nfsd: fix management of listener transports

Currently, when no active threads are running, a root user using nfsdctl
command can try to remove a particular listener from the list of previously
added ones, then start the server by increasing the number of threads,
it leads to the following problem:

[  158.835354] refcount_t: addition on 0; use-after-free.
[  158.835603] WARNING: CPU: 2 PID: 9145 at lib/refcount.c:25 refcount_warn_saturate+0x160/0x1a0
[  158.836017] Modules linked in: rpcrdma rdma_cm iw_cm ib_cm ib_core nfsd auth_rpcgss nfs_acl lockd grace overlay isofs uinput snd_seq_dummy snd_hrtimer nft_fib_inet nft_fib_ipv4 nft_fib_ipv6 nft_fib nft_reject_inet nf_reject_ipv4 nf_reject_ipv6 nft_reject nft_ct nft_chain_nat nf_nat nf_conntrack nf_defrag_ipv6 nf_defrag_ipv4 rfkill ip_set nf_tables qrtr sunrpc vfat fat uvcvideo videobuf2_vmalloc videobuf2_memops uvc videobuf2_v4l2 videodev videobuf2_common snd_hda_codec_generic mc e1000e snd_hda_intel snd_intel_dspcfg snd_hda_codec snd_hda_core snd_hwdep snd_seq snd_seq_device snd_pcm snd_timer snd soundcore sg loop dm_multipath dm_mod nfnetlink vsock_loopback vmw_vsock_virtio_transport_common vmw_vsock_vmci_transport vmw_vmci vsock xfs libcrc32c crct10dif_ce ghash_ce vmwgfx sha2_ce sha256_arm64 sr_mod sha1_ce cdrom nvme drm_client_lib drm_ttm_helper ttm nvme_core drm_kms_helper nvme_auth drm fuse
[  158.840093] CPU: 2 UID: 0 PID: 9145 Comm: nfsd Kdump: loaded Tainted: G    B   W          6.13.0-rc6+ #7
[  158.840624] Tainted: [B]=BAD_PAGE, [W]=WARN
[  158.840802] Hardware name: VMware, Inc. VMware20,1/VBSA, BIOS VMW201.00V.24006586.BA64.2406042154 06/04/2024
[  158.841220] pstate: 61400005 (nZCv daif +PAN -UAO -TCO +DIT -SSBS BTYPE=--)
[  158.841563] pc : refcount_warn_saturate+0x160/0x1a0
[  158.841780] lr : refcount_warn_saturate+0x160/0x1a0
[  158.842000] sp : ffff800089be7d80
[  158.842147] x29: ffff800089be7d80 x28: ffff00008e68c148 x27: ffff00008e68c148
[  158.842492] x26: ffff0002e3b5c000 x25: ffff600011cd1829 x24: ffff00008653c010
[  158.842832] x23: ffff00008653c000 x22: 1fffe00011cd1829 x21: ffff00008653c028
[  158.843175] x20: 0000000000000002 x19: ffff00008653c010 x18: 0000000000000000
[  158.843505] x17: 0000000000000000 x16: 0000000000000000 x15: 0000000000000000
[  158.843836] x14: 0000000000000000 x13: 0000000000000001 x12: ffff600050a26493
[  158.844143] x11: 1fffe00050a26492 x10: ffff600050a26492 x9 : dfff800000000000
[  158.844475] x8 : 00009fffaf5d9b6e x7 : ffff000285132493 x6 : 0000000000000001
[  158.844823] x5 : ffff000285132490 x4 : ffff600050a26493 x3 : ffff8000805e72bc
[  158.845174] x2 : 0000000000000000 x1 : 0000000000000000 x0 : ffff000098588000
[  158.845528] Call trace:
[  158.845658]  refcount_warn_saturate+0x160/0x1a0 (P)
[  158.845894]  svc_recv+0x58c/0x680 [sunrpc]
[  158.846183]  nfsd+0x1fc/0x348 [nfsd]
[  158.846390]  kthread+0x274/0x2f8
[  158.846546]  ret_from_fork+0x10/0x20
[  158.846714] ---[ end trace 0000000000000000 ]---

nfsd_nl_listener_set_doit() would manipulate the list of transports of
server's sv_permsocks and close the specified listener but the other
list of transports (server's sp_xprts list) would not be changed leading
to the problem above.

Instead, determined if the nfsdctl is trying to remove a listener, in
which case, delete all the existing listener transports and re-create
all-but-the-removed ones.

Fixes: 16a4711 ("NFSD: add listener-{set,get} netlink command")
Signed-off-by: Olga Kornievskaia <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Cc: [email protected]
Signed-off-by: Chuck Lever <[email protected]>

Author: Olga Kornievskaia

fs/nfsd/nfsctl.c

@@ -1917,6 +1917,7 @@ int nfsd_nl_listener_set_doit(struct sk_buff *skb, struct genl_info *info)
 	struct svc_serv *serv;
 	LIST_HEAD(permsocks);
 	struct nfsd_net *nn;
+	bool delete = false;
 	int err, rem;
 
 	mutex_lock(&nfsd_mutex);
@@ -1977,34 +1978,28 @@ int nfsd_nl_listener_set_doit(struct sk_buff *skb, struct genl_info *info)
 		}
 	}
 
-	/* For now, no removing old sockets while server is running */
-	if (serv->sv_nrthreads && !list_empty(&permsocks)) {
+	/*
+	 * If there are listener transports remaining on the permsocks list,
+	 * it means we were asked to remove a listener.
+	 */
+	if (!list_empty(&permsocks)) {
 		list_splice_init(&permsocks, &serv->sv_permsocks);
-		spin_unlock_bh(&serv->sv_lock);
-		err = -EBUSY;
-		goto out_unlock_mtx;
+		delete = true;
 	}
+	spin_unlock_bh(&serv->sv_lock);
 
-	/* Close the remaining sockets on the permsocks list */
-	while (!list_empty(&permsocks)) {
-		xprt = list_first_entry(&permsocks, struct svc_xprt, xpt_list);
-		list_move(&xprt->xpt_list, &serv->sv_permsocks);
-
-		/*
-		 * Newly-created sockets are born with the BUSY bit set. Clear
-		 * it if there are no threads, since nothing can pick it up
-		 * in that case.
-		 */
-		if (!serv->sv_nrthreads)
-			clear_bit(XPT_BUSY, &xprt->xpt_flags);
-
-		set_bit(XPT_CLOSE, &xprt->xpt_flags);
-		spin_unlock_bh(&serv->sv_lock);
-		svc_xprt_close(xprt);
-		spin_lock_bh(&serv->sv_lock);
+	/* Do not remove listeners while there are active threads. */
+	if (serv->sv_nrthreads) {
+		err = -EBUSY;
+		goto out_unlock_mtx;
 	}
 
-	spin_unlock_bh(&serv->sv_lock);
+	/*
+	 * Since we can't delete an arbitrary llist entry, destroy the
+	 * remaining listeners and recreate the list.
+	 */
+	if (delete)
+		svc_xprt_destroy_all(serv, net);
 
 	/* walk list of addrs again, open any that still don't exist */
 	nlmsg_for_each_attr(attr, info->nlhdr, GENL_HDRLEN, rem) {
@@ -2031,6 +2026,9 @@ int nfsd_nl_listener_set_doit(struct sk_buff *skb, struct genl_info *info)
 
 		xprt = svc_find_listener(serv, xcl_name, net, sa);
 		if (xprt) {
+			if (delete)
+				WARN_ONCE(1, "Transport type=%s already exists\n",
+					  xcl_name);
 			svc_xprt_put(xprt);
 			continue;
 		}