On 8/24/22 11:29 AM, Xin Long wrote:
On Wed, Aug 17, 2022 at 4:11 PM Adel Abouchaev adel.abushaev@gmail.com wrote:
QUIC requires end to end encryption of the data. The application usually prepares the data in clear text, encrypts and calls send() which implies multiple copies of the data before the packets hit the networking stack. Similar to kTLS, QUIC kernel offload of cryptography reduces the memory pressure by reducing the number of copies.
The scope of kernel support is limited to the symmetric cryptography, leaving the handshake to the user space library. For QUIC in particular, the application packets that require symmetric cryptography are the 1RTT packets with short headers. Kernel will encrypt the application packets on transmission and decrypt on receive. This series implements Tx only, because in QUIC server applications Tx outweighs Rx by orders of magnitude.
Supporting the combination of QUIC and GSO requires the application to correctly place the data and the kernel to correctly slice it. The encryption process appends an arbitrary number of bytes (tag) to the end of the message to authenticate it. The GSO value should include this overhead, the offload would then subtract the tag size to parse the input on Tx before chunking and encrypting it.
With the kernel cryptography, the buffer copy operation is conjoined with the encryption operation. The memory bandwidth is reduced by 5-8%. When devices supporting QUIC encryption in hardware come to the market, we will be able to free further 7% of CPU utilization which is used today for crypto operations.
Adel Abouchaev (6): Documentation on QUIC kernel Tx crypto. Define QUIC specific constants, control and data plane structures Add UDP ULP operations, initialization and handling prototype functions. Implement QUIC offload functions Add flow counters and Tx processing error counter Add self tests for ULP operations, flow setup and crypto tests
Documentation/networking/index.rst | 1 + Documentation/networking/quic.rst | 185 ++++ include/net/inet_sock.h | 2 + include/net/netns/mib.h | 3 + include/net/quic.h | 63 ++ include/net/snmp.h | 6 + include/net/udp.h | 33 + include/uapi/linux/quic.h | 60 + include/uapi/linux/snmp.h | 9 + include/uapi/linux/udp.h | 4 + net/Kconfig | 1 + net/Makefile | 1 + net/ipv4/Makefile | 3 +- net/ipv4/udp.c | 15 + net/ipv4/udp_ulp.c | 192 ++++ net/quic/Kconfig | 16 + net/quic/Makefile | 8 + net/quic/quic_main.c | 1417 ++++++++++++++++++++++++ net/quic/quic_proc.c | 45 + tools/testing/selftests/net/.gitignore | 4 +- tools/testing/selftests/net/Makefile | 3 +- tools/testing/selftests/net/quic.c | 1153 +++++++++++++++++++ tools/testing/selftests/net/quic.sh | 46 + 23 files changed, 3267 insertions(+), 3 deletions(-) create mode 100644 Documentation/networking/quic.rst create mode 100644 include/net/quic.h create mode 100644 include/uapi/linux/quic.h create mode 100644 net/ipv4/udp_ulp.c create mode 100644 net/quic/Kconfig create mode 100644 net/quic/Makefile create mode 100644 net/quic/quic_main.c create mode 100644 net/quic/quic_proc.c create mode 100644 tools/testing/selftests/net/quic.c create mode 100755 tools/testing/selftests/net/quic.sh
base-commit: fd78d07c7c35de260eb89f1be4a1e7487b8092ad
2.30.2
Hi, Adel,
I don't see how the key update(rfc9001#section-6) is handled on the TX path, which is not using TLS Key update, and "Key Phase" indicates which key will be used after rekeying. Also, I think it is almost impossible to handle the peer rekeying on the RX path either based on your current model in the future.
The update is not present in these patches, but it is an important part of the QUIC functionality. As this patch is only storing a single key, you are correct that this approach does not handle the key rotation. To implement re-keying on Tx and on Rx a rolling secret will need to be stored in kernel. In that case, the subsequent 1RTT (Application space) keys will be refreshed by the kernel. After all, when the hardware is mature enough to support QUIC encryption and decryption - the secret will need to be kept in the hardware to react on time on Rx, especially. Tx path could solicit the re-key at any point or by the exhaustion of the counter of GCM (packet number in this case). The RFC expects the implementation to retain 2 keys, at least, while keeping 3 (old, current and next) is not prohibited either. Keeping more is not necessary.
The patch seems to get the crypto_ctx by doing a connection hash table lookup in the sendmsg(), which is not good from the performance side. One QUIC connection can go over multiple UDP sockets, but I don't think one socket can be used by multiple QUIC connections. So why not save the ctx in the socket instead?
A single socket could have multiple connections originated from it, having different destinations, if the socket is not connected. An optimization could be made for connected sockets to cache the context and save time on a lookup. The measurement of kernel operations timing did not reveal a significant amount of time spent in this lookup due to a relatively small number of connections per socket in general. A shared table across multiple sockets might experience a different performance grading.
The patch is to reduce the copying operations between user space and the kernel. I might miss something in your user space code, but the msg to send is *already packed* into the Stream Frame in user space, what's the difference if you encrypt it in userspace and then sendmsg(udp_sk) with zero-copy to the kernel.
It is possible to do it this way. Zero-copy works best with packet sizes starting at 32K and larger. Anything less than that would consume the improvements of zero-copy by zero-copy pre/post operations and needs to align memory. The other possible obstacle would be that eventual support of QUIC encryption and decryption in hardware would integrate well with this current approach.
Didn't really understand the "GSO" you mentioned, as I don't see any code about kernel GSO, I guess it's just "Fragment size", right? BTW, it‘s not common to use "//" for the kernel annotation.
Once the payload arrives into the kernel, the GSO on the interface would instruct L3/L4 stack on fragmentation. In this case, the plaintext QUIC packets should be aligned on the GSO marks less the tag size that would be added by encryption. For GSO size 1000, the QUIC packets in the batch for transmission should all be 984 bytes long, except maybe the last one. Once the tag is attached, the new size of 1000 will correctly split the QUIC packets further down the stack for transmission in individual IP/UDP packets. The code is also saving processing time by sending all packets at once to UDP in a single call, when GSO is enabled.
I'm not sure if it's worth adding a ULP layer over UDP for this QUIC TX only. Honestly, I'm more supporting doing a full QUIC stack in the kernel independently with socket APIs to use it: https://github.com/lxin/tls_hs.
Thanks.