While KVM_XEN_EVTCHN_RESET is usually called with no vCPUs running,
if that happened it could cause a deadlock. This is due to
kvm_xen_eventfd_reset() doing a synchronize_srcu() inside
a kvm->lock critical section.
To avoid this, first collect all the evtchnfd objects in an
array and free all of them once the kvm->lock critical section
is over and th SRCU grace period has expired.
Reported-by: Michal Luczaj <[email protected]>
Cc: David Woodhouse <[email protected]>
Signed-off-by: Paolo Bonzini <[email protected]>
---
arch/x86/kvm/xen.c | 30 +++++++++++++++++--
.../selftests/kvm/x86_64/xen_shinfo_test.c | 6 ++++
2 files changed, 33 insertions(+), 3 deletions(-)
diff --git a/arch/x86/kvm/xen.c b/arch/x86/kvm/xen.c
index b178f40bd863..2e29bdc2949c 100644
--- a/arch/x86/kvm/xen.c
+++ b/arch/x86/kvm/xen.c
@@ -1942,18 +1942,42 @@ static int kvm_xen_eventfd_deassign(struct kvm *kvm, u32 port)
static int kvm_xen_eventfd_reset(struct kvm *kvm)
{
- struct evtchnfd *evtchnfd;
+ struct evtchnfd *evtchnfd, **all_evtchnfds;
int i;
+ int n = 0;
mutex_lock(&kvm->lock);
+
+ /*
+ * Because synchronize_srcu() cannot be called inside the
+ * critical section, first collect all the evtchnfd objects
+ * in an array as they are removed from evtchn_ports.
+ */
+ idr_for_each_entry(&kvm->arch.xen.evtchn_ports, evtchnfd, i)
+ n++;
+
+ all_evtchnfds = kmalloc_array(n, sizeof(struct evtchnfd *), GFP_KERNEL);
+ if (!all_evtchnfds) {
+ mutex_unlock(&kvm->lock);
+ return -ENOMEM;
+ }
+
+ n = 0;
idr_for_each_entry(&kvm->arch.xen.evtchn_ports, evtchnfd, i) {
+ all_evtchnfds[n++] = evtchnfd;
idr_remove(&kvm->arch.xen.evtchn_ports, evtchnfd->send_port);
- synchronize_srcu(&kvm->srcu);
+ }
+ mutex_unlock(&kvm->lock);
+
+ synchronize_srcu(&kvm->srcu);
+
+ while (n--) {
+ evtchnfd = all_evtchnfds[n];
if (!evtchnfd->deliver.port.port)
eventfd_ctx_put(evtchnfd->deliver.eventfd.ctx);
kfree(evtchnfd);
}
- mutex_unlock(&kvm->lock);
+ kfree(all_evtchnfds);
return 0;
}
diff --git a/tools/testing/selftests/kvm/x86_64/xen_shinfo_test.c b/tools/testing/selftests/kvm/x86_64/xen_shinfo_test.c
index 721f6a693799..dae510c263b4 100644
--- a/tools/testing/selftests/kvm/x86_64/xen_shinfo_test.c
+++ b/tools/testing/selftests/kvm/x86_64/xen_shinfo_test.c
@@ -962,6 +962,12 @@ int main(int argc, char *argv[])
}
done:
+ struct kvm_xen_hvm_attr evt_reset = {
+ .type = KVM_XEN_ATTR_TYPE_EVTCHN,
+ .u.evtchn.flags = KVM_XEN_EVTCHN_RESET,
+ };
+ vm_ioctl(vm, KVM_XEN_HVM_SET_ATTR, &evt_reset);
+
alarm(0);
clock_gettime(CLOCK_REALTIME, &max_ts);
--
2.31.1
On 12/28/22 12:55, Woodhouse, David wrote:
>
> IIRC the send_port and type aren't used from hcall_send so we could
> stick those in a union with a 'next' pointer and use it to build a list
> instead of having to allocate the array (pathological worst case 32KiB).
Yeah I thought about using a list. The union is doable as you say, but
I wasn't sure if things were going to change in kvm_xen_evtchn_send later.
And the worst case is order 3, which is within PAGE_ALLOC_COSTLY_ORDER,
so I decided to go with the array which only consumes memory on reset
rather than always; it's much more likely that the array will be smaller
than a page.
Paolo
> Or if the union is a bit icky, we could just add the 'next' pointer
> unconditionally.
>
Hi Paolo,
> diff --git a/tools/testing/selftests/kvm/x86_64/xen_shinfo_test.c b/tools/testing/selftests/kvm/x86_64/xen_shinfo_test.c
> index 721f6a693799..dae510c263b4 100644
> --- a/tools/testing/selftests/kvm/x86_64/xen_shinfo_test.c
> +++ b/tools/testing/selftests/kvm/x86_64/xen_shinfo_test.c
> @@ -962,6 +962,12 @@ int main(int argc, char *argv[])
> }
>
> done:
> + struct kvm_xen_hvm_attr evt_reset = {
> + .type = KVM_XEN_ATTR_TYPE_EVTCHN,
> + .u.evtchn.flags = KVM_XEN_EVTCHN_RESET,
> + };
> + vm_ioctl(vm, KVM_XEN_HVM_SET_ATTR, &evt_reset);
> +
> alarm(0);
> clock_gettime(CLOCK_REALTIME, &max_ts);
>
This change generates a build failure with error message:
"error: a label can only be part of a statement and a declaration is not a statement".
Moving the definition of evt_reset to the beginning of main() can fix it:
diff --git a/tools/testing/selftests/kvm/x86_64/xen_shinfo_test.c b/tools/testing/selftests/kvm/x86_64/xen_shinfo_test.c
index dae510c263b4..d71f1508bb21 100644
--- a/tools/testing/selftests/kvm/x86_64/xen_shinfo_test.c
+++ b/tools/testing/selftests/kvm/x86_64/xen_shinfo_test.c
@@ -623,6 +623,10 @@ int main(int argc, char *argv[])
rs->state = 0x5a;
bool evtchn_irq_expected = false;
+ struct kvm_xen_hvm_attr evt_reset = {
+ .type = KVM_XEN_ATTR_TYPE_EVTCHN,
+ .u.evtchn.flags = KVM_XEN_EVTCHN_RESET,
+ };
for (;;) {
volatile struct kvm_run *run = vcpu->run;
@@ -962,10 +966,6 @@ int main(int argc, char *argv[])
}
done:
- struct kvm_xen_hvm_attr evt_reset = {
- .type = KVM_XEN_ATTR_TYPE_EVTCHN,
- .u.evtchn.flags = KVM_XEN_EVTCHN_RESET,
- };
vm_ioctl(vm, KVM_XEN_HVM_SET_ATTR, &evt_reset);
alarm(0);
B.R.
Yu
On Thu, Jan 19, 2023, Yu Zhang wrote:
> Hi Paolo,
>
> > diff --git a/tools/testing/selftests/kvm/x86_64/xen_shinfo_test.c b/tools/testing/selftests/kvm/x86_64/xen_shinfo_test.c
> > index 721f6a693799..dae510c263b4 100644
> > --- a/tools/testing/selftests/kvm/x86_64/xen_shinfo_test.c
> > +++ b/tools/testing/selftests/kvm/x86_64/xen_shinfo_test.c
> > @@ -962,6 +962,12 @@ int main(int argc, char *argv[])
> > }
> >
> > done:
> > + struct kvm_xen_hvm_attr evt_reset = {
> > + .type = KVM_XEN_ATTR_TYPE_EVTCHN,
> > + .u.evtchn.flags = KVM_XEN_EVTCHN_RESET,
> > + };
> > + vm_ioctl(vm, KVM_XEN_HVM_SET_ATTR, &evt_reset);
> > +
> > alarm(0);
> > clock_gettime(CLOCK_REALTIME, &max_ts);
> >
>
> This change generates a build failure with error message:
> "error: a label can only be part of a statement and a declaration is not a statement".
And other flavors too, e.g.
x86_64/xen_shinfo_test.c:965:2: error: expected expression
struct kvm_xen_hvm_attr evt_reset = {
^
x86_64/xen_shinfo_test.c:969:38: error: use of undeclared identifier 'evt_reset'
vm_ioctl(vm, KVM_XEN_HVM_SET_ATTR, &evt_reset);
^
x86_64/xen_shinfo_test.c:969:38: error: use of undeclared identifier 'evt_reset'
3 errors generated.
make: *** [../lib.mk:145: tools/testing/selftests/kvm/x86_64/xen_shinfo_test] Error 1
make: *** Waiting for unfinished jobs....
I'm surprised bots haven't complained about this, haven't seen any reports.
> Moving the definition of evt_reset to the beginning of main() can fix it:
I'll queue a patch, this is already in Linus' tree and I've collected a few other
tiny fixes for v6.2-rcwhatever that I'll send to Paolo.
Thanks!
On Thu, Jan 19, 2023 at 7:04 PM Sean Christopherson <[email protected]> wrote:
> > It's clang only; GCC only warns with -Wpedantic. Plus, bots probably
> > don't compile tools/ that much.
>
> /wave
>
> Want to queue Yu's fix directly Paolo? I was assuming you'd be offline until
> sometime tomorrow.
Yes, I can, but what other patches were you meaning to send?
Paolo
On Thu, Jan 19, 2023 at 6:57 PM Sean Christopherson <[email protected]> wrote:
> > This change generates a build failure with error message:
> > "error: a label can only be part of a statement and a declaration is not a statement".
>
> And other flavors too, e.g.
>
> x86_64/xen_shinfo_test.c:965:2: error: expected expression
> struct kvm_xen_hvm_attr evt_reset = {
> ^
> x86_64/xen_shinfo_test.c:969:38: error: use of undeclared identifier 'evt_reset'
> vm_ioctl(vm, KVM_XEN_HVM_SET_ATTR, &evt_reset);
> ^
> x86_64/xen_shinfo_test.c:969:38: error: use of undeclared identifier 'evt_reset'
> 3 errors generated.
> make: *** [../lib.mk:145: tools/testing/selftests/kvm/x86_64/xen_shinfo_test] Error 1
> make: *** Waiting for unfinished jobs....
>
> I'm surprised bots haven't complained about this, haven't seen any reports.
It's clang only; GCC only warns with -Wpedantic. Plus, bots probably
don't compile tools/ that much.
Paolo
On Thu, Jan 19, 2023 at 7:15 PM Sean Christopherson <[email protected]> wrote:
> A minor selftest fix
>
> https://lore.kernel.org/all/[email protected]
>
> and a fix for a longstanding VMX bug that seems problematic enough that it
> warrants going into this cycle.
>
> https://lore.kernel.org/all/[email protected]
Ok, I had seen the latter so I'll put together a pull request.
Paolo
On Thu, Jan 19, 2023, Paolo Bonzini wrote:
> On Thu, Jan 19, 2023 at 7:04 PM Sean Christopherson <[email protected]> wrote:
> > > It's clang only; GCC only warns with -Wpedantic. Plus, bots probably
> > > don't compile tools/ that much.
> >
> > /wave
> >
> > Want to queue Yu's fix directly Paolo? I was assuming you'd be offline until
> > sometime tomorrow.
>
> Yes, I can, but what other patches were you meaning to send?
A minor selftest fix
https://lore.kernel.org/all/[email protected]
and a fix for a longstanding VMX bug that seems problematic enough that it
warrants going into this cycle.
https://lore.kernel.org/all/[email protected]
On Thu, Jan 19, 2023, Paolo Bonzini wrote:
> On Thu, Jan 19, 2023 at 6:57 PM Sean Christopherson <[email protected]> wrote:
> > > This change generates a build failure with error message:
> > > "error: a label can only be part of a statement and a declaration is not a statement".
> >
> > And other flavors too, e.g.
> >
> > x86_64/xen_shinfo_test.c:965:2: error: expected expression
> > struct kvm_xen_hvm_attr evt_reset = {
> > ^
> > x86_64/xen_shinfo_test.c:969:38: error: use of undeclared identifier 'evt_reset'
> > vm_ioctl(vm, KVM_XEN_HVM_SET_ATTR, &evt_reset);
> > ^
> > x86_64/xen_shinfo_test.c:969:38: error: use of undeclared identifier 'evt_reset'
> > 3 errors generated.
> > make: *** [../lib.mk:145: tools/testing/selftests/kvm/x86_64/xen_shinfo_test] Error 1
> > make: *** Waiting for unfinished jobs....
> >
> > I'm surprised bots haven't complained about this, haven't seen any reports.
>
> It's clang only; GCC only warns with -Wpedantic. Plus, bots probably
> don't compile tools/ that much.
/wave
Want to queue Yu's fix directly Paolo? I was assuming you'd be offline until
sometime tomorrow.
On 28/12/2022 12:04, Paolo Bonzini wrote:
> While KVM_XEN_EVTCHN_RESET is usually called with no vCPUs running,
> if that happened it could cause a deadlock. This is due to
> kvm_xen_eventfd_reset() doing a synchronize_srcu() inside
> a kvm->lock critical section.
>
> [...]
>
> + /*
> + * Because synchronize_srcu() cannot be called inside the
> + * critical section, first collect all the evtchnfd objects
> + * in an array as they are removed from evtchn_ports.
> + */
With the recent changes regarding the locking order (locking.rst:
"synchronize_srcu(&kvm->srcu) is called inside critical sections for kvm->lock,
vcpu->mutex and kvm->slots_lock"), is this comment still valid?
Or is there a rule that forbids synchronize_srcu() under the newly introduced
kvm->arch.xen.xen_lock?
thanks,
Michal