KVM: async_pf: avoid recursive flushing of work items

This was reported by syzkaller:

    [ INFO: possible recursive locking detected ]
    4.9.0-rc4+ #49 Not tainted
    ---------------------------------------------
    kworker/2:1/5658 is trying to acquire lock:
     ([ 1644.769018] (&work->work)
    [<     inline     >] list_empty include/linux/compiler.h:243
    [<ffffffff8128dd60>] flush_work+0x0/0x660 kernel/workqueue.c:1511

    but task is already holding lock:
     ([ 1644.769018] (&work->work)
    [<ffffffff812916ab>] process_one_work+0x94b/0x1900 kernel/workqueue.c:2093

    stack backtrace:
    CPU: 2 PID: 5658 Comm: kworker/2:1 Not tainted 4.9.0-rc4+ #49
    Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011
    Workqueue: events async_pf_execute
     ffff8800676ff630 ffffffff81c2e46b ffffffff8485b930 ffff88006b1fc480
     0000000000000000 ffffffff8485b930 ffff8800676ff7e0 ffffffff81339b27
     ffff8800676ff7e8 0000000000000046 ffff88006b1fcce8 ffff88006b1fccf0
    Call Trace:
    ...
    [<ffffffff8128ddf3>] flush_work+0x93/0x660 kernel/workqueue.c:2846
    [<ffffffff812954ea>] __cancel_work_timer+0x17a/0x410 kernel/workqueue.c:2916
    [<ffffffff81295797>] cancel_work_sync+0x17/0x20 kernel/workqueue.c:2951
    [<ffffffff81073037>] kvm_clear_async_pf_completion_queue+0xd7/0x400 virt/kvm/async_pf.c:126
    [<     inline     >] kvm_free_vcpus arch/x86/kvm/x86.c:7841
    [<ffffffff810b728d>] kvm_arch_destroy_vm+0x23d/0x620 arch/x86/kvm/x86.c:7946
    [<     inline     >] kvm_destroy_vm virt/kvm/kvm_main.c:731
    [<ffffffff8105914e>] kvm_put_kvm+0x40e/0x790 virt/kvm/kvm_main.c:752
    [<ffffffff81072b3d>] async_pf_execute+0x23d/0x4f0 virt/kvm/async_pf.c:111
    [<ffffffff8129175c>] process_one_work+0x9fc/0x1900 kernel/workqueue.c:2096
    [<ffffffff8129274f>] worker_thread+0xef/0x1480 kernel/workqueue.c:2230
    [<ffffffff812a5a94>] kthread+0x244/0x2d0 kernel/kthread.c:209
    [<ffffffff831f102a>] ret_from_fork+0x2a/0x40 arch/x86/entry/entry_64.S:433

The reason is that kvm_put_kvm is causing the destruction of the VM, but
the page fault is still on the ->queue list.  The ->queue list is owned
by the VCPU, not by the work items, so we cannot just add list_del to
the work item.

Instead, use work->vcpu to note async page faults that have been resolved
and will be processed through the done list.  There is no need to flush
those.

Cc: Dmitry Vyukov <dvyukov@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Radim Krčmář <rkrcmar@redhat.com>

virt/kvm/async_pf.c

@@ -91,6 +91,7 @@ static void async_pf_execute(struct work_struct *work)
 
 	spin_lock(&vcpu->async_pf.lock);
 	list_add_tail(&apf->link, &vcpu->async_pf.done);
+	apf->vcpu = NULL;
 	spin_unlock(&vcpu->async_pf.lock);
 
 	/*
@@ -113,13 +114,23 @@ static void async_pf_execute(struct work_struct *work)
 
 void kvm_clear_async_pf_completion_queue(struct kvm_vcpu *vcpu)
 {
+	spin_lock(&vcpu->async_pf.lock);
+
 	/* cancel outstanding work queue item */
 	while (!list_empty(&vcpu->async_pf.queue)) {
 		struct kvm_async_pf *work =
 			list_first_entry(&vcpu->async_pf.queue,
 					 typeof(*work), queue);
 		list_del(&work->queue);
 
+		/*
+		 * We know it's present in vcpu->async_pf.done, do
+		 * nothing here.
+		 */
+		if (!work->vcpu)
+			continue;
+
+		spin_unlock(&vcpu->async_pf.lock);
 #ifdef CONFIG_KVM_ASYNC_PF_SYNC
 		flush_work(&work->work);
 #else
@@ -129,9 +140,9 @@ void kvm_clear_async_pf_completion_queue(struct kvm_vcpu *vcpu)
 			kmem_cache_free(async_pf_cache, work);
 		}
 #endif
+		spin_lock(&vcpu->async_pf.lock);
 	}
 
-	spin_lock(&vcpu->async_pf.lock);
 	while (!list_empty(&vcpu->async_pf.done)) {
 		struct kvm_async_pf *work =
 			list_first_entry(&vcpu->async_pf.done,