On Mon, Dec 03, 2012 at 10:08:32AM +0800, Li Zhong wrote:

On Fri, 2012-11-30 at 12:26 +0200, Gleb Natapov wrote:

quoted

On Fri, Nov 30, 2012 at 05:18:41PM +0800, Li Zhong wrote:

quoted

This patch adds user eqs exception hooks for async page fault page not
present code path, to exit the user eqs and re-enter it as necessary. 

Async page fault is different from other exceptions that it may be
triggered from idle process, so we still need rcu_irq_enter() and
rcu_irq_exit() to exit cpu idle eqs when needed, to protect the code
that needs use rcu.

As Frederic pointed out it would be safest and simplest to protect the
whole kvm_async_pf_task_wait(). Otherwise, "we need to check all the
code there deeply for potential RCU uses and ensure it will never be
extended later to use RCU.".

However, We'd better re-enter the cpu idle eqs if we get the exception
in cpu idle eqs, by calling rcu_irq_exit() before native_safe_halt(). 

So the patch does what Frederic suggested for rcu_irq_*() API usage
here, except that I moved the rcu_irq_*() pair originally in
do_async_page_fault() into kvm_async_pf_task_wait(). 

That's because, I think it's better to have rcu_irq_*() pairs to be in
one function ( rcu_irq_exit() after rcu_irq_enter() ), especially here,
kvm_async_pf_task_wait() has other callers, which might cause
rcu_irq_exit() be called without a matching rcu_irq_enter() before it,
which is illegal if the cpu happens to be in rcu idle state. 

Signed-off-by: Li Zhong <redacted>
---
 arch/x86/kernel/kvm.c | 12 ++++++++++--
 1 file changed, 10 insertions(+), 2 deletions(-)

diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c
index 4180a87..342b00b 100644
--- a/arch/x86/kernel/kvm.c
+++ b/arch/x86/kernel/kvm.c

@@ -42,6 +42,7 @@
 #include <asm/apic.h>
 #include <asm/apicdef.h>
 #include <asm/hypervisor.h>
+#include <asm/rcu.h>
 
 static int kvmapf = 1;
 

@@ -112,6 +113,8 @@ void kvm_async_pf_task_wait(u32 token)
 	DEFINE_WAIT(wait);
 	int cpu, idle;
 
+	rcu_irq_enter();
+

Why move rcu_irq_*() calls into kvm_async_pf_task_wait()?

I think it is not good for a function to have a rcu_irq_exit(), which
needs a matching rcu_irq_enter() in its caller. 

Here, if not move rcu_irq_* in, then the rcu_irq_exit() before
native_safe_halt() in kvm_async_pf_task_wait() is the one that needs the
matching rcu_irq_enter() in do_async_page_fault(). And, for this case,
kvm_async_pf_task_wait() even has other caller - pf_interception().
Maybe it will always be rcu non-idle for pf_interception (so a matching
rcu_irq_enter() is not needed), or maybe we could (or need) add
rcu_irq_*() in pf_interception().  But I still think it's good to have
those function calls that need to be matched be contained in one
function.

kvm_async_pf_task_wait() call from pf_interception() will always go to
schedule() path. I get your point and am fine with the patch as is.

Thanks, Zhong

quoted

quoted

 	cpu = get_cpu();
 	idle = idle_cpu(cpu);
 	put_cpu();

@@ -123,6 +126,8 @@ void kvm_async_pf_task_wait(u32 token)
 		hlist_del(&e->link);
 		kfree(e);
 		spin_unlock(&b->lock);
+
+		rcu_irq_exit();

We can skip that if  rcu_irq_*() will stay outside.

quoted

 		return;
 	}
 

@@ -147,13 +152,16 @@ void kvm_async_pf_task_wait(u32 token)
 			/*
 			 * We cannot reschedule. So halt.
 			 */
+			rcu_irq_exit();
 			native_safe_halt();
+			rcu_irq_enter();
 			local_irq_disable();
 		}
 	}
 	if (!n.halted)
 		finish_wait(&n.wq, &wait);
 
+	rcu_irq_exit();
 	return;
 }
 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);

@@ -247,10 +255,10 @@ do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
 		break;
 	case KVM_PV_REASON_PAGE_NOT_PRESENT:
 		/* page is swapped out by the host. */
-		rcu_irq_enter();
+		exception_enter(regs);
 		exit_idle();
 		kvm_async_pf_task_wait((u32)read_cr2());
-		rcu_irq_exit();
+		exception_exit(regs);
 		break;
 	case KVM_PV_REASON_PAGE_READY:
 		rcu_irq_enter();

-- 
1.7.11.4

--
			Gleb.

--
			Gleb.