1 PublicInbox::DS - event loop and async I/O base class
3 Our PublicInbox::DS event loop which powers public-inbox-nntpd
4 and public-inbox-httpd diverges significantly from the
5 unmaintained Danga::Socket package we forked from. In fact,
6 it's probably different from most other event loops out there.
10 * There is one and only one callback: ->event_step. Unlike other
11 event loops, there are no separate callbacks for read, write,
12 error or hangup events. In fact, we never care which kevent
13 filter or poll/epoll event flag (e.g. POLLIN/POLLOUT/POLLHUP)
16 The lack of read/write callback distinction is driven by the
17 fact TLS libraries (e.g. OpenSSL via IO::Socket::SSL) may
18 declare SSL_WANT_READ on SSL_write(), and SSL_WANT_READ on
19 SSL_read(). So we end up having to let each user object decide
20 whether it wants to make read or write calls depending on its
21 internal state, completely independent of the event loop.
23 Error and hangup (POLLERR and POLLHUP) callbacks are redundant and
24 only triggered in rare cases. They're redundant because the
25 result of every read and write call in ->event_step must be
26 checked, anyways. At best, callbacks for POLLHUP and POLLERR can
27 save one syscall per socket lifetime and not worth the extra code
30 Reducing the user-supplied code down to a single callback allows
31 subclasses to keep their logic self-contained. The combination
32 of this change and one-shot wakeups (see below) for bidirectional
33 data flows make asynchronous code easier to reason about.
37 * ->write buffering uses temporary files whereas Danga::Socket used
38 the heap. The rationale for this is the kernel already provides
39 ample (and configurable) space for socket buffers. Modern kernels
40 also cache FS operations aggressively, so systems with ample RAM
41 are unlikely to notice degradation, while small systems are less
42 likely to suffer unpredictable heap fragmentation, swap and OOM
45 In the future, we may introduce sendfile and mmap+SSL_write to
46 reduce data copies, and use FALLOC_FL_PUNCH_HOLE on Linux to
47 release space after the buffer is partially cleared.
51 * obj->write(CODEREF) passes the object itself to the CODEREF
52 Being able to enqueue subroutine calls is a powerful feature in
53 Danga::Socket for keeping linear logic in an asynchronous environment.
54 Unfortunately, each subroutine takes several kilobytes of memory.
55 One small change to Danga::Socket is to pass the receiver object
56 (aka "$self") to the CODEREF. $self can store any necessary
57 state it needs for a normal (named) subroutine. This allows us to
58 put the same sub into multiple queues without paying a large
59 memory penalty for each one.
61 This idea is also more easily ported to C or other languages which
62 lack anonymous subroutines (aka "closures").
64 * ->requeue support. An optimization of the AddTimer(0, ...) idiom
65 for immediately dispatching code at the next event loop iteration.
66 public-inbox uses this for fairly generating large responses
67 iteratively (see PublicInbox::NNTP::long_response or ibx_async_cat
72 * One-shot wakeups allowed via EPOLLONESHOT or EV_DISPATCH. These
73 flags allow us to simplify code in ->event_step callbacks for
74 bidirectional sockets (NNTP and HTTP). Instead of merely reacting
75 to events, control is handed over at ->event_step in one-shot scenarios.
76 The event_step caller (NNTP || HTTP) then becomes proactive in declaring
77 which (if any) events it's interested in for the next loop iteration.
79 * Edge-triggering available via EPOLLET or EV_CLEAR. These reduce wakeups
80 for unidirectional classes when throughput is more important than fairness.
82 * IO::Socket::SSL support (for NNTPS, STARTTLS+NNTP, HTTPS)
84 * dwaitpid (waitpid wrapper) support for reaping dead children
86 * reliable signal wakeups are supported via signalfd on Linux,
87 EVFILT_SIGNAL on *BSDs via IO::KQueue.
91 * Many fields removed or moved to subclasses, so the underlying
92 hash is smaller and suitable for FDs other than stream sockets.
93 Some fields we enforce (e.g. wbuf, wbuf_off) are autovivified
94 on an as-needed basis to save memory when they're not needed.
96 * TCP_CORK support removed, instead we use MSG_MORE on non-TLS sockets
97 and we may use vectored I/O support via GnuTLS in the future
100 * per-FD PLCMap (post-loop callback) removed, we got ->requeue
101 support where no extra hash lookups or assignments are necessary.
103 * read push backs removed. Some subclasses use a read buffer ({rbuf})
104 but they control it, not this event loop.
106 * Profiling and debug logging removed. Perl and OS-specific tracers
107 and profilers are sufficient.
109 * ->AddOtherFds support removed, everything watched is a subclass of
110 PublicInbox::DS, but we've slimmed down the fields to eliminate
111 the memory penalty for objects.