1 /*
2 * Copyright 2016, The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17 #include <fcntl.h>
18 #include <stdio.h>
19 #include <stdlib.h>
20 #include <sys/stat.h>
21 #include <sys/types.h>
22 #include <unistd.h>
23
24 #include <array>
25 #include <deque>
26 #include <string>
27 #include <unordered_map>
28 #include <utility>
29
30 #include <event2/event.h>
31 #include <event2/listener.h>
32 #include <event2/thread.h>
33
34 #include <android-base/cmsg.h>
35 #include <android-base/logging.h>
36 #include <android-base/properties.h>
37 #include <android-base/stringprintf.h>
38 #include <android-base/unique_fd.h>
39 #include <cutils/sockets.h>
40
41 #include "debuggerd/handler.h"
42 #include "dump_type.h"
43 #include "protocol.h"
44 #include "util.h"
45
46 #include "intercept_manager.h"
47
48 using android::base::GetIntProperty;
49 using android::base::SendFileDescriptors;
50 using android::base::StringPrintf;
51 using android::base::unique_fd;
52
53 static InterceptManager* intercept_manager;
54
55 enum CrashStatus {
56 kCrashStatusRunning,
57 kCrashStatusQueued,
58 };
59
60 // Ownership of Crash is a bit messy.
61 // It's either owned by an active event that must have a timeout, or owned by
62 // queued_requests, in the case that multiple crashes come in at the same time.
63 struct Crash {
~CrashCrash64 ~Crash() { event_free(crash_event); }
65
66 std::string crash_tombstone_path;
67 unique_fd crash_tombstone_fd;
68 unique_fd crash_socket_fd;
69 pid_t crash_pid;
70 event* crash_event = nullptr;
71
72 DebuggerdDumpType crash_type;
73 };
74
75 class CrashQueue {
76 public:
CrashQueue(const std::string & dir_path,const std::string & file_name_prefix,size_t max_artifacts,size_t max_concurrent_dumps)77 CrashQueue(const std::string& dir_path, const std::string& file_name_prefix, size_t max_artifacts,
78 size_t max_concurrent_dumps)
79 : file_name_prefix_(file_name_prefix),
80 dir_path_(dir_path),
81 dir_fd_(open(dir_path.c_str(), O_DIRECTORY | O_RDONLY | O_CLOEXEC)),
82 max_artifacts_(max_artifacts),
83 next_artifact_(0),
84 max_concurrent_dumps_(max_concurrent_dumps),
85 num_concurrent_dumps_(0) {
86 if (dir_fd_ == -1) {
87 PLOG(FATAL) << "failed to open directory: " << dir_path;
88 }
89
90 // NOTE: If max_artifacts_ <= max_concurrent_dumps_, then theoretically the
91 // same filename could be handed out to multiple processes.
92 CHECK(max_artifacts_ > max_concurrent_dumps_);
93
94 find_oldest_artifact();
95 }
96
for_crash(const Crash * crash)97 static CrashQueue* for_crash(const Crash* crash) {
98 return (crash->crash_type == kDebuggerdJavaBacktrace) ? for_anrs() : for_tombstones();
99 }
100
for_tombstones()101 static CrashQueue* for_tombstones() {
102 static CrashQueue queue("/data/tombstones", "tombstone_" /* file_name_prefix */,
103 GetIntProperty("tombstoned.max_tombstone_count", 10),
104 1 /* max_concurrent_dumps */);
105 return &queue;
106 }
107
for_anrs()108 static CrashQueue* for_anrs() {
109 static CrashQueue queue("/data/anr", "trace_" /* file_name_prefix */,
110 GetIntProperty("tombstoned.max_anr_count", 64),
111 4 /* max_concurrent_dumps */);
112 return &queue;
113 }
114
get_output()115 std::pair<std::string, unique_fd> get_output() {
116 std::string path;
117 unique_fd result(openat(dir_fd_, ".", O_WRONLY | O_APPEND | O_TMPFILE | O_CLOEXEC, 0640));
118 if (result == -1) {
119 // We might not have O_TMPFILE. Try creating with an arbitrary filename instead.
120 static size_t counter = 0;
121 std::string tmp_filename = StringPrintf(".temporary%zu", counter++);
122 result.reset(openat(dir_fd_, tmp_filename.c_str(),
123 O_WRONLY | O_APPEND | O_CREAT | O_TRUNC | O_CLOEXEC, 0640));
124 if (result == -1) {
125 PLOG(FATAL) << "failed to create temporary tombstone in " << dir_path_;
126 }
127
128 path = StringPrintf("%s/%s", dir_path_.c_str(), tmp_filename.c_str());
129 }
130 return std::make_pair(std::move(path), std::move(result));
131 }
132
get_next_artifact_path()133 std::string get_next_artifact_path() {
134 std::string file_name =
135 StringPrintf("%s/%s%02d", dir_path_.c_str(), file_name_prefix_.c_str(), next_artifact_);
136 next_artifact_ = (next_artifact_ + 1) % max_artifacts_;
137 return file_name;
138 }
139
maybe_enqueue_crash(Crash * crash)140 bool maybe_enqueue_crash(Crash* crash) {
141 if (num_concurrent_dumps_ == max_concurrent_dumps_) {
142 queued_requests_.push_back(crash);
143 return true;
144 }
145
146 return false;
147 }
148
maybe_dequeue_crashes(void (* handler)(Crash * crash))149 void maybe_dequeue_crashes(void (*handler)(Crash* crash)) {
150 while (!queued_requests_.empty() && num_concurrent_dumps_ < max_concurrent_dumps_) {
151 Crash* next_crash = queued_requests_.front();
152 queued_requests_.pop_front();
153 handler(next_crash);
154 }
155 }
156
on_crash_started()157 void on_crash_started() { ++num_concurrent_dumps_; }
158
on_crash_completed()159 void on_crash_completed() { --num_concurrent_dumps_; }
160
161 private:
find_oldest_artifact()162 void find_oldest_artifact() {
163 size_t oldest_tombstone = 0;
164 time_t oldest_time = std::numeric_limits<time_t>::max();
165
166 for (size_t i = 0; i < max_artifacts_; ++i) {
167 std::string path = StringPrintf("%s/%s%02zu", dir_path_.c_str(), file_name_prefix_.c_str(), i);
168 struct stat st;
169 if (stat(path.c_str(), &st) != 0) {
170 if (errno == ENOENT) {
171 oldest_tombstone = i;
172 break;
173 } else {
174 PLOG(ERROR) << "failed to stat " << path;
175 continue;
176 }
177 }
178
179 if (st.st_mtime < oldest_time) {
180 oldest_tombstone = i;
181 oldest_time = st.st_mtime;
182 }
183 }
184
185 next_artifact_ = oldest_tombstone;
186 }
187
188 const std::string file_name_prefix_;
189
190 const std::string dir_path_;
191 const int dir_fd_;
192
193 const size_t max_artifacts_;
194 int next_artifact_;
195
196 const size_t max_concurrent_dumps_;
197 size_t num_concurrent_dumps_;
198
199 std::deque<Crash*> queued_requests_;
200
201 DISALLOW_COPY_AND_ASSIGN(CrashQueue);
202 };
203
204 // Whether java trace dumps are produced via tombstoned.
205 static constexpr bool kJavaTraceDumpsEnabled = true;
206
207 // Forward declare the callbacks so they can be placed in a sensible order.
208 static void crash_accept_cb(evconnlistener* listener, evutil_socket_t sockfd, sockaddr*, int, void*);
209 static void crash_request_cb(evutil_socket_t sockfd, short ev, void* arg);
210 static void crash_completed_cb(evutil_socket_t sockfd, short ev, void* arg);
211
perform_request(Crash * crash)212 static void perform_request(Crash* crash) {
213 unique_fd output_fd;
214 bool intercepted =
215 intercept_manager->GetIntercept(crash->crash_pid, crash->crash_type, &output_fd);
216 if (!intercepted) {
217 if (crash->crash_type == kDebuggerdNativeBacktrace) {
218 // Don't generate tombstones for native backtrace requests.
219 output_fd.reset(open("/dev/null", O_WRONLY | O_CLOEXEC));
220 } else {
221 std::tie(crash->crash_tombstone_path, output_fd) = CrashQueue::for_crash(crash)->get_output();
222 crash->crash_tombstone_fd.reset(dup(output_fd.get()));
223 }
224 }
225
226 TombstonedCrashPacket response = {
227 .packet_type = CrashPacketType::kPerformDump
228 };
229 ssize_t rc =
230 SendFileDescriptors(crash->crash_socket_fd, &response, sizeof(response), output_fd.get());
231 output_fd.reset();
232
233 if (rc == -1) {
234 PLOG(WARNING) << "failed to send response to CrashRequest";
235 goto fail;
236 } else if (rc != sizeof(response)) {
237 PLOG(WARNING) << "crash socket write returned short";
238 goto fail;
239 } else {
240 // TODO: Make this configurable by the interceptor?
241 struct timeval timeout = { 10, 0 };
242
243 event_base* base = event_get_base(crash->crash_event);
244 event_assign(crash->crash_event, base, crash->crash_socket_fd, EV_TIMEOUT | EV_READ,
245 crash_completed_cb, crash);
246 event_add(crash->crash_event, &timeout);
247 }
248
249 CrashQueue::for_crash(crash)->on_crash_started();
250 return;
251
252 fail:
253 delete crash;
254 }
255
crash_accept_cb(evconnlistener * listener,evutil_socket_t sockfd,sockaddr *,int,void *)256 static void crash_accept_cb(evconnlistener* listener, evutil_socket_t sockfd, sockaddr*, int,
257 void*) {
258 event_base* base = evconnlistener_get_base(listener);
259 Crash* crash = new Crash();
260
261 // TODO: Make sure that only java crashes come in on the java socket
262 // and only native crashes on the native socket.
263 struct timeval timeout = { 1, 0 };
264 event* crash_event = event_new(base, sockfd, EV_TIMEOUT | EV_READ, crash_request_cb, crash);
265 crash->crash_socket_fd.reset(sockfd);
266 crash->crash_event = crash_event;
267 event_add(crash_event, &timeout);
268 }
269
crash_request_cb(evutil_socket_t sockfd,short ev,void * arg)270 static void crash_request_cb(evutil_socket_t sockfd, short ev, void* arg) {
271 ssize_t rc;
272 Crash* crash = static_cast<Crash*>(arg);
273
274 TombstonedCrashPacket request = {};
275
276 if ((ev & EV_TIMEOUT) != 0) {
277 LOG(WARNING) << "crash request timed out";
278 goto fail;
279 } else if ((ev & EV_READ) == 0) {
280 LOG(WARNING) << "tombstoned received unexpected event from crash socket";
281 goto fail;
282 }
283
284 rc = TEMP_FAILURE_RETRY(read(sockfd, &request, sizeof(request)));
285 if (rc == -1) {
286 PLOG(WARNING) << "failed to read from crash socket";
287 goto fail;
288 } else if (rc != sizeof(request)) {
289 LOG(WARNING) << "crash socket received short read of length " << rc << " (expected "
290 << sizeof(request) << ")";
291 goto fail;
292 }
293
294 if (request.packet_type != CrashPacketType::kDumpRequest) {
295 LOG(WARNING) << "unexpected crash packet type, expected kDumpRequest, received "
296 << StringPrintf("%#2hhX", request.packet_type);
297 goto fail;
298 }
299
300 crash->crash_type = request.packet.dump_request.dump_type;
301 if (crash->crash_type < 0 || crash->crash_type > kDebuggerdAnyIntercept) {
302 LOG(WARNING) << "unexpected crash dump type: " << crash->crash_type;
303 goto fail;
304 }
305
306 if (crash->crash_type != kDebuggerdJavaBacktrace) {
307 crash->crash_pid = request.packet.dump_request.pid;
308 } else {
309 // Requests for java traces are sent from untrusted processes, so we
310 // must not trust the PID sent down with the request. Instead, we ask the
311 // kernel.
312 ucred cr = {};
313 socklen_t len = sizeof(cr);
314 int ret = getsockopt(sockfd, SOL_SOCKET, SO_PEERCRED, &cr, &len);
315 if (ret != 0) {
316 PLOG(ERROR) << "Failed to getsockopt(..SO_PEERCRED)";
317 goto fail;
318 }
319
320 crash->crash_pid = cr.pid;
321 }
322
323 LOG(INFO) << "received crash request for pid " << crash->crash_pid;
324
325 if (CrashQueue::for_crash(crash)->maybe_enqueue_crash(crash)) {
326 LOG(INFO) << "enqueueing crash request for pid " << crash->crash_pid;
327 } else {
328 perform_request(crash);
329 }
330
331 return;
332
333 fail:
334 delete crash;
335 }
336
crash_completed_cb(evutil_socket_t sockfd,short ev,void * arg)337 static void crash_completed_cb(evutil_socket_t sockfd, short ev, void* arg) {
338 ssize_t rc;
339 Crash* crash = static_cast<Crash*>(arg);
340 TombstonedCrashPacket request = {};
341
342 CrashQueue::for_crash(crash)->on_crash_completed();
343
344 if ((ev & EV_READ) == 0) {
345 goto fail;
346 }
347
348 rc = TEMP_FAILURE_RETRY(read(sockfd, &request, sizeof(request)));
349 if (rc == -1) {
350 PLOG(WARNING) << "failed to read from crash socket";
351 goto fail;
352 } else if (rc != sizeof(request)) {
353 LOG(WARNING) << "crash socket received short read of length " << rc << " (expected "
354 << sizeof(request) << ")";
355 goto fail;
356 }
357
358 if (request.packet_type != CrashPacketType::kCompletedDump) {
359 LOG(WARNING) << "unexpected crash packet type, expected kCompletedDump, received "
360 << uint32_t(request.packet_type);
361 goto fail;
362 }
363
364 if (crash->crash_tombstone_fd != -1) {
365 std::string fd_path = StringPrintf("/proc/self/fd/%d", crash->crash_tombstone_fd.get());
366 std::string tombstone_path = CrashQueue::for_crash(crash)->get_next_artifact_path();
367
368 // linkat doesn't let us replace a file, so we need to unlink first.
369 int rc = unlink(tombstone_path.c_str());
370 if (rc != 0 && errno != ENOENT) {
371 PLOG(ERROR) << "failed to unlink tombstone at " << tombstone_path;
372 goto fail;
373 }
374
375 rc = linkat(AT_FDCWD, fd_path.c_str(), AT_FDCWD, tombstone_path.c_str(), AT_SYMLINK_FOLLOW);
376 if (rc != 0) {
377 PLOG(ERROR) << "failed to link tombstone";
378 } else {
379 if (crash->crash_type == kDebuggerdJavaBacktrace) {
380 LOG(ERROR) << "Traces for pid " << crash->crash_pid << " written to: " << tombstone_path;
381 } else {
382 // NOTE: Several tools parse this log message to figure out where the
383 // tombstone associated with a given native crash was written. Any changes
384 // to this message must be carefully considered.
385 LOG(ERROR) << "Tombstone written to: " << tombstone_path;
386 }
387 }
388
389 // If we don't have O_TMPFILE, we need to clean up after ourselves.
390 if (!crash->crash_tombstone_path.empty()) {
391 rc = unlink(crash->crash_tombstone_path.c_str());
392 if (rc != 0) {
393 PLOG(ERROR) << "failed to unlink temporary tombstone at " << crash->crash_tombstone_path;
394 }
395 }
396 }
397
398 fail:
399 CrashQueue* queue = CrashQueue::for_crash(crash);
400 delete crash;
401
402 // If there's something queued up, let them proceed.
403 queue->maybe_dequeue_crashes(perform_request);
404 }
405
main(int,char * [])406 int main(int, char* []) {
407 umask(0137);
408
409 // Don't try to connect to ourselves if we crash.
410 struct sigaction action = {};
411 action.sa_handler = [](int signal) {
412 LOG(ERROR) << "received fatal signal " << signal;
413 _exit(1);
414 };
415 debuggerd_register_handlers(&action);
416
417 int intercept_socket = android_get_control_socket(kTombstonedInterceptSocketName);
418 int crash_socket = android_get_control_socket(kTombstonedCrashSocketName);
419
420 if (intercept_socket == -1 || crash_socket == -1) {
421 PLOG(FATAL) << "failed to get socket from init";
422 }
423
424 evutil_make_socket_nonblocking(intercept_socket);
425 evutil_make_socket_nonblocking(crash_socket);
426
427 event_base* base = event_base_new();
428 if (!base) {
429 LOG(FATAL) << "failed to create event_base";
430 }
431
432 intercept_manager = new InterceptManager(base, intercept_socket);
433
434 evconnlistener* tombstone_listener =
435 evconnlistener_new(base, crash_accept_cb, CrashQueue::for_tombstones(), LEV_OPT_CLOSE_ON_FREE,
436 -1 /* backlog */, crash_socket);
437 if (!tombstone_listener) {
438 LOG(FATAL) << "failed to create evconnlistener for tombstones.";
439 }
440
441 if (kJavaTraceDumpsEnabled) {
442 const int java_trace_socket = android_get_control_socket(kTombstonedJavaTraceSocketName);
443 if (java_trace_socket == -1) {
444 PLOG(FATAL) << "failed to get socket from init";
445 }
446
447 evutil_make_socket_nonblocking(java_trace_socket);
448 evconnlistener* java_trace_listener =
449 evconnlistener_new(base, crash_accept_cb, CrashQueue::for_anrs(), LEV_OPT_CLOSE_ON_FREE,
450 -1 /* backlog */, java_trace_socket);
451 if (!java_trace_listener) {
452 LOG(FATAL) << "failed to create evconnlistener for java traces.";
453 }
454 }
455
456 LOG(INFO) << "tombstoned successfully initialized";
457 event_base_dispatch(base);
458 }
459