[patch] allow dynamic sizing the kvm vmm area on ia64

[patch] allow dynamic sizing the kvm vmm area on ia64

[Jes Sorensen]

[-- Attachment #1: Type: text/plain, Size: 326 bytes --]

Hi,

This patch allows dynamically sizing the vmm area on ia64. I have left
it with the same size as we used to have with the constants, but it
seems to work going higher. I put in an artificial limit of 64 vcpus for
now anyway as scaling wise it doesn't work well going higher anyway.....
we'll get there :-)

Cheers,
Jes




[-- Attachment #2: 5001-kvm-ia64-scale-vcpu.patch --]
[-- Type: text/plain, Size: 4638 bytes --]

Make KVM_VM_DATA_SHIFT a variable, making it possible allocate a
larger vm data area, making more space for the struct vcpu_data array.
This allows for increasing the number of vcpus a guest can support.

Signed-off-by: Jes Sorensen <jes@sgi.com>

---
 arch/ia64/include/asm/kvm_host.h |   14 +++++++++-----
 arch/ia64/kvm/kvm-ia64.c         |   32 ++++++++++++++++----------------
 2 files changed, 25 insertions(+), 21 deletions(-)

Index: linux-2.6.git/arch/ia64/include/asm/kvm_host.h
===================================================================
--- linux-2.6.git.orig/arch/ia64/include/asm/kvm_host.h
+++ linux-2.6.git/arch/ia64/include/asm/kvm_host.h
@@ -101,8 +101,7 @@
 #define KVM_VM_DATA_OFFSET_SHIFT	30
 #define KVM_VM_DATA_OFFSET	(__IA64_UL_CONST(1) << KVM_VM_DATA_OFFSET_SHIFT)
 
-#define KVM_VM_DATA_SHIFT	26
-#define KVM_VM_DATA_SIZE	(__IA64_UL_CONST(1) << KVM_VM_DATA_SHIFT)
+#define KVM_VM_DATA_SIZE	(__IA64_UL_CONST(1) << kvm_vm_data_shift)
 #define KVM_VM_DATA_BASE	(KVM_VMM_BASE + KVM_VM_DATA_OFFSET)
 
 #define KVM_P2M_BASE		KVM_VM_DATA_BASE
@@ -132,9 +131,14 @@
 
 #ifndef __ASSEMBLY__
 
+extern int kvm_vm_data_shift;
+
 /*Define the max vcpus and memory for Guests.*/
-#define KVM_MAX_VCPUS	(KVM_VM_DATA_SIZE - KVM_P2M_SIZE - KVM_VM_STRUCT_SIZE -\
-			KVM_MEM_DIRTY_LOG_SIZE) / sizeof(struct kvm_vcpu_data)
+#define KVM_MAX_POSSIBLE_VCPUS (KVM_VM_DATA_SIZE - KVM_P2M_SIZE - \
+				KVM_VM_STRUCT_SIZE - KVM_MEM_DIRTY_LOG_SIZE)/ \
+				sizeof(struct kvm_vcpu_data)
+#define KVM_MAX_VCPUS		64
+
 #define KVM_MAX_MEM_SIZE (KVM_P2M_SIZE >> 3 << PAGE_SHIFT)
 
 #define VMM_LOG_LEN 256
@@ -160,7 +164,7 @@
 	char kvm_p2m[KVM_P2M_SIZE];
 	char kvm_vm_struct[KVM_VM_STRUCT_SIZE];
 	char kvm_mem_dirty_log[KVM_MEM_DIRTY_LOG_SIZE];
-	struct kvm_vcpu_data vcpu_data[KVM_MAX_VCPUS];
+	struct kvm_vcpu_data vcpu_data[0];
 };
 
 #define VCPU_BASE(n)	(KVM_VM_DATA_BASE + \
Index: linux-2.6.git/arch/ia64/kvm/kvm-ia64.c
===================================================================
--- linux-2.6.git.orig/arch/ia64/kvm/kvm-ia64.c
+++ linux-2.6.git/arch/ia64/kvm/kvm-ia64.c
@@ -54,6 +54,7 @@
 static unsigned long kvm_vm_buffer;
 static unsigned long kvm_vm_buffer_size;
 unsigned long kvm_vmm_gp;
+int kvm_vm_data_shift = 26;	/* currently 26 limits us to 156 vcpus */
 
 static long vp_env_info;
 
@@ -583,8 +584,7 @@
 	vcpu->arch.vmm_tr_slot = r;
 	/*Insert a pairt of tr to map data of vm*/
 	pte = pte_val(mk_pte_phys(__pa(kvm->arch.vm_base), PAGE_KERNEL));
-	r = ia64_itr_entry(0x3, KVM_VM_DATA_BASE,
-					pte, KVM_VM_DATA_SHIFT);
+	r = ia64_itr_entry(0x3, KVM_VM_DATA_BASE, pte, kvm_vm_data_shift);
 	if (r < 0)
 		goto out;
 	vcpu->arch.vm_tr_slot = r;
@@ -749,14 +749,14 @@
 
 	BUG_ON(sizeof(struct kvm) > KVM_VM_STRUCT_SIZE);
 
-	vm_base = __get_free_pages(GFP_KERNEL, get_order(KVM_VM_DATA_SIZE));
+	vm_base = __get_free_pages(GFP_KERNEL, kvm_vm_data_shift - PAGE_SHIFT);
 
 	if (!vm_base)
 		return ERR_PTR(-ENOMEM);
 
-	memset((void *)vm_base, 0, KVM_VM_DATA_SIZE);
+	memset((void *)vm_base, 0, (1 << kvm_vm_data_shift));
 	kvm = (struct kvm *)(vm_base +
-			offsetof(struct kvm_vm_data, kvm_vm_struct));
+			     offsetof(struct kvm_vm_data, kvm_vm_struct));
 	kvm->arch.vm_base = vm_base;
 	printk(KERN_DEBUG"kvm: vm's data area:0x%lx\n", vm_base);
 
@@ -1039,7 +1039,8 @@
 {
 	if (!kvm_vmm_base && (kvm_vm_buffer_size < KVM_VM_BUFFER_SIZE)) {
 		kvm_vmm_base = __get_free_pages(GFP_KERNEL,
-				get_order(KVM_VMM_SIZE));
+						KVM_VMM_SHIFT - PAGE_SHIFT);
+
 		if (!kvm_vmm_base)
 			return -ENOMEM;
 
@@ -1286,20 +1287,19 @@
 	return r;
 }
 
-struct kvm_vcpu *kvm_arch_vcpu_create(struct kvm *kvm,
-		unsigned int id)
+struct kvm_vcpu *kvm_arch_vcpu_create(struct kvm *kvm, unsigned int id)
 {
 	struct kvm_vcpu *vcpu;
 	unsigned long vm_base = kvm->arch.vm_base;
-	int r;
-	int cpu;
+	int r, cpu, max_cpu;
 
 	BUG_ON(sizeof(struct kvm_vcpu) > VCPU_STRUCT_SIZE/2);
 
 	r = -EINVAL;
-	if (id >= KVM_MAX_VCPUS) {
-		printk(KERN_ERR"kvm: Can't configure vcpus > %ld",
-				KVM_MAX_VCPUS);
+
+	max_cpu = min_t(long, KVM_MAX_VCPUS, KVM_MAX_POSSIBLE_VCPUS);
+	if (id >= max_cpu) {
+		printk(KERN_ERR "kvm: Can't configure vcpus > %i", max_cpu);
 		goto fail;
 	}
 
@@ -1353,12 +1353,12 @@
 static void free_kvm(struct kvm *kvm)
 {
 	unsigned long vm_base = kvm->arch.vm_base;
+	unsigned long vm_size = 1UL << kvm_vm_data_shift;
 
 	if (vm_base) {
-		memset((void *)vm_base, 0, KVM_VM_DATA_SIZE);
-		free_pages(vm_base, get_order(KVM_VM_DATA_SIZE));
+		memset((void *)vm_base, 0, vm_size);
+		free_pages(vm_base, kvm_vm_data_shift - PAGE_SHIFT);
 	}
-
 }
 
 static void kvm_release_vm_pages(struct kvm *kvm)

RE: [patch] allow dynamic sizing the kvm vmm area on ia64

[Zhang, Xiantao]

Hi, Jes
     I am in holidays these days, sorry for late response.  I think you are saying vm area not vmm area.  In addition,  do you have plan to add the part to dynamicly set the value of kvm_vm_data_shift ?  You know, it has no difference with before until this part is ready.  Maybe we can dynamicly cacluate the value in upspace according to the amount of vcups, and transfer it to kernel at the stage of vm creation.  Another thing we need to consider is its value should follow tlb's page size, otherwise, it may result in other issues.  Thanks for your effort! 
Xiantao

Jes Sorensen wrote:
> Hi,
> 
> This patch allows dynamically sizing the vmm area on ia64. I have left
> it with the same size as we used to have with the constants, but it
> seems to work going higher. I put in an artificial limit of 64 vcpus
> for now anyway as scaling wise it doesn't work well going higher
> anyway..... we'll get there :-)
> 
> Cheers,
> Jes

Re: [patch] allow dynamic sizing the kvm vmm area on ia64

[Jes Sorensen]

Zhang, Xiantao wrote:
> Hi, Jes
>      I am in holidays these days, sorry for late response.  I think
 > you are saying vm area not vmm area.  In addition,  do you have plan
 > to add the part to dynamicly set the value of kvm_vm_data_shift ?
 > You know, it has no difference with before until this part is ready.
> Maybe we can dynamicly cacluate the value in upspace according to the
> amount of vcups, and transfer it to kernel at the stage of vm
> creation.  Another thing we need to consider is its value should
> follow tlb's page size, otherwise, it may result in other issues. 
> Thanks for your effort! 

Hi Xiantao,

Thanks for taking a look at this. Yes I do plan to add some logic for
setting this from QEMU. Basically what I talked to Hollis about was to
create some KVM_CONTEXT_CREATE or similar ioctl, where one requests a
maximum size of the KVM instance, like maximum possible number of vcpus,
memory etc.

Cheers,
Jes


> Xiantao
> 
> Jes Sorensen wrote:
>> Hi,
>>
>> This patch allows dynamically sizing the vmm area on ia64. I have left
>> it with the same size as we used to have with the constants, but it
>> seems to work going higher. I put in an artificial limit of 64 vcpus
>> for now anyway as scaling wise it doesn't work well going higher
>> anyway..... we'll get there :-)
>>
>> Cheers,
>> Jes