cifs: fix data corruption in read after invalidate

When invalidating a file as part of breaking a lease, the folios holding
the file data are disposed of, and truncate calls ->invalidate_folio()
to get rid of them rather than calling ->release_folio().  This means
that the netfs_inode::zero_point value didn't get updated in current
upstream code to reflect the point after which we can assume that the
server will only return zeroes, and future reads will then return blocks
of zeroes if the file got extended for any region beyond the old zero
point.

Fix this by updating zero_point before invalidating the inode in
cifs_revalidate_mapping().

Suggested-by: David Howells <dhowells@redhat.com>
Fixes: 3ee1a1fc3981 ("cifs: Cut over to using netfslib")
Reviewed-by: David Howells <dhowell@redhat.com>
Signed-off-by: Steve French <stfrench@microsoft.com>

diff --git a/fs/smb/client/inode.c b/fs/smb/client/inode.c
index 5d4b0fd3a59e45de4abac84cad9e51d88211fe36..262576573eb5182b8a1729677e1b98e7bdee66a3 100644 (file)
--- a/fs/smb/client/inode.c
+++ b/fs/smb/client/inode.c
@@ -2465,7 +2465,8 @@ int
 cifs_revalidate_mapping(struct inode *inode)
 {
        int rc;
-       unsigned long *flags = &CIFS_I(inode)->flags;
+       struct cifsInodeInfo *cifs_inode = CIFS_I(inode);
+       unsigned long *flags = &cifs_inode->flags;
        struct cifs_sb_info *cifs_sb = CIFS_SB(inode->i_sb);
 
        /* swapfiles are not supposed to be shared */
@@ -2482,6 +2483,7 @@ cifs_revalidate_mapping(struct inode *inode)
                if (cifs_sb->mnt_cifs_flags & CIFS_MOUNT_RW_CACHE)
                        goto skip_invalidate;
 
+               cifs_inode->netfs.zero_point = cifs_inode->netfs.remote_i_size;
                rc = filemap_invalidate_inode(inode, true, 0, LLONG_MAX);
                if (rc) {
                        cifs_dbg(VFS, "%s: invalidate inode %p failed with rc %d\n",