On Wed 27-02-19 16:12:30, Laura Abbott wrote:
> On 2/26/19 6:29 AM, Christoph Hellwig wrote:
> > I don't think this is a good idea. The whole concept of just passing
> > random GFP_ flags to dma_alloc_attrs / dma_alloc_coherent can't work,
> > given that on many architectures we need to set up new page tables
> > to remap the allocated memory, and we can't use arbitrary gfp flags
> > for pte allocations.
> >
> > So instead of trying to pass them further down again we need to instead
> > work to fix all callers of dma_alloc_attrs / dma_alloc_coherent
> > that don't just pass GFP_KERNEL.
> >
>
> What's the expected approach to fix callers? It's not clear how
> you would fix the callers for the case that prompted this series
> (context correctly used GFP_NOIO but it was not passed to
> dma_alloc_coherent)
Use the scope API (memalloc_noio_{save,restore}) at the scope boundary
(lock or other restriction that prohibids IO path recursion) and you do
not have to care about the specific allocation because it will inherit
GFP_IO automagically.
--
Michal Hocko
SUSE Labs
