#12011 Speed up small TLS writes

[itamarst]

Fixes #12011.

Benchmark results with Python 3.11:

trunk:

$ time python tls_benchmark.py small

real    0m2.425s
user    0m2.302s
sys     0m0.069s

This branch:

$ time python tls_benchmark.py small

real    0m1.705s
user    0m1.562s
sys     0m0.088s

Benchmark:

import sys

from twisted.internet.protocol import Protocol
from twisted.internet.ssl import CertificateOptions, PrivateCertificate
from twisted.protocols.loopback import loopbackTCP
from twisted.internet.task import react
from twisted.internet import reactor

class Client(Protocol):
    def connectionMade(self):
        self.count = 10_000
        self.transport.startTLS(CertificateOptions(verify=False, requireCertificate=False))
        assert sys.argv[1] in ("small", "big")
        self.next_iteration()

    def next_iteration(self):
        self.count -= 1
        if sys.argv[1] == "small":
            for i in range(1_000):
                self.transport.write(b"X")
        else:
            self.transport.write(b"X" * 1000)

        if self.count == 0:
            self.transport.loseConnection()
        else:
            reactor.callLater(0, self.next_iteration)


class Server(Protocol):
    received = 0

    def connectionMade(self):
        with open("server.pem", "rb") as f:
            options = PrivateCertificate.loadPEM(f.read()).options()
        self.transport.startTLS(options)

    def dataReceived(self, data):
        self.received += len(data)

    def connectionLost(self, reason):
        assert self.received == 10_000_000, self.received


assert len(sys.argv) == 2
react(lambda reactor: loopbackTCP(Server(), Client(), noisy=True))

[adiroiban]

Hi Itamar, Thanks for looking into this. Is this PR ready for review? Regards

[itamarst]

Yes, it's ready for review.

[adiroiban]

Looks good. Thanks. Only minor comments.

[adiroiban]

It's not pretty, but I think is the best option.

I think that, in the end, having the _AggregateSmallWrites code using composition is nicer.

[adiroiban]

Maybe leave this test as a reminder that passing bytes should never be supported again.

But I am also ok to have this removed.

[itamarst]

The problem is it doesn't fail in a useful way now.

[adiroiban]

I think the important part is to make sure that it fails... in any way. Rather than ending up with a case in which the data is ignored... or something like that.

So even if the error is not explicit. I think that it still important to make sure that it fails.

The previous explicit error was a huge help for my in the past... but I am ok to have something less helful, in the name of performance.

I hope in the future mypy can help here.

[itamarst]

OK I restored the test.