Re: [PATCH v2 0/8] sparc: use asm-generic/dma-mapping-common.h and pci-dma-compat.h

[Ingo Molnar <mingo@elte.hu>]

* David Miller <davem@davemloft.net> wrote:

> From: Joerg Roedel <joerg.roedel@amd.com>
> Date: Tue, 14 Jul 2009 11:23:55 +0200
> 
> > On Tue, Jul 14, 2009 at 10:40:16AM +0900, FUJITA Tomonori wrote:
> >> On Mon, 13 Jul 2009 20:56:21 -0400
> >> Robert Reif <reif@earthlink.net> wrote:
> >> 
> >> > The bad address is within the kernel so it looks like
> >> > it's catching a real bug.
> >> > 
> >> > cat kallsyms | grep f0007000
> >> > f0007000 T trapbase_cpu3
> >> > 
> >> > WARNING: at lib/dma-debug.c:873 check_for_illegal_area+0xc8/0x100()
> >> > esp ffd7ba30: DMA-API: device driver maps memory from kernel text or 
> >> > rodata [addr=f0007000] [len=4096]
> >> > Modules linked in: ext3 jbd sd_mod sun_esp esp_scsi scsi_transport_spi 
> >> 
> >> Ok, I looked at check_for_illegal_area() in dma-debug.
> >> 
> >> What check_for_illegal_area() does looks bogus to me with some of I/O
> >> remapping hardware.
> > 
> > Can you be more specific about this one? 
> > check_for_illegal_area() should not depend on any hardware 
> > because all it does is checking the machine addresses to be 
> > mapped.
> 
> The check can't work properly on sparc32.
> 
> Sparc32 always maps the kernel to a fixed physical location, and 
> it therefore can execute in the identity mapping area of physical 
> memory like where all the free pages and kmalloc areas live 
> virtually.
> 
> So if we free up some pages within the kernel image (because the 
> memory is unused, for exmple that's what's happening here with the 
> extra trap table pages on Robert's machine) we have pages in the 
> free page pool that are located right inside of the kernel text, 
> data, etc.
> 
> We'll thus need a way to turn off these checks somehow.  You could 
> also augment this check by seeing if there is a backing page, and 
> if so, whether it is PageReserved or not.  That's just one idea.

Hm, note, this sparc32 behavior might break certain aspects of 
lockdep as well, see kernel/lockdep.c::static_obj(). You could get 
spurious non-printing of 'trying to register non-static key'.

I'm wondering why sparc32 frees from the middle of the kernel image. 
The way architectures generally do it is to put freeable pages into 
a separate section. That way it does not get mingled with the kernel 
core image area (which stays nicely continuous).

	Ingo