1 | /**************************************************************************** |
2 | ** |
3 | ** Copyright (C) 2020 Intel Corporation. |
4 | ** Copyright (C) 2015 Klarälvdalens Datakonsult AB, a KDAB Group company, info@kdab.com |
5 | ** |
6 | ** Permission is hereby granted, free of charge, to any person obtaining a copy |
7 | ** of this software and associated documentation files (the "Software"), to deal |
8 | ** in the Software without restriction, including without limitation the rights |
9 | ** to use, copy, modify, merge, publish, distribute, sublicense, and/or sell |
10 | ** copies of the Software, and to permit persons to whom the Software is |
11 | ** furnished to do so, subject to the following conditions: |
12 | ** |
13 | ** The above copyright notice and this permission notice shall be included in |
14 | ** all copies or substantial portions of the Software. |
15 | ** |
16 | ** THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
17 | ** IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
18 | ** FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
19 | ** AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
20 | ** LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
21 | ** OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN |
22 | ** THE SOFTWARE. |
23 | ** |
24 | ****************************************************************************/ |
25 | |
26 | #ifndef _GNU_SOURCE |
27 | # define _GNU_SOURCE |
28 | #endif |
29 | |
30 | #include "forkfd.h" |
31 | |
32 | /* Macros fine-tuning the build: */ |
33 | //#define FORKFD_NO_FORKFD 1 /* disable the forkfd() function */ |
34 | //#define FORKFD_NO_SPAWNFD 1 /* disable the spawnfd() function */ |
35 | //#define FORKFD_DISABLE_FORK_FALLBACK 1 /* disable falling back to fork() from system_forkfd() */ |
36 | |
37 | #include <sys/types.h> |
38 | #if defined(__OpenBSD__) || defined(__NetBSD__) |
39 | # include <sys/param.h> |
40 | #endif |
41 | #include <sys/time.h> |
42 | #include <sys/resource.h> |
43 | #include <sys/wait.h> |
44 | #include <assert.h> |
45 | #include <errno.h> |
46 | #include <pthread.h> |
47 | #include <signal.h> |
48 | #include <stdlib.h> |
49 | #include <string.h> |
50 | #include <time.h> |
51 | #include <unistd.h> |
52 | |
53 | #ifdef __linux__ |
54 | # define HAVE_WAIT4 1 |
55 | # if defined(__BIONIC__) || (defined(__GLIBC__) && (__GLIBC__ << 8) + __GLIBC_MINOR__ >= 0x208 && \ |
56 | (!defined(__UCLIBC__) || ((__UCLIBC_MAJOR__ << 16) + (__UCLIBC_MINOR__ << 8) + __UCLIBC_SUBLEVEL__ > 0x90201))) |
57 | # include <sys/eventfd.h> |
58 | # ifdef EFD_CLOEXEC |
59 | # define HAVE_EVENTFD 1 |
60 | # endif |
61 | # endif |
62 | # if defined(__BIONIC__) || (defined(__GLIBC__) && (__GLIBC__ << 8) + __GLIBC_MINOR__ >= 0x209 && \ |
63 | (!defined(__UCLIBC__) || ((__UCLIBC_MAJOR__ << 16) + (__UCLIBC_MINOR__ << 8) + __UCLIBC_SUBLEVEL__ > 0x90201))) |
64 | # define HAVE_PIPE2 1 |
65 | # endif |
66 | #endif |
67 | |
68 | #if _POSIX_VERSION-0 >= 200809L || _XOPEN_VERSION-0 >= 500 |
69 | # define HAVE_WAITID 1 |
70 | #endif |
71 | #if !defined(WEXITED) || !defined(WNOWAIT) |
72 | # undef HAVE_WAITID |
73 | #endif |
74 | |
75 | #if (defined(__FreeBSD__) && defined(__FreeBSD_version) && __FreeBSD_version >= 1000032) || \ |
76 | (defined(__OpenBSD__) && OpenBSD >= 201505) || \ |
77 | (defined(__NetBSD__) && __NetBSD_Version__ >= 600000000) |
78 | # define HAVE_PIPE2 1 |
79 | #endif |
80 | #if defined(__FreeBSD__) || defined(__DragonFly__) || defined(__FreeBSD_kernel__) || \ |
81 | defined(__OpenBSD__) || defined(__NetBSD__) || defined(__APPLE__) |
82 | # define HAVE_WAIT4 1 |
83 | #endif |
84 | |
85 | #if defined(__APPLE__) |
86 | /* Up until OS X 10.7, waitid(P_ALL, ...) will return success, but will not |
87 | * fill in the details of the dead child. That means waitid is not useful to us. |
88 | * Therefore, we only enable waitid() support if we're targetting OS X 10.8 or |
89 | * later. |
90 | */ |
91 | # include <Availability.h> |
92 | # include <AvailabilityMacros.h> |
93 | # if MAC_OS_X_VERSION_MIN_REQUIRED <= 1070 |
94 | # define HAVE_BROKEN_WAITID 1 |
95 | # endif |
96 | #endif |
97 | |
98 | #include "forkfd_atomic.h" |
99 | |
100 | static int system_has_forkfd(void); |
101 | static int system_forkfd(int flags, pid_t *ppid, int *system); |
102 | static int system_vforkfd(int flags, pid_t *ppid, int (*)(void *), void *, int *system); |
103 | static int system_forkfd_wait(int ffd, struct forkfd_info *info, int ffdwoptions, struct rusage *rusage); |
104 | |
105 | static int disable_fork_fallback(void) |
106 | { |
107 | #ifdef FORKFD_DISABLE_FORK_FALLBACK |
108 | /* if there's no system forkfd, we have to use the fallback */ |
109 | return system_has_forkfd(); |
110 | #else |
111 | return 0; |
112 | #endif |
113 | } |
114 | |
115 | #define CHILDREN_IN_SMALL_ARRAY 16 |
116 | #define CHILDREN_IN_BIG_ARRAY 256 |
117 | #define sizeofarray(array) (sizeof(array)/sizeof(array[0])) |
118 | #define EINTR_LOOP(ret, call) \ |
119 | do { \ |
120 | ret = call; \ |
121 | } while (ret == -1 && errno == EINTR) |
122 | |
123 | struct pipe_payload |
124 | { |
125 | struct forkfd_info info; |
126 | struct rusage rusage; |
127 | }; |
128 | |
129 | typedef struct process_info |
130 | { |
131 | ffd_atomic_int pid; |
132 | int deathPipe; |
133 | } ProcessInfo; |
134 | |
135 | struct BigArray; |
136 | typedef struct |
137 | { |
138 | ffd_atomic_pointer(struct BigArray) ; |
139 | ffd_atomic_int ; |
140 | } ; |
141 | |
142 | typedef struct BigArray |
143 | { |
144 | Header ; |
145 | ProcessInfo entries[CHILDREN_IN_BIG_ARRAY]; |
146 | } BigArray; |
147 | |
148 | typedef struct SmallArray |
149 | { |
150 | Header ; |
151 | ProcessInfo entries[CHILDREN_IN_SMALL_ARRAY]; |
152 | } SmallArray; |
153 | static SmallArray children; |
154 | |
155 | static struct sigaction old_sigaction; |
156 | static pthread_once_t forkfd_initialization = PTHREAD_ONCE_INIT; |
157 | static ffd_atomic_int forkfd_status = FFD_ATOMIC_INIT(0); |
158 | |
159 | #ifdef HAVE_BROKEN_WAITID |
160 | static int waitid_works = 0; |
161 | #else |
162 | static const int waitid_works = 1; |
163 | #endif |
164 | |
165 | static ProcessInfo *(Header *, ProcessInfo entries[], int maxCount) |
166 | { |
167 | /* we use ACQUIRE here because the signal handler might have released the PID */ |
168 | int busyCount = ffd_atomic_add_fetch(&header->busyCount, 1, FFD_ATOMIC_ACQUIRE); |
169 | if (busyCount <= maxCount) { |
170 | /* there's an available entry in this section, find it and take it */ |
171 | int i; |
172 | for (i = 0; i < maxCount; ++i) { |
173 | /* if the PID is 0, it's free; mark it as used by swapping it with -1 */ |
174 | int expected_pid = 0; |
175 | if (ffd_atomic_compare_exchange(&entries[i].pid, &expected_pid, |
176 | -1, FFD_ATOMIC_RELAXED, FFD_ATOMIC_RELAXED)) |
177 | return &entries[i]; |
178 | } |
179 | } |
180 | |
181 | /* there isn't an available entry, undo our increment */ |
182 | (void)ffd_atomic_add_fetch(&header->busyCount, -1, FFD_ATOMIC_RELAXED); |
183 | return NULL; |
184 | } |
185 | |
186 | static ProcessInfo *(Header **) |
187 | { |
188 | Header * = &children.header; |
189 | |
190 | /* try to find an available entry in the small array first */ |
191 | ProcessInfo *info = |
192 | tryAllocateInSection(header: currentHeader, entries: children.entries, sizeofarray(children.entries)); |
193 | |
194 | /* go on to the next arrays */ |
195 | while (info == NULL) { |
196 | BigArray *array = ffd_atomic_load(¤tHeader->nextArray, FFD_ATOMIC_ACQUIRE); |
197 | if (array == NULL) { |
198 | /* allocate an array and try to use it */ |
199 | BigArray *allocatedArray = (BigArray *)calloc(nmemb: 1, size: sizeof(BigArray)); |
200 | if (allocatedArray == NULL) |
201 | return NULL; |
202 | |
203 | if (ffd_atomic_compare_exchange(¤tHeader->nextArray, &array, allocatedArray, |
204 | FFD_ATOMIC_RELEASE, FFD_ATOMIC_ACQUIRE)) { |
205 | /* success */ |
206 | array = allocatedArray; |
207 | } else { |
208 | /* failed, the atomic updated 'array' */ |
209 | free(ptr: allocatedArray); |
210 | } |
211 | } |
212 | |
213 | currentHeader = &array->header; |
214 | info = tryAllocateInSection(header: currentHeader, entries: array->entries, sizeofarray(array->entries)); |
215 | } |
216 | |
217 | *header = currentHeader; |
218 | return info; |
219 | } |
220 | |
221 | #ifdef HAVE_WAITID |
222 | static int isChildReady(pid_t pid, siginfo_t *info) |
223 | { |
224 | info->si_pid = 0; |
225 | return waitid(idtype: P_PID, id: pid, infop: info, WEXITED | WNOHANG | WNOWAIT) == 0 && info->si_pid == pid; |
226 | } |
227 | #endif |
228 | |
229 | static void convertStatusToForkfdInfo(int status, struct forkfd_info *info) |
230 | { |
231 | if (WIFEXITED(status)) { |
232 | info->code = CLD_EXITED; |
233 | info->status = WEXITSTATUS(status); |
234 | } else if (WIFSIGNALED(status)) { |
235 | info->code = CLD_KILLED; |
236 | # ifdef WCOREDUMP |
237 | if (WCOREDUMP(status)) |
238 | info->code = CLD_DUMPED; |
239 | # endif |
240 | info->status = WTERMSIG(status); |
241 | } |
242 | } |
243 | |
244 | #ifdef __GNUC__ |
245 | __attribute__((unused)) |
246 | #endif |
247 | static int convertForkfdWaitFlagsToWaitFlags(int ffdoptions) |
248 | { |
249 | int woptions = WEXITED; |
250 | if (ffdoptions & FFDW_NOWAIT) |
251 | woptions |= WNOWAIT; |
252 | if (ffdoptions & FFDW_NOHANG) |
253 | woptions |= WNOHANG; |
254 | return woptions; |
255 | } |
256 | |
257 | static int tryReaping(pid_t pid, struct pipe_payload *payload) |
258 | { |
259 | /* reap the child */ |
260 | #if defined(HAVE_WAIT4) |
261 | int status; |
262 | if (wait4(pid: pid, stat_loc: &status, WNOHANG, usage: &payload->rusage) <= 0) |
263 | return 0; |
264 | convertStatusToForkfdInfo(status, info: &payload->info); |
265 | #else |
266 | # if defined(HAVE_WAITID) |
267 | if (waitid_works) { |
268 | /* we have waitid(2), which gets us some payload values on some systems */ |
269 | siginfo_t info; |
270 | info.si_pid = 0; |
271 | int ret = waitid(P_PID, pid, &info, WEXITED | WNOHANG) == 0 && info.si_pid == pid; |
272 | if (!ret) |
273 | return ret; |
274 | |
275 | payload->info.code = info.si_code; |
276 | payload->info.status = info.si_status; |
277 | # ifdef __linux__ |
278 | payload->rusage.ru_utime.tv_sec = info.si_utime / CLOCKS_PER_SEC; |
279 | payload->rusage.ru_utime.tv_usec = info.si_utime % CLOCKS_PER_SEC; |
280 | payload->rusage.ru_stime.tv_sec = info.si_stime / CLOCKS_PER_SEC; |
281 | payload->rusage.ru_stime.tv_usec = info.si_stime % CLOCKS_PER_SEC; |
282 | # endif |
283 | return 1; |
284 | } |
285 | # endif // HAVE_WAITID |
286 | int status; |
287 | if (waitpid(pid, &status, WNOHANG) <= 0) |
288 | return 0; // child did not change state |
289 | convertStatusToForkfdInfo(status, &payload->info); |
290 | #endif // !HAVE_WAIT4 |
291 | |
292 | return 1; |
293 | } |
294 | |
295 | static void (Header *, ProcessInfo *entry) |
296 | { |
297 | entry->deathPipe = -1; |
298 | ffd_atomic_store(&entry->pid, 0, FFD_ATOMIC_RELEASE); |
299 | |
300 | (void)ffd_atomic_add_fetch(&header->busyCount, -1, FFD_ATOMIC_RELEASE); |
301 | assert(header->busyCount >= 0); |
302 | } |
303 | |
304 | static void notifyAndFreeInfo(Header *, ProcessInfo *entry, |
305 | const struct pipe_payload *payload) |
306 | { |
307 | ssize_t ret; |
308 | EINTR_LOOP(ret, write(entry->deathPipe, payload, sizeof(*payload))); |
309 | EINTR_LOOP(ret, close(entry->deathPipe)); |
310 | |
311 | freeInfo(header, entry); |
312 | } |
313 | |
314 | static void reapChildProcesses(); |
315 | static void sigchld_handler(int signum, siginfo_t *handler_info, void *handler_context) |
316 | { |
317 | /* |
318 | * This is a signal handler, so we need to be careful about which functions |
319 | * we can call. See the full, official listing in the POSIX.1-2008 |
320 | * specification at: |
321 | * http://pubs.opengroup.org/onlinepubs/9699919799/functions/V2_chap02.html#tag_15_04_03 |
322 | * |
323 | * The handler_info and handler_context parameters may not be valid, if |
324 | * we're a chained handler from another handler that did not use |
325 | * SA_SIGINFO. Therefore, we must obtain the siginfo ourselves directly by |
326 | * calling waitid. |
327 | * |
328 | * But we pass them anyway. Let's call the chained handler first, while |
329 | * those two arguments have a chance of being correct. |
330 | */ |
331 | if (old_sigaction.sa_handler != SIG_IGN && old_sigaction.sa_handler != SIG_DFL) { |
332 | if (old_sigaction.sa_flags & SA_SIGINFO) |
333 | old_sigaction.sa_sigaction(signum, handler_info, handler_context); |
334 | else |
335 | old_sigaction.sa_handler(signum); |
336 | } |
337 | |
338 | if (ffd_atomic_load(&forkfd_status, FFD_ATOMIC_RELAXED) == 1) { |
339 | int saved_errno = errno; |
340 | reapChildProcesses(); |
341 | errno = saved_errno; |
342 | } |
343 | } |
344 | |
345 | static inline void reapChildProcesses() |
346 | { |
347 | /* is this one of our children? */ |
348 | BigArray *array; |
349 | siginfo_t info; |
350 | struct pipe_payload payload; |
351 | int i; |
352 | |
353 | memset(s: &info, c: 0, n: sizeof info); |
354 | memset(s: &payload, c: 0, n: sizeof payload); |
355 | |
356 | #ifdef HAVE_WAITID |
357 | if (waitid_works) { |
358 | /* be optimistic: try to see if we can get the child that exited */ |
359 | search_next_child: |
360 | /* waitid returns -1 ECHILD if there are no further children at all; |
361 | * it returns 0 and sets si_pid to 0 if there are children but they are not ready |
362 | * to be waited (we're passing WNOHANG). We should not get EINTR because |
363 | * we're passing WNOHANG and we should definitely not get EINVAL or anything else. |
364 | * That means we can actually ignore the return code and only inspect si_pid. |
365 | */ |
366 | info.si_pid = 0; |
367 | waitid(idtype: P_ALL, id: 0, infop: &info, WNOHANG | WNOWAIT | WEXITED); |
368 | if (info.si_pid == 0) { |
369 | /* there are no further un-waited-for children, so we can just exit. |
370 | */ |
371 | return; |
372 | } |
373 | |
374 | for (i = 0; i < (int)sizeofarray(children.entries); ++i) { |
375 | /* acquire the child first: swap the PID with -1 to indicate it's busy */ |
376 | int pid = info.si_pid; |
377 | if (ffd_atomic_compare_exchange(&children.entries[i].pid, &pid, -1, |
378 | FFD_ATOMIC_ACQUIRE, FFD_ATOMIC_RELAXED)) { |
379 | /* this is our child, send notification and free up this entry */ |
380 | /* ### FIXME: what if tryReaping returns false? */ |
381 | if (tryReaping(pid, payload: &payload)) |
382 | notifyAndFreeInfo(header: &children.header, entry: &children.entries[i], payload: &payload); |
383 | goto search_next_child; |
384 | } |
385 | } |
386 | |
387 | /* try the arrays */ |
388 | array = ffd_atomic_load(&children.header.nextArray, FFD_ATOMIC_ACQUIRE); |
389 | while (array != NULL) { |
390 | for (i = 0; i < (int)sizeofarray(array->entries); ++i) { |
391 | int pid = info.si_pid; |
392 | if (ffd_atomic_compare_exchange(&array->entries[i].pid, &pid, -1, |
393 | FFD_ATOMIC_ACQUIRE, FFD_ATOMIC_RELAXED)) { |
394 | /* this is our child, send notification and free up this entry */ |
395 | /* ### FIXME: what if tryReaping returns false? */ |
396 | if (tryReaping(pid, payload: &payload)) |
397 | notifyAndFreeInfo(header: &array->header, entry: &array->entries[i], payload: &payload); |
398 | goto search_next_child; |
399 | } |
400 | } |
401 | |
402 | array = ffd_atomic_load(&array->header.nextArray, FFD_ATOMIC_ACQUIRE); |
403 | } |
404 | |
405 | /* if we got here, we couldn't find this child in our list. That means this child |
406 | * belongs to one of the chained SIGCHLD handlers. However, there might be another |
407 | * child that exited and does belong to us, so we need to check each one individually. |
408 | */ |
409 | } |
410 | #endif |
411 | |
412 | for (i = 0; i < (int)sizeofarray(children.entries); ++i) { |
413 | int pid = ffd_atomic_load(&children.entries[i].pid, FFD_ATOMIC_ACQUIRE); |
414 | if (pid <= 0) |
415 | continue; |
416 | #ifdef HAVE_WAITID |
417 | if (waitid_works) { |
418 | /* The child might have been reaped by the block above in another thread, |
419 | * so first check if it's ready and, if it is, lock it */ |
420 | if (!isChildReady(pid, info: &info) || |
421 | !ffd_atomic_compare_exchange(&children.entries[i].pid, &pid, -1, |
422 | FFD_ATOMIC_RELAXED, FFD_ATOMIC_RELAXED)) |
423 | continue; |
424 | } |
425 | #endif |
426 | if (tryReaping(pid, payload: &payload)) { |
427 | /* this is our child, send notification and free up this entry */ |
428 | notifyAndFreeInfo(header: &children.header, entry: &children.entries[i], payload: &payload); |
429 | } |
430 | } |
431 | |
432 | /* try the arrays */ |
433 | array = ffd_atomic_load(&children.header.nextArray, FFD_ATOMIC_ACQUIRE); |
434 | while (array != NULL) { |
435 | for (i = 0; i < (int)sizeofarray(array->entries); ++i) { |
436 | int pid = ffd_atomic_load(&array->entries[i].pid, FFD_ATOMIC_ACQUIRE); |
437 | if (pid <= 0) |
438 | continue; |
439 | #ifdef HAVE_WAITID |
440 | if (waitid_works) { |
441 | /* The child might have been reaped by the block above in another thread, |
442 | * so first check if it's ready and, if it is, lock it */ |
443 | if (!isChildReady(pid, info: &info) || |
444 | !ffd_atomic_compare_exchange(&array->entries[i].pid, &pid, -1, |
445 | FFD_ATOMIC_RELAXED, FFD_ATOMIC_RELAXED)) |
446 | continue; |
447 | } |
448 | #endif |
449 | if (tryReaping(pid, payload: &payload)) { |
450 | /* this is our child, send notification and free up this entry */ |
451 | notifyAndFreeInfo(header: &array->header, entry: &array->entries[i], payload: &payload); |
452 | } |
453 | } |
454 | |
455 | array = ffd_atomic_load(&array->header.nextArray, FFD_ATOMIC_ACQUIRE); |
456 | } |
457 | } |
458 | |
459 | static void ignore_sigpipe() |
460 | { |
461 | #ifdef O_NOSIGPIPE |
462 | static ffd_atomic_int done = FFD_ATOMIC_INIT(0); |
463 | if (ffd_atomic_load(&done, FFD_ATOMIC_RELAXED)) |
464 | return; |
465 | #endif |
466 | |
467 | struct sigaction action; |
468 | memset(s: &action, c: 0, n: sizeof action); |
469 | sigemptyset(set: &action.sa_mask); |
470 | action.sa_handler = SIG_IGN; |
471 | action.sa_flags = 0; |
472 | sigaction(SIGPIPE, act: &action, NULL); |
473 | |
474 | #ifdef O_NOSIGPIPE |
475 | ffd_atomic_store(&done, 1, FFD_ATOMIC_RELAXED); |
476 | #endif |
477 | } |
478 | |
479 | #if defined(__GNUC__) && (!defined(__FreeBSD__) || __FreeBSD__ < 10) |
480 | __attribute((destructor, unused)) static void cleanup(); |
481 | #endif |
482 | |
483 | static void cleanup() |
484 | { |
485 | BigArray *array; |
486 | /* This function is not thread-safe! |
487 | * It must only be called when the process is shutting down. |
488 | * At shutdown, we expect no one to be calling forkfd(), so we don't |
489 | * need to be thread-safe with what is done there. |
490 | * |
491 | * But SIGCHLD might be delivered to any thread, including this one. |
492 | * There's no way to prevent that. The correct solution would be to |
493 | * cooperatively delete. We don't do that. |
494 | */ |
495 | if (ffd_atomic_load(&forkfd_status, FFD_ATOMIC_RELAXED) == 0) |
496 | return; |
497 | |
498 | /* notify the handler that we're no longer in operation */ |
499 | ffd_atomic_store(&forkfd_status, 0, FFD_ATOMIC_RELAXED); |
500 | |
501 | /* free any arrays we might have */ |
502 | array = ffd_atomic_load(&children.header.nextArray, FFD_ATOMIC_ACQUIRE); |
503 | while (array != NULL) { |
504 | BigArray *next = ffd_atomic_load(&array->header.nextArray, FFD_ATOMIC_ACQUIRE); |
505 | free(ptr: array); |
506 | array = next; |
507 | } |
508 | } |
509 | |
510 | static void forkfd_initialize() |
511 | { |
512 | #if defined(HAVE_BROKEN_WAITID) |
513 | pid_t pid = fork(); |
514 | if (pid == 0) { |
515 | _exit(0); |
516 | } else if (pid > 0) { |
517 | siginfo_t info; |
518 | waitid(P_ALL, 0, &info, WNOWAIT | WEXITED); |
519 | waitid_works = (info.si_pid != 0); |
520 | info.si_pid = 0; |
521 | |
522 | // now really reap the child |
523 | waitid(P_PID, pid, &info, WEXITED); |
524 | waitid_works = waitid_works && (info.si_pid != 0); |
525 | } |
526 | #endif |
527 | |
528 | /* install our signal handler */ |
529 | struct sigaction action; |
530 | memset(s: &action, c: 0, n: sizeof action); |
531 | sigemptyset(set: &action.sa_mask); |
532 | action.sa_flags = SA_NOCLDSTOP | SA_SIGINFO; |
533 | action.sa_sigaction = sigchld_handler; |
534 | |
535 | /* ### RACE CONDITION |
536 | * The sigaction function does a memcpy from an internal buffer |
537 | * to old_sigaction, which we use in the SIGCHLD handler. If a |
538 | * SIGCHLD is delivered before or during that memcpy, the handler will |
539 | * see an inconsistent state. |
540 | * |
541 | * There is no solution. pthread_sigmask doesn't work here because the |
542 | * signal could be delivered to another thread. |
543 | */ |
544 | sigaction(SIGCHLD, act: &action, oact: &old_sigaction); |
545 | |
546 | #ifndef O_NOSIGPIPE |
547 | /* disable SIGPIPE too */ |
548 | ignore_sigpipe(); |
549 | #endif |
550 | |
551 | #ifdef __GNUC__ |
552 | (void) cleanup; /* suppress unused static function warning */ |
553 | #else |
554 | atexit(cleanup); |
555 | #endif |
556 | |
557 | ffd_atomic_store(&forkfd_status, 1, FFD_ATOMIC_RELAXED); |
558 | } |
559 | |
560 | static int create_pipe(int filedes[], int flags) |
561 | { |
562 | int ret = -1; |
563 | #ifdef HAVE_PIPE2 |
564 | /* use pipe2(2) whenever possible, since it can thread-safely create a |
565 | * cloexec pair of pipes. Without it, we have a race condition setting |
566 | * FD_CLOEXEC |
567 | */ |
568 | |
569 | # ifdef O_NOSIGPIPE |
570 | /* try first with O_NOSIGPIPE */ |
571 | ret = pipe2(filedes, O_CLOEXEC | O_NOSIGPIPE); |
572 | if (ret == -1) { |
573 | /* O_NOSIGPIPE not supported, ignore SIGPIPE */ |
574 | ignore_sigpipe(); |
575 | } |
576 | # endif |
577 | if (ret == -1) |
578 | ret = pipe2(pipedes: filedes, O_CLOEXEC); |
579 | if (ret == -1) |
580 | return ret; |
581 | |
582 | if ((flags & FFD_CLOEXEC) == 0) |
583 | fcntl(fd: filedes[0], F_SETFD, 0); |
584 | #else |
585 | ret = pipe(filedes); |
586 | if (ret == -1) |
587 | return ret; |
588 | |
589 | fcntl(filedes[1], F_SETFD, FD_CLOEXEC); |
590 | if (flags & FFD_CLOEXEC) |
591 | fcntl(filedes[0], F_SETFD, FD_CLOEXEC); |
592 | #endif |
593 | if (flags & FFD_NONBLOCK) |
594 | fcntl(fd: filedes[0], F_SETFL, fcntl(fd: filedes[0], F_GETFL) | O_NONBLOCK); |
595 | return ret; |
596 | } |
597 | |
598 | #ifndef FORKFD_NO_FORKFD |
599 | static int forkfd_fork_fallback(int flags, pid_t *ppid) |
600 | { |
601 | Header *; |
602 | ProcessInfo *info; |
603 | pid_t pid; |
604 | int fd = -1; |
605 | int death_pipe[2]; |
606 | int sync_pipe[2]; |
607 | int ret; |
608 | #ifdef __linux__ |
609 | int efd; |
610 | #endif |
611 | |
612 | (void) pthread_once(once_control: &forkfd_initialization, init_routine: forkfd_initialize); |
613 | |
614 | info = allocateInfo(header: &header); |
615 | if (info == NULL) { |
616 | errno = ENOMEM; |
617 | return -1; |
618 | } |
619 | |
620 | /* create the pipes before we fork */ |
621 | if (create_pipe(filedes: death_pipe, flags) == -1) |
622 | goto err_free; /* failed to create the pipes, pass errno */ |
623 | |
624 | #ifdef HAVE_EVENTFD |
625 | /* try using an eventfd, which consumes less resources */ |
626 | efd = eventfd(count: 0, EFD_CLOEXEC); |
627 | if (efd == -1) |
628 | #endif |
629 | { |
630 | /* try a pipe */ |
631 | if (create_pipe(filedes: sync_pipe, FFD_CLOEXEC) == -1) { |
632 | /* failed both at eventfd and pipe; fail and pass errno */ |
633 | goto err_close; |
634 | } |
635 | } |
636 | |
637 | /* now fork */ |
638 | pid = fork(); |
639 | if (pid == -1) |
640 | goto err_close2; /* failed to fork, pass errno */ |
641 | if (ppid) |
642 | *ppid = pid; |
643 | |
644 | /* |
645 | * We need to store the child's PID in the info structure, so |
646 | * the SIGCHLD handler knows that this child is present and it |
647 | * knows the writing end of the pipe to pass information on. |
648 | * However, the child process could exit before we stored the |
649 | * information (or the handler could run for other children exiting). |
650 | * We prevent that from happening by blocking the child process in |
651 | * a read(2) until we're finished storing the information. |
652 | */ |
653 | if (pid == 0) { |
654 | /* this is the child process */ |
655 | /* first, wait for the all clear */ |
656 | #ifdef HAVE_EVENTFD |
657 | if (efd != -1) { |
658 | eventfd_t val64; |
659 | EINTR_LOOP(ret, eventfd_read(efd, &val64)); |
660 | EINTR_LOOP(ret, close(efd)); |
661 | } else |
662 | #endif |
663 | { |
664 | char c; |
665 | EINTR_LOOP(ret, close(sync_pipe[1])); |
666 | EINTR_LOOP(ret, read(sync_pipe[0], &c, sizeof c)); |
667 | EINTR_LOOP(ret, close(sync_pipe[0])); |
668 | } |
669 | |
670 | /* now close the pipes and return to the caller */ |
671 | EINTR_LOOP(ret, close(death_pipe[0])); |
672 | EINTR_LOOP(ret, close(death_pipe[1])); |
673 | fd = FFD_CHILD_PROCESS; |
674 | } else { |
675 | /* parent process */ |
676 | info->deathPipe = death_pipe[1]; |
677 | fd = death_pipe[0]; |
678 | ffd_atomic_store(&info->pid, pid, FFD_ATOMIC_RELEASE); |
679 | |
680 | /* release the child */ |
681 | #ifdef HAVE_EVENTFD |
682 | if (efd != -1) { |
683 | eventfd_t val64 = 42; |
684 | EINTR_LOOP(ret, eventfd_write(efd, val64)); |
685 | EINTR_LOOP(ret, close(efd)); |
686 | } else |
687 | #endif |
688 | { |
689 | /* |
690 | * Usually, closing would be enough to make read(2) return and the child process |
691 | * continue. We need to write here: another thread could be calling forkfd at the |
692 | * same time, which means auxpipe[1] might be open in another child process. |
693 | */ |
694 | EINTR_LOOP(ret, close(sync_pipe[0])); |
695 | EINTR_LOOP(ret, write(sync_pipe[1], "" , 1)); |
696 | EINTR_LOOP(ret, close(sync_pipe[1])); |
697 | } |
698 | } |
699 | |
700 | return fd; |
701 | |
702 | err_close2: |
703 | #ifdef HAVE_EVENTFD |
704 | if (efd != -1) { |
705 | EINTR_LOOP(ret, close(efd)); |
706 | } else |
707 | #endif |
708 | { |
709 | EINTR_LOOP(ret, close(sync_pipe[0])); |
710 | EINTR_LOOP(ret, close(sync_pipe[1])); |
711 | } |
712 | err_close: |
713 | EINTR_LOOP(ret, close(death_pipe[0])); |
714 | EINTR_LOOP(ret, close(death_pipe[1])); |
715 | err_free: |
716 | /* free the info pointer */ |
717 | freeInfo(header, entry: info); |
718 | return -1; |
719 | } |
720 | |
721 | /** |
722 | * @brief forkfd returns a file descriptor representing a child process |
723 | * @return a file descriptor, or -1 in case of failure |
724 | * |
725 | * forkfd() creates a file descriptor that can be used to be notified of when a |
726 | * child process exits. This file descriptor can be monitored using select(2), |
727 | * poll(2) or similar mechanisms. |
728 | * |
729 | * The @a flags parameter can contain the following values ORed to change the |
730 | * behaviour of forkfd(): |
731 | * |
732 | * @li @c FFD_NONBLOCK Set the O_NONBLOCK file status flag on the new open file |
733 | * descriptor. Using this flag saves extra calls to fnctl(2) to achieve the same |
734 | * result. |
735 | * |
736 | * @li @c FFD_CLOEXEC Set the close-on-exec (FD_CLOEXEC) flag on the new file |
737 | * descriptor. You probably want to set this flag, since forkfd() does not work |
738 | * if the original parent process dies. |
739 | * |
740 | * @li @c FFD_USE_FORK Tell forkfd() to actually call fork() instead of a |
741 | * different system implementation that may be available. On systems where a |
742 | * different implementation is available, its behavior may differ from that of |
743 | * fork(), such as not calling the functions registered with pthread_atfork(). |
744 | * If that's necessary, pass this flag. |
745 | * |
746 | * The file descriptor returned by forkfd() supports the following operations: |
747 | * |
748 | * @li read(2) When the child process exits, then the buffer supplied to |
749 | * read(2) is used to return information about the status of the child in the |
750 | * form of one @c siginfo_t structure. The buffer must be at least |
751 | * sizeof(siginfo_t) bytes. The return value of read(2) is the total number of |
752 | * bytes read. |
753 | * |
754 | * @li poll(2), select(2) (and similar) The file descriptor is readable (the |
755 | * select(2) readfds argument; the poll(2) POLLIN flag) if the child has exited |
756 | * or signalled via SIGCHLD. |
757 | * |
758 | * @li close(2) When the file descriptor is no longer required it should be closed. |
759 | */ |
760 | int forkfd(int flags, pid_t *ppid) |
761 | { |
762 | int fd; |
763 | if (disable_fork_fallback()) |
764 | flags &= ~FFD_USE_FORK; |
765 | |
766 | if ((flags & FFD_USE_FORK) == 0) { |
767 | int system_forkfd_works; |
768 | fd = system_forkfd(flags, ppid, system: &system_forkfd_works); |
769 | if (system_forkfd_works || disable_fork_fallback()) |
770 | return fd; |
771 | } |
772 | |
773 | return forkfd_fork_fallback(flags, ppid); |
774 | } |
775 | |
776 | /** |
777 | * @brief vforkfd returns a file descriptor representing a child process |
778 | * @return a file descriptor, or -1 in case of failure |
779 | * |
780 | * vforkfd() operates in the same way as forkfd() and the @a flags and @a ppid |
781 | * arguments are the same. See the forkfd() documentation for details on the |
782 | * possible values and information on the returned file descriptor. |
783 | * |
784 | * This function does not return @c FFD_CHILD_PROCESS. Instead, the function @a |
785 | * childFn is called in the child process with the @a token parameter as |
786 | * argument. If that function returns, its return value will be passed to |
787 | * _exit(2). |
788 | * |
789 | * This function differs from forkfd() the same way that vfork() differs from |
790 | * fork(): the parent process may be suspended while the child is has not yet |
791 | * called _exit(2) or execve(2). Additionally, on some systems, the child |
792 | * process may share memory with the parent process the same way an auxiliary |
793 | * thread would, so extreme care should be employed on what functions the child |
794 | * process uses before termination. |
795 | * |
796 | * The @c FFD_USE_FORK flag retains its behavior as described in the forkfd() |
797 | * documentation, including that of actually using fork(2) and no other |
798 | * implementation. |
799 | * |
800 | * Currently, only on Linux will this function have any behavior different from |
801 | * forkfd(). In all other systems, it is equivalent to the following code: |
802 | * |
803 | * @code |
804 | * int ffd = forkfd(flags, &pid); |
805 | * if (ffd == FFD_CHILD_PROCESS) |
806 | * _exit(childFn(token)); |
807 | * @endcode |
808 | */ |
809 | int vforkfd(int flags, pid_t *ppid, int (*childFn)(void *), void *token) |
810 | { |
811 | int fd; |
812 | if ((flags & FFD_USE_FORK) == 0) { |
813 | int system_forkfd_works; |
814 | fd = system_vforkfd(flags, ppid, childFn, token, system: &system_forkfd_works); |
815 | if (system_forkfd_works || disable_fork_fallback()) |
816 | return fd; |
817 | } |
818 | |
819 | fd = forkfd_fork_fallback(flags, ppid); |
820 | if (fd == FFD_CHILD_PROCESS) { |
821 | /* child process */ |
822 | _exit(status: childFn(token)); |
823 | } |
824 | return fd; |
825 | } |
826 | #endif // FORKFD_NO_FORKFD |
827 | |
828 | #if _POSIX_SPAWN > 0 && !defined(FORKFD_NO_SPAWNFD) |
829 | int spawnfd(int flags, pid_t *ppid, const char *path, const posix_spawn_file_actions_t *file_actions, |
830 | posix_spawnattr_t *attrp, char *const argv[], char *const envp[]) |
831 | { |
832 | Header *header; |
833 | ProcessInfo *info; |
834 | struct pipe_payload payload; |
835 | pid_t pid; |
836 | int death_pipe[2]; |
837 | int ret = -1; |
838 | /* we can only do work if we have a way to start the child in stopped mode; |
839 | * otherwise, we have a major race condition. */ |
840 | |
841 | assert(!system_has_forkfd()); |
842 | |
843 | (void) pthread_once(&forkfd_initialization, forkfd_initialize); |
844 | |
845 | info = allocateInfo(&header); |
846 | if (info == NULL) { |
847 | errno = ENOMEM; |
848 | goto out; |
849 | } |
850 | |
851 | /* create the pipe before we spawn */ |
852 | if (create_pipe(death_pipe, flags) == -1) |
853 | goto err_free; /* failed to create the pipes, pass errno */ |
854 | |
855 | /* start the process */ |
856 | if (flags & FFD_SPAWN_SEARCH_PATH) { |
857 | /* use posix_spawnp */ |
858 | if (posix_spawnp(&pid, path, file_actions, attrp, argv, envp) != 0) |
859 | goto err_close; |
860 | } else { |
861 | if (posix_spawn(&pid, path, file_actions, attrp, argv, envp) != 0) |
862 | goto err_close; |
863 | } |
864 | |
865 | if (ppid) |
866 | *ppid = pid; |
867 | |
868 | /* Store the child's PID in the info structure. |
869 | */ |
870 | info->deathPipe = death_pipe[1]; |
871 | ffd_atomic_store(&info->pid, pid, FFD_ATOMIC_RELEASE); |
872 | |
873 | /* check if the child has already exited */ |
874 | if (tryReaping(pid, &payload)) |
875 | notifyAndFreeInfo(header, info, &payload); |
876 | |
877 | ret = death_pipe[0]; |
878 | return ret; |
879 | |
880 | err_close: |
881 | EINTR_LOOP(ret, close(death_pipe[0])); |
882 | EINTR_LOOP(ret, close(death_pipe[1])); |
883 | |
884 | err_free: |
885 | /* free the info pointer */ |
886 | freeInfo(header, info); |
887 | |
888 | out: |
889 | return -1; |
890 | } |
891 | #endif // _POSIX_SPAWN && !FORKFD_NO_SPAWNFD |
892 | |
893 | int forkfd_wait4(int ffd, struct forkfd_info *info, int options, struct rusage *rusage) |
894 | { |
895 | struct pipe_payload payload; |
896 | int ret; |
897 | |
898 | if (system_has_forkfd()) { |
899 | /* if this is one of our pipes, not a procdesc/pidfd, we'll get an EBADF */ |
900 | ret = system_forkfd_wait(ffd, info, ffdwoptions: options, rusage); |
901 | if (disable_fork_fallback() || ret != -1 || errno != EBADF) |
902 | return ret; |
903 | } |
904 | |
905 | ret = read(fd: ffd, buf: &payload, nbytes: sizeof(payload)); |
906 | if (ret == -1) |
907 | return ret; /* pass errno, probably EINTR, EBADF or EWOULDBLOCK */ |
908 | |
909 | assert(ret == sizeof(payload)); |
910 | if (info) |
911 | *info = payload.info; |
912 | if (rusage) |
913 | *rusage = payload.rusage; |
914 | |
915 | return 0; /* success */ |
916 | } |
917 | |
918 | |
919 | int forkfd_close(int ffd) |
920 | { |
921 | return close(fd: ffd); |
922 | } |
923 | |
924 | #if defined(__FreeBSD__) && __FreeBSD__ >= 9 |
925 | # include "forkfd_freebsd.c" |
926 | #elif defined(__linux__) |
927 | # include "forkfd_linux.c" |
928 | #else |
929 | int system_has_forkfd() |
930 | { |
931 | return 0; |
932 | } |
933 | |
934 | int system_forkfd(int flags, pid_t *ppid, int *system) |
935 | { |
936 | (void)flags; |
937 | (void)ppid; |
938 | *system = 0; |
939 | return -1; |
940 | } |
941 | |
942 | int system_forkfd_wait(int ffd, struct forkfd_info *info, int options, struct rusage *rusage) |
943 | { |
944 | (void)ffd; |
945 | (void)info; |
946 | (void)options; |
947 | (void)rusage; |
948 | return -1; |
949 | } |
950 | #endif |
951 | #ifndef SYSTEM_FORKFD_CAN_VFORK |
952 | int system_vforkfd(int flags, pid_t *ppid, int (*childFn)(void *), void *token, int *system) |
953 | { |
954 | /* we don't have a way to vfork(), so fake it */ |
955 | int ret = system_forkfd(flags, ppid, system); |
956 | if (ret == FFD_CHILD_PROCESS) { |
957 | /* child process */ |
958 | _exit(childFn(token)); |
959 | } |
960 | return ret; |
961 | } |
962 | #endif |
963 | #undef SYSTEM_FORKFD_CAN_VFORK |
964 | |