Re: [PATCH v2 1/4] mm/mlock: return EINVAL if len overflows for mlock/munlock

[Andrew Morton]

On Mon, 16 Jan 2023 19:58:10 +0800 Wupeng Ma <mawupeng1@xxxxxxxxxx> wrote:

> From: Ma Wupeng <mawupeng1@xxxxxxxxxx>
> 
> While testing mlock, we have a problem if the len of mlock is ULONG_MAX.
> The return value of mlock is zero. But nothing will be locked since the
> len in do_mlock overflows to zero due to the following code in mlock:
> 
>   len = PAGE_ALIGN(len + (offset_in_page(start)));
> 
> The same problem happens in munlock.
> 
> Add new check and return -EINVAL to fix this overflowing scenarios since
> they are absolutely wrong.
> 
> ...
>
> --- a/mm/mlock.c
> +++ b/mm/mlock.c
> @@ -569,6 +569,7 @@ static __must_check int do_mlock(unsigned long start, size_t len, vm_flags_t fla
>  	unsigned long locked;
>  	unsigned long lock_limit;
>  	int error = -ENOMEM;
> +	size_t old_len = len;

I'm not sure that "old_len" is a good identifier.  It reads to me like
"the length of the old mlocked region" or something.

I really don't like it when functions modify the values of the incoming
argument like this.  It would be better to leave `len' alone and create
a new_len or something.

>  	start = untagged_addr(start);
>  
> @@ -578,6 +579,9 @@ static __must_check int do_mlock(unsigned long start, size_t len, vm_flags_t fla
>  	len = PAGE_ALIGN(len + (offset_in_page(start)));
>  	start &= PAGE_MASK;
>  
> +	if (old_len != 0 && len == 0)
> +		return -EINVAL;

It would be clearer to do this immediately after calculating the new
value of `len'.  Before going on to play with `start'.

Can we do something like this?

--- a/mm/mlock.c~a
+++ a/mm/mlock.c
@@ -575,7 +575,12 @@ static __must_check int do_mlock(unsigne
 	if (!can_do_mlock())
 		return -EPERM;
 
-	len = PAGE_ALIGN(len + (offset_in_page(start)));
+	if (len) {
+		len = PAGE_ALIGN(len + (offset_in_page(start)));
+		if (len == 0)	/* overflow */
+			return -EINVAL;
+	}
+
 	start &= PAGE_MASK;
 
 	lock_limit = rlimit(RLIMIT_MEMLOCK);
_

That depends on how we handle len==0.  afaict, mlock(len==0) will
presently burn a bunch of cpu cycles (not that we want to optimize this
case), do nothing then return 0?