Re: [xml] win32 thread safety patch

["Robert Collins" <robert collins itdomain com au>]

===
----- Original Message -----
From: "Serguei Narojnyi" <serguein opencola com>
To: <xml gnome org>
Sent: Tuesday, January 15, 2002 2:02 AM
Subject: RE: [xml] win32 thread safety patch


> > My objection is to the thread-per-thread overhead.
> >
> > Now it's obvious (now I'm a little more awake) that a cached
> > copy of the
> > pointer is needed, along with the thread HANDLE from
> > syncronisation. But
> > a thread-per thread approach is very wasteful, only one cleanup
thread
> > should be needed.
>
> the solution is very obvious, have some sort of protected queue. then
> xmlGlobalStateCleanupHelperParams gets pushed down there from the
> xmlGetGlobalState and there is some thread that checks that queue say
once
> every 5 seconds or so and cleans up all stuff in queue for threads
that are
> dead. i'll come around implementing it once i have some free time if
someone
> else wouldn't.

Thats a polled approach. What I was actually thinking was (pseudoish
code)

HANDLE aMutex; /* protects the next three variables */
unsigned int threads = 0;
HANDLE *cleanuphandles; /* has 1 more entry than cleanupptrs - an Event
Handle */
void *cleanupptrs;

cleanupthread ()
{
  while (threads) {
    object = WaitOnMultipleObjects (cleanuphandles, INFINITY);
    /* grab aMutex */
    if (object == WAIT_OBJECT_0)
      continue;
    else if (object > WAIT_OBJECT_0 && object < WAIT_OBJECT_0 + threads)
      {

        docleanup (cleanupptrs[object - WAIT_OBJECT_0])
        removefromarray (cleanuphandles, object - WAIT_OBJECT_0,
threads)
        removefromarray (cleanupptrs, object - WAIT_OBJECT_0, threads)
        --threads;
      }
    /* else if (object == WAIT_FAILED... */
    /* release aMutex */
}

I'll leave the queue addition/removal code as an exercise to the reader
:}.

This is better than a polled approach as it will stay completely
inactive until a new thread is added (and the Event is triggered), or a
thread quits.

Rob