1/*
2 * Copyright 2006, 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 <stdio.h>
18#include <errno.h>
19#include <signal.h>
20#include <pthread.h>
21#include <stdarg.h>
22#include <fcntl.h>
23#include <sys/types.h>
24#include <dirent.h>
25#include <time.h>
26
27#include <sys/ptrace.h>
28#include <sys/wait.h>
29#include <elf.h>
30#include <sys/stat.h>
31#include <sys/poll.h>
32
33#include <log/logger.h>
34
35#include <cutils/sockets.h>
36#include <cutils/properties.h>
37#include <cutils/debugger.h>
38
39#include <linux/input.h>
40
41#include <private/android_filesystem_config.h>
42
43#include "backtrace.h"
44#include "getevent.h"
45#include "tombstone.h"
46#include "utility.h"
47
48struct debugger_request_t {
49  debugger_action_t action;
50  pid_t pid, tid;
51  uid_t uid, gid;
52  uintptr_t abort_msg_address;
53  int32_t original_si_code;
54};
55
56static void wait_for_user_action(const debugger_request_t &request) {
57  // Find out the name of the process that crashed.
58  char path[64];
59  snprintf(path, sizeof(path), "/proc/%d/exe", request.pid);
60
61  char exe[PATH_MAX];
62  int count;
63  if ((count = readlink(path, exe, sizeof(exe) - 1)) == -1) {
64    ALOGE("readlink('%s') failed: %s", path, strerror(errno));
65    strlcpy(exe, "unknown", sizeof(exe));
66  } else {
67    exe[count] = '\0';
68  }
69
70  // Explain how to attach the debugger.
71  ALOGI("********************************************************\n"
72        "* Process %d has been suspended while crashing.\n"
73        "* To attach gdbserver for a gdb connection on port 5039\n"
74        "* and start gdbclient:\n"
75        "*\n"
76        "*     gdbclient %s :5039 %d\n"
77        "*\n"
78        "* Wait for gdb to start, then press the VOLUME DOWN key\n"
79        "* to let the process continue crashing.\n"
80        "********************************************************\n",
81        request.pid, exe, request.tid);
82
83  // Wait for VOLUME DOWN.
84  if (init_getevent() == 0) {
85    while (true) {
86      input_event e;
87      if (get_event(&e, -1) == 0) {
88        if (e.type == EV_KEY && e.code == KEY_VOLUMEDOWN && e.value == 0) {
89          break;
90        }
91      }
92    }
93    uninit_getevent();
94  }
95
96  ALOGI("debuggerd resuming process %d", request.pid);
97}
98
99static int get_process_info(pid_t tid, pid_t* out_pid, uid_t* out_uid, uid_t* out_gid) {
100  char path[64];
101  snprintf(path, sizeof(path), "/proc/%d/status", tid);
102
103  FILE* fp = fopen(path, "r");
104  if (!fp) {
105    return -1;
106  }
107
108  int fields = 0;
109  char line[1024];
110  while (fgets(line, sizeof(line), fp)) {
111    size_t len = strlen(line);
112    if (len > 6 && !memcmp(line, "Tgid:\t", 6)) {
113      *out_pid = atoi(line + 6);
114      fields |= 1;
115    } else if (len > 5 && !memcmp(line, "Uid:\t", 5)) {
116      *out_uid = atoi(line + 5);
117      fields |= 2;
118    } else if (len > 5 && !memcmp(line, "Gid:\t", 5)) {
119      *out_gid = atoi(line + 5);
120      fields |= 4;
121    }
122  }
123  fclose(fp);
124  return fields == 7 ? 0 : -1;
125}
126
127static int read_request(int fd, debugger_request_t* out_request) {
128  ucred cr;
129  socklen_t len = sizeof(cr);
130  int status = getsockopt(fd, SOL_SOCKET, SO_PEERCRED, &cr, &len);
131  if (status != 0) {
132    ALOGE("cannot get credentials\n");
133    return -1;
134  }
135
136  ALOGV("reading tid\n");
137  fcntl(fd, F_SETFL, O_NONBLOCK);
138
139  pollfd pollfds[1];
140  pollfds[0].fd = fd;
141  pollfds[0].events = POLLIN;
142  pollfds[0].revents = 0;
143  status = TEMP_FAILURE_RETRY(poll(pollfds, 1, 3000));
144  if (status != 1) {
145    ALOGE("timed out reading tid (from pid=%d uid=%d)\n", cr.pid, cr.uid);
146    return -1;
147  }
148
149  debugger_msg_t msg;
150  memset(&msg, 0, sizeof(msg));
151  status = TEMP_FAILURE_RETRY(read(fd, &msg, sizeof(msg)));
152  if (status < 0) {
153    ALOGE("read failure? %s (pid=%d uid=%d)\n", strerror(errno), cr.pid, cr.uid);
154    return -1;
155  }
156  if (status != sizeof(debugger_msg_t)) {
157    ALOGE("invalid crash request of size %d (from pid=%d uid=%d)\n", status, cr.pid, cr.uid);
158    return -1;
159  }
160
161  out_request->action = msg.action;
162  out_request->tid = msg.tid;
163  out_request->pid = cr.pid;
164  out_request->uid = cr.uid;
165  out_request->gid = cr.gid;
166  out_request->abort_msg_address = msg.abort_msg_address;
167  out_request->original_si_code = msg.original_si_code;
168
169  if (msg.action == DEBUGGER_ACTION_CRASH) {
170    // Ensure that the tid reported by the crashing process is valid.
171    char buf[64];
172    struct stat s;
173    snprintf(buf, sizeof buf, "/proc/%d/task/%d", out_request->pid, out_request->tid);
174    if (stat(buf, &s)) {
175      ALOGE("tid %d does not exist in pid %d. ignoring debug request\n",
176          out_request->tid, out_request->pid);
177      return -1;
178    }
179  } else if (cr.uid == 0
180            || (cr.uid == AID_SYSTEM && msg.action == DEBUGGER_ACTION_DUMP_BACKTRACE)) {
181    // Only root or system can ask us to attach to any process and dump it explicitly.
182    // However, system is only allowed to collect backtraces but cannot dump tombstones.
183    status = get_process_info(out_request->tid, &out_request->pid,
184                              &out_request->uid, &out_request->gid);
185    if (status < 0) {
186      ALOGE("tid %d does not exist. ignoring explicit dump request\n", out_request->tid);
187      return -1;
188    }
189  } else {
190    // No one else is allowed to dump arbitrary processes.
191    return -1;
192  }
193  return 0;
194}
195
196static bool should_attach_gdb(debugger_request_t* request) {
197  if (request->action == DEBUGGER_ACTION_CRASH) {
198    char value[PROPERTY_VALUE_MAX];
199    property_get("debug.db.uid", value, "-1");
200    int debug_uid = atoi(value);
201    return debug_uid >= 0 && request->uid <= (uid_t)debug_uid;
202  }
203  return false;
204}
205
206static void handle_request(int fd) {
207  ALOGV("handle_request(%d)\n", fd);
208
209  debugger_request_t request;
210  memset(&request, 0, sizeof(request));
211  int status = read_request(fd, &request);
212  if (!status) {
213    ALOGV("BOOM: pid=%d uid=%d gid=%d tid=%d\n",
214         request.pid, request.uid, request.gid, request.tid);
215
216    // At this point, the thread that made the request is blocked in
217    // a read() call.  If the thread has crashed, then this gives us
218    // time to PTRACE_ATTACH to it before it has a chance to really fault.
219    //
220    // The PTRACE_ATTACH sends a SIGSTOP to the target process, but it
221    // won't necessarily have stopped by the time ptrace() returns.  (We
222    // currently assume it does.)  We write to the file descriptor to
223    // ensure that it can run as soon as we call PTRACE_CONT below.
224    // See details in bionic/libc/linker/debugger.c, in function
225    // debugger_signal_handler().
226    if (ptrace(PTRACE_ATTACH, request.tid, 0, 0)) {
227      ALOGE("ptrace attach failed: %s\n", strerror(errno));
228    } else {
229      bool detach_failed = false;
230      bool attach_gdb = should_attach_gdb(&request);
231      if (TEMP_FAILURE_RETRY(write(fd, "\0", 1)) != 1) {
232        ALOGE("failed responding to client: %s\n", strerror(errno));
233      } else {
234        char* tombstone_path = NULL;
235
236        if (request.action == DEBUGGER_ACTION_CRASH) {
237          close(fd);
238          fd = -1;
239        }
240
241        int total_sleep_time_usec = 0;
242        for (;;) {
243          int signal = wait_for_signal(request.tid, &total_sleep_time_usec);
244          if (signal < 0) {
245            break;
246          }
247
248          switch (signal) {
249            case SIGSTOP:
250              if (request.action == DEBUGGER_ACTION_DUMP_TOMBSTONE) {
251                ALOGV("stopped -- dumping to tombstone\n");
252                tombstone_path = engrave_tombstone(request.pid, request.tid,
253                                                   signal, request.original_si_code,
254                                                   request.abort_msg_address, true,
255                                                   &detach_failed, &total_sleep_time_usec);
256              } else if (request.action == DEBUGGER_ACTION_DUMP_BACKTRACE) {
257                ALOGV("stopped -- dumping to fd\n");
258                dump_backtrace(fd, -1, request.pid, request.tid, &detach_failed,
259                               &total_sleep_time_usec);
260              } else {
261                ALOGV("stopped -- continuing\n");
262                status = ptrace(PTRACE_CONT, request.tid, 0, 0);
263                if (status) {
264                  ALOGE("ptrace continue failed: %s\n", strerror(errno));
265                }
266                continue; // loop again
267              }
268              break;
269
270            case SIGABRT:
271            case SIGBUS:
272            case SIGFPE:
273            case SIGILL:
274            case SIGPIPE:
275            case SIGSEGV:
276#ifdef SIGSTKFLT
277            case SIGSTKFLT:
278#endif
279            case SIGTRAP:
280              ALOGV("stopped -- fatal signal\n");
281              // Send a SIGSTOP to the process to make all of
282              // the non-signaled threads stop moving.  Without
283              // this we get a lot of "ptrace detach failed:
284              // No such process".
285              kill(request.pid, SIGSTOP);
286              // don't dump sibling threads when attaching to GDB because it
287              // makes the process less reliable, apparently...
288              tombstone_path = engrave_tombstone(request.pid, request.tid,
289                                                 signal, request.original_si_code,
290                                                 request.abort_msg_address, !attach_gdb,
291                                                 &detach_failed, &total_sleep_time_usec);
292              break;
293
294            default:
295              ALOGE("process stopped due to unexpected signal %d\n", signal);
296              break;
297          }
298          break;
299        }
300
301        if (request.action == DEBUGGER_ACTION_DUMP_TOMBSTONE) {
302          if (tombstone_path) {
303            write(fd, tombstone_path, strlen(tombstone_path));
304          }
305          close(fd);
306          fd = -1;
307        }
308        free(tombstone_path);
309      }
310
311      ALOGV("detaching\n");
312      if (attach_gdb) {
313        // stop the process so we can debug
314        kill(request.pid, SIGSTOP);
315
316        // detach so we can attach gdbserver
317        if (ptrace(PTRACE_DETACH, request.tid, 0, 0)) {
318          ALOGE("ptrace detach from %d failed: %s\n", request.tid, strerror(errno));
319          detach_failed = true;
320        }
321
322        // if debug.db.uid is set, its value indicates if we should wait
323        // for user action for the crashing process.
324        // in this case, we log a message and turn the debug LED on
325        // waiting for a gdb connection (for instance)
326        wait_for_user_action(request);
327      } else {
328        // just detach
329        if (ptrace(PTRACE_DETACH, request.tid, 0, 0)) {
330          ALOGE("ptrace detach from %d failed: %s\n", request.tid, strerror(errno));
331          detach_failed = true;
332        }
333      }
334
335      // resume stopped process (so it can crash in peace).
336      kill(request.pid, SIGCONT);
337
338      // If we didn't successfully detach, we're still the parent, and the
339      // actual parent won't receive a death notification via wait(2).  At this point
340      // there's not much we can do about that.
341      if (detach_failed) {
342        ALOGE("debuggerd committing suicide to free the zombie!\n");
343        kill(getpid(), SIGKILL);
344      }
345    }
346
347  }
348  if (fd >= 0) {
349    close(fd);
350  }
351}
352
353static int do_server() {
354  // debuggerd crashes can't be reported to debuggerd.
355  // Reset all of the crash handlers.
356  signal(SIGABRT, SIG_DFL);
357  signal(SIGBUS, SIG_DFL);
358  signal(SIGFPE, SIG_DFL);
359  signal(SIGILL, SIG_DFL);
360  signal(SIGSEGV, SIG_DFL);
361#ifdef SIGSTKFLT
362  signal(SIGSTKFLT, SIG_DFL);
363#endif
364  signal(SIGTRAP, SIG_DFL);
365
366  // Ignore failed writes to closed sockets
367  signal(SIGPIPE, SIG_IGN);
368
369  int logsocket = socket_local_client("logd", ANDROID_SOCKET_NAMESPACE_ABSTRACT, SOCK_DGRAM);
370  if (logsocket < 0) {
371    logsocket = -1;
372  } else {
373    fcntl(logsocket, F_SETFD, FD_CLOEXEC);
374  }
375
376  struct sigaction act;
377  act.sa_handler = SIG_DFL;
378  sigemptyset(&act.sa_mask);
379  sigaddset(&act.sa_mask,SIGCHLD);
380  act.sa_flags = SA_NOCLDWAIT;
381  sigaction(SIGCHLD, &act, 0);
382
383  int s = socket_local_server(DEBUGGER_SOCKET_NAME, ANDROID_SOCKET_NAMESPACE_ABSTRACT, SOCK_STREAM);
384  if (s < 0)
385    return 1;
386  fcntl(s, F_SETFD, FD_CLOEXEC);
387
388  ALOGI("debuggerd: " __DATE__ " " __TIME__ "\n");
389
390  for (;;) {
391    sockaddr addr;
392    socklen_t alen = sizeof(addr);
393
394    ALOGV("waiting for connection\n");
395    int fd = accept(s, &addr, &alen);
396    if (fd < 0) {
397      ALOGV("accept failed: %s\n", strerror(errno));
398      continue;
399    }
400
401    fcntl(fd, F_SETFD, FD_CLOEXEC);
402
403    handle_request(fd);
404  }
405  return 0;
406}
407
408static int do_explicit_dump(pid_t tid, bool dump_backtrace) {
409  fprintf(stdout, "Sending request to dump task %d.\n", tid);
410
411  if (dump_backtrace) {
412    fflush(stdout);
413    if (dump_backtrace_to_file(tid, fileno(stdout)) < 0) {
414      fputs("Error dumping backtrace.\n", stderr);
415      return 1;
416    }
417  } else {
418    char tombstone_path[PATH_MAX];
419    if (dump_tombstone(tid, tombstone_path, sizeof(tombstone_path)) < 0) {
420      fputs("Error dumping tombstone.\n", stderr);
421      return 1;
422    }
423    fprintf(stderr, "Tombstone written to: %s\n", tombstone_path);
424  }
425  return 0;
426}
427
428static void usage() {
429  fputs("Usage: -b [<tid>]\n"
430        "  -b dump backtrace to console, otherwise dump full tombstone file\n"
431        "\n"
432        "If tid specified, sends a request to debuggerd to dump that task.\n"
433        "Otherwise, starts the debuggerd server.\n", stderr);
434}
435
436int main(int argc, char** argv) {
437  if (argc == 1) {
438    return do_server();
439  }
440
441  bool dump_backtrace = false;
442  bool have_tid = false;
443  pid_t tid = 0;
444  for (int i = 1; i < argc; i++) {
445    if (!strcmp(argv[i], "-b")) {
446      dump_backtrace = true;
447    } else if (!have_tid) {
448      tid = atoi(argv[i]);
449      have_tid = true;
450    } else {
451      usage();
452      return 1;
453    }
454  }
455  if (!have_tid) {
456    usage();
457    return 1;
458  }
459  return do_explicit_dump(tid, dump_backtrace);
460}
461