[dpdk-dev] [PATCH] eal: forcing IOVA as PA in ppc

[Burakov, Anatoly]

On 01-Aug-19 2:01 PM, David Marchand wrote:
> Hello Anatoly,
> 
> On Wed, Jul 31, 2019 at 5:36 AM Takeshi Yoshimura <tyos at jp.ibm.com> wrote:
>>
>> Commit b76fafb174d2 ("eal: fix IOVA mode selection as VA for PCI
>> drivers") breaks ppc apps with no IOVA configs (or RTE_IOVA_DC)
>> because of the inconsistency of user's request and the result of
>> device capability for IOVA mode. I updated the code to force IOVA as
>> PA in ppc as before because current ppc driver does not support VA
>> mode.
>>
>> Theoretically, ppc can support VA mode, but I suspect that ppc with
>> VA mode may have performance issues to create a big DMA window
>> (VA often uses higher addresses than PA). So, I didn't change the
>> code to check device capability in ppc.
>>
>> Signed-off-by: Takeshi Yoshimura <tyos at jp.ibm.com>
>> ---
>>   lib/librte_eal/linux/eal/eal.c | 6 ++++++
>>   1 file changed, 6 insertions(+)
>>
>> diff --git a/lib/librte_eal/linux/eal/eal.c b/lib/librte_eal/linux/eal/eal.c
>> index 946222ccd..db2dec922 100644
>> --- a/lib/librte_eal/linux/eal/eal.c
>> +++ b/lib/librte_eal/linux/eal/eal.c
>> @@ -1121,6 +1121,12 @@ rte_eal_init(int argc, char **argv)
>>                                  RTE_LOG(DEBUG, EAL, "KNI can not work since physical addresses are unavailable\n");
>>                          }
>>                  }
>> +#endif
>> +#ifdef RTE_ARCH_PPC_64
>> +               if (iova_mode == RTE_IOVA_VA) {
>> +                       iova_mode = RTE_IOVA_PA;
>> +                       RTE_LOG(WARNING, EAL, "Forcing IOVA as 'PA' because PPC uses PA mode.\n");
>> +               }
>>   #endif
>>                  rte_eal_get_configuration()->iova_mode = iova_mode;
>>          } else {
>> --
>> 2.17.1
>>
> 
> Afaiu, this problem is linked to a specific iommu implementation.
> Does it make sense to introduce a check "iommu_support_va" in eal, a
> bit like what we have in the pci bus ?
> 

It probably does. Some platforms simply don't support VA mode (FreeBSD, 
for one), so that would be useful instead of hardcoding IOVA as PA/VA as 
defaults.

-- 
Thanks,
Anatoly