Re: [RFC] pci: Rework config space blocking services

[Jan Kiszka <jan.kiszka@siemens.com>]

On 2011-09-06 10:04, Michael S. Tsirkin wrote:
> On Tue, Sep 06, 2011 at 09:18:13AM +0200, Jan Kiszka wrote:
>>>> @@ -401,36 +403,58 @@ int pci_vpd_truncate(struct pci_dev *dev, size_t size)
>>>>  EXPORT_SYMBOL(pci_vpd_truncate);
>>>>
>>>>  /**
>>>> - * pci_block_user_cfg_access - Block userspace PCI config reads/writes
>>>> + * pci_block_cfg_access - Block PCI config reads/writes
>>>
>>> This comment seems confusing. We don't in fact block all config
>>> reads writes. Instead we block userspace accesses and
>>> concurrent block requests.
>>
>> I'm open for a better suggestion that summarize the more verbose (and
>> hopefully clearer) explanation below.
> 
> I think the problem is, it doesn't block config access
> and we call it pci_block_cfg_access.
> 
> Thinking about it, doesn't this behave somewhat like a lock?
> How about
> 
> pci_user_cfg_access_trylock
> pci_user_cfg_access_lock
> pci_user_cfg_access_unlock
> 
> And then:
>  * pci_user_cfg_access_lock - Lock userspace PCI config access

Except that the "userspace" here is still only half of the truth and I
prefer to drop it, the naming locks good to me.

>  *
>  * When locked, any userspace reads or writes to config space
>  * and concurrent lock requests will sleep, and trylock requests
>  * will fail, until pci_user_cfg_access_unlock is called.
> 
> I had a brief thought of using an rwsem internally, but
> this would make trylock fail if userspace does config read,
> changing semantics.

Also, I bet we would make lockdep unhappy when calling
down_write_trylock from IRQ context.

Jan

-- 
Siemens AG, Corporate Technology, CT T DE IT 1
Corporate Competence Center Embedded Linux