1 use std::cell::RefCell;
2 use std::collections::hash_map::Entry;
7 use rustc_data_structures::fx::{FxHashMap, FxHashSet};
9 use rustc_target::abi::{HasDataLayout, Size};
13 #[derive(Copy, Clone, Debug, PartialEq, Eq)]
14 pub enum ProvenanceMode {
15 /// We support `expose_addr`/`from_exposed_addr` via "wildcard" provenance.
16 /// However, we want on `from_exposed_addr` to alert the user of the precision loss.
18 /// Like `Default`, but without the warning.
20 /// We error on `from_exposed_addr`, ensuring no precision loss.
24 pub type GlobalState = RefCell<GlobalStateInner>;
26 #[derive(Clone, Debug)]
27 pub struct GlobalStateInner {
28 /// This is used as a map between the address of each allocation and its `AllocId`.
29 /// It is always sorted
30 int_to_ptr_map: Vec<(u64, AllocId)>,
31 /// The base address for each allocation. We cannot put that into
32 /// `AllocExtra` because function pointers also have a base address, and
33 /// they do not have an `AllocExtra`.
34 /// This is the inverse of `int_to_ptr_map`.
35 base_addr: FxHashMap<AllocId, u64>,
36 /// Whether an allocation has been exposed or not. This cannot be put
37 /// into `AllocExtra` for the same reason as `base_addr`.
38 exposed: FxHashSet<AllocId>,
39 /// This is used as a memory address when a new pointer is casted to an integer. It
40 /// is always larger than any address that was previously made part of a block.
42 /// The provenance to use for int2ptr casts
43 provenance_mode: ProvenanceMode,
46 impl GlobalStateInner {
47 pub fn new(config: &MiriConfig) -> Self {
49 int_to_ptr_map: Vec::default(),
50 base_addr: FxHashMap::default(),
51 exposed: FxHashSet::default(),
52 next_base_addr: STACK_ADDR,
53 provenance_mode: config.provenance_mode,
58 impl<'mir, 'tcx> GlobalStateInner {
59 // Returns the exposed `AllocId` that corresponds to the specified addr,
60 // or `None` if the addr is out of bounds
61 fn alloc_id_from_addr(ecx: &MiriEvalContext<'mir, 'tcx>, addr: u64) -> Option<AllocId> {
62 let global_state = ecx.machine.intptrcast.borrow();
63 assert!(global_state.provenance_mode != ProvenanceMode::Strict);
65 let pos = global_state.int_to_ptr_map.binary_search_by_key(&addr, |(addr, _)| *addr);
67 // Determine the in-bounds provenance for this pointer.
68 // (This is only called on an actual access, so in-bounds is the only possible kind of provenance.)
69 let alloc_id = match pos {
70 Ok(pos) => Some(global_state.int_to_ptr_map[pos].1),
73 // This is the largest of the adresses smaller than `int`,
74 // i.e. the greatest lower bound (glb)
75 let (glb, alloc_id) = global_state.int_to_ptr_map[pos - 1];
76 // This never overflows because `addr >= glb`
77 let offset = addr - glb;
78 // If the offset exceeds the size of the allocation, don't use this `alloc_id`.
79 let size = ecx.get_alloc_info(alloc_id).0;
80 if offset <= size.bytes() { Some(alloc_id) } else { None }
84 // We only use this provenance if it has been exposed, *and* is still live.
85 if global_state.exposed.contains(&alloc_id) {
86 let (_size, _align, kind) = ecx.get_alloc_info(alloc_id);
88 AllocKind::LiveData | AllocKind::Function => return Some(alloc_id),
96 pub fn expose_ptr(ecx: &mut MiriEvalContext<'mir, 'tcx>, alloc_id: AllocId, sb: SbTag) {
97 let global_state = ecx.machine.intptrcast.get_mut();
98 // In strict mode, we don't need this, so we can save some cycles by not tracking it.
99 if global_state.provenance_mode != ProvenanceMode::Strict {
100 trace!("Exposing allocation id {alloc_id:?}");
101 global_state.exposed.insert(alloc_id);
102 if ecx.machine.stacked_borrows.is_some() {
103 ecx.expose_tag(alloc_id, sb);
108 pub fn ptr_from_addr_transmute(
109 ecx: &MiriEvalContext<'mir, 'tcx>,
111 ) -> Pointer<Option<Tag>> {
112 trace!("Transmuting 0x{:x} to a pointer", addr);
114 let provenance = if ecx.machine.allow_ptr_int_transmute {
115 // When we allow transmutes, treat them like casts: generating a wildcard pointer.
118 // Usually, we consider transmuted pointers to be "invalid" (`None` provenance).
121 Pointer::new(provenance, Size::from_bytes(addr))
124 pub fn ptr_from_addr_cast(
125 ecx: &MiriEvalContext<'mir, 'tcx>,
127 ) -> InterpResult<'tcx, Pointer<Option<Tag>>> {
128 trace!("Casting 0x{:x} to a pointer", addr);
130 let global_state = ecx.machine.intptrcast.borrow();
132 match global_state.provenance_mode {
133 ProvenanceMode::Default => {
134 // The first time this happens at a particular location, print a warning.
136 // `Span` is non-`Send`, so we use a thread-local instead.
137 static PAST_WARNINGS: RefCell<FxHashSet<Span>> = RefCell::default();
139 PAST_WARNINGS.with_borrow_mut(|past_warnings| {
140 let first = past_warnings.is_empty();
141 if past_warnings.insert(ecx.cur_span()) {
142 // Newly inserted, so first time we see this span.
143 register_diagnostic(NonHaltingDiagnostic::Int2Ptr { details: first });
147 ProvenanceMode::Strict => {
149 "integer-to-pointer casts and `from_exposed_addr` are not supported with `-Zmiri-strict-provenance`; use `with_addr` instead"
152 ProvenanceMode::Permissive => {}
155 // This is how wildcard pointers are born.
156 Ok(Pointer::new(Some(Tag::Wildcard), Size::from_bytes(addr)))
159 fn alloc_base_addr(ecx: &MiriEvalContext<'mir, 'tcx>, alloc_id: AllocId) -> u64 {
160 let mut global_state = ecx.machine.intptrcast.borrow_mut();
161 let global_state = &mut *global_state;
163 match global_state.base_addr.entry(alloc_id) {
164 Entry::Occupied(entry) => *entry.get(),
165 Entry::Vacant(entry) => {
166 // There is nothing wrong with a raw pointer being cast to an integer only after
167 // it became dangling. Hence we allow dead allocations.
168 let (size, align, _kind) = ecx.get_alloc_info(alloc_id);
170 // This allocation does not have a base address yet, pick one.
171 // Leave some space to the previous allocation, to give it some chance to be less aligned.
173 let mut rng = ecx.machine.rng.borrow_mut();
174 // This means that `(global_state.next_base_addr + slack) % 16` is uniformly distributed.
177 // From next_base_addr + slack, round up to adjust for alignment.
178 let base_addr = global_state.next_base_addr.checked_add(slack).unwrap();
179 let base_addr = Self::align_addr(base_addr, align.bytes());
180 entry.insert(base_addr);
182 "Assigning base address {:#x} to allocation {:?} (size: {}, align: {}, slack: {})",
190 // Remember next base address. Leave a gap of at least 1 to avoid two zero-sized allocations
191 // having the same base address, and to avoid ambiguous provenance for the address between two
192 // allocations (also see https://github.com/rust-lang/unsafe-code-guidelines/issues/313).
193 let size_plus_1 = size.bytes().checked_add(1).unwrap();
194 global_state.next_base_addr = base_addr.checked_add(size_plus_1).unwrap();
195 // Given that `next_base_addr` increases in each allocation, pushing the
196 // corresponding tuple keeps `int_to_ptr_map` sorted
197 global_state.int_to_ptr_map.push((base_addr, alloc_id));
204 /// Convert a relative (tcx) pointer to an absolute address.
205 pub fn rel_ptr_to_addr(ecx: &MiriEvalContext<'mir, 'tcx>, ptr: Pointer<AllocId>) -> u64 {
206 let (alloc_id, offset) = ptr.into_parts(); // offset is relative (AllocId provenance)
207 let base_addr = GlobalStateInner::alloc_base_addr(ecx, alloc_id);
209 // Add offset with the right kind of pointer-overflowing arithmetic.
210 let dl = ecx.data_layout();
211 dl.overflowing_offset(base_addr, offset.bytes()).0
214 /// When a pointer is used for a memory access, this computes where in which allocation the
216 pub fn abs_ptr_to_rel(
217 ecx: &MiriEvalContext<'mir, 'tcx>,
219 ) -> Option<(AllocId, Size)> {
220 let (tag, addr) = ptr.into_parts(); // addr is absolute (Tag provenance)
222 let alloc_id = if let Tag::Concrete { alloc_id, .. } = tag {
225 // A wildcard pointer.
226 GlobalStateInner::alloc_id_from_addr(ecx, addr.bytes())?
229 let base_addr = GlobalStateInner::alloc_base_addr(ecx, alloc_id);
231 // Wrapping "addr - base_addr"
232 let dl = ecx.data_layout();
233 let neg_base_addr = (base_addr as i64).wrapping_neg();
236 Size::from_bytes(dl.overflowing_signed_offset(addr.bytes(), neg_base_addr).0),
240 /// Shifts `addr` to make it aligned with `align` by rounding `addr` to the smallest multiple
241 /// of `align` that is larger or equal to `addr`
242 fn align_addr(addr: u64, align: u64) -> u64 {
245 rem => addr.checked_add(align).unwrap() - rem,
255 fn test_align_addr() {
256 assert_eq!(GlobalStateInner::align_addr(37, 4), 40);
257 assert_eq!(GlobalStateInner::align_addr(44, 4), 44);