app

Examples and exercises for the Nucleo STM32F401re/STM32F11re devkits.

Dependencies

• Rust 1.40, or later. Run the following commands to update you Rust tool-chain and add the target for Arm Cortex M4 with hardware floating point support.

> rustup update
> rustup target add thumbv7em-none-eabihf

• For programming (flashing) and debugging

  • openocd debug host, (install using your package manager).

  If you are under Ubuntu or ÒSX you may need to change the openocd.cfg and .vscode/launch.json to use the deprecated (older) script (source [find interface/stlink-v2-1.cfg], and "configFiles": [ "interface/stlink-v2-1.cfg", ... ], respectively).

  • arm-none-eabi tool-chain (install using your package manager). In the following we refer the arm-none-eabi-gdb as just gdb for brevity.

  If using OSX you can download and install the latest official GNU ARM toolchain for OSX. You need to unpack the toolchain to your computer and make sure the bin directory is in your path. Under OSX the path is defined in ~/.bash_profile, add export PATH="<INSTALL>/bin/:$PATH", where <INSTALL> is the path to the GNU tools ($HOME/Prog/ARM/gcc-arm-none-eabi-9-2019-q4-major or similar). (This setting will take effect for all new terminals opened.)

  • stlink (optional) tools for erasing and programming ST microcontrollers (install using your package manager).

• itm tools for ITM trace output, install by:

> cargo install itm

• vscode editor/ide and cortex-debug plugin. Install vscode using your package manager and follow the instructions at cortex-debug (optional for an integrated debugging experience).

• rust-analyzer install following instructions at rust-analyzer (optional for Rust support in vscode). You may first need to install npm (through you package manager).

Examples

Hello World! Building and Debugging an Application

1. Connect your devkit using USB. To check that it is found you can run:

> lsusb
...
Bus 001 Device 004: ID 0483:374b STMicroelectronics ST-LINK/V2.1
...

(Bus/Device/ID may vary.)

2. Run in a terminal (in the app project folder):

> openocd -f openocd.cfg
...
Info : Listening on port 6666 for tcl connections
Info : Listening on port 4444 for telnet connections
Info : clock speed 2000 kHz
Info : STLINK V2J20M4 (API v2) VID:PID 0483:374B
Info : Target voltage: 3.254773
Info : stm32f4x.cpu: hardware has 6 breakpoints, 4 watchpoints
Info : Listening on port 3333 for gdb connections

openocd should connect to your target using the stlink programmer (onboard your Nucleo devkit). See the Trouble Shooting section if you run into trouble.

3. In another terminal (in the same app folder) run:

> cargo run --example hello

The cargo sub-command run looks in the .cargo/config file on the configuration (runner = "arm-none-eabi-gdb -q -x openocd.gdb"), if you use gdb-multiarch, you can change the configuration to runner = "gdb-multiarch -q -x openocd.gdb".

We can also do this manually.

> cargo build --example hello
> arm-none-eabi-gdb target/thumbv7em-none-eabihf/debug/examples/hello -x openocd.gdb

This starts gdb with file being the hello (elf) binary, and runs the openocd.gdb script, which loads (flashes) the binary to the target (our devkit). The script connects to the openocd server, enables semihosting and ITM tracing, sets breakpoints at main (as well as some exception handlers, more on those later), finally it flashes the binary and runs the first instruction (stepi). (You can change the startup behavior in the openocd.gdb script, e.g., to continue instead of stepi.)

4. You can now continue debugging of the program:

...
Note: automatically using hardware breakpoints for read-only addresses.
halted: PC: 0x08000a72
DefaultPreInit ()
    at /home/pln/.cargo/registry/src/github.com-1ecc6299db9ec823/cortex-m-rt-0.6.12/src/lib.rs:571
571     pub unsafe extern "C" fn DefaultPreInit() {}

(gdb) c
Continuing.

Breakpoint 1, main () at examples/hello.rs:12
12      #[entry]

The cortex-m-rt run-time initializes the system and your global variables (in this case there are none). After that it calls the [entry] function. Here you hit a breakpoint.

5. You can continue debugging:

(gdb) c
Continuing.
halted: PC: 0x0800043a
^C
Program received signal SIGINT, Interrupt.
hello::__cortex_m_rt_main () at examples/hello.rs:15
15          loop {

At this point, the openocd terminal should read something like:

 Thread
xPSR: 0x01000000 pc: 0x08000a1a msp: 0x20008000, semihosting
Info : halted: PC: 0x08000a72
Info : halted: PC: 0x0800043a
Hello, world!

Your program is now stuck in an infinite loop (doing nothing).

6. Press CTRL-c in the gdb terminal:

Program received signal SIGINT, Interrupt.
0x08000624 in main () at examples/hello.rs:14
14          loop {}
(gdb)

You have now compiled and debugged a minimal Rust hello example. gdb is a very useful tool so lookup some tutorials/docs (e.g., gdb-doc, and the GDB Cheat Sheet.

ITM Tracing

The hello.rs example uses the semihosting interface to emit the trace information (appearing in the openocd terminal). The drawback is that semihosting is incredibly slow as it involves a lot of machinery to process each character. (Essentially, it writes a character to a given position in memory, runs a dedicated break instruction, openocd detects the break, reads the character at the given position in memory and emits the character to the console.)

A better approach is to use the ARM ITM (Instrumentation Trace Macrocell), designed to more efficiently implement tracing. The onboard stlink programmer can put up to 4 characters into an ITM package, and transmit that to the host (openocd). openocd can process the incoming data and send it to a file or FIFO queue. The ITM package stream needs to be decoded (header + data). To this end we use the itmdump tool.

In a separate terminal, create a named fifo. If you have started openocd without first creating the fifo, openocd would have created a file tmp/itm.fifo for you. In general a fifo is the better option since it is a zero sized "socket" that won't blow up over time, so first remove the tmp/itm.fifo (if existing).

> rm /tmp/itm.fifo
> mkfifo /tmp/itm.fifo
> itmdump -f /tmp/itm.fifo
Hello, again!

Now you can compile and run the itm.rs application using the same steps as the hello program. In the itmdump console you should now have the trace output.

> cargo run --example itm

Under the hood there is much less overhead, the serial transfer rate is set to 2Mbit/s in between the ITM (inside of the MCU) and stlink programmer (onboard the Nucleo devkit). So in theory we can transmit some 200kByte/s data over ITM. However, we are limited by the USB interconnection and openocd to receive and forward packages.

The stlink programmer, buffers packages but has limited buffer space. Hence in practice, you should keep tracing to short messages, else the buffer will overflow. See trouble shooting section if you run into trouble.

Rust panic Handling

The rust compiler statically analyses your code, but in cases some errors cannot be detected at compile time (e.g., array indexing out of bounds, division by zero etc.). The rust compiler generates code checking such faults at run-time, instead of just crashing (or even worse, continuing with faulty/undefined values like a C program would) . A fault in Rust will render a panic, with an associated error message (useful to debugging the application). We can choose how such panics should be treated, e.g., transmitting the error message using semihosting, ITM, some other channel (e.g. a serial port), or simply aborting the program.

The panic example demonstrates some possible use cases.

The openocd.gdb script sets a breakpoint at rust_begin_unwind (a function in the rust core library, used to recover errors.)

When running the example (see above how to compile and run), the gdb terminal will show:

...
Breakpoint 2, main () at examples/panic.rs:27
27          panic!("Oops")
(gdb) c
Continuing.
halted: PC: 0x08000404

Breakpoint 1, rust_begin_unwind (_info=0x20017fb4)
    at /home/pln/.cargo/registry/src/github.com-1ecc6299db9ec823/panic-halt-0.2.0/src/lib.rs:33
33              atomic::compiler_fence(Ordering::SeqCst);
(gdb) p *_info
$1 = core::panic::PanicInfo {payload: core::any::&Any {pointer: 0x8000760 <.Lanon.21a036e607595cc96ffa1870690e4414.142> "\017\004\000", vtable: 0x8000760 <.Lanon.21a036e607595cc96ffa1870690e4414.142>}, message: core::option::Option<&core::fmt::Arguments>::Some(0x20017fd0), location: core::panic::Location {file: <error reading variable>, line: 27, col: 5}}

Here p *_info prints the argument to rust_begin_unwind, at the far end you will find line: 27, col 5, which corresponds to the source code calling panic("Ooops"). (gdb is not (yet) Rust aware enough to figure out how the file field should be interpreted, but at least we get some useful information).

Alternatively we can trace the panic message over semihosting (comment out extern crate panic_halt and uncomment extern crate panic_semihosting).

The openocd console should now show:

Info : halted: PC: 0x080011a0
panicked at 'Oops', examples/panic.rs:27:5

Under the hood, this approach involves formatting of the panic message, which implementation occupies a bit of flash memory (in our case we have 512kB so plenty enough, but for the smallest of MCUs this may be a problem). Another drawback is that it requires a debugger to be connected and active.

Another alternative is to use ITM (uncomment extern crate panic_itm), this is faster, but be aware, the message may overflow the ITM buffer, so it may be unreliable. Also it assumes, that the ITM stream is actively monitored.

A third alternative would be to store the panic message in some non-volatile memory (flash, eeprom, etc.). This allows for true post-mortem debugging of a unit put in production. This approach is used e.g. in automotive applications where the workshop can read-out error codes of your vehicle.

Exception Handling and Core Peripheral Access

The ARM Cortex-M processors features a set of core peripherals and exception handlers. These offer basic functionality independent of vendor (NXP, STM, ...). The SysTick peripheral is a 24-bit countdown timer, that raises a SysTick exception when hitting 0 and reloads the set value. Seen as a real-time system, we can dispatch the SysTick task in a periodic fashion (without accumulated drift under some additional constraints).

In the exception.rs example a . is emitted by the SysTick handler using semihosting. Running the example should give you a periodic updated of the openocd console.

The exception_itm.rs and exception_itm_raw.rs uses the ITM instead. The difference is the way they gain access to the ITM peripheral. In the first case we steal the whole set of core peripherals, while the in the second case we use raw pointer access to the ITM. In both cases, the code is unsafe, as there is no guarantee that other tasks may access the peripheral simultaneously (causing a conflict/race). Later we will see how the concurrency problem is solved in RTFM to offer safe access to peripherals.

Crash - Analyzing the Exception Frame

In case the execution of an instruction fails, a HardFault exception is raised by the hardware, and the HardFault handler is executed. We can define our own handler as in example crash.rs. In main we attempt to read an illegal address, causing a HardFault, and we hit a breakpoint (openocd.gdb script sets a breakpoint at the HardFault handler). From there you can print the exception frame, reflecting the state of the MCU when the error occurred. You can use gdb to give a back trace of the call-stack leading up to the error. See the example for detailed information.

Most crash conditions trigger a hard fault exception, whose handler is defined via

#[exception]
fn HardFault(ef: &cortex_m_rt::ExceptionFrame) -> ! {
...

cortex-m-rt generates a trampoline, that calls into your user defined HardFault handler. We can use cargo expand to view the expanded code:

> cargo expand --example crash > crash_expand.rs

In the generated file we find:

#[doc(hidden)]
#[export_name = "HardFault"]
#[link_section = ".HardFault.user"]
pub unsafe extern "C" fn __cortex_m_rt_HardFault_trampoline(frame: &::cortex_m_rt::ExceptionFrame) {
   __cortex_m_rt_HardFault(frame)
}

The HardFault handler has access to the exception frame, a snapshot of the CPU registers at the moment of the exception.

To better see what is happening we make a --release build (It reduces the amount of redundant code.)

> cargo run --example crash --release
...
Breakpoint 2, HardFault (frame=0x20007fe0) at examples/crash.rs:28
28      #[exception]
(gdb) p/x *frame
$1 = cortex_m_rt::ExceptionFrame {r0: 0x2fffffff, r1: 0xf00000, r2: 0x0, r3: 0x0, r12: 0x0, lr: 0x800051f, pc: 0x8000524, xpsr: 0x61000000}
(gdb) disassemble frame.pc
Dump of assembler code for function crash::__cortex_m_rt_main:
   0x08000520 <+0>:     mvn.w   r0, #3489660928 ; 0xd0000000
   0x08000524 <+4>:     ldr     r0, [r0, #0]
   0x08000526 <+6>:     b.n     0x8000526 <crash::__cortex_m_rt_main+6>
End of assembler dump.

The program counter (frame.pc) contains the address of the instruction that caused the exception. In GDB one can disassemble the program around this address to observe the instruction that caused the exception. In our case its the ldr r0, [r0, #0] caused the exception. This instruction tried to load (read) a 32-bit word from the address stored in the register r0. Looking again at the contents of ExceptionFrame we find that r0 contained the address 0x2FFF_FFFF when this instruction was executed.

Looking at the assembly mvn.w r0, #3489660928 ; 0xd0000000. This is a move and not instruction, so the resulting value here is actually 0x2fffffff. Why did it not do it straight up then as 0x2FFF_FFFF?

Well a 32 bit constant cannot be stored in a 32 bit instruction. So under the hood it stores 0xd0, bit shifts it and bit wise inversion. This is the level of optimization Rust + LLVM is capable of.

We can further backtrace the calls leading up to the fault.

(gdb) bt
#0  HardFault (frame=0x20007fe0) at examples/crash.rs:79
#1  <signal handler called>
#2  core::ptr::read_volatile (src=0x2fffffff)
    at /rustc/73528e339aae0f17a15ffa49a8ac608f50c6cf14/src/libcore/ptr/mod.rs:948
#3  crash::__cortex_m_rt_main () at examples/crash.rs:71
#4  0x08000404 in main () at examples/crash.rs:66

Here we see that on frame #2 we are doing the read causing havoc.

We can also use panic!("Exception frame {:?}", ef); to format and print the exception frame, e.g., over semihosting or ITM. In the example we use semihosting, so when continuing debugging you will eventually the exception frame printed in the openocd console. In the openocd.gdb file we set breakpoints to the exception handlers:

# detect unhandled exceptions, hard faults and panics
break DefaultHandler
break HardFault
break rust_begin_unwind

So in case, you want to go directly to a panic! printout of the exception frame comment out the breakpoints.

Notice. panic!("Exception frame {:?}", ef); will bring in the formatting code from the core library (which is kind of large), so in case you are scarce on flash memory, you may want use some other method.

Device Crates and System View Descriptions (SVDs)

Besides the ARM provided core peripherals the STM32F401re/STM32F411re MCUs has numerous vendor specific peripherals (GPIOs, Timers, USARTs etc.). The vendor provides a System View Description (SVD) specifying the register block layouts (fields, enumerated values, etc.). Using the svd2rust tool we can derive a Peripheral Access Crate (PAC) providing an API for the device that allow us to access each register according to the vendors specification. The device.rs example showcase how a PAC for the STM32F401re/STM32F411re MCUs can be added. (These MCUs have the same set of peripherals, only the maximum clock rating differs.)

> cargo run --example device --features stm32f4

The example output a . each second over semihosting and ITM.

The Cargo.toml file

Looking at the Cargo.toml file we find:

...
[dependencies.stm32f4]
version         = "0.9.0"
features        = ["stm32f401", "rt"]
optional        = true

...

# Built options for different examples
[[example]]
name                = "device"
required-features   = ["stm32f4"]
...

We compile stm32f4 (a generic library for all STMF4 MCUs) with features = ["stm32f401", "rt"], which indicates the specific MCU with rt (so we get the interrupt vector etc.). By having the PAC as an optional dependency, we did not need to compile it (unless we need it, and as you might have experienced already compiling the PAC takes a bit of time to compile initially). (An SVD file is typically > 50k lines, amounting to the same (or more) lines of Rust code.)

By compiling with --features stm32f4 we "opt-in" this dependency.

Hardware Abstraction Layer

For convenience common functionality can be implemented for a specific MCU (or family of MCUs). The stm32f4xx-hal is a Work In Progress, implementing a Hardware Abstraction Layer (hal) for the stm32f4 family. It implements the https://crates.io/search?q=embedded-hal serial trait (interface), to read and write single bytes over a serial port. However, setting up communication is out of scope for the embedded-hal.

The serial.rs example showcase a simple echo application, repeating back incoming data over a serial port (byte by byte). You will also get trace output over the ITM.

Looking closer at the example, rcc is a singleton (constrain consumes the RCC and returns a singleton. The freeze consumes the singleton (rcc) and sets the MCU clock tree according to the (default) cfgr. (Later in the exercises you will change this.)

This pattern ensures that the clock configuration will remain unchanged (the freeze function cannot be called again, as the rcc is consumed, also you cannot get a new rcc as the RCC was consumed by constrain).

Why is this important you may ask? Well, this pattern allows the compiler to check and ensure that your code (or some library that you use) does not make changes to the system (in this case the clocking), which reduces the risk of errors and improves robustness.

Similarly, we split the GPIOA into its parts (pins), and select the operating mode to af7 for tx (the transmit pin pa2), and rx (the receive pin pa3). For details see, RM0368, figure 17 (page 151), and table 9 in STM32F401xD STM32F401xE. The GPIO pins pa2 and pa3 are (by default) connected to the stlink programmer, see section 6.8 of the Nucleo64 user manual UM1724. When the stlink programmer is connected to a linux host, the device \dev\ttyACM0 appears as a virtual com port.

Now we can call Serial::usart2 to setup the serial communication, (according to table 9 in STM32F401xD STM32F401xE documentation it is USART2).

Following the singleton pattern it consumes the USART2 peripheral (to ensure that only one configuration can be active at any time). The second parameter is the pair of pins (tx, px) that we setup earlier. The third parameter is the USART configuration. By default it is set to 8 bit data, and one stop bit. We set the baudrate to 115200. We also pass the clocks (holding information about the MCU clock setup).

At this point tx and rx is owned by serial. We can get access to them again by serial.split().

In the loop we match on the result of block!(rx.read()). block! repeatedly calls rx.read() until ether a byte is received or an error returned. In case rx.read() succeeded, we trace the received byte over the ITM, and echo it by tx.write(byte), ignoring the result (we just assume sending will always succeed). In case rx.read returned with an error, we trace the error message.

As the underlying hardware implementation buffers only a single byte, the input buffer may overflow (resulting in tx.read returning an error).

You can now compile and run the example. Start moserial (or some other serial communication tool), connect to /dev/ttyACM0 with 115200 8N1). Now write a single character a in the Outgoing pane, followed by pressing . You should receive back an a from the target. In the ITM trace output you should see:

Ok 97

Depending if was encoded as CR+LF, CR, LF, TAB, ... you will get additional bytes sent (and received). Try sending multiple characters at once, e.g. abcd, you will see that the you well get a buffer overflow.

This is an example of a bad programming pattern, typically leading to serious problems in (real-time) embedded programming (so it takes more than just Rust to get it right). Later in the exercises we will see how better patterns can be adopted.

Real Time For the Masses (RTFM)

RTFM allows for safe concurrency, sharing resources between different tasks running at different priorities. The resource management and scheduling follow the Stack Resource Policy, which gives us outstanding properties of race- and deadlock free scheduling, single blocking, stack sharing etc.

We start by a simple example. For the full documentation see the RTFM book.

RTFM ITM, with Interrupt

Key here is that we share the ITM peripheral in between the init task (that runs first), the exti0 task (that runs preemptively with a default priority of 1), and the background task idle (that runs on priority 0). Since the highest priority task cannot be interrupted we can safely access the shared resource directly (in exti0). In idle however, we need to lock the resource (which returns itm (a reference to the ITM) peripheral to the closure).

By rtfm::pend we can simulate we trigger an interrupt, (more realistically, interrupts are triggered by the environment, e.g., a peripheral has received some data).

> cargo run --example rtfm_itm --features rtfm

For more information see app.

Looking at the Cargo.toml file we find:

...

[dependencies.stm32f4xx-hal]
version         = "0.6.0"
features        = ["stm32f401", "rt"]
optional        = true

[dependencies.cortex-m-rtfm]
version         = "0.5.1"
optional        = true

[features]
rtfm            = ["cortex-m-rtfm", "stm32f4xx-hal"]
...

[[example]]
name                = "rtfm_itm"
required-features   = ["rtfm"]

The rtfm feature opt-in the dependencies to cortex-m-rtfm and stm32f4xx-hal (which in turn opt-in the dependency to stm32f4 under the stm32f401 and rt features). Through the hal we can get access to the underlying device/PAC (peripherals, interrupts etc.).

RTFM ITM, using Spawn

In the previous example we triggered the exti0 task manually. We can let RTFM do that for us using spawn with an optional payload. Thus we have simple way to do message passing.

> cargo run --example rtfm_itm_spawn --features rtfm

The spawn.unwrap() panics if the message could not be delivered (i.e, the queue is full). The size (capacity) of queues are 1 by default, but can be for each task individually, see spawn.

The example shows that the message to task1 is queued while idle is holding the itm resource.

RTFM Schedule

Similarly to spawn, RTFM allows for to schedule messages to be spawned at a specific point in time.

$ cargo run --example rtfm_schedule --features rtfm

For this to work, we have to provide an implementation of a timer for RTFM to use (e.g, monotonic = rtfm::cyccnt::CYCCNT, which is a built in implementation using the ARM core DWT unit). You may opt to provide your own implementation as well. Since the underlying timer hardware may differ between platforms, it's up to you to make sure the timer is initialized. For more information see schedule

RTFM Blinky

No example suit is complete without the mandatory blinky demo, so here we go! The examples rtfm_blinky.rs, rtfm_blinky_msg1.rs, rtfm_blinky_msg2.rs, and rtfm_blinky_msg3.rs showcase different approaches.

• rtfm_blinky.rs stores the GPIOA peripheral as a resource, and uses a state local variable to hold the TOGGLE state. The toggle task is bound the SysTick exception handler, setup in init to fire SysTick with a period of 1s.

The msg examples uses the scheduling primitives provided by RTFM.

• rtfm_blinky_msg1.rs stores the GPIOA peripheral as a resource, and uses a state local variable to hold the TOGGLE state.

• rtfm_blinky_msg2.rs stores the GPIOA peripheral as a resource, but uses the message payload to represent current state.

• rtfm_blinky_msg3.rs uses messages to pass around both current state and the owned peripheral.

For all cases, RTFM ensures memory safety. Which approach to take depends on the use case.

• If your intention/design requires concurrent tasks to access a shared resource (e.g., a peripheral) you need to use the Resources approach.

• If your intention is that a resource should be accessed sequential manner, the message passing of owned resources is the way.

The latter approach actually proves the sequential accesses pattern, so besides its simplicity it gives you a guarantee. This comes in handy e.g. when juggling buffers between owners of memory for DMA operations, handling of secure data etc, as you are in full control over resource ownership at all times.

Regarding the HW access.

• In init we:

  • configure the RCC (enabling the GPIOA peripheral),
  • configure the PA5 pin (connected to the Green LED on the Nucleo) as an output, and finally
  • delegate access to GPIO as a "late" resource.

• In toggle we:

• either define a task local resource TOGGLE to hold the current state, or pass it along as boolean argument.

• either access GPIOA as a resource (provided by the context), or as an owned resources passed as parameter.

• set/clear the PA5 pin correspondingly. (The bs5 field sets the PA5 high, while br5 clears the corresponding bit controlling the led.)

• finally schedule a message to invoke toggle at a later time.

Trouble Shooting

Working with embedded targets involves a lot of tooling, and many things can go wrong.

openocd fails to connect

If you end up with a program that puts the MCU in a bad state.

• Hold the RESET button (black), while starting openocd. If that does not work, disconnect the USB cable, hold the RESET button, re-connect the USB, start openocd then then let go of the button.

• However even a reset might not help you. In that case you can erase the flash memory. st-flash connects to the target directly (bypassing gdb and openocd) and hence more likely to get access to the target even if its in a bad state.

> st-flash erase

• Make sure that the st-link firmware is up to date, you can use the java application: stsw-link007 to check/update the firmware.

gdb fails to connect

openocd acts as a gdb server, while gdb is a gdb client. By default they connect over port :3333 (: indicates that the port is on the localhost, not a remote connection). In cases you might have another gdb connection blocking the port.

$ ps -all
F S   UID   PID  PPID  C PRI  NI ADDR SZ WCHAN  TTY          TIME CMD
0 S  1000  5659  8712  0  80   0 -  6139 -      pts/1    00:00:28 openocd
0 S  1000  7549 16215  0  80   0 - 25930 se_sys pts/4    00:00:00 arm-none-eabi-g
...

In this case you can try killing gdb by:

$ kill -9 7549

or even

$ killall -9 arm-none-eabi-g

Notice, the process name is truncated for some reason...

If this did not help you can check if some other client has acquired the port, and kill the intruder accordingly. (In this case it was the gdb process so the above method would have worked, but in general it could be another process blocking the port.)

$ lsof -i :3333
COMMAND    PID USER   FD   TYPE DEVICE SIZE/OFF NODE NAME
openocd   5659  pln   12u  IPv4 387143      0t0  TCP localhost:dec-notes (LISTEN)
openocd   5659  pln   13u  IPv4 439988      0t0  TCP localhost:dec-notes->localhost:59560 (ESTABLISHED)
arm-none- 7825  pln   14u  IPv4 442734      0t0  TCP localhost:59560->localhost:dec-notes (ESTABLISHED)
$ kill -9 7825

itmdump no tracing or faulty output

There can be a number of reasons ITM tracing fails.

• The openocd.gdb script (or launch.json if in vscode using external ITM logging) enables ITM tracing assuming that the /tmp/itm.fifo and itmdump has been correctly setup before gdb is launched. So the first thing is to check that you follow the sequence suggested above.

• openocd.gdbsets enables ITM tracing by:

# 16000000 must match the core clock frequency
monitor tpiu config internal /tmp/itm.fifo uart off 16000000
monitor itm port 0 on

The transfer speed (baud rate) is automatically negotiated, however you can set it explicitly (maximum 2000000).

monitor tpiu config internal /tmp/itm.fifo uart off 16000000 2000000

You may try a lower value.

• The stm32f401re/stm32f411re defaults to 16000000 (16MHz) as the core clock frequency, based on an internal oscillator. If your application sets another core clock frequency the openocd.gdb script (tpiu setting) must be changed accordingly.

• openocd implements a number of events which might be called by gdb, e.g.:

(gdb) monitor reset init
adapter speed: 2000 kHz
target halted due to debug-request, current mode: Thread
xPSR: 0x01000000 pc: 0x08001298 msp: 0x20018000, semihosting
adapter speed: 8000 kHz

This invokes the init event, which sets the core clock to 64MHz. If you intend to run the MCU at 64MHz (using this approach), ITM will not work unless the tpiu setting matches 64MHz.

monitor tpiu config internal /tmp/itm.fifo uart off 64000000 2000000

If you on the other hand want to use monitor reset init but not having the core clock set to 64MHz, you can use a custom .cfg (instead of the one shipped with openocd). The original /usr/share/openocd/scripts/target/stm32f4x.cfg looks like this:

...
$_TARGETNAME configure -event reset-init {
	# Configure PLL to boost clock to HSI x 4 (64 MHz)
	mww 0x40023804 0x08012008   ;# RCC_PLLCFGR 16 Mhz /8 (M) * 128 (N) /4(P)
	mww 0x40023C00 0x00000102   ;# FLASH_ACR = PRFTBE | 2(Latency)
	mmw 0x40023800 0x01000000 0 ;# RCC_CR |= PLLON
	sleep 10                    ;# Wait for PLL to lock
	mmw 0x40023808 0x00001000 0 ;# RCC_CFGR |= RCC_CFGR_PPRE1_DIV2
	mmw 0x40023808 0x00000002 0 ;# RCC_CFGR |= RCC_CFGR_SW_PLL

	# Boost JTAG frequency
	adapter_khz 8000
}

Make a copy of the original and comment out the clock configuration, and store the file as stm32f4x.cfg:

...
$_TARGETNAME configure -event reset-init {
	# # Configure PLL to boost clock to HSI x 4 (64 MHz)
	# mww 0x40023804 0x08012008   ;# RCC_PLLCFGR 16 Mhz /8 (M) * 128 (N) /4(P)
	# mww 0x40023C00 0x00000102   ;# FLASH_ACR = PRFTBE | 2(Latency)
	# mmw 0x40023800 0x01000000 0 ;# RCC_CR |= PLLON
	# sleep 10                    ;# Wait for PLL to lock
	# mmw 0x40023808 0x00001000 0 ;# RCC_CFGR |= RCC_CFGR_PPRE1_DIV2
	# mmw 0x40023808 0x00000002 0 ;# RCC_CFGR |= RCC_CFGR_SW_PLL

	# Boost JTAG frequency
	adapter_khz 8000
}

You can start openocd to use these (local) settings by:

> openocd -f interface/stlink.cfg -f stm32f4x.cfg

A possible advantage of monitor reset init is that the adapter speed is set to 8MHz, which at least in theory gives better transfer rate between openocd and the stlink programmer (default is 2Mbit/s). I'm not sure the improvement is noticeable.

• ITM buffer overflow

In case the ITM buffer is saturated, ITM tracing stops working (and might be hard to recover). In such case:

1. Correct and recompile the program,

2. Erase the flash (using st-flash),

3. Power cycle the Nucleo (disconnect-and-re-connect),

4. Remove/re-make fifo, and finally re-start openocd/gdb.

This ensures 1) the program will not yet again overflow the ITM buffer, 2) the faulty program is gone (and not restarted accidentally on a RESET), 3) the programmer firmware is restarted and does not carry any persistent state, notice a RESET applies only to the target, not the programmer, so if the programmer crashes it needs to be power cycled), 4) the FIFO /tmp/itm.fifo, openocd and gdb will have fresh states.

• Check/update the Nucleo st-link firmware (as mentioned above).

Visual Studio Code

vscode is highly configurable, (keyboard shortcuts, keymaps, plugins etc.) Besides rust-analyzer, there is also Rust support through the rls-vscode plugin. Both should work nicely.

It is possible to run arm-none-eabi-gdb from within the vscode using the cortex-debug plugin.

For general information regarding debugging in vscode, see debugging.

Some useful (default) shortcuts:

• CTRL+SHIFT+b compilation tasks, (e.g., compile all examples cargo build --examples). Cargo is smart and just re-compiles what is changed.

• CTRL+SHIFT+d debug launch configurations, enter debug mode to choose a binary (e.g., itm internal (debug))

• Select the application you wan't to debug as the active window in vscode.

• F5 to start. It will run the progam up to main (your entry point). In the launch.json you can comment out the "runToMain": true if you want to the target to halt directly after reset. In this case it will open the cortex_m_rt/src/lib.rs file, which contains the startup code. From there you can continue F5 again.

• F6 to break. The program will now be in the infinite loop (for this example). In general it will just break wherever the program counter happens to be.

• You can view the ITM trace in the OUTPUT tab (if using the internal ITM viewer). Choose the dropdown SWO: ITM [port 0, type console]. It should now display:

[2019-01-02T21:35:26.457Z]   Hello, world!

• SHIFT-F5 shuts down the debugger.

You may step, view the current context variables, add watches, inspect the call stack, add breakpoints, inspect peripherals and registers. Read more in the documentation for the plugin.

Caveats

Visual Studio Code is not an "IDE", its a text editor with plugin support, with an API somewhat limiting what can be done from within a plugin (in comparison to Eclipse, IntelliJ...) regarding panel layouts etc. E.g., as far as I know you cannot view the adapter output (openocd) at the same time as the ITM trace, they are both under the OUTPUT tab. Moreover, each time you re-start a debug session, you need to re-select the SWO: Name [port 0, type console] to view the ITM output. You can work around this problem:

• Never shut down the debug session. Instead use the DEBUG CONSOLE (CTRL+SHIFT+Y) to get to the gdb console. This is not the full gdb interactive shell with some limitations (no tab completion e.g.). Make sure the MCU is stopped (F6). The console should show something like:

Program
 received signal SIGINT, Interrupt.
itm::__cortex_m_rt_main () at examples/itm.rs:26
26	    loop {

• Now you can edit an re-compile your program, e.g. changing the text:

  iprintln!(stim, "Hello, again!!!!");

• In the DEBUG CONSOLE, write load press ENTER write monitor reset init press ENTER.

load
{"token":97,"outOfBandRecord":[{"isStream":false,"type":"status","asyncClass":"download","output":[]}]}
`/home/pln/rust/app/target/thumbv7em-none-eabihf/debug/examples/itm' has changed; re-reading symbols.
Loading section .vector_table, size 0x400 lma 0x8000000
Loading section .text, size 0x10c8 lma 0x8000400
Loading section .rodata, size 0x2a8 lma 0x80014d0
Start address 0x8001298, load size 6000
Transfer rate: 9 KB/sec, 2000 bytes/write.

mon reset init
{"token":147,"outOfBandRecord":[],"resultRecords":{"resultClass":"done","results":[]}}
adapter speed: 2000 kHz
target halted due to debug-request, current mode: Thread 
xPSR: 0x01000000 pc: 0x08001298 msp: 0x20018000
adapter speed: 8000 kHz

• The newly compiled binary is now loaded and you can continue (F5). Switching to the OUTPUT window now preserves the ITM view and displays both traces:

[2019-01-02T21:43:27.988Z]   Hello, world!
[2019-01-02T22:07:29.090Z]   Hello, again!

• Using the gdb terminal (DEBUG CONSOLE) from within vscode is somewhat unstable/experimental. E.g., CTRL+c does not break the target (use F6, or write interrupt). The contiune command, indeed continues execution (and the control bar changes mode, but you cannot break using neither F6 nor interrupt). So it seems that the state of the cortex-debug plugin is not correctly updated. Moreover setting breakpoints from the gdb terminal indeed informs gdb about the breakpoint, but the state in vscode is not updated, so be aware.

Vscode Launch Configurations

Some example launch configurations from the .vscode/launch.json file:

        ...
        // Launch configuration for `examples`
        // - debug
        // - semihosting
        // - internal ITM/SWO tracing 
        // - run to main 
        {
            "type": "cortex-debug",
            "request": "launch",
            "servertype": "openocd",
            "name": "itm internal (debug)",
            "preLaunchTask": "cargo build --examples",
            "executable": "./target/thumbv7em-none-eabihf/debug/examples/${fileBasenameNoExtension}",
            "configFiles": [
                "interface/stlink.cfg",
                "target/stm32f4x.cfg"
            ],
            "postLaunchCommands": [
                "monitor arm semihosting enable",
            ],
            "swoConfig": {
                "enabled": true,
                "cpuFrequency": 16000000,
                "swoFrequency": 2000000,
                "source": "probe",
                "decoders": [
                    {
                        "type": "console",
                        "label": "ITM",
                        "port": 0
                    }
                ]
            },
            "runToMain": true,
            "cwd": "${workspaceRoot}"
        },
        // Launch configuration for `examples`
        // - debug
        // - semihosting
        // - ITM/SWO tracing to file/fifo `/tmp/itm.fifo`
        // - run to main   
        {
            "type": "cortex-debug",
            "request": "launch",
            "servertype": "openocd",
            "name": "itm fifo (debug)",
            "preLaunchTask": "cargo build --examples",
            "executable": "./target/thumbv7em-none-eabihf/debug/examples/${fileBasenameNoExtension}",
            "configFiles": [
                "interface/stlink.cfg",
                "target/stm32f4x.cfg"
            ],
            "postLaunchCommands": [
                "monitor arm semihosting enable",
                "monitor tpiu config internal /tmp/itm.fifo uart off 16000000",
                "monitor itm port 0 on"
            ],
            "runToMain": true,
            "cwd": "${workspaceRoot}"
        },

We see some similarities to the openocd.gdb file, we don't need to explicitly connect to the target (that is automatic). Also launching openocd is automatic (for good and bad, its re-started each time, unless you use the gdb prompt to load).

postLaunchCommands allows arbitrary commands to be executed by gdb once the session is up. E.g. in the app case we enable semihosting, while in the itm case we run monitor reset init to get the MCU in 64MHz (first example) or 16MHz (third example), before running the application (continue). Notice the first example uses the "stock" openocd configuration files, while the third example uses our local configuration files (that does not change the core frequency).

Rust-Analyzer

In order to properly track symbols from the core and/or std libraries, you need to add the Rust source. This can be done by:

> rustup component add rust-src

GDB Advanced Usage

There are numerous ways to automate gdb. Scripts can be run by the gdb command source (so for short). Scripting common tasks like setting breakpoints, dumping some memory region etc. can be really helpful.

For now: Copyright Per Lindgren per.lindgren@ltu.se