[GDBServer] Send SIGINT using process group ID
Commit Message
Hi,
linux_request_interrupt is supposed to send SIGINT to the process group,
but it passes the process ID to kill() instead of the process group ID,
which may not be the same as the process ID. The patch calls getpgid
first to get the process group ID.
Patch tested on arm-linux.
gdb/gdbserver:
2016-05-11 Jim Chen <nchen@mozilla.com>
* linux-low.c (linux_request_interrupt): Use process group ID for
sending SIGINT
---
gdb/gdbserver/linux-low.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
Comments
On 05/11/2016 10:51 PM, Jim Chen wrote:
> Hi,
>
> linux_request_interrupt is supposed to send SIGINT to the process group,
> but it passes the process ID to kill() instead of the process group ID,
> which may not be the same as the process ID.
> The patch calls getpgid
> first to get the process group ID.
>
> Patch tested on arm-linux.
Can you expand on the use case you see this happening on, please?
I can imagine some, but I'd like to hear it from you.
Some have expressed desire to _not_ send the SIGINT to the whole
process group, which may make sense when you're attached to a
process rather than having started it. IIRC, there's a bug filed in
bugzilla about this.
Looking at the code, not-sending-to-process-group-when-attached
is what native GNU/Linux does too (inflow.c:pass_signal).
Seems like "c&" -> "interrupt" doesn't consider "attach" either,
as inf-ptrace.c:inf_ptrace_interrupt sends the SIGINT
to the process, and like gdbserver, assumes the inferior's
PID is the process group id.
Thanks,
Pedro Alves
The use case was attaching, like you mentioned, on Android. And I just
found PR/18945 [1], which describes the problem. I now see that my patch
is not a good solution either because on Android, the pgid is the zygote
process, and all running apps on the system belong to that group.
Seems like the best solution is to handle attaching separately? Or maybe
send SIGINT to the process first, and if that fails, send it to the
process group, in order to address the original bug of a dead main
thread when pausing?
Thanks,
Jim
[1] https://sourceware.org/bugzilla/show_bug.cgi?id=18945
On 5/18/2016 8:10 AM, Pedro Alves wrote:
> On 05/11/2016 10:51 PM, Jim Chen wrote:
>> Hi,
>>
>> linux_request_interrupt is supposed to send SIGINT to the process group,
>> but it passes the process ID to kill() instead of the process group ID,
>> which may not be the same as the process ID.
>> The patch calls getpgid
>> first to get the process group ID.
>>
>> Patch tested on arm-linux.
>
> Can you expand on the use case you see this happening on, please?
> I can imagine some, but I'd like to hear it from you.
>
> Some have expressed desire to _not_ send the SIGINT to the whole
> process group, which may make sense when you're attached to a
> process rather than having started it. IIRC, there's a bug filed in
> bugzilla about this.
>
> Looking at the code, not-sending-to-process-group-when-attached
> is what native GNU/Linux does too (inflow.c:pass_signal).
>
> Seems like "c&" -> "interrupt" doesn't consider "attach" either,
> as inf-ptrace.c:inf_ptrace_interrupt sends the SIGINT
> to the process, and like gdbserver, assumes the inferior's
> PID is the process group id.
>
> Thanks,
> Pedro Alves
>
@@ -5745,17 +5745,17 @@ linux_look_up_symbols (void)
static void
linux_request_interrupt (void)
{
extern unsigned long signal_pid;
/* Send a SIGINT to the process group. This acts just like the user
typed a ^C on the controlling terminal. */
- kill (-signal_pid, SIGINT);
+ kill (-getpgid(signal_pid), SIGINT);
}
/* Copy LEN bytes from inferior's auxiliary vector starting at OFFSET
to debugger memory starting at MYADDR. */
static int
linux_read_auxv (CORE_ADDR offset, unsigned char *myaddr, unsigned int len)
{