Le 04/12/2021 à 15:10, Maxime Bizon a écrit :
>
> On Saturday 04 Dec 2021 à 10:01:07 (+0000), Christophe Leroy wrote:
>
>> In fact BAT4 is wrong. Both virtual and physical address of a 64M BAT
>> must be 64M aligned. I think the display is wrong as well (You took it
>
> oh so hardware does simple bitmask after all
>
> I got fooled by the lack of guard in the bat setup code, so I assumed
> magical hardware
I guess all the guard is in the comment ...
/*
* Set up one of the I/D BAT (block address translation) register pairs.
* The parameters are not checked; in particular size must be a power
* of 2 between 128k and 256M.
*/
void __init setbat(int index, unsigned long virt, phys_addr_t phys,
unsigned int size, pgprot_t prot)
>
>> from ptdump ?), BEPI and BRPN must be anded with complement of BL.
>
> yes that was ptdump code with seq_printf replaced by printk
>
> ptdump code is correct but iif the bat addresses are correctly
> aligned, maybe add a safeguard like this ?
>
> index 85062ce2d849..f7c5cf62ef41 100644
> --- a/arch/powerpc/mm/book3s32/mmu.c
> +++ b/arch/powerpc/mm/book3s32/mmu.c
> @@ -275,6 +279,10 @@ void __init setbat(int index, unsigned long virt,
> phys_addr_t phys,
> (unsigned long long)phys);
> return;
> }
> +
> + WARN_ON(!is_power_of_2(size));
> + WARN_ON((phys & (size - 1)));
> + WARN_ON((virt & (size - 1)));
> bat = BATS[index];
>
Yes we could add some check allthough I'd go for a 'pr_err()' like when
no BAT is available.
>
>> So here your 64M BAT maps 0xf8000000-0xfbffffff, therefore the address
>> 0xfd3fce00 is not mapped by any BAT hence the OOPS.
>
> ok I think I found the issue:
>
> diff --git a/arch/powerpc/mm/kasan/book3s_32.c
> b/arch/powerpc/mm/kasan/book3s_32.c
> index 35b287b0a8da..fcbb9a136c1a 100644
> --- a/arch/powerpc/mm/kasan/book3s_32.c
> +++ b/arch/powerpc/mm/kasan/book3s_32.c
> @@ -12,14 +12,14 @@ int __init kasan_init_region(void *start, size_t size)
> unsigned long k_end = (unsigned long)kasan_mem_to_shadow(start +
> size);
> unsigned long k_cur = k_start;
> int k_size = k_end - k_start;
> - int k_size_base = 1 << (ffs(k_size) - 1);
> + int k_size_base = 1 << (fls(k_size) - 1);
> int ret;
> void *block;
>
> block = memblock_alloc(k_size, k_size_base);
>
> if (block && k_size_base >= SZ_128K && k_start == ALIGN(k_start,
> k_size_base)) {
> - int shift = ffs(k_size - k_size_base);
> + int shift = fls(k_size - k_size_base);
> int k_size_more = shift ? 1 << (shift - 1) : 0;
>
> setbat(-1, k_start, __pa(block), k_size_base, PAGE_KERNEL);
>
>
>
Not sure it is that simple.
I'm cooking a patch reusing the block_size() function in mm/book3s32/mmu.c
Christophe
