Re: [PATCH v3 7/8] ext4: Use rbtrees to manage PAs instead of inode i_prealloc_list

From: Ojaswin Mujoo
Date: Fri Mar 17 2023 - 08:41:23 EST


On Thu, Feb 16, 2023 at 10:07:39AM -0700, Andreas Dilger wrote:
> On Feb 10, 2023, at 7:37 AM, Jan Kara <jack@xxxxxxx> wrote:
> > So I belive mballoc tries to align everything (offsets & lengths)
> > to powers of two to reduce fragmentation and simplify the work for
> > the buddy allocator. If ac->ac_b_ex.fe_len is a power-of-two, the
> > alignment makes sense. But once we had to resort to higher allocator
> > passes and just got some random-length extent, the alignment stops
> > making sense.
>
> In addition to optimizing for the buddy allocator, the other reason that
> the allocations are aligned to power-of-two offsets is to better align
> with underlying RAID stripes. Otherwise, unaligned writes will cause
> parity read-modify-write updates to multiple RAID stripes. This alignment
> can also help (though to a lesser degree) with NAND flash erase blocks.
>
> Cheers, Andreas
>
Got it, thanks. So from my limited understanding of RAID, if the write
is stripe aligned and the (length % stripe == 0) then we won't need a
RMW cycle for parity bits and thats one of the reasons to pay attention
to alignment and length in mballoc code.

Then I think Jan's reasoning still holds that if ac_b_ex.fe_len is already
not of a proper size then we'll anyways be ending with a RMW write in
RAID so no point of paying attention to its alignment, right?

Regards,
ojaswin