Hello,

Am Thu, Jan 09, 2025 at 05:27:58PM +0100 schrieb Alexander Dahl:

Hello Bence,

I had another round of intense looking at the code, this time I
focused on pm_runtime handling.  Although I just learned about the
basic concepts, I think the ported patch has some mistakes.  I'll
comment here, because I don't have a SAMA7G5 to test, and I'm not
confident enough to fix the code, but I think it's worth reporting.
See below.

Am Thu, Nov 28, 2024 at 06:43:15PM +0100 schrieb Csókás, Bence:

quoted

From: Tudor Ambarus <redacted>

[…]

quoted

+	/* Release the chip-select. */
+	ret = atmel_qspi_reg_sync(aq);
+	if (ret) {
+		pm_runtime_mark_last_busy(&aq->pdev->dev);
+		pm_runtime_put_autosuspend(&aq->pdev->dev);
+		return ret;
+	}
+	atmel_qspi_write(QSPI_CR_LASTXFER, aq, QSPI_CR);
+
+	return atmel_qspi_wait_for_completion(aq, QSPI_SR_CSRA);
+}

This function atmel_qspi_sama7g5_transfer() seems to be called from
atmel_qspi_exec_op() through ops->transfer() only.  I think the two
lines in the error handling of atmel_qspi_reg_sync() lead to
unbalanced calls of pm_runtime_xxx.  Compare with
atmel_qspi_transfer() which has no calls to pm_runtime, everything is
covered by atmel_qspi_exec_op() in this case where the pm_runtime
calls surround ->set_cfg() and ->transfer().  Right?

This problem has been addressed in downstream kernel (linux4sam) by
Claudiu Beznea back in 2023 already:

https://github.com/linux4sam/linux-at91/commit/e59f646f516088fdab6d8213d8acda0c1063ec21

[…]

The whole call tree from atmel_qspi_sama7g5_setup() downwards is not
covered by pm_runtime get and put calls, although heavily doing i/o.
Further down in atmel_qspi_setup() there's a write to QSPI_SCR which
seems to be handled correctly.

Same for this:

https://github.com/linux4sam/linux-at91/commit/5ff0e74c1d548599fe85113e2f1817cb8a052b15

Some hunks of that seem to have made it to upstream, not sure?

Maybe Microchip should upstream those fixes, now that SAMA7G5 support
was ported to mainline?

Greets
Alex