devices: add support for offloads
Type: improvement
This patch adds support for:
1) GSO
2) checksum offload
Signed-off-by: Mohsin Kazmi <sykazmi@cisco.com>
Change-Id: Ib00629888c62da04d58db36ce021993769e736c9
diff --git a/src/vnet/devices/af_packet/node.c b/src/vnet/devices/af_packet/node.c
index 229e050..a42e3e0 100644
--- a/src/vnet/devices/af_packet/node.c
+++ b/src/vnet/devices/af_packet/node.c
@@ -28,6 +28,7 @@
#include <vnet/ethernet/packet.h>
#include <vnet/devices/af_packet/af_packet.h>
+#include <vnet/devices/virtio/virtio_std.h>
#define foreach_af_packet_input_error \
_(PARTIAL_PKT, "partial packet")
@@ -51,10 +52,11 @@
u32 next_index;
u32 hw_if_index;
int block;
- u32 num_pkts;
+ u32 pkt_num;
void *block_start;
block_desc_t bd;
tpacket3_hdr_t tph;
+ vnet_virtio_net_hdr_t vnet_hdr;
} af_packet_input_trace_t;
static u8 *
@@ -68,12 +70,10 @@
s = format (s, "af_packet: hw_if_index %d next-index %d",
t->hw_if_index, t->next_index);
- s = format (s,
- "\n%Ublock %u:\n%Uaddress %p version %u seq_num %lu"
- " num_pkts %u",
- format_white_space, indent + 2, t->block, format_white_space,
- indent + 4, t->block_start, t->bd.version, t->bd.hdr.bh1.seq_num,
- t->num_pkts);
+ s = format (
+ s, "\n%Ublock %u:\n%Uaddress %p version %u seq_num %lu pkt_num %u",
+ format_white_space, indent + 2, t->block, format_white_space, indent + 4,
+ t->block_start, t->bd.version, t->bd.hdr.bh1.seq_num, t->pkt_num);
s =
format (s,
"\n%Utpacket3_hdr:\n%Ustatus 0x%x len %u snaplen %u mac %u net %u"
@@ -91,16 +91,21 @@
t->tph.hv1.tp_vlan_tpid
#endif
);
+
+ s = format (s,
+ "\n%Uvnet-hdr:\n%Uflags 0x%02x gso_type 0x%02x hdr_len %u"
+ "\n%Ugso_size %u csum_start %u csum_offset %u",
+ format_white_space, indent + 2, format_white_space, indent + 4,
+ t->vnet_hdr.flags, t->vnet_hdr.gso_type, t->vnet_hdr.hdr_len,
+ format_white_space, indent + 4, t->vnet_hdr.gso_size,
+ t->vnet_hdr.csum_start, t->vnet_hdr.csum_offset);
return s;
}
always_inline void
-buffer_add_to_chain (vlib_main_t * vm, u32 bi, u32 first_bi, u32 prev_bi)
+buffer_add_to_chain (vlib_buffer_t *b, vlib_buffer_t *first_b,
+ vlib_buffer_t *prev_b, u32 bi)
{
- vlib_buffer_t *b = vlib_get_buffer (vm, bi);
- vlib_buffer_t *first_b = vlib_get_buffer (vm, first_bi);
- vlib_buffer_t *prev_b = vlib_get_buffer (vm, prev_bi);
-
/* update first buffer */
first_b->total_length_not_including_first_buffer += b->current_length;
@@ -109,11 +114,11 @@
prev_b->flags |= VLIB_BUFFER_NEXT_PRESENT;
/* update current buffer */
- b->next_buffer = 0;
+ b->next_buffer = ~0;
}
static_always_inline void
-fill_gso_buffer_flags (vlib_buffer_t *b, u32 gso_size, u8 l4_hdr_sz)
+fill_gso_offload (vlib_buffer_t *b, u32 gso_size, u8 l4_hdr_sz)
{
b->flags |= VNET_BUFFER_F_GSO;
vnet_buffer2 (b)->gso_size = gso_size;
@@ -121,43 +126,66 @@
}
static_always_inline void
-mark_tcp_udp_cksum_calc (vlib_buffer_t *b, u8 *l4_hdr_sz)
+fill_cksum_offload (vlib_buffer_t *b, u8 *l4_hdr_sz, u8 is_ip)
{
- ethernet_header_t *eth = vlib_buffer_get_current (b);
vnet_buffer_oflags_t oflags = 0;
- if (clib_net_to_host_u16 (eth->type) == ETHERNET_TYPE_IP4)
+ u16 l2hdr_sz = 0;
+ u16 ethertype = 0;
+ u8 l4_proto = 0;
+
+ if (is_ip)
{
- ip4_header_t *ip4 =
- (vlib_buffer_get_current (b) + sizeof (ethernet_header_t));
- b->flags |= VNET_BUFFER_F_IS_IP4;
- if (ip4->protocol == IP_PROTOCOL_TCP)
+ switch (b->data[0] & 0xf0)
{
- oflags |= VNET_BUFFER_OFFLOAD_F_TCP_CKSUM;
- tcp_header_t *tcp = (tcp_header_t *) (vlib_buffer_get_current (b) +
- sizeof (ethernet_header_t) +
- ip4_header_bytes (ip4));
- *l4_hdr_sz = tcp_header_bytes (tcp);
+ case 0x40:
+ ethertype = ETHERNET_TYPE_IP4;
+ break;
+ case 0x60:
+ ethertype = ETHERNET_TYPE_IP6;
+ break;
}
- else if (ip4->protocol == IP_PROTOCOL_UDP)
- {
- oflags |= VNET_BUFFER_OFFLOAD_F_UDP_CKSUM;
- udp_header_t *udp = (udp_header_t *) (vlib_buffer_get_current (b) +
- sizeof (ethernet_header_t) +
- ip4_header_bytes (ip4));
- *l4_hdr_sz = sizeof (*udp);
- }
- vnet_buffer (b)->l3_hdr_offset = sizeof (ethernet_header_t);
- vnet_buffer (b)->l4_hdr_offset =
- sizeof (ethernet_header_t) + ip4_header_bytes (ip4);
- if (oflags)
- vnet_buffer_offload_flags_set (b, oflags);
}
- else if (clib_net_to_host_u16 (eth->type) == ETHERNET_TYPE_IP6)
+ else
{
- ip6_header_t *ip6 =
- (vlib_buffer_get_current (b) + sizeof (ethernet_header_t));
- b->flags |= VNET_BUFFER_F_IS_IP6;
+ ethernet_header_t *eth = vlib_buffer_get_current (b);
+ ethertype = clib_net_to_host_u16 (eth->type);
+ l2hdr_sz = sizeof (ethernet_header_t);
+ if (ethernet_frame_is_tagged (ethertype))
+ {
+ ethernet_vlan_header_t *vlan = (ethernet_vlan_header_t *) (eth + 1);
+
+ ethertype = clib_net_to_host_u16 (vlan->type);
+ l2hdr_sz += sizeof (*vlan);
+ if (ethertype == ETHERNET_TYPE_VLAN)
+ {
+ vlan++;
+ ethertype = clib_net_to_host_u16 (vlan->type);
+ l2hdr_sz += sizeof (*vlan);
+ }
+ }
+ }
+
+ vnet_buffer (b)->l2_hdr_offset = 0;
+ vnet_buffer (b)->l3_hdr_offset = l2hdr_sz;
+
+ if (ethertype == ETHERNET_TYPE_IP4)
+ {
+ ip4_header_t *ip4 = (vlib_buffer_get_current (b) + l2hdr_sz);
+ vnet_buffer (b)->l4_hdr_offset = l2hdr_sz + ip4_header_bytes (ip4);
+ b->flags |= (VNET_BUFFER_F_IS_IP4 | VNET_BUFFER_F_L2_HDR_OFFSET_VALID |
+ VNET_BUFFER_F_L3_HDR_OFFSET_VALID |
+ VNET_BUFFER_F_L4_HDR_OFFSET_VALID);
+
+ l4_proto = ip4->protocol;
+ }
+ else if (ethertype == ETHERNET_TYPE_IP6)
+ {
+ ip6_header_t *ip6 = (vlib_buffer_get_current (b) + l2hdr_sz);
+ b->flags |= (VNET_BUFFER_F_IS_IP6 | VNET_BUFFER_F_L2_HDR_OFFSET_VALID |
+ VNET_BUFFER_F_L3_HDR_OFFSET_VALID |
+ VNET_BUFFER_F_L4_HDR_OFFSET_VALID);
u16 ip6_hdr_len = sizeof (ip6_header_t);
+
if (ip6_ext_hdr (ip6->protocol))
{
ip6_ext_header_t *p = (void *) (ip6 + 1);
@@ -167,34 +195,34 @@
ip6_hdr_len += ip6_ext_header_len (p);
p = ip6_ext_next_header (p);
}
+ l4_proto = p->next_hdr;
}
- if (ip6->protocol == IP_PROTOCOL_TCP)
- {
- oflags |= VNET_BUFFER_OFFLOAD_F_TCP_CKSUM;
- tcp_header_t *tcp =
- (tcp_header_t *) (vlib_buffer_get_current (b) +
- sizeof (ethernet_header_t) + ip6_hdr_len);
- *l4_hdr_sz = tcp_header_bytes (tcp);
- }
- else if (ip6->protocol == IP_PROTOCOL_UDP)
- {
- oflags |= VNET_BUFFER_OFFLOAD_F_UDP_CKSUM;
- udp_header_t *udp =
- (udp_header_t *) (vlib_buffer_get_current (b) +
- sizeof (ethernet_header_t) + ip6_hdr_len);
- *l4_hdr_sz = sizeof (*udp);
- }
- vnet_buffer (b)->l3_hdr_offset = sizeof (ethernet_header_t);
- vnet_buffer (b)->l4_hdr_offset =
- sizeof (ethernet_header_t) + ip6_hdr_len;
- if (oflags)
- vnet_buffer_offload_flags_set (b, oflags);
+ else
+ l4_proto = ip6->protocol;
+ vnet_buffer (b)->l4_hdr_offset = l2hdr_sz + ip6_hdr_len;
}
+
+ if (l4_proto == IP_PROTOCOL_TCP)
+ {
+ oflags |= VNET_BUFFER_OFFLOAD_F_TCP_CKSUM;
+ tcp_header_t *tcp = (tcp_header_t *) (vlib_buffer_get_current (b) +
+ vnet_buffer (b)->l4_hdr_offset);
+ *l4_hdr_sz = tcp_header_bytes (tcp);
+ }
+ else if (l4_proto == IP_PROTOCOL_UDP)
+ {
+ oflags |= VNET_BUFFER_OFFLOAD_F_UDP_CKSUM;
+ *l4_hdr_sz = sizeof (udp_header_t);
+ }
+
+ if (oflags)
+ vnet_buffer_offload_flags_set (b, oflags);
}
always_inline uword
-af_packet_device_input_fn (vlib_main_t * vm, vlib_node_runtime_t * node,
- vlib_frame_t * frame, af_packet_if_t * apif)
+af_packet_device_input_fn (vlib_main_t *vm, vlib_node_runtime_t *node,
+ vlib_frame_t *frame, af_packet_if_t *apif,
+ u8 is_cksum_gso_enabled)
{
af_packet_main_t *apm = &af_packet_main;
tpacket3_hdr_t *tph;
@@ -210,19 +238,16 @@
u32 thread_index = vm->thread_index;
u32 n_buffer_bytes = vlib_buffer_get_default_data_size (vm);
u32 min_bufs = apif->rx_req->tp_frame_size / n_buffer_bytes;
- u32 eth_header_size = 0;
u32 num_pkts = 0;
u32 rx_frame_offset = 0;
block_desc_t *bd = 0;
vlib_buffer_t bt;
+ u8 is_ip = (apif->mode == AF_PACKET_IF_MODE_IP);
- if (apif->mode == AF_PACKET_IF_MODE_IP)
- {
- next_index = VNET_DEVICE_INPUT_NEXT_IP4_INPUT;
- }
+ if (is_ip)
+ next_index = VNET_DEVICE_INPUT_NEXT_IP4_INPUT;
else
{
- eth_header_size = sizeof (ethernet_header_t);
next_index = VNET_DEVICE_INPUT_NEXT_ETHERNET_INPUT;
if (PREDICT_FALSE (apif->per_interface_next_index != ~0))
next_index = apif->per_interface_next_index;
@@ -257,10 +282,11 @@
while (num_pkts && (n_free_bufs > min_bufs))
{
- vlib_buffer_t *b0 = 0, *first_b0 = 0;
u32 next0 = next_index;
u32 n_left_to_next;
+
vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next);
+
while (num_pkts && n_left_to_next && (n_free_bufs > min_bufs))
{
tph = (tpacket3_hdr_t *) (block_start + rx_frame_offset);
@@ -269,19 +295,36 @@
CLIB_PREFETCH (block_start + rx_frame_offset +
tph->tp_next_offset,
2 * CLIB_CACHE_LINE_BYTES, LOAD);
+
+ vlib_buffer_t *b0 = 0, *first_b0 = 0, *prev_b0 = 0;
+ vnet_virtio_net_hdr_t *vnet_hdr = 0;
u32 data_len = tph->tp_snaplen;
u32 offset = 0;
- u32 bi0 = 0, first_bi0 = 0, prev_bi0;
+ u32 bi0 = ~0, first_bi0 = ~0;
u8 l4_hdr_sz = 0;
+ if (is_cksum_gso_enabled)
+ vnet_hdr =
+ (vnet_virtio_net_hdr_t *) ((u8 *) tph + tph->tp_mac -
+ sizeof (vnet_virtio_net_hdr_t));
+
+ if (PREDICT_FALSE (((data_len / n_buffer_bytes) + 1) >
+ vec_len (apm->rx_buffers[thread_index])))
+ {
+ vec_validate (apm->rx_buffers[thread_index],
+ VLIB_FRAME_SIZE + n_free_bufs - 1);
+ n_free_bufs += vlib_buffer_alloc (
+ vm, &apm->rx_buffers[thread_index][n_free_bufs],
+ VLIB_FRAME_SIZE);
+ _vec_len (apm->rx_buffers[thread_index]) = n_free_bufs;
+ }
+
while (data_len)
{
/* grab free buffer */
u32 last_empty_buffer =
vec_len (apm->rx_buffers[thread_index]) - 1;
- prev_bi0 = bi0;
bi0 = apm->rx_buffers[thread_index][last_empty_buffer];
- b0 = vlib_get_buffer (vm, bi0);
_vec_len (apm->rx_buffers[thread_index]) = last_empty_buffer;
n_free_bufs--;
@@ -290,7 +333,10 @@
data_len > n_buffer_bytes ? n_buffer_bytes : data_len;
u32 vlan_len = 0;
u32 bytes_copied = 0;
+
+ b0 = vlib_get_buffer (vm, bi0);
b0->current_data = 0;
+
/* Kernel removes VLAN headers, so reconstruct VLAN */
if (PREDICT_FALSE (tph->tp_status & TP_STATUS_VLAN_VALID))
{
@@ -328,20 +374,22 @@
vnet_buffer (b0)->sw_if_index[VLIB_RX] =
apif->sw_if_index;
vnet_buffer (b0)->sw_if_index[VLIB_TX] = (u32) ~0;
+ first_b0 = b0;
first_bi0 = bi0;
- first_b0 = vlib_get_buffer (vm, first_bi0);
- if (tph->tp_status & TP_STATUS_CSUMNOTREADY)
- mark_tcp_udp_cksum_calc (first_b0, &l4_hdr_sz);
- /* This is a trade-off for GSO. As kernel isn't passing
- * us the GSO state or size, we guess it by comparing it
- * to the host MTU of the interface */
- if (tph->tp_snaplen > (apif->host_mtu + eth_header_size))
- fill_gso_buffer_flags (first_b0, apif->host_mtu,
- l4_hdr_sz);
+ if (is_cksum_gso_enabled)
+ {
+ if (vnet_hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM)
+ fill_cksum_offload (first_b0, &l4_hdr_sz, is_ip);
+ if (vnet_hdr->gso_type & (VIRTIO_NET_HDR_GSO_TCPV4 |
+ VIRTIO_NET_HDR_GSO_TCPV6))
+ fill_gso_offload (first_b0, vnet_hdr->gso_size,
+ l4_hdr_sz);
+ }
}
else
- buffer_add_to_chain (vm, bi0, first_bi0, prev_bi0);
+ buffer_add_to_chain (b0, first_b0, prev_b0, bi0);
+ prev_b0 = b0;
offset += bytes_to_copy;
data_len -= bytes_to_copy;
}
@@ -398,9 +446,15 @@
tr->hw_if_index = apif->hw_if_index;
tr->block = block;
tr->block_start = bd;
- tr->num_pkts = num_pkts;
+ tr->pkt_num = bd->hdr.bh1.num_pkts - num_pkts;
clib_memcpy_fast (&tr->bd, bd, sizeof (block_desc_t));
clib_memcpy_fast (&tr->tph, tph, sizeof (tpacket3_hdr_t));
+ if (is_cksum_gso_enabled)
+ clib_memcpy_fast (&tr->vnet_hdr, vnet_hdr,
+ sizeof (vnet_virtio_net_hdr_t));
+ else
+ clib_memset_u8 (&tr->vnet_hdr, 0,
+ sizeof (vnet_virtio_net_hdr_t));
}
/* enque and take next packet */
@@ -446,9 +500,15 @@
af_packet_if_t *apif;
apif = vec_elt_at_index (apm->interfaces, pv[i].dev_instance);
if (apif->is_admin_up)
- n_rx_packets += af_packet_device_input_fn (vm, node, frame, apif);
+ {
+ if (apif->is_cksum_gso_enabled)
+ n_rx_packets +=
+ af_packet_device_input_fn (vm, node, frame, apif, 1);
+ else
+ n_rx_packets +=
+ af_packet_device_input_fn (vm, node, frame, apif, 0);
+ }
}
-
return n_rx_packets;
}