These fixes fix two regressions reported by Red Hat QE. This first
issue reported was that writing a file larger than the available space
would result in a client hang instead of returning -ENOSPC. The second
issue reported was that the client was returning -EIO instead of -EDQUOT
when a quota is exceeded.
On further investigation, I found that in the first issue the client
wasn't really hung. Instead it was performing all of the requested
writes before it would eventually return -ENOSPC. The dd command was
writing 400TB, which would have taken a few weeks to complete.
Adjusting the 'bs' and 'count' arguments so that the resulting file
would be much smaller (but still enough to fill up the space on the NFS
server) showed that the dd command would indeed complete with -ENOSPC.
But on older kernels even the 400TB dd would return -ENOSPC quickly.
In the second issue, I found that adding 'conv=fsync' would make the dd
command return -EDQUOT. So what was happening with the original
command was that it was doing a close() without first calling fsync()
and the close() was returning -EIO.
I bisected both of these down to commit aded8d7b54f2 ("NFS: Don't
inadvertently clear writeback errors"). HOWEVER, as a test I reverted
that commit and it wasn't sufficient to bring back the old behavior. I
turns out that was due to commit 6fbda89b257f ("NFS: Replace custom
error reporting mechanism with generic one"). This is why I listed the
second commit in the 'Fixes:' tag of both of my patches.
The first patch fixes the -EDQUOT issue and the second one fixes the
-ENOSPC issue.
-Scott
v2:
- Use filemap_sample_wb_err() & filemap_check_wb_err() instead of
file_check_and_advance_wb_err() so that the error is not consumed.
Scott Mayhew (2):
nfs: ensure correct writeback errors are returned on close()
nfs: nfs_file_write() should check for writeback errors
fs/nfs/file.c | 17 +++++++++++++----
fs/nfs/nfs4file.c | 5 ++++-
2 files changed, 17 insertions(+), 5 deletions(-)
--
2.25.4
The NFS_CONTEXT_ERROR_WRITE flag (as well as the check of said flag) was
removed by commit 6fbda89b257f. The absence of an error check allows
writes to be continually queued up for a server that may no longer be
able to handle them. Fix it by adding an error check using the generic
error reporting functions.
Fixes: 6fbda89b257f ("NFS: Replace custom error reporting mechanism with
generic one")
Signed-off-by: Scott Mayhew <[email protected]>
---
fs/nfs/file.c | 12 +++++++++---
1 file changed, 9 insertions(+), 3 deletions(-)
diff --git a/fs/nfs/file.c b/fs/nfs/file.c
index d72496efa17b..63940a7a70be 100644
--- a/fs/nfs/file.c
+++ b/fs/nfs/file.c
@@ -590,12 +590,14 @@ static const struct vm_operations_struct nfs_file_vm_ops = {
.page_mkwrite = nfs_vm_page_mkwrite,
};
-static int nfs_need_check_write(struct file *filp, struct inode *inode)
+static int nfs_need_check_write(struct file *filp, struct inode *inode,
+ int error)
{
struct nfs_open_context *ctx;
ctx = nfs_file_open_context(filp);
- if (nfs_ctx_key_to_expire(ctx, inode))
+ if (nfs_error_is_fatal_on_server(error) ||
+ nfs_ctx_key_to_expire(ctx, inode))
return 1;
return 0;
}
@@ -606,6 +608,8 @@ ssize_t nfs_file_write(struct kiocb *iocb, struct iov_iter *from)
struct inode *inode = file_inode(file);
unsigned long written = 0;
ssize_t result;
+ errseq_t since;
+ int error;
result = nfs_key_timeout_notify(file, inode);
if (result)
@@ -630,6 +634,7 @@ ssize_t nfs_file_write(struct kiocb *iocb, struct iov_iter *from)
if (iocb->ki_pos > i_size_read(inode))
nfs_revalidate_mapping(inode, file->f_mapping);
+ since = filemap_sample_wb_err(file->f_mapping);
nfs_start_io_write(inode);
result = generic_write_checks(iocb, from);
if (result > 0) {
@@ -648,7 +653,8 @@ ssize_t nfs_file_write(struct kiocb *iocb, struct iov_iter *from)
goto out;
/* Return error values */
- if (nfs_need_check_write(file, inode)) {
+ error = filemap_check_wb_err(file->f_mapping, since);
+ if (nfs_need_check_write(file, inode, error)) {
int err = nfs_wb_all(inode);
if (err < 0)
result = err;
--
2.25.4