> > Would it be possible to add a syscall to "zero" a range of memory? I mean
> > not only clear it to zero, as with memset, but if any complete pages are
> > within the range, unmap them, and replace them with zero mappings (as
> > mmap()'ing /dev/zero gives you).
>
> You could just mmap /dev/zero instead. No need to unmap what was previously
> there, it will get automatically discarded. The malloc implementation would
> need to keep a file descriptor for /dev/zero to avoid the overhead of
> opening the file over and over again. Other than that, there would be no
> overhead compared to a special-case system call.
Ah, so a mmap of /dev/zero doesn't involve any of the overhead of a
"normal" mmapping of a file, then? If so, and you can successfully mmap in
blank pages over sbrk()'d space with this, without consuming any
resources, then it does indeed match my request.
> I think one should perform measurements first to find out:
> - how many, and which applications suffer from this phenomen. E.g. after
> running Netscape for one day, how many pages could be returned to the
> system. LD_PRELOAD is your friend, here.
Agreed.
> - for those applications, how much overhead is introduced by inserting the
> additional mappings. For example, drivers/char/mem.c:read_zero only changes
> the mapping if the user requested at least four pages.
And whether the implementation should wait for eight pages, or whatever,
to be free before discarding them in this manner.
> Once those data are available, we could decide whether this algorithm
> becomes part of the standard C library, or a special malloc library that
> is only linked into selected applications.
>
> Regards,
> Martin
-- Kenneth Albanowski (kjahds@kjahds.com, CIS: 70705,126)