[PATCH 20/19] ARM: LPAE: Invalidate the TLB before freeing the PMD

On Wed, 2011-05-11 at 11:54 +0100, Russell King - ARM Linux wrote:

On Wed, May 11, 2011 at 11:23:19AM +0100, Catalin Marinas wrote:

quoted

+static inline void __pmd_free_tlb(struct mmu_gather *tlb, pmd_t *pmdp,
+                               unsigned long addr)
+{
+#ifdef CONFIG_ARM_LPAE
+     tlb_add_flush(tlb, addr);
+     tlb_flush(tlb);
+     pmd_free((tlb)->mm, pmdp);
+#endif
+}

You're:

1. tlb_add_flush() - Adding the address which covers the PMD to the range
   of virtual addresses which need flushing - ok.
2. tlb_flush() - You're then forcing a flush.
3. pmd_free() - You're now freeing the page.

One of the points about the shootdown interface is that it batches things
up.  So what's wrong with:

static inline void __pmd_free_tlb(struct mmu_gather *tlb, pmd_t *pmdp,
        unsigned long addr)
{
#ifdef CONFIG_ARM_LPAE
        tlb_add_flush(tlb, addr);
        tlb_remove_page(tlb, virt_to_page(pmdp));
#endif
}

and leave the tlb invalidate and actual page freeing to the batching code
to deal with?

There isn't a big overhead with my initial code as a pmd covers 1GB and
we only have 1 or 2 pmds per process that we can free.

Is there any room for optimising the mmu_gather range? I think this only
matters for case 1 in your tlb_flush() comment - unmapping a page range
with a few pages in one pmd and a few other pages in the next pmd we get
over 1GB range when we actually only need to flush the TLB for a few
pages.

If tlb_add_flush would get a start/end range (or addr/size), we know
that any TLB flush within the start..end range would be enough and thus
we avoid artificially increasing the range.

We could also modify flush_tlb_range() to branch to flush_tlb_mm() for
big ranges.

-- 
Catalin