Re: [PATCH v2 linux-trace 1/8] tracing: attach eBPF programs to tracepoints and syscalls

On Thu, Jan 29, 2015 at 1:40 PM, Alexei Starovoitov [off-list ref] wrote:

On Wed, Jan 28, 2015 at 4:46 PM, Namhyung Kim [off-list ref] wrote:

quoted

quoted

+static int event_filter_release(struct inode *inode, struct file *filp)
+{
+     struct ftrace_event_file *file;
+     char buf[2] = "0";
+
+     mutex_lock(&event_mutex);
+     file = event_file_data(filp);
+     if (file) {
+             if (file->flags & TRACE_EVENT_FL_BPF) {
+                     /* auto-disable the filter */
+                     ftrace_event_enable_disable(file, 0);

Hmm.. what if user already enabled an event, attached a bpf filter and
then detached the filter - I'm not sure we can always auto-disable
it..

why not?
I think it makes sense auto enable/disable, since that
is cleaner user experience.
Otherwise Ctrl-C of the user process will have bpf program dangling.
not good. If we auto-unload bpf program only, it's equally bad.
Since Ctrl-C of the process will auto-onload only
and will keep tracepoint enabled which will be spamming
the trace buffer.

I think it's not a problem of bpf.  An user process can be killed
anytime while it enabed events without bpf.  The only thing it should
care is the auto-unload IMHO.

quoted

quoted

+unsigned int trace_filter_call_bpf(struct event_filter *filter, void *ctx)
+{
+     unsigned int ret;
+
+     if (in_nmi()) /* not supported yet */
+             return 0;

But doesn't this mean to auto-disable all attached events during NMI
as returning 0 will prevent the event going to ring buffer?

well, it means that if tracepoint fired during nmi the program
won't be called and event won't be sent to trace buffer.
The program might be broken (like divide by zero) and
it will self-terminate with 'return 0'
so zero should be the safest return value that
causes minimum disturbance to the whole system overall.

I'm okay for not calling bpf program in NMI but not for disabling events.

Suppose an user was collecting an event (including in NMI) and then
[s]he also wanted to run a bpf program.  So [s]he wrote a program
always return 1.  But after attaching the program, it didn't record
the event in NMI..  Isn't that a problem?

quoted

I think it'd be better to keep an attached event in a soft-disabled
state like event trigger and give control of enabling to users..

I think it suffers from the same Ctrl-C issue.
Say, attaching bpf program activates tracepoint and keeps
it in soft-disabled. Then user space clears soft-disabled.
Then user Ctrl-C it. Now bpf program must auto-detach
and unload, since prog_fd is closing.
If we don't completely deactivate tracepoint, then
Ctrl-C will leave the state of the system in the state
different from it was before user process started running.
I think we must avoid such situation.
'kill pid' should be completely cleaning all resources
that user process was using.
Yes. It's different from typical usage of /sys/.../tracing
that has all global knobs, but, imo, it's cleaner this way.

Right.  I think bpf programs belong to a user process but events are
global resource.  Maybe you also need to consider attaching bpf
program via perf (ioctl?) interface..

Thanks,
Namhyung