Skip to contentSkip to conversation

net: fec: add a check for CONFIG_PM to avoid clock count mis-match

PATCH In review Chuhong Yuan 3 participants 7 messages Nov 12, 2019 – Nov 18, 2019

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 net: fec: add a check for CONFIG_PM to avoid clock count mis-match

If CONFIG_PM is enabled, runtime pm will work and call runtime_suspend automatically to disable clks.
Therefore, remove only needs to disable clks when CONFIG_PM is disabled. Add this check to avoid clock count mis-match caused by double-disable.

Fixes c43eab3eddb4 ("net: fec: add missed clk_disable_unprepare in remove")

---
Changes in v2:
- Add fixes tag.

 drivers/net/ethernet/freescale/fec_main.c | 2 ++
 1 file changed, 2 insertions(+)
drivers/net/ethernet/freescale/fec_main.c +2 -0
index a9c386b63581..696550f4972f 100644
@@ -3645,8 +3645,10 @@ fec_drv_remove(struct platform_device *pdev)
 		regulator_disable(fep->reg_phy);
 	pm_runtime_put(&pdev->dev);
 	pm_runtime_disable(&pdev->dev);
+#ifndef CONFIG_PM
 	clk_disable_unprepare(fep->clk_ahb);
 	clk_disable_unprepare(fep->clk_ipg);
+#endif
 	if (of_phy_is_fixed_link(np))
 		of_phy_deregister_fixed_link(np);
 	of_node_put(fep->phy_node);
Signature · 1 line
2.23.0
maintainer
#1 lore raw

From: Chuhong Yuan <[email protected]>
Date: Tue, 12 Nov 2019 19:28:30 +0800

Quoted 7 lines

If CONFIG_PM is enabled, runtime pm will work and call runtime_suspend automatically to disable clks.
Therefore, remove only needs to disable clks when CONFIG_PM is disabled. Add this check to avoid clock count mis-match caused by double-disable.

Fixes c43eab3eddb4 ("net: fec: add missed clk_disable_unprepare in remove")

I don't understand this at all.

The clk disables here match the unconditional clk enables in the probe function.

And that is how this is supposed to work, probe enables match remove disables. And suspend disables match resume enables.

Why isn't the probe enable taking the correct count, which the remove function must match with an appropriate disable? There is no CONFIG_PM guarding the probe time clk enables.

#2 lore raw

From: David Miller <[email protected]> Sent: Wednesday, November 13, 2019 3:13 AM

Quoted 17 lines

From: Chuhong Yuan <[email protected]>
Date: Tue, 12 Nov 2019 19:28:30 +0800

Quoted 8 lines

If CONFIG_PM is enabled, runtime pm will work and call runtime_suspend automatically to disable clks.
Therefore, remove only needs to disable clks when CONFIG_PM is disabled. Add this check to avoid clock count mis-match caused by double-disable.

Fixes c43eab3eddb4 ("net: fec: add missed clk_disable_unprepare in

remove")

I don't understand this at all.

The clk disables here match the unconditional clk enables in the probe function.

Quoted 6 lines

And that is how this is supposed to work, probe enables match remove disables. And suspend disables match resume enables.

Why isn't the probe enable taking the correct count, which the remove function must match with an appropriate disable? There is no CONFIG_PM guarding the probe time clk enables.

Current driver runtime pm callback enable/disable clk_ipg/clk_ahb two clks. CONFIG_PM is a optional config, if CONFIG_PM is disabled, runtime callbacks will Not be called.
The driver enable clk_ipg/clk_ahb two clks during probe, and depends runtime suspend to disable the two clks if CONFIG_PM is enabled.

In driver remove() also need to disable the two clks if CONIFG_PM is disabled. So the patch c43eab3eddb4 ("net: fec: add missed clk_disable_unprepare in emove") target the fixes if CONFIG_PM is not enabled, but the patch ignore to check the CONFIG_PM that make clock count mismatch in CONFIG_PM enabled case.

Andy

maintainer
#3 lore raw
In reply to Chuhong Yuan

From: Chuhong Yuan <[email protected]>
Date: Tue, 12 Nov 2019 19:28:30 +0800

Quoted 7 lines

If CONFIG_PM is enabled, runtime pm will work and call runtime_suspend automatically to disable clks.
Therefore, remove only needs to disable clks when CONFIG_PM is disabled. Add this check to avoid clock count mis-match caused by double-disable.

Fixes c43eab3eddb4 ("net: fec: add missed clk_disable_unprepare in remove")

Your explanation in your reply to my feedback still doesn't explain the situation to me.

For every clock enable done during probe, there must be a matching clock disable during remove.

Period.

There is no CONFIG_PM guarding the clock enables during probe in this driver, therefore there should be no reason to require CONFIG_PM guards to the clock disables during the remove method,

You have to explain clearly, and in detail, why my logic and analysis of this situation is not correct.

And when you do so, you will need to add those important details to the commit message of this change and submit a v3.

Thank you.

#4 lore raw

From: David Miller <[email protected]> Sent: Saturday, November 16, 2019 4:11 AM

Quoted 24 lines

From: Chuhong Yuan <[email protected]>
Date: Tue, 12 Nov 2019 19:28:30 +0800

Quoted 8 lines

If CONFIG_PM is enabled, runtime pm will work and call runtime_suspend automatically to disable clks.
Therefore, remove only needs to disable clks when CONFIG_PM is disabled. Add this check to avoid clock count mis-match caused by double-disable.

Fixes c43eab3eddb4 ("net: fec: add missed clk_disable_unprepare in

remove")

Your explanation in your reply to my feedback still doesn't explain the situation to me.

For every clock enable done during probe, there must be a matching clock disable during remove.

Period.

There is no CONFIG_PM guarding the clock enables during probe in this driver, therefore there should be no reason to require CONFIG_PM guards to the clock disables during the remove method,

Quoted 7 lines

You have to explain clearly, and in detail, why my logic and analysis of this situation is not correct.

And when you do so, you will need to add those important details to the commit message of this change and submit a v3.

Thank you.

I agree with David. Below fixes is more reasonable.
Chuhong, if there has no voice about below fixes, you can submit v3 later.

@@ -3636,6 +3636,11 @@ fec_drv_remove(struct platform_device *pdev)

        struct net_device *ndev = platform_get_drvdata(pdev);
        struct fec_enet_private *fep = netdev_priv(ndev);
        struct device_node *np = pdev->dev.of_node;

+ int ret;
+
+ ret = pm_runtime_get_sync(&pdev->dev);
+ if (ret < 0)
+ return ret;

        cancel_work_sync(&fep->tx_timeout_work);
        fec_ptp_stop(pdev);
@@ -3643,15 +3648,17 @@ fec_drv_remove(struct platform_device *pdev)
        fec_enet_mii_remove(fep);
        if (fep->reg_phy)
                regulator_disable(fep->reg_phy);
-       pm_runtime_put(&pdev->dev);
-       pm_runtime_disable(&pdev->dev);
-       clk_disable_unprepare(fep->clk_ahb);
-       clk_disable_unprepare(fep->clk_ipg);
+
        if (of_phy_is_fixed_link(np))
                of_phy_deregister_fixed_link(np);
        of_node_put(fep->phy_node);
        free_netdev(ndev);
+       clk_disable_unprepare(fep->clk_ahb);
+       clk_disable_unprepare(fep->clk_ipg);
+       pm_runtime_put_noidle(&pdev->dev);
+       pm_runtime_disable(&pdev->dev);
+
        return 0;

}

Regards,
Fugang Duan

#5 lore raw
Quoted 29 lines · Andy Duan

From: David Miller <[email protected]> Sent: Saturday, November 16, 2019 4:11 AM

Quoted 27 lines

From: Chuhong Yuan <[email protected]>
Date: Tue, 12 Nov 2019 19:28:30 +0800

Quoted 8 lines

If CONFIG_PM is enabled, runtime pm will work and call runtime_suspend automatically to disable clks.
Therefore, remove only needs to disable clks when CONFIG_PM is disabled. Add this check to avoid clock count mis-match caused by double-disable.

Fixes c43eab3eddb4 ("net: fec: add missed clk_disable_unprepare in

remove")

Your explanation in your reply to my feedback still doesn't explain the situation to me.

For every clock enable done during probe, there must be a matching clock disable during remove.

Period.

There is no CONFIG_PM guarding the clock enables during probe in this driver, therefore there should be no reason to require CONFIG_PM guards to the clock disables during the remove method,

You have to explain clearly, and in detail, why my logic and analysis of this situation is not correct.

Quoted 8 lines
Quoted 4 lines

And when you do so, you will need to add those important details to the commit message of this change and submit a v3.

Thank you.

I agree with David. Below fixes is more reasonable.
Chuhong, if there has no voice about below fixes, you can submit v3 later.

I get confused that how does this work to solve the CONFIG_PM problem. And why do we need to adjust the position of the latter four functions? I need to explain them in the commit message.

Quoted 36 lines
(quoted diff) +11 -4
@@ -3636,6 +3636,11 @@ fec_drv_remove(struct platform_device *pdev)
        struct net_device *ndev = platform_get_drvdata(pdev);
        struct fec_enet_private *fep = netdev_priv(ndev);
        struct device_node *np = pdev->dev.of_node;
+       int ret;
+
+       ret = pm_runtime_get_sync(&pdev->dev);
+       if (ret < 0)
+               return ret;
 
        cancel_work_sync(&fep->tx_timeout_work);
        fec_ptp_stop(pdev);
@@ -3643,15 +3648,17 @@ fec_drv_remove(struct platform_device *pdev)
        fec_enet_mii_remove(fep);
        if (fep->reg_phy)
                regulator_disable(fep->reg_phy);
-       pm_runtime_put(&pdev->dev);
-       pm_runtime_disable(&pdev->dev);
-       clk_disable_unprepare(fep->clk_ahb);
-       clk_disable_unprepare(fep->clk_ipg);
+
        if (of_phy_is_fixed_link(np))
                of_phy_deregister_fixed_link(np);
        of_node_put(fep->phy_node);
        free_netdev(ndev);
 
+       clk_disable_unprepare(fep->clk_ahb);
+       clk_disable_unprepare(fep->clk_ipg);
+       pm_runtime_put_noidle(&pdev->dev);
+       pm_runtime_disable(&pdev->dev);
+
        return 0;
 }
 

Regards,
Fugang Duan

#6 lore raw

From: Chuhong Yuan <[email protected]> Sent: Saturday, November 16, 2019 10:00 PM

Quoted 36 lines · Andy Duan
Quoted 9 lines

From: David Miller <[email protected]> Sent: Saturday, November 16, 2019 4:11 AM

Quoted 6 lines

From: Chuhong Yuan <[email protected]>
Date: Tue, 12 Nov 2019 19:28:30 +0800

Quoted 3 lines

If CONFIG_PM is enabled, runtime pm will work and call runtime_suspend automatically to disable clks. Therefore, remove only needs to disable clks when CONFIG_PM is

disabled.

Quoted 25 lines
Quoted 25 lines
Quoted 6 lines

Add this check to avoid clock count mis-match caused by double-disable.

Fixes c43eab3eddb4 ("net: fec: add missed clk_disable_unprepare

in
remove")

Your explanation in your reply to my feedback still doesn't explain the situation to me.

For every clock enable done during probe, there must be a matching clock disable during remove.

Period.

There is no CONFIG_PM guarding the clock enables during probe in this driver, therefore there should be no reason to require CONFIG_PM guards to the clock disables during the remove method,

You have to explain clearly, and in detail, why my logic and analysis of this situation is not correct.

And when you do so, you will need to add those important details to the commit message of this change and submit a v3.

Quoted 8 lines
Quoted 5 lines
Quoted 1 line

Thank you.

I agree with David. Below fixes is more reasonable.
Chuhong, if there has no voice about below fixes, you can submit v3 later.

I get confused that how does this work to solve the CONFIG_PM problem. And why do we need to adjust the position of the latter four functions?

Just looks better, no function change.

Quoted 1 line

I need to explain them in the commit message.

Please see below logic in remove():
If CONFIG_PM is enabled:
.probe()
enable clks
pm_runtime_mark_last_busy()
pm_runtime_put_autosuspend()
disable clks

.remove():
pm_runtime_get_sync()
runtime resume callback is called, enable clks
disable clks
pm_runtime_put_noidle()
runtime suspend callback is not called

If CONFIG_PM is disabled, runtime pm is NULL operation:
.probe()
enable clks
.remove():
disable clks

Quoted 38 lines

Review fec_drv_remove(struct platform_device

Quoted 14 lines
(quoted diff) +5 -0
@@ -3636,6 +3636,11 @@ fec_drv_remove(struct platform_device *pdev)
        struct net_device *ndev = platform_get_drvdata(pdev);
        struct fec_enet_private *fep = netdev_priv(ndev);
        struct device_node *np = pdev->dev.of_node;
+       int ret;
+
+       ret = pm_runtime_get_sync(&pdev->dev);
+       if (ret < 0)
+               return ret;
 
        cancel_work_sync(&fep->tx_timeout_work);
        fec_ptp_stop(pdev);
@@ -3643,15 +3648,17 @@ fec_drv_remove(struct platform_device

*pdev)

Quoted 23 lines
(quoted diff) +6 -4
        fec_enet_mii_remove(fep);
        if (fep->reg_phy)
                regulator_disable(fep->reg_phy);
-       pm_runtime_put(&pdev->dev);
-       pm_runtime_disable(&pdev->dev);
-       clk_disable_unprepare(fep->clk_ahb);
-       clk_disable_unprepare(fep->clk_ipg);
+
        if (of_phy_is_fixed_link(np))
                of_phy_deregister_fixed_link(np);
        of_node_put(fep->phy_node);
        free_netdev(ndev);
 
+       clk_disable_unprepare(fep->clk_ahb);
+       clk_disable_unprepare(fep->clk_ipg);
+       pm_runtime_put_noidle(&pdev->dev);
+       pm_runtime_disable(&pdev->dev);
+
        return 0;
 }
 

Regards,
Fugang Duan