On Wed, Feb 7, 2024 at 3:42 PM Trond Myklebust [off-list ref] wrote:

On Wed, 2024-02-07 at 14:58 -0500, Olga Kornievskaia wrote:

quoted

On Wed, Feb 7, 2024 at 2:51 PM Olga Kornievskaia
[off-list ref] wrote:

quoted

On Wed, Feb 7, 2024 at 2:12 PM Trond Myklebust
[off-list ref] wrote:

quoted

On Wed, 2024-02-07 at 13:29 -0500, Olga Kornievskaia wrote:

quoted

From: Olga Kornievskaia <redacted>

Currently, if the server returns a partial layout, the client
gets
stuck asking for a layout indefinitely. Until we add support
for
partial layouts, treat partial layout as layout unavailable
error.

Signed-off-by: Olga Kornievskaia <redacted>
---
 fs/nfs/nfs4proc.c | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/fs/nfs/nfs4proc.c b/fs/nfs/nfs4proc.c
index dae4c1b6cc1c..108bc7f3e8c2 100644
--- a/fs/nfs/nfs4proc.c
+++ b/fs/nfs/nfs4proc.c

@@ -9790,6 +9790,12 @@ nfs4_proc_layoutget(struct

nfs4_layoutget
*lgp,
      if (status != 0)
              goto out;

+     /* Since client does not support partial pnfs layout,
then
treat
+      * getting a partial layout as LAYOUTUNAVAILABLE error
+      */
+     if (lgp->args.range.length != lgp->res.range.length)
+             task->tk_status = -NFS4ERR_LAYOUTUNAVAILABLE;

I think this case is better handled by allowing the caller to set
lgp-

quoted

args.minlength to an appropriate minimum value.

I do not understand what this suggestion means. What I can think of
is
that the caller would set an appropriate minimum value and the code
here would check that the result is at least as large?

A follow up question on a "minimum value". It seems that since the
client would then need to set it to the same value as the "length"
(ie
whole file layout value), yes? So it shifts the responsibility to the
server, disallowing it from returning a partial layout.

My expectation is that we use 'minimum length' to mean the length that
is the smallest value that is acceptable to the client (i.e. the
"loga_minlength" as described in RFC5661 Section 18.43). If the client
cannot handle a layout that is smaller than a whole file layout, then
that's the value we should set for loga_minlength. The server then gets
to decide if it can honour the request, or should return
NFS4ERR_LAYOUTTRYLATER.

The other value is the desired length (i.e. the "loga_length" as
described in RFC5661, Section 18.43). Currently, in the flexfiles
client, we set that to the length of the I/O request that we're trying
to obey. If the server can meet or exceed that value, it will still do
so, but it doesn't need to return an error if it cannot meet that value
provided that it can still return a layout with a length that meets the
"loga_minlength" requirement.

I think I'm misunderstanding a partial layout and the client's ability
to support them. Is partial layout support something that generic pnfs
code does or is that something that each layout driver supposed to
support?  I was under the impression that it is or will be generic.
I'm specifically confused because in the previous email you said, the
minimal (layout) length is set to a PAGE (or I/O request size) and
that each driver is supposed to be able to handle that. That seems to
imply to me that if the server returned back a layout of 1 page
(which is a partial layout) then the layoutdriver should be OK with it
(mean handle partial layouts) (which filalyout driver does not, it
needs a whole file layout), is that correct?

--
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trond.myklebust@hammerspace.com