vmxnet3 driver's XDP handling is buggy for packet sizes using ring0 (that
is, packet sizes between 128 - 3k bytes).
We noticed MTU-related connectivity issues with Cilium's service load-
balancing in case of vmxnet3 as NIC underneath. A simple curl to a HTTP
backend service where the XDP LB was doing IPIP encap led to overly large
packet sizes but only for *some* of the packets (e.g. HTTP GET request)
while others (e.g. the prior TCP 3WHS) looked completely fine on the wire.
In fact, the pcap recording on the backend node actually revealed that the
node with the XDP LB was leaking uninitialized kernel data onto the wire
for the affected packets, for example, while the packets should have been
152 bytes their actual size was 1482 bytes, so the remainder after 152 bytes
was padded with whatever other data was in that page at the time (e.g. we
saw user/payload data from prior processed packets).
We only noticed this through an MTU issue, e.g. when the XDP LB node and
the backend node both had the same MTU (e.g. 1500) then the curl request
got dropped on the backend node's NIC given the packet was too large even
though the IPIP-encapped packet normally would never even come close to
the MTU limit. Lowering the MTU on the XDP LB (e.g. 1480) allowed to let
the curl request succeed (which also indicates that the kernel ignored the
padding, and thus the issue wasn't very user-visible).
Commit e127ce7699 ("vmxnet3: Fix missing reserved tailroom") was too eager
to also switch xdp_prepare_buff() from rcd->len to rbi->len. It really needs
to stick to rcd->len which is the actual packet length from the descriptor.
The latter we also feed into vmxnet3_process_xdp_small(), by the way, and
it indicates the correct length needed to initialize the xdp->{data,data_end}
parts. For e127ce7699 ("vmxnet3: Fix missing reserved tailroom") the
relevant part was adapting xdp_init_buff() to address the warning given the
xdp_data_hard_end() depends on xdp->frame_sz. With that fixed, traffic on
the wire looks good again.
Fixes: e127ce7699 ("vmxnet3: Fix missing reserved tailroom")
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Tested-by: Andrew Sauber <andrew.sauber@isovalent.com>
Cc: Anton Protopopov <aspsk@isovalent.com>
Cc: William Tu <witu@nvidia.com>
Cc: Martin Zaharinov <micron10@gmail.com>
Cc: Ronak Doshi <ronak.doshi@broadcom.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20250423133600.176689-1-daniel@iogearbox.net
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
If XDP traffic runs on a CPU which is greater than or equal to
the number of the Tx queues of the NIC, then vmxnet3_xdp_get_tq()
always picks up queue 0 for transmission as it uses reciprocal scale
instead of simple modulo operation.
vmxnet3_xdp_xmit() and vmxnet3_xdp_xmit_frame() use the above
returned queue without any locking which can lead to race conditions
when multiple XDP xmits run in parallel on different CPU's.
This patch uses a simple module scheme when the current CPU equals or
exceeds the number of Tx queues on the NIC. It also adds locking in
vmxnet3_xdp_xmit() and vmxnet3_xdp_xmit_frame() functions.
Fixes: 54f00cce11 ("vmxnet3: Add XDP support.")
Signed-off-by: Sankararaman Jayaraman <sankararaman.jayaraman@broadcom.com>
Signed-off-by: Ronak Doshi <ronak.doshi@broadcom.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20250131042340.156547-1-sankararaman.jayaraman@broadcom.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Andrew and Nikolay reported connectivity issues with Cilium's service
load-balancing in case of vmxnet3.
If a BPF program for native XDP adds an encapsulation header such as
IPIP and transmits the packet out the same interface, then in case
of vmxnet3 a corrupted packet is being sent and subsequently dropped
on the path.
vmxnet3_xdp_xmit_frame() which is called e.g. via vmxnet3_run_xdp()
through vmxnet3_xdp_xmit_back() calculates an incorrect DMA address:
page = virt_to_page(xdpf->data);
tbi->dma_addr = page_pool_get_dma_addr(page) +
VMXNET3_XDP_HEADROOM;
dma_sync_single_for_device(&adapter->pdev->dev,
tbi->dma_addr, buf_size,
DMA_TO_DEVICE);
The above assumes a fixed offset (VMXNET3_XDP_HEADROOM), but the XDP
BPF program could have moved xdp->data. While the passed buf_size is
correct (xdpf->len), the dma_addr needs to have a dynamic offset which
can be calculated as xdpf->data - (void *)xdpf, that is, xdp->data -
xdp->data_hard_start.
Fixes: 54f00cce11 ("vmxnet3: Add XDP support.")
Reported-by: Andrew Sauber <andrew.sauber@isovalent.com>
Reported-by: Nikolay Nikolaev <nikolay.nikolaev@isovalent.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Tested-by: Nikolay Nikolaev <nikolay.nikolaev@isovalent.com>
Acked-by: Anton Protopopov <aspsk@isovalent.com>
Cc: William Tu <witu@nvidia.com>
Cc: Ronak Doshi <ronak.doshi@broadcom.com>
Link: https://patch.msgid.link/a0888656d7f09028f9984498cc698bb5364d89fc.1728931137.git.daniel@iogearbox.net
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The patch adds native-mode XDP support: XDP DROP, PASS, TX, and REDIRECT.
Background:
The vmxnet3 rx consists of three rings: ring0, ring1, and dataring.
For r0 and r1, buffers at r0 are allocated using alloc_skb APIs and dma
mapped to the ring's descriptor. If LRO is enabled and packet size larger
than 3K, VMXNET3_MAX_SKB_BUF_SIZE, then r1 is used to mapped the rest of
the buffer larger than VMXNET3_MAX_SKB_BUF_SIZE. Each buffer in r1 is
allocated using alloc_page. So for LRO packets, the payload will be in one
buffer from r0 and multiple from r1, for non-LRO packets, only one
descriptor in r0 is used for packet size less than 3k.
When receiving a packet, the first descriptor will have the sop (start of
packet) bit set, and the last descriptor will have the eop (end of packet)
bit set. Non-LRO packets will have only one descriptor with both sop and
eop set.
Other than r0 and r1, vmxnet3 dataring is specifically designed for
handling packets with small size, usually 128 bytes, defined in
VMXNET3_DEF_RXDATA_DESC_SIZE, by simply copying the packet from the backend
driver in ESXi to the ring's memory region at front-end vmxnet3 driver, in
order to avoid memory mapping/unmapping overhead. In summary, packet size:
A. < 128B: use dataring
B. 128B - 3K: use ring0 (VMXNET3_RX_BUF_SKB)
C. > 3K: use ring0 and ring1 (VMXNET3_RX_BUF_SKB + VMXNET3_RX_BUF_PAGE)
As a result, the patch adds XDP support for packets using dataring
and r0 (case A and B), not the large packet size when LRO is enabled.
XDP Implementation:
When user loads and XDP prog, vmxnet3 driver checks configurations, such
as mtu, lro, and re-allocate the rx buffer size for reserving the extra
headroom, XDP_PACKET_HEADROOM, for XDP frame. The XDP prog will then be
associated with every rx queue of the device. Note that when using dataring
for small packet size, vmxnet3 (front-end driver) doesn't control the
buffer allocation, as a result we allocate a new page and copy packet
from the dataring to XDP frame.
The receive side of XDP is implemented for case A and B, by invoking the
bpf program at vmxnet3_rq_rx_complete and handle its returned action.
The vmxnet3_process_xdp(), vmxnet3_process_xdp_small() function handles
the ring0 and dataring case separately, and decides the next journey of
the packet afterward.
For TX, vmxnet3 has split header design. Outgoing packets are parsed
first and protocol headers (L2/L3/L4) are copied to the backend. The
rest of the payload are dma mapped. Since XDP_TX does not parse the
packet protocol, the entire XDP frame is dma mapped for transmission
and transmitted in a batch. Later on, the frame is freed and recycled
back to the memory pool.
Performance:
Tested using two VMs inside one ESXi vSphere 7.0 machine, using single
core on each vmxnet3 device, sender using DPDK testpmd tx-mode attached
to vmxnet3 device, sending 64B or 512B UDP packet.
VM1 txgen:
$ dpdk-testpmd -l 0-3 -n 1 -- -i --nb-cores=3 \
--forward-mode=txonly --eth-peer=0,<mac addr of vm2>
option: add "--txonly-multi-flow"
option: use --txpkts=512 or 64 byte
VM2 running XDP:
$ ./samples/bpf/xdp_rxq_info -d ens160 -a <options> --skb-mode
$ ./samples/bpf/xdp_rxq_info -d ens160 -a <options>
options: XDP_DROP, XDP_PASS, XDP_TX
To test REDIRECT to cpu 0, use
$ ./samples/bpf/xdp_redirect_cpu -d ens160 -c 0 -e drop
Single core performance comparison with skb-mode.
64B: skb-mode -> native-mode
XDP_DROP: 1.6Mpps -> 2.4Mpps
XDP_PASS: 338Kpps -> 367Kpps
XDP_TX: 1.1Mpps -> 2.3Mpps
REDIRECT-drop: 1.3Mpps -> 2.3Mpps
512B: skb-mode -> native-mode
XDP_DROP: 863Kpps -> 1.3Mpps
XDP_PASS: 275Kpps -> 376Kpps
XDP_TX: 554Kpps -> 1.2Mpps
REDIRECT-drop: 659Kpps -> 1.2Mpps
Demo: https://youtu.be/4lm1CSCi78Q
Future work:
- XDP frag support
- use napi_consume_skb() instead of dev_kfree_skb_any at unmap
- stats using u64_stats_t
- using bitfield macro BIT()
- optimization for DMA synchronization using actual frame length,
instead of always max_len
Signed-off-by: William Tu <u9012063@gmail.com>
Reviewed-by: Alexander Duyck <alexanderduyck@fb.com>
Reviewed-by: Alexander Lobakin <alexandr.lobakin@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
