Re: [PATCH] local_irq_disable removal

[Esben Nielsen]

On Sat, 11 Jun 2005, Sven-Thorsten Dietrich wrote:

> On Sat, 2005-06-11 at 16:32 +0200, Esben Nielsen wrote:
> 
> 
> > > > The more I think about it the more dangerous I think it is. What does 
> > > > local_irq_disable() protect against? All local threads as well as 
> > > > irq-handlers. If these sections keeped mutual exclusive but preemtible 
> > > > we will not have protected against a irq-handler.
> > > 
> 
> The more Dangerous you perieve it to be. Can you point to real damage?
> After all this is experimental code.
> 
> > That is exactly my point: We can't make a per-cpu mutex to replace
> > local_irq_disable(). We have to make real lock for each subsystem now
> > relying on local_irq_disable(). A global lock will not work. We could have
> > a temporary lock all non-RT can share but that would be a hack similar to
> > BKL.
> > 
> 
> Why do we need any of this?

If we want deterministic task-latencies without worrying about
proof-reading the code of every subsystem which might be using
local_irq_disable() in some non-deterministic way, we need to make another
locking mechanism.

> 
> > The current soft-irq states only gives us better hard-irq latency but
> > nothing else. I think the overhead runtime and the complication of the
> > code is way too big for gaining only that. 
> 
> 
> Real numbers please, not speculation! Science, not religion.
> 
Well, isn't enough to see that the code contains more instructions and
looks somewhat more complicated?

> > What is the aim of PREEMPT_RT? Low irq-latency, low task latency or
> > determnistic task latency? The soft irq-state helps on the first but
> > harms the two others indirectly by introducing extra overhead. To be
> > honest I think that approach should be abandoned.
> 
> The purpose of the irq fix is deterministic interrupt response at a low
> cost.
> 
> Quite honestly, if it does get abandoned in the community, Daniel and I
> will continue to maintain it.

I think it should be in as an obtion. And it should not be marketed as
fixing the problem of wide spread use of local_irq_disable().
> 
> >  
> > > what seems a better is to rewrite per-CPU-local-irq-disable users to 
> > > make use of the DEFINE_PER_CPU_LOCKED/per_cpu_locked/get_cpu_lock 
> > > primitives to use preemptible per-CPU data structures. In this case 
> > > these sections would be truly preemptible. I've done this for a couple 
> > > of cases already, where it was unavoidable for lock-dependency reasons.
> > >
> > 
> > I'll continue that work then but in a way where !PREEMPT_RT will make it
> > back into local-irq-disable such it wont hurt performance there.
> >  
> > I.e. I will try to make a macro system, and try to turn references to
> > local_irq_disable() into these - or raw_local_irq_disable().
> 
> Esben, are there any bugs you are finding, or do you just dislike the
> overhead?
> 

Both the overhead and the complication. You instroduce yet another kind of
context into the system. I was seeing PREEMPT_RT as a cleanup: There is
basicly _only_ the thread state.

> Your argument is identical to the one that people made (and still make)
> about preemption in the first place. 
> 

Yes it is. Preemption is expensive and if you don't need it avoid it.
This is basicly the same thing - just on another level in the system.
Preemption turns 100 ms latencies into 10 ms latencies. PREEMPT_RT turns
them into 100 us. You can probably take interrupt latencies down to a few
us. That is good. But there is a price to pay in performance at each step.
I do not see any reason to go below 100 us in latencies. Others might want
to pay to get the last bit.

> A lot of issues were raised about simple preempt disable that was
> necessary to make the kernel preemptable outside critical sections with
> preemption, even though the concept was easy to understand from an SMP
> POV. A lot of FUD is STILL floating around from that, 5 YEARS AGO.
> 
> A lot of discussion has been taking place about the constant overhead
> necessary to improve preemption down to the level where we are now, in
> PREEMPT_RT (with or without the IRQ relief you are concerned about)
> 
> The overhead of the mutex, which IS somewhat significant.
> 
> The people who need the preemption response-time absolutely don't care
> about the overhead, because the cost of the CPUs required to make
> vanilla Linux (with its high preemption latency) perform at latencies as
> low as RT, is much higher, than the cost of a middle-of-the-road-
> upgrade, to mask the constant mutex overhead.
> 
> In other words, the people who need RT preemption are willing to accept
> the overhead absolutely, because it still saves them money, time, power,
> whatever.
> 
> Daniel has now gone another step, very similar to the first step in
> preemption, but taking advantage of a new environment, IRQs in threads.
> 
> The price here isn't that high, and the gain is astronomical, in terms
> of the confidence in the interrupt subsystem that can be established
> with this.
> 
> In a similar context as the argument I made above, people who need this
> level of performance, will be real happy that they can leverage it, and
> they WILL accept the overhead.
> 
> Daniel made this configurable, which is conventional in add-on patches.
> Ingo removed that, and I do not disagree with either approach. I am just
> glad that some folks understand the value of this concept, which Daniel
> and I have been digesting since we opensourced the Mutex concept in the
> first place.
> 
I think Ingo made a wrong move there. It should remain optional - just as
the whole PREEMPT_RT and even PREEMPT is optional.

> I would prefer REAL performance data on the overhead you are suggesting,
> rather than pages and pages of speculation, that make me question if you
> have fully digested what is being done here.
> 
> I will be producing those ASAP, but I am REALLY busy with people who are
> excited about the new era in Linux development that we are
> (collectively) introducing. We are not expecting everyone to understand
> it or agree with it.
> 
> But this is scientific, and we do need scientific approaches to
> analysis.
>

Please, you are taking this too personal. I started by asking questions
because what I saw wasn't living up to my expectations (getting around
"random" local_irq_disable()) but "only" gave good irq-latencies, which
I to be frank have no interrest in.

No, I don't understand _all_ the details of the patch but I think I
understand the principles. From my chair it is just an unwanted
complication :-(
 
> Thanks,
> 
> Sven
> 
> 
Esben

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