Re: [PATCH RFC 1/1] ocfs2: fix write IO performance improvement for high fragmentation

[Heming Zhao]

Hello Gautham,

Thanks for locating the issue and submitting a patch.
Is it possible to share your test case for this bug?

The key of this bug is ocfs2_cluster_group_search() comparing with wrong
bits size. I have another fix for this bug and will send it to this mailing
list later.

Thanks,
Heming

On 3/24/25 13:48, Gautham Ananthakrishna wrote:
> The commit 4eb7b93e03101fd3f35e69affe566e4b1e3e3dca caused a regression
> in our test suite in discontig extent tests. Upon troubleshooting I found
> The following issues.
>
> 1. The function ocfs2_cluster_group_search() was called for discontig allocations
> as well. But it checks only the contiguous bits 'bg_contig_free_bits'.
> It hit the ENOSPC in the following case in one of the tests.
>
> ocfs2_mkdir()
>   ocfs2_reserve_new_inode()
>    ocfs2_reserve_suballoc_bits()
>     ocfs2_block_group_alloc()
>      ocfs2_block_group_alloc_discontig()
>       __ocfs2_claim_clusters()
>        ocfs2_claim_suballoc_bits()
>         ocfs2_search_chain()
>          ocfs2_cluster_group_search()
>
> Looked like the commit did not consider discontig searches. To fix this,
> I have split ocfs2_cluster_group_search() into *_common(), *_contig() and
> *_discontig()
>
> 2. That commit enforced ocfs2_cluster_group_search() to search only the
> 'bits_wanted' number of bits whereas ocfs2_block_group_find_clear_bits()
> fills the best available size and the function ocfs2_cluster_group_search()
> itself is supposed to search 'min_bits' at the minimum and need not be
> 'bits_wanted' always.
>
> Fixed the above issues in this patch.
> This patch fixes 4eb7b93e03101fd3f35e69affe566e4b1e3e3dca
>
> Signed-off-by: Gautham Ananthakrishna <gautham.ananthakrishna@xxxxxxxxxx>
> ---
>   fs/ocfs2/suballoc.c | 146 ++++++++++++++++++++++++++++----------------
>   1 file changed, 95 insertions(+), 51 deletions(-)