Chuck Silvers <chuq@chuq.com> writes:
> the "bad" aliases we're generating now should not cause any problems,
> because all mappings of loaned pages are read-only and thus safe to cache
> if all the changes made to the page before the page is loaned are flushed
> to main memory when the loan is created.  that's why the cache flush in
> pmap_protect() fixes the problem.  are you sure NEW_PIPE still has
> problems on mips?

I couldn't figure out the scenario, but I can confirm that there is
virtual alias wich was detected by hardware(R4400 with S-cache).

My experiment was;

1) tweak pmap.c to enforce virtual alias elimination is enabled for
   R4400 with SC.
   (it is disabled by default for CPUs with SC.)

2) compile kernel with -DDEBUG.

3) make sure NEW_PIPE is used.
   % cat opt_pipe.h
/* option `PIPE_NODIRECT' not defined */
/* option `PIPE_SOCKETPAIR' not defined */

4) boot kernel.

5) make sure VCED_count's are all 0 with ddb.
   this means there was no virtual alias (no VCED was detected by hardware).

6) perform simple tar test.
   % cd /; tar cf - bin | (cd /tmp; tar xfpv -)

7) examine VCED_count's and see there was virtual alias.

db> x/x VCED_count,12
VCED_count:     0           0           0           0           0           0
VCED_count+0x18:        0           0           c6bc        0           0
VCED_count+0x2c:        0           0           0           0           0
VCED_epc:       8012d83c    ca55e7c0

--- Takao Shinohara