On Thu, Dec 14, 2017 at 01:55:14PM -0500, J. Bruce Fields wrote:
> So actually what happens is if you do a direct io write where some
> WRITEs succeed and the one fails, then this:
>
> if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
> dreq->flags = 0;
> dreq->error = hdr->error;
> }
>
> clears the NFS_ODIRECT_DO_COMMIT flag, so nfs_direct_write_complete
> never scheduels the commit calls. It looks like that leaves a bunch of
> nfs_pages on some to-be-committed list, so we end up leaking a bunch of
> stuff, with the most visible symptom being an unnecessarily sillyrename
> on close.
>
> I can just remove that clear of dreq_flags and that fixes the problem,
> but I doubt that's correct.
Or, maybe it is. If *any* WRITE(s) involved in this write might need
a commit or reschedule, then surely we should run
nfs_direct_write_schedule_work and let it sort them out? I'm having
trouble seeing why clearing that field partway through could ever be
correct.
Trond, Anna, does the following look right?
--b.
>From a5ad80d2f239ed47eef7ad224db5eb6afba84062 Mon Sep 17 00:00:00 2001
From: "J. Bruce Fields" <bfields@xxxxxxxxxx>
Date: Fri, 15 Dec 2017 15:25:45 -0500
Subject: [PATCH] NFS: commit direct writes even if they fail partially
If some of the WRITE calls making up an O_DIRECT write syscall fail,
we neglect to commit, even if some of the WRITEs succeed.
We also depend on the commit code to free the reference count on the
nfs_page taken in the "if (request_commit)" case at the end of
nfs_direct_write_completion(). The problem was originally noticed
because ENOSPC's encountered partway through a write would result in a
closed file being sillyrenamed when it should have been unlinked.
Signed-off-by: J. Bruce Fields <bfields@xxxxxxxxxx>
---
fs/nfs/direct.c | 4 +---
1 file changed, 1 insertion(+), 3 deletions(-)
diff --git a/fs/nfs/direct.c b/fs/nfs/direct.c
index d2972d537469..8c10b0562e75 100644
--- a/fs/nfs/direct.c
+++ b/fs/nfs/direct.c
@@ -775,10 +775,8 @@ static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
spin_lock(&dreq->lock);
- if (test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
- dreq->flags = 0;
+ if (test_bit(NFS_IOHDR_ERROR, &hdr->flags))
dreq->error = hdr->error;
- }
if (dreq->error == 0) {
nfs_direct_good_bytes(dreq, hdr);
if (nfs_write_need_commit(hdr)) {
--
2.14.3
--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
