Re: [RFC][PATCH 3/7] kprobes: Remove kretprobe hash

[Masami Hiramatsu]

On Fri, 28 Aug 2020 16:19:17 +0200
peterz@xxxxxxxxxxxxx wrote:

> On Fri, Aug 28, 2020 at 02:11:18PM +0000, Eddy_Wu@xxxxxxxxxxxxxx wrote:
> > > From: Masami Hiramatsu <mhiramat@xxxxxxxxxx>
> > >
> > > OK, schedule function will be the key. I guess the senario is..
> > >
> > > 1) kretporbe replace the return address with kretprobe_trampoline on task1's kernel stack
> > > 2) the task1 forks task2 before returning to the kretprobe_trampoline
> > > 3) while copying the process with the kernel stack, task2->kretprobe_instances.first = NULL
> > 
> > I think new process created by fork/clone uses a brand new kernel
> > stack? I thought only user stack are copied.  Otherwise any process
> > launch should crash in the same way
> 
> I was under the same impression, we create a brand new stack-frame for
> the new task, this 'fake' frame we can schedule into.
> 
> It either points to ret_from_fork() for new user tasks, or
> kthread_frame_init() for kernel threads.

Ah sorry, then it's my misreading... anyway, I could reproduce the crash with
probing on schedule(). Hmm, it is better to dump the current comm with
BUG().

Thank you,



-- 
Masami Hiramatsu <mhiramat@xxxxxxxxxx>