Coding for a Smaller Memory Footprint (Practical mod

14.2. Coding for a Smaller Memory Footprint

The following sections present proactive techniques that prevent processes from growing large in the first place.

14.2.1. Memory Reuse

Example 14-1. memory_hog.pl

use GTop ( );
my $gtop = GTop->new;
my $proc = $gtop->proc_mem($$);
print "size before:  ", $gtop->proc_mem($$)->size( ), " B\n";
{
    my $x = 'a' x 10**7;
    print "size inside: ", $gtop->proc_mem($$)->size( ), " B\n";
}
print "size  after: ", $gtop->proc_mem($$)->size( ), " B\n";

When executed, it prints:

size before:  1830912 B
size inside: 21852160 B
size  after: 21852160 B

This script starts by printing the size of the memory it occupied when it was first loaded. The opening curly brace starts a new block, in which a lexical variable $x is populated with a string 10,000,000 bytes in length. The script then prints the new size of the process and exits from the block. Finally, the script again prints the size of the process.

Since the variable $x is lexical, it is destroyed at the end of the block, before the final print statement, thus releasing all the memory that it was occupying. But from the output we can clearly see that a huge chunk of memory wasn't released to the OS—the process's memory usage didn't change. Perl reuses this released memory internally. For example, let's modify the script as shown in Example 14-2.

Example 14-2. memory_hog2.pl

use GTop ( );
my $gtop = GTop->new;
my $proc = $gtop->proc_mem($$);
print "size before :  ", $gtop->proc_mem($$)->size( ), " B\n";
{
    my $x = 'a' x 10**7;
    print "size inside : ", $gtop->proc_mem($$)->size( ), " B\n";
}
print "size  after : ", $gtop->proc_mem($$)->size( ), " B\n";
{
    my $x = 'a' x 10;
    print "size inside2: ", $gtop->proc_mem($$)->size( ), " B\n";
}
print "size  after2: ", $gtop->proc_mem($$)->size( ), " B\n";

When we execute this script, we will see the following output:

size before :  1835008 B
size inside : 21852160 B
size  after : 21852160 B
size inside2: 21852160 B
size  after2: 21852160 B

As you can see, the memory usage of this script was no more than that of the previous one.

So we have just learned that Perl programs don't return memory to the OS until they quit. If variables go out of scope, the memory they occupied is reused by Perl for newly created or growing variables.

Suppose your code does memory-intensive operations and the processes grow fast at first, but after a few requests the sizes of the processes stabilize as Perl starts to reuse the acquired memory. In this case, the wisest approach is to find this limiting size and set the upper memory limit to a slightly higher value. If you set the limit lower, processes will be killed unnecessarily and lots of redundant operations will be performed by the OS.

14.2.2. Big Input, Big Damage

This section demonstrates how a malicious user can bring the service down or cause problems by submitting unexpectedly big data.

Imagine that you have a guestbook script/handler, which works fine. But you've forgotten about a small nuance: you don't check the size of the submitted message. A 10 MB core file copied and pasted into the HTML textarea entry box intended for a guest's message and submitted to the server will make the server grow by at least 10 MB. (Not to mention the horrible experience users will go through when trying to view the guest book, since the contents of the binary core file will be displayed.) If your server is short of memory, after a few more submissions like this one it will start swapping, and it may be on its way to crashing once all the swap memory is exhausted.

To prevent such a thing from happening, you could check the size of the submitted argument, like this:

my $r = shift;
my %args = $r->args;
my $message = exists $args{message} ? $args{message} : '';
die "the message is too big" 
    unless length $message > 8192; # 8KB

While this prevents your program from adding huge inputs into the guest book, the size of the process will grow anyway, since you have allowed the code to process the submitted form's data. The only way to really protect your server from accepting huge inputs is not to read data above some preset limit. However, you cannot safely rely on the Content-Length header, since that can easily be spoofed.

You don't have to worry about GET requests, since their data is submitted via the query string of the URI, which has a hard limit of about 8 KB.

Think about disabling file uploads if you don't use them. Remember that a user can always write an HTML form from scratch and submit it to your program for processing, which makes it easy to submit huge files. If you don't limit the size of the form input, even if your program rejects the faulty input, the data will be read in by the server and the process will grow as a result. Here is a simple example that will readily accept anything submitted by the form, including fields that you didn't create, which a malicious user may have added by mangling the original form:

use CGI;
my $q = CGI->new;
my %args = map {$_ => $q->param($_)} $q->params;

If you are using CGI.pm, you can set the maximum allowed POSTsize and disable file uploads using the following setting:

use CGI;
$CGI::POST_MAX = 1048576;  # max 1MB allowed
$CGI::DISABLE_UPLOADS = 1; # disable file uploads

The above setting will reject all submitted forms whose total size exceeds 1 MB. Only non-file upload inputs will be processed.

If you are using the Apache::Request module, you can disable file uploads and limit the maximum POSTsize by passing the appropriate arguments to the new( ) function. The following example has the same effect as the CGI.pm example shown above:

my $apr = Apache::Request->new($r, 
                               POST_MAX        => 1048576,
                               DISABLE_UPLOADS => 1
                              );

Another alternative is to use the LimitRequestBody directive in httpd.conf to limit the size of the request body. This directive can be set per-server, per-directory, per-file, or per-location. The default value is 0, which means unlimited. As an example, to limit the size of the request body to 2 MB, you should add:

LimitRequestBody 2097152

The value is set in bytes (2097152 bytes = = 2 MB).

In this section, we have presented only a single example among many that can cause your server to use more memory than planned. It helps to keep an open mind and to explore what other things a creative user might try to do with your service. Don't assume users will only click where you intend them to.

14.2.3. Small Input, Big Damage

This section demonstrates how a small input submitted by a malicious user may hog the whole server.

Imagine an online service that allows users to create a canvas on the server side and do some fancy image processing. Among the inputs that are to be submitted by the user are the width and the height of the canvas. If the program doesn't restrict the maximum values for them, some smart user may ask your program to create a canvas of 1,000,000 × 1,000,000 pixels. In addition to working the CPU rather heavily, the processes that serve this request will probably eat all the available memory (including the swap space) and kill the server.

How can the user do this, if you have prepared a form with a pull-down list of possible choices? Simply by saving the form and later editing it, or by using a GET request. Don't forget that what you receive is merely an input from a user agent, and it can very easily be spoofed by anyone knowing how to use LWP::UserAgent or something equivalent. There are various techniques to prevent users from fiddling with forms, but it's much simpler to make your code check that the submitted values are acceptable and then move on.

If you do some relational database processing, you will often encounter the need to read lots of records from the database and then print them to the browser after they are formatted. Let's look at an example.

We will use DBI and CGI.pm for this example. Assume that we are already connected to the database server (refer to the DBI manpage for a complete reference to the DBI module):

my $q = new CGI;
my $default_hits = 10;
my $hits = int $q->param("hints") || $default_hits;

my $do_sql = "SELECT from foo LIMIT 0,$hits";
my $sth = $dbh->prepare($do_sql);
$sth->execute;

while (@row_ary = $sth->fetchrow_array) {
    # do DB accumulation into some variable
}
# print the data
...

In this example, the records are accumulated in the program data before they are printed. The variables that are used to store the records that matched the query will grow by the size of the data, in turn causing the httpd process to grow by the same amount.

Imagine a search engine interface that allows a user to choose to display 10, 50, or 100 results. What happens if the user modifies the form to ask for 1,000,000 hits? If you have a big enough database, and if you rely on the fact that the only valid choices would be 10, 50, or 100 without actually checking, your database engine may unexpectedly return a million records. Your process will grow by many megabytes, possibly eating all the available memory and swap space.

The obvious solution is to disallow arbitrary inputs for critical variables like this one. Another improvement is to avoid the accumulation of matched records in the program data. Instead, you could use DBI::bind_columns( ) or a similar function to print each record as it is fetched from the database. In Chapter 20 we will talk about this technique in depth.

14.2.4. Think Production, Not Development

Developers often use sample inputs for testing their new code. But sometimes they forget that the real inputs can be much bigger than those they used in development.

Consider code like this, which is common enough in Perl scripts:

{
    open IN, $file or die $!;
    local $/;
    $content = <IN>; # slurp the whole file in
    close IN;
}

If you know for sure that the input will always be small, the code we have presented here might be fine. But if the file is 5 MB, the child process that executes this script when serving the request will grow by that amount. Now if you have 20 children, and each one executes this code, together they will consume 20 × 5 MB = 100 MB of RAM! If, when the code was developed and tested, the input file was very small, this potential excessive memory usage probably went unnoticed.

Try to think about the many situations in which your code might be used. For example, it's possible that the input will originate from a source you did not envisage. Your code might behave badly as a result. To protect against this possibility, you might want to try to use other approaches to processing the file. If it has lines, perhaps you can process one line at a time instead of reading them all into a variable at once. If you need to modify the file, use a temporary file. When the processing is finished, you can overwrite the source file. Make sure that you lock the files when you modify them.

Often you just don't expect the input to grow. For example, you may want to write a birthday reminder process intended for your own personal use. If you have 100 friends and relatives about whom you want to be reminded, slurping the whole file in before processing it might be a perfectly reasonable way to approach the task.

But what happens if your friends (who know you as one who usually forgets their birthdays) are so surprised by your timely birthday greetings that they ask you to allow them to use your cool invention as well? If all 100 friends have yet another 100 friends, you could end up with 10,000 records in your database. The code may not work well with input of this size. Certainly, the answer is to rewrite the code to use a DBM file or a relational database. If you continue to store the records in a flat file and read the whole database into memory, your code will use a lot of memory and be very slow.

14.2.5. Passing Variables

Let's talk about passing varia bles to a subroutine. There are two ways to do this: you can pass a copy of the variable to the subroutine (this is called passing by value) or you can instead pass a reference to it (a reference is just a pointer, so the variable itself is not copied). Other things being equal, if the copy of the variable is larger than a pointer to it, it will be more efficient to pass a reference.

Let's use the example from the previous section, assuming we have no choice but to read the whole file before any data processing takes place and its size is 5 MB. Suppose you have some subroutine called process( ) that processes the data and returns it. Now say you pass $content by value and process( ) makes a copy of it in the familiar way:

my $content = qq{foobarfoobar};
$content = process($content);
sub process {
    my $content = shift;
    $content =~ s/foo/bar/gs;
    return $content;
}

You have just copied another 5 MB, and the child has grown in size by another 5 MB. Assuming 20 Apache children, you can multiply this growth again by factor of 20—now you have 200 MB of wasted RAM! This will eventually be reused, but it's still a waste. Whenever you think the variable may grow bigger than a few kilobytes, definitely pass it by reference.

There are several forms of syntax you can use to pass and use variables passed by reference. For example:

my $content = qq{foobarfoobar};
process(\$content);
sub process {
    my $r_content = shift; 
    $$r_content =~ s/foo/bar/gs;
}

Here $content is populated with some data and then passed by reference to the subroutine process( ), which replaces all occurrences of the string foo with the string bar. process( ) doesn't have to return anything—the variable $content was modified directly, since process( ) took a reference to it.

If the hashes or arrays are passed by reference, their individual elements are still accessible. You don't need to dereference them:

$var_lr->[$index]  get $index'th element of an array via a ref
$var_hr->{$key}    get $key'th element of a hash via a ref

Note that if you pass the variable by reference but then dereference it to copy it to a new string, you don't gain anything, since a new chunk of memory will be acquired to make a copy of the original variable. The perlref manpage provides extensive information about working with references.

Another approach is to use the @_ array directly. Internally, Perl always passes these variables by reference and dereferences them when they are copied from the @_ array. This is an efficiency mechanism to allow you to write subroutines that take a variable passed as a value, without copying it.

process($content);
sub process {
  $_[0] =~ s/foo/bar/gs;
}

From perldoc perlsub:

The array @_ is a local array, but its elements are aliases for the actual scalar 
parameters.  In particular, if an element $_[0] is updated, the corresponding 
argument is updated (or an error occurs if it is not possible to update)...

Be careful when you write this kind of subroutine for use by someone else; it can be confusing. It's not obvious that a call like process($content); modifies the passed variable. Programmers (the users of your library, in this case) are used to subroutines that either modify variables passed by reference or expressly return a result, like this:

$content = process($content);

You should also be aware that if the user tries to submit a read-only value, this code won't work and you will get a runtime error. Perl will refuse to modify a read-only value:

$content = process("string foo");

14.2.6. Memory Leakage

It's normal for a process to grow when it processes its first few requests. They may be different requests, or the same requests processing different data. You may try to reload the same request a few times, and in many cases the process will stop growing after only the second reload. In any case, once a representative selection of requests and inputs has been executed by a process, it won't usually grow any more unless the code leaks memory. If it grows after each reload of an identical request, there is probably a memory leak.

The experience might be different if the code works with some external resource that can change between requests. For example, if the code retrieves database records matching some query, it's possible that from time to time the database will be updated and that a different number of records will match the same query the next time it is issued. Depending on the techniques you use to retrieve the data, format it, and send it to the user, the process may increase or decrease in size, reflecting the changes in the data.

The easiest way to see whether the code is leaking is to run the server in single-process mode (httpd -X), issuing the same request a few times to see whether the process grows after each request. If it does, you probably have a memory leak. If the code leaks 5 KB per request, then after 1,000 requests to run the leaking code, 5 MB of memory will have leaked. If in production you have 20 processes, this could possibly lead to 100 MB of leakage after a few tens of thousands of requests.

This technique to detect leakage can be misleading if you are not careful. Suppose your process first runs some clean (non-leaking) code that acquires 100 KB of memory. In an attempt to make itself more efficient, Perl doesn't give the 100 KB of memory back to the operating system. The next time the process runs any script, some of the 100 KB will be reused. But if this time the process runs a script that needs to acquire only 5 KB, you won't see the process grow even if the code has actually leaked these 5 KB. Now it might take 20 or more requests for the leaking script served by the same process before you would see that process start growing again.

A process may leak memory for several reasons: badly written system C/C++ libraries used in the httpd binary and badly written Perl code are the most common. Perl modules may also use C libraries, and these might leak memory as well. Also, some operating systems have been known to have problems with their memory-management functions.

If you know that you have no leaks in your code, then for detecting leaks in C/C++ libraries you should either use the technique of sampling the memory usage described above, or use C/C++ developer tools designed for this purpose. This topic is beyond the scope of this book.

The Apache::Leak module (derived from Devel::Leak) might help you to detect leaks in your code. Consider the script in Example 14-3.

Example 14-3. leaktest.pl

use Apache::Leak;

my $global = "FooA";

leak_test {
    $$global = 1;
    ++$global;
};

You do not need to be inside mod_perl to use this script. The argument to leak_test( ) is an anonymous sub or a block, so you can just throw in any code you suspect might be leaking. The script will run the code twice. The first time, new scalar values (SVs) are created, but this does not mean the code is leaking. The second pass will give better evidence.

From the command line, the above script outputs:

ENTER: 1482 SVs
new c28b8 : new c2918 : 
LEAVE: 1484 SVs
ENTER: 1484 SVs
new db690 : new db6a8 : 
LEAVE: 1486 SVs
!!! 2 SVs leaked !!!

This module uses the simple approach of walking the Perl internal table of allocated SVs. It records them before entering the scope of the code under test and after leaving the scope. At the end, a comparison of the two sets is performed, sv_dump( ) is called for anything that did not exist in the first set, and the difference in counts is reported. Note that you will see the dumps of SVs only if Perl was built with the -DDEBUGGING option. In our example the script will dump two SVs twice, since the same code is run twice. The volume of output is too great to be presented here.

Our example leaks because $$global = 1; creates a new global variable, FooA (with the value of 1), which will not be destroyed until this module is destroyed. Under mod_perl the module doesn't get destroyed until the process quits. When the code is run the second time, $global will contain FooB because of the increment operation at the end of the first run. Consider:

$foo = "AAA";
print "$foo\n";
$foo++;
print "$foo\n";

which prints:

AAA
AAB

So every time the code is executed, a new variable (FooC, FooD, etc.) will spring into existence.

Apache::Leak is not very user-friendly. You may want to take a look at B::LexInfo. It is possible to see something that might appear to be a leak, but is actually just a Perl optimization. Consider this code, for example:

sub test { my ($string) = @_;}
test("a string");

B::LexInfo will show you that Perl does not release the value from $string unless you undef( ) it. This is because Perl anticipates that the memory will be needed for another string, the next time the subroutine is entered. You'll see similar behavior for @array lengths, %hash keys, and scratch areas of the padlist for operations such as join( ), ., etc.

Let's look at how B::LexInfo works. The code in Example 14-4 creates a new B::LexInfo object, then runs cvrundiff( ), which creates two snapshots of the lexical variables' padlists—one before the call to LeakTest1::test( ) and the other, in this case, after it has been called with the argument "a string". Then it calls diff -u to generate the difference between the snapshots.

Example 14-4. leaktest1.pl

package LeakTest1;
use B::LexInfo ( );

sub test { my ($string) = @_;}

my $lexi = B::LexInfo->new;
my $diff = $lexi->cvrundiff('LeakTest1::test', "a string");
print $$diff;

In case you aren't familiar with how diff works, - at the beginning of the line means that that line was removed, + means that a line was added, and other lines are there to show the context in which the difference was found. Here is the output:

--- /tmp/B_LexInfo_3099.before        Tue Feb 13 20:09:52 2001
+++ /tmp/B_LexInfo_3099.after        Tue Feb 13 20:09:52 2001
@@ -2,9 +2,11 @@
   {
     'LeakTest1::test' => {
       '$string' => {
-        'TYPE' => 'NULL',
+        'TYPE' => 'PV',
+        'LEN' => 9,
         'ADDRESS' => '0x8146d80',
-        'NULL' => '0x8146d80'
+        'PV' => 'a string',
+        'CUR' => 8
       },
       '_ _SPECIAL_ _1' => {
         'TYPE' => 'NULL',

Perl tries to optimize the speed by keeping the memory allocated for $string, even after the variable is destroyed.

Let's run the script from Example 14-3 with B::LexInfo (see Example 14-5).

Example 14-5. leaktest2.pl

package LeakTest2;
use B::LexInfo ( );

my $global = "FooA";

sub test {
    $$global = 1;
    ++$global;
}

my $lexi = B::LexInfo->new;
my $diff = $lexi->cvrundiff('LeakTest2::test');
print $$diff;

Here's the result:

--- /tmp/B_LexInfo_3103.before Tue Feb 13 20:12:04 2001
+++ /tmp/B_LexInfo_3103.after         Tue Feb 13 20:12:04 2001
@@ -5,7 +5,7 @@
         'TYPE' => 'PV',
         'LEN' => 5,
         'ADDRESS' => '0x80572ec',
-        'PV' => 'FooA',
+        'PV' => 'FooB',
         'CUR' => 4
       }
     }

We can clearly see the leakage, since the value of the PV entry has changed from one string to a different one. Compare this with the previous example, where a variable didn't exist and sprang into existence for optimization reasons. If you find this confusing, probably the best approach is to run diff twice when you test your code.

Now let's run the cvrundiff( ) function on this example, as shown in Example 14-6.

Example 14-6. leaktest3.pl

package LeakTest2;
use B::LexInfo ( );

my $global = "FooA";

sub test {
    $$global = 1;
    ++$global;
}

my $lexi = B::LexInfo->new;
my $diff = $lexi->cvrundiff('LeakTest2::test');
$diff    = $lexi->cvrundiff('LeakTest2::test');
print $$diff;

Here's the output:

--- /tmp/B_LexInfo_3103.before Tue Feb 13 20:12:04 2001
+++ /tmp/B_LexInfo_3103.after         Tue Feb 13 20:12:04 2001
@@ -5,7 +5,7 @@
         'TYPE' => 'PV',
         'LEN' => 5,
         'ADDRESS' => '0x80572ec',
-        'PV' => 'FooB',
+        'PV' => 'FooC',
         'CUR' => 4
       }
     }

We can see the leak again, since the value of PV has changed again, from FooB to FooC. Now let's run cvrundiff( ) on the second example script, as shown in Example 14-7.

Example 14-7. leaktest4.pl

package LeakTest1;
use B::LexInfo ( );

sub test { my ($string) = @_;}

my $lexi = B::LexInfo->new;
my $diff = $lexi->cvrundiff('LeakTest1::test', "a string");
   $diff = $lexi->cvrundiff('LeakTest1::test', "a string");
print $$diff;

No output is produced, since there is no difference between the second and third runs. All the data structures are allocated during the first execution, so we are sure that no memory is leaking here.

Apache::Status includes a StatusLexInfo option that can show you the internals of your code via B::LexInfo. See Chapter 21 for more information.