Skip to contentSkip to conversation

acpi/hmat: Fix lockdep warning for hmem_register_resource()

PATCH In review Dave Jiang 2 participants 6 messages Oct 15, 2025 – Oct 16, 2025

Subsystems (2)
lore
Share & tools
lore

Dave Jiang

author maintainer
#0 lore raw

PATCH v2 1/1 acpi/hmat: Fix lockdep warning for hmem_register_resource()

The following lockdep splat was observed while kernel auto-online a CXL memory region:

======================================================
WARNING: possible circular locking dependency detected
6.17.0djtest+ #53 Tainted: G W
------------------------------------------------------
systemd-udevd/3334 is trying to acquire lock:
ffffffff90346188 (hmem_resource_lock){+.+.}-{4:4}, at: hmem_register_resource+0x31/0x50

but task is already holding lock:
ffffffff90338890 ((node_chain).rwsem){++++}-{4:4}, at: blocking_notifier_call_chain+0x2e/0x70

which lock already depends on the new lock.
[..]

Chain exists of:
  hmem_resource_lock --> mem_hotplug_lock --> (node_chain).rwsem

 Possible unsafe locking scenario:
       CPU0                    CPU1
       ----                    ----
  rlock((node_chain).rwsem);
                               lock(mem_hotplug_lock);
                               lock((node_chain).rwsem);

lock(hmem_resource_lock);

The lock ordering can cause potential deadlock. There are instances where hmem_resource_lock is taken after (node_chain).rwsem, and vice versa.

Remove registering of target devices from the hmat_callback(). By the time the hmat hotplug notifier is being called, there should not be hmem targets that still need to be registered.

Fixes cf8741ac57ed ("ACPI: NUMA: HMAT: Register "soft reserved" memory as an "hmem" device")

---
v2:
- Drop target registering in hmat_callback instead. (Dan)

---
 drivers/acpi/numa/hmat.c | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)
drivers/acpi/numa/hmat.c +6 -4
index 5a36d57289b4..5084ae1688f6 100644
@@ -874,7 +874,8 @@ static void hmat_register_target_devices(struct memory_target *target)
 	}
 }
 
-static void hmat_register_target(struct memory_target *target)
+static void hmat_register_target(struct memory_target *target,
+				 bool register_devices)
 {
 	int nid = pxm_to_node(target->memory_pxm);
 
@@ -882,7 +883,8 @@ static void hmat_register_target(struct memory_target *target)
 	 * Devices may belong to either an offline or online
 	 * node, so unconditionally add them.
 	 */
-	hmat_register_target_devices(target);
+	if (register_devices)
+		hmat_register_target_devices(target);
 
 	/*
 	 * Register generic port perf numbers. The nid may not be
@@ -921,7 +923,7 @@ static void hmat_register_targets(void)
 	struct memory_target *target;
 
 	list_for_each_entry(target, &targets, node)
-		hmat_register_target(target);
+		hmat_register_target(target, true);
 }
 
 static int hmat_callback(struct notifier_block *self,
@@ -939,7 +941,7 @@ static int hmat_callback(struct notifier_block *self,
 	if (!target)
 		return NOTIFY_OK;
 
-	hmat_register_target(target);
+	hmat_register_target(target, false);
 	return NOTIFY_OK;
 }
 

base-commit: 3a8660878839faadb4f1a6dd72c3179c1df56787

Signature · 1 line
2.51.0
#1 lore raw
Quoted 61 lines · Dave Jiang

The following lockdep splat was observed while kernel auto-online a CXL memory region:

======================================================
WARNING: possible circular locking dependency detected
6.17.0djtest+ #53 Tainted: G W
------------------------------------------------------
systemd-udevd/3334 is trying to acquire lock:
ffffffff90346188 (hmem_resource_lock){+.+.}-{4:4}, at: hmem_register_resource+0x31/0x50

but task is already holding lock:
ffffffff90338890 ((node_chain).rwsem){++++}-{4:4}, at: blocking_notifier_call_chain+0x2e/0x70

which lock already depends on the new lock.
[..]

Chain exists of:
  hmem_resource_lock --> mem_hotplug_lock --> (node_chain).rwsem

 Possible unsafe locking scenario:
       CPU0                    CPU1
       ----                    ----
  rlock((node_chain).rwsem);
                               lock(mem_hotplug_lock);
                               lock((node_chain).rwsem);

lock(hmem_resource_lock);

The lock ordering can cause potential deadlock. There are instances where hmem_resource_lock is taken after (node_chain).rwsem, and vice versa.

Remove registering of target devices from the hmat_callback(). By the time the hmat hotplug notifier is being called, there should not be hmem targets that still need to be registered.

Fixes cf8741ac57ed ("ACPI: NUMA: HMAT: Register "soft reserved" memory as an "hmem" device")

---
v2:
- Drop target registering in hmat_callback instead. (Dan)

---
 drivers/acpi/numa/hmat.c | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)
drivers/acpi/numa/hmat.c +2 -1
index 5a36d57289b4..5084ae1688f6 100644
@@ -874,7 +874,8 @@ static void hmat_register_target_devices(struct memory_target *target)
 	}
 }
 
-static void hmat_register_target(struct memory_target *target)
+static void hmat_register_target(struct memory_target *target,
+				 bool register_devices)
 {
 	int nid = pxm_to_node(target->memory_pxm);
 

drivers/acpi/numa/hmat.c static void hmat_register_target(struct memory_target *target) · line 887

Quoted 6 lines
drivers/acpi/numa/hmat.c +2 -1
@@ -882,7 +883,8 @@ static void hmat_register_target(struct memory_target *target)
 	 * Devices may belong to either an offline or online
 	 * node, so unconditionally add them.
 	 */
-	hmat_register_target_devices(target);
+	if (register_devices)
+		hmat_register_target_devices(target);

Why a new flag to pass around and not something like:

drivers/acpi/numa/hmat.c +3 -0
index 5a36d57289b4..9f9f09480765 100644
@@ -867,6 +867,9 @@ static void hmat_register_target_devices(struct memory_target *target)
        if (!IS_ENABLED(CONFIG_DEV_DAX_HMEM))
                return;
 
+       if (target->registered)
+               return;
+
        for (res = target->memregions.child; res; res = res->sibling) {
                int target_nid = pxm_to_node(target->memory_pxm);
 

...?

author maintainer
#2 lore raw
Quoted 61 lines · On 10/15/25 1:19 PM, [email protected] wrote:
Quoted 60 lines · Dave Jiang

The following lockdep splat was observed while kernel auto-online a CXL memory region:

======================================================
WARNING: possible circular locking dependency detected
6.17.0djtest+ #53 Tainted: G W
------------------------------------------------------
systemd-udevd/3334 is trying to acquire lock:
ffffffff90346188 (hmem_resource_lock){+.+.}-{4:4}, at: hmem_register_resource+0x31/0x50

but task is already holding lock:
ffffffff90338890 ((node_chain).rwsem){++++}-{4:4}, at: blocking_notifier_call_chain+0x2e/0x70

which lock already depends on the new lock.
[..]

Chain exists of:
  hmem_resource_lock --> mem_hotplug_lock --> (node_chain).rwsem

 Possible unsafe locking scenario:
       CPU0                    CPU1
       ----                    ----
  rlock((node_chain).rwsem);
                               lock(mem_hotplug_lock);
                               lock((node_chain).rwsem);

lock(hmem_resource_lock);

The lock ordering can cause potential deadlock. There are instances where hmem_resource_lock is taken after (node_chain).rwsem, and vice versa.

Remove registering of target devices from the hmat_callback(). By the time the hmat hotplug notifier is being called, there should not be hmem targets that still need to be registered.

Fixes cf8741ac57ed ("ACPI: NUMA: HMAT: Register "soft reserved" memory as an "hmem" device")

---
v2:
- Drop target registering in hmat_callback instead. (Dan)

---
 drivers/acpi/numa/hmat.c | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)
drivers/acpi/numa/hmat.c +2 -1
index 5a36d57289b4..5084ae1688f6 100644
@@ -874,7 +874,8 @@ static void hmat_register_target_devices(struct memory_target *target)
 	}
 }
 
-static void hmat_register_target(struct memory_target *target)
+static void hmat_register_target(struct memory_target *target,
+				 bool register_devices)
 {
 	int nid = pxm_to_node(target->memory_pxm);
 
Quoted 9 lines

drivers/acpi/numa/hmat.c static void hmat_register_target(struct memory_target *target) · line 887

Quoted 7 lines
drivers/acpi/numa/hmat.c +2 -1
@@ -882,7 +883,8 @@ static void hmat_register_target(struct memory_target *target)
 	 * Devices may belong to either an offline or online
 	 * node, so unconditionally add them.
 	 */
-	hmat_register_target_devices(target);
+	if (register_devices)
+		hmat_register_target_devices(target);

Why a new flag to pass around and not something like:

That works. I was tunnel visioned in trying to not mess up the current code flow.

DJ>

Quoted 16 lines
drivers/acpi/numa/hmat.c +3 -0
index 5a36d57289b4..9f9f09480765 100644
@@ -867,6 +867,9 @@ static void hmat_register_target_devices(struct memory_target *target)
        if (!IS_ENABLED(CONFIG_DEV_DAX_HMEM))
                return;
 
+       if (target->registered)
+               return;
+
        for (res = target->memregions.child; res; res = res->sibling) {
                int target_nid = pxm_to_node(target->memory_pxm);
 

...?

author maintainer
#3 lore raw
In reply to [email protected]
Quoted 76 lines · On 10/15/25 1:19 PM, [email protected] wrote:
Quoted 61 lines · Dave Jiang

The following lockdep splat was observed while kernel auto-online a CXL memory region:

======================================================
WARNING: possible circular locking dependency detected
6.17.0djtest+ #53 Tainted: G W
------------------------------------------------------
systemd-udevd/3334 is trying to acquire lock:
ffffffff90346188 (hmem_resource_lock){+.+.}-{4:4}, at: hmem_register_resource+0x31/0x50

but task is already holding lock:
ffffffff90338890 ((node_chain).rwsem){++++}-{4:4}, at: blocking_notifier_call_chain+0x2e/0x70

which lock already depends on the new lock.
[..]

Chain exists of:
  hmem_resource_lock --> mem_hotplug_lock --> (node_chain).rwsem

 Possible unsafe locking scenario:
       CPU0                    CPU1
       ----                    ----
  rlock((node_chain).rwsem);
                               lock(mem_hotplug_lock);
                               lock((node_chain).rwsem);

lock(hmem_resource_lock);

The lock ordering can cause potential deadlock. There are instances where hmem_resource_lock is taken after (node_chain).rwsem, and vice versa.

Remove registering of target devices from the hmat_callback(). By the time the hmat hotplug notifier is being called, there should not be hmem targets that still need to be registered.

Fixes cf8741ac57ed ("ACPI: NUMA: HMAT: Register "soft reserved" memory as an "hmem" device")

---
v2:
- Drop target registering in hmat_callback instead. (Dan)

---
 drivers/acpi/numa/hmat.c | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)
drivers/acpi/numa/hmat.c +2 -1
index 5a36d57289b4..5084ae1688f6 100644
@@ -874,7 +874,8 @@ static void hmat_register_target_devices(struct memory_target *target)
 	}
 }
 
-static void hmat_register_target(struct memory_target *target)
+static void hmat_register_target(struct memory_target *target,
+				 bool register_devices)
 {
 	int nid = pxm_to_node(target->memory_pxm);
 

drivers/acpi/numa/hmat.c static void hmat_register_target(struct memory_target *target) · line 887

Quoted 6 lines
drivers/acpi/numa/hmat.c +2 -1
@@ -882,7 +883,8 @@ static void hmat_register_target(struct memory_target *target)
 	 * Devices may belong to either an offline or online
 	 * node, so unconditionally add them.
 	 */
-	hmat_register_target_devices(target);
+	if (register_devices)
+		hmat_register_target_devices(target);

Why a new flag to pass around and not something like:

drivers/acpi/numa/hmat.c +0 -0
index 5a36d57289b4..9f9f09480765 100644

drivers/acpi/numa/hmat.c static void hmat_register_target_devices(struct memory_target *target) · line 872

Quoted 6 lines
drivers/acpi/numa/hmat.c +3 -0
@@ -867,6 +867,9 @@ static void hmat_register_target_devices(struct memory_target *target)
        if (!IS_ENABLED(CONFIG_DEV_DAX_HMEM))
                return;
 
+       if (target->registered)
+               return;
+

So this still triggers the lockdep warning. I don't think it's smart enough to know that it gets around the issue. My changes with a new flag does not trigger the lockdep.

DJ

Quoted 5 lines
drivers/acpi/numa/hmat.c +0 -0
        for (res = target->memregions.child; res; res = res->sibling) {
                int target_nid = pxm_to_node(target->memory_pxm);
 

...?

#4 lore raw

Dave Jiang wrote:
[..]

Quoted 8 lines
Quoted 8 lines
drivers/acpi/numa/hmat.c +0 -0
index 5a36d57289b4..9f9f09480765 100644
@@ -867,6 +867,9 @@ static void hmat_register_target_devices(struct memory_target *target)
        if (!IS_ENABLED(CONFIG_DEV_DAX_HMEM))
                return;
 
Quoted 7 lines

drivers/acpi/numa/hmat.c static void hmat_register_target_devices(struct memory_target *target) · line 872

Quoted 3 lines
drivers/acpi/numa/hmat.c +3 -0
+       if (target->registered)
+               return;
+

So this still triggers the lockdep warning. I don't think it's smart enough to know that it gets around the issue. My changes with a new flag does not trigger the lockdep.

You have a case where target->registered is false in the hmat_callback() path? How does that happen?

...and the splat is the same hmem_resource lock entanglement?

author maintainer
#5 lore raw
Quoted 18 lines · On 10/16/25 12:06 PM, [email protected] wrote:

Dave Jiang wrote:
[..]

Quoted 8 lines
Quoted 8 lines
drivers/acpi/numa/hmat.c +0 -0
index 5a36d57289b4..9f9f09480765 100644
@@ -867,6 +867,9 @@ static void hmat_register_target_devices(struct memory_target *target)
        if (!IS_ENABLED(CONFIG_DEV_DAX_HMEM))
                return;
 
Quoted 7 lines

drivers/acpi/numa/hmat.c static void hmat_register_target_devices(struct memory_target *target) · line 872

Quoted 3 lines
drivers/acpi/numa/hmat.c +3 -0
+       if (target->registered)
+               return;
+

So this still triggers the lockdep warning. I don't think it's smart enough to know that it gets around the issue. My changes with a new flag does not trigger the lockdep.

Quoted 2 lines

You have a case where target->registered is false in the hmat_callback() path? How does that happen?

It seems the ELC nodes are not on CPU node and also not generic port nodes. This platform has socket 0 and 1 where the memory targets are registered. However, node 2 and 3 gets hot-plugged and does not target registered at hmat_callback() time. >

Quoted 1 line

...and the splat is the same hmem_resource lock entanglement?

yes

Yes.