Re: [PATCH v6 1/4] uio: introduce UIO_MEM_DMA_COHERENT type

[Guenter Roeck]

On 3/22/24 07:30, Greg Kroah-Hartman wrote:
> On Fri, Mar 22, 2024 at 07:16:19AM -0700, Guenter Roeck wrote:
> > On Mon, Feb 05, 2024 at 12:01:37PM -0800, Chris Leech wrote:
> > > Add a UIO memtype specifically for sharing dma_alloc_coherent
> > > memory with userspace, backed by dma_mmap_coherent.
> > >
> > > This is mainly for the bnx2/bnx2x/bnx2i "cnic" interface, although there
> > > are a few other uio drivers which map dma_alloc_coherent memory and will
> > > be converted to use dma_mmap_coherent as well.
> > >
> > > Signed-off-by: Nilesh Javali <njavali@xxxxxxxxxxx>
> > > Signed-off-by: Chris Leech <cleech@xxxxxxxxxx>
> > > ---
> >
> > Building i386:allyesconfig ... failed
> > --------------
> > Error log:
> > drivers/uio/uio.c: In function 'uio_mmap_dma_coherent':
> > drivers/uio/uio.c:795:16: error: cast to pointer from integer of different size [-Werror=int-to-pointer-cast]
> >    795 |         addr = (void *)mem->addr;
> >        |                ^
>
> So on 32bit systems phys_addr_t != the same size as (void *)?  How is
> that possible?  We also are doing an explicit cast here, how does this
> not work?

phys_addr_t is not always equivalent to the size of a pointer.
PHYS_ADDR_T_64BIT is a configuration option, after all, and it
is independent of "config 64BIT".

> Ah, do you have CONFIG_X86_PAE enabled?  That would cause that mess,
> ick.
>
>
> > cc1: all warnings being treated as errors
> > make[5]: [scripts/Makefile.build:244: drivers/uio/uio.o] Error 1 (ignored)
> > drivers/uio/uio_dmem_genirq.c: In function 'uio_dmem_genirq_open':
> > drivers/uio/uio_dmem_genirq.c:63:39: error: cast from pointer to integer of different size [-Werror=pointer-to-int-cast]
> >     63 |                 uiomem->addr = addr ? (phys_addr_t) addr : DMEM_MAP_ERROR;
> >        |                                       ^
> > drivers/uio/uio_dmem_genirq.c: In function 'uio_dmem_genirq_release':
> > drivers/uio/uio_dmem_genirq.c:92:43: error: cast to pointer from integer of different size [-Werror=int-to-pointer-cast]
> >     92 |                                           (void *) uiomem->addr,
> >        |                                           ^
> > cc1: all warnings being treated as errorsphys_addr_t
> > make[5]: [scripts/Makefile.build:244: drivers/uio/uio_dmem_genirq.o] Error 1 (ignored)
> > drivers/uio/uio_pruss.c: In function 'pruss_probe':
> > drivers/uio/uio_pruss.c:194:34: error: cast from pointer to integer of different size [-Werror=pointer-to-int-cast]
> >    194 |                 p->mem[2].addr = (phys_addr_t) gdev->ddr_vaddr;
> >        |                                  ^
> > cc1: all warnings being treated as errors
> >
> > Caused by this patch and "uio_dmem_genirq: UIO_MEM_DMA_COHERENT conversion" as well
> > as "uio_pruss: UIO_MEM_DMA_COHERENT conversion".
> >
> > I'd suggest to make uio dependent on 64 bit if 32 bit is no longer supported
> > to prevent waste of test builds resources.
>
> Perhaps disable it if PHYS_ADDR_T_64BIT is not enabled?

i386:allyesconfig sets

CONFIG_X86_HAVE_PAE=y
CONFIG_X86_PAE=y
CONFIG_PHYS_ADDR_T_64BIT=y

so that would not really help. The problem here is that this is a 32-bit build,
meaning pointers are 32 bit, while phys_addr_t is 64 bit. I did not check the
code, but it will simply not work if the pointer is supposed to reflect a
physical address. The new dependency would have to check if sizeof(void *) is
larger or equal to sizeof(phys_addr_t). Even then the code would need a double
cast since it is also possible that sizeof(void *) > sizeof(phys_addr_t).

Guenter