1 // Copyright (c) 2013 The Chromium Authors. All rights reserved. 2 // Use of this source code is governed by a BSD-style license that can be 3 // found in the LICENSE file. 4 5 #include "base/process/kill.h" 6 7 #include <signal.h> 8 #include <sys/event.h> 9 #include <sys/types.h> 10 #include <sys/wait.h> 11 12 #include "base/file_util.h" 13 #include "base/logging.h" 14 #include "base/posix/eintr_wrapper.h" 15 16 namespace base { 17 18 namespace { 19 20 const int kWaitBeforeKillSeconds = 2; 21 22 // Reap |child| process. This call blocks until completion. 23 void BlockingReap(pid_t child) { 24 const pid_t result = HANDLE_EINTR(waitpid(child, NULL, 0)); 25 if (result == -1) { 26 DPLOG(ERROR) << "waitpid(" << child << ", NULL, 0)"; 27 } 28 } 29 30 // Waits for |timeout| seconds for the given |child| to exit and reap it. If 31 // the child doesn't exit within the time specified, kills it. 32 // 33 // This function takes two approaches: first, it tries to use kqueue to 34 // observe when the process exits. kevent can monitor a kqueue with a 35 // timeout, so this method is preferred to wait for a specified period of 36 // time. Once the kqueue indicates the process has exited, waitpid will reap 37 // the exited child. If the kqueue doesn't provide an exit event notification, 38 // before the timeout expires, or if the kqueue fails or misbehaves, the 39 // process will be mercilessly killed and reaped. 40 // 41 // A child process passed to this function may be in one of several states: 42 // running, terminated and not yet reaped, and (apparently, and unfortunately) 43 // terminated and already reaped. Normally, a process will at least have been 44 // asked to exit before this function is called, but this is not required. 45 // If a process is terminating and unreaped, there may be a window between the 46 // time that kqueue will no longer recognize it and when it becomes an actual 47 // zombie that a non-blocking (WNOHANG) waitpid can reap. This condition is 48 // detected when kqueue indicates that the process is not running and a 49 // non-blocking waitpid fails to reap the process but indicates that it is 50 // still running. In this event, a blocking attempt to reap the process 51 // collects the known-dying child, preventing zombies from congregating. 52 // 53 // In the event that the kqueue misbehaves entirely, as it might under a 54 // EMFILE condition ("too many open files", or out of file descriptors), this 55 // function will forcibly kill and reap the child without delay. This 56 // eliminates another potential zombie vector. (If you're out of file 57 // descriptors, you're probably deep into something else, but that doesn't 58 // mean that zombies be allowed to kick you while you're down.) 59 // 60 // The fact that this function seemingly can be called to wait on a child 61 // that's not only already terminated but already reaped is a bit of a 62 // problem: a reaped child's pid can be reclaimed and may refer to a distinct 63 // process in that case. The fact that this function can seemingly be called 64 // to wait on a process that's not even a child is also a problem: kqueue will 65 // work in that case, but waitpid won't, and killing a non-child might not be 66 // the best approach. 67 void WaitForChildToDie(pid_t child, int timeout) { 68 DCHECK(child > 0); 69 DCHECK(timeout > 0); 70 71 // DON'T ADD ANY EARLY RETURNS TO THIS FUNCTION without ensuring that 72 // |child| has been reaped. Specifically, even if a kqueue, kevent, or other 73 // call fails, this function should fall back to the last resort of trying 74 // to kill and reap the process. Not observing this rule will resurrect 75 // zombies. 76 77 int result; 78 79 int kq = HANDLE_EINTR(kqueue()); 80 if (kq == -1) { 81 DPLOG(ERROR) << "kqueue()"; 82 } else { 83 file_util::ScopedFD auto_close_kq(&kq); 84 85 struct kevent change = {0}; 86 EV_SET(&change, child, EVFILT_PROC, EV_ADD, NOTE_EXIT, 0, NULL); 87 result = HANDLE_EINTR(kevent(kq, &change, 1, NULL, 0, NULL)); 88 89 if (result == -1) { 90 if (errno != ESRCH) { 91 DPLOG(ERROR) << "kevent (setup " << child << ")"; 92 } else { 93 // At this point, one of the following has occurred: 94 // 1. The process has died but has not yet been reaped. 95 // 2. The process has died and has already been reaped. 96 // 3. The process is in the process of dying. It's no longer 97 // kqueueable, but it may not be waitable yet either. Mark calls 98 // this case the "zombie death race". 99 100 result = HANDLE_EINTR(waitpid(child, NULL, WNOHANG)); 101 102 if (result != 0) { 103 // A positive result indicates case 1. waitpid succeeded and reaped 104 // the child. A result of -1 indicates case 2. The child has already 105 // been reaped. In both of these cases, no further action is 106 // necessary. 107 return; 108 } 109 110 // |result| is 0, indicating case 3. The process will be waitable in 111 // short order. Fall back out of the kqueue code to kill it (for good 112 // measure) and reap it. 113 } 114 } else { 115 // Keep track of the elapsed time to be able to restart kevent if it's 116 // interrupted. 117 TimeDelta remaining_delta = TimeDelta::FromSeconds(timeout); 118 TimeTicks deadline = TimeTicks::Now() + remaining_delta; 119 result = -1; 120 struct kevent event = {0}; 121 while (remaining_delta.InMilliseconds() > 0) { 122 const struct timespec remaining_timespec = remaining_delta.ToTimeSpec(); 123 result = kevent(kq, NULL, 0, &event, 1, &remaining_timespec); 124 if (result == -1 && errno == EINTR) { 125 remaining_delta = deadline - TimeTicks::Now(); 126 result = 0; 127 } else { 128 break; 129 } 130 } 131 132 if (result == -1) { 133 DPLOG(ERROR) << "kevent (wait " << child << ")"; 134 } else if (result > 1) { 135 DLOG(ERROR) << "kevent (wait " << child << "): unexpected result " 136 << result; 137 } else if (result == 1) { 138 if ((event.fflags & NOTE_EXIT) && 139 (event.ident == static_cast<uintptr_t>(child))) { 140 // The process is dead or dying. This won't block for long, if at 141 // all. 142 BlockingReap(child); 143 return; 144 } else { 145 DLOG(ERROR) << "kevent (wait " << child 146 << "): unexpected event: fflags=" << event.fflags 147 << ", ident=" << event.ident; 148 } 149 } 150 } 151 } 152 153 // The child is still alive, or is very freshly dead. Be sure by sending it 154 // a signal. This is safe even if it's freshly dead, because it will be a 155 // zombie (or on the way to zombiedom) and kill will return 0 even if the 156 // signal is not delivered to a live process. 157 result = kill(child, SIGKILL); 158 if (result == -1) { 159 DPLOG(ERROR) << "kill(" << child << ", SIGKILL)"; 160 } else { 161 // The child is definitely on the way out now. BlockingReap won't need to 162 // wait for long, if at all. 163 BlockingReap(child); 164 } 165 } 166 167 } // namespace 168 169 void EnsureProcessTerminated(ProcessHandle process) { 170 WaitForChildToDie(process, kWaitBeforeKillSeconds); 171 } 172 173 } // namespace base 174
