Re: Make 2.5.17 TLB even more friendlier

[Paul Mackerras]

David S. Miller writes:

> You get called via pte_free_tlb() and pmd_free_tlb() for every
> operation performed by clear_page_tables().  The PTEs themselves are
> all cleared out at the point that clear_page_tables, so you can't
> possibly need the PTE contents.  I am assuming therefore you just
> need to get at the linkage, and those two pte/pmd hooks give you
> that.

I have a bit in the PTE that tells me if there is an MMU hash table
entry for the virtual address represented by the PTE.  This bit is not
affected by set_pte or ptep_get_and_clear etc. and it is not part of
the swap-entry fields of the PTE.  Thus I need to have the PTE page
still around at the point where I flush stuff from the MMU hash
table, even though all the PTEs in it have been cleared, so that I can
avoid searching the hash table for PTEs for virtual addresses for
which I have not put a PTE in the hash table.

Now I could of course try to come up with some other way to store this
information, such as a bitmap associated with the mmu context.  The
bitmap would be 96kB in size for a 3GB userspace, though.  It could be
allocated on a page-by-page basis, I guess.  But all the schemes I
have thought of so far end up being more complex than using a special
bit in the PTE.

Regards,
Paul.