Re: [PATCH v5 16/22] target/arm: Implement data cache set allocation tags

[Richard Henderson]

On 12/5/19 10:17 AM, Peter Maydell wrote:
> On Fri, 11 Oct 2019 at 14:50, Richard Henderson
> <address@hidden> wrote:
>>
>> This is DC GVA and DC GZVA.
>>
>> Signed-off-by: Richard Henderson <address@hidden>
>> ---
>> v2: Use allocation_tag_mem + memset.
>> v3: Require pre-cleaned addresses.
>> ---
> 
>> diff --git a/target/arm/mte_helper.c b/target/arm/mte_helper.c
>> index f1315bae37..e8d8a6bedb 100644
>> --- a/target/arm/mte_helper.c
>> +++ b/target/arm/mte_helper.c
>> @@ -510,3 +510,31 @@ void HELPER(stzgm)(CPUARMState *env, uint64_t ptr, 
>> uint64_t val)
>>          }
>>      }
>>  }
>> +
>> +void HELPER(dc_gva)(CPUARMState *env, uint64_t ptr)
>> +{
>> +    ARMCPU *cpu = env_archcpu(env);
>> +    size_t blocklen = 4 << cpu->dcz_blocksize;
>> +    int el;
>> +    uint64_t sctlr;
>> +    uint8_t *mem;
>> +    int rtag;
>> +
>> +    ptr = QEMU_ALIGN_DOWN(ptr, blocklen);
>> +
>> +    /* Trap if accessing an invalid page.  */
>> +    mem = allocation_tag_mem(env, ptr, true, GETPC());
>> +
>> +    /* No action if page does not support tags, or if access is disabled.  
>> */
>> +    el = arm_current_el(env);
>> +    sctlr = arm_sctlr(env, el);
>> +    if (!mem || !allocation_tag_access_enabled(env, el, sctlr)) {
>> +        return;
>> +    }
>> +
>> +    rtag = allocation_tag_from_addr(ptr);
>> +    rtag |= rtag << 4;
>> +
>> +    assert(QEMU_IS_ALIGNED(blocklen, 2 * TAG_GRANULE));
> 
> Could we assert this on CPU init rather than in this helper?
> That way if anybody tries to create a CPU whose dcz_blocksize
> doesn't work with the TAG_GRANULE they'll realize immediately
> rather than only if they happen to run a guest workload that
> use DC GVA or DC GZVA.

Sure.  I've moved it to realize.

> I also had to think a bit to work out which way round this
> assert is checking: it's testing that the ZVA block length
> (usually 64 bytes) is a multiple of (twice the TAG_GRANULE),
> which is to say a multiple of 32. Given that the blocksize
> is stored as a log2 value, this can only fail for blocksizes
> 16, 8, 4, 2 or 1, which are all fairly unlikely.

Indeed.

>> +    case ARM_CP_DC_GVA:
>> +        tcg_rt = clean_data_tbi(s, cpu_reg(s, rt), false);
>> +        gen_helper_dc_gva(cpu_env, tcg_rt);
>> +        return;
>> +    case ARM_CP_DC_GZVA:
>> +        tcg_rt = clean_data_tbi(s, cpu_reg(s, rt), false);
>> +        gen_helper_dc_zva(cpu_env, tcg_rt);
>> +        gen_helper_dc_gva(cpu_env, tcg_rt);
> 
> I think this means that if there's a watchpoint set on the memory
> partway through the block we're zeroing then we'll take the
> watchpoint with some of the memory zeroed but without the
> corresponding tags having been updated. But that's probably OK
> (at any rate I wouldn't worry about it for now...)

The relatively new probe_{access,write} functions take care of watchpoints.  On
the next round I'll have another look at what may or may not be missing to use
that here.  It's something that I've been thinking about to get right for SVE
as well.


r~