On Mon 22 Jul 02:30 PDT 2019, Christoph Hellwig wrote:
> On Mon, Jul 22, 2019 at 10:38:55AM +0200, Marc Gonzalez wrote:
> > > In file included from drivers/firmware/qcom_scm.c:12:0:
> > > ./include/linux/dma-mapping.h:636:21: note: expected ‘dma_addr_t * {aka long long unsigned int *}’ but argument is of type ‘phys_addr_t * {aka unsigned int *}’
> > > static inline void *dma_alloc_coherent(struct device *dev, size_t size,
> > > ^~~~~~~~~~~~~~~~~~
> > > ```
> > >
> > > We just can cast phys_addr_t to dma_addr_t here.
> >
> > IME, casting is rarely a proper solution.
>
> *nod*
>
> ptr_phys probably should be a dma_addr_t. Unless this driver is so
> magic that it really wants a physical and not a dma address, in which
> case it needs to use alloc_pages instead of dma_alloc_coherent
> and then call page_to_phys on the returned page, and a very big comment
> explaining why it is so special.
The scm call takes physical addresses (which happens to be 1:1 with DMA
addresses for this driver).
This allocation started off (downstream) as a simple kmalloc(), but
while the scm call is being executed an access from Linux will cause a
security violation (that's not handled gracefully). The properties of
dma_alloc is closer, so that's where the code is today.
Optimally this should be something like alloc_pages() and some mechanism
for unmapping the pages during the call. But no one has come up with a
suitable patch for that.
But there's a patch from Stephen for this already (not doing a
typecast). Apparently I missed merging this, so I'll do that.
https://lore.kernel.org/linux-arm-msm/20190517210923.202131-2-swboyd@xxxxxxxxxxxx/
Regards,
Bjorn
