CVS commit: src/sys/arch

["Maxime Villard" <maxv%netbsd.org@localhost>]

Module Name:    src
Committed By:   maxv
Date:           Thu May 26 07:24:55 UTC 2016

Modified Files:
        src/sys/arch/amd64/amd64: locore.S
        src/sys/arch/i386/i386: locore.S

Log Message:
There is an issue in the way the fillkpt macro sets up pages on both
amd64 and i386.

The fillkpt loop is equivalent to the following:

        do {
                /* fill in the slot */
                /* increment %ebx to the next slot */
                /* increment %eax to the next pa */
        } while (%ecx > 0)

The issue here is that if %ecx = 0 (i.e., the chunk we are trying to
map is zero-sized), there is still one entry created in the page table.
The kernel expects the va<->pa translation to be linear in low memory.
If there is a zero-sized chunk, the dead entry creates a +4096 offset in
the virtual space, with two consecutive entries that point to the same
physical address. In other words, the mappings are not linear anymore,
which causes the kernel to die.

Before my recent changes, there were only two big chunks that were
mapped, and neither of these could be zero-sized. Now, with multiple,
fine-grained chunks, it is possible that the [SYMS]+[PRELOADED_MODULES]
chunk could be zero-sized.

[PRELOADED_MODULES] is almost never here, and [SYMS] is always here on
default kernels. Except for floppies, where the bootloader does not load
[SYMS].

Should fix PR 51148.


To generate a diff of this commit:
cvs rdiff -u -r1.93 -r1.94 src/sys/arch/amd64/amd64/locore.S
cvs rdiff -u -r1.124 -r1.125 src/sys/arch/i386/i386/locore.S

Please note that diffs are not public domain; they are subject to the
copyright notices on the relevant files.