kvm, rt: change async pagefault code locking for PREEMPT_RT
authorRik van Riel <riel@redhat.com>
Mon, 21 Mar 2016 14:13:27 +0000 (15:13 +0100)
committerPaolo Bonzini <pbonzini@redhat.com>
Tue, 22 Mar 2016 15:38:38 +0000 (16:38 +0100)
The async pagefault wake code can run from the idle task in exception
context, so everything here needs to be made non-preemptible.

Conversion to a simple wait queue and raw spinlock does the trick.

Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
arch/x86/kernel/kvm.c

index 47190bd399e7ef5680172859fa78d9ae1ba0e595..807950860fb7028e28fe1e98d8a2cddeccfa8063 100644 (file)
@@ -36,6 +36,7 @@
 #include <linux/kprobes.h>
 #include <linux/debugfs.h>
 #include <linux/nmi.h>
+#include <linux/swait.h>
 #include <asm/timer.h>
 #include <asm/cpu.h>
 #include <asm/traps.h>
@@ -91,14 +92,14 @@ static void kvm_io_delay(void)
 
 struct kvm_task_sleep_node {
        struct hlist_node link;
-       wait_queue_head_t wq;
+       struct swait_queue_head wq;
        u32 token;
        int cpu;
        bool halted;
 };
 
 static struct kvm_task_sleep_head {
-       spinlock_t lock;
+       raw_spinlock_t lock;
        struct hlist_head list;
 } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
 
@@ -122,17 +123,17 @@ void kvm_async_pf_task_wait(u32 token)
        u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
        struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
        struct kvm_task_sleep_node n, *e;
-       DEFINE_WAIT(wait);
+       DECLARE_SWAITQUEUE(wait);
 
        rcu_irq_enter();
 
-       spin_lock(&b->lock);
+       raw_spin_lock(&b->lock);
        e = _find_apf_task(b, token);
        if (e) {
                /* dummy entry exist -> wake up was delivered ahead of PF */
                hlist_del(&e->link);
                kfree(e);
-               spin_unlock(&b->lock);
+               raw_spin_unlock(&b->lock);
 
                rcu_irq_exit();
                return;
@@ -141,13 +142,13 @@ 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;
-       init_waitqueue_head(&n.wq);
+       init_swait_queue_head(&n.wq);
        hlist_add_head(&n.link, &b->list);
-       spin_unlock(&b->lock);
+       raw_spin_unlock(&b->lock);
 
        for (;;) {
                if (!n.halted)
-                       prepare_to_wait(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
+                       prepare_to_swait(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
                if (hlist_unhashed(&n.link))
                        break;
 
@@ -166,7 +167,7 @@ void kvm_async_pf_task_wait(u32 token)
                }
        }
        if (!n.halted)
-               finish_wait(&n.wq, &wait);
+               finish_swait(&n.wq, &wait);
 
        rcu_irq_exit();
        return;
@@ -178,8 +179,8 @@ static void apf_task_wake_one(struct kvm_task_sleep_node *n)
        hlist_del_init(&n->link);
        if (n->halted)
                smp_send_reschedule(n->cpu);
-       else if (waitqueue_active(&n->wq))
-               wake_up(&n->wq);
+       else if (swait_active(&n->wq))
+               swake_up(&n->wq);
 }
 
 static void apf_task_wake_all(void)
@@ -189,14 +190,14 @@ static void apf_task_wake_all(void)
        for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
                struct hlist_node *p, *next;
                struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
-               spin_lock(&b->lock);
+               raw_spin_lock(&b->lock);
                hlist_for_each_safe(p, next, &b->list) {
                        struct kvm_task_sleep_node *n =
                                hlist_entry(p, typeof(*n), link);
                        if (n->cpu == smp_processor_id())
                                apf_task_wake_one(n);
                }
-               spin_unlock(&b->lock);
+               raw_spin_unlock(&b->lock);
        }
 }
 
@@ -212,7 +213,7 @@ void kvm_async_pf_task_wake(u32 token)
        }
 
 again:
-       spin_lock(&b->lock);
+       raw_spin_lock(&b->lock);
        n = _find_apf_task(b, token);
        if (!n) {
                /*
@@ -225,17 +226,17 @@ again:
                         * Allocation failed! Busy wait while other cpu
                         * handles async PF.
                         */
-                       spin_unlock(&b->lock);
+                       raw_spin_unlock(&b->lock);
                        cpu_relax();
                        goto again;
                }
                n->token = token;
                n->cpu = smp_processor_id();
-               init_waitqueue_head(&n->wq);
+               init_swait_queue_head(&n->wq);
                hlist_add_head(&n->link, &b->list);
        } else
                apf_task_wake_one(n);
-       spin_unlock(&b->lock);
+       raw_spin_unlock(&b->lock);
        return;
 }
 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
@@ -486,7 +487,7 @@ void __init kvm_guest_init(void)
        paravirt_ops_setup();
        register_reboot_notifier(&kvm_pv_reboot_nb);
        for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
-               spin_lock_init(&async_pf_sleepers[i].lock);
+               raw_spin_lock_init(&async_pf_sleepers[i].lock);
        if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
                x86_init.irqs.trap_init = kvm_apf_trap_init;