[pointer] Document validity safety invariant

Document the semantics of `Ptr` validity invariants on the `Validity`
trait.

Makes progress on #2226, #1866

gherrit-pr-id: Ie66db9044be1dc310a6b7280a73652a357878376

src/impls.rs

@@ -101,19 +101,11 @@ safety_comment! {
     ///   - Given `t: *mut bool` and `let r = *mut u8`, `r` refers to an object
     ///     of the same size as that referred to by `t`. This is true because
     ///     `bool` and `u8` have the same size (1 byte) [1]. Neither `r` nor `t`
-    ///     contain `UnsafeCell`s because neither `bool` nor `u8` do [4].
-    ///   - Since the closure takes a `&u8` argument, given a `Maybe<'a,
-    ///     bool>` which satisfies the preconditions of
-    ///     `TryFromBytes::<bool>::is_bit_valid`, it must be guaranteed that the
-    ///     memory referenced by that `MaybeValid` always contains a valid `u8`.
-    ///     Since `bool`'s single byte is always initialized, `is_bit_valid`'s
-    ///     precondition requires that the same is true of its argument. Since
-    ///     `u8`'s only bit validity invariant is that its single byte must be
-    ///     initialized, this memory is guaranteed to contain a valid `u8`.
+    ///     contain `UnsafeCell`s because neither `bool` nor `u8` do [3].
     ///   - The impl must only return `true` for its argument if the original
-    ///     `Maybe<bool>` refers to a valid `bool`. We only return true if
-    ///     the `u8` value is 0 or 1, and both of these are valid values for
-    ///     `bool`. [3]
+    ///     `Maybe<bool>` refers to a valid `bool`. We only return true if the
+    ///     `u8` value is 0 or 1, and both of these are valid values for `bool`.
+    ///     [2]
     ///
     /// [1] Per https://doc.rust-lang.org/1.81.0/reference/type-layout.html#primitive-data-layout:
     ///
@@ -124,16 +116,12 @@ safety_comment! {
     ///   | `bool`    | 1                    |
     ///   | `u8`/`i8` | 1                    |
     ///
-    /// [2] Per https://doc.rust-lang.org/1.81.0/reference/type-layout.html#size-and-alignment:
-    ///
-    ///   The size of a value is always a multiple of its alignment.
-    ///
-    /// [3] Per https://doc.rust-lang.org/1.81.0/reference/types/boolean.html:
+    /// [2] Per https://doc.rust-lang.org/1.81.0/reference/types/boolean.html:
     ///
     ///   The value false has the bit pattern 0x00 and the value true has the
     ///   bit pattern 0x01.
     ///
-    /// [4] TODO(#429): Justify this claim.
+    /// [3] TODO(#429): Justify this claim.
     unsafe_impl!(bool: TryFromBytes; |byte: MaybeAligned<u8>| *byte.unaligned_as_ref() < 2);
 }
 safety_comment! {
@@ -155,20 +143,10 @@ safety_comment! {
     ///   - Given `t: *mut char` and `let r = *mut u32`, `r` refers to an object
     ///     of the same size as that referred to by `t`. This is true because
     ///     `char` and `u32` have the same size [1]. Neither `r` nor `t` contain
-    ///     `UnsafeCell`s because neither `char` nor `u32` do [4].
-    ///   - Since the closure takes a `&u32` argument, given a `Maybe<'a,
-    ///     char>` which satisfies the preconditions of
-    ///     `TryFromBytes::<char>::is_bit_valid`, it must be guaranteed that the
-    ///     memory referenced by that `MaybeValid` always contains a valid
-    ///     `u32`. Since `char`'s bytes are always initialized [2],
-    ///     `is_bit_valid`'s precondition requires that the same is true of its
-    ///     argument. Since `u32`'s only bit validity invariant is that its
-    ///     bytes must be initialized, this memory is guaranteed to contain a
-    ///     valid `u32`.
+    ///     `UnsafeCell`s because neither `char` nor `u32` do [3].
     ///   - The impl must only return `true` for its argument if the original
-    ///     `Maybe<char>` refers to a valid `char`. `char::from_u32`
-    ///     guarantees that it returns `None` if its input is not a valid
-    ///     `char`. [3]
+    ///     `Maybe<char>` refers to a valid `char`. `char::from_u32` guarantees
+    ///     that it returns `None` if its input is not a valid `char`. [2]
     ///
     /// [1] Per https://doc.rust-lang.org/nightly/reference/types/textual.html#layout-and-bit-validity:
     ///
@@ -177,14 +155,10 @@ safety_comment! {
     ///
     /// [2] Per https://doc.rust-lang.org/core/primitive.char.html#method.from_u32:
     ///
-    ///   Every byte of a `char` is guaranteed to be initialized.
-    ///
-    /// [3] Per https://doc.rust-lang.org/core/primitive.char.html#method.from_u32:
-    ///
     ///   `from_u32()` will return `None` if the input is not a valid value for
     ///   a `char`.
     ///
-    /// [4] TODO(#429): Justify this claim.
+    /// [3] TODO(#429): Justify this claim.
     unsafe_impl!(char: TryFromBytes; |candidate: MaybeAligned<u32>| {
         let candidate = candidate.read_unaligned::<BecauseImmutable>();
         char::from_u32(candidate).is_some()
@@ -215,19 +189,9 @@ safety_comment! {
     ///     `str` and `[u8]` have the same representation. [1] Neither `t` nor
     ///     `r` contain `UnsafeCell`s because `[u8]` doesn't, and both `t` and
     ///     `r` have that representation.
-    ///   - Since the closure takes a `&[u8]` argument, given a `Maybe<'a,
-    ///     str>` which satisfies the preconditions of
-    ///     `TryFromBytes::<str>::is_bit_valid`, it must be guaranteed that the
-    ///     memory referenced by that `MaybeValid` always contains a valid
-    ///     `[u8]`. Since `str`'s bytes are always initialized [1],
-    ///     `is_bit_valid`'s precondition requires that the same is true of its
-    ///     argument. Since `[u8]`'s only bit validity invariant is that its
-    ///     bytes must be initialized, this memory is guaranteed to contain a
-    ///     valid `[u8]`.
     ///   - The impl must only return `true` for its argument if the original
-    ///     `Maybe<str>` refers to a valid `str`. `str::from_utf8`
-    ///     guarantees that it returns `Err` if its input is not a valid `str`.
-    ///     [2]
+    ///     `Maybe<str>` refers to a valid `str`. `str::from_utf8` guarantees
+    ///     that it returns `Err` if its input is not a valid `str`. [2]
     ///
     /// [1] Per https://doc.rust-lang.org/1.81.0/reference/types/textual.html:
     ///
@@ -296,18 +260,9 @@ safety_comment! {
     ///     because `NonZeroXxx` and `xxx` have the same size. [1] Neither `r`
     ///     nor `t` refer to any `UnsafeCell`s because neither `NonZeroXxx` [2]
     ///     nor `xxx` do.
-    ///   - Since the closure takes a `&xxx` argument, given a `Maybe<'a,
-    ///     NonZeroXxx>` which satisfies the preconditions of
-    ///     `TryFromBytes::<NonZeroXxx>::is_bit_valid`, it must be guaranteed
-    ///     that the memory referenced by that `MabyeValid` always contains a
-    ///     valid `xxx`. Since `NonZeroXxx`'s bytes are always initialized [1],
-    ///     `is_bit_valid`'s precondition requires that the same is true of its
-    ///     argument. Since `xxx`'s only bit validity invariant is that its
-    ///     bytes must be initialized, this memory is guaranteed to contain a
-    ///     valid `xxx`.
     ///   - The impl must only return `true` for its argument if the original
-    ///     `Maybe<NonZeroXxx>` refers to a valid `NonZeroXxx`. The only
-    ///     `xxx` which is not also a valid `NonZeroXxx` is 0. [1]
+    ///     `Maybe<NonZeroXxx>` refers to a valid `NonZeroXxx`. The only `xxx`
+    ///     which is not also a valid `NonZeroXxx` is 0. [1]
     ///
     /// [1] Per https://doc.rust-lang.org/1.81.0/core/num/type.NonZeroU16.html:
     ///

src/lib.rs

@@ -2845,7 +2845,12 @@ unsafe fn try_read_from<S, T: TryFromBytes>(
     let c_ptr = Ptr::from_mut(&mut candidate);
     let c_ptr = c_ptr.transparent_wrapper_into_inner();
     // SAFETY: `c_ptr` has no uninitialized sub-ranges because it derived from
-    // `candidate`, which the caller promises is entirely initialized.
+    // `candidate`, which the caller promises is entirely initialized. Since
+    // `candidate` is a `MaybeUninit`, it has no validity requirements, and so
+    // no values written to `c_ptr` can violate its validity. Since `c_ptr` has
+    // `Exclusive` aliasing, no mutations may happen except via `c_ptr` so long
+    // as it is live, so we don't need to worry about the fact that `c_ptr` may
+    // have more restricted validity than `candidate`.
     let c_ptr = unsafe { c_ptr.assume_validity::<invariant::Initialized>() };
 
     // This call may panic. If that happens, it doesn't cause any soundness
@@ -4603,7 +4608,12 @@ pub unsafe trait FromBytes: FromZeros {
 
         let ptr = Ptr::from_mut(&mut buf);
         // SAFETY: After `buf.zero()`, `buf` consists entirely of initialized,
-        // zeroed bytes.
+        // zeroed bytes. Since `MaybeUninit` has no validity requirements, `ptr`
+        // cannot be used to write values which will violate `buf`'s bit
+        // validity. Since `ptr` has `Exclusive` aliasing, nothing other than
+        // `ptr` may be used to mutate `ptr`'s referent, and so its bit validity
+        // cannot be violated even though `buf` may have more permissive bit
+        // validity than `ptr`.
         let ptr = unsafe { ptr.assume_validity::<invariant::Initialized>() };
         let ptr = ptr.as_bytes::<BecauseExclusive>();
         src.read_exact(ptr.as_mut())?;

src/pointer/invariant.rs

@@ -42,7 +42,48 @@ pub trait Aliasing: Sealed {
 pub trait Alignment: Sealed {}
 
 /// The validity invariant of a [`Ptr`][super::Ptr].
-pub trait Validity: Sealed {}
+///
+/// # Safety
+///
+/// In this section, we will use `Ptr<T, V>` as a shorthand for `Ptr<T, I:
+/// Invariants<Validity = V>>` for brevity.
+///
+/// Each `V: Validity` defines a set of bit values which may appear in the
+/// referent of a `Ptr<T, V>`, denoted `S(T, V)`. Each `V: Validity`, in its
+/// documentation, provides a definition of `S(T, V)` which must be valid for
+/// all `T: ?Sized`. Any `V: Validity` must guarantee that this set is only a
+/// function of the *bit validity* of the referent type, `T`, and not of any
+/// other property of `T`. As a consequence, given `V: Validity`, `T`, and `U`
+/// where `T` and `U` have the same bit validity, `S(V, T) = S(V, U)`.
+///
+/// It is guaranteed that the referent of any `ptr: Ptr<T, V>` is a member of
+/// `S(T, V)`. Unsafe code must ensure that this guarantee will be upheld for
+/// any existing `Ptr`s or any `Ptr`s that that code creates.
+///
+/// An important implication of this guarantee is that it restricts what
+/// transmutes are sound, where "transmute" is used in this context to refer to
+/// changing the referent type or validity invariant of a `Ptr`, as either
+/// change may change the set of bit values permitted to appear in the referent.
+/// In particular, the following are necessary (but not sufficient) conditions
+/// in order for a transmute from `src: Ptr<T, V>` to `dst: Ptr<U, W>` to be
+/// sound:
+/// - If `S(T, V) = S(U, W)`, then no restrictions apply; otherwise,
+/// - If `dst` permits mutation of its referent (e.g. via `Exclusive` aliasing
+///   or interior mutation under `Shared` aliasing), then it must hold that
+///   `S(T, V) ⊇ S(U, W)` - in other words, the transmute must not expand the
+///   set of allowed referent bit patterns. A violation of this requirement
+///   would permit using `dst` to write `x` where `x ∈ S(U, W)` but `x ∉ S(T,
+///   V)`, which would violate the guarantee that `src`'s referent may only
+///   contain values in `S(T, V)`.
+/// - If the referent may be mutated without going through `dst` while `dst` is
+///   live (e.g. via interior mutation on a `Shared`-aliased `Ptr` or `&`
+///   reference), then it must hold that `S(T, V) ⊆ S(U, W)` - in other words,
+///   the transmute must not shrink the set of allowed referent bit patterns. A
+///   violation of this requirement would permit using `src` or another
+///   mechanism (e.g. a `&` reference used to derive `src`) to write `x` where
+///   `x ∈ S(T, V)` but `x ∉ S(U, W)`, which would violate the guarantee that
+///   `dst`'s referent may only contain values in `S(U, W)`.
+pub unsafe trait Validity: Sealed {}
 
 /// An [`Aliasing`] invariant which is either [`Shared`] or [`Exclusive`].
 ///
@@ -90,9 +131,14 @@ impl Alignment for Aligned {}
 /// Any bit pattern is allowed in the `Ptr`'s referent, including uninitialized
 /// bytes.
 pub enum Uninit {}
-impl Validity for Uninit {}
-
-/// The byte ranges initialized in `T` are also initialized in the referent.
+// SAFETY: `Uninit`'s validity is well-defined for all `T: ?Sized`, and is not a
+// function of any property of `T` other than its bit validity (in fact, it's
+// not even a property of `T`'s bit validity, but this is more than we are
+// required to uphold).
+unsafe impl Validity for Uninit {}
+
+/// The byte ranges initialized in `T` are also initialized in the referent of a
+/// `Ptr<T>`.
 ///
 /// Formally: uninitialized bytes may only be present in `Ptr<T>`'s referent
 /// where they are guaranteed to be present in `T`. This is a dynamic property:
@@ -119,16 +165,24 @@ impl Validity for Uninit {}
 ///   enum type, in which case the same rules apply depending on the state of
 ///   its discriminant, and so on recursively).
 pub enum AsInitialized {}
-impl Validity for AsInitialized {}
+// SAFETY: `AsInitialized`'s validity is well-defined for all `T: ?Sized`, and
+// is not a function of any property of `T` other than its bit validity.
+unsafe impl Validity for AsInitialized {}
 
 /// The byte ranges in the referent are fully initialized. In other words, if
 /// the referent is `N` bytes long, then it contains a bit-valid `[u8; N]`.
 pub enum Initialized {}
-impl Validity for Initialized {}
+// SAFETY: `Initialized`'s validity is well-defined for all `T: ?Sized`, and is
+// not a function of any property of `T` other than its bit validity (in fact,
+// it's not even a property of `T`'s bit validity, but this is more than we are
+// required to uphold).
+unsafe impl Validity for Initialized {}
 
-/// The referent is bit-valid for `T`.
+/// The referent of a `Ptr<T>` is bit-valid for `T`.
 pub enum Valid {}
-impl Validity for Valid {}
+// SAFETY: `Valid`'s validity is well-defined for all `T: ?Sized`, and is not a
+// function of any property of `T` other than its bit validity.
+unsafe impl Validity for Valid {}
 
 /// # Safety
 ///

src/pointer/mod.rs

@@ -50,7 +50,7 @@ where
         T: Copy,
         T: invariant::Read<Aliasing, R>,
     {
-        // SAFETY: By invariant on `MaybeAligned`, `raw` contains
+        // SAFETY: By invariant on `MaybeAligned`, `self` contains
         // validly-initialized data for `T`. By `T: Read<Aliasing>`, we are
         // permitted to perform a read of `self`'s referent.
         unsafe { self.as_inner().read_unaligned() }