From: Marc Orr <marcorr@xxxxxxxxxx>
This patch introduces split_large_page(), which is used to force the
APIC-access address to be a 4k page. Otherwise, the apic_reg_virt_test
fails on upstream.
Signed-off-by: Marc Orr <marcorr@xxxxxxxxxx>
Reviewed-by: Jim Mattson <jmattson@xxxxxxxxxx>
Signed-off-by: Paolo Bonzini <pbonzini@xxxxxxxxxx>
---
lib/x86/processor.h | 10 +++++++
lib/x86/vm.c | 83 +++++++++++++++++++++++++++++++++++++++++++++++++++++
lib/x86/vm.h | 3 ++
x86/vmx_tests.c | 6 ++++
4 files changed, 102 insertions(+)
diff --git a/lib/x86/processor.h b/lib/x86/processor.h
index 916e67d..6f77148 100644
--- a/lib/x86/processor.h
+++ b/lib/x86/processor.h
@@ -36,6 +36,7 @@
#define X86_CR4_PSE 0x00000010
#define X86_CR4_PAE 0x00000020
#define X86_CR4_MCE 0x00000040
+#define X86_CR4_PGE 0x00000080
#define X86_CR4_PCE 0x00000100
#define X86_CR4_UMIP 0x00000800
#define X86_CR4_VMXE 0x00002000
@@ -477,4 +478,13 @@ static inline void set_bit(int bit, u8 *addr)
: "+m" (*addr) : "Ir" (bit) : "cc", "memory");
}
+static inline void flush_tlb(void)
+{
+ ulong cr4;
+
+ cr4 = read_cr4();
+ write_cr4(cr4 ^ X86_CR4_PGE);
+ write_cr4(cr4);
+}
+
#endif
diff --git a/lib/x86/vm.c b/lib/x86/vm.c
index 73d9be4..edbbe82 100644
--- a/lib/x86/vm.c
+++ b/lib/x86/vm.c
@@ -177,3 +177,86 @@ phys_addr_t virt_to_pte_phys(pgd_t *cr3, void *mem)
{
return (*get_pte(cr3, mem) & PT_ADDR_MASK) + ((ulong)mem & (PAGE_SIZE - 1));
}
+
+/*
+ * split_large_page: Split a 2M/1G large page into 512 smaller PTEs.
+ * @ptep : large page table entry to split
+ * @level : level of ptep (2 or 3)
+ */
+void split_large_page(unsigned long *ptep, int level)
+{
+ unsigned long *new_pt;
+ unsigned long pa;
+ unsigned long pte;
+ unsigned long prototype;
+ int i;
+
+ pte = *ptep;
+ assert(pte & PT_PRESENT_MASK);
+ assert(pte & PT_PAGE_SIZE_MASK);
+ assert(level == 2 || level == 3);
+
+ new_pt = alloc_page();
+ assert(new_pt);
+
+ prototype = pte & ~PT_ADDR_MASK;
+ if (level == 2)
+ prototype &= ~PT_PAGE_SIZE_MASK;
+
+ pa = pte & PT_ADDR_MASK;
+ for (i = 0; i < (1 << PGDIR_WIDTH); i++) {
+ new_pt[i] = prototype | pa;
+ pa += 1ul << PGDIR_BITS(level - 1);
+ }
+
+ pte &= ~PT_PAGE_SIZE_MASK;
+ pte &= ~PT_ADDR_MASK;
+ pte |= virt_to_phys(new_pt);
+
+ /* Modify the relevant paging-structure entry */
+ *ptep = pte;
+
+ /*
+ * Flush the TLB to eradicate stale mappings.
+ *
+ * Note: Removing specific TLB mappings is tricky because
+ * split_large_page() can be called to split the active code page
+ * backing the next set of instructions to be fetched and executed.
+ * Furthermore, Intel SDM volume 3 recommends to clear the present bit
+ * for the page being split, before invalidating any mappings.
+ *
+ * But clearing the mapping from the page table and removing it from the
+ * TLB (where it's not actually guaranteed to reside anyway) makes it
+ * impossible to continue fetching instructions!
+ */
+ flush_tlb();
+}
+
+/*
+ * force_4k_page: Ensures that addr translate to a 4k page.
+ *
+ * This function uses split_large_page(), as needed, to ensure that target
+ * address, addr, translates to a 4k page.
+ *
+ * @addr: target address that should be mapped to a 4k page
+ */
+void force_4k_page(void *addr)
+{
+ unsigned long *ptep;
+ unsigned long pte;
+ unsigned long *cr3 = current_page_table();
+
+ ptep = get_pte_level(cr3, addr, 3);
+ assert(ptep);
+ pte = *ptep;
+ assert(pte & PT_PRESENT_MASK);
+ if (pte & PT_PAGE_SIZE_MASK)
+ split_large_page(ptep, 3);
+
+ ptep = get_pte_level(cr3, addr, 2);
+ assert(ptep);
+ pte = *ptep;
+ assert(pte & PT_PRESENT_MASK);
+ if (pte & PT_PAGE_SIZE_MASK)
+ split_large_page(ptep, 2);
+}
diff --git a/lib/x86/vm.h b/lib/x86/vm.h
index 729f172..8750a1e 100644
--- a/lib/x86/vm.h
+++ b/lib/x86/vm.h
@@ -42,4 +42,7 @@ static inline void *current_page_table(void)
{
return phys_to_virt(read_cr3());
}
+
+void split_large_page(unsigned long *ptep, int level);
+void force_4k_page(void *addr);
#endif
diff --git a/x86/vmx_tests.c b/x86/vmx_tests.c
index 73aca95..092e70e 100644
--- a/x86/vmx_tests.c
+++ b/x86/vmx_tests.c
@@ -5673,7 +5673,13 @@ static void apic_reg_virt_test(void)
test_set_guest(apic_reg_virt_guest);
+ /*
+ * From the SDM: The 1-setting of the "virtualize APIC accesses"
+ * VM-execution is guaranteed to apply only if translations to the
+ * APIC-access address use a 4-KByte page.
+ */
apic_access_address = alloc_page();
+ force_4k_page(apic_access_address);
vmcs_write(APIC_ACCS_ADDR, virt_to_phys(apic_access_address));
virtual_apic_page = alloc_page();
--
1.8.3.1
