(Not saying I agree or disagree, just saying that this isn't so much an ancient artifact as it is a fairly recent and purposeful addition.)
-- NT is missing sendfile()? Also syscalls are fast on x86_64, most stuff is dynamic now a days and serving static files straight from page cache is fast enough.
• Kernel-mode request queuing. Requests cause less overhead in context switching
-- This one is somewhat valid but given you can ctx switch pretty fast on modern CPUs it is not clear how big the gains are.
Yeah but there would be some gains - that's not deniable. Question is whether they are big enough to risk something like the vulnerability at hand.
They were in use (at least TransmitFile was) by IIS long before http.sys was conceived.
HTTP.sys obviously does a lot more, but especially back in the Windows 2000 days, the performance difference was substantial.
I am curious as to what the reasons are - what are the benefits of this that outweigh the huge risk (as demonstrated repeatedly by the various vulnerabilities) of running in HTTP code in the kernel? Only thing I can think of is they are trying to squeeze the last bit of performance / scalability but surely that could only mean that there are parts of the kernel (ctx switching overhead, threading scalability, page cache performance, etc.) that need to be fixed in order to not need these type of hacks in 2015?