Benjamin Herrenschmidt [off-list ref] wrote on 03/10/2009 12:57:28:

On Sat, 2009-10-03 at 11:24 +0200, Joakim Tjernlund wrote:

quoted

So yes, there is a missing _tlbil_va() missing for 8xx somewhere
but there is something more too.
Maybe your new filter functions and my
 powerpc, 8xx: DTLB Error must check for more errors.
will do the trick?

Well, if we can't tell between a load and a store on a TLB miss, then
we should probably let it create an unpopulated entry in all cases,
so that we do take a proper DSI/ISI the second time around, which
would then tell us where we come from...

Not quite sure what you mean here?
Always branch to DSI at TLB Miss and create a unpopulated
entry? That does not feel right.

The current method is better. The only odd thing is
the null pmd entry and what to load into RPN. I am not convinced
that it is a problem but I would like to see a generic impl. of
a null pmd and use that instead. A 8xx impl. of that
looks like this(tested on 2.4, works fine)

	lwz	r11, 0(r10)	/* Get the level 1 entry */
	rlwinm.	r10, r11,0,0,19	/* Extract page descriptor page address */
	bne+	4f		/* If zero, use a null pmd */
	lis	r11, null_pmd_dir@h
	ori	r11, r11, null_pmd_dir@l
4:	tophys(r11,r11)

...
	.data
	.globl	null_pmd_dir
null_pmd_dir:
	.space	4096

If the pmd_none() and friends are updated to test for a null_pmd_dir
we can loose the test above and save 4 insn :)

Anyhow did you post a patch(can't find one) about your suggested
filter functions for 8xx? I sure would like to see one :)