[RFC 06/18] arm: msm: implement proper dmb() for 7x27

[Russell King - ARM Linux]

On Mon, Jan 11, 2010 at 03:45:16PM -0800, Daniel Walker wrote:
> On Mon, 2010-01-11 at 23:39 +0000, Russell King - ARM Linux wrote:
> > On Mon, Jan 11, 2010 at 02:47:25PM -0800, Daniel Walker wrote:
> > > From: Larry Bassel <lbassel at quicinc.com>
> > > 
> > > For 7x27 it is necessary to write to strongly
> > > ordered memory after executing the coprocessor 15
> > > instruction dmb instruction.
> > > 
> > > This is only for data barrier dmb().
> > > Note that the test for 7x27 is done on all MSM platforms
> > > (even ones such as 7201a whose kernel is distinct from
> > > that of 7x25/7x27).
> > > 
> > > Acked-by: Willie Ruan <wruan at quicinc.com>
> > > Signed-off-by: Larry Bassel <lbassel at quicinc.com>
> > > Signed-off-by: Daniel Walker <dwalker at codeaurora.org>
> > 
> > Can only see half of this change - what's the actual implementation of
> > arch_barrier_extra()?
> > 
> > I'd prefer not to include asm/memory.h into asm/system.h to avoid
> > needlessly polluting headers.
> 
> I don't have a real patch for it yet, but here are the pieces ..
> 
> +#define arch_barrier_extra() do \
> +       { if (machine_is_msm7x27_surf() || machine_is_msm7x27_ffa())  \
> +               write_to_strongly_ordered_memory(); \
> +       } while (0)
> 
> (btw, the machine types above registered either..)

Hmm.  We can do far better than this.  Rather than do two tests and call
a function, wouldn't it be better to do something like:

#ifdef CONFIG_ARM_DMB_MEM
extern int *dmb_mem;
#define dmb_extra() do { if (dmb_mem) *dmb_mem = 0; } while (0)
#else
#define dmb_extra() do { } while (0)
#endif

in asm/system.h, and only set dmb_mem for the affected platforms?

> static void map_zero_page_strongly_ordered(void)
> {
> 	if (zero_page_strongly_ordered)
> 		return;
> 
> 	zero_page_strongly_ordered =
> 		ioremap_strongly_ordered(page_to_pfn(empty_zero_page)
> 		<< PAGE_SHIFT, PAGE_SIZE);

This can't work.  You're not allowed to map the same memory with differing
memory types from ARMv7.  This ends up mapping 'empty_zero_page' as both
cacheable memory and strongly ordered.  That's illegal according to the
ARM ARM.

You need to find something else to map - allocating a page of system
memory for this won't work either (it'll have the same issue.)

(This is a new problem to the ARM architecture, one which we're only just
getting to grips with - many of our old tricks with remapping DMA memory
no longer work on these latest CPUs.  You really must not take the
remapping which the kernel does today as a good idea anymore.)

> void flush_axi_bus_buffer(void)
> {
> 	__asm__ __volatile__ ("mcr p15, 0, %0, c7, c10, 5" \
> 				    : : "r" (0) : "memory");
> 	write_to_strongly_ordered_memory();

Isn't this just one of your modified dmb()s ?

> }
> 
> void *alloc_bootmem_aligned(unsigned long size, unsigned long alignment)
> {
> 	void *unused_addr = NULL;
> 	unsigned long addr, tmp_size, unused_size;
> 
> 	/* Allocate maximum size needed, see where it ends up.
> 	 * Then free it -- in this path there are no other allocators
> 	 * so we can depend on getting the same address back
> 	 * when we allocate a smaller piece that is aligned
> 	 * at the end (if necessary) and the piece we really want,
> 	 * then free the unused first piece.
> 	 */
> 
> 	tmp_size = size + alignment - PAGE_SIZE;
> 	addr = (unsigned long)alloc_bootmem(tmp_size);
> 	free_bootmem(__pa(addr), tmp_size);
> 
> 	unused_size = alignment - (addr % alignment);
> 	if (unused_size)
> 		unused_addr = alloc_bootmem(unused_size);
> 
> 	addr = (unsigned long)alloc_bootmem(size);
> 	if (unused_size)
> 		free_bootmem(__pa(unused_addr), unused_size);
> 
> 	return (void *)addr;

Erm, there is __alloc_bootmem(size, align, 0) - the bootmem allocator
already does alignment.