On Mon, Jan 4, 2021 at 6:13 AM Kishon Vijay Abraham I [off-list ref] wrote:

Hi Rob,

On 15/12/20 9:31 pm, Rob Herring wrote:

quoted

On Wed, Nov 11, 2020 at 9:37 AM Kishon Vijay Abraham I [off-list ref] wrote:

quoted

Implement ->msi_map_irq() ops in order to map physical address to
MSI address and return MSI data.

Signed-off-by: Kishon Vijay Abraham I <redacted>
---
 .../pci/controller/cadence/pcie-cadence-ep.c  | 53 +++++++++++++++++++
 1 file changed, 53 insertions(+)

diff --git a/drivers/pci/controller/cadence/pcie-cadence-ep.c b/drivers/pci/controller/cadence/pcie-cadence-ep.c
index 84cc58dc8512..1fe6b8baca97 100644
--- a/drivers/pci/controller/cadence/pcie-cadence-ep.c
+++ b/drivers/pci/controller/cadence/pcie-cadence-ep.c

@@ -382,6 +382,57 @@ static int cdns_pcie_ep_send_msi_irq(struct cdns_pcie_ep *ep, u8 fn,
        return 0;
 }

+static int cdns_pcie_ep_map_msi_irq(struct pci_epc *epc, u8 fn,
+                                   phys_addr_t addr, u8 interrupt_num,
+                                   u32 entry_size, u32 *msi_data,
+                                   u32 *msi_addr_offset)
+{
+       struct cdns_pcie_ep *ep = epc_get_drvdata(epc);
+       u32 cap = CDNS_PCIE_EP_FUNC_MSI_CAP_OFFSET;
+       struct cdns_pcie *pcie = &ep->pcie;
+       u64 pci_addr, pci_addr_mask = 0xff;
+       u16 flags, mme, data, data_mask;
+       u8 msi_count;
+       int ret;
+       int i;
+

quoted

+       /* Check whether the MSI feature has been enabled by the PCI host. */
+       flags = cdns_pcie_ep_fn_readw(pcie, fn, cap + PCI_MSI_FLAGS);
+       if (!(flags & PCI_MSI_FLAGS_ENABLE))
+               return -EINVAL;
+
+       /* Get the number of enabled MSIs */
+       mme = (flags & PCI_MSI_FLAGS_QSIZE) >> 4;
+       msi_count = 1 << mme;
+       if (!interrupt_num || interrupt_num > msi_count)
+               return -EINVAL;
+
+       /* Compute the data value to be written. */
+       data_mask = msi_count - 1;
+       data = cdns_pcie_ep_fn_readw(pcie, fn, cap + PCI_MSI_DATA_64);
+       data = data & ~data_mask;
+
+       /* Get the PCI address where to write the data into. */
+       pci_addr = cdns_pcie_ep_fn_readl(pcie, fn, cap + PCI_MSI_ADDRESS_HI);
+       pci_addr <<= 32;
+       pci_addr |= cdns_pcie_ep_fn_readl(pcie, fn, cap + PCI_MSI_ADDRESS_LO);
+       pci_addr &= GENMASK_ULL(63, 2);

Wouldn't all of the above be the same code for any endpoint driver? We
just need endpoint config space accessors for the same 32-bit only
access issues. Not asking for that in this series, but if that's the
direction we should go.

Do you mean "endpoint" variant of pci_generic_config_read() which takes
function number and capability offset? That could be done but we have to
add support to traverse the linked list of capabilities though the
capabilities are going to be at a fixed location for a given IP.

Well, the above code would call the equivalent of
pci_bus_read_config_*() functions which then calls driver specific
read/write ops like pci_generic_config_read().

Once we have common accessors, then functions to get the capability
offsets would be common too. It shouldn't matter that they happen to
be fixed, walking the linked list should work either way. Getting rid
of fixed offsets for the host side drivers is something I've been
doing too.

Also in some cases, the writes are to a different register than the
configuration space registers like vendor_id in Cadence EP should be
written to Local Management register instead of the configuration space
register.

We have the same issue on the host side as well. That just means we
need to wrap the generic ops functions.

Rob