Skip to contentSkip to conversation

spi: add runtime PM for transfer_one_message

PATCH In review Chunyan Zhang 2 participants 4 messages Nov 2, 2020 – Nov 3, 2020 SPI

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 spi: add runtime PM for transfer_one_message

From: Chunyan Zhang <[email protected]>

Before transfer message, spi devices probably have been in runtime suspended, that would cause the kernel crash on some platforms once access spi registers, such as on Unisoc's SoCs. The spi devices can be suspended until message transfer completed.

Also this patch move the API spi_idle_runtime_pm() above to spi_transfer_one_message() which need to call that API.

Fixes b158935f70b9 ("spi: Provide common spi_message processing loop")
---
 drivers/spi/spi.c | 28 ++++++++++++++++++++--------
 1 file changed, 20 insertions(+), 8 deletions(-)
drivers/spi/spi.c +20 -8
index 0cab239d8e7f..63f7ebea7076 100644
@@ -1224,6 +1224,14 @@ static void _spi_transfer_cs_change_delay(struct spi_message *msg,
 	}
 }
 
+static void spi_idle_runtime_pm(struct spi_controller *ctlr)
+{
+	if (ctlr->auto_runtime_pm) {
+		pm_runtime_mark_last_busy(ctlr->dev.parent);
+		pm_runtime_put_autosuspend(ctlr->dev.parent);
+	}
+}
+
 /*
  * spi_transfer_one_message - Default implementation of transfer_one_message()
  *
@@ -1240,6 +1248,16 @@ static int spi_transfer_one_message(struct spi_controller *ctlr,
 	struct spi_statistics *statm = &ctlr->statistics;
 	struct spi_statistics *stats = &msg->spi->statistics;
 
+	if (ctlr->auto_runtime_pm) {
+		ret = pm_runtime_get_sync(ctlr->dev.parent);
+		if (ret < 0) {
+			pm_runtime_put_noidle(ctlr->dev.parent);
+			dev_err(&ctlr->dev, "Failed to power device: %d\n",
+				ret);
+			return ret;
+		}
+	}
+
 	spi_set_cs(msg->spi, true);
 
 	SPI_STATISTICS_INCREMENT_FIELD(statm, messages);
@@ -1329,6 +1347,8 @@ static int spi_transfer_one_message(struct spi_controller *ctlr,
 
 	spi_finalize_current_message(ctlr);
 
+	spi_idle_runtime_pm(ctlr);
+
 	return ret;
 }
 
@@ -1346,14 +1366,6 @@ void spi_finalize_current_transfer(struct spi_controller *ctlr)
 }
 EXPORT_SYMBOL_GPL(spi_finalize_current_transfer);
 
-static void spi_idle_runtime_pm(struct spi_controller *ctlr)
-{
-	if (ctlr->auto_runtime_pm) {
-		pm_runtime_mark_last_busy(ctlr->dev.parent);
-		pm_runtime_put_autosuspend(ctlr->dev.parent);
-	}
-}
-
 /**
  * __spi_pump_messages - function which processes spi message queue
  * @ctlr: controller to process queue for
Signature · 1 line
2.20.1
maintainer
#1 lore raw
Quoted 6 lines · Chunyan Zhang

From: Chunyan Zhang <[email protected]>

Before transfer message, spi devices probably have been in runtime suspended, that would cause the kernel crash on some platforms once access spi registers, such as on Unisoc's SoCs. The spi devices can be suspended until message transfer completed.

This commit message is a bit hard to follow so I don't really understand what the issue is. We only ever call transfer_one_message() from within __spi_pump_messages() which already handles auto_runtime_pm so I'm not seeing the situation where we might get to transfer_one_message() without having already runtime resumed the controller. What exactly is the error situation here? This code has been around for a while and I'm not aware of reports of issues here and I can't see anything unusual that the Spreadtrum driver is doing.

Also why are we doing this in transfer_one_message() where it will only work for controllers using that? If we're missing runtime PM in some paths then presumably controllers with a custom implementation are also going to be affected as well, auto_runtime_pm is supposed to work for them as well.

#2 lore raw
Quoted 9 lines · Mark Brown
Quoted 6 lines · Chunyan Zhang

From: Chunyan Zhang <[email protected]>

Before transfer message, spi devices probably have been in runtime suspended, that would cause the kernel crash on some platforms once access spi registers, such as on Unisoc's SoCs. The spi devices can be suspended until message transfer completed.

Quoted 8 lines

This commit message is a bit hard to follow so I don't really understand what the issue is. We only ever call transfer_one_message() from within __spi_pump_messages() which already handles auto_runtime_pm so I'm not seeing the situation where we might get to transfer_one_message() without having already runtime resumed the controller. What exactly is the error situation here? This code has been around for a while and I'm not aware of reports of issues here and I can't see anything unusual that the Spreadtrum driver is doing.

With further tests and looking into this part of code, the problem we found recently on our platform which runs kernel 4.14 can be fixed by this commit [1].
In a word, there's no issue here indeed, this patch should be ignored, sorry for the noise and thanks for your review.

Chunyan

[1] https://lkml.org/lkml/2019/10/30/173

Quoted 6 lines

Also why are we doing this in transfer_one_message() where it will only work for controllers using that? If we're missing runtime PM in some paths then presumably controllers with a custom implementation are also going to be affected as well, auto_runtime_pm is supposed to work for them as well.

maintainer
#3 lore raw
Quoted 11 lines · Chunyan Zhang
Quoted 3 lines · Mark Brown

the error situation here? This code has been around for a while and I'm not aware of reports of issues here and I can't see anything unusual that the Spreadtrum driver is doing.

With further tests and looking into this part of code, the problem we found recently on our platform which runs kernel 4.14 can be fixed by this commit [1].
In a word, there's no issue here indeed, this patch should be ignored, sorry for the noise and thanks for your review.

Ah, great - glad it's sorted.