Re: [PATCH v8 7/9] livepatch: Improve the search performance of module_kallsyms_on_each_symbol()

On Mon, Nov 14, 2022 at 07:30:16PM +0800, Leizhen (ThunderTown) wrote:

On 2022/11/14 18:31, Jiri Olsa wrote:

quoted

On Mon, Nov 14, 2022 at 06:00:38PM +0800, Leizhen (ThunderTown) wrote:

quoted

On 2022/11/14 17:27, Jiri Olsa wrote:

quoted

On Mon, Nov 14, 2022 at 04:50:25PM +0800, Leizhen (ThunderTown) wrote:

quoted

On 2022/11/14 15:47, Jiri Olsa wrote:

quoted

On Wed, Nov 02, 2022 at 04:49:19PM +0800, Zhen Lei wrote:

quoted

Currently we traverse all symbols of all modules to find the specified
function for the specified module. But in reality, we just need to find
the given module and then traverse all the symbols in it.

hi,
sorry for delayed answer, I did not notice this until Stephen's email
about merge issue with recent bpf change [1]

quoted

Let's add a new parameter 'const char *modname' to function
module_kallsyms_on_each_symbol(), then we can compare the module names

we have use case for iterating all modules and their symbols when we
want to resolve passed addresses for tracing

we don't have 'modname' that we could pass, we need to iterate all modules

so perhaps this could be made optional like with passing NULL for modname?

The deletion of modname was suggested by Petr Mladek. The reason is that
everyone passes modname as NULL, there was no actual demand at the time.
https://lkml.org/lkml/2022/9/20/682

quoted

quoted

directly in this function and call hook 'fn' after matching. And the
parameter 'struct module *' in the hook 'fn' can also be deleted.

we need 'struct module *' argument in the callback as well because we are
taking the module reference if we trace function in it, so it wont get
unloaded

please let me know if I should do the change or can help in any way

It seems that we should take the module reference before invoking callback
and put it after it is called, without passing modname.

we take the module ref only if we (callback) find the traced address in
the module, we don't have the module object before

jirka

Do it in function module_kallsyms_on_each_symbol()?

But I just saw that mutex_lock(&module_mutex) protection is already
provided in this function. So reference counting protection may not
be required.

we take the module ref so it won't unload even outside of the
module_kallsyms_on_each_symbol function

There's another way to do it, but it's more time consuming.

struct module *__module_text_address(unsigned long addr);
struct module *__module_address(unsigned long addr);

Which way do you think is more appropriate?

I think it'd be best to keep the module argument in the callback,
I don't see the harm.. AFAICS it was removed because you thought
nobody was using it, but that's not the case

using __module_text_address/__module_address might be doable and
it might be fast enough, because it's using rbtree

I'll check on that, meanwhile if we could keep the module argument,
that'd be great

thanks,
jirka