MX682X
commented
2 years ago I've started with debugging right now and I stumbled upon 3 things so far that might create problems:
- Do you make sure that Serial finishes before putting the CPU into sleep? - It's not impacting in IDLE, but will probably in the deeper sleep states, since USART can wake the MCU only on SOF. the MCU might go to sleep with bytes still left in the buffer or maybe a half-finished transmitted byte.
- I noticed that a BUSERR is thrown by the TWI SLAVE in the last byte of a transaction when in STANDBY (and I expect POWER_DOWN). I think this is kind of a silicon bug. When in Standby or Power Down, all High-Frequency clocks are disabled, and thus the clock for TWI MASTER. But the TWI MASTER has to be enabled for BUSERR to work. So, the BUSERR circuitry does not register the Start condition, but only a STOP condition and thus creates the BUSERR. This leads the Wire library to virtually throw away the whole transaction. I'll have to think about how to work around that, if someone has any suggestions, like a "this interrupt has woken up the CPU flag", if something like this exists, feel free to write a comment.
- I'm not clearing the BUSERR flag, so any consecutive transactions will be lost too
SpenceKonde
ObviousInRetrospect
I've noticed unreliable slave behavior in SLEEP_MODE_STANDBY (and SLEEP_MODE_POWER_DOWN). This seems to be more of an issue when transferring larger buffers. I thought guarding the sleep_cpu() call with !Wire.slaveTransactionOpen() was sufficient. The twi supports wake from STANDBY and even POWER_DOWN but every i2c slave I have written with DxCore and megaTinyCore has been flakey unless set for SLEEP_MODE_IDLE.
The code here: https://github.com/ObviousInRetrospect/DualMode/tree/main/DualModeExample
works as expected
but if SLEEP_MODE_IDLE is changed to SLEEP_MODE_STANDBY the slave starts corrupting data: https://github.com/ObviousInRetrospect/DualMode/blob/10dfca6b5bc6c8fedf1605d2d8c5981e2e4aca34/DualModeExample/DualModeExample.ino#L319
this is most easily seen (especially without an ina3221) by uncommenting the test pattern data fill (L184-188): https://github.com/ObviousInRetrospect/DualMode/blob/10dfca6b5bc6c8fedf1605d2d8c5981e2e4aca34/DualModeExample/DualModeExample.ino#L184
When I raised this in https://github.com/SpenceKonde/DxCore/discussions/316 @MX682X suggested opening a separate issue with more details.
The output on the master https://github.com/ObviousInRetrospect/DualMode/tree/main/DualModeExampleClient in SLEEP_MODE_IDLE is a long string of correct reads:
(the master does a 32-byte read follow by a 25-byte read. the last 4 bytes are a crc32 covering the entire data)
omitting thousands of similar lines with very few errors. In an overnight run ran this overnight, 99.2% of transfers were clean, 0.8% corruption due to race condition around crc calculation where data looks fine but the data changed during crc calculation.
when the slave is changed to SLEEP_MODE_STANDBY, the output changes to:
[content in square brackets are comments I added]
This is with test pattern data in which each byte contains its own address