On Thu, Jun 20, 2024 at 3:36 PM Willem de Bruijn willemdebruijn.kernel@gmail.com wrote:
Chengen Du wrote:
On Tue, Jun 18, 2024 at 5:24 PM Willem de Bruijn willemdebruijn.kernel@gmail.com wrote:
Jiri Pirko wrote:
Mon, Jun 17, 2024 at 09:07:58PM CEST, willemdebruijn.kernel@gmail.com wrote:
Jiri Pirko wrote:
Mon, Jun 17, 2024 at 07:45:14AM CEST, chengen.du@canonical.com wrote: >The issue initially stems from libpcap. The ethertype will be overwritten >as the VLAN TPID if the network interface lacks hardware VLAN offloading. >In the outbound packet path, if hardware VLAN offloading is unavailable, >the VLAN tag is inserted into the payload but then cleared from the sk_buff >struct. Consequently, this can lead to a false negative when checking for >the presence of a VLAN tag, causing the packet sniffing outcome to lack >VLAN tag information (i.e., TCI-TPID). As a result, the packet capturing >tool may be unable to parse packets as expected. > >The TCI-TPID is missing because the prb_fill_vlan_info() function does not >modify the tp_vlan_tci/tp_vlan_tpid values, as the information is in the >payload and not in the sk_buff struct. The skb_vlan_tag_present() function >only checks vlan_all in the sk_buff struct. In cooked mode, the L2 header >is stripped, preventing the packet capturing tool from determining the >correct TCI-TPID value. Additionally, the protocol in SLL is incorrect, >which means the packet capturing tool cannot parse the L3 header correctly. > >Link: https://github.com/the-tcpdump-group/libpcap/issues/1105 >Link: https://lore.kernel.org/netdev/20240520070348.26725-1-chengen.du@canonical.c... >Fixes: 393e52e33c6c ("packet: deliver VLAN TCI to userspace") >Cc: stable@vger.kernel.org >Signed-off-by: Chengen Du chengen.du@canonical.com >--- > net/packet/af_packet.c | 86 +++++++++++++++++++++++++++++++++++++++++- > 1 file changed, 84 insertions(+), 2 deletions(-) > >diff --git a/net/packet/af_packet.c b/net/packet/af_packet.c >index ea3ebc160e25..84e8884a77e3 100644 >--- a/net/packet/af_packet.c >+++ b/net/packet/af_packet.c >@@ -538,6 +538,61 @@ static void *packet_current_frame(struct packet_sock *po, > return packet_lookup_frame(po, rb, rb->head, status); > } > >+static u16 vlan_get_tci(struct sk_buff *skb, struct net_device *dev) >+{ >+ struct vlan_hdr vhdr, *vh; >+ u8 *skb_orig_data = skb->data; >+ int skb_orig_len = skb->len; >+ unsigned int header_len; >+ >+ if (!dev) >+ return 0; >+ >+ /* In the SOCK_DGRAM scenario, skb data starts at the network >+ * protocol, which is after the VLAN headers. The outer VLAN >+ * header is at the hard_header_len offset in non-variable >+ * length link layer headers. If it's a VLAN device, the >+ * min_header_len should be used to exclude the VLAN header >+ * size. >+ */ >+ if (dev->min_header_len == dev->hard_header_len) >+ header_len = dev->hard_header_len; >+ else if (is_vlan_dev(dev)) >+ header_len = dev->min_header_len; >+ else >+ return 0; >+ >+ skb_push(skb, skb->data - skb_mac_header(skb)); >+ vh = skb_header_pointer(skb, header_len, sizeof(vhdr), &vhdr); >+ if (skb_orig_data != skb->data) { >+ skb->data = skb_orig_data; >+ skb->len = skb_orig_len; >+ } >+ if (unlikely(!vh)) >+ return 0; >+ >+ return ntohs(vh->h_vlan_TCI); >+} >+ >+static __be16 vlan_get_protocol_dgram(struct sk_buff *skb) >+{ >+ __be16 proto = skb->protocol; >+ >+ if (unlikely(eth_type_vlan(proto))) { >+ u8 *skb_orig_data = skb->data; >+ int skb_orig_len = skb->len; >+ >+ skb_push(skb, skb->data - skb_mac_header(skb)); >+ proto = __vlan_get_protocol(skb, proto, NULL); >+ if (skb_orig_data != skb->data) { >+ skb->data = skb_orig_data; >+ skb->len = skb_orig_len; >+ } >+ } >+ >+ return proto; >+} >+ > static void prb_del_retire_blk_timer(struct tpacket_kbdq_core *pkc) > { > del_timer_sync(&pkc->retire_blk_timer); >@@ -1007,10 +1062,16 @@ static void prb_clear_rxhash(struct tpacket_kbdq_core *pkc, > static void prb_fill_vlan_info(struct tpacket_kbdq_core *pkc, > struct tpacket3_hdr *ppd) > { >+ struct packet_sock *po = container_of(pkc, struct packet_sock, rx_ring.prb_bdqc); >+ > if (skb_vlan_tag_present(pkc->skb)) { > ppd->hv1.tp_vlan_tci = skb_vlan_tag_get(pkc->skb); > ppd->hv1.tp_vlan_tpid = ntohs(pkc->skb->vlan_proto); > ppd->tp_status = TP_STATUS_VLAN_VALID | TP_STATUS_VLAN_TPID_VALID; >+ } else if (unlikely(po->sk.sk_type == SOCK_DGRAM && eth_type_vlan(pkc->skb->protocol))) { >+ ppd->hv1.tp_vlan_tci = vlan_get_tci(pkc->skb, pkc->skb->dev); >+ ppd->hv1.tp_vlan_tpid = ntohs(pkc->skb->protocol); >+ ppd->tp_status = TP_STATUS_VLAN_VALID | TP_STATUS_VLAN_TPID_VALID; > } else { > ppd->hv1.tp_vlan_tci = 0; > ppd->hv1.tp_vlan_tpid = 0; >@@ -2428,6 +2489,10 @@ static int tpacket_rcv(struct sk_buff *skb, struct net_device *dev, > h.h2->tp_vlan_tci = skb_vlan_tag_get(skb); > h.h2->tp_vlan_tpid = ntohs(skb->vlan_proto); > status |= TP_STATUS_VLAN_VALID | TP_STATUS_VLAN_TPID_VALID; >+ } else if (unlikely(sk->sk_type == SOCK_DGRAM && eth_type_vlan(skb->protocol))) { >+ h.h2->tp_vlan_tci = vlan_get_tci(skb, skb->dev); >+ h.h2->tp_vlan_tpid = ntohs(skb->protocol); >+ status |= TP_STATUS_VLAN_VALID | TP_STATUS_VLAN_TPID_VALID; > } else { > h.h2->tp_vlan_tci = 0; > h.h2->tp_vlan_tpid = 0; >@@ -2457,7 +2522,8 @@ static int tpacket_rcv(struct sk_buff *skb, struct net_device *dev, > sll->sll_halen = dev_parse_header(skb, sll->sll_addr); > sll->sll_family = AF_PACKET; > sll->sll_hatype = dev->type; >- sll->sll_protocol = skb->protocol; >+ sll->sll_protocol = (sk->sk_type == SOCK_DGRAM) ? >+ vlan_get_protocol_dgram(skb) : skb->protocol; > sll->sll_pkttype = skb->pkt_type; > if (unlikely(packet_sock_flag(po, PACKET_SOCK_ORIGDEV))) > sll->sll_ifindex = orig_dev->ifindex; >@@ -3482,7 +3548,8 @@ static int packet_recvmsg(struct socket *sock, struct msghdr *msg, size_t len, > /* Original length was stored in sockaddr_ll fields */ > origlen = PACKET_SKB_CB(skb)->sa.origlen; > sll->sll_family = AF_PACKET; >- sll->sll_protocol = skb->protocol; >+ sll->sll_protocol = (sock->type == SOCK_DGRAM) ? >+ vlan_get_protocol_dgram(skb) : skb->protocol; > } > > sock_recv_cmsgs(msg, sk, skb); >@@ -3539,6 +3606,21 @@ static int packet_recvmsg(struct socket *sock, struct msghdr *msg, size_t len, > aux.tp_vlan_tci = skb_vlan_tag_get(skb); > aux.tp_vlan_tpid = ntohs(skb->vlan_proto); > aux.tp_status |= TP_STATUS_VLAN_VALID | TP_STATUS_VLAN_TPID_VALID; >+ } else if (unlikely(sock->type == SOCK_DGRAM && eth_type_vlan(skb->protocol))) {
I don't understand why this would be needed here. We spent quite a bit of efford in the past to make sure vlan header is always stripped. Could you fix that in tx path to fulfill the expectation?
Doesn't that require NETIF_F_HW_VLAN_CTAG_TX?
I also wondered whether we should just convert the skb for this case with skb_vlan_untag, to avoid needing new PF_PACKET logic to handle unstripped tags in the packet socket code. But it seems equally complex.
Correct. skb_vlan_untag() as a preparation of skb before this function is called is exactly what I was suggesting.
It's not necessarily simpler, as that function expects skb->data to point to the (outer) VLAN header.
It will pull that one, but not any subsequent ones.
SOCK_DGRAM expects skb->data to point to the network layer header. And we only want to make this change for SOCK_DGRAM and if auxdata is requested.
Not sure that it will be simpler. But worth a look at least.
Thank you for your suggestion.
I have analyzed the code and considered a feasible approach. We could call skb_vlan_untag() in packet_rcv before pushing skb into sk->sk_receive_queue.
Only for SOCK_DGRAM.
And there is some user risk, as they will see different packets on the same devices as before. A robust program should work for both vlan stripped and unstripped, and the unstripped case is already broken wrt sll_protocol returned, so I suppose this is acceptable.
We would also need to determine if auxdata is required to maintain performance, which might cause the logic of judging PACKET_SOCK_AUXDATA to be spread across both the packet_rcv() and packet_recvmsg() functions.
You mean to only make the above change if SOCK_DGRAM and auxdata is requested?
Yes, we can constrain the performance overhead to specific scenarios this way.
Btw, also tpacket_rcv, where auxdata is always returned.
The skb_vlan_untag() function handles VLANs in a more comprehensive way, but it seems to have a greater performance impact compared to our current approach.
I was afraid of that too. The skb_share_check is fine, as this also exists in packet_rcv, before we would call skb_vlan_untag.
A bigger issue: this only pulls the outer tag. So we still need to handle the vlan stacking case correctly manually.
It seems we are on the same page. The need to manually handle VLAN stacking is a significant concern. Since the code is in the critical path, we must carefully manage the performance overhead. Given that the current method is more efficient than calling skb_vlan_untag(), I propose retaining the patch as is. Please let me know if there are any other concerns.
As I am not an expert in this domain, I might have overlooked some important aspects. I would appreciate it if you could share your thoughts on this decision.