From: Bhavana Nagendra <bnagendr@redhat.com>
Date: Mon, 17 Dec 2007 16:41:53 -0500
Subject: [xen] enable nested paging by default on amd-v
Message-id: 20071217214156.1638.1797.sendpatchset@localhost.localdomain
O-Subject: Re: [RHEL5.2 PATCH] Enable Nested Paging by default on AMD-V
Bugzilla: 247190

Resolves BZ 247190

This patch enables AMD's Nested Paging that allows the guest virtual pages
to map to host physical pages through hardware, to be the default.
Benefits of Nested Paging
- Reduces the complexity of memory management in virtualized environment
(less s/w bugs)
- Improves performance by avoiding a vast amount of #VMEXIT for memory
intensive guests (less CPU utilized)
Nested Paging can only be enabled on AMD-V platforms and does not affect the
VT code paths.Testing:
There has been a lot of performance experiments done with Nested Paging
enabled, and with Oracle workloads.  Shak and I can provide data offline.

This patch has been applied to Bill Burns' 3.1.1 HV upgrade and tested at
AMD (Austin team, and BDC (me)).

Chris Lalancette wrote:
>
> OK, great.  Yeah, I'm much more comfortable with this new patch.  I also did
> some basic smoke-testing on my Barcelona and an Intel SDV, and it seemed to do
> the right thing.
>
Great, thanks.   Chris, I looked into the boolean_param patch and found a changeset that's
not in R5.2 code base.   This allows for users to not only use hap=off and hap=no, which are
both functional, but also to use hap=false and hap=0.   These are supported upstream, and
something that should really work in R5.2 as well.

Upstream c/s 15885.

Please review and ACK this change.

Acked-by: Chris Lalancette <clalance@redhat.com>
Acked-by: "Stephen C. Tweedie" <sct@redhat.com>
Acked-by: Bill Burns <bburns@redhat.com>

diff --git a/arch/x86/hvm/hvm.c b/arch/x86/hvm/hvm.c
index a15e74a..66fc9a5 100644
--- a/arch/x86/hvm/hvm.c
+++ b/arch/x86/hvm/hvm.c
@@ -49,6 +49,15 @@
 #include <public/version.h>
 #include <public/memory.h>
 
+/*
+ * Xen command-line option to allow/disallow hardware-assisted paging.
+ * Since the phys-to-machine table of AMD NPT is in host format, 32-bit Xen
+ * could only support guests using NPT with up to a 4GB memory map. Therefore
+ * we only allow HAP by default on 64-bit Xen.
+ */
+static int opt_hap_permitted = (CONFIG_PAGING_LEVELS != 3);
+boolean_param("hap", opt_hap_permitted);
+
 int hvm_enabled __read_mostly;
 
 unsigned int opt_hvm_debug_level __read_mostly;
@@ -74,6 +83,14 @@ void hvm_enable(struct hvm_function_table *fns)
 
     hvm_funcs   = *fns;
     hvm_enabled = 1;
+
+    if ( hvm_funcs.hap_supported )
+    {
+        if ( !opt_hap_permitted )
+            hvm_funcs.hap_supported = 0;
+        printk("HVM: Hardware Assisted Paging detected %s.\n",
+               hvm_funcs.hap_supported ? "and enabled" : "but disabled");
+    }
 }
 
 void hvm_disable(void)
diff --git a/arch/x86/hvm/svm/svm.c b/arch/x86/hvm/svm/svm.c
index 4bf131f..23a44cd 100644
--- a/arch/x86/hvm/svm/svm.c
+++ b/arch/x86/hvm/svm/svm.c
@@ -66,9 +66,6 @@ static void *hsa[NR_CPUS] __read_mostly;
 /* vmcb used for extended host state */
 static void *root_vmcb[NR_CPUS] __read_mostly;
 
-/* hardware assisted paging bits */
-extern int opt_hap_enabled;
-
 static void svm_inject_exception(struct vcpu *v, int trap, 
                                         int ev, int error_code)
 {
@@ -1008,20 +1005,6 @@ static struct hvm_function_table svm_function_table = {
     .event_injection_faulted = svm_event_injection_faulted
 };
 
-static void svm_npt_detect(void)
-{
-    u32 eax, ebx, ecx, edx;
-
-    /* Check CPUID for nested paging support. */
-    cpuid(0x8000000A, &eax, &ebx, &ecx, &edx);
-
-    if ( !(edx & 1) && opt_hap_enabled )
-    {
-        printk("SVM: Nested paging is not supported by this CPU.\n");
-        opt_hap_enabled = 0;
-    }
-}
-
 int start_svm(struct cpuinfo_x86 *c)
 {
     u32 eax, ecx, edx;
@@ -1050,8 +1033,6 @@ int start_svm(struct cpuinfo_x86 *c)
 
     write_efer(read_efer() | EFER_SVME);
 
-    svm_npt_detect();
-
     /* Initialize the HSA for this core. */
     phys_hsa = (u64) virt_to_maddr(hsa[cpu]);
     phys_hsa_lo = (u32) phys_hsa;
@@ -1066,11 +1047,10 @@ int start_svm(struct cpuinfo_x86 *c)
 
     setup_vmcb_dump();
 
+    svm_function_table.hap_supported = (cpuid_edx(0x8000000A) & 1);
+
     hvm_enable(&svm_function_table);
 
-    if ( opt_hap_enabled )
-        printk("SVM: Nested paging enabled.\n");
-        
     return 1;
 }
 
diff --git a/arch/x86/mm/paging.c b/arch/x86/mm/paging.c
index 4cf1c06..0be022f 100644
--- a/arch/x86/mm/paging.c
+++ b/arch/x86/mm/paging.c
@@ -27,9 +27,7 @@
 #include <asm/hap.h>
 #include <asm/guest_access.h>
 
-/* Xen command-line option to enable hardware-assisted paging */
-int opt_hap_enabled;
-boolean_param("hap", opt_hap_enabled);
+#define hap_enabled(d) (hvm_funcs.hap_supported && is_hvm_domain(d))
 
 /* Printouts */
 #define PAGING_PRINTK(_f, _a...)                                     \
@@ -362,14 +360,14 @@ void paging_domain_init(struct domain *d)
     shadow_domain_init(d);
 
     /* ... but we will use hardware assistance if it's available. */
-    if ( opt_hap_enabled && is_hvm_domain(d) )
+    if ( hap_enabled(d) )
         hap_domain_init(d);
 }
 
 /* vcpu paging struct initialization goes here */
 void paging_vcpu_init(struct vcpu *v)
 {
-    if ( opt_hap_enabled && is_hvm_vcpu(v) )
+    if ( hap_enabled(v->domain) )
         hap_vcpu_init(v);
     else
         shadow_vcpu_init(v);
@@ -429,7 +427,7 @@ int paging_domctl(struct domain *d, xen_domctl_shadow_op_t *sc,
     }
 	
     /* Here, dispatch domctl to the appropriate paging code */
-    if ( opt_hap_enabled && is_hvm_domain(d) )
+    if ( hap_enabled(d) )
 	return hap_domctl(d, sc, u_domctl);
     else
 	return shadow_domctl(d, sc, u_domctl);
@@ -438,7 +436,7 @@ int paging_domctl(struct domain *d, xen_domctl_shadow_op_t *sc,
 /* Call when destroying a domain */
 void paging_teardown(struct domain *d)
 {
-    if ( opt_hap_enabled && is_hvm_domain(d) )
+    if ( hap_enabled(d) )
         hap_teardown(d);
     else
         shadow_teardown(d);
@@ -450,7 +448,7 @@ void paging_teardown(struct domain *d)
 /* Call once all of the references to the domain have gone away */
 void paging_final_teardown(struct domain *d)
 {
-    if ( opt_hap_enabled && is_hvm_domain(d) )
+    if ( hap_enabled(d) )
         hap_final_teardown(d);
     else
         shadow_final_teardown(d);
@@ -460,7 +458,7 @@ void paging_final_teardown(struct domain *d)
  * creation. */
 int paging_enable(struct domain *d, u32 mode)
 {
-    if ( opt_hap_enabled && is_hvm_domain(d) )
+    if ( hap_enabled(d) )
         return hap_enable(d, mode | PG_HAP_enable);
     else
         return shadow_enable(d, mode | PG_SH_enable);
diff --git a/common/kernel.c b/common/kernel.c
index 9bfba6b..034ff21 100644
--- a/common/kernel.c
+++ b/common/kernel.c
@@ -80,7 +80,10 @@ void cmdline_parse(char *cmdline)
                 break;
             case OPT_BOOL:
             case OPT_INVBOOL:
-                if ( !strcmp("no", optval) || !strcmp("off", optval) )
+                if ( !strcmp("no", optval) ||
+                     !strcmp("off", optval) ||
+                     !strcmp("false", optval) ||
+                     !strcmp("0", optval) )
                     bool_assert = !bool_assert;
                 if ( param->type == OPT_INVBOOL )
                     bool_assert = !bool_assert;
diff --git a/include/asm-x86/hvm/hvm.h b/include/asm-x86/hvm/hvm.h
index 4562e8c..117f167 100644
--- a/include/asm-x86/hvm/hvm.h
+++ b/include/asm-x86/hvm/hvm.h
@@ -63,6 +63,9 @@ typedef struct segment_register {
 struct hvm_function_table {
     char *name;
 
+    /* Support Hardware-Assisted Paging? */
+    int hap_supported;
+
     /*
      *  Disable HVM functionality
      */