ContextQMD
Back to Tock

Tock Core Notes 2021-08-13

doc/wg/core/notes/core-notes-2021-08-13.md

latest4.0 KB
Original Source

Tock Core Notes 2021-08-13

Attendees

  • Pat Pannuto
  • Leon Schuermann
  • Jett Rink
  • Philip Levis
  • Alexandru Radovici
  • Vadim Sukhomlinov
  • Johnathan Van Why
  • Hudson Ayers
  • Brad Campbell
  • Amit Levy

Updates

Changelog

Panic Print Code Overhead

  • PR: https://github.com/tock/tock/pull/2759
  • Able to remove 7 kB from imix build, if you aren't going to use the debug print in panic handler.
  • Works because variable is const.
  • Still need change to kernel crate. Hard for out-of-tree boards.
  • Jett: idea: control config struct from rustc features.
  • Phil: hard to know what was actually built and what code is actually there.
  • Johnathan: might be able to use const generic.
  • Hudson: maybe not easy to use. Trait/object safety.
  • Alexandru: unless returns self should be safe.
  • Phil: still strange to have option set in kernel not boards. Maybe have some debug object.
  • Amit: agree, but need LLVM to do the right thing to remove unused code.
  • Amit: could have board define static variable that kernel is dependent on.
  • Johnathan: would work, but gets awkward.
  • discussion on compiler options, and how those compare or not to rustc config
  • Amit: if we want to test all combinations of flags, easier on command line.
  • Johnathan: hard to test all combinations.
  • Brad: do not agree with using a kernel::config option for this. Core problem is Rust does not remove unused trait functions. This is not the same as the syscall tracing and process load debug features.
  • Phil: pull this out of ProcessStandard?
  • Brad: agree, could be in different trait.
  • Phil: should be in different object.
  • Amit: could use different impl of Process.
  • Trait if never used would be elided.
  • Phil: having printing object could be useful in other ways.
  • Amit: is this a bug/limitation of LLVM/Rust that should fix this for us?
  • Hudson: Dead virtual function elimination is what we want, LLVM supports it, Rust does not. Trying to see if we can get this added to Rust.
  • Amit: are we all at least in support of the config in the short term?
  • Brad: no, why not just a very small, easy to cherry-pick commit.
  • Amit: strongly in favor of config option rather than that.
  • Phil: opposed to config because it is a short-term fix.
  • Hudson: config options at least all compile.
  • Brad: that's a low bar.
  • Hudson: traits can be a combination nightmare too.
  • Brad: let's get 2.0 done.
  • Phil: supportive of software redesign (decomposition) of process trait.

HIL buffer lengths too short

  • What happens if length longer than buffer in HILs?
    • Truncate length to size of buffer?
    • Return error (fail)?
  • Which option should we go with?
  • Leon: fail.
  • Alex: fail.
  • Amit: fail.
  • Amit: length field is needed because we need to give up ownership of buffer, but we need to get the entire buffer back. Could use LeaseableBuffer instead.
  • Leon: nice to use leaseable buffer so do not have to shift bytes around.
  • Phil: disagree, harder to use leaseable buffer.
  • Phil: SPI truncates in case want to read fewer bytes than write.
  • Leon: LeaseableBuffer nice for users, maybe less so for implementors.
  • Phil: it would be good to get more experience with LeaseableBuffer.
  • Amit: seems consensus is fail fast.

2.0

  • Hudson: goal still mid August for release?
  • Amit: what testing are we still missing?
  • Hudson: several boards missing tests.
  • Alex: I will ping my student.
  • Amit: shouldn't block release on untested boards we don't have.
  • Brad: What stability guarantees do we want? https://github.com/tock/tock/blob/master/doc/syscalls/README.md