There are many functions invoke apic_debug(), which is defined
a null function by default, and that's incovenient for debuging
lapic.
This patch allows setting apic debug according to add a apic_dbg
parameter of kvm.
Signed-off-by: Yi Wang <[email protected]>
---
v2: change apic_dbg to bool and tag __read_mostly. Thanks to Sean.
arch/x86/kvm/lapic.c | 7 ++++++-
1 file changed, 6 insertions(+), 1 deletion(-)
diff --git a/arch/x86/kvm/lapic.c b/arch/x86/kvm/lapic.c
index 9bf70cf..0827e7c 100644
--- a/arch/x86/kvm/lapic.c
+++ b/arch/x86/kvm/lapic.c
@@ -54,8 +54,13 @@
#define PRIu64 "u"
#define PRIo64 "o"
+static bool apic_dbg __read_mostly;
+module_param(apic_dbg, bool, 0644);
+
/* #define apic_debug(fmt,arg...) printk(KERN_WARNING fmt,##arg) */
-#define apic_debug(fmt, arg...) do {} while (0)
+#define apic_debug(fmt, arg...) do { if (apic_dbg) \
+ printk(KERN_DEBUG fmt, ##arg); \
+} while (0)
/* 14 is the version for Xeon and Pentium 8.4.8*/
#define APIC_VERSION (0x14UL | ((KVM_APIC_LVT_NUM - 1) << 16))
--
1.8.3.1
On Thu, May 09, 2019 at 06:47:57PM +0800, Yi Wang wrote:
> There are many functions invoke apic_debug(), which is defined
> a null function by default, and that's incovenient for debuging
> lapic.
>
> This patch allows setting apic debug according to add a apic_dbg
> parameter of kvm.
>
> Signed-off-by: Yi Wang <[email protected]>
> ---
> v2: change apic_dbg to bool and tag __read_mostly. Thanks to Sean.
>
> arch/x86/kvm/lapic.c | 7 ++++++-
> 1 file changed, 6 insertions(+), 1 deletion(-)
>
> diff --git a/arch/x86/kvm/lapic.c b/arch/x86/kvm/lapic.c
> index 9bf70cf..0827e7c 100644
> --- a/arch/x86/kvm/lapic.c
> +++ b/arch/x86/kvm/lapic.c
> @@ -54,8 +54,13 @@
> #define PRIu64 "u"
> #define PRIo64 "o"
>
> +static bool apic_dbg __read_mostly;
> +module_param(apic_dbg, bool, 0644);
Probably don't need to shorten "debug".
> +
> /* #define apic_debug(fmt,arg...) printk(KERN_WARNING fmt,##arg) */
> -#define apic_debug(fmt, arg...) do {} while (0)
> +#define apic_debug(fmt, arg...) do { if (apic_dbg) \
> + printk(KERN_DEBUG fmt, ##arg); \
> +} while (0)
Pulling in your comment regarding sched_debug and noirqdebug...
On Thu, May 09, 2019 at 08:29:38AM +0800, [email protected] wrote:
> Also, we have some similar parameters already, such like sched_debug,
> noirqdebug :)
The IRQ debug hook is a completely different beast than the APIC debug
messages.
sched_debug is a much better comparison. The param only exists if
CONFIG_SCHED_DEBUG=y, which is "default y" but "depends on DEBUG_KERNEL".
That seems like the route to go if we want the ability to toggle APIC
debugging at runtime. And if we go with an all encompassing config,
e.g. CONFIG_KVM_DEBUG, we can use it to wrap x86/mmu.c's debug param as
well (and rename it to mmu_debug).
>
> /* 14 is the version for Xeon and Pentium 8.4.8*/
> #define APIC_VERSION (0x14UL | ((KVM_APIC_LVT_NUM - 1) << 16))
> --
> 1.8.3.1
>
On Fri, May 10, 2019 at 12:54:21PM +0800, [email protected] wrote:
> I grep "debug" in arch/x86/kvm, and find these *_debug:
> ioapic_debug
> apic_debug
>
> and dbg in mmu.c, which is better to be renamed to mmu_debug as you said.
>
> and vcpu_debug, which uses kvm_debug macro.
>
> kvm_debug macro uses pr_debug which can be dynamically set during running
> time, so, how about change all *_debug in kvm to pr_debug like vcpu_debug?
It's still the same end result, we're bloating and slowing KVM with code
and conditionals that aren't useful in normal operation. grep vcpu_debug
a bit further and you'll see that the only uses in x86 are when the guest
has crashed, is being reset, or is accessing an unhandled MSR and KVM is
injecting a #GP in response. In other words, the existing uses are only
in code that isn't remotely performance critical.
hyperv.c: vcpu_debug(vcpu, "hv crash (0x%llx 0x%llx 0x%llx 0x%llx 0x%llx)\n",
hyperv.c: vcpu_debug(vcpu, "hyper-v reset requested\n");
x86.c: vcpu_debug_ratelimited(vcpu, "unhandled wrmsr: 0x%x data 0x%llx\n",
x86.c: vcpu_debug_ratelimited(vcpu, "unhandled rdmsr: 0x%x\n",
pr_debug does have more direct uses, notably in nested VMX and KVM TSC
handling. Similar to the above vcpu_debug case, the nVMX uses are all
failing paths and not performance critical. The TSC code does have one
path that may affect performance (get_kvmclock_ns()->kvm_get_time_scale()),
but I don't think that should be considered as setting the precedent. In
fact, it may make sense to convert the TSC pr_debugs to be gated by
CONFIG_DEBUG_KVM as well.
Paolo, do you have any thoughts?