Re: On "ratelimit: Use per ratelimit context locking"

[Ingo Molnar <mingo@elte.hu>]

* Stefan Richter <stefanr@s5r6.in-berlin.de> wrote:

> Hi Ingo,
> 
> in pre 2.6.33-rc1 commit 979f693d you wrote: "I'd like to use
> printk_ratelimit() in atomic context, but that's not possible right now
> due to the spinlock usage this commit introduced more than a year ago:
> 717115e: printk ratelimiting rewrite".?,?
> 
> By not possible, do you mean it would be an outright bug to call
> printk_ratelimit in atomic context prior to 979f693d, or merely a
> performance issue?  If the former, under which circumstances would the
> bug hit?
> 
> I'm asking because one of my commits actually introduced a
> printk_ratelimit use in an interrupt handler shortly before 2.6.32 was
> released.  In my testing, it didn't occur to me that there might be a
> problem.
> 
> ?http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h=979f693def9084a452846365dfde5dcb28366333
> ?http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h=717115e1a5856b57af0f71e1df7149108294fc10

There used to be a global lock:

  static DEFINE_SPINLOCK(ratelimit_lock);

which is now per ratelimit context. The case i was after was to use the 
ratelimit state from NMI context. With NMIs we can lock up if an NMI hits when 
some other code uses the ratelimit code. It's a small but existing race 
window.

OTOH, IRQ context use of printk ratelimit was safe before (and after) this 
commit - so i think your code should be safe too.

Hope that helps,

	Ingo