docs: revise restrictions on volatile operations

A distinction between usage on Rust memory vs. non-Rust memory was
introduced. Documentation was reworded to explain what that means, and
make explicit that:

- No trapping can occur from volatile operations;
- On Rust memory, all safety rules must be respected;
- On Rust memory, the primary difference from regular access is that
  volatile always involves a memory dereference;
- On Rust memory, the only data affected by an operation is the one
  pointed to in the argument(s) of the function;
- On Rust memory, provenance follows the same rules as non-volatile
  access;
- On non-Rust memory, any address known to not contain Rust memory is
  valid (including 0 and usize::MAX);
- On non-Rust memory, no Rust memory may be affected (it is implicit
  that any other non-Rust memory may be affected, though, even if not
  referenced by the pointer). This should be relevant when, for example,
  reading register A causes a flag to change in register B, or writing
  to A causes B to change in some way. Everything affected mustn't be
  inside an allocation.
- On non-Rust memory, provenance is irrelevant and a pointer with none
  can be used in a valid way.

Author: LuigiPiucco

library/core/src/ptr/mod.rs

@@ -28,7 +28,8 @@
 //!   undefined behavior to perform two concurrent accesses to the same location from different
 //!   threads unless both accesses only read from memory. Notice that this explicitly
 //!   includes [`read_volatile`] and [`write_volatile`]: Volatile accesses cannot
-//!   be used for inter-thread synchronization.
+//!   be used for inter-thread synchronization, regardless of whether it is acting on
+//!   Rust memory or not.
 //! * The result of casting a reference to a pointer is valid for as long as the
 //!   underlying allocation is live and no reference (just raw pointers) is used to
 //!   access the same memory. That is, reference and pointer accesses cannot be
@@ -114,6 +115,10 @@
 //! fully contiguous (i.e., has no "holes"), there is no guarantee that this
 //! will not change in the future.
 //!
+//! Allocations must behave like "normal" memory: in particular, reads must not have
+//! side-effects, and writes must become visible to other threads using the usual synchronization
+//! primitives.
+//!
 //! For any allocation with `base` address, `size`, and a set of
 //! `addresses`, the following are guaranteed:
 //! - For all addresses `a` in `addresses`, `a` is in the range `base .. (base +
@@ -2021,54 +2026,68 @@ pub const unsafe fn write_unaligned<T>(dst: *mut T, src: T) {
     }
 }
 
-/// Performs a volatile read of the value from `src` without moving it. This
-/// leaves the memory in `src` unchanged.
-///
-/// Volatile operations are intended to act on I/O memory, and are guaranteed
-/// to not be elided or reordered by the compiler across other volatile
-/// operations.
-///
-/// # Notes
-///
-/// Rust does not currently have a rigorously and formally defined memory model,
-/// so the precise semantics of what "volatile" means here is subject to change
-/// over time. That being said, the semantics will almost always end up pretty
-/// similar to [C11's definition of volatile][c11].
-///
-/// The compiler shouldn't change the relative order or number of volatile
-/// memory operations. However, volatile memory operations on zero-sized types
-/// (e.g., if a zero-sized type is passed to `read_volatile`) are noops
-/// and may be ignored.
+/// Performs a volatile read of the value from `src` without moving it.
+///
+/// Rust does not currently have a rigorously and formally defined memory model, so the precise
+/// semantics of what "volatile" means here is subject to change over time. That being said, the
+/// semantics will almost always end up pretty similar to [C11's definition of volatile][c11].
+///
+/// Volatile operations are intended to act on I/O memory. As such, they are considered externally
+/// observable events (just like syscalls), and are guaranteed to not be elided or reordered by the
+/// compiler across other externally observable events. With this in mind, there are two cases of
+/// usage that need to be distinguished:
+///
+/// - When a volatile operation is used for memory inside an [allocation], it behaves exactly like
+///   [`read`], except for the additional guarantee that it won't be elided or reordered (see
+///   above). This implies that the operation will actually access memory and not e.g. be lowered to
+///   reusing data from a previous read, such as a register that performed a previous load on that
+///   memory. Other than that, all the usual rules for memory accesses apply (including provenance).
+///   In particular, just like in C, whether an operation is volatile has no bearing whatsoever on
+///   questions involving concurrent access from multiple threads. Volatile accesses behave exactly
+///   like non-atomic accesses in that regard.
+///
+/// - Volatile operations, however, may also be used to access memory that is _outside_ of any Rust
+///   allocation. In this use-case, the pointer does *not* have to be [valid] for reads. This is
+///   typically used for CPU and peripheral registers that must be accessed via an I/O memory
+///   mapping, most commonly at fixed addresses reserved by the hardware. These often have special
+///   semantics associated to their manipulation, and cannot be used as general purpose memory.
+///   Here, any address value is possible, including 0 and [`usize::MAX`], so long as the semantics
+///   of such a read are well-defined by the target hardware. The provenance of the pointer is
+///   irrelevant, and it can be created with [`without_provenance`]. The access must not trap. It
+///   can cause side-effects, but those must not affect Rust-allocated memory in in any way. This
+///   access is still not considered [atomic], and as such it cannot be used for inter-thread
+///   synchronization.
+///
+/// Note that volatile memory operations on zero-sized types (e.g., if a zero-sized type is passed
+/// to `read_volatile`) are noops and may be ignored.
 ///
 /// [c11]: http://www.open-std.org/jtc1/sc22/wg14/www/docs/n1570.pdf
+/// [allocation]: crate::ptr#allocated-object
+/// [atomic]: crate::sync::atomic#memory-model-for-atomic-accesses
 ///
 /// # Safety
 ///
+/// Like [`read`], `read_volatile` creates a bitwise copy of `T`, regardless of whether `T` is
+/// [`Copy`]. If `T` is not [`Copy`], using both the returned value and the value at `*src` can
+/// [violate memory safety][read-ownership]. However, storing non-[`Copy`] types in volatile memory
+/// is almost certainly incorrect.
+///
 /// Behavior is undefined if any of the following conditions are violated:
 ///
-/// * `src` must be [valid] for reads.
+/// * `src` must be either [valid] for reads, or it must point to memory outside of all Rust
+///   allocations and reading from that memory must:
+///   - not trap, and
+///   - not cause any memory inside a Rust allocation to be modified.
 ///
 /// * `src` must be properly aligned.
 ///
-/// * `src` must point to a properly initialized value of type `T`.
-///
-/// Like [`read`], `read_volatile` creates a bitwise copy of `T`, regardless of
-/// whether `T` is [`Copy`]. If `T` is not [`Copy`], using both the returned
-/// value and the value at `*src` can [violate memory safety][read-ownership].
-/// However, storing non-[`Copy`] types in volatile memory is almost certainly
-/// incorrect.
+/// * Reading from `src` must produce a properly initialized value of type `T`.
 ///
 /// Note that even if `T` has size `0`, the pointer must be properly aligned.
 ///
 /// [valid]: self#safety
 /// [read-ownership]: read#ownership-of-the-returned-value
 ///
-/// Just like in C, whether an operation is volatile has no bearing whatsoever
-/// on questions involving concurrent access from multiple threads. Volatile
-/// accesses behave exactly like non-atomic accesses in that regard. In particular,
-/// a race between a `read_volatile` and any write operation to the same location
-/// is undefined behavior.
-///
 /// # Examples
 ///
 /// Basic usage:
@@ -2100,52 +2119,67 @@ pub unsafe fn read_volatile<T>(src: *const T) -> T {
     }
 }
 
-/// Performs a volatile write of a memory location with the given value without
-/// reading or dropping the old value.
-///
-/// Volatile operations are intended to act on I/O memory, and are guaranteed
-/// to not be elided or reordered by the compiler across other volatile
-/// operations.
-///
-/// `write_volatile` does not drop the contents of `dst`. This is safe, but it
-/// could leak allocations or resources, so care should be taken not to overwrite
-/// an object that should be dropped.
-///
-/// Additionally, it does not drop `src`. Semantically, `src` is moved into the
-/// location pointed to by `dst`.
-///
-/// # Notes
-///
-/// Rust does not currently have a rigorously and formally defined memory model,
-/// so the precise semantics of what "volatile" means here is subject to change
-/// over time. That being said, the semantics will almost always end up pretty
-/// similar to [C11's definition of volatile][c11].
-///
-/// The compiler shouldn't change the relative order or number of volatile
-/// memory operations. However, volatile memory operations on zero-sized types
-/// (e.g., if a zero-sized type is passed to `write_volatile`) are noops
-/// and may be ignored.
+/// Performs a volatile write of a memory location with the given value without reading or dropping
+/// the old value.
+///
+/// Rust does not currently have a rigorously and formally defined memory model, so the precise
+/// semantics of what "volatile" means here is subject to change over time. That being said, the
+/// semantics will almost always end up pretty similar to [C11's definition of volatile][c11].
+///
+/// Volatile operations are intended to act on I/O memory. As such, they are considered externally
+/// observable events (just like syscalls), and are guaranteed to not be elided or reordered by the
+/// compiler across other externally observable events. With this in mind, there are two cases of
+/// usage that need to be distinguished:
+///
+/// - When a volatile operation is used for memory inside an [allocation], it behaves exactly like
+///   [`write()`], except for the additional guarantee that it won't be elided or reordered (see
+///   above). This implies that the operation will actually access memory and not e.g. be lowered to
+///   a register access or stack pop. Other than that, all the usual rules for memory accesses apply
+///   (including provenance). In particular, just like in C, whether an operation is volatile has no
+///   bearing whatsoever on questions involving concurrent access from multiple threads. Volatile
+///   accesses behave exactly like non-atomic accesses in that regard.
+///
+/// - Volatile operations, however, may also be used to access memory that is _outside_ of any Rust
+///   allocation. In this use-case, the pointer does *not* have to be [valid] for writes. This is
+///   typically used for CPU and peripheral registers that must be accessed via an I/O memory
+///   mapping, most commonly at fixed addresses reserved by the hardware. These often have special
+///   semantics associated to their manipulation, and cannot be used as general purpose memory.
+///   Here, any address value is possible, including 0 and [`usize::MAX`], so long as the semantics
+///   of such a write are well-defined by the target hardware. The provenance of the pointer is
+///   irrelevant, and it can be created with [`without_provenance`]. The access must not trap. It
+///   can cause side-effects, but those must not affect Rust-allocated memory in any way. This
+///   access is still not considered [atomic], and as such it cannot be used for inter-thread
+///   synchronization.
+///
+/// Note that volatile memory operations on zero-sized types (e.g., if a zero-sized type is passed
+/// to `write_volatile`) are noops and may be ignored.
+///
+/// `write_volatile` does not drop the contents of `dst`. This is safe, but it could leak
+/// allocations or resources, so care should be taken not to overwrite an object that should be
+/// dropped when operating on Rust memory.
+///
+/// Additionally, it does not drop `src`. Semantically, `src` is moved into the location pointed to
+/// by `dst`.
 ///
 /// [c11]: http://www.open-std.org/jtc1/sc22/wg14/www/docs/n1570.pdf
+/// [allocation]: crate::ptr#allocated-object
+/// [atomic]: crate::sync::atomic#memory-model-for-atomic-accesses
 ///
 /// # Safety
 ///
 /// Behavior is undefined if any of the following conditions are violated:
 ///
-/// * `dst` must be [valid] for writes.
+/// * `dst` must be either [valid] for writes, or it must point to memory outside of all Rust
+///   allocations and writing to that memory must:
+///   - not trap, and
+///   - not cause any memory inside a Rust allocation to be modified.
 ///
 /// * `dst` must be properly aligned.
 ///
 /// Note that even if `T` has size `0`, the pointer must be properly aligned.
 ///
 /// [valid]: self#safety
 ///
-/// Just like in C, whether an operation is volatile has no bearing whatsoever
-/// on questions involving concurrent access from multiple threads. Volatile
-/// accesses behave exactly like non-atomic accesses in that regard. In particular,
-/// a race between a `write_volatile` and any other operation (reading or writing)
-/// on the same location is undefined behavior.
-///
 /// # Examples
 ///
 /// Basic usage: