Skip to contentSkip to conversation

x86: Look up the PTEs rather than assuming them

PATCH In review Aaron Lewis 2 participants 2 messages Nov 2, 2021 – Nov 9, 2021

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 x86: Look up the PTEs rather than assuming them

Rather than assuming which PTEs the SMEP test runs on, look them up to ensure they are correct. If this test were to run on a different page table (ie: run in an L2 test) the wrong PTEs would be set. Switch to looking up the PTEs to avoid this from happening.

---
 lib/libcflat.h |  1 +
 lib/x86/vm.c   | 19 +++++++++++++++++++
 lib/x86/vm.h   |  3 +++
 x86/access.c   | 23 +++++++++++++++--------
 x86/cstart64.S |  1 -
 x86/flat.lds   |  1 +
 6 files changed, 39 insertions(+), 9 deletions(-)
lib/libcflat.h +1 -0
index 9bb7e08..c1fd31f 100644
@@ -35,6 +35,7 @@
 #define __ALIGN_MASK(x, mask)	(((x) + (mask)) & ~(mask))
 #define __ALIGN(x, a)		__ALIGN_MASK(x, (typeof(x))(a) - 1)
 #define ALIGN(x, a)		__ALIGN((x), (a))
+#define ALIGN_DOWN(x, a)	__ALIGN((x) - ((a) - 1), (a))
 #define IS_ALIGNED(x, a)	(((x) & ((typeof(x))(a) - 1)) == 0)
 
 #define MIN(a, b)		((a) < (b) ? (a) : (b))
lib/x86/vm.c +19 -0
index 5cd2ee4..cbecddc 100644
@@ -281,3 +281,22 @@ void force_4k_page(void *addr)
 	if (pte & PT_PAGE_SIZE_MASK)
 		split_large_page(ptep, 2);
 }
+
+/*
+ * Call the callback, function, on each page from va_start to va_end.
+ */
+void walk_pte(void *va_start, void *va_end,
+	      void (*function)(pteval_t *pte, void *va)){
+    struct pte_search search;
+    uintptr_t curr_addr;
+    u64 page_size;
+
+    for (curr_addr = (uintptr_t)va_start; curr_addr < (uintptr_t)va_end;
+         curr_addr = ALIGN_DOWN(curr_addr + page_size, page_size)) {
+        search = find_pte_level(current_page_table(), (void *)curr_addr, 1);
+        assert(found_leaf_pte(search));
+        page_size = 1ul << PGDIR_BITS(search.level);
+
+        function(search.pte, (void *)curr_addr);
+    }
+}
lib/x86/vm.h +3 -0
index d9753c3..f4ac2c5 100644
@@ -52,4 +52,7 @@ struct vm_vcpu_info {
         u64 cr0;
 };
 
+void walk_pte(void *va_start, void *va_end,
+	      void (*function)(pteval_t *pte, void *va));
+
 #endif
x86/access.c +15 -8
index 4725bbd..17a6ef9 100644
@@ -201,10 +201,21 @@ static void set_cr0_wp(int wp)
     }
 }
 
+static void clear_user_mask(pteval_t *pte, void *va) {
+	*pte &= ~PT_USER_MASK;
+}
+
+static void set_user_mask(pteval_t *pte, void *va) {
+	*pte |= PT_USER_MASK;
+
+	/* Flush to avoid spurious #PF */
+	invlpg(va);
+}
+
 static unsigned set_cr4_smep(int smep)
 {
+    extern char stext, etext;
     unsigned long cr4 = shadow_cr4;
-    extern u64 ptl2[];
     unsigned r;
 
     cr4 &= ~CR4_SMEP_MASK;
@@ -214,14 +225,10 @@ static unsigned set_cr4_smep(int smep)
         return 0;
 
     if (smep)
-        ptl2[2] &= ~PT_USER_MASK;
+        walk_pte(&stext, &etext, clear_user_mask);
     r = write_cr4_checking(cr4);
-    if (r || !smep) {
-        ptl2[2] |= PT_USER_MASK;
-
-	/* Flush to avoid spurious #PF */
-	invlpg((void *)(2 << 21));
-    }
+    if (r || !smep)
+        walk_pte(&stext, &etext, set_user_mask);
     if (!r)
         shadow_cr4 = cr4;
     return r;
x86/cstart64.S +0 -1
index 5c6ad38..4ba9943 100644
@@ -26,7 +26,6 @@ ring0stacktop:
 .data
 
 .align 4096
-.globl ptl2
 ptl2:
 i = 0
 	.rept 512 * 4
x86/flat.lds +1 -0
index a278b56..337bc44 100644
@@ -3,6 +3,7 @@ SECTIONS
     . = 4M + SIZEOF_HEADERS;
     stext = .;
     .text : { *(.init) *(.text) *(.text.*) }
+    etext = .;
     . = ALIGN(4K);
     .data : {
           *(.data)
Signature · 1 line
2.34.0.rc0.344.g81b53c2807-goog
#1 lore raw
Quoted 32 lines · Aaron Lewis

Rather than assuming which PTEs the SMEP test runs on, look them up to ensure they are correct. If this test were to run on a different page table (ie: run in an L2 test) the wrong PTEs would be set. Switch to looking up the PTEs to avoid this from happening.

---
 lib/libcflat.h |  1 +
 lib/x86/vm.c   | 19 +++++++++++++++++++
 lib/x86/vm.h   |  3 +++
 x86/access.c   | 23 +++++++++++++++--------
 x86/cstart64.S |  1 -
 x86/flat.lds   |  1 +
 6 files changed, 39 insertions(+), 9 deletions(-)
lib/libcflat.h +1 -0
index 9bb7e08..c1fd31f 100644
@@ -35,6 +35,7 @@
 #define __ALIGN_MASK(x, mask)	(((x) + (mask)) & ~(mask))
 #define __ALIGN(x, a)		__ALIGN_MASK(x, (typeof(x))(a) - 1)
 #define ALIGN(x, a)		__ALIGN((x), (a))
+#define ALIGN_DOWN(x, a)	__ALIGN((x) - ((a) - 1), (a))
 #define IS_ALIGNED(x, a)	(((x) & ((typeof(x))(a) - 1)) == 0)
 
 #define MIN(a, b)		((a) < (b) ? (a) : (b))
lib/x86/vm.c +0 -0
index 5cd2ee4..cbecddc 100644

lib/x86/vm.c void force_4k_page(void *addr) · line 288

Quoted 8 lines
lib/x86/vm.c +5 -0
@@ -281,3 +281,22 @@ void force_4k_page(void *addr)
 	if (pte & PT_PAGE_SIZE_MASK)
 		split_large_page(ptep, 2);
 }
+
+/*
+ * Call the callback, function, on each page from va_start to va_end.
+ */
+void walk_pte(void *va_start, void *va_end,

To align with other helpers in vm.c, "void *virt, size_t len" would be more approriate. That would also avoid having to document whether or not va_end is inclusive or exclusive.

lib/x86/vm.c void force_4k_page(void *addr) · line 289

Quoted 1 line
lib/x86/vm.c +1 -0
+	      void (*function)(pteval_t *pte, void *va)){

Curly brace goes on a separate line for functions.

lib/x86/vm.c void force_4k_page(void *addr) · line 291

Quoted 2 lines
lib/x86/vm.c +2 -0
+    struct pte_search search;
+    uintptr_t curr_addr;

Maybe just curr? To match other code and maybe squeeze the for-loop on a single line?

	for (curr = virt; curr < end; curr = ALIGN_DOWN(curr + page_size, page_size)) {

	}

If you do split the loop, it's usually easier to read if all three parts are on separate lines, e.g.

	for (curr = virt;
	     curr < end;
	     curr = curr = ALIGN_DOWN(curr + page_size, page_size)) {

	}

lib/x86/vm.c void force_4k_page(void *addr) · line 292

Quoted 1 line
lib/x86/vm.c +1 -0
+    u64 page_size;

Probably should be size_t.

lib/x86/vm.c void force_4k_page(void *addr) · line 296

Quoted 5 lines
lib/x86/vm.c +3 -0
+    for (curr_addr = (uintptr_t)va_start; curr_addr < (uintptr_t)va_end;
+         curr_addr = ALIGN_DOWN(curr_addr + page_size, page_size)) {
 
 
+        search = find_pte_level(current_page_table(), (void *)curr_addr, 1);

Probably worth caching current_page_table(). On x86 with shadow paging, that'll trigger an exit on every iteration to read CR3.

Quoted 11 lines
lib/x86/vm.c +6 -0
+        assert(found_leaf_pte(search));
+        page_size = 1ul << PGDIR_BITS(search.level);
+
+        function(search.pte, (void *)curr_addr);
+    }
+}
lib/x86/vm.h +0 -0
index d9753c3..f4ac2c5 100644

lib/x86/vm.h struct vm_vcpu_info { · line 56

Quoted 5 lines
lib/x86/vm.h +2 -0
@@ -52,4 +52,7 @@ struct vm_vcpu_info {
         u64 cr0;
 };
 
+void walk_pte(void *va_start, void *va_end,
+	      void (*function)(pteval_t *pte, void *va));

A typedef for the function pointer would be helpful. Maybe pte_callback_t? And pass "struct pte_search" instead of pteval_t so that future users can get at the level?

x86/access.c static void set_cr0_wp(int wp) · line 204

Quoted 10 lines
lib/x86/vm.h +0 -0
 #endif
x86/access.c +1 -0
index 4725bbd..17a6ef9 100644
@@ -201,10 +201,21 @@ static void set_cr0_wp(int wp)
     }
 }
 
+static void clear_user_mask(pteval_t *pte, void *va) {

Curly brace thing again.

x86/access.c static void set_cr0_wp(int wp) · line 208

Quoted 4 lines
x86/access.c +4 -0
+	*pte &= ~PT_USER_MASK;
+}
+
+static void set_user_mask(pteval_t *pte, void *va) {

And here.

Quoted 14 lines
x86/access.c +7 -1
+	*pte |= PT_USER_MASK;
+
+	/* Flush to avoid spurious #PF */
+	invlpg(va);
+}
+
 static unsigned set_cr4_smep(int smep)
 {
+    extern char stext, etext;
     unsigned long cr4 = shadow_cr4;
-    extern u64 ptl2[];
     unsigned r;
 
     cr4 &= ~CR4_SMEP_MASK;