jermoluk@hppcgo.UUCP (jermoluk) (03/02/85)
xx
We have the code for the new AT&T Sys 5.2.2 mods and have found a couple
of what we believe to be major bugs. If anyone else has this up perhaps
they can confirm/deny/explain it.
1. The segment map, page table (pde), and disk map (dbd) are not
swapped with the process. This would tie up a minimum of 4.5k
(3 pages per region, 3 regions per process) per incore process.
2. In exec.c 410 files are loaded text and data in a single segment
and the segment is made r/w. No text sharing is therefore accomplished.
This makes loading kind of slow, so to speak. COFF format files
of 570/575 will get ENOEXEC no matter what you do.
3. If the TWOPASS variable is not set then regions are aged before
calling getpages() to prune the working sets. This clears the
reference bits, and will result in getpages() stealing all of
the pages of a process. This will thrash on the free list. Code
should steal pages then clear the reference bits.
If we have missed something then I apoligize in advance, but it appears
there are some problems here. If anyone wants to discuss further in length
drop me some mail. Anyone has solutions, etc probably good stuff for the
net.
tom jermoluk
hplabs!hppcgo!jermoluk
UNIX is a trademark of AT&T Bell Laboratories (etc. , etc.)