spi: axi-spi-engine: fix sleep calculation

The sleep calculation was not taking into account increased delay when
the SPI device is not running at the maximum SCLK frequency.

Rounding down when one SCLK tick was the same as the instruction
execution time was fine, but it rounds down too much when SCLK is
slower. This changes the rounding to round up instead while still
taking into account the instruction execution time so that small
delays remain accurate.

Fixes: be9070bc ("spi: axi-spi-engine: fix sleep ticks calculation")
Signed-off-by: David Lechner <dlechner@baylibre.com>
Link: https://patch.msgid.link/20240620-spi-axi-spi-engine-fix-sleep-time-v1-1-b20b527924a0@baylibre.comSigned-off-by: Mark Brown <broonie@kernel.org>

drivers/spi/spi-axi-spi-engine.c

@@ -164,16 +164,20 @@ static void spi_engine_gen_xfer(struct spi_engine_program *p, bool dry,
 }
 
 static void spi_engine_gen_sleep(struct spi_engine_program *p, bool dry,
-				 int delay_ns, u32 sclk_hz)
+				 int delay_ns, int inst_ns, u32 sclk_hz)
 {
 	unsigned int t;
 
-	/* negative delay indicates error, e.g. from spi_delay_to_ns() */
-	if (delay_ns <= 0)
+	/*
+	 * Negative delay indicates error, e.g. from spi_delay_to_ns(). And if
+	 * delay is less that the instruction execution time, there is no need
+	 * for an extra sleep instruction since the instruction execution time
+	 * will already cover the required delay.
+	 */
+	if (delay_ns < 0 || delay_ns <= inst_ns)
 		return;
 
-	/* rounding down since executing the instruction adds a couple of ticks delay */
-	t = DIV_ROUND_DOWN_ULL((u64)delay_ns * sclk_hz, NSEC_PER_SEC);
+	t = DIV_ROUND_UP_ULL((u64)(delay_ns - inst_ns) * sclk_hz, NSEC_PER_SEC);
 	while (t) {
 		unsigned int n = min(t, 256U);
 
@@ -220,10 +224,16 @@ static void spi_engine_compile_message(struct spi_message *msg, bool dry,
 	struct spi_device *spi = msg->spi;
 	struct spi_controller *host = spi->controller;
 	struct spi_transfer *xfer;
-	int clk_div, new_clk_div;
+	int clk_div, new_clk_div, inst_ns;
 	bool keep_cs = false;
 	u8 bits_per_word = 0;
 
+	/*
+	 * Take into account instruction execution time for more accurate sleep
+	 * times, especially when the delay is small.
+	 */
+	inst_ns = DIV_ROUND_UP(NSEC_PER_SEC, host->max_speed_hz);
+
 	clk_div = 1;
 
 	spi_engine_program_add_cmd(p, dry,
@@ -252,7 +262,7 @@ static void spi_engine_compile_message(struct spi_message *msg, bool dry,
 
 		spi_engine_gen_xfer(p, dry, xfer);
 		spi_engine_gen_sleep(p, dry, spi_delay_to_ns(&xfer->delay, xfer),
-				     xfer->effective_speed_hz);
+				     inst_ns, xfer->effective_speed_hz);
 
 		if (xfer->cs_change) {
 			if (list_is_last(&xfer->transfer_list, &msg->transfers)) {
@@ -262,7 +272,7 @@ static void spi_engine_compile_message(struct spi_message *msg, bool dry,
 					spi_engine_gen_cs(p, dry, spi, false);
 
 				spi_engine_gen_sleep(p, dry, spi_delay_to_ns(
-					&xfer->cs_change_delay, xfer),
+					&xfer->cs_change_delay, xfer), inst_ns,
 					xfer->effective_speed_hz);
 
 				if (!list_next_entry(xfer, transfer_list)->cs_off)