Le 24/06/2019 à 05:12, Nicholas Piggin a écrit :
Christophe Leroy's on June 20, 2019 2:25 am:
Le 19/06/2019 à 05:59, Nicholas Piggin a écrit :
Christophe Leroy's on June 11, 2019 4:46 pm:
Le 10/06/2019 à 05:08, Nicholas Piggin a écrit :
I would like to remove the early ioremap or make it into its own
function. Re-implement map_kernel_page with ioremap_page_range,
allow page tables that don't use slab to avoid the early check,
unbolt the hptes mapped in early boot, etc.
Getting early ioremap out of the picture is a very good idea, it will
help making things more common between all platform types. Today we face
the fact that PPC32 allocates early io from the top of memory while
PPC64 allocates it from the bottom of memory.
Any idea on how to proceed ?
I have to have a bit better look at other arches and our platform
code. Without having looked closely at all the details, I would hope
we could use GENERIC_EARLY_IOREMAP without too much pain.
Good idea.
I have looked at it and implemented it for PPC32. In its own it works
well, now the challenge is to move all early call sites of ioremap() to
early_ioremap().
I point however is that early_ioremap() expects all maps being released
by the time we do paging_init(), whereas several of early PPC ioremap()
users never release the mapped area. I think we have to dig into this in
more details.
Christophe
