RE: [Xen-ia64-devel] [PATCH] bug fix new_tlbflush_clock_period() - Xu, Anthony - com.xensource.lists.xen-ia64-devel

13 messages in com.xensource.lists.xen-ia64-develRE: [Xen-ia64-devel] [PATCH] bug fix ...

From	Sent On	Attachments
Kouya SHIMURA	01 Feb 2007 22:57	.patch
Alex Williamson	04 Feb 2007 12:13
Isaku Yamahata	04 Feb 2007 17:45
Xu, Anthony	04 Feb 2007 18:02
Isaku Yamahata	04 Feb 2007 18:55	.patch
Xu, Anthony	04 Feb 2007 19:34
Isaku Yamahata	04 Feb 2007 19:42
Xu, Anthony	04 Feb 2007 19:47
Isaku Yamahata	04 Feb 2007 20:58
Xu, Anthony	04 Feb 2007 21:11
Alex Williamson	05 Feb 2007 15:34
Isaku Yamahata	05 Feb 2007 20:54
Xu, Anthony	05 Feb 2007 21:14

Subject:	RE: [Xen-ia64-devel] [PATCH] bug fix new_tlbflush_clock_period()
From:	Xu, Anthony (anth...@intel.com)
Date:	02/04/2007 07:47:05 PM
List:	com.xensource.lists.xen-ia64-devel

Isaku Yamahata write on 2007年2月5日 11:43:

On Mon, Feb 05, 2007 at 11:34:59AM +0800, Xu, Anthony wrote:

Before calling local_vhpt_flush, We need to make sure it is not VTI-domain and it is not per-VCPU VHPT,

local_vhpt_flush() always flushes vhpt associated to pcpu. not vcpu.

If HAS_PERVCPU_VHPT(current->domain) is true, or VMX_DOMAIN (current) is true.

Vhpt_paddr may be 0, So It erases 64KB memory content that starts from machine address 0.

- Anthony

DEFINE_PER_CPU (unsigned long, vhpt_paddr); DEFINE_PER_CPU (unsigned long, vhpt_pend); local_vhpt_flush(void) __vhpt_flush(__ia64_per_cpu_var(vhpt_paddr));

+static void +tlbflush_clock_local_flush(void *unused) +{

if(!VMX_DOMAIN(current)&&!HAS_PERVCPU_VHPT(current->domain){ + local_vhpt_flush(); }

+ local_flush_tlb_all(); +} +

- Anthony

Isaku Yamahata write on 2007年2月5日 10:56:

On Mon, Feb 05, 2007 at 10:02:53AM +0800, Xu, Anthony wrote:

Isaku Yamahata write on 2007年2月5日 9:45:

Hi Kouya. Good catch! Although this patch already commited and I made the bug, vti domain also relies on tlb flush lock. (See flush_vtlb_for_context_switch())

So we should do if (!test_bit(_VCPUF_initialize)) continue if (VMX_DOMAIN(v)) <flush all hash and collision chain of v> else vcpu_vhpt_flush()

disable the tlb flush clock usage in flush_vtlb_for_context_switch().

Hi Isaku,

Why do we need to call vcpu_vhpt_flush? IMO we only need to call __local_flush_tlb_all, if we use per-vcpu VHPT. Can you elaborate it?

That's right. when I wrote that, I tried to apply tlb flush clock to not only mTLB and vhpt but also per-vcpu vhpt. but it isn't used for per-vcpu vhpt. So I removed the related bogus code.