Message-ID: <d3cad11c-a65d-4faf-a636-3d85474d7175@189.cn>
Date: Wed, 20 Nov 2024 10:17:10 +0800
From: Song Chen <chensong_2000@....cn>
To: Petr Pavlu <petr.pavlu@...e.com>
Cc: mcgrof@...nel.org, samitolvanen@...gle.com, da.gomez@...sung.com,
 linux-kernel@...r.kernel.org, linux-modules@...r.kernel.org
Subject: Re: [PATCH] kmod: verify module name before invoking modprobe

Hi Petr,

在 2024/11/18 20:54, Petr Pavlu 写道:
> On 11/13/24 03:15, Song Chen wrote:
>> 在 2024/11/12 20:56, Petr Pavlu 写道:
>>> On 11/10/24 12:42, Song Chen wrote:
>>>> Sometimes when kernel calls request_module to load a module
>>>> into kernel space, it doesn't pass the module name appropriately,
>>>> and request_module doesn't verify it as well.
>>>>
>>>> As a result, modprobe is invoked anyway and spend a lot of time
>>>> searching a nonsense name.
>>>>
>>>> For example reported from a customer, he runs a user space process
>>>> to call ioctl(fd, SIOCGIFINDEX, &ifr), the callstack in kernel is
>>>> like that:
>>>> dev_ioctl(net/core/dev_iovtl.c)
>>>>     dev_load
>>>>        request_module("netdev-%s", name);
>>>>        or request_module("%s", name);
>>>>
>>>> However if name of NIC is empty, neither dev_load nor request_module
>>>> checks it at the first place, modprobe will search module "netdev-"
>>>> in its default path, env path and path configured in etc for nothing,
>>>> increase a lot system overhead.
>>>>
>>>> To address this problem, this patch copies va_list and introduces
>>>> a helper is_module_name_valid to verify the parameters validity
>>>> one by one, either null or empty. if it fails, no modprobe invoked.
>>>
>>> I'm not sure if I fully follow why this should be addressed at the
>>> request_module() level. If the user repeatedly invokes SIOCGIFINDEX with
>>> an empty name and this increases their system load, wouldn't it be
>>> better to update the userspace to prevent this non-sense request in the
>>> first place?
>>
>> If the user process knew, it wouldn't make the mistake.
> 
> The user process should be able to check that the ifr_name passed to
> SIOCGIFINDEX is empty and avoid the syscall altogether, or am I missing
> something? Even if the kernel gets improved in some way to handle this
> case better, I would still suggest looking at what the application is
> doing and how it ends up making this call.
> 

yes, agree, it's the user space process's fault after all.

>> moreover, what
>> happened in dev_load was quite confusing, please see the code below:
>>
>>       no_module = !dev;
>>       if (no_module && capable(CAP_NET_ADMIN))
>>           no_module = request_module("netdev-%s", name);
>>       if (no_module && capable(CAP_SYS_MODULE))
>>           request_module("%s", name);
>>
>> Running the same process, sys admin or root user spends more time than
>> normal user, it took a while for us to find the cause, that's why i
>> tried to fix it in kernel.
>>
>> Similarly, if something should be done in the kernel,
>>> wouldn't it be more straightforward for dev_ioctl()/dev_load() to check
>>> this case?
>>
>> I thought about it at the beginning, not only dev_ioctl/dev_load but
>> also other request_module callers should check this case as well, that
>> would be too much effort, then I switched to check it at the beginning
>> of request_module which every caller goes through.
>>
>>>
>>> I think the same should in principle apply to other places that might
>>> invoke request_module() with "%s" and a bogus value. The callers can
>>> appropriately decide if their request makes sense and should be
>>> fixed/improved.
>>>
>>
>> Callees are obliged to do fault tolerance for callers, or at least let
>> them know what is going on inside, what kinds of mistake they are
>> making, there are a lot of such cases in kernel, such as call_modprobe
>> in kernel/module/kmod.c, it checks if orig_module_name is NULL.
> 
> Ok, I see the idea behind checking that a value passed to
> request_module() to format "%s" is non-NULL.
> 
> I'm however not sure about rejecting empty strings as is also done by
> the patch. Consider a call to request_module("mod%s", suffix) where the
> suffix could be empty to select the default variant, or non-empty to
> select e.g. some optimized version of the module. Only the caller knows
> if the suffix being empty is valid or not.
> 
> I've checked if this pattern is currently used in the kernel and wasn't
> able to find anything, so that is good. However, I'm not sure if
> request_module() should flat-out reject this use.
> 

I accidentally found another problem in request_module when i was 
testing this patch again, if the caller just passes a empty pointer to 
request_module, like request_module(NULL), the process will be broken:

[    2.336160]  ? asm_exc_page_fault+0x2b/0x30
[    2.336160]  ? __pfx_crc64_rocksoft_notify+0x10/0x10
[    2.336160]  ? vsnprintf+0x5a/0x4f0
[    2.336160]  __request_module+0x93/0x2b0
[    2.336160]  ? __pfx_crc64_rocksoft_notify+0x10/0x10
[    2.336160]  ? notifier_call_chain+0x65/0xd0
[    2.336160]  ? __pfx_crc64_rocksoft_notify+0x10/0x10
[    2.336160]  crypto_probing_notify+0x43/0x60

(please ignore the caller, that is a testing code.)

I searched kernel code if this patter exists, and found in 
__trace_bprintk of kernel/trace/trace_printk.c, it checks fmt at the 
beginning of the function:

      va_list ap;

      if (unlikely(!fmt))
          return 0;

Therefore, i would like to suggest we should at least add this check in 
request_module too. In that sense, why don't we do a little further to 
verify every parameter's validity to provide better fault tolerance, 
besides, it costs almost nothing.

If you like this idea, i will send a v2.

Many thanks.

Song

Powered by blists - more mailing lists