On Sat, 2022-08-20 at 08:31 +0800, Baoquan He wrote:
> Add hooks arch_ioremap() and arch_iounmap() for s390's special
> operation when ioremap() and iounmap(), then ioremap_[wc|wt]() are
> converted to use ioremap_prot() from GENERIC_IOREMAP.
>
> Signed-off-by: Baoquan He <bhe@xxxxxxxxxx>
> Cc: Heiko Carstens <hca@xxxxxxxxxxxxx>
> Cc: Vasily Gorbik <gor@xxxxxxxxxxxxx>
> Cc: Alexander Gordeev <agordeev@xxxxxxxxxxxxx>
> Cc: Christian Borntraeger <borntraeger@xxxxxxxxxxxxx>
> Cc: Sven Schnelle <svens@xxxxxxxxxxxxx>
> Cc: linux-s390@xxxxxxxxxxxxxxx
> ---
> arch/s390/Kconfig | 1 +
> arch/s390/include/asm/io.h | 26 +++++++++++------
> arch/s390/pci/pci.c | 60 +++++---------------------------------
> 3 files changed, 26 insertions(+), 61 deletions(-)
>
> diff --git a/arch/s390/Kconfig b/arch/s390/Kconfig
> index 318fce77601d..c59e1b25f59d 100644
> --- a/arch/s390/Kconfig
> +++ b/arch/s390/Kconfig
> @@ -135,6 +135,7 @@ config S390
> select GENERIC_SMP_IDLE_THREAD
> select GENERIC_TIME_VSYSCALL
> select GENERIC_VDSO_TIME_NS
> + select GENERIC_IOREMAP
> select HAVE_ALIGNED_STRUCT_PAGE if SLUB
> select HAVE_ARCH_AUDITSYSCALL
> select HAVE_ARCH_JUMP_LABEL
> diff --git a/arch/s390/include/asm/io.h b/arch/s390/include/asm/io.h
> index e3882b012bfa..f837e20b7bbd 100644
> --- a/arch/s390/include/asm/io.h
> +++ b/arch/s390/include/asm/io.h
> @@ -22,11 +22,23 @@ void unxlate_dev_mem_ptr(phys_addr_t phys, void *addr);
>
> #define IO_SPACE_LIMIT 0
>
> -void __iomem *ioremap_prot(phys_addr_t addr, size_t size, unsigned long prot);
> -void __iomem *ioremap(phys_addr_t addr, size_t size);
> -void __iomem *ioremap_wc(phys_addr_t addr, size_t size);
> -void __iomem *ioremap_wt(phys_addr_t addr, size_t size);
> -void iounmap(volatile void __iomem *addr);
> +
Checkpatch nitpick, remove the empty line addition above so as not to
create two consecutive empty lines.
> +/*
> + * I/O memory mapping functions.
> + */
> +void __iomem *
> +arch_ioremap(phys_addr_t *paddr, size_t size, unsigned long *prot_val);
> +#define arch_ioremap arch_ioremap
> +
> +int arch_iounmap(void __iomem *addr);
> +#define arch_iounmap arch_iounmap
> +
> +#define _PAGE_IOREMAP pgprot_val(PAGE_KERNEL)
> +
> +#define ioremap_wc(addr, size) \
> + ioremap_prot((addr), (size), pgprot_val(pgprot_writecombine(PAGE_KERNEL)))
> +#define ioremap_wt(addr, size) \
> + ioremap_prot((addr), (size), pgprot_val(pgprot_writethrough(PAGE_KERNEL)))
>
> static inline void __iomem *ioport_map(unsigned long port, unsigned int nr)
> {
> @@ -51,10 +63,6 @@ static inline void ioport_unmap(void __iomem *p)
> #define pci_iomap_wc pci_iomap_wc
> #define pci_iomap_wc_range pci_iomap_wc_range
>
> -#define ioremap ioremap
> -#define ioremap_wt ioremap_wt
> -#define ioremap_wc ioremap_wc
> -
> #define memcpy_fromio(dst, src, count) zpci_memcpy_fromio(dst, src, count)
> #define memcpy_toio(dst, src, count) zpci_memcpy_toio(dst, src, count)
> #define memset_io(dst, val, count) zpci_memset_io(dst, val, count)
> diff --git a/arch/s390/pci/pci.c b/arch/s390/pci/pci.c
> index 73cdc5539384..984cad9cd5a1 100644
> --- a/arch/s390/pci/pci.c
> +++ b/arch/s390/pci/pci.c
> @@ -244,64 +244,20 @@ void __iowrite64_copy(void __iomem *to, const void *from, size_t count)
> zpci_memcpy_toio(to, from, count);
> }
>
> -static void __iomem *__ioremap(phys_addr_t addr, size_t size, pgprot_t prot)
> +void __iomem *
> +arch_ioremap(phys_addr_t *paddr, size_t size, unsigned long *prot_val)
> {
> - unsigned long offset, vaddr;
> - struct vm_struct *area;
> - phys_addr_t last_addr;
> -
> - last_addr = addr + size - 1;
> - if (!size || last_addr < addr)
> - return NULL;
> -
> if (!static_branch_unlikely(&have_mio))
> - return (void __iomem *) addr;
> -
> - offset = addr & ~PAGE_MASK;
> - addr &= PAGE_MASK;
> - size = PAGE_ALIGN(size + offset);
> - area = get_vm_area(size, VM_IOREMAP);
> - if (!area)
> - return NULL;
> -
> - vaddr = (unsigned long) area->addr;
> - if (ioremap_page_range(vaddr, vaddr + size, addr, prot)) {
> - free_vm_area(area);
> - return NULL;
> - }
> - return (void __iomem *) ((unsigned long) area->addr + offset);
> -}
> -
> -void __iomem *ioremap_prot(phys_addr_t addr, size_t size, unsigned long prot)
> -{
> - return __ioremap(addr, size, __pgprot(prot));
> + return (void __iomem *) *paddr;
Another checkpatch nitpick no space after the cast.
> + return NULL;
> }
> -EXPORT_SYMBOL(ioremap_prot);
>
> -void __iomem *ioremap(phys_addr_t addr, size_t size)
> +int arch_iounmap(void __iomem *addr)
> {
> - return __ioremap(addr, size, PAGE_KERNEL);
> -}
> -EXPORT_SYMBOL(ioremap);
> -
> -void __iomem *ioremap_wc(phys_addr_t addr, size_t size)
> -{
> - return __ioremap(addr, size, pgprot_writecombine(PAGE_KERNEL));
> -}
> -EXPORT_SYMBOL(ioremap_wc);
> -
> -void __iomem *ioremap_wt(phys_addr_t addr, size_t size)
> -{
> - return __ioremap(addr, size, pgprot_writethrough(PAGE_KERNEL));
> -}
> -EXPORT_SYMBOL(ioremap_wt);
> -
> -void iounmap(volatile void __iomem *addr)
> -{
> - if (static_branch_likely(&have_mio))
> - vunmap((__force void *) ((unsigned long) addr & PAGE_MASK));
> + if (!static_branch_likely(&have_mio))
> + return -EINVAL;
As Christoph suggested this might be a good opportunity to add a
comment for this branch.
One other nitpick. The return value doesn't really matter here since
anything != NULL turns iounmap() into a no-op so this looks correct but
semantically I think returning -EINVAL wrongly suggests that addr was
invalid. Maybe -ENXIO would be better at conveying that there is
nothing to unmap.
Looking at your patch 1 another idea would be to have 3 kinds of return
values for arch_iounmap() too e.g.:
arch_iounmap() return an __iomem pointer
- IS_ERR means skip vunmap and return directly
- NULL means continue to vunmap
- a non-NULL, non-IS_ERR pointer has been unmapped successfully
Then we would simply return addr in case of
!static_branch_likely(&have_mio) and NULL otherwise.
What do you think? Either way no strong opinion on my side,
functionally it makes no difference.
> + return 0;
> }
> -EXPORT_SYMBOL(iounmap);
>
> /* Create a virtual mapping cookie for a PCI BAR */
> static void __iomem *pci_iomap_range_fh(struct pci_dev *pdev, int bar,
Apart from the above nitpicks and suggestion this looks good to me.
I did also test this with and without PCI MIO support including use of
PCI MIO instructions in user-space (added in rdma-core v40).
So feel free to add:
Tested-by: Niklas Schnelle <schnelle@xxxxxxxxxxxxx>
Acked-by: Niklas Schnelle <schnelle@xxxxxxxxxxxxx>
Since it looks like there will be a v3 due to other comments anyway
please Cc me on that directly and I'm sure I can upgrade the Acked-by
to Reviewed-by when we're closer to the final code.
