Auto merge of #114330 - RalfJung:dagling-ptr-deref, r=oli-obk

don't UB on dangling ptr deref, instead check inbounds on projections

This implements rust-lang/reference#1387 in Miri. See that PR for what the change is about.

Detecting dangling references in `let x = &...;` is now done by validity checking only, so some tests need to have validity checking enabled. There is no longer inherently a "nodangle" check in evaluating the expression `&*ptr` (aside from the aliasing model).

r? `@oli-obk`

Based on:
- rust-lang/reference#1387
- rust-lang/rust#115524

Author: bors

src/borrow_tracker/stacked_borrows/mod.rs

@@ -14,7 +14,7 @@ use log::trace;
 use rustc_data_structures::fx::FxHashSet;
 use rustc_middle::mir::{Mutability, RetagKind};
 use rustc_middle::ty::{self, layout::HasParamEnv, Ty};
-use rustc_target::abi::{Abi, Align, Size};
+use rustc_target::abi::{Abi, Size};
 
 use crate::borrow_tracker::{
     stacked_borrows::diagnostics::{AllocHistory, DiagnosticCx, DiagnosticCxBuilder},
@@ -616,7 +616,7 @@ trait EvalContextPrivExt<'mir: 'ecx, 'tcx: 'mir, 'ecx>: crate::MiriInterpCxExt<'
     ) -> InterpResult<'tcx, Option<Provenance>> {
         let this = self.eval_context_mut();
         // Ensure we bail out if the pointer goes out-of-bounds (see miri#1050).
-        this.check_ptr_access_align(place.ptr(), size, Align::ONE, CheckInAllocMsg::InboundsTest)?;
+        this.check_ptr_access(place.ptr(), size, CheckInAllocMsg::InboundsTest)?;
 
         // It is crucial that this gets called on all code paths, to ensure we track tag creation.
         let log_creation = |this: &MiriInterpCx<'mir, 'tcx>,

src/borrow_tracker/tree_borrows/mod.rs

@@ -1,6 +1,6 @@
 use log::trace;
 
-use rustc_target::abi::{Abi, Align, Size};
+use rustc_target::abi::{Abi, Size};
 
 use crate::borrow_tracker::{
     AccessKind, GlobalState, GlobalStateInner, ProtectorKind, RetagFields,
@@ -206,10 +206,9 @@ trait EvalContextPrivExt<'mir: 'ecx, 'tcx: 'mir, 'ecx>: crate::MiriInterpCxExt<'
         // Make sure the new permission makes sense as the initial permission of a fresh tag.
         assert!(new_perm.initial_state.is_initial());
         // Ensure we bail out if the pointer goes out-of-bounds (see miri#1050).
-        this.check_ptr_access_align(
+        this.check_ptr_access(
             place.ptr(),
             ptr_size,
-            Align::ONE,
             CheckInAllocMsg::InboundsTest,
         )?;
 

src/concurrency/data_race.rs

@@ -1017,11 +1017,9 @@ trait EvalContextPrivExt<'mir, 'tcx: 'mir>: MiriInterpCxExt<'mir, 'tcx> {
         // even if the type they wrap would be less aligned (e.g. AtomicU64 on 32bit must
         // be 8-aligned).
         let align = Align::from_bytes(place.layout.size.bytes()).unwrap();
-        this.check_ptr_access_align(
+        this.check_ptr_align(
             place.ptr(),
-            place.layout.size,
             align,
-            CheckInAllocMsg::MemoryAccessTest,
         )?;
         // Ensure the allocation is mutable. Even failing (read-only) compare_exchange need mutable
         // memory on many targets (i.e., they segfault if taht memory is mapped read-only), and

src/helpers.rs

@@ -697,27 +697,7 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
     ) -> InterpResult<'tcx, MPlaceTy<'tcx, Provenance>> {
         let this = self.eval_context_ref();
         let ptr = this.read_pointer(op)?;
-
-        let mplace = MPlaceTy::from_aligned_ptr(ptr, layout);
-
-        this.check_mplace(&mplace)?;
-
-        Ok(mplace)
-    }
-
-    /// Deref' a pointer *without* checking that the place is dereferenceable.
-    fn deref_pointer_unchecked(
-        &self,
-        val: &ImmTy<'tcx, Provenance>,
-        layout: TyAndLayout<'tcx>,
-    ) -> InterpResult<'tcx, MPlaceTy<'tcx, Provenance>> {
-        let this = self.eval_context_ref();
-        let mut mplace = this.ref_to_mplace(val)?;
-
-        mplace.layout = layout;
-        mplace.align = layout.align.abi;
-
-        Ok(mplace)
+        Ok(this.ptr_to_mplace(ptr, layout))
     }
 
     /// Calculates the MPlaceTy given the offset and layout of an access on an operand
@@ -805,7 +785,7 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
         loop {
             // FIXME: We are re-getting the allocation each time around the loop.
             // Would be nice if we could somehow "extend" an existing AllocRange.
-            let alloc = this.get_ptr_alloc(ptr.offset(len, this)?, size1, Align::ONE)?.unwrap(); // not a ZST, so we will get a result
+            let alloc = this.get_ptr_alloc(ptr.offset(len, this)?, size1)?.unwrap(); // not a ZST, so we will get a result
             let byte = alloc.read_integer(alloc_range(Size::ZERO, size1))?.to_u8()?;
             if byte == 0 {
                 break;
@@ -845,13 +825,13 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
     fn read_wide_str(&self, mut ptr: Pointer<Option<Provenance>>) -> InterpResult<'tcx, Vec<u16>> {
         let this = self.eval_context_ref();
         let size2 = Size::from_bytes(2);
-        let align2 = Align::from_bytes(2).unwrap();
+        this.check_ptr_align(ptr, Align::from_bytes(2).unwrap())?;
 
         let mut wchars = Vec::new();
         loop {
             // FIXME: We are re-getting the allocation each time around the loop.
             // Would be nice if we could somehow "extend" an existing AllocRange.
-            let alloc = this.get_ptr_alloc(ptr, size2, align2)?.unwrap(); // not a ZST, so we will get a result
+            let alloc = this.get_ptr_alloc(ptr, size2)?.unwrap(); // not a ZST, so we will get a result
             let wchar = alloc.read_integer(alloc_range(Size::ZERO, size2))?.to_u16()?;
             if wchar == 0 {
                 break;
@@ -887,8 +867,9 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
         // Store the UTF-16 string.
         let size2 = Size::from_bytes(2);
         let this = self.eval_context_mut();
+        this.check_ptr_align(ptr, Align::from_bytes(2).unwrap())?;
         let mut alloc = this
-            .get_ptr_alloc_mut(ptr, size2 * string_length, Align::from_bytes(2).unwrap())?
+            .get_ptr_alloc_mut(ptr, size2 * string_length)?
             .unwrap(); // not a ZST, so we will get a result
         for (offset, wchar) in wide_str.iter().copied().chain(iter::once(0x0000)).enumerate() {
             let offset = u64::try_from(offset).unwrap();

src/machine.rs

@@ -1285,6 +1285,7 @@ impl<'mir, 'tcx> Machine<'mir, 'tcx> for MiriMachine<'mir, 'tcx> {
         // We do need to write `uninit` so that even after the call ends, the former contents of
         // this place cannot be observed any more. We do the write after retagging so that for
         // Tree Borrows, this is considered to activate the new tag.
+        // Conveniently this also ensures that the place actually points to suitable memory.
         ecx.write_uninit(&protected_place)?;
         // Now we throw away the protected place, ensuring its tag is never used again.
         Ok(())

src/shims/foreign_items.rs

@@ -807,9 +807,7 @@ trait EvalContextExtPriv<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
 
                 this.mem_copy(
                     ptr_src,
-                    Align::ONE,
                     ptr_dest,
-                    Align::ONE,
                     Size::from_bytes(n),
                     true,
                 )?;
@@ -830,9 +828,7 @@ trait EvalContextExtPriv<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
                 let n = this.read_c_str(ptr_src)?.len().checked_add(1).unwrap();
                 this.mem_copy(
                     ptr_src,
-                    Align::ONE,
                     ptr_dest,
-                    Align::ONE,
                     Size::from_bytes(n),
                     true,
                 )?;

src/shims/unix/fs.rs

@@ -13,7 +13,7 @@ use log::trace;
 
 use rustc_data_structures::fx::FxHashMap;
 use rustc_middle::ty::TyCtxt;
-use rustc_target::abi::{Align, Size};
+use rustc_target::abi::Size;
 
 use crate::shims::os_str::bytes_to_os_str;
 use crate::*;
@@ -756,10 +756,9 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
         trace!("Reading from FD {}, size {}", fd, count);
 
         // Check that the *entire* buffer is actually valid memory.
-        this.check_ptr_access_align(
+        this.check_ptr_access(
             buf,
             Size::from_bytes(count),
-            Align::ONE,
             CheckInAllocMsg::MemoryAccessTest,
         )?;
 
@@ -810,10 +809,9 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
         // Isolation check is done via `FileDescriptor` trait.
 
         // Check that the *entire* buffer is actually valid memory.
-        this.check_ptr_access_align(
+        this.check_ptr_access(
             buf,
             Size::from_bytes(count),
-            Align::ONE,
             CheckInAllocMsg::MemoryAccessTest,
         )?;
 
@@ -1370,7 +1368,7 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
                         ("d_reclen", size.into()),
                         ("d_type", file_type.into()),
                     ],
-                    &MPlaceTy::from_aligned_ptr(entry, dirent64_layout),
+                    &this.ptr_to_mplace(entry, dirent64_layout),
                 )?;
 
                 let name_ptr = entry.offset(Size::from_bytes(d_name_offset), this)?;

src/shims/unix/linux/sync.rs

@@ -34,7 +34,7 @@ pub fn futex<'tcx>(
 
     let thread = this.get_active_thread();
     // This is a vararg function so we have to bring our own type for this pointer.
-    let addr = MPlaceTy::from_aligned_ptr(addr, this.machine.layouts.i32);
+    let addr = this.ptr_to_mplace(addr, this.machine.layouts.i32);
     let addr_usize = addr.ptr().addr().bytes();
 
     let futex_private = this.eval_libc_i32("FUTEX_PRIVATE_FLAG");
@@ -85,9 +85,8 @@ pub fn futex<'tcx>(
                 return Ok(());
             }
 
-            // `read_timespec` will check the place when it is not null.
-            let timeout = this.deref_pointer_unchecked(
-                &this.read_immediate(&args[3])?,
+            let timeout = this.deref_pointer_as(
+                &args[3],
                 this.libc_ty_layout("timespec"),
             )?;
             let timeout_time = if this.ptr_is_null(timeout.ptr())? {

src/shims/windows/sync.rs

@@ -322,8 +322,8 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
 
         let layout = this.machine.layouts.uint(size).unwrap();
         let futex_val = this
-            .read_scalar_atomic(&MPlaceTy::from_aligned_ptr(ptr, layout), AtomicReadOrd::Relaxed)?;
-        let compare_val = this.read_scalar(&MPlaceTy::from_aligned_ptr(compare, layout))?;
+            .read_scalar_atomic(&this.ptr_to_mplace(ptr, layout), AtomicReadOrd::Relaxed)?;
+        let compare_val = this.read_scalar(&this.ptr_to_mplace(compare, layout))?;
 
         if futex_val == compare_val {
             // If the values are the same, we have to block.

src/shims/x86/sse3.rs

@@ -1,6 +1,5 @@
 use rustc_middle::mir;
 use rustc_span::Symbol;
-use rustc_target::abi::Align;
 use rustc_target::spec::abi::Abi;
 
 use super::horizontal_bin_op;
@@ -76,9 +75,7 @@ pub(super) trait EvalContextExt<'mir, 'tcx: 'mir>:
 
                 this.mem_copy(
                     src_ptr,
-                    Align::ONE,
                     dest.ptr(),
-                    Align::ONE,
                     dest.layout.size,
                     /*nonoverlapping*/ true,
                 )?;