Re: [PATCH v15 3/9] PCI/AER: Handle ERR_FATAL with removal and re-enumeration of devices

[Bjorn Helgaas]

On Wed, May 09, 2018 at 06:44:53PM +0530, poza@xxxxxxxxxxxxxx wrote:
> On 2018-05-09 18:37, Bjorn Helgaas wrote:
> > On Tue, May 08, 2018 at 06:53:30PM -0500, Bjorn Helgaas wrote:
> > > On Thu, May 03, 2018 at 01:03:52AM -0400, Oza Pawandeep wrote:
> > > > This patch alters the behavior of handling of ERR_FATAL, where removal
> > > > of devices is initiated, followed by reset link, followed by
> > > > re-enumeration.
> > > >
> > > > So the errors are handled in a different way as follows:
> > > > ERR_NONFATAL => call driver recovery entry points
> > > > ERR_FATAL    => remove and re-enumerate
> > > >
> > > > please refer to Documentation/PCI/pci-error-recovery.txt for more details.
> > > >
> > > > Signed-off-by: Oza Pawandeep <poza@xxxxxxxxxxxxxx>
> > > >
> > > > diff --git a/drivers/pci/pcie/aer/aerdrv.c b/drivers/pci/pcie/aer/aerdrv.c
> > > > index 779b387..206f590 100644
> > > > --- a/drivers/pci/pcie/aer/aerdrv.c
> > > > +++ b/drivers/pci/pcie/aer/aerdrv.c
> > > > @@ -330,6 +330,13 @@ static pci_ers_result_t aer_root_reset(struct pci_dev *dev)
> > > >         reg32 |= ROOT_PORT_INTR_ON_MESG_MASK;
> > > >         pci_write_config_dword(dev, pos + PCI_ERR_ROOT_COMMAND, reg32);
> > > >
> > > > +       /*
> > > > +        * This function is called only on ERR_FATAL now, and since
> > > > +        * the pci_report_resume is called only in ERR_NONFATAL case,
> > > > +        * the clearing part has to be taken care here.
> > > > +        */
> > > > +       aer_error_resume(dev);
> > > 
> > > I don't understand this part.  Previously the ERR_FATAL path looked
> > > like
> > > this:
> > > 
> > >   do_recovery
> > >     reset_link
> > >       driver->reset_link
> > >         aer_root_reset
> > >           pci_reset_bridge_secondary_bus                # <-- reset
> > >     broadcast_error_message(..., report_resume)
> > >       pci_walk_bus(..., report_resume, ...)
> > >         report_resume
> > >       if (cb == report_resume)
> > >         pci_cleanup_aer_uncorrect_error_status
> > >           pci_write_config_dword(PCI_ERR_UNCOR_STATUS)  # <-- clear
> > > status
> > > 
> > > After this patch, it will look like this:
> > > 
> > >   do_recovery
> > >     do_fatal_recovery
> > >       pci_cleanup_aer_uncorrect_error_status
> > >         pci_write_config_dword(PCI_ERR_UNCOR_STATUS)    # <-- clear
> > > status
> > >       reset_link
> > >         driver->reset_link
> > >           aer_root_reset
> > >             pci_reset_bridge_secondary_bus              # <-- reset
> > >             aer_error_resume
> > >               pcie_capability_write_word(PCI_EXP_DEVSTA)        #
> > > <-- clear more
> > >               pci_write_config_dword(PCI_ERR_UNCOR_STATUS)      #
> > > <-- clear status
> > > 
> > > So if I'm understanding correctly, the new path clears the status too
> > > early, then clears it again (plus clearing DEVSTA, which we didn't do
> > > before) later.
> > > 
> > > I would think we would want to leave aer_root_reset() alone, and
> > > just move
> > > the pci_cleanup_aer_uncorrect_error_status() in do_fatal_recovery()
> > > down so
> > > it happens after we call reset_link().  That way the reset/clear
> > > sequence
> > > would be the same as it was before.
> > 
> > I've been fiddling with this a bit myself and will post the results to
> > see
> > what you think.
> 
> 
> ok so you are suggesting to move pci_cleanup_aer_uncorrect_error_status down
> which I can do.
> 
> And not to call aer_error_resume, because you think its clearing the status
> again.
> 
> following code: calls aer_error_resume.
> pci_broadcast_error_message()
>  /*
>                  * If the error is reported by an end point, we think this
>                  * error is related to the upstream link of the end point.
>                  */
>                 if (state == pci_channel_io_normal)
>                         /*
>                          * the error is non fatal so the bus is ok, just
> invoke
>                          * the callback for the function that logged the
> error.
>                          */
>                         cb(dev, &result_data);
>                 else
>                         pci_walk_bus(dev->bus, cb, &result_data);

Holy crap, I thought this could not possibly get any more complicated,
but you're right; we do actually call aer_error_resume() today via an
extremely convoluted path:

  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)

aerdriver is the only port service driver that implements
.error_resume(), and aerdriver only binds to root ports.  I can't
really believe all these device_for_each_child()/resume_iter()
gyrations are necessary when this is AER code calling AER code.

Bjorn