[PATCH v5 02/19] ARM: LPAE: add ISBs around MMU enabling code

[Russell King - ARM Linux]

On Mon, May 09, 2011 at 11:59:54AM +0100, Catalin Marinas wrote:
> On Mon, 2011-05-09 at 11:32 +0100, Russell King - ARM Linux wrote:
> > On Mon, May 09, 2011 at 11:22:19AM +0100, Catalin Marinas wrote:
> > > Alternatively an exception return would do as well (like movs pc, lr)
> > > but I think we still add some code for setting up the SPSR.
> > 
> > That gives us a way out of both of these without introducing any CPU
> > specific code.  We can setup the SPSR before this block of code, and
> > call it with two movs pc, reg instructions which will provide the
> > necessary synchronization.
> 
> We still need an ISB before enabling the MMU to make sure that the TTBR
> changing is visible. We may run with the MMU enabled (in the identity
> mapping) before the exception return but with random data in TTBR.

  Changes to CP15 registers and the memory order model
  All changes to CP15 registers that appear in program order after any
  explicit memory operations are guaranteed not to affect those memory
  operations.

  Any change to CP15 registers is guaranteed to be visible to subsequent
  instructions only after one of:
  • the execution of an ISB instruction
  • the taking of an exception
  • the return from an exception.

  To guarantee the visibility of changes to some CP15 registers, additional
  operations might be required, on a case by case basis, before the ISB
  instruction, exception or return from exception. These cases are
  identified specifically in the definition of the registers.

  However, for CP15 register accesses, all MRC and MCR instructions to
  the same register using the same register number appear to occur in
  program order relative to each other without context synchronization.

So, my reading of this suggests that ISB and returning from an exception
(iow, movs pc, reg) have the same properties.  So:

        mcr     p15, 0, r5, c3, c0, 0           @ load domain access register
        mcr     p15, 0, r4, c2, c0, 0           @ load page table pointer
-       b       __turn_mmu_on
+	mrs	r4, cpsr			@ copy cpsr to spsr
+	msr	spsr, r4
+	adr	r4, BSYM(__turn_mmu_on)
+	movs	pc, r4				@ synchronizing

        .align  5
__turn_mmu_on:
        mov     r0, r0
        mcr     p15, 0, r0, c1, c0, 0           @ write control reg
        mrc     p15, 0, r3, c0, c0, 0           @ read id reg
        mov     r3, r3
        mov     r3, r13
-       mov     pc, r3
+       movs    pc, r3				@ synchronizing

should be sufficient - and has the advantage that it should work on
existing CPUs.