On Fri, 2023-04-07 at 16:48 -0400, Anna Schumaker wrote:
> From: Anna Schumaker <Anna.Schumaker@xxxxxxxxxx>
>
> Instead of using a tiny scratch buffer, we should use a full scratch
> page to match how other NFSv4 operations handle scratch data. This
> patch
> also lets us set the scratch page before decoding any part of the
> READ_PLUS operation instead of setting the buffer right before
> segment
> decoding, which feels a little more robust to me.
>
> Signed-off-by: Anna Schumaker <Anna.Schumaker@xxxxxxxxxx>
> ---
> fs/nfs/nfs42xdr.c | 4 ++--
> fs/nfs/nfs4proc.c | 17 ++++++++++++-----
> include/linux/nfs_xdr.h | 1 +
> 3 files changed, 15 insertions(+), 7 deletions(-)
>
> diff --git a/fs/nfs/nfs42xdr.c b/fs/nfs/nfs42xdr.c
> index d80ee88ca996..702567d5b1db 100644
> --- a/fs/nfs/nfs42xdr.c
> +++ b/fs/nfs/nfs42xdr.c
> @@ -1122,7 +1122,6 @@ static int decode_read_plus(struct xdr_stream
> *xdr, struct nfs_pgio_res *res)
> uint32_t segments;
> struct read_plus_segment *segs;
> int status, i;
> - char scratch_buf[16];
> __be32 *p;
>
> status = decode_op_hdr(xdr, OP_READ_PLUS);
> @@ -1143,7 +1142,6 @@ static int decode_read_plus(struct xdr_stream
> *xdr, struct nfs_pgio_res *res)
> if (!segs)
> return -ENOMEM;
>
> - xdr_set_scratch_buffer(xdr, &scratch_buf,
> sizeof(scratch_buf));
> status = -EIO;
> for (i = 0; i < segments; i++) {
> status = decode_read_plus_segment(xdr, &segs[i]);
> @@ -1348,6 +1346,8 @@ static int nfs4_xdr_dec_read_plus(struct
> rpc_rqst *rqstp,
> struct compound_hdr hdr;
> int status;
>
> + xdr_set_scratch_page(xdr, res->scratch);
> +
> status = decode_compound_hdr(xdr, &hdr);
> if (status)
> goto out;
> diff --git a/fs/nfs/nfs4proc.c b/fs/nfs/nfs4proc.c
> index 5607b1e2b821..2b5e62713bdd 100644
> --- a/fs/nfs/nfs4proc.c
> +++ b/fs/nfs/nfs4proc.c
> @@ -5439,6 +5439,8 @@ static bool nfs4_read_plus_not_supported(struct
> rpc_task *task,
>
> static int nfs4_read_done(struct rpc_task *task, struct
> nfs_pgio_header *hdr)
> {
> + if (hdr->res.scratch)
> + __free_page(hdr->res.scratch);
> if (!nfs4_sequence_done(task, &hdr->res.seq_res))
> return -EAGAIN;
> if (nfs4_read_stateid_changed(task, &hdr->args))
> @@ -5452,17 +5454,22 @@ static int nfs4_read_done(struct rpc_task
> *task, struct nfs_pgio_header *hdr)
> }
>
> #if defined CONFIG_NFS_V4_2 && defined CONFIG_NFS_V4_2_READ_PLUS
> -static void nfs42_read_plus_support(struct nfs_pgio_header *hdr,
> +static bool nfs42_read_plus_support(struct nfs_pgio_header *hdr,
> struct rpc_message *msg)
> {
> /* Note: We don't use READ_PLUS with pNFS yet */
> - if (nfs_server_capable(hdr->inode, NFS_CAP_READ_PLUS) &&
> !hdr->ds_clp)
> + if (nfs_server_capable(hdr->inode, NFS_CAP_READ_PLUS) &&
> !hdr->ds_clp) {
> msg->rpc_proc =
> &nfs4_procedures[NFSPROC4_CLNT_READ_PLUS];
> + hdr->res.scratch = alloc_page(GFP_KERNEL);
> + return hdr->res.scratch != NULL;
Hmm... Does it make sense to do this if we're only requesting a read of
a few bytes? Perhaps we should look at a minimum size when deciding
whether or not to do read plus?
> + }
> + return false;
> }
> #else
> -static void nfs42_read_plus_support(struct nfs_pgio_header *hdr,
> +static bool nfs42_read_plus_support(struct nfs_pgio_header *hdr,
> struct rpc_message *msg)
> {
> + return false;
> }
> #endif /* CONFIG_NFS_V4_2 */
>
> @@ -5472,8 +5479,8 @@ static void nfs4_proc_read_setup(struct
> nfs_pgio_header *hdr,
> hdr->timestamp = jiffies;
> if (!hdr->pgio_done_cb)
> hdr->pgio_done_cb = nfs4_read_done_cb;
> - msg->rpc_proc = &nfs4_procedures[NFSPROC4_CLNT_READ];
> - nfs42_read_plus_support(hdr, msg);
> + if (!nfs42_read_plus_support(hdr, msg))
> + msg->rpc_proc = &nfs4_procedures[NFSPROC4_CLNT_READ];
> nfs4_init_sequence(&hdr->args.seq_args, &hdr->res.seq_res, 0,
> 0);
> }
>
> diff --git a/include/linux/nfs_xdr.h b/include/linux/nfs_xdr.h
> index e86cf6642d21..6d821aaf0b1a 100644
> --- a/include/linux/nfs_xdr.h
> +++ b/include/linux/nfs_xdr.h
> @@ -670,6 +670,7 @@ struct nfs_pgio_res {
> struct {
> unsigned int replen; /*
> used by read */
> int eof; /*
> used by read */
> + struct page *scratch; /*
> used by read */
> };
> struct {
> struct nfs_writeverf * verf; /*
> used by write */
--
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trond.myklebust@xxxxxxxxxxxxxxx
