[PATCH v3 2/3] dma-mapping: replace zone_dma_bits by zone_dma_limit

[Nathan Chancellor]

On Tue, Jul 30, 2024 at 05:34:50PM +0200, Christoph Hellwig wrote:
> On Mon, Jul 29, 2024 at 07:12:08PM -0700, Nathan Chancellor wrote:
> > >          |             ~~~~~~~~~~~~~~                   ^~~~~~~~~~~~~~~~
> > >    include/linux/dma-mapping.h:77:40: note: expanded from macro 'DMA_BIT_MASK'
> > >       77 | #define DMA_BIT_MASK(n) (((n) == 64) ? ~0ULL : ((1ULL<<(n))-1))
> > >          |                                        ^~~~~
> > >    2 warnings generated.
> > 
> > FWIW, this is likely a false positive due to an issue in Clang with the
> > control flow graph for global variables:
> > 
> > https://github.com/ClangBuiltLinux/linux/issues/92
> > 
> > DMA_BIT_MASK() has been the biggest offender :/ If there is any way to
> > refactor this code to avoid this, that would be great (as that has been
> > one of our longest outstanding issues and getting it fixed in the
> > compiler does not seem super easy at this point).
> 
> I have no idea what you'd want changed here, but I'll happily take
> patches.

Unfortunately, I am not sure either... I do not see anything obviously,
so perhaps it could just be avoided with the __diag() infrastructure?

diff --git a/kernel/dma/direct.c b/kernel/dma/direct.c
index 3dbc0b89d6fb..b58e7eb9c8f1 100644
--- a/kernel/dma/direct.c
+++ b/kernel/dma/direct.c
@@ -20,7 +20,12 @@
  * it for entirely different regions. In that case the arch code needs to
  * override the variable below for dma-direct to work properly.
  */
+__diag_push();
+__diag_ignore(clang, 13, "-Wconstant-conversion",
+	      "Clang incorrectly thinks the n == 64 case in DMA_BIT_MASK() can happen here,"
+	      "which would truncate with a 32-bit phys_addr_t");
 phys_addr_t zone_dma_limit __ro_after_init = DMA_BIT_MASK(24);
+__diag_pop();
 
 static inline dma_addr_t phys_to_dma_direct(struct device *dev,
 		phys_addr_t phys)