Re: [PATCHv4 02/10] mm: convert mm->nr_ptes to atomic_t

[Dave Hansen <dave.hansen@intel.com>]

On 09/27/2013 01:46 PM, Cody P Schafer wrote:
> On 09/27/2013 06:16 AM, Kirill A. Shutemov wrote:
>> @@ -339,6 +339,7 @@ struct mm_struct {
>>       pgd_t * pgd;
>>       atomic_t mm_users;            /* How many users with user space? */
>>       atomic_t mm_count;            /* How many references to "struct
>> mm_struct" (users count as 1) */
>> +    atomic_t nr_ptes;            /* Page table pages */
>>       int map_count;                /* number of VMAs */
...
> 
> Will 32bits always be enough here? Should atomic_long_t be used instead?

There are 48 bits of virtual address space on x86 today.  12 bits of
that is the address inside the page, so we've at *most* 2^36 pages.  2^9
(512) pages are mapped by a pte page, so that means the page tables only
hold 2^27 pte pages in a single process.

We've got 31 bits of usable space in the atomic_t, so that definitely
works _today_.  If the virtual address space ever gets bigger, we might
have problems, though.

In practice, though, we steal a big chunk of that virtual address space
for the kernel, and that doesn't get accounted in mm->nr_ptes, so we've
got a _bit_ more wiggle room than just 4 bits.  Also, anybody that's
mapping >4 petabytes of memory with 4k ptes is just off their rocker.

I'm also not sure what the virtual address limits are for the more
obscure architectures, so I guess it's also possible they'll hit this.
I guess it wouldn't hurt to stick an overflow check in there for VM
debugging purposes.