I'm not Doshita-san, but it was me who translated with confusion, so...

> > Pre-cache load length exceeds source memory area in memcpy.
> > Unrelated area is written back to memory, which causes
> > some troubles.
>
> Can you elaborate on the "some troubles" part? The source cacheline will
> not be marked dirty by pld, assuming it was already clean or not
> resident in the cache to begin with, so a write-back will not occur.

"Write-back" was wrong.  This is all about read (cache fill).

> The only case I can think of where this *may* be an issue would be
> copying from a cacheable memory region which is shared with a
> DMA-capable device...

The place he had problem was a memcpy() in m_dup(9).  The memcpy() does
its job correctly, but something slow happens there and the system loses
lots of packets.  We guess it's kind of exception, but have not figured out yet.

Masao