+37
−0
Loading
Merge branch 'TLS-offload-rx-netdev-and-mlx5'
Boris Pismenny says:
====================
TLS offload rx, netdev & mlx5
The following series provides TLS RX inline crypto offload.
v5->v4:
- Remove the Kconfig to mutually exclude both IPsec and TLS
v4->v3:
- Remove the iov revert for zero copy send flow
v2->v3:
- Fix typo
- Adjust cover letter
- Fix bug in zero copy flows
- Use network byte order for the record number in resync
- Adjust the sequence provided in resync
v1->v2:
- Fix bisectability problems due to variable name changes
- Fix potential uninitialized return value
This series completes the generic infrastructure to offload TLS crypto to
a network devices. It enables the kernel TLS socket to skip decryption and
authentication operations for SKBs marked as decrypted on the receive
side of the data path. Leaving those computationally expensive operations
to the NIC.
This infrastructure doesn't require a TCP offload engine. Instead, the
NIC decrypts a packet's payload if the packet contains the expected TCP
sequence number. The TLS record authentication tag remains unmodified
regardless of decryption. If the packet is decrypted successfully and it
contains an authentication tag, then the authentication check has passed.
Otherwise, if the authentication fails, then the packet is provided
unmodified and the KTLS layer is responsible for handling it.
Out-Of-Order TCP packets are provided unmodified. As a result,
in the slow path some of the SKBs are decrypted while others remain as
ciphertext.
The GRO and TCP layers must not coalesce decrypted and non-decrypted SKBs.
At the worst case a received TLS record consists of both plaintext
and ciphertext packets. These partially decrypted records must be
reencrypted, only to be decrypted.
The notable differences between SW KTLS and NIC offloaded TLS
implementations are as follows:
1. Partial decryption - Software must handle the case of a TLS record
that was only partially decrypted by HW. This can happen due to packet
reordering.
2. Resynchronization - tls_read_size calls the device driver to
resynchronize HW whenever it lost track of the TLS record framing in
the TCP stream.
The infrastructure should be extendable to support various NIC offload
implementations. However it is currently written with the
implementation below in mind:
The NIC identifies packets that should be offloaded according to
the 5-tuple and the TCP sequence number. If these match and the
packet is decrypted and authenticated successfully, then a syndrome
is provided to software. Otherwise, the packet is unmodified.
Decrypted and non-decrypted packets aren't coalesced by the network stack,
and the KTLS layer decrypts and authenticates partially decrypted records.
The NIC provides an indication whenever a resync is required. The resync
operation is triggered by the KTLS layer while parsing TLS record headers.
Finally, we measure the performance obtained by running single stream
iperf with two Intel(R) Xeon(R) CPU E5-2620 v3 @ 2.40GHz machines connected
back-to-back with Innova TLS (40Gbps) NICs. We compare TCP (upper bound)
and KTLS-Offload running both in Tx and Rx. The results show that the
performance of offload is comparable to TCP.
| Bandwidth (Gbps) | CPU Tx (%) | CPU rx (%)
TCP | 28.8 | 5 | 12
KTLS-Offload-Tx-Rx | 28.6 | 7 | 14
Paper: https://netdevconf.org/2.2/papers/pismenny-tlscrypto-talk.pdf
====================
Signed-off-by: 
David S. Miller <davem@davemloft.net>