Skip to contentSkip to conversation

KVM: x86: remove irq disablement around KVM_SET_CLOCK/KVM_GET_CLOCK

PATCH Superseded Marcelo Tosatti 5 participants 11 messages Apr 12, 2017 – Apr 21, 2017

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 KVM: x86: remove irq disablement around KVM_SET_CLOCK/KVM_GET_CLOCK

The disablement of interrupts at KVM_SET_CLOCK/KVM_GET_CLOCK attempts to disable interrupts in that section to protect
the values that are calculated in that section from interrupt interference.

now_ns is calculated inside the irq protected region, user_ns.clock is passed from userspace (therefore not susceptible to interrupt variation).

About the line

                now_ns = __get_kvmclock_ns(kvm);  (1)

Interrupts can happen afterwards local_irq_enable(), rendering "now_ns" relative to its execution time PLUS interrupt time.

Therefore the local_irq_disable() / local_irq_enable() protection is not necessary (that is: interrupts triggering after local_irq_enable cause the same problem that the protection is trying to avoid).

With this reasoning, and the -RT bug that the irq disablement causes (because spin_lock is now a sleeping lock), remove the IRQ protection as it causes:

[ 1064.668109] in_atomic(): 0, irqs_disabled(): 1, pid: 15296, name:m
[ 1064.668110] INFO: lockdep is turned off.
[ 1064.668110] irq event stamp: 0
[ 1064.668112] hardirqs last  enabled at (0): [<          (null)>]  )
[ 1064.668116] hardirqs last disabled at (0): [<ffffffff9308184a>] c0
[ 1064.668118] softirqs last  enabled at (0): [<ffffffff9308184a>] c0
[ 1064.668118] softirqs last disabled at (0): [<          (null)>]  )
[ 1064.668121] CPU: 13 PID: 15296 Comm: qemu-kvm Not tainted 3.10.0-1
[ 1064.668121] Hardware name: Dell Inc. PowerEdge R730/0H21J3, BIOS 5
[ 1064.668123]  ffff8c1796b88000 00000000afe7344c ffff8c179abf3c68 f3
[ 1064.668125]  ffff8c179abf3c90 ffffffff930ccb3d ffff8c1b992b3610 f0
[ 1064.668126]  00007ffc1a26fbc0 ffff8c179abf3cb0 ffffffff9375f694 f0
[ 1064.668126] Call Trace:
[ 1064.668132]  [<ffffffff93757413>] dump_stack+0x19/0x1b
[ 1064.668135]  [<ffffffff930ccb3d>] __might_sleep+0x12d/0x1f0
[ 1064.668138]  [<ffffffff9375f694>] rt_spin_lock+0x24/0x60
[ 1064.668155]  [<ffffffffc06ab996>] __get_kvmclock_ns+0x36/0x110 [k]
[ 1064.668159]  [<ffffffff93112993>] ? futex_wait_queue_me+0x103/0x10
[ 1064.668171]  [<ffffffffc06b8782>] kvm_arch_vm_ioctl+0xa2/0xd70 [k]
[ 1064.668173]  [<ffffffff9311333c>] ? futex_wait+0x1ac/0x2a0

On -RT kernels.

---

arch/x86/kvm/x86.c | 4 ----

 1 file changed, 4 deletions(-)
arch/x86/kvm/x86.c +0 -4
index 7c39b8a..935143b 100644
@@ -4062,10 +4062,8 @@ long kvm_arch_vm_ioctl(struct file *filp,
 			goto out;
 
 		r = 0;
-		local_irq_disable();
 		now_ns = __get_kvmclock_ns(kvm);
 		kvm->arch.kvmclock_offset += user_ns.clock - now_ns;
-		local_irq_enable();
 		kvm_gen_update_masterclock(kvm);
 		break;
 	}
@@ -4073,11 +4071,9 @@ long kvm_arch_vm_ioctl(struct file *filp,
 		struct kvm_clock_data user_ns;
 		u64 now_ns;
 
-		local_irq_disable();
 		now_ns = __get_kvmclock_ns(kvm);
 		user_ns.clock = now_ns;
 		user_ns.flags = kvm->arch.use_master_clock ? KVM_CLOCK_TSC_STABLE : 0;
-		local_irq_enable();
 		memset(&user_ns.pad, 0, sizeof(user_ns.pad));
 
 		r = -EFAULT;
#1 lore raw

2017-04-12 14:23-0300, Marcelo Tosatti:

Quoted 11 lines

The disablement of interrupts at KVM_SET_CLOCK/KVM_GET_CLOCK attempts to disable interrupts in that section to protect
the values that are calculated in that section from interrupt interference.

now_ns is calculated inside the irq protected region, user_ns.clock is passed from userspace (therefore not susceptible to interrupt variation).

About the line

                now_ns = __get_kvmclock_ns(kvm);  (1)
Quoted 7 lines

Interrupts can happen afterwards local_irq_enable(), rendering "now_ns" relative to its execution time PLUS interrupt time.

Therefore the local_irq_disable() / local_irq_enable() protection is not necessary (that is: interrupts triggering after local_irq_enable cause the same problem that the protection is trying to avoid).

Makes sense.

get_kvmclock_ns() has the same problem and this patch should cover it as well.

#2 lore raw
In reply to Marcelo Tosatti
Quoted 10 lines

The disablement of interrupts at KVM_SET_CLOCK/KVM_GET_CLOCK attempts to disable interrupts in that section to protect
the values that are calculated in that section from interrupt interference.

now_ns is calculated inside the irq protected region, user_ns.clock is passed from userspace (therefore not susceptible to interrupt variation).

Quoted 6 lines

About the line

                now_ns = __get_kvmclock_ns(kvm);  (1)

Interrupts can happen afterwards local_irq_enable(), rendering "now_ns" relative to its execution time PLUS interrupt time.

Are you saying interrupt disable code here is to avoid any interrupt latency before we update time?

Quoted 4 lines

Therefore the local_irq_disable() / local_irq_enable() protection is not necessary (that is: interrupts triggering after local_irq_enable cause the same problem that the protection is trying to avoid).

I am not sure because this could be different use-case for RT and mainline. RT can afford to preempt, but if we do will we get desired results?

Maybe if we want this code path to be "irq safe", can we use 'raw_spin_lock' in place of 'spin_lock'.

Quoted 17 lines

With this reasoning, and the -RT bug that the irq disablement causes (because spin_lock is now a sleeping lock), remove the IRQ protection as it causes:

[ 1064.668109] in_atomic(): 0, irqs_disabled(): 1, pid: 15296, name:m
[ 1064.668110] INFO: lockdep is turned off.
[ 1064.668110] irq event stamp: 0
[ 1064.668112] hardirqs last  enabled at (0): [<          (null)>]  )
[ 1064.668116] hardirqs last disabled at (0): [] c0
[ 1064.668118] softirqs last  enabled at (0): [] c0
[ 1064.668118] softirqs last disabled at (0): [<          (null)>]  )
[ 1064.668121] CPU: 13 PID: 15296 Comm: qemu-kvm Not tainted 3.10.0-1
[ 1064.668121] Hardware name: Dell Inc. PowerEdge R730/0H21J3, BIOS 5
[ 1064.668123]  ffff8c1796b88000 00000000afe7344c ffff8c179abf3c68 f3
[ 1064.668125]  ffff8c179abf3c90 ffffffff930ccb3d ffff8c1b992b3610 f0
[ 1064.668126]  00007ffc1a26fbc0 ffff8c179abf3cb0 ffffffff9375f694 f0
Quoted 8 lines
[ 1064.668126] Call Trace:
[ 1064.668132]  [] dump_stack+0x19/0x1b
[ 1064.668135]  [] __might_sleep+0x12d/0x1f0
[ 1064.668138]  [] rt_spin_lock+0x24/0x60
[ 1064.668155]  [] __get_kvmclock_ns+0x36/0x110 [k]
[ 1064.668159]  [] ? futex_wait_queue_me+0x103/0x10
[ 1064.668171]  [] kvm_arch_vm_ioctl+0xa2/0xd70 [k]
[ 1064.668173]  [] ? futex_wait+0x1ac/0x2a0

Even with this patch, we might get same error for:

get_kvmclock_ns {

 local_irq_save(flags);
        ns = __get_kvmclock_ns(kvm);
        local_irq_restore(flags);

}

Quoted 39 lines

On -RT kernels.

---

arch/x86/kvm/x86.c | 4 ----

 1 file changed, 4 deletions(-)
arch/x86/kvm/x86.c +0 -4
index 7c39b8a..935143b 100644
@@ -4062,10 +4062,8 @@ long kvm_arch_vm_ioctl(struct file *filp,
                         goto out;
 
                 r = 0;
-                local_irq_disable();
                 now_ns = __get_kvmclock_ns(kvm);
                 kvm->arch.kvmclock_offset += user_ns.clock - now_ns;
-                local_irq_enable();
                 kvm_gen_update_masterclock(kvm);
                 break;
         }
@@ -4073,11 +4071,9 @@ long kvm_arch_vm_ioctl(struct file *filp,
                 struct kvm_clock_data user_ns;
                 u64 now_ns;
 
-                local_irq_disable();
                 now_ns = __get_kvmclock_ns(kvm);
                 user_ns.clock = now_ns;
                 user_ns.flags = kvm->arch.use_master_clock ? KVM_CLOCK_TSC_STABLE : 0;
-                local_irq_enable();
                 memset(&user_ns.pad, 0, sizeof(user_ns.pad));
 
                 r = -EFAULT;
#3 lore raw
Quoted 24 lines · Pankaj Gupta
Quoted 10 lines

The disablement of interrupts at KVM_SET_CLOCK/KVM_GET_CLOCK attempts to disable interrupts in that section to protect
the values that are calculated in that section from interrupt interference.

now_ns is calculated inside the irq protected region, user_ns.clock is passed from userspace (therefore not susceptible to interrupt variation).

Quoted 6 lines

About the line

                now_ns = __get_kvmclock_ns(kvm);  (1)

Interrupts can happen afterwards local_irq_enable(), rendering "now_ns" relative to its execution time PLUS interrupt time.

Are you saying interrupt disable code here is to avoid any interrupt latency before we update time?

Quoted 4 lines

Therefore the local_irq_disable() / local_irq_enable() protection is not necessary (that is: interrupts triggering after local_irq_enable cause the same problem that the protection is trying to avoid).

Quoted 5 lines

I am not sure because this could be different use-case for RT and mainline. RT can afford to preempt, but if we do will we get desired results?

Maybe if we want this code path to be "irq safe", can we use 'raw_spin_lock' in place of 'spin_lock'.

commit 395c6b0a9d56fe7fdb7aeda12795d0eb02475d24

Author: Avi Kivity <[email protected]>
Date:   Mon Oct 4 12:55:49 2010 +0200

KVM: Disable interrupts around get_kernel_ns()

get_kernel_ns() wants preemption disabled. It doesn't make a lot of
sense
during the get/set ioctls (no way to make them non-racy) but the
callee wants
it.

And

static inline u64 get_kernel_ns(void)

{               
        struct timespec ts;

        WARN_ON(preemptible());
        ktime_get_ts(&ts);
        monotonic_to_bootbased(&ts);
        return timespec_to_ns(&ts);
}

Which was added by:
Add a helper function to compute the kernel time and convert nanoseconds back to CPU specific cycles. Note that these must not be called in preemptible context, as that would mean the kernel could enter software suspend state, which would cause non-atomic operation.

However this comment makes no sense: ktime_get_ts(&ts) returns a clock value relative to either before or after software suspend (either case is fine).

Quoted 27 lines
Quoted 16 lines

With this reasoning, and the -RT bug that the irq disablement causes (because spin_lock is now a sleeping lock), remove the IRQ protection as it causes:

[ 1064.668109] in_atomic(): 0, irqs_disabled(): 1, pid: 15296, name:m
[ 1064.668110] INFO: lockdep is turned off.
[ 1064.668110] irq event stamp: 0
[ 1064.668112] hardirqs last  enabled at (0): [<          (null)>]  )
[ 1064.668116] hardirqs last disabled at (0): [] c0
[ 1064.668118] softirqs last  enabled at (0): [] c0
[ 1064.668118] softirqs last disabled at (0): [<          (null)>]  )
[ 1064.668121] CPU: 13 PID: 15296 Comm: qemu-kvm Not tainted 3.10.0-1
[ 1064.668121] Hardware name: Dell Inc. PowerEdge R730/0H21J3, BIOS 5
[ 1064.668123]  ffff8c1796b88000 00000000afe7344c ffff8c179abf3c68 f3
[ 1064.668125]  ffff8c179abf3c90 ffffffff930ccb3d ffff8c1b992b3610 f0
[ 1064.668126]  00007ffc1a26fbc0 ffff8c179abf3cb0 ffffffff9375f694 f0
Quoted 8 lines
[ 1064.668126] Call Trace:
[ 1064.668132]  [] dump_stack+0x19/0x1b
[ 1064.668135]  [] __might_sleep+0x12d/0x1f0
[ 1064.668138]  [] rt_spin_lock+0x24/0x60
[ 1064.668155]  [] __get_kvmclock_ns+0x36/0x110 [k]
[ 1064.668159]  [] ? futex_wait_queue_me+0x103/0x10
[ 1064.668171]  [] kvm_arch_vm_ioctl+0xa2/0xd70 [k]
[ 1064.668173]  [] ? futex_wait+0x1ac/0x2a0

Even with this patch, we might get same error for:

Quoted 8 lines

get_kvmclock_ns {

 local_irq_save(flags);
        ns = __get_kvmclock_ns(kvm);
        local_irq_restore(flags);

}

Yes, should fix that as well.

Quoted 39 lines
Quoted 39 lines

On -RT kernels.

---

arch/x86/kvm/x86.c | 4 ----

 1 file changed, 4 deletions(-)
arch/x86/kvm/x86.c +0 -4
index 7c39b8a..935143b 100644
@@ -4062,10 +4062,8 @@ long kvm_arch_vm_ioctl(struct file *filp,
                         goto out;
 
                 r = 0;
-                local_irq_disable();
                 now_ns = __get_kvmclock_ns(kvm);
                 kvm->arch.kvmclock_offset += user_ns.clock - now_ns;
-                local_irq_enable();
                 kvm_gen_update_masterclock(kvm);
                 break;
         }
@@ -4073,11 +4071,9 @@ long kvm_arch_vm_ioctl(struct file *filp,
                 struct kvm_clock_data user_ns;
                 u64 now_ns;
 
-                local_irq_disable();
                 now_ns = __get_kvmclock_ns(kvm);
                 user_ns.clock = now_ns;
                 user_ns.flags = kvm->arch.use_master_clock ? KVM_CLOCK_TSC_STABLE : 0;
-                local_irq_enable();
                 memset(&user_ns.pad, 0, sizeof(user_ns.pad));
 
                 r = -EFAULT;
#4 lore raw

PATCH v2 1/1 KVM: x86: remove irq disablement around KVM_SET_CLOCK/KVM_GET_CLOCK

In reply to Pankaj Gupta

[PATCH v2] KVM: x86: remove irq disablement around KVM_SET_CLOCK/KVM_GET_CLOCK

The disablement of interrupts at KVM_SET_CLOCK/KVM_GET_CLOCK attempts to disable software suspend from causing "non atomic behaviour" of the operation:

Add a helper function to compute the kernel time and convert nanoseconds back to CPU specific cycles. Note that these must not be called in preemptible context, as that would mean the kernel could enter software suspend state, which would cause non-atomic operation.

However, assume the kernel can enter software suspend at the following 2 points:

        ktime_get_ts(&ts);
1.
						hypothetical_ktime_get_ts(&ts)
        monotonic_to_bootbased(&ts);

2.

monotonic_to_bootbased() should be correct relative to a ktime_get_ts(&ts) performed after point 1 (that is after resuming from software suspend), hypothetical_ktime_get_ts()

Therefore it is also correct for the ktime_get_ts(&ts) before point 1, which is

ktime_get_ts(&ts) = hypothetical_ktime_get_ts(&ts) + time-to-execute-suspend-code

Note CLOCK_MONOTONIC does not count during suspension.

So remove the irq disablement, which causes the following warning on -RT kernels:

With this reasoning, and the -RT bug that the irq disablement causes (because spin_lock is now a sleeping lock), remove the IRQ protection as it causes:

 [ 1064.668109] in_atomic(): 0, irqs_disabled(): 1, pid: 15296, name:m
 [ 1064.668110] INFO: lockdep is turned off.
 [ 1064.668110] irq event stamp: 0
 [ 1064.668112] hardirqs last  enabled at (0): [<          (null)>]  )
 [ 1064.668116] hardirqs last disabled at (0): [] c0
 [ 1064.668118] softirqs last  enabled at (0): [] c0
 [ 1064.668118] softirqs last disabled at (0): [<          (null)>]  )
 [ 1064.668121] CPU: 13 PID: 15296 Comm: qemu-kvm Not tainted 3.10.0-1
 [ 1064.668121] Hardware name: Dell Inc. PowerEdge R730/0H21J3, BIOS 5
 [ 1064.668123]  ffff8c1796b88000 00000000afe7344c ffff8c179abf3c68 f3
 [ 1064.668125]  ffff8c179abf3c90 ffffffff930ccb3d ffff8c1b992b3610 f0
 [ 1064.668126]  00007ffc1a26fbc0 ffff8c179abf3cb0 ffffffff9375f694 f0
 [ 1064.668126] Call Trace:
 [ 1064.668132]  [] dump_stack+0x19/0x1b
 [ 1064.668135]  [] __might_sleep+0x12d/0x1f0
 [ 1064.668138]  [] rt_spin_lock+0x24/0x60
 [ 1064.668155]  [] __get_kvmclock_ns+0x36/0x110 [k]
 [ 1064.668159]  [] ? futex_wait_queue_me+0x103/0x10
 [ 1064.668171]  [] kvm_arch_vm_ioctl+0xa2/0xd70 [k]
 [ 1064.668173]  [] ? futex_wait+0x1ac/0x2a0

v2: notice get_kvmclock_ns with the same problem (Pankaj).

arch/x86/kvm/x86.c +0 -7
index 1faf620..4901178 100644
@@ -1798,12 +1798,9 @@ static u64 __get_kvmclock_ns(struct kvm *kvm)
 
 u64 get_kvmclock_ns(struct kvm *kvm)
 {
-	unsigned long flags;
 	s64 ns;
 
-	local_irq_save(flags);
 	ns = __get_kvmclock_ns(kvm);
-	local_irq_restore(flags);
 
 	return ns;
 }
@@ -4196,10 +4193,8 @@ long kvm_arch_vm_ioctl(struct file *filp,
 			goto out;
 
 		r = 0;
-		local_irq_disable();
 		now_ns = __get_kvmclock_ns(kvm);
 		kvm->arch.kvmclock_offset += user_ns.clock - now_ns;
-		local_irq_enable();
 		kvm_gen_update_masterclock(kvm);
 		break;
 	}
@@ -4207,11 +4202,9 @@ long kvm_arch_vm_ioctl(struct file *filp,
 		struct kvm_clock_data user_ns;
 		u64 now_ns;
 
-		local_irq_disable();
 		now_ns = __get_kvmclock_ns(kvm);
 		user_ns.clock = now_ns;
 		user_ns.flags = kvm->arch.use_master_clock ? KVM_CLOCK_TSC_STABLE : 0;
-		local_irq_enable();
 		memset(&user_ns.pad, 0, sizeof(user_ns.pad));
 
 		r = -EFAULT;
#5 lore raw
Quoted 87 lines

[PATCH v2] KVM: x86: remove irq disablement around KVM_SET_CLOCK/KVM_GET_CLOCK

The disablement of interrupts at KVM_SET_CLOCK/KVM_GET_CLOCK attempts to disable software suspend from causing "non atomic behaviour" of the operation:

Add a helper function to compute the kernel time and convert nanoseconds back to CPU specific cycles. Note that these must not be called in preemptible
context, as that would mean the kernel could enter software suspend state,
which would cause non-atomic operation.

However, assume the kernel can enter software suspend at the following 2 points:

        ktime_get_ts(&ts);
1.
						hypothetical_ktime_get_ts(&ts)
        monotonic_to_bootbased(&ts);

2.

monotonic_to_bootbased() should be correct relative to a ktime_get_ts(&ts) performed after point 1 (that is after resuming from software suspend), hypothetical_ktime_get_ts()

Therefore it is also correct for the ktime_get_ts(&ts) before point 1, which is

ktime_get_ts(&ts) = hypothetical_ktime_get_ts(&ts) + time-to-execute-suspend-code

Note CLOCK_MONOTONIC does not count during suspension.

So remove the irq disablement, which causes the following warning on -RT kernels:

With this reasoning, and the -RT bug that the irq disablement causes (because spin_lock is now a sleeping lock), remove the IRQ protection as it causes:

 [ 1064.668109] in_atomic(): 0, irqs_disabled(): 1, pid: 15296, name:m
 [ 1064.668110] INFO: lockdep is turned off.
 [ 1064.668110] irq event stamp: 0
 [ 1064.668112] hardirqs last  enabled at (0): [<          (null)>]  )
 [ 1064.668116] hardirqs last disabled at (0): [] c0
 [ 1064.668118] softirqs last  enabled at (0): [] c0
 [ 1064.668118] softirqs last disabled at (0): [<          (null)>]  )
 [ 1064.668121] CPU: 13 PID: 15296 Comm: qemu-kvm Not tainted 3.10.0-1
 [ 1064.668121] Hardware name: Dell Inc. PowerEdge R730/0H21J3, BIOS 5
 [ 1064.668123]  ffff8c1796b88000 00000000afe7344c ffff8c179abf3c68 f3
 [ 1064.668125]  ffff8c179abf3c90 ffffffff930ccb3d ffff8c1b992b3610 f0
 [ 1064.668126]  00007ffc1a26fbc0 ffff8c179abf3cb0 ffffffff9375f694 f0
 [ 1064.668126] Call Trace:
 [ 1064.668132]  [] dump_stack+0x19/0x1b
 [ 1064.668135]  [] __might_sleep+0x12d/0x1f0
 [ 1064.668138]  [] rt_spin_lock+0x24/0x60
 [ 1064.668155]  [] __get_kvmclock_ns+0x36/0x110 [k]
 [ 1064.668159]  [] ? futex_wait_queue_me+0x103/0x10
 [ 1064.668171]  [] kvm_arch_vm_ioctl+0xa2/0xd70 [k]
 [ 1064.668173]  [] ? futex_wait+0x1ac/0x2a0

v2: notice get_kvmclock_ns with the same problem (Pankaj).

arch/x86/kvm/x86.c +0 -3
index 1faf620..4901178 100644
@@ -1798,12 +1798,9 @@ static u64 __get_kvmclock_ns(struct kvm *kvm)
 
 u64 get_kvmclock_ns(struct kvm *kvm)
 {
-	unsigned long flags;
 	s64 ns;
 
-	local_irq_save(flags);
 	ns = __get_kvmclock_ns(kvm);
-	local_irq_restore(flags);
 
 	return ns;
 }

arch/x86/kvm/x86.c long kvm_arch_vm_ioctl(struct file *filp, · line 4196

Quoted 5 lines
arch/x86/kvm/x86.c +0 -1
@@ -4196,10 +4193,8 @@ long kvm_arch_vm_ioctl(struct file *filp,
 			goto out;
 
 		r = 0;
-		local_irq_disable();
 		now_ns = __get_kvmclock_ns(kvm);

As we have wrapper 'get_kvmclock_ns' above '__get_kvmclock_ns(kvm)' we can use that directly.

arch/x86/kvm/x86.c long kvm_arch_vm_ioctl(struct file *filp, · line 4205

Quoted 11 lines
arch/x86/kvm/x86.c +0 -2
 		kvm->arch.kvmclock_offset += user_ns.clock - now_ns;
-		local_irq_enable();
 		kvm_gen_update_masterclock(kvm);
 		break;
 	}
@@ -4207,11 +4202,9 @@ long kvm_arch_vm_ioctl(struct file *filp,
 		struct kvm_clock_data user_ns;
 		u64 now_ns;
 
-		local_irq_disable();
 		now_ns = __get_kvmclock_ns(kvm);

wrapper here as well.

arch/x86/kvm/x86.c long kvm_arch_vm_ioctl(struct file *filp, · line 4210

Quoted 6 lines
arch/x86/kvm/x86.c +0 -1
 		user_ns.clock = now_ns;
 		user_ns.flags = kvm->arch.use_master_clock ? KVM_CLOCK_TSC_STABLE : 0;
-		local_irq_enable();
 		memset(&user_ns.pad, 0, sizeof(user_ns.pad));
 
 		r = -EFAULT;

This patch looks okay to me. Just a small suggestion above.

Thanks,
Pankaj

Quoted 1 line
#6 lore raw

PATCH v3 1/1 KVM: x86: remove irq disablement around KVM_SET_CLOCK/KVM_GET_CLOCK

The disablement of interrupts at KVM_SET_CLOCK/KVM_GET_CLOCK attempts to disable software suspend from causing "non atomic behaviour" of the operation:

Add a helper function to compute the kernel time and convert nanoseconds back to CPU specific cycles. Note that these must not be called in preemptible context, as that would mean the kernel could enter software suspend state, which would cause non-atomic operation.

However, assume the kernel can enter software suspend at the following 2 points:

        ktime_get_ts(&ts);
1.
						hypothetical_ktime_get_ts(&ts)
        monotonic_to_bootbased(&ts);

2.

monotonic_to_bootbased() should be correct relative to a ktime_get_ts(&ts) performed after point 1 (that is after resuming from software suspend), hypothetical_ktime_get_ts()

Therefore it is also correct for the ktime_get_ts(&ts) before point 1, which is

ktime_get_ts(&ts) = hypothetical_ktime_get_ts(&ts) + time-to-execute-suspend-code

Note CLOCK_MONOTONIC does not count during suspension.

So remove the irq disablement, which causes the following warning on -RT kernels:

With this reasoning, and the -RT bug that the irq disablement causes (because spin_lock is now a sleeping lock), remove the IRQ protection as it causes:

 [ 1064.668109] in_atomic(): 0, irqs_disabled(): 1, pid: 15296, name:m
 [ 1064.668110] INFO: lockdep is turned off.
 [ 1064.668110] irq event stamp: 0
 [ 1064.668112] hardirqs last  enabled at (0): [<          (null)>]  )
 [ 1064.668116] hardirqs last disabled at (0): [] c0
 [ 1064.668118] softirqs last  enabled at (0): [] c0
 [ 1064.668118] softirqs last disabled at (0): [<          (null)>]  )
 [ 1064.668121] CPU: 13 PID: 15296 Comm: qemu-kvm Not tainted 3.10.0-1
 [ 1064.668121] Hardware name: Dell Inc. PowerEdge R730/0H21J3, BIOS 5
 [ 1064.668123]  ffff8c1796b88000 00000000afe7344c ffff8c179abf3c68 f3
 [ 1064.668125]  ffff8c179abf3c90 ffffffff930ccb3d ffff8c1b992b3610 f0
 [ 1064.668126]  00007ffc1a26fbc0 ffff8c179abf3cb0 ffffffff9375f694 f0
 [ 1064.668126] Call Trace:
 [ 1064.668132]  [] dump_stack+0x19/0x1b
 [ 1064.668135]  [] __might_sleep+0x12d/0x1f0
 [ 1064.668138]  [] rt_spin_lock+0x24/0x60
 [ 1064.668155]  [] __get_kvmclock_ns+0x36/0x110 [k]
 [ 1064.668159]  [] ? futex_wait_queue_me+0x103/0x10
 [ 1064.668171]  [] kvm_arch_vm_ioctl+0xa2/0xd70 [k]
 [ 1064.668173]  [] ? futex_wait+0x1ac/0x2a0

v2: notice get_kvmclock_ns with the same problem (Pankaj).
v3: remove useless helper function (Pankaj).

arch/x86/kvm/x86.c +3 -19
index 1faf620..deacbe1 100644
@@ -1775,7 +1775,7 @@ static void kvm_gen_update_masterclock(struct kvm *kvm)
 #endif
 }
 
-static u64 __get_kvmclock_ns(struct kvm *kvm)
+u64 get_kvmclock_ns(struct kvm *kvm)
 {
 	struct kvm_arch *ka = &kvm->arch;
 	struct pvclock_vcpu_time_info hv_clock;
@@ -1796,18 +1796,6 @@ static u64 __get_kvmclock_ns(struct kvm *kvm)
 	return __pvclock_read_cycles(&hv_clock, rdtsc());
 }
 
-u64 get_kvmclock_ns(struct kvm *kvm)
-{
-	unsigned long flags;
-	s64 ns;
-
-	local_irq_save(flags);
-	ns = __get_kvmclock_ns(kvm);
-	local_irq_restore(flags);
-
-	return ns;
-}
-
 static void kvm_setup_pvclock_page(struct kvm_vcpu *v)
 {
 	struct kvm_vcpu_arch *vcpu = &v->arch;
@@ -4196,10 +4184,8 @@ long kvm_arch_vm_ioctl(struct file *filp,
 			goto out;
 
 		r = 0;
-		local_irq_disable();
-		now_ns = __get_kvmclock_ns(kvm);
+		now_ns = get_kvmclock_ns(kvm);
 		kvm->arch.kvmclock_offset += user_ns.clock - now_ns;
-		local_irq_enable();
 		kvm_gen_update_masterclock(kvm);
 		break;
 	}
@@ -4207,11 +4193,9 @@ long kvm_arch_vm_ioctl(struct file *filp,
 		struct kvm_clock_data user_ns;
 		u64 now_ns;
 
-		local_irq_disable();
-		now_ns = __get_kvmclock_ns(kvm);
+		now_ns = get_kvmclock_ns(kvm);
 		user_ns.clock = now_ns;
 		user_ns.flags = kvm->arch.use_master_clock ? KVM_CLOCK_TSC_STABLE : 0;
-		local_irq_enable();
 		memset(&user_ns.pad, 0, sizeof(user_ns.pad));
 
 		r = -EFAULT;
Reviewed-by on v3 1/1
#7 lore raw
Quoted 114 lines

The disablement of interrupts at KVM_SET_CLOCK/KVM_GET_CLOCK attempts to disable software suspend from causing "non atomic behaviour" of the operation:

Add a helper function to compute the kernel time and convert nanoseconds back to CPU specific cycles. Note that these must not be called in preemptible
context, as that would mean the kernel could enter software suspend state,
which would cause non-atomic operation.

However, assume the kernel can enter software suspend at the following 2 points:

        ktime_get_ts(&ts);
1.
						hypothetical_ktime_get_ts(&ts)
        monotonic_to_bootbased(&ts);

2.

monotonic_to_bootbased() should be correct relative to a ktime_get_ts(&ts) performed after point 1 (that is after resuming from software suspend), hypothetical_ktime_get_ts()

Therefore it is also correct for the ktime_get_ts(&ts) before point 1, which is

ktime_get_ts(&ts) = hypothetical_ktime_get_ts(&ts) + time-to-execute-suspend-code

Note CLOCK_MONOTONIC does not count during suspension.

So remove the irq disablement, which causes the following warning on -RT kernels:

With this reasoning, and the -RT bug that the irq disablement causes (because spin_lock is now a sleeping lock), remove the IRQ protection as it causes:

 [ 1064.668109] in_atomic(): 0, irqs_disabled(): 1, pid: 15296, name:m
 [ 1064.668110] INFO: lockdep is turned off.
 [ 1064.668110] irq event stamp: 0
 [ 1064.668112] hardirqs last  enabled at (0): [<          (null)>]  )
 [ 1064.668116] hardirqs last disabled at (0): [] c0
 [ 1064.668118] softirqs last  enabled at (0): [] c0
 [ 1064.668118] softirqs last disabled at (0): [<          (null)>]  )
 [ 1064.668121] CPU: 13 PID: 15296 Comm: qemu-kvm Not tainted 3.10.0-1
 [ 1064.668121] Hardware name: Dell Inc. PowerEdge R730/0H21J3, BIOS 5
 [ 1064.668123]  ffff8c1796b88000 00000000afe7344c ffff8c179abf3c68 f3
 [ 1064.668125]  ffff8c179abf3c90 ffffffff930ccb3d ffff8c1b992b3610 f0
 [ 1064.668126]  00007ffc1a26fbc0 ffff8c179abf3cb0 ffffffff9375f694 f0
 [ 1064.668126] Call Trace:
 [ 1064.668132]  [] dump_stack+0x19/0x1b
 [ 1064.668135]  [] __might_sleep+0x12d/0x1f0
 [ 1064.668138]  [] rt_spin_lock+0x24/0x60
 [ 1064.668155]  [] __get_kvmclock_ns+0x36/0x110 [k]
 [ 1064.668159]  [] ? futex_wait_queue_me+0x103/0x10
 [ 1064.668171]  [] kvm_arch_vm_ioctl+0xa2/0xd70 [k]
 [ 1064.668173]  [] ? futex_wait+0x1ac/0x2a0

v2: notice get_kvmclock_ns with the same problem (Pankaj).
v3: remove useless helper function (Pankaj).

arch/x86/kvm/x86.c +2 -16
index 1faf620..deacbe1 100644
@@ -1775,7 +1775,7 @@ static void kvm_gen_update_masterclock(struct kvm *kvm)
 #endif
 }
 
-static u64 __get_kvmclock_ns(struct kvm *kvm)
+u64 get_kvmclock_ns(struct kvm *kvm)
 {
 	struct kvm_arch *ka = &kvm->arch;
 	struct pvclock_vcpu_time_info hv_clock;
@@ -1796,18 +1796,6 @@ static u64 __get_kvmclock_ns(struct kvm *kvm)
 	return __pvclock_read_cycles(&hv_clock, rdtsc());
 }
 
-u64 get_kvmclock_ns(struct kvm *kvm)
-{
-	unsigned long flags;
-	s64 ns;
-
-	local_irq_save(flags);
-	ns = __get_kvmclock_ns(kvm);
-	local_irq_restore(flags);
-
-	return ns;
-}
-
 static void kvm_setup_pvclock_page(struct kvm_vcpu *v)
 {
 	struct kvm_vcpu_arch *vcpu = &v->arch;
@@ -4196,10 +4184,8 @@ long kvm_arch_vm_ioctl(struct file *filp,
 			goto out;
 
 		r = 0;
-		local_irq_disable();
-		now_ns = __get_kvmclock_ns(kvm);
+		now_ns = get_kvmclock_ns(kvm);
 		kvm->arch.kvmclock_offset += user_ns.clock - now_ns;
-		local_irq_enable();
 		kvm_gen_update_masterclock(kvm);
 		break;
 	}
@@ -4207,11 +4193,9 @@ long kvm_arch_vm_ioctl(struct file *filp,
 		struct kvm_clock_data user_ns;
 		u64 now_ns;

arch/x86/kvm/x86.c long kvm_arch_vm_ioctl(struct file *filp, · line 4201

Quoted 10 lines
arch/x86/kvm/x86.c +1 -3
 
-		local_irq_disable();
-		now_ns = __get_kvmclock_ns(kvm);
+		now_ns = get_kvmclock_ns(kvm);
 		user_ns.clock = now_ns;
 		user_ns.flags = kvm->arch.use_master_clock ? KVM_CLOCK_TSC_STABLE : 0;
-		local_irq_enable();
 		memset(&user_ns.pad, 0, sizeof(user_ns.pad));
 
 		r = -EFAULT;

Over all patch looks good to me. I think explanation provided by Marcelo is valid.

Quoted 1 line
maintainer Applied
#8 lore raw
In reply to Marcelo Tosatti
Quoted 110 lines · Marcelo Tosatti

The disablement of interrupts at KVM_SET_CLOCK/KVM_GET_CLOCK attempts to disable software suspend from causing "non atomic behaviour" of the operation:

Add a helper function to compute the kernel time and convert nanoseconds back to CPU specific cycles. Note that these must not be called in preemptible context, as that would mean the kernel could enter software suspend state, which would cause non-atomic operation.

However, assume the kernel can enter software suspend at the following 2 points:

        ktime_get_ts(&ts);
1.
						hypothetical_ktime_get_ts(&ts)
        monotonic_to_bootbased(&ts);

2.

monotonic_to_bootbased() should be correct relative to a ktime_get_ts(&ts) performed after point 1 (that is after resuming from software suspend), hypothetical_ktime_get_ts()

Therefore it is also correct for the ktime_get_ts(&ts) before point 1, which is

ktime_get_ts(&ts) = hypothetical_ktime_get_ts(&ts) + time-to-execute-suspend-code

Note CLOCK_MONOTONIC does not count during suspension.

So remove the irq disablement, which causes the following warning on -RT kernels:

With this reasoning, and the -RT bug that the irq disablement causes (because spin_lock is now a sleeping lock), remove the IRQ protection as it causes:

 [ 1064.668109] in_atomic(): 0, irqs_disabled(): 1, pid: 15296, name:m
 [ 1064.668110] INFO: lockdep is turned off.
 [ 1064.668110] irq event stamp: 0
 [ 1064.668112] hardirqs last  enabled at (0): [<          (null)>]  )
 [ 1064.668116] hardirqs last disabled at (0): [] c0
 [ 1064.668118] softirqs last  enabled at (0): [] c0
 [ 1064.668118] softirqs last disabled at (0): [<          (null)>]  )
 [ 1064.668121] CPU: 13 PID: 15296 Comm: qemu-kvm Not tainted 3.10.0-1
 [ 1064.668121] Hardware name: Dell Inc. PowerEdge R730/0H21J3, BIOS 5
 [ 1064.668123]  ffff8c1796b88000 00000000afe7344c ffff8c179abf3c68 f3
 [ 1064.668125]  ffff8c179abf3c90 ffffffff930ccb3d ffff8c1b992b3610 f0
 [ 1064.668126]  00007ffc1a26fbc0 ffff8c179abf3cb0 ffffffff9375f694 f0
 [ 1064.668126] Call Trace:
 [ 1064.668132]  [] dump_stack+0x19/0x1b
 [ 1064.668135]  [] __might_sleep+0x12d/0x1f0
 [ 1064.668138]  [] rt_spin_lock+0x24/0x60
 [ 1064.668155]  [] __get_kvmclock_ns+0x36/0x110 [k]
 [ 1064.668159]  [] ? futex_wait_queue_me+0x103/0x10
 [ 1064.668171]  [] kvm_arch_vm_ioctl+0xa2/0xd70 [k]
 [ 1064.668173]  [] ? futex_wait+0x1ac/0x2a0

v2: notice get_kvmclock_ns with the same problem (Pankaj).
v3: remove useless helper function (Pankaj).

arch/x86/kvm/x86.c +2 -16
index 1faf620..deacbe1 100644
@@ -1775,7 +1775,7 @@ static void kvm_gen_update_masterclock(struct kvm *kvm)
 #endif
 }
 
-static u64 __get_kvmclock_ns(struct kvm *kvm)
+u64 get_kvmclock_ns(struct kvm *kvm)
 {
 	struct kvm_arch *ka = &kvm->arch;
 	struct pvclock_vcpu_time_info hv_clock;
@@ -1796,18 +1796,6 @@ static u64 __get_kvmclock_ns(struct kvm *kvm)
 	return __pvclock_read_cycles(&hv_clock, rdtsc());
 }
 
-u64 get_kvmclock_ns(struct kvm *kvm)
-{
-	unsigned long flags;
-	s64 ns;
-
-	local_irq_save(flags);
-	ns = __get_kvmclock_ns(kvm);
-	local_irq_restore(flags);
-
-	return ns;
-}
-
 static void kvm_setup_pvclock_page(struct kvm_vcpu *v)
 {
 	struct kvm_vcpu_arch *vcpu = &v->arch;
@@ -4196,10 +4184,8 @@ long kvm_arch_vm_ioctl(struct file *filp,
 			goto out;
 
 		r = 0;
-		local_irq_disable();
-		now_ns = __get_kvmclock_ns(kvm);
+		now_ns = get_kvmclock_ns(kvm);
 		kvm->arch.kvmclock_offset += user_ns.clock - now_ns;
-		local_irq_enable();
 		kvm_gen_update_masterclock(kvm);
 		break;
 	}
@@ -4207,11 +4193,9 @@ long kvm_arch_vm_ioctl(struct file *filp,
 		struct kvm_clock_data user_ns;
 		u64 now_ns;

arch/x86/kvm/x86.c long kvm_arch_vm_ioctl(struct file *filp, · line 4201

Quoted 10 lines
arch/x86/kvm/x86.c +1 -3
 
-		local_irq_disable();
-		now_ns = __get_kvmclock_ns(kvm);
+		now_ns = get_kvmclock_ns(kvm);
 		user_ns.clock = now_ns;
 		user_ns.flags = kvm->arch.use_master_clock ? KVM_CLOCK_TSC_STABLE : 0;
-		local_irq_enable();
 		memset(&user_ns.pad, 0, sizeof(user_ns.pad));
 
 		r = -EFAULT;

Applied, thanks.

Paolo

#9 lore raw
In reply to Marcelo Tosatti
Quoted 44 lines · Marcelo Tosatti

The disablement of interrupts at KVM_SET_CLOCK/KVM_GET_CLOCK attempts to disable interrupts in that section to protect
the values that are calculated in that section from interrupt interference.

now_ns is calculated inside the irq protected region, user_ns.clock is passed from userspace (therefore not susceptible to interrupt variation).

About the line

                now_ns = __get_kvmclock_ns(kvm);  (1)

Interrupts can happen afterwards local_irq_enable(), rendering "now_ns" relative to its execution time PLUS interrupt time.

Therefore the local_irq_disable() / local_irq_enable() protection is not necessary (that is: interrupts triggering after local_irq_enable cause the same problem that the protection is trying to avoid).

With this reasoning, and the -RT bug that the irq disablement causes (because spin_lock is now a sleeping lock), remove the IRQ protection as it causes:

[ 1064.668109] in_atomic(): 0, irqs_disabled(): 1, pid: 15296, name:m
[ 1064.668110] INFO: lockdep is turned off.
[ 1064.668110] irq event stamp: 0
[ 1064.668112] hardirqs last  enabled at (0): [<          (null)>]  )
[ 1064.668116] hardirqs last disabled at (0): [<ffffffff9308184a>] c0
[ 1064.668118] softirqs last  enabled at (0): [<ffffffff9308184a>] c0
[ 1064.668118] softirqs last disabled at (0): [<          (null)>]  )
[ 1064.668121] CPU: 13 PID: 15296 Comm: qemu-kvm Not tainted 3.10.0-1
[ 1064.668121] Hardware name: Dell Inc. PowerEdge R730/0H21J3, BIOS 5
[ 1064.668123]  ffff8c1796b88000 00000000afe7344c ffff8c179abf3c68 f3
[ 1064.668125]  ffff8c179abf3c90 ffffffff930ccb3d ffff8c1b992b3610 f0
[ 1064.668126]  00007ffc1a26fbc0 ffff8c179abf3cb0 ffffffff9375f694 f0
[ 1064.668126] Call Trace:
[ 1064.668132]  [<ffffffff93757413>] dump_stack+0x19/0x1b
[ 1064.668135]  [<ffffffff930ccb3d>] __might_sleep+0x12d/0x1f0
[ 1064.668138]  [<ffffffff9375f694>] rt_spin_lock+0x24/0x60
[ 1064.668155]  [<ffffffffc06ab996>] __get_kvmclock_ns+0x36/0x110 [k]
[ 1064.668159]  [<ffffffff93112993>] ? futex_wait_queue_me+0x103/0x10
[ 1064.668171]  [<ffffffffc06b8782>] kvm_arch_vm_ioctl+0xa2/0xd70 [k]
[ 1064.668173]  [<ffffffff9311333c>] ? futex_wait+0x1ac/0x2a0
Quoted 1 line

On -RT kernels.

Hmm, __get_kvmclock_ns used not to need a spinlock back when it was added... Why does it now?

Looking at its current state, I'm not sure I understand what it's supposed to do: it uses the host tsc rate rather than the guest one, which seems to just defeat the purpose of originally introducing it: to have a way to obtain the clock value exactly the same as the guest would see...

Am I missing anything obvious?

Thanks,
Roman.

#10 lore raw

2017-04-13 10:32+0300, Roman Kagan:

Quoted 2 lines

Hmm, __get_kvmclock_ns used not to need a spinlock back when it was added... Why does it now?

It accessed vcpu to get the data it now takes from the master clock. The lock is there to protect master clock's consistency.

There were problems when accessing VCPUs from VM ioctls -- you first had to conjure a VCPU out of thin air and then be extremely careful when accessing it, which we didn't manage to and it resulted in an invalid VMCS access bug, IIRC. (I would bet there we other bugs.)

Quoted 7 lines

Looking at its current state, I'm not sure I understand what it's supposed to do: it uses the host tsc rate rather than the guest one, which seems to just defeat the purpose of originally introducing it: to have a way to obtain the clock value exactly the same as the guest would see...

Am I missing anything obvious?

It is a bit obscure ... the current state is simulating what the vcpu does when setting up its kvmclock, and should therefore yield the same result as the guest kvmclock would see.

Without master clock, the guest kvmclock follows ktime_get_boot_ns().

With master clock, the guest follows host TSC (different frequency than ktime_get_boot_ns()). The guest might have extra scaling and offset on that TSC, but both are known to KVM when computing the kvmclock variables to get the 1GHz result.

We could first convert the host TSC to guest TSC and then compute kvmclock from it, but we should obtain the same result when converting to 1GHz directly from host TSC (within rouding errors, ofc).

This is possible as guest and host use the same physical TSC.