hi,

while it would be possible to optimize for applications which read from but
don't write to zero-fill memory, or write only zeroes to zero-fill memory,
that would greatly reduce the performance of all real applications.
real applications invariably write non-zero values to memory that they
allocate, so optimizing for this case is preferred.

-Chuck


On Sun, Mar 17, 2002 at 07:31:00PM +0100, Reinoud Zandijk wrote:
> Dear all,
> 
> just reading the UVM dissitation and I noted the following remark om page 
> 11 :
> 
> ``This (backing object for a page) is usually a file or `zero-fill' memory
> wich means that the area should be filled with zeroed memeory as its
> accessed. (....) For zero-fill memory, the copy-on-write flag doesn't
> matter''
> 
> I disagree with this ... why shouldn't zero filled memory be backed with
> just *one* by all processes shared R/O page when it faults for reading that
> uses copy-on-write to get a private one if something changed ?
> 
> Is this allready done in UVM ? or is it explained later on ? seems spilling
> pages otherwise.... if a process would only just _READ_ that memory it
> would get heaps of zeroed memory pages ... where it could be just one zero 
> page it reads over and over again ... writing one of the pages then gives a 
> new fresh page for its modifications.
> 
> I dont know how our `calloc' works but maybe that could be integrated too ? 
> since you _know_ that those pages are zero anyway.
> 
> A further optimalisation could even be to check if the page is zero still 
> and rename that page to the anonymous _one_ page when memory gets tight ...
> 
> Cheers,
> Reinoud
> 
> (trying to understand UVM/pmap better)