On Thu, 25 Nov 2021, Petr Mladek wrote:
> On Fri 2021-11-19 10:03:27, Miroslav Benes wrote:
> > Add a test for the API which allows the user to specify functions which
> > are then searched for on any tasks's stack during a transition process.
> >
> > --- /dev/null
> > +++ b/lib/livepatch/test_klp_funcstack_mod.c
> > @@ -0,0 +1,72 @@
> > +// SPDX-License-Identifier: GPL-2.0
> > +// Copyright (C) 2021 Miroslav Benes <mbenes@xxxxxxx>
> > +
> > +#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
> > +
> > +#include <linux/module.h>
> > +#include <linux/kernel.h>
> > +#include <linux/debugfs.h>
> > +#include <linux/delay.h>
> > +
> > +static int sleep_length = 10000;
> > +module_param(sleep_length, int, 0644);
> > +MODULE_PARM_DESC(sleep_length, "length of sleep in seconds (default=10)");
> > +
> > +static noinline void child_function(void)
> > +{
> > + pr_info("%s enter\n", __func__);
> > + msleep(sleep_length);
>
> The hardcoded sleep is not ideal. It might be too low or non-necessary high.
It is not.
> If I get it correctly, we are trying to achieve here the same as
> busymod_work_func() in test_klp_callbacks_busy.c.
Yes.
> The approach with debugfs is an interesting trick. Though, I slightly
> prefer using the scheduled work. The workqueue API looks less tricky
> to me than sysfs/debugfs API. Also it does not block the module
> in the init() callback[*]. But I might be biased.
It seemed to me that debugfs gave us more control over the process than
workqueues, but I do not really care. Could you explain the blocking in
the init callback? I do not follow.
> Anyway, it might make sense to use the same trick in both situations.
> It would make it easier to maintain the test modules.
True. So I will rewrite it to workqueues as you are proposing below.
> [*] There is actually a race in the workqueue approach. The module
> init() callback should wait until the work is really scheduled
> and sleeping. It might be achieved by similar hand-shake like
> with @block_transition variable. Or completion API might be
> even more elegant.
>
>
> > + pr_info("%s exit\n", __func__);
> > +}
> > +
> > +static noinline void child2_function(void)
> > +{
> > + pr_info("%s\n", __func__);
> > +}
> > +
> > +static noinline void parent_function(void)
> > +{
> > + pr_info("%s enter\n", __func__);
> > + child_function();
> > + child2_function();
>
> This would deserve some explanation what we try to simulate here
> and how it is achieved. It is not easy for me even with the background
> that I have freshly in my mind.
>
> Also I think about more descriptive names ;-)
Hey, I thought it was self-explaining :). So, yes, I started with the
example given in the .fixup thread, but it is not really tied to .cold
section, jumps or whatever. The setup is just used to test a new API.
Moreover, the .fixup example is just a one scenario the new API tries to
solve.
What you propose below, that is function names and comments, is a bit
confusing for me. Especially if I did not know anything about the original
issue (which will be the case in a couple of weeks when I forget
everything).
So I think it I will stick to brevity unless you or someone else really
insist.
I can improve tests description in
tools/testing/selftests/livepatch/test-func-stack.sh if it helps anything.
Miroslav
> What about something like this (using workqueue work and completion):
>
> /*
> * Simulate part of the caller code that is in another .elf section
> * and is reached via jump. It this was really the case then the stack
> * unwinder might not be able to detect that the process is sleeping
> * in the caller.
> */
> static void simulate_jump_part(void)
> {
> pr_info("%s enter\n", __func__);
>
> /* Stay in the jump part unless told to leave. */
> wait_for_completion(finish_jump);
>
> pr_info("%s exit\n", __func__);
> }
>
> /*
> * Simulate modified part of the caller code. It should never get
> * livepatched when the caller is sleeping in the just_part().
> */
> static void simulate_modified_part(void)
> {
> pr_info("%s\n", __func__);
> }
>
> static void test_not_on_stack_func_work(struct work_struct *work)
> {
> pr_info("%s enter\n", __func__);
>
> /* Simulation ready */
> complete(work_started);
>
> simulate_jump_part();
> simulate_modified_part();
>
> pr_info("%s exit\n", __func__);
> }
>
> static int test_klp_no_on_stack_init(void)
> {
> pr_info("%s\n", __func__);
>
> schedule_work(&work);
> wait_for_completion(&work_started);
>
> return 0;
> }
>
> static void test_not_on_stack_exit(void)
> {
> complete(&finish_jump);
> flush_work(&work);
> pr_info("%s\n", __func__);
> }
>
> module_init(test_klp_not_on_stack_init);
> module_exit(test_klp_not_on_stack_exit);
>
> > + pr_info("%s exit\n", __func__);
> > +}
> > +
