> It seems like Yury's commit 7165754b6b5f3b7c07050d921fa1c58bba5f0ff1 made sendfile tests of test_asyncio 6x slower on FreeBSD :-(

Well, yeah, I kind of expected to see a performance regression here.  Before, we were reading from a file directly, which is a blocking operation and can potentially block the event loop.  Now we're using a thread pool to do the read, but its more overhead.

In this case I think we have to do it the *correct* way, even if it means sacrificing the performance.  As a workaround, I propose to rise the read buffer size to 1Mb.