[Qemu-devel] [PATCH] s390x/kvm: Fix potential deadlock in sigp handling

[Christian Borntraeger]

FYI, unless I find a better alternative,  I plan to go with this patch.
------snip------
[PATCH] s390x/kvm: Fix potential deadlock in sigp handling

If two VCPUs exit at the same time and target each other
with a sigp, both could run into a deadlock as run_on_cpu
on CPU0 will free the BQL when starting the CPU1 target routine.
CPU1 will run its sigp initiater for CPU0 before handling
the run_on_cpu requests, thus resulting in a dead lock.

As all qemu SIGPs are slow path anway we can use a big sigp
lock and allow only one SIGP for the guest at a time. We will
return condition code 2 (BUSY) on contention to the guest.

Reported-by: Paolo Bonzini <address@hidden>
Signed-off-by: Christian Borntraeger <address@hidden>
Reviewed-by: Cornelia Huck <address@hidden>
---
 target-s390x/kvm.c | 11 +++++++++++
 1 file changed, 11 insertions(+)

diff --git a/target-s390x/kvm.c b/target-s390x/kvm.c
index f348745..a9fa831 100644
--- a/target-s390x/kvm.c
+++ b/target-s390x/kvm.c
@@ -133,6 +133,8 @@ const KVMCapabilityInfo kvm_arch_required_capabilities[] = {
     KVM_CAP_LAST_INFO
 };
 
+static QemuMutex qemu_sigp_mutex;
+
 static int cap_sync_regs;
 static int cap_async_pf;
 static int cap_mem_op;
@@ -358,6 +360,8 @@ int kvm_arch_init(MachineState *ms, KVMState *s)
         rc = compat_disable_facilities(s, fac_mask, ARRAY_SIZE(fac_mask));
     }
 
+    qemu_mutex_init(&qemu_sigp_mutex);
+
     return rc;
 }
 
@@ -1845,6 +1849,11 @@ static int handle_sigp(S390CPU *cpu, struct kvm_run 
*run, uint8_t ipa1)
     status_reg = &env->regs[r1];
     param = (r1 % 2) ? env->regs[r1] : env->regs[r1 + 1];
 
+    if (qemu_mutex_trylock(&qemu_sigp_mutex)) {
+        ret = SIGP_CC_BUSY;
+        goto out;
+    }
+
     switch (order) {
     case SIGP_SET_ARCH:
         ret = sigp_set_architecture(cpu, param, status_reg);
@@ -1854,7 +1863,9 @@ static int handle_sigp(S390CPU *cpu, struct kvm_run *run, 
uint8_t ipa1)
         dst_cpu = s390_cpu_addr2state(env->regs[r3]);
         ret = handle_sigp_single_dst(dst_cpu, order, param, status_reg);
     }
+    qemu_mutex_unlock(&qemu_sigp_mutex);
 
+out:
     trace_kvm_sigp_finished(order, CPU(cpu)->cpu_index,
                             dst_cpu ? CPU(dst_cpu)->cpu_index : -1, ret);
 
-- 
2.5.5