Skip to contentSkip to conversation

net: phy: clear phydev->devlink when the link is deleted

PATCH Applied Wei Fang 7 participants 13 messages May 23, 2025 – Jun 6, 2025

Subsystems (4)
lore
Share & tools
lore

Wei Fang

author
#0 lore raw

PATCH v2 1/1 net: phy: clear phydev->devlink when the link is deleted

There is a potential crash issue when disabling and re-enabling the network port. When disabling the network port, phy_detach() calls device_link_del() to remove the device link, but it does not clear phydev->devlink, so phydev->devlink is not a NULL pointer. Then the network port is re-enabled, but if phy_attach_direct() fails before calling device_link_add(), the code jumps to the "error" label and calls phy_detach(). Since phydev->devlink retains the old value from the previous attach/detach cycle, device_link_del() uses the old value, which accesses a NULL pointer and causes a crash. The simplified crash log is as follows.

[   24.702421] Call trace:
[   24.704856]  device_link_put_kref+0x20/0x120
[   24.709124]  device_link_del+0x30/0x48
[   24.712864]  phy_detach+0x24/0x168
[   24.716261]  phy_attach_direct+0x168/0x3a4
[   24.720352]  phylink_fwnode_phy_connect+0xc8/0x14c
[   24.725140]  phylink_of_phy_connect+0x1c/0x34

Therefore, phydev->devlink needs to be cleared when the device link is deleted.

Fixes bc66fa87d4fd ("net: phy: Add link between phy dev and mac dev")

---
v2: Improve the commit message.

---
 drivers/net/phy/phy_device.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)
drivers/net/phy/phy_device.c +3 -1
index cc1bfd22fb81..7d5e76a3db0e 100644
@@ -1727,8 +1727,10 @@ void phy_detach(struct phy_device *phydev)
 	struct module *ndev_owner = NULL;
 	struct mii_bus *bus;
 
-	if (phydev->devlink)
+	if (phydev->devlink) {
 		device_link_del(phydev->devlink);
+		phydev->devlink = NULL;
+	}
 
 	if (phydev->sysfs_links) {
 		if (dev)
Signature · 1 line
2.34.1
maintainer Reviewed-by on v2 1/1
#1 lore raw
Quoted 19 lines · Wei Fang

There is a potential crash issue when disabling and re-enabling the network port. When disabling the network port, phy_detach() calls device_link_del() to remove the device link, but it does not clear phydev->devlink, so phydev->devlink is not a NULL pointer. Then the network port is re-enabled, but if phy_attach_direct() fails before calling device_link_add(), the code jumps to the "error" label and calls phy_detach(). Since phydev->devlink retains the old value from the previous attach/detach cycle, device_link_del() uses the old value, which accesses a NULL pointer and causes a crash. The simplified crash log is as follows.

[   24.702421] Call trace:
[   24.704856]  device_link_put_kref+0x20/0x120
[   24.709124]  device_link_del+0x30/0x48
[   24.712864]  phy_detach+0x24/0x168
[   24.716261]  phy_attach_direct+0x168/0x3a4
[   24.720352]  phylink_fwnode_phy_connect+0xc8/0x14c
[   24.725140]  phylink_of_phy_connect+0x1c/0x34
Quoted 5 lines

Therefore, phydev->devlink needs to be cleared when the device link is deleted.

Fixes bc66fa87d4fd ("net: phy: Add link between phy dev and mac dev")

Andrew

Reviewed-by on v2 1/1
#2 lore raw
In reply to Wei Fang
Quoted 19 lines · Wei Fang

There is a potential crash issue when disabling and re-enabling the network port. When disabling the network port, phy_detach() calls device_link_del() to remove the device link, but it does not clear phydev->devlink, so phydev->devlink is not a NULL pointer. Then the network port is re-enabled, but if phy_attach_direct() fails before calling device_link_add(), the code jumps to the "error" label and calls phy_detach(). Since phydev->devlink retains the old value from the previous attach/detach cycle, device_link_del() uses the old value, which accesses a NULL pointer and causes a crash. The simplified crash log is as follows.

[   24.702421] Call trace:
[   24.704856]  device_link_put_kref+0x20/0x120
[   24.709124]  device_link_del+0x30/0x48
[   24.712864]  phy_detach+0x24/0x168
[   24.716261]  phy_attach_direct+0x168/0x3a4
[   24.720352]  phylink_fwnode_phy_connect+0xc8/0x14c
[   24.725140]  phylink_of_phy_connect+0x1c/0x34
Quoted 5 lines

Therefore, phydev->devlink needs to be cleared when the device link is deleted.

Fixes bc66fa87d4fd ("net: phy: Add link between phy dev and mac dev")
Signature · 1 line
Florian
#3 lore raw
Quoted 22 lines · Florian Fainelli
Quoted 19 lines · Wei Fang

There is a potential crash issue when disabling and re-enabling the network port. When disabling the network port, phy_detach() calls device_link_del() to remove the device link, but it does not clear phydev->devlink, so phydev->devlink is not a NULL pointer. Then the network port is re-enabled, but if phy_attach_direct() fails before calling device_link_add(), the code jumps to the "error" label and calls phy_detach(). Since phydev->devlink retains the old value from the previous attach/detach cycle, device_link_del() uses the old value, which accesses a NULL pointer and causes a crash. The simplified crash log is as follows.

[   24.702421] Call trace:
[   24.704856]  device_link_put_kref+0x20/0x120 [   24.709124]  device_link_del+0x30/0x48
[   24.712864]  phy_detach+0x24/0x168
[   24.716261]  phy_attach_direct+0x168/0x3a4 [   24.720352]  phylink_fwnode_phy_connect+0xc8/0x14c [   24.725140]  phylink_of_phy_connect+0x1c/0x34

Quoted 6 lines
Quoted 5 lines

Therefore, phydev->devlink needs to be cleared when the device link is deleted.

Fixes bc66fa87d4fd ("net: phy: Add link between phy dev and mac dev")

@Wei
What happens in case of shared mdio ?

1. Device 23040000 has the mdio node of both the ethernet phy and device 23000000 references the phy-handle present in the Device 23040000
2. When rmmod of the driver happens
3. the parent devlink is already deleted.
4. This cause the child mdio to access an entry causing a corruption.
5. Thought this fix would help but i see that its not helping the case.

Wondering if this is a legacy issue with shared mdio framework.

 43369.232799:   qcom-ethqos 23040000.ethernet eth1: stmmac_dvr_remove: removing driver
 43369.233782:   stmmac_pcs: Link Down
 43369.258337:   qcom-ethqos 23040000.ethernet eth1: FPE workqueue stop
 43369.258522:   br1: port 1(eth1) entered disabled state
 43369.758779:   qcom-ethqos 23040000.ethernet eth1: Timeout accessing MAC_VLAN_Tag_Filter
 43369.758789:   qcom-ethqos 23040000.ethernet eth1: failed to kill vid 0081/0
 43369.759270:   qcom-ethqos 23040000.ethernet eth1 (unregistering): left allmulticast mode
 43369.759275:   qcom-ethqos 23040000.ethernet eth1 (unregistering): left promiscuous mode
 43369.759301:   br1: port 1(eth1) entered disabled state
 43370.259618:   qcom-ethqos 23040000.ethernet eth1 (unregistering): Timeout accessing MAC_VLAN_Tag_Filter
 43370.309863:   qcom-ethqos 23000000.ethernet eth0: stmmac_dvr_remove: removing driver
 43370.310019:   list_del corruption, ffffff80c6ec9408->prev is LIST_POISON2 (dead000000000122)
 43370.310034:   ------------[ cut here ]------------
 43370.310035:   kernel BUG at lib/list_debug.c:59!
 43370.310119:   CPU: 4 PID: 3067767 Comm: rmmod Tainted: G        W  OE      6.6.65-rt47-debug #1
 43370.310122:   Hardware name: Qualcomm Technologies, Inc. SA8775P Ride (DT)
 43370.310165:   Call trace:
 43370.310166:    __list_del_entry_valid_or_report+0xa8/0xe0
 43370.310168:    __device_link_del+0x40/0xf0
 43370.310172:    device_link_put_kref+0xb4/0xc8
 43370.310174:    device_link_del+0x38/0x58
 43370.310176:    phy_detach+0x2c/0x170
 43370.310181:    phy_disconnect+0x4c/0x70
 43370.310184:    phylink_disconnect_phy+0x6c/0xc0 [phylink]
 43370.310194:    stmmac_release+0x60/0x358 [stmmac]
 43370.310210:    __dev_close_many+0xb4/0x160
 43370.310213:    dev_close_many+0xbc/0x1a0
 43370.310215:    unregister_netdevice_many_notify+0x178/0x870
 43370.310218:    unregister_netdevice_queue+0xf8/0x140
 43370.310221:    unregister_netdev+0x2c/0x48
 43370.310223:    stmmac_dvr_remove+0xd0/0x1b0 [stmmac]
 43370.310233:    devm_stmmac_pltfr_remove+0x2c/0x58 [stmmac_platform]
Quoted 1 line
#4 lore raw
Quoted 34 lines · Abhishek Chauhan (ABC)
Quoted 22 lines · Florian Fainelli
Quoted 19 lines · Wei Fang

There is a potential crash issue when disabling and re-enabling the network port. When disabling the network port, phy_detach() calls device_link_del() to remove the device link, but it does not clear phydev->devlink, so phydev->devlink is not a NULL pointer. Then the network port is re-enabled, but if phy_attach_direct() fails before calling device_link_add(), the code jumps to the "error" label and calls phy_detach(). Since phydev->devlink retains the old value from the previous attach/detach cycle, device_link_del() uses the old value, which accesses a NULL pointer and causes a crash. The simplified crash log is as follows.

[   24.702421] Call trace:
[   24.704856]  device_link_put_kref+0x20/0x120 [   24.709124]  device_link_del+0x30/0x48
[   24.712864]  phy_detach+0x24/0x168
[   24.716261]  phy_attach_direct+0x168/0x3a4 [   24.720352]  phylink_fwnode_phy_connect+0xc8/0x14c [   24.725140]  phylink_of_phy_connect+0x1c/0x34

Quoted 6 lines
Quoted 5 lines

Therefore, phydev->devlink needs to be cleared when the device link is deleted.

Fixes bc66fa87d4fd ("net: phy: Add link between phy dev and mac dev")

@Wei
What happens in case of shared mdio ?

Quoted 7 lines

1. Device 23040000 has the mdio node of both the ethernet phy and device 23000000 references the phy-handle present in the Device 23040000
2. When rmmod of the driver happens
3. the parent devlink is already deleted.
4. This cause the child mdio to access an entry causing a corruption.
5. Thought this fix would help but i see that its not helping the case.

Wondering if this is a legacy issue with shared mdio framework.

The device link does nothing for this as it has DL_FLAG_STATELESS set, which only affects suspend/resume/shutdown ordering, and with DL_FLAG_PM_RUNTIME also set, runtime PM.

The device probe/removal ordering is unaffected. Maybe that's a problem, but it needs careful consideration to change.

Signature · 2 lines
RMK's Patch system: https://www.armlinux.org.uk/developer/patches/
FTTP is here! 80Mbps down 10Mbps up. Decent connectivity at last!
author
#5 lore raw
In reply to Abhishek Chauhan (ABC)
Quoted 29 lines
Quoted 20 lines · Wei Fang
Quoted 19 lines

There is a potential crash issue when disabling and re-enabling the network port. When disabling the network port, phy_detach() calls device_link_del() to remove the device link, but it does not clear phydev->devlink, so phydev->devlink is not a NULL pointer. Then the network port is re-enabled, but if phy_attach_direct() fails before calling device_link_add(), the code jumps to the "error" label and calls phy_detach(). Since phydev->devlink retains the old value from the previous attach/detach cycle, device_link_del() uses the old value, which accesses a NULL pointer and causes a crash. The simplified crash log is as follows.

[   24.702421] Call trace:
[   24.704856]  device_link_put_kref+0x20/0x120 [   24.709124]  device_link_del+0x30/0x48
[   24.712864]  phy_detach+0x24/0x168
[   24.716261]  phy_attach_direct+0x168/0x3a4 [   24.720352]  phylink_fwnode_phy_connect+0xc8/0x14c [   24.725140]  phylink_of_phy_connect+0x1c/0x34

Quoted 6 lines
Quoted 5 lines

Therefore, phydev->devlink needs to be cleared when the device link is deleted.

Fixes bc66fa87d4fd ("net: phy: Add link between phy dev and mac dev")

@Wei
What happens in case of shared mdio ?

Quoted 7 lines

1. Device 23040000 has the mdio node of both the ethernet phy and device 23000000 references the phy-handle present in the Device 23040000
2. When rmmod of the driver happens
3. the parent devlink is already deleted.
4. This cause the child mdio to access an entry causing a corruption.
5. Thought this fix would help but i see that its not helping the case.

My patch is only to fix the potential crash issue when re-enabling the network interface. phy_detach() is not called when the MDIO controller driver is removed. So phydev->devlink is not cleared, but actually the device link has been removed by phy_device_remove() --> device_del(). Therefore, it will cause the crash when the MAC controller driver is removed.

Quoted 2 lines

Wondering if this is a legacy issue with shared mdio framework.

I think this issue is also introduced by the commit bc66fa87d4fd ("net: phy: Add link between phy dev and mac dev"). I suggested to change the DL_FLAG_STATELESS flag to DL_FLAG_AUTOREMOVE_SUPPLIER to solve this issue, so that the consumer (MAC controller) driver will be automatically removed when the link is removed. The changes are as follows.

drivers/net/phy/phy_device.c +11 -8
index 73f9cb2e2844..a6d7acd73391 100644
@@ -1515,6 +1515,7 @@ int phy_attach_direct(struct net_device *dev, struct phy_device *phydev,
        struct mii_bus *bus = phydev->mdio.bus;
        struct device *d = &phydev->mdio.dev;
        struct module *ndev_owner = NULL;
+       struct device_link *devlink;
        bool using_genphy = false;
        int err;
 
@@ -1646,9 +1647,16 @@ int phy_attach_direct(struct net_device *dev, struct phy_device *phydev,
         * another mac interface, so we should create a device link between
         * phy dev and mac dev.
         */
-       if (dev && phydev->mdio.bus->parent && dev->dev.parent != phydev->mdio.bus->parent)
-               phydev->devlink = device_link_add(dev->dev.parent, &phydev->mdio.dev,
-                                                 DL_FLAG_PM_RUNTIME | DL_FLAG_STATELESS);
+       if (dev && phydev->mdio.bus->parent &&
+           dev->dev.parent != phydev->mdio.bus->parent) {
+               devlink = device_link_add(dev->dev.parent, &phydev->mdio.dev,
+                                         DL_FLAG_PM_RUNTIME |
+                                         DL_FLAG_AUTOREMOVE_SUPPLIER);
+               if (!devlink) {
+                       err = -ENOMEM;
+                       goto error;
+               }
+       }
 
        return err;
 
@@ -1749,11 +1757,6 @@ void phy_detach(struct phy_device *phydev)
        struct module *ndev_owner = NULL;
        struct mii_bus *bus;
 
-       if (phydev->devlink) {
-               device_link_del(phydev->devlink);
-               phydev->devlink = NULL;
-       }
-
        if (phydev->sysfs_links) {
                if (dev)
                        sysfs_remove_link(&dev->dev.kobj, "phydev");
include/linux/phy.h +0 -4
index e194dad1623d..cc1f45c3ff21 100644
@@ -505,8 +505,6 @@ struct macsec_ops;
  *
  * @mdio: MDIO bus this PHY is on
  * @drv: Pointer to the driver for this PHY instance
- * @devlink: Create a link between phy dev and mac dev, if the external phy
- *           used by current mac interface is managed by another mac interface.
  * @phyindex: Unique id across the phy's parent tree of phys to address the PHY
  *           from userspace, similar to ifindex. A zero index means the PHY
  *           wasn't assigned an id yet.
@@ -610,8 +608,6 @@ struct phy_device {
        /* And management functions */
        const struct phy_driver *drv;
 
-       struct device_link *devlink;
-
        u32 phyindex;
        u32 phy_id;
 
#6 lore raw
Quoted 47 lines · Wei Fang
Quoted 29 lines
Quoted 20 lines · Wei Fang
Quoted 19 lines

There is a potential crash issue when disabling and re-enabling the network port. When disabling the network port, phy_detach() calls device_link_del() to remove the device link, but it does not clear phydev->devlink, so phydev->devlink is not a NULL pointer. Then the network port is re-enabled, but if phy_attach_direct() fails before calling device_link_add(), the code jumps to the "error" label and calls phy_detach(). Since phydev->devlink retains the old value from the previous attach/detach cycle, device_link_del() uses the old value, which accesses a NULL pointer and causes a crash. The simplified crash log is as follows.

[   24.702421] Call trace:
[   24.704856]  device_link_put_kref+0x20/0x120 [   24.709124]  device_link_del+0x30/0x48
[   24.712864]  phy_detach+0x24/0x168
[   24.716261]  phy_attach_direct+0x168/0x3a4 [   24.720352]  phylink_fwnode_phy_connect+0xc8/0x14c [   24.725140]  phylink_of_phy_connect+0x1c/0x34

Quoted 6 lines
Quoted 5 lines

Therefore, phydev->devlink needs to be cleared when the device link is deleted.

Fixes bc66fa87d4fd ("net: phy: Add link between phy dev and mac dev")

@Wei
What happens in case of shared mdio ?

Quoted 7 lines

1. Device 23040000 has the mdio node of both the ethernet phy and device 23000000 references the phy-handle present in the Device 23040000
2. When rmmod of the driver happens
3. the parent devlink is already deleted.
4. This cause the child mdio to access an entry causing a corruption.
5. Thought this fix would help but i see that its not helping the case.

My patch is only to fix the potential crash issue when re-enabling the network interface. phy_detach() is not called when the MDIO controller driver is removed. So phydev->devlink is not cleared, but actually the device link has been removed by phy_device_remove() --> device_del(). Therefore, it will cause the crash when the MAC controller driver is removed.

Quoted 2 lines

Wondering if this is a legacy issue with shared mdio framework.

Quoted 7 lines

I think this issue is also introduced by the commit bc66fa87d4fd ("net: phy: Add link between phy dev and mac dev"). I suggested to change the DL_FLAG_STATELESS flag to DL_FLAG_AUTOREMOVE_SUPPLIER to solve this issue, so that the consumer (MAC controller) driver will be automatically removed when the link is removed. The changes are as follows.

thanks a lot , Russell and Wei for your prompt response. I appreciate your help. let me test this change and get back.

Quoted 66 lines
drivers/net/phy/phy_device.c +11 -8
index 73f9cb2e2844..a6d7acd73391 100644
@@ -1515,6 +1515,7 @@ int phy_attach_direct(struct net_device *dev, struct phy_device *phydev,
        struct mii_bus *bus = phydev->mdio.bus;
        struct device *d = &phydev->mdio.dev;
        struct module *ndev_owner = NULL;
+       struct device_link *devlink;
        bool using_genphy = false;
        int err;
 
@@ -1646,9 +1647,16 @@ int phy_attach_direct(struct net_device *dev, struct phy_device *phydev,
         * another mac interface, so we should create a device link between
         * phy dev and mac dev.
         */
-       if (dev && phydev->mdio.bus->parent && dev->dev.parent != phydev->mdio.bus->parent)
-               phydev->devlink = device_link_add(dev->dev.parent, &phydev->mdio.dev,
-                                                 DL_FLAG_PM_RUNTIME | DL_FLAG_STATELESS);
+       if (dev && phydev->mdio.bus->parent &&
+           dev->dev.parent != phydev->mdio.bus->parent) {
+               devlink = device_link_add(dev->dev.parent, &phydev->mdio.dev,
+                                         DL_FLAG_PM_RUNTIME |
+                                         DL_FLAG_AUTOREMOVE_SUPPLIER);
+               if (!devlink) {
+                       err = -ENOMEM;
+                       goto error;
+               }
+       }
 
        return err;
 
@@ -1749,11 +1757,6 @@ void phy_detach(struct phy_device *phydev)
        struct module *ndev_owner = NULL;
        struct mii_bus *bus;
 
-       if (phydev->devlink) {
-               device_link_del(phydev->devlink);
-               phydev->devlink = NULL;
-       }
-
        if (phydev->sysfs_links) {
                if (dev)
                        sysfs_remove_link(&dev->dev.kobj, "phydev");
include/linux/phy.h +0 -4
index e194dad1623d..cc1f45c3ff21 100644
@@ -505,8 +505,6 @@ struct macsec_ops;
  *
  * @mdio: MDIO bus this PHY is on
  * @drv: Pointer to the driver for this PHY instance
- * @devlink: Create a link between phy dev and mac dev, if the external phy
- *           used by current mac interface is managed by another mac interface.
  * @phyindex: Unique id across the phy's parent tree of phys to address the PHY
  *           from userspace, similar to ifindex. A zero index means the PHY
  *           wasn't assigned an id yet.
@@ -610,8 +608,6 @@ struct phy_device {
        /* And management functions */
        const struct phy_driver *drv;
 
-       struct device_link *devlink;
-
        u32 phyindex;
        u32 phy_id;
 
#7 lore raw
Quoted 58 lines · Abhishek Chauhan (ABC)
Quoted 47 lines · Wei Fang
Quoted 29 lines
Quoted 20 lines · Wei Fang
Quoted 19 lines

There is a potential crash issue when disabling and re-enabling the network port. When disabling the network port, phy_detach() calls device_link_del() to remove the device link, but it does not clear phydev->devlink, so phydev->devlink is not a NULL pointer. Then the network port is re-enabled, but if phy_attach_direct() fails before calling device_link_add(), the code jumps to the "error" label and calls phy_detach(). Since phydev->devlink retains the old value from the previous attach/detach cycle, device_link_del() uses the old value, which accesses a NULL pointer and causes a crash. The simplified crash log is as follows.

[   24.702421] Call trace:
[   24.704856]  device_link_put_kref+0x20/0x120 [   24.709124]  device_link_del+0x30/0x48
[   24.712864]  phy_detach+0x24/0x168
[   24.716261]  phy_attach_direct+0x168/0x3a4 [   24.720352]  phylink_fwnode_phy_connect+0xc8/0x14c [   24.725140]  phylink_of_phy_connect+0x1c/0x34

Quoted 6 lines
Quoted 5 lines

Therefore, phydev->devlink needs to be cleared when the device link is deleted.

Fixes bc66fa87d4fd ("net: phy: Add link between phy dev and mac dev")

@Wei
What happens in case of shared mdio ?

Quoted 7 lines

1. Device 23040000 has the mdio node of both the ethernet phy and device 23000000 references the phy-handle present in the Device 23040000
2. When rmmod of the driver happens
3. the parent devlink is already deleted.
4. This cause the child mdio to access an entry causing a corruption.
5. Thought this fix would help but i see that its not helping the case.

My patch is only to fix the potential crash issue when re-enabling the network interface. phy_detach() is not called when the MDIO controller driver is removed. So phydev->devlink is not cleared, but actually the device link has been removed by phy_device_remove() --> device_del(). Therefore, it will cause the crash when the MAC controller driver is removed.

Quoted 2 lines

Wondering if this is a legacy issue with shared mdio framework.

Quoted 7 lines

I think this issue is also introduced by the commit bc66fa87d4fd ("net: phy: Add link between phy dev and mac dev"). I suggested to change the DL_FLAG_STATELESS flag to DL_FLAG_AUTOREMOVE_SUPPLIER to solve this issue, so that the consumer (MAC controller) driver will be automatically removed when the link is removed. The changes are as follows.

Quoted 3 lines

thanks a lot , Russell and Wei for your prompt response. I appreciate your help. let me test this change and get back.

Myself and Abhishek tested this suggestion from Wei and we now observe that the driver removal is successful without a crash. It looks like the flag 'DL_FLAG_AUTOREMOVE_SUPPLIER' is helping. Wei and Russell, please suggest the next steps.

Quoted 1 line

Review

Quoted 1 line
(quoted diff) +0 -0

b/drivers/net/phy/phy_device.c

Quoted 65 lines
Quoted 65 lines

index 73f9cb2e2844..a6d7acd73391 100644

drivers/net/phy/phy_device.c +11 -8
@@ -1515,6 +1515,7 @@ int phy_attach_direct(struct net_device *dev, struct phy_device *phydev,
         struct mii_bus *bus = phydev->mdio.bus;
         struct device *d = &phydev->mdio.dev;
         struct module *ndev_owner = NULL;
+       struct device_link *devlink;
         bool using_genphy = false;
         int err;
 
@@ -1646,9 +1647,16 @@ int phy_attach_direct(struct net_device *dev, struct phy_device *phydev,
          * another mac interface, so we should create a device link between
          * phy dev and mac dev.
          */
-       if (dev && phydev->mdio.bus->parent && dev->dev.parent != phydev->mdio.bus->parent)
-               phydev->devlink = device_link_add(dev->dev.parent, &phydev->mdio.dev,
-                                                 DL_FLAG_PM_RUNTIME | DL_FLAG_STATELESS);
+       if (dev && phydev->mdio.bus->parent &&
+           dev->dev.parent != phydev->mdio.bus->parent) {
+               devlink = device_link_add(dev->dev.parent, &phydev->mdio.dev,
+                                         DL_FLAG_PM_RUNTIME |
+                                         DL_FLAG_AUTOREMOVE_SUPPLIER);
+               if (!devlink) {
+                       err = -ENOMEM;
+                       goto error;
+               }
+       }
 
         return err;
 
@@ -1749,11 +1757,6 @@ void phy_detach(struct phy_device *phydev)
         struct module *ndev_owner = NULL;
         struct mii_bus *bus;
 
-       if (phydev->devlink) {
-               device_link_del(phydev->devlink);
-               phydev->devlink = NULL;
-       }
-
         if (phydev->sysfs_links) {
                 if (dev)
                         sysfs_remove_link(&dev->dev.kobj, "phydev");
include/linux/phy.h +0 -4
index e194dad1623d..cc1f45c3ff21 100644
@@ -505,8 +505,6 @@ struct macsec_ops;
   *
   * @mdio: MDIO bus this PHY is on
   * @drv: Pointer to the driver for this PHY instance
- * @devlink: Create a link between phy dev and mac dev, if the external phy
- *           used by current mac interface is managed by another mac interface.
   * @phyindex: Unique id across the phy's parent tree of phys to address the PHY
   *           from userspace, similar to ifindex. A zero index means the PHY
   *           wasn't assigned an id yet.
@@ -610,8 +608,6 @@ struct phy_device {
         /* And management functions */
         const struct phy_driver *drv;
 
-       struct device_link *devlink;
-
         u32 phyindex;
         u32 phy_id;
 
#8 lore raw
In reply to Wei Fang
Quoted 6 lines · Wei Fang

I think this issue is also introduced by the commit bc66fa87d4fd ("net: phy: Add link between phy dev and mac dev"). I suggested to change the DL_FLAG_STATELESS flag to DL_FLAG_AUTOREMOVE_SUPPLIER to solve this issue, so that the consumer (MAC controller) driver will be automatically removed when the link is removed. The changes are as follows.

I suspect this still has problems. This is fine if the PHY device is going away and as you say device_del() is called.

However, you need to consider the case where a MAC driver attaches the PHY during .ndo_open and releases it during .ndo_release. These will happen multiple times.

Each time the MAC driver attaches to the PHY via .ndo_open, we will call device_link_add(), but the device link will not be removed when .ndo_release is called.

Either device_link_add() will fail, or we will eat memory each time the device is closed and re-opened.

If that is correct, then we're trading one problem for another.

Signature · 2 lines
RMK's Patch system: https://www.armlinux.org.uk/developer/patches/
FTTP is here! 80Mbps down 10Mbps up. Decent connectivity at last!
author
#9 lore raw
Quoted 8 lines · Wei Fang
Quoted 6 lines

I think this issue is also introduced by the commit bc66fa87d4fd ("net: phy: Add link between phy dev and mac dev"). I suggested to change the DL_FLAG_STATELESS flag to DL_FLAG_AUTOREMOVE_SUPPLIER to solve this issue, so that the consumer (MAC controller) driver will be automatically removed when the link is removed. The changes are as follows.

Quoted 6 lines

I suspect this still has problems. This is fine if the PHY device is going away and as you say device_del() is called.

However, you need to consider the case where a MAC driver attaches the PHY during .ndo_open and releases it during .ndo_release. These will happen multiple times.

.ndo_release? Do you mean .ndo_stop?

Quoted 7 lines

Each time the MAC driver attaches to the PHY via .ndo_open, we will call device_link_add(), but the device link will not be removed when .ndo_release is called.

Either device_link_add() will fail, or we will eat memory each time the device is closed and re-opened.

Below is what I find in the kernel doc of device_link_add().
https://elixir.bootlin.com/linux/v6.15/source/drivers/base/core.c#L711

if a device link between the given @consumer and @supplier pair exists already when this function is called for them, the existing link will be returned regardless of its current type and status.

Therefore, it will not create new link each time the netdev is re-opened.

Quoted 3 lines

If that is correct, then we're trading one problem for another.

#10 lore raw
Quoted 25 lines · Wei Fang
Quoted 8 lines · Wei Fang
Quoted 6 lines

I think this issue is also introduced by the commit bc66fa87d4fd ("net: phy: Add link between phy dev and mac dev"). I suggested to change the DL_FLAG_STATELESS flag to DL_FLAG_AUTOREMOVE_SUPPLIER to solve this issue, so that the consumer (MAC controller) driver will be automatically removed when the link is removed. The changes are as follows.

Quoted 6 lines

I suspect this still has problems. This is fine if the PHY device is going away and as you say device_del() is called.

However, you need to consider the case where a MAC driver attaches the PHY during .ndo_open and releases it during .ndo_release. These will happen multiple times.

.ndo_release? Do you mean .ndo_stop?

Quoted 7 lines

Each time the MAC driver attaches to the PHY via .ndo_open, we will call device_link_add(), but the device link will not be removed when .ndo_release is called.

Either device_link_add() will fail, or we will eat memory each time the device is closed and re-opened.

Quoted 9 lines

Below is what I find in the kernel doc of device_link_add().
https://elixir.bootlin.com/linux/v6.15/source/drivers/base/core.c#L711

if a device link between the given @consumer and @supplier pair exists already when this function is called for them, the existing link will be returned regardless of its current type and status.

Therefore, it will not create new link each time the netdev is re-opened.

@Wei,

We were able to verify the suggestion what you gave us. No crash is seen now.
Should we raise a patch or shall we wait until this is clarified and Russell has no more further questions. ?

I am okay with anything.

Quoted 4 lines
Quoted 3 lines

If that is correct, then we're trading one problem for another.

author
#11 lore raw

Best Regards,
Wei Fang

Quoted 47 lines
Quoted 22 lines
Quoted 8 lines · Wei Fang
Quoted 6 lines

I think this issue is also introduced by the commit bc66fa87d4fd ("net: phy: Add link between phy dev and mac dev"). I suggested to change the DL_FLAG_STATELESS flag to DL_FLAG_AUTOREMOVE_SUPPLIER to solve this issue, so that the consumer (MAC controller) driver will be automatically removed when the link is removed. The changes are as follows.

Quoted 6 lines

I suspect this still has problems. This is fine if the PHY device is going away and as you say device_del() is called.

However, you need to consider the case where a MAC driver attaches the PHY during .ndo_open and releases it during .ndo_release. These will happen multiple times.

.ndo_release? Do you mean .ndo_stop?

Quoted 5 lines

Each time the MAC driver attaches to the PHY via .ndo_open, we will call device_link_add(), but the device link will not be removed when .ndo_release is called.

Quoted 6 lines
Quoted 2 lines

Either device_link_add() will fail, or we will eat memory each time the device is closed and re-opened.

Below is what I find in the kernel doc of device_link_add().
https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Felix

ir.bootlin.com%2Flinux%2Fv6.15%2Fsource%2Fdrivers%2Fbase%2Fcore.c%23L7

Quoted 1 line

11&data=05%7C02%7Cwei.fang%40nxp.com%7C078b59e45bd54f13cfea08dda 49477a

Quoted 1 line

4%7C686ea1d3bc2b4c6fa92cd99c5c301635%7C0%7C0%7C638847679768356 722%7CUn

Quoted 1 line

known%7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCI sIlAiOi

Quoted 1 line

JXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ%3D%3D%7C0%7C%7C%7C&sdata=iFtK o%2F

Quoted 6 lines

NfwPVYbxLBtZM8Y15Aa9eomUNhPTRfnEb0L7c%3D&reserved=0

if a device link between the given @consumer and @supplier pair exists already when this function is called for them, the existing link will be returned regardless of its current type and status.

Quoted 9 lines
Quoted 2 lines

Therefore, it will not create new link each time the netdev is re-opened.

@Wei,

We were able to verify the suggestion what you gave us.
No crash is seen now.
Should we raise a patch or shall we wait until this is clarified and Russell has no more further questions. ?

I'm okay, but I don't know if Russel has any other concerns. And if there are no more comments, you can raise a patch. :)

Quoted 6 lines

I am okay with anything.

Quoted 4 lines
Quoted 3 lines

If that is correct, then we're trading one problem for another.

#12 lore raw
In reply to Wei Fang
This patch was applied to netdev/net.git (main) by Jakub Kicinski <[email protected]>:

Hello:

This patch was applied to netdev/net.git (main)
by Jakub Kicinski <[email protected]>:

Quoted 11 lines · On Fri, 23 May 2025 16:37:59 +0800 you wrote:

There is a potential crash issue when disabling and re-enabling the network port. When disabling the network port, phy_detach() calls device_link_del() to remove the device link, but it does not clear phydev->devlink, so phydev->devlink is not a NULL pointer. Then the network port is re-enabled, but if phy_attach_direct() fails before calling device_link_add(), the code jumps to the "error" label and calls phy_detach(). Since phydev->devlink retains the old value from the previous attach/detach cycle, device_link_del() uses the old value, which accesses a NULL pointer and causes a crash. The simplified crash log is as follows.

Quoted 1 line

[...]

Here is the summary with links:
- [v2,net] net: phy: clear phydev->devlink when the link is deleted

    https://git.kernel.org/netdev/net/c/0795b05a59b1

You are awesome, thank you!

Signature · 2 lines
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html