On Monday, 6 June 2022 10:27:20 PM AEST Bharat Kumar Gogada wrote:
> Hi All,
>
> In 5.4 kernel hmm_range structure has pfn_shift parameter.
> This parameter is removed in the latest kernel.
>
> Can any one help me understand usage of this parameter ?
I think originally it was intended to help hmm_range_fault() format PFNs into
some kind of HW specific value, but it turned out not to be very useful - see
2733ea144dcc ("2733ea144dcce789de20988c1056e228a07b1bff") for more background.
I think just setting it to PAGE_SHIFT would be fine.
> In 5.4 kernel, if this parameter set to PAGE_SHIFT, does it mean HMM
framework
> expects user buffers to be 4K aligned, i.e which are allocated via mmap/
aligned_alloc ?
It doesn't matter what the pfn_shift parameter is set to as HMM just copies
(and reformats) CPU page tables so minimum alignment would be PAGE_SIZE (ie.
generally 4K). HMM doesn't add any extra alignment constraints.
> Does HMM have any alignment requirements in 5.4 or current kernel ? Will it
work
> with buffers allocated via malloc ?
Yes. It should work with any normal page with present PTEs although I haven't
personally tested it with v5.4 in a long time. There's been a number of
changes in this area so I'd recommend trying to use a more recent kernel if
possible.
- Alistair
> Regards,
> bharat
>
>
>
>
>
