> > > This patch is seriously broken. Request a very large file (100MB or greater)
>and watch what
> happens
> > > to memory usage.
> > >
> > > The problem is this loop. We basically read the entire content of the file into
>memory before
> > > sending it out on the network. Haven't given much though on the best way to fix
>this.
> > >
> > > > + APR_BRIGADE_FOREACH(bucket, b) {
> > > > + const char *str;
> > > > + apr_size_t n;
> > > > +
> > > > + rv = apr_bucket_read(bucket, &str, &n, APR_BLOCK_READ);
> > > > + apr_brigade_write(ctx->b, NULL, NULL, str, n);
> > > > + }
> >
> > I don't see how that could happen. We only enter that section of the
> > core_output_filter if we are saving some data off to the side for
> > keepalive requests. In fact, we specifically do not enter this loop if we
> > are serving a file from disk.
>
> Attach a debugger and watch what happens. I am seeing the following buckets...
>
> 1 heap bucket containing the headers
> 1 file bucket with the file descriptor
> 1 eos bucket
>
> The following code is hit and we enter the conditional because the last bucket was
>an eos and the
> connection is keep-alive.
>
> if ((!fd && !more &&
> (nbytes < AP_MIN_BYTES_TO_WRITE) && !APR_BUCKET_IS_FLUSH(e))
> || (APR_BUCKET_IS_EOS(e) && c->keepalive)) {
>
> I think the logic in the conditional is just wrong.
I agree completely. I think I can fix this in a few minutes. Watch for a
patch.
Ryan
_______________________________________________________________________________
Ryan Bloom [EMAIL PROTECTED]
406 29th St.
San Francisco, CA 94131
-------------------------------------------------------------------------------
