Re: [ PATCH RESEND ] PCI-AER: Do not report successful error recoveryfor devices with AER-unaware drivers

[Bjorn Helgaas]

On Thu, Nov 15, 2012 at 12:09 AM, Pandarathil, Vijaymohan R
<vijaymohan.pandarathil@xxxxxx> wrote:
> Thanks for the comments. See my response below.
>
>> -----Original Message-----
>> From: Bjorn Helgaas [mailto:bhelgaas@xxxxxxxxxx]
>> Sent: Wednesday, November 14, 2012 4:51 PM
>> To: Pandarathil, Vijaymohan R
>> Cc: linux-kernel@xxxxxxxxxxxxxxx; linux-pci@xxxxxxxxxxxxxxx;
>> linasvepstas@xxxxxxxxx; Myron Stowe; Ortiz, Lance E; Huang Ying; Hidetoshi
>> Seto; Patterson, Andrew D (LeftHand Networks); Zhang Yanmin
>> Subject: Re: [ PATCH RESEND ] PCI-AER: Do not report successful error
>> recovery for devices with AER-unaware drivers
>>
>> [+cc Lance, Huang, Hidetoshi, Andrew, Zhang]
>>
>> On Sat, Nov 10, 2012 at 07:41:04AM +0000, Pandarathil, Vijaymohan R wrote:
>> > When an error is detected on a PCIe device which does not have an
>> > AER-aware driver, prevent AER infrastructure from reporting
>> > successful error recovery.
>> >
>> > This is because the report_error_detected() function that gets
>> > called in the first phase of recovery process allows forward
>> > progress even when the driver for the device does not have AER
>> > capabilities. It seems that all callbacks (in pci_error_handlers
>> > structure) registered by drivers that gets called during error
>> > recovery are not mandatory. So the intention of the infrastructure
>> > design seems to be to allow forward progress even when a specific
>> > callback has not been registered by a driver. However, if error
>> > handler structure itself has not been registered, it doesn't make
>> > sense to allow forward progress.
>> >
>> > As a result of the current design, in the case of a single device
>> > having an AER-unaware driver or in the case of any function in a
>> > multi-function card having an AER-unaware driver, a successful
>> > recovery is reported.
>> >
>> > Typical scenario this happens is when a PCI device is detached
>> > from a KVM host and the pci-stub driver on the host claims the
>> > device. The pci-stub driver does not have error handling capabilities
>> > but the AER infrastructure still reports that the device recovered
>> > successfully.
>> >
>> > The changes proposed here leaves the device in an unrecovered state
>> > if the driver for the device or for any function in a multi-function
>> > card does not have error handler structure registered. This reflects
>> > the true state of the device and prevents any partial recovery (or no
>> > recovery at all) reported as successful.
>> >
>> > Please also see comments from Linas Vepstas at the following link
>> > http://www.spinics.net/lists/linux-pci/msg18572.html
>> >
>> > Reviewed-by: Linas Vepstas <linasvepstas <at> gmail.com>
>> > Reviewed-by: Myron Stowe <mstowe <at> redhat.com>
>> > Signed-off-by: Vijay Mohan Pandarathil <vijaymohan.pandarathil <at>
>> hp.com>
>> >
>> > ---
>> >
>> > drivers/pci/pcie/aer/aerdrv_core.c | 6 ++++++
>> >  1 file changed, 6 insertions(+)
>> >
>> > diff --git a/drivers/pci/pcie/aer/aerdrv_core.c
>> b/drivers/pci/pcie/aer/aerdrv_core.c
>> > index 06bad96..030b229 100644
>> > --- a/drivers/pci/pcie/aer/aerdrv_core.c
>> > +++ b/drivers/pci/pcie/aer/aerdrv_core.c
>> > @@ -215,6 +215,12 @@ static int report_error_detected(struct pci_dev
>> *dev, void *data)
>> >
>> >     dev->error_state = result_data->state;
>> >
>> > +   if ((!dev->driver || !dev->driver->err_handler) &&
>> > +           !(dev->hdr_type & PCI_HEADER_TYPE_BRIDGE)) {
>> > +           dev_info(&dev->dev, "AER: Error detected but no driver has
>> claimed this device or the driver is AER-unaware\n");
>> > +           result_data->result = PCI_ERS_RESULT_NONE;
>> > +           return 1;
>>
>> This doesn't seem right because returning 1 here causes pci_walk_bus()
>> to terminate, which means we won't set dev->error_state or notify
>> drivers for any devices we haven't visited yet.
>>
>> > +   }
>> >     if (!dev->driver ||
>> >             !dev->driver->err_handler ||
>> >             !dev->driver->err_handler->error_detected) {
>>
>> If none of the drivers in the affected hierarchy supports error handling,
>> I think the call tree looks like this:
>>
>>     do_recovery                                 # uncorrectable only
>>         broadcast_error_message(dev, ..., report_error_detected)
>>             result_data.result = CAN_RECOVER
>>             pci_walk_bus(..., report_error_detected)
>>                 report_error_detected           # (each dev in subtree)
>>                     return 0                    # no change to result
>>             return result_data.result
>>         broadcast_error_message(dev, ..., report_mmio_enabled)
>>             result_data.result = PCI_ERS_RESULT_RECOVERED
>>             pci_walk_bus(..., report_mmio_enabled)
>>                 report_mmio_enabled             # (each dev in subtree)
>>                     return 0                    # no change to result
>>         dev_info("recovery successful")
>>
>> Specifically, there are no error_handler functions, so we never change
>> result_data.result, and the default is that we treat the error as
>> "recovered successfully."  That seems broken.  An uncorrectable error
>> is by definition recoverable only by device-specific software, i.e.,
>> the driver.  We didn't call any drivers, so we can't have recovered
>> anything.
>>
>> What do you think of the following alternative?  I don't know why you
>> checked for bridge devices in your patch, so I don't know whether
>> that's important here or not.
>>
>> diff --git a/drivers/pci/pcie/aer/aerdrv_core.c
>> b/drivers/pci/pcie/aer/aerdrv_core.c
>> index 06bad96..a109c68 100644
>> --- a/drivers/pci/pcie/aer/aerdrv_core.c
>> +++ b/drivers/pci/pcie/aer/aerdrv_core.c
>> @@ -231,11 +231,11 @@ static int report_error_detected(struct pci_dev *dev,
>> void *data)
>>                                  dev->driver ?
>>                                  "no AER-aware driver" : "no driver");
>>               }
>> -             return 0;
>> +             vote = PCI_ERS_RESULT_DISCONNECT;
>> +     } else {
>> +             err_handler = dev->driver->err_handler;
>> +             vote = err_handler->error_detected(dev, result_data->state);
>>       }
>> -
>> -     err_handler = dev->driver->err_handler;
>> -     vote = err_handler->error_detected(dev, result_data->state);
>>       result_data->result = merge_result(result_data->result, vote);
>>       return 0;
>>  }
>
> This would definitely set the error_state for all devices correctly. However, with the
> current implementation of merge_result(), won't we still end up reporting successful
> recovery ? The following case statement in merge_result() can set back the result
> from PCI_ERS_RESULT_DISCONNECT to PCI_ERS_RESULT_NEED_RESET for a subsequent device
> on the bus which returned PCI_ERS_RESULT_NEED_RESET from its error_detected callback .
>
> merge_result()
> ...
>         case PCI_ERS_RESULT_DISCONNECT:
>                 if (new == PCI_ERS_RESULT_NEED_RESET)
>                         orig = new;
>                 break;
>
> This would mean do_recovery() proceeds along to the next broadcast_message and
> ultimately report success. Right ? Let me know if I am missing something.

Yes, I think you're right.  I only looked at the case where none of
the devices in the subtree had drivers.

> I looked at a few options and the following looked more promising. Thoughts ?
>
> Introduce a new pci_ers_result enum PCI_ERS_RESULT_NO_AER_DRIVER and make changes as follows.

I don't really like this new enum because it's not really obvious how
it's different from PCI_ERS_RESULT_NONE and, more importantly, it
makes merge_result() even more of a kludge than it already is.

It's hard to write a nice simple description of this algorithm
(visiting all the devices in a subtree, collecting error handler
results, and merging them).  It would be a lot easier to describe if
merge_result() could be written simply as "max()", but I'm not sure
the pci_ers_result codes could be ordered in a way that would make
that possible.  And the desired semantics might make it impossible,
too.

I think the intent of your patch is that if there's any device in the
subtree that lacks an .error_detected() method, we do not call
.mmio_enabled() or .slot_reset() or .resume() for *any* device in the
subtree.  Right?

So maybe this is the best we can do, and it certainly seems better
than what we have now.  Can you repost this as a fresh v2 patch?

> diff --git a/drivers/pci/pcie/aer/aerdrv.h b/drivers/pci/pcie/aer/aerdrv.h
> index 94a7598..149ba79 100644
> --- a/drivers/pci/pcie/aer/aerdrv.h
> +++ b/drivers/pci/pcie/aer/aerdrv.h
> @@ -87,6 +87,9 @@ struct aer_broadcast_data {
>  static inline pci_ers_result_t merge_result(enum pci_ers_result orig,
>                 enum pci_ers_result new)
>  {
> +       if (new == PCI_ERS_RESULT_NO_AER_DRIVER)
> +               return new;

BTW, if you keep this, please just use "return
PCI_ERS_RESULT_NO_AER_DRIVER" rather than "return new" since we *know*
what we're returning.  I think there's another instance of this in
merge_result() that you could fix, too.

> +
>         if (new == PCI_ERS_RESULT_NONE)
>                 return orig;
>
> diff --git a/drivers/pci/pcie/aer/aerdrv_core.c b/drivers/pci/pcie/aer/aerdrv_core.c
> index 06bad96..729580a 100644
> --- a/drivers/pci/pcie/aer/aerdrv_core.c
> +++ b/drivers/pci/pcie/aer/aerdrv_core.c
> @@ -231,11 +231,12 @@ static int report_error_detected(struct pci_dev *dev, void *data)
>                                    dev->driver ?
>                                    "no AER-aware driver" : "no driver");
>                 }
> -               return 0;
> +               vote = PCI_ERS_RESULT_NO_AER_DRIVER;
> +       } else {
> +               err_handler = dev->driver->err_handler;
> +               vote = err_handler->error_detected(dev, result_data->state);
>         }
>
> -       err_handler = dev->driver->err_handler;
> -       vote = err_handler->error_detected(dev, result_data->state);
>         result_data->result = merge_result(result_data->result, vote);
>         return 0;
>  }
> diff --git a/include/linux/pci.h b/include/linux/pci.h
> index ee21795..fb7e869 100644
> --- a/include/linux/pci.h
> +++ b/include/linux/pci.h
> @@ -538,6 +538,9 @@ enum pci_ers_result {
>
>         /* Device driver is fully recovered and operational */
>         PCI_ERS_RESULT_RECOVERED = (__force pci_ers_result_t) 5,
> +
> +       /* No AER capabilities registered for the driver */
> +       PCI_ERS_RESULT_NO_AER_DRIVER = (__force pci_ers_result_t) 6,
>  };
>
>  /* PCI bus error event callbacks */
>
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/