acrn-kernel/Documentation/driver-api/soundwire/error_handling.rst

66 lines
3.8 KiB
ReStructuredText

========================
SoundWire Error Handling
========================
The SoundWire PHY was designed with care and errors on the bus are going to
be very unlikely, and if they happen it should be limited to single bit
errors. Examples of this design can be found in the synchronization
mechanism (sync loss after two errors) and short CRCs used for the Bulk
Register Access.
The errors can be detected with multiple mechanisms:
1. Bus clash or parity errors: This mechanism relies on low-level detectors
that are independent of the payload and usages, and they cover both control
and audio data. The current implementation only logs such errors.
Improvements could be invalidating an entire programming sequence and
restarting from a known position. In the case of such errors outside of a
control/command sequence, there is no concealment or recovery for audio
data enabled by the SoundWire protocol, the location of the error will also
impact its audibility (most-significant bits will be more impacted in PCM),
and after a number of such errors are detected the bus might be reset. Note
that bus clashes due to programming errors (two streams using the same bit
slots) or electrical issues during the transmit/receive transition cannot
be distinguished, although a recurring bus clash when audio is enabled is a
indication of a bus allocation issue. The interrupt mechanism can also help
identify Slaves which detected a Bus Clash or a Parity Error, but they may
not be responsible for the errors so resetting them individually is not a
viable recovery strategy.
2. Command status: Each command is associated with a status, which only
covers transmission of the data between devices. The ACK status indicates
that the command was received and will be executed by the end of the
current frame. A NAK indicates that the command was in error and will not
be applied. In case of a bad programming (command sent to non-existent
Slave or to a non-implemented register) or electrical issue, no response
signals the command was ignored. Some Master implementations allow for a
command to be retransmitted several times. If the retransmission fails,
backtracking and restarting the entire programming sequence might be a
solution. Alternatively some implementations might directly issue a bus
reset and re-enumerate all devices.
3. Timeouts: In a number of cases such as ChannelPrepare or
ClockStopPrepare, the bus driver is supposed to poll a register field until
it transitions to a NotFinished value of zero. The MIPI SoundWire spec 1.1
does not define timeouts but the MIPI SoundWire DisCo document adds
recommendation on timeouts. If such configurations do not complete, the
driver will return a -ETIMEOUT. Such timeouts are symptoms of a faulty
Slave device and are likely impossible to recover from.
Errors during global reconfiguration sequences are extremely difficult to
handle:
1. BankSwitch: An error during the last command issuing a BankSwitch is
difficult to backtrack from. Retransmitting the Bank Switch command may be
possible in a single segment setup, but this can lead to synchronization
problems when enabling multiple bus segments (a command with side effects
such as frame reconfiguration would be handled at different times). A global
hard-reset might be the best solution.
Note that SoundWire does not provide a mechanism to detect illegal values
written in valid registers. In a number of cases the standard even mentions
that the Slave might behave in implementation-defined ways. The bus
implementation does not provide a recovery mechanism for such errors, Slave
or Master driver implementers are responsible for writing valid values in
valid registers and implement additional range checking if needed.