Subject: [PATCH] uvm_map.c
To: None <,,>
From: J. Scott Kasten <>
List: netbsd-bugs
Date: 02/29/2000 18:42:37

I'm a NetBSD newbie.  I recently installed 1.4 for arch Sun3 on a Sun3/80
(MC68030) with 16MB ram with 1Gig HD.  When the machine was under heavy load
compiling c++ code, I was periodically getting a kernel panic:

"Panic: Malloc: out of space in kmem_map"

After recompiling the kernal with both egcs and gcc- and playing with
the hardware, I finally convinced myself due to the consistency of the error
that this was neither a freak hardware problem, nor a kernel compile problem.

I researched the archives, and the port-m68k list in May of 1999 had discussion
about exactly this type of issue from Chuck Cranor and Ken Nakata.  Chuck
found an issue with pointer wrap around in uvm_amap.c.  His patch was applied
back at release 1.9 of the file.  I verified that his patch was still in the
current release, however, I was still having the problem.  This indicated to
me that there must be other places where this can happen.

I beleive that I found one such place in uvm_map.c release  After
inserting a one line conditional for the wrap problem and recompiling the
kernel, I am pleased to say that I have not seen the error in over 4 days
now.  I've kept the box compiling c++ code or otherwise abusing it for
the entire 4 days, 24 hours a day.  I was getting the problem every 4 to
8 hours before.  Therefore, I beleive the problem is fixed and offer the
enclosed diff to the port maintainers for consideration.

(There is a second spot in uvm_map.c that appears similar to the one
 I patched, however, the outer code block seems to have sufficient
 conditionals to avoid the problem there, so I left it untouched.
 This second block is much further down in the file from the patch.)

*** uvm_map.c.orig      Fri Feb 25 21:03:18 2000
--- uvm_map.c   Fri Feb 25 21:10:10 2000
*** 1671,1676 ****
--- 1671,1677 ----
        if (uvm_map_lookup_entry(map, start, &entry)) {
                UVM_MAP_CLIP_START(map, entry, start);
+                 if (end >= start)
                UVM_MAP_CLIP_END(map, entry, end);              /* to be safe */
        else {

J. Scott Kasten

jsk AT tetracon-eng DOT net

"That wasn't an attack.  It was preemptive retaliation!"