Re: [PATCH] PCI/MSI: pci-xgene-msi: Enable MSI support in ACPI boot for X-Gene v1

[Lorenzo Pieralisi]

[+ Rafael]

On Thu, May 26, 2016 at 01:49:23PM -0700, Duc Dang wrote:
> Hi Lorenzo,
> 
> On Thu, May 26, 2016 at 5:34 AM, Lorenzo Pieralisi
> <lorenzo.pieralisi@xxxxxxx> wrote:
> > Hi Duc,
> >
> > On Wed, May 25, 2016 at 04:13:35PM -0700, Duc Dang wrote:
> >> On Thu, Feb 25, 2016 at 9:38 AM, Lorenzo Pieralisi
> >> <lorenzo.pieralisi@xxxxxxx> wrote:
> >> > On Wed, Feb 24, 2016 at 02:28:10PM -0800, Duc Dang wrote:
> >> >> On Wed, Feb 24, 2016 at 8:16 AM, Marc Zyngier <marc.zyngier@xxxxxxx> wrote:
> >> >> > On 24/02/16 16:09, Mark Salter wrote:
> >> >> >> On Tue, 2016-02-09 at 17:56 -0800, Duc Dang wrote:
> >> >> >>> This patch makes pci-xgene-msi driver ACPI-aware and provides
> >> >> >>> MSI capability for X-Gene v1 PCIe controllers in ACPI boot mode.
> >> >> >>>
> >> >> >>> Signed-off-by: Duc Dang <dhdang@xxxxxxx>
> >> >> >>> ---
> >> >> >>>  drivers/pci/host/pci-xgene-msi.c | 35 ++++++++++++++++++++++++++++++++---
> >> >> >>>  1 file changed, 32 insertions(+), 3 deletions(-)
> >> >> >>>
> >> >> >>> diff --git a/drivers/pci/host/pci-xgene-msi.c b/drivers/pci/host/pci-xgene-msi.c
> >> >> >>> index a6456b5..466aa93 100644
> >> >> >>> --- a/drivers/pci/host/pci-xgene-msi.c
> >> >> >>> +++ b/drivers/pci/host/pci-xgene-msi.c
> >> >> >>> @@ -24,6 +24,7 @@
> >> >> >>>  #include <linux/pci.h>
> >> >> >>>  #include <linux/platform_device.h>
> >> >> >>>  #include <linux/of_pci.h>
> >> >> >>> +#include <linux/acpi.h>
> >> >> >>>
> >> >> >>>  #define MSI_IR0                     0x000000
> >> >> >>>  #define MSI_INT0            0x800000
> >> >> >>> @@ -39,7 +40,7 @@ struct xgene_msi_group {
> >> >> >>>  };
> >> >> >>>
> >> >> >>>  struct xgene_msi {
> >> >> >>> -    struct device_node      *node;
> >> >> >>> +    struct fwnode_handle    *fwnode;
> >> >> >>>      struct irq_domain       *inner_domain;
> >> >> >>>      struct irq_domain       *msi_domain;
> >> >> >>>      u64                     msi_addr;
> >> >> >>> @@ -249,6 +250,13 @@ static const struct irq_domain_ops msi_domain_ops = {
> >> >> >>>      .free   = xgene_irq_domain_free,
> >> >> >>>  };
> >> >> >>>
> >> >> >>> +#ifdef CONFIG_ACPI
> >> >> >>> +static struct fwnode_handle *xgene_msi_get_fwnode(struct device *dev)
> >> >> >>> +{
> >> >> >>> +    return xgene_msi_ctrl.fwnode;
> >> >> >>> +}
> >> >> >>> +#endif
> >> >> >>> +
> >> >> >>>  static int xgene_allocate_domains(struct xgene_msi *msi)
> >> >> >>>  {
> >> >> >>>      msi->inner_domain = irq_domain_add_linear(NULL, NR_MSI_VEC,
> >> >> >>> @@ -256,7 +264,7 @@ static int xgene_allocate_domains(struct xgene_msi *msi)
> >> >> >>>      if (!msi->inner_domain)
> >> >> >>>              return -ENOMEM;
> >> >> >>>
> >> >> >>> -    msi->msi_domain = pci_msi_create_irq_domain(of_node_to_fwnode(msi->node),
> >> >> >>> +    msi->msi_domain = pci_msi_create_irq_domain(msi->fwnode,
> >> >> >>>                                                  &xgene_msi_domain_info,
> >> >> >>>                                                  msi->inner_domain);
> >> >> >>
> >> >> >> This doesn't work for me (ACPI probing on Mustang) unless I change this
> >> >> >> to be pci_msi_create_default_irq_domain(). The problem seems to be that
> >> >> >> the MSI probe happens after the PCIe RC is probed so there is no MSI domain
> >> >> >> at the time the PCIe root is initialized by ACPI.
> >> >> >
> >> >> > pci_msi_create_default_irq_domain is the wrong thing do use, specially
> >> >> > if you have multiple MSI controllers in the system. I certainly wouldn't
> >> >> > want to see it being used on arm64.
> >> >> >
> >> >> > This is the usual dependency hell. You try moving the probing earlier,
> >> >> > but that may break something else in the process.
> >> >>
> >> >> Hi Mark and Marc,
> >> >>
> >> >> I have modified Tianocore firmware to have MSI node declared before
> >> >> PCIe node in Dsdt table. With this modification, the MSI driver will
> >> >> be loaded before PCIe driver and MSI domain is available at the time
> >> >> PCIe root is initialized.
> >> >
> >> > I am totally against this. We should not hack ACPI tables to make
> >> > the kernel work (on top of that with unwritten ordering rules that may
> >> > well require changes as kernel evolves), we should have a standard set
> >> > of bindings that people use to describe HW in the DSDT and the kernel(s)
> >> > has to cope with that. If there is a dependency problem in the description
> >> > we may solve it at bindings level, but I absolutely do not want to rely
> >> > on DSDT nodes ordering for things to work, that's fragile, no shortcuts
> >> > please.
> >>
> >> Hi Lorenzo, Bjorn,
> >>
> >> (Refresh this thread on this merge cycle)
> >>
> >> I tried to use _DEP method to describe the dependency of PCIe node to
> >> MSI node but it turns out that PCIe driver does not check for the
> >> dependency (only EC (ec.c) and I2C core (i2c-core.c) use
> >> acpi_walk_dep_device_list to check for dependency before loading
> >> dependent drivers)
> >>
> >> Do you have other suggestion/example about how to ensure strict
> >> ordering on driver loading in ACPI boot mode without changing the
> >> order of ACPI nodes in DSDT?
> >
> > Before doing anything else, I would like to ask you to report the code
> > paths leading to failures (what fails actually ?) please, and I want
> > to understand what "this does not work" means.
> >
> > In particular:
> >
> > - Why using pci_msi_create_default_irq_domain() works
> > - Why, if you change the DSDT nodes ordering, things work (please add
> >   some debugging to MSI allocation functions to detect the code paths
> >   leading to correct MSI allocation)
> > - What's the difference wrt the DT probe path
> 
> The MSI allocation happens with following code path (I use PMC pm80xx
> SAS driver as an example (drivers/scsi/pm8001/pm8001_init.c)):
> pci_enable_msix_exact
>     pci_enable_msix_range
>         pci_enable_msix
>             msix_capbility_init
>                 pci_msi_setup_msi_irqs
>                     pci_msi_get_domain
>                         dev_get_msi_domain
>                         arch_get_pci_msi_domain
>                             return pci_msi_default_domain;
> 
> In my failing case: pci_msi_get_domain return NULL, so MSI allocation
> function (pci_enable_msix_exact) fails.
> 
> The expected MSI domain is registered when X-Gene MSI driver is
> initialized and is attached to each PCI bus by calling
> pci_set_bus_msi_domain during PCIe bus-scan:
> pci_scan_root_bus
>     pci_scan_root_bus_msi
>         pci_scan_child_bus
>             pci_scan_bridge
>                 pci_add_new_bus
>                     pci_alloc_child_bus
>                         pci_set_bus_msi_domain
>                             pci_host_bridge_msi_domain
>                                 pci_host_bridge_of_msi_domain
>                                 pci_host_bridge_acpi_msi_domain
> 
> + When booting with DT: X-Gene MSI driver is initialized early
> (subsys_initcall level); X-Gene PCIe driver is initialized later at
> device_initcall level (using module_platform_driver). Because of that
> when PCIe enumeration happens, the X-Gene MSI domain was already
> registered and available. So when the driver of PCIe Endpoint devices
> gets loaded, the call to pci_enable_msix_exact (to allocate MSI)
> succeeds.
> 
> + When booting with ACPI: acpi_scan_init is called at subsys_initcall
> level (the same level with X-Gene MSI driver):
> acpi_init
>     acpi_bus_init (GIC is initialized here)
>         acpi_scan_init
>             acpi_pci_root_init
>             acpi_pci_link_init
>             ...
>             acpi_bus_scan
>                 acpi_bus_check_add
>  ...
> 
> Currently, in DSDT table, X-Gene MSI node is defined right after
> X-Gene PCIe nodes. So PCIe driver initialization and PCIe enumeration
> happens before MSI driver is initialized. During PCIe bus scan, the
> pci_set_bus_msi_domain gets a NULL  irq_domain value from
> pci_host_bridge_msi_domain (as MSI driver is not initialized), so
> there is no MSI support for PCIe bus.
> 
> + Using pci_msi_create_default_irq_domain() helps MSI allocation
> succeed as pci_msi_get_domain (in the code path of
> pci_enable_msix_exact above) can fall back to get a valid domain from
> arch_get_pci_msi_domain (please note that PCIe endpoint driver gets
> called later (at device_initcall level (using module_init)), so at the
> time PCIe endpoint driver probe function executes, the MSI driver
> already register the global MSI domain with
> pci_msi_create_default_irq_domain)
> 
> + If changing DSDT node ordering, to define MSI node before PCIe
> nodes, then X-Gene MSI driver will be loaded before PCIe host driver.
> At the time PCIe enumeration happens, X-Gene MSI driver already
> registered a valid MSI domain (by calling
> pci_msi_register_fwnode_provider to register the fwnode provider
> function), so all the detected PCIe buses (and PCIe endpoint devices)
> will have a valid MSI domain attached to them (as a result of function
> pci_set_bus_msi_domain)

This still works out of sheer luck, given that relying on initcall
ordering at same init level is undefined (ie it works if your msi driver
registration call is executed before acpi_init() is called).

I though about using a scan handler like:

See arch/ia64/hp/common/sba_iommu.c acpi_sba_ioc_handler

but that won't solve your problem either since, given that PNP0A03 (ie
PCI host bridge) is handled through a scan handler too, IIUC you will
still have issues related to namespace device nodes ordering.

I suspect we will have to add a hook walking the ACPI namespace before
acpi_bus_scan() is called in acpi_scan_init() (aka MSI controllers must
be probed before any other device is, that's the gist), I CC'ed Rafael
to see if there is any other suitable way to sort out this issue.

Thanks,
Lorenzo