Re: [BUG] long freezes on thinkpad t60

[Eric Dumazet <dada1@cosmosbay.com>]

Linus Torvalds a écrit :
> On Thu, 21 Jun 2007, Linus Torvalds wrote:
> > We don't do nesting locking either, for exactly the same reason. Are 
> > nesting locks "easier"? Absolutely. They are also almost always a sign of 
> > a *bug*. So making spinlocks and/or mutexes nest by default is just a way 
> > to encourage bad programming!
>
> Side note, and as a "truth in advertising" section: I'll have to admit 
> that I argued against fair semaphores on the same grounds. I was wrong 
> then (and eventually admitted it, and we obviously try to make our mutexes 
> and semaphores fair these days!), and maybe I'm wrong now.
>
> If somebody can actually come up with a sequence where we have spinlock 
> starvation, and it's not about an example of bad locking, and nobody 
> really can come up with any other way to fix it, we may eventually have to 
> add the notion of "fair spinlocks".

I tried to find such a sequence, but I think its more a matter of hardware 
evolution, and some degenerated cases.

In some years (months ?), it might possible to starve say the file struct 
spinlock of a process in a open()/close() infinite loop. This because the 
number of instruction per 'memory cache line transfert between cpus/core' is 
raising.

But then one can say its a bug in user code :)

Another way to starve kernel might be a loop doing settime() , since seqlock 
are quite special in serialization :

Only seqlock's writers perform atomic ops, readers could be starved because of 
some hardware 'optimization'.


> So my arguments are purely pragmatic. It's not that I hate fairness per 
> se. I dislike it only when it's used to "solve" (aka hide) other problems.
>
> In the end, some situations do need fairness, and the fact that aiming for 
> fairness is often harder, slower, and more complicated than not doing so 
> at that point turns into a non-argument. If you need it, you need it.

Maybe some *big* NUMA machines really want this fairness (even if it cost some 
cycles as pointed by Davide in http://lkml.org/lkml/2007/3/29/246 ) , I am 
just guessing since I cannot test such monsters. I tested Davide program on a 
Dual Opteron and got some perf difference.


$ ./qspins  -n 2
now testing: TICKLOCK
timeres=4000
uscycles=1991
AVG[0]: 2195.250000 cycles/loop
SIG[0]: 11.813657
AVG[1]: 2212.312500 cycles/loop
SIG[1]: 38.038991

$ ./qspins  -n 2 -s
now testing: SPINLOCK
timeres=4000
uscycles=1991
AVG[0]: 2066.000000 cycles/loop
SIG[0]: 0.000000
AVG[1]: 2115.687500 cycles/loop
SIG[1]: 63.083000


> I just don't think we need it, and we're better off solving problems other 
> ways.
>
> (For example, we might also solve such problems by creating a separate
> "fair_spin_lock" abstraction, and only making the particular users that 
> need it actually use it. It would depend a bit on whether the cost of 
> implementing the fairness is noticeable enough for it to be worth having 
> a separate construct for it).
>
> 		Linus

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/