Re: [PATCH v2] PCIe AER: report uncorrectable errors only to the functions that logged the errors

[Bjorn Helgaas]

On Fri, Sep 01, 2017 at 11:39:35AM +0000, Gabriele Paoloni wrote:
> Hi Bjorn
> 
> Many thanks for looking at this
> 
> > -----Original Message-----
> > From: Bjorn Helgaas [mailto:helgaas@xxxxxxxxxx]
> > Sent: 01 September 2017 05:43
> > To: Gabriele Paoloni
> > Cc: Linuxarm; liudongdong (C); linux-pci@xxxxxxxxxxxxxxx; linux-
> > kernel@xxxxxxxxxxxxxxx
> > Subject: Re: [PATCH v2] PCIe AER: report uncorrectable errors only to
> > the functions that logged the errors
> > 
> > On Thu, Aug 31, 2017 at 03:03:44PM -0500, Bjorn Helgaas wrote:
> > > On Fri, Aug 18, 2017 at 12:02:21PM +0100, Gabriele Paoloni wrote:
> > > > Currently if an uncorrectable error is reported by an EP the AER
> > > > driver walks over all the devices connected to the upstream port
> > > > bus and in turns call the report_error_detected() callback.
> > > > If any of the devices connected to the bus does not implement
> > > > dev->driver->err_handler->error_detected() do_recovery() will fail
> > > > leaving all the bus hierarchy devices unrecovered.
> > > >
> > > > However for non fatal errors the PCIe link should not be considered
> > > > compromised, therefore it makes sense to report the error only to
> > > > all the functions that logged an error.
> > >
> > > Can you include a pointer to the relevant part of the spec here?
> 
> Sure
> According to section "6.2.2.2.2. Non-Fatal Errors" 
> << Non-fatal errors are uncorrectable errors which cause a particular
> transaction to be unreliable but the Link is otherwise fully functional.
> Isolating Non-fatal from Fatal errors provides Requester/Receiver logic
> in a device or system management software the opportunity to recover
> from the error without resetting the components on the Link and
> disturbing other transactions in progress. Devices not associated with
> the transaction in error are not impacted by the error.>>
> 
> I will add it to the commit msg:
> 
> > 
> > Also, I forgot to ask: can you outline the problem this fixes?  I'm
> > curious about why this hasn't been an issue in the past.  My guess is
> > there's something new about your configuration, and the config and the
> > symptoms might help connect this fix to similar problems.
> 
> I already replied about this in the previous patch...
> << In Hi1620 we have some integrated controllers that appear as PCIe EPs
> under the same bus. Some of these controllers (e.g. the SATA 
> controller) are missing the err_handler callbacks.
> 
> If one device reports a non-fatal uncorrectable error with the current
> AER core code the callbacks for all the devices under the same bus will
> be called and, if any of the devices is missing the callback all the
> devices in the subtree are left in error state without recovery... 
> This patch is needed to sort out a situation like this one.>>
> 
> Should I add this to the commit msg?

Thanks for the reminder.  I thought I remembered some details but
hadn't dug them out again.  Yes, I was hoping for something we could
include in the commit message.  I'm still not sure what specifically
is *new* about this situation.  Maybe the particular mix of functions
in a multi-function device?  Maybe the fact that you're seeing more
AER errors than before (or maybe just doing more testing?)

Since this is actually a bug fix, this might be a good opportunity to
open a bugzilla for it.  Then we have a place to attach the complete
"lspci -vv" output, dmesg, etc., that are of interest but too much for
the commit message.

> > > > This patch implements this new behaviour for non fatal errors.
> > > >
> > > > Signed-off-by: Gabriele Paoloni <gabriele.paoloni@xxxxxxxxxx>
> > > > Signed-off-by: Dongdong Liu <liudongdong3@xxxxxxxxxx>
> > > > ---
> > > > Changes from v1:
> > > >    - now errors are reported only to the fucntions that logged the
> > error
> > > >      instead of all the functions in the same device.
> > > >    - the patch subject has changed to match the new implementation
> > > > ---
> > > >  drivers/pci/pcie/aer/aerdrv_core.c | 9 ++++++++-
> > > >  1 file changed, 8 insertions(+), 1 deletion(-)
> > > >
> > > > diff --git a/drivers/pci/pcie/aer/aerdrv_core.c
> > b/drivers/pci/pcie/aer/aerdrv_core.c
> > > > index b1303b3..057465ad 100644
> > > > --- a/drivers/pci/pcie/aer/aerdrv_core.c
> > > > +++ b/drivers/pci/pcie/aer/aerdrv_core.c
> > > > @@ -390,7 +390,14 @@ static pci_ers_result_t
> > broadcast_error_message(struct pci_dev *dev,
> > > >  		 * If the error is reported by an end point, we think this
> > > >  		 * error is related to the upstream link of the end point.
> > > >  		 */
> > > > -		pci_walk_bus(dev->bus, cb, &result_data);
> > > > +		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);
> > >
> > > I think the concept of this change makes sense, but I don't like the
> > > implicit connection of PCI_ERR_ROOT_UNCOR_RCV -> AER_NONFATAL ->
> > > pci_channel_io_normal.  That makes it harder than it should be to
> > read
> > > the code.
> > >
> > > What would you think of changing the signature of do_recovery() and
> > > broadcast_error_message() so they take the struct aer_err_info
> > pointer
> > > instead of just the severity and pci_channel_state?  Then we could
> > > check directly for AER_NONFATAL here.
> 
> I think the main issue of this approach is 
> aer_recover_work_func()->do_recovery()
> If we modify the signature of do_recovery() we also need to modify
> struct aer_recover_entry to embed aer_err_info sub struct (and anyway
> so far there is no hook to aer_err_info in ghes.c...so it seems to
> me like unfeasible...)
> 
> I think we can either add a severity field to broadcast_error_message()
> or move 
> 
> 	enum pci_channel_state state;
> 
> 	if (severity == AER_FATAL)
> 		state = pci_channel_io_frozen;
> 	else
> 		state = pci_channel_io_normal;
> 
> inside broadcast_error_message()
> 
> In both cases we make the code more readable but we add redundancy...
> Thoughts?

Hmmm, it's not as simple to solve as I hoped.

The "native" Linux AER driver reads registers directly from the AER
capability, produces logs, and does some recovery.

>From 100 miles away, APEI is basically a way for the platform (i.e.,
firmware or manageability software) to insert itself in the middle:
*it* reads the AER registers, does whatever it wants to do (e.g.,
OS-independent logging), and then passes a copy of the AER capability
(and the PCIe capability for good measure) on to the OS via APEI.

The Linux APEI code then stuffs that AER info into the native Linux
AER path where we do our own recovery.

So we basically have two entry points to the Linux AER code: (1) the
get_device_error_info() path that reads the AER registers directly,
and (2) the APEI path that gets the AER register info from the
firmware.

In my mind, these paths ought to be far more unified than they are.
The transition from APEI to the native Linux AER code throws away
almost all the information we got from the platform:
aer_recover_work_func() in the APEI path gets a copy of the AER
capability, but all it passes to do_recovery() in the native AER code
is "severity" -- a single int.

I think we should revamp the native AER code so it collects the AER
register info a little higher up, maybe in aer_isr_one_error() (it's
kind of stupid that aer_process_err_devices() currently reads the AER
registers *twice*).

Then we could make the APEI code copy the information we care about
from the CPER into a struct aer_err_info, then stuff it into
aer_process_err_devices().  That would have the nice side-effect of
using exactly the same logging code for both paths (currently the APEI
path uses cper_print_aer(), while the native path uses
aer_print_error()).

This is a lot of work.

Bjorn