Ralf Baechle 写道:
On Tue, May 18, 2010 at 11:23:39AM +0800, Yang Shi wrote:
It seems CPU_PREFETCH caused this issue. See commit:
Is this a commit ID of a Wind River tree? The commit ID for the kernel.org
kernel tree is 634286f127bef8799cd04799d3e1d5471e8fd91c, for the lmo tree
Yes, that commit is cherry-picked from mainline commit
634286f127bef8799cd04799d3e1d5471e8fd91c. If I revert this commit,
module loading works well.
Author: Ralf Baechle <firstname.lastname@example.org>
Date: Wed Jan 28 17:48:40 2009 +0000
MIPS: IP27: Switch from DMA_IP27 to DMA_COHERENT
commit 0d356eaa6316cbb3e89b4607de20b2f2d0ceda25 from linux-mips
The special IP27 DMA code selected by DMA_IP27 has been removed a while
ago turning DMA_IP27 into almost a nop. Also fixup the broken logic of
its last users memcpy.S and memcpy-inatomic.s.
Signed-off-by: Ralf Baechle <email@example.com>
If undef CPU_PREFETCH for SB1250, module can be loaded correctly.
This patch did below fix:
-#if !defined(CONFIG_DMA_COHERENT) || !defined(CONFIG_DMA_IP27)
Before the fix, CONFIG_DMA_IP27 is undefined for all of boards except
IP27, so CONFIG_CPU_HAS_PREFETCH is undefined always.
That's an old issue striking here then. Memcpy will prefetch beyond the
end of the source or destination area, if the use of prefetch is enabled.
On non-coherent systems this is a problem because these prefetches might
bring back data that was just flushed from the cache back into the cache
resulting in corrupted DMA transfers.
Some systems that do very tight checking of address but do not differenciate
between actual loads / stores and prefetch accesses may also also throw
address error exceptions when prefetching from non-decoded address ranges.
That would typically be just beyond the end of a RAM range. For those
systems the solution is to either disable prefetching or to not use the
last page of every physically contiguous area of memory.
I'm still looking at why we this wasn't triggered earlier. Something else
must have changed to suddenly trigger this issue - and your system seems
to be the only system affected.
As I said above, before 634286f127bef8799cd04799d3e1d5471e8fd91c
checking in, actually prefetch should be *undefined* for all of boards
except IP27, so maybe no one noticed this issue.
Seems yes, just Sibyte is affected :(