[PATCH] x86/xen: avoid updating TLS descriptors if they haven't changed

[PATCH] x86/xen: avoid updating TLS descriptors if they haven't changed

[David Vrabel]

From: David Vrabel <david.vrabel@citrix.com>

When switching tasks in a Xen PV guest, avoid updating the TLS
descriptors if they haven't changed.  This improves the speed of
context switches by almost 10% as much of the time the descriptors are
the same or only one is different.

The descriptors written into the GDT by Xen are modified from the
values passed in the update_descriptor hypercall so we keep shadow
copies of the three TLS descriptors to compare against.

lmbench3 test     Before  After  Improvement
--------------------------------------------
lat_ctx -s 32 24   7.19    6.52  9%
lat_pipe          12.56   11.66  7%

Signed-off-by: David Vrabel <david.vrabel@citrix.com>
---
I note that the comment in asm/desc_defs.h says the 'a' and 'b' fields
in desc_struct as deprecated but there seems to be no suitable
alternatives.
---
 arch/x86/xen/enlighten.c |   30 +++++++++++++++++++++++++++---
 1 files changed, 27 insertions(+), 3 deletions(-)

diff --git a/arch/x86/xen/enlighten.c b/arch/x86/xen/enlighten.c
index e74df95..18e14af 100644
--- a/arch/x86/xen/enlighten.c
+++ b/arch/x86/xen/enlighten.c
@@ -124,6 +124,19 @@ struct shared_info *HYPERVISOR_shared_info = (void *)&xen_dummy_shared_info;
  */
 static int have_vcpu_info_placement = 1;
 
+struct tls_descs {
+	struct desc_struct desc[3];
+};
+
+/*
+ * Updating the 3 TLS descriptors in the GDT on every task switch is
+ * surprisingly expensive so we avoid updating them if they haven't
+ * changed.  Since Xen writes different descriptors than the one
+ * passed in the update_descriptor hypercall we keep shadow copies to
+ * compare against.
+ */
+static DEFINE_PER_CPU(struct tls_descs, shadow_tls_desc);
+
 static void clamp_max_cpus(void)
 {
 #ifdef CONFIG_SMP
@@ -535,9 +548,20 @@ static void __init xen_load_gdt_boot(const struct desc_ptr *dtr)
 static void load_TLS_descriptor(struct thread_struct *t,
 				unsigned int cpu, unsigned int i)
 {
-	struct desc_struct *gdt = get_cpu_gdt_table(cpu);
-	xmaddr_t maddr = arbitrary_virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
-	struct multicall_space mc = __xen_mc_entry(0);
+	struct desc_struct *shadow = &per_cpu(shadow_tls_desc, cpu).desc[i];
+	struct desc_struct *gdt;
+	xmaddr_t maddr;
+	struct multicall_space mc;
+
+	if (shadow->a == t->tls_array[i].a && shadow->b == t->tls_array[i].b)
+		return;
+
+	shadow->a = t->tls_array[i].a;
+	shadow->b = t->tls_array[i].b;
+
+	gdt = get_cpu_gdt_table(cpu);
+	maddr = arbitrary_virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
+	mc = __xen_mc_entry(0);
 
 	MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
 }
-- 
1.7.2.5

Re: [PATCH] x86/xen: avoid updating TLS descriptors if they haven't changed

[David Vrabel]

On 07/06/12 18:01, David Vrabel wrote:
> From: David Vrabel <david.vrabel@citrix.com>
> 
> When switching tasks in a Xen PV guest, avoid updating the TLS
> descriptors if they haven't changed.  This improves the speed of
> context switches by almost 10% as much of the time the descriptors are
> the same or only one is different.
> 
> The descriptors written into the GDT by Xen are modified from the
> values passed in the update_descriptor hypercall so we keep shadow
> copies of the three TLS descriptors to compare against.
> 
> lmbench3 test     Before  After  Improvement
> --------------------------------------------
> lat_ctx -s 32 24   7.19    6.52  9%
> lat_pipe          12.56   11.66  7%
> 
> Signed-off-by: David Vrabel <david.vrabel@citrix.com>
> ---
> I note that the comment in asm/desc_defs.h says the 'a' and 'b' fields
> in desc_struct as deprecated but there seems to be no suitable
> alternatives.

ping?  Any opinion on this patch from the x86 side?  If it's okay can we
get an ack so Konrad can take the patch via his tree.

Thanks.

David

> ---
>  arch/x86/xen/enlighten.c |   30 +++++++++++++++++++++++++++---
>  1 files changed, 27 insertions(+), 3 deletions(-)
> 
> diff --git a/arch/x86/xen/enlighten.c b/arch/x86/xen/enlighten.c
> index e74df95..18e14af 100644
> --- a/arch/x86/xen/enlighten.c
> +++ b/arch/x86/xen/enlighten.c
> @@ -124,6 +124,19 @@ struct shared_info *HYPERVISOR_shared_info = (void *)&xen_dummy_shared_info;
>   */
>  static int have_vcpu_info_placement = 1;
>  
> +struct tls_descs {
> +	struct desc_struct desc[3];
> +};
> +
> +/*
> + * Updating the 3 TLS descriptors in the GDT on every task switch is
> + * surprisingly expensive so we avoid updating them if they haven't
> + * changed.  Since Xen writes different descriptors than the one
> + * passed in the update_descriptor hypercall we keep shadow copies to
> + * compare against.
> + */
> +static DEFINE_PER_CPU(struct tls_descs, shadow_tls_desc);
> +
>  static void clamp_max_cpus(void)
>  {
>  #ifdef CONFIG_SMP
> @@ -535,9 +548,20 @@ static void __init xen_load_gdt_boot(const struct desc_ptr *dtr)
>  static void load_TLS_descriptor(struct thread_struct *t,
>  				unsigned int cpu, unsigned int i)
>  {
> -	struct desc_struct *gdt = get_cpu_gdt_table(cpu);
> -	xmaddr_t maddr = arbitrary_virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
> -	struct multicall_space mc = __xen_mc_entry(0);
> +	struct desc_struct *shadow = &per_cpu(shadow_tls_desc, cpu).desc[i];
> +	struct desc_struct *gdt;
> +	xmaddr_t maddr;
> +	struct multicall_space mc;
> +
> +	if (shadow->a == t->tls_array[i].a && shadow->b == t->tls_array[i].b)
> +		return;
> +
> +	shadow->a = t->tls_array[i].a;
> +	shadow->b = t->tls_array[i].b;
> +
> +	gdt = get_cpu_gdt_table(cpu);
> +	maddr = arbitrary_virt_to_machine(&gdt[GDT_ENTRY_TLS_MIN+i]);
> +	mc = __xen_mc_entry(0);
>  
>  	MULTI_update_descriptor(mc.mc, maddr.maddr, t->tls_array[i]);
>  }

Re: [PATCH] x86/xen: avoid updating TLS descriptors if they haven't changed

[Konrad Rzeszutek Wilk]

On Thu, Jun 14, 2012 at 03:52:19PM +0100, David Vrabel wrote:
> On 07/06/12 18:01, David Vrabel wrote:
> > From: David Vrabel <david.vrabel@citrix.com>
> > 
> > When switching tasks in a Xen PV guest, avoid updating the TLS
> > descriptors if they haven't changed.  This improves the speed of
> > context switches by almost 10% as much of the time the descriptors are
> > the same or only one is different.
> > 
> > The descriptors written into the GDT by Xen are modified from the
> > values passed in the update_descriptor hypercall so we keep shadow
> > copies of the three TLS descriptors to compare against.
> > 
> > lmbench3 test     Before  After  Improvement
> > --------------------------------------------
> > lat_ctx -s 32 24   7.19    6.52  9%
> > lat_pipe          12.56   11.66  7%
> > 
> > Signed-off-by: David Vrabel <david.vrabel@citrix.com>
> > ---
> > I note that the comment in asm/desc_defs.h says the 'a' and 'b' fields
> > in desc_struct as deprecated but there seems to be no suitable
> > alternatives.
> 
> ping?  Any opinion on this patch from the x86 side?  If it's okay can we
> get an ack so Konrad can take the patch via his tree.

It breaks my all my bootup tests - so NACK until at least that is fixed.
I think I sent you the whole serial log - is there something else that would help
narrow it down?

Re: [PATCH] x86/xen: avoid updating TLS descriptors if they haven't changed

[David Vrabel]

On 14/06/12 17:49, Konrad Rzeszutek Wilk wrote:
> On Thu, Jun 14, 2012 at 03:52:19PM +0100, David Vrabel wrote:
>> On 07/06/12 18:01, David Vrabel wrote:
>>> From: David Vrabel <david.vrabel@citrix.com>
>>>
>>> When switching tasks in a Xen PV guest, avoid updating the TLS
>>> descriptors if they haven't changed.  This improves the speed of
>>> context switches by almost 10% as much of the time the descriptors are
>>> the same or only one is different.
>>>
>>> The descriptors written into the GDT by Xen are modified from the
>>> values passed in the update_descriptor hypercall so we keep shadow
>>> copies of the three TLS descriptors to compare against.
>>>
>>> lmbench3 test     Before  After  Improvement
>>> --------------------------------------------
>>> lat_ctx -s 32 24   7.19    6.52  9%
>>> lat_pipe          12.56   11.66  7%
>>>
>>> Signed-off-by: David Vrabel <david.vrabel@citrix.com>
>>> ---
>>> I note that the comment in asm/desc_defs.h says the 'a' and 'b' fields
>>> in desc_struct as deprecated but there seems to be no suitable
>>> alternatives.
>>
>> ping?  Any opinion on this patch from the x86 side?  If it's okay can we
>> get an ack so Konrad can take the patch via his tree.
> 
> It breaks my all my bootup tests - so NACK until at least that is fixed.
> I think I sent you the whole serial log - is there something else that would help
> narrow it down?

You're mixing up this patch with "xen/mm: do direct hypercall in
xen_set_pte() if batching is unavailable" which is broken on 64-bit and
I am looking into it.

David