lkml.org 
[lkml]   [2009]   [Oct]   [27]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [KVM PATCH v3 1/3] KVM: fix race in irq_routing logic
On Tue, Oct 27, 2009 at 09:39:03AM -0400, Gregory Haskins wrote:
> Gleb Natapov wrote:
> > On Mon, Oct 26, 2009 at 12:21:57PM -0400, Gregory Haskins wrote:
> >> The current code suffers from the following race condition:
> >>
> >> thread-1 thread-2
> >> -----------------------------------------------------------
> >>
> >> kvm_set_irq() {
> >> rcu_read_lock()
> >> irq_rt = rcu_dereference(table);
> >> rcu_read_unlock();
> >>
> >> kvm_set_irq_routing() {
> >> mutex_lock();
> >> irq_rt = table;
> >> rcu_assign_pointer();
> >> mutex_unlock();
> >> synchronize_rcu();
> >>
> >> kfree(irq_rt);
> >>
> >> irq_rt->entry->set(); /* bad */
> >>
> > This is not what happens. irq_rt is never accessed outside read-side
> > critical section.
>
> Sorry, I was generalizing to keep the comments short. I figured it
> would be clear what I was actually saying, but realize in retrospect
> that I was a little ambiguous.
>
A little is underestimation :) There is not /* bad */ line in the code!

> Yes, irq_rt is not accessed outside the RSCS. However, the entry
> pointers stored in the irq_rt->map are, and this is equally problematic
> afaict.
The pointer is in text and can't disappear without kvm_set_irq()
disappearing too.

>
> In this particular case we seem to never delete entries at run-time once
> they are established. Therefore, while perhaps sloppy, its technically
> safe to leave them unprotected from this perspective. The issue is more
> related to shutdown since a kvm_set_irq() caller could be within the
> aforementioned race-region and call entry->set() after the guest is
> gone. Or did I miss something?
>
The caller of kvm_set_irq() should hold reference to kvm instance, so it
can't disappear while you are inside kvm_set_irq(). RCU protects only
kvm->irq_routing not kvm structure itself.

> > Data is copied from irq_rt onto the stack and this copy is accessed
> > outside critical section.
>
> As mentioned above, I do not believe this really protect us. And even
I don't see the prove it doesn't, so I assume it does.

> if it did, the copy is just a work-around to avoid sleeping within the
It is not a work-around. There was two solutions to the problem one is
to call ->set() outside rcu critical section another is to use SRCU. I
decided to use the first one. This way the code is much simpler and I
remember asking Paul what are the disadvantages of using SRCU and there
was something.

> standard RCU RSCS, which is what SRCU is designed for. So rather than
> inventing an awkward two-phased stack based solution, it's better to
> reuse the provided tools, IMO.
>
> To flip it around: Is there any reason why an SRCU would not work here,
> and thus we were forced to use something like the stack-copy approach?
>
If SRCU has no disadvantage comparing to RCU why not use it always? :)

--
Gleb.


\
 
 \ /
  Last update: 2009-10-27 15:05    [W:0.076 / U:1.468 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site