On 04/14/2012 03:38 AM, Russell King - ARM Linux wrote:On Fri, Apr 13, 2012 at 04:05:50PM +0200, Marek Szyprowski wrote:This patch changes dma-mapping subsystem to use generic vmalloc areas
for all consistent dma allocations. This increases the total size limit
of the consistent allocations and removes platform hacks and a lot of
duplicated code.
NAK. I don't think you appreciate the contexts from which the dma coherent
code can be called from, and the reason why we pre-allocate the page
tables (so that IRQ-based allocations work.)
The vmalloc region doesn't allow that because page tables are allocated
using GFP_KERNEL not GFP_ATOMIC.
Sorry.
Off-topic.
I don't know why vmalloc functions have gfp_t argument.
As Russel pointed out, we allocates page tables with GFP_KERNEL regardless of gfp_t passed.
It means gfp_t passed is useless.
I see there are many cases calling __vmalloc with GFP_NOFS, even GFP_ATOMIC. Then, it could end up deadlocking in reclaim context or schedule bug.
I'm not sure why we can't see such bugs until now.
If I didn't miss something, Shouldn't we fix it?