Re: avoiding pci_disable_device()...

[Jeff Garzik]

Greg KH wrote:
> On Sun, Feb 13, 2005 at 08:42:55PM -0500, Jeff Garzik wrote:
>
> > Currently, in almost every PCI driver, if pci_request_regions() fails -- 
> > indicating another driver is using the hardware -- then 
> > pci_disable_device() is called on the error path, disabling a device 
> > that another driver is using
> >
> > To call this "rather rude" is an understatement :)
> >
> > Fortunately, the ugliness is mitigated in large part by the PCI layer 
> > helping to make sure that no two drivers bind to the same PCI device. 
> > Thus, in the vast majority of cases, pci_request_regions() -should- be 
> > guaranteed to succeed.
> >
> > However, there are oddball cases like mixed PCI/ISA devices (hello IDE) 
> > or cases where a driver refers a pci_dev other than the primary, where 
> > pci_request_regions() and request_regions() still matter.
>
>
> But this is a very small subset of pci devices, correct?

No.  You also need to consider situations such as out-of-tree drivers 
for the same hardware (might not use PCI API), and situations where you 
have peer devices discovered and used (PCI API doesn't have "hey, <this> 
device is associated with <current driver>, too" capability)


> > As a result, I have committed the attached patch to libata-2.6.  In many 
> > cases, it is a "semantic fix", addressing the case
> >
> > 	* pci_request_regions() indicates hardware is in use
> > 	* we rudely disable the in-use hardware
> >
> > that would not occur in practice.
> >
> > But better safe than sorry.  Code cuts cut-n-pasted all over the place.
> >
> > I'm hoping one or two things will happen now:
> > * janitors fix up the other PCI drivers along these lines
> > * improve the PCI API so that pci_request_regions() is axiomatic
>
>
> Do you have any suggestions for how to do this?

I'm glad you asked ;-)  As the author of pci_disable_device() and 
pci_request_regions(), I recognized their inadequacy almost immediately.

There are some fundamental flaws in the API that need correcting:

* pci_disable_device() should perform exactly the opposite of 
pci_enable_device(), no more, no less.   It should NOT unconditionally 
disable the device, but instead restore the hardware to the state it was 
in prior to pci_enable_device().

* pci_request_regions() should be axiomatic.  By that I mean, 
pci_enable_device() should
	(a) handle pci_request_regions() completely
	(b) fail if regions are not available

* pci_enable_device() may touch the hardware when it should not.  In an 
ideal world, pci_enable_device() would
	* assign resources to device, if necessary
	* request_resource()s [aka pci_request_regions()]
	* enable device by setting bits in PCI_COMMAND, etc.
but since the request-resource step is assumed to occur after 
pci_enable_device() returns to the driver, this is impossible.


The solution?  I am still thinking.  My gut feeling is that we want a 
slightly higher level PCI API for drivers.  Drivers pass in an 'info' 
structure to pci_up().  pci_up() enables the device, requests resources 
(not just irq), maps resources as necessary, enables irqs and/or MSI as 
necessary, and similar housekeeping.  pci_down() does the precise 
opposite. Essentially, pci_up() is a lib function that kills a ton of 
duplicate code from the vast majority of PCI drivers.


OTOH, Alan's suggestion seems sane and a lot more simple, but doesn't 
address the flaws in the API.

	Jeff


-
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/