.. | ||
async.h | ||
concurrent.h | ||
concurrent_for_each.h | ||
condition_variable.h | ||
context.h | ||
continuation.h | ||
critical_assertion.h | ||
critical_indicator.h | ||
ctx.h | ||
dock.h | ||
exception_handler.h | ||
fault.h | ||
future.h | ||
latch.h | ||
list.h | ||
mutex.h | ||
ole.h | ||
pool.h | ||
prof.h | ||
promise.h | ||
queue.h | ||
README.md | ||
scope_notify.h | ||
shared_mutex.h | ||
shared_state.h | ||
slice_usage_warning.h | ||
stack_usage_assertion.h | ||
syscall_usage_warning.h | ||
this_ctx.h | ||
uninterruptible.h | ||
unlock_guard.h | ||
upgrade_lock.h | ||
view.h | ||
when.h |
Userspace Context Switching
The ircd::ctx
subsystem is a userspace threading library meant to regress
the asynchronous callback pattern back to synchronous suspensions. These are
stackful coroutines which provide developers with more intuitive control in
environments which conduct frequent I/O which would otherwise break up a single
asynchronous stack into callback-hell.
Motivation
Userspace threads are an alternative to using posix kernel threads as a way to develop intuitively-stackful programs in applications which are primarily I/O-bound rather than CPU-bound. This is born out of a recognition that a single CPU core has enough capacity to compute the entirety of all requests for an efficiently-written network daemon if I/O were instantaneous; if one could use a single thread it is advantageous to do so right up until the compute-bound is reached, rather than introducing more threads for any other reason. The limits to single-threading and scaling beyond a single CPU is then pushed to higher-level application logic: either message-passing between multiple processes (or machines in a cluster), and/or threads which have extremely low interference.
ircd::ctx
allows for a very large number of contexts to exist, on the order
of thousands or more, and still efficiently make progress without the overhead
of kernel context switches. As an anecdotal example, a kernel context switch
from a contended mutex could realistically be five to ten times more costly
than a userspace context switch if not significantly more, and with effects
that are less predictable. Contexts will accomplish as much work as possible
in a "straight line" before yielding to the kernel to wait for the completion
of any I/O event.
Foundation
This library is based in boost::coroutine / boost::context
which wraps
the register save/restores in a cross-platform way in addition to providing
properly mmap(NOEXEC)'ed
etc memory appropriate for stacks on each platform.
boost::asio
has then added its own comprehensive integration with the above
libraries eliminating the need for us to worry about a lot of boilerplate to
de-async the asio networking calls. See: boost::asio::spawn.
This is a nice boost, but that's as far as it goes. The rest is on us here to actually make a threading library.
Interface
We mimic the standard library std::thread
suite as much as possible (which
mimics the boost::thread
library) and offer alternative threading primitives
for these userspace contexts rather than those for operating system threads in
std::
such as ctx::mutex
and ctx::condition_variable
and ctx::future
among others.
- The primary user object is
ircd::context
(orircd::ctx::context
) which has anstd::thread
interface.
Context Switching
A context switch has the overhead of a heavy function call -- a function with a bunch of arguments (i.e the registers being saved and restored). We consider this fast and our philosophy is to not think about the context switch itself as a bad thing to be avoided for its own sake.
This system is also fully integrated both with the IRCd core
boost::asio::io_service
event loop and networking systems. There are actually
several types of context switches going on here built on two primitives:
-
Direct jump: This is the fastest switch. Context
A
can yield to contextB
directly ifA
knows aboutB
and if it knows thatB
is in a state ready to resume from a direct jump and thatA
will also be further resumed somehow. This is not always suitable in practice so other techniques may be used instead. -
Queued wakeup: This is the common default and safe switch. This is where the context system integrates with the
boost::asio::io_service
event loop. The execution of a "slice" as we'll call a yield-to-yield run of non-stop computation is analogous to a function posted to theio_service
in the asynchronous pattern. ContextA
can enqueue contextB
if it knows aboutB
and then choose whether to yield or not to yield. In any case theio_service
queue will simply continue to the next task which isn't guaranteed to beB
.
When does Context Switching (yielding) occur?
Bottom line is that this is simply not javascript. There are no
stack-clairvoyant keywords like await
which explicitly indicate to everyone
everywhere that the overall state of the program before and after any totally
benign-looking function call will be different. This is indeed multi-threaded
programming but in a very PG-13 rated way. You have to assume that if you
aren't sure some function has a "deep yield" somewhere way up the stack that
there is a potential for yielding in that function. Unlike real concurrent
threading everything beyond this is much easier.
-
Anything directly on your stack is safe (same with real MT anyway).
-
static
and global assets are safe if you can assert no yielding. Such an assertion can be made with an instance ofircd::ctx::critical_assertion
. Note that we try to usethread_local
rather thanstatic
to still respect any real multi-threading that may occur now or in the future. -
Calls which may yield and do IO may be marked with
[GET]
and[SET]
conventional labels but they may not be. Some reasoning about obvious yields and a zen-like awareness is always recommended.