Re: [RFC][PATCH] x86/mm: Sync all vmalloc mappings before text_poke()

[Joerg Roedel <jroedel@suse.de>]

On Thu, Apr 30, 2020 at 12:11:36PM -0400, Steven Rostedt wrote:
> bool trace_event_ignore_this_pid(struct trace_event_file *trace_file)
> {
> 	struct trace_array *tr = trace_file->tr;
> 	struct trace_array_cpu *data;
> 	struct trace_pid_list *no_pid_list;
> 	struct trace_pid_list *pid_list;
> 
> 	pid_list = rcu_dereference_raw(tr->filtered_pids);
> 	no_pid_list = rcu_dereference_raw(tr->filtered_no_pids);
> 
> 	if (!pid_list && !no_pid_list)
> 		return false;
> 
> 	data = this_cpu_ptr(tr->array_buffer.data);
> 
> 	return data->ignore_pid;
> }
>
> [...]
> 
> tl;dr; It's not an issue with the vmalloc, it's an issue with per_cpu
> allocations!

Yes, looks like you are right, it faults on the return statement, so the
data-pointer seems not to be mapped. I looked at another instance:

The trapping instruction is:

	movzbl 0x7c(%rax),%eax

Which is a pointer-dereference, with %rax=0xffffe8ffffccc870. The 'data'
pointer is allocated with alloc_percpu().

And alloc_percpu() calls down into pcpu_alloc(), which allocates new
percpu chunks using vmalloc() on x86. And there we are again in the
vmalloc area.

Regards,

	Joerg