Terminate replayer processes when parent process (unexpectedly) dies.

cli/fossilize_replay_linux.hpp

@@ -525,7 +525,7 @@ static int run_master_process(const VulkanDevice::Options &opts,
 
 	// Create an epoll instance and add the signal fd to it.
 	// The signalfd will signal when SIGCHLD is pending.
-	Global::epoll_fd = epoll_create(2 * int(processes) + 1);
+	Global::epoll_fd = epoll_create(2 * int(processes) + 2);
 	if (Global::epoll_fd < 0)
 	{
 		LOGE("Failed to create epollfd. Too old Linux kernel?\n");
@@ -543,6 +543,25 @@ static int run_master_process(const VulkanDevice::Options &opts,
 		}
 	}
 
+	// If the master process is the process group master, then STDIN is used as a sentinel.
+	// If it closes, treat this as the parent process having terminated without
+	// cleanly shutting down the Fossilize process and we should take down this process group.
+	if (getpgrp() == getpid())
+	{
+		LOGI("Using STDIN as sentinel FD.\n");
+
+		epoll_event event = {};
+		event.events = 0;
+		event.data.u32 = UINT32_MAX - 1;
+		if (epoll_ctl(Global::epoll_fd, EPOLL_CTL_ADD, STDIN_FILENO, &event) < 0)
+		{
+			LOGE("Failed to add STDIN_FILENO to epoll.\n");
+			return EXIT_FAILURE;
+		}
+	}
+	else
+		LOGI("Not using STDIN as sentinel FD.\n");
+
 	// fork() and pipe() strategy.
 	for (unsigned i = 0; i < processes; i++)
 	{
@@ -577,9 +596,9 @@ static int run_master_process(const VulkanDevice::Options &opts,
 		for (int i = 0; i < ret; i++)
 		{
 			auto &e = events[i];
-			if (e.events & (EPOLLIN | EPOLLRDHUP))
+			if (e.events & (EPOLLIN | EPOLLHUP | EPOLLRDHUP))
 			{
-				if (e.data.u32 != UINT32_MAX)
+				if (e.data.u32 < UINT32_MAX - 1)
 				{
 					auto &proc = child_processes[e.data.u32 & 0x7fffffffu];
 
@@ -604,7 +623,7 @@ static int run_master_process(const VulkanDevice::Options &opts,
 						}
 					}
 				}
-				else
+				else if (e.data.u32 == UINT32_MAX)
 				{
 					// Read from signalfd to clear the pending flag.
 					signalfd_siginfo info = {};
@@ -643,6 +662,16 @@ static int run_master_process(const VulkanDevice::Options &opts,
 						}
 					}
 				}
+				else if (e.data.u32 == UINT32_MAX - 1)
+				{
+					LOGI("Parent process died, terminating early ...\n");
+					// STDIN in parent process was hung, kill every child process and ourselves and then exit.
+					if (killpg(getpid(), SIGKILL) < 0)
+					{
+						LOGE("Failed to kill process group ... This should not happen.\n");
+						abort();
+					}
+				}
 			}
 			else if (e.events & EPOLLERR)
 			{

fossilize_external_replayer_linux.hpp

@@ -59,6 +59,7 @@ struct ExternalReplayer::Impl
 	pid_t pid = -1;
 	int fd = -1;
 	int kill_fd = -1;
+	int close_fd = -1;
 	SharedControlBlock *shm_block = nullptr;
 	size_t shm_block_size = 0;
 	int wstatus = 0;
@@ -95,6 +96,8 @@ ExternalReplayer::Impl::~Impl()
 		close(fd);
 	if (kill_fd >= 0)
 		close(kill_fd);
+	if (close_fd >= 0)
+		close(close_fd);
 
 	if (shm_block)
 		munmap(shm_block, shm_block_size);
@@ -111,6 +114,9 @@ void ExternalReplayer::Impl::reset_pid()
 	if (kill_fd >= 0)
 		close(kill_fd);
 	kill_fd = -1;
+	if (close_fd >= 0)
+		close(close_fd);
+	close_fd = -1;
 }
 
 ExternalReplayer::PollResult ExternalReplayer::Impl::poll_progress(ExternalReplayer::Progress &progress)
@@ -601,6 +607,10 @@ bool ExternalReplayer::Impl::start(const ExternalReplayer::Options &options)
 	if (pipe(fds) < 0)
 		return false;
 
+	int close_fds[2] = { -1, -1 };
+	if (!options.inherit_process_group && pipe(close_fds) < 0)
+		return false;
+
 	pid_t new_pid = fork();
 	if (new_pid > 0)
 	{
@@ -609,18 +619,34 @@ bool ExternalReplayer::Impl::start(const ExternalReplayer::Options &options)
 		fd = -1;
 		pid = new_pid;
 		kill_fd = fds[0];
+
+		if (!options.inherit_process_group)
+		{
+			close(close_fds[0]);
+			close_fd = close_fds[1];
+		}
 	}
 	else if (new_pid == 0)
 	{
 		close(fds[0]);
 		if (!options.inherit_process_group)
 		{
+			close(close_fds[1]);
+
 			// Set the process group ID so we can kill all the child processes as needed.
 			if (setpgid(0, 0) < 0)
 			{
 				LOGE("Failed to set PGID in child.\n");
 				exit(1);
 			}
+
+			if (dup2(close_fds[0], STDIN_FILENO) < 0)
+			{
+				LOGE("Failed to dup FD to stdin in child process.\n");
+				exit(1);
+			}
+
+			close(close_fds[0]);
 		}
 
 		// Notify parent process that it can return.

fossilize_external_replayer_windows.hpp

@@ -575,24 +575,21 @@ bool ExternalReplayer::Impl::start(const ExternalReplayer::Options &options)
 		si.hStdError = GetStdHandle(STD_ERROR_HANDLE);
 	}
 
-	if (options.inherit_process_group)
+	job_handle = CreateJobObjectA(nullptr, nullptr);
+	if (!job_handle)
 	{
-		job_handle = CreateJobObjectA(nullptr, nullptr);
-		if (!job_handle)
-		{
-			LOGE("Failed to create job handle.\n");
-			// Not fatal, we just won't bother with this.
-		}
-		else
+		LOGE("Failed to create job handle.\n");
+		// Not fatal, we just won't bother with this.
+	}
+	else
+	{
+		// Kill all child processes if the parent dies.
+		JOBOBJECT_EXTENDED_LIMIT_INFORMATION jeli = {};
+		jeli.BasicLimitInformation.LimitFlags = JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE;
+		if (!SetInformationJobObject(job_handle, JobObjectExtendedLimitInformation, &jeli, sizeof(jeli)))
 		{
-			// Kill all child processes if the parent dies.
-			JOBOBJECT_EXTENDED_LIMIT_INFORMATION jeli = {};
-			jeli.BasicLimitInformation.LimitFlags = JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE;
-			if (!SetInformationJobObject(job_handle, JobObjectExtendedLimitInformation, &jeli, sizeof(jeli)))
-			{
-				LOGE("Failed to set information for job object.\n");
-				// Again, not fatal.
-			}
+			LOGE("Failed to set information for job object.\n");
+			// Again, not fatal.
 		}
 	}