Theodore Ts'o <[email protected]> 于2023年8月18日周五 01:31写道：
>
> On Mon, Aug 07, 2023 at 09:26:54AM +0800, zhangshida wrote:
> > From: Shida Zhang <[email protected]>
> >
> > Following Andreas' suggestion, it is time to adapt these helpers
> > to handle larger records during runtime, especially in preparation
> > for the eventual support of ext4 with a block size greater than
> > PAGE_SIZE.
>
> Is there a reason for landing this now? We don't have support for
> block_size > PAGE_SIZE yet, and this patch doesn't come for free, at
> least not systems with page_size < 64k. These inline functions are
> *very* hot and get used in a large number of places. Have you looked
> to see what it might do to text size of the ext4 code? And whether
> the expansion to the icache might actually impact performance on CPU
> bound workloads with very large directories?
>
> I will note that there are some opportunities to optimize how often we
> use ext4_rec_len_from_disk. For example, it gets called from
> ext4_check_dir_entry(), and often the callers of that function will
> need the directory record length. So having ext4_check_dir_entry()
> optionally fill in the rec_len via a passed-in pointer might be
> worthwhile.
Yep, the best way to do it is to leave it unmerged until it is necessary.
At the same time, I will try to eliminate these regression concerns based
on these suggestions.
Cheers,
Shida
>
> Cheers,
>
> - Ted