On Fri, 2021-05-14 at 11:58 +0800, Nick Huang wrote:
> From: Yu Hsiang Huang <nickhuang@xxxxxxxxxxxx>
>
> Truncation of an unlinked inode may take a long time for I/O waiting,
> and
> it doesn't have to prevent access to the directory. Thus, let
> truncation
> occur outside the directory's mutex, just like do_unlinkat() does.
>
> Signed-off-by: Yu Hsiang Huang <nickhuang@xxxxxxxxxxxx>
> Signed-off-by: Bing Jing Chang <bingjingc@xxxxxxxxxxxx>
> Signed-off-by: Robbie Ko <robbieko@xxxxxxxxxxxx>
> ---
> fs/nfsd/vfs.c | 5 +++++
> 1 file changed, 5 insertions(+)
>
> diff --git a/fs/nfsd/vfs.c b/fs/nfsd/vfs.c
> index 15adf1f6ab21..39948f130712 100644
> --- a/fs/nfsd/vfs.c
> +++ b/fs/nfsd/vfs.c
> @@ -1859,6 +1859,7 @@ nfsd_unlink(struct svc_rqst *rqstp, struct
> svc_fh *fhp, int type,
> {
> struct dentry *dentry, *rdentry;
> struct inode *dirp;
> + struct inode *rinode;
> __be32 err;
> int host_err;
>
> @@ -1887,6 +1888,8 @@ nfsd_unlink(struct svc_rqst *rqstp, struct
> svc_fh *fhp, int type,
> host_err = -ENOENT;
> goto out_drop_write;
> }
> + rinode = d_inode(rdentry);
> + ihold(rinode);
>
> if (!type)
> type = d_inode(rdentry)->i_mode & S_IFMT;
> @@ -1902,6 +1905,8 @@ nfsd_unlink(struct svc_rqst *rqstp, struct
> svc_fh *fhp, int type,
> if (!host_err)
> host_err = commit_metadata(fhp);
Why leave the commit_metadata() call under the lock? If you're
concerned about latency, then it makes more sense to call fh_unlock()
before flushing those metadata updates to disk.
This is, BTW, an optimisation that appears to be possible in several
other cases in fs/nfsd/vfs.c.
> dput(rdentry);
> + fh_unlock(fhp);
> + iput(rinode); /* truncate the inode here */
>
> out_drop_write:
> fh_drop_write(fhp);
--
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trond.myklebust@xxxxxxxxxxxxxxx
