On 08/23/2016 03:49 PM, Yuanhan Liu wrote:
> On Tue, Aug 23, 2016 at 03:25:33PM +0200, Maxime Coquelin wrote:
>>
>>
>> On 08/23/2016 02:32 PM, Yuanhan Liu wrote:
>>>>> +
>>>>>>> + /* FIXME */
>>>>>>> + RTE_LOG(INFO, VHOST_CONFIG, ":: %u ::\n", pre_read);
>>>>> For my information, what is the purpose of pre_read?
>>> Again, I put a FIXME here, but I forgot to add some explanation.
>>>
>>> Here is the thing: the read will make sure the kernel populate the
>>> corresponding PTE entry, so that rte_mem_virt2phy() will return proper
>>> physical address, otherwise, invalid value is returned.
>>>
>>> I can't simply do the read but do not actually reference/consume it.
>>> Otherwise, the compiler will treat it as some noops and remove it.
>>>
>>> An ugly RTE_LOG will make sure the read operation is not eliminated.
>>> I'm seeking a more proper way to achieve that. Maybe I can add a new
>>> field in virtio_net structure and store it there.
>>>
>>> Or, do you have better ideas?
>>
>> This behavior is pretty twisted, no?
>
> I have to say, yes, kind of.
>
>> Shouldn't be rte_mem_virt2phy() role to ensure returning a valid value?
>
> Not exactly. I think rte_mem_virt2phy() is more likely to fetch the
> physical address of huge pages. And for those huge pages, EAL makes
> sure they will be populated: it used to do a zero memset before to
> achieve that. Since 5ce3ace1de45 ("eal: remove unnecessary hugepage
> zero-filling"), it uses MAP_POPULATE option instead.
>
> So, thank you that you just remind me of the MAP_POPULATE option.
> I just had a quick try, it worked like a charm :)
Excellent! :)
