The hardware has a 128 byte RX FIFO buffer for each independent UART.
Previously, the code was always reading that byte-by-byte via
independent SPI transactions and the associated overhead. In practice,
this led to up to eight bytes over SPI for just one byte in the UART's
RX FIFO:
- reading the global IRQ register (two bytes, one for command, the other
for data)
- reading one UART's ISR (again two bytes)
- reading the byte count (two bytes yet again)
- finally, reading one byte of the FIFO via another two-byte transaction
We cannot always use a batched read. If the TTY is set to intercept
break conditions or report framing or parity errors, then it is required
to check the Line Status Register (LSR) for each byte which is read from
the RX FIFO. The documentation does not show a way of doing that in a
single SPI transaction; registers 0x00 and 0x04 are separate.
In my testing, this is no silver bullet. I was feeding 2MB of random
data over four daisy-chaned UARTs of MAX14830, and this is the
distribution that I was getting:
- R <= 1: 7437322
- R <= 2: 162093
- R <= 4: 4093
- R <= 8: 4196
- R <= 16: 645
- R <= 32: 165
- R <= 64: 58
- R <= 128: 0
For a reference, batching the write operations works much better:
- W <= 1: 2664
- W <= 2: 1305
- W <= 4: 627
- W <= 8: 371
- W <= 16: 121
- W <= 32: 68
- W <= 64: 33
- W <= 128: 63139
That's probably because this HW/SW combination (Clearfog Base, Armada
388) is probably "good enough" to react to the chip's IRQ "fast enough"
most of the time. Still, I was getting RX overruns every now and then.
In future, I plan to improve this by letting the RX FIFO be filled a
little more (the chip has support for that and also for a "stale
timeout" to prevent additional starvation).
Signed-off-by: Jan Kundrát <jan.kundrat@cesnet.cz>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2b4bac48c1