pci error recovery procedure

Zhang, Yanmin yanmin_zhang at linux.intel.com
Wed Sep 6 11:26:56 EST 2006

On Wed, 2006-09-06 at 03:01, Linas Vepstas wrote:
> On Tue, Sep 05, 2006 at 10:32:08AM +0800, Zhang, Yanmin wrote:
> > Is it the exclusive reason to have multi-steps?
> I don't understand the question. A previous email explained the reason
> to have mutiple steps.
The question is against Ben's comments. Pls. don't delete his comments
in your reply. 

> > 1) Here link reset and hard reset are hardware operations, not the
> > link_reset and slot_reset callback in pci_error_handlers.
> I don't understand the comment.
I wanted to clarify that we need differentiate link/hard reset from
callback link_reset and slot_reset when discussing the API.

> > 2) Callback error_detected will notify drivers there is PCI errors. Drivers
> > shouldn't do any I/O in error_detected.
> It shouldn't matter. If it is truly important for a particular platform
> to make sure that there is no i/o, then the low-level i/o routines
> could be modified to drop any accidentally issued i/o on the floor.
> This doesn't require a change to either the API or the policy.
> > 3) If both the link and slot are reset after all error_detected are called,
> > the device should go back to initial status and all DMA should be stopped
> > automatically. Why does the driver still need a chance to stop DMA? 
> As explained previously, not all drivers may want to have a full
> electrical device reset.
I need repeat my idea.
1) My new pci_error_handlers doesn't always choose to reset slot. It
still depends on the return value of error_detected.
2) As a matter of fact, most cases of specific device's error_detected callback
will choose to return PCI_ERS_RESULT_NEED_RESET. Like what you did for

> > The
> > error_detected of the drivers in the latest kernel who support err handlers
> > always returns PCI_ERS_RESULT_NEED_RESET. They are typical examples.
> Just because the current drivers do it this way does not mean that this is
> the best way to do things.
If it's not the best way, why did you choose to reset slot for e1000/e100/ipr
error handlers? They are typical widely-used devices. To make it easier to
add error handlers?

>  A full reset is time-consuming. Some drivers
> may want to implement a faster and quicker reset.
> --linas

More information about the Linuxppc-dev mailing list