[v5,21/22] target/arm: Add mmu indexes for tag memory

The process by which one goes from an address space plus physical
address to a host pointer is complex.  It is easiest to reuse the
mechanism already present within cputlb, and letting that cache
the results.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
 target/arm/cpu-param.h |  2 +-
 target/arm/cpu.h       | 12 +++++++++---
 target/arm/internals.h |  2 ++
 target/arm/helper.c    | 25 +++++++++++++++++++++++--
 4 files changed, 35 insertions(+), 6 deletions(-)

On Fri, 11 Oct 2019 at 14:50, Richard Henderson
<richard.henderson@linaro.org> wrote:
>
> The process by which one goes from an address space plus physical
> address to a host pointer is complex.  It is easiest to reuse the
> mechanism already present within cputlb, and letting that cache
> the results.
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
>  target/arm/cpu-param.h |  2 +-
>  target/arm/cpu.h       | 12 +++++++++---
>  target/arm/internals.h |  2 ++
>  target/arm/helper.c    | 25 +++++++++++++++++++++++--
>  4 files changed, 35 insertions(+), 6 deletions(-)
>
> diff --git a/target/arm/cpu-param.h b/target/arm/cpu-param.h
> index 6e6948e960..18ac562346 100644
> --- a/target/arm/cpu-param.h
> +++ b/target/arm/cpu-param.h
> @@ -29,6 +29,6 @@
>  # define TARGET_PAGE_BITS_MIN  10
>  #endif
>
> -#define NB_MMU_MODES 8
> +#define NB_MMU_MODES 9
>
>  #endif
> diff --git a/target/arm/cpu.h b/target/arm/cpu.h
> index faca43ea78..c3609ef9d5 100644
> --- a/target/arm/cpu.h
> +++ b/target/arm/cpu.h
> @@ -2854,8 +2854,8 @@ static inline bool arm_excp_unmasked(CPUState *cs, unsigned int excp_idx,
>  #define ARM_MMU_IDX_M_NEGPRI 0x2
>  #define ARM_MMU_IDX_M_S 0x4
>
> -#define ARM_MMU_IDX_TYPE_MASK (~0x7)
> -#define ARM_MMU_IDX_COREIDX_MASK 0x7
> +#define ARM_MMU_IDX_TYPE_MASK (~0xf)
> +#define ARM_MMU_IDX_COREIDX_MASK 0xf
>
>  typedef enum ARMMMUIdx {
>      ARMMMUIdx_S12NSE0 = 0 | ARM_MMU_IDX_A,
> @@ -2865,6 +2865,9 @@ typedef enum ARMMMUIdx {
>      ARMMMUIdx_S1SE0 = 4 | ARM_MMU_IDX_A,
>      ARMMMUIdx_S1SE1 = 5 | ARM_MMU_IDX_A,
>      ARMMMUIdx_S2NS = 6 | ARM_MMU_IDX_A,
> +    ARMMMUIdx_TagNS = 7 | ARM_MMU_IDX_A,
> +    ARMMMUIdx_TagS = 8 | ARM_MMU_IDX_A,
> +
>      ARMMMUIdx_MUser = 0 | ARM_MMU_IDX_M,
>      ARMMMUIdx_MPriv = 1 | ARM_MMU_IDX_M,
>      ARMMMUIdx_MUserNegPri = 2 | ARM_MMU_IDX_M,
> @@ -2891,6 +2894,8 @@ typedef enum ARMMMUIdxBit {
>      ARMMMUIdxBit_S1SE0 = 1 << 4,
>      ARMMMUIdxBit_S1SE1 = 1 << 5,
>      ARMMMUIdxBit_S2NS = 1 << 6,
> +    ARMMMUIdxBit_TagNS = 1 << 7,
> +    ARMMMUIdxBit_TagS = 1 << 8,
>      ARMMMUIdxBit_MUser = 1 << 0,
>      ARMMMUIdxBit_MPriv = 1 << 1,
>      ARMMMUIdxBit_MUserNegPri = 1 << 2,
> @@ -3254,7 +3259,8 @@ enum {
>  /* Return the address space index to use for a memory access */
>  static inline int arm_asidx_from_attrs(CPUState *cs, MemTxAttrs attrs)
>  {
> -    return attrs.secure ? ARMASIdx_S : ARMASIdx_NS;
> +    return ((attrs.target_tlb_bit2 ? ARMASIdx_TagNS : ARMASIdx_NS)
> +            + attrs.secure);

If you want to do the "just add attrs.secure" can we have
a build-time assert that ARMASIdx_S is ARMASIdx_NS + 1, and
ditto for TagNS/TagS, please? It seems like the kind of thing
that will catch us out later on.

>  }

>      if (env->cp15.hcr_el2 & HCR_TGE) {
> @@ -10662,7 +10671,9 @@ bool get_phys_addr(CPUARMState *env, target_ulong address,
>                     target_ulong *page_size,
>                     ARMMMUFaultInfo *fi, ARMCacheAttrs *cacheattrs)
>  {
> -    if (mmu_idx == ARMMMUIdx_S12NSE0 || mmu_idx == ARMMMUIdx_S12NSE1) {
> +    switch (mmu_idx) {
> +    case ARMMMUIdx_S12NSE0:
> +    case ARMMMUIdx_S12NSE1:
>          /* Call ourselves recursively to do the stage 1 and then stage 2
>           * translations.
>           */
> @@ -10713,6 +10724,16 @@ bool get_phys_addr(CPUARMState *env, target_ulong address,
>               */
>              mmu_idx = stage_1_mmu_idx(mmu_idx);
>          }
> +        break;
> +
> +    case ARMMMUIdx_TagS:
> +    case ARMMMUIdx_TagNS:
> +        /* Indicate tag memory to arm_asidx_from_attrs.  */
> +        attrs->target_tlb_bit2 = true;
> +        break;

So here we fall through to the "handle a stage 1 lookup" code, which:
 * sets attrs->secure
 * sets attrs->user (always false, so we could have left it alone)
 * skips the FCSE handling (as we're v8)
 * skips the PMSA handling
 * hits the regime_translation_disabled() check, which fills in
   *phys_ptr, *prot and *page_size and returns

Maybe it would be clearer if this case here just did all of that:

    case ARMMMUIdx_TagS:
        attrs->secure = true;
        /* fall through */
    case ARMMMUIdx_TagNS:
        /* Indicate tag memory to arm_asidx_from_attrs.  */
        attrs->target_tlb_bit2 = true;
        *phys_ptr = address;
        *prot = PAGE_READ | PAGE_WRITE | PAGE_EXEC;
        *page_size = TARGET_PAGE_SIZE;
        return 0;

Did I miss anything out?
Or are we going to want more things which are common between
the stage 1 codepath and the "just a tag ram access" case
in future?

thanks
-- PMM

On 12/6/19 3:46 AM, Peter Maydell wrote:
>> +    case ARMMMUIdx_TagS:
>> +    case ARMMMUIdx_TagNS:
>> +        /* Indicate tag memory to arm_asidx_from_attrs.  */
>> +        attrs->target_tlb_bit2 = true;
>> +        break;
> 
> So here we fall through to the "handle a stage 1 lookup" code, which:
>  * sets attrs->secure
>  * sets attrs->user (always false, so we could have left it alone)
>  * skips the FCSE handling (as we're v8)
>  * skips the PMSA handling
>  * hits the regime_translation_disabled() check, which fills in
>    *phys_ptr, *prot and *page_size and returns

Exactly.

> Maybe it would be clearer if this case here just did all of that:
> 
>     case ARMMMUIdx_TagS:
>         attrs->secure = true;
>         /* fall through */
>     case ARMMMUIdx_TagNS:
>         /* Indicate tag memory to arm_asidx_from_attrs.  */
>         attrs->target_tlb_bit2 = true;
>         *phys_ptr = address;
>         *prot = PAGE_READ | PAGE_WRITE | PAGE_EXEC;
>         *page_size = TARGET_PAGE_SIZE;
>         return 0;
> 
> Did I miss anything out?

I think that's about it.  I thought about doing exactly this.

Also, this is a better location if I ever do something about the TODO in tne
next patch, wherein I talk about mapping not from physical address but from the
normal ram's ramaddr_t, so as to
cache that mapping step as well.


r~