The I/O buffer caches are allocated by the percentage of
physical memory from the look of

	src/sys/arch/$MACHINE/$MACHINE/machdep.c  and
	src/sys/arch/mips/mips/mips_machdep.c  (for MIPS ports).

Is it more reasonable to use the size of memory
excluding kernel text + data + bss (+ symbol for DDB)
instead of whole of the physical memory
for calculating the buffer size?

If the kernel is huge (like an installation kernel which has md miniroot),
the kernel may run out of memory by the allocation of buffer caches.

How do you think about this?
How can we implement this correct?
--
ITOH, Yasufumi <yasufu-i@is.aist-nara.ac.jp>, <itohy@netbsd.org>