kvm/x86: Avoid async PF preempting the kernel incorrectly
authorBoqun Feng <boqun.feng@gmail.com>
Tue, 3 Oct 2017 13:36:51 +0000 (21:36 +0800)
committerRadim Krčmář <rkrcmar@redhat.com>
Wed, 4 Oct 2017 16:28:53 +0000 (18:28 +0200)
Currently, in PREEMPT_COUNT=n kernel, kvm_async_pf_task_wait() could call
schedule() to reschedule in some cases.  This could result in
accidentally ending the current RCU read-side critical section early,
causing random memory corruption in the guest, or otherwise preempting
the currently running task inside between preempt_disable and
preempt_enable.

The difficulty to handle this well is because we don't know whether an
async PF delivered in a preemptible section or RCU read-side critical section
for PREEMPT_COUNT=n, since preempt_disable()/enable() and rcu_read_lock/unlock()
are both no-ops in that case.

To cure this, we treat any async PF interrupting a kernel context as one
that cannot be preempted, preventing kvm_async_pf_task_wait() from choosing
the schedule() path in that case.

To do so, a second parameter for kvm_async_pf_task_wait() is introduced,
so that we know whether it's called from a context interrupting the
kernel, and the parameter is set properly in all the callsites.

Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Wanpeng Li <wanpeng.li@hotmail.com>
Cc: stable@vger.kernel.org
Signed-off-by: Boqun Feng <boqun.feng@gmail.com>
Signed-off-by: Radim Krčmář <rkrcmar@redhat.com>
arch/x86/include/asm/kvm_para.h
arch/x86/kernel/kvm.c
arch/x86/kvm/mmu.c

index bc62e7cbf1b1f883fc9acb0a145305384a3bf062..59ad3d132353280cae5a1ba4f3a8ae0efe996331 100644 (file)
@@ -88,7 +88,7 @@ static inline long kvm_hypercall4(unsigned int nr, unsigned long p1,
 bool kvm_para_available(void);
 unsigned int kvm_arch_para_features(void);
 void __init kvm_guest_init(void);
-void kvm_async_pf_task_wait(u32 token);
+void kvm_async_pf_task_wait(u32 token, int interrupt_kernel);
 void kvm_async_pf_task_wake(u32 token);
 u32 kvm_read_and_reset_pf_reason(void);
 extern void kvm_disable_steal_time(void);
@@ -103,7 +103,7 @@ static inline void kvm_spinlock_init(void)
 
 #else /* CONFIG_KVM_GUEST */
 #define kvm_guest_init() do {} while (0)
-#define kvm_async_pf_task_wait(T) do {} while(0)
+#define kvm_async_pf_task_wait(T, I) do {} while(0)
 #define kvm_async_pf_task_wake(T) do {} while(0)
 
 static inline bool kvm_para_available(void)
index e675704fa6f7d89ed919631284156717e342979f..8bb9594d076166ee0f6bd4f70350fe3ecf8c3d8b 100644 (file)
@@ -117,7 +117,11 @@ static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
        return NULL;
 }
 
-void kvm_async_pf_task_wait(u32 token)
+/*
+ * @interrupt_kernel: Is this called from a routine which interrupts the kernel
+ *                   (other than user space)?
+ */
+void kvm_async_pf_task_wait(u32 token, int interrupt_kernel)
 {
        u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
        struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
@@ -140,8 +144,10 @@ void kvm_async_pf_task_wait(u32 token)
 
        n.token = token;
        n.cpu = smp_processor_id();
-       n.halted = is_idle_task(current) || preempt_count() > 1 ||
-                  rcu_preempt_depth();
+       n.halted = is_idle_task(current) ||
+                  (IS_ENABLED(CONFIG_PREEMPT_COUNT)
+                   ? preempt_count() > 1 || rcu_preempt_depth()
+                   : interrupt_kernel);
        init_swait_queue_head(&n.wq);
        hlist_add_head(&n.link, &b->list);
        raw_spin_unlock(&b->lock);
@@ -269,7 +275,7 @@ do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
        case KVM_PV_REASON_PAGE_NOT_PRESENT:
                /* page is swapped out by the host. */
                prev_state = exception_enter();
-               kvm_async_pf_task_wait((u32)read_cr2());
+               kvm_async_pf_task_wait((u32)read_cr2(), !user_mode(regs));
                exception_exit(prev_state);
                break;
        case KVM_PV_REASON_PAGE_READY:
index eca30c1eb1d97cd367c6b2b76396991b79328b1a..106d4a029a8a933ad3865a11a5e2c1a11810e714 100644 (file)
@@ -3837,7 +3837,7 @@ int kvm_handle_page_fault(struct kvm_vcpu *vcpu, u64 error_code,
        case KVM_PV_REASON_PAGE_NOT_PRESENT:
                vcpu->arch.apf.host_apf_reason = 0;
                local_irq_disable();
-               kvm_async_pf_task_wait(fault_address);
+               kvm_async_pf_task_wait(fault_address, 0);
                local_irq_enable();
                break;
        case KVM_PV_REASON_PAGE_READY: