Re: [pcre-dev] An issue from a pcre user

Top Page
Delete this message
Author: Philip Hazel
Date:  
To: 徐?z
CC: pcre-dev
Subject: Re: [pcre-dev] An issue from a pcre user
On Mon, 22 Nov 2010, 徐?z wrote:

> Hi,
>
> When I used the pattern "((.|\n)+)" to match a huge buffer which size was
> about 32Kbytes, I got a segmentation fault. Is it a bug?


This is not a bug, though it is frequently mis-reported as one, so much
so that I keep this standard response on file: (I think I posted it yet
again only last week).

------------------------------------------------------------------------
1. Matching a regular expression is like finding your way through a
forest with many branching paths. As PCRE passes each junction, it has
to remember data so that it can backtrack to that point if necessary. By
default, it uses recursion to store this data on the process stack,
because that is fast. However, it can alternatively be compiled to use
the heap instead (run ./configure with --disable-stack-for-recursion),
but that slows performance.

2. It is very easy to write a regular expression that has a very large
number of branches (unlimited repetition of a group, for example). When
PCRE goes deep into such a tree, it may use a lot of memory.

3. Even in these days of gigabyte main memories, some operating system
environments set small default limits on the maximum size of the process
stack, for example, 8Mb. Thus, it is often the case that there is more
heap than stack available (by default). A matching operation that needs
a lot of memory may succeed if the heap is used, but run out of memory
if the stack is used.

4. Running out of stack often causes a segfault. Because of this, PCRE
contains the facility to limit the depth of recursion so as to return an
error code instead. However, the default value is large, so it does not
normally come into play unless you explicitly set a smaller value.

5. If you are running into a problem of stack overflow, you have the
following choices:

  (a) Work on your regular expression pattern so that it uses less 
      memory. Sometimes using atomic groups can help with this.
  (b) Increase the size of your process stack.
  (c) Compile PCRE to use the heap instead of the stack.
  (d) Set PCRE's recursion limit small enough so that it gives an error
      before the stack overflows.    


6. There is more discussion of some of these ideas in the pcrestack.3
man page.
------------------------------------------------------------------------

In your case, using a repeated parenthesised subpattern to match just
one character many times uses a lot of memory.

Philip

--
Philip Hazel