On Sat, Feb 10, 2024, at 20:38, Linus Walleij wrote:
> On Sat, Feb 10, 2024 at 12:58 PM Arnd Bergmann <arnd@xxxxxxxx> wrote:
>> On Sat, Feb 10, 2024, at 00:41, Linus Walleij wrote:
>
>> > The only difference is where the CPU have to read/write the
>> > buffers really, before the change those were all guaranteed to
>> > be in lowmem (bounced there by the block core), now they can
>> > also be in highmem, but sg_miter will deal with it for sure.
>>
>> Yes, that was my point: The sg_miter() code is meant to
>> handle exactly this case with highmem data, but as far
>> as I can tell, that code path has never been tested on
>> 32-bit systems with highmem but without BLK_BOUNCE_HIGH.
>
> It's actually possible to enforce testing of highmem scatterlists
> to an MMC card (one need to be careful as this is destructive
> testing!)
> drivers/mmc/core/mmc_test.c
>
> ...but the one relevant target I have is a Kirkwood and it only
> has 128 MB of memory so highmem won't be exercised.
I think you can pass a vmalloc= command line option to the
kernel that will increase the size of the vmalloc are at
the expense of lowmem and give you some highmem instead.
Arnd
