Re: [PATCH v3 02/11] mm: Hardened usercopy

From: Laura Abbott
Date: Wed Jul 20 2016 - 11:36:54 EST

On 07/20/2016 03:24 AM, Balbir Singh wrote:
On Tue, 2016-07-19 at 11:48 -0700, Kees Cook wrote:
On Mon, Jul 18, 2016 at 6:06 PM, Laura Abbott <labbott@xxxxxxxxxx> wrote:

On 07/15/2016 02:44 PM, Kees Cook wrote:

This doesn't work when copying CMA allocated memory since CMA purposely
allocates larger than a page block size without setting head pages.
Given CMA may be used with drivers doing zero copy buffers, I think it
should be permitted.

Something like the following lets it pass (I can clean up and submit
the is_migrate_cma_page APIs as a separate patch for review)
Yeah, this would be great. I'd rather use an accessor to check this
than a direct check for MIGRATE_CMA.

for (; ptr <= end ; ptr += PAGE_SIZE, page = virt_to_head_page(ptr))
- if (!PageReserved(page))
+ if (!PageReserved(page) && !is_migrate_cma_page(page))
return "<spans multiple pages>";
Yeah, I'll modify this a bit so that which type it starts as is
maintained for all pages (rather than allowing to flip back and forth
-- even though that is likely impossible).

Sorry, I completely missed the MIGRATE_CMA bits. Could you clarify if you
caught this in testing/review?

Balbir Singh.

I caught it while looking at the code and then wrote a test case to confirm
I was correct because I wasn't sure how to easily find an in tree user.