Suspend-to-ram not working when ftrace is enabled, again!

Hi,

If tracing is enabled and we are tracing low-level suspend-to-ram related functions like restore_processor_state() etc (which are included by default in the list of traced functions), and we try suspending the machine, the machine doesn't resume. It reboots instead.
(If we trace some unrelated functions like kzalloc() for example, there is no problem with suspend/resume).

Looking at https://lkml.org/lkml/2008/8/27/177, it appears that this is an old problem and also had a workaround (disabling tracing around suspend). The above patch corresponds to commit id: f42ac38c59 (ftrace:
disable tracing for suspend to ram), which went in around 2.6.27 I think. But then commit cbe2f5a6e84 (tracing: allow tracing of suspend/resume & hibernation code again) reverted that commit.

And from https://lkml.org/lkml/2008/8/21/349, it looks like 2.6.28 and further was supposed to be problem-free. But unfortunately this problem has resurfaced.

I tested kernel 2.6.32.54 and I observed that the machine reboots during resume, which looks exactly like the problem discussed in the link above.

In another machine, I tested 3.3-rc6 and it doesn't seem to respond to resume events (like button press, lid open) at all. It just seems to remain suspended forever.

Should we resort to disabling ftrace around suspend again? Or do we have a better solution this time around?

Quoted 8 lines · Srivatsa S. Bhat

Hi,

If tracing is enabled and we are tracing low-level suspend-to-ram related functions like restore_processor_state() etc (which are included by default in the list of traced functions), and we try suspending the machine, the machine doesn't resume. It reboots instead.
(If we trace some unrelated functions like kzalloc() for example, there is no problem with suspend/resume).

Yeah, this is a know issue. I need to look at the suspend code and add notrace annotations, or keep entire files from being traced.

The problem is that on resume, there's functions that are called that do not have all kernel setup initialized. For example, smp_processor_id() uses the %gs register to access the per_cpu data which also contains the cpu id. On resume, the %gs register is not yet set up, and calling the function tracer, which uses smp_processor_id() to find out what buffer to write to causes a page fault. Then the page fault handling also calls the function tracer which it too will page fault, and we end up with a triple fault and the machine reboots.

Quoted 19 lines

Looking at https://lkml.org/lkml/2008/8/27/177, it appears that this is an old problem and also had a workaround (disabling tracing around suspend). The above patch corresponds to commit id: f42ac38c59 (ftrace:
disable tracing for suspend to ram), which went in around 2.6.27 I think. But then commit cbe2f5a6e84 (tracing: allow tracing of suspend/resume & hibernation code again) reverted that commit.

And from https://lkml.org/lkml/2008/8/21/349, it looks like 2.6.28 and further was supposed to be problem-free. But unfortunately this problem has resurfaced.

I tested kernel 2.6.32.54 and I observed that the machine reboots during resume, which looks exactly like the problem discussed in the link above.

In another machine, I tested 3.3-rc6 and it doesn't seem to respond to resume events (like button press, lid open) at all. It just seems to remain suspended forever.

Quoted 3 lines

Should we resort to disabling ftrace around suspend again? Or do we have a better solution this time around?

No the real solution is to find the functions that break and fix them. Probably requires more notrace annotations.