Re: [Query] ACS enablement in the DT based boot flow

[Will Deacon]

On Wed, Apr 10, 2024 at 02:28:40PM -0500, Bjorn Helgaas wrote:
> [+cc Will, Joerg]
> 
> On Mon, Apr 01, 2024 at 10:40:15AM +0000, Vidya Sagar wrote:
> > Hi folks,
> > ACS (Access Control Services) is configured for a PCI device through
> > pci_enable_acs().  The first thing pci_enable_acs() checks for is
> > whether the global flag 'pci_acs_enable' is set or not.  The global
> > flag 'pci_acs_enable' is set by the function pci_request_acs().
> > 
> > pci_enable_acs() function is called whenever a new PCI device is
> > added to the system
> > 
> >  pci_enable_acs+0x4c/0x2a4
> >  pci_acs_init+0x38/0x60
> >  pci_device_add+0x1a0/0x670
> >  pci_scan_single_device+0xc4/0x100
> >  pci_scan_slot+0x6c/0x1e0
> >  pci_scan_child_bus_extend+0x48/0x2e0
> >  pci_scan_root_bus_bridge+0x64/0xf0
> >  pci_host_probe+0x18/0xd0
> > 
> > In the case of a system that boots using device-tree blob,
> > pci_request_acs() is called when the device driver binds with the
> > respective device
> > 
> > of_iommu_configure+0xf4/0x230
> > of_dma_configure_id+0x110/0x340
> > pci_dma_configure+0x54/0x120
> > really_probe+0x80/0x3e0
> > __driver_probe_device+0x88/0x1c0
> > driver_probe_device+0x3c/0x140
> > __device_attach_driver+0xe8/0x1e0
> > bus_for_each_drv+0x78/0xf0
> > __device_attach+0x104/0x1e0
> > device_attach+0x14/0x30
> > pci_bus_add_device+0x50/0xd0
> > pci_bus_add_devices+0x38/0x90
> > pci_host_probe+0x40/0xd0
> > 
> > Since the device addition always happens first followed by the
> > driver binding, this flow effectively makes sure that ACS never gets
> > enabled.
> > 
> > Ideally, I would expect the pci_request_acs() get called (probably
> > by the OF framework itself) before calling pci_enable_acs().
> > 
> > This happens in the ACPI flow where pci_request_acs() is called
> > during IORT node initialization (i.e. iort_init_platform_devices()
> > function).
> > 
> > Is this understanding correct? If yes, would it make sense to call
> > pci_request_acs() during OF initialization (similar to IORT
> > initialization in ACPI flow)?
> 
> Your understanding looks correct to me.  My call graph notes, FWIW:
> 
>   mem_init
>     pci_iommu_alloc                   # x86 only
>       amd_iommu_detect                # init_state = IOMMU_START_STATE
>         iommu_go_to_state(IOMMU_IVRS_DETECTED)
>           state_next
>             switch (init_state)
>             case IOMMU_START_STATE:
>               detect_ivrs
>                 pci_request_acs
>                   pci_acs_enable = 1  # <--
>       detect_intel_iommu
>         pci_request_acs
>           pci_acs_enable = 1          # <--
> 
>   pci_scan_single_device              # PCI enumeration
>     ...
>       pci_init_capabilities
>         pci_acs_init
>           pci_enable_acs
>             if (pci_acs_enable)       # <--
>               pci_std_enable_acs
> 
>   __driver_probe_device
>     really_probe
>       pci_dma_configure               # pci_bus_type.dma_configure
>         if (OF)
>           of_dma_configure
>             of_dma_configure_id
>               of_iommu_configure
>                 pci_request_acs       # <-- 6bf6c24720d3
>                 iommu_probe_device
>         else if (ACPI)
>           acpi_dma_configure
>             acpi_dma_configure_id
>               acpi_iommu_configure_id
>                 iommu_probe_device
> 
> The pci_request_acs() in of_iommu_configure(), which happens too late
> to affect pci_enable_acs(), was added by 6bf6c24720d3 ("iommu/of:
> Request ACS from the PCI core when configuring IOMMU linkage"), so I
> cc'd Will and Joerg.  I don't know if that *used* to work and got
> broken somehow, or if it never worked as intended.

I don't have any way to test this, but I'm supportive of having the same
flow for DT and ACPI-based flows. Vidya, are you able to cook a patch?

Will