2.6.22-rc4: wait4() waiting for SIGSTOP may not return if the target process is ptraced
Hi,
I found a bug on signal code. wait4() after sending SIGSTOP may not return if the target process is ptraced. It's caused by signal redirection mechainism. Because of the bug, sometimes strace can't detach target process correctly.
I confirmed the following kernel have this bug.
- i386 2.6.22-rc3
- i386 2.6.22-rc4
- ia64 2.6.22-rc3
- ia64 2.6.22-rc4
Assuming that there are following processes:
- P1: a multi thread process. It consists of thread T1 and thread T2.
- strace: attaching P1 with -f option
When you try to kill strace by signal, strace does the following for each thread of P1. I regard T1 as the target thread in the following process.
1. call ptrace(PTRACE_DETACH) to detach T1. It only succeed if the T1 already traced or stopped. If succeed, no more work is needed for the thread.
2. Send SIGSTOP to the T1.
3. call wait4() to wait for the target thread's status change.
4. call ptrace(PTRACE_DETACH) again. It succeed because T1 is stopped.
If the target thread is TASK_RUNNING on process 1, and is TASK_TRACED on process 2, T2 receives SIGSTOP instead of T1. Then T2 try to wake up strace on receiving SIGSTOP, but strace is waiting for T1, so strace can't wake up eternally. In addition, traced process can't restart after that.
I have some ideas to fix this problem. How should I do?
1. Fix __group_complete_signal()
Don't redirect SIGSTOP to other thread.
2. Fix do_notify_parent_cldstop()
use the pid of original target thread for __wake_up_parent
3. Fix wait4()
Wait for tgid instead of pid.
4. Bypass the problem (or other porcesses using above rendezvous)
Change user land program such as strace to bypass this problem.
5. others
Does anyone know the better solution?
I attached the recreate program as usual.
Thanks,
Satoru
-------------------------------------------------------------------------------
/*
* recreate-ptrace-mt-wait-for-sigstop - recreate a signal bug
*
* --------------------------------------------------------------------------- *
* Problem
* =======
*
* wait4() after sending SIGSTOP may not return if the target process is
* ptraced.
*
*
* Assuming that there are following processes:
*
* - P1: a multi thread process. It consists of thread T1 and thread T2.
* - strace: attaching P1 with -f option
*
* When you try to kill strace by signal, strace does the following for each
* thread of P1. I regard T1 as the target thread in the following process. *
* 1. call ptrace(PTRACE_DETACH) to detach T1. It only succeed if the T1
* already traced or stopped. If succeed, no more work is needed for
* the thread.
* 2. Send SIGSTOP to the T1. * 3. call wait4() to wait for the target thread's status change. * 4. call ptrace(PTRACE_DETACH) again. It succeed because T1 is stopped.
*
* If the target thread is TASK_RUNNING on process 1, and is TASK_TRACED on
* process 2, T2 recives SIGSTOP instead of T1. Then T2 try to wake up strace
* on receiving SIGSTOP, but strace is waiting for T1, so strace can't wake
* up eternally. In addition, traced process can't restart after that. *
* How to recreate
* ===============
*
* 1. run this program
*
* $ recreate-ptrace-mt-wait-for-sigstop &
*
* 2. run strace and attach this program
*
* $ strace -f -p $!
*
* 3. C-c on your terminal
*
* Expected Result
* ===============
*
* All threads of this program was detached safely
*
* Actual Result
* =============
*
* strace sleeps on wait4() eternally.
*
* Note
* ====
*
* This program can't always recreate a problem. However recreate
* possibility is very high.
*
*----------------------------------------------------------------------
*
* Copyright 2007 Satoru Takeuchi <[email protected]> *
* This software may be used and distributed according to the terms
* of the GNU General Public License, incorporated herein by reference. *
*/
#include <sys/types.h>
#include <unistd.h>
#include <pthread.h>
#include <stdio.h>
#include <stdlib.h>
#include <err.h>
void *thread_fn(void *arg)
{
int i;
for (;;)
sleep(1);
}#define NTHREAD 64
int main(int argc, char **argv)
{
pthread_t t[NTHREAD];
int i;
for (i = 0; i < NTHREAD; i++)
if (pthread_create(&t[i], NULL, thread_fn, NULL)) {
fprintf(stderr, "pthread_create(%d) failed\n", i);
exit(EXIT_FAILURE);
}
thread_fn(NULL);
for (i = 0; i < NTHREAD; i++)
if (pthread_join(t[i], NULL))
warn("pthread_join(%d) failed", i);
exit(EXIT_SUCCESS);
}The bug is in strace (and is being fixed). It needs to use tgkill instead of kill to send that SIGSTOP to the right thread only.
Thanks,
Roland
Hi,
At Thu, 7 Jun 2007 02:54:32 -0700 (PDT),
Quoted 3 lines · Roland McGrath
The bug is in strace (and is being fixed). It needs to use tgkill instead of kill to send that SIGSTOP to the right thread only.
You mean kill() to specifig thread is fundamentally wrong? If so, current kill() is wrong because it can receive pid of non main thread.
Thanks,
Satoru
Quoted 5 lines
Quoted 2 lines
The bug is in strace (and is being fixed). It needs to use tgkill instead of kill to send that SIGSTOP to the right thread only.
You mean
kill()to specifig thread is fundamentally wrong? If so, currentkill()is wrong because it can receive pid of non main thread.
The kill call should only be used on process IDs (tgid in linuspeak). This is what POSIX defines it for. I'm not sure off hand if it's technically a violation of POSIX that it works for numbers that aren't a POSIX process ID (aka tgid) but happen happen to be a Linux thread ID (aka pid).
Thanks,
Roland