On 08/02/2024 18:02, Andy Lutomirski wrote:

I’ve been using OPT_ID-style timestamping for years, but for some
reason this issue only bit me last week: if sendmsg() fails on a UDP
or ping socket, sk_tskey is poorly.  It may or may not get incremented
by the failed sendmsg().

Well, there are several error paths, for sure. For the sockets you
mention the increment of tskey happens at __ip{,6}_append_data. There
are 2 different types of failures which can happen after the increment.
The first is MTU check fail, another one is memory allocation failures.
I believe we can move increment to a later position, after MTU check in
both functions to avoid first type of problem.

I can think of at least three ways to improve this:

1. Make it so that the sequence number is genuinely only incremented
on success. This may be tedious to implement and may be nearly
impossible if there are multiple concurrent sendmsg() calls on the
same socket.

Multiple concurrent sendmsg() should bring a lot of problems on user-
space side. With current implementation the application has to track the
value of tskey to check incoming TX timestamp later. But with parallel
sendmsg() the app cannot be sure which value is assigned to which call
even in case of proper track value synchronization. That brings us to
the other solutions if we consider having parallel threads working with
same socket. Or we can simply pretend that it's impossible and then fix
error path to decrement tskey value.

2. Allow the user program to specify an explicit ID.  cmsg values are
variable length, so for datagram sockets, extending the
SO_TIMESTAMPING cmsg with 64 bits of sequence number to be used for
the TX timestamp on that particular packet might be a nice solution.

This option can be really useful in case of really parallel work with
sockets.

3. Add a getsockopt to read sk_tskey, which user code could use to
determine the next sequence number after a failed sendmsg() call.

I don't believe it's usable interface. Especially if we think about
multiple threads working with the same socket.

Thanks,
Andy