[PATCH v11 14/21] virt: geniezone: Optimize performance of protected VM memory

[Liju-clr Chen]

From: Yi-De Wu <yi-de.wu@xxxxxxxxxxxx>

From: "Yingshiuan Pan" <yingshiuan.pan@xxxxxxxxxxxx>

The memory protection mechanism performs better with batch operations on
memory pages. To leverage this, we pre-allocate memory for VMs that are
set to protected mode. As a result, the memory protection mechanism can
proactively protect the pre-allocated memory in advance through batch
operations, leading to improved performance during VM booting.

Signed-off-by: Yingshiuan Pan <yingshiuan.pan@xxxxxxxxxxxx>
Signed-off-by: Jerry Wang <ze-yu.wang@xxxxxxxxxxxx>
Signed-off-by: Liju Chen <liju-clr.chen@xxxxxxxxxxxx>
Signed-off-by: Yi-De Wu <yi-de.wu@xxxxxxxxxxxx>
---
 arch/arm64/geniezone/vm.c             | 127 ++++++++++++++++++++++++++
 include/linux/soc/mediatek/gzvm_drv.h |   1 +
 2 files changed, 128 insertions(+)

diff --git a/arch/arm64/geniezone/vm.c b/arch/arm64/geniezone/vm.c
index 00c74a4bcfd7..db16716f5b8d 100644
--- a/arch/arm64/geniezone/vm.c
+++ b/arch/arm64/geniezone/vm.c
@@ -11,6 +11,8 @@
 #include <linux/soc/mediatek/gzvm_drv.h>
 #include "gzvm_arch_common.h"
 
+#define PAR_PA47_MASK GENMASK_ULL(47, 12)
+
 /**
  * gzvm_hypcall_wrapper() - the wrapper for hvc calls
  * @a0: arguments passed in registers 0
@@ -201,6 +203,126 @@ static int gzvm_vm_ioctl_get_pvmfw_size(struct gzvm *gzvm,
 	return 0;
 }
 
+/**
+ * fill_constituents() - Populate pa to buffer until full
+ * @consti: Pointer to struct mem_region_addr_range.
+ * @consti_cnt: Constituent count.
+ * @max_nr_consti: Maximum number of constituent count.
+ * @gfn: Guest frame number.
+ * @total_pages: Total page numbers.
+ * @slot: Pointer to struct gzvm_memslot.
+ *
+ * Return: how many pages we've fill in, negative if error
+ */
+static int fill_constituents(struct mem_region_addr_range *consti,
+			     int *consti_cnt, int max_nr_consti, u64 gfn,
+			     u32 total_pages, struct gzvm_memslot *slot)
+{
+	u64 pfn = 0, prev_pfn = 0, gfn_end = 0;
+	int nr_pages = 0;
+	int i = -1;
+
+	if (unlikely(total_pages == 0))
+		return -EINVAL;
+	gfn_end = gfn + total_pages;
+
+	while (i < max_nr_consti && gfn < gfn_end) {
+		if (pfn == (prev_pfn + 1)) {
+			consti[i].pg_cnt++;
+		} else {
+			i++;
+			if (i >= max_nr_consti)
+				break;
+			consti[i].address = PFN_PHYS(pfn);
+			consti[i].pg_cnt = 1;
+		}
+		prev_pfn = pfn;
+		gfn++;
+		nr_pages++;
+	}
+	if (i != max_nr_consti)
+		i++;
+	*consti_cnt = i;
+
+	return nr_pages;
+}
+
+/**
+ * gzvm_vm_populate_mem_region() - Iterate all mem slot and populate pa to
+ * buffer until it's full
+ * @gzvm: Pointer to struct gzvm.
+ * @slot_id: Memory slot id to be populated.
+ *
+ * Return: 0 if it is successful, negative if error
+ */
+int gzvm_vm_populate_mem_region(struct gzvm *gzvm, int slot_id)
+{
+	struct gzvm_memslot *memslot = &gzvm->memslot[slot_id];
+	struct gzvm_memory_region_ranges *region;
+	int max_nr_consti, remain_pages;
+	u64 gfn, gfn_end;
+	u32 buf_size;
+
+	buf_size = PAGE_SIZE * 2;
+	region = alloc_pages_exact(buf_size, GFP_KERNEL);
+	if (!region)
+		return -ENOMEM;
+
+	max_nr_consti = (buf_size - sizeof(*region)) /
+			sizeof(struct mem_region_addr_range);
+
+	region->slot = memslot->slot_id;
+	remain_pages = memslot->npages;
+	gfn = memslot->base_gfn;
+	gfn_end = gfn + remain_pages;
+
+	while (gfn < gfn_end) {
+		int nr_pages;
+
+		nr_pages = fill_constituents(region->constituents,
+					     &region->constituent_cnt,
+					     max_nr_consti, gfn,
+					     remain_pages, memslot);
+
+		if (nr_pages < 0) {
+			pr_err("Failed to fill constituents\n");
+			free_pages_exact(region, buf_size);
+			return -EFAULT;
+		}
+
+		region->gpa = PFN_PHYS(gfn);
+		region->total_pages = nr_pages;
+		remain_pages -= nr_pages;
+		gfn += nr_pages;
+
+		if (gzvm_arch_set_memregion(gzvm->vm_id, buf_size,
+					    virt_to_phys(region))) {
+			pr_err("Failed to register memregion to hypervisor\n");
+			free_pages_exact(region, buf_size);
+			return -EFAULT;
+		}
+	}
+	free_pages_exact(region, buf_size);
+
+	return 0;
+}
+
+static int populate_all_mem_regions(struct gzvm *gzvm)
+{
+	int ret, i;
+
+	for (i = 0; i < GZVM_MAX_MEM_REGION; i++) {
+		if (gzvm->memslot[i].npages == 0)
+			continue;
+
+		ret = gzvm_vm_populate_mem_region(gzvm, i);
+		if (ret != 0)
+			return ret;
+	}
+
+	return 0;
+}
+
 /**
  * gzvm_vm_ioctl_cap_pvm() - Proceed GZVM_CAP_PROTECTED_VM's subcommands
  * @gzvm: Pointer to struct gzvm.
@@ -222,6 +344,11 @@ static int gzvm_vm_ioctl_cap_pvm(struct gzvm *gzvm,
 	case GZVM_CAP_PVM_SET_PVMFW_GPA:
 		fallthrough;
 	case GZVM_CAP_PVM_SET_PROTECTED_VM:
+		/*
+		 * To improve performance for protected VM, we have to populate VM's memory
+		 * before VM booting
+		 */
+		populate_all_mem_regions(gzvm);
 		ret = gzvm_vm_arch_enable_cap(gzvm, cap, &res);
 		return ret;
 	case GZVM_CAP_PVM_GET_PVMFW_SIZE:
diff --git a/include/linux/soc/mediatek/gzvm_drv.h b/include/linux/soc/mediatek/gzvm_drv.h
index 982463eea4f6..54ac91670611 100644
--- a/include/linux/soc/mediatek/gzvm_drv.h
+++ b/include/linux/soc/mediatek/gzvm_drv.h
@@ -159,6 +159,7 @@ int gzvm_vm_ioctl_arch_enable_cap(struct gzvm *gzvm,
 
 int gzvm_gfn_to_hva_memslot(struct gzvm_memslot *memslot, u64 gfn,
 			    u64 *hva_memslot);
+int gzvm_vm_populate_mem_region(struct gzvm *gzvm, int slot_id);
 
 int gzvm_vm_ioctl_create_vcpu(struct gzvm *gzvm, u32 cpuid);
 int gzvm_arch_vcpu_update_one_reg(struct gzvm_vcpu *vcpu, __u64 reg_id,
-- 
2.18.0