Re: Question regarding MAX_ARG_STRLEN with execve()

[Anshuman Khandual <khandual@linux.vnet.ibm.com>]

On 07/03/2017 02:51 PM, Michal Hocko wrote:
> On Mon 03-07-17 13:58:59, Anshuman Khandual wrote:
>> On 06/30/2017 07:52 PM, Michal Hocko wrote:
>>> On Fri 30-06-17 11:59:37, Anshuman Khandual wrote:
>>>> Hello,
>>>>
>>>> execve() system call should support argument length of
>>>> MAX_ARG_STRLEN (PAGE_SIZE * 32). On 64K page size systems, we
>>>> are not able to pass 32 * PAGE_SIZE arguments into the execve()
>>>> system call because of the following reasons.
>>>>
>>>> * struct linux_binprm's vma starts with a size of PAGE_SIZE
>>>>
>>>> 	vma->vm_end = STACK_TOP_MAX;
>>>> 	vma->vm_start = vma->vm_end - PAGE_SIZE;
>>>>
>>>> * The VMA expands as much depending upon the argument size. So
>>>>   for 32 * PAGE_SIZE argument, it becomes 33 * PAGE_SIZE.
>>>>
>>>> * 33 * PAGE_SIZE with 64K pages fails the following test in
>>>>   get_arg_page() function. 33 * PAGE_SIZE is more than 2MB
>>>>   (8 MB /4) with 64K page size.
>>>>
>>>>    if (size > READ_ONCE(rlim[RLIMIT_STACK].rlim_cur) / 4)
>>>>
>>>> * Right now RLIMIT_STACK is hard coded 8MB which does not take
>>>>   PAGE_SIZE into account. 
>>>>
>>>> Wondering what should be the solution for this problem ?
>>>>
>>>> * Change the default stack size from 8MB ?
>>> just increase the ulimit if you want to use such a large arguments.
>>>
>>
>> Yeah that is possible but it does not still offset the fact that
>> the calculation is broken on the page size of 64K. I mean, yeah
>> its not practical to have such a large argument. But the point
>> is whether we would want to support the MAX_ARG_STRLEN semantic
>> for execve system call or not. At present its broken for 64K
>> and I am asking whether we will be willing to revisit the
>> '1/4th of the stack' condition.
> 
> I dunno. We have this 1/4 of RLIMIT semantic for years and it doesn't
> seem there were any bug reports. Yes, MAX_ARG_STRLEN being PAGE_SIZE
> dependent is unfortunate because it makes an arch independent default
> ulimit hard to get right but I am not sure we actually have to lose
> sleep over this.

I understand your point.

> 
> Or do you have any specific proposal how to "fix" this limitation which
> wouldn't break other userspace?

There are three variables here MAX_ARG_STRLEN, RLIMIT_STACK and the 25%
condition. Execve() is supporting MAX_ARG_STRLEN for a long time, hence
it cannot be changed now. That leaves us to change either the default
RLIMIT_STACK value or the 25% condition. Both are kernel internal
implementation. But I am not sure how changing them might affect any
other userspace behavior, hence asking for suggestions. I just wanted
to explore the possibilities of a fix here.