Skip to contentSkip to conversation

Allow signaling a process by all its thread ids?

DISCUSSION Simon Holm Thøgersen 3 participants 4 messages May 21, 2009

lore
Share & tools
lore
#0 lore raw

There is a bug report at https://bugs.launchpad.net/bugs/341239 where the question is asked. Should it be possible to signal a process with kill(2) by passing any of the thread ids belong to the process as argument to kill?

Current behaviour allows that but the reporter argues that it shouldn't be allowed by the following argument:

"I'm no expert in interpreting POSIX. A casual (but careful) reading of POSIX 1003.1-2004 suggests that this isn't the intended behavior. It is implied in a few places, most notably the getpid[1] documentation, that a process has only one PID (by use of the definite article when referring to it - ie, "the process ID"). The kill function[2] operates on "a process or a group of processes". It sends a signal "to the process whose process ID is equal to pid". Combined with the previous point, this suggests there should be only one process ID which can be passed to kill to send a signal to a particular process.

It may be that POSIX is weakly worded enough that either allowing or disallowing this behavior is valid. I think that allowing it is a violation of POSIX, but I can find no single, explicit, direct statement in POSIX to back this up.

There is at least one other argument to disallow this behavior even if it is not technically illegal. Consider the traditional, widespread convention of pid files. This behavior drastically reduces their utility, as it greatly increases the chance of PID collisions, a case in which it is difficult to automatically (that is, without human intervention) recognize that a pid file no longer corresponds to a running process which created it. This is the case in which I encountered the behavior - when it repeatedly caused services not to restart because they encountered a thread of another process and misinterpreted it to mean the service was already running.

I hope this is convincing. If not, what kind of argument would be?

[1]: http://www.opengroup.org/onlinepubs/009695399/functions/getpid.html
[2]: http://www.opengroup.org/onlinepubs/009695399/functions/kill.html"

Simon Holm Thøgersen

#1 lore raw
Quoted 4 lines

It may be that POSIX is weakly worded enough that either allowing or disallowing this behavior is valid. I think that allowing it is a violation of POSIX, but I can find no single, explicit, direct statement in POSIX to back this up.

If there was a specific explicit statement it might be an argument to add a separate poxix_kill(), but the existing kill(2) interface the kernel has is API, its been that way forever and you need *very* strong reasons to make such changes that might break things.

You need to show that
- The current API breaks stuff
- The current API is absolutely invalid in posix
- Changing it improves the functionality and power of the kernel
- Changing it doesn't break existing applications

Quoted 10 lines

There is at least one other argument to disallow this behavior even if it is not technically illegal. Consider the traditional, widespread convention of pid files. This behavior drastically reduces their utility, as it greatly increases the chance of PID collisions, a case in which it is difficult to automatically (that is, without human intervention) recognize that a pid file no longer corresponds to a running process which created it. This is the case in which I encountered the behavior - when it repeatedly caused services not to restart because they encountered a thread of another process and misinterpreted it to mean the service was already running.

Either your pid files are secure, or they are not. If collisions can occur by accident I can cause them deliberately. If I can cause them deliberately your system is insecure and you need to fix it.

#2 lore raw
Quoted 4 lines · Alan Cox

You need to show that
- The current API breaks stuff
- The current API is absolutely invalid in posix
- Changing it improves the functionality and power of the kernel

I think an argument could be made for these. A task is not necessarily a process, and kill() is only defined to send a signal to one or more processes. This introduces potential problems if a) it is possible to send a signal to an individual tid, and b) a tid can be reused in a different process.

We already have the tgkill() syscall, which fixes the ambiguity by specifying both the pid and tid. This is what pthread_kill() uses under the hood. The fact that it was seen as necessary (back in 2.5) indicates that there are problems with kill() as currently implemented.

Quoted 1 line

- Changing it doesn't break existing applications

This last one is the kicker. As I mentioned in my other reply, I suspect that making such a change would break a lot of (not-quite-POSIX) applications that assume they can send a signal to particular threads using kill().

I see analogies to the whole fsync() issue.

Chris

#3 lore raw
In reply to Simon Holm Thøgersen
Quoted 4 lines · Simon Holm Thøgersen

There is a bug report at https://bugs.launchpad.net/bugs/341239 where the question is asked. Should it be possible to signal a process with kill(2) by passing any of the thread ids belong to the process as argument to kill?

Based on POSIX, from the perspective of the kernel I think that a tid that is not a tgid should result in ESRCH for the kill() syscall.

Such a tid would however be valid for the tgkill or tkill syscalls.

Makes sense to me. I suspect such a change would break a lot of not-strictly-compliant programs though.

In the meantime, the pid file scenario could be modified to store both the tid and pid (using userspace terminology now) and use the tgkill() syscall. It should be noted that tgkill() is linux-specific.

Chris