On Wed, Mar 29, 2000 at 06:53:24PM +0900, Toru Nishimura wrote:

 > Isn't it ridiculous pmap_protect() does following, is it?

Yes :-)

 >         while (sva < eva) {
 > 		...
 >                 pte = ...
 >                 for (; sva < nssva; sva += NBPG, pte++) {
 >                         entry = pte->pt_entry;
 >                         if (!mips_pg_v(entry))
 >                                 continue;
 >                         entry = (entry & ~(mips_pg_m_bit() |
 >                             mips_pg_ro_bit())) | p;
 >                         pte->pt_entry = entry;
 >                       	...
 >                         MachTLBUpdate(sva | asid, entry);
 >                 }
 >         }
 > 
 > When pmap_protect() is called by uvmspace_fork1(), code above fills a
 > bunch of entries into TLB which has only the capacity of 56, I think.

Right, if it were me, I'd instead flush the entries from the TLB, and let
them get refilled via a miss.

Alternatively, MachTLBUpdate() could be changed to only update an entry
if it exists in the TLB, which might make some sense; that could be less
expensive than just nuking it and then letting a miss occur.

-- 
        -- Jason R. Thorpe <thorpej@zembu.com>