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
|
=pod
=head1 NAME
B<XPA Race Conditions>
=head1 SYNOPSIS
Potential XPA race conditions and how to avoid them.
=head1 DESCRIPTION
Currently, there is only one known circumstance in which XPA can get
(temporarily) deadlocked in a race condition: if two or more XPA
servers send messages to one another using an XPA client routine such
as XPASet(), they can deadlock while each waits for the other server
to respond. (This can happen if the servers call XPAPoll() with a
time limit, and send messages in between the polling call.) The
reason this happens is that both client routines send a string to the
other server to establish the handshake and then wait for the server
response. Since each client is waiting for a response, neither is able
to enter its event-handling loop and respond to the other's
request. This deadlock will continue until one of the timeout periods
expire, at which point an error condition will be triggered and the
timed-out server will return to its event loop.
Starting with version 2.1.6, this rare race condition can be
avoided by setting the XPA_IOCALLSXPA environment variable for servers
that will make client calls. Setting this variable causes all XPA
socket IO calls to process outstanding XPA requests whenever the
primary socket is not ready for IO. This means that a server making a
client call will (recursively) process incoming server requests while
waiting for client completion. It also means that a server callback
routine can handle incoming XPA messages if it makes its own XPA call.
The semi-public routine oldvalue=XPAIOCallsXPA(newvalue) can be used
to turn this behavior off and on temporarily. Passing a 0 will turn
off IO processing, 1 will turn it back on. The old value is returned
by the call.
By default, the XPA_IOCALLSXPA option is turned off, because we judge
that the added code complication and overhead involved will not be
justified by the amount of its use. Moreover, processing XPA requests
within socket IO can lead to non-intuitive results, since incoming
server requests will not necessarily be processed to completion in the
order in which they are received.
Aside from setting XPA_IOCALLSXPA, the simplest way to avoid this race
condition is to multi-process: when you want to send a client message,
simply start a separate process to call the client routine, so that
the server is not stopped. It probably is fastest and easiest to use
fork() and then have the child call the client routine and exit. But
you also can use either the system() or popen() routine to start one
of the command line programs and do the same thing. Alternatively, you
can use XPA's internal launch() routine instead of system(). Based on
fork() and exec(), this routine is more secure than system() because
it does not call /bin/sh.
Starting with version 2.1.5, you also can send an XPAInfo() message with
the mode string "ack=false". This will cause the client to send a message
to the server and then exit without waiting for any return message from
the server. This UDP-like behavior will avoid the server deadlock when
sending short XPAInfo messages.
=head1 SEE ALSO
See xpa(n) for a list of XPA help pages
=cut
|
