Re: [PATCH] HPET driver

[Jeff Garzik <jgarzik@pobox.com>]

Andrew Morton wrote:
> Jeff Garzik <jgarzik@pobox.com> wrote:
> 
>>>	[PATCH] HPET driver
>>>	
>>
>>Was this posted on lkml, or simply snuck in?
> 
> 
> Was posted on lkml, was fairly widely reviewed, had comments from hch and
> others, had several fixes from myself and from Robert and a long discussion
> wrt the readq() implementation.

I'm surprised it was reviewed, but I apologize for my harsh words in any 
case.


> wrt the readq() implementation: I reverted the generic implementation based
> on concerns raised on lkml by Eric Biederman.  As a generic readq/writeq
> implementation seems to be a new R&D project I decided to leave the
> implementation private to the HPET driver until someone takes all of this
> on.

All the readq/writeq users at the moment don't give a crap about atomicity.


> wrt the hpets list locking: yeah, I noticed that, mentioned it to Robert
> wrt the request_irq() bug: yipes.  Robert, please fix.
> wrt the new miscdev minor: yes, devices.txt should be updated.  When the

And:

1) a merge issue, we shouldn't be merging new procfs stuff

2) build breaks if CONFIG_ACPI is not set, but this driver is selected

3) shared interrupt causes very incorrect behavior, look at the last few 
lines of hpet_interrupt().

4) return EINVAL in open(2) if FMODE_WRITE isn't set.  Yes, vfs_write() 
will return EINVAL if you actually attempt to write(2), but other areas 
of the kernel check FMODE_WRITE.  I consider this a security bug, if the 
driver does not support writing, but does not prevent FMODE_WRITE from 
being set.  I do not know for sure, but I strongly suspect you can use 
this to cause incorrect behavior _somewhere_.

5) use of "__set_current_state" _and_ "current->state ="

6) race:
	spin-lock
	set HPET_IE
	spin-unlock

	doh!  we shouldn't have set HPET_IE

	spin-lock
	clear HPET_IE
	spin-unlock