Re: [RFC PATCH v2] arm64: cpufeatures: add support for tlbi range instructions

[Zhenyu Ye]

On 2019/11/11 22:04, Marc Zyngier wrote:
> On 2019-11-11 14:56, Zhenyu Ye wrote:
>> On 2019/11/11 21:27, Will Deacon wrote:
>>> On Mon, Nov 11, 2019 at 09:23:55PM +0800, Zhenyu Ye wrote:
>>>> ARMv8.4-TLBI provides TLBI invalidation instruction that apply to a
>>>> range of input addresses. This patch adds support for this feature.
>>>> This is the second version of the patch.
>>>>
>>>> I traced the __flush_tlb_range() for a minute and get some statistical
>>>> data as below:
>>>>
>>>>     PAGENUM        COUNT
>>>>     1        34944
>>>>     2        5683
>>>>     3        1343
>>>>     4        7857
>>>>     5        838
>>>>     9        339
>>>>     16        933
>>>>     19        427
>>>>     20        5821
>>>>     23        279
>>>>     41        338
>>>>     141        279
>>>>     512        428
>>>>     1668        120
>>>>     2038        100
>>>>
>>>> Those data are based on kernel-5.4.0, where PAGENUM = end - start, COUNT
>>>> shows number of calls to the __flush_tlb_range() in a minute. There only
>>>> shows the data which COUNT >= 100. The kernel is started normally, and
>>>> transparent hugepage is opened. As we can see, though most user TLBI
>>>> ranges were 1 pages long, the num of long-range can not be ignored.
>>>>
>>>> The new feature of TLB range can improve lots of performance compared to
>>>> the current implementation. As an example, flush 512 ranges needs only 1
>>>> instruction as opposed to 512 instructions using current implementation.
>>>>
>>>> And for a new hardware feature, support is better than not.
>>>>
>>>> Signed-off-by: Zhenyu Ye <yezhenyu2@xxxxxxxxxx>
>>>> ---
>>>> ChangeLog v1 -> v2:
>>>> - Change the main implementation of this feature.
>>>> - Add some comments.
>>>
>>> How does this address my concerns here:
>>>
>>>
>>> https://lore.kernel.org/linux-arm-kernel/20191031131649.GB27196@willie-the-truck/
>>>
>>> ?
>>>
>>> Will
>>>
>>> .
>>>
>>
>> I think your concern is more about the hardware level, and we can do
>> nothing about
>> this at all. The interconnect/DVM implementation is not exposed to
>> software layer
>> (and no need), and may should be constrained at hardware level.
> 
> You're missing the point here: the instruction may be implemented
> and perfectly working at the CPU level, and yet not carried over
> the interconnect. In this situation, other CPUs may not observe
> the DVM messages instructing them of such invalidation, and you'll end
> up with memory corruption.
> 
> So, in the absence of an architectural guarantee that range invalidation
> is supported and observed by all the DVM agents in the system, there must
> be a firmware description for it on which the kernel can rely.
> 
>         M.

OK, sounds like an architectural defect. Some hardware-level improvements
may should be done here, which is unfamiliar to me.

However, We can discuss other aspects of the patch. Anyway, this should be
support by the kernel in the future, isn't it?