Oleg,
first of all thank you for a good review, please see my comments below.
On Wed, Jul 13, 2011 at 6:09 PM, Oleg Nesterov [off-list ref] wrote:
On 07/12, Vladimir Zapolskiy wrote:
quoted
Note, a detach signal is not emitted, if a tracer process terminates
without explicit PTRACE_DETACH request. Such cases can be covered
listening to PROC_EVENT_EXIT connector events.
Hmm. More and more reasons to make the implicit detach sleepable...
But. There is another case. The (dead) tracee can be detached via
do_wait(). Perhaps this falls into "covered listening to EXIT" too,
but imho makes sense to document in the changelog. Oh, and probably
we will add the ability to detach a zombie...
I don't really understand why do you need this, but I won't argue.
I found that implicit ptrace detach codepath is quite mutable and
vast, and I don't want to interfere in that changes without knowing
even basic pitfalls. Somehow the sending a connector signal on
explicit detach is quite sufficient at least for the most of the proc
connector usecases I can imagine, because hopefully almost(?) all
implicit detach cases are related to tracer or tracee thread
completion, and that is supposed to be reported to userspace via
do_exit()/proc_exit_connector() path.
As for the patch,
quoted
+void proc_ptrace_connector(struct task_struct *task)
+{
+ struct cn_msg *msg;
+ struct proc_event *ev;
+ struct timespec ts;
+ __u8 buffer[CN_PROC_MSG_SIZE];
+ struct task_struct *tracer;
+
+ if (atomic_read(&proc_event_num_listeners) < 1)
+ return;
+
+ msg = (struct cn_msg *)buffer;
+ ev = (struct proc_event *)msg->data;
+ get_seq(&msg->seq, &ev->cpu);
+ ktime_get_ts(&ts); /* get high res monotonic timestamp */
+ put_unaligned(timespec_to_ns(&ts), (__u64 *)&ev->timestamp_ns);
+ ev->what = PROC_EVENT_PTRACE;
+ ev->event_data.ptrace.process_pid = task->pid;
+ ev->event_data.ptrace.process_tgid = task->tgid;
+
+ rcu_read_lock();
+ tracer = tracehook_tracer_task(task);
+ if (tracer) {
+ ev->event_data.ptrace.tracer_pid = tracer->pid;
+ ev->event_data.ptrace.tracer_tgid = tracer->tgid;
+ } else {
+ ev->event_data.ptrace.tracer_pid = 0;
+ ev->event_data.ptrace.tracer_tgid = 0;
+ }
This doesn't look right. The code uses tracehook_tracer_task() to
figure out whether this task traced or not. But this is racy.
ptrace_attach:
...attach...
/* WINDOW */
proc_ptrace_connector(task);
The task can exit in between, and the caller's subthread can do
wait4() and release it. In this case proc_ptrace_connector() will
see tracehook_tracer_task() == NULL and report "detach".
The similar race in ptrace_detach() path. Another tracer can attach
to this task before we proc_ptrace_connector().
I think proc_ptrace_connector() needs the explicit "task_struct *tracer"
argument, NULL if ptrace_detach(). Or a simple boolean, the tracer is
current.
If you think this is fine - I won't argue.
Fixed in the second version of the change, thanks.
But in any case, please rediff against
git://git.kernel.org/pub/scm/linux/kernel/git/oleg/misc.git ptrace
tracehook_tracer_task() was removed, and
quoted
if (wait_trap)
wait_event(current->signal->wait_chldexit,
!(task->group_stop & GROUP_STOP_TRAPPING));
+ if (!retval)
+ proc_ptrace_connector(task);
+
return retval;
}
this chunk probably should be updated.
Rebased, thanks a lot.
Vladimir