[PATCH net v5 0/2] tun: bound the NAPI receive queue

[PATCH net v5 0/2] tun: bound the NAPI receive queue

[Antonio Quartulli]

With IFF_NAPI, tun queues written packets for NAPI on an unbounded
list, which an unprivileged user can grow until the system runs out of
memory. Patch 1 bounds it by netdev_max_backlog; patch 2 adds a
kselftest for the vhost-net paths into that queue.

v5:
- Take at most budget packets in the NAPI poll instead of keeping the
  leftovers on a separate queue (Willem)
- Keep only the lockless check, before allocating, also in
  tun_xdp_one() (Willem)
- Say that an OOM is still possible with large packets and many queues,
  if hard to reach in practice (Sashiko)
- selftest: little-endian ring, buffers returned in any order, lower
  netdev_max_backlog only after setup and restore it first, any non-zero
  experimental_zcopytx, .gitignore entry (Sashiko)
- Rebase on net, which conflicted in selftests/net/config

v4: https://lore.kernel.org/netdev/20261103142232.1869185-1-antonio@mandelbit.com/
- Keep the poll's leftovers on a separate queue, like process_backlog():
  v3 did not bound the queue when the poll ran on another CPU (Sashiko)
- Schedule NAPI when dropping: with netdev_max_backlog below
  NAPI_POLL_WEIGHT, MSG_MORE senders (vhost-net) stalled the queue
  (Sashiko)
- Use goto free_skb for the locked drop (Willem)
- Add a vhost-net kselftest (patch 2); the tun kselftest passes (Willem)
- Cc vhost, kselftest and XDP maintainers and the Fixes author/acker

v3: https://lore.kernel.org/netdev/20261031231444.1528758-1-antonio@mandelbit.com/
- Also bound the sendmsg() paths, which share the queue: zerocopy
  buffers are completed when dropped early, and the batched XDP buffers
  of tun_xdp_one() are covered too (Willem)
- Drop the IFF_NAPI_FRAGS exclusion (Willem)
- Recheck under the queue lock before queueing (Willem)
- Compare with > like enqueue_to_backlog() (Willem)
- Keep the includes sorted (Willem)
- Add the reproducer and a testing statement to the commit message

v2: https://lore.kernel.org/netdev/20261030152540.1409884-1-antonio@mandelbit.com/
- Cc netdev and the maintainers

Testing, in an x86-64 VM with 1 and 2 vCPUs, on net ead744fd1ce0:
- Reproducer from patch 1, 15 seconds, 500 and 3000 iptables rules,
  1400 to 65000-byte packets: memory use stays flat, no OOM; forwarding
  rate unchanged (152577 packets in 3 s without the patch, 151277 with
  it).
- tun_vhost (patch 2), with vhost_net experimental_zcopytx=0 and =1:
  passes with patch 1. Without it, every variant fails (nothing is
  dropped); removing the tun_xdp_one() check fails the batch variant,
  removing the zerocopy completion fails the zerocopy one, and removing
  the NAPI scheduling on drop fails the copy and zerocopy ones.
- tun kselftest: pass:91 fail:0 xfail:12, as without the patch.

Notes:
- Affected since 943170998b20 (v4.15).
- Mitigation: disallow unprivileged user namespaces or restrict
  access to /dev/net/tun.

Antonio Quartulli (2):
  tun: bound the NAPI receive queue
  selftests: net: add a vhost-net test for the tun NAPI queue bound

 drivers/net/tun.c                       |  59 +++-
 tools/testing/selftests/net/.gitignore  |   1 +
 tools/testing/selftests/net/Makefile    |   1 +
 tools/testing/selftests/net/config      |   1 +
 tools/testing/selftests/net/tun_vhost.c | 409 ++++++++++++++++++++++++
 5 files changed, 463 insertions(+), 8 deletions(-)
 create mode 100644 tools/testing/selftests/net/tun_vhost.c

-- 
2.56.0

[PATCH net v5 1/2] tun: bound the NAPI receive queue

[Antonio Quartulli]

With IFF_NAPI, tun_get_user() queues every written packet for NAPI on
an unbounded list. When the poll runs asynchronously (threaded NAPI)
and is slower than the writer, the queue grows until the system runs
out of memory. An unprivileged user can do this from its own user and
network namespaces, and the queued skbs are not charged to its memcg:
in a test VM, a uid 1000 process limited to 128 MB caused a global OOM
in ~6 seconds. The socket send buffer cannot bound this: tun sets it to
INT_MAX, and the build_skb() and XDP batch paths do not check it.

Drop packets once more than netdev_max_backlog of them are queued, like
enqueue_to_backlog() does, and count them as rx_dropped, before
allocating and copying them. This covers write(), sendmsg() and the
batches of XDP buffers vhost-net passes to tun_sendmsg(), which share
the queue. Zerocopy buffers of dropped packets are completed. The check
is lockless, so concurrent writers can each still add one packet.

The NAPI poll moved the whole queue out, processed up to its budget and
put the rest back: while it ran, the queue looked empty to writers, so
with the poll on another CPU the backlog could still grow without limit.
Take at most budget packets instead, so that the queue never looks
emptier than it is.

Dropping because the queue is full also schedules NAPI. vhost-net sends
with MSG_MORE while it has more packets, and tun_get_user() schedules
NAPI only after the last one or once more than NAPI_POLL_WEIGHT are
queued. With a netdev_max_backlog below that, the queue would not drain
until vhost-net ran out of packets; with zerocopy, vhost-net never
would, as it stops sending when too many queued buffers are pending.

Reproducer, as an unprivileged user in a new user and network
namespace, without kernel modifications or fault injection:
  ip tuntap add t0 mode tun; ip addr add 10.11.0.1/24 dev t0
  ip link add d0 type dummy; ip addr add 10.20.0.1/24 dev d0
  ip link set t0 up; ip link set d0 up; sysctl net.ipv4.ip_forward=1
  install ~500 iptables FORWARD rules and a MASQUERADE rule on d0
  open t0 with IFF_TUN | IFF_NO_PI | IFF_NAPI
  echo 1 > /sys/class/net/t0/threaded
  write() 1400-byte IPv4 packets for 10.20.0.2 in a loop

Tested in an x86-64 VM with 1 and 2 vCPUs: with the reproducer running
for 15 seconds, with 1400 to 65000-byte packets, memory use stays flat
and the forwarding rate is unchanged; the tun kselftest passes, as it
does without the patch; and the tun_vhost kselftest added by the next
patch, which drives the vhost-net batch, copy and zerocopy paths,
passes, while it fails without the patch.

Still open: the limit counts packets rather than bytes and applies per
queue, so with large packets and many queues (IFF_MULTI_QUEUE works in a
user namespace too) the queued skbs, which are not charged to the
writer's memcg, can still add up to an OOM. This is hard to reach in
practice: with large packets the writer's copy is slower than the NAPI
poll, so the queue does not fill. With 64 KB writes the reproducer never
reached the limit and memory use stayed flat.

Found while analysing the performance of the tun write path, and tested,
with the help of an AI coding assistant (Claude Code).

Fixes: 943170998b20 ("tun: enable NAPI for TUN/TAP driver")
Signed-off-by: Antonio Quartulli <antonio@mandelbit.com>
---
 drivers/net/tun.c | 59 ++++++++++++++++++++++++++++++++++++++++-------
 1 file changed, 51 insertions(+), 8 deletions(-)

diff --git a/drivers/net/tun.c b/drivers/net/tun.c
index dcec7311e601..396e4295589f 100644
--- a/drivers/net/tun.c
+++ b/drivers/net/tun.c
@@ -58,6 +58,7 @@
 #include <linux/nsproxy.h>
 #include <linux/virtio_net.h>
 #include <linux/rcupdate.h>
+#include <net/hotdata.h>
 #include <net/net_namespace.h>
 #include <net/netns/generic.h>
 #include <net/rtnetlink.h>
@@ -227,6 +228,16 @@ struct veth {
 static void tun_flow_init(struct tun_struct *tun);
 static void tun_flow_uninit(struct tun_struct *tun);
 
+/* In IFF_NAPI mode written packets wait in the tfile's sk_write_queue for the
+ * NAPI poll, which may run later in a kthread or ksoftirqd. Bound that queue
+ * like the per-CPU backlog of netif_rx().
+ */
+static bool tun_napi_queue_full(const struct sk_buff_head *queue)
+{
+	return skb_queue_len_lockless(queue) >
+	       READ_ONCE(net_hotdata.max_backlog);
+}
+
 static int tun_napi_receive(struct napi_struct *napi, int budget)
 {
 	struct tun_file *tfile = container_of(napi, struct tun_file, napi);
@@ -237,21 +248,24 @@ static int tun_napi_receive(struct napi_struct *napi, int budget)
 
 	__skb_queue_head_init(&process_queue);
 
+	/* Take at most budget skbs, so that the queue never looks emptier to
+	 * writers than it is.
+	 */
 	spin_lock(&queue->lock);
-	skb_queue_splice_tail_init(queue, &process_queue);
+	if (skb_queue_len(queue) <= budget) {
+		skb_queue_splice_tail_init(queue, &process_queue);
+	} else {
+		while (skb_queue_len(&process_queue) < budget &&
+		       (skb = __skb_dequeue(queue)))
+			__skb_queue_tail(&process_queue, skb);
+	}
 	spin_unlock(&queue->lock);
 
-	while (received < budget && (skb = __skb_dequeue(&process_queue))) {
+	while ((skb = __skb_dequeue(&process_queue))) {
 		napi_gro_receive(napi, skb);
 		++received;
 	}
 
-	if (!skb_queue_empty(&process_queue)) {
-		spin_lock(&queue->lock);
-		skb_queue_splice(&process_queue, queue);
-		spin_unlock(&queue->lock);
-	}
-
 	return received;
 }
 
@@ -1856,6 +1870,24 @@ static ssize_t tun_get_user(struct tun_struct *tun, struct tun_file *tfile,
 			zerocopy = true;
 	}
 
+	/* NAPI queue full: drop before allocating, and schedule NAPI, which a
+	 * MSG_MORE sender may not have done yet. Lockless: concurrent writers
+	 * can each still add one.
+	 */
+	if (tfile->napi_enabled &&
+	    unlikely(tun_napi_queue_full(&tfile->sk.sk_write_queue))) {
+		local_bh_disable();
+		napi_schedule(&tfile->napi);
+		local_bh_enable();
+		if (msg_control) {
+			struct ubuf_info *uarg = msg_control;
+
+			uarg->ops->complete(NULL, uarg, false);
+		}
+		dev_core_stats_rx_dropped_inc(tun->dev);
+		return total_len;
+	}
+
 	if (!frags && tun_can_build_skb(tun, tfile, len, noblock, zerocopy)) {
 		/* For the packet that is not easy to be processed
 		 * (e.g gso or jumbo packet), we will do it at after
@@ -2561,6 +2593,17 @@ static int tun_xdp_one(struct tun_struct *tun,
 		return -EIO;
 	}
 
+	/* NAPI queue full: drop, and schedule NAPI so that it drains even if
+	 * nothing in this batch gets queued
+	 */
+	if (tfile->napi_enabled &&
+	    unlikely(tun_napi_queue_full(&tfile->sk.sk_write_queue))) {
+		napi_schedule(&tfile->napi);
+		put_page(virt_to_head_page(xdp->data));
+		dev_core_stats_rx_dropped_inc(tun->dev);
+		return 0;
+	}
+
 	xdp_prog = rcu_dereference(tun->xdp_prog);
 	if (xdp_prog) {
 		if (gso->gso_type) {
-- 
2.56.0

[PATCH net v5 2/2] selftests: net: add a vhost-net test for the tun NAPI queue bound

[Antonio Quartulli]

Send 16k packets from vhost-net into a tap device with IFF_NAPI and
netdev_max_backlog set to 8, in the three ways vhost-net passes packets
to tun: batches of XDP buffers, one packet per sendmsg() with MSG_MORE,
and zerocopy buffers. Check that excess packets are dropped and counted,
that the queue keeps draining while vhost-net keeps sending, that every
packet is either delivered or dropped, and that every buffer comes back
to the ring. If a dropped zerocopy buffer is never completed, the test
fails, and closing vhost-net then blocks until the buffer is released.

The zerocopy variant needs vhost_net loaded with experimental_zcopytx
set and is skipped otherwise; with it, vhost-net does not batch and the
batch variant is skipped instead.

The test drives a minimal split virtqueue itself, in little-endian as
VIRTIO_F_VERSION_1 requires. It needs root and CONFIG_VHOST_NET.
netdev_max_backlog is global: the test lowers it only once setup has
succeeded and restores it first in teardown, and should not run in
parallel with other tests.

Written and tested with the help of an AI coding assistant (Claude
Code).

Signed-off-by: Antonio Quartulli <antonio@mandelbit.com>
---
 tools/testing/selftests/net/.gitignore  |   1 +
 tools/testing/selftests/net/Makefile    |   1 +
 tools/testing/selftests/net/config      |   1 +
 tools/testing/selftests/net/tun_vhost.c | 409 ++++++++++++++++++++++++
 4 files changed, 412 insertions(+)
 create mode 100644 tools/testing/selftests/net/tun_vhost.c

diff --git a/tools/testing/selftests/net/.gitignore b/tools/testing/selftests/net/.gitignore
index 21412510ac26..aafd07910058 100644
--- a/tools/testing/selftests/net/.gitignore
+++ b/tools/testing/selftests/net/.gitignore
@@ -54,6 +54,7 @@ timestamping
 tls
 tools
 tun
+tun_vhost
 txring_overwrite
 txtimestamp
 udp_splice_checksum
diff --git a/tools/testing/selftests/net/Makefile b/tools/testing/selftests/net/Makefile
index 9180b9bec515..babb684d6f77 100644
--- a/tools/testing/selftests/net/Makefile
+++ b/tools/testing/selftests/net/Makefile
@@ -202,6 +202,7 @@ TEST_GEN_PROGS := \
 	tap \
 	tcp_port_share \
 	tls \
+	tun_vhost \
 	udp_splice_checksum \
 # end of TEST_GEN_PROGS
 
diff --git a/tools/testing/selftests/net/config b/tools/testing/selftests/net/config
index d4416f8bdcfd..9afa82dea11d 100644
--- a/tools/testing/selftests/net/config
+++ b/tools/testing/selftests/net/config
@@ -135,6 +135,7 @@ CONFIG_TRACEPOINTS=y
 CONFIG_TUN=y
 CONFIG_USER_NS=y
 CONFIG_VETH=y
+CONFIG_VHOST_NET=m
 CONFIG_VLAN_8021Q=y
 CONFIG_VXLAN=m
 CONFIG_XFRM_INTERFACE=m
diff --git a/tools/testing/selftests/net/tun_vhost.c b/tools/testing/selftests/net/tun_vhost.c
new file mode 100644
index 000000000000..ed35435df90e
--- /dev/null
+++ b/tools/testing/selftests/net/tun_vhost.c
@@ -0,0 +1,409 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Send packets from vhost-net into a tap device with IFF_NAPI and check
+ * that the NAPI queue is bounded by netdev_max_backlog: excess packets
+ * are dropped and counted, and every buffer is returned to the driver,
+ * zerocopy ones included.
+ *
+ * vhost-net passes packets to tun in three ways:
+ *  - batch: batches of XDP buffers (tun_xdp_one()), the default when the
+ *    socket send buffer is unlimited and zerocopy is disabled
+ *  - copy: one packet per sendmsg(), when the send buffer is limited
+ *  - zerocopy: one packet per sendmsg() with a ubuf_info to complete,
+ *    only with vhost_net experimental_zcopytx set (else skipped)
+ *
+ * Tun schedules NAPI after a batch, after the last of a series of MSG_MORE
+ * packets, once more than NAPI_POLL_WEIGHT (64) packets are queued, and
+ * when it drops a packet because the queue is full. With netdev_max_backlog
+ * below 64, the queue always overflows and only the last of these keeps it
+ * draining while the sender keeps sending.
+ *
+ * netdev_max_backlog is global: the test lowers it for the whole system
+ * while it runs, so it should not run in parallel with other tests.
+ *
+ * Author: Antonio Quartulli <antonio@mandelbit.com>
+ */
+
+#define _GNU_SOURCE
+#include <endian.h>
+#include <errno.h>
+#include <fcntl.h>
+#include <poll.h>
+#include <sched.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <unistd.h>
+#include <linux/if.h>
+#include <linux/if_arp.h>
+#include <linux/if_ether.h>
+#include <linux/if_tun.h>
+#include <linux/vhost.h>
+#include <linux/virtio_config.h>
+#include <linux/virtio_net.h>
+#include <linux/virtio_ring.h>
+#include <sys/eventfd.h>
+#include <sys/ioctl.h>
+#include <sys/socket.h>
+
+#include "kselftest_harness.h"
+
+#define TV_IFNAME	"tvh0"
+#define TV_RING_NUM	256
+#define TV_BUF_SIZE	2048
+#define TV_HDR_LEN	sizeof(struct virtio_net_hdr_v1)
+#define TV_PKTS		(64 * TV_RING_NUM)
+#define TV_BACKLOG	"8"
+#define TV_TX_VQ	1
+
+static const unsigned char tv_mac_tap[ETH_ALEN] = { 2, 0, 0, 0, 0, 1 };
+static const unsigned char tv_mac_dst[ETH_ALEN] = { 2, 0, 0, 0, 0, 2 };
+static const unsigned char tv_mac_src[ETH_ALEN] = { 2, 0, 0, 0, 0, 3 };
+
+/* TX virtqueue, little-endian as negotiated with VIRTIO_F_VERSION_1 */
+struct tv_ring {
+	void *mem;
+	struct vring vr;
+	char *bufs;
+	int kick;
+	int call;
+	unsigned short avail_idx;
+	unsigned short used_idx;
+	/* descriptors not owned by vhost: buffers may come back in any order */
+	unsigned short free[TV_RING_NUM];
+	int nfree;
+};
+
+struct tv_stats {
+	unsigned long long rx_packets;
+	unsigned long long rx_dropped;
+};
+
+static int tv_read_stats(const char *ifname, struct tv_stats *st)
+{
+	char line[512], *p;
+	FILE *f;
+	int ret = -1;
+
+	f = fopen("/proc/net/dev", "r");
+	if (!f)
+		return -1;
+	while (fgets(line, sizeof(line), f)) {
+		p = strchr(line, ':');
+		if (!p)
+			continue;
+		*p = '\0';
+		if (strcmp(line + strspn(line, " "), ifname))
+			continue;
+		if (sscanf(p + 1, "%*u %llu %*u %llu", &st->rx_packets,
+			   &st->rx_dropped) == 2)
+			ret = 0;
+		break;
+	}
+	fclose(f);
+	return ret;
+}
+
+FIXTURE(tun_vhost)
+{
+	int backlog_fd;
+	char backlog[32];
+	int tap;
+	int vhost;
+	struct tv_ring ring;
+};
+
+FIXTURE_VARIANT(tun_vhost)
+{
+	bool batch;
+	bool limit_sndbuf;
+	bool zerocopy;
+	int pktlen;
+};
+
+FIXTURE_VARIANT_ADD(tun_vhost, batch)
+{
+	.batch = true,
+	.pktlen = 64,
+};
+
+FIXTURE_VARIANT_ADD(tun_vhost, copy)
+{
+	.limit_sndbuf = true,
+	.pktlen = 64,
+};
+
+FIXTURE_VARIANT_ADD(tun_vhost, zerocopy)
+{
+	.limit_sndbuf = true,
+	.zerocopy = true,
+	.pktlen = 1000,
+};
+
+/* vhost-net uses zerocopy for any non-zero value */
+static bool tv_zerocopy_enabled(void)
+{
+	char buf[16] = {};
+	int fd;
+
+	fd = open("/sys/module/vhost_net/parameters/experimental_zcopytx",
+		  O_RDONLY);
+	if (fd < 0)
+		return false;
+	if (read(fd, buf, sizeof(buf) - 1) <= 0)
+		buf[0] = '\0';
+	close(fd);
+	return atoi(buf) != 0;
+}
+
+static int tv_vhost_setup(int vhost, int tap, struct tv_ring *r)
+{
+	struct vhost_vring_state state = { .index = TV_TX_VQ };
+	struct vhost_vring_file file = { .index = TV_TX_VQ };
+	struct vhost_vring_addr addr = { .index = TV_TX_VQ };
+	size_t memsz = sizeof(struct vhost_memory) +
+		       sizeof(struct vhost_memory_region);
+	unsigned long long features;
+	struct vhost_memory *mem;
+	int ret;
+
+	if (ioctl(vhost, VHOST_SET_OWNER, NULL) ||
+	    ioctl(vhost, VHOST_GET_FEATURES, &features) ||
+	    !(features & (1ULL << VIRTIO_F_VERSION_1)))
+		return -1;
+	features = 1ULL << VIRTIO_F_VERSION_1;
+	if (ioctl(vhost, VHOST_SET_FEATURES, &features))
+		return -1;
+
+	/* buffer addresses are passed as is: map them 1:1 */
+	mem = calloc(1, memsz);
+	if (!mem)
+		return -1;
+	mem->nregions = 1;
+	mem->regions[0].guest_phys_addr = (unsigned long)r->bufs;
+	mem->regions[0].userspace_addr = (unsigned long)r->bufs;
+	mem->regions[0].memory_size = TV_RING_NUM * TV_BUF_SIZE;
+	ret = ioctl(vhost, VHOST_SET_MEM_TABLE, mem);
+	free(mem);
+	if (ret)
+		return -1;
+
+	state.num = TV_RING_NUM;
+	if (ioctl(vhost, VHOST_SET_VRING_NUM, &state))
+		return -1;
+	state.num = 0;
+	if (ioctl(vhost, VHOST_SET_VRING_BASE, &state))
+		return -1;
+	addr.desc_user_addr = (unsigned long)r->vr.desc;
+	addr.avail_user_addr = (unsigned long)r->vr.avail;
+	addr.used_user_addr = (unsigned long)r->vr.used;
+	if (ioctl(vhost, VHOST_SET_VRING_ADDR, &addr))
+		return -1;
+	file.fd = r->kick;
+	if (ioctl(vhost, VHOST_SET_VRING_KICK, &file))
+		return -1;
+	file.fd = r->call;
+	if (ioctl(vhost, VHOST_SET_VRING_CALL, &file))
+		return -1;
+	file.fd = tap;
+	return ioctl(vhost, VHOST_NET_SET_BACKEND, &file);
+}
+
+static int tv_ring_init(struct tv_ring *r)
+{
+	int i;
+
+	if (posix_memalign((void **)&r->bufs, 4096, TV_RING_NUM * TV_BUF_SIZE))
+		return -1;
+	memset(r->bufs, 0, TV_RING_NUM * TV_BUF_SIZE);
+
+	/* every buffer: zeroed vnet header + IPv4 frame for another host,
+	 * which the stack drops without counting it in rx_dropped
+	 */
+	for (i = 0; i < TV_RING_NUM; i++) {
+		unsigned char *eth = (unsigned char *)r->bufs +
+				     i * TV_BUF_SIZE + TV_HDR_LEN;
+
+		memcpy(eth, tv_mac_dst, ETH_ALEN);
+		memcpy(eth + ETH_ALEN, tv_mac_src, ETH_ALEN);
+		eth[12] = ETH_P_IP >> 8;
+		eth[13] = ETH_P_IP & 0xff;
+		eth[14] = 0x45;
+		r->free[i] = i;
+	}
+	r->nfree = TV_RING_NUM;
+
+	if (posix_memalign(&r->mem, 4096, vring_size(TV_RING_NUM, 4096)))
+		return -1;
+	memset(r->mem, 0, vring_size(TV_RING_NUM, 4096));
+	vring_init(&r->vr, TV_RING_NUM, r->mem, 4096);
+
+	r->kick = eventfd(0, EFD_NONBLOCK);
+	r->call = eventfd(0, EFD_NONBLOCK);
+	return r->kick < 0 || r->call < 0 ? -1 : 0;
+}
+
+/* save the global netdev_max_backlog, to restore it in teardown */
+static int tv_backlog_save(int *fd, char *old, size_t size)
+{
+	ssize_t n;
+
+	/* only visible from init_net: keep it open across unshare() */
+	*fd = open("/proc/sys/net/core/netdev_max_backlog", O_RDWR);
+	if (*fd < 0)
+		return -1;
+	n = read(*fd, old, size - 1);
+	if (n <= 0)
+		return -1;
+	old[n] = '\0';
+	return 0;
+}
+
+FIXTURE_SETUP(tun_vhost)
+{
+	struct ifreq ifr = {};
+	int hdrlen = TV_HDR_LEN;
+	int sndbuf = 1 << 20;
+	int fd;
+
+	self->backlog_fd = -1;
+	self->vhost = open("/dev/vhost-net", O_RDWR);
+	if (self->vhost < 0)
+		SKIP(return, "no /dev/vhost-net");
+	if (variant->zerocopy && !tv_zerocopy_enabled())
+		SKIP(return, "vhost_net experimental_zcopytx not enabled");
+	if (variant->batch && tv_zerocopy_enabled())
+		SKIP(return, "vhost_net does not batch with zerocopy");
+
+	ASSERT_EQ(tv_backlog_save(&self->backlog_fd, self->backlog,
+				  sizeof(self->backlog)), 0);
+	ASSERT_EQ(unshare(CLONE_NEWNET), 0);
+
+	self->tap = open("/dev/net/tun", O_RDWR);
+	ASSERT_GE(self->tap, 0);
+	ifr.ifr_flags = IFF_TAP | IFF_NO_PI | IFF_VNET_HDR | IFF_NAPI;
+	memcpy(ifr.ifr_name, TV_IFNAME, sizeof(TV_IFNAME));
+	ASSERT_EQ(ioctl(self->tap, TUNSETIFF, &ifr), 0);
+	ASSERT_EQ(ioctl(self->tap, TUNSETVNETHDRSZ, &hdrlen), 0);
+	if (variant->limit_sndbuf)
+		ASSERT_EQ(ioctl(self->tap, TUNSETSNDBUF, &sndbuf), 0);
+
+	fd = socket(AF_INET, SOCK_DGRAM, 0);
+	ASSERT_GE(fd, 0);
+	ifr.ifr_hwaddr.sa_family = ARPHRD_ETHER;
+	memcpy(ifr.ifr_hwaddr.sa_data, tv_mac_tap, ETH_ALEN);
+	ASSERT_EQ(ioctl(fd, SIOCSIFHWADDR, &ifr), 0);
+	ASSERT_EQ(ioctl(fd, SIOCGIFFLAGS, &ifr), 0);
+	ifr.ifr_flags |= IFF_UP;
+	ASSERT_EQ(ioctl(fd, SIOCSIFFLAGS, &ifr), 0);
+	close(fd);
+
+	ASSERT_EQ(tv_ring_init(&self->ring), 0);
+	ASSERT_EQ(tv_vhost_setup(self->vhost, self->tap, &self->ring), 0);
+}
+
+FIXTURE_TEARDOWN(tun_vhost)
+{
+	/* restore first: closing vhost may block if buffers are not returned */
+	if (self->backlog_fd >= 0) {
+		if (pwrite(self->backlog_fd, self->backlog,
+			   strlen(self->backlog), 0) < 0)
+			TH_LOG("cannot restore netdev_max_backlog to %s",
+			       self->backlog);
+		close(self->backlog_fd);
+	}
+	close(self->vhost);
+	close(self->tap);
+}
+
+/* queue up to @n buffers, return how many */
+static int tv_post(struct tv_ring *r, int pktlen, int n)
+{
+	unsigned long long one = 1;
+	int i;
+
+	for (i = 0; i < n && r->nfree; i++) {
+		unsigned short id = r->free[--r->nfree];
+
+		r->vr.desc[id].addr = htole64((unsigned long)r->bufs +
+					      id * TV_BUF_SIZE);
+		r->vr.desc[id].len = htole32(TV_HDR_LEN + pktlen);
+		r->vr.desc[id].flags = 0;
+		r->vr.avail->ring[r->avail_idx % TV_RING_NUM] = htole16(id);
+		r->avail_idx++;
+	}
+	__atomic_store_n(&r->vr.avail->idx, htole16(r->avail_idx),
+			 __ATOMIC_RELEASE);
+	if (i && write(r->kick, &one, sizeof(one)) != sizeof(one))
+		return -1;
+	return i;
+}
+
+/* reap returned buffers, waiting up to @ms for at least one */
+static int tv_reap(struct tv_ring *r, int ms)
+{
+	struct pollfd pfd = { .fd = r->call, .events = POLLIN };
+	unsigned short used;
+	unsigned long long v;
+	int n = 0;
+
+	for (;;) {
+		used = le16toh(__atomic_load_n(&r->vr.used->idx,
+					       __ATOMIC_ACQUIRE));
+		while (r->used_idx != used) {
+			struct vring_used_elem *e;
+
+			e = &r->vr.used->ring[r->used_idx % TV_RING_NUM];
+			r->free[r->nfree++] = le32toh(e->id);
+			r->used_idx++;
+			n++;
+		}
+		if (n)
+			return n;
+		if (ms <= 0 || poll(&pfd, 1, 100) < 0)
+			return 0;
+		if (pfd.revents & POLLIN && read(r->call, &v, sizeof(v)) < 0)
+			return -1;
+		ms -= 100;
+	}
+}
+
+TEST_F(tun_vhost, napi_queue_bounded)
+{
+	struct tv_stats before, after;
+	unsigned long long delivered, dropped;
+	int posted = 0, reaped = 0, n;
+
+	ASSERT_EQ(tv_read_stats(TV_IFNAME, &before), 0);
+	/* lowered only now, so that a failed setup leaves it alone */
+	ASSERT_EQ(pwrite(self->backlog_fd, TV_BACKLOG, strlen(TV_BACKLOG), 0),
+		  strlen(TV_BACKLOG));
+
+	while (reaped < TV_PKTS) {
+		n = tv_post(&self->ring, variant->pktlen, TV_PKTS - posted);
+		ASSERT_GE(n, 0);
+		posted += n;
+		n = tv_reap(&self->ring, 5000);
+		ASSERT_GE(n, 0);
+		if (!n) {
+			TH_LOG("%d of %d buffers not returned",
+			       posted - reaped, posted);
+			ASSERT_GT(n, 0);
+		}
+		reaped += n;
+	}
+
+	ASSERT_EQ(tv_read_stats(TV_IFNAME, &after), 0);
+	delivered = after.rx_packets - before.rx_packets;
+	dropped = after.rx_dropped - before.rx_dropped;
+	TH_LOG("sent %d, delivered %llu, dropped %llu", TV_PKTS, delivered,
+	       dropped);
+	/* bounded, */
+	EXPECT_GT(dropped, 0);
+	/* still draining, */
+	EXPECT_GE(delivered, TV_PKTS / 64);
+	/* and every packet accounted for */
+	EXPECT_EQ(delivered + dropped, TV_PKTS);
+}
+
+TEST_HARNESS_MAIN
-- 
2.56.0