Re: xfs_alloc_ag_vextent_near() takes minutes to complete

[Christoph Hellwig <hch@infradead.org>]

On Sun, May 07, 2017 at 11:00:40AM +0300, Alex Lyakas wrote:
> We are using the sync mount option, because the XFS instances that we
> have are exposed via nfsd. Without the "sync" mount option, data from
> NFS write command would end up sitting in page cache, and the nfs
> client would not know when it has been finally persisted on disk.
> Is the any other way that you can recommend to provide data integrity
> guarantee for nfs clients? Even if it requires some development, like
> placing the data in a short-term persistent space, this is something
> we will look at.

You don't need any mount option for NFS exports.  Only NFSv2 requires
symchronous writes, while NFS3+ have the concept of unstable writes
that the client needs to explicitly committ using the COMMIT on the
write operation.  Both the NFSv2 synchronous and the NFSv3+ unstable
semantics are managed by NFSD.

(I'll speak here as a XFS and NFSD developer, and someone involved in
NFS protocol development, just in case you're doubting)

> 
> Thanks,
> Alex.
---end quoted text---