Re: [RFC/PATCH] Use mmu_gather for fork() instead of flush_tlb_mm()

[Nick Piggin <nickpiggin@yahoo.com.au>]

Benjamin Herrenschmidt wrote:
> Use mmu_gather for fork() instead of flush_tlb_mm()
> 
> This patch uses an mmu_gather for copying page tables instead of
> flush_tlb_mm(). This allows archs like ppc32 with hash table to
> avoid walking the page tables a second time to invalidate hash
> entries, and to only flush PTEs that have actually been changed
> from RW to RO.
> 
> Note that this contain a small change to the mmu gather stuff,
> it must not call free_pages_and_swap_cache() if no page have been
> queued up for freeing (if we are only invalidating PTEs). Calling
> it on fork can deadlock (I haven't dug why but it looks like a
> good idea to test anyway if we're going to use the mmu_gather for
> more than just removing pages).
> 
> If the patch gets accepted, I will split that bit from the rest
> of the patch and send it separately.
> 
> The main possible issue I see is with huge pages. Arch code might
> have relied on flush_tlb_mm() and might not cope with
> tlb_remove_tlb_entry() called for huge PTEs.
> 
> Other possible issues are if archs make assumptions about
> flush_tlb_mm() being called in fork for different unrelated reasons.
> 
> Ah also, we could probably improve the tracking of start/end, in
> the case of lock breaking, the outside function will still finish
> the batch with the entire range. It doesn't matter on ppc and x86
> I think though.

Would it be better off to start off with a new API for this? The
mmu gather I think is traditionally entirely for dealing with
page removal...

-- 
SUSE Labs, Novell Inc.