On 9/27/22 10:12 AM, Willem de Bruijn wrote:
On Tue, Sep 27, 2022 at 12:45 PM Adel Abouchaev <adel.abushaev@xxxxxxxxx> wrote:
On 9/25/22 11:04 AM, Willem de Bruijn wrote:
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.
I'm late to this patch series, sorry. High quality implementation. I
have a few design questions similar to Xin.
If multiplexing, instead of looking up a connection by { address, port
variable length connection ID }, perhaps return a connection table
index on setsockopt and use that in sendmsg.
It was deliberate to not to return anything other than 0 from
setsockopt() as defined in the spec for the function. Despite that it
says "shall", the doc says that 0 is the only value for successful
operation. This was the reason not to use setsockopt() for any
bidirectional transfers of data and or status. A more sophisticated
approach with netlink sockets would be more suitable for it. The second
reason is the API asymmetry for Tx and Rx which will be introduced - the
I thought the cover letter indicated that due to asymmetry of most
QUIC workloads, only Tx offload is implemented. You do intend to add
Rx later?
We are planning to include Rx later as well, Tx is more compelling from
use case perspective as the main application of this would go to an
Edge, where Tx is heavy and Rx is much lighter.
Rx will still need to match on the address, port and cid. The third
reason is that in current implementations there are no more than a few
connections per socket,
This is very different from how we do QUIC at Google. There we
definitely multiplex many connections across essentially a socket per
CPU IIRC.
Ian mentioned that for such a case, while using a zero length QUIC CID,
the remote ends are ephemeral ports and are different. In which case
each connection will have its own context and yes, it will load the hash
table quite a bit for a socket. The limiting factor is still a BSD4.4
interface and returning a value from setsockopt() with the entry ID. It
might be more suitable once a standard netlink interface is available
for this to easily plug in. However, this will further complicate a
userspace library integration which is a not so trivial task today too.
which does not abuse the rhashtable that does a
lookup, although it takes time to hash the key into a hash for a seek.
The performance measurement ran against the runtime and did not flag
this path as underperforming either, there were other parts that
substantially add to the runtime, not the key lookup though.
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.
Part of the cost of MSG_ZEROCOPY is in mapping and unmapping user
pages. This series re-implements that with its own get_user_pages.
That is duplicative non-trivial code. And it will incur the same cost.
What this implementation saves is the (indeed non-trivial)
asynchronous completion notification over the error queue.
The cover letter gives some performance numbers against a userspace
implementation that has to copy from user to kernel. It might be more
even to compare against an implementation using MSG_ZEROCOPY and
UDP_SEGMENT. A userspace crypto implementation may have other benefits
compared to a kernel implementation, such as not having to convert to
crypto API scatter-gather arrays and back to network structures.
A few related points
- The implementation support multiplexed connections, but only one
crypto sendmsg can be outstanding at any time:
+ /**
+ * To synchronize concurrent sendmsg() requests through the same socket
+ * and protect preallocated per-context memory.
+ **/
+ struct mutex sendmsg_mux;
That is quite limiting for production workloads.
The use case that we have with MVFST library currently runs a single
worker for a connection and has a single socket attached to it. QUIC
allows simultaneous use of multiple connection IDs to swap them in
runtime, and implementation would request only a handful of these. The
MVFST batches writes into a block of about 8Kb and then uses GSO to send
them all at once.
- Crypto operations are also executed synchronously, using
crypto_wait_req after each operationn. This limits throughput by using
at most one core per UDP socket. And adds sendmsg latency (which may
or may not be important to the application). Wireguard shows an
example of how to parallelize software crypto across cores.
- The implementation avoids dynamic allocation of cipher text pages by
using a single ctx->cipher_page. This is protected by sendmsg_mux (see
above). Is that safe when packets leave the protocol stack and are
then held in a qdisc or when being processed by the NIC?
quic_sendmsg_locked will return, but the cipher page is not free to
reuse yet.
There is currently no use case that we have in hands that requires
parallel transmission of data for the same connection. Multiple
connections would have no issue running in parallel as each of them will
have it's own preallocated cipher_page in the context.
This still leaves the point that sendmsg may return and the mutex
released while the cipher_page is still associated with an skb in the
transmit path.
Correct, there is a further copy from the cipher buffer into fragmented
pieces done by ip_generic_getfrag(). Am I reading it wrong that it would
not need the cipher text allocated buffer after that is done and all the
data is copied from it into further structures using copy_from_iter()?
The skb would be built upon a different buffer space than the encrypted
data. The udp_sendmsg() assumes that it receives the memory from
userspace and does all these ops to move stuff from userspace. While
doing that, it would omit tons of it's checks as the memory is already
in kernel and still execute quickly.
There is a fragmentation further down the stack with
ip_generic_getfrag() that eventually does copy_from_iter() and makea a
copy of the data. This is executed as part of __ip_append_data() called
from udp_sendmsg() in ipv4/udp.c. The assumption was that this is
executed synchronously and the queues and NIC will see a mapping of a
different memory area than the ciphertext in the pre-allocated page.
- The real benefit of kernel QUIC will come from HW offload. Would it
be better to avoid the complexity of an in-kernel software
implementation and only focus on HW offload? Basically, pass the
plaintext QUIC packets over a standard UDP socket and alongside in a
cmsg pass either an index into a HW security association database or
the immediate { key, iv } connection_info (for stateless sockets), to
be encoded into the descriptor by the device driver.
Hardware usually targets a single ciphersuite such as AES-GCM-128/256,
while QUIC also supports Chacha20-Poly1305 and AES-CCM. The generalized
support for offload prompted implementation of these ciphers in kernel
code.
All userspace libraries also support all protocols as fall-back. No
need for two fall-backs if HW support is missing?
Could be done that way. Looking at TLS - it has fallback to kernel
though in tls_sw, although that is a dual purpose code.
The kernel code could also engage if the future hardware has
capacity caps preventing it from handling all requests in the hardware.
- With such a simpler path, could we avoid introducing ULP and just
have udp [gs]etsockopt CRYPTO_STATE. Where QUIC is the only defined
state type yet.
- Small aside: as the series introduces new APIs with non-trivial
parsing in the kernel, it's good to run a fuzzer like syzkaller on it
(if not having done so yet).
Agreed.
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.
minor point: fragment has meaning in IPv4. For GSO, prefer gso_size.
Sure, will change it to gso_size.
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.
