stackdump integration for better backtraces

[japaric]

the set of stackdump crates provide a mechanism to dump the contents of a device's memory (stackdump-capture) and analyze those stack dumps (stackdump-trace). the trace crate, in particular, can produce stack backtraces more complete than probe-run's as they include information about function arguments (see #62)

we'd like to make use of the trace crate to improve probe-run's backtraces but we'd like to avoid copying the whole stack memory from the device to the host as that's rather slow (see --measure-stack feature in early v0.3.x), which is the mechanism provided by the capture crate.

the integration work would consist of two parts

• implementing the RegisterData and MemoryRegion traits for the probe_rs API. I suggest using a newtype over &mut probe_rs::Core, e.g. pub Stackdump<'a>(pub &'a mut probe_rs::Core). rationale: some of the probe-rs API uses the Mutex<Session> (e.g. gdb-server); a newtype over the value is not compatible with such API. as this is a reusable component it may be better for it to reside outside of probe-run as a library

• replacing probe-runs's backtrace implementation with stackdump-trace's API. the API is generic; it needs to be instantiated with the newtype implemented in (1)

potential integration problems

• RegisterData and MemoryRegion are infallible. a probe_rs implementation can fail due to IO errors but there's no proper way to bubble up the result so one can map the result into None or unwrap all internal Results
• the 'static bound in stackdump_core::DeviceMemory::add_memory_region may be incompatible with the newtype over a reference; see (1)

[diondokter]

I made a branch with some initial changes here: https://github.com/tweedegolf/stackdump/tree/probe

[diondokter]

So, I made an implementation that works for my CLI, see the stackdump PR above. Small problem, though, is that probe-rs doesn't support reading the FPU registers in the latest release. I made a PR to fix that here: https://github.com/probe-rs/probe-rs/pull/1122

But even if that gets merged, that won't be in an official release until probe-rs 0.13. Theoretically this update isn't absolutely required. But if you miss the fpu registers, then most/all floats will not be able to be read. What do you guys think?

[japaric]

@diondokter nice! re: fpu registers, I think it's fine if they are not read and stackdump-capture-probe is published with a dependency on probe-rs v0.12.x. as you mentioned unwinding does not depend on the contents of the fpu registers so whenever their contents are missing probe-run could report a ? (unknown) value for those function arguments. probe-run currently does not report the contents of any register (function arguments) in the backtrace so having the data of the general purpose registers would already be a huge improvement.

[Urhengulas]

So, I made an implementation that works for my CLI, see the stackdump PR above. Small problem, though, is that probe-rs doesn't support reading the FPU registers in the latest release. I made a PR to fix that here: probe-rs/probe-rs#1122

But even if that gets merged, that won't be in an official release until probe-rs 0.13. Theoretically this update isn't absolutely required. But if you miss the fpu registers, then most/all floats will not be able to be read. What do you guys think?

That is great news! 🎉 Firstly, I agree with Jorge, that having non-float arguments being reported is already a good addition which is worth integrating. Secondly, speaking from past experience, it might take some time until the next probe-rs version releases, so I think we can move on without float support and add that as soon as the next probe-rs gets released.

[diondokter]

I have released stackdump-capture-probe: https://crates.io/crates/stackdump-capture-probe This is the adaptor that probe-run should be able to use.

I've got my own CLI as well and this is the usage: https://github.com/tweedegolf/stackdump/blob/master/cli/src/probe.rs

Printing is done like here: https://github.com/tweedegolf/stackdump/blob/master/cli/src/main.rs#L123-L163 This has some line wrapping logic because some data can be annoyingly long.