[PATCH] arm64: Add flush_cache_vmap call in __early_set_fixmap

[Will Deacon]

On Mon, Jun 09, 2014 at 02:24:29PM +0100, Leif Lindholm wrote:
> On Mon, Jun 09, 2014 at 12:03:56PM +0100, Catalin Marinas wrote:
> > So I'm proposing an alternative patch (which needs some benchmarking as
> > well to see if anything is affected, maybe application startup time).

I don't like the proposed patch at all -- keeping the dsb out of set_pte is
worthwhile if we can manage it. That said, it would be interesting to know
how often we get a subsequent page fault after mapping invalid -> valid
because of the missing dsb. It could be that the cost of the benign fault is
hitting us more than we think.

> I'm happy for any fix which can be included in 3.16.
> 
> But is the dsb(ishst) sufficient? We need to also prevent reads from
> overtaking the set_pte(). i.e.:
> 
> ptr = early_ioremap(phys_addr, size);
> if (ptr && strcmp(ptr, "magic") == 0)
>    ...
> 
> Does it not require a dsb(ish)?

I don't think so. Crudely, the sequence above would look like:

  STR	x0, [PTEP]
  DSB	ISHST
  LDR	x0, [MAGIC]

The DSB can't complete until the STR is globally observed within the
inner-shareable domain, so the LDR cannot execute until the page table
update is visible to the walker.

If it was a DMB, we'd have a problem. Interestingly, the asm-generic
page table allocators (e.g. __pmd_alloc) *do* use dmb for ordering
observability of page-table updates via smp_wmb. I'm struggling to decide
whether that's broken or not.

Will