What is PERL_COPY_ON_WRITE Compile Option About

[DeadPoet]

To All:

Please enlighten me on a curious perl compile option. The option is know as PERL_COPY_ON_WRITE. What is it and what does it do?

Sincerely,
DeadPoet

[BrowserUk]

PERL_COPY_ON_WRITE has nothing to do with threads or process as far as I can see.

It is used by code within the pp_hot.c, regcomp.h and other places to avoid copying data until (and unless) it is written to.

An example with informative comment comes from pp_hot.c:

#ifdef PERL_COPY_ON_WRITE
/* The match may make the string COW. If so, brilliant, because that's
   just saved us one malloc, copy and free - the regexp has donated
   the old buffer, and we malloc an entirely new one, rather than the
   regexp malloc()ing a buffer and copying our original, only for
   us to throw it away here during the substitution.  */
if (SvIsCOW(TARG)) {
    sv_force_normal_flags(TARG, SV_COW_DROP_PV);
} else
#endif

[DeadPoet]

To All:

Here is what I just found at
http://yapc.mongueurs.net/yapc/talk/?tid=85

"Currently in perl when you write $a = $b, the contents of $b are immediately copied into $a. In theory perl could just pretend that it did the copy, and only actually do it if one of $a or $b is modified. If $b is a large string, and $a is never modified this could potentially be a big efficiency gain. This talk consists of simple pictures - how I didn't do it, and how I did."



DeadPoet

[BrowserUk]

It appears that currently, COW is only used for certain copy operations--some of those in the regex engine, shared hash keys (or values?), and a few other places--but not currently simple scalars.

It looks like the proposal is to extend it's reach.

[nobull]

That was an excelent talk. Highly informative and imaginatively illustrated with several people holding up cards in a carefully choreographed sequence.

[hardburn]

CoW refers to a method of optimizing memory usage for multithreaded programs. Consider a program that fork()s after intitlizing a variable. Under normal usage, those variables are copied into the memory space for each program. Under CoW, they point to the orginal variable in the parent. As long as each thread only reads the variables, this is safe. When a modification occurs, the variable is copied into the specific thread's memory space, thus keeping the change from other threads.

[ikegami]

I'm confused by your answer. You talk of threads created by fork, but fork creates new processes, not new threads (except in Windows). Does CoW affect multitasking or multithreading or both?

[hardburn]

"Processes are threads with an attitude. Or the other way around." (I belive this is attributed to Larry, and may not be an exact quote).

Forks and threads are conceputally similar, and for the purposes of CoW, are basically the same thing.