1 // Copyright (C) 2015 The Android Open Source Project
2 //
3 // Licensed under the Apache License, Version 2.0 (the "License");
4 // you may not use this file except in compliance with the License.
5 // You may obtain a copy of the License at
6 //
7 // http://www.apache.org/licenses/LICENSE-2.0
8 //
9 // Unless required by applicable law or agreed to in writing, software
10 // distributed under the License is distributed on an "AS IS" BASIS,
11 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 // See the License for the specific language governing permissions and
13 // limitations under the License.
14
15 // makeparallel communicates with the GNU make jobserver
16 // (http://make.mad-scientist.net/papers/jobserver-implementation/)
17 // in order claim all available jobs, and then passes the number of jobs
18 // claimed to a subprocess with -j<jobs>.
19
20 #include <errno.h>
21 #include <fcntl.h>
22 #include <getopt.h>
23 #include <poll.h>
24 #include <signal.h>
25 #include <stdio.h>
26 #include <stdlib.h>
27 #include <string.h>
28 #include <unistd.h>
29 #include <sys/resource.h>
30 #include <sys/time.h>
31 #include <sys/types.h>
32 #include <sys/wait.h>
33
34 #include <string>
35 #include <vector>
36
37 #ifdef __linux__
38 #include <error.h>
39 #endif
40
41 #ifdef __APPLE__
42 #include <err.h>
43 #define error(code, eval, fmt, ...) errc(eval, code, fmt, ##__VA_ARGS__)
44 // Darwin does not interrupt syscalls by default.
45 #define TEMP_FAILURE_RETRY(exp) (exp)
46 #endif
47
48 // Throw an error if fd is not valid.
CheckFd(int fd)49 static void CheckFd(int fd) {
50 int ret = fcntl(fd, F_GETFD);
51 if (ret < 0) {
52 if (errno == EBADF) {
53 error(errno, 0, "no jobserver pipe, prefix recipe command with '+'");
54 } else {
55 error(errno, errno, "fnctl failed");
56 }
57 }
58 }
59
60 // Extract flags from MAKEFLAGS that need to be propagated to subproccess
ReadMakeflags()61 static std::vector<std::string> ReadMakeflags() {
62 std::vector<std::string> args;
63
64 const char* makeflags_env = getenv("MAKEFLAGS");
65 if (makeflags_env == nullptr) {
66 return args;
67 }
68
69 // The MAKEFLAGS format is pretty useless. The first argument might be empty
70 // (starts with a leading space), or it might be a set of one-character flags
71 // merged together with no leading space, or it might be a variable
72 // definition.
73
74 std::string makeflags = makeflags_env;
75
76 // Split makeflags into individual args on spaces. Multiple spaces are
77 // elided, but an initial space will result in a blank arg.
78 size_t base = 0;
79 size_t found;
80 do {
81 found = makeflags.find_first_of(" ", base);
82 args.push_back(makeflags.substr(base, found - base));
83 base = found + 1;
84 } while (found != makeflags.npos);
85
86 // Drop the first argument if it is empty
87 while (args.size() > 0 && args[0].size() == 0) {
88 args.erase(args.begin());
89 }
90
91 // Prepend a - to the first argument if it does not have one and is not a
92 // variable definition
93 if (args.size() > 0 && args[0][0] != '-') {
94 if (args[0].find('=') == makeflags.npos) {
95 args[0] = '-' + args[0];
96 }
97 }
98
99 return args;
100 }
101
ParseMakeflags(std::vector<std::string> & args,int * in_fd,int * out_fd,bool * parallel,bool * keep_going)102 static bool ParseMakeflags(std::vector<std::string>& args,
103 int* in_fd, int* out_fd, bool* parallel, bool* keep_going) {
104
105 std::vector<char*> getopt_argv;
106 // getopt starts reading at argv[1]
107 getopt_argv.reserve(args.size() + 1);
108 getopt_argv.push_back(strdup(""));
109 for (std::string& v : args) {
110 getopt_argv.push_back(strdup(v.c_str()));
111 }
112
113 opterr = 0;
114 optind = 1;
115 while (1) {
116 const static option longopts[] = {
117 {"jobserver-fds", required_argument, 0, 0},
118 {0, 0, 0, 0},
119 };
120 int longopt_index = 0;
121
122 int c = getopt_long(getopt_argv.size(), getopt_argv.data(), "kj",
123 longopts, &longopt_index);
124
125 if (c == -1) {
126 break;
127 }
128
129 switch (c) {
130 case 0:
131 switch (longopt_index) {
132 case 0:
133 {
134 // jobserver-fds
135 if (sscanf(optarg, "%d,%d", in_fd, out_fd) != 2) {
136 error(EXIT_FAILURE, 0, "incorrect format for --jobserver-fds: %s", optarg);
137 }
138 // TODO: propagate in_fd, out_fd
139 break;
140 }
141 default:
142 abort();
143 }
144 break;
145 case 'j':
146 *parallel = true;
147 break;
148 case 'k':
149 *keep_going = true;
150 break;
151 case '?':
152 // ignore unknown arguments
153 break;
154 default:
155 abort();
156 }
157 }
158
159 for (char *v : getopt_argv) {
160 free(v);
161 }
162
163 return true;
164 }
165
166 // Read a single byte from fd, with timeout in milliseconds. Returns true if
167 // a byte was read, false on timeout. Throws away the read value.
168 // Non-reentrant, uses timer and signal handler global state, plus static
169 // variable to communicate with signal handler.
170 //
171 // Uses a SIGALRM timer to fire a signal after timeout_ms that will interrupt
172 // the read syscall if it hasn't yet completed. If the timer fires before the
173 // read the read could block forever, so read from a dup'd fd and close it from
174 // the signal handler, which will cause the read to return EBADF if it occurs
175 // after the signal.
176 // The dup/read/close combo is very similar to the system described to avoid
177 // a deadlock between SIGCHLD and read at
178 // http://make.mad-scientist.net/papers/jobserver-implementation/
ReadByteTimeout(int fd,int timeout_ms)179 static bool ReadByteTimeout(int fd, int timeout_ms) {
180 // global variable to communicate with the signal handler
181 static int dup_fd = -1;
182
183 // dup the fd so the signal handler can close it without losing the real one
184 dup_fd = dup(fd);
185 if (dup_fd < 0) {
186 error(errno, errno, "dup failed");
187 }
188
189 // set up a signal handler that closes dup_fd on SIGALRM
190 struct sigaction action = {};
191 action.sa_flags = SA_SIGINFO,
192 action.sa_sigaction = [](int, siginfo_t*, void*) {
193 close(dup_fd);
194 };
195 struct sigaction oldaction = {};
196 int ret = sigaction(SIGALRM, &action, &oldaction);
197 if (ret < 0) {
198 error(errno, errno, "sigaction failed");
199 }
200
201 // queue a SIGALRM after timeout_ms
202 const struct itimerval timeout = {{}, {0, timeout_ms * 1000}};
203 ret = setitimer(ITIMER_REAL, &timeout, NULL);
204 if (ret < 0) {
205 error(errno, errno, "setitimer failed");
206 }
207
208 // start the blocking read
209 char buf;
210 int read_ret = read(dup_fd, &buf, 1);
211 int read_errno = errno;
212
213 // cancel the alarm in case it hasn't fired yet
214 const struct itimerval cancel = {};
215 ret = setitimer(ITIMER_REAL, &cancel, NULL);
216 if (ret < 0) {
217 error(errno, errno, "reset setitimer failed");
218 }
219
220 // remove the signal handler
221 ret = sigaction(SIGALRM, &oldaction, NULL);
222 if (ret < 0) {
223 error(errno, errno, "reset sigaction failed");
224 }
225
226 // clean up the dup'd fd in case the signal never fired
227 close(dup_fd);
228 dup_fd = -1;
229
230 if (read_ret == 0) {
231 error(EXIT_FAILURE, 0, "EOF on jobserver pipe");
232 } else if (read_ret > 0) {
233 return true;
234 } else if (read_errno == EINTR || read_errno == EBADF) {
235 return false;
236 } else {
237 error(read_errno, read_errno, "read failed");
238 }
239 abort();
240 }
241
242 // Measure the size of the jobserver pool by reading from in_fd until it blocks
GetJobserverTokens(int in_fd)243 static int GetJobserverTokens(int in_fd) {
244 int tokens = 0;
245 pollfd pollfds[] = {{in_fd, POLLIN, 0}};
246 int ret;
247 while ((ret = TEMP_FAILURE_RETRY(poll(pollfds, 1, 0))) != 0) {
248 if (ret < 0) {
249 error(errno, errno, "poll failed");
250 } else if (pollfds[0].revents != POLLIN) {
251 error(EXIT_FAILURE, 0, "unexpected event %d\n", pollfds[0].revents);
252 }
253
254 // There is probably a job token in the jobserver pipe. There is a chance
255 // another process reads it first, which would cause a blocking read to
256 // block forever (or until another process put a token back in the pipe).
257 // The file descriptor can't be set to O_NONBLOCK as that would affect
258 // all users of the pipe, including the parent make process.
259 // ReadByteTimeout emulates a non-blocking read on a !O_NONBLOCK socket
260 // using a SIGALRM that fires after a short timeout.
261 bool got_token = ReadByteTimeout(in_fd, 10);
262 if (!got_token) {
263 // No more tokens
264 break;
265 } else {
266 tokens++;
267 }
268 }
269
270 // This process implicitly gets a token, so pool size is measured size + 1
271 return tokens;
272 }
273
274 // Return tokens to the jobserver pool.
PutJobserverTokens(int out_fd,int tokens)275 static void PutJobserverTokens(int out_fd, int tokens) {
276 // Return all the tokens to the pipe
277 char buf = '+';
278 for (int i = 0; i < tokens; i++) {
279 int ret = TEMP_FAILURE_RETRY(write(out_fd, &buf, 1));
280 if (ret < 0) {
281 error(errno, errno, "write failed");
282 } else if (ret == 0) {
283 error(EXIT_FAILURE, 0, "EOF on jobserver pipe");
284 }
285 }
286 }
287
main(int argc,char * argv[])288 int main(int argc, char* argv[]) {
289 int in_fd = -1;
290 int out_fd = -1;
291 bool parallel = false;
292 bool keep_going = false;
293 bool ninja = false;
294 int tokens = 0;
295
296 if (argc > 1 && strcmp(argv[1], "--ninja") == 0) {
297 ninja = true;
298 argv++;
299 argc--;
300 }
301
302 if (argc < 2) {
303 error(EXIT_FAILURE, 0, "expected command to run");
304 }
305
306 const char* path = argv[1];
307 std::vector<char*> args({argv[1]});
308
309 std::vector<std::string> makeflags = ReadMakeflags();
310 if (ParseMakeflags(makeflags, &in_fd, &out_fd, ¶llel, &keep_going)) {
311 if (in_fd >= 0 && out_fd >= 0) {
312 CheckFd(in_fd);
313 CheckFd(out_fd);
314 fcntl(in_fd, F_SETFD, FD_CLOEXEC);
315 fcntl(out_fd, F_SETFD, FD_CLOEXEC);
316 tokens = GetJobserverTokens(in_fd);
317 }
318 }
319
320 std::string jarg;
321 if (parallel) {
322 if (tokens == 0) {
323 if (ninja) {
324 // ninja is parallel by default
325 jarg = "";
326 } else {
327 // make -j with no argument, guess a reasonable parallelism like ninja does
328 jarg = "-j" + std::to_string(sysconf(_SC_NPROCESSORS_ONLN) + 2);
329 }
330 } else {
331 jarg = "-j" + std::to_string(tokens + 1);
332 }
333 }
334
335
336 if (ninja) {
337 if (!parallel) {
338 // ninja is parallel by default, pass -j1 to disable parallelism if make wasn't parallel
339 args.push_back(strdup("-j1"));
340 } else {
341 if (jarg != "") {
342 args.push_back(strdup(jarg.c_str()));
343 }
344 }
345 if (keep_going) {
346 args.push_back(strdup("-k0"));
347 }
348 } else {
349 if (jarg != "") {
350 args.push_back(strdup(jarg.c_str()));
351 }
352 }
353
354 args.insert(args.end(), &argv[2], &argv[argc]);
355
356 args.push_back(nullptr);
357
358 static pid_t pid;
359
360 // Set up signal handlers to forward SIGTERM to child.
361 // Assume that all other signals are sent to the entire process group,
362 // and that we'll wait for our child to exit instead of handling them.
363 struct sigaction action = {};
364 action.sa_flags = SA_RESTART;
365 action.sa_handler = [](int signal) {
366 if (signal == SIGTERM && pid > 0) {
367 kill(pid, signal);
368 }
369 };
370
371 int ret = 0;
372 if (!ret) ret = sigaction(SIGHUP, &action, NULL);
373 if (!ret) ret = sigaction(SIGINT, &action, NULL);
374 if (!ret) ret = sigaction(SIGQUIT, &action, NULL);
375 if (!ret) ret = sigaction(SIGTERM, &action, NULL);
376 if (!ret) ret = sigaction(SIGALRM, &action, NULL);
377 if (ret < 0) {
378 error(errno, errno, "sigaction failed");
379 }
380
381 pid = fork();
382 if (pid < 0) {
383 error(errno, errno, "fork failed");
384 } else if (pid == 0) {
385 // child
386 unsetenv("MAKEFLAGS");
387 unsetenv("MAKELEVEL");
388
389 // make 3.81 sets the stack ulimit to unlimited, which may cause problems
390 // for child processes
391 struct rlimit rlim{};
392 if (getrlimit(RLIMIT_STACK, &rlim) == 0 && rlim.rlim_cur == RLIM_INFINITY) {
393 rlim.rlim_cur = 8*1024*1024;
394 setrlimit(RLIMIT_STACK, &rlim);
395 }
396
397 int ret = execvp(path, args.data());
398 if (ret < 0) {
399 error(errno, errno, "exec %s failed", path);
400 }
401 abort();
402 }
403
404 // parent
405
406 siginfo_t status = {};
407 int exit_status = 0;
408 ret = waitid(P_PID, pid, &status, WEXITED);
409 if (ret < 0) {
410 error(errno, errno, "waitpid failed");
411 } else if (status.si_code == CLD_EXITED) {
412 exit_status = status.si_status;
413 } else {
414 exit_status = -(status.si_status);
415 }
416
417 if (tokens > 0) {
418 PutJobserverTokens(out_fd, tokens);
419 }
420 exit(exit_status);
421 }
422