This bug causes the last PTE page not freed when a task's mmap is
freed.  The last PTE page corresponds to stack area which is usually
valid.  The bug causes slow memory leak, roughly 4k (one page) per died
process.

Pete Popov narrowed down the problem to PTE page leak before I further
spotted the root.

Jun

diff -Nru linux-test/include/asm-mips/pgtable.h.orig linux-test/include/asm-mips/pgtable.h
--- linux-test/include/asm-mips/pgtable.h.orig	Wed Jan  2 14:00:57 2002
+++ linux-test/include/asm-mips/pgtable.h	Thu Jan  3 13:04:51 2002
@@ -104,7 +104,7 @@
 #define PTRS_PER_PTE	1024
 #define PTRS_PER_PMD	1
 #define PTRS_PER_PGD	1024
-#define USER_PTRS_PER_PGD	(TASK_SIZE/PGDIR_SIZE)
+#define USER_PTRS_PER_PGD	((TASK_SIZE-1)/PGDIR_SIZE + 1)
 #define FIRST_USER_PGD_NR	0
 
 #define VMALLOC_START     KSEG2