PCI Power management (was: Re: [PATCH 4/13]: PCI Err: e100 ethernet driver recovery

PCI Power management (was: Re: [PATCH 4/13]: PCI Err: e100 ethernet driver recovery

[Linas Vepstas]

On Wed, Jun 29, 2005 at 06:58:29PM +0200, Andi Kleen was heard to remark:
> > Yep, OK. Pushig the timer would in fact break if the device was marked
> > perm disabled.
> 
> I think for network drivers you should just write a generic error handler
> (perhaps in net/core/dev.c) that calls the watchdog handler. 
> Then all drivers could be easily converted without much code duplication.

Well, there's no watchdog per-se in "struct net_device" -- are you
suggesting I add one?

It looks like I can almost create generic handlers for net devices; 
looks like calling netdev->stop() is enough to handle the error
detection. 

However, a generic bringup would need to call pci_enable_device(), 
and net/core/dev.c does not include pci.h so I can't really do it 
there.  Other than that, a generic recovry routine looks like it might
be possible; I'll have to experiment; its hard to tell by reading code.

This might be the wrong paradigm, though.  The pci error recovery 
routines are *almost identical* to the power-management suspend/resume
routines.  From what I can tell, the only real difference is that 
I want to not actually turn off/on the power. 

Thus, the right thing to do might be to split up the 
struct pci_dev->suspend() and pci_dev->resume() calls into

   suspend()
   poweroff()
   poweron()
   resume()

and then have the generic pci error recovery routines call
suspend/resume only, skipping the poweroff-on calls.  Does that 
sound good?

I'm not sure I can pull this off without having someone from 
the power-management world throw a brick at me.

--linas

Re: PCI Power management (was: Re: [PATCH 4/13]: PCI Err: e100 ethernet driver recovery

[Linas Vepstas]

Hm,

Scratch the idea I outline below, seems like its not a good idea.

I'm reading the e100, e1000 and the ixgb power management code, and they
go through all sorts of steps I don't need to do for PCI device reset.
There's no clear abstraction that would serve both needs.

On Thu, Jun 30, 2005 at 03:39:31PM -0500, Linas Vepstas was heard to remark:
> On Wed, Jun 29, 2005 at 06:58:29PM +0200, Andi Kleen was heard to remark:
> > > Yep, OK. Pushig the timer would in fact break if the device was marked
> > > perm disabled.
> > 
> > I think for network drivers you should just write a generic error handler
> > (perhaps in net/core/dev.c) that calls the watchdog handler. 
> > Then all drivers could be easily converted without much code duplication.
> 
> Well, there's no watchdog per-se in "struct net_device" -- are you
> suggesting I add one?
> 
> It looks like I can almost create generic handlers for net devices; 
> looks like calling netdev->stop() is enough to handle the error
> detection. 
> 
> However, a generic bringup would need to call pci_enable_device(), 
> and net/core/dev.c does not include pci.h so I can't really do it 
> there.  Other than that, a generic recovry routine looks like it might
> be possible; I'll have to experiment; its hard to tell by reading code.
> 
> This might be the wrong paradigm, though.  The pci error recovery 
> routines are *almost identical* to the power-management suspend/resume
> routines.  From what I can tell, the only real difference is that 
> I want to not actually turn off/on the power. 
> 
> Thus, the right thing to do might be to split up the 
> struct pci_dev->suspend() and pci_dev->resume() calls into
> 
>    suspend()
>    poweroff()
>    poweron()
>    resume()
> 
> and then have the generic pci error recovery routines call
> suspend/resume only, skipping the poweroff-on calls.  Does that 
> sound good?
> 
> I'm not sure I can pull this off without having someone from 
> the power-management world throw a brick at me.
> 
> --linas
> 
> 

Re: PCI Power management (was: Re: [PATCH 4/13]: PCI Err: e100 ethernet driver recovery

[Benjamin Herrenschmidt]

On Thu, 2005-06-30 at 15:39 -0500, Linas Vepstas wrote:

> Thus, the right thing to do might be to split up the 
> struct pci_dev->suspend() and pci_dev->resume() calls into
> 
>    suspend()
>    poweroff()
>    poweron()
>    resume()

No. There are very good reasons not to do that split at the pci_dev
level.
 
> and then have the generic pci error recovery routines call
> suspend/resume only, skipping the poweroff-on calls.  Does that 
> sound good?
> 
> I'm not sure I can pull this off without having someone from 
> the power-management world throw a brick at me.

Just keep the error recovery callbacks for now, and we might be able to
provide a generic "helper" doing the watchdog thing (yes, there is a
watchdog in the net core)

Ben.