1 |
=head1 NAME |
2 |
|
3 |
Async::Interrupt - allow C/XS libraries to interrupt perl asynchronously |
4 |
|
5 |
=head1 SYNOPSIS |
6 |
|
7 |
use Async::Interrupt; |
8 |
|
9 |
=head1 DESCRIPTION |
10 |
|
11 |
This module implements a single feature only of interest to advanced perl |
12 |
modules, namely asynchronous interruptions (think "UNIX signals", which |
13 |
are very similar). |
14 |
|
15 |
Sometimes, modules wish to run code asynchronously (in another thread, |
16 |
or from a signal handler), and then signal the perl interpreter on |
17 |
certain events. One common way is to write some data to a pipe and use an |
18 |
event handling toolkit to watch for I/O events. Another way is to send |
19 |
a signal. Those methods are slow, and in the case of a pipe, also not |
20 |
asynchronous - it won't interrupt a running perl interpreter. |
21 |
|
22 |
This module implements asynchronous notifications that enable you to |
23 |
signal running perl code from another thread, asynchronously, and |
24 |
sometimes even without using a single syscall. |
25 |
|
26 |
=head2 USAGE SCENARIOS |
27 |
|
28 |
=over 4 |
29 |
|
30 |
=item Race-free signal handling |
31 |
|
32 |
There seems to be no way to do race-free signal handling in perl: to |
33 |
catch a signal, you have to execute Perl code, and between entering the |
34 |
interpreter C<select> function (or other blocking functions) and executing |
35 |
the select syscall is a small but relevant timespan during which signals |
36 |
will be queued, but perl signal handlers will not be executed and the |
37 |
blocking syscall will not be interrupted. |
38 |
|
39 |
You can use this module to bind a signal to a callback while at the same |
40 |
time activating an event pipe that you can C<select> on, fixing the race |
41 |
completely. |
42 |
|
43 |
This can be used to implement the signal hadling in event loops, |
44 |
e.g. L<AnyEvent>, L<POE>, L<IO::Async::Loop> and so on. |
45 |
|
46 |
=item Background threads want speedy reporting |
47 |
|
48 |
Assume you want very exact timing, and you can spare an extra cpu core |
49 |
for that. Then you can run an extra thread that signals your perl |
50 |
interpreter. This means you can get a very exact timing source while your |
51 |
perl code is number crunching, without even using a syscall to communicate |
52 |
between your threads. |
53 |
|
54 |
For example the deliantra game server uses a variant of this technique |
55 |
to interrupt background processes regularly to send map updates to game |
56 |
clients. |
57 |
|
58 |
Or L<EV::Loop::Async> uses an interrupt object to wake up perl when new |
59 |
events have arrived. |
60 |
|
61 |
L<IO::AIO> and L<BDB> could also use this to speed up result reporting. |
62 |
|
63 |
=item Speedy event loop invocation |
64 |
|
65 |
One could use this module e.g. in L<Coro> to interrupt a running coro-thread |
66 |
and cause it to enter the event loop. |
67 |
|
68 |
Or one could bind to C<SIGIO> and tell some important sockets to send this |
69 |
signal, causing the event loop to be entered to reduce network latency. |
70 |
|
71 |
=back |
72 |
|
73 |
=head2 HOW TO USE |
74 |
|
75 |
You can use this module by creating an C<Async::Interrupt> object for each |
76 |
such event source. This object stores a perl and/or a C-level callback |
77 |
that is invoked when the C<Async::Interrupt> object gets signalled. It is |
78 |
executed at the next time the perl interpreter is running (i.e. it will |
79 |
interrupt a computation, but not an XS function or a syscall). |
80 |
|
81 |
You can signal the C<Async::Interrupt> object either by calling it's C<< |
82 |
->signal >> method, or, more commonly, by calling a C function. There is |
83 |
also the built-in (POSIX) signal source. |
84 |
|
85 |
The C<< ->signal_func >> returns the address of the C function that is to |
86 |
be called (plus an argument to be used during the call). The signalling |
87 |
function also takes an integer argument in the range SIG_ATOMIC_MIN to |
88 |
SIG_ATOMIC_MAX (guaranteed to allow at least 0..127). |
89 |
|
90 |
Since this kind of interruption is fast, but can only interrupt a |
91 |
I<running> interpreter, there is optional support for signalling a pipe |
92 |
- that means you can also wait for the pipe to become readable (e.g. via |
93 |
L<EV> or L<AnyEvent>). This, of course, incurs the overhead of a C<read> |
94 |
and C<write> syscall. |
95 |
|
96 |
=head1 USAGE EXAMPLES |
97 |
|
98 |
=head2 Async::Interrupt to implement race-free signal handling |
99 |
|
100 |
This example uses a single event pipe for all signals, and one |
101 |
Async::Interrupt per signal. |
102 |
|
103 |
First, create the event pipe and hook it into the event loop (this code is |
104 |
actually what L<AnyEvent> uses itself): |
105 |
|
106 |
$SIGPIPE = new Async::Interrupt::EventPipe; |
107 |
$SIGPIPE_W = AnyEvent->io ( |
108 |
fh => $SIGPIPE->fileno, |
109 |
poll => "r", |
110 |
cb => \&_signal_check, |
111 |
); |
112 |
|
113 |
Then, for each signal to hook, create an Async::Interrupt object. The |
114 |
callback just sets a global variable, as we are only interested in |
115 |
synchronous signals (i.e. when the event loop polls), which is why the |
116 |
pipe draining is not done automatically. |
117 |
|
118 |
my $interrupt = new Async::Interrupt |
119 |
cb => sub { undef $SIGNAL_RECEIVED{$signum} } |
120 |
signal => $signal, |
121 |
pipe => [$SIGPIPE_R->filenos], |
122 |
pipe_autodrain => 0, |
123 |
; |
124 |
|
125 |
Finally, the I/O callback for the event pipe handles the signals: |
126 |
|
127 |
sub _signal_check { |
128 |
# drain the pipe first |
129 |
$SIGPIPE->drain; |
130 |
|
131 |
# two loops, just to be sure |
132 |
while (%SIGNAL_RECEIVED) { |
133 |
for (keys %SIGNAL_RECEIVED) { |
134 |
delete $SIGNAL_RECEIVED{$_}; |
135 |
warn "signal $_ received\n"; |
136 |
} |
137 |
} |
138 |
} |
139 |
|
140 |
|
141 |
|
142 |
=head1 THE Async::Interrupt CLASS |
143 |
|
144 |
=over 4 |
145 |
|
146 |
=cut |
147 |
|
148 |
package Async::Interrupt; |
149 |
|
150 |
use common::sense; |
151 |
|
152 |
BEGIN { |
153 |
# the next line forces initialisation of internal |
154 |
# signal handling # variables |
155 |
$SIG{KILL} = sub { }; |
156 |
|
157 |
our $VERSION = '0.6'; |
158 |
|
159 |
require XSLoader; |
160 |
XSLoader::load ("Async::Interrupt", $VERSION); |
161 |
} |
162 |
|
163 |
our $DIED = sub { warn "$@" }; |
164 |
|
165 |
=item $async = new Async::Interrupt key => value... |
166 |
|
167 |
Creates a new Async::Interrupt object. You may only use async |
168 |
notifications on this object while it exists, so you need to keep a |
169 |
reference to it at all times while it is used. |
170 |
|
171 |
Optional constructor arguments include (normally you would specify at |
172 |
least one of C<cb> or C<c_cb>). |
173 |
|
174 |
=over 4 |
175 |
|
176 |
=item cb => $coderef->($value) |
177 |
|
178 |
Registers a perl callback to be invoked whenever the async interrupt is |
179 |
signalled. |
180 |
|
181 |
Note that, since this callback can be invoked at basically any time, it |
182 |
must not modify any well-known global variables such as C<$/> without |
183 |
restoring them again before returning. |
184 |
|
185 |
The exceptions are C<$!> and C<$@>, which are saved and restored by |
186 |
Async::Interrupt. |
187 |
|
188 |
If the callback should throw an exception, then it will be caught, |
189 |
and C<$Async::Interrupt::DIED> will be called with C<$@> containing |
190 |
the exception. The default will simply C<warn> about the message and |
191 |
continue. |
192 |
|
193 |
=item c_cb => [$c_func, $c_arg] |
194 |
|
195 |
Registers a C callback the be invoked whenever the async interrupt is |
196 |
signalled. |
197 |
|
198 |
The C callback must have the following prototype: |
199 |
|
200 |
void c_func (pTHX_ void *c_arg, int value); |
201 |
|
202 |
Both C<$c_func> and C<$c_arg> must be specified as integers/IVs, and |
203 |
C<$value> is the C<value> passed to some earlier call to either C<$signal> |
204 |
or the C<signal_func> function. |
205 |
|
206 |
Note that, because the callback can be invoked at almost any time, you |
207 |
have to be careful at saving and restoring global variables that Perl |
208 |
might use (the exception is C<errno>, which is saved and restored by |
209 |
Async::Interrupt). The callback itself runs as part of the perl context, |
210 |
so you can call any perl functions and modify any perl data structures (in |
211 |
which case the requirements set out for C<cb> apply as well). |
212 |
|
213 |
=item var => $scalar_ref |
214 |
|
215 |
When specified, then the given argument must be a reference to a |
216 |
scalar. The scalar will be set to C<0> initially. Signalling the interrupt |
217 |
object will set it to the passed value, handling the interrupt will reset |
218 |
it to C<0> again. |
219 |
|
220 |
Note that the only thing you are legally allowed to do is to is to check |
221 |
the variable in a boolean or integer context (e.g. comparing it with a |
222 |
string, or printing it, will I<destroy> it and might cause your program to |
223 |
crash or worse). |
224 |
|
225 |
=item signal => $signame_or_value |
226 |
|
227 |
When this parameter is specified, then the Async::Interrupt will hook the |
228 |
given signal, that is, it will effectively call C<< ->signal (0) >> each time |
229 |
the given signal is caught by the process. |
230 |
|
231 |
Only one async can hook a given signal, and the signal will be restored to |
232 |
defaults when the Async::Interrupt object gets destroyed. |
233 |
|
234 |
=item pipe => [$fileno_or_fh_for_reading, $fileno_or_fh_for_writing] |
235 |
|
236 |
Specifies two file descriptors (or file handles) that should be signalled |
237 |
whenever the async interrupt is signalled. This means a single octet will |
238 |
be written to it, and before the callback is being invoked, it will be |
239 |
read again. Due to races, it is unlikely but possible that multiple octets |
240 |
are written. It is required that the file handles are both in nonblocking |
241 |
mode. |
242 |
|
243 |
The object will keep a reference to the file handles. |
244 |
|
245 |
This can be used to ensure that async notifications will interrupt event |
246 |
frameworks as well. |
247 |
|
248 |
Note that C<Async::Interrupt> will create a suitable signal fd |
249 |
automatically when your program requests one, so you don't have to specify |
250 |
this argument when all you want is an extra file descriptor to watch. |
251 |
|
252 |
If you want to share a single event pipe between multiple Async::Interrupt |
253 |
objects, you can use the C<Async::Interrupt::EventPipe> class to manage |
254 |
those. |
255 |
|
256 |
=back |
257 |
|
258 |
=cut |
259 |
|
260 |
sub new { |
261 |
my ($class, %arg) = @_; |
262 |
|
263 |
bless \(_alloc $arg{cb}, @{$arg{c_cb}}[0,1], @{$arg{pipe}}[0,1], $arg{signal}, $arg{var}), $class |
264 |
} |
265 |
|
266 |
=item ($signal_func, $signal_arg) = $async->signal_func |
267 |
|
268 |
Returns the address of a function to call asynchronously. The function |
269 |
has the following prototype and needs to be passed the specified |
270 |
C<$signal_arg>, which is a C<void *> cast to C<IV>: |
271 |
|
272 |
void (*signal_func) (void *signal_arg, int value) |
273 |
|
274 |
An example call would look like: |
275 |
|
276 |
signal_func (signal_arg, 0); |
277 |
|
278 |
The function is safe to call from within signal and thread contexts, at |
279 |
any time. The specified C<value> is passed to both C and Perl callback. |
280 |
|
281 |
C<$value> must be in the valid range for a C<sig_atomic_t>, except C<0> |
282 |
(1..127 is portable). |
283 |
|
284 |
If the function is called while the Async::Interrupt object is already |
285 |
signaled but before the callbacks are being executed, then the stored |
286 |
C<value> is either the old or the new one. Due to the asynchronous |
287 |
nature of the code, the C<value> can even be passed to two consecutive |
288 |
invocations of the callback. |
289 |
|
290 |
=item $address = $async->c_var |
291 |
|
292 |
Returns the address (cast to IV) of an C<IV> variable. The variable is set |
293 |
to C<0> initially and gets set to the passed value whenever the object |
294 |
gets signalled, and reset to C<0> once the interrupt has been handled. |
295 |
|
296 |
Note that it is often beneficial to just call C<PERL_ASYNC_CHECK ()> to |
297 |
handle any interrupts. |
298 |
|
299 |
Example: call some XS function to store the address, then show C code |
300 |
waiting for it. |
301 |
|
302 |
my_xs_func $async->c_var; |
303 |
|
304 |
static IV *valuep; |
305 |
|
306 |
void |
307 |
my_xs_func (void *addr) |
308 |
CODE: |
309 |
valuep = (IV *)addr; |
310 |
|
311 |
// code in a loop, waiting |
312 |
while (!*valuep) |
313 |
; // do something |
314 |
|
315 |
=item $async->signal ($value=1) |
316 |
|
317 |
This signals the given async object from Perl code. Semi-obviously, this |
318 |
will instantly trigger the callback invocation (it does not, as the name |
319 |
might imply, do anything with POSIX signals). |
320 |
|
321 |
C<$value> must be in the valid range for a C<sig_atomic_t>, except C<0> |
322 |
(1..127 is portable). |
323 |
|
324 |
=item $async->signal_hysteresis ($enable) |
325 |
|
326 |
Enables or disables signal hysteresis (default: disabled). If a POSIX |
327 |
signal is used as a signal source for the interrupt object, then enabling |
328 |
signal hysteresis causes Async::Interrupt to reset the signal action to |
329 |
C<SIG_IGN> in the signal handler and restore it just before handling the |
330 |
interruption. |
331 |
|
332 |
When you expect a lot of signals (e.g. when using SIGIO), then enabling |
333 |
signal hysteresis can reduce the number of handler invocations |
334 |
considerably, at the cost of two extra syscalls. |
335 |
|
336 |
Note that setting the signal to C<SIG_IGN> can have unintended side |
337 |
effects when you fork and exec other programs, as often they do nto expect |
338 |
signals to be ignored by default. |
339 |
|
340 |
=item $async->block |
341 |
|
342 |
=item $async->unblock |
343 |
|
344 |
Sometimes you need a "critical section" of code that will not be |
345 |
interrupted by an Async::Interrupt. This can be implemented by calling C<< |
346 |
$async->block >> before the critical section, and C<< $async->unblock >> |
347 |
afterwards. |
348 |
|
349 |
Note that there must be exactly one call of C<unblock> for every previous |
350 |
call to C<block> (i.e. calls can nest). |
351 |
|
352 |
Since ensuring this in the presence of exceptions and threads is |
353 |
usually more difficult than you imagine, I recommend using C<< |
354 |
$async->scoped_block >> instead. |
355 |
|
356 |
=item $async->scope_block |
357 |
|
358 |
This call C<< $async->block >> and installs a handler that is called when |
359 |
the current scope is exited (via an exception, by canceling the Coro |
360 |
thread, by calling last/goto etc.). |
361 |
|
362 |
This is the recommended (and fastest) way to implement critical sections. |
363 |
|
364 |
=item ($block_func, $block_arg) = $async->scope_block_func |
365 |
|
366 |
Returns the address of a function that implements the C<scope_block> |
367 |
functionality. |
368 |
|
369 |
It has the following prototype and needs to be passed the specified |
370 |
C<$block_arg>, which is a C<void *> cast to C<IV>: |
371 |
|
372 |
void (*block_func) (void *block_arg) |
373 |
|
374 |
An example call would look like: |
375 |
|
376 |
block_func (block_arg); |
377 |
|
378 |
The function is safe to call only from within the toplevel of a perl XS |
379 |
function and will call C<LEAVE> and C<ENTER> (in this order!). |
380 |
|
381 |
=item $async->pipe_enable |
382 |
|
383 |
=item $async->pipe_disable |
384 |
|
385 |
Enable/disable signalling the pipe when the interrupt occurs (default is |
386 |
enabled). Writing to a pipe is relatively expensive, so it can be disabled |
387 |
when you know you are not waiting for it (for example, with L<EV> you |
388 |
could disable the pipe in a check watcher, and enable it in a prepare |
389 |
watcher). |
390 |
|
391 |
Note that currently, while C<pipe_disable> is in effect, no attempt to |
392 |
read from the pipe will be done when handling events. This might change as |
393 |
soon as I realize why this is a mistake. |
394 |
|
395 |
=item $fileno = $async->pipe_fileno |
396 |
|
397 |
Returns the reading side of the signalling pipe. If no signalling pipe is |
398 |
currently attached to the object, it will dynamically create one. |
399 |
|
400 |
Note that the only valid oepration on this file descriptor is to wait |
401 |
until it is readable. The fd might belong currently to a pipe, a tcp |
402 |
socket, or an eventfd, depending on the platform, and is guaranteed to be |
403 |
C<select>able. |
404 |
|
405 |
=item $async->pipe_autodrain ($enable) |
406 |
|
407 |
Enables (C<1>) or disables (C<0>) automatic draining of the pipe (default: |
408 |
enabled). When automatic draining is enabled, then Async::Interrupt will |
409 |
automatically clear the pipe. Otherwise the user is responsible for this |
410 |
draining. |
411 |
|
412 |
This is useful when you want to share one pipe among many Async::Interrupt |
413 |
objects. |
414 |
|
415 |
=item $async->post_fork |
416 |
|
417 |
The object will not normally be usable after a fork (as the pipe fd is |
418 |
shared between processes). Calling this method after a fork in the child |
419 |
ensures that the object will work as expected again. It only needs to be |
420 |
called when the async object is used in the child. |
421 |
|
422 |
This only works when the pipe was created by Async::Interrupt. |
423 |
|
424 |
Async::Interrupt ensures that the reading file descriptor does not change |
425 |
it's value. |
426 |
|
427 |
=item $signum = Async::Interrupt::sig2num $signame_or_number |
428 |
|
429 |
=item $signame = Async::Interrupt::sig2name $signame_or_number |
430 |
|
431 |
These two convenience functions simply convert a signal name or number to |
432 |
the corresponding name or number. They are not used by this module and |
433 |
exist just because perl doesn't have a nice way to do this on its own. |
434 |
|
435 |
They will return C<undef> on illegal names or numbers. |
436 |
|
437 |
=back |
438 |
|
439 |
=head1 THE Async::Interrupt::EventPipe CLASS |
440 |
|
441 |
Pipes are the predominent utility to make asynchronous signals |
442 |
synchronous. However, pipes are hard to come by: they don't exist on the |
443 |
broken windows platform, and on GNU/Linux systems, you might want to use |
444 |
an C<eventfd> instead. |
445 |
|
446 |
This class creates selectable event pipes in a portable fashion: on |
447 |
windows, it will try to create a tcp socket pair, on GNU/Linux, it will |
448 |
try to create an eventfd and everywhere else it will try to use a normal |
449 |
pipe. |
450 |
|
451 |
=over 4 |
452 |
|
453 |
=item $epipe = new Async::Interrupt::EventPipe |
454 |
|
455 |
This creates and returns an eventpipe object. This object is simply a |
456 |
blessed array reference: |
457 |
|
458 |
=item ($r_fd, $w_fd) = $epipe->filenos |
459 |
|
460 |
Returns the read-side file descriptor and the write-side file descriptor. |
461 |
|
462 |
Example: pass an eventpipe object as pipe to the Async::Interrupt |
463 |
constructor, and create an AnyEvent watcher for the read side. |
464 |
|
465 |
my $epipe = new Async::Interrupt::EventPipe; |
466 |
my $asy = new Async::Interrupt pipe => [$epipe->filenos]; |
467 |
my $iow = AnyEvent->io (fh => $epipe->fileno, poll => 'r', cb => sub { }); |
468 |
|
469 |
=item $r_fd = $epipe->fileno |
470 |
|
471 |
Return only the reading/listening side. |
472 |
|
473 |
=item $epipe->signal |
474 |
|
475 |
Write something to the pipe, in a portable fashion. |
476 |
|
477 |
=item $epipe->drain |
478 |
|
479 |
Drain (empty) the pipe. |
480 |
|
481 |
=item $epipe->renew |
482 |
|
483 |
Recreates the pipe (useful after a fork). The reading side will not change |
484 |
it's file descriptor number, but the writing side might. |
485 |
|
486 |
=back |
487 |
|
488 |
=cut |
489 |
|
490 |
1; |
491 |
|
492 |
=head1 EXAMPLE |
493 |
|
494 |
There really should be a complete C/XS example. Bug me about it. Better |
495 |
yet, create one. |
496 |
|
497 |
=head1 IMPLEMENTATION DETAILS AND LIMITATIONS |
498 |
|
499 |
This module works by "hijacking" SIGKILL, which is guaranteed to always |
500 |
exist, but also cannot be caught, so is always available. |
501 |
|
502 |
Basically, this module fakes the occurance of a SIGKILL signal and |
503 |
then intercepts the interpreter handling it. This makes normal signal |
504 |
handling slower (probably unmeasurably, though), but has the advantage |
505 |
of not requiring a special runops function, nor slowing down normal perl |
506 |
execution a bit. |
507 |
|
508 |
It assumes that C<sig_atomic_t>, C<int> and C<IV> are all async-safe to |
509 |
modify. |
510 |
|
511 |
=head1 AUTHOR |
512 |
|
513 |
Marc Lehmann <schmorp@schmorp.de> |
514 |
http://home.schmorp.de/ |
515 |
|
516 |
=cut |
517 |
|