Re: Make 2.5.17 TLB even more friendlier

[Linus Torvalds]

On Wed, 22 May 2002, Paul Mackerras wrote:
> I'm thinking now that I would be better off doing the batching by
> storing a list of virtual addresses needing flushing in the
> mmu_gather_t, instead of relying on getting at the special bits in the
> PTEs at some time after the tlb_remove_tlb_entry call.

Well, you can combine that with something like

	static inline void tlb_remove_tlb_entry(tlb, pte, address)
	{
		if (pte_tlb_hash(pte)) {
			if (tlb->start_addr == NOSTART)
				tlb->start_addr = address;
			tlb->end_addr = address+PAGE_SIZE;
		}
	}

and then have the tlb_end_vma() do something like

	/* No pages mapped? */
	if (tlb->start_addr == NOADDR)
		return;
	pte_remove_range(vma, tlb->start_addr, tlb->end_addr)
	tlb->start_addr = NOADDR;

which will bunch them up on a vma granularity (if there is any reason to 
do that), while still retaining the optimization that if a VMA was mostly 
unmapped you wouldn't need to do a lot of hash table searching because of 
the start/end thing.

		Linus