Re: [PATCH ats_vtd v5 20/22] pci: add a pci-level API for ATS

[Minwoo Im]

On 24-07-10 05:17:42, CLEMENT MATHIEU--DRIF wrote:
> 
> 
> On 09/07/2024 23:17, Minwoo Im wrote:
> > Caution: External email. Do not open attachments or click links, unless 
> > this email comes from a known sender and you know the content is safe.
> >
> >
> > On 24-07-09 11:58:53, CLEMENT MATHIEU--DRIF wrote:
> >>
> >> On 09/07/2024 12:15, Minwoo Im wrote:
> >>> Caution: External email. Do not open attachments or click links, unless 
> >>> this email comes from a known sender and you know the content is safe.
> >>>
> >>>
> >>> On 24-07-02 05:52:45, CLEMENT MATHIEU--DRIF wrote:
> >>>> From: Clément Mathieu--Drif <clement.mathieu--drif@eviden.com>
> >>>>
> >>>> Devices implementing ATS can send translation requests using
> >>>> pci_ats_request_translation_pasid.
> >>>>
> >>>> The invalidation events are sent back to the device using the iommu
> >>>> notifier managed with pci_register_iommu_tlb_event_notifier and
> >>>> pci_unregister_iommu_tlb_event_notifier
> >>>>
> >>>> Signed-off-by: Clément Mathieu--Drif <clement.mathieu--drif@eviden.com>
> >>>> ---
> >>>>    hw/pci/pci.c         | 44 +++++++++++++++++++++++++++++++++++++
> >>>>    include/hw/pci/pci.h | 52 ++++++++++++++++++++++++++++++++++++++++++++
> >>>>    2 files changed, 96 insertions(+)
> >>>>
> >>>> diff --git a/hw/pci/pci.c b/hw/pci/pci.c
> >>>> index 7a483dd05d..93b816aff2 100644
> >>>> --- a/hw/pci/pci.c
> >>>> +++ b/hw/pci/pci.c
> >>>> @@ -2833,6 +2833,50 @@ void pci_device_unset_iommu_device(PCIDevice *dev)
> >>>>        }
> >>>>    }
> >>>>
> >>>> +ssize_t pci_ats_request_translation_pasid(PCIDevice *dev, uint32_t 
> >>>> pasid,
> >>>> +                                          bool priv_req, bool exec_req,
> >>>> +                                          hwaddr addr, size_t length,
> >>>> +                                          bool no_write, IOMMUTLBEntry 
> >>>> *result,
> >>>> +                                          size_t result_length,
> >>>> +                                          uint32_t *err_count)
> >>>> +{
> >>>> +    assert(result_length);
> >>>> +    IOMMUMemoryRegion *iommu_mr = 
> >>>> pci_device_iommu_memory_region_pasid(dev,
> >>>> +                                                                        
> >>>> pasid);
> >>>> +    if (!iommu_mr || !pcie_ats_enabled(dev)) {
> >>>> +        return -EPERM;
> >>>> +    }
> >>>> +    return memory_region_iommu_ats_request_translation(iommu_mr, 
> >>>> priv_req,
> >>>> +                                                       exec_req, addr, 
> >>>> length,
> >>>> +                                                       no_write, result,
> >>>> +                                                       result_length,
> >>>> +                                                       err_count);
> >>>> +}
> >>> Can we use this function not from the endpoint PCI device, but inside of 
> >>> the pci
> >>> subsystem (hw/pci/pci.c) to make transparent abstraction for ATS request 
> >>> from
> >>> PCI endpoint device POV?  I guess it would be better to have PCI 
> >>> subsystem to
> >>> issue ATS request if pcie_ats_enabled(dev) rather than calling from the 
> >>> endpoint
> >>> side.
> >> Hi,
> >>
> >> This series aims to bring support for SVM (we are trying to integrate
> >> the patches bit by bit).
> >>   From a spec point of view, I don't know if it would make sense to
> >> implement the SVM logic at the PCI level
> >> as it's supposed to be implemented by endpoint devices.
> > Understood that this series is targeting the SVM usage.  But ATS feature is
> > something general to PCI devices, not only just for SVM, so I guess it 
> > would be
> > better to have caller to `pci_ats_request_translation_pasid()` in pci 
> > subsystem
> > like pci_dma_rw() to avoid duplicated implementation in the future for the
> > other PCI enpoint devices.
> 
> Would we store the ATC directly in the PCI subsytem?

Yes, endpoint device (e.g., svm.c) should call pci_* helpers in PCI subsystem
with `PCIDevice *pdev instance` which represents the endpoint device itself.
By the instance, we can look up the IOTLB entry from the ATC in the PCI
subsystem, not the current caller side.

> >
> >> However, we could consider providing a reference/reusable/encapsulated
> >> implementation of SVM with a simplified API
> >> that would call the pci_* functions under the hood.
> > I would prefer that PCI devices which want to request ATS translation has no
> > additional implementation for ATS, but only pcie_ats_init().
> Hi,
> 
> I think both strategies can coexist.
> Keeping control can be interesting for people who use Qemu for hardware 
> prototyping and who generally want to experiment.
> We can keep the current PCI-level API for devices that want to 
> reimplement the logic themselves
> and add a kind of "DMA module"/"ATS+PRI module" that works out of the box.

I think we should proivde hybrid mode on this.  One for a `generic` cache
policy mode for every PCI endpoint devices which can be controlled in the PCI
subsystem for ATC, the other one is that device-specific cache policy mode
which let each device implement their own ATC lookup behaviors to optimize
their own caching impact.

> That module could be called "struct PciDmaModule" and expose a simple 
> set of functions like pci_dma_module_init, pci_dma_module_read, 
> pci_dma_module_write.
> I think it's important to keep existing DMA API as is to allow devices 
> to do both "with ATS" and "without ATS" operations.
> 
> Do you agree with that?

Indeed.  Keeping the existing APIs is a good choice, but I would like to have
endpoint devices code much more simpler for the generic usages :)

> >
> >> Do you have a specific use case in mind?
> > ATS/PRI is the actual use case, and it's not that different what you are
> > targeting for :)
> >
> >>   >cmd
> >>
> >>>