[embassy-executor]: Upstream "Earliest Deadline First" Scheduler (version 2)

[diondokter]

An updated #4035 that makes use of the new metadata API.

The example output is still good.

I've also retested with #4374 and these are the numbers:

cycles per poll
              nrf52832
master         78
edf            78
edf enabled    91

I don't have a nice rp2040 to test with...

[lulf]

Could you add a build combination for edf-scheduler in [package.metadata.embassy] as well? This will make the release tool check it for semver break as well.

[diondokter]

@lulf like this?

[lulf]

@diondokter Yup, that looks good for the example. Also add this to the same section in embassy-executor/Cargo.toml:

[package.metadata.embassy]
build = [
...
    {target = "thumbv7em-none-eabi", features = ["arch-cortex-m", "executor-interrupt", "executor-thread", "edf-scheduler"]},
...
]

[jamesmunns]

Updated numbers using the bench2 test from #4374:

Details

Main

Test OK, count=109655, cycles=15300
  R0: 00000001   R1: 200200e0   R2: 00000000   R3: 20020249
  R4: 0001ac57   R5: 00000004   R6: 00000000   R7: 2003ff60
  R8: ffffffff   R9: ffffffff  R10: ffffffff  R11: ffffffff
 R12: ffffffff   SP: 2003ff58   LR: 200002ab   PC: 20001094
XPSR: 6100000f

upstream-drs-2, SHA 1736fd055

EDF NOT active (default)

Test OK, count=106185, cycles=15799
  R0: 00000001   R1: 200200e0   R2: 00000000   R3: 20020249
  R4: 00019ec9   R5: 00000004   R6: 00000000   R7: 2003ff58
  R8: ffffffff   R9: ffffffff  R10: ffffffff  R11: ffffffff
 R12: ffffffff   SP: 2003ff50   LR: 200002af   PC: 20001098
XPSR: 6100000f

EDF active (feat=embassy-executor/edf-scheduler)

Test OK, count=95870, cycles=17499
  R0: 00000001   R1: 200200e8   R2: 00000000   R3: 20020251
  R4: 0001767e   R5: 00000004   R6: 20020088   R7: 2003ff40
  R8: ffffffff   R9: ffffffff  R10: ffffffff  R11: ffffffff
 R12: ffffffff   SP: 2003ff38   LR: 20000313   PC: 20001124
XPSR: 6100000f

cycles per poll
               nrf52832  rp2040
master         78        153
edf            78        158
edf enabled    91        175

edit: this matches the previous test run (a bit faster with edf enabled): #4035 (comment)

[diondokter]

Alright, so perf looks good and I just added the extra metadata and fixed the merge conflict

[Dirbaio]

Implementation looks good, and it's great that perf doesn't regress now 🚀

1. The public API is a bit strange though:

   • Metadata gives you a &Deadline.
   • Deadline has pub fn instant_ticks(&self) -> u64, to call that you must get the &Deadline from the Metadata.
   • Then it has a few pub fn *_current_task_deadline() -> impl Future methods, to call those you do not go through Metadata.

   Maybe the API should be just methods on Metadata? More consistent with the current name/set_name.

   • fn deadline(&self) -> u64
   • fn set_deadline(&self) -> u64
   • fn clear_deadline(&self)
   • etc

2. set_current_task_deadline_after: it requires embassy-time-driver, which should be optional.

3. about cargo features:

   • I'd name the feature scheduler-deadline.
   • I'm not sure if metadata-deadline is useful as a separate feature. What it'd be used for? I'd just make it all controlled by 'scheduler-deadline'.

[diondokter]

Alright! Made the changes.

You're right about the weird API. I didn't critically review it after the metadata refactoring.

• I'm not sure if metadata-deadline is useful as a separate feature. What it'd be used for? I'd just make it all controlled by 'scheduler-deadline'.

Idk, I just mirrored the existing metadata feature which looked nice to me. But I've removed it as per your preference.