1 |
NAME |
2 |
AnyEvent::Fork - everything you wanted to use fork() for, but couldn't |
3 |
|
4 |
SYNOPSIS |
5 |
use AnyEvent::Fork; |
6 |
|
7 |
################################################################## |
8 |
# create a single new process, tell it to run your worker function |
9 |
|
10 |
AnyEvent::Fork |
11 |
->new |
12 |
->require ("MyModule") |
13 |
->run ("MyModule::worker, sub { |
14 |
my ($master_filehandle) = @_; |
15 |
|
16 |
# now $master_filehandle is connected to the |
17 |
# $slave_filehandle in the new process. |
18 |
}); |
19 |
|
20 |
# MyModule::worker might look like this |
21 |
sub MyModule::worker { |
22 |
my ($slave_filehandle) = @_; |
23 |
|
24 |
# now $slave_filehandle is connected to the $master_filehandle |
25 |
# in the original prorcess. have fun! |
26 |
} |
27 |
|
28 |
################################################################## |
29 |
# create a pool of server processes all accepting on the same socket |
30 |
|
31 |
# create listener socket |
32 |
my $listener = ...; |
33 |
|
34 |
# create a pool template, initialise it and give it the socket |
35 |
my $pool = AnyEvent::Fork |
36 |
->new |
37 |
->require ("Some::Stuff", "My::Server") |
38 |
->send_fh ($listener); |
39 |
|
40 |
# now create 10 identical workers |
41 |
for my $id (1..10) { |
42 |
$pool |
43 |
->fork |
44 |
->send_arg ($id) |
45 |
->run ("My::Server::run"); |
46 |
} |
47 |
|
48 |
# now do other things - maybe use the filehandle provided by run |
49 |
# to wait for the processes to die. or whatever. |
50 |
|
51 |
# My::Server::run might look like this |
52 |
sub My::Server::run { |
53 |
my ($slave, $listener, $id) = @_; |
54 |
|
55 |
close $slave; # we do not use the socket, so close it to save resources |
56 |
|
57 |
# we could go ballistic and use e.g. AnyEvent here, or IO::AIO, |
58 |
# or anything we usually couldn't do in a process forked normally. |
59 |
while (my $socket = $listener->accept) { |
60 |
# do sth. with new socket |
61 |
} |
62 |
} |
63 |
|
64 |
DESCRIPTION |
65 |
This module allows you to create new processes, without actually forking |
66 |
them from your current process (avoiding the problems of forking), but |
67 |
preserving most of the advantages of fork. |
68 |
|
69 |
It can be used to create new worker processes or new independent |
70 |
subprocesses for short- and long-running jobs, process pools (e.g. for |
71 |
use in pre-forked servers) but also to spawn new external processes |
72 |
(such as CGI scripts from a web server), which can be faster (and more |
73 |
well behaved) than using fork+exec in big processes. |
74 |
|
75 |
Special care has been taken to make this module useful from other |
76 |
modules, while still supporting specialised environments such as |
77 |
App::Staticperl or PAR::Packer. |
78 |
|
79 |
WHAT THIS MODULE IS NOT |
80 |
This module only creates processes and lets you pass file handles and |
81 |
strings to it, and run perl code. It does not implement any kind of RPC |
82 |
- there is no back channel from the process back to you, and there is no |
83 |
RPC or message passing going on. |
84 |
|
85 |
If you need some form of RPC, you can either implement it yourself in |
86 |
whatever way you like, use some message-passing module such as |
87 |
AnyEvent::MP, some pipe such as AnyEvent::ZeroMQ, use AnyEvent::Handle |
88 |
on both sides to send e.g. JSON or Storable messages, and so on. |
89 |
|
90 |
PROBLEM STATEMENT |
91 |
There are two ways to implement parallel processing on UNIX like |
92 |
operating systems - fork and process, and fork+exec and process. They |
93 |
have different advantages and disadvantages that I describe below, |
94 |
together with how this module tries to mitigate the disadvantages. |
95 |
|
96 |
Forking from a big process can be very slow (a 5GB process needs 0.05s |
97 |
to fork on my 3.6GHz amd64 GNU/Linux box for example). This overhead is |
98 |
often shared with exec (because you have to fork first), but in some |
99 |
circumstances (e.g. when vfork is used), fork+exec can be much faster. |
100 |
This module can help here by telling a small(er) helper process to |
101 |
fork, or fork+exec instead. |
102 |
|
103 |
Forking usually creates a copy-on-write copy of the parent process. |
104 |
Memory (for example, modules or data files that have been will not take |
105 |
additional memory). When exec'ing a new process, modules and data files |
106 |
might need to be loaded again, at extra CPU and memory cost. Likewise |
107 |
when forking, all data structures are copied as well - if the program |
108 |
frees them and replaces them by new data, the child processes will |
109 |
retain the memory even if it isn't used. |
110 |
This module allows the main program to do a controlled fork, and |
111 |
allows modules to exec processes safely at any time. When creating a |
112 |
custom process pool you can take advantage of data sharing via fork |
113 |
without risking to share large dynamic data structures that will |
114 |
blow up child memory usage. |
115 |
|
116 |
Exec'ing a new perl process might be difficult and slow. For example, it |
117 |
is not easy to find the correct path to the perl interpreter, and all |
118 |
modules have to be loaded from disk again. Long running processes might |
119 |
run into problems when perl is upgraded for example. |
120 |
This module supports creating pre-initialised perl processes to be |
121 |
used as template, and also tries hard to identify the correct path |
122 |
to the perl interpreter. With a cooperative main program, exec'ing |
123 |
the interpreter might not even be necessary. |
124 |
|
125 |
Forking might be impossible when a program is running. For example, |
126 |
POSIX makes it almost impossible to fork from a multi-threaded program |
127 |
and do anything useful in the child - strictly speaking, if your perl |
128 |
program uses posix threads (even indirectly via e.g. IO::AIO or |
129 |
threads), you cannot call fork on the perl level anymore, at all. |
130 |
This module can safely fork helper processes at any time, by calling |
131 |
fork+exec in C, in a POSIX-compatible way. |
132 |
|
133 |
Parallel processing with fork might be inconvenient or difficult to |
134 |
implement. For example, when a program uses an event loop and creates |
135 |
watchers it becomes very hard to use the event loop from a child |
136 |
program, as the watchers already exist but are only meaningful in the |
137 |
parent. Worse, a module might want to use such a system, not knowing |
138 |
whether another module or the main program also does, leading to |
139 |
problems. |
140 |
This module only lets the main program create pools by forking |
141 |
(because only the main program can know when it is still safe to do |
142 |
so) - all other pools are created by fork+exec, after which such |
143 |
modules can again be loaded. |
144 |
|
145 |
CONCEPTS |
146 |
This module can create new processes either by executing a new perl |
147 |
process, or by forking from an existing "template" process. |
148 |
|
149 |
Each such process comes with its own file handle that can be used to |
150 |
communicate with it (it's actually a socket - one end in the new |
151 |
process, one end in the main process), and among the things you can do |
152 |
in it are load modules, fork new processes, send file handles to it, and |
153 |
execute functions. |
154 |
|
155 |
There are multiple ways to create additional processes to execute some |
156 |
jobs: |
157 |
|
158 |
fork a new process from the "default" template process, load code, run |
159 |
it |
160 |
This module has a "default" template process which it executes when |
161 |
it is needed the first time. Forking from this process shares the |
162 |
memory used for the perl interpreter with the new process, but |
163 |
loading modules takes time, and the memory is not shared with |
164 |
anything else. |
165 |
|
166 |
This is ideal for when you only need one extra process of a kind, |
167 |
with the option of starting and stopping it on demand. |
168 |
|
169 |
Example: |
170 |
|
171 |
AnyEvent::Fork |
172 |
->new |
173 |
->require ("Some::Module") |
174 |
->run ("Some::Module::run", sub { |
175 |
my ($fork_fh) = @_; |
176 |
}); |
177 |
|
178 |
fork a new template process, load code, then fork processes off of it |
179 |
and run the code |
180 |
When you need to have a bunch of processes that all execute the same |
181 |
(or very similar) tasks, then a good way is to create a new template |
182 |
process for them, loading all the modules you need, and then create |
183 |
your worker processes from this new template process. |
184 |
|
185 |
This way, all code (and data structures) that can be shared (e.g. |
186 |
the modules you loaded) is shared between the processes, and each |
187 |
new process consumes relatively little memory of its own. |
188 |
|
189 |
The disadvantage of this approach is that you need to create a |
190 |
template process for the sole purpose of forking new processes from |
191 |
it, but if you only need a fixed number of processes you can create |
192 |
them, and then destroy the template process. |
193 |
|
194 |
Example: |
195 |
|
196 |
my $template = AnyEvent::Fork->new->require ("Some::Module"); |
197 |
|
198 |
for (1..10) { |
199 |
$template->fork->run ("Some::Module::run", sub { |
200 |
my ($fork_fh) = @_; |
201 |
}); |
202 |
} |
203 |
|
204 |
# at this point, you can keep $template around to fork new processes |
205 |
# later, or you can destroy it, which causes it to vanish. |
206 |
|
207 |
execute a new perl interpreter, load some code, run it |
208 |
This is relatively slow, and doesn't allow you to share memory |
209 |
between multiple processes. |
210 |
|
211 |
The only advantage is that you don't have to have a template process |
212 |
hanging around all the time to fork off some new processes, which |
213 |
might be an advantage when there are long time spans where no extra |
214 |
processes are needed. |
215 |
|
216 |
Example: |
217 |
|
218 |
AnyEvent::Fork |
219 |
->new_exec |
220 |
->require ("Some::Module") |
221 |
->run ("Some::Module::run", sub { |
222 |
my ($fork_fh) = @_; |
223 |
}); |
224 |
|
225 |
FUNCTIONS |
226 |
my $pool = new AnyEvent::Fork key => value... |
227 |
Create a new process pool. The following named parameters are |
228 |
supported: |
229 |
|
230 |
my $proc = new AnyEvent::Fork |
231 |
Create a new "empty" perl interpreter process and returns its |
232 |
process object for further manipulation. |
233 |
|
234 |
The new process is forked from a template process that is kept |
235 |
around for this purpose. When it doesn't exist yet, it is created by |
236 |
a call to "new_exec" and kept around for future calls. |
237 |
|
238 |
When the process object is destroyed, it will release the file |
239 |
handle that connects it with the new process. When the new process |
240 |
has not yet called "run", then the process will exit. Otherwise, |
241 |
what happens depends entirely on the code that is executed. |
242 |
|
243 |
$new_proc = $proc->fork |
244 |
Forks $proc, creating a new process, and returns the process object |
245 |
of the new process. |
246 |
|
247 |
If any of the "send_" functions have been called before fork, then |
248 |
they will be cloned in the child. For example, in a pre-forked |
249 |
server, you might "send_fh" the listening socket into the template |
250 |
process, and then keep calling "fork" and "run". |
251 |
|
252 |
my $proc = new_exec AnyEvent::Fork |
253 |
Create a new "empty" perl interpreter process and returns its |
254 |
process object for further manipulation. |
255 |
|
256 |
Unlike the "new" method, this method *always* spawns a new perl |
257 |
process (except in some cases, see AnyEvent::Fork::Early for |
258 |
details). This reduces the amount of memory sharing that is |
259 |
possible, and is also slower. |
260 |
|
261 |
You should use "new" whenever possible, except when having a |
262 |
template process around is unacceptable. |
263 |
|
264 |
The path to the perl interpreter is divined using various methods - |
265 |
first $^X is investigated to see if the path ends with something |
266 |
that sounds as if it were the perl interpreter. Failing this, the |
267 |
module falls back to using $Config::Config{perlpath}. |
268 |
|
269 |
$pid = $proc->pid |
270 |
Returns the process id of the process *iff it is a direct child of |
271 |
the process* running AnyEvent::Fork, and "undef" otherwise. |
272 |
|
273 |
Normally, only processes created via "AnyEvent::Fork->new_exec" and |
274 |
AnyEvent::Fork::Template are direct children, and you are |
275 |
responsible to clean up their zombies when they die. |
276 |
|
277 |
All other processes are not direct children, and will be cleaned up |
278 |
by AnyEvent::Fork. |
279 |
|
280 |
$proc = $proc->eval ($perlcode, @args) |
281 |
Evaluates the given $perlcode as ... perl code, while setting @_ to |
282 |
the strings specified by @args. |
283 |
|
284 |
This call is meant to do any custom initialisation that might be |
285 |
required (for example, the "require" method uses it). It's not |
286 |
supposed to be used to completely take over the process, use "run" |
287 |
for that. |
288 |
|
289 |
The code will usually be executed after this call returns, and there |
290 |
is no way to pass anything back to the calling process. Any |
291 |
evaluation errors will be reported to stderr and cause the process |
292 |
to exit. |
293 |
|
294 |
Returns the process object for easy chaining of method calls. |
295 |
|
296 |
$proc = $proc->require ($module, ...) |
297 |
Tries to load the given module(s) into the process |
298 |
|
299 |
Returns the process object for easy chaining of method calls. |
300 |
|
301 |
$proc = $proc->send_fh ($handle, ...) |
302 |
Send one or more file handles (*not* file descriptors) to the |
303 |
process, to prepare a call to "run". |
304 |
|
305 |
The process object keeps a reference to the handles until this is |
306 |
done, so you must not explicitly close the handles. This is most |
307 |
easily accomplished by simply not storing the file handles anywhere |
308 |
after passing them to this method. |
309 |
|
310 |
Returns the process object for easy chaining of method calls. |
311 |
|
312 |
Example: pass a file handle to a process, and release it without |
313 |
closing. It will be closed automatically when it is no longer used. |
314 |
|
315 |
$proc->send_fh ($my_fh); |
316 |
undef $my_fh; # free the reference if you want, but DO NOT CLOSE IT |
317 |
|
318 |
$proc = $proc->send_arg ($string, ...) |
319 |
Send one or more argument strings to the process, to prepare a call |
320 |
to "run". The strings can be any octet string. |
321 |
|
322 |
The protocol is optimised to pass a moderate number of relatively |
323 |
short strings - while you can pass up to 4GB of data in one go, this |
324 |
is more meant to pass some ID information or other startup info, not |
325 |
big chunks of data. |
326 |
|
327 |
Returns the process object for easy chaining of method calls. |
328 |
|
329 |
$proc->run ($func, $cb->($fh)) |
330 |
Enter the function specified by the fully qualified name in $func in |
331 |
the process. The function is called with the communication socket as |
332 |
first argument, followed by all file handles and string arguments |
333 |
sent earlier via "send_fh" and "send_arg" methods, in the order they |
334 |
were called. |
335 |
|
336 |
If the called function returns, the process exits. |
337 |
|
338 |
Preparing the process can take time - when the process is ready, the |
339 |
callback is invoked with the local communications socket as |
340 |
argument. |
341 |
|
342 |
The process object becomes unusable on return from this function. |
343 |
|
344 |
If the communication socket isn't used, it should be closed on both |
345 |
sides, to save on kernel memory. |
346 |
|
347 |
The socket is non-blocking in the parent, and blocking in the newly |
348 |
created process. The close-on-exec flag is set on both. Even if not |
349 |
used otherwise, the socket can be a good indicator for the existence |
350 |
of the process - if the other process exits, you get a readable |
351 |
event on it, because exiting the process closes the socket (if it |
352 |
didn't create any children using fork). |
353 |
|
354 |
Example: create a template for a process pool, pass a few strings, |
355 |
some file handles, then fork, pass one more string, and run some |
356 |
code. |
357 |
|
358 |
my $pool = AnyEvent::Fork |
359 |
->new |
360 |
->send_arg ("str1", "str2") |
361 |
->send_fh ($fh1, $fh2); |
362 |
|
363 |
for (1..2) { |
364 |
$pool |
365 |
->fork |
366 |
->send_arg ("str3") |
367 |
->run ("Some::function", sub { |
368 |
my ($fh) = @_; |
369 |
|
370 |
# fh is nonblocking, but we trust that the OS can accept these |
371 |
# extra 3 octets anyway. |
372 |
syswrite $fh, "hi #$_\n"; |
373 |
|
374 |
# $fh is being closed here, as we don't store it anywhere |
375 |
}); |
376 |
} |
377 |
|
378 |
# Some::function might look like this - all parameters passed before fork |
379 |
# and after will be passed, in order, after the communications socket. |
380 |
sub Some::function { |
381 |
my ($fh, $str1, $str2, $fh1, $fh2, $str3) = @_; |
382 |
|
383 |
print scalar <$fh>; # prints "hi 1\n" and "hi 2\n" |
384 |
} |
385 |
|
386 |
PERFORMANCE |
387 |
Now for some unscientific benchmark numbers (all done on an amd64 |
388 |
GNU/Linux box). These are intended to give you an idea of the relative |
389 |
performance you can expect, they are not meant to be absolute |
390 |
performance numbers. |
391 |
|
392 |
OK, so, I ran a simple benchmark that creates a socket pair, forks, |
393 |
calls exit in the child and waits for the socket to close in the parent. |
394 |
I did load AnyEvent, EV and AnyEvent::Fork, for a total process size of |
395 |
5100kB. |
396 |
|
397 |
2079 new processes per second, using manual socketpair + fork |
398 |
|
399 |
Then I did the same thing, but instead of calling fork, I called |
400 |
AnyEvent::Fork->new->run ("CORE::exit") and then again waited for the |
401 |
socket form the child to close on exit. This does the same thing as |
402 |
manual socket pair + fork, except that what is forked is the template |
403 |
process (2440kB), and the socket needs to be passed to the server at the |
404 |
other end of the socket first. |
405 |
|
406 |
2307 new processes per second, using AnyEvent::Fork->new |
407 |
|
408 |
And finally, using "new_exec" instead "new", using vforks+execs to exec |
409 |
a new perl interpreter and compile the small server each time, I get: |
410 |
|
411 |
479 vfork+execs per second, using AnyEvent::Fork->new_exec |
412 |
|
413 |
So how can "AnyEvent->new" be faster than a standard fork, even though |
414 |
it uses the same operations, but adds a lot of overhead? |
415 |
|
416 |
The difference is simply the process size: forking the 6MB process takes |
417 |
so much longer than forking the 2.5MB template process that the overhead |
418 |
introduced is canceled out. |
419 |
|
420 |
If the benchmark process grows, the normal fork becomes even slower: |
421 |
|
422 |
1340 new processes, manual fork in a 20MB process |
423 |
731 new processes, manual fork in a 200MB process |
424 |
235 new processes, manual fork in a 2000MB process |
425 |
|
426 |
What that means (to me) is that I can use this module without having a |
427 |
very bad conscience because of the extra overhead required to start new |
428 |
processes. |
429 |
|
430 |
TYPICAL PROBLEMS |
431 |
This section lists typical problems that remain. I hope by recognising |
432 |
them, most can be avoided. |
433 |
|
434 |
"leaked" file descriptors for exec'ed processes |
435 |
POSIX systems inherit file descriptors by default when exec'ing a |
436 |
new process. While perl itself laudably sets the close-on-exec flags |
437 |
on new file handles, most C libraries don't care, and even if all |
438 |
cared, it's often not possible to set the flag in a race-free |
439 |
manner. |
440 |
|
441 |
That means some file descriptors can leak through. And since it |
442 |
isn't possible to know which file descriptors are "good" and |
443 |
"necessary" (or even to know which file descriptors are open), there |
444 |
is no good way to close the ones that might harm. |
445 |
|
446 |
As an example of what "harm" can be done consider a web server that |
447 |
accepts connections and afterwards some module uses AnyEvent::Fork |
448 |
for the first time, causing it to fork and exec a new process, which |
449 |
might inherit the network socket. When the server closes the socket, |
450 |
it is still open in the child (which doesn't even know that) and the |
451 |
client might conclude that the connection is still fine. |
452 |
|
453 |
For the main program, there are multiple remedies available - |
454 |
AnyEvent::Fork::Early is one, creating a process early and not using |
455 |
"new_exec" is another, as in both cases, the first process can be |
456 |
exec'ed well before many random file descriptors are open. |
457 |
|
458 |
In general, the solution for these kind of problems is to fix the |
459 |
libraries or the code that leaks those file descriptors. |
460 |
|
461 |
Fortunately, most of these leaked descriptors do no harm, other than |
462 |
sitting on some resources. |
463 |
|
464 |
"leaked" file descriptors for fork'ed processes |
465 |
Normally, AnyEvent::Fork does start new processes by exec'ing them, |
466 |
which closes file descriptors not marked for being inherited. |
467 |
|
468 |
However, AnyEvent::Fork::Early and AnyEvent::Fork::Template offer a |
469 |
way to create these processes by forking, and this leaks more file |
470 |
descriptors than exec'ing them, as there is no way to mark |
471 |
descriptors as "close on fork". |
472 |
|
473 |
An example would be modules like EV, IO::AIO or Gtk2. Both create |
474 |
pipes for internal uses, and Gtk2 might open a connection to the X |
475 |
server. EV and IO::AIO can deal with fork, but Gtk2 might have |
476 |
trouble with a fork. |
477 |
|
478 |
The solution is to either not load these modules before use'ing |
479 |
AnyEvent::Fork::Early or AnyEvent::Fork::Template, or to delay |
480 |
initialising them, for example, by calling "init Gtk2" manually. |
481 |
|
482 |
exit runs destructors |
483 |
This only applies to users of Lc<AnyEvent::Fork:Early> and |
484 |
AnyEvent::Fork::Template. |
485 |
|
486 |
When a process created by AnyEvent::Fork exits, it might do so by |
487 |
calling exit, or simply letting perl reach the end of the program. |
488 |
At which point Perl runs all destructors. |
489 |
|
490 |
Not all destructors are fork-safe - for example, an object that |
491 |
represents the connection to an X display might tell the X server to |
492 |
free resources, which is inconvenient when the "real" object in the |
493 |
parent still needs to use them. |
494 |
|
495 |
This is obviously not a problem for AnyEvent::Fork::Early, as you |
496 |
used it as the very first thing, right? |
497 |
|
498 |
It is a problem for AnyEvent::Fork::Template though - and the |
499 |
solution is to not create objects with nontrivial destructors that |
500 |
might have an effect outside of Perl. |
501 |
|
502 |
PORTABILITY NOTES |
503 |
Native win32 perls are somewhat supported (AnyEvent::Fork::Early is a |
504 |
nop, and ::Template is not going to work), and it cost a lot of blood |
505 |
and sweat to make it so, mostly due to the bloody broken perl that |
506 |
nobody seems to care about. The fork emulation is a bad joke - I have |
507 |
yet to see something useful that you can do with it without running into |
508 |
memory corruption issues or other braindamage. Hrrrr. |
509 |
|
510 |
Cygwin perl is not supported at the moment, as it should implement fd |
511 |
passing, but doesn't, and rolling my own is hard, as cygwin doesn't |
512 |
support enough functionality to do it. |
513 |
|
514 |
SEE ALSO |
515 |
AnyEvent::Fork::Early (to avoid executing a perl interpreter), |
516 |
AnyEvent::Fork::Template (to create a process by forking the main |
517 |
program at a convenient time). |
518 |
|
519 |
AUTHOR |
520 |
Marc Lehmann <schmorp@schmorp.de> |
521 |
http://home.schmorp.de/ |
522 |
|