miri native-call support: all previously exposed provenance is accessible to the callee

Author: RalfJung

Diff for: compiler/rustc_const_eval/src/interpret/memory.rs

@@ -955,18 +955,13 @@ impl<'tcx, M: Machine<'tcx>> InterpCx<'tcx, M> {
 
     /// Handle the effect an FFI call might have on the state of allocations.
     /// This overapproximates the modifications which external code might make to memory:
-    /// We set all reachable allocations as initialized, mark all provenances as exposed
+    /// We set all reachable allocations as initialized, mark all reachable provenances as exposed
     /// and overwrite them with `Provenance::WILDCARD`.
-    pub fn prepare_for_native_call(
-        &mut self,
-        id: AllocId,
-        initial_prov: M::Provenance,
-    ) -> InterpResult<'tcx> {
-        // Expose provenance of the root allocation.
-        M::expose_provenance(self, initial_prov)?;
-
+    ///
+    /// The allocations in `ids` are assumed to be already exposed.
+    pub fn prepare_for_native_call(&mut self, ids: Vec<AllocId>) -> InterpResult<'tcx> {
         let mut done = FxHashSet::default();
-        let mut todo = vec![id];
+        let mut todo = ids;
         while let Some(id) = todo.pop() {
             if !done.insert(id) {
                 // We already saw this allocation before, don't process it again.

Diff for: src/tools/miri/src/alloc_addresses/mod.rs

@@ -285,9 +285,19 @@ trait EvalContextExtPriv<'tcx>: crate::MiriInterpCxExt<'tcx> {
 
 impl<'tcx> EvalContextExt<'tcx> for crate::MiriInterpCx<'tcx> {}
 pub trait EvalContextExt<'tcx>: crate::MiriInterpCxExt<'tcx> {
-    fn expose_ptr(&self, alloc_id: AllocId, tag: BorTag) -> InterpResult<'tcx> {
+    fn expose_provenance(&self, provenance: Provenance) -> InterpResult<'tcx> {
         let this = self.eval_context_ref();
         let mut global_state = this.machine.alloc_addresses.borrow_mut();
+
+        let (alloc_id, tag) = match provenance {
+            Provenance::Concrete { alloc_id, tag } => (alloc_id, tag),
+            Provenance::Wildcard => {
+                // No need to do anything for wildcard pointers as
+                // their provenances have already been previously exposed.
+                return interp_ok(());
+            }
+        };
+
         // In strict mode, we don't need this, so we can save some cycles by not tracking it.
         if global_state.provenance_mode == ProvenanceMode::Strict {
             return interp_ok(());
@@ -422,6 +432,19 @@ pub trait EvalContextExt<'tcx>: crate::MiriInterpCxExt<'tcx> {
         let rel_offset = this.truncate_to_target_usize(addr.bytes().wrapping_sub(base_addr));
         Some((alloc_id, Size::from_bytes(rel_offset)))
     }
+
+    /// Prepare all exposed memory for a native call.
+    /// This overapproximates the modifications which external code might make to memory:
+    /// We set all reachable allocations as initialized, mark all reachable provenances as exposed
+    /// and overwrite them with `Provenance::WILDCARD`.
+    fn prepare_exposed_for_native_call(&mut self) -> InterpResult<'tcx> {
+        let this = self.eval_context_mut();
+        // We need to make a deep copy of this list, but it's fine; it also serves as scratch space
+        // for the search within `prepare_for_native_call`.
+        let exposed: Vec<AllocId> =
+            this.machine.alloc_addresses.get_mut().exposed.iter().copied().collect();
+        this.prepare_for_native_call(exposed)
+    }
 }
 
 impl<'tcx> MiriMachine<'tcx> {

Diff for: src/tools/miri/src/machine.rs

@@ -1291,18 +1291,12 @@ impl<'tcx> Machine<'tcx> for MiriMachine<'tcx> {
     /// Called on `ptr as usize` casts.
     /// (Actually computing the resulting `usize` doesn't need machine help,
     /// that's just `Scalar::try_to_int`.)
+    #[inline(always)]
     fn expose_provenance(
         ecx: &InterpCx<'tcx, Self>,
         provenance: Self::Provenance,
     ) -> InterpResult<'tcx> {
-        match provenance {
-            Provenance::Concrete { alloc_id, tag } => ecx.expose_ptr(alloc_id, tag),
-            Provenance::Wildcard => {
-                // No need to do anything for wildcard pointers as
-                // their provenances have already been previously exposed.
-                interp_ok(())
-            }
-        }
+        ecx.expose_provenance(provenance)
     }
 
     /// Convert a pointer with provenance into an allocation-offset pair and extra provenance info.

Diff for: src/tools/miri/src/shims/native_lib.rs

@@ -160,16 +160,12 @@ pub trait EvalContextExt<'tcx>: crate::MiriInterpCxExt<'tcx> {
             }
             let imm = this.read_immediate(arg)?;
             libffi_args.push(imm_to_carg(&imm, this)?);
-            // If we are passing a pointer, prepare the memory it points to.
+            // If we are passing a pointer, expose its provenance. Below, all exposed memory
+            // (previously exposed and new exposed) will then be properly prepared.
             if matches!(arg.layout.ty.kind(), ty::RawPtr(..)) {
                 let ptr = imm.to_scalar().to_pointer(this)?;
                 let Some(prov) = ptr.provenance else {
-                    // Pointer without provenance may not access any memory.
-                    continue;
-                };
-                // We use `get_alloc_id` for its best-effort behaviour with Wildcard provenance.
-                let Some(alloc_id) = prov.get_alloc_id() else {
-                    // Wildcard pointer, whatever it points to must be already exposed.
+                    // Pointer without provenance may not access any memory anyway, skip.
                     continue;
                 };
                 // The first time this happens, print a warning.
@@ -178,12 +174,12 @@ pub trait EvalContextExt<'tcx>: crate::MiriInterpCxExt<'tcx> {
                     this.emit_diagnostic(NonHaltingDiagnostic::NativeCallSharedMem);
                 }
 
-                this.prepare_for_native_call(alloc_id, prov)?;
+                this.expose_provenance(prov)?;
             }
         }
 
-        // FIXME: In the future, we should also call `prepare_for_native_call` on all previously
-        // exposed allocations, since C may access any of them.
+        // Prepare all exposed memory.
+        this.prepare_exposed_for_native_call()?;
 
         // Convert them to `libffi::high::Arg` type.
         let libffi_args = libffi_args

Diff for: src/tools/miri/tests/native-lib/pass/ptr_write_access.rs

@@ -6,7 +6,7 @@
 #![feature(box_as_ptr)]
 
 use std::mem::MaybeUninit;
-use std::ptr::null;
+use std::ptr;
 
 fn main() {
     test_increment_int();
@@ -20,6 +20,8 @@ fn main() {
     test_pass_dangling();
     test_swap_ptr_triple_dangling();
     test_return_ptr();
+    test_pass_ptr_as_int();
+    test_pass_ptr_via_previously_shared_mem();
 }
 
 /// Test function that modifies an int.
@@ -112,7 +114,7 @@ fn test_swap_ptr() {
     }
 
     let x = 61;
-    let (mut ptr0, mut ptr1) = (&raw const x, null());
+    let (mut ptr0, mut ptr1) = (&raw const x, ptr::null());
 
     unsafe { swap_ptr(&mut ptr0, &mut ptr1) };
     assert_eq!(unsafe { *ptr1 }, x);
@@ -131,7 +133,7 @@ fn test_swap_ptr_tuple() {
     }
 
     let x = 71;
-    let mut tuple = Tuple { ptr0: &raw const x, ptr1: null() };
+    let mut tuple = Tuple { ptr0: &raw const x, ptr1: ptr::null() };
 
     unsafe { swap_ptr_tuple(&mut tuple) }
     assert_eq!(unsafe { *tuple.ptr1 }, x);
@@ -148,7 +150,7 @@ fn test_overwrite_dangling() {
     drop(b);
 
     unsafe { overwrite_ptr(&mut ptr) };
-    assert_eq!(ptr, null());
+    assert_eq!(ptr, ptr::null());
 }
 
 /// Test function that passes a dangling pointer.
@@ -200,3 +202,33 @@ fn test_return_ptr() {
     let ptr = unsafe { return_ptr(ptr) };
     assert_eq!(unsafe { *ptr }, x);
 }
+
+/// Test casting a pointer to an integer and passing that to C.
+fn test_pass_ptr_as_int() {
+    extern "C" {
+        fn pass_ptr_as_int(ptr: usize, set_to_val: i32);
+    }
+
+    let mut m: MaybeUninit<i32> = MaybeUninit::uninit();
+    unsafe { pass_ptr_as_int(m.as_mut_ptr() as usize, 42) };
+    assert_eq!(unsafe { m.assume_init() }, 42);
+}
+
+fn test_pass_ptr_via_previously_shared_mem() {
+    extern "C" {
+        fn set_shared_mem(ptr: *mut *mut i32);
+        fn init_ptr_stored_in_shared_mem(val: i32);
+    }
+
+    let mut m: *mut i32 = ptr::null_mut();
+    let ptr_to_m = &raw mut m;
+    unsafe { set_shared_mem(&raw mut m) };
+
+    let mut m2: MaybeUninit<i32> = MaybeUninit::uninit();
+    // Store a pointer to m2 somewhere that C code can access it.
+    unsafe { ptr_to_m.write(m2.as_mut_ptr()) };
+    // Have C code write there.
+    unsafe { init_ptr_stored_in_shared_mem(42) };
+    // Ensure this memory is now considered initialized.
+    assert_eq!(unsafe { m2.assume_init() }, 42);
+}

Diff for: src/tools/miri/tests/native-lib/ptr_write_access.c

@@ -1,4 +1,5 @@
 #include <stddef.h>
+#include <stdint.h>
 
 // See comments in build_native_lib()
 #define EXPORT __attribute__((visibility("default")))
@@ -88,3 +89,21 @@ EXPORT void swap_ptr_triple_dangling(Triple *t_ptr) {
 EXPORT const int *return_ptr(const int *ptr) {
   return ptr;
 }
+
+/* Test: test_pass_ptr_as_int */
+
+EXPORT void pass_ptr_as_int(uintptr_t ptr, int set_to_val) {
+  *(int*)ptr = set_to_val;
+}
+
+/* Test: test_pass_ptr_via_previously_shared_mem */
+
+int** shared_place;
+
+EXPORT void set_shared_mem(int** ptr) {
+  shared_place = ptr;
+}
+
+EXPORT void init_ptr_stored_in_shared_mem(int val) {
+  **shared_place = val;
+}