On Jan 27, 2023, at 10:30 AM, Jeff Layton [off-list ref] wrote:

On Fri, 2023-01-27 at 15:21 +0000, Chuck Lever III wrote:

quoted

quoted

On Jan 27, 2023, at 7:09 AM, Jeff Layton [off-list ref] wrote:

We had a bug report that xfstest generic/355 was failing on NFSv4.0.
This test sets various combinations of setuid/setgid modes and tests
whether DIO writes will cause them to be stripped.

What I found was that the server did properly strip those bits, but
the client didn't notice because it held a delegation that was not
recalled. The recall didn't occur because the client itself was the
one generating the activity and we avoid recalls in that case.

Clearing setuid bits is an "implicit" activity. The client didn't
specifically request that we do that, so we need the server to issue a
CB_RECALL, or avoid the situation entirely by not issuing a delegation.

The easiest fix here is to simply not give out a delegation if the file
is being opened for write, and the mode has the setuid and/or setgid bit
set. Note that there is a potential race between the mode and lease
being set, so we test for this condition both before and after setting
the lease.

This patch fixes generic/355, generic/683 and generic/684 for me.

generic/355 2s ...  1s

I should note that 355 only fails with vers=4.0. On 4.1+ the client
specifies that it doesn't want a delegation (as this test is doing DIO).

I used a NFSv4.0 mount for the test.

quoted

That's good.

generic/683 2s ... [not run] xfs_io falloc  failed (old kernel/wrong fs?)
generic/684 2s ... [not run] xfs_io fpunch  failed (old kernel/wrong fs?)

What am I doing wrong?

Not sure here. This test requires v4.2, but the client and server should
negotiate that. You might need to run the test by hand and see what it
outputs. i.e.:

   $ sudo ./tests/generic/683

Then, these two failed only for NFSv4.2 and are not run for other
minor versions. For some reason I thought this was an NFSv4.0-only
bug.

quoted

quoted

Reported-by: Boyang Xue <redacted>
Signed-off-by: Jeff Layton <jlayton@kernel.org>
---
fs/nfsd/nfs4state.c | 27 +++++++++++++++++++++++++++
1 file changed, 27 insertions(+)

diff --git a/fs/nfsd/nfs4state.c b/fs/nfsd/nfs4state.c
index e61b878a4b45..ace02fd0d590 100644
--- a/fs/nfsd/nfs4state.c
+++ b/fs/nfsd/nfs4state.c

@@ -5421,6 +5421,23 @@ nfsd4_verify_deleg_dentry(struct nfsd4_open *open, struct nfs4_file *fp,

	return 0;
}

+/*
+ * We avoid breaking delegations held by a client due to its own activity, but
+ * clearing setuid/setgid bits on a write is an implicit activity and the client
+ * may not notice and continue using the old mode. Avoid giving out a delegation
+ * on setuid/setgid files when the client is requesting an open for write.
+ */
+static int
+nfsd4_verify_setuid_write(struct nfsd4_open *open, struct nfsd_file *nf)
+{
+	struct inode *inode = file_inode(nf->nf_file);
+
+	if ((open->op_share_access & NFS4_SHARE_ACCESS_WRITE) &&
+	    (inode->i_mode & (S_ISUID|S_ISGID)))
+		return -EAGAIN;
+	return 0;
+}
+
static struct nfs4_delegation *
nfs4_set_delegation(struct nfsd4_open *open, struct nfs4_ol_stateid *stp,
		    struct svc_fh *parent)

@@ -5454,6 +5471,8 @@ nfs4_set_delegation(struct nfsd4_open *open, struct nfs4_ol_stateid *stp,

	spin_lock(&fp->fi_lock);
	if (nfs4_delegation_exists(clp, fp))
		status = -EAGAIN;
+	else if (nfsd4_verify_setuid_write(open, nf))
+		status = -EAGAIN;
	else if (!fp->fi_deleg_file) {
		fp->fi_deleg_file = nf;
		/* increment early to prevent fi_deleg_file from being

@@ -5494,6 +5513,14 @@ nfs4_set_delegation(struct nfsd4_open *open, struct nfs4_ol_stateid *stp,

	if (status)
		goto out_unlock;

+	/*
+	 * Now that the deleg is set, check again to ensure that nothing
+	 * raced in and changed the mode while we weren't lookng.
+	 */
+	status = nfsd4_verify_setuid_write(open, fp->fi_deleg_file);
+	if (status)
+		goto out_unlock;
+
	spin_lock(&state_lock);
	spin_lock(&fp->fi_lock);
	if (fp->fi_had_conflict)
-- 
2.39.1

--
Chuck Lever

-- 
Jeff Layton [off-list ref]

--
Chuck Lever