1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
| | PublicInbox::DS - event loop and async I/O base class
Our PublicInbox::DS event loop which powers most of our long-lived
processes(*) diverges significantly from the unmaintained Danga::Socket
package we forked from. In fact, it's probably different from most
other event loops out there.
Most notably, it uses one-shot, level-trigger, and edge-trigger mode
modes of kqueue|epoll depending on the situation.
(*) public-inbox-netd,(-httpd,-imapd,-nntpd,-pop3d,-watch) + lei-daemon
Most notably:
* There is one and only one callback: ->event_step. Unlike other
event loops, there are no separate callbacks for read, write,
error or hangup events. In fact, we never care which kevent
filter or poll/epoll event flag (e.g. POLLIN/POLLOUT/POLLHUP)
triggers a call.
The lack of read/write callback distinction is driven by the
fact that TLS libraries (e.g. OpenSSL via IO::Socket::SSL) may
declare SSL_WANT_READ on SSL_write(), and SSL_WANT_READ on
SSL_read(). So we end up having to let each user object decide
whether it wants to make read or write calls depending on its
internal state, completely independent of the event loop.
Error and hangup (POLLERR and POLLHUP) callbacks are redundant and
only triggered in rare cases. They're redundant because the
result of every read and write call in ->event_step must be
checked, anyways. At best, callbacks for POLLHUP and POLLERR can
save one syscall per socket lifetime and not worth the extra code
it imposes.
Reducing the user-supplied code down to a single callback allows
subclasses to keep their logic self-contained. The combination
of this change and one-shot wakeups (see below) for bidirectional
data flows makes asynchronous code easier to reason about.
Other divergences:
* ->write buffering uses temporary files whereas Danga::Socket used
the heap. The rationale for this is the kernel already provides
ample (and configurable) space for socket buffers. Modern kernels
also cache FS operations aggressively, so systems with ample RAM
are unlikely to notice degradation, while small systems are less
likely to suffer unpredictable heap fragmentation, swap and OOM
penalties.
In the future, we may introduce sendfile and mmap+SSL_write to
reduce data copies, and use FALLOC_FL_PUNCH_HOLE on Linux to
release space after the buffer is partially cleared.
Augmented features:
* obj->write(CODEREF) passes the object itself to the CODEREF.
Being able to enqueue subroutine calls is a powerful feature in
Danga::Socket for keeping linear logic in an asynchronous environment.
Unfortunately, each subroutine takes several kilobytes of memory.
One small change to Danga::Socket is to pass the receiver object
(aka "$self") to the CODEREF. $self can store any necessary
state it needs for a normal (named) subroutine. This allows us to
put the same sub into multiple queues without paying a large
memory penalty for each one.
This idea is also more easily ported to C or other languages which
lack anonymous subroutines (aka "closures").
* ->requeue support. An optimization of the AddTimer(0, ...) idiom
for immediately dispatching code at the next event loop iteration.
public-inbox uses this for fairly generating large responses
iteratively (see PublicInbox::NNTP::long_response or ibx_async_cat
for blob retrievals).
New features
* One-shot wakeups allowed via EPOLLONESHOT or EV_DISPATCH. These
flags allow us to simplify code in ->event_step callbacks for
bidirectional sockets (NNTP and HTTP). Instead of merely reacting
to events, control is handed over at ->event_step in one-shot scenarios.
The event_step caller (NNTP || HTTP) then becomes proactive in declaring
which (if any) events it's interested in for the next loop iteration.
* Edge-triggering available via EPOLLET or EV_CLEAR. These reduce wakeups
for unidirectional classes when throughput is more important than fairness.
* IO::Socket::SSL support (for NNTPS, STARTTLS+NNTP, HTTPS)
* awaitpid (waitpid wrapper) support for reaping dead children
* reliable signal wakeups are supported via signalfd on Linux,
EVFILT_SIGNAL on *BSDs via IO::KQueue.
Removed features
* Many fields removed or moved to subclasses, so the underlying
hash is smaller and suitable for FDs other than stream sockets.
Some fields we enforce (e.g. wbuf, wbuf_off) are autovivified
on an as-needed basis to save memory when they're not needed.
* TCP_CORK support removed, instead we use MSG_MORE on non-TLS sockets
and we may use vectored I/O support via GnuTLS in the future
for TLS sockets.
* per-FD PLCMap (post-loop callback) removed, we got ->requeue
support where no extra hash lookups or assignments are necessary.
* read push backs removed. Some subclasses use a read buffer ({rbuf})
but they control it, not this event loop.
* Profiling and debug logging removed. Perl and OS-specific tracers
and profilers are sufficient.
* ->AddOtherFds support removed, everything watched is a subclass of
PublicInbox::DS, but we've slimmed down the fields to eliminate
the memory penalty for objects.
|