Re: [dpdk-dev] [PATCH] pci/linux: use RTE_IOVA_VA whenever possible

["Burakov, Anatoly" <anatoly.burakov@intel.com>]

On 11-Oct-18 11:00 AM, Alejandro Lucero wrote:
> On Fri, Sep 7, 2018 at 3:55 PM Darek Stojaczyk <dariusz.stojaczyk@intel.com>
> wrote:
> 
>> This allows DPDK to use RTE_IOVA_VA with VFIO/UIO-bound PCI
>> devices present on the system, but not attached to any
>> rte_pci_driver at the time of init.
>>
>> So far we used RTE_IOVA_VA whenever there was at least one
>> device attached to a driver with an RTE_PCI_DRV_IOVA_AS_VA flag,
>> meaning that other drivers which didn't explicitly report such
>> flag could have been forced to work in RTE_IOVA_VA as well.
>>
> 
> This is the opposite. Just one device not being able to use IOVA VA makes
> all to use IOVA PA.
> 
> 
>>
>> This patch makes the RTE_PCI_DRV_IOVA_AS_VA explicitly a hint.
>> If it's set, but RTE_IOVA_VA cannot be used, then EAL will print
>> a proper warning.
>>
>> Signed-off-by: Darek Stojaczyk <dariusz.stojaczyk@intel.com>
>> ---
>>   drivers/bus/pci/linux/pci.c | 11 +++++------
>>   1 file changed, 5 insertions(+), 6 deletions(-)
>>
>> diff --git a/drivers/bus/pci/linux/pci.c b/drivers/bus/pci/linux/pci.c
>> index 04648ac93..961e24024 100644
>> --- a/drivers/bus/pci/linux/pci.c
>> +++ b/drivers/bus/pci/linux/pci.c
>> @@ -534,7 +534,7 @@ pci_one_device_bound_uio(void)
>>    * Any one of the device has iova as va
>>    */
>>   static inline int
>> -pci_one_device_has_iova_va(void)
>> +pci_one_device_want_iova_va(void)
>>   {
>>          struct rte_pci_device *dev = NULL;
>>          struct rte_pci_driver *drv = NULL;
>> @@ -635,7 +635,7 @@ rte_pci_get_iommu_class(void)
>>   {
>>          bool is_bound;
>>          bool is_vfio_noiommu_enabled = true;
>> -       bool has_iova_va;
>> +       bool want_iova_va;
>>          bool is_bound_uio;
>>          bool iommu_no_va;
>>
>> @@ -643,7 +643,7 @@ rte_pci_get_iommu_class(void)
>>          if (!is_bound)
>>                  return RTE_IOVA_DC;
>>
>> -       has_iova_va = pci_one_device_has_iova_va();
>> +       want_iova_va = pci_one_device_want_iova_va();
>>          is_bound_uio = pci_one_device_bound_uio();
>>          iommu_no_va = !pci_devices_iommu_support_va();
>>   #ifdef VFIO_PRESENT
>> @@ -651,11 +651,10 @@ rte_pci_get_iommu_class(void)
>>                                          true : false;
>>   #endif
>>
>> -       if (has_iova_va && !is_bound_uio && !is_vfio_noiommu_enabled &&
>> -                       !iommu_no_va)
>> +       if (!is_bound_uio && !is_vfio_noiommu_enabled && !iommu_no_va)
>>                  return RTE_IOVA_VA;
>>
> 
> This is wrong. A device not able to work with IOVA VA will fail.
> 
> 
>>
>> -       if (has_iova_va) {
>> +       if (want_iova_va) {
>>                  RTE_LOG(WARNING, EAL, "Some devices want iova as va but pa
>> will be used because.. ");
>>                  if (is_vfio_noiommu_enabled)
>>                          RTE_LOG(WARNING, EAL, "vfio-noiommu mode
>> configured\n");
>> --
>> 2.17.1
>>
>>
> 

For these cases, i think the explicit IOVA mode on command line should 
work better. If the device has not reported IOVA as VA capability, it is 
to be assumed unsupported.

-- 
Thanks,
Anatoly