spi: stm32-qspi: Increase read throughput in indirect mode

When WATCHDOG_RESET() was replaced by schedule() in commit
29caf93 ("cyclic: Use schedule() instead of WATCHDOG_RESET()")
we not only reset the watchdog but also call the cyclic infrastructure
which takes time and has impact on read accesses performances.
Schedule() is called every u8/u16 or u32 read accesses which is overkill.

Move schedule() from _stm32_qspi_read_fifo() to _stm32_qspi_poll()
and call schedule() only every 1MB chunk of data.

Test performed by reading 64MB on sNOR on stm32mp157c-ev1 board:

          before      after    ratio
Read  :  201 KB/s    520KB/s   +258%

Signed-off-by: Patrice Chotard <patrice.chotard@foss.st.com>
Change-Id: I24d8196604e6c8f690a0b2024ee562d344d78f33
Reviewed-on: https://gerrit.st.com/c/mpu/oe/st/u-boot/+/518997
ACI: CIBUILD <MDG-smet-aci-builds@list.st.com>
Domain-Review: Christophe KERELLO <christophe.kerello@st.com>

Author: Patrice Chotard

drivers/spi/stm32_qspi.c

@@ -175,7 +175,6 @@ static void _stm32_qspi_read_fifo(void *val, void __iomem *addr, u8 len)
 	case sizeof(u8):
 		*((u8 *)val) = readb_relaxed(addr);
 	};
-	schedule();
 }
 
 static void _stm32_qspi_write_fifo(void *val, void __iomem *addr, u8 len)
@@ -236,6 +235,9 @@ static int _stm32_qspi_poll(struct stm32_qspi_priv *priv,
 		fifo(buf, &priv->regs->dr, step);
 		len -= step;
 		buf += step;
+
+		if (!(len % SZ_1M))
+			schedule();
 	}
 
 	return 0;