Merge #365

365: spi: allow impls returning early, add flush.  r=therealprof a=Dirbaio

Fixes #264 
Depends on #351 

- Document that all methods may return early, before bus is idle.
- Add a `flush` method to wait for bus to be idle.

The alternative is to document that methods MUST wait for bus idle before returning, but that would be bad for performance IMO. Allowing early return means multiple successive writes can "overlap" so that bytes come out back-to-back without gaps, which is great for workloads with tons of small writes, like SPI displays with `embedded-graphics`.

Drivers using `SpiDevice` won't have to worry about this, it'll Just Work since Device impls must flush before deasserting CS. Drivers using SpiBus will have to care if they coordinate SPI activity with other GPIO activity.

Co-authored-by: Dario Nieuwenhuis <dirbaio@dirbaio.net>

CHANGELOG.md

@@ -11,6 +11,7 @@ and this project adheres to [Semantic Versioning](http://semver.org/).
 - The Minimum Supported Rust Version (MSRV) is now 1.54.0
 - `spi`: unify all traits into `SpiReadBus`, `SpiWriteBus` and `SpiBus` (read-write).
 - `spi`: Add `SpiDevice` trait to represent a single device in a (possibly shared) bus, with managed chip-select (CS) pin.
+- `spi`: Clarify that implementations are allowed to return before operations are finished, add `flush` to wait until finished.
 
 ## [v1.0.0-alpha.7] - 2022-02-09
 

src/spi/blocking.rs

@@ -140,6 +140,32 @@
 //!
 //! HALs **must not** add infrastructure for sharing at the [`SpiBus`] level. User code owning a [`SpiBus`] must have the guarantee
 //! of exclusive access.
+//!
+//! # Flushing
+//!
+//! To improve performance, [`SpiBus`] implementations are allowed to return before the operation is finished, i.e. when the bus is still not
+//! idle.
+//!
+//! When using a [`SpiBus`], call [`flush`](SpiBusFlush::flush) to wait for operations to actually finish. Examples of situations
+//! where this is needed are:
+//! - To synchronize SPI activity and GPIO activity, for example before deasserting a CS pin.
+//! - Before deinitializing the hardware SPI peripheral.
+//!
+//! When using a [`SpiDevice`], you can still call [`flush`](SpiBusFlush::flush) on the bus within a transaction.
+//! It's very rarely needed, because [`transaction`](SpiDevice::transaction) already flushes for you
+//! before deasserting CS. For example, you may need it to synchronize with GPIOs other than CS, such as DCX pins
+//! sometimes found in SPI displays.
+//!
+//! For example, for [`write`](SpiBusWrite::write) operations, it is common for hardware SPI peripherals to have a small
+//! FIFO buffer, usually 1-4 bytes. Software writes data to the FIFO, and the peripheral sends it on MOSI at its own pace,
+//! at the specified SPI frequency. It is allowed for an implementation of [`write`](SpiBusWrite::write) to return as soon
+//! as all the data has been written to the FIFO, before it is actually sent. Calling [`flush`](SpiBusFlush::flush) would
+//! wait until all the bits have actually been sent, the FIFO is empty, and the bus is idle.
+//!
+//! This still applies to other operations such as [`read`](SpiBusRead::read) or [`transfer`](SpiBus::transfer). It is less obvious
+//! why, because these methods can't return before receiving all the read data. However it's still technically possible
+//! for them to return before the bus is idle. For example, assuming SPI mode 0, the last bit is sampled on the first (rising) edge
+//! of SCK, at which point a method could return, but the second (falling) SCK edge still has to happen before the bus is idle.
 
 use core::fmt::Debug;
 
@@ -162,6 +188,7 @@ pub trait SpiDevice: ErrorType {
  /// - Locks the bus
  /// - Asserts the CS (Chip Select) pin.
  /// - Calls `f` with an exclusive reference to the bus, which can then be used to do transfers against the device.
+ /// - [Flushes](SpiBusFlush::flush) the bus.
  /// - Deasserts the CS pin.
  /// - Unlocks the bus.
  ///
@@ -236,12 +263,29 @@ impl<T: SpiDevice> SpiDevice for &mut T {
  }
 }
 
+/// Flush support for SPI bus
+pub trait SpiBusFlush: ErrorType {
+ /// Blocks until all operations have completed and the bus is idle.
+ ///
+ /// See the [module-level documentation](self) for important usage information.
+ fn flush(&mut self) -> Result<(), Self::Error>;
+}
+
+impl<T: SpiBusFlush> SpiBusFlush for &mut T {
+ fn flush(&mut self) -> Result<(), Self::Error> {
+ T::flush(self)
+ }
+}
+
 /// Read-only SPI bus
-pub trait SpiBusRead<Word: Copy = u8>: ErrorType {
+pub trait SpiBusRead<Word: Copy = u8>: SpiBusFlush {
  /// Reads `words` from the slave.
  ///
  /// The word value sent on MOSI during reading is implementation-defined,
  /// typically `0x00`, `0xFF`, or configurable.
+ ///
+ /// Implementations are allowed to return before the operation is
+ /// complete. See the [module-level documentation](self) for detials.
  fn read(&mut self, words: &mut [Word]) -> Result<(), Self::Error>;
 }
 
@@ -252,8 +296,11 @@ impl<T: SpiBusRead<Word>, Word: Copy> SpiBusRead<Word> for &mut T {
 }
 
 /// Write-only SPI bus
-pub trait SpiBusWrite<Word: Copy = u8>: ErrorType {
+pub trait SpiBusWrite<Word: Copy = u8>: SpiBusFlush {
  /// Writes `words` to the slave, ignoring all the incoming words
+ ///
+ /// Implementations are allowed to return before the operation is
+ /// complete. See the [module-level documentation](self) for detials.
  fn write(&mut self, words: &[Word]) -> Result<(), Self::Error>;
 }
 
@@ -277,11 +324,17 @@ pub trait SpiBus<Word: Copy = u8>: SpiBusRead<Word> + SpiBusWrite<Word> {
  /// incoming words after `read` has been filled will be discarded. If `write` is shorter,
  /// the value of words sent in MOSI after all `write` has been sent is implementation-defined,
  /// typically `0x00`, `0xFF`, or configurable.
+ ///
+ /// Implementations are allowed to return before the operation is
+ /// complete. See the [module-level documentation](self) for detials.
  fn transfer(&mut self, read: &mut [Word], write: &[Word]) -> Result<(), Self::Error>;
 
  /// Writes and reads simultaneously. The contents of `words` are
  /// written to the slave, and the received words are stored into the same
  /// `words` buffer, overwriting it.
+ ///
+ /// Implementations are allowed to return before the operation is
+ /// complete. See the [module-level documentation](self) for detials.
  fn transfer_in_place(&mut self, words: &mut [Word]) -> Result<(), Self::Error>;
 }
 
@@ -335,15 +388,15 @@ impl<BUS, CS> ExclusiveDevice<BUS, CS> {
 
 impl<BUS, CS> ErrorType for ExclusiveDevice<BUS, CS>
 where
- BUS: ErrorType,
+ BUS: SpiBusFlush,
  CS: OutputPin,
 {
  type Error = ExclusiveDeviceError<BUS::Error, CS::Error>;
 }
 
 impl<BUS, CS> SpiDevice for ExclusiveDevice<BUS, CS>
 where
- BUS: ErrorType,
+ BUS: SpiBusFlush,
  CS: OutputPin,
 {
  type Bus = BUS;
@@ -356,10 +409,12 @@ where
 
  let f_res = f(&mut self.bus);
 
- // If the closure fails, it's important to still deassert CS.
+ // On failure, it's important to still flush and deassert CS.
+ let flush_res = self.bus.flush();
  let cs_res = self.cs.set_high();
 
  let f_res = f_res.map_err(ExclusiveDeviceError::Spi)?;
+ flush_res.map_err(ExclusiveDeviceError::Spi)?;
  cs_res.map_err(ExclusiveDeviceError::Cs)?;
 
  Ok(f_res)