ptrace_detach: the wrong wakeup breaks the ERESTARTxxx logic
Oleg Nesterov
maintainerPATCH 1/3 ptrace_detach: the wrong wakeup breaks the ERESTARTxxx logic
Another ancient bug. Consider this trivial test-case,
int main(void)
{
int pid = fork();
if (pid) {
ptrace(PTRACE_ATTACH, pid, NULL, NULL);
wait(NULL);
ptrace(PTRACE_DETACH, pid, NULL, NULL);
} else {
pause();
printf("WE HAVE A KERNEL BUG!!!\n");
}
return 0;
}the child must not "escape" for sys_pause(), but it can and this was seen in practice.
This is because ptrace_detach does:
if (!child->exit_state) wake_up_process(child);
this wakeup can happen after this child has already restarted sys_pause(), because it gets another wakeup from ptrace_untrace().
It is not clear to me what was the rationale, but this is obviously wrong. Remove this wakeup. The caller saw this task in TASK_TRACED state, and unless it was SIGKILL'ed in between __ptrace_unlink()->ptrace_untrace() should handle this case correctly. If it was SIGKILL'ed, we don't need to wakup the dying tracee too.
kernel/ptrace.c renamed from kernel/ptrace.c~1_KILL_WAKE +0 -4
@@ -250,11 +250,7 @@ int ptrace_detach(struct task_struct *ch/* protect against de_thread()->release_task() */ if (child->ptrace) { child->exit_code = data; - dead =__ptrace_detach (current, child); - -if (!child->exit_state) -wake_up_process (child); }write_unlock_irq (&tasklist_lock);
Quoted 4 lines
This is because
ptrace_detachdoes:if (!child->exit_state) wake_up_process(child);
I'm pretty sure that all these uses of wake_up_process were just blindly copied from an original use in ptrace code (what's now ptrace_resume). That original use just dates from the beforetime, the long long ago. (I don't think it indicates any coherent original intent.)
It's many kinds of wrong. It's also always been wrong in case of a simultaneous SIGKILL that already woke the child, which has then blocked on some mutex or semaphore or whatnot. I don't know what the stated general policy about spurious wakeups from schedule() is supposed to be. Perhaps to be pedantic, the sys_pause() code has been wrong to return without checking signal_pending().
It's also about as wrong to use blind wake_up_process in ptrace_resume. It ought to be wake_up_state(__TASK_TRACED|__TASK_STOPPED).
Frankly, I've always been afraid of strange cruft that might unexpectedly turn out to rely on this "wrong" (unconditional) wake-up. Probably the things like that historically were all just to do with the stopped/traced bookkeeping and would be covered by explicitly dealing with PTRACE_CONT vs group stop et al. But FWIW my reaction to fiddling the wake_up_process bogons in the past has been, "Be afraid."
Thanks,
Roland
Oleg Nesterov
maintainerQuoted 11 lines · Roland McGrath
Quoted 5 lines
This is because
ptrace_detachdoes:if (!child->exit_state) wake_up_process(child);I'm pretty sure that all these uses of
wake_up_processwere just blindly copied from an original use in ptrace code (what's nowptrace_resume). That original use just dates from the beforetime, the long long ago. (I don't think it indicates any coherent original intent.)
Quoted 6 lines
It's many kinds of wrong. It's also always been wrong in case of a simultaneous SIGKILL that already woke the child, which has then blocked on some mutex or semaphore or whatnot. I don't know what the stated general policy about spurious wakeups from
schedule()is supposed to be. Perhaps to be pedantic, thesys_pause()code has been wrong to return without checkingsignal_pending().
Yes, I thought about fixing sys_pause() too, but I'm afraid we can have the similar code.
Quoted 6 lines
Frankly, I've always been afraid of strange cruft that might unexpectedly turn out to rely on this "wrong" (unconditional) wake-up. Probably the things like that historically were all just to do with the stopped/traced bookkeeping and would be covered by explicitly dealing with
PTRACE_CONTvs group stop et al. But FWIW my reaction to fiddling thewake_up_processbogons in the past has been, "Be afraid."
Yes, I am afraid, seriously ;)
This can reveal other subtle problems, of course. But there is another reason why this wakeup is wrong. It clearly breaks the SIGNAL_GROUP_STOPPED logic in ptrace_untrace().
Oleg.