On Wed, 16 Oct 2024 13:07:37 -0500
Terry Bowman <Terry.Bowman@xxxxxxx> wrote:
> Hi Jonathan,
>
> On 10/16/24 11:54, Jonathan Cameron wrote:
> > On Tue, 8 Oct 2024 17:16:49 -0500
> > Terry Bowman <terry.bowman@xxxxxxx> wrote:
> >
> >> The current pcie_do_recovery() handles device recovery as result of
> >> uncorrectable errors (UCE). But, CXL port devices require unique
> >> recovery handling.
> >>
> >> Create a cxl_do_recovery() function parallel to pcie_do_recovery(). Add CXL
> >> specific handling to the new recovery function.
> >>
> >> The CXL port UCE recovery must invoke the AER service driver's CXL port
> >> UCE callback. This is different than the standard pcie_do_recovery()
> >> recovery that calls the pci_driver::err_handler UCE handler instead.
> >>
> >> Treat all CXL PCIe port UCE errors as fatal and call kernel panic to
> >> "recover" the error. A panic is called instead of attempting recovery
> >> to avoid potential system corruption.
> >>
> >> The uncorrectable support added here will be used to complete CXL PCIe
> >> port error handling in the future.
> >>
> >> Signed-off-by: Terry Bowman <terry.bowman@xxxxxxx>
> >
> > Hi Terry,
> >
> > I'm a little bothered by the subtle difference in the bus walks
> > in here vs the existing cases. If we need them, comments needed
> > to explain why.
> >
>
> Yes, I will add more details in the commit message about "why".
> I added explanation following your below comment.
>
> > If we are going to have separate handling, see if you can share
> > a lot more of the code by factoring out common functions for
> > the pci and cxl handling with callbacks to handle the differences.
> >
>
> Dan requested separate paths for the PCIe and CXL recovery. The intent,
> as I understand, is to isolate the handling of PCIe and CXL protocol
> errors. This is to create 2 different classes of protocol errors.
Function call chain wise I'm reasonably convinced that might be a good
idea. But not code wise if it means we end up with more hard to review
code.
>
> > I've managed to get my head around this code a few times in the past
> > (I think!) and really don't fancy having two subtle variants to
> > consider next time we get a bug :( The RC_EC additions hurt my head.
> >
> > Jonathan
>
> Right, the UCE recovery logic is not straightforward. The code can be
> refactored to take advantage of reuse. I'm interested in your thoughts
> after I have provided some responses here.
>
> >
> >> static int handles_cxl_error_iter(struct pci_dev *dev, void *data)
> >> diff --git a/drivers/pci/pcie/err.c b/drivers/pci/pcie/err.c
> >> index 31090770fffc..de12f2eb19ef 100644
> >> --- a/drivers/pci/pcie/err.c
> >> +++ b/drivers/pci/pcie/err.c
> >> @@ -86,6 +86,63 @@ static int report_error_detected(struct pci_dev *dev,
> >> return 0;
> >> }
> >>
> >> +static int cxl_report_error_detected(struct pci_dev *dev,
> >> + pci_channel_state_t state,
> >> + enum pci_ers_result *result)
> >> +{
> >> + struct cxl_port_err_hndlrs *cxl_port_hndlrs;
> >> + struct pci_driver *pdrv;
> >> + pci_ers_result_t vote;
> >> +
> >> + device_lock(&dev->dev);
> >> + cxl_port_hndlrs = find_cxl_port_hndlrs();
> >
> > Can we refactor to have a common function under this and report_error_detected()?
> >
>
> Sure, this can be refactored.
>
> The difference between cxl_report_error_detected() and report_error_detected() is the
> handlers that are called.
>
> cxl_report_error_detected() calls the CXL driver's registered port error handler.
>
> report_error_recovery() calls the pcie_dev::err_handlers.
>
> Let me know if I should refactor for common code here?
It certainly makes sense to do that somewhere in here. Just have light
wrappers that provide callbacks so the bulk of the code is shared.
>
>
> >> + pdrv = dev->driver;
> >> + if (pci_dev_is_disconnected(dev)) {
> >> + vote = PCI_ERS_RESULT_DISCONNECT;
> >> + } else if (!pci_dev_set_io_state(dev, state)) {
> >> + pci_info(dev, "can't recover (state transition %u -> %u invalid)\n",
> >> + dev->error_state, state);
> >> + vote = PCI_ERS_RESULT_NONE;
> >> + } else if (!cxl_port_hndlrs || !cxl_port_hndlrs->error_detected) {
> >> + if (dev->hdr_type != PCI_HEADER_TYPE_BRIDGE) {
> >> + vote = PCI_ERS_RESULT_NO_AER_DRIVER;
> >> + pci_info(dev, "can't recover (no error_detected callback)\n");
> >> + } else {
> >> + vote = PCI_ERS_RESULT_NONE;
> >> + }
> >> + } else {
> >> + vote = cxl_port_hndlrs->error_detected(dev, state);
> >> + }
> >> + pci_uevent_ers(dev, vote);
> >> + *result = merge_result(*result, vote);
> >> + device_unlock(&dev->dev);
> >> + return 0;
> >> +}
> >
> >> static int pci_pm_runtime_get_sync(struct pci_dev *pdev, void *data)
> >> {
> >> pm_runtime_get_sync(&pdev->dev);
> >> @@ -188,6 +245,28 @@ static void pci_walk_bridge(struct pci_dev *bridge,
> >> cb(bridge, userdata);
> >> }
> >>
> >> +/**
> >> + * cxl_walk_bridge - walk bridges potentially AER affected
> >> + * @bridge: bridge which may be a Port, an RCEC, or an RCiEP
> >> + * @cb: callback to be called for each device found
> >> + * @userdata: arbitrary pointer to be passed to callback
> >> + *
> >> + * If the device provided is a bridge, walk the subordinate bus, including
> >> + * the device itself and any bridged devices on buses under this bus. Call
> >> + * the provided callback on each device found.
> >> + *
> >> + * If the device provided has no subordinate bus, e.g., an RCEC or RCiEP,
> >> + * call the callback on the device itself.
> > only call the callback on the device itself.
> >
> > (as you call it as stated above either way).
> >
>
> Thanks. I will update the function header to include "only".
>
> >> + */
> >> +static void cxl_walk_bridge(struct pci_dev *bridge,
> >> + int (*cb)(struct pci_dev *, void *),
> >> + void *userdata)
> >> +{
> >> + cb(bridge, userdata);
> >> + if (bridge->subordinate)
> >> + pci_walk_bus(bridge->subordinate, cb, userdata);
> > The difference between this and pci_walk_bridge() is subtle and
> > I'd like to avoid having both if we can.
> >
>
> The cxl_walk_bridge() was added because pci_walk_bridge() does not report
> CXL errors as needed. If the erroring device is a bridge then pci_walk_bridge()
> does not call report_error_detected() for the root port itself. If the bridge
> is a CXL root port then the CXL port error handler is not called. This has 2
> problems: 1. Error logging is not provided, 2. A result vote is not provided
> by the root port's CXL port handler.
So what happens for PCIe errors on the root port? How are they reported?
What I'm failing to understand is why these should be different.
Maybe there is something missing on the PCIe side though!
That code plays a game with what bridge and I thought that was there to handle
this case.
>
> >> +}
> >> +
> >> pci_ers_result_t pcie_do_recovery(struct pci_dev *dev,
> >> pci_channel_state_t state,
> >> pci_ers_result_t (*reset_subordinates)(struct pci_dev *pdev))
> >> @@ -276,3 +355,74 @@ pci_ers_result_t pcie_do_recovery(struct pci_dev *dev,
> >>
> >> return status;
> >> }
> >> +
> >> +pci_ers_result_t cxl_do_recovery(struct pci_dev *bridge,
> >> + pci_channel_state_t state,
> >> + pci_ers_result_t (*reset_subordinates)(struct pci_dev *pdev))
> >> +{
> >> + struct pci_host_bridge *host = pci_find_host_bridge(bridge->bus);
> >> + pci_ers_result_t status = PCI_ERS_RESULT_CAN_RECOVER;
> >> + int type = pci_pcie_type(bridge);
> >> +
> >> + if ((type != PCI_EXP_TYPE_ROOT_PORT) &&
> >> + (type != PCI_EXP_TYPE_RC_EC) &&
> >> + (type != PCI_EXP_TYPE_DOWNSTREAM) &&
> >> + (type != PCI_EXP_TYPE_UPSTREAM)) {
> >> + pci_dbg(bridge, "Unsupported device type (%x)\n", type);
> >> + return status;
> >> + }
> >> +
> >
> > Would similar trick to in pcie_do_recovery work here for the upstream
> > and downstream ports use pci_upstream_bridge() and for the others pass the dev into
> > pci_walk_bridge()?
> >
>
> Yes, that would be a good starting point to begin reuse refactoring.
> I'm interested in getting yours and others feedback on the separation of the
> PCI and CXL protocol errors and how much separation is or not needed.
Separation may make sense (I'm still thinking about it) for separate passes
through the topology and separate callbacks / handling when an error is seen.
What I don't want to see is two horribly complex separate walking codes if
we can possibly avoid it. Long term to me that just means two sets of bugs
and problem corners instead of one.
Jonathan
>
>
> Regards,
> Terry
>
