Re: Re: modules, add_kallsyms() && DEFINE_PER_CPU

[Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>]

(2013/11/27 22:35), Oleg Nesterov wrote:
> On 11/27, Masami Hiramatsu wrote:
>>
>> (2013/11/27 2:50), Oleg Nesterov wrote:
>>> On 11/26, Oleg Nesterov wrote:
>>>>
>>>> Note: this doesn't work for modules, but module's per-cpu data is
>>>> not visible for kallsyms_lookup_name() anyway.
>>>
>>> Rusty, I am just curious if it makes sense to change this or not...
>>>
>>> But DEFINE_PER_CPU'ed symbols are ignored by add_kallsyms(). I guess
>>> this is because is_core_symbol() requires "sh_flags & SHF_ALLOC".
>>> And probably because of INIT_OFFSET_MASK.
>>
>> Oleg, I think you can do it by using is_module_percpu_address(). :)
> 
> Not only is_module_percpu_address() can't help. We can solve the
> is-it-percpu problem (although the check won't be cheap).

I think you can do this at registering phase (by adding a flag
in symbol_cache). That is in a slow path.

> The problem is, sc->addr is always NULL if DEFINE_PER_CPU() was used
> in a module. kallsyms_lookup_name() can never see it because it is
> ignored by add_kallsyms(), it is nacked by is_core_symbol().
>
> IOW, with or without this patch @modular_percpu_sym never reads the
> memory.

Hmm, in that case, should we support it? It looks non-existed
symbol is given. I think until the module.c support per-cpu in
kallsyms, we can delay to support it. I mean we can do 3 steps
1. support percpu symbol in the kernel except for drivers
2. update module.c to store the percpu symbols into kallsyms.
3. support percpu symbol in module .


Thank you,

-- 
Masami HIRAMATSU
IT Management Research Dept. Linux Technology Center
Hitachi, Ltd., Yokohama Research Laboratory
E-mail: masami.hiramatsu.pt@hitachi.com