netvsp: Add VLAN support

Cargo.lock

@@ -4627,6 +4627,7 @@ dependencies = [
  "mesh",
  "net_backend",
  "pal_async",
+ "parking_lot",
  "pci_core",
  "safeatomic",
  "test_with_tracing",

vm/devices/net/gdma/src/bnic.rs

@@ -139,6 +139,11 @@ impl BufferAccess for GuestBuffers {
             },
         }
 
+        if let Some(vlan) = &metadata.vlan {
+            flags.set_rx_vlantag_present(true);
+            flags.set_rx_vlan_id(vlan.vlan_id as u32);
+        }
+
         let packet = &mut self.rx_packets[id.0 as usize];
 
         let cqe_type = if metadata.len > packet.len as usize {
@@ -548,6 +553,11 @@ impl TxRxTask {
 
         let sge0 = sqe.sgl().first().context("no sgl")?;
         let total_len: usize = sqe.sgl().iter().map(|sge| sge.size as usize).sum();
+        let l2_len = if oob.l_oob.inject_vlan_pri_tag() {
+            net_backend::ETHERNET_VLAN_HEADER_LEN
+        } else {
+            net_backend::ETHERNET_HEADER_LEN
+        };
         let mut meta = TxMetadata {
             id: TxId(0),
             segment_count: sqe.sgl().len().try_into().unwrap(),
@@ -558,10 +568,19 @@ impl TxRxTask {
                 .with_offload_udp_checksum(oob.s_oob.comp_udp_csum())
                 .with_is_ipv4(oob.s_oob.is_outer_ipv4())
                 .with_is_ipv6(oob.s_oob.is_outer_ipv6() && !oob.s_oob.is_outer_ipv4()),
-            l2_len: 14,
-            l3_len: oob.s_oob.trans_off().clamp(14, 255) - 14,
+            l2_len: l2_len as u8,
+            l3_len: oob.s_oob.trans_off().clamp(l2_len as u16, 255) - l2_len as u16,
             l4_len: 0,
+            transport_header_offset: 0,
             max_segment_size: 0,
+            vlan: oob
+                .l_oob
+                .inject_vlan_pri_tag()
+                .then(|| net_backend::VlanMetadata {
+                    priority: oob.l_oob.pcp(),
+                    drop_eligible_indicator: oob.l_oob.dei(),
+                    vlan_id: oob.l_oob.vlan_id(),
+                }),
         };
 
         if sqe.header.params.client_oob_in_sgl() {

vm/devices/net/net_backend/src/lib.rs

@@ -281,6 +281,24 @@ pub trait BufferAccess {
     }
 }
 
+pub const ETHERNET_HEADER_LEN: u32 = 14;
+pub const ETHERNET_VLAN_HEADER_LEN: u32 = 18;
+
+pub const IPV4_MIN_HEADER_LEN: u16 = 20;
+pub const IPV6_MIN_HEADER_LEN: u16 = 40;
+
+#[derive(Debug, Copy, Clone)]
+pub struct VlanMetadata {
+    /// Priority for 802.1Q. Actually a 3-bit value.
+    pub priority: u8,
+    /// In pretty much every circumstance this is false. When
+    /// it is used, setting DEI will inform switches/routing infra
+    /// that this can be dropped before higher priority traffic.
+    pub drop_eligible_indicator: bool,
+    /// The 802.1Q ID for this transmission. Actually a 12-bit value.
+    pub vlan_id: u16,
+}
+
 /// A receive buffer ID.
 #[derive(Debug, Copy, Clone)]
 #[repr(transparent)]
@@ -308,6 +326,10 @@ pub struct RxMetadata {
     pub l4_checksum: RxChecksumState,
     /// The L4 protocol.
     pub l4_protocol: L4Protocol,
+    /// Information about 802.1Q VLAN tagging. When a vlan is in use, this structure
+    /// is populated. Only applies when traffic is being received over an L2 connection,
+    /// so L3-only or above traffic will not use this option.
+    pub vlan: Option<VlanMetadata>,
 }
 
 impl Default for RxMetadata {
@@ -318,6 +340,7 @@ impl Default for RxMetadata {
             ip_checksum: RxChecksumState::Unknown,
             l4_checksum: RxChecksumState::Unknown,
             l4_protocol: L4Protocol::Unknown,
+            vlan: None,
         }
     }
 }
@@ -396,10 +419,17 @@ pub struct TxMetadata {
     /// The length of the TCP header. Only guaranteed to be set if various
     /// offload flags are set.
     pub l4_len: u8,
+    /// The offset into the buffer where the L4 header begins (TCP or UDP). Only
+    /// expected to be set if offload (checksum and/or segmentation) flags are set.
+    pub transport_header_offset: u16,
     /// The maximum segment size, used for segmentation offload (TSO or USO).
     /// Only guaranteed to be set if [`TxFlags::offload_tcp_segmentation`] or
     /// [`TxFlags::offload_udp_segmentation`] is set.
     pub max_segment_size: u16,
+    /// Information about 802.1Q VLAN tagging. When a vlan is in use, this structure
+    /// is populated. Only applies when traffic is being sent over an L2 connection,
+    /// so L3-only or above traffic will not use this option.
+    pub vlan: Option<VlanMetadata>,
 }
 
 /// Flags affecting transmit behavior.
@@ -444,7 +474,9 @@ impl Default for TxMetadata {
             l2_len: 0,
             l3_len: 0,
             l4_len: 0,
+            transport_header_offset: 0,
             max_segment_size: 0,
+            vlan: None,
         }
     }
 }

vm/devices/net/net_backend/src/tests.rs

@@ -75,3 +75,10 @@ impl BufferAccess for Bufs {
         *self.inner.rx_metadata[id.0 as usize].lock() = Some(*metadata);
     }
 }
+
+impl Bufs {
+    /// Returns the [`RxMetadata`] written for the given receive buffer, if any.
+    pub fn rx_metadata(&self, id: RxId) -> Option<RxMetadata> {
+        *self.inner.rx_metadata[id.0 as usize].lock()
+    }
+}

vm/devices/net/net_consomme/src/lib.rs

@@ -522,6 +522,7 @@ impl consomme::Client for Client<'_> {
                     } else {
                         L4Protocol::Unknown
                     },
+                    vlan: None,
                 },
                 data,
             );

vm/devices/net/net_mana/Cargo.toml

@@ -32,6 +32,7 @@ zerocopy.workspace = true
 [dev-dependencies]
 chipset_device.workspace = true
 gdma.workspace = true
+parking_lot.workspace = true
 pci_core.workspace = true
 test_with_tracing.workspace = true
 user_driver_emulated_mock.workspace = true

vm/devices/net/net_mana/src/lib.rs

@@ -58,6 +58,7 @@ use net_backend::TxId;
 use net_backend::TxOffloadSupport;
 use net_backend::TxSegment;
 use net_backend::TxSegmentType;
+use net_backend::VlanMetadata;
 use pal_async::task::Spawn;
 use safeatomic::AtomicSliceOps;
 use std::collections::VecDeque;
@@ -965,6 +966,11 @@ impl<T: DeviceBacking + Send> Queue for ManaQueue<T> {
                         } else {
                             (L4Protocol::Unknown, RxChecksumState::Unknown)
                         };
+                        let vlantag = rx_oob.flags.rx_vlantag_present().then(|| VlanMetadata {
+                            drop_eligible_indicator: false,
+                            priority: 0,
+                            vlan_id: rx_oob.flags.rx_vlan_id() as u16,
+                        });
                         let len = rx_oob.ppi[0].pkt_len.into();
                         pool.write_header(
                             rx.id,
@@ -974,6 +980,7 @@ impl<T: DeviceBacking + Send> Queue for ManaQueue<T> {
                                 ip_checksum,
                                 l4_checksum,
                                 l4_protocol,
+                                vlan: vlantag,
                             },
                         );
                         if rx.bounced_len_with_padding > 0 {
@@ -1163,10 +1170,16 @@ impl<T: DeviceBacking> ManaQueue<T> {
             .set_comp_tcp_csum(meta.flags.offload_tcp_checksum());
         oob.s_oob
             .set_comp_udp_csum(meta.flags.offload_udp_checksum());
-        if meta.flags.offload_tcp_checksum() {
+        if meta.flags.offload_tcp_checksum() || meta.flags.offload_udp_checksum() {
             oob.s_oob.set_trans_off(meta.l2_len as u16 + meta.l3_len);
         }
-        let short_format = self.vp_offset <= 0xff;
+        if let Some(vlan) = &meta.vlan {
+            oob.l_oob.set_inject_vlan_pri_tag(true);
+            oob.l_oob.set_vlan_id(vlan.vlan_id);
+            oob.l_oob.set_pcp(vlan.priority);
+            oob.l_oob.set_dei(vlan.drop_eligible_indicator);
+        }
+        let short_format = self.vp_offset <= 0xff && meta.vlan.is_none();
         if short_format {
             oob.s_oob.set_pkt_fmt(MANA_SHORT_PKT_FMT);
             oob.s_oob.set_short_vp_offset(self.vp_offset as u8);