PS3 early lock-up

Geoff Levand geoffrey.levand at am.sony.com
Wed Aug 6 09:24:02 EST 2008


Benjamin Herrenschmidt wrote:
>> arch/powerpc/platfroms/ps3/htab.c:ps3_hpte_updatepp() uses `htab[slot].v'.
> 
> Ok, that leaves us with 2 options:
> 
>  - Change ps3_hpte_updatepp() to not read from the hash table via that
> mapping (ie, do you have an LV1 call to read an HPTE ? Do you measure
> any significant performance loss using that instead ? updatepp shouldn't
> be something called -that- often).

Yes, we have lv1_read_htab_entries().  Mokuno-san started some work to
convert to using it and removing the htab mapping:

  http://git.kernel.org/?p=linux/kernel/git/geoff/ps3-linux-patches.git;a=blob;f=ps3-wip/ps3-htab-rework.diff;hb=HEAD

Unfortunately, this week Mokuno-san is on holiday, and I am busy preparing
for SIGGRAPH (next week).

>  - Add a way to setup HPTEs using 3 PPP bits. I'm not going to implement
> that for the main hash code just yet though (the assembly) but it might
> be possible to implement it specifically for mappings bolted. That
> means it would only work when the mapping is done early but on PS3, we
> know that the hash table is always mapped early.
> 
> The later would be a matter of taking my htab_convert_pte_flags() function
> and making it capable, when _PAGE_USER is _not_ set and _PAGE_RW is not
> set neither, to set PPP to 110.
> 
> You could do that by adding:
> 
> 	if (!(pteflags & (_PAGE_USER | _PAGE_RW)))
> 		rflags |= (1 << 1) | (1 << 63);
> 
> Dbl check that the resulting mapping isn't accessible to user space though.

If we can't remove the htab mapping with lv1_read_htab_entries(), I'll
look into this way.

-Geoff





More information about the Linuxppc-dev mailing list