[RFC/PATCH 0/12] WIP mmu_gather and PTE accessors work

From: Benjamin Herrenschmidt
Date: Tue Aug 07 2007 - 03:20:09 EST


This is a snapshot of my current work on PTE accessors and
mmu_gather. It's not complete but it should show the direction
I'm heading toward.

The main goals are:

- Make mmu_gather used for all page table walk operations that
also need to invalidate TLB entries, thus obsoleting flush_tlb_range()
and flush_tlb_mm() as generic APIs to the TLB flushing.

- Make mmu_gather stack based. The cleans quite a bit of stuff up,
and once fully done, should allow to reduce latencies caused by the
need to use get_cpu() for a long time.

- Make mmu_gather more flexible so that archs who need to do more than
what the standard implementation does don't have to copy all of it and
do their own implementation.

- Make mmu_gather suitable for batching on powerpc :-) This involves
mostly adding a hook before PTE pages are unlocked and some work on
the interaction between PTE accessors and tlb_remove_tlb_entry()

- Remove other remaings of flush_tlb_*, keeping only for now
flush_tlb_kernel_range() which will be harder to "fix" (provided we
want to do it at all)

- Go through all remaining page table accessors, remove all the unused
ones (there's still a few), there should be only a handful left and redo
the documentation accordingly.

These goals are _NOT_ yet met by this patch serie and some of the bits
in there may want to be done a bit differently. As I did the patches,
and hacked various archs, I got a better visibility on what is done and
why and thus some of my initial directions end up not looking so good.

Most notably, the MMF_DEAD flag doesn't sound like such a good idea
anymore and I'm considering instead replacing ptep_get_and_clear()
and tlb_remove_tlb_entry() with a version that takes the batch as an
argument.

Also, I haven't fully moved the batch off the per-cpu, only "part 1" is
there at this stage.

However, I'll be travelling for a while and won't have much time to work
on it until I'm back mid september, so I decided now was a good time to
post what I have for comments and discussions on the approach taken.

Cheers,
Ben.
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/