On Thu, Sep 4, 2025 at 12:59 AM Nathan Chancellor nathan@kernel.org wrote:
On Wed, Sep 03, 2025 at 02:04:10PM +0200, Arnd Bergmann wrote:
On Wed, Sep 3, 2025, at 12:08, Naresh Kamboju wrote:
Build error: drivers/dma/mmp_pdma.c:1188:14: error: shift count >= width of type [-Werror,-Wshift-count-overflow] 1188 | .dma_mask = DMA_BIT_MASK(64), /* force 64-bit DMA addr capability */ | ^~~~~~~~~~~~~~~~ include/linux/dma-mapping.h:73:54: note: expanded from macro 'DMA_BIT_MASK' 73 | #define DMA_BIT_MASK(n) (((n) == 64) ? ~0ULL : ((1ULL<<(n))-1)) | ^ ~~~
I see two separate issues:
- The current DMA_BIT_MASK() definition seems unfortunate, as the
'(n) == 64' check is meant to avoid this problem, but I think this only works inside of a function, not in a static structure definition.
Right, this is one of our longest outstanding issues :/
https://github.com/ClangBuiltLinux/linux/issues/92 https://github.com/llvm/llvm-project/issues/38137
This only happens at global scope.
This could perhaps be avoided by replacing the ?: operator with __builtin_choose_expr(), but that likely causes other build failures.
Yeah, that makes the problem worse somehow even though GCC says the non-taken option should not be evaluated...
drivers/dma/mmp_pdma.c:1188:14: error: shift count >= width of type [-Werror,-Wshift-count-overflow] 1188 | .dma_mask = DMA_BIT_MASK(64), /* force 64-bit DMA addr capability */ | ^~~~~~~~~~~~~~~~ include/linux/dma-mapping.h:73:70: note: expanded from macro 'DMA_BIT_MASK' 73 | #define DMA_BIT_MASK(n) __builtin_choose_expr((n) == 64, ~0ULL, (1ULL<<(n))-1) | ^ ~~~ drivers/dma/mmp_pdma.c:1323:27: error: shift count >= width of type [-Werror,-Wshift-count-overflow] 1323 | dma_set_mask(pdev->dev, DMA_BIT_MASK(64)); | ^~~~~~~~~~~~~~~~ include/linux/dma-mapping.h:73:70: note: expanded from macro 'DMA_BIT_MASK' 73 | #define DMA_BIT_MASK(n) __builtin_choose_expr((n) == 64, ~0ULL, (1ULL<<(n))-1) | ^ ~~~
Thanks Nathan for the information here and above.
Guodong, how about a patch to drop all the custom dma_mask handling and instead just use dma_set_mask_and_coherent(DMA_BIT_MASK(64)) or dma_set_mask_and_coherent(DMA_BIT_MASK(32)) here? Instead of passing the mask in the mmp_pdma_ops, you can replace it e.g. with a 'bool addr64' flag, or an 'int dma_width' number that gets passed into the DMA_MASK_MASK().
Thanks, Arnd. I'll send a patch to clean up and simplify the logic.
If this works, I think it is worth pursuing to avoid this bogus warning/error.
Cheers, Nathan