| 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
| ||
|
Message-ID: <alpine.DEB.2.02.1207181851010.23783@kaball.uk.xensource.com>
Date: Wed, 18 Jul 2012 19:17:11 +0100
From: Stefano Stabellini <stefano.stabellini@...citrix.com>
To: Konrad Rzeszutek Wilk <konrad.wilk@...cle.com>
CC: Stefano Stabellini <Stefano.Stabellini@...citrix.com>,
"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
"xen-devel@...ts.xensource.com" <xen-devel@...ts.xensource.com>
Subject: Re: [Xen-devel] [PATCH] xen/events: fix unmask_evtchn for PV on HVM
guests
On Mon, 16 Jul 2012, Konrad Rzeszutek Wilk wrote:
> On Fri, Jul 13, 2012 at 06:48:35PM +0100, Stefano Stabellini wrote:
> > On Mon, 9 Jul 2012, Konrad Rzeszutek Wilk wrote:
> > > On Fri, Jun 22, 2012 at 05:26:07PM +0100, Stefano Stabellini wrote:
> > > > When unmask_evtchn is called, if we already have an event pending, we
> > > > just set evtchn_pending_sel waiting for local_irq_enable to be called.
> > > > That is because PV guests set the irq_enable pvops to
> > >
> > > Can you point out where the PV guests do that please? Even just
> > > including a snippet of code would be nice so that somebody
> > > in the future has an idea of where it was/is.
> >
> > Do you mean where PV guests set the irq_enable pvop?
> >
> > That would be in xen_setup_vcpu_info_placement.
> > irq_enable is set to xen_irq_enable_direct that is implemented in
> > assembly in arch/x86/xen/xen-asm.S: it tests for XEN_vcpu_info_pending
> > and call xen_force_evtchn_callback.
>
> Excellent. Pls include that comment in the git commit.
OK
> > > > diff --git a/drivers/xen/events.c b/drivers/xen/events.c
> > > > index eae0d0b..0132505 100644
> > > > --- a/drivers/xen/events.c
> > > > +++ b/drivers/xen/events.c
> > > > @@ -372,8 +372,11 @@ static void unmask_evtchn(int port)
> > > >
> > > > BUG_ON(!irqs_disabled());
> > > >
> > > > - /* Slow path (hypercall) if this is a non-local port. */
> > > > - if (unlikely(cpu != cpu_from_evtchn(port))) {
> > > > + /* Slow path (hypercall) if this is a non-local port or if this is
> > > > + * an hvm domain and an event is pending (hvm domains don't have
> > > > + * their own implementation of irq_enable). */
> > > > + if (unlikely((cpu != cpu_from_evtchn(port)) ||
> > > > + (xen_hvm_domain() && sync_test_bit(port, &s->evtchn_pending[0])))) {
> > > > struct evtchn_unmask unmask = { .port = port };
> > >
> > > We already have two seperate acks - for when there is an GMFN APIC bitmap and
> > > when there is not. Can we also have to seperate unmask_evtchn then? And
> > > just have the HVM and ARM just do a straightforward unmaks_evtchn while
> > > the PV remains the same?
> >
> > Do you mean HVM and ARM do a straightforward EVTCHNOP_unmask hypercall?
>
> I was thinking of some way to lessen the impact of the 'if (..)' statement.
> There is already a check from the cpu, and now there is a bit check
> and another check for domain. Was wondering if it would make more sense
> to abstract the code the unmask_evtchn calls, and provide two variants
> of the unmask_evtchn: a one that is mostly called on PV/PVHVM on x86 and
> then the ARM version?
>
> Or won't that really give us any performance benefits and that
> extra check for hvm_domain and test_bit is negligible?
>
> Perhaps a better question is - do you have further plans for this
> function? As in expanding it with more 'if' conditionals?
Nope, I certainly don't. In fact I agree with you on the fact that is
not very readable as it is.
> > In that case we would lose performances because most of the time an
> > hypercall won't be necessary.
> > If we keep the code as it is, it makes sense to have the PV and PVHVM
> > cases in the same function.
>
> The two things that roam my mind is:
> - performance impact
> - code readability.
>
> Granted this code is the slow patch so maybe the performance part is
> not an issue. But that 'sync_test_bit' isn't that an atomic locked
> call so it flushes the bus? There is a 'xen_hvm_domain()' condition
> before it so that does lessen the impact to be only done on HVM.
>
> If we do run this under HVM, we would do:
> 1) cpu == cpu_from_evtchn, so
> 2).sync_test_bit .. say it returns false
> 3). sync_clear_bit
> 4). sync_test_bit on the same word that 2) was done.
>
> If this was re-organized a bit differently could we remove 2)
> out of the picture so that under HVM we just do 1) 3) and 4) ?
I see what you mean now. It might make sense.
> And for that we might have to have two implementations of unmaks_evtchn - were
> both of them might call the same underlaying functions that do the
> bit-operations, but the 'if' conditionals are differently organized.
> Or is this scenario really unlikely and I am just thinking to hard about this?
Let just say that I only managed to reproduce it with a buggy hypervisor
port to a new architecture :-)
But I don't like the idea of writing obfuscated and inefficient code, so
let me give it a second try.
---
xen/events: fix unmask_evtchn for PV on HVM guests
When unmask_evtchn is called, if we already have an event pending, we
just set evtchn_pending_sel waiting for local_irq_enable to be called.
That is because PV guests set the irq_enable pvops to
xen_irq_enable_direct in xen_setup_vcpu_info_placement:
xen_irq_enable_direct is implemented in assembly in
arch/x86/xen/xen-asm.S and call xen_force_evtchn_callback if
XEN_vcpu_info_pending is set.
However HVM guests (and ARM guests) do not change or do not have the
irq_enable pvop, so evtchn_unmask cannot work properly for them.
Considering that having the pending_irq bit set when unmask_evtchn is
called is not very common, and it is simpler to keep the
native_irq_enable implementation for HVM guests (and ARM guests), the
best thing to do is just use the EVTCHNOP_unmask hypercall (Xen
re-injects pending events in response).
Signed-off-by: Stefano Stabellini <stefano.stabellini@...citrix.com>
---
drivers/xen/events.c | 17 ++++++++++++++---
1 files changed, 14 insertions(+), 3 deletions(-)
diff --git a/drivers/xen/events.c b/drivers/xen/events.c
index 0a8a17c..d75cc39 100644
--- a/drivers/xen/events.c
+++ b/drivers/xen/events.c
@@ -373,11 +373,22 @@ static void unmask_evtchn(int port)
{
struct shared_info *s = HYPERVISOR_shared_info;
unsigned int cpu = get_cpu();
+ int do_hypercall = 0, evtchn_pending = 0;
BUG_ON(!irqs_disabled());
- /* Slow path (hypercall) if this is a non-local port. */
- if (unlikely(cpu != cpu_from_evtchn(port))) {
+ if (unlikely((cpu != cpu_from_evtchn(port))))
+ do_hypercall = 1;
+ else
+ evtchn_pending = sync_test_bit(port, &s->evtchn_pending[0]);
+
+ if (unlikely(evtchn_pending && xen_hvm_domain()))
+ do_hypercall = 1;
+
+ /* Slow path (hypercall) if this is a non-local port or if this is
+ * an hvm domain and an event is pending (hvm domains don't have
+ * their own implementation of irq_enable). */
+ if (do_hypercall) {
struct evtchn_unmask unmask = { .port = port };
(void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
} else {
@@ -390,7 +401,7 @@ static void unmask_evtchn(int port)
* 'hw_resend_irq'. Just like a real IO-APIC we 'lose
* the interrupt edge' if the channel is masked.
*/
- if (sync_test_bit(port, &s->evtchn_pending[0]) &&
+ if (evtchn_pending &&
!sync_test_and_set_bit(port / BITS_PER_LONG,
&vcpu_info->evtchn_pending_sel))
vcpu_info->evtchn_upcall_pending = 1;
--
1.7.2.5
--
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/
Powered by blists - more mailing lists