Re: patch for kernel 2.6.10-rc2

[Flavio Stanchina <flavio at stanchina dot net>]

Jan De Luyck wrote:
> Upgraded today to 2.6.10-rc2, the fglrx driver compiles but fails to insert 
> because of missing pci_find_class() function. This has been superseded by 
> pci_get_class(), changing the calls makes it work again.

Alexandru Fomin did the same patch, but I peeked at 2.6.10-rc1's 
drivers/pci/search.c on www.kernel.org and I'm not sure it's correct to 
just replace the old function with the new one: the comment on 
pci_get_class() says that "the reference count to the device is 
incremented" so I guess that calling it two times is not good. I'd 
expect that we need to store the device pointer returned in pass 1 and 
reuse it in pass 3.

Alex answered my concerns like this:

"I looked into it a little last night and I think we should decrement 
the reference count by running pci_dev_put() after we finish useing the 
device, or if it doesn't have the capabilities that we need. I didn't 
understood yet verry well for what the refcount is used, but from 
logical point of view, it's not right to let the kernel think a device 
has more users then it really has."

"Personally, because I couldn't find the devices returned by 
pci_get_class() reused in any other function, I would be tempted to run 
pci_dev_put() as the last command of the while loops. I'm runing now a 
version that just implements this, and I don't have any bugs yet, but I 
want to test/learn more until I publish any patch."

Until we can get a firm answer on this (any kernel gurus on the list?) 
I'm not including this patch in my sources. Also, AFAIK your patch would 
not work on 2.6.9 because pci_get_class() doesn't exist before 
2.6.10-rc1. Note however that the patch from Alex is included in a 
subdirectory called "limbo" and it conditionally uses pci_get_class() or 
pci_find_class() so it should compile correctly. I'll let you decide if 
it also runs correctly. :)

--
Ciao, Flavio