Re: [PATCH v4 1/3] target/arm: Add support for FEAT_TLBIRANGE

[Richard Henderson]

On 3/16/21 9:49 AM, Rebecca Cran wrote:
> +    for (page = addr; page < (addr + length); page += TARGET_PAGE_SIZE) {

This test means that it's impossible to flush the last page of the address 
space (addr + length == 0).  I think better to do

  for (l = 0; l < length; l += TARGET_PAGE_SIZE)
      page = addr + l;
      ...

> +        for (mmu_idx = 0; mmu_idx < NB_MMU_MODES; mmu_idx++) {
> +            if ((idxmap >> mmu_idx) & 1) {
> +                tlb_flush_page_bits_locked(env, mmu_idx, page, bits);

Hmm.  I'm not keen on this.  You're not able to notice the special cases 
within, where we flush the entire tlb -- and therefore you do not need to 
continue the outer loop for this mmuidx.

> +                tb_flush_jmp_cache(cpu, page);

This does not need to be in the mmuidx loop.  But since above means that the 
mmuidx loop should be the outer loop, this would go in a separate page loop by 
itself.

> +void tlb_flush_page_range_bits_by_mmuidx(CPUState *cpu,
> +                                         target_ulong addr,
> +                                         target_ulong length,
> +                                         uint16_t idxmap,
> +                                         unsigned bits)
> +{
> +    TLBFlushPageRangeBitsByMMUIdxData d;
> +    TLBFlushPageRangeBitsByMMUIdxData *p;
> +
> +    /* This should already be page aligned */
> +    addr &= TARGET_PAGE_BITS;
> +
> +    /* If all bits are significant, this devolves to tlb_flush_page. */
> +    if (bits >= TARGET_LONG_BITS) {
> +        tlb_flush_page_by_mmuidx(cpu, addr, idxmap);
> +        return;
> +    }

This case is incorrect.

The cputlb changes should have remained a separate patch.

> @@ -4759,6 +4759,241 @@ static void tlbi_aa64_vae3is_write(CPUARMState *env, 
> const ARMCPRegInfo *ri,
>                                                     ARMMMUIdxBit_SE3, bits);
>   }
>   
> +#ifdef TARGET_AARCH64
> +static uint64_t tlbi_aa64_range_get_length(CPUARMState *env,
> +                                           uint64_t value)
> +{
> +    unsigned int page_size;
> +    unsigned int page_size_granule;
> +    uint64_t num;
> +    uint64_t scale;
> +    uint64_t exponent;
> +    uint64_t length;
> +
> +    num = extract64(value, 39, 4);
> +    scale = extract64(value, 44, 2);
> +    page_size_granule = extract64(value, 46, 2);
> +
> +    switch (page_size_granule) {
> +    case 1:
> +      page_size = 4096;

Indentation is off?

> +      break;
> +    case 2:
> +      page_size = 16384;
> +      break;
> +    case 3:
> +      page_size = 65536;

You might as well have this as page_shift = {12,14,16}, or perhaps page_shift = 
page_size_granule * 2 + 10 instead of the full switch.

> +    exponent = (5 * scale) + 1;
> +    length = ((num + 1) << exponent) * page_size;

  length = (num + 1) << (exponent + page_shift);

> +    mask = vae1_tlbmask(env);
> +    if (regime_has_2_ranges(mask)) {

You can't pass in mask.

All of the mmuidx will have the same form, so ctz32(mask) would pick out the 
mmuidx for the first bit.

> +    if (regime_has_2_ranges(secure ? ARMMMUIdxBit_SE2 : ARMMMUIdxBit_E2)) {

again.  Only this time we know that E2 & SE2 have one range.  Only (S)EL1&0 and 
(S)EL2&0 have two ranges.

> +    if (regime_has_2_ranges(ARMMMUIdxBit_SE3)) {

Likewise, E3 has only one range.


r~