sysdeps/linux-gnu/events.c

#include "config.h"

#define	_GNU_SOURCE	1
#include <stdlib.h>
#include <sys/types.h>
#include <sys/wait.h>
#include <errno.h>
#include <signal.h>
#include <string.h>
#include <sys/ptrace.h>
#include <assert.h>
#include <unistd.h>

#include "common.h"
#include "breakpoint.h"

static Event event;

/* A queue of events that we missed while enabling the
 * breakpoint in one of tasks.  */
static Event * delayed_events = NULL;
static Event * end_delayed_events = NULL;

static enum pcb_status
first (Process * proc, void * data)
{
	return pcb_stop;
}

void
enque_event(Event * event)
{
	debug(DEBUG_FUNCTION, "%d: queuing event %d for later",
	      event->proc->pid, event->type);
	Event * ne = malloc(sizeof(*ne));
	if (ne == NULL) {
		perror("event will be missed: malloc");
		return;
	}

	*ne = *event;
	ne->next = NULL;
	if (end_delayed_events == NULL) {
		assert(delayed_events == NULL);
		end_delayed_events = delayed_events = ne;
	}
	else {
		assert(delayed_events != NULL);
		end_delayed_events = end_delayed_events->next = ne;
	}
}

Event *
each_qd_event(enum ecb_status (*pred)(Event *, void *), void * data)
{
	Event * prev = delayed_events;
	Event * event;
	for (event = prev; event != NULL; ) {
		switch ((*pred)(event, data)) {
		case ecb_cont:
			prev = event;
			event = event->next;
			continue;

		case ecb_deque:
			debug(DEBUG_FUNCTION, "dequeuing event %d for %d",
			      event->type,
			      event->proc != NULL ? event->proc->pid : -1);
			/*
			printf("dequeuing event %d for %d\n", event->type,
			       event->proc != NULL ? event->proc->pid : -1) ;
			*/
			if (end_delayed_events == event)
				end_delayed_events = prev;
			if (delayed_events == event)
				delayed_events = event->next;
			else
				prev->next = event->next;
			if (delayed_events == NULL)
				end_delayed_events = NULL;
			/* fall-through */

		case ecb_yield:
			return event;
		}
	}

	return NULL;
}

static enum ecb_status
event_process_not_reenabling(Event * event, void * data)
{
	if (event->proc == NULL
	    || event->proc->leader == NULL
	    || event->proc->leader->event_handler == NULL)
		return ecb_deque;
	else
		return ecb_cont;
}

static Event *
next_qd_event(void)
{
	return each_qd_event(&event_process_not_reenabling, NULL);
}

int linux_in_waitpid = 0;

Event *
next_event(void)
{
	pid_t pid;
	int status;
	int tmp;
	int stop_signal;

	debug(DEBUG_FUNCTION, "next_event()");
	Event * ev;
	if ((ev = next_qd_event()) != NULL) {
		event = *ev;
		free(ev);
		return &event;
	}

	if (!each_process(NULL, &first, NULL)) {
		debug(DEBUG_EVENT, "event: No more traced programs: exiting");
		exit(0);
	}

	linux_in_waitpid = 1;
	pid = waitpid(-1, &status, __WALL);
	linux_in_waitpid = 0;

	if (pid == -1) {
		if (errno == ECHILD) {
			debug(DEBUG_EVENT, "event: No more traced programs: exiting");
			exit(0);
		} else if (errno == EINTR) {
			debug(DEBUG_EVENT, "event: none (wait received EINTR?)");
			event.type = EVENT_NONE;
			return &event;
		}
		perror("wait");
		exit(1);
	}
	event.proc = pid2proc(pid);
	if (!event.proc || event.proc->state == STATE_BEING_CREATED) {
		/* Work around (presumably) a bug on some kernels,
		 * where we are seeing a waitpid event even though the
		 * process is still reported to be running.  Wait for
		 * the tracing stop to propagate.  But don't get stuck
		 * here forever.
		 *
		 * We need the process in T, because there's a lot of
		 * ptracing going on all over the place, and these
		 * calls fail when the process is not in T.
		 *
		 * N.B. This was observed on RHEL 5 Itanium, but I'm
		 * turning this on globally, to save some poor soul
		 * down the road (which could well be me a year from
		 * now) the pain of figuring this out all over again.
		 * Petr Machata 2011-11-22.  */
		int i = 0;
		for (; i < 100 && process_status(pid) != ps_tracing_stop; ++i) {
			debug(2, "waiting for %d to stop", pid);
			usleep(10000);
		}
		event.type = EVENT_NEW;
		event.e_un.newpid = pid;
		debug(DEBUG_EVENT, "event: NEW: pid=%d", pid);
		return &event;
	}
	get_arch_dep(event.proc);
	debug(3, "event from pid %u", pid);
	if (event.proc->breakpoints_enabled == -1)
		trace_set_options(event.proc, event.proc->pid);
	Process *leader = event.proc->leader;
	if (leader == event.proc) {
		if (event.proc->breakpoints_enabled == -1) {
			event.type = EVENT_NONE;
			enable_all_breakpoints(event.proc);
			continue_process(event.proc->pid);
			debug(DEBUG_EVENT,
			      "event: NONE: pid=%d (enabling breakpoints)",
			      pid);
			return &event;
		} else if (!event.proc->libdl_hooked) {
			/* debug struct may not have been written yet.. */
			if (linkmap_init(event.proc, &main_lte) == 0) {
				event.proc->libdl_hooked = 1;
			}
		}
	}

	/* The process should be stopped after the waitpid call.  But
	 * when the whole thread group is terminated, we see
	 * individual tasks spontaneously transitioning from 't' to
	 * 'R' and 'Z'.  Calls to ptrace fail and /proc/pid/status may
	 * not even be available anymore, so we can't check in
	 * advance.  So we just drop the error checking around ptrace
	 * calls.  We check for termination ex post when it fails,
	 * suppress the event, and let the event loop collect the
	 * termination in the next iteration.  */
#define CHECK_PROCESS_TERMINATED					\
	do {								\
		int errno_save = errno;					\
		switch (process_stopped(pid))				\
		case 0:							\
		case -1: {						\
			debug(DEBUG_EVENT,				\
			      "process not stopped, is it terminating?"); \
			event.type = EVENT_NONE;			\
			continue_process(event.proc->pid);		\
			return &event;					\
		}							\
		errno = errno_save;					\
	} while (0)

	event.proc->instruction_pointer = (void *)(uintptr_t)-1;

	/* Check for task termination now, before we have a need to
	 * call CHECK_PROCESS_TERMINATED later.  That would suppress
	 * the event that we are processing.  */
	if (WIFSIGNALED(status)) {
		event.type = EVENT_EXIT_SIGNAL;
		event.e_un.signum = WTERMSIG(status);
		debug(DEBUG_EVENT, "event: EXIT_SIGNAL: pid=%d, signum=%d", pid, event.e_un.signum);
		return &event;
	}
	if (WIFEXITED(status)) {
		event.type = EVENT_EXIT;
		event.e_un.ret_val = WEXITSTATUS(status);
		debug(DEBUG_EVENT, "event: EXIT: pid=%d, status=%d", pid, event.e_un.ret_val);
		return &event;
	}

	event.proc->instruction_pointer = get_instruction_pointer(event.proc);
	if (event.proc->instruction_pointer == (void *)(uintptr_t)-1) {
		CHECK_PROCESS_TERMINATED;
		if (errno != 0)
			perror("get_instruction_pointer");
	}

	switch (syscall_p(event.proc, status, &tmp)) {
		case 1:
			event.type = EVENT_SYSCALL;
			event.e_un.sysnum = tmp;
			debug(DEBUG_EVENT, "event: SYSCALL: pid=%d, sysnum=%d", pid, tmp);
			return &event;
		case 2:
			event.type = EVENT_SYSRET;
			event.e_un.sysnum = tmp;
			debug(DEBUG_EVENT, "event: SYSRET: pid=%d, sysnum=%d", pid, tmp);
			return &event;
		case 3:
			event.type = EVENT_ARCH_SYSCALL;
			event.e_un.sysnum = tmp;
			debug(DEBUG_EVENT, "event: ARCH_SYSCALL: pid=%d, sysnum=%d", pid, tmp);
			return &event;
		case 4:
			event.type = EVENT_ARCH_SYSRET;
			event.e_un.sysnum = tmp;
			debug(DEBUG_EVENT, "event: ARCH_SYSRET: pid=%d, sysnum=%d", pid, tmp);
			return &event;
		case -1:
			CHECK_PROCESS_TERMINATED;
			if (errno != 0)
				perror("syscall_p");
	}
	if (WIFSTOPPED(status)) {
		int what = status >> 16;
		if (what == PTRACE_EVENT_VFORK
		    || what == PTRACE_EVENT_FORK
		    || what == PTRACE_EVENT_CLONE) {
			unsigned long data;
			event.type = what == PTRACE_EVENT_VFORK
				? EVENT_VFORK : EVENT_CLONE;
			ptrace(PTRACE_GETEVENTMSG, pid, NULL, &data);
			event.e_un.newpid = data;
			debug(DEBUG_EVENT, "event: CLONE: pid=%d, newpid=%d",
			      pid, (int)data);
			return &event;
		}
	}
	if (WIFSTOPPED(status) && (status>>16 == PTRACE_EVENT_EXEC)) {
		event.type = EVENT_EXEC;
		debug(DEBUG_EVENT, "event: EXEC: pid=%d", pid);
		return &event;
	}
	if (!WIFSTOPPED(status)) {
		/* should never happen */
		event.type = EVENT_NONE;
		debug(DEBUG_EVENT, "event: NONE: pid=%d (wait error?)", pid);
		return &event;
	}

	stop_signal = WSTOPSIG(status);

	/* On some targets, breakpoints are signalled not using
	   SIGTRAP, but also with SIGILL, SIGSEGV or SIGEMT.  SIGEMT
	   is not defined on Linux, but check for the others.

	   N.B. see comments in GDB's infrun.c for details.  I've
	   actually seen this on an Itanium machine on RHEL 5, I don't
	   remember the exact kernel version anymore.  ia64-sigill.s
	   in the test suite tests this.  Petr Machata 2011-06-08.  */
	void * break_address
		= event.proc->instruction_pointer - DECR_PC_AFTER_BREAK;
	if ((stop_signal == SIGSEGV || stop_signal == SIGILL)
	    && leader != NULL
	    && address2bpstruct(leader, break_address))
			stop_signal = SIGTRAP;

	if (stop_signal != (SIGTRAP | event.proc->tracesysgood)
			&& stop_signal != SIGTRAP) {
		event.type = EVENT_SIGNAL;
		event.e_un.signum = stop_signal;
		debug(DEBUG_EVENT, "event: SIGNAL: pid=%d, signum=%d", pid, stop_signal);
		return &event;
	}

	/* last case [by exhaustion] */
	event.type = EVENT_BREAKPOINT;

	event.e_un.brk_addr = break_address;
	debug(DEBUG_EVENT, "event: BREAKPOINT: pid=%d, addr=%p", pid, event.e_un.brk_addr);

	return &event;
}