Re: [PATCH] scripts/gdb: fix list_for_each

[Kieran Bingham <kbingham@kernel.org>]

Hi George,

On 22/09/2020 18:17, Prekas, George wrote:
> 
> On 9/22/2020 9:32 AM, Jan Kiszka wrote:
>>
>> On 22.09.20 16:28, George Prekas wrote:
>>> If the next pointer is NULL, list_for_each gets stuck in an infinite
>>> loop.
>>>
>>> Signed-off-by: George Prekas <prekageo@amazon.com>
>>> ---
>>>   scripts/gdb/linux/lists.py | 2 ++
>>>   1 file changed, 2 insertions(+)
>>>
>>> diff --git a/scripts/gdb/linux/lists.py b/scripts/gdb/linux/lists.py
>>> index c487ddf09d38..424a91c1aa8b 100644
>>> --- a/scripts/gdb/linux/lists.py
>>> +++ b/scripts/gdb/linux/lists.py
>>> @@ -27,6 +27,8 @@ def list_for_each(head):
>>>           raise TypeError("Must be struct list_head not {}"
>>>                              .format(head.type))
>>>
>>> +    if head['next'] == 0:
>>> +        return
>>>       node = head['next'].dereference()
>>>       while node.address != head.address:
>>>           yield node.address
>>
>> Obviously, infinite loops are bad and should be avoided. But NULL is
>> bug, isn't it? Shouldn't we report such a corruption?
>>
> 
> Hi Jan,
> 
> Is it a bug? Or does it mean that the list is empty?

A correctly initialised (empty) list_head has the next, and prev
pointers pointing to itself

See:

 https://elixir.bootlin.com/linux/latest/source/include/linux/list.h#L33

You can see that the implementation of list_empty() checks for this at:

https://elixir.bootlin.com/linux/latest/source/include/linux/list.h#L280

/**
 * list_empty - tests whether a list is empty
 * @head: the list to test.
 */
static inline int list_empty(const struct list_head *head)
{
	return READ_ONCE(head->next) == head;
}

> Let me give some background. If you do the following:
> 
> $ qemu-system-x86_64 -nographic -m 1024 -kernel
> build/arch/x86/boot/bzImage -s -S < /dev/null > /dev/null &
> $ gdb -q build/vmlinux -ex 'target remote localhost:1234' -iex 'set
> auto-load safe-path /' -ex 'lx-symbols'

I suspect this is trying to load modules before the kernel is actually
fully loaded and running, so nothing is yet initialised.


> You will see:
> 
> loading vmlinux
> scanning for modules in /home/ubuntu/linux-5.8.10
> no module object found for ''
> 
> And the last line repeats forever. This happens because modules.next ==
> NULL. This is the Python stack trace:
> 
>   File ".../symbols.py", line 174, in invoke
>     self.load_all_symbols()
>   File ".../symbols.py", line 161, in load_all_symbols
>     [self.load_module_symbols(module) for module in module_list]
>   File ".../symbols.py", line 161, in <listcomp>
>     [self.load_module_symbols(module) for module in module_list]
>   File ".../modules.py", line 30, in module_list
>     for module in lists.list_for_each_entry(modules, module_ptr_type,
> "list"):
>   File ".../lists.py", line 41, in list_for_each_entry
>     for node in list_for_each(head):
>   File ".../lists.py", line 31, in list_for_each
>     traceback.print_stack()
> 
> This patch tries to fix the above problem.

Does it fix it for you ?

I expect it allows the boot process to continue, but the lx-symbols
command will not have completed successfully (or rather I expect it will
not have found anything to load).

I suspect adding defensive checks in here might be helpful but I think
the reality is the code is being called at the wrong time.

The fact that it 'can' be called at the wrong time is where we might
need to be more defensive.


> George

--
Kieran