spi: bcm2835: Workaround/fix for zero-length transfers

A relatively recent commit ([1]) contained optimisation for the PIO
SPI FIFO-filling functions. The commit message includes the phrase
"[t]he blind and counted loops are always called with nonzero count".
This is technically true, but it is still possible for count to become
zero before the loop is entered - if tfr->len is zero. Moving the loop
exit condition to the end of the loop saves a few cycles, but results
in a near-infinite loop should the revised count be zero on entry.

Strangely, zero-lengthed transfers aren't filtered by the SPI framework
and, even more strangely, the Python3 spidev library is triggering them
for no obvious reason.

Avoid the problem completely by bailing out of the main transfer
function early if trf->len is zero, although there may be a case for
moving the mitigation into the framework.

See: raspberrypi#4100

Signed-off-by: Phil Elwell <[email protected]>

[1] 26751de ("spi: bcm2835: Micro-optimise FIFO loops")

drivers/spi/spi-bcm2835.c

@@ -1090,6 +1090,16 @@ static int bcm2835_spi_transfer_one(struct spi_controller *ctlr,
 	unsigned long hz_per_byte, byte_limit;
 	u32 cs = bs->prepare_cs[spi->chip_select];
 
+	if (unlikely(!tfr->len)) {
+		static int warned;
+
+		if (!warned)
+			dev_warn(&spi->dev,
+				 "zero-length SPI transfer ignored\n");
+		warned = 1;
+		return 0;
+	}
+
 	/* set clock */
 	spi_hz = tfr->speed_hz;
 	clk_hz = clk_get_rate(bs->clk);