On Thu, Sep 14, 2023 at 07:56:21AM +0000, Yoshihiro Shimoda wrote:
> > From: Bjorn Helgaas, Sent: Thursday, September 14, 2023 8:31 AM
> > On Fri, Aug 25, 2023 at 06:32:04PM +0900, Yoshihiro Shimoda wrote:
> > > Add support for triggering INTx IRQs by using outbound iATU.
> > > Outbound iATU is utilized to send assert and de-assert INTA TLPs
> > > as simulated edge IRQ for INTA. (Other INT[BCD] are not asserted.)
> > > This INTx support is optional (if there is no memory for INTx,
> > > probe will not fail).
> > >
> > > The message is generated based on the payloadless Msg TLP with type
> > > 0x14, where 0x4 is the routing code implying the Terminate at
> > > Receiver message. The message code is specified as b1000xx for
> > > the INTx assertion and b1001xx for the INTx de-assertion.
> > >
> > > Signed-off-by: Yoshihiro Shimoda <yoshihiro.shimoda.uh@xxxxxxxxxxx>
> > > Reviewed-by: Serge Semin <fancer.lancer@xxxxxxxxx>
> > > ---
> > > .../pci/controller/dwc/pcie-designware-ep.c | 70 +++++++++++++++++--
> > > drivers/pci/controller/dwc/pcie-designware.h | 2 +
> > > 2 files changed, 68 insertions(+), 4 deletions(-)
> > >
> > > diff --git a/drivers/pci/controller/dwc/pcie-designware-ep.c b/drivers/pci/controller/dwc/pcie-designware-ep.c
> > > index 747d5bc07222..91e3c4335031 100644
> > > --- a/drivers/pci/controller/dwc/pcie-designware-ep.c
> > > +++ b/drivers/pci/controller/dwc/pcie-designware-ep.c
> > > @@ -6,9 +6,11 @@
> > > * Author: Kishon Vijay Abraham I <kishon@xxxxxx>
> > > */
> > >
> > > +#include <linux/delay.h>
> > > #include <linux/of.h>
> > > #include <linux/platform_device.h>
> > >
> > > +#include "../../pci.h"
> > > #include "pcie-designware.h"
> > > #include <linux/pci-epc.h>
> > > #include <linux/pci-epf.h>
> > > @@ -484,14 +486,61 @@ static const struct pci_epc_ops epc_ops = {
> > > .get_features = dw_pcie_ep_get_features,
> > > };
> > >
> > > +static int dw_pcie_ep_send_msg(struct dw_pcie_ep *ep, u8 func_no, u8 code,
> > > + u8 routing)
> > > +{
> > > + struct dw_pcie_ob_atu_cfg atu = { 0 };
> > > + struct pci_epc *epc = ep->epc;
> > > + int ret;
> > > +
> > > + atu.func_no = func_no;
> > > + atu.code = code;
> > > + atu.routing = routing;
> > > + atu.type = PCIE_ATU_TYPE_MSG;
> > > + atu.cpu_addr = ep->intx_mem_phys;
> > > + atu.size = epc->mem->window.page_size;
> > > +
> > > + ret = dw_pcie_ep_outbound_atu(ep, &atu);
> > > + if (ret)
> > > + return ret;
> > > +
> > > + /* A dummy-write ep->intx_mem is converted to a Msg TLP */
> > > + writel(0, ep->intx_mem);
> > > +
> > > + dw_pcie_ep_unmap_addr(epc, func_no, 0, ep->intx_mem_phys);
> > > +
> > > + return 0;
> > > +}
> > > +
> > > int dw_pcie_ep_raise_legacy_irq(struct dw_pcie_ep *ep, u8 func_no)
> > > {
> > > struct dw_pcie *pci = to_dw_pcie_from_ep(ep);
> > > struct device *dev = pci->dev;
> > > + int ret;
> > >
> > > - dev_err(dev, "EP cannot trigger legacy IRQs\n");
> > > + if (!ep->intx_mem) {
> > > + dev_err(dev, "legacy IRQs not supported\n");
> > > + return -EOPNOTSUPP;
> > > + }
> > >
> > > - return -EINVAL;
> > > + /*
> > > + * Even though the PCI bus specification implies the level-triggered
> > > + * INTx interrupts the kernel PCIe endpoint framework has a single
> > > + * PCI_EPC_IRQ_INTx flag defined for the legacy IRQs simulation. Thus
> > > + * this function sends the Deassert_INTx PCIe TLP after the Assert_INTx
> > > + * message with the 50 usec duration basically implementing the
> > > + * rising-edge triggering IRQ. Hopefully the interrupt controller will
> > > + * still be able to register the incoming IRQ event...
> >
> > I'm not really convinced about this "assert INTA, wait 50us, deassert
> > INTA" thing. All the INTx language in the spec is like this:
> >
> > ... the virtual INTx wire must be asserted whenever and *as long as*
> > the following conditions are satisfied:
> >
> > - The Interrupt Disable bit in the Command register is set to 0b.
> >
> > - The <feature> Interrupt Enable bit in the <feature> Control
> > Register is set to 1b.
> >
> > - The <feature> Status bit in the <feature> Status register is
> > set.
> >
> > E.g., sec PCIe r6.0, sec 5.5.6 (Link Activation), 6.1.6 (Native PME),
> > 6.2.4.1.2 (AER Interrupt Generation), 6.2.11.1 (DPC Interrupts),
> > 6.7.3.4 (Software Notification of Hot-Plug Events).
> >
> > So it seems to me like the endpoint needs an "interrupt status" bit,
> > and the Deassert_INTx message would be sent when the host interrupt
> > handler clears that bit.
>
> Thank you very much for your comments! About this topic,
> Frank also has a similar opinion before [1]. So, I asked Kishon
> about this, but I didn't get any comment from Kishon at that time.
> Anyway, to handle INTx on PCIe endpoint framework properly,
> we need to modify the PCIe Endpoint framework, IIUC.
>
> Should I modify the PCIe Endpoint framework at first?
> Or, can this patch be applied as-is?
> I guess that such modification of the PCIe Endpoint framework
> need much time. So, if I should modify the framework at first,
> I would like to drop adding INTx support [2] from my patch series
> because supporting INTx on my PCIe controller is not mandatory.
>
> [1]
> https://lore.kernel.org/linux-pci/TYBPR01MB5341EFAC471AEBB9100D6051D8719@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/
>
> [2]
> The following patches are not needed if INTx support should be dropped:
>
> eb185e1e628a PCI: designware-ep: Add INTx IRQs support
> 5d0e51f85b23 PCI: dwc: Add outbound MSG TLPs support
> 4758bef61cc2 PCI: dwc: Change arguments of dw_pcie_prog_outbound_atu()
> 44938b13046b PCI: Add INTx Mechanism Messages macros
Since INTx support isn't mandatory at this time, I think we should
drop these patches for now. If/when somebody definitely needs INTx
support, we can resurrect them, and then we'll have more clarity on
how it should work and we'll be better able to test it.
Bjorn
