On Wed, May 16, 2018 at 08:28:39PM +0530, poza@xxxxxxxxxxxxxx wrote:
> On 2018-05-16 18:34, Bjorn Helgaas wrote:
> > On Wed, May 16, 2018 at 05:45:58PM +0530, poza@xxxxxxxxxxxxxx wrote:
> > > On 2018-05-16 16:22, Bjorn Helgaas wrote:
> > > > On Wed, May 16, 2018 at 01:46:25PM +0530, poza@xxxxxxxxxxxxxx wrote:
> > > > > I am sorry I pasted the wrong snippet.
> > > > > following needs to be fixed in v17.
> > > > > from:
> > > > > if (dev->hdr_type == PCI_HEADER_TYPE_BRIDGE) {
> > > > > /*
> > > > > * If the error is reported by a bridge, we think
> > > > > this error
> > > > > * is related to the downstream link of the bridge,
> > > > > so we
> > > > > * do error recovery on all subordinates of the bridge
> > > > > instead
> > > > > * of the bridge and clear the error status of the
> > > > > bridge.
> > > > > */
> > > > > pci_walk_bus(dev->subordinate, report_resume,
> > > > > &result_data);
> > > > > pci_cleanup_aer_uncorrect_error_status(dev);
> > > > > }
> > > > >
> > > > >
> > > > > to
> > > > >
> > > > > if (service==AER && dev->hdr_type == PCI_HEADER_TYPE_BRIDGE) {
> > > > > /*
> > > > > * If the error is reported by a bridge, we think
> > > > > this error
> > > > > * is related to the downstream link of the bridge,
> > > > > so we
> > > > > * do error recovery on all subordinates of the bridge
> > > > > instead
> > > > > * of the bridge and clear the error status of the
> > > > > bridge.
> > > > > */
> > > > > pci_walk_bus(dev->subordinate, report_resume,
> > > > > &result_data);
> > > > > pci_cleanup_aer_uncorrect_error_status(dev);
> > > > > }
> > > > >
> > > > > this is only needed in case of AER.
> > > >
> > > > Oh, I missed this before. It makes sense to clear the AER status
> > > > here, but why do we need to call report_resume()? We just called all
> > > > the driver .remove() methods and detached the drivers from the
> > > > devices. So I don't think report_resume() will do anything
> > > > ("dev->driver" should be NULL) except set the dev->error_state to
> > > > pci_channel_io_normal. We probably don't need that because we didn't
> > > > change error_state in this fatal error path.
> > >
> > > if you remember, the path ends up calling
> > > aer_error_resume
> > >
> > > the existing code ends up calling aer_error_resume as follows.
> > >
> > > do_recovery(pci_dev)
> > > broadcast_error_message(..., error_detected, ...)
> > > if (AER_FATAL)
> > > reset_link(pci_dev)
> > > udev = BRIDGE ? pci_dev : pci_dev->bus->self
> > > driver->reset_link(udev)
> > > aer_root_reset(udev)
> > > if (CAN_RECOVER)
> > > broadcast_error_message(..., mmio_enabled, ...)
> > > if (NEED_RESET)
> > > broadcast_error_message(..., slot_reset, ...)
> > > broadcast_error_message(dev, ..., report_resume, ...)
> > > if (BRIDGE)
> > > report_resume
> > > driver->resume
> > > pcie_portdrv_err_resume
> > > device_for_each_child(..., resume_iter)
> > > resume_iter
> > > driver->error_resume
> > > aer_error_resume
> > > pci_cleanup_aer_uncorrect_error_status(pci_dev) # only
> > > if
> > > BRIDGE
> > > pci_write_config_dword(PCI_ERR_UNCOR_STATUS)
> > >
> > > hence I think we have to call it in order to clear the root port
> > > PCI_ERR_UNCOR_STATUS and PCI_EXP_DEVSTA.
> > > makes sense ?
> >
> > I know I sent you the call graph above, so you would think I might
> > understand it, but you would be mistaken :) This still doesn't make
> > sense to me.
> >
> > I think your point is that we need to call aer_error_resume(). That
> > is the aerdriver.error_resume() method. The AER driver only binds to
> > root ports.
> >
> > This path:
> >
> > pcie_do_fatal_recovery
> > pci_walk_bus(dev->subordinate, report_resume, &result_data)
> >
> > calls report_resume() for every device on the dev->subordinate bus
> > (and for anything below those devices). There are no root ports on
> > that dev->subordinate bus, because root ports are always on a root
> > bus, never on a subordinate bus.
> >
> > So I don't see how report_resume() can ever get to aer_error_resume().
> > Can you instrument that path and verify that it actually does get
> > there somehow?
>
> you are right....the call
> pci_walk_bus(dev->subordinate, report_resume, &result_data);
> does not call aer_error_resume()
>
> but
> pci_walk_bus(udev->bus, report_resume, &result_data);
> does call aer_error_resume()
>
> now if you look at the comment of the function:
> /**
> * aer_error_resume - clean up corresponding error status bits
> * @dev: pointer to Root Port's pci_dev data structure
> *
> * Invoked by Port Bus driver during nonfatal recovery.
> */
>
> it is invoked during nonfatal recovery.
> but the code handles both fatal and nonfatal clearing of error bits.
>
> if (dev->error_state == pci_channel_io_normal)
> status &= ~mask; /* Clear corresponding nonfatal bits */
> else
> status &= mask; /* Clear corresponding fatal bits */
> pci_write_config_dword(dev, pos + PCI_ERR_UNCOR_STATUS, status);
>
>
> so the question is, should we not call aer_error_resume during fatal
> recovery ?
> so that it clears the root port status, if of course the error is triggered
> by AER running agent (RP, Switch)
I'm sure we *should* clear AER status bits somewhere during ERR_FATAL
recovery.
As far as I can tell, the current code (before your patches) never
calls aer_error_resume(). That might be a bug, but even if it is,
it's something that should be fixed separately from *this* series.
I think in this series, you should probably adjust the patch that adds
do_fatal_recovery() so it doesn't call pci_walk_bus(..., report_resume).
I don't think that does anything useful anyway, and that patch already
changes AER so it doesn't call the pci_error_handlers callbacks
(except .resume()). I think it would be cleaner to remove the
ERR_FATAL use of .resume() at the same time you remove the others.
