Re: [PATCH] uprobe: Add uprobe_pre/post_sstep_notifier to NOKPROBE_SYMBOL

[Pratyush Anand]

On Wed, Aug 3, 2016 at 9:54 AM, Pratyush Anand <panand@xxxxxxxxxx> wrote:
> Hi Masami,
>
> On 03/08/2016:12:45:24 AM, Masami Hiramatsu wrote:
>> On Tue,  2 Aug 2016 12:14:06 +0530
>> Pratyush Anand <panand@xxxxxxxxxx> wrote:
>>
>> > uprobe_pre_sstep_notifier and uprobe_post_sstep_notifier are called from
>> > debug exception handler, so blacklist them for kprobing.
>>
>> Actually, these exception notifers are kicked only if the debug exception
>> is not related to kprobes (at least on x86). In that case, we don't have
>> to take care about that. Or, would you hit any problem on it?
>
> Well, I have faced issue on ARM64. So, if I have a kprobe instrumented at these
> functions and then if I hit a uprobe then kernel goes into an infinite loop of
> "Unexpected kernel single-step exception at EL1".
>
> On x86 I have not tested, but I see that all functions except
> arch_uprobe_exception_notify() in the call stack of
> uprobe_pre/post_sstep_notifier() are blacklisted for kprobe. So, I am unable to
> understand that why arch_uprobe_exception_notify() and
> uprobe_pre/post_sstep_notifier() are not blacklisted.
>
>>
>> IOW, where do we have to prohibit kprobes are, the code path from where
>> right after the breakpoint (debug) exception is occurred, to where right
>> before the kprobe is handled. After that, it should be safe.
>
> Hummmm...My understanding was that if a function a() is not good to be kprobed
> then we can not kprobe any function called by a() as well. Thanks for the
> clarification. So, if I go with your definition then, something is still wrong on
> ARM64 which is causing issue when I kprobe uprobe_pre/post_sstep_notifier().

I found that one modification in ARM64 kprobe code allows me to kprobe
uprobe_pre/post_sstep_notifier(). So, taking back this patch. Will
discuss ARM64 modification on arm mailing list.

Thanks Masami for your input.

~Pratyush