | |
1026 | const char *path) | |
1027 | { | |
1028 | const char *rel; | |
1029 | ||
1030 | if (fspathncmp(path, state->path_worktree_watch.buf, | |
1031 | state->path_worktree_watch.len)) | |
1032 | return IS_OUTSIDE_CONE; | |
1033 | ||
1034 | rel = path + state->path_worktree_watch.len; | |
1035 | ||
1036 | if (!*rel) | |
1037 | return IS_WORKDIR_PATH; /* it is the root dir exactly */ | |
1038 | if (*rel != '/') | |
1039 | return IS_OUTSIDE_CONE; | |
1040 | rel++; | |
1041 | ||
1042 | return fsmonitor_classify_path_workdir_relative(rel); | |
1043 | } | |
1044 | ||
1045 | enum fsmonitor_path_type fsmonitor_classify_path_absolute( | |
1046 | struct fsmonitor_daemon_state *state, | |
1047 | const char *path) | |
1048 | { | |
1049 | const char *rel; | |
1050 | enum fsmonitor_path_type t; | |
1051 | ||
1052 | t = try_classify_workdir_abs_path(state, path); | |
1053 | if (state->nr_paths_watching == 1) | |
1054 | return t; | |
1055 | if (t != IS_OUTSIDE_CONE) | |
1056 | return t; | |
1057 | ||
1058 | if (fspathncmp(path, state->path_gitdir_watch.buf, | |
1059 | state->path_gitdir_watch.len)) | |
1060 | return IS_OUTSIDE_CONE; | |
1061 | ||
1062 | rel = path + state->path_gitdir_watch.len; | |
1063 | ||
1064 | if (!*rel) | |
1065 | return IS_GITDIR; /* it is the <gitdir> exactly */ | |
1066 | if (*rel != '/') | |
1067 | return IS_OUTSIDE_CONE; | |
1068 | rel++; | |
1069 | ||
1070 | return fsmonitor_classify_path_gitdir_relative(rel); | |
1071 | } | |
1072 | ||
bec486b9 JH |
1073 | /* |
1074 | * We try to combine small batches at the front of the batch-list to avoid | |
1075 | * having a long list. This hopefully makes it a little easier when we want | |
1076 | * to truncate and maintain the list. However, we don't want the paths array | |
1077 | * to just keep growing and growing with realloc, so we insert an arbitrary | |
1078 | * limit. | |
1079 | */ | |
1080 | #define MY_COMBINE_LIMIT (1024) | |
1081 | ||
1082 | void fsmonitor_publish(struct fsmonitor_daemon_state *state, | |
1083 | struct fsmonitor_batch *batch, | |
1084 | const struct string_list *cookie_names) | |
1085 | { | |
1086 | if (!batch && !cookie_names->nr) | |
1087 | return; | |
1088 | ||
1089 | pthread_mutex_lock(&state->main_lock); | |
1090 | ||
1091 | if (batch) { | |
1092 | struct fsmonitor_batch *head; | |
1093 | ||
1094 | head = state->current_token_data->batch_head; | |
1095 | if (!head) { | |
1096 | BUG("token does not have batch"); | |
1097 | } else if (head->pinned_time) { | |
1098 | /* | |
1099 | * We cannot alter the current batch list | |
1100 | * because: | |
1101 | * | |
1102 | * [a] it is being transmitted to at least one | |
1103 | * client and the handle_client() thread has a | |
1104 | * ref-count, but not a lock on the batch list | |
1105 | * starting with this item. | |
1106 | * | |
1107 | * [b] it has been transmitted in the past to | |
1108 | * at least one client such that future | |
1109 | * requests are relative to this head batch. | |
1110 | * | |
1111 | * So, we can only prepend a new batch onto | |
1112 | * the front of the list. | |
1113 | */ | |
1114 | batch->batch_seq_nr = head->batch_seq_nr + 1; | |
1115 | batch->next = head; | |
1116 | state->current_token_data->batch_head = batch; | |
1117 | } else if (!head->batch_seq_nr) { | |
1118 | /* | |
1119 | * Batch 0 is unpinned. See the note in | |
1120 | * `fsmonitor_new_token_data()` about why we | |
1121 | * don't need to accumulate these paths. | |
1122 | */ | |
1123 | fsmonitor_batch__free_list(batch); | |
1124 | } else if (head->nr + batch->nr > MY_COMBINE_LIMIT) { | |
1125 | /* | |
1126 | * The head batch in the list has never been | |
1127 | * transmitted to a client, but folding the | |
1128 | * contents of the new batch onto it would | |
1129 | * exceed our arbitrary limit, so just prepend | |
1130 | * the new batch onto the list. | |
1131 | */ | |
1132 | batch->batch_seq_nr = head->batch_seq_nr + 1; | |
1133 | batch->next = head; | |
1134 | state->current_token_data->batch_head = batch; | |
1135 | } else { | |
1136 | /* | |
1137 | * We are free to add the paths in the given | |
1138 | * batch onto the end of the current head batch. | |
1139 | */ | |
1140 | fsmonitor_batch__combine(head, batch); | |
1141 | fsmonitor_batch__free_list(batch); | |
1142 | } | |
1143 | } | |
1144 | ||
b05880d3 JH |
1145 | if (cookie_names->nr) |
1146 | with_lock__mark_cookies_seen(state, cookie_names); | |
1147 | ||
bec486b9 JH |
1148 | pthread_mutex_unlock(&state->main_lock); |
1149 | } | |
1150 | ||
d0605550 JH |
1151 | static void *fsm_health__thread_proc(void *_state) |
1152 | { | |
1153 | struct fsmonitor_daemon_state *state = _state; | |
1154 | ||
1155 | trace2_thread_start("fsm-health"); | |
1156 | ||
1157 | fsm_health__loop(state); | |
1158 | ||
1159 | trace2_thread_exit(); | |
1160 | return NULL; | |
1161 | } | |
1162 | ||
9dcba0ba JH |
1163 | static void *fsm_listen__thread_proc(void *_state) |
1164 | { | |
1165 | struct fsmonitor_daemon_state *state = _state; | |
1166 | ||
1167 | trace2_thread_start("fsm-listen"); | |
1168 | ||
1169 | trace_printf_key(&trace_fsmonitor, "Watching: worktree '%s'", | |
1170 | state->path_worktree_watch.buf); | |
1171 | if (state->nr_paths_watching > 1) | |
1172 | trace_printf_key(&trace_fsmonitor, "Watching: gitdir '%s'", | |
1173 | state->path_gitdir_watch.buf); | |
1174 | ||
1175 | fsm_listen__loop(state); | |
1176 | ||
bec486b9 JH |
1177 | pthread_mutex_lock(&state->main_lock); |
1178 | if (state->current_token_data && | |
1179 | state->current_token_data->client_ref_count == 0) | |
1180 | fsmonitor_free_token_data(state->current_token_data); | |
1181 | state->current_token_data = NULL; | |
1182 | pthread_mutex_unlock(&state->main_lock); | |
1183 | ||
9dcba0ba JH |
1184 | trace2_thread_exit(); |
1185 | return NULL; | |
1186 | } | |
1187 | ||
1188 | static int fsmonitor_run_daemon_1(struct fsmonitor_daemon_state *state) | |
1189 | { | |
1190 | struct ipc_server_opts ipc_opts = { | |
1191 | .nr_threads = fsmonitor__ipc_threads, | |
1192 | ||
1193 | /* | |
1194 | * We know that there are no other active threads yet, | |
1195 | * so we can let the IPC layer temporarily chdir() if | |
1196 | * it needs to when creating the server side of the | |
1197 | * Unix domain socket. | |
1198 | */ | |
1199 | .uds_disallow_chdir = 0 | |
1200 | }; | |
d0605550 | 1201 | int health_started = 0; |
802aa318 JH |
1202 | int listener_started = 0; |
1203 | int err = 0; | |
9dcba0ba JH |
1204 | |
1205 | /* | |
1206 | * Start the IPC thread pool before the we've started the file | |
1207 | * system event listener thread so that we have the IPC handle | |
1208 | * before we need it. | |
1209 | */ | |
1210 | if (ipc_server_run_async(&state->ipc_server_data, | |
39664e93 | 1211 | state->path_ipc.buf, &ipc_opts, |
9dcba0ba JH |
1212 | handle_client, state)) |
1213 | return error_errno( | |
1214 | _("could not start IPC thread pool on '%s'"), | |
39664e93 | 1215 | state->path_ipc.buf); |
9dcba0ba JH |
1216 | |
1217 | /* | |
1218 | * Start the fsmonitor listener thread to collect filesystem | |
1219 | * events. | |
1220 | */ | |
1221 | if (pthread_create(&state->listener_thread, NULL, | |
786e6761 | 1222 | fsm_listen__thread_proc, state)) { |
9dcba0ba | 1223 | ipc_server_stop_async(state->ipc_server_data); |
802aa318 JH |
1224 | err = error(_("could not start fsmonitor listener thread")); |
1225 | goto cleanup; | |
9dcba0ba | 1226 | } |
802aa318 | 1227 | listener_started = 1; |
9dcba0ba | 1228 | |
d0605550 JH |
1229 | /* |
1230 | * Start the health thread to watch over our process. | |
1231 | */ | |
1232 | if (pthread_create(&state->health_thread, NULL, | |
786e6761 | 1233 | fsm_health__thread_proc, state)) { |
d0605550 JH |
1234 | ipc_server_stop_async(state->ipc_server_data); |
1235 | err = error(_("could not start fsmonitor health thread")); | |
1236 | goto cleanup; | |
1237 | } | |
1238 | health_started = 1; | |
1239 | ||
9dcba0ba JH |
1240 | /* |
1241 | * The daemon is now fully functional in background threads. | |
802aa318 JH |
1242 | * Our primary thread should now just wait while the threads |
1243 | * do all the work. | |
1244 | */ | |
1245 | cleanup: | |
1246 | /* | |
9dcba0ba | 1247 | * Wait for the IPC thread pool to shutdown (whether by client |
802aa318 | 1248 | * request, from filesystem activity, or an error). |
9dcba0ba JH |
1249 | */ |
1250 | ipc_server_await(state->ipc_server_data); | |
1251 | ||
1252 | /* | |
1253 | * The fsmonitor listener thread may have received a shutdown | |
1254 | * event from the IPC thread pool, but it doesn't hurt to tell | |
1255 | * it again. And wait for it to shutdown. | |
1256 | */ | |
802aa318 JH |
1257 | if (listener_started) { |
1258 | fsm_listen__stop_async(state); | |
1259 | pthread_join(state->listener_thread, NULL); | |
1260 | } | |
9dcba0ba | 1261 | |
d0605550 JH |
1262 | if (health_started) { |
1263 | fsm_health__stop_async(state); | |
1264 | pthread_join(state->health_thread, NULL); | |
1265 | } | |
1266 | ||
802aa318 JH |
1267 | if (err) |
1268 | return err; | |
207534e4 JH |
1269 | if (state->listen_error_code) |
1270 | return state->listen_error_code; | |
d0605550 JH |
1271 | if (state->health_error_code) |
1272 | return state->health_error_code; | |
802aa318 | 1273 | return 0; |
9dcba0ba JH |
1274 | } |
1275 | ||
1276 | static int fsmonitor_run_daemon(void) | |
1277 | { | |
1278 | struct fsmonitor_daemon_state state; | |
39664e93 | 1279 | const char *home; |
9dcba0ba JH |
1280 | int err; |
1281 | ||
1282 | memset(&state, 0, sizeof(state)); | |
1283 | ||
b05880d3 | 1284 | hashmap_init(&state.cookies, cookies_cmp, NULL, 0); |
9dcba0ba | 1285 | pthread_mutex_init(&state.main_lock, NULL); |
b05880d3 | 1286 | pthread_cond_init(&state.cookies_cond, NULL); |
207534e4 | 1287 | state.listen_error_code = 0; |
d0605550 | 1288 | state.health_error_code = 0; |
aeef767a | 1289 | state.current_token_data = fsmonitor_new_token_data(); |
9dcba0ba JH |
1290 | |
1291 | /* Prepare to (recursively) watch the <worktree-root> directory. */ | |
1292 | strbuf_init(&state.path_worktree_watch, 0); | |
1293 | strbuf_addstr(&state.path_worktree_watch, absolute_path(get_git_work_tree())); | |
1294 | state.nr_paths_watching = 1; | |
1295 | ||
12fd27df ED |
1296 | strbuf_init(&state.alias.alias, 0); |
1297 | strbuf_init(&state.alias.points_to, 0); | |
1298 | if ((err = fsmonitor__get_alias(state.path_worktree_watch.buf, &state.alias))) | |
1299 | goto done; | |
1300 | ||
9dcba0ba JH |
1301 | /* |
1302 | * We create and delete cookie files somewhere inside the .git | |
1303 | * directory to help us keep sync with the file system. If | |
1304 | * ".git" is not a directory, then <gitdir> is not inside the | |
1305 | * cone of <worktree-root>, so set up a second watch to watch | |
1306 | * the <gitdir> so that we get events for the cookie files. | |
1307 | */ | |
1308 | strbuf_init(&state.path_gitdir_watch, 0); | |
1309 | strbuf_addbuf(&state.path_gitdir_watch, &state.path_worktree_watch); | |
1310 | strbuf_addstr(&state.path_gitdir_watch, "/.git"); | |
1311 | if (!is_directory(state.path_gitdir_watch.buf)) { | |
1312 | strbuf_reset(&state.path_gitdir_watch); | |
1313 | strbuf_addstr(&state.path_gitdir_watch, absolute_path(get_git_dir())); | |
1314 | state.nr_paths_watching = 2; | |
1315 | } | |
1316 | ||
b05880d3 JH |
1317 | /* |
1318 | * We will write filesystem syncing cookie files into | |
1319 | * <gitdir>/<fsmonitor-dir>/<cookie-dir>/<pid>-<seq>. | |
1320 | * | |
1321 | * The extra layers of subdirectories here keep us from | |
1322 | * changing the mtime on ".git/" or ".git/foo/" when we create | |
1323 | * or delete cookie files. | |
1324 | * | |
1325 | * There have been problems with some IDEs that do a | |
1326 | * non-recursive watch of the ".git/" directory and run a | |
1327 | * series of commands any time something happens. | |
1328 | * | |
1329 | * For example, if we place our cookie files directly in | |
1330 | * ".git/" or ".git/foo/" then a `git status` (or similar | |
1331 | * command) from the IDE will cause a cookie file to be | |
1332 | * created in one of those dirs. This causes the mtime of | |
1333 | * those dirs to change. This triggers the IDE's watch | |
1334 | * notification. This triggers the IDE to run those commands | |
1335 | * again. And the process repeats and the machine never goes | |
1336 | * idle. | |
1337 | * | |
1338 | * Adding the extra layers of subdirectories prevents the | |
1339 | * mtime of ".git/" and ".git/foo" from changing when a | |
1340 | * cookie file is created. | |
1341 | */ | |
1342 | strbuf_init(&state.path_cookie_prefix, 0); | |
1343 | strbuf_addbuf(&state.path_cookie_prefix, &state.path_gitdir_watch); | |
1344 | ||
1345 | strbuf_addch(&state.path_cookie_prefix, '/'); | |
1346 | strbuf_addstr(&state.path_cookie_prefix, FSMONITOR_DIR); | |
1347 | mkdir(state.path_cookie_prefix.buf, 0777); | |
1348 | ||
1349 | strbuf_addch(&state.path_cookie_prefix, '/'); | |
1350 | strbuf_addstr(&state.path_cookie_prefix, FSMONITOR_COOKIE_DIR); | |
1351 | mkdir(state.path_cookie_prefix.buf, 0777); | |
1352 | ||
1353 | strbuf_addch(&state.path_cookie_prefix, '/'); | |
1354 | ||
39664e93 JH |
1355 | /* |
1356 | * We create a named-pipe or unix domain socket inside of the | |
1357 | * ".git" directory. (Well, on Windows, we base our named | |
1358 | * pipe in the NPFS on the absolute path of the git | |
1359 | * directory.) | |
1360 | */ | |
1361 | strbuf_init(&state.path_ipc, 0); | |
6beb2688 ED |
1362 | strbuf_addstr(&state.path_ipc, |
1363 | absolute_path(fsmonitor_ipc__get_path(the_repository))); | |
39664e93 | 1364 | |
9dcba0ba JH |
1365 | /* |
1366 | * Confirm that we can create platform-specific resources for the | |
1367 | * filesystem listener before we bother starting all the threads. | |
1368 | */ | |
1369 | if (fsm_listen__ctor(&state)) { | |
1370 | err = error(_("could not initialize listener thread")); | |
1371 | goto done; | |
1372 | } | |
1373 | ||
d0605550 JH |
1374 | if (fsm_health__ctor(&state)) { |
1375 | err = error(_("could not initialize health thread")); | |
1376 | goto done; | |
1377 | } | |
1378 | ||
39664e93 JH |
1379 | /* |
1380 | * CD out of the worktree root directory. | |
1381 | * | |
1382 | * The common Git startup mechanism causes our CWD to be the | |
1383 | * root of the worktree. On Windows, this causes our process | |
1384 | * to hold a locked handle on the CWD. This prevents the | |
1385 | * worktree from being moved or deleted while the daemon is | |
1386 | * running. | |
1387 | * | |
1388 | * We assume that our FS and IPC listener threads have either | |
1389 | * opened all of the handles that they need or will do | |
1390 | * everything using absolute paths. | |
1391 | */ | |
1392 | home = getenv("HOME"); | |
1393 | if (home && *home && chdir(home)) | |
1394 | die_errno(_("could not cd home '%s'"), home); | |
1395 | ||
9dcba0ba JH |
1396 | err = fsmonitor_run_daemon_1(&state); |
1397 | ||
1398 | done: | |
b05880d3 | 1399 | pthread_cond_destroy(&state.cookies_cond); |
9dcba0ba JH |
1400 | pthread_mutex_destroy(&state.main_lock); |
1401 | fsm_listen__dtor(&state); | |
d0605550 | 1402 | fsm_health__dtor(&state); |
9dcba0ba JH |
1403 | |
1404 | ipc_server_free(state.ipc_server_data); | |
1405 | ||
1406 | strbuf_release(&state.path_worktree_watch); | |
1407 | strbuf_release(&state.path_gitdir_watch); | |
b05880d3 | 1408 | strbuf_release(&state.path_cookie_prefix); |
39664e93 | 1409 | strbuf_release(&state.path_ipc); |
12fd27df ED |
1410 | strbuf_release(&state.alias.alias); |
1411 | strbuf_release(&state.alias.points_to); | |
9dcba0ba JH |
1412 | |
1413 | return err; | |
1414 | } | |
1415 | ||
c284e27b | 1416 | static int try_to_run_foreground_daemon(int detach_console) |
9dcba0ba JH |
1417 | { |
1418 | /* | |
1419 | * Technically, we don't need to probe for an existing daemon | |
1420 | * process, since we could just call `fsmonitor_run_daemon()` | |
1421 | * and let it fail if the pipe/socket is busy. | |
1422 | * | |
1423 | * However, this method gives us a nicer error message for a | |
1424 | * common error case. | |
1425 | */ | |
1426 | if (fsmonitor_ipc__get_state() == IPC_STATE__LISTENING) | |
1427 | die(_("fsmonitor--daemon is already running '%s'"), | |
1428 | the_repository->worktree); | |
1429 | ||
1430 | if (fsmonitor__announce_startup) { | |
1431 | fprintf(stderr, _("running fsmonitor-daemon in '%s'\n"), | |
1432 | the_repository->worktree); | |
1433 | fflush(stderr); | |
1434 | } | |
1435 | ||
c284e27b JH |
1436 | #ifdef GIT_WINDOWS_NATIVE |
1437 | if (detach_console) | |
1438 | FreeConsole(); | |
1439 | #endif | |
1440 | ||
9dcba0ba JH |
1441 | return !!fsmonitor_run_daemon(); |
1442 | } | |
1443 | ||
c284e27b JH |
1444 | static start_bg_wait_cb bg_wait_cb; |
1445 | ||
1446 | static int bg_wait_cb(const struct child_process *cp, void *cb_data) | |
1447 | { | |
1448 | enum ipc_active_state s = fsmonitor_ipc__get_state(); | |
1449 | ||
1450 | switch (s) { | |
1451 | case IPC_STATE__LISTENING: | |
1452 | /* child is "ready" */ | |
1453 | return 0; | |
1454 | ||
1455 | case IPC_STATE__NOT_LISTENING: | |
1456 | case IPC_STATE__PATH_NOT_FOUND: | |
1457 | /* give child more time */ | |
1458 | return 1; | |
1459 | ||
1460 | default: | |
1461 | case IPC_STATE__INVALID_PATH: | |
1462 | case IPC_STATE__OTHER_ERROR: | |
1463 | /* all the time in world won't help */ | |
1464 | return -1; | |
1465 | } | |
1466 | } | |
1467 | ||
1468 | static int try_to_start_background_daemon(void) | |
1469 | { | |
1470 | struct child_process cp = CHILD_PROCESS_INIT; | |
1471 | enum start_bg_result sbgr; | |
1472 | ||
1473 | /* | |
1474 | * Before we try to create a background daemon process, see | |
1475 | * if a daemon process is already listening. This makes it | |
1476 | * easier for us to report an already-listening error to the | |
1477 | * console, since our spawn/daemon can only report the success | |
1478 | * of creating the background process (and not whether it | |
1479 | * immediately exited). | |
1480 | */ | |
1481 | if (fsmonitor_ipc__get_state() == IPC_STATE__LISTENING) | |
1482 | die(_("fsmonitor--daemon is already running '%s'"), | |
1483 | the_repository->worktree); | |
1484 | ||
1485 | if (fsmonitor__announce_startup) { | |
1486 | fprintf(stderr, _("starting fsmonitor-daemon in '%s'\n"), | |
1487 | the_repository->worktree); | |
1488 | fflush(stderr); | |
1489 | } | |
1490 | ||
1491 | cp.git_cmd = 1; | |
1492 | ||
1493 | strvec_push(&cp.args, "fsmonitor--daemon"); | |
1494 | strvec_push(&cp.args, "run"); | |
1495 | strvec_push(&cp.args, "--detach"); | |
1496 | strvec_pushf(&cp.args, "--ipc-threads=%d", fsmonitor__ipc_threads); | |
1497 | ||
1498 | cp.no_stdin = 1; | |
1499 | cp.no_stdout = 1; | |
1500 | cp.no_stderr = 1; | |
1501 | ||
1502 | sbgr = start_bg_command(&cp, bg_wait_cb, NULL, | |
1503 | fsmonitor__start_timeout_sec); | |
1504 | ||
1505 | switch (sbgr) { | |
1506 | case SBGR_READY: | |
1507 | return 0; | |
1508 | ||
1509 | default: | |
1510 | case SBGR_ERROR: | |
1511 | case SBGR_CB_ERROR: | |
1512 | return error(_("daemon failed to start")); | |
1513 | ||
1514 | case SBGR_TIMEOUT: | |
1515 | return error(_("daemon not online yet")); | |
1516 | ||
1517 | case SBGR_DIED: | |
1518 | return error(_("daemon terminated")); | |
1519 | } | |
1520 | } | |
1521 | ||
16d9d617 JH |
1522 | int cmd_fsmonitor__daemon(int argc, const char **argv, const char *prefix) |
1523 | { | |
1524 | const char *subcmd; | |
62a62a28 | 1525 | enum fsmonitor_reason reason; |
c284e27b | 1526 | int detach_console = 0; |
16d9d617 JH |
1527 | |
1528 | struct option options[] = { | |
c284e27b | 1529 | OPT_BOOL(0, "detach", &detach_console, N_("detach from console")), |
9dcba0ba JH |
1530 | OPT_INTEGER(0, "ipc-threads", |
1531 | &fsmonitor__ipc_threads, | |
1532 | N_("use <n> ipc worker threads")), | |
c284e27b JH |
1533 | OPT_INTEGER(0, "start-timeout", |
1534 | &fsmonitor__start_timeout_sec, | |
1535 | N_("max seconds to wait for background daemon startup")), | |
1536 | ||
16d9d617 JH |
1537 | OPT_END() |
1538 | }; | |
1539 | ||
9dcba0ba | 1540 | git_config(fsmonitor_config, NULL); |
16d9d617 JH |
1541 | |
1542 | argc = parse_options(argc, argv, prefix, options, | |
1543 | builtin_fsmonitor__daemon_usage, 0); | |
1544 | if (argc != 1) | |
1545 | usage_with_options(builtin_fsmonitor__daemon_usage, options); | |
1546 | subcmd = argv[0]; | |
1547 | ||
9dcba0ba JH |
1548 | if (fsmonitor__ipc_threads < 1) |
1549 | die(_("invalid 'ipc-threads' value (%d)"), | |
1550 | fsmonitor__ipc_threads); | |
1551 | ||
62a62a28 JH |
1552 | prepare_repo_settings(the_repository); |
1553 | /* | |
1554 | * If the repo is fsmonitor-compatible, explicitly set IPC-mode | |
1555 | * (without bothering to load the `core.fsmonitor` config settings). | |
1556 | * | |
1557 | * If the repo is not compatible, the repo-settings will be set to | |
1558 | * incompatible rather than IPC, so we can use one of the __get | |
1559 | * routines to detect the discrepancy. | |
1560 | */ | |
1561 | fsm_settings__set_ipc(the_repository); | |
1562 | ||
1563 | reason = fsm_settings__get_reason(the_repository); | |
1564 | if (reason > FSMONITOR_REASON_OK) | |
1565 | die("%s", | |
1566 | fsm_settings__get_incompatible_msg(the_repository, | |
1567 | reason)); | |
1568 | ||
c284e27b JH |
1569 | if (!strcmp(subcmd, "start")) |
1570 | return !!try_to_start_background_daemon(); | |
1571 | ||
9dcba0ba | 1572 | if (!strcmp(subcmd, "run")) |
c284e27b | 1573 | return !!try_to_run_foreground_daemon(detach_console); |
9dcba0ba | 1574 | |
abc9dbc0 JH |
1575 | if (!strcmp(subcmd, "stop")) |
1576 | return !!do_as_client__send_stop(); | |
1577 | ||
1578 | if (!strcmp(subcmd, "status")) | |
1579 | return !!do_as_client__status(); | |
1580 | ||
16d9d617 JH |
1581 | die(_("Unhandled subcommand '%s'"), subcmd); |
1582 | } | |
1583 | ||
1584 | #else | |
5247b762 | 1585 | int cmd_fsmonitor__daemon(int argc, const char **argv, const char *prefix UNUSED) |
16d9d617 JH |
1586 | { |
1587 | struct option options[] = { | |
1588 | OPT_END() | |
1589 | }; | |
1590 | ||
1591 | if (argc == 2 && !strcmp(argv[1], "-h")) | |
1592 | usage_with_options(builtin_fsmonitor__daemon_usage, options); | |
1593 | ||
1594 | die(_("fsmonitor--daemon not supported on this platform")); | |
1595 | } | |
1596 | #endif |