spi: cadence-qspi: add no-IRQ mode to indirect reads

Support reads through polling, without any IRQ. The main reason is
performance; profiling shows that the first IRQ comes quickly on our
specific hardware. Once this IRQ arrives, we poll until all data is
retrieved. Avoid initial sleep to reduce IRQ count.

Hide this behavior behind a quirk flag.

This is confirmed through micro-benchmarks, but also end-to-end
performance tests. Mobileye EyeQ5, octal flash, reading 235M on a UBIFS
filesystem:
 - No optimizations, ~10.34s, ~22.7 MB/s, 199230 IRQs
 - CQSPI_SLOW_SRAM,  ~10.34s, ~22.7 MB/s,  70284 IRQs
 - CQSPI_RD_NO_IRQ,   ~9.37s, ~25.1 MB/s,    521 IRQs
Signed-off-by: Théo Lebrun <theo.lebrun@bootlin.com>
Link: https://lore.kernel.org/r/20240423-cdns-qspi-mbly-v4-2-3d2a7b535ad0@bootlin.comSigned-off-by: Mark Brown <broonie@kernel.org>

drivers/spi/spi-cadence-quadspi.c

@@ -42,6 +42,7 @@ static_assert(CQSPI_MAX_CHIPSELECT <= SPI_CS_CNT_MAX);
 #define CQSPI_NO_SUPPORT_WR_COMPLETION	BIT(3)
 #define CQSPI_SLOW_SRAM		BIT(4)
 #define CQSPI_NEEDS_APB_AHB_HAZARD_WAR	BIT(5)
+#define CQSPI_RD_NO_IRQ			BIT(6)
 
 /* Capabilities */
 #define CQSPI_SUPPORTS_OCTAL		BIT(0)
@@ -702,6 +703,7 @@ static int cqspi_indirect_read_execute(struct cqspi_flash_pdata *f_pdata,
 				       const size_t n_rx)
 {
 	struct cqspi_st *cqspi = f_pdata->cqspi;
+	bool use_irq = !(cqspi->ddata && cqspi->ddata->quirks & CQSPI_RD_NO_IRQ);
 	struct device *dev = &cqspi->pdev->dev;
 	void __iomem *reg_base = cqspi->iobase;
 	void __iomem *ahb_base = cqspi->ahb_base;
@@ -725,17 +727,20 @@ static int cqspi_indirect_read_execute(struct cqspi_flash_pdata *f_pdata,
 	 * all the read interrupts disabled for max performance.
 	 */
 
-	if (!cqspi->slow_sram)
+	if (use_irq && cqspi->slow_sram)
+		writel(CQSPI_REG_IRQ_WATERMARK, reg_base + CQSPI_REG_IRQMASK);
+	else if (use_irq)
 		writel(CQSPI_IRQ_MASK_RD, reg_base + CQSPI_REG_IRQMASK);
 	else
-		writel(CQSPI_REG_IRQ_WATERMARK, reg_base + CQSPI_REG_IRQMASK);
+		writel(0, reg_base + CQSPI_REG_IRQMASK);
 
 	reinit_completion(&cqspi->transfer_complete);
 	writel(CQSPI_REG_INDIRECTRD_START_MASK,
 	       reg_base + CQSPI_REG_INDIRECTRD);
 
 	while (remaining > 0) {
-		if (!wait_for_completion_timeout(&cqspi->transfer_complete,
+		if (use_irq &&
+		    !wait_for_completion_timeout(&cqspi->transfer_complete,
 						 msecs_to_jiffies(CQSPI_READ_TIMEOUT_MS)))
 			ret = -ETIMEDOUT;
 
@@ -777,7 +782,7 @@ static int cqspi_indirect_read_execute(struct cqspi_flash_pdata *f_pdata,
 			bytes_to_read = cqspi_get_rd_sram_level(cqspi);
 		}
 
-		if (remaining > 0) {
+		if (use_irq && remaining > 0) {
 			reinit_completion(&cqspi->transfer_complete);
 			if (cqspi->slow_sram)
 				writel(CQSPI_REG_IRQ_WATERMARK, reg_base + CQSPI_REG_IRQMASK);