safe to access shared data structure from signal handler

Question

I'm trying to decide wether it's safe to access a common (read: shared between handler-code and rest of the programm) data structure from a signal handler in perl (v5.14.2) built for x86_64-linux-thread-multi, but target platform is solaris11).

perlipc has the following sample code:

use POSIX ":sys_wait_h"; # for nonblocking read
my %children;
$SIG{CHLD} = sub {
    # don't change $! and $? outside handler
    local ($!, $?);
    my $pid = waitpid(-1, WNOHANG);
    return if $pid == -1;
    return unless defined $children{$pid};
    delete $children{$pid};
    cleanup_child($pid, $?);
};
while (1) {
    my $pid = fork();
    die "cannot fork" unless defined $pid;
    if ($pid == 0) {
        # ...
        exit 0;
    } else {
        $children{$pid}=1;
        # ...
        system($command);
        # ...
   }
}

So, %children is accessed from the while-loop and the handler. This seems to be no problem as:

1. There won't be two processes having the same pid
2. Access is keyed by pid (I am not sure if $childer{pid}=1 is atomic and interruptible without causing corruption, though.)

Now, i'm trying to do even more in my handler:

my %categoryForPid;
my %childrenPerCategory;

$SIG{CHLD} = sub {
    # ... acquire pid like above
    my $category = $categoryForPid{$pid};
    $childrenPerCategory{$category}--;
    delete $categoryForPid{$pid};
}

while (1) {
    # ... same as above
    } else {
        $children{$pid}=1;
        my $category = # ... chose some how
        $childrenPerCategory{$category}++;
        $categoryForPid{$pid} = $category;
        # ...
    }
}

The idea here is: every child belongs to a certain category (N to 1). I want to keep track of how many children per category exist. That information could be derived from $categoryForPid, but i think that might be problematic also (e.g., when the subroutine doing the computation gets interrupted while summing up).

So my question is:

• Do I need to synchronize here somehow?

And on a side note:

• Are nested invocations of the signal handler possible in perl 5.12, or are they linearized by the interpreter?

Update

In addition to the problem spotted by @goldilocks and his proposed solution I block signals now while updating the data structures to ensure "atomicity":

my $sigset = POSIX::SigSet->new(SIGCHLD);

sub ublk {
    unless (defined sigprocmask(SIG_UNBLOCK, $sigset)) {
        die "Could not unblock SIGCHLD\n";
    }
}

sub blk {
    unless (defined sigprocmask(SIG_BLOCK, $sigset)) {
        die "Could not block SIGCHLD\n";
    }
}

while (1) {
    # ... same as above
    } else {
         blk;
         $children{$pid}=1;
         my $category = # ... chose some how
         $childrenPerCategory{$category}++;
         $categoryForPid{$pid} = $category;
         ublk;
         # ...
    }
}

Answer 1

Seems like a bad idea to me. IPC::Semaphore might solve the problem, if you can get them to work properly in a signal handler -- if control does not return until the handler exits, you're out of luck. However, you could get around that by locking in the parent and having the child wait on the lock until initialization is complete; the handler is not involved with the semaphore. You'd only actually need one lock for that, I think. Anyway:

my @array = (1..10);
my $x = 'x';

$SIG{'USR1'} = sub {
    print "SIGUSER1\n";
    undef @array;
    $x = '!!';
};

print "$$\n";

foreach (@array) {
    print "$_:\n";
    sleep(2);
    print "\t$x\n";
    print "\t$array[$_ - 1]\n";
}

Not surprisingly, does this:

2482
1:
    x
    1
2:
    x
    2
3:
SIGUSER1
    !!
Use of uninitialized value within @array in concatenation (.) or string at ./test.pl line 42.

Implying that if you catch the signal at this point:

    my $category = # ... chose some how

$categoryForPid{$pid} will be non-existent in the handler. Etc. Ie, yes you have to synchronize.