Skip to contentSkip to conversation

arm: spi: atmel_qspi: fix race condition in transfer completion check

PATCH In review Ramin Moussavi 2 participants 2 messages Jul 22, 2025 – Aug 13, 2025

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 arm: spi: atmel_qspi: fix race condition in transfer completion check

In atmel_qspi_transfer(), the status register is polled with:

imr = QSPI_SR_INSTRE | QSPI_SR_CSR; return readl_poll_timeout(aq->regs + QSPI_SR, sr,

                            (sr & imr) == imr,
                            ATMEL_QSPI_TIMEOUT);

However, this is racy: QSPI_SR_INSTRE can be set before QSPI_SR_CSR, and will then be cleared by the read. If that happens, the condition "(sr & imr) == imr" can never be true, and the function times out.

This race condition is avoided in at91bootstrap by accumulating the status bits across reads until both bits have been observed:

  /* Poll INSTruction End and Chip Select Rise flags. */
  imr = (QSPI_SR_INSTRE | QSPI_SR_CSR);
  sr = 0;
  do {
    udelay(1);
    sr |= qspi_readl(qspi, QSPI_SR) & imr;
  } while ((--timeout) && (sr != imr));

Update U-Boot's atmel_qspi_transfer() to use the same pattern, ensuring that both flags are observed even if they are not set simultaneously.

---
 drivers/spi/atmel-quadspi.c | 14 ++++++++++++--
 1 file changed, 12 insertions(+), 2 deletions(-)
drivers/spi/atmel-quadspi.c +12 -2
index 8aa7a83aef4..0f9ce409a0c 100644
@@ -616,6 +616,8 @@ static int atmel_qspi_transfer(struct atmel_qspi *aq,
 			       const struct spi_mem_op *op, u32 offset)
 {
 	u32 sr, imr;
+	u32 val = 0;
+	unsigned long timeout;
 
 	/* Skip to the final steps if there is no data */
 	if (op->data.nbytes) {
@@ -636,8 +638,16 @@ static int atmel_qspi_transfer(struct atmel_qspi *aq,
 
 	/* Poll INSTruction End and Chip Select Rise flags. */
 	imr = QSPI_SR_INSTRE | QSPI_SR_CSR;
-	return readl_poll_timeout(aq->regs + QSPI_SR, sr, (sr & imr) == imr,
-				  ATMEL_QSPI_TIMEOUT);
+
+	timeout = timer_get_us() + ATMEL_QSPI_TIMEOUT;
+	while (1) {
+		val |= readl(aq->regs + QSPI_SR) & imr;
+		if ((val & imr) == imr)
+			return 0;
+
+		if (time_after(timer_get_us(), timeout))
+			return -ETIMEDOUT;
+	}
 }
 
 static int atmel_qspi_sama7g5_set_cfg(struct atmel_qspi *aq,
Signature · 1 line
2.50.1
#1 lore raw
Quoted 22 lines · Ramin Moussavi

In atmel_qspi_transfer(), the status register is polled with:

imr = QSPI_SR_INSTRE | QSPI_SR_CSR; return readl_poll_timeout(aq->regs + QSPI_SR, sr,

                            (sr & imr) == imr,
                            ATMEL_QSPI_TIMEOUT);

However, this is racy: QSPI_SR_INSTRE can be set before QSPI_SR_CSR, and will then be cleared by the read. If that happens, the condition "(sr & imr) == imr" can never be true, and the function times out.

This race condition is avoided in at91bootstrap by accumulating the status bits across reads until both bits have been observed:

  /* Poll INSTruction End and Chip Select Rise flags. */
  imr = (QSPI_SR_INSTRE | QSPI_SR_CSR);
  sr = 0;
  do {
    udelay(1);
    sr |= qspi_readl(qspi, QSPI_SR) & imr;
  } while ((--timeout) && (sr != imr));
Quoted 6 lines

Update U-Boot's atmel_qspi_transfer() to use the same pattern, ensuring that both flags are observed even if they are not set simultaneously.

---

Applied to u-boot-at91/master, thanks !

For future, please send out patch updates with 'git send-email -v 2' (e.g. for a second version) and include a changelog below the '---' line

Eugen