[PATCH] NFS: don't raise i_size for DEALLOCATE in _nfs42_proc_fallocate()

[Joseph Qi]

_nfs42_proc_fallocate() raises i_size to offset + len on success, so
that the cached size is not left stale when a delegation blocks the
post-op attribute refresh.  The helper is shared by all three
fallocate modes, but DEALLOCATE always carries FALLOC_FL_KEEP_SIZE
and never extends the file on the server, so punching a hole that
reaches past EOF now wrongly enlarges the cached i_size.

The post-op update would normally restore the size from the returned
attributes, but nfs_update_inode() applies a differing size only when
no delegation is held, and then only when no write is in flight or the
new size is larger:

	if (new_isize != cur_isize && !have_delegation) {
		if (!nfs_have_writebacks(inode) || new_isize > cur_isize)
			i_size_write(inode, new_isize);
	}

With a concurrent mmap writer (nrequests != 0) or any outstanding
delegation, the shrinking server size is therefore ignored and the
client keeps reporting a file larger than the server has, in violation
of KEEP_SIZE, until some later revalidation without writebacks happens
to fix it up.

Skip the i_size bump for the DEALLOCATE procedure.  ALLOCATE (mode 0)
and ZERO_RANGE (which nfs42_fallocate() accepts only without
FALLOC_FL_KEEP_SIZE) do extend the file when the range reaches past
EOF, so they keep the update.

Fixes: 35168eb947f2 ("NFS: fix eof updates after NFSv4.2 fallocate/zero-range")
Cc: stable@xxxxxxxxxxxxxxx
Signed-off-by: Joseph Qi <joseph.qi@xxxxxxxxxxxxxxxxx>
---
 fs/nfs/nfs42proc.c | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)

diff --git a/fs/nfs/nfs42proc.c b/fs/nfs/nfs42proc.c
index ab86246fc364f..914499d30d47e 100644
--- a/fs/nfs/nfs42proc.c
+++ b/fs/nfs/nfs42proc.c
@@ -84,7 +84,16 @@ static int _nfs42_proc_fallocate(struct rpc_message *msg, struct file *filep,
 		loff_t newsize = offset + len;
 
 		spin_lock(&inode->i_lock);
-		if (newsize > i_size_read(inode))
+		/*
+		 * DEALLOCATE always carries FALLOC_FL_KEEP_SIZE and never
+		 * extends the file on the server, so only ALLOCATE and
+		 * ZERO_RANGE may raise i_size here.  A wrongly raised size
+		 * can persist: nfs_update_inode() ignores a shrinking
+		 * post-op size while writebacks are outstanding or a
+		 * delegation is held.
+		 */
+		if (msg->rpc_proc != &nfs4_procedures[NFSPROC4_CLNT_DEALLOCATE] &&
+		    newsize > i_size_read(inode))
 			i_size_write(inode, newsize);
 		nfs_set_cache_invalid(inode, NFS_INO_INVALID_BLOCKS);
 		if (nfs_should_remove_suid(inode))
-- 
2.39.3