linux-kernel - Re: [PATCH] x86/irq: assign vectors from numa

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Message-ID: <20101222004010.GI4709@sgi.com>
Date:	Tue, 21 Dec 2010 16:40:10 -0800
From:	Arthur Kepner <akepner@....com>
To:	Thomas Gleixner <tglx@...utronix.de>
Cc:	linux-kernel@...r.kernel.org, x86@...nel.org
Subject: Re: [PATCH] x86/irq: assign vectors from numa_node

On Fri, Dec 17, 2010 at 10:04:48AM +0100, Thomas Gleixner wrote:
> B1;2401;0cOn Thu, 16 Dec 2010, Arthur Kepner wrote:
> 
> > On Fri, Dec 10, 2010 at 11:55:12AM +0100, Thomas Gleixner wrote:
> > > .....
> > > There is absolutely no point assigning 4096 interrupts to a single
> > > node in the first place. 
> > 
> > I'm not arguing that it's a good idea for a driver to request so 
> > many resources. But some do. ...
> 
> Right, some do and we now make efforts to let them do that nonsense no
> matter what ?

To me, it seems better (and easier) to let drivers do non-sensical 
things, than to determine what is sensible for a particluar driver. 

How would you choose a limit of interrupt allocations? 

> 
> >  .... And what we do now is even worse than 
> > assigning all the interrupts to a single node.
> 
> We assign it to a single core, which is wrong. But what the heck is
> the difference between assinging 4k irqs to a single core or to a
> single node ?
> 
> Nothing. And that's the whole problem.
> 
> I agree that the per node assignement needs to be resolved, but not in
> the way that we just ignore the underlying problems and solve them at
> the wrong place. You did not answer my comment further down in the
> patch:
> 
> > > I agree, that __assign_irq_vector() should honour the request for a
> > > node, but I don't agree, that we should magically spread stuff on
> > > whatever node we find, when the node ran out of vectors.
> > > 
> > > There is probably a reason, why you want an interrupt on a specific
> > > node and just silently pushing it somewhere else feels very wrong.
> > > 
> > > This needs to be solved from ground up with proper rules about failure
> > > modes and fallback decisions.
> 
> Where is the answer to this ?

The patch I sent made a 'best effort' to assign an interrupt to the 
node where the device is installed. If no vectors were available on 
that node, it tried elsewhere. 

> 
> > > Especially not, if we end up using only a few
> > > of them in the end. And those you use are not necessarily on that very
> > > node.
> > 
> > OK. Eric mentioned in a related thread that vector allocation 
> > might be deferred until request_irq(). That sounds like a good 
> > idea. But unless we change the way initial vector assignment is 
> 
> It does NOT sound like a good idea. 

This seems at odds with what you said here:

http://marc.info/?l=linux-kernel&m=128566136604152&w=2

(specifically this bit:

8<---------------------------- [snip] ----------------------------

..... So the solution would be:

create_irq allocates an irq number + irq descriptor, nothing else

chip->startup() will setup the vector and chip->shutdown releases
it. ......

8<---------------------------- [snip] ----------------------------

)

Can you clarify?

> ... Again, I agree that we need to fix
> the per node assignment, but we need to discuss the problem of
> anything which goes beyond the node.

Suppose that a interrupt is initially assigned to a different 
node than the device. Is that such a big deal? They can be 
redistributed later by a user-space irq balancer.

> 
> > done, it just defers the problem (assuming that request_irq() 
> > is done for every irq allocated in create_irq_nr()).
> 
> There is neither a reason for a driver to create so many interrupts in
> the first place nor a re reason to request them right away. This is
> just horrible crap, nothing else. Why the hell would a driver need to
> startup 4k interrupts just to load itself w/o a single user?
> 
> We do _NOT_ work around such insanity somewhere else even if we can.
> 
> > Using the numa_node of the device to do the initial vector 
> > assignment seems like a reasonable default choice, no? 
> 
> I still agree that we should honour the node request, but that has a
> totally different scope than what your patch is trying to do.
> 

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