Re: [PATCH] signal/ptrace: Fix the bug of ptrace attach and signal handling concurrency

[Oleg Nesterov <oleg@redhat.com>]

Hi Shuzhi,

On 05/01, Shuzhi Zu wrote:
>
> Example 1：
>     A                             B
>     get_signal
>     dequeue_signal (SIGCHLD）
>                                   ptrace_attach ( A->ptrace |= PT_PTRACED)
>
>     ptrace_signal
>     ->ptrace_stop(TASK_TRACED)
>                                   ptrace_attach ( Send SIGSTOP to A)
>                                   ptrace_waitpid( return 0)
>                                   ptrace_detach (A->ptrace=0, wakeup A)
>     dequeue_signal(SIGSTOP)
>     sig_kernel_stop(SIGSTOP)
>     do_signal_stop (TASK_STOPPED)
>
> then：
> A (other threads of the process received signal)
> get_signal-> do_signal_stop(0))->TASK_STOPPED

Yes, there are a lot of known problems with send_sig_info(SIGSTOP) in
ptrace_attach(). This one of the reasons for PTRACE_SEIZE which doesn't
abuse SIGSTOP. Please use it instead of PTRACE_ATTACH, PTRACE_SEIZE has
more features.


As for your particular example, this is an application bug. Debugger (if
it uses PTRACE_ATTACH) should not detach until the tracee reports SIGSTOP
sent by PTRACE_ATTACH.

This can lead to other problems, say we can miss a "real" SIGSTOP from
another application, but again PTRACE_ATTACH is hopeless wrt SIGSTOP,
please consider PTRACE_SEIZE.

Oleg.