On 2024-02-13 14:31, Eric Dumazet wrote:
On Tue, Feb 13, 2024 at 7:39 PM Paolo Abeni <pabeni@redhat.com> wrote:
On Tue, 2024-02-13 at 16:49 +0100, Eric Dumazet wrote:
On Tue, Feb 13, 2024 at 4:28 PM Paolo Abeni <pabeni@redhat.com> wrote:
On Tue, 2024-02-13 at 14:34 +0100, Eric Dumazet wrote:
This sk_peek_offset protocol, needing  sk_peek_offset_bwd() in the non
MSG_PEEK case is very strange IMO.

Ideally, we should read/write over sk_peek_offset only when MSG_PEEK
is used by the caller.

That would only touch non fast paths.

Since the API is mono-threaded anyway, the caller should not rely on
the fact that normal recvmsg() call
would 'consume' sk_peek_offset.
Storing in sk_peek_seq the tcp next sequence number to be peeked should
avoid changes in the non MSG_PEEK cases.

AFAICS that would need a new get_peek_off() sock_op and a bit somewhere
(in sk_flags?) to discriminate when sk_peek_seq is actually set. Would
that be acceptable?
We could have a parallel SO_PEEK_OFFSET option, reusing the same socket field.

The new semantic would be : Supported by TCP (so far), and tcp
recvmsg() only reads/writes this field when MSG_PEEK is used.
Applications would have to clear the values themselves.
I feel like there is some misunderstanding, or at least I can't follow.
Let me be more verbose, to try to clarify my reasoning.

Two consecutive recvmsg(MSG_PEEK) calls for TCP after SO_PEEK_OFF will
return adjacent data. AFAICS this is the same semantic currently
implemented by UDP and unix sockets.

Currently 'sk_peek_off' maintains the next offset to be peeked into the
current receive queue. To implement the above behaviour, tcp_recvmsg()
has to update 'sk_peek_off' after MSG_PEEK, to move the offset to the
next data, and after a plain read, to account for the data removed from
the receive queue.

I proposed to let introduce a tcp-specific set_peek_off doing something
alike:

        WRTIE_ONCE(sk->sk_peek_off, tcp_sk(sk)->copied_seq + val);

so that the recvmsg will need to update sk_peek_off only for MSG_PEEK,
while retaining the semantic described above.

To keep the userspace interface unchanged that will need a paired
tcp_get_peek_off(), so that getsockopt(SO_PEEK_OFF) could return to the
user a plain offset. An additional bit flag will be needed to store the
information "the user-space enabled peek with offset".

I don't understand how a setsockopt(PEEK_OFFSET) variant would help
avoiding touching sk->sk_peek_offset?

I was trying to avoid using an extra storage, I was not trying to
implement the alternative myself :0)

If the recvmsg( MSG_PEEK) is supposed to auto-advance the peek_offset,
we probably need more than a mere 32bit field.

I wonder if the following could be acceptable:

if (flags & MSG_PEEK)
       sk_peek_offset_fwd(sk, used);
else if (peek_offset > 0)
      sk_peek_offset_bwd(sk, used);

peek_offset is already present in the data cache, and if it has the value
zero it means either that that sk->sk_peek_off is unused (-1) or actually is zero.
Either way, no rewind is needed in that case.

///jon



      
Thanks!

Paolo