On Wed, 28 Jan 2026 20:27:31 +0800 Shuai Xue <[email protected]> wrote:
> On 1/27/26 6:24 PM, Jonathan Cameron wrote: > > On Sat, 24 Jan 2026 15:45:54 +0800 > > Shuai Xue <[email protected]> wrote: > > > >> The current implementation of pcie_do_recovery() assumes that the > >> recovery process is executed for the device that detected the error. > >> However, the DPC driver currently passes the error port that experienced > >> the DPC event to pcie_do_recovery(). > >> > >> Use the SOURCE ID register to correctly identify the device that > >> detected the error. When passing the error device, the > >> pcie_do_recovery() will find the upstream bridge and walk bridges > >> potentially AER affected. And subsequent commits will be able to > >> accurately access AER status of the error device. > >> > >> Should not observe any functional changes. > >> > >> Reviewed-by: Kuppuswamy Sathyanarayanan > >> <[email protected]> > >> Signed-off-by: Shuai Xue <[email protected]> Hi Shuai, > >> --- > >> drivers/pci/pci.h | 2 +- > >> drivers/pci/pcie/dpc.c | 25 +++++++++++++++++++++---- > >> drivers/pci/pcie/edr.c | 7 ++++--- > >> 3 files changed, 26 insertions(+), 8 deletions(-) > >> ... > >> -void dpc_process_error(struct pci_dev *pdev) > >> +/** > >> + * dpc_process_error - handle the DPC error status > >> + * @pdev: the port that experienced the containment event > >> + * > >> + * Return: the device that detected the error. > >> + * > >> + * NOTE: The device reference count is increased, the caller must > >> decrement > >> + * the reference count by calling pci_dev_put(). > >> + */ > >> +struct pci_dev *dpc_process_error(struct pci_dev *pdev) > > > > Maybe it makes sense to carry the err_port naming for the pci_dev > > in here as well? Seems stronger than just relying on people > > reading the documentation you've added. > > Good point. I think renaming the parameter would improve clarity. However, > I'd prefer to handle it in a separate patch to keep this change focused on > the functional modification. Would that work for you? > Sure. Ideal would be a precursor patch, but if it's much easier to do on top of this I'm fine with that. You are absolutely correct that it should be a separate patch! > > > >> { > >> u16 cap = pdev->dpc_cap, status, source, reason, ext_reason; > >> struct aer_err_info info = {}; > >> + struct pci_dev *err_dev; > >> > >> pci_read_config_word(pdev, cap + PCI_EXP_DPC_STATUS, &status); > >> > >> @@ -279,6 +289,7 @@ void dpc_process_error(struct pci_dev *pdev) > >> pci_aer_clear_nonfatal_status(pdev); > >> pci_aer_clear_fatal_status(pdev); > >> } > >> + err_dev = pci_dev_get(pdev); > >> break; > >> case PCI_EXP_DPC_STATUS_TRIGGER_RSN_NFE: > >> case PCI_EXP_DPC_STATUS_TRIGGER_RSN_FE: > >> @@ -290,6 +301,8 @@ void dpc_process_error(struct pci_dev *pdev) > >> "ERR_FATAL" : "ERR_NONFATAL", > >> pci_domain_nr(pdev->bus), PCI_BUS_NUM(source), > >> PCI_SLOT(source), PCI_FUNC(source)); > >> + err_dev = pci_get_domain_bus_and_slot(pci_domain_nr(pdev->bus), > >> + PCI_BUS_NUM(source), source & > >> 0xff); > > > > Bunch of replication in her with the pci_warn(). Maybe some local variables? > > Maybe even rebuild the final parameter from PCI_DEVFN(slot, func) just to > > make the > > association with the print really obvious? > > Agreed. Here's the improved version: > > --- a/drivers/pci/pcie/dpc.c > +++ b/drivers/pci/pcie/dpc.c > @@ -293,17 +293,28 @@ struct pci_dev *dpc_process_error(struct pci_dev *pdev) > break; > case PCI_EXP_DPC_STATUS_TRIGGER_RSN_NFE: > case PCI_EXP_DPC_STATUS_TRIGGER_RSN_FE: > - pci_read_config_word(pdev, cap + PCI_EXP_DPC_SOURCE_ID, > - &source); > + { > + int domain, bus; > + u8 slot, func, devfn; > + const char *err_type; > + > + pci_read_config_word(pdev, cap + PCI_EXP_DPC_SOURCE_ID, > &source); > + > + /* Extract source device location */ > + domain = pci_domain_nr(pdev->bus); > + bus = PCI_BUS_NUM(source); > + slot = PCI_SLOT(source); > + func = PCI_FUNC(source); > + devfn = PCI_DEVFN(slot, func); > + > + err_type = (reason == PCI_EXP_DPC_STATUS_TRIGGER_RSN_FE) ? > + "ERR_FATAL" : "ERR_NONFATAL"; > + > pci_warn(pdev, "containment event, status:%#06x, %s received > from %04x:%02x:%02x.%d\n", > - status, > - (reason == PCI_EXP_DPC_STATUS_TRIGGER_RSN_FE) ? > - "ERR_FATAL" : "ERR_NONFATAL", > - pci_domain_nr(pdev->bus), PCI_BUS_NUM(source), > - PCI_SLOT(source), PCI_FUNC(source)); > - err_dev = > pci_get_domain_bus_and_slot(pci_domain_nr(pdev->bus), > - PCI_BUS_NUM(source), source & > 0xff); > + status, err_type, domain, bus, slot, func); > + err_dev = pci_get_domain_bus_and_slot(domain, bus, devfn); Maybe don't bother with local variables for the things only used once. e.g. err_dev = pci_get_domain_bus_and_slot(domain, bus, PCI_DEVFN(slot, func)); and possibly the same for err_type. I don't mind though if you prefer to use locals for everything. > break; > + } > case PCI_EXP_DPC_STATUS_TRIGGER_RSN_IN_EXT: > ext_reason = status & PCI_EXP_DPC_STATUS_TRIGGER_RSN_EXT; > pci_warn(pdev, "containment event, status:%#06x: %s > detected\n", > > > > > Is there any chance that this might return NULL? Feels like maybe that's > > only a possibility on a broken setup, but I'm not sure of all the wonderful > > races around hotplug and DPC occurring before the OS has caught up. > > Surprise Down events are handled separately in > dpc_handle_surprise_removal() and won't reach dpc_process_error(). > Please correct me if I missed anything. Probably fine. I just didn't check particularly closely. Jonathan > > Thanks for valuable comments. > > Best Regards, > Shuai
