[U-Boot] [PATCH v3] armv8: caches: Added routine to set non cacheable region

[Mark Rutland]

On Thu, Jun 11, 2015 at 08:17:15AM +0100, Siva Durga Prasad Paladugu wrote:
> 
> Hi Mark,
> 
> > -----Original Message-----
> > From: Mark Rutland [mailto:mark.rutland at arm.com]
> > Sent: Thursday, May 28, 2015 3:10 PM
> > To: Siva Durga Prasad Paladugu
> > Cc: u-boot at lists.denx.de; Michal Simek; Siva Durga Prasad Paladugu
> > Subject: Re: [PATCH v3] armv8: caches: Added routine to set non cacheable
> > region
> >
> > Hi,
> >
> > > +void mmu_set_region_dcache_behaviour(phys_addr_t start, size_t size,
> > > +                                enum dcache_option option)
> > > +{
> > > +   u64 *page_table = arch_get_page_table();
> > > +   u64 upto, end;
> > > +
> > > +   if (page_table == NULL)
> > > +           return;
> > > +
> > > +   end = ALIGN(start + size, (1 << MMU_SECTION_SHIFT)) >>
> > > +         MMU_SECTION_SHIFT;
> > > +   start = start >> MMU_SECTION_SHIFT;
> > > +   for (upto = start; upto < end; upto++) {
> > > +           page_table[upto] &= ~PMD_ATTRINDX_MASK;
> > > +           page_table[upto] |= PMD_ATTRINDX(option);
> > > +   }
> >
> > These writes might not be visible to the page table walkers immediately, and
> > the TLBs might still contain stale values for a while afterwards.
> > That could render the cache maintenance useless (as speculative fetches
> > could still occur due to cacheable attributes still being in place).
> >
> > You need a DSB to ensure writes are visible to the page table walkers (with a
> > compiler barrier to ensure that the writes actually occur before the DSB), and
> > some TLB maintenance (complete with another DSB) to ensure that the TLBs
> > don't contain stale values by the time to get to the cache afterwards.
> The flush_dcache _range() below contains a dsb. Isn't it fine enough?
> Or we need a separte dsb in the for loopafter we changed the cache
> attribute.

The DSB in flush_dcache_range() is not sufficient. You need a DSB
between the page table modifications and the TLB invalidation, and the
TLB invalidation must be completed before the cache maintenance begins.

> Regarding the TLB maintenance if we have _asm_invalidate_tlb_all()
> after the flush dcache range below it should be fine right?

No. The TLB maintenance must be complete _before_ the cache maintenance,
or the cache can be refilled while the maintenance is ongoing (e.g. the
CPU could make speculative prefetches).

You need a strictly-ordered sequence:

1) Modify the page tables

2) DSB

   This ensures the updates are visible to the page table walker(s).

3) TLB invalidation

4) DSB

   This ensures that the TLB invalidation is complete (i.e. from this
   point on the TLBs cannot hold entries for the region with cacheable
   attributes).

5) ISB

   This ensures that the effects of TLB invalidation are visible to
   later instructions. Otherwise instructions later could be using stale
   attributes fetched earlier by the CPU from the TLB, before the TLB
   invalidation completed (and hence could allocate in the caches).

6) Cache maintenance 

7) DSB to complete cache maintenance

Thanks,
Mark.