On Fri, 22 Nov 2024 17:43:34 +0100 Laurent Vivier <lvivier(a)redhat.com> wrote:+/** + * tcp_vu_send_flag() - Send segment with flags to vhost-user (no payload) + * @c: Execution context + * @conn: Connection pointer + * @flags: TCP flags: if not set, send segment only if ACK is due + * + * Return: negative error code on connection reset, 0 otherwise + */ +int tcp_vu_send_flag(const struct ctx *c, struct tcp_tap_conn *conn, int flags) +{ + struct vu_dev *vdev = c->vdev; + struct vu_virtq *vq = &vdev->vq[VHOST_USER_RX_QUEUE]; + const struct flowside *tapside = TAPFLOW(conn); + size_t l2len, l4len, optlen, hdrlen; + struct vu_virtq_element flags_elem[2]; + struct tcp_payload_t *payload; + struct ipv6hdr *ip6h = NULL; + struct iovec flags_iov[2]; + struct iphdr *iph = NULL; + struct ethhdr *eh; + uint32_t seq; + int elem_cnt; + int nb_ack; + int ret; + + hdrlen = tcp_vu_hdrlen(CONN_V6(conn)); + + vu_set_element(&flags_elem[0], NULL, &flags_iov[0]); + + elem_cnt = vu_collect(vdev, vq, &flags_elem[0], 1, + hdrlen + sizeof(struct tcp_syn_opts), NULL);Oops, I made this crash, by starting a number of iperf3 client threads on the host: $ iperf3 -c localhost -p 6001 -Z -l 500 -w 256M -t 600 -P20 with matching server in the guest, then terminating QEMU while the test is running. Details (I saw it first, then I reproduced it under gdb): accepted connection from PID 3115463 NDP: received RS, sending RA DHCP: offer to discover from 52:54:00:12:34:56 DHCP: ack to request from 52:54:00:12:34:56 NDP: sending unsolicited RA, next in 212s Client connection closed Program received signal SIGSEGV, Segmentation fault. 0x00005555555884f5 in vring_avail_idx (vq=0x555559343f10 <vdev_storage+1296>) at virtio.c:138 138 vq->shadow_avail_idx = le16toh(vq->vring.avail->idx); (gdb) list 133 * 134 * Return: the available ring index of the given virtqueue 135 */ 136 static inline uint16_t vring_avail_idx(struct vu_virtq *vq) 137 { 138 vq->shadow_avail_idx = le16toh(vq->vring.avail->idx); 139 140 return vq->shadow_avail_idx; 141 } 142 (gdb) bt #0 0x00005555555884f5 in vring_avail_idx (vq=0x555559343f10 <vdev_storage+1296>) at virtio.c:138 #1 vu_queue_empty (vq=vq@entry=0x555559343f10 <vdev_storage+1296>) at virtio.c:290 #2 vu_queue_pop (dev=dev@entry=0x555559343a00 <vdev_storage>, vq=vq@entry=0x555559343f10 <vdev_storage+1296>, elem=elem@entry=0x7ffffff6f510) at virtio.c:505 #3 0x0000555555588c8c in vu_collect (vdev=vdev@entry=0x555559343a00 <vdev_storage>, vq=vq@entry=0x555559343f10 <vdev_storage+1296>, elem=elem@entry=0x7ffffff6f510, max_elem=max_elem@entry=1, size=size@entry=74, frame_size=frame_size@entry=0x0) at vu_common.c:86 #4 0x000055555557e00e in tcp_vu_send_flag (c=0x7ffffff6f7a0, conn=0x5555555bd2d0 <flowtab+2160>, flags=4) at tcp_vu.c:116 #5 0x0000555555578125 in tcp_send_flag (flags=4, conn=0x5555555bd2d0 <flowtab+2160>, c=0x7ffffff6f7a0) at tcp.c:1278 #6 tcp_rst_do (conn=<optimized out>, c=<optimized out>) at tcp.c:1293 #7 tcp_timer_handler (c=c@entry=0x7ffffff6f7a0, ref=..., ref@entry=...) at tcp.c:2266 #8 0x0000555555558f26 in main (argc=<optimized out>, argv=<optimized out>) at passt.c:342 (gdb) p *vq $1 = {vring = {num = 256, desc = 0x0, avail = 0x0, used = 0x0, log_guest_addr = 4338774592, flags = 0}, last_avail_idx = 35133, shadow_avail_idx = 35133, used_idx = 35133, signalled_used = 0, signalled_used_valid = false, notification = true, inuse = 0, call_fd = -1, kick_fd = -1, err_fd = -1, enable = 1, started = false, vra = {index = 0, flags = 0, desc_user_addr = 139660501995520, used_user_addr = 139660502000192, avail_user_addr = 139660501999616, log_guest_addr = 4338774592}} (gdb) p *vq->vring.avail Cannot access memory at address 0x0 ...so we're sending a RST segment to the guest, but the ring doesn't exist anymore. By the way, I still have the gdb session running, if you need something else out of it. Now, I guess we should eventually introduce a more comprehensive handling of the case where the guest suddenly terminates (not specific to vhost-user), but given that we have vu_cleanup() working as expected in this case, I wonder if we shouldn't simply avoid calling vring_avail_idx() (it has a single caller) by checking for !vring.avail in the caller, or something like that. We can also handle this as a follow-up patch, it's not something that would stand in the way of a bisect, but I wouldn't push the series without a fix immediately following it. -- Stefano